text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
null | null |
null | null |
# Introduction and Statement of Results {#s1}
The smallest parts function \({\mathrm{spt}}(n)\) of Andrews is defined for any integer \(n\geq1\) as the number of smallest parts among the integer partitions of size \(n\). For example, the partitions of \(n=4\) are (with the smallest parts underlined): \[\begin{aligned}
&\underline{4},\\ &3+\underline{1},\\ &\underline{2}+\underline{2},\\ &2+\underline{1}+\underline{1},\\ &\underline{1}+\underline{1}+\underline{1}+\underline{1}, \end{aligned}\] and so \({\mathrm{spt}}(4)=10\). The spt-function has many remarkable properties. For example, Andrews proved the following analogues of the well-known Ramanujan congruences for the partition function \(p(n)\): \[\begin{aligned}
{\mathrm{spt}}(5n+4)&\equiv0\pmod{5},\\ {\mathrm{spt}}(7n+5)&\equiv0\pmod{7},\\ {\mathrm{spt}}(13n+6)&\equiv0\pmod{13}. \end{aligned}\] One can compute \({\mathrm{spt}}(n)\) by making use of the generating function \[\sum_{n=1}^{\infty}{\mathrm{spt}}(n)q^n=\sum_{n=1}^\infty\frac{q^n}{\left(1-q^n\right)^2\left(q^{n+1};q\right)_\infty},\] where \[\begin{aligned}
(a;q)_\infty=\prod_{n=0}^\infty\left(1-aq^n\right). \end{aligned}\] We use this generating function to compute the values of \({\mathrm{spt}}(n)\) required for this paper. In this paper, we will prove the following asymptotic formula for \({\mathrm{spt}}(n)\) with an effective bound on the error term.
Our interest in proving the effective error bounds of Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} was motivated in part by the following recent conjectures of Chen concerning inequalities which involve \(p(n)\) and \({\mathrm{spt}}(n)\).
**Remark**. Conjectures ([\[1\]](#1){reference-type="ref" reference="1"}) and ([\[2\]](#2){reference-type="ref" reference="2"}) are slight modifications of Chen's original claims. By combining Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} with classical work of Lehmer which gives effective error bounds for \(p(n)\), we will prove the following result.
We will also use Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} to prove the following more precise version of Theorem [\[main\]](#main){reference-type="ref" reference="main"} regarding Conjecture ([\[1\]](#1){reference-type="ref" reference="1"}).
**Remark**. The constant \(N(\epsilon)\) of Theorem [\[better\]](#better){reference-type="ref" reference="better"} can be computed in practice. For example, by letting \(\epsilon = 1-\sqrt{6}/\pi\) in Theorem [\[better\]](#better){reference-type="ref" reference="better"}, we get Theorem [\[main\]](#main){reference-type="ref" reference="main"} (1) for \(n \geq N(1-\sqrt{6}/\pi)\) with \(N(1-\sqrt{6}/\pi)=5310\). We then use a computer to verify Theorem [\[main\]](#main){reference-type="ref" reference="main"} (1) in the exceptional range \(5 \leq n < 5310\). **Remark**. In analogy with the Hardy-Ramanujan asymptotic for \(p(n)\), Bringmann used the circle method to establish the asymptotic \[{\mathrm{spt}}(n) \sim \frac{1}{\pi\sqrt{8n}} e^{\pi\sqrt{\frac{2n}{3}}}\] as \(n\rightarrow\infty\). Bringmann's asymptotic for \({\mathrm{spt}}(n)\) implies that \[\begin{aligned}
\label{sptpn} {\mathrm{spt}}(n) \sim \frac{\sqrt{6}}{\pi} \sqrt{n} p(n) \end{aligned}\] as \(n \rightarrow \infty\). Theorem [\[better\]](#better){reference-type="ref" reference="better"} refines the asymptotic ([\[sptpn\]](#sptpn){reference-type="ref" reference="sptpn"}). We now describe our approach to Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"}. In particular, we explain some of the difficulties involved in proving effective error bounds for \({\mathrm{spt}}(n)\). In, Rademacher established an exact formula for \(p(n)\) as the absolutely convergent infinite sum \[\begin{aligned}
\label{rademacher} p(n)=\frac{2\pi}{(24n-1)^{3/4}}\sum_{c=1}^\infty\frac{A_c(n)}{c}I_{3/2}\left(\frac{\pi\sqrt{24n-1}}{6c}\right), \end{aligned}\] where \(I_{\nu}\) is the \(I\)-Bessel function, \(A_c(n)\) is the Kloosterman-type sum \[A_c(n):=\sum_{\substack{d \ensuremath{\negthickspace \negthickspace \negthickspace \pmod} c\\[1pt](d,c)=1}}e^{\pi is(d,c)}e^{-\frac{2\pi idn}{c}},\] and \(s(d,c)\) is the classical Dedekind sum \[\begin{aligned}
s(d,c):=\sum_{r=1}^{c-1}\frac{r}{c}\left(\frac{dr}{c}-\left\lfloor\frac{dr}{c}\right\rfloor-\frac{1}{2}\right). \end{aligned}\] Recently, Ahlgren and Andersen gave a Rademacher-type exact formula for \({\mathrm{spt}}(n)\) as the conditionally convergent infinite sum \[\begin{aligned}
\label{AA} {\mathrm{spt}}(n)=\frac{\pi}{6}(24n-1)^\frac{1}{4}\sum_{c=1}^\infty\frac{A_c(n)}{c}\left(I_{1/2}-I_{3/2}\right)\left(\frac{\pi\sqrt{24n-1}}{6c}\right). \end{aligned}\] In order to give an effective bound on the error term for \(p(n)\), Lehmer truncated the absolutely convergent sum ([\[rademacher\]](#rademacher){reference-type="ref" reference="rademacher"}) and applied bounds for the Kloosterman-type sum \(A_c(n)\). On the other hand, since the formula ([\[AA\]](#AA){reference-type="ref" reference="AA"}) is only conditionally convergent, bounding \({\mathrm{spt}}(n)\) is a much more delicate matter. In fact, to resolve the difficult problem of proving that ([\[AA\]](#AA){reference-type="ref" reference="AA"}) converges, Ahlgren and Andersen used advanced methods from the spectral theory of automorphic forms. To give an effective bound on the error term for \({\mathrm{spt}}(n)\), we will instead use different types of formulas for \(p(n)\) and \({\mathrm{spt}}(n)\) which express these functions as traces of singular moduli. To state these formulas, consider the weight \(-2\) weakly holomorphic modular form for \(\Gamma_0(6)\) defined by \[\begin{aligned}
g(z):=\frac{1}{2}\frac{E_2(z)-2E_2(2z)-3E_2(3z)+6E_2(6z)}{\big(\eta(z)\eta(2z)\eta(3z)\eta(6z)\big)^2}, \quad z=x+iy \in \mathbb{H}. \end{aligned}\] By applying the Maass weight-raising operator to \(g(z)\), one gets the following weight zero weak Maass form for \(\Gamma_0(6)\), \[\begin{aligned}
P(z):=-\left(\frac{1}{2\pi i} \frac{d}{dz} + \frac{1}{2\pi y}\right)g(z). \end{aligned}\] Bruinier and Ono proved the following formula for \(p(n)\).
Similarly, consider the weight zero weakly holomorphic modular form for \(\Gamma_0(6)\) defined by \[\begin{aligned}
\label{fdefinition} f(z):=\frac{1}{24}\frac{E_4(z)-4E_4(2z)-9E_4(3z)+36E_4(6z)}{\big(\eta(z)\eta(2z)\eta(3z)\eta(6z)\big)^2}. \end{aligned}\] Ahlgren and Andersen proved the following analogue of [\[boformula\]](#boformula){reference-type="eqref" reference="boformula"} for \({\mathrm{spt}}(n)\).
Identities which express Fourier coefficients of weak Maass forms as traces of singular moduli have been used in many contexts to give strong asymptotic formulas. For example, Bringmann and Ono expressed \(p(n)\) as a twisted trace of singular moduli by arithmetically reformulating Rademacher's exact formula ([\[rademacher\]](#rademacher){reference-type="ref" reference="rademacher"}) for \(p(n)\). Folsom and the second author then combined the Bringmann-Ono formula with spectral methods and subconvexity bounds for quadratic twists of modular \(L\)--functions to give an asymptotic formula for \(p(n)\) with a power-saving error term. In particular, by calculating the main term in this asymptotic formula in terms of the truncated main term in Rademacher's exact formula for \(p(n)\), these authors improved the exponent in Lehmer's bound. This exponent was further improved by Ahlgren and Andersen. In the works, spectral methods and subconvexity bounds were again used to give asymptotic formulas with power-saving error terms for twisted traces of singular moduli. These results were applied to study a variety of arithmetic problems, including the distribution of \(p(n)\) and \({\mathrm{spt}}(n)\), and the distribution of partition ranks. Note that although the constants in the error terms of these results are effective, it would be very difficult to actually give explicit numerical values for these constants because of the techniques involved in the proofs. However, there is an alternative approach which we now describe. Using ([\[boformula\]](#boformula){reference-type="ref" reference="boformula"}), the formula ([\[AAspt\]](#AAspt){reference-type="ref" reference="AAspt"}) can be written as \[\begin{aligned}
\label{reorganized} {\mathrm{spt}}(n)=\frac{1}{12}S(n)-\frac{24n-1}{12}p(n), \end{aligned}\] where \(S(n)\) is the trace of singular moduli for \(f(z)\) given by \[\begin{aligned}
S(n):=\sum_{[Q]}f(\tau_Q). \end{aligned}\] By applying Lehmer's effective error bounds for \(p(n)\) in ([\[reorganized\]](#reorganized){reference-type="ref" reference="reorganized"}), we will reduce the proof of Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} to the following asymptotic formula for the trace \(S(n)\) with an effective bound on the error term.
Our proof of Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"} is inspired by work of Dewar and Murty, who used the formula ([\[boformula\]](#boformula){reference-type="ref" reference="boformula"}) to derive the Hardy-Ramanujan asymptotic formula for \(p(n)\) without using the circle method. In order to give effective bounds, additional care must be taken. For instance, we must give effective bounds for the Fourier coefficients of \(f(z)\). **Organization**. The paper is organized as follows. In Section [2](#heegner){reference-type="ref" reference="heegner"}, we review some facts regarding quadratic forms and Heegner points. In Section [3](#trace){reference-type="ref" reference="trace"}, we prove Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"}. In Section [4](#s2){reference-type="ref" reference="s2"}, we prove Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"}. In Section [5](#s3){reference-type="ref" reference="s3"}, we prove Theorem [\[better\]](#better){reference-type="ref" reference="better"}. Finally, in Section [6](#s4){reference-type="ref" reference="s4"}, we prove the remaining conjectures. **Acknowledgments**. We thank Adrian Barquero-Sanchez, Sheng-Chi Liu, Karl Mahlburg, Ken Ono, Wei-Lun Tsai, and Matt Young for very helpful discussions regarding this work, and Michael Griffin and Lea Beneish for help computing values of \({\mathrm{spt}}(n)\). We also thank the referees for many detailed comments and corrections, leading to simplifications of some arguments, sharper estimates, and an improved exposition.
# Quadratic forms and Heegner points {#heegner}
Let \(N \geq 1\) be a positive integer and \(D < 0\) be a negative discriminant coprime to \(N\). Let \(\mathcal{Q}_{D,N}\) be the set of positive definite, integral binary quadratic forms \[Q(X,Y)=[a_Q,b_Q,c_Q](X,Y)=a_Q X^2+b_Q XY+c_Q Y^2\] of discriminant \(b_Q ^2-4a_Q c_Q= D < 0\) with \(a_Q \equiv 0 \pmod N\). There is a (right) action of \(\Gamma_0(N)\) on \(\mathcal{Q}_{D,N}\) defined by \[\begin{aligned}
Q=\left[a_Q, b_Q, c_Q\right] \mapsto Q \circ \sigma = \left[a_Q^{\sigma}, b_Q^{\sigma}, c_Q^{\sigma}\right], \end{aligned}\] where for \(\sigma = \begin{pmatrix} \alpha & \beta \\ \gamma & \delta \end{pmatrix} \in \Gamma_0(N)\) we have \[\begin{aligned}
a_Q^{\sigma}&= a_Q \alpha^2 + b_Q \alpha \gamma + c_Q \gamma^2 ,\\ b_Q^{\sigma}&= 2 a_Q \alpha \beta + b_Q(\alpha \delta + \beta \gamma) + 2 c_Q \gamma \delta ,\\ c_Q^{\sigma}&= a_Q \beta^2 + b_Q \beta \delta +c_Q \delta^2. \end{aligned}\] Given a solution \(r \pmod {2N}\) of \(r^2 \equiv D \pmod {4N}\), we define the subset of forms \[\begin{aligned}
\mathcal{Q}_{D,N,r}:=\left\{Q=\left[a_Q,b_Q,c_Q\right] \in \mathcal{Q}_{D,N}:~b_Q \equiv r \ensuremath{\negthickspace \negthickspace \negthickspace \pmod} {2N}\right\}. \end{aligned}\] Then the group \(\Gamma_0(N)\) also acts on \(\mathcal{Q}_{D,N,r}\). The number of \(\Gamma_0(N)\) equivalence classes in \(\mathcal{Q}_{D,N,r}\) is given by the Hurwitz-Kronecker class number \(H(D)\). The preceding facts remain true if we restrict to the subset \(\mathcal{Q}_{D,N}^{\textrm{prim}}\) of primitive forms in \(\mathcal{Q}_{D,N}\); i.e., those forms with \[(a_Q, b_Q, c_Q)=1.\] In this case, the number of \(\Gamma_0(N)\) equivalence classes in \(\mathcal{Q}_{D,N,r}^{\textrm{prim}}\) is given by the class number \(h(D)\). To each form \(Q \in \mathcal{Q}_{D,N}\), we associate a Heegner point \(\tau_Q\) which is the root of \(Q\left(X,1\right)\) given by \[\begin{aligned}
\tau_Q=\frac{-b_Q+\sqrt{D}}{2a_Q} \in \mathbb{H}. \end{aligned}\] The Heegner points \(\tau_Q\) are compatible with the action of \(\Gamma_0(N)\) in the sense that if \(\sigma \in \Gamma_0(N)\), then \[\begin{aligned}
\label{compatible} \sigma (\tau_{Q}) = \tau_{Q \circ \sigma^{-1}}. \end{aligned}\]
# Proof of Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"} {#trace}
In this section, we prove Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"}, which gives an asymptotic formula with an effective bound on the error term for the trace of the weight zero weakly holomorphic modular form \(f(z)\) for \(\Gamma_0(6)\) defined by ([\[fdefinition\]](#fdefinition){reference-type="ref" reference="fdefinition"}). This will be used crucially in the proof of Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"}. Let \(D_n:=-24n+1\) for \(n \in \Z^{+}\) and define the trace of \(f(z)\) by \[\begin{aligned}
S(n):=\sum_{[Q] \in \mathcal{Q}_{D_n, 6, 1}/\Gamma_0(6)}f(\tau_Q). \end{aligned}\] First, we decompose \(S(n)\) as a linear combination of traces involving primitive forms. Let \(\Delta < 0\) be any discriminant with \(\Delta \equiv 1 \pmod {24}\) and define the class polynomials \[\begin{aligned}
H_n(X):=\prod_{[Q] \in \mathcal{Q}_{D_n, 6, 1}/\Gamma_0(6)}(X-f(\tau_Q)) \end{aligned}\] and \[\begin{aligned}
\widehat{H}_{\Delta}(X):=\prod_{[Q] \in \mathcal{Q}_{\Delta, 6, 1}^{\textrm{prim}}/\Gamma_0(6)}(X-f(\tau_Q)). \end{aligned}\] Let \(\{W_{\ell}\}_{\ell|6}\) be the group of Atkin-Lehner involutions for \(\Gamma_0(6)\). Since \[\begin{aligned}
\label{ALid} f|_0W_{\ell}=\lambda_{\ell} f \end{aligned}\] with \(\lambda_{\ell}=1\) for \(\ell=1,6\) and \(\lambda_{\ell}=-1\) for \(\ell=2,3\), then arguing exactly as in the proof of, we get the identity \[\begin{aligned}
\label{relation} H_n(X)=\prod_{\substack{u > 0 \\ u^2 | D_n}}\varepsilon(u)^{h(D_{n}/u^2)}\widehat{H}_{D_{n}/u^2}(\varepsilon(u)X), \end{aligned}\] where \(\varepsilon(u)=1\) if \(u \equiv \pm 1 \pmod{12}\) and \(\varepsilon(u)=-1\) otherwise. Comparing terms on both sides of ([\[relation\]](#relation){reference-type="ref" reference="relation"}) yields the class number relation \[\begin{aligned}
H(D_n)=\sum_{\substack{u > 0 \\ u^2 | D_n}}h(D_{n}/u^2) \end{aligned}\] and the decomposition \[\begin{aligned}
\label{Sdecomp} S(n)=\sum_{\substack{u > 0 \\ u^2 | D_n}}\varepsilon(u)S_u(n), \end{aligned}\] where \[\begin{aligned}
S_u(n):=\sum_{[Q] \in \mathcal{Q}_{D_n/u^2, 6, 1}^{\textrm{prim}}/\Gamma_0(6)}f(\tau_Q). \end{aligned}\] Next, following we express \(S_u(n)\) as a trace involving primitive forms of level 1. The group \(\Gamma_0(6)\) has index 12 in \(SL_2(\Z)\). We choose the following 12 right coset representatives: \[\begin{aligned}
\gamma_{\infty}&:=\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix},\\ \gamma_{1/3,r}&:=\begin{pmatrix} 1 & 0 \\ 3 & 1 \end{pmatrix} \hspace{-0.05in} \begin{pmatrix} 1 & r \\ 0 & 1 \end{pmatrix}, \quad r=0,1; \\ \gamma_{1/2,s}&:=\begin{pmatrix} 1 & 1 \\ 2 & 3 \end{pmatrix} \hspace{-0.05in} \begin{pmatrix} 1 & s \\ 0 & 1 \end{pmatrix}, \quad s=0,1,2; \\ \gamma_{0,t}&:=\begin{pmatrix} 0 &-1 \\ 1 & 0 \end{pmatrix} \hspace{-0.05in} \begin{pmatrix} 1 & t \\ 0 & 1 \end{pmatrix}, \quad t=0,1,2, 3, 4, 5. \end{aligned}\] We denote this set of coset representatives by \(\mathbf{C}_6\). Each matrix \(\gamma \in \mathbf{C}_6\) maps the cusp \(\infty\) to one of the four cusps \(\{\infty, 1/3, 1/2, 0\}\) of the modular curve \(X_0(6)\), which have widths 1, 2, 3, and 6, respectively. In particular, we have \(\gamma_{\infty}(\infty)=\infty\), \(\gamma_{1/3,r}(\infty)=1/3\), \(\gamma_{1/2, s}(\infty)=1/2\), and \(\gamma_{0,t}(\infty)=0\). Recall that a form \(Q=[a_Q, b_Q, c_Q] \in \mathcal{Q}_{\Delta,1}\) is reduced if \[\begin{aligned}
|b_Q| \leq a_Q \leq c_Q, \end{aligned}\] and if either \(|b_Q|=a_Q\) or \(a_Q=c_Q\), then \(b_Q \geq 0\). Let \(\mathcal{Q}_{\Delta}\) denote a set of primitive, reduced forms representing the equivalence classes in \(\mathcal{Q}_{\Delta, 1}^{\textrm{prim}}/SL_2(\Z)\). For each \(Q \in \mathcal{Q}_{\Delta}\), there is a unique choice of coset representative \(\gamma_Q \in \mathbf{C}_6\) such that \[\begin{aligned}
\in \mathcal{Q}_{\Delta, 6, 1}^{\textrm{prim}}/\Gamma_0(6). \end{aligned}\] This induces a bijection \[\begin{aligned}
\label{bijection} \mathcal{Q}_{\Delta} & \longrightarrow \mathcal{Q}_{\Delta, 6, 1}^{\textrm{prim}}/\Gamma_0(6)\\ Q &\longmapsto [Q \circ \gamma_Q^{-1}];\notag \end{aligned}\] see the Proposition on page 505 in, or more concretely,, where an explicit list of the matrices \(\gamma_Q \in \mathbf{C}_6\) is given. Using the bijection ([\[bijection\]](#bijection){reference-type="ref" reference="bijection"}) and the compatibility relation ([\[compatible\]](#compatible){reference-type="ref" reference="compatible"}) for Heegner points, the trace \(S_u(n)\) can be expressed as \[\begin{aligned}
\label{newtrace} S_u(n)=\sum_{[Q] \in \mathcal{Q}_{D_n/u^2, 6, 1}^{\textrm{prim}}/\Gamma_0(6)}f(\tau_Q)=\sum_{Q \in \mathcal{Q}_{D_{n}/u^2}}f\left(\gamma_Q(\tau_Q)\right). \end{aligned}\] Therefore, to study the asymptotic distribution of \(S_u(n)\), we need the Fourier expansion of \(f(z)\) with respect to the matrices \(\gamma_{\infty}, \gamma_{1/3,r}, \gamma_{1/2,s}\), and \(\gamma_{0,t}\). In, Ahlgren and Andersen compute the Fourier expansion of \(f(z)\) at the cusp \(\infty\). The basic idea is as follows. The weakly holomorphic modular form \(f(z)\) has a Fourier expansion of the form \[\begin{aligned}
f(z)=e(-z) + b(0) + \sum_{m=1}^{\infty}b(m)e(mz), \quad e(z):=e^{2\pi i z} \end{aligned}\] for some integers \(b(m)\) for \(m \geq 0\). One can construct a weight zero weak Maass form \(f(z,s)\) for \(\Gamma_0(6)\) with eigenvalue \(s(1-s)\) whose analytic continuation at \(s=1\) is a harmonic function on \(\mathbb{H}\) with the Fourier expansion \[\begin{aligned}
f(z,1)=e(-z) + a(0) + \sum_{m=1}^{\infty}\frac{a(m)}{\sqrt{m}}e(mz)-e(-\overline{z}) + \sum_{m=1}^{\infty}\frac{a(-m)}{\sqrt{m}} e(-m\overline{z}), \end{aligned}\] where \[\begin{aligned}
a(0)& =4\pi^2\sum_{\ell|6}\frac{\mu(\ell)}{\ell}\sum_{\substack{0 < c \equiv 0 \mod (6/\ell) \\ (c,\ell)=1}} \frac{S\left(-\overline{\ell}, 0;c\right)}{c^{2}},\\ \vspace{0.05in} a(m)&=2\pi \sum_{\ell|6}\frac{\mu(\ell)}{\sqrt{\ell}}\sum_{\substack{0 < c \equiv 0 \mod (6/\ell) \\ (c,\ell)=1}} \frac{S\left(-\overline{\ell}, m ;c\right)}{c} I_1\left(\frac{4\pi\sqrt{m}}{\sqrt{\ell}c}\right), \quad m \geq 1 \\ \vspace{0.05in} a(-m)&=2\pi \sum_{\ell|6}\frac{\mu(\ell)}{\sqrt{\ell}}\sum_{\substack{0 < c \equiv 0 \mod (6/\ell) \\ (c,\ell)=1}} \frac{S\left(-\overline{\ell},-m ;c\right)}{c} J_1\left(\frac{4\pi\sqrt{m}}{\sqrt{\ell}c}\right), \quad m \geq 1. \end{aligned}\] Here \(\mu(\ell)\) is the Möbius function, \(S(a,b;c)\) is the Kloosterman sum \[\begin{aligned}
S(a,b;c):=\sum_{\substack{d \ensuremath{\negthickspace \negthickspace \negthickspace \pmod} c \\ (c,d)=1}}e\left(\frac{a\overline{d} + bd}{c}\right), \end{aligned}\] and \(I_1, J_1\) are the Bessel functions of order 1 (note that \(\overline{d}\) is the multiplicative inverse of \(d \pmod c\)). From these Fourier expansions, one can see that the functions \(f(z)\) and \(f(z,1)\) have the same principal parts in the cusps \(\{\infty, 1/3, 1/2, 0\}\), hence the function \(f(z)-f(z,1)\) is bounded on the compact Riemann surface \(X_0(6)\). Since a bounded harmonic function on a compact Riemann surface is constant, the function \(f(z)-f(z,1)\) is constant. Now, using the Fourier expansions of \(E_4(z)\) and \(\eta(z)\), we use `SageMath` to compute \[\begin{aligned}
f(z)=q^{-1} + 12 + 77 q + 376 q^2 + 1299q^3 + 4600q^4 + 12025q^5 + \cdots, \quad q:=e(z). \end{aligned}\] In particular, \(b(0)=12\). On the other hand, in Lemma [\[fourierbound\]](#fourierbound){reference-type="ref" reference="fourierbound"} we show by a direct calculation that \(a(0)=12\). Since \(f(z)-f(z,1)\) is constant, we have \[\begin{aligned}
f(z)-f(z,1)=b(0)-a(0)=12-12=0. \end{aligned}\] Finally, since \(f(z)=f(z,1)\), then by uniqueness of Fourier expansions we have \(b(m)=m^{-1/2}a(m)\) for \(m \geq 1\), \(a(-1)=1\), and \(a(-m)=0\) for \(m \geq 2\). We next use the Fourier expansion \[\begin{aligned}
f|_{0}\gamma_{\infty}(z) = e(-z) + 12 + \sum_{m=1}^{\infty}\frac{a(m)}{\sqrt{m}}e(mz) \end{aligned}\] to compute the Fourier expansion of \(f(z)\) with respect to the matrices \(\gamma_{\infty}, \gamma_{1/3,r}, \gamma_{1/2,s}\), and \(\gamma_{0,t}\). The Atkin-Lehner involutions for \(\Gamma_0(6)\) are given by \[\begin{aligned}
W_1= \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}, \quad W_2= \frac{1}{\sqrt{2}}\begin{pmatrix} 2 &-1 \\ 6 &-2 \end{pmatrix}, \quad W_3= \frac{1}{\sqrt{3}}\begin{pmatrix} 3 & 1 \\ 6 & 3 \end{pmatrix}, \quad W_6= \frac{1}{\sqrt{6}}\begin{pmatrix} 0 &-1 \\ 6 & 0 \end{pmatrix}. \end{aligned}\] For each \(\ell|6\) and \(v=6/\ell\), let \(V_{\ell} = \sqrt{\ell} W_{\ell}\) and \[\begin{aligned}
A_{\ell} = \begin{pmatrix} \frac{1}{\textrm{width of the cusp \(1/v\)}} & 0 \\ 0 & 1 \end{pmatrix}. \end{aligned}\] We have
Note that \(V_{\ell}A_{\ell} \in SL_2(\mathbb{Z})\) and \[\begin{aligned}
V_{\ell}A_{\ell}(\infty)=\frac{1}{v}. \end{aligned}\] Let \(\gamma \in SL_2(\Z)\) be any matrix such that \(\gamma(\infty)=1/v\). Then \[\begin{aligned}
\left(V_{\ell}A_{\ell}\right)^{-1}(\gamma(\infty)) = \infty, \end{aligned}\] so that \[\begin{aligned}
\left(V_{\ell}A_{\ell}\right)^{-1}\gamma \in \Gamma_{\infty}:=\left\{\begin{pmatrix} 1 & n \\ 0 & 1 \end{pmatrix}:~n \in \Z \right\} \end{aligned}\] where \(\Gamma_\infty\) is the stabilizer of the cusp \(\infty\). In particular, there is an integer \(n \in \mathbb{Z}\) such that \[\begin{aligned}
\gamma=V_{\ell}A_{\ell} \begin{pmatrix} 1 & n \\ 0 & 1 \end{pmatrix}. \end{aligned}\] By solving for \(n\) for each cusp, we have \[\begin{aligned}
\gamma_{\infty} = V_1A_1, \quad \gamma_{1/3,r} = V_2A_2 \begin{pmatrix} 1 & r+1 \\ 0 & 1 \end{pmatrix}, \quad \gamma_{1/2,s} & = V_3A_3 \begin{pmatrix} 1 & s \\ 0 & 1 \end{pmatrix}, \quad \gamma_{0,t}= V_6A_6 \begin{pmatrix} 1 & t \\ 0 & 1 \end{pmatrix}. \end{aligned}\] Now, by ([\[ALid\]](#ALid){reference-type="ref" reference="ALid"}) we have \(f(V_{\ell}z)=f(z)\) for \(\ell=1,6\) and \(f(V_{\ell}z)=-f(z)\) for \(\ell=2,3\). Hence \[\begin{aligned}
f|_{0}\gamma_{\infty} (z) & = f(z),\\ f|_{0}\gamma_{1/3,r} (z) & = f\left(V_2A_2\begin{pmatrix} 1 & r+1 \\ 0 & 1 \end{pmatrix} z\right) = f\left(V_2\left(\frac{z+r+1}{2}\right)\right) =-f\left(\frac{z+r+1}{2}\right),\\ f|_{0}\gamma_{1/2,s} (z) &= f\left(V_3A_3\begin{pmatrix} 1 & s \\ 0 & 1 \end{pmatrix} z\right) = f\left(V_3\left(\frac{z+s}{3}\right)\right) =-f\left(\frac{z+s}{3}\right),\\ f|_{0}\gamma_{0,t} (z) &= f\left(V_6A_6\begin{pmatrix} 1 & t \\ 0 & 1 \end{pmatrix} z\right) = f\left(V_6\left(\frac{z+t}{6}\right)\right) = f\left(\frac{z+t}{6}\right). \end{aligned}\] The Fourier expansion of \(f(z)\) with respect to the matrices \(\gamma_{1/3,r}, \gamma_{1/2,s}, \gamma_{0,t}\) can now be determined from the Fourier expansion at \(\infty\) using these identities. In particular, if \(\zeta_6:=e(1/6)\) is a primitive sixth root of unity, we have \[\begin{aligned}
f|_{0}\gamma_{1/3,r}(z) & = \zeta_6^{3r}e(-z/2)-12 + \sum_{m=1}^{\infty}\zeta_6^{3+3m(r+1)}\frac{a(m)}{\sqrt{m}}e(mz/2),\\ f|_{0}\gamma_{1/2,s}(z) & = \zeta_6^{3-2s}e(-z/3)-12 + \sum_{m=1}^{\infty}\zeta_6^{3+2ms}\frac{a(m)}{\sqrt{m}}e(mz/3),\\ f|_{0}\gamma_{0,t}(z) &= \zeta_6^{-t}e(-z/6) + 12 + \sum_{m=1}^{\infty}\zeta_6^{mt}\frac{a(m)}{\sqrt{m}}e(mz/6). \end{aligned}\] Given a form \(Q \in Q_\Delta\) and corresponding coset representative \(\gamma_{Q} \in \mathbf{C}_6\), let \(h_Q \in \{1,2,3,6\}\) be the width of the cusp \(\gamma_{Q}(\infty)\), and let \(\zeta_Q\) and \(\phi_{m,Q}\) be the sixth roots of unity defined as follows:
[\[table:rootU\]]{#table:rootU label="table:rootU"}
Then we can write \[\begin{aligned}
\label{formexpand} f|_{0}\gamma_{Q}(z) & = \zeta_Qe(-z/h_Q) + 12\mu(h_Q) + \sum_{m=1}^{\infty}\phi_{m,Q}\frac{a(m)}{\sqrt{m}}e(mz/h_Q). \end{aligned}\] In the following lemma we evaluate \(a(0)\) and give effective bounds for the Fourier coefficients \(a(m)\) for \(m \geq 1\).
We are now in position to prove Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"}, which we restate for the convenience of the reader.
# Proof of Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} {#s2}
In this section, we prove Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"}. We will require an asymptotic formula for \(p(n)\) with an effective bound on the error term due to Lehmer. For convenience, define \[{\lambda}(n):=\frac{\pi}{6}\sqrt{24n-1}.\] Inspired by the Hardy-Ramanujan asymptotic for \(p(n)\), Rademacher obtained the exact formula \[p(n)=\frac{2\pi}{(24n-1)^{3/4}}\sum_{c=1}^\infty\frac{A_c(n)}{c}I_{3/2}\left(\frac{{\lambda}(n)}{c}\right),\] where \(A_c(n)\) is the Kloosterman-type sum \[A_c(n):=\sum_{\substack{d \ensuremath{\negthickspace \negthickspace \negthickspace \pmod} c\\[1pt](d,c)=1}}e^{\pi is(d,c)}e^{-\frac{2\pi idn}{c}}\] and \(s(d,c)\) is the Dedekind sum \[s(d,c):=\sum_{r=1}^{c-1}\frac{r}{c}\left(\frac{dr}{c}-\left\lfloor\frac{dr}{c}\right\rfloor-\frac{1}{2}\right).\] Using Rademacher's formula, Lehmer proved the following result.
We first use Theorem [\[Lehmer\]](#Lehmer){reference-type="ref" reference="Lehmer"} to deduce the following effective bound.
## Proof of Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} {#proof-of-theorem-riad}
Using ([\[boformula\]](#boformula){reference-type="ref" reference="boformula"}), the formula ([\[AAspt\]](#AAspt){reference-type="ref" reference="AAspt"}) can be written as \[\begin{aligned}
\label{sptformula} \textrm{spt}(n)=\frac{1}{12}\left[S(n)-(24n-1)p(n)\right]. \end{aligned}\] Then using ([\[sptformula\]](#sptformula){reference-type="ref" reference="sptformula"}), Theorem [\[Sprop\]](#Sprop){reference-type="ref" reference="Sprop"}, and Lemma [\[plemma\]](#plemma){reference-type="ref" reference="plemma"}, a straightforward calculation yields \[\begin{aligned}
\textrm{spt}(n) &= \frac{\sqrt{3}}{\pi \sqrt{24n-1}}e^{{\lambda}(n)} + E_s(n), \end{aligned}\] where the error term \[\begin{aligned}
E_s(n) & := \frac{E(n)}{12}-\frac{24n-1}{12}E_p(n) \end{aligned}\] satisfies the bound \[\begin{aligned}
|E_s(n)| < (3.59 \times 10^{22})2^{q(n)}(24n-1)^2e^{{\lambda}(n)/2}. \end{aligned}\] 0◻ As pointed out by Bessenrodt and Ono, it is straightforward to obtain from Theorem [\[Lehmer\]](#Lehmer){reference-type="ref" reference="Lehmer"} that \[\frac{\sqrt{3}}{12n}\left(1-\frac{1}{\sqrt{n}}\right)e^{{\lambda}(n)}<p(n)<\frac{\sqrt{3}}{12n}\left(1+\frac{1}{\sqrt{n}}\right)e^{{\lambda}(n)}\] for all \(n \geq 1\). We will use Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} to prove the following analogous statement for \({\mathrm{spt}}(n)\), where \(\sqrt{n}\) is replaced by any positive integral power of \(n\).
# Proof of Theorem [\[better\]](#better){reference-type="ref" reference="better"} {#s3}
By Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} and Lemma [\[plemma\]](#plemma){reference-type="ref" reference="plemma"}, we may write \[\begin{aligned}
\label{eqn1} \textrm{spt}(n)&=\alpha(n)e^{{\lambda}(n)} + E_s(n) \end{aligned}\] and \[\begin{aligned}
\label{eqn2} p(n)&=\beta(n)e^{{\lambda}(n)} + E_p(n), \end{aligned}\] where \[\begin{aligned}
\alpha(n):=\frac{\sqrt{3}}{\pi \sqrt{24n-1}}, \quad \beta(n):=\frac{2\sqrt{3}}{24n-1}\left(1-\frac{6}{\pi \sqrt{24n-1}}\right). \end{aligned}\] Also, for \(\epsilon > 0\) we define \[\begin{aligned}
\gamma(n):=\frac{\sqrt{6}}{\pi}\sqrt{n}, \quad \gamma(n,\epsilon):=\left(\frac{\sqrt{6}}{\pi} + \epsilon\right)\sqrt{n}. \end{aligned}\] We must prove that there exists an effectively computable positive constant \(N(\epsilon) > 0\) such that for all \(n \geq N(\epsilon)\), we have \[\begin{aligned}
\label{chenbound2} \gamma(n) p(n) < \textrm{spt}(n) < \gamma(n, \epsilon)p(n). \end{aligned}\] First, using ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) and ([\[eqn2\]](#eqn2){reference-type="ref" reference="eqn2"}) we find that the lower bound in ([\[chenbound2\]](#chenbound2){reference-type="ref" reference="chenbound2"}) is equivalent to \[\begin{aligned}
\label{lower} c_1(n)e^{{\lambda}(n)} > \gamma(n)E_p(n)-E_s(n), \end{aligned}\] where \(c_1(n):=\alpha(n)-\beta(n)\gamma(n)\). Now, the error bounds in Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} and Lemma [\[plemma\]](#plemma){reference-type="ref" reference="plemma"} imply that \[\begin{aligned}
\left|\gamma(n)E_p(n)-E_s(n)\right| \leq c_2(n)e^{{\lambda}(n)/2}, \end{aligned}\] where \[\begin{aligned}
c_2(n):=(1313)\gamma(n) + (3.59 \times 10^{22})2^{q(n)}(24n-1)^2. \end{aligned}\] Then noting that \(c_1(n) > 0\) for all \(n \geq 1\), we find that ([\[lower\]](#lower){reference-type="ref" reference="lower"}) is implied by the bound \[\begin{aligned}
e^{{\lambda}(n)/2} > c_3(n):=\frac{c_2(n)}{c_1(n)}, \end{aligned}\] or equivalently, the bound \[\begin{aligned}
\label{lower2} n > \frac{1}{24}\left[\left(\frac{12}{\pi}\log(c_3(n))\right)^2 + 1\right]. \end{aligned}\] A calculation shows that ([\[lower2\]](#lower2){reference-type="ref" reference="lower2"}) holds for all \(n \geq N:=5310\). Similarly, using ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) and ([\[eqn2\]](#eqn2){reference-type="ref" reference="eqn2"}) we find that the upper bound in ([\[chenbound2\]](#chenbound2){reference-type="ref" reference="chenbound2"}) is equivalent to \[\begin{aligned}
\label{upper} c_4(n, \epsilon)e^{{\lambda}(n)} > E_s(n)-\gamma(n,\epsilon)E_p(n), \end{aligned}\] where \(c_4(n,\epsilon):=\beta(n)\gamma(n,\epsilon)-\alpha(n)\). The error bounds in Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"} and Lemma [\[plemma\]](#plemma){reference-type="ref" reference="plemma"} imply that \[\begin{aligned}
|E_s(n)-\gamma(n, \epsilon)E_p(n)| \leq c_5(n, \epsilon)e^{{\lambda}(n)/2}, \end{aligned}\] where \[\begin{aligned}
c_5(n, \epsilon):=(1313)\gamma(n, \epsilon) + (3.59 \times 10^{22})2^{q(n)}(24n-1)^2. \end{aligned}\] Moreover, there exists an effectively computable positive constant \(N_1(\epsilon) > 0\) such that \(c_4(n, \epsilon) > 0\) for all \(n \geq N_1(\epsilon)\). Then arguing as above, we find that if \(n \geq N_1(\epsilon)\), the bound ([\[upper\]](#upper){reference-type="ref" reference="upper"}) is implied by the bound \[\begin{aligned}
\label{upper2} n > \frac{1}{24}\left[\left(\frac{12}{\pi}\log(c_6(n, \epsilon))\right)^2 + 1\right], \end{aligned}\] where \(c_6(n,\epsilon):=c_5(n, \epsilon)/c_4(n,\epsilon)\). Clearly, there exists an effectively computable positive constant \(N_2(\epsilon) \geq N_1(\epsilon)\) such that ([\[upper2\]](#upper2){reference-type="ref" reference="upper2"}) holds for all \(n \geq N_2(\epsilon)\). Let \(N(\epsilon):=\max\{N, N_2(\epsilon)\}\). Then the inequalities ([\[chenbound2\]](#chenbound2){reference-type="ref" reference="chenbound2"}) hold for all \(n \geq N(\epsilon)\). 0◻
# Proof of Theorem [\[main\]](#main){reference-type="ref" reference="main"} {#s4}
## Proof of Conjecture ([\[1\]](#1){reference-type="ref" reference="1"})
Let \(\epsilon=1-\sqrt{6}/\pi\) in Theorem [\[better\]](#better){reference-type="ref" reference="better"}. We need to determine the constant \(N(1-\sqrt{6}/\pi)\). A calculation shows that the inequality \(c_4(n, 1-\sqrt{6}/\pi) > 0\) holds if \(n \geq N_1(1-\sqrt{6}/\pi)\) where \(N_1(1-\sqrt{6}/\pi)=4\). Next, we need to find the smallest positive integer \(N_2(1-\sqrt{6}/\pi) \geq 4\) such that the bound \[\begin{aligned}
n > \frac{1}{24}\left[\left(\frac{12}{\pi}\log(c_6(n, 1-\sqrt{6}/\pi))\right)^2 + 1\right] \end{aligned}\] holds for all \(n \geq N_2(1-\sqrt{6}/\pi)\). A calculation shows that this constant is given by \(N_2(1-\sqrt{6}/\pi)=4845\). We now have \[\begin{aligned}
N(1-\sqrt{6}/\pi):=\max\{N, N_2(1-\sqrt{6}/\pi)\}=\max \{5310, 4845\}=5310. \end{aligned}\] Therefore, the inequalities \[\begin{aligned}
\frac{\sqrt{6}}{\pi} \sqrt{n} p(n) < \mathrm{spt}(n) < \sqrt{n} p(n) \end{aligned}\] hold for all \(n \geq 5310\). Finally, one can verify with a computer that these inequalities also hold for \(5 \leq n < 5310\). 0◻
## Proof of Conjecture ([\[2\]](#2){reference-type="ref" reference="2"})
We follow closely the proof of. By taking \(\alpha=k=1\) in Theorem [\[bounds\]](#bounds){reference-type="ref" reference="bounds"} (recall that \(B_1(1)=5729\)), we find that \[\label{boundseq} \frac{\sqrt{3}}{\pi\sqrt{24n-1}}\left(1-\frac{1}{n}\right)e^{{\lambda}(n)}< {\mathrm{spt}}(n) <\frac{\sqrt{3}}{\pi\sqrt{24n-1}}\left(1+\frac{1}{n}\right)e^{{\lambda}(n)}\] holds for all \(n \geq 5729\). One can verify with a computer that [\[boundseq\]](#boundseq){reference-type="eqref" reference="boundseq"} also holds for \(1 \leq n < 5729\). Now, assume that \(1 < a \leq b\), and let \(b=Ca\) where \(C \geq 1\). From ([\[boundseq\]](#boundseq){reference-type="ref" reference="boundseq"}) we get the inequalities \[\begin{aligned}
{\mathrm{spt}}(a){\mathrm{spt}}(Ca) > \frac{3}{\pi^2 \sqrt{24a-1}\sqrt{24Ca-1}}\left(1-\frac{1}{a}\right)\left(1-\frac{1}{Ca}\right)e^{{\lambda}(a)+{\lambda}(Ca)} \end{aligned}\] and \[\begin{aligned}
{\mathrm{spt}}(a+Ca) < \frac{\sqrt{3}}{\pi\sqrt{24(a+Ca)-1}}\left(1+\frac{1}{a+Ca}\right)e^{{\lambda}(a+Ca)}. \end{aligned}\] Hence, for all but finitely many cases, it suffices to find conditions on \(a>1\) and \(C\geq1\) such that \[\label{Tinequality1} e^{{\lambda}(a)+{\lambda}(Ca)-{\lambda}(a+Ca)} > \frac{ \pi \sqrt{24a-1}\sqrt{24Ca-1}}{\sqrt{3} \sqrt{24(a+Ca)-1}}\cdot\frac{\left(1+\frac{1}{a+Ca}\right)}{\left(1-\frac{1}{a}\right)\left(1-\frac{1}{Ca}\right)}.\] For convenience, define \[T_a(C):={\lambda}(a)+{\lambda}(Ca)-{\lambda}(a+Ca)\hspace{.5cm}\text{and}\hspace{.5cm}S_a(C):=\frac{\left(1+\frac{1}{a+Ca}\right)}{\left(1-\frac{1}{a}\right)\left(1-\frac{1}{Ca}\right)}.\] Then by taking logarithms, we find that ([\[Tinequality1\]](#Tinequality1){reference-type="ref" reference="Tinequality1"}) is equivalent to \[\label{T>S} T_a(C)>\log\left(\frac{\pi \sqrt{24a-1}\sqrt{24Ca-1}}{\sqrt{3}\sqrt{24(a+Ca)-1}}\right)+\log(S_a(C)).\] As functions of \(C\), it can be shown that \(T_a(C)\) is increasing and \(S_a(C)\) is decreasing for \(C\geq1\), and thus \[T_a(C)\geq T_a(1)\] and \[\log(S_a(1)) \geq \log (S_a(C)).\] Hence it suffices to show that \[T_a(1) > \log\left(\frac{\pi \sqrt{24a-1}\sqrt{24Ca-1}}{\sqrt{3}\sqrt{24(a+Ca)-1}}\right)+\log(S_a(1)).\] Moreover, since \[\frac{\sqrt{24Ca-1}}{\sqrt{24(a+Ca)-1}} \leq 1\] for all \(C\geq1\) and all \(a>1\), it suffices to show that \[\label{T>S1} T_a(1) > \log\left(\frac{\pi \sqrt{24a-1}}{\sqrt{3}}\right)+\log (S_a(1)).\] By computing the values \(T_a(1)\) and \(S_a(1)\), we find that [\[T\>S1\]](#T>S1){reference-type="eqref" reference="T>S1"} holds for all \(a \geq 6\). To complete the proof, assume that \(2 \leq a\leq 5\). For each such integer \(a\), we calculate the real number \(C_a\) for which \[T_a\left(C_a\right)=\log\left(\frac{ \pi \sqrt{24a-1}}{\sqrt{3}}\right)+\log (S_a\left(C_a\right)).\] The values \(C_a\) are listed in the table below.
By the discussion above, if \(b=Ca\geq a\) is an integer for which \(C > C_a\), then ([\[T\>S\]](#T>S){reference-type="ref" reference="T>S"}) holds, which in turn gives the theorem in these cases. Only finitely many cases remain, namely the pairs of integers where \(2 \leq a \leq 5\) and \(1 \leq b/a \leq C_a\). We compute \({\mathrm{spt}}(a),{\mathrm{spt}}(b)\), and \({\mathrm{spt}}(a+b)\) in these cases to complete the proof. 0◻
## Proof of Conjecture ([\[3\]](#3){reference-type="ref" reference="3"})
We require some lemmas and a proposition analogous to those of Desalvo and Pak in order to prove the remaining conjectures. The following is.
By Theorem [\[riad\]](#riad){reference-type="ref" reference="riad"}, we may write \[{\mathrm{spt}}(n)=f(n)+E_s(n)\] where \[\label{sptsumdetails} f(n):=\frac{\sqrt{3}}{\pi\sqrt{24n-1}}e^{{\lambda}(n)}\] and \[\begin{aligned}
|E_s(n)| < (3.59 \times 10^{22})2^{q(n)}(24n-1)^2e^{{\lambda}(n)/2}. \end{aligned}\]
To prove Conjecture ([\[3\]](#3){reference-type="ref" reference="3"}), we must show that \[{\mathrm{spt}}(n)^2 > {\mathrm{spt}}(n-1){\mathrm{spt}}(n+1)\] for \(n \geq 36\). Taking logarithms, we see that this is equivalent to \({\mathrm{spt}}_2(n) > 0.\) By the lower bound in Proposition [\[prop\]](#prop){reference-type="ref" reference="prop"}, we have \({\mathrm{spt}}_2(n)>0\) for all \(n \geq 6553\). Finally, one can verify with a computer that \({\mathrm{spt}}_2(n)>0\) for all \(36 \leq n < 6553\). This completes the proof. 0◻
## Proof of Conjecture ([\[4\]](#4){reference-type="ref" reference="4"})
We follow closely the proof of. Recall that a sequence \(\{a(k)\}_{k=0}^{\infty}\) of non-negative integers is log-concave if \[\begin{aligned}
a(k)^2 \geq a(k-1)a(k+1) \end{aligned}\] for all \(k \geq 1\). Moreover, it is known that log-concavity implies strong log-concavity \[\begin{aligned}
a(\ell-i)a(k+i) \geq a(k)a(\ell), \end{aligned}\] for all \(0\leq k \leq \ell \leq n\) and \(0 \leq i \leq \ell-k\) (see e.g. ). Now, we have proved that \[\begin{aligned}
{\mathrm{spt}}(n)^2 > {\mathrm{spt}}(n-1){\mathrm{spt}}(n+1) \end{aligned}\] for all \(n \geq 36\). Therefore, if we take \(k=n-m\), \(\ell=n+m\), and \(i=m\), then \[{\mathrm{spt}}(n)^2 > {\mathrm{spt}}(n-m){\mathrm{spt}}(n+m)\] for all \(n>m>1\) with \(n-m > 36\). We next consider the case \(n>m>1\) with \(1 \leq n-m \leq 36\). We will prove that \[\begin{aligned}
\label{string} {\mathrm{spt}}(n)^2 \geq {\mathrm{spt}}(m+1)^2 > {\mathrm{spt}}(36){\mathrm{spt}}(36+2m) \geq {\mathrm{spt}}(n-m){\mathrm{spt}}(n+m) \end{aligned}\] for all \(1 \leq n-m \leq 36\) with \(m \geq 6244\). On the other hand, one can verify with a computer that \[\begin{aligned}
{\mathrm{spt}}(n)^2 > {\mathrm{spt}}(n-m){\mathrm{spt}}(n+m) \end{aligned}\] for all \(1 \leq n-m \leq 36\) with \(m < 6244\). This completes the proof of Conjecture ([\[4\]](#4){reference-type="ref" reference="4"}), subject to verifying the inequalities ([\[string\]](#string){reference-type="ref" reference="string"}). Since \(n \geq m+1\), we have \[\begin{aligned}
{\mathrm{spt}}(n)^2 \geq {\mathrm{spt}}(m+1)^2. \end{aligned}\] Moreover, since \(n-m \leq 36\) we have \[\begin{aligned}
{\mathrm{spt}}(n-m)< {\mathrm{spt}}(36), \end{aligned}\] and thus \[\begin{aligned}
{\mathrm{spt}}(36){\mathrm{spt}}(36+2m) \geq {\mathrm{spt}}(n-m){\mathrm{spt}}(n+m). \end{aligned}\] This verifies the first and third inequalities in ([\[string\]](#string){reference-type="ref" reference="string"}). It remains to prove that \[\begin{aligned}
\label{sptlast} {\mathrm{spt}}(m+1)^2 > {\mathrm{spt}}(36){\mathrm{spt}}(36+2m) \end{aligned}\] for all \(m \geq 6244\). Taking logarithms in ([\[sptlast\]](#sptlast){reference-type="ref" reference="sptlast"}), we see that it suffices to prove \[\begin{aligned}
\label{sptlast2} 2\log ({\mathrm{spt}}(m+1))-\log ({\mathrm{spt}}(36))-\log ({\mathrm{spt}}(36+2m)) > 0 \end{aligned}\] for all \(m \geq 6244\). By and, respectively, we have the lower and upper bounds \[\begin{aligned}
\frac{e^{2\sqrt{m}}}{2\pi m e^{1/6m}} < p(m) < e^{\pi \sqrt{(2/3)m}} \end{aligned}\] for all \(m \geq 1\). Then by the inequality stated in Conjecture ([\[1\]](#1){reference-type="ref" reference="1"}) (which is true by Theorem [\[main\]](#main){reference-type="ref" reference="main"}), we have \[\label{sptsqueeze} \frac{\sqrt{6}}{\pi} \sqrt{m} \frac{e^{2\sqrt{m}}}{2\pi me^{1/6m}} < {\mathrm{spt}}(m) < \sqrt{m}e^{\pi \sqrt{(2/3)m}}\] for all \(m \geq 5\). Using the inequalities ([\[sptsqueeze\]](#sptsqueeze){reference-type="ref" reference="sptsqueeze"}) and \({\mathrm{spt}}(36)<90000\), we see that the left hand side of ([\[sptlast2\]](#sptlast2){reference-type="ref" reference="sptlast2"}) is bounded below by the function \[2\log\left(\frac{\sqrt{6(m+1)}}{2\pi^2(m+1)e^{1/6(m+1)}}\right)+4\sqrt{m+1}-\log(90000)-\log(\sqrt{36+2m})-\frac{\pi\sqrt{2(36+2m)}}{\sqrt{3}}\] for all \(m \geq 4\). A calculation shows that this function is positive for all \(m \geq 6244\). 0◻
## Proof of Conjecture ([\[5\]](#5){reference-type="ref" reference="5"})
Taking logarithms, we find that Conjecture ([\[5\]](#5){reference-type="ref" reference="5"}) is equivalent to \[{\mathrm{spt}}_2(n)<\log\left(1+\frac{1}{n}\right)\] for all \(n \geq 13\). By the upper bound in Proposition [\[prop\]](#prop){reference-type="ref" reference="prop"} and some straightforward estimates, we have \[{\mathrm{spt}}_2(n)<\frac{2}{n^{3/2}}<\frac{1}{n+1}<\log\left(1+\frac{1}{n}\right)\] for all \(n \geq 6445\). Finally, one can verify with a computer that the conjectured inequality holds for all \(13 \leq n < 6445\). This completes the proof. 0◻
## Proof of Conjecture ([\[6\]](#6){reference-type="ref" reference="6"})
We follow closely the proof of. Taking logarithms, we find that Conjecture ([\[6\]](#6){reference-type="ref" reference="6"}) is equivalent to \[{\mathrm{spt}}_2(n)<\log\left(1+\frac{\pi}{\sqrt{24}n^{3/2}}\right)\] for all \(n \geq 73\). By ([\[spt2upper\]](#spt2upper){reference-type="ref" reference="spt2upper"}) we have \[\begin{aligned}
{\mathrm{spt}}_2(n) < \frac{24\pi}{\big(24(n-1)-1\big)^{3/2}}-\frac{288}{\big(24(n+1)-1\big)^2} + 2 M(n) + g(n+1) +g(n-1) \end{aligned}\] for all \(n \geq 4\). On the other hand, by we have \[\frac{24\pi}{\big(24(n+1)-1\big)^{3/2}}<\frac{24\pi}{(24n)^{3/2}}-\left(\frac{24\pi}{(24n)^{3/2}}\right)^2+\frac{3}{2n^{5/2}}\] for all \(n \geq 50\), and by we have \[-\frac{288}{\big(24(n+1)-1\big)^2}<\frac{1}{6n^{5/2}}-\frac{1}{2n^2}\] for all \(n \geq 50\). Therefore, for all \(n \geq 50\) we have \[{\mathrm{spt}}_2(n)<\frac{24\pi}{(24n)^{3/2}}-\left(\frac{24\pi}{(24n)^{3/2}}\right)^2 + \frac{5}{3n^{5/2}}-\frac{1}{2n^2} + 2 M(n) + g(n+1) +g(n-1).\] Now, a calculation shows that \[\begin{aligned}
\frac{5}{3n^{5/2}}-\frac{1}{2n^2} + 2 M(n) + g(n+1) +g(n-1) < 0 \end{aligned}\] for all \(n \geq 7211\). Hence \[\begin{aligned}
{\mathrm{spt}}_2(n) < \frac{24\pi}{(24n)^{3/2}}-\left(\frac{24\pi}{(24n)^{3/2}}\right)^2 = \frac{24\pi}{(24n)^{3/2}}\left(1-\frac{24\pi}{(24n)^{3/2}}\right) \end{aligned}\] for all \(n \geq 7211\). Then using the inequality \[\begin{aligned}
x(1-x)<\log(1+x) \quad \textrm{for} \quad x>0, \end{aligned}\] we get \[\begin{aligned}
{\mathrm{spt}}_2(n) < \log\left(1+ \frac{24\pi}{(24n)^{3/2}}\right)=\log\left(1+\frac{\pi}{\sqrt{24}n^{3/2}}\right) \end{aligned}\] for all \(n \geq 7211\). Finally, one can verify with a computer that this inequality also holds for all \(73 \leq n < 7211\). This completes the proof. 0◻ | {'timestamp': '2019-05-29T02:20:08', 'yymm': '1706', 'arxiv_id': '1706.01814', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01814'} |
null | null |
# \... | {'timestamp': '2017-06-14T02:07:23', 'yymm': '1706', 'arxiv_id': '1706.01907', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01907'} |
# Introduction {#sec1}
One-dimensional quantum systems with point interactions are quite non-trivial. The point interaction in one-dimensional quantum systems has a relatively large parameter space, in comparison with those in higher dimensions. It has been known that a point interaction in one dimension is parametrized by the group \(U(2)\), while that in two or three dimensions is parametrized by \(U(1)\). The parameters characterize connection conditions for a wavefunction and its derivative. In one dimension, a variety of connection conditions leads to various intriguing physical properties such as duality, anholonomy, supersymmetry, geometric phase, and scale anomaly. We consider the scattering of a quantum particle by point interactions in one dimension. Several authors [@SC16; @SC2; @SC3; @SC4; @SC15; @SC5; @SC6; @SC7; @SC8; @SC10; @SC11; @SC12; @SC13; @SC14] have been investigated the scattering properties by potential barriers made of the Dirac delta functions and its (higher) derivatives. Essential properties of the scattering by a single point interaction parametrized by the \(U(2)\) were investigated in. The authors of discussed the scattering by scale-invariant point interactions, which are considered to be a subclass of the point interactions. The parameter space of a scale-invariant point interaction is given by a sphere \(S^2\), which is described by two parameters. They showed that the quantum transmission through arbitrary \(N\) scale invariant point interactions exhibits random quantum dynamics. In our previous paper, we investigated the scattering by two independent, successive parity-invariant point interactions in one dimension. The parameter space of a parity-invariant point interaction is given by a torus \(T^2 = S^1 \times S^1\). Thus the parameter space of two independent parity-invariant point interactions is given by \(T^2 \times T^2\), which is described by four real parameters. Even in the reduced parameter space, it was shown that non-trivial resonant conditions for perfect transmission appear. In this paper, we extend our previous work to the cases of the scattering by two independent point interactions in one dimension without any restriction, that is, on the whole parameter space \(U(2)\times U(2)\). The main purposes of this paper are to investigate the conditions for perfect resonant transmission on the whole parameter space and to provide its physical interpretation. This paper is organized as follows. In section [2](#sec2){reference-type="ref" reference="sec2"}, we review the scattering of plane wave by a single point interaction, and give the scattering matrix formula. In section [\[sec3\]](#sec3){reference-type="ref" reference="sec3"}, we consider the scattering by two independent point interactions, derive the scattering amplitudes and the transmission probability, and investigate the conditions for the parameter space under which perfect resonant transmission occurs. Furthermore, the physical interpretation of the perfect resonant transmission condition is discussed. Finally, section [\[sec4\]](#sec4){reference-type="ref" reference="sec4"} is devoted to a summary.
# One-dimensional quantum systems with a point interaction {#sec2}
## Connection conditions and parametrization
In this section, we discuss quantum mechanics in one dimension (\(x\)-axis) with a point interaction located at \(x= A point interaction is specified by a characteristic matrix\)UU(2)\(, and a wavefunction\)(x)\(and its derivative\)'(x) (= (x))\(are required to obey the connection conditions \begin{eqnarray} (U-I) \Phi( \end{eqnarray} where\)I\(is the\)`<!-- -->`{=html}2\(identity matrix. The parameter\)L_0\(is an arbitrary nonzero constant with the dimension of length, and \begin{eqnarray} \Phi( \varphi( \varphi( \end{array} \right), \ \ \Phi'( \begin{array}{c} \varphi'( -\varphi'( \end{array} \right), \end{eqnarray} where\) | {'timestamp': '2017-09-21T02:04:21', 'yymm': '1706', 'arxiv_id': '1706.01595', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01595'} |
null | null |
null | null |
null | null |
# Introduction
Most European countries are now engaged in the energy transition whose ultimate goal is to meet energy demand from human activities solely with renewable energy sources (RES). In its current intermediate stages, the transition steadily increases the penetration of nondispatchable electricity productions, which results in large uncontrolled fluctuations in power generation. The development of RES in Europe follows from strong public investments and incentives which temporarily bias the electricity market. RES have negligible marginal cost and consequently their increasing penetration artificially lowers electricity prices below production costs for many other power generations. Simultaneously, new RES such as solar photovoltaics and wind turbines have undispatchable, strongly fluctuating productions which need to be counterbalanced by controllable, dispatchable productions and electrical energy storage solutions. In the context of the energy transition it is therefore of key importance to understand how RES penetrations can be increased without jeopardizing the dispatchable productions required by the next level of RES penetration. One of our main interests here is the hydroelectric sector, with its fast dispatchable and fully controllable dam productions as well as pumped-storage (PS), the only sizeable, mature storage solution to date. Hydroelectricity seems like an ideal partner to RES in the context of the energy transition and one may anticipate that further investments in new PS or higher power dam facilities would significantly help to absorb increased production fluctuations from larger RES penetrations. However, somewhat ironically, the current low electricity prices penalize investments in new hydroelectric facilities--with today's economic conditions in the European electric sector, RES jeopardize the future of hydroelectricity, arguably one of its main and most reliable future partner. To evaluate scenarios for the energy transition it is therefore of paramount importance to evaluate whether this trend will continue, and if yes, for how long, and determine if and when the precious flexibility of hydroelectric production will be again rewarded. To achieve that, one needs a reliable dispatch model for all types of electric productions as well as a reliable economic indicator. The current electricity market requires tools for financial analysis with increased precision to identify the need for further production investments and the returns they will generate. Of particular interest is to try and implement transparent dispatch models at the level of the pan-European grid that rely as weakly as possible on highly volatile political, economical or financial predictions. In this manuscript, we develop an integrated physico-economical power dispatch model relying on physical constraints for electric power productions, on published future production capacity developments and on basic, demand-supply economical laws only. We demonstrate the validity of our model by reproducing rather accurately historical 2015 electricity production profiles for all power production types in nineteen European countries. We argue that our model will become more and more accurate as the energy transition progresses and investigate European dispatches as well as intercountry exchanges for 2030. This allows us to identify the needs for increased grid capacity, for further storage capacity as well as future rules of engagement for hydroelectric dam power plants. Additionally, we clarify the financial conditions prevailing in the electricity market in the forthcoming stages of the energy transition. The main limitation of our approach is related to uncertainties in future installed production capacities in European countries, which will depend on more general economic and financial conditions in Europe and in the world as well as on future political and societal decisions. However, regardless of these mostly unpredictable conditions, we argue that our dispatch and revenue evaluation model retains its validity, provided production capacities are adapted to their true evolution. In other words, our model qualitatively predicts production dispatches and revenues for given production capacities. Accordingly, feasibility studies should consider various scenarios for future production capacities to investigate which one presents the best operational and financial perspectives. The manuscript is organized as follows. In Section II we discuss our aggregated pan-European power grid model, the optimal power flow and the parameters on which our dispatch model is based. In Section III we calibrate these parameters by reproducing historical data for the year 2015. In Section IV, we apply our model to one ENTSO-E scenario for future European electric power production capacities for the year 2030. Our results show how electric power is transferred across the continent as different meteorological conditions prevail, and from this, we infer the magnitude of intercountry power flows. In Section V we introduce the residual load as an economic indicator which allows us to evaluate normalized future revenues. As an example we calculate future revenues for pumped-storage hydroelectric power plants in Germany. Conclusions and discussions of our results and our model are presented in Section VI.
# An aggregated model for future pan-European electricity dispatch
We develop an equivalent model to determine future power dispatches in the pan-European power grid at different stages of the energy transition. Equivalent aggregated models have a relatively long history. They are standardly used for systemic investigations such as ours, where precise details of power flows are not crucial (as opposed to, say, grid stability investigations) and exact, geographically resolved production and consumption data are hard to obtain.
## An aggregated pan-European electric grid
Fig. [\[euro_grid\]](#euro_grid){reference-type="ref" reference="euro_grid"} shows our aggregated European grid, with each node representing an independent dispatch zone (Portuguese consumption and production are included in the Spain node). Aggregated lines have admittances obtained via a standard reduction method and thermal limits given by the sum of the physical lines they represent. The power flows are computed in the DC lossless approximation.
## Productions and Consumptions
Consumptions and productions are aggregated within each dispatch region and attributed to the corresponding node. Power productions are subdivided into two sets. They are,
- Non-flexible productions, mostly consisting of run-of-the-river (RoR), solar photovoltaics (PV) and wind turbine productions. The remaining non-flexible productions are grouped into \"miscellaneous productions\". Note that RoR is in principle flexible, at least to some extent, however we neglect curtailment and consider that, as for PV and wind turbines, RoR production is determined by weather/seasonal conditions only.
- Flexible productions: We classify them into 6 types, which are (i) dam hydroelectricity, (ii) pumped-storage hydroelectricity (which can be positive as well as negative, but always counted as a production), (iii) gas and oil, (iv) nuclear, (v) hard coal and (vi) lignite productions.
For each zone and at each time, we define the residual loads \(R_i(t)\) as the difference between the consumption and the non-flexible productions, \[\label{eq:rl} R_i(t) = L_{i}(t)-P_i^{\mathrm{inflex}}(t),\] where \(L_i(t)\) and \(P_i^{\mathrm{inflex}}(t)\) respectively give the load and the sum of the non-flexible productions at time \(t\) in the \(i^{th}\) zone. In our approach, non-flexible sources produce according to weather and seasonal conditions, and only flexible productions are dispatchable. Our task is therefore to dispatch all flexible productions so that their production is equal to the total residual load at all times-this is equivalent to satisfy the balance condition that consumption is equal to production at all times. The association of European transmission grid operators (ENTSO-E) provides data on historical production and load profiles and installed capacities in the different countries and forecasts for annual RES productions that we use to set up our model.
## Economic dispatch
A large number of different optimized power flows exist. Our dispatch algorithm follows a merit order. The latter is based, first, on marginal costs, \(a^k\), specific to each production type, \(k\). Second, we introduce effective parameters in the form of repulsion costs, \(b^k\), which progressively increase the total production cost as the production increases and reaches its maximal possible value. Such repulsion costs do not directly correspond to any real economic cost, however we found that they are necessary to smoothen production curves and reproduce historical time series faithfully. With these two parameters for each of the six different flexible productions, our model has a total of 12 parameters that need to be calibrated. The production cost in the \(i^{th}\) zone at each time step \(\Delta t=1\mathrm{h}\) is given by a sum over the marginal and repulsion costs for all production types as \[W_i(t) = \sum_k \left[a^kP_i^k(t) + b^k\frac{P_i^k(t)^2}{P_{\max i}^k}\right]\Delta t,\] where \(P_i^k(t)\) is the power generated by a given production type labelled \(k\), in a geographical zone labelled \(i\), at time \(t\), and \(P_{\max i}^k\) is the corresponding installed capacity. Our algorithm is based on an optimal power flow which determines the production profiles \(\{P_i^k(t)\}\) minimizing the total, annual generation cost \[\label{eq:gains} W(\{P_i^k(t)\}) = \sum_{i,t} W_i(t),\] under the following technical constraints:
#### Power limits
\(P_i^k(t) \le P_{\max i}^k\), \(\forall t\); the power generated never exceeds its maximal installed capacity.
#### Ramp rates
\(| \partial P_i^k(t)/\partial t| \le \Gamma_i^k\), \(\forall t\); each production type has a maximal ramp rate \(\Gamma_i^k\) at which the production increases or decreases. These ramp rates are similar, but not exactly equal, to the real, technical rates. We adapted them slightly when calibrating our model, to reproduce historical production time series better.
#### Internodal power flows
\(|P_{ij}(t)| \le P_{ij}^{\rm therm}\); they should never exceed the thermal limit \(P_{ij}^{\rm therm}\) of the aggregated line between node \(i\) and \(j\) that carries them; when they do, a different dispatch must be implemented to correct this.
#### Dam storage
Dam hydroelectric plants are constrained by the finiteness of their reservoir and the annual water intake into the latter.
# Model Calibration {#model_cal}
To calibrate the parameters in our model, we fixed non-flexible productions to those of 2015 and optimized the 12 parameters in our model to reproduce true 2015 production data as faithfully as possible. In Fig. [\[DE_dispatch\]](#DE_dispatch){reference-type="ref" reference="DE_dispatch"} we show the result for a winter and a summer week in Germany and Italy, after the 12 parameters have been optimized. The agreement between dispatched and actual productions is excellent. We found comparable agreement between calculated and real 2015 productions for all other countries in our aggregated model. Another level of complexity is brought about by dam hydroelectricity with its great flexibility. To illustrate that our dispatch model works even in that case, we show in Fig. [\[hydro_dispatch\]](#hydro_dispatch){reference-type="ref" reference="hydro_dispatch"} the productions of Swiss and Norwegian dam hydroelectric plants during one week in summer and winter. Despite the inherent difficulty to dispatch dam hydro production, we see that our model captures most features of the 2015 production rather faithfully. Few discrepancies exist, in particular our calculation overuses flexibility in Norway in winter, which we attribute to mid-and long-term supply contracts whose effect cannot be captured by our model. Even with these few discrepancies, we are unaware of another model that captures the national productions up to this level of detail at a European scale, including hydroelectric productions. From Fig. [\[DE_dispatch\]](#DE_dispatch){reference-type="ref" reference="DE_dispatch"} and [\[hydro_dispatch\]](#hydro_dispatch){reference-type="ref" reference="hydro_dispatch"}, we conclude that our model is calibrated and fully valid. Its 12 free parameters having been fixed, we next use the model to investigate future scenarios of the energy transition.
# Future power dispatch
Having calibrated our model, we next investigate how the flexible productions are dispatched, both geographically and in time, to handle future large penetration of RES in later stages of the energy transition. Our results are based on three assumptions. First, for production capacities in each country, we use the ENTSO-E scenario *2030 Vision 4 of European Green Revolution* . Second, non-flexible productions are obtained by rescaling their 2015 production profiles in direct proportion to their capacity evolution. Third, we assume that consumption profiles will not be too different in 2030 from what they are now and use 2015 consumption profiles for each country in the aggregated model of Fig. [\[euro_grid\]](#euro_grid){reference-type="ref" reference="euro_grid"}. Obviously, our model can be used to check any other production and consumption scenario one may wish to implement. Fig. [\[dispatch_30_W\]](#dispatch_30_W){reference-type="ref" reference="dispatch_30_W"} shows the productions of Germany, Italy, Switzerland and Norway for two consecutive weeks in the winter of 2030. One sees first that when RES have low production (first five days), dam hydro productions are high to help supplying the demand for electricity. When RES productions are high, dam hydro production is significantly lowered. In particular, one sees that, with large RES productions, Switzerland continuously imports electricity during several consecutive days, which is never the case nowadays. Pump-storage hydro is additionally intensively used, as it produces a lot when RES produce little and consumes (pumps) when RES productions are high. The yearly dam hydro production corresponds to the yearly water intake and as it is not expected to change significantly in the next two decades, the annual productions of Norway and Switzerland are comparable to the 2015 productions. Note that the total Swiss production diminishes a bit compared to 2015, which is due to the incomplete substitution of dismantled nuclear power by RES in the chosen ENTSO-E scenario for Switzerland. Fig. [\[power_flows\]](#power_flows){reference-type="ref" reference="power_flows"} shows the power flow of three important interconnects for the same two weeks. For comparison we added the power flows obtained for 2015 for the same period. We observe that more flexibility is asked of dispatchable productions. In particular, the flows in 2015 tend to have a dominant direction. For instance, the CH-IT connection is used for Italian import only. In 2030, however, the increased Italian PV capacity results in a reversed flow across the CH-IT interconnect. It is clear that large RES productions induce increased power exchanges between European countries, often reversing the direction of the power flows and leading the latter regularly close to their thermal limits and sometimes in an unexpected direction.
# Effective electricity price
To anticipate changes and necessary upgrades to electric power systems in the light of the energy transition, a reliable economic indicator is needed which gives a qualitatively reliable estimate for the price of electricity. Here we deliberately choose to use an indicator solely based on technico-physical conditions and not on highly speculative economic forecasts. As a matter of fact, such an economic indicator exists, which reflects quite clearly the law of supply and demand: it is the residual load \(R_i(t)\) defined in Eq. ([\[eq:rl\]](#eq:rl){reference-type="ref" reference="eq:rl"}). At a qualitative level, it is easily understood that when \(R_i(t)\) is low (high), the demand for flexible electricity and thus the price one is ready to pay for it are also low (high). More surprising is that the correlation between electricity prices and residual load is almost perfect quantitatively. This correlation was already observed in Ref. for the case of Germany. Fig. [\[rl_da\]](#rl_da){reference-type="ref" reference="rl_da"} shows that the residual load is strongly correlated with the day-ahead prices in Germany for two weeks, one in summer and one in winter. The correlation remains strong even when considering the whole year and we found a correlation coefficient between residual load and day-ahead electricity price distributed as \(r \in [0.5,0.9]\) for all countries in our aggregated model. Given that nowadays a major part of electricity transactions occur on the day-ahead market, and that in all likelihood, with the expiration and non-renewal of many long-term electricity contracts, this trend will be strengthen in the foreseeable future, such large correlations suggests to introduce an effective electricity price based on the residual load as
with two parameters \(\alpha_i\) and \(\beta_i\) to be empirically determined from historical data. We obtained estimates \(\alpha_i \simeq\) 1 \[EUR/(MWh \(\cdot\) GW)\] and \(\beta_i \simeq\) 20 \[EUR/MWh\] from recent historical data for Germany. Having introduced this effective electricity price, it is now possible to investigate future economic conditions and opportunities with our model. To illustrate this, we evaluate future economic conditions for pumped-storage (PS) power plants. The revenue generated by a PS plant depends on its pump/turbine powers \(P_{\mathrm{p}i}(t)\) and \(P_{\mathrm{t}i}(t)\) and the filling \(S_{\mathrm{PS}_i}(t)\) of its reservoirs as
At each time step \(\Delta t=\)`<!-- -->`{=html}1h, the reservoir filling evolves as \[\label{eq:ps2} S_{\mathrm{PS}_i}(t+\Delta t) = S_{\mathrm{PS}_i}(t) + [\eta P_{\mathrm{p}_i}(t)-\eta^{-1} P_{\mathrm{t}_i}(t)]\Delta t\] with a typical pump/turbine efficiency of \(\eta = 0.9\). Including hydro pumped-storage defined by Eqs. ([\[eq:ps0\]](#eq:ps0){reference-type="ref" reference="eq:ps0"}--[\[eq:ps2\]](#eq:ps2){reference-type="ref" reference="eq:ps2"}), our pan-European aggregated model is similar to the power-node model of Ref. . To obtain the PS production/consumption profile, we include its effective revenue, \(G\) in Eq. ([\[eq:ps0\]](#eq:ps0){reference-type="ref" reference="eq:ps0"}), in the total gain \(W\) to optimize \[see Eq. ([\[eq:gains\]](#eq:gains){reference-type="ref" reference="eq:gains"})\], and the constraints of Eqs. ([\[eq:ps1\]](#eq:ps1){reference-type="ref" reference="eq:ps1"}) and ([\[eq:ps2\]](#eq:ps2){reference-type="ref" reference="eq:ps2"}) into our aggregated model. Fig. [\[fig:ps\]](#fig:ps){reference-type="ref" reference="fig:ps"} shows time profiles for PS production/consumption, electricity prices and PS reservoir level for a fictitious 1 GW, 32 GWh PS plant. The production profile is, as expected, clearly correlated with the electricity price, and the constraints on the reservoir level are met. The PS revenue is further calculated using Eq. ([\[eq:ps0\]](#eq:ps0){reference-type="ref" reference="eq:ps0"}) and we plot it in Fig. [\[fig:psgains\]](#fig:psgains){reference-type="ref" reference="fig:psgains"} for the special case of Germany. Data are superimposed on histograms depicting the annual RES electricity production for PV (yellow) and wind turbines (light blue). We normalized the revenue with the revenue obtained from our dispatch model for the year 2000. We correctly obtain a significant revenue reduction from 2008 on, with a minimum around 2013, after which the revenue increases again to get back to its pre-2008 value at around 2015-2016. The latter behavior is likely a bit premature, however, overall, our data qualitatively suggest that, (i) after a period of difficulties, PS will get back to its pre-2008 profit level rather soon, at least in Germany, and (ii) how soon PS gets back to larger profit margins depends mostly on how fast RES are substituted for fossil productions.
# Conclusions
We have constructed a pan-European model for the future electricity market. Using a mathematically well-defined merit order, we calibrated it so that it reproduces 2015 production profiles. We investigated how productions will change up to 2030 and found that enhanced intercountry power exchanges will help absorbing large fluctuations of productions from PV and wind turbines. We introduced an effective electricity price and illustrated its predictive power by investigating revenues generated by pump-storage facilities in Germany. Our results suggest that hydro pump-storage power plants will again generate comfortable profits in the future. How soon that will be depends mostly on the pace at which the energy transition proceeds. | {'timestamp': '2017-06-07T02:04:43', 'yymm': '1706', 'arxiv_id': '1706.01666', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01666'} |
# Introduction {#introduction .unnumbered}
Bibliometric indices are ubiquitous in measuring the scientific impact of both journals, institutions, research groups and individual researchers. At some stage in their evaluation process, scientific committees, administrators or policy makers often rely on citation data to assess scientific output. Among the variety of measures that can be derived from raw citation data, the \(h\)-index became the most popular bibliometric criterion for the ranking of research accomplishment . The \(h\)-index of an individual is given by the total number of published papers with citations \(\geq h\). Due to its simple definition, its ease of computation from existing bibliography databases, its robustness against errors in the long tails of the citations-rank distribution and its good properties when quantifying the scientific production and its impact, the \(h\)-index has been largely adopted as a reliable measure of research output . Nowadays, the automatic calculation of \(h\)-indices is a built-in feature of major bibliographic databases such as Google Scholar, Researchgate, Scopus and Web of Science. While it is quite dependent on the database being used, the typical value of the \(h\)-index is robust when it comes to ranking both individuals and research institutions. Shortcomings of the \(h\)-index have been pointed out, although many other available indicators do suffer from similar biases . Among the arguments raised against using the \(h\)-index one can enumerate: it does not allow to compare scientists from different disciplines , it does not take into account multi-authored papers , it is a time-dependent quantity , it does not highlight the citation scores of top articles , it does not take into account the context of the citations in the papers. Consequently, several subsequent variations of the \(h\)-index have been proposed to overcome some of these drawbacks . A non exhaustive list of such bibliometric indices include the \(g\)-index , the \(h^{(2)}\)-index , the \(A\)-\(R\)-and \(AR\)-index , the \(m\)-index , the \(h_m\)-index , etc. Interestingly, all of them were based on an increasingly sophisticated analysis of raw citation data, which makes them more difficult to access. Moreover, although being designed to surmount its perceived shortcomings, they were found to be quite positively correlated with the \(h\)-index . This fact yielded the general consensus that no other bibliometric indicator of research output is clearly preferable to the \(h\)-index . Since it is desirable to evaluate researchers and/or institutions on more than a single quality, we propose to supplement the \(h\)-index with a new quantitative indicator specialized to interdisciplinary research and readership extent. This journal-based index is by construction independent of citation counts and measures different qualities that could be useful to assess, for example in the process of hiring individuals who are expected to teach and to advise, for which high specialization values are not required.
# A new two-dimensional index {#a-new-two-dimensional-index .unnumbered}
Many refinements of the \(h\)-index, on both theoretical and empirical sides, were based on alternative analysis of citation data. We believe that this is the reason for their strong positive correlation with the latter. Therefore, the quest for other bibliometric indices should be aimed at complementing the \(h\)-index rather than replacing it. A new index should be a quantitative indicator with a core that is independent of citation data and as easy to compute from existing bibliographic databases as the \(h\)-index. It should concern both individuals, research groups and institutions and should highlight different achievements than research output, which is already well quantified by the \(h\)-index. Noticing that the \(h\)-index is bounded by the total number of published papers, \(h\leq N\), we have looked for another possibly interesting and simple quantity that satisfies the same property. A straightforward one is the number of different journals in which these papers were published. This number, denoted by \(N_j\), is not directly available in existing bibliographic databases but can be easily extracted from them, either by direct counting or using simple script codes. Obviously, \(N_j\leq N\) and is not based on citation count, thus one would expect that it is unlikely correlated to research output. Indeed, common publishing habits show that a paper is submitted to a given journal because of the studied area of research, to reach a specific scientific community, to fit a specific format or to be faced with a given more or less selective refereeing procedure. One could argue that a such choice is mainly dictated by the journal's impact factor, its ranking by a target institution or its interdisciplinary readership. Nevertheless, if the published research is worthwhile it will positively impact the \(h\)-index and consequently research output, regardless of the journal's ranking. Moreover, a paper in a high impact journal with an interdisciplinary readership is often followed by a series of detailed papers in the same subject intended for a specific scientific community, which thus contributes to increase \(N_j\). However, what type of research achievement is embedded in this new number? On the one hand, publishing in a small set of journals could indicate mono-thematic research interests while a multidisciplinary researcher is prone to publish in a wide range of journals. On the other hand, a researcher could impact a single field by publishing in the same journal on a single subject, and conversely, an author may publish about various subjects in many journals by achieving a moderate impact. These two extreme and antagonistic examples are in favour of taking into account both \(N_j\) and the \(h\)-index for more elaborate ranking purposes. Although \(N_j\) may exhibit some flaws, which will be discussed later, one can reasonably assume that it carries information about either the multidisciplinary nature of the research, the diversity of interests of the researcher, or the extent of his readership. While these features are indicators of research quality, they are clearly not directly provided by the \(h\)-index. To test the relevance of this two-dimensional index, we have chosen as a panel 95 permanent physicists belonging to a same department, namely the Physics Department of ENS Paris, which is divided in three sub-departments ("Laboratoires") with more or less specific research areas: condensed matter physics (Laboratoire Pierre Aigrain, LPA, 31 researchers), theoretical and statistical physics (Laboratoire de Physique Théorique, LPT, 26 researchers), statistical physics, biophysics and nonlinear physics (Laboratoire de Physique Statistique, LPS, 38 researchers). The \(h\)-index and \(N_j\) of each individual have been retrieved from Web of Science in July 2016. All types of publications were taken into account. Fig. [\[fig:example1\]](#fig:example1){reference-type="ref" reference="fig:example1"} shows the results of this survey: a scatter plot of these quantities, though containing a global trend, is dominated by the scatter noise, illustrating the signals low correlation. This confirms that when the department is taken as a whole, individuals bibliometric indicators \(h\) and \(N_j\) carry complementary information. This example justifies the use of both \(h\) and \(N_j\) to characterise research accomplishment. In order to sharpen the comparison between individuals, we propose a different representation of these two bibliometric indicators. Since both \(h\) and \(N_j\) are bounded by \(N\), one can define a "complex" representation of the two-dimensional index \((h,N_j)\) as follow \[H \exp\left(\frac{i\pi}{2}M\right)= \frac{h}{\sqrt{2}}+i\,\frac{N_j}{\sqrt{2}}\;. \label{eq:first}\] or, equivalently \[H=\sqrt{\frac{h^2+N_j^2}{2}}\,,\qquad M=\frac{2}{\pi}\,\arctan\left(\frac{N_j}{h}\right)\;. \label{eq:second}\] Obviously, the two-dimensional index \((H,M)\) is defined such that \(0<H\leq N\) and \(0<M\leq 1\). This representation separates the extensive contribution characterizing the volume of quality output, from an intensive characterization of its multidisciplinarity: small values of the argument \(M\) indicates strong specialization, while \(M\) close to one indicates thematic dispersion with rather low impact. Fig. [\[fig:example2\]](#fig:example2){reference-type="ref" reference="fig:example2"} reproduces the same set of data as Fig. [\[fig:example1\]](#fig:example1){reference-type="ref" reference="fig:example1"} in these two variables. Interestingly, while \(H\) is widely distributed, the average of \(M\) discriminates more sharply between the different "Laboratoires". One can even identify that the overlap between LPT and LPS is ensured by LPT researchers working in statistical physics. The overlap between data of LPA and LPS is due to the fact that LPA benefits from large publication material due the applied part of its research area (readership extent), while LPS researchers often work in different fields (multidisciplinarity). Clearly the \(M\)-index differentiates between communities and highlights readership diversity and research multidisciplinarity. The two indices together allow for a better qualification of research achievement compared to a single evaluator. Moreover, this alternative representation increases the level of independance between the two components, with a coefficient of correlation of 0.7 for \(h\) and \(N_j\) and \(0.14\) for \(H\) and \(M\) in our dataset.
# Discussion {#discussion .unnumbered}
The use of the \(h\)-index to compare scientists from different research areas is a quite difficult task due to the inherent differences among different research fields. Although it is not an exclusive problem of the \(h\)-index, there have been several different efforts in the literature to supersede it. Interestingly, our work quantitatively demonstrates a bias associated with a citation-only index, namely its inability to capture research diversity. This is in favour to define \(N_j\) as an additional index to supplement the \(h\)-index. However, the representation using the two-dimensional index \([H,M]\) is preferable since it yields weakly correlated components. Furthermore, while \(M\) is bounded by 1 and separates finely the different disciplines, \(H\), which is bounded by the number of papers, balances impact and diversity. In the following, we discuss questions that this two-dimensional index may rise and propose refinements to tackle its possible drawbacks.
- There are generalist journals that encompass different research areas and that could be downgraded by the proposed index. In order to overcome this difficulty, one could for example modify \(N_j\) to categorize the paper using its research topic. Unfortunately, there is no uniform nomenclature between journal publishers (PACS numbers, keywords, subject classification\...). To do that, one would need a standard classification of research topics that is equivalent to Digital Object Identifier (DOI) system for publishers.
- Let's try to compare two individuals with the same \(H\). In an extreme case, an individual with \(h=1\) and \(N_j=N_0\), though carrying out multidisciplinary research, is not very successful since his research is not followed up by others. On the other side of the spectrum, a scientist with \(h=N_0\) and \(N_j=1\), though very exclusive in his publication choices, can be considered much more successful due to his citation rate. Hence having \(M\approx1\) is never a good thing, while having \(M\approx0\) can be of interest in some situations. That is why we claim that it would be ideal to have \(h\approx N_j\), then \(M\approx 1/2\), thus balancing recognition and curiosity.
- The choice of the "perfect" \(M\) would of course strongly depend on the kind of target one aims at, but it lifts a degeneracy and provides deciders with an additional lever. The existence of a two-dimensional index refines the judgement depending of what one looks for. For example, hiring a researcher for a specific research position or for his teaching abilities is not the same, the former would require a smaller \(M\) than the latter. Furthermore, evaluating an institution through citation metrics only is not fair because interdisciplinary research is an indisputable quality of an institution as it illustrates its local and international attractiveness and the diversity of education offered to students.
We claim that it is too limiting of an approach to categorize individuals along a single axis. Additional and independent indicators should complement such an approach. Our main line of axiomatics is that, independently of how "good or bad" a researcher is ranked through the \(h\)-index, he may actually exhibit alternative qualities, like being a specialist, or conversely foraging results in various scientific areas. We claim that the two-dimensional index we introduced in Eq. ([\[eq:second\]](#eq:second){reference-type="ref" reference="eq:second"}) is orthogonal to the usually accepted notion of "good or bad". Finally, we hope our proposition will trigger more bibliometric studies. Full-scale analysis of the \((H,M)\)-index using different databases is an interesting question that will test its robustness and ease of computation. Before being adopted, it is necessary to apply it to real evaluation problems and situations: for example, additional studies comparing institutions should be developed. Nevertheless, we believe that combining the two numbers will refine degenerate situations and balance between impact and scope. Eventually, adding indices as legitimate as the \(h\)-index that quantify different facets of research achievements limit the misuse of the \(h\)-index. | {'timestamp': '2017-10-06T02:07:52', 'yymm': '1706', 'arxiv_id': '1706.01882', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01882'} |
# Introduction
The following document is a comprehensive guide to the operation of the program. It describes in detail the structure of the required input files and resulting output files, and explains the meaning of the individual input parameters. To help the user, the explanations are supplemented by a number of actual examples of input files for model atmospheres of various stellar types that may be used as templates for a construction of similar models. The program allows for a large number of options, controlling both the physical as well as numerical setup of the calculations. While some options are very important or even critical for the numerical performance of the code and the quality of the model, there are many options and parameters that are important only in special cases. This document is therefore divided into two parts. The first part deals with the basic numerical implementation and provides a guide for understanding the essential features and a basic operation of the program. Additional details, a comprehensive description of all physical and numerical options, and a description of all input parameters, most of which needed only in special cases, are described in Part II.
# Compiling and linking {#compile}
## Compilation {#comp_comp}
The program is distributed as several files. The largest is `tlusty[nnn].f` where `nnn` represents the current version number. In the following text, we take 205 for the current version number. Communication between subprograms is principally carried out through labeled common blocks. To allow for the program to be scaled (re-dimensioned) easily, arrays are dimensioned by parameter constants. The arrays and parameters are defined using INCLUDE files
IMPLIC.FOR BASICS.FOR ATOMIC.FOR MODELQ.FOR ITERAT.FOR ARRAY1.FOR ODFPAR.FOR ALIPAR.FOR
The INCLUDE files have to reside in the same directory as the `tlusty205.f` file, and, under UNIX/LINUX, their names *must* be in capital letters, e.g. `BASICS.FOR`. We stress that the INCLUDE files generally evolve together with tlusty (more parameters are being added), so that the user is advised to copy these files along with any new version of tlusty. In the standard distribution, described in Paper I, § 3.1, the source file as well as the INCLUDE files are indeed located in the same directory. The compilation is done as follows:\
\(\bullet\) Under Mac OSX, or modern versions of Linux
gfortran-fno-automatic [-O3] [-o tlusty] tlusty205.f
where the option `"-fno-automatic"` indicates the static allocation of memory, which is a mandatory option. The level-3 optimization (`"-O3"`) should be switched on since it improves the performance of the code considerably. Reaming the executable as specified by `"-o tlusty"` allows the user to use script file `RTlusty` to conveniently run tlusty.\
\(\bullet\) Under older versions of LINUX
f77-fno-automatic [-O3] [-o tlusty] tlusty205.f
\(\bullet\) Under generic UNIX
f77 [-O4] [-static] [-Nl100] [-o tlusty] tlusty205.f
where the option `"-Nl100"` is sometimes needed under older variants of the SUN operation system (increases the number of continuation lines to 100); the option `"-static"`, which is equivalent to `` "`-fno-automatic" ``, is needed for some variants of UNIX. It is a default in most implementations. Similarly, the optimization (the option `"-O4"`) is a default on most workstations. If not, the optimization should be switched on since it improves the performance of the code considerably.\
\(\bullet\) One may also use (within LINUX) the commercial Portland Group compiler, pgf77, which provides better optimization (the code is typically 20-30% faster);
pgf77 [-Bstatic] [-fastsse] [-o tlusty] tlusty205.f
## Reducing the size of the executable file {#reduce}
The strategy adopted in early days of the development of tlusty, before FORTRAN90 was introduced, was to code the maximum dimensions of the important arrays that determine the overall memory consumption of the code in a few special files, communicated through the `INCLUDE` statements. This way it became possible to easily change array dimensions, and thus the memory needed, by simple changing a few lines in the INCLUDE files. Although the memory consumption is no longer such a critical issue as it used to be in the past, it may still be sometimes necessary to change the overall memory requirements. For instance, one has to be careful not to exceed 2 GB of core memory, a typical value of a medium-class Mac or a LINUX box, when computing NLTE metal line-blanketed models. The basic parameters defining array dimensions appear in the INCLUDE file `BASICS.FOR`. The first `PARAMETER` statement there contains the most important parameters; which we list below. Their exact meaning is explained through this paper.
There are also secondary parameters; their values are rarely changed. Here is their list:
The above parameters are universal. There are also other parameters which specify dimensions of arrays which are being used only under special circumstances (including using the opacity tables, Compton scattering; tri-diagonal approximate operator, etc.), which are set to 1 for the bulk of applications. These are the following:
In addition, there are several parameters in the INCLUDE file `ODFPAR.FOR` that specify array dimensions for a treatment of metal line blanketing. They are the following:
The program checks whether the current values are less than or equal to the corresponding maximum dimension, and stops if there is a conflict. The program then issues a brief message (on the standard output and on the special performance and error message file), showing the corresponding current and maximum values. The program should then be recompiled with the corresponding parameter appropriately modified in the file `BASICS.FOR` or `ODFPAR.FOR`. The convention is that the names beginning with `M` designate the maximum dimension; the analogous names beginning with `N` then denote the current values; for instance `MATOM` is the dimension of the arrays containing information about explicit atoms (stored in the INCLUDE file `ATOMIC.FOR`), while `NATOM` is the current, actual number of explicit atoms, etc. The exception from this rule is `MDEPTH` and the corresponding `ND` for the maximum and actual number of discretized depth points. A reduction of the size of the executable file is accomplished by reducing some of the above parameters specifying array dimensions in the INCLUDE files `BASICS.FOR` and `ODFPAR.FOR`). We stress again that when one computes a model without the Opacity Sampling treatment of line blanketing due to the iron-peak species, the parameters `MKULEV`, `MLINE`, and `MCFE` in the file `ODFPAR.FOR` should be set to 2 which leads to a significant reduction of the size of the executable file. We stress that the standard distribution of tlusty, as specified in Paper I, § 3.1 contains the INCLUDE files `BASICS.FOR` and `ODFPAR.FOR` that specify the basic array dimensions in such a way that one can run all the test cases described in Chapter [6](#new_ex){reference-type="ref" reference="new_ex"}, including a metal line-blanketed model atmosphere of a B star (§ [6.3](#examp_bs07){reference-type="ref" reference="examp_bs07"}). In other words, the dimensions are relatively large, and the executable takes about 1.7 GB of memory. If the user does not have a computer with 2 GB of core memory or more, then he/she have to modify the files accordingly, for instance by decreasing the values of the critical parameters in `ODFPAR.FOR` to 1, which would allow to run all tests cases in Chapter [6](#new_ex){reference-type="ref" reference="new_ex"} except the one for a B star.
## Program pretlus {#pretlus}
Setting the maximum dimensions of arrays may be a delicate task because in many instances the actual dimensions depend on input data in a rather non-trivial manner. In order to assist the user in setting the dimensions of arrays, and consequently to reduce the size of the executable file, we have developed a utility program pretlus, which is distributed along with the the main tlusty program. Program pretlus is easy to run. It accepts exactly the same input as tlusty, and it essentially performs the initialization part of tlusty without any actual calculations, and outputs the list of actual dimensions of all important arrays. The user has too check files `BASIC.FOR` and `ODFPAR.FOR` to make sure that the values of the dimension parameters given there are equal or larger than the actual values listed in the standard output from program pretlus. An example of the output produced by pretlus is given in § [6.3](#examp_bs07){reference-type="ref" reference="examp_bs07"}.
# General scheme of the input {#ginput}
## Overview {#ginput_over}
The essential feature of the input data format is that there is only a very short standard input file, which specifies (i) the very basic parameters (\(T_{\rm eff}, \log g\)) for which no reasonable default values can be specified; (ii) the name of the file where the optional, or keyword, parameters are set up; and (iii) the names of files where the atomic data for the individual ions are stored. Keyword parameters are defined as those for which the program assigns default values, which are optimum for most applications, but could be changed as required for a particular case. Keyword parameters also allow the user to choose among several alternative numerical schemes, or to cope with convergence problems. The most important ones are described in Chap. [\[nonst\]](#nonst){reference-type="ref" reference="nonst"}, the remaining ones in Chaps. [\[nst2_phys\]](#nst2_phys){reference-type="ref" reference="nst2_phys"} and [\[nst2_num\]](#nst2_num){reference-type="ref" reference="nst2_num"}, and the full list of keyword parameters together with their default values is presented in Chap. [\[list\]](#list){reference-type="ref" reference="list"}. Here is a list of input files.
- `fort.1`---The basic control file, containing just one single number, specifying whether one calculates a stellar atmosphere or accretion disk model. If this number is 0, or if the file is missing altogether, a stellar atmosphere model is to be computed. Otherwise, a disk vertical structure is computed.
- Standard input file (unit 5)---Main control data. It is a short file with only the most important parameters, and filenames of other files. The structure of the file and the meaning of the individual input parameters are explained in Chap. [4](#new){reference-type="ref" reference="new"}.
- File that specifies non-default values of the keyword parameters. Its name is specified in the standard input file.
- `fort.8`---A starting model atmosphere (if the calculation does not start from the scratch, that is, with an LTE-gray model)-see Chap. [\[unit8\]](#unit8){reference-type="ref" reference="unit8"}.
These are universal input files that need to be supplied for any mode of operation. In the standard mode, where the opacities are computed on the fly, and where one introduces the concept of explicit ions and levels (which is mandatory for NLTE models), there is an important set of input files, namely
- Files containing atomic data for the individual ions. These files are described in detail in Chap. [\[ions\]](#ions){reference-type="ref" reference="ions"}. A collection of such files is a part of the standard distribution. It is also available on the tlusty website.
Further, depending on the mode of application, which is set up by appropriate keyword parameters, there may be additional input files, which are also a part of the standard distribution of tlusty, namely:
- `lemke.dat` or `tremblay.dat`--special hydrogen line broadening tables --see also § [\[nst2_hyd\]](#nst2_hyd){reference-type="ref" reference="nst2_hyd"}
- Special tables for hydrogen quasi-molecular data--see also § [\[nst2_quas\]](#nst2_quas){reference-type="ref" reference="nst2_quas"}
- `tsuji.molec`--a table with necessary parameters for the molecular state equation--see also § [6.4](#examp_optab){reference-type="ref" reference="examp_optab"}, § [\[nonst_glob\]](#nonst_glob){reference-type="ref" reference="nonst_glob"} and § [\[nst_exa_optab\]](#nst_exa_optab){reference-type="ref" reference="nst_exa_optab"}
- `irwin.dat`--a table of Irwin partition functions--see also § [\[nst2_eos\]](#nst2_eos){reference-type="ref" reference="nst2_eos"}
In the case where some or all opacities are supplied from the pre-calculated opacity tables, one needs the following input files:
- `absopac.dat`--opacity table, required if IOPTAB=\(-1\) --see also § [6.4](#examp_optab){reference-type="ref" reference="examp_optab"} and § [\[nst2_optab\]](#nst2_optab){reference-type="ref" reference="nst2_optab"}
- Rayleigh scattering opacity table; if required--see also § [\[nst2_ray\]](#nst2_ray){reference-type="ref" reference="nst2_ray"}
- State equation tables, if required. Notice that when using the full opacity table option, one can still solve the equation of state and compute the thermodynamic parameters needed for treating convection on the fly, so these tables may not be required.
All the input files are ASCII files to enable easy portability. All the `READ` statements use a free format. Moreover, Unit 5 may contain comment lines; tlusty understands a line beginning with `` or `!` as comment. The structure of these files is explained in detail below.
## What needs to be specified? {#ginput_what}
There are two types of input parameters; the true *physical quantities* (such as the effective temperature, \(T_{\rm eff}\)), and *control parameters*--typically flags that switch on/off various numerical procedures, control the choice of adopted numerical scheme, etc. We stress that a NLTE stellar atmosphere code is not a black box. The setup of a model depends to a large extent of user's judgment (e.g., determining a degree of sophistication of the model). The user has to understand the meaning of a number of input parameters. The philosophy behind introducing several input files is an attempt to help the user in such a way that important parameters have to be set up in the standard input file, while other, either less important, or newly introduced parameters, assume their default values unless the user specifically requires to change them. This is done through the keyword parameters file. As the code evolves, it is often necessary to introduce new input parameters. These are set through the keyword parameter file, even if they may sometimes be quite important. A practical reason for setting newly introduced parameters in the keyword parameter file instead of the standard input is that there is a large library of standard input files for various models, and it would be cumbersome to be forced to change them whenever a new important input parameter is introduced. We list the essential categories of the input parameters below.
- Basic model parameters.\
These are \(T_{\rm eff}\) and \(\log g\) for stellar atmospheres, and analogous parameters for accretion disks. These are of course mandatory, and are communicated to tlusty through the standard input--see § [4.1](#new1){reference-type="ref" reference="new1"}.
- Setting LTE or NLTE.\
This is done in the standard input, see § [4.1](#new1){reference-type="ref" reference="new1"}.
- Chemical composition and the choice of explicit species.\
We recall that the *explicit* atom is defined as such for which a selected set of energy levels of a selected set of its ionization states are considered explicitly, i.e., their populations are determined by solving the kinetic equilibrium equations. These species are the only ones that are allowed to contribute to the total opacity, so their choice is important because when neglecting an important opacity source the quality of the resulting model will suffer. An *implicit* atom is not allowed to contribute to the opacity that is calculated on the fly, but is allowed to contribute to the total number of particles and to the total charge; the latter is evaluated assuming LTE ionization balance, i.e., by solving a set of Saha equations.\
The choice of explicit and implicit atoms, together with their abundances, is given in the standard input--see § [4.2](#new_at){reference-type="ref" reference="new_at"}.
- Choice of explicit ions.\
Again, the choice has to be made judiciously, as it significantly influences the overall quality of a model. The choice is also given in the standard input--see § [4.3](#new_ion){reference-type="ref" reference="new_ion"}. Notice that if the full opacity table option is adopted, there are no explicit ions defined.
- Atomic data for explicit ions.\
This is an important, and sometimes decisive, ingredient of good NLTE models. The standard input file specifies the filenames of the individual atomic data for each explicit ion; the atomic/ionic data files are separate. This enables us to store these individual files and distribute them together with the program. This is actually done on the tlusty website, where a collection of such files, in varying degree of sophistication and complexity, is presented. The user can thus use those files without being required to construct them. The atomic data files also contain a mixture of physical parameters (such as the level energies, oscillator strengths, etc.), and also control parameters for switching particular options. Therefore, in some cases one may need to modify the atomic data file to set up a different option if required (for instance, setting a different choice of level grouping-see § [\[tricks_num\]](#tricks_num){reference-type="ref" reference="tricks_num"}). Again, in the case of using full opacity tables, there is no need to have these files.
- Auxiliary physical parameters.\
Here we have for instance convection parameters, microturbulent velocity, possible external irradiation intensity, etc. As explained in the next chapters, the user has a choice of including or not including these physical mechanisms at all, and if so, to set up corresponding numerical values. These parameters are communicated through the keyword parameters file, discussed in detail in Chap. [\[nonst\]](#nonst){reference-type="ref" reference="nonst"}.
- Basic discretization parameters.\
This category contains the number of depth, frequency, and angle points used when discretizing the corresponding structural equations. The number of depth and angle points are set up by corresponding keyword parameters. The actual values of the individual column masses (which is the default depth coordinate) are either given in the input model, or are constructed in the starting LTE-grey model atmosphere. The corresponding parameters are also set up by the keyword parameters-see § [\[nonst_gr\]](#nonst_gr){reference-type="ref" reference="nonst_gr"}. If none are specified, the default values are adopted. The total number of frequency points is set up by tlusty, and depends on the actual selection of lines and continua that are treated explicitly. The user can influence some aspects of the selection via the standard input--see § [4.1](#new1){reference-type="ref" reference="new1"}, and through several optional keyword parameters--see § [\[nonst_freq\]](#nonst_freq){reference-type="ref" reference="nonst_freq"} and § [\[nonst_blank\]](#nonst_blank){reference-type="ref" reference="nonst_blank"}.
- Numerical options\
There is a large number of flags for controlling many numerical aspects, usually communicated as keyword parameters--see Chaps. [\[nonst\]](#nonst){reference-type="ref" reference="nonst"} and [\[tricks\]](#tricks){reference-type="ref" reference="tricks"}. As mentioned above, the atomic data files also contain some flags of this category. If the model computation proceeds well without any special keyword parameters file, that is if the default values of keyword parameters provide a satisfactory numerical strategy, the user does not have to set up this file. However, in a majority of cases one needs to set up some optional keyword parameters, so one is recommended to obtain at least a rudimentary knowledge from Chap. [\[nonst\]](#nonst){reference-type="ref" reference="nonst"}; more dedicated users are encouraged to study also Chaps. [\[nst2_phys\]](#nst2_phys){reference-type="ref" reference="nst2_phys"} and [\[nst2_num\]](#nst2_num){reference-type="ref" reference="nst2_num"} in detail. A practical advice how to use various options for troubleshooting is briefly summarized in Chap. [\[tricks\]](#tricks){reference-type="ref" reference="tricks"}.
# Standard input file {#new}
We now turn to a detailed description of the individual input parameters. The standard input file is composed of four basic blocks:
## First block--Basic Parameters {#new1}
This block contains only four lines of input.
### *First line:* {#first-line .unnumbered}
This is the only input record that differs for atmospheres and disks; all the other input values have the same meaning for both basic options. :
TEFF
: --effective temperature \[K\]
GRAV
: --log \(g\) \[cm s\(^{-2}\)\]
:
XMSTAR
: --mass of the central object. There are still three possibilities:
- XMSTAR \(>\ 0\)--in this case the central object is supposed to be a star; computation of the structure is done in the classical approximation. The mass can be expressed either in grams, or in solar masses.
- XMSTAR \(<\ 0\)--in this case the central object is a black hole; computation of the structure is done using general relativistic corrections. The mass of the black hole is then abs(XMSTAR); again, it is expressed either in grams or in solar units.
- XMSTAR = 0--in this case the meaning of the subsequent three quantities of the input is different; instead of using basic parameters \(M_\ast\), \(\dot M\), \(R_\ast\), and \(R/R_\ast\) or \(R/R_g\), one uses specific parameters for an annulus without explicit reference to the central object, namely \(T_{\rm eff}\), \(Q\), and \(m_0\).
Specifically, if XMSTAR \(>0\), then the subsequent three parameters are:
XMDOT
: --mass accretion rate. It can be expressed in g\(/\)s, or in \(M_\odot/{\rm yr}\).
RSTAR
: --again, two different meanings depending whether one considers a classical or general relativistic central object.\
\(\bullet\) In the classical case, RSTAR is the radius of the central star (in cm or in \(R_\odot\));\
\(\bullet\) in the case of a black hole disk, RSTAR has the meaning of the spin (angular momentum) of the black hole, expressed in geometrized units (i.e. =0 for a Schwarzschild black hole; =0.998 for a Kerr black hole with maximum stable rotation).
RELDST
: --relative distance of a given annulus with respect to the stellar radius (classical case), or the gravitational radius (for the black hole case). Gravitational radius is defined by \(R_g = G M/c^2\). Caution: the Schwarzschild radius, which is being used by some authors as a reference radius, is given by \(R_S = 2 R_g\)
In the alternative case, when XMSTAR = 0, the subsequent three numbers in the first line of input have the following meaning:
TEFF
: --effective temperature \[K\]
QGRAV
: --the proportionality coefficient, \(Q\), to evaluate the \(z\)-dependent gravity acceleration, \(g = Q z\). In the classical case, \(Q = \Omega_{\rm K}^2\), where \(\Omega_{\rm K}\) is the Keplerian velocity, \(\Omega_{\rm K} = (G M/R^3)^{1/2}\).
DMTOT
: --column mass at the midplane (=1/2 of the total column mass) \[g\]
### *Second line:* {#second-line .unnumbered}
LTE
: --a logical variable indicating whether an LTE model is going to be calculated.\
=.TRUE.--LTE model is calculated\
=.FALSE.--NLTE model is calculated
LTGRAY
: --a logical variable indicating whether an LTE-gray model is calculated at the beginning as a starting approximation for the linearization iterations.\
=.TRUE.--LTE-gray model is calculated as a starting model;\
=.FALSE.--LTE-gray model is not calculated; the user has to supply a starting model atmosphere--the Unit 8 (`fort.8`) input.
### *Third line:* {#third-line .unnumbered}
FINSTD
: --a character variable (up to 20 characters) with the name of file containing the values of keyword parameters.\
= `''` (null string)--all keyword parameters are taken with their default values--see the next section.
### *Fourth line:* {#fourth-line .unnumbered}
NFREAD
: --an indicator of the number of frequency points and their setup:\
\(>\) 0--the program sets up the *continuum* frequency points. We stress that the frequency points in the lines are set up separately based on the input data contained in the input atomic data files. The details of setup are governed by several keyword parameters (FRCMAX, FRCMIN, CFRMAX, NFTAIL, and DFTAIL--see Sect.[\[nonst_freq\]](#nonst_freq){reference-type="ref" reference="nonst_freq"}). In the default case, the program sets two frequencies near discontinuities corresponding to the bound-free transitions form all explicit levels, plus approximately NFREAD frequencies in between, plus a number of points in the high-and low-frequency tails of the spectrum. The endpoints of the high-and low-frequency tails are specified by the keyword parameters FRCMAX and FRCMIN, respectively; the number of points in the high-frequency tail are specified by the keyword parameters NFTAIL and DFTAIL. (i.e. they may be changed by a corresponding specification in the input file FINSTD).\
\(<\) 0--the frequency points are set up logarithmically equidistant between FRCMIN and FRCMAX, with the total number abs(NFREQ).
## Second block--Selection of, and basic data for, chemical species {#new_at}
The block contains one record with a value of NATOMS, and then NATOMS analogous records for the individual species; each containing three parameters: MODE, ABN, MODPF. The order of individual records *must* exactly follow the atomic number (i.e. H, He, Li, Be, B, C, N, O, etc.). tlusty can potentially treat elements with atomic number 1--99, but in practice only a smaller number of elements (typically 30) are taken into account. Another concern is that the values of the partition functions for species with atomic number higher than 30 are hardwired only for the first two ionization stages, so when treating higher ions of such species the user has to supply the corresponding expressions or tables. The individual input parameters have the following meaning:
NATOMS
: --the highest atomic number of an element that is considered (explicitly or non-explicitly). \(<\) 0--then abs(NATOMS) has the meaning as above, but all the partition functions of all species considered by the Opacity Project are evaluated from the Opacity Project ionization fraction tables, regardless of the parameter MODPF (see below).\
= 0--no explicit atoms are selected. This option only makes sense in the case of using complete pre-calculated opacity and state equation tables (set by the keyword parameter IOPTAB = \(-2\)).
MODE
: --a specification of the mode of treatment of the given species:\
= 0 --the element is not considered at all;\
= 1 --he element is treated *implicitly*. In this case, the species does not contribute to the opacity; but it is allowed to contribute to the total number of particles and to the total charge; the latter is evaluated assuming LTE ionization balance, i.e., by solving a set of Saha equations.\
= 2 --the element is treated *explicitly*, i.e., selected energy levels of of the selected ionization states are considered explicitly; i.e., their populations are determined by solving the corresponding kinetic equilibrium equations.
ABN
: --a specification of the abundance of the given species:\
= 0--the solar abundance is assumed (Grevesse & Sauval 1998)--see Paper II, Table 1;\
\(<\) 0--a non-solar abundance is assumed, abs(ABN) has now the meaning of the abundance expressed as a multiple of the solar abundance (i.e. \(-0.1\) means 1/10 of solar, \(-5\) means 5 times solar abundance, etc.);\
\(>\) 0--a non-solar abundance is assumed, expressed as \(N({\rm elem})/N({\rm ref})\), i.e. relative by number to the reference species. The reference atom is H by default, but the reference species can be changed by means of the optional parameter IATREF (see Sect. [\[nonst_ese\]](#nonst_ese){reference-type="ref" reference="nonst_ese"})\
\(>10^6\)--non-homogeneous (depth-dependent) abundance is assumed. In this case, the immediately following \(N\!D\) lines should be added that contain the individual values of the abundance (relative to hydrogen by number), for all depth points \(d=1,\ldots,N\!D\).
MODPF
: --a flag indicating a mode of evaluation of the partition functions for the given species. Notice that this may be overwritten by coding NATOMS as negative--see above.\
= 0--a standard evaluation of the partition functions--see Paper II, § 2.7\
\(>\) 0--the partition functions evaluated from the Opacity Project ionization fraction tables.\
\(<\) 0--non-standard evaluation by a user-supplied formula to be implemented in subroutine PFSPEC. For details, refer to § [\[pfspec\]](#pfspec){reference-type="ref" reference="pfspec"}.
## Third block--Explicit ions {#new_ion}
For each ion, including the highest ionization degree of a given species, there must be one input record for each of then relevant ionization stages containing the following parameters:
IATII
: -the atomic number of the parent species of the ion (i.e. 1 for hydrogen, 2 for all ions of helium, etc.).
IZII
: --the charge of the ion (0 for neutrals, 1 for once ionized, etc.).
NLEVS
: --a number of energy levels considered explicitly.
ILAST
: --an indicator whether the given ion is the highest considered ionization degree:\
= 0--the ion is not the highest ion of the parent species; the subsequent input record has to contain parameters for the next higher ion;\
\(>\) 0--the ion is the highest ionization degree of the parent species.\
= 1--the program assigns the correct statistical weight of the ground state of this ion automatically;\
\(\neq\) 1--has the meaning of the statistical weight of the ground state of this ion;\
\(<\) 0--indicates the last record of the explicit ions input block.
ILVLIN
: --an indicator of changing the treatment of a whole group of bound-bound transitions, regardless of the input communicated by the atomic data file-see below. ILVLIN has the meaning that all lines with the relative index of the lower level smaller than ILVLIN are considered in detailed radiative balance and their opacity is neglected. The relative index counts the levels within the ion; i.e. the ground state of the ion has relative index 1, the last considered level the index NLEVS. For instance, setting ILVLIN=2 will put all lines originating from the ground state to the detailed radiative balance, which is often a useful option. Setting ILVLIN \(>\) NLEVS will put all lines of the ion to detailed balance.\
This option enables one to consider the same atomic input files for LTE, NLTE/C (NLTE with continua only), and NLTE/L (NLTE with lines) models--see below.
NONSTD
: --an indicator of an additional input record, to change specific "non-standard" parameters for the ion (those having assigned default values that provide optimum for most applications), or to provide necessary filenames for treating ions for which the superlevel and superline formalism is used (typically the iron-peak elements).\
\(= 0\)--no change of non-standard parameters is required;\
\(> 0\) --additional record with "non-standard" parameters. This is a more or less outdated option; for completeness see § [\[new_ion2\]](#new_ion2){reference-type="ref" reference="new_ion2"};\
\(< 0\) --additional record(s) with filenames for evaluating cross sections for superlines--see below. Moreover, when superlines are treated in the Opacity Sampling option, the actual numerical value of NONSTD defines the formation of superlines, namely:\
\(=-1\) --all internal lines contribute to an appropriate superline;\
\(=-2\)-- all internal lines except autoionizing lines (with upper levels above the ionization potential) contribute; autoionizing lines are neglected;\
\(<-2\)--only lines between observed levels (internal levels with measured energies) are considered.
TYPION
: --a character``4 variable containing a descriptive label, e.g. `'He 2'` for He\(^+\), etc.
FILEI
: --a character variable containing the filename where the detailed input of parameters for explicit level, bound-free transitions, and bound-bound transitions are stored. The structure of this file is described in detail in Chap. [\[ions\]](#ions){reference-type="ref" reference="ions"}.
Note: The number of levels considered for the ion, NLEVS, *must not* exceed the number of levels given in the file `FILEI`. However, NLEVS may be smaller; in such a case the current run will select NLEVS lowest levels from the file `FILEI`.\
If the parameter NONSTD of the standard input (see above) is coded as negative, the program reads an additional record with the following four parameters:
INODF1, INODF2
: --unit numbers for pre-computed data for superline cross sections treated through Opacity Distribution Function (ODF). This is now an outdated option, these two numbers should bet set to zero,\
\({\rm INODF}1={\rm INODF}2=0\).
FIODF1, FIODF2
: --filenames for evaluating superline cross sections:\
--in the standard case of the Opacity Sampling (OS), the first file is the Kurucz level data file (e.g., `gf2601.gam`, for Fe II), and the second file is the line data file (`gf2601.lin`).\
--in the (outdated) case of ODF treatment of lines, the filenames of the corresponding ODF input.
FIBFCS
: --the name of the file containing the photoionization cross sections for the individual superlevels. The parameters IFANCY (see Sect. [\[ion_bf\]](#ion_bf){reference-type="ref" reference="ion_bf"}) has to be set to a value between 50 and 99 to switch on reading cross sections from a previously created file.
# General strategy of model construction {#strateg}
As explained in Paper II, § 3.1 --3.3, any variant of the linearization scheme that is used for constructing model atmospheres is quite sensitive to a quality of the initial estimate. This feature dictates the strategy of constructing models, which often consists in a series of consecutive steps. We shall describe several possible approaches below.
## LTE models {#strateg_lte}
Constructing an LTE model is usually relatively straightforward. The strategy depends on whether an LTE model to be computed is used as a starting solution for a subsequent NLTE model, or whether it is the final result on its own right. In the former case, the model can be quite simple, for instance without taking into account lines. In the latter case, one aims at achieving a much higher degree of complexity of the model, which usually requires more refined modeling techniques. As mentioned earlier, an attractive option to construct an LTE model is using a pre-calculated opacity table. This can in principle be done at any temperature range, but at the moment our opacity table only covers temperatures between 4000 and 10,000 K, so it can be used only for G and K star atmospheres. We plan to extend the opacity table to lower temperatures (say, down to 100 K). An extension to higher temperatures is possible as well, and in fact is relatively easy, but it would not have much practical value because LTE model atmospheres are not accurate enough for higher temperatures (say, for effective temperatures above 15,000 K). There are several possibilities to construct an LTE model atmosphere: (i) The standard way is to start from the scratch; that is, first to construct an initial LTE-gray model, and immediately continue to converge the desired LTE model. As shown in the previous chapter, this is done in one step. Some examples are shown in § [6.2](#examp_hhe){reference-type="ref" reference="examp_hhe"} (a simple H-He model), and § [6.4](#examp_optab){reference-type="ref" reference="examp_optab"} (a complex line-blanketed model, computed using an opacity table). (ii) One can use an existing LTE model, either constructed earlier by tlusty, or taking a Kurucz model atmosphere, as a starting mode atmosphere, and converge the desired LTE model from it. Obviously, the basic input parameters, \(T_{\rm eft}\) and \(\log g\) for the input model must not be very different from \(T_{\rm eff}\) and \(\log g\) of the model to be constructed. A conservative estimate is to change \(T_{\rm eff}\) by no more than 5-10%, and \(\log g\) by no more than 0.2 dex. Otherwise, the model may still converge, but could easily take more computer time than when starting from the scratch. (iii) When constructing a line-blanketed model without using an opacity table, one has to proceed in at least two steps. The first step consists of computing a simple model starting from the scratch (typically without any lines), and the next step includes all the lines. The reason for this procedure is that the adopted treatment of metal line blanketing needs an input starting model in order to construct the appropriate cross sections for superlines. (iv) For complex or difficult models, it is possible to construct the desired model by a sequence of intermediate models. Typically, one first converges a model with a few or no lines considered explicitly, and then one adds lines. This approach is analogous to an often used strategy of constructing NLTE models, which will be described below.
## NLTE models {#strateg_nlte}
A usual strategy, adopted already in the pioneering study of Auer & Mihalas (1969) is to proceed in three steps: LTE \(\rightarrow\) NLTE/C \(\rightarrow\) NLTE/L,\
where NLTE/C represents a NLTE model with continua only; i.e., without lines. In other words, all lines are taken in detailed radiative balance (their radiative rates are set to 0), and their opacity is neglected. NLTE/L denotes a NLTE with lines. We stress again that the LTE model used for this purposes does not have to be a sophisticated LTE model; typically an LTE without lines is sufficient. The reason for adopting this procedure is the following: We know that with increasing depth in the atmosphere departures from LTE generally decrease, so deep enough the state of the atmosphere is well described by an LTE model. The continua (bound-free atomic transitions) are typically less opaque than the bound-bound transitions (lines), so they are formed deeper in the atmosphere. Therefore, including continua only in an NLTE model yields an essentially exact atmospheric structure in deep, continuum-forming, layers. Including lines in the final step then corrects the structure closer to the surface. By this strategy, one gradually improves the structure from deep layers to the surface, which is exactly how the energy in the atmosphere typically flows. (This also hints why a construction of an atmospheric model with a strong external irradiation may be numerically difficult). The above reasoning also explains why it is sometimes advantageous to split the third step into several parts: NLTE/L1 \(\rightarrow\) NLTE/L2 \(\rightarrow\,\ldots\,\rightarrow\) NLTE/L,\
where the individual steps consist of adding more and more lines. We stress that in some cases this is not necessary, and a final NLTE/L model can be constructed in one step. But often this is a good way to proceed. The simplest possibility, always worth trying in case of convergence problems with an NLTE/L model, is first to set a detailed balance in all the resonance lines (those with lower levels being the ground states of the corresponding ions), and then switch them on, perhaps in several steps starting with weaker resonance lines (those of less abundant ions), and ending with the strongest ones. The reason for this strategy is exactly the same as that put forward for the general strategy of improving the atmospheric structure going from deep to upper layers. In some cases, in particular for models where departures from LTE are not crucial, say for late B stars and cooler, and also for hot white dwarfs that contain significant amounts of metals, the following procedure often proves advantageous: LTE/C \(\rightarrow\) LTE/L \(\rightarrow\) NLTE/L,\
that is, constructing first a simple LTE model, than a fully blanketed LTE model (denoted LTE/L), and from it to proceed directly to fully blanketed NLTE/L model with all lines. This strategy reflects the physical fact that when NLTE effects are week, the most important mechanism for determining the atmospheric structure is the metal line blanketing.
## Setting a detailed balance in lines {#strateg_db}
One may set up a detailed radiative balance in a line, and therefore exclude the corresponding transition rate from the kinetic equilibrium equation, and removing the corresponding opacity, in several different ways: (i) individually; in which case one has to modify the corresponding entry in the atomic data file (see § [\[ion_bb\]](#ion_bb){reference-type="ref" reference="ion_bb"}). This is the most flexible way, but also the most cumbersome since one has to modify the atomic data file. (ii) excluding all lines of an atom/ion that originate from a selected set of lower levels. This is driven by the parameter ILVLIN (see § [4.3](#new_ion){reference-type="ref" reference="new_ion"}) for a given atom/ion, which is set in the standard input. The meaning of ILVLIN is that all lines originating at level with index lower than ILVLIN are set to detailed radiative balance. For instance, ILVLIN=2 excludes all resonance lines (i.e., originating from the ground state with index 1); setting ILVLIN=99 (or some other large number) will effectively set all lines of the given atom/ion to detailed balance. (iii) The third way to set lines to detailed balance is based on their frequency. This is driven by keyword parameters FRLMAX and FRLMIN (see § [\[nonst_freq\]](#nonst_freq){reference-type="ref" reference="nonst_freq"}), which put all lines with frequency smaller than FRLMIN and/or larger than FRLMAX to detailed balance. Setting FRLMIN to values around \(10^{13}\) will not only save computer time because it excludes all far-infrared lines that typically do not influence the atmospheric structure, but also remove possible problems with the laser effect in lines, which may lead to numerical problems.
# Examples of standard input; test cases {#new_ex}
## General remarks about running tlusty {#examp_gen}
As already mentioned in Paper I, § 4.4, while the names of the tlusty input and output files can be arbitrary, it is advantageous to use the following convention: Any name of an input or output file is composed of a *core name* that may label the basic physical parameters of a model, with an extension identifying the unit number. For example, let us take a H-He model for \(T_{\rm eff} = 35,000\) K and \(\log g = 4\), in LTE. Let the core name be `hhe35lt`, so the standard input file is then `hhe35lt.5`. This example is considered below in § [6.2](#examp_hhe){reference-type="ref" reference="examp_hhe"}. As is shown in Paper I, § 4.4, tlusty can be run as
tlusty.exe < [std.input_file] > [std. output_file]
but in this case the executable file `tlusty.exe` has to be present in (or liked to) the current directory, and also the necessary atomic data have to be copied or linked to the files specified by the standard input. An easier and in fact a safer way is to use the shell script `RTlusty`, which is also a part of the standard distribution of tlusty. It is called with one or two parameters,
RTlusty model_core_name [core_name_of_starting_model]
The second parameter does not have to be present if the model is calculated from scratch. We stress that the script `RTlusty` is designed to run a model in any directory provided that the two following requirements are fulfilled:
- One sets the environment variable `TLUSTY` that specifies the main tlusty directory. For instance, using the tar files downloaded from the Arizona site[^1], and assuming that the tar file is extracted in the home directory, so the main tlusty directory is generated as `\widetilde{}/tlusty205`, one sets
setenv TLUSTY ~/tlusty205
- The universal directory containing atomic data is a subdirectory of the main tlusty directory, that is \$`TLUSTY/data`
The script `RTlusty` has the following content:\
------------------------------------------------------------------------
------------------------------------------------------------------------
#!/bin/bash # # shell script to run tlusty # # print syntax if no parameters specified # if [ \(#-lt 1 ] then echo Usage: echo RTlusty model core name [core name of starting model] exit fi # # check that the starting model exists # if [ \)#-eq 2 ] then if [-e \(2.7 ]; then echo STARTING MODEL: \)2.7; else echo FILE \(2.7 does not exist, therefore quitting; exit; fi # rm-f fort.8 cp \)2.7 fort.8 fi # # link the "data" directory # ln-s-f \(TLUSTY/data data # # run tlusty # MOD=\)1 \(TLUSTY/tlusty/tlusty.exe < \)MOD.5 > \(MOD.6 # # save important output files # cp fort.7 \)MOD.7 cp fort.9 \(MOD.9 cp fort.69 \)MOD.69 cp fort.13 \(MOD.13 # echo "COMPUTED MODEL \)MOD FINISHED!" date #
------------------------------------------------------------------------
------------------------------------------------------------------------
Let us take an example from the next section, § [6.2](#examp_hhe){reference-type="ref" reference="examp_hhe"}, with the name of the input file being `hhe35lt.5`. Then the code can be either as
tlusty.exe < hhe35lt.5 > hhe35lt.6
where the standard output is redirected to a file `hhe25lt.6` for further inspection; or one can use the script `RTlusty` and to run the test as (this time in the background)
RTlusty hhe35lt &
which runs the code and stores not only the standard output (unit 6) but also other important output files, all with the same core name, and with the suffix that corresponds to the unit number. They are explained in more detail in § [\[out_bas\]](#out_bas){reference-type="ref" reference="out_bas"}. Briefly, `hhe35lt.7` contains a condensed model atmosphere, `hhe35lt.9` the convergence log, `hhe35lt.13` the emergent flux, and `hhe35lt.69` the timing log. There are more output files; if they are intended to be kept for future use, they should be renamed, perhaps with the same core name. After a completed run, it is very important to inspect the convergence log to make sure that all the relative changes of the components of the state vector are sufficiently small. The standard distribution of tlusty also contains an IDL program `pconv.pro` which plots the contents of the convergence log. The program needs three output files, units 7, 9, and 69, and assumes that the file names are constructed using the above convention, that is the core names are the same. The program It is called as
pconv,'hhe35lt'
It can be called also as
pconv
in which case the files are supposed to have generic names `fort.*`. Several examples of plots generated by `pconv` are presented throughout this chapter. It is also important to inspect the last table of the standard output; in the present case `hhe35lt.6`, which summarizes the computed model atmosphere. The last four columns show the computed total flux (radiative plus, if applicable, convective), and the ratios of the radiative, convective, and computed total fluxes with respect to the theoretical total flux, \(\sigma T_{\rm eft}^4\), respectively. Therefore, the values of the last column should be very close to unity (up to at most 1% departures from it); otherwise the model cannot be viewed as sufficiently accurate, even if the convergence log may show that the model is formally converged.
## Simple H-He model atmosphere from scratch {#examp_hhe}
A simple LTE model atmosphere with \(T_{\rm eff}\) = 35,000 K, \(\log g\) = 4, composed of H and He (treated as explicit), and C, N, O (implicit; taken only for particle and charge conservation), is constructed by tlusty by setting the standard input as follows--file `hhe35lt.5`:\
------------------------------------------------------------------------
------------------------------------------------------------------------
35000. 4.0 ! TEFF, GRAV T T ! LTE, LTGRAY '' ! no change of general optional parameters *-----------------------------------------------------------------* frequencies 50 ! NFREAD *-----------------------------------------------------------------* data for atoms * 8 ! NATOMS * mode abn modpf 2 0 0 2 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 1 0 0 *-----------------------------------------------------------------* data for ions * *iat iz nlevs ilast ilvlin nonstd typion filei * 1 0 9 0 100 0 ' H 1' './data/h1.dat' 1 1 1 1 0 0 ' H 2' ' ' 2 0 14 0 100 0 'He 1' './data/he1.dat' 2 1 14 0 100 0 'He 2' './data/he2.dat' 2 2 1 1 0 0 'He 3' ' ' 0 0 0 -1 0 0 ' ' ' ' * * end
------------------------------------------------------------------------
------------------------------------------------------------------------
As mentioned before, one has to make a link from a standard \"data\" directory to `./data` refereed to in the input file. When using the `RTlusty` script, this is done automatically. The files `h1.dat, he1.dat, he2.dat` contain necessary atomic data for H I, He I, and He II, respectively. Structure of these, as well as other atomic data files, is explained in detail in Chap. [\[ions\]](#ions){reference-type="ref" reference="ions"}. Here we assume that the model atoms contain 9, 14, and 14 levels of these ions, respectively. H II and He III are taken as 1-level ions. Notice that there are no additional input files associated with the highest ions, H II and He III, because these ions have no internal structure. Since the `ILVLIN` parameters are set to 100, higher than the number of levels, all lines are set to detailed radiative balance, and their opacity is neglected. This model, and its convergence properties were already shown in Paper 1, § 3.4. The corresponding NLTE/C model is constructed with virtually the same standard input; the only change is replacing the second record by:
F F ! LTE, LTGRAY
The file is stored as `hhe35nc.5`, The model is run, using the script \]`RTlusty` as
RTlusty hhe35nc hhe35lt
where the second parameter, `hhe35lt`, stipulates that the starting model is taken as the result of the previous run, `hhe35lt.7`. The frequency points in both previous models are set up with the default values of the highest frequency, (such as \(\nu_{\rm max}= 10^{11} T_{\rm eff}\)--see § [\[nonst_freq\]](#nonst_freq){reference-type="ref" reference="nonst_freq"}), the lowest frequency \(\nu_{\rm min} = 10^{12}\). The number of frequency points is approximately 50+2\(\times\)(9+14+14)+20 = 144 (i.e. NFREAD + twice the number of levels from which a bound-free transition can occur + about 20 points at the short-and long-wavelength tails). The actual number of points set up by the program is 128 because unnecessary frequency points for the He II edges that coincide with H edges are removed. The final NLTE/L model (NLTE with lines), considering all lines of H and He explicitly, is calculated by modifying the fourth block of the standard input file (stored as `hhe35nl.5`) as follows:\
------------------------------------------------------------------------
------------------------------------------------------------------------
* *iat iz nlevs ilast ilvlin nonstd typion filei * 1 0 9 0 0 0 ' H 1' './data/h1.dat' 1 1 1 1 0 0 ' H 2' ' ' 2 0 14 0 0 0 'He 1' './data/he1.dat' 2 1 14 0 0 0 'He 2' './data/he2.dat' 2 2 1 1 0 0 'He 3' ' ' 0 0 0 -1 0 0 ' ' ' '
------------------------------------------------------------------------
------------------------------------------------------------------------
The frequency points in all cases are set up with the default values of the highest frequency, (such as \(\nu_{\rm max}= 10^{11} T_{\rm eff}\)--see the next Section), the lowest frequency \(\nu_{\rm min} = 10^{12}\). The number of frequency points is approximately 50+2\(\times\)(9+14+14)+20 = 144 (i.e. NFREAD + twice the number of levels from which a bound-free transition can occur + about 20 points at the short-and long-wavelength tails). The actual number of points set up by the program is 128 because unnecessary frequency points for the He II edges that coincide with H edges are removed. The run is generated analogously as before,
RTlusty hhe35nl hhe35nc
The convergence log, displayed using the IDL program `pconv`, is shown in in Fig. 1. As in the analogous Fig. 1 of Paper I, there are six panels on the figure, the upper three display the relative changes of temperature (which is usually the most important component of the state vector), while the lower three panels show the maximum relative change of all components of the state vector. The leftmost panels show the relative change as a function of column mass in the absolute scale, and the middle panels the same in the logarithmic scale. We display both, because the linear scale shows the sign of the relative changes but does not properly show those that are small. These are in turn clearly seen on the logarithmic scale. The rightmost panels show the maximum relative change over all depth points. This is an indicator of the global convergence of the model. Ideally, the relative changes should gradually decrease, which was clearly the case here. The convergence is slower than for the LTE corresponding model, but still stable. The effects of the Ng acceleration are seen in the 7th iteration, and then after each 4 iteration steps. The figure also contains a header that displays the core name and the total execution time. In the present case, it is 23 s, longer than for the LTE model presented in Paper 1, which was less than 1 s. This, and all other calculations reported below, were done on MacBook Pro, with 2.2.GHz Intel Core i7. Figure 2 displays the temperature structure for all the three models. The most interesting feature is a temperature rise at the surface, first discovered by Auer & Mihalas (1969), and explained as an indirect effect of the Lyman \(\alpha\) line on the heating rate in the Lyman continuum. The plot also shows that the temperature structure for both NLTE models approaches that of the LTE model deep in the atmospheres (around \(\log m \approx 0.5\) in the present case), and that the temperature structure for the NLTE/C and NLTE/L models start to differ around \(\log m \approx-1\) where lines become transparent. Figure 3 shows the emergent flux (better speaking, \(H_\nu\)) as a function of frequency, which is given by the first two columns of the file `hhe35nl.13`. We stress that this is the flux produced directly by tlusty, and which is only a schematic representation of the exact emergent flux because lines of only H and He are taken into account. One can construct a detailed spectrum, including all lines, using the associated program synspec. This was shown in Paper I, § 4.7.
## NLTE line-blanketed B star model atmosphere {#examp_bs07}
A complex example is provided for a NLTE line-blanketed B star model atmosphere, with \(T_{\rm eff}\) = 20,000 K, \(\log g\) = 4, with a solar metallicity. This example corresponds to the data used for calculating the bstar2006 grid (Lanz & Hubeny 2007), file `BGA20000g400v2.5`, which is called here `BGA20000g400v2a.5` to stress that one will construct a different model.\
------------------------------------------------------------------------
------------------------------------------------------------------------
20000. 4.0 F F ! LTE, LTGRAY 'nst' ! keyword parameters filename * * frequencies * 2000 * * data for atoms * 30 ! NATOMS * mode abn modpf 2 0. 0 ! H 2 0. 0 ! He 0 0. 0 0 0. 0 0 0. 0 2 0. 0 ! C 2 0. 0 ! N 2 0. 0 ! O 1 0. 0 2 0. 0 ! Ne 1 0. 0 2 0. 0 ! Mg 2 0. 0 ! Al 2 0. 0 ! Si 1 0. 0 2 0. 0 ! S 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 2 0. 0 ! Fe 1 0. 0 1 0. 0 1 0. 0 1 0. 0 * * data for ions * *iat iz nlevs ilast ilvlin nonstd typion filei * 1 0 9 0 0 0 ' H 1' 'data/h1.dat' 1 1 1 1 0 0 ' H 2' ' ' 2 0 24 0 0 0 'He 1' 'data/he1.dat' 2 1 20 0 0 0 'He 2' 'data/he2.dat' 2 2 1 1 0 0 'He 3' ' ' 6 0 40 0 0 0 ' C 1' 'data/c1.dat' 6 1 22 0 0 0 ' C 2' 'data/c2.dat' 6 2 46 0 0 0 ' C 3' 'data/c3_34+12lev.dat' 6 3 25 0 0 0 ' C 4' 'data/c4.dat' 6 4 1 1 0 0 ' C 5' ' ' 7 0 34 0 0 0 ' N 1' 'data/n1.dat' 7 1 42 0 0 0 ' N 2' 'data/n2_32+10lev.dat' 7 2 32 0 0 0 ' N 3' 'data/n3.dat' 7 3 48 0 0 0 ' N 4' 'data/n4_34+14lev.dat' 7 4 16 0 0 0 ' N 5' 'data/n5.dat' 7 5 1 1 0 0 ' N 6' ' ' 8 0 33 0 0 0 ' O 1' 'data/o1_23+10lev.dat' 8 1 48 0 0 0 ' O 2' 'data/o2_36+12lev.dat' 8 2 41 0 0 0 ' O 3' 'data/o3_28+13lev.dat' 8 3 39 0 0 0 ' O 4' 'data/o4.dat' 8 4 6 0 0 0 ' O 5' 'data/o5.dat' 8 5 1 1 0 0 ' O 6' ' ' 10 0 35 0 0 0 'Ne 1' 'data/ne1_23+12lev.dat' 10 1 32 0 0 0 'Ne 2' 'data/ne2_23+9lev.dat' 10 2 34 0 0 0 'Ne 3' 'data/ne3_22+12lev.dat' 10 3 12 0 0 0 'Ne 4' 'data/ne4.dat' 10 4 1 1 0 0 'Ne 5' ' ' 12 1 25 0 0 0 'Mg 2' 'data/mg2.dat' 12 2 1 1 0 0 'Mg 3' ' ' 13 1 29 0 0 0 'Al 2' 'data/al2_20+9lev.dat' 13 2 23 0 0 0 'Al 3' 'data/al3_19+4lev.dat' 13 3 1 1 0 0 'Al 4' ' ' 14 1 40 0 0 0 'Si 2' 'data/si2_36+4lev.dat' 14 2 30 0 0 0 'Si 3' 'data/si3.dat' 14 3 23 0 0 0 'Si 4' 'data/si4.dat' 14 4 1 1 0 0 'Si 5' ' ' 16 1 33 0 0 0 ' S 2' 'data/s2_23+10lev.dat' 16 2 41 0 0 0 ' S 3' 'data/s3_29+12lev.dat' 16 3 38 0 0 0 ' S 4' 'data/s4_33+5lev.dat' 16 4 25 0 0 0 ' S 5' 'data/s5_20+5lev.dat' 16 5 1 1 0 0 ' S 6' ' ' 26 1 36 0 0 -1 'Fe 2' 'data/fe2va.dat' 0 0 'data/gf2601.gam' 'data/gf2601.lin' 'data/fe2p_14+11lev.rap' 26 2 50 0 0 -1 'Fe 3' 'data/fe3va.dat' 0 0 'data/gf2602.gam' 'data/gf2602.lin' 'data/fe3p_22+7lev.rap' 26 3 43 0 0 -1 'Fe 4' 'data/fe4va.dat' 0 0 'data/gf2603.gam' 'data/gf2603.lin' 'data/fe4p_21+11lev.rap' 26 4 42 0 0 -1 'Fe 5' 'data/fe5va.dat' 0 0 'data/gf2604.gam' 'data/gf2604.lin' 'data/fe5p_19+11lev.rap' 26 5 1 1 0 0 'Fe 6' ' ' 0 0 0 -1 0 0 ' ' ' ' * * end
------------------------------------------------------------------------
------------------------------------------------------------------------
File `nst`, described in detail in § [\[nst_exa_bs07\]](#nst_exa_bs07){reference-type="ref" reference="nst_exa_bs07"}, contains the necessary keyword parameters:
ND=50,NLAMBD=3,VTB=2.,ISPODF=1,DDNU=50.,CNU1=6.,NITER=0
For the testing purposes, this model is not identical to the corresponding model from thebstar2006 grid. It is calculated with a lower sampling step in frequency, 50 fiducial Doppler widths, and no iteration (`NITER=0`) of the hybrid CL/ALI scheme is done. This run therefore only initializes the superline cross sections and performs a global formal solution (i.e., a simultaneous solution of the radiative transfer and the kinetic equilibrium equations to obtain new values of mean intensities of radiation and atomic level populations), so it tests most of the features specific to NLTE metal line blanketed models. The user can obviously increase the number of iterations by skipping the `NITER=0` statement, in which case the run would however take several hours. The first 30 elements, except Li, Be, and B, are included, all with the solar abundances. The model atmosphere includes 11 explicit atoms (H, He, C, N, O, Ne, Mg, Al, Si, S, and Fe), 35 explicit ions, and 1127 explicit NLTE (super)levels. Opacity Sampling is used (ISPODF=1). Unlike the models from the bstar2006 grid that were calculated with a sampling step of 0.75 fiducial Doppler width for iron, here we use for simplicity a lower resolution of 50 fiducial Doppler widths. The whole spectrum is represented by 123,396 frequencies. The mean intensity of radiation is linearized at 16 frequencies. Again, a subdirectory `"./data"`, in which all the atomic data files referred to in the standard input are located, must exist. When using the standard directory tree and the script `RTlusty`, the link to the standard \"`data`\" directory is done automatically. The model is then run simply as
RTlusty BGA20000g400v2a BGA20000g400v2
where `BGA2000g400v2a` is the core name of the new model, and the second parameter, `BG20000g400v2` sets the file `BG20000g400v2.7` as a starting model for the present run. Here we also demonstrate the use of the program pretlus. It is run as
cd \(TLUSTY/examples/bstar ls-s-f \)TLUSTY/data data $TLUSTY/pretlus/pretlus.exe <BGA20000g40v2a.5
The standard output from the program is the following:\
------------------------------------------------------------------------
------------------------------------------------------------------------
MATOM = 99 MION = 35 MLEVEL = 1127 MLVEXP = 222 MTRANS = 20476 MDEPTH = 50 MFREQ = 123396 MFREQP = 215637 MFREQC = 4015 MFREX = 16 MFREQL = 4935 MTOT = 241 MMU = 3 MFIT = 157 MITJ = 199 MMCDW = 6 MMER = 1 MVOIGT = 79 MZZ = 6 NLMX = 80 MSMX = 1 MFREQ1 = 1 MFRTAB = 0 MTABT = 0 MTABR = 0 MCROSS = 1132 MBF = 1127 MDEPTC = 0 MMUC = 0 parameters in ODFPAR.FOR: MDODF = 3 MKULEV = 6870 MLINE = 1133840 MCFE = 7823728
------------------------------------------------------------------------
------------------------------------------------------------------------
This is to be compared to the content of the files `BASICS.FOR` and `ODFPAR.FOR` to check that the actual values of the parameters are greater than or equal to the values listed above. If not, the files `BASICS.FOR` and/or `ODFPAR.FOR` have to be modified accordingly.
## LTE model atmosphere of a K star {#examp_optab}
Here is an example of using the pre-calculated opacity table, while solving the equation of state on the fly. This example also shows how to deal with convection. The standard input file, called `s4545.5` is as follows:\
------------------------------------------------------------------------
------------------------------------------------------------------------
4500. 4.5 ! TEFF, GRAV T T ! LTE, LTGRAY 's45.param' ! keyword parameters filename * * frequencies * -15000 * * data for atoms * 30 ! NATOMS * mode abn modpf 1 0. 0 ! H 1 0. 0 ! He 0 0. 0 0 0. 0 0 0. 0 1 0. 0 ! C 1 0. 0 ! N 1 0. 0 ! O 1 0. 0 1 0. 0 ! Ne 1 0. 0 1 0. 0 ! Mg 1 0. 0 ! Al 1 0. 0 ! Si 1 0. 0 1 0. 0 ! S 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 1 0. 0 ! Fe 1 0. 0 1 0. 0 ! Ni 1 0. 0 1 0. 0 * * data for ions * *iat iz nlevs ilast ilvlin nonstd typion filei * 0 0 0 -1 0 0 ' ' ' '
------------------------------------------------------------------------
------------------------------------------------------------------------
The file `s45.param`, described in detail in § [\[nst_exa_optab\]](#nst_exa_optab){reference-type="ref" reference="nst_exa_optab"}, contains the necessary keyword parameters, most importantly `IOPTAB=-1` which sets the mode of evaluation to the pre-calculated opacity table, but solves the equation of state and evaluates the thermodynamic parameters on the fly. The file looks like this:
IOPTAB=-1,IFRYB=1,IFMOL=1,IDLST=0 FRCMAX=3.2e15,FRCMIN=1.5e13,IBINOP=0, ITEK=50,IACC=50 TAUFIR=1.e-7,TAULAS=1.0e2 HMIX0=1,NDCGAP=10,ICONRE=0,IDEEPC=3,CRFLIM=-10
The frequency number parameter NFREAD is set to \(-\)`<!-- -->`{=html}15000, which sets logarithmically equidistant frequencies between FRCMIN and FRCMAX, which are set in the `s45.param` file. The first 30 elements, except Li, Be, and B, are considered in the equation of state with solar abundances. In this mode, there are no explicit atoms and ions; the input block for ions thus contains only the final line indicating its end. Due to the specification of the keyword parameters, several additional input files are needed:
Using the standard tlusty directory tree, these files are readily available. The convergence of the model is excellent--see Fig. 4. It can hardly be ever better; this is essentially thanks to the use of the Rybicki scheme which linearizes the mean intensities of radiation in all 15,000 frequency points, so the adopted iteration scheme is essentially an exact Newton-Raphson method. The execution was also quite fast; the complete model was calculated in 60 s on a 2.2 GHz Mac. Actually, we used here a binary form of the opacity table; its reading took 20 out of 60 s. When using an ASCII table, the time for its reading increases to about 50s. In Fig. 5, we plot the temperature structure, together with the temperature structure of the Kurucz model with the same \(T_{\rm eff}\) and \(\log g\). Overall temperature structure is quite similar, but there are small differences, in particular close to the surface. They are most likely caused by our neglecting the contribution of molecules in the opacity table. We stress that the current opacity table is not meant to provide the exact opacity, but rather to be used for testing purposes. Figure 6 shows an analogous comparison of the emergent flux. Again, an agreement between the two models is quite good, in particular in view of the fact that our opacity table is just a beta-version of a more complete opacity table that is currently under construction.
## Moderately cool DA white dwarf {#examp_cwd}
We take an example of a DA (pure-hydrogen) white dwarf with \(T_{\rm eff}=10,000\) K and \(\log g =8\). The atmosphere is convective; we assume the ML2-type convection with the mixing length equal to 0.6 pressure scale height. Hydrogen line broadening is treated using the Tremblay tables (see § [\[nst2_hyd\]](#nst2_hyd){reference-type="ref" reference="nst2_hyd"}). The standard input for an LTE model is as follows (file `t100g80l.5`):\
------------------------------------------------------------------------
------------------------------------------------------------------------
10000.0 8.0 T T 'cwd.flag' 200 * * data for atoms * 1 * mode abn modpf 2 0 0 * *iat iz nlv ilst ilvln nonst typion filei * 1 0 16 0 0 0 ' H 1' './data/h1s16.dat' 1 1 1 1 0 0 ' H 2' ' ' 0 0 0 -1 0 0 ' ' ' '
------------------------------------------------------------------------
------------------------------------------------------------------------
[^1]: http://aegis.as.arizona.edu/\(\,\,\widetilde{\ }\,\)hubeny/pub/tlusty205.tar.gz | {'timestamp': '2017-06-08T02:01:29', 'yymm': '1706', 'arxiv_id': '1706.01937', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01937'} |
# Introduction
In hand with observational progress, the understanding of high redshift galaxies has progressed immensely in the last few decades. Almost 20 years ago, the only efficient way of observing these galaxies was by detecting breaks in their spectra with broadband photometry. For example, the two main classes of rest-frame UV-selected galaxies are Lyman Break Galaxies (LBGs; @steidel1996 [@steidel1999; @steidel2003; @shapley2003; @stark2009; @bouwens2015]) and Lyman-alpha Emitters (LAEs; @cowie1996 [@cowiehu1998; @hu1998; @kashikawa2006; @shimasaku2006; @gronwall2007; @ouchi2008; @blanc2011; @sobral2015]). While the former are observed by their Lyman break at 912 Å, the latter are detected by their Lyman alpha (Ly\(\alpha\)) emission at 1216 Å, which is produced in star forming regions and is subject to resonant scattering in the neutral hydrogen of the ISM. After being, at first, used as a galaxy detection method, Ly\(\alpha\) emission from galaxies is now widely used to derive a wide variety of information about the physical properties of galaxies and the universe as a whole. A lot of effort has been dedicated to study the process of Ly\(\alpha\) emission within galaxies. Radiative transfer analysis focuses on the dependence of Ly\(\alpha\) escape on ISM kinematics, clumpiness, and outflows (@dijkstra2012 [@duval2014; @rivera-thorsen2015; @gronke2016; @gronke2016b]). These studies yield elementary insights that, when combined with observations of the Ly\(\alpha\) escape fraction (\(f_{esc}\); @hayes2011 [@ciardullo2014]), provide a coherent picture for Ly\(\alpha\) emission. Still, despite the complex modeling required to explain this particular type of radiation at small scales, Ly\(\alpha\) is actively used as a technique for solving cosmological scale paradigms. Among other things, Ly\(\alpha\) emission is being used to study the properties of neutral hydrogen in the interstellar, circumgalactic and intergalactic mediums. For instance, the fraction of LAEs as a function of redshift is used as a proxy for the fraction of neutral gas in the IGM (@stark2011 [@ono2012; @tilvi2014]). If successful, this insight can be used to trace the epoch of reionization, which we have been able to constrain using QSO sightlines (@fan2006) and cosmic microwave background (CMB) measurements (@bennett2013 [@planck2016a; @planck2016b]). Moreover, with the hope of tracing the sources responsible for reionization, Ly\(\alpha\) emission is also conveniently used for characterizing luminosity functions (LFs) at high redshift (@ouchi2008 [@sobral2009; @dressler2015; @santos2016]). Nevertheless, the cosmological questions that we are trying to answer by means of this complex emission are not restricted to the reionization of the universe. For instance, Ly\(\alpha\) emission can be used to trace the cosmic web (@gould1996 [@cantalupo2005; @cantalupo2012; @cantalupo2014; @kollmeier2010]) and young, metal-poor stellar populations in the early universe (@sobral2015). Even more, Ly\(\alpha\) emission will also be used to study dark energy through baryonic acoustic oscillations in the clustering of LAEs (@adams2011 [@blanc2011]). As evidenced by small-scale studies, Ly\(\alpha\) emission is highly stochastic and complex. The observed equivalent width (\(W_{Ly\alpha}\)) of Ly\(\alpha\) lines is highly dependent on neutral gas opacity and kinematics, clumpiness of the ISM, dust distribution, and, therefore, line of sight toward the observer. Furthermore, typical Ly\(\alpha\) emitters are extreme objects in terms of their luminosity, metallicity, and emission line diagnostics (@trainor2016). As a consequence, if we are to use Ly\(\alpha\) emission as a tracer, a careful and thorough approach is required. However, limitations in survey design make such an approach difficult. For instance, spectroscopic studies of UV continuum detected galaxies, typically LBGs (e.g. @shapley2003 [@stark2010; @ono2012]), reject galaxies with no significant Lyman break and/or red rest-frame UV spectral energy distributions (SEDs), i.e., these studies do not account for passive and heavily extincted galaxies. Observationally, this technique is limited by the M\(_{UV}\) sensitivity of the survey, which translates to incompleteness at low-SFR. On the other hand, samples where galaxies are detected directly in Ly\(\alpha\) using narrowband imaging (e.g. @gronwall2007 [@ouchi2008]) or blind spectroscopy (e.g. @blanc2011 [@dressler2011]) select on Ly\(\alpha\) equivalent width (W\(_{Ly\alpha}\)) and emission line flux. Such methodology does not require continuum selection, which allows for line detection in faint objects. However, follow-up rarely allows for more than spectroscopic confirmation of the line, limiting the use of this technique to the measurement of line fluxes, \(W_{Ly\alpha}\), and emission profiles. Even more, such surveys can have non-negligible amounts of line interlopers (@sobral2016). Comprehensive studies of Ly\(\alpha\) emission in the overall galaxy population are, therefore, of crucial importance. Correlations between galaxy properties and emission are the statistical manifestation of the radiative escape of Ly\(\alpha\) photons. Properties such as stellar mass (M\(_*\)), star formation rate (SFR), UV slope (\(\beta\)), and merger rate reveal information of how the escape of Ly\(\alpha\) radiation is affected by stellar population ages, gas fraction, dust content, and ISM turbulence, respectively. Acknowledgment of such relations is required to draw conclusions from high-redshift Ly\(\alpha\) surveys. show how the normalization and e-folding scale of the W\(_{Ly\alpha}\) distribution of \(3<z<4.6\) galaxies anti-correlate with M\(_*\). In other words, higher M\(_*\) objects typically have lower W\(_{Ly\alpha}\). More massive galaxies typically have lower gas fractions, but higher gas mass (@tacconi2013 [@song2014; @genzel2015]). More neutral gas contributes to increase the scatter of Ly\(\alpha\) photons, spreading this radiation throughout the galaxy and toward the circumgalactic medium. Higher M\(_*\) galaxies also have more dust extinction (@franx2003 [@daddi2004; @daddi2007; @forster-schreiber2004; @perez-gonzalez2008]), which leads to more Ly\(\alpha\) photon absorption. The escape fraction of this radiation is, therefore, severely affected by the stellar mass of galaxies. These results even explain some discrepancies between LBG and narrowband studies of Ly\(\alpha\) emission that originated in sample selection effects. This example highlights the need for proper assessment. In this work, we present a comprehensive analysis of Ly\(\alpha\) emission in the \(3<z<4.6\) galaxy population. To this end, we further analyze the data introduced by. This sample is composed of 625 galaxies in the M\(_*\) range \(7.6<\log{(M_*[M_{\odot}])}<10.6\) from the 3D-HST/CANDELS survey (@grogin2011 [@koekemoer2011]). Our work is based on spectroscopic observations of the sample using the Michigan/Magellan Fiber System (M2FS; @mateo2012), allowing us to measure Ly\(\alpha\) fluxes and use 3D-HST/CANDELS ancillary data. In particular, we study Ly\(\alpha\) emission dependence on the M\(_*\), SFR, UV luminosity, and dust extinction of galaxies. To do so, we introduce a Bayesian approach to properly compare \(W_{Ly\alpha}\) distribution models, account for incompleteness, and quantify the significance of observed correlations. We also use our results to simulate high-redshift Ly\(\alpha\) emission surveys, allowing us to infer their biases and imprints on the resulting samples. We finally use the correlations we recover to predict the fraction of LAEs as a function of redshift. This semi-analytic approach provides a baseline for comparing observational drops in the fraction of LAEs at high redshift. This work is structured as follows. In Section [2](#2){reference-type="ref" reference="2"}, we describe our sample and data set. In Section [3](#3){reference-type="ref" reference="3"}, we describe our line detection and measurement methodologies. We explain our Bayesian \(W_{Ly\alpha}\) distribution analysis in Section [4](#4){reference-type="ref" reference="4"}. We show our results on the \(W_{Ly\alpha}\) distribution dependence on different properties and selection techniques in Sections [5](#5){reference-type="ref" reference="5"} and [6](#6){reference-type="ref" reference="6"}. We state in Section [7](#7){reference-type="ref" reference="7"} our inferences on higher redshift Ly\(\alpha\) emission. We present our conclusions in Section [8](#8){reference-type="ref" reference="8"}. Throughout this work, all magnitudes are in the AB system (@oke1983). A CDM cosmology with \(H_{0}\) = 70 km s\(^{-1}\) Mpc\(^{-1}\), \(\Omega_{m}\) = 0.3, and \(\Omega_{\Lambda}\) = 0.7 was assumed whenever needed.
# Data Set {#2}
## Sample Selection {#2.1}
Our sample was initially composed of 629 galaxies in the COSMOS, GOODS-S and UDS fields. Every object was observed under the 3D-HST/CANDELS program, providing HST and *Spitzer* photometry from 3800 Å to 7.9 \(\mu m\) (9 bands for COSMOS, 14 for GOODS-S, and 9 for UDS). We construct our sample using 3D-HST outputs (@skelton2014). According to these, our 629 photometric redshifts satisfy \(3.25<z_{3D-HST}<4.25\) and have a 95% probability of \(2.9<z<4.25\). Every galaxy also complies with a photometric redshift reliability parameter \(Q_{z}\leqslant3\) selection to remove catastrophic outliers (@brammer2008). In terms of M\(_*\), our galaxies are homogeneously distributed in the range \(8<\log({\mbox{M}_*[\mbox{M}_{\odot}]})_{3D-HST}<10.4\). These 3D-HST output values are based on the stellar population synthesis model library with a IMF and solar metallicity. Exponentially declining star formation histories (SFHs) with a minimum e-folding time of \(log_{10}(\tau/yr) = 7\) and a dust attenuation law were also assumed for the calculation (@skelton2014). We further emphasize that sample selection was performed based on the values of \(z_{3D-HST}\) and M\(_*\) detailed in this section, whereas all further analysis in this paper is based on our own estimates (Section [2.3](#2.3){reference-type="ref" reference="2.3"}).
## Data {#2.2}
Spectroscopy of the full sample was conducted at the Magellan Clay 6.5 m telescope during 2014 December and 2015 February. To this end, we used the M2FS, a multi-object fiber-fed spectrograph. The \(1\farcs2\) fibers of this instrument allow the observation of 256 targets within 30 arcmin in a single exposure. We are then able to observe 210 targets in each field simultaneously, while using 40 fibers for sky apertures and five for calibration stars. We used this spectrograph in LoRes mode, which features an expected resolution of R=2000 and a continuum sensitivity of V=24 with S/N=5 in 2 hr (@mateo2012). The final data set consists of six exposure hours on each of the three fields with an average seeing of \(0\farcs6\). For data reduction, we developed a custom M2FS pipeline. This routine features standard bias subtraction and dark correction. For wavelength calibration, we use HgArNe lamps observed on each night. The wavelength solution is obtained separately for each fiber, with a typical rms uncertainty of 0.03Å. We further correct the solution for each fiber using the sky lines in the science spectra. For flat-fielding, we use sky-flats obtained during either twilight or dawn. The correction is calculated separately for each fiber, and features illumination, fiber profile, and a correction to account for the shifting of fiber spectra on the detector due to thermal effects. Despite the fact that we are not using dome-flats, we estimate the uncertainties in our flat-fielding to be \(<5\)%. Sky subtraction is performed using the 40 sky fibers homogeneously distributed over the field of view and across the detector, where fibers are grouped in blocks. We find the sky solution to be more dependent on fiber location in the CCD than in the sky, leading us to perform sky subtraction separately for each frame fiber block. The solution is computed using a non-parametric spline fit, yielding satisfactory sky subtraction for most sky lines. For the bright sky lines that we are not able to properly subtract, we build emission line masks. For consistency, we use the same mask for all spectra, except for particularly noisy fibers. In those rare cases (11), we mask broad sections of the spectrum. Due to fiber malfunction, we could not obtain spectra for 4 of the 629 targets, leaving our final sample in 625 objects. Flux calibration is performed using five M\(_V=19-22\) calibration stars on each exposure. Due to atmospheric differential refraction (ADR) comparable to the \(1\farcs2\) fiber size, we need to recover the intrinsic spectrum of each star. To do so, we fit stellar templates from the Pickles library (@pickles1998) to the continuum-normalized instrumental spectra of the calibration stars. We then use the five stars on each exposure to obtain an average sensitivity curve. We estimate the rms uncertainty of our method to be about \(\sim 15\%\). After this calibration, we correct the fluxes in our spectra for Galactic extinction. Samples of sky-subtracted spectra are shown in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. Considering the three fields, the science area we survey is of \(\sim 550\) arcmin\(^2\). The resulting spectral FWHM line resolution is of \(\sim 2\)Å, and we reach a 1\(\sigma\) continuum flux density limit of \(\sim 4\times10^{-19}\) erg s\(^{-1}\) cm\(^{-2}\) Å\(^{-1}\) per pixel in our 6 hr of exposure. We estimate a 5\(\sigma\) emission line-flux sensitivity of \(\sim 8\times10^{-18}\) erg s\(^{-1}\) cm\(^{-2}\) in our final spectra (see Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}). We identify 120 Ly\(\alpha\) emission lines with S/N\(\geqslant 5.5\) in our data (details in Section [3.1](#3.1){reference-type="ref" reference="3.1"}). Given that we have 625 observed targets, then we are then recovering a Ly\(\alpha\)-emitting galaxy with S/N\(\geqslant 5.5\) every \(\sim\)`<!-- -->`{=html}5 objects.
## Sample Properties {#2.3}
As stated by, 3D-HST outputs for these galaxies were obtained using FAST (@kriek2009). These calculations assume exponentially declining SFHs with a minimum e-folding time of \(log_{10}(\tau/yr) = 7\) (@skelton2014). However, since recent studies suggest that it is more adequate to reproduce high-redshift observations with constant SFHs (cSFHs; @gonzalez2014), or even rising SFHs (@maraston2010), we perform our own executions of FAST assuming cSFHs. For a detailed discussion on this topic, we refer the reader to. Our FAST executions also adopt the stellar population synthesis model library, a IMF, and solar metallicity, similarly to. We do not account for nebular emission lines in our SED fitting, which, in principle, can overestimate our reported M\(_*\) by a factor of up to 4 (@atek2011 [@conroy2013; @stark2013]) or even \(\sim7\) for strong emitters (@deBarros2014). However, our galaxies are at \(z\sim4\). At this redshift, the overestimate is within a factor of \(\lesssim 1.1\) (@stark2013 [@salmon2015]), since the 4.5\(\mu m\) *Spitzer*/IRAC band is unaffected[^1]. Our FAST outputs also include extinction-corrected SFRs. The SFRs are derived from the SED, and the extinction correction assumes a attenuation law with \(R_{V}=4.05\). We are also required to run EAZY (@brammer2008) on the photometry from CANDELS/IRAC, since FAST requires redshifts as input. The executions of EAZY yield a most probable redshift \(z_{EAZY}\). Our 625 objects satisfy \(3<z_{EAZY}<4.25\), with a median uncertainty \(\sigma_{EAZY}=0.1\) (Figure [\[fig:2_5\]](#fig:2_5){reference-type="ref" reference="fig:2_5"}). EAZY outputs also include \(2\sigma\) constraints, which for our sample are limited to \(2.95<z<4.5\). Our FAST executions yield a mass coverage of \(7.6<\log{(\mbox{M}_*[\mbox{M}_{\odot}])}<10.6\) (Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}), with a characteristic uncertainty of \(\log{(\mbox{M}_*[\mbox{M}_{\odot}])}\sim 0.3\). Most of our SFR values are in the range \(1-100\mbox{M}_{\odot}/yr\) (further analysis and plots below). From now on, we use the FAST outputs based on the spectroscopic redshifts (\(z_{Ly\alpha}\)) for the 120 detections (S/N\(\geqslant 5.5\); Section [3.1](#3.1){reference-type="ref" reference="3.1"}) and the outputs based on \(z_{EAZY}\) for the 505 non-detections.
## Considerations {#2.4}
The flux calibration procedure performed in our spectra is based on using stars; therefore, it corrects for a \(\sim32\)% fiber flux loss, which corresponds to a point-source Ly\(\alpha\) surface brightness distribution. In cases of extended Ly\(\alpha\) emission halos (@steidel2011 [@matsuda2012; @feldmeier2013; @hayes2013; @momose2014; @caminha2016; @patricio2016; @wisotzki2016]), the fluxes we derive are mostly associated with the galaxies themselves and the inner parts of their Ly\(\alpha\) halos. Furthermore, Ly\(\alpha\) emission from a galaxy can show significant misalignment with the UV continuum (@rauch2011), although such cases are not the norm (@finkelstein2011 [@jiang2013b]). At the cosmic time of our sample, the fiber diameter corresponds to a scale of \(\sim8\) kpc, roughly a factor of four larger than the typical effective diameter of galaxies (@bond2012 [@law2012]). We find a median redshift offset of \(\Delta z= z_{Ly\alpha}-z_{EAZY}=0.24\) in our detections (see Figure [\[fig:2_5\]](#fig:2_5){reference-type="ref" reference="fig:2_5"}). We show in that this offset does not have any noticeable effects on our sample dependence on M\(_*\), which is the primary selection criterion for our galaxies. A very similar offset has also been found in the MUSE-Wide Survey (see @herenz2017). We find the offset to correlate with \(W_{Ly\alpha}\) (see Figure [\[fig:2_5\]](#fig:2_5){reference-type="ref" reference="fig:2_5"}), hinting at biases in photometric redshift fitting when a strong Ly\(\alpha\) emission line is present. A plausible scenario is one in which the Lyman break from the template is fitted slightly blueshifted to account for the flux excess in the redder band. Thorough simulations beyond the scope of this work are required to explore the causes behind this bias. We stress that the 3D-HST/CANDELS mass incompleteness is restricted to \(\log{(\mbox{M}_*[\mbox{M}_{\odot}])}<8.5\) at \(z\sim 4\) (at least for GOODS-S; @duncan2014), which corresponds to about one-quarter of the sample. We want to stress that this is a homogeneously M\(_*\)-selected sample, designed to study Ly\(\alpha\) emission statistics dependence on galaxy properties. As a consequence, it is by no means representative of the M\(_*\) or Ly\(\alpha\) LFs at \(3<z<4.6\). This must be taken into account when comparing this sample to analogues directly drawn from the galaxy population (i.e. LBGs or narrowband samples).
# Ly\(\alpha\) Measurements {#3}
## Line Detection {#3.1}
For line detection, we use an automated maximum likelihood fitting routine after continuum subtraction. We assume intrinsic Gaussian profiles of the form \[\begin{gathered}
\label{gaussian_line} f_{rest}(\lambda)=\frac{F}{\sqrt{2\pi \sigma_{\lambda}^{2}}} \mbox{exp} \left[-\frac{(\lambda-\lambda_{0})^{2}}{2\sigma_{\lambda}^{2}}\right] \end{gathered}\] where \(F\), \(\lambda_{0}\), and \(\sigma_{\lambda}\) compose the parameter space explored by the maximum likelihood. In order to account for false positives, we run the line detection routine on the 115 sky fibers. The results are shown in Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}. We detect four lines above 4\(\sigma\) and none above 5\(\sigma\). Therefore, down to 5\(\sigma\), we expect at most two false detections in our \(120\) lines, translating into \(\lesssim 2\%\) contamination using signal-to-noise S/N\(^{*}=5.5\) as our threshold. We also characterize our detection completeness (see Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}). To obtain it, we use \(p(\)S/N\(_{i}>\)S/N\(^{*}|\)S/N), with S/N\(_{i}\) the measured signal-to-noise ratio and S/N\(^{*}\) the imposed detection threshold. We define the simulated signal to noise as S/N\(_{sim}\)=F\(/\sqrt{\sum d\lambda^2 \sigma_k^2}\), with \(d\lambda\) the wavelength dispersion in the spectrum and \(\sigma_k\) the flux uncertainty for pixel *k*. We find the most accurate representation by summing over an interval of \(20\)Å centered at 5500 Å. To recover the actual completeness shown in Figure [\[fig:3\]](#fig:3){reference-type="ref" reference="fig:3"}, we simulate \(\sim 10^3\) lines on the 115 sky-spectra sampling fluxes of \(10^{-19}-10^{-17}\) ergs s\(^{-1}\) cm\(^{-2}\), FWHMs between \(5-13\) Å, and wavelengths of 4800-6700 Å. The fraction for which we measure S/N\(_{i}>\)S/N\(^{*}\) is our completeness. We use S/N\(^{*}=5.5\) as our detection threshold, which corresponds to a line-flux sensitivity in the final spectra of \(\sim 1\times10^{-17}\) erg s\(^{-1}\) cm\(^{-2}\) (see Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}).
## Line Profiles {#3.2}
The radiative transfer and escape of Ly\(\alpha\) radiation from galaxies can be highly complicated. As a matter of fact, the resonant nature of this line has led to thorough modeling of its radiative escape (e.g. @verhamme2006 [@dijkstra2012]). Such complications imply that the flux profile of a Ly\(\alpha\) line is not always well reproduced by the usual Gaussian profile (@chonis2013 [@trainor2015]). Hence, for flux measurements, we adopt a more sophisticated model. Similarly to and, we fit double-peaked Gaussian profiles of the form \[\begin{gathered}
\label{double-peak} f_{rest}(\lambda)= f_{blue}(\lambda) + f_{red}(\lambda) \end{gathered}\] where \(f_{blue}(\lambda)\) represents the blue emission component and \(f_{red}(\lambda)\) represents the red component. We assume each of these components to be asymmetric, i.e., they follow Equation ([\[gaussian_line\]](#gaussian_line){reference-type="ref" reference="gaussian_line"}), with \(\sigma\) defined as \[\begin{gathered}
\label{sigma} \sigma_\lambda \equiv \sigma_{b} \mbox{ if } \lambda < \lambda_0 \\ \nonumber \sigma_\lambda \equiv \sigma_{r} \mbox{ if } \lambda > \lambda_0 \end{gathered}\] Before fitting, we convolve the profiles given by Equations ([\[double-peak\]](#double-peak){reference-type="ref" reference="double-peak"}) and ([\[sigma\]](#sigma){reference-type="ref" reference="sigma"}) with the spectral resolution. This allows us to properly characterize the errors in our measurements and methodology. In case that there is some sky contamination (which happens for 28 of the 120 detections), we only fit single-peaked profiles. The resulting fits to six of our emission lines are shown in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}.
## Ly\(\alpha\) Equivalent Width and Escape Fraction {#3.3}
There are typically two diagnostics used to characterize the prominence of Ly\(\alpha\) emission in galaxies: the equivalent width (\(W_{Ly\alpha}\)) and the Ly\(\alpha\) escape fraction (\(f_{esc}\)). The rest-frame \(W_{Ly\alpha}\) is defined as the fraction between line flux and UV continuum flux in the rest frame of the galaxy. Explicitly, \[\label{W_{Ly_alpha}} W_{Ly\alpha}=\frac{F}{f_{\lambda}}\frac{1}{(1+z_{Ly\alpha})}\] with \(F\) the Ly\(\alpha\) flux we measure in the spectra and \(f_{\lambda}\) the observed flux at rest frame \(1700\)Å from 3D-HST rest-frame colors (@skelton2014). It must be noted that 3D-HST rest-frame colors come from the best-fit template to the photometry, i.e., they are not direct measurements. Still, this allows us to use, in principle, the same rest-frame wavelength for every object, regardless of its actual redshift. In addition, we can also derive a value of \(f_{\lambda}\) even if the rest frame \(1700\)Å photometry is missing. For reference, we find that the templates to match the photometry at \(\sim 1700\)Å of every object to a typical agreement of \(\sim 90\%\). For the uncertainties on \(f_{\lambda}\), however, we use the errors on the photometry. We do not use any \(f_{\lambda}\) directly measured from our data, since every galaxy has a continuum fainter or comparable to the 1\(\sigma\) errors in the spectra. On the other hand, \(f_{esc}\) is the fraction of the number of Ly\(\alpha\) photons that escape the galaxy from the number produced. This diagnostic is typically indirectly recovered using SFRs derived from the Ly\(\alpha\) line and intrinsic SFRs. The latter are typically calculated from UV continuum measurements or H\(\alpha\) fluxes, subject to extinction correction. In our case, we use the intrinsic (i.e., extinction-corrected) SFRs from FAST. The observed SFRs are derived from the SED, whereas the extinction corrections assume a attenuation law with \(R_{V}=4.05\). The explicit definition is then (e.g. @blanc2011) \[\begin{gathered}
\nonumber f_{esc}(Ly\alpha)= \frac{L(Ly\alpha)_{obs}}{L(Ly\alpha)_{intrinsic}}= \\ \label{fesc} \frac{\mbox{SFR}(Ly\alpha)}{\mbox{SFR}(\mbox{UV})_{corr}}=\frac{4.4\times 10^{-42}L_{Ly\alpha}/8.7}{\mbox{SFR}(\mbox{UV})_{corr}} \end{gathered}\] with \(L_{Ly\alpha}\) the luminosity associated with the Ly\(\alpha\) line and SFR(UV)\(_{corr}\) the extinction-corrected SFRs. This assumes a Ly\(\alpha\) to H\(\alpha\) ratio of 8.7 and the H\(\alpha\) SFR calibration for a IMF of \(4.4\times 10^{-42}\) (@mathee2016). In case of non-detections, we use the corresponding S/N\(^*=5.5\) line fluxes to derive upper limits on \(W_{Ly\alpha}\) and \(f_{esc}\). Our calculations do not account for the UV excess associated with binaries, which induce uncertainties on \(f_{esc}\) that can be very difficult to account for (@stanway2016).
## AGN contamination
We do not measure any objects to have more than one emission line with S/N\(_{i}>\)S/N\(^{*}\), dismissing the existence of evident active galactic nuclei (AGNs) in our sample. We do not find any of the 120 Ly\(\alpha\)-emitting galaxies to have S/N\(>1\) emission potentially associated with the C IV\(\lambda1549\) line. Moreover, we find no broad-line Ly\(\alpha\) emission (\(\Delta v>1000\) km s\(^{-1}\)), consistent with the rarity of Type I AGNs at high redshift (@dawson2004), and in strong contrast with the AGN fractions at low redshift (\(\sim 0.4\); @finkelstein2009). This is consistent with the fact that bright AGNs are found in \(\lesssim\) 1% of \(z>3\) LAEs (@malhotra2003 [@wang2004; @gawiser2007; @ouchi2008; @zheng2010]). Low-luminosity AGN contamination in LAE samples is more difficult to constrain, but estimate the total AGN fraction to be \(<17\%\). We use our \(f_{esc}\) measurements to further dismiss the existence of evident AGNs. We perform cross-matching with the NASA/IPAC Extragalactic Database (NED[^2]) for all objects with \(f_{esc}>1\), and we find none of them to be a reported X-ray source.
# The Ly\(\alpha\) Equivalent Width Distribution {#4}
## Bayesian Inference {#4.1}
Measurement of \(W_{Ly\alpha}\) for a galaxy sample yields the \(W_{Ly\alpha}\) distribution. Since \(W_{Ly\alpha}\) is directly measured from the data, characterization of the \(W_{Ly\alpha}\) distribution can be naively considered straightforward. However, careful consideration of uncertainties and completeness can yield important insights into the underlying information. Therefore, for proper characterization of uncertainties, significance, and trends in our results, we use Bayesian statistics. In this section, we explain how to recover the \(W_{Ly\alpha}\) distribution within this framework, complementary to the one introduced in. Different probability distribution models can be adopted to reproduce \(W_{Ly\alpha}\) distribution measurements. For instance, studies use Gaussian (e.g. @guaita2010), exponential (e.g. @jiang2013a [@zheng2014]), and log-normal distributions. For our analysis, we define the probability distribution as \(p(W_{Ly\alpha}|W_n)\). Let \(W_n\) be the parameter space associated with the model. From now on, we describe the Bayesian approach to recover the posterior distribution of \(W_n\). By means of this approach, we include the uncertainties in sample size, flux measurements, and photometry in the estimation of the posterior. The description we provide is not limited to this particular work, allowing for further application in similar datasets. We only present here the fundamental equations, as this procedure is already described in detail in. Our Bayesian analysis is based on the Ly\(\alpha\) line flux \(F\) instead of \(W_{Ly\alpha}\), as introduced in equation ([\[W\_{Ly_alpha}\]](#W_{Ly_alpha}){reference-type="ref" reference="W_{Ly_alpha}"}). This approach simplifies the equations, since we can assume \(F\) and \(f_{\lambda}\) to be normally distributed, which cannot be done for \(W_{Ly\alpha}\). According to Bayes' theorem, the posterior distribution \(p(W_n| \{{\mbox{\)F\(}}\} )\), i.e., the parameter space probability distribution given our data set \(\{ F\}\), is \[\begin{gathered}
p(W_n| \{ F\} )= \frac{p(\{ F\}|W_n)p(W_n)}{p(\{ F\} )} \label{bayes} \end{gathered}\] The likelihood is just the product of the individual likelihood for every galaxy, i.e., \(p(\{F\} |W_n)= \displaystyle \prod p(F_{i} |W_n)\). For a detection, it is given by \[\begin{gathered}
p(F_{i}|W_n) = \int_{0}^{\infty} p(F_{i}|F)p(F|W_n)dF \label{single} \end{gathered}\] where \(p(F_{i}|F)\) is the line-flux probability distribution for the corresponding galaxy valued at a flux \(F\), which we consider to distribute normally. On the other hand, the term \(p(F|W_n)\) is just the probability distribution of \(F\) given \(W_n\). Using the definition of \(W_{Ly\alpha}\) from Equation ([\[W\_{Ly_alpha}\]](#W_{Ly_alpha}){reference-type="ref" reference="W_{Ly_alpha}"}), the term translates to the product distribution \(p(F|W_n)=p(W_{Ly\alpha}|W_n) p(f_{\lambda})\). At this point, we include the probability distribution for the continuum, which we also assume to be Gaussian. The limiting line flux \(F^{*}_{i}\) for discerning detections from noise is given by our S/N threshold, i.e., \(F^{*}_{i}=\)S/N\(^{*}\sigma_{i}\). For galaxies with no detections above \(F^{*}_{i}\), we adopt the following value for the likelihood: \[\begin{gathered}
\nonumber p(F_{i}<F^{*}_{i}|W_n)= \\ \int_{0}^{\infty}\left(1-p(F_{i}>F^{*}_{i}|F)\right)p(F|W_n)dF \label{nondet} \end{gathered}\] with \(p(F_{i}>F^{*}_{i}|F)\) the detection completeness at a line flux \(F\) (see Section [3.1](#3.1){reference-type="ref" reference="3.1"}). Using the expressions for detections and non-detections, the posterior distribution takes the final form: \[\begin{gathered}
\nonumber p(W_n| \{F\})= \\ \frac{p(W_n)}{p(\{F\})}\displaystyle \prod_{D}p(F_{i}|W_n) \displaystyle \prod_{ND}p(F_{i}<F^{*}_{i}|W_n) \label{final} \end{gathered}\] with \(p(W_n)\) the priors of the model parameters and \(p(\{F\})\) a normalization constant reflecting the likelihood of the model. The form of Equation ([\[final\]](#final){reference-type="ref" reference="final"}) is general and will be used as a starting point for multiple analysis throughout.
## Model Comparison {#4.2}
As we introduced in the previous section, multiple probability distributions can be adopted for the representation of the \(W_{Ly\alpha}\) distribution. In order to perform model selection, several elements are taken into account, such as model complexity and number of parameters. In this section, we describe our methodology to perform such a selection from a quantitative standpoint. By means of a Bayesian approach, we recover probability ratios for the different models, providing insight into how we perform the selection given our measurements. The analysis presented here is a quantitative implementation of Occam's Razor and is not unique to our dataset, i.e., it can be applied to any dataset modeling. Every model we discuss here is composed of a scaled probability distribution and a Dirac delta, as proposed in. If we define the standard probability distributions as \(p_0(W_{Ly\alpha}|W_n)\), the modified counterparts we consider are given by \[\begin{gathered}
p(W_{Ly\alpha}|W_n)=A\times H(W_{Ly\alpha}) \times p_0(W_{Ly\alpha}|W_n)\\ \nonumber +(1-A)\times \delta(W_{Ly\alpha}) \end{gathered}\] where the first term is the scaled probability distribution. It is multiplied by the Heaviside \(H(W_{Ly\alpha})\) to ensure it only represents positive \(W_{Ly\alpha}\) values. Hence, this scaled term integrates \(A\), i.e., \(A\) can only adopt values between 0 and 1. Note that this term is different from the fraction of detections in our sample, as we consider upper limits for our non-detections. The second term groups the fraction of galaxies that do not emit in Ly\(\alpha\) (i.e., no line and/or absorption). As our data is restricted to emission lines, we represent this term using the Dirac delta. We explore here exponential-, Gaussian-, and log-normal-based distributions. The first two are two-parameter models, while the log-normal is three-parameter dependent. Hence, we generalize our parameter space as \(W_n=(A, W_{0}, \mu)\). Then, the expressions for our exponential, Gaussian, and log-normal models are, respectively, \[\begin{gathered}
\nonumber p(W_{Ly\alpha}|\mbox{exp})=\frac{A}{W_{0}}e^{-W_{Ly\alpha}/W_{0}}H(W_{Ly\alpha})\\ \label{exponential} +(1-A)\delta(W_{Ly\alpha}) \end{gathered}\] \[\begin{gathered}
\label{gaussian} \nonumber p(W_{Ly\alpha}|\mbox{Gauss})=\frac{2A}{\sqrt{2\pi W_0^2}}e^{-W_{Ly\alpha}^{2}/2W_0^{2}}H(W_{Ly\alpha})\\ +(1-A)\delta (W_{Ly\alpha}) \end{gathered}\] \[\begin{gathered}
\label{lognormal} \nonumber p(W_{Ly\alpha}|\mbox{log n})=\frac{A}{\sqrt{2\pi W_0^2W_{Ly\alpha}^2}}e^{-(ln(W_{Ly\alpha})-\mu)^2/2W_0^2}\\ +(1-A)\delta(W_{Ly\alpha}) \end{gathered}\] We now describe our approach for comparing the three models. For a set of measurements, Bayes' theorem gives the probability of model \(M_i\) in the model space \(\{M\}\) \[\begin{gathered}
p(M_i | \{ F\})= \frac{p(\{ F\}|M_i)p(M_i)}{p(\{ F\})} \end{gathered}\] with \(p(M_i)\) the prior for model \(M_i\) in the set \(\{M\}\), which we assume to be equal for the three models. Once again, \(p(\{ F\})\) is a normalization constant. Therefore, the probability for model \(M_i\) is proportional to the likelihood of the model, i.e., \[\begin{gathered}
p(M_i | \{ F\}) \propto p(\{ F\}|M_i) = \int_{n}p(W_n)p(\{F\}|W_n)dW_n \label{model_likelihood} \end{gathered}\] The absolute probability of each model \(M_i\) within the set \(\{M\}\) is obtained by imposing that the models explored cover all possible choices, i.e., \(\sum p(M_i | \{ F\})=1\). Hence, the probability of \(M_i\) given our dataset is \[\begin{gathered}
p(M_i | \{ F\}) = \frac{ \int_{n}p(W_n)p(\{F\}|W_n)dW_n}{\sum p(M_i | \{ F\})} \label{model_theorem} \end{gathered}\] For an analytically correct model comparison, analysis of Equation ([\[model_theorem\]](#model_theorem){reference-type="ref" reference="model_theorem"}) is required. Still, the term in Equation ([\[model_likelihood\]](#model_likelihood){reference-type="ref" reference="model_likelihood"}) is strongly dependent on the priors assumed for the parameter space \(W_n\) of every model. Therefore, different prior selections can have significant effects on the odds for each model. As a workaround, we rewrite the model probabilities as \[\begin{gathered}
p(M_i | \{ F\}) = \frac{ \int_{n}p(\{F\}|W_n)dW_n}{\sum p(M_i | \{ F\})} \label{odds} \end{gathered}\] Effectively, this simplification conveniently limits our analysis to a pure likelihood comparison, i.e., we adopt constant, uninformative priors. This is equivalent to assuming ignorance in linear scales for every parameter. Since our distributions are smooth and single peaked, we are confident in this assumption. The dimensions of our parameter spaces do not go beyond three, and the uncertainties in our parameters are of the order of the most probable values, providing further assurance to our assumptions. In, we show that a galaxy sample with a broad M\(_*\) range yields a composite \(W_{Ly\alpha}\) distribution. For the rest of this section, we divide our sample into three M\(_*\) bins, as shown in Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}. We use the complete sample and these three subsamples to contrast the models. The outcomes are presented in Table [\[table:1\]](#table:1){reference-type="ref" reference="table:1"} and Figure [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}. Table [\[table:1\]](#table:1){reference-type="ref" reference="table:1"} gives evidence that the best likelihood is obtained with the log-normal model for three of the four distributions. This can be verified in our distribution simulations for the complete sample in Figure [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}, especially toward the high \(W_{Ly\alpha}\) tail. Still, when integrating the likelihoods, the lognormal distribution is the least probable. This is a consequence of the extra parameter needed by the model, which penalizes the likelihood when integrating over the parameter space. Then, according to our analysis, the preferred model is the exponential. While it models the distribution better than the Gaussian, it also reproduces our \(W_{Ly\alpha}\) measurements fairly well, despite depending on only two parameters. In addition, the uncertainties in Figure [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}, especially for low \(W_{Ly\alpha}\), confirm this model is the most adequate to reproduce our measurements. We remark that the procedure for model selection described here considers the lower \(W_{Ly\alpha}\) end of the distribution, which includes our completeness and non-detections. Nonetheless, further \(W_{Ly\alpha}\) distribution analysis in this paper is mostly focused on the higher \(W_{Ly\alpha}\) end and is not strongly dependent on model preference. From now on, we perform our \(W_{Ly\alpha}\) distribution analysis using the exponential model of Equation ([\[exponential\]](#exponential){reference-type="ref" reference="exponential"}). We stress that this expression is dependent on the parameters \(A\) and \(W_0\), with the first being the fraction of galaxies showing emission and the second the e-folding scale of the distribution. We advise caution when using \(A\) as a proxy for the fraction of line emitters in the parent population, since it is tied to the adequacy of an exponential profile.
# Ly\(\alpha\) Emission Dependence on Galaxy Properties {#5}
## Stellar Mass {#5.1}
Evidence suggests Ly\(\alpha\) emission is strongly dependent on the M\(_*\) of galaxies. Galaxies with higher M\(_*\) have been forming stars for longer, leading to greater ISM dust that presumably forms in supernovae and AGB stars (@silva1998). A greater dust content leads to more Ly\(\alpha\) photon absorption, decreasing \(W_{Ly\alpha}\). This effect has already been observed, at least for high \(W_{Ly\alpha}\), in and. Similarly, the bulk of M\(_*\) is dominated by older stars, which do not contribute significantly to the Ly\(\alpha\) photon budget of galaxies. As a matter of fact, Ly\(\alpha\) emission decreases steadily with the age of stellar populations, as seen in and. Ly\(\alpha\) radiative transfer is also severely affected by the neutral gas structure and kinematics of the ISM and circumgalactic medium (@verhamme2006). Since more massive star-forming galaxies are bound to have higher gas mass (e.g. @keres2005 [@finlator2007]), Ly\(\alpha\) photons should be subject to more resonant scattering, therefore decreasing their \(W_{Ly\alpha}\). The trends we find in confirm this qualitative scheme at \(z\sim 4\). In this section, we perform a more detailed and robust characterization of Ly\(\alpha\) emission dependence on M\(_*\). Our sample is especially designed to study the dependence of Ly\(\alpha\) emission in M\(_*\). As shown in Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}, our objects are selected in redshift and M\(_*\), homogeneously covering the range \(7.6<log(\mbox{M}_*[\mbox{M}_{\odot}])<10.6\). For further clarity, we plot in Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} the M\(_*\) and SFRs of our complete sample and detections. Comparison of the M\(_*\) histograms slightly hints at an anti-correlation between the LAE fraction and M\(_*\), at least down to our detection limit. However, since our completeness depends on M\(_*\), a more thorough analysis is required (see below). The existence of any Ly\(\alpha\) emission dependence on M\(_*\) becomes clearer in Figure [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"}, where we plot \(W_{Ly\alpha}\) and \(f_{esc}\) as a function of M\(_*\). We also plot in Figure [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"} our upper limits for non-detections. The regions sampled by our non-detections reveal how our completeness is not independent of M\(_*\). Our detections are flux limited, so we achieve lower \(W_{Ly\alpha}\) for galaxies with brighter UV continuum. Therefore, even though we observe lower M\(_*\) galaxies to have higher \(W_{Ly\alpha}\) and \(f_{esc}\), any qualitative conclusions we can draw involving the fraction of LAEs as a function of M\(_*\) are affected by our completeness. Still, there is a clear upper envelope to the distribution of galaxies in this plot, where we are not affected by incompleteness. For M\(_*<10^{8.5}\), there is a clear anti-correlation between \(W_{Ly\alpha}\) and M\(_*\). For more massive galaxies, however, the trend is mostly flat, except for the presence of a few interlopers. Still, our qualitative result is that both Ly\(\alpha\) emission diagnostics show an anti-correlation with M\(_*\). For the rest of this section, we focus on how to thoroughly quantify the effect of M\(_*\) on \(W_{Ly\alpha}\). The overall dependence of Ly\(\alpha\) emission on M\(_*\) implies that \(W_{Ly\alpha}\) distributions in the literature (e.g., @gronwall2007 [@zheng2014]) are influenced by the M\(_*\) distribution of the sample. In comparison to deeper M\(_{UV}\) surveys, shallower samples are bound to observe lower \(W_{Ly\alpha}\). This can lead to incorrect contrast of surveys and misinterpretation of trends. To verify these claims, we divide our sample into three M\(_*\) bins (see Figure [\[fig:2\]](#fig:2){reference-type="ref" reference="fig:2"}) and plot the resulting \(W_{Ly\alpha}\) distributions in Figure [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}. As expected, there is an apparent anti-correlation between M\(_*\) and both, the tail of the \(W_{Ly\alpha}\) distribution and the normalization. We perform our first quantitative characterization of the \(W_{Ly\alpha}\) distribution dependence on M\(_*\) in. In that study, we divide our sample into three M\(_*\) bins and obtain the posterior distribution for the exponential parameters separately. This procedure allowed us to fit a linear relation to the final parameters, recovering \(A\)(M\(_*\)) and \(W_0\)(M\(_*\)) using expressions of the form \[\begin{gathered}
\label{mass1} A(\mbox{M}_*) = A_{\mbox{\scriptsize M}_*}\log(\mbox{M}_*[\mbox{M}_{\odot}])+A_C \\ W_{0} (\mbox{M}_*) = W_{\mbox{\scriptsize M}_*}\log(\mbox{M}_*[\mbox{M}_{\odot}]) + W_C \label{mass2} \end{gathered}\] In this section, we recover these linear relations directly from the complete sample, i.e., we recover the posterior distribution for the four-parameter space composed of the linear coefficients in Equations ([\[mass1\]](#mass1){reference-type="ref" reference="mass1"}) and ([\[mass2\]](#mass2){reference-type="ref" reference="mass2"}). This more robust methodology does not rely on binning, while also allowing us to constrain the errors on the coefficients directly from the model and measurements. We once again start from Equation ([\[final\]](#final){reference-type="ref" reference="final"}). As mentioned, our parameter space is now \(W_n=(A_{\mbox{\scriptsize M}_*}, A_C, W_{\mbox{\scriptsize M}_*}, W_C)\). As these linear coefficients represent the exponential parameters of Equation ([\[exponential\]](#exponential){reference-type="ref" reference="exponential"}), deriving non-informative priors is highly complicated. Therefore, in order to determine our priors, we only consider linear scales ignorance. Therefore, the posterior translates to \[\begin{gathered}
p(A_{\mbox{\scriptsize M}_*}, A_C, W_{\mbox{\scriptsize M}_*}, W_C| \{F\})= C\times \\ \nonumber \displaystyle \prod_{D}p(F_{i}|A(\mbox{M}_*), W_{0}(\mbox{M}_*)) \displaystyle \prod_{ND}p(F_{i}<F^{*}_{i}|A(\mbox{M}_*),W_{0}(\mbox{M}_*)) \end{gathered}\] with \(C\) a normalization constant. We use MCMC simulations to characterize this four-parameter posterior. Its maximum gives the best solution for \(A(\mbox{M}_*)\) and \(W_0(\mbox{M}_*)\), while the collapsed posteriors yield the uncertainties on the parameters. We can then write Equations ([\[mass1\]](#mass1){reference-type="ref" reference="mass1"}) and ([\[mass2\]](#mass2){reference-type="ref" reference="mass2"}) as \[\begin{gathered}
\label{mass12} A(\mbox{M}_*) =-0.28_{-.02}^{+.12}\log{(\mbox{M}_*[\mbox{M}_{\odot}])}+ 3.1^{+0.15}_{-1.1} \\ W_{0} (\mbox{M}_*) =-17.7_{-2.0}^{+2.6} \log{(\mbox{M}_*[\mbox{M}_{\odot}])} + 190^{+20}_{-25} \label{mass22} \end{gathered}\] In the framework of an exponential profile, these relations we recover yield the \(W_{Ly\alpha}\) probability distribution for an object with known M\(_*\). Hence, we can simulate the expected \(W_{Ly\alpha}\) distribution for each of the three M\(_*\) subsamples and compare with our direct measurements. The results are presented in Figure [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}. Our constraints are consistent with the observed \(W_{Ly\alpha}\) distributions. Our results regarding the dependence of Ly\(\alpha\) on M\(_*\) are conclusive. In the range \(10^{8}-10^{10.5}\mbox{M}_{\odot}\), the \(W_{Ly\alpha}\) probability distribution extends to higher \(W_{Ly\alpha}\) for lower M\(_*\) galaxies. In other words, more massive galaxies tend to have lower \(W_{Ly\alpha}\). A similar trend is observed for \(f_{esc}\), further highlighting the role of dust and gas mass in the escape of Ly\(\alpha\) photons. At \(z\sim 2.2\), also observe the \(f_{esc}\) anti-correlation from Figure [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"}, although only when stacking galaxies. Their massive objects showing high \(f_{esc}\), which they associate with dusty gas outflows, seem to lie below the M\(_*\)-SFR sequence at \(z\sim 2\). These inferences, combined with the significantly higher \(f_{esc}\) they measure for larger apertures, make sense in a Ly\(\alpha\) diffuse halo scheme, which we do not observe due to our aperture size. Studies on the dependence of \(f_{esc}\) on M\(_*\) also explore \(1.9<z<3.6\) (@hagen2014). Their \(W_{Ly\alpha}\)-selected LAEs follow a trend similar to the one we find at \(z\sim 4\). In summary, the evidence for an anti-correlation between \(W_{Ly\alpha}\) (or \(f_{esc}\)) and M\(_*\) is significant, but the scatter seems to depend on measurement methodology and sample selection. Inferences on the M\(_*\) distribution of LAEs are not as evident. do not find their \(1.9<z<3.6\) Ly\(\alpha\) luminosity-selected LAE number distribution to depend on M\(_*\). Their results agree with the \(z\sim 3.1\) narrowband-selected survey of. However, we show in Figure [\[fig:6\]](#fig:6){reference-type="ref" reference="fig:6"} that spectroscopic completeness is not independent of M\(_*\). Therefore, most LAEs survey follow-ups could have higher incompleteness toward lower M\(_*\). Since our Bayesian analysis takes into account our completeness for every object, we can test the significance of this claim. The coefficient \(A_{\mbox{\scriptsize M}_*}\) in Equation ([\[mass1\]](#mass1){reference-type="ref" reference="mass1"}) represents the exponential fraction of LAE dependence on M\(_*\). As evidenced by Equation ([\[mass12\]](#mass12){reference-type="ref" reference="mass12"}), our measurements are more than \(2\sigma\) consistent with a decrease in the fraction going to higher M\(_*\). In the scheme of an exponential model, this translates to an LAE distribution dominated by lower M\(_*\) galaxies. This result complements the much more significant anti-correlation between \(W_{0}\) and M\(_*\) we find (see Equation [\[mass22\]](#mass22){reference-type="ref" reference="mass22"}).
## SFR {#5.2}
High-redshift galaxies have been observed to follow a correlation between SFR and M\(_*\), known as the star-forming main sequence (e.g., @keres2005 [@finlator2007; @stark2009; @gonzalez2011; @whitaker2012]; see our Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}). In terms of the underlying physics, more massive objects dominate gas accretion in their neighborhood, feeding and triggering star-formation. Such gas infall seems to dominate over galaxy growth at high redshift (@keres2005 [@finlator2007]). This scheme implies that more massive objects form stars at higher rates, at least down to our observational limitations and modeling of high-redshift ISM. Given our results on M\(_*\) from the previous section, we expect similar trends between \(W_{Ly\alpha}\) (\(f_{esc}\)) and SFR (Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}). Even more, star-forming galaxies have a higher neutral gas mass, which can hamper the escape of Ly\(\alpha\) photons from galaxies (@verhamme2006). In fact, it has also been suggested that photoelectric absorption rules Ly\(\alpha\) depletion, even over dust attenuation (@reddy2016). In this section, we explore any Ly\(\alpha\) dependence on SFR within our dataset. We remark that our SFRs come from SED fitting of 3D-HST photometry using FAST (see Section [2.3](#2.3){reference-type="ref" reference="2.3"}), i.e., they have typical associated timescales of 100 Myr (@kennicutt1998). We stress that our derived SFRs differ from 3D-HST SFRs, since our calculation assumes cSFHs instead of exponentially declining SFHs (@skelton2014). We show the \(W_{Ly\alpha}\) and \(f_{esc}\) dependence on SFR in Figures [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} and [\[fig:9\]](#fig:9){reference-type="ref" reference="fig:9"}. In the latter, we include upper limits for our non-detections to give an insight into how our incompleteness depends on SFR. A clear anti-correlation between \(W_{Ly\alpha}\) (\(f_{esc}\)) and SFR is observed. These results come as no surprise, as they have been previously reported. Most studies of \(W_{Ly\alpha}\) dependence on SFR involve uncorrected SFRs (@pettini2002 [@shapley2003; @yamada2005; @gronwall2007; @tapken2007; @ouchi2008]; all compiled in @verhamme2008). Even without dust correction, the anti-correlation is still present in these studies (refer to Figure 19 in @verhamme2008 and Section [5.4](#5.4){reference-type="ref" reference="5.4"} of this work). Based on a \(z\sim 2\) H\(\alpha\) emitters sample, also observe a clear anti-correlation between Ly\(\alpha\) \(f_{esc}\) and SFR. Interestingly, they do not only observe such a trend in their individual objects, but likewise on their stacks when using different apertures (galaxy diameters of 12 and 24 kpc). As their dataset includes H\(\alpha\) fluxes, they can recover SFRs and \(f_{esc}\) using H\(\alpha\) luminosities. The fact that they observe similar trends with such a different sample suggests that the anti-correlation between \(W_{Ly\alpha}\) (\(f_{esc}\)) and SFR is not only independent of redshift, but also observational constraints like aperture and methodology for recovering SFRs. Their comparison, however, is restricted to SFRs higher than \(\sim 5\mbox{M}_{\odot}\)yr\(^{-1}\). Most of our low-mass objects have SFRs lower than \(\sim 5\mbox{M}_{\odot}\)yr\(^{-1}\), but they seem to follow the same regime as the rest of our sample. Even though \(f_{esc}\) uncertainties and incompleteness increase toward lower-SFR, UV-fainter galaxies (Figure [\[fig:9\]](#fig:9){reference-type="ref" reference="fig:9"}), our results suggest that \(f_{esc}\) reaches values of 100% toward SFR\(\sim 1-3\mbox{M}_{\odot}\)yr\(^{-1}\). These numbers are consistent with the analysis by. They compare their \(z<0.5\) SFR(Ly\(\alpha\))/SFR(UV) measurements with the literature at \(z>2\) (@taniguchi2005 [@gronwall2007; @guaita2010; @curtis-lake2012; @jiang2013a]). Our finding that \(f_{esc}\) reaches values of 100% toward SFR\(\sim 1-3\mbox{M}_{\odot}\)/yr overestimates SFR(Ly\(\alpha\))/SFR(UV) for \(z<0.5\), but is consistent with higher redshift. As pointed by, it seems that \(f_{esc}\) at fixed SFR(UV) increases with redshift. Then again, it must be kept in mind that these literature results consider uncorrected SFRs.\
## UV Luminosity {#5.3}
In this section, we analyze the M\(_{UV}\) distribution of our sample, while also exploring any correlations between \(W_{Ly\alpha}\) and UV luminosities. It must be noted, though, that our sample is not representative of the galaxy population at \(3<z<4\). First, our galaxies are homogeneously distributed in M\(_*\), i.e., they are not a random sample from \(3<z<4\) CANDELS objects. Second, we are affected by CANDELS completeness, which decreases toward lower M\(_*\) galaxies. Since more massive galaxies tend to have higher UV luminosities (@stark2009 [@gonzalez2014]), our sample has a higher contribution of bright M\(_{UV}\) galaxies than a population-representative subsample. In order to determine M\(_{UV}\) for our objects, we use the CANDELS \(i_{775}\) band. We show in Figure [\[fig:11\]](#fig:11){reference-type="ref" reference="fig:11"} the corresponding distribution of the complete sample and detections. We also present the dependence of \(W_{Ly\alpha}\) on M\(_{UV}\) in this figure. As expected from the SFR-\(W_{Ly\alpha}\) anti-correlations we recover in Section [5.2](#5.2){reference-type="ref" reference="5.2"}, a similar trend is observed for UV luminosities. This anti-correlation comes as no surprise, since brighter UV galaxies tend to have higher M\(_*\) at the cosmic time of our sample (@stark2009 [@gonzalez2014]). Galaxies brighter in the UV have been subject to more intensive star-formation events in 100 Myr timescales. Typically, higher neutral gas, turbulence, and bulk gas motions are associated with higher SFRs, boosting the scatter of Ly\(\alpha\) photons. In combination, higher dust extinction, older age of stellar populations, and greater neutral gas mass in UV brighter galaxies seem to rule Ly\(\alpha\) statistics dependence on SFR. Most analyses of Ly\(\alpha\) emission dependence on M\(_{UV}\) have been performed using uncorrected SFRs. The literature compilation shown in reveals how observed UV SFRs anti-correlate with \(W_{Ly\alpha}\) from z\(\sim2.7\) to 5.7 (@pettini2002 [@shapley2003; @yamada2005; @gronwall2007; @tapken2007; @ouchi2008]). Analyses explicitly M\(_{UV}\) have also been performed in the surveys of, ,, ,, , and, yielding similar trends up to \(z\sim 6\). More recent studies confirm these trends (@jiang2013a [@jiang2016; @zheng2014]). Simulations likewise predict such correlations (see @shimizu2011). However, regardless of the methodology, the scatter in these correlations is non-negligible. Moreover, question the existence of the correlation in their sample. We argue that the scatter observed in the \(W_{Ly\alpha}\) dependence on M\(_{UV}\)/SFR\(_{obs}\) can be a consequence of the role played by dust. Galaxies brighter in the UV naturally have a greater Ly\(\alpha\) photon production, but the anti-correlation with M\(_*\) affects the escape fraction (probably through increased dust extinction). In this scenario, Ly\(\alpha\) photon escape is a complex process simultaneously ruled by different properties of high-redshift galaxies. We explore such property space in our analysis of the \(W_{Ly\alpha}\) distribution dependence on the M\(_{UV}-\beta\) relationship in Section [5.5](#5.5){reference-type="ref" reference="5.5"}. Implications involving the fraction of detections as a function of M\(_{UV}\) are not straightforward. In principle, this fraction seems to correlate with the UV luminosities of our galaxies, as opposed to what is observed for characteristic \(W_{Ly\alpha}\). Nevertheless, our Ly\(\alpha\) measurements are flux limited. Therefore, our detection completeness in \(W_{Ly\alpha}\) is higher for brighter objects, leading to biases difficult to account for, as noted in. Under these circumstances, the ideal approach is to consider both detections and non-detections, while taking into account the uncertainties for line and continuum fluxes (\(F\) and \(f_\lambda\), respectively). Hence, we encourage further interpretations of these results to focus on the analysis performed on the M\(_{UV}-\beta\) plane (Section [5.5](#5.5){reference-type="ref" reference="5.5"}).
## UV slope {#5.4}
Measurement of the UV slopes of high-redshift galaxies is a direct way of tracing the amount of dust inside galaxies, given the assumption of an extinction law and an intrinsic spectral shape. This is of particular interest for Ly\(\alpha\) surveys, since simulations (e.g., @verhamme2008) and observations at low redshift (@hayes2011 [@atek2014]) suggest that dust plays an important role in the escape of Ly\(\alpha\) photons. Since we have rest-frame UV photometry from CANDELS for all our objects, we can determine their UV slopes and study their effect on Ly\(\alpha\) emission at \(z\sim 4\). In this section, we detail our method to estimate the UV slopes for our objects and show our results on the Ly\(\alpha\) dependence on this galaxy property. To determine UV slopes, we fit a power law \(f_\lambda= f_0\lambda^{\beta}\) (@calzetti1994) to the photometry of each object. For fitting, we just use a standard least-squares routine on the photometry between rest frame 1400 and 3500Å. These calculations correspond, in principle, to 8-15 bands between 1500 and 3600Å. However, since we only use fluxes with S/N\(>3\), our median number of bands is eight. Naturally, we further require at least two bands to associate a slope to our targets, which means we can measure the UV slope for 611 of the 625 observed objects composing our sample. Our FAST outputs include dust extinction in the \(V\)-band, A\(_V\), for every object. However, if we want to associate a reddening \(E(B-V)\) with every galaxy, we need to use our derived UV slopes. We use the relation \[\begin{gathered}
\label{ebv} f_\lambda= f_0\lambda^{\beta} \propto \lambda^{-2.23}10^{0.4k_{\lambda}E(B-V)} \end{gathered}\] where we assume a pristine slope of \(\beta=-2.23\) (@meurer1999). We also require the adoption of an attenuation law \(k_{\lambda}\) (e.g. @calzetti2000). recover a more appropriate attenuation law than for high-redshift galaxies, so we use the \(k_{\lambda}\) of the former. It is worth noting, nevertheless, that both attenuation laws yield almost identical results. Since our targets are M\(_*\) selected, we have a higher contribution of massive objects in comparison to the M\(_*\) distribution of the galaxy population. As more massive objects tend to have higher \(E(B-V)\) and redder UV slopes, we expect \(z\sim 4\) samples representative of the galaxy population to have a lower contribution from such galaxies. In any case, we show in Figure [\[fig:13\]](#fig:13){reference-type="ref" reference="fig:13"} the UV slope histogram of our sample and our Ly\(\alpha\) emitters. Just by comparing both distributions, it is clear that the fraction of emitters increases toward bluer galaxies. We present a quantitative analysis of this claim in Section [5.5](#5.5){reference-type="ref" reference="5.5"}. We also show in Figure [\[fig:13\]](#fig:13){reference-type="ref" reference="fig:13"} our results on \(W_{Ly\alpha}\) and \(f_{esc}\) as a function of \(\beta\) and \(E(B-V)\). Since lower mass galaxies have bluer UV slopes than more massive ones, the trends we find are complementary to our previous results. There is a correlation between the steepness of the UV spectrum and \(W_{Ly\alpha}\)/\(f_{esc}\), although with significant scatter. As extinction seems to play a major role in Ly\(\alpha\) photon escape from galaxies, mainly through scattering and absorption (@blanc2011 [@hagen2014]), these correlations come as no surprise. Qualitatively, our results agree with the measurements from, ,, and. Regarding the scatter we find at fixed \(\beta\) (see also @blanc2011), it is consistent with a scenario where the observed Ly\(\alpha\) flux is mostly affected by the dependence of the dust-covering fraction on the line of sight. We discuss this picture in detail in Section [5.5](#5.5){reference-type="ref" reference="5.5"}. When comparing the more dusty Ly\(\alpha\) emitters in our sample with results from the literature, however, some differences show up. The \(1.9<z<3.8\) survey from and the \(z<0.5\) study from find Ly\(\alpha\) emitters up to \(E(B-V)\sim 1\). Similarly, find a population of dusty LAEs at \(z\sim 2.23\), and they speculate on how dusty gas outflows might be the feature driving the escape of Ly\(\alpha\) radiation. Ly\(\alpha\) sources up to \(z\sim3\) are *Herschel* (@oteo2011 [@oteo2012; @casey2012; @sandberg2015]) and SCUBA (@geach2005 [@hine2016]) detected. Moreover, Ly\(\alpha\) emission has also been measured in sub-millimeter galaxies (e.g. @chapman2005). On the other hand, our 625 object sample features \(\sim5\) objects with \(E(B-V)>0.4\), but none of them qualifies as a detection. We cannot state whether the absence of such LAEs in our sample is representative of \(z\sim 4\). True enough, the fraction of dusty galaxies at \(z\sim 4\) is expected to be lower than at \(z<4\). Along such lines, cosmic evolution in the fraction of dusty LAEs has already been discussed in the literature. study any dust evolution in their \(1.9<z<3.8\) LAEs sample and find no significant trend. study the same sample, and show that there is little anti-correlation, if any, between \(E(B-V)\) and redshift. do observe such evolution in the range \(2<z<3.5\), but question the validity given the selection effects associated with LBG surveys. We conclude that even though Ly\(\alpha\) emitting galaxies are mostly low-dust objects (e.g. @song2014), there is also a population of dustier, low-\(W_{Ly\alpha}\) LAEs. Still, the significance of their numbers at \(z\gtrsim4\) is still an open question. Our results also confirm at \(z\sim 4\) a trend in \(f_{esc}\) that has already been observed at lower redshift. perform a \(f_{esc}\) study at \(z<0.5\), and recover a similar anti-correlation to the one we show in Figure [\[fig:13\]](#fig:13){reference-type="ref" reference="fig:13"}. at \(z\sim2.2\), at \(1.9<z<3.8\), at \(z\sim2.1-2.5\), and at \(z\sim 2.23\) also observe the same trends. replicate such trends using radiative simulations of galaxies in the range \(2.8<z<5\), confirming that qualitative explanations for these observational relations are well supported by theory. We show several best-fit relations from the literature in our lower plot of Figure [\[fig:13\]](#fig:13){reference-type="ref" reference="fig:13"}. Considering that our results are dominated by upper limits, the two higher redshift relations are roughly consistent with our measurements. Still, of particular interest might be the potential redshift evolution suggested by these relations. There is a clear decrease in \(f_{esc}\) at high dust contents when going from low to high redshift. If real, this trend could back our previous analysis on the fraction of more dusty LAEs and their evolution as a function of cosmic time. At higher redshift (@verhamme2008; this work), low-\(W_{Ly\alpha}\), very dusty LAEs do not seem to be common, driving the \(E(B-V)-f_{esc}\) relation down significantly for \(E(B-V)>0.2\). However, at lower redshift, such objects are actually observed, driving the relation up for \(E(B-V)>0.2\).
## M\(_{UV}\)-\(\beta\) sequence {#5.5}
We have characterized in this paper the dependence of Ly\(\alpha\) emission on M\(_*\), SFR, UV luminosity, and UV slope. We find \(W_{Ly\alpha}\) and \(f_{esc}\) to anti-correlate with these four properties. However, these results might not be independent, since these properties are correlated (see Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}). For instance, M\(_*\) and SFR are known to follow a relation at high redshift referred to as the main sequence (@keres2005 [@finlator2007; @noeske2007; @daddi2007]). Similarly, a relation between M\(_{UV}\) and \(\beta\) has been studied in LBGs (@bouwens2009 [@bouwens2014]). As we discuss throughout this work, Ly\(\alpha\) escape from galaxies is likely to be a process ruled by many parameters, such as age of the population, gas column density, extinction, and SFR. Therefore, it is interesting to explore Ly\(\alpha\) emission dependence on multi-dimensional spaces. In this section, we study \(W_{Ly\alpha}\) dependence on M\(_{UV}\) and \(\beta\). There are two reasons to justify exploring this parameter space in particular. First, both properties are observables, i.e., the amount of assumptions involved in their calculation is kept to a minimum (unlike, for example, M\(_*\) and SFR). Second, these two properties are directly related to elements that rule Ly\(\alpha\) escape. The UV luminosity of a galaxy traces its ionizing output and gas content, whereas the UV slope is a proxy for dust. Hence, in this section, we focus on characterizing Ly\(\alpha\) emission in the M\(_{UV}-\beta\) plane. Similarly to previous analyses in this paper, we take advantage of a Bayesian approach to obtain our results. We first present in Figure [\[fig:14\]](#fig:14){reference-type="ref" reference="fig:14"} the location of our detections and non-detections location in this plane. In qualitative terms, our detections sample most of the sequence initially covered by our targets. For further insight into how our observations depend on this sequence, we construct four subsamples for visualization. The anti-correlation we observe between \(W_{Ly\alpha}\) and M\(_{UV}\) is still clearly observed when comparing the low M\(_{UV}\) with the high M\(_{UV}\) subsamples, independent of \(\beta\). For the UV slope, however, the trend between \(W_{Ly\alpha}\) and \(\beta\) does not seem that clear anymore. The normalization of the distribution, however, looks to be highly dependent on the UV slope. To characterize the significance of these insights, we simultaneously model the exponential profile parameters of Equation ([\[exponential\]](#exponential){reference-type="ref" reference="exponential"}) as a function of M\(_{UV}\) and \(\beta\). We start again with linear expressions of the form \[\begin{gathered}
\label{A} A(M_{UV}, \beta)= A_{M_{UV}} M_{UV}+ A_{\beta} \beta+A_{C} \\ \nonumber \\ \label{W_0} W_0(M_{UV}, \beta)= W_{M_{UV}} M_{UV}+ W_{\beta} \beta+W_{C} \\ \nonumber \end{gathered}\] Once the parameterization and priors are set, we can obtain the posterior distribution of \(\{W_n\}=(A_{M_{UV}}, A_{\beta}, A_C, W_{M_{UV}}, W_{\beta}, W_C)\) using Equation ([\[final\]](#final){reference-type="ref" reference="final"}). We assume the priors to be independent, which translates to \(p(\{W_n\})\propto \displaystyle \prod p(W_i)\). Once again, we impose ignorance on the parameters, i.e., we adopt non-informative priors. We study the posterior distribution assuming parameter ignorance in both linear and logarithmic scales. We decide for the first, since logarithmic priors diverge for parameters than can adopt values close to zero (\(A_{M_{UV}}\), \(A_{C}\) and \(W_{\beta}\)). Therefore, our prior is simply a constant \(p(\{W_n\})\). Considering that \(p(\{F\})\) is just a normalization factor in Equation ([\[final\]](#final){reference-type="ref" reference="final"}), the posterior distribution of the six-parameter space can be obtained: \[\begin{gathered}
p(W_n| \{F\})\propto \displaystyle \prod_{D}p(F_{i}|W_n) \displaystyle \prod_{ND}p(F_{i}<F^{*}_{i}|W_n) \end{gathered}\] The value of the parameters \(A\) and \(W_0\) is, by definition, constrained to the intervals \(A=[0,1]\) and \(W_0=(0,\infty)\). Still, in the case of extreme luminosities and/or slopes, our linear parameterizations can yield values outside these intervals. As a solution, we just impose \(A\) and \(W_0\) to saturate outside their corresponding ranges. This condition can be considered simply as an indirect prior on the parameters. For the particular case of \(W_0<0\), we impose \(A=0\), i.e., the \(W_{Ly\alpha}\) distribution is a Dirac delta (i.e., no Ly\(\alpha\) emission and/or absorption). We recover our best solution from Monte Carlo simulations and obtain the uncertainties on our parameters using the collapsed distributions. The results are as follows: \[\begin{gathered}
\nonumber A(M_{UV}, \beta)= \\ \label{Afinal} 0.08^{+0.1}_{-0.06} M_{UV}-0.6^{+0.1}_{-1.2}\beta+1.1^{+1.5}_{-1.4} \\ \nonumber \\ \nonumber W_0(M_{UV}, \beta)= \\ \label{W_0final} 11^{+2.0}_{-1.8}M_{UV}-7.8^{+5.2}_{-5.5}\beta+235^{+45}_{-42}\\ \nonumber \end{gathered}\] We show with more detail the results for these coefficients in Figure [\[fig:15\]](#fig:15){reference-type="ref" reference="fig:15"}. The early analysis based on the sample binning from Figure [\[fig:14\]](#fig:14){reference-type="ref" reference="fig:14"} is verified in Figure [\[fig:15\]](#fig:15){reference-type="ref" reference="fig:15"}. The normalization factor \(A\) anti-correlates with \(\beta\) with a significance \(\gtrsim2\sigma\). There also seems to be an anti-correlation between \(A\) and M\(_{UV}\), but to a level of \(\sim 1\sigma\). For \(W_{0}\), the e-folding scale of the distribution, the behavior is the opposite. The extent of the distribution correlates with M\(_{UV}\) and shows a weak anti-correlation with \(\beta\). We now focus on the interpretation of these results. The fact that beta correlates mostly with \(A\) suggests that the UV slope is somehow related to a stochastic probability of being able to observe or not a particular object in Ly\(\alpha\). On the other hand, the stronger correlation of M\(_{UV}\) with \(W_{Ly\alpha}\) suggests that the UV luminosity is associated with physical processes that determine the magnitude of the resulting \(W_{Ly\alpha}\). A possible scenario is one in which M\(_{UV}\) traces SFR and, therefore, the total cold gas mass of galaxies. Hence, higher UV luminosities translate into increased levels of Ly\(\alpha\) photon scattering into an extended halo, smoothly reducing the \(W_{Ly\alpha}\) of the central source. At the same time, in a significantly clumpy ISM in which dust is well mixed with the gas, a large contrast in gas density/column will imply that dust can effectively block the bulk of both UV and Ly\(\alpha\) radiation in some parts of the disk. Such impact will be significant toward certain lines of sight, with little to no effect on others. Such "covering factor\" scenario could explain that \(\beta\) correlates more significantly with the normalization (\(A\)) than with the shape of the distribution (\(W_{0}\)). This is consistent with the finding that Ly\(\alpha\) and UV suffer from similar levels of extinction by dust in LAEs (e.g. @blanc2011). Moreover, the scatter in the \(f_{esc}\) of Ly\(\alpha\) photons at fixed reddening (see Figure [\[fig:13\]](#fig:13){reference-type="ref" reference="fig:13"} and @blanc2011) is also consistent with this scenario where wide ranges of dust absorption lines of sight and photon scattering halos rule the escape of Ly\(\alpha\) radiation. Naturally, this interpretation is a simplified description of Ly\(\alpha\) escape from galaxies. Going no further, Figure [\[fig:15\]](#fig:15){reference-type="ref" reference="fig:15"} reveals how both distribution parameters do not solely depend on one property. Ideally, studies of \(W_{Ly\alpha}\) and \(f_{esc}\) in three-parameter spaces (e.g. M\(_*\), M\(_{UV}\), \(\beta\)) can yield predictive and more accurate parameterization of Ly\(\alpha\) emission. However, such analysis must be performed in much larger datasets than ours, at least if significant enough results are to be obtained. Furthermore, hydrodynamical simulations can also give insight into how Ly\(\alpha\) escape depends on the line of sight and the distribution of gas, dust, and star-forming regions (e.g. @verhamme2012), especially if performed in a statistically significant sample.
# Ly\(\alpha\) Dependence on Sample Selection {#6}
## LBG Samples {#6.1}
The Lyman break selection technique has proven to be a very efficient method for detecting high-redshift galaxies (e.g. @steidel1996 [@shapley2003; @stark2010; @ono2012]). The fact that the Lyman break is in the optical region of the observed spectrum for galaxies at redshift \(z\geqslant3\) allows for efficient detection from ground telescopes. By only requiring the use of a few broadband filters, several galaxies that can be detected in a deep single exposure. Still, to avoid aliasing with the Balmer break, there are unavoidable biases associated with this technique. First, galaxies with no prominent Lyman break are, by construction, not detected. As a consequence, either extremely young or passive, heavily extincted galaxies are underrepresented in LBG surveys. Second, these surveys also impose color restrictions on the slope of galaxies, further increasing the selection toward, in principle, bluer UV objects. Third, this technique is limited by the M\(_{UV}\) sensitivity of the survey, creating incompleteness at low-SFR. In the case of Ly\(\alpha\) emission, this observational limit can have a significant downside. As shown in Section [5.2](#5.2){reference-type="ref" reference="5.2"}, there is a clear anti-correlation between UV luminosity and \(W_{Ly\alpha}\), leading to the possibility that Ly\(\alpha\) studies in LBG samples are missing the highest \(W_{Ly\alpha}\) galaxies in the universe. Indeed, the deepest surveys nowadays reach \(\sim 50\%\) completeness at observed m\(_{UV}\sim29.5\) (e.g @bouwens2015 [@bowler2015; @bowler2017; @ishigaki2017]). These limits translate to higher redshift \(\sim 50\%\) completeness at M\(_{UV}\sim\)-18.3,-19,-19.5, and-19.7 (\(z\sim\) 4, 5, 6, and 7, respectively). These limitations become problematic when comparing the results between LBG and narrowband samples. Presumably due to not requiring a continuum detection, the \(W_{Ly\alpha}\) observed in narrowband surveys are systematically higher (e.g. @zheng2014). Our dataset and results can be useful for characterizing the effects that selection techniques can have on Ly\(\alpha\) emission. Even though it is possible to identify these selections directly in our sample, any comprehensive analysis must consider our galaxy selection procedure. Since our sample follows neither the M\(_*\) or M\(_{UV}\) functions of the galaxy population, correcting requires assumption of M\(_*\) and/or M\(_{UV}\) distributions. Nevertheless, we can still simulate Ly\(\alpha\) emission samples on CANDELS galaxies using our results of \(W_{Ly\alpha}\) dependence on the M\(_{UV}-\beta\) plane given by Equations ([\[Afinal\]](#Afinal){reference-type="ref" reference="Afinal"}) and ([\[W_0final\]](#W_0final){reference-type="ref" reference="W_0final"}). In this section, we describe our simulation of a \(z \sim 4\) LBG sample from the 3D-HST catalogs. We then compare the properties of LBGs with the parent distribution, focusing on M\(_*\), SFR, M\(_{UV}\), and \(\beta\). We finally simulate Ly\(\alpha\) fractions for each sample and conclude on the effects of LBG selection at \(z \sim 4\). LBGs at \(z\sim 4\) are typically selected using B-dropouts and imposing color selections on redder filters. As we are simulating this survey in CANDELS, our detection limits are given by the depth of their \(H_{160}\) images. For the \(z\sim 4\) LBG selection, we adopt the same methodology applied in: \[\begin{gathered}
\nonumber \mbox{B-dropouts: } (B_{435}-V_{606}>1.1)\ \wedge (B_{435}-V_{606} \\ \label{color} >(V_{606}-z_{850})+1.1) \wedge (V_{606}-z_{850}<1.6). \end{gathered}\] After color selection, we impose \(5\sigma\) detections in the \(V_{606}\) filter, to which we now refer to as the detection band. As commonly applied in LBG surveys, we also require all candidates to have \(<1.5\sigma\) measurements in the \(U\) band, since the Lyman break must be redshifted out of this filter. Before performing any analysis on the CANDELS LBG sample, we run EAZY and FAST according to our prescriptions (see Section [2.3](#2.3){reference-type="ref" reference="2.3"}). Of the \(\sim 3200\) CANDELS galaxies that comply with our selection, \(\sim 800\) classify as low-redshift interlopers (\(z_{EAZY}<3\)), according to our outputs. We check EAZY 3\(\sigma\) constraints for these interlopers and find most to have \(z_{99}<4\). Hence, we find a low-redshift contamination in \(z\sim 4\) CANDELS LBGs that is higher than the typically reported number of \(\sim 10 \%\) (@bouwens2007). From now on, we remove these presumed \(z_{EAZY}<3\) contaminants and work with a \(z\sim 4\) LBG sample of nearly 2300 objects. Using our outputs, we plot relevant properties of CANDELS LBGs in Figure [\[fig:16\]](#fig:16){reference-type="ref" reference="fig:16"}. To ensure that we do not venture way beyond CANDELS completeness, we restrict the upcoming analysis to galaxies with M\(_*>10^{8.5}\mbox{M}_{\odot}\). We construct two \(3<z<4.2\) CANDELS samples for comparison: all galaxies with M\(_*>10^{8.5}\mbox{M}_{\odot}\) and all \(V\)-band detected M\(_*>10^{8.5}\mbox{M}_{\odot}\) objects. As seen in the central panel, the major LBG selection effect is associated with the detection band threshold. These left-out objects are either red, heavily extincted galaxies, or blue, intrinsically faint objects (@quadri2007). Indeed, as revealed by the comparison between LBGs and the detection-band-selected sample, any selection imprints associated to color requirements are minor. It is still worth noting, however, that these color criteria seem to neglect blue rather than red galaxies at \(z\sim 4\) (center of Figure [\[fig:16\]](#fig:16){reference-type="ref" reference="fig:16"}). We find the \((B-V>1.1)\) Lyman break cut to be the driving criterion behind this selection effect (see Equation [\[color\]](#color){reference-type="ref" reference="color"}). We remark that these insights might not be valid at different redshifts or lower M\(_*\), however. Some studies have indeed explored differences at higher redshift (e.g. @jiang2013a [@jiang2013b; @jiang2016]). In analyses that explore Ly\(\alpha\) emission strength, UV continuum properties, morphologies, and sizes, they find no significant differences between LBGs and LAEs at \(z\geqslant 6\). The insights we present can be further tested by simulating the fraction of LAEs above an \(W_{Ly\alpha}\) threshold. We use our results from Section [5.5](#5.5){reference-type="ref" reference="5.5"} to estimate the Ly\(\alpha\) fraction for each sample and plot our results in the right panel of Figure [\[fig:16\]](#fig:16){reference-type="ref" reference="fig:16"}. Indeed, depending on the depth of the detection band, the fraction of low UV luminosity galaxies composing the sample can vary significantly. As a consequence, deeper surveys can potentially recover a higher fraction of line emitters at high \(W_{Ly\alpha}\). To deal with this selection, it has been proposed to compare the Ly\(\alpha\) fractions over narrow UV luminosity samples (@stark2011 [@mallery2012; @ono2012; @schenker2012]). Our results in this paper further emphasize the need to compare galaxies of the same luminosity and properly characterize completeness when studying the evolution of the Ly\(\alpha\) fraction with cosmic time. Regarding color selections, we find the Ly\(\alpha\) fraction to be slightly lower for LBGs than for the \(V\)-band-detected sample. As we suggested, this is a consequence of LBG color cuts leaving out some of the bluest galaxies.
## Narrowband Samples {#6.2}
Ly\(\alpha\)-emitting galaxies can also be selected using narrowband imaging or blind spectroscopy (e.g. @gronwall2007 [@ouchi2008; @adams2011]). By establishing a \(W_{Ly\alpha}\) detection threshold between the narrow-and broadband flux measurement, this technique allows for efficient line-emitter selection. Since only line detection is required, such surveys can trace fainter objects than the LBG technique, possibly leading to selection of the youngest and faintest galaxies at high redshift. However, even though most narrowband measurements of LAEs are followed up by spectroscopy, any sample selection effects induced by the narrowband technique are already present in the sample. The \(W_{Ly\alpha}\) threshold used for detection, which is determined by the ability to separate low-redshift interlopers from high-redshift LAEs, can adopt a wide range of values. Depending on redshift, observer frame thresholds translate into different rest-frame \(W_{Ly\alpha}\) cuts. For instance, select sources with \(W_{Ly\alpha}>20\)Å at \(z\sim 2.1\), select \(W_{Ly\alpha}>9\)Å at \(z\sim 4.5\), and go as low as \(W_{Ly\alpha}>5\)Å at \(z\sim 2.23\). If \(W_{Ly\alpha}\) selections induce important biases on galaxy samples, the comparison of different surveys is not straightforward. In this section, we explore the effects that narrowband selections have on the population of LAEs, focusing on the M\(_*\)-SFR plane. The insights we present here are based on Equations ([\[Afinal\]](#Afinal){reference-type="ref" reference="Afinal"}) and ([\[W_0final\]](#W_0final){reference-type="ref" reference="W_0final"}), i.e., our M\(_{UV}-\beta\) model. We show in Figure [\[fig:17\]](#fig:17){reference-type="ref" reference="fig:17"} the outcome of \(W_{Ly\alpha}\) (top) and line-flux selections (bottom) on the M\(_*\)-SFR sequence. The red contours show \(3.5<z<4.5\) CANDELS galaxies with M\(>10^{8.5}\mbox{M}_{\odot}\). For clarity, however, we concentrate our analysis on the main sequence, removing objects with young and old ages (red dots). This approach allows our results to be dominated by objects optimally fitted by our FAST executions. To explore the effect of \(W_{Ly\alpha}\) selections (top panel of Figure [\[fig:17\]](#fig:17){reference-type="ref" reference="fig:17"}), we base our analysis on the \(z\sim 2.23\) narrowband survey of, whose galaxies are limited to \(W_{Ly\alpha}>5\)Å and fluxes \(>2\times 10^{-17}\) erg cm\(^{-2}\) s\(^{-1}\) (5\(\sigma\)). We simulate a flux-limited-only Ly\(\alpha\) sample and use it as baseline (black). We then show the region where \(W_{Ly\alpha}>5\)Å (1\(\sigma\); blue) and \(W_{Ly\alpha}>20\)Å (1\(\sigma\); green) objects lie. The \(W_{Ly\alpha}>5\)Å contours are intended to reproduce selections, whereas the more restrictive selection \(W_{Ly\alpha}>20\)Å is representative of multiple surveys (@gronwall2007 [@hagen2014; @vargas2014]). As confirmed by this plot, galaxies with higher characteristic \(W_{Ly\alpha}\) (i.e., low M\(_*\), low-SFR) are more likely to be selected by narrowband samples, even though the effect is minor for the typical narrowband cuts of \(W_{Ly\alpha}\leqslant20\)Å. Still, selections based solely on \(W_{Ly\alpha}\) systematically fail to remove AGNs and neglect bright Ly\(\alpha\) emitters (@sobral2016). All these insights remark on the importance of using low \(W_{Ly\alpha}\) cuts and thorough interloper controls. After narrowband outcomes are used for sample selection, spectroscopic observations follow. However, these follow-ups of Ly\(\alpha\) emitting galaxies are flux limited, just like the ones presented in this work. Still, as we consider completeness in our modeling and simulations of Section [5.5](#5.5){reference-type="ref" reference="5.5"}, we can still make inferences on flux limited studies. We use our Monte Carlo simulation outputs to also assess the effects of line luminosity selections. Using the corresponding M\(_*\), SFR, and \(f_{UV}\) for every object, we obtain the probability of \(L_{Ly\alpha}>L_{Ly\alpha}^*\) for every galaxy. The outcome (bottom panel of Figure [\[fig:17\]](#fig:17){reference-type="ref" reference="fig:17"}) reveals that flux selections bias samples toward high-SFR LAEs. Discrepancies in the location of LAEs in the M\(_*\)-SFR plane have already been observed in the literature. In Figure 10 of their paper, plot the M\(_*\)-SFR relation of their \(z\sim 2-3\) LAEs alongside the \(z\sim 2.1\) counterparts of. The objects are part of the HETDEX Pilot Survey (@adams2011), for which detections are constrained to line fluxes \(>7-10\times 10^{-17}\) erg cm\(^{-2}\) s\(^{-1}\) (5\(\sigma\)) and \(W_{Ly\alpha}>20\)Å. In contrast, the flux depth is higher than that of HETDEX, reaching \(2\times 10^{-17}\) erg cm\(^{-2}\) s\(^{-1}\) (5\(\sigma\)), while also selecting sources with \(W_{Ly\alpha}>20\) Å. Therefore, are comparing two \(W_{Ly\alpha}\)-selected surveys with different line-flux depths. Our results in this section would, then, predict LAEs to sample lower M\(_*\) and SFRs because they go deeper (see \(\sim 10^9\) M\(_{\odot}\) galaxies in Figure [\[fig:17\]](#fig:17){reference-type="ref" reference="fig:17"}). This is in fact the pattern observed in Figure 10 of, i.e., we can qualitatively reproduce their comparison with our simulations from the results of Section [5.5](#5.5){reference-type="ref" reference="5.5"}. This explanation is backed by the fact that these discrepancies are not caused by inconsistent M\(_*\) or SFR derivations. Both studies assume a IMF, cSFHs, and a fixed metallicity of \(Z=0.2Z_{\odot}\). and find M\(_*<10^9\)M\(_{\odot}\) LAEs to lie above the M\(_*\)-SFR relation. A similar trend is reported in the work of, who study Ly\(\alpha\) emitters down to \(10^6\)M\(_{\odot}\). point out that this offset can be a consequence of how uncertain SFRs are for low M\(_*\), starburst galaxies. This explanation is consistent with the fact that most of our M\(_*<10^9\)M\(_{\odot}\) Ly\(\alpha\) emitters are constrained to the lowest age locus of the M\(_*\)-SFR plane (see Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"}). Moreover, observe the same lowest age, low M\(_*\) Ly\(\alpha\) emitters at \(z\sim 4.5\). To answer whether low M\(_*\) LAEs lie above starforming galaxies or the slope of the sequence changes toward M\(_*<10^9\)M\(_{\odot}\), different approaches are required. For instance, no discrepancies are found between \(z\sim 2\) LAEs and H\(\alpha\) sources (see @mathee2016), although most of their galaxies have M\(_*>10^9\) M\(_{\odot}\). In summary, evidence suggests that \(z>2\), M\(_*<10^9\) M\(_{\odot}\) LAEs lie above extrapolations of the M\(_*\)-SFR relation. It is unclear whether this offset is real or the position of the relation at low M\(_*\) is far from certain.
# Inferences on the \(4\leqslant \lowercase{z}\leqslant 7\) Ly\(\alpha\) Fraction {#7}
Probably the most important use for Ly\(\alpha\) emission is tracing the neutral hydrogen fraction in the IGM. Several studies over the last decade have constrained the fraction of Ly\(\alpha\)-emitting galaxies as a function of redshift (@stark2010 [@stark2011; @ono2012; @schenker2012; @tilvi2014; @cassata2015]), bringing us closer to the goal of constraining the epoch of reionization. The Ly\(\alpha\) fraction is fairly well understood at \(z<5\) (@stark2011 [@cassata2015]), with most efforts nowadays focusing on \(z\sim7,8\) (@ono2012 [@treu2013; @tilvi2014; @furusawa2016]). Along these lines, our characterization of \(W_{Ly\alpha}\) on the M\(_{UV}-\beta\) plane can be used to simulate \(W_{Ly\alpha}\) distributions at higher redshift. In this section, we apply the observed M\(_{UV}-\beta\) relations from and LFs from to simulate high-redshift \(W_{Ly\alpha}\) distributions. By means of these simulations, we can predict the Ly\(\alpha\) fraction in galaxies up to \(z\sim 7\), providing the first semi-analytical constraint to this tracer toward the reionization epoch. Furthermore, we also simulate dropouts from CANDELS photometry to explore the effects of observational limitations on the inferred Ly\(\alpha\) fractions at high redshift. Our results from this section are based on assuming the same \(z\sim 4\) \(W_{Ly\alpha}\) dependence on M\(_{UV}-\beta\) at every redshift. We remark that testing this assumption with currently available datasets is challenging, since different sample selections and incompleteness levels can affect the observed \(W_{Ly\alpha}\) dependence on the M\(_{UV}-\beta\) plane. It is also worth noting that the analysis described in this section does not account for any effects related to changes in the merger fraction, IGM opacity (e.g. @gunn-peterson1965 [@becker2001; @fan2006]), and/or conditions of the ISM with cosmic time (e.g. @carilli2013). As the number of complete, unbiased Ly\(\alpha\) surveys grows (e.g. @cassata2015 [@hathi2016]; this work), we will be able to test these assumptions and explore their dependence on redshift. We now detail our simulation of high-redshift LBG samples. For the LFs, we use 's best Schechter parameters (excluding CANDELS-EGS). We start by drawing objects following the LF and then associate a UV slope according to the best-fit relations from Table 3 of. We finish by adding an intrinsic scatter of 0.35 to the slopes at every redshift (@bouwens2012 [@castellano2012; @bouwens2014]). Similarly, we perform an analogue procedure following the M\(_{UV}\) distribution of CANDELS LBGs, which allows us to assess the effect of magnitude incompleteness. In order to do so, we make use of the optical and IR photometry publicly available from the 3D-HST catalogs (@skelton2014). For every redshift, we use the selections from, since they are based on CANDELS photometry: \[\begin{gathered}
\nonumber \mbox{\(z\sim\) 4: } (B_{435}-V_{606}>1) \wedge (i_{775}-J_{125}<1)\\ \wedge (B_{435}-V_{606}>1.6(i_{775}-J_{125})+1) \end{gathered}\] \[\begin{gathered}
\nonumber \mbox{\(z\sim\) 5: } (V_{606}-I_{814}>1.3) \wedge (I_{814}-H_{160}<1.25)\\ \wedge (V_{606}-I_{814}>0.72(I_{814}-H_{160})+1.3) \end{gathered}\] \[\begin{gathered}
\nonumber \mbox{\(z\sim\) 6: } (I_{814}-J_{125}>0.8) \wedge (J_{125}-H_{160}<0.4) \\ \wedge (I_{814}-J_{125}>2(J_{125}-H_{160})+0.8) \end{gathered}\] \[\begin{gathered}
\nonumber \mbox{\(z\sim\) 7: }(I_{814}-J_{125}>2.2) \wedge (J_{125}-H_{160}<0.4) \\ \wedge (I_{814}-J_{125}>2(J_{125}-H_{160})+2.2). \end{gathered}\] Similarly to our Section [6.1](#6.1){reference-type="ref" reference="6.1"} selection, we impose S/N cuts in the detection bands. For \(z\sim 4\) and \(5\) dropouts, we require \(3\sigma\) detections in the \(V_{606}\), \(z_{850}\) bands. For \(z\sim 6\) and 7 dropouts, we impose at least \(3\sigma\) detections in the \(Y_{102}\) and \(J_{125}\) bands, respectively. To associate a dropout redshift with every galaxy, however, we use the photometric redshifts from 3D-HST instead of the actual dropout from where it was selected. By doing so, we are in agreement with the methodology of. We finally complete these CANDELS LBG samples by associating UV slopes to every galaxy. Just as for the complete sample, we associate the slopes following the M\(_{UV}-\beta\) relations from. We present in Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} the assumed M\(_{UV}-\beta\) sequences and the UV luminosity distribution for both sets under consideration. Since we are starting from the M\(_{UV}-\beta\) characterization, the steepness of the linear relation increases with redshift. The contours we show in Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} (left panel) are obtained after incorporating the intrinsic scatter, which we assume to be of 0.35. Note that the addition of such a scatter does not render the redshift evolution of the M\(_{UV}-\beta\) distribution negligible. According to our Equations ([\[Afinal\]](#Afinal){reference-type="ref" reference="Afinal"}) and ([\[W_0final\]](#W_0final){reference-type="ref" reference="W_0final"}), this implies that the probability of observing \(W_{Ly\alpha}\) in emission at fixed M\(_{UV}\) should increase with redshift. The right panel of Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} shows the absolute magnitude distribution of \(4\leqslant z \leqslant 7\) LBGs as found by. Note the growth of the LF as a function of redshift. Since the LF becomes steeper as the redshift increases, the probability of observing a high-\(W_{Ly\alpha}\) galaxy should also increase. One more important point regarding Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} must be noted. As initially suggested, CANDELS incompleteness affects the observed M\(_{UV}\) distribution of LBGs (shown as crosses). The fact that there is more contribution from brighter galaxies could, in principle, bias samples toward lower \(W_{Ly\alpha}\) when compared to the LFs. We now focus on quantifying all these effects within the framework of our model. We simulate the \(W_{Ly\alpha}\) distributions for the two sample sets we constructed. To do so, we use the M\(_{UV}\) and \(\beta\) of each object to draw \(W_{Ly\alpha}\) distributions following our sampling of Equations ([\[Afinal\]](#Afinal){reference-type="ref" reference="Afinal"}) and ([\[W_0final\]](#W_0final){reference-type="ref" reference="W_0final"}). This methodology allows us to simulate the \(W_{Ly\alpha}\) distribution for the number of galaxies in every redshift subsample, properly taking into account the uncertainties on each M\(_{UV}-\beta\) domain and the errors in our modeling. We present the corresponding \(4\leqslant z \leqslant 7\) Ly\(\alpha\) fractions for such galaxies in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"}. To this end, we select using the thresholds \(W_{Ly\alpha}^*\)= 25 Å and \(W_{Ly\alpha}^*\)=55 Å (@stark2011). Still, as we note in Section [5.3](#5.3){reference-type="ref" reference="5.3"}, LBGs with M\(_{UV}\) distributions dominated by fainter objects can yield higher fractions of Ly\(\alpha\). In recent works, characterization has been performed separately in galaxies with M\(_{UV}<-20.25\) and M\(_{UV}>-20.25\) (@stark2011 [@ono2012; @schenker2012; @tilvi2014; @cassata2015]). Following the same approach, we now focus our analysis on samples in the ranges \(-21.75<\)M\(_{UV}<-20.25\) and \(-20.25<\)M\(_{UV}<-18.75\) (Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"}), which are well covered by our dataset (see Figure [\[fig:14\]](#fig:14){reference-type="ref" reference="fig:14"}). We plot in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"} the resulting Ly\(\alpha\) fractions. We find the complete fractions (i.e., derived from the LFs; black circles) to be consistent with a steady increase up to \(z\sim 7\), which has been previously assumed as baseline for measuring drops in the Ly\(\alpha\) fraction (@stark2010 [@stark2011; @ono2012]). Apart from backing this assumption, our result provides a more robust baseline for comparison. We reiterate that a closer look at Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} reveals the explanation for such trend, at least within our approach. First, the LF of LBGs becomes steeper as redshift increases (@bouwens2015). Second, the M\(_{UV}-\beta\) relation favors bluer slopes as redshift increases. The combination of both relationships explain an increase in the Ly\(\alpha\) fraction as a function of redshift for LBGs. It has already been discussed (@blanc2011 [@cassata2015]) how the growth in the Ly\(\alpha\) fraction is likely tied to changes in the \(f_{esc}\) of Ly\(\alpha\) photons, at least for \(3<z<5\). This change in the \(f_{esc}\) is likely tied to evolution in the dust content and ISM of galaxies, which we are empirically accounting for in this model. Under this picture, Lyman continuum (LyC) photon escape should follow a similar trend. Indeed, empirically predicts that LyC photon escape fraction increases with redshift at fixed M\(_*\). In summary, our results are in strong agreement with the picture that UV photon escape fraction from galaxies tends to decrease with cosmic time. From the literature, we include in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"} the measurements from, ,, ,, , and. The overall literature trend is an increase in the Ly\(\alpha\) fraction up to \(z\sim5\), which flattens at \(z\sim6\) and drops toward \(z\sim7\). Departures from this increasing regime are typically associated with changes in the transparency of the IGM, with drops toward \(z\sim 7\) associated with a more neutral universe (@stark2010 [@ono2012]). Even though such inferences might be correct, the contrast between the completeness-corrected (black circles) and-uncorrected (black crosses) fractions in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"} suggests that the driving factor behind these \(z\sim7\) drops might be sample M\(_{UV}\) incompleteness. Our simulated \(z\geqslant 6\) CANDELS LBG samples do not follow the corresponding LF as a result of survey depth (see Figure [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"}), leading them to include more luminous objects than a population-representative sample. Therefore, as seen in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"}, drops in the Ly\(\alpha\) fraction are induced. As we point out in Section [6.1](#6.1){reference-type="ref" reference="6.1"}, the deepest LBG surveys nowadays (e.g @bouwens2015 [@bowler2015; @bowler2017; @ishigaki2017]) reach \(\sim 50\%\) completeness at M\(_{UV}\sim\)-18.3,-19,-19.5, and-19.7 for \(z\sim\) 4, 5, 6, and 7, respectively. Therefore, Ly\(\alpha\) studies of even the deepest LBG surveys are sufficiently incomplete to bias the reported Ly\(\alpha\) fractions (see Figures [\[fig:18\]](#fig:18){reference-type="ref" reference="fig:18"} and [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"}). Taking into account this effect is essential if we are to use Ly\(\alpha\) to further constrain the reionization epoch during the next decade. It is important to clarify the limitations of our modeling. First, our characterization is based on the whole \(W_{Ly\alpha}\) distribution, which may misrepresent the actual fraction of high-\(W_{Ly\alpha}\) galaxies. In fact, the analysis we perform in Section [4](#4){reference-type="ref" reference="4"} reveals that a log-normal profile might be better for reproducing the tail of the distribution (Figure [\[fig:4\]](#fig:4){reference-type="ref" reference="fig:4"}). For example, find an exponential profile to only adequately represent the low-\(W_{Ly\alpha}\) end of the distribution, leading them to model the high-\(W_{Ly\alpha}\) tail separately. As a consequence, we encourage the reader to focus most of their analysis on our \(W_{Ly\alpha}^*\)= 25 Å results (top panel in Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"}). Along those lines, we expect our modeling to underestimate the fraction for populations with significantly higher-\(W_{Ly\alpha}\) tails. Indeed, our fractions for galaxies with \(-20.25<\)M\(_{UV}<-18.75\) are systematically lower than measurements (see Figure [\[fig:19\]](#fig:19){reference-type="ref" reference="fig:19"}). In summary, our results suggest that the Ly\(\alpha\) fraction steadily increases up to \(z\sim 7\) in an ionized universe. However, drops in this fraction are not necessarily related to changes in the opacity of the IGM, since we find that sample M\(_{UV}\) incompleteness can alternatively explain them. In fact, the uncertainties in both literature measurements and our predictions are still large enough to set assertive constraints on the Ly\(\alpha\) trends toward \(z\sim 7\) (see @robertson2015). Our results, therefore, further highlight that Ly\(\alpha\) constraints on \(z_{re}\) (the redshift at which the fraction of ionized hydrogen is 0.5) are still inconclusive. It is worth discussing whether \(z_{re}>7\) is in agreement with observational reionization constraints besides Ly\(\alpha\). Recently, the analysis of CMB anisotropies has yielded \(z_{re}\sim 8.2^{+1.0}_{-1.2}\). Systematic detection of Ly\(\alpha\)-emitting galaxies in the range \(z\sim 7-9\) (@zitrin2015 [@furusawa2016; @stark2017]) still leads to the question whether reionization is already in place at \(z\sim 7\). Several studies have also explored Ly\(\alpha\) LFs toward \(z\sim 9\), although observational limitations require deeper/wider surveys to obtain more meaningful conclusions (@sobral2009). Therefore, and since reionization is predicted to happen so rapidly (@finlator2009), claims of \(z_{re}>7\) are not in tension with measurements of the Ly\(\alpha\) fraction at \(z\sim 7\). Interestingly enough, it is not even clear whether reionization happens first around fainter, lower density environments ("outside-in,\" @kashikawa2006), or brighter, higher density regions ("inside-out,\" @santos2016). This overall picture suggests that the role of M\(_{UV}\) in reionization studies will no longer restrict itself to sample control, but it will also extend to parameterizing property. Therefore, the natural approach for future studies would be to constrain reionization as a smooth function of both redshift and UV luminosity/environment (e.g. @robertson2013).
# Conclusions {#8}
In this work, we present an exhaustive analysis of Ly\(\alpha\) emission at \(3<z<4.6\). To this end, we M\(_*\)-select 625 galaxies from the CANDELS survey, allowing us to study Ly\(\alpha\) emission over a diverse and heterogeneous galaxy sample. We conduct spectroscopic observations of our targets with the M2FS, a multi-fiber spectrograph at the Clay 6.5m telescope. We then use a Bayesian approach for proper statistical handling of our results. By means of this framework, we are capable of characterizing Ly\(\alpha\) emission in the high-redshift galaxy population. In summary, our conclusions are the following. 1. We present a Bayesian methodology to measure the \(W_{Ly\alpha}\) distribution considering the completeness in fluxes measured and the uncertainties in both spectroscopy and photometry. We combine this approach with Monte Carlo simulations for robust modeling of observed trends. Combining all of these features allows us to properly state the significance and uncertainties in our results. 2. We take advantage of the Bayesian framework to give an insight into how to compare multiple \(W_{Ly\alpha}\) distribution models from a quantitative standpoint. In this paper, we explore the extent to which exponential, Gaussian, and log-normal probability distributions can reproduce \(W_{Ly\alpha}\) measurements. We conclude that, in our case, an exponential profile is the most adequate. However, as we find in Section [5](#5){reference-type="ref" reference="5"}, this profile can struggle to reproduce the tail of high-\(W_{Ly\alpha}\) populations (e.g. samples with low M\(_*\), SFR, UV luminosity, or \(\beta\)). 3. Our measured \(W_{Ly\alpha}\) and \(f_{esc}\) strongly anti-correlate with M\(_*\). We associate these trends with the higher dust fraction and gas mass in more massive galaxies, boosting the scattering and absorption of Ly\(\alpha\) photons. We model both exponential \(W_{Ly\alpha}\) distribution parameters (\(A, W_0\)) using linear relations dependent on M\(_*\), and find them both to anti-correlate with M\(_*\). Our modeling is also capable of reproducing our observed \(W_{Ly\alpha}\) distributions when binning the sample (Figure [\[fig:7\]](#fig:7){reference-type="ref" reference="fig:7"}). 4. We also explore the dependence of Ly\(\alpha\) emission on M\(_{UV}\). We find \(W_{Ly\alpha}\) and \(f_{esc}\) to be typically higher for UV faint objects, which have been previously observed in the literature (@stark2010 [@schaerer2011]). Since \(z\sim 4\) galaxies seem to follow an M\(_*-\)M\(_{UV}\) sequence (@gonzalez2014), this result is consistent with our M\(_*\) trends. This confirms that the role played by higher dust fraction and gas mass rules \(W_{Ly\alpha}\) dependence on UV luminosity. We also observe \(W_{Ly\alpha}\) and \(f_{esc}\) to anti-correlate with the SFR and UV slope, confirming the qualitative Ly\(\alpha\) escape scheme presented. 5. The relatively uniform location of our targets in the M\(_{UV}-\beta\) plane allows us to characterize the \(W_{Ly\alpha}\) distribution in this space. Once more, we use a Bayesian framework and Monte Carlo simulations to obtain linear representations of exponential \(W_{Ly\alpha}\) distribution parameters. Our results suggest that the probability of observing Ly\(\alpha\) emission from a galaxy is determined by its dust content, whereas the actual magnitude of \(W_{Ly\alpha}\) is mostly ruled by the UV luminosity of the object. This characterization also allows us to simulate \(W_{Ly\alpha}\) distributions from M\(_{UV}-\beta\) galaxy samples (see below). 6. Using CANDELS, we mimic a \(z\sim4\) LBG survey based on B-dropouts. We simulate \(W_{Ly\alpha}\) distributions from the LBG sample and verify that the lower probability of measuring high \(W_{Ly\alpha}\) when compared to the population is mostly a consequence of detection band limitations. In other words, shallower LBG surveys are bound to include fewer low-M\(_*\) galaxies in their samples, decreasing the probability of including high-\(W_{Ly\alpha}\) objects. We find color constraints imposed by the LBG technique to also lower the Ly\(\alpha\) fraction, although the effect is minor. These results highlight the importance of comparing surveys with similar sensitivities and M\(_{UV}\) distributions. 7. Our results on the M\(_{UV}-\beta\) plane also allow us to explore the effects \(W_{Ly\alpha}\) an line-flux limitations induce on Ly\(\alpha\) narrowband surveys. We find \(W_{Ly\alpha}\) cuts to bias samples toward low-M\(_*\) objects, whereas flux limitations seem to preferentially leave out low-SFR galaxies. These insights contribute to explain the location of LAEs within the main sequence (e.g. some of the differences between @vargas2014 and @hagen2014). 8. We generate \(4\leqslant z \leqslant 7\) LBG samples following reported M\(_{UV}-\beta\) relations and LFs from the literature (@bouwens2014 [@bouwens2015]). Assuming the same \(W_{Ly\alpha}\) dependence on M\(_{UV}-\beta\) that we find at \(z\sim 4\), we estimate the fraction of Ly\(\alpha\) emitting galaxies in LBG samples above an \(W_{Ly\alpha}\) threshold. Our findings are consistent with observational measurements in the literature that suggest an increase in the fraction up to \(z\sim 6\) (@stark2010 [@stark2011; @cassata2015]). This result constitutes the first semi-analytical constraint to the Ly\(\alpha\) fraction at \(z\sim7\), replacing extrapolations of lower redshift regimes. 9. We simulate the \(4 \leqslant z \leqslant 7\) Ly\(\alpha\) fraction in LBG samples drawing from CANDELS M\(_{UV}\) distribution. We conclude that M\(_{UV}\) incompleteness can lower the Ly\(\alpha\) fraction, reproducing drops at \(z\sim 7\). Claims of drops at this redshift need to thoroughly account for this effect if they are to be used to constrain the reionization epoch. This result highlights that reported drops in the Ly\(\alpha\) fraction are not inconsistent with \(z_{re}>7\). Given the uncertainties in both predictions and measurements, more work is needed to conclude on the actual Ly\(\alpha\) trends toward \(z\sim 7\) and beyond. We thank the referee for useful feedback that improved the quality of this paper. G.O. was supported by CONICYT, Beca Magíster Nacional 2014, Folio 22140924. G.B. was supported by CONICYT/FONDECYT, Programa de Iniciación, Folio 11150220. V.G. was supported by CONICYT/FONDECYT, Programa de Iniciación, Folio 11160832. We also acknowledge the NSF/MRI grant for the M2FS, number AST 0923160. This work is based on observations taken by the 3D-HST Treasury Program (GO 12177 and 12328) with the NASA/ESA HST, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. This research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile.
[^1]: For the redshift range of our galaxies, the H\(\beta\), O II , and O III emission lines fall between the \(H_{160}\) and 3.6\(\mu m\) bands. H\(\alpha\) only contributes to the 3.6\(\mu m\) *Spitzer*/IRAC band for \(z>3.75\).
[^2]: The NASA/IPAC Extragalactic Database (NED) is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. | {'timestamp': '2017-06-12T02:01:37', 'yymm': '1706', 'arxiv_id': '1706.01886', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01886'} |
# Introduction
The experimental discovery of large magnetic anisotropy in single Co atoms on a Pt(111) surface has given hope that systems as small as one atom can be used for information storage in real-life applications. The magnetic properties of single adsorbed atoms have since been investigated for a large range of atom-substrate combinations using low-temperature scanning tunneling microscopy (STM) and X-ray magnetic circular dichroism (XMCD). Both techniques allow to obtain information on the magnetic anisotropy, i.e. the zero field splitting (ZFS) of the magnetic states, caused by the crystal field, and the magnetization decay, i.e. the lifetime \(T_1\), caused by magnetization tunneling . In most of the cases reported, the lifetime of the magnetic state was found to be short, with the atom switching between two degenerate antiparallel configurations on a timescale of pico-to microseconds. The interaction of the atom with conduction electrons of the substrate is an important effect limiting the stability of these states . Two ways of reducing the interaction were proposed: one by physically separating the atom and the conducting substrate by a thin insulator , the other by using rare-earth atoms, where localized 4\(f\) electrons, responsible for the magnetic moment, are shielded from the substrate electrons . Both approaches extended the magnetic lifetimes. Lifetimes to the order of minutes at zero magnetic field were found for Holmium atoms on Pt(111) followed by an even higher stability in Ho atoms on MgO , persisting in the presence of magnetic field. Dy atoms on graphene on Ir were also found to be stable up to 1000 s . In the latter three reports, the role of the symmetry of the crystal field for stability was highlighted. In this work we focus on the influence of symmetries on magnetic stability and provide a theoretical framework of stability criteria for rare-earth systems that follow logically from the combination of symmetries of the magnetic atom and its environment. We start with the analysis of \(C_q\) point group (\(q\)-fold rotational symmetry) as the one relevant for atoms adsorbed on a crystal surface, and later extend the results to \(C_{qv}, C_{qh}, D_q, D_{qh}, D_{qd}\) and \(S_{2q}\) point groups. The results are thus applicable not only to single magnetic atoms on surfaces but also to larger complexes (e.g. magnetic molecules) containing a single spin. One can also imagine removing the limitation of rare-earth atoms by considering an effective spin. Further, we do not consider phonon-mediated processes due to their ill-defined angular momentum and for the moment neglect the role of the nuclear spin. During the preparation of this manuscript several other theoretical works analysing magnetic stability have been published . This work, however, allows stability analysis without restricting the ground state spin or the magnetic field direction. Further, we extend the stability analysis to higher-order processes and higher symmetries. Interestingly, unconstrained analysis gives different predictions about stability, although identical Hamiltonians are studied. We discuss the differences in the results at the end of the paper.
# Symmetries of the Hamiltonian {#symmetries}
A magnetic atom in an anisotropic environment can be generally described by the following Hamiltonian: \[\mathcal{H} = \mathcal{H}_\mathrm{el} + \mathcal{H}_\mathrm{SO} + \Hs{CF},\] where \(\mathcal{H}_\mathrm{el}\) describes the Coulomb interactions within the atom, \(\mathcal{H}_\mathrm{SO}\) is the spin-orbit interaction and is the crystal field, i.e. the electrostatic interaction of the atom with the environment. For rare earth atoms, the spin-orbit coupling is generally stronger than the effect of the crystal field. Therefore, the last term can be treated as a perturbation on the free atom's lowest \(J\) multiplet. We can denote the components of the multiplet as \(\ket{L,S,J,m}\equiv\ket{m}\) with fixed values for \(L,S\) and \(J\) and \(m =-J,-J+1, \hdots, J\). We choose the quantization axis \(z\) along the highest order rotation axis, i.e. perpendicular to the surface when atom on a surface is considered. Note that for transition metals, the crystal field usually exceeds the spin-orbit interaction leading to the partial or full quenching of the orbital momentum. Further, the atomic states can be a complex mixture of several electronic configurations. Thus, often an effective spin is used to describe the low energy part of the atomic states. In this case, care has to be taken when analyzing magnetic stability. As has already been reported , two inherent symmetries of play an important role for the properties of the atom wavefunction: the rotational symmetry and the time-inversion symmetry. **Time-inversion symmetry:** The Hamiltonian is due to the electrostatic interactions and thus has to commute with the time reversal operator (see ). is antiunitary, i.e. for any states \(\ket{a}\) and \(\ket{b}\), \(\bracket{a}{b}=\bracket{\T b}{\T a}\). further anticommutes with all components of the total angular momentum \(\vec{J}\), such that \(\T\vec{J} =-\vec{J}\T\) and \(\T\hat{J}_\pm =-\hat{J}_\mp \T\). The angular momentum eigenstates transform according to \(\T\ket{m} = (-1)^{m}\ket{-m}\) which implies \(\T^2 = 1\) for integer \(J\) and \(\T^2 =-1\) for half-integer \(J\) (see e.g. Refs. ). Since commutes with , for any eigenstate \(\ket{\Psi}\) also the time-reversed state \(\T\ket{\Psi}\) is an eigenstate of with the same energy. As time reversal also changes the sign of the expectation value of of such a state, we conclude that the states with non-zero \(\expect{\J{z}}\) form degenerate pairs [^1]. If singlet states exist, their \(\expect{\J{z}}\) is zero, i.e. they are effectively non-magnetic and do not change under time reversal (modulo a phase). Such states are only possible in systems with integer \(J\), as for half-integer \(J\) the spectrum consists exclusively of Kramers' doublets. Since we are interested in magnetization tunneling, we only consider doublet states in the following. Higher order degeneracies are only possible with fine-tuning the parameters of , which we explicitly exclude [^2]. We note that singlet states also usally come in pairs, as symmetric and antisymmetric combinations of two magnetic states with opposite \(\expect{\J{z}}\), and additional energy terms might reverse singlet formation. We will consider this case later in the section on the influence of magnetic field. **Rotational symmetry:** In \(C_q\) symmetry, the environment of the magnetic atom is invariant under a rotation by \(2\pi/q\), and so also commutes with the rotation operator (see ). Therefore, we can choose the eigenstates of the system to be eigenstates of both and . Such eigenstates are superpositions of \(\ket{m + nq}\) kets, as all such kets acquire the same phase \(\phi = 2\pi m/q\) under rotation. In the following we use the example of \(C_5\) symmetry (Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}), although it is not found on crystal surfaces. Five-fold rotational symmetry does, however, occur in magnetic molecules, and has recently attracted interest regarding magnetic stability . Most importanly, it is the lowest symmetry that demonstrates all the relevant effects. A more practically-minded reader can use Fig. [\[fig_c\_all\]](#fig_c_all){reference-type="ref" reference="fig_c_all"} as a reference, that also contains examples of all effects, spread over more practically relevant systems. In the aforementioned figures the \(q\) rotation classes are visualized as points on a circle in the complex plane (Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}a & b). The points represent the phases \(\phi\) gained by the application of the rotation operator. A given eigenstate belongs to one and only one class. Note that for states \(\ket{a}\) and \(\ket{b}\) belonging to different classes, \(\matx{a}{\J{z}}{b}\) is zero. This follows from the fact that and commute. The same relation can be achieved for two degenerate states from the same class by choosing the corresponding eigenvector basis. This choice removes the ambiguity in the definition of eigenstates and maximizes \(|\expect{\J{z}}|\) for each state [^3], mimicking the interaction between the single spin and the electron bath . Such a choice of eigenstates has one more important consequence. Since an eigenstate \(\ket{\Psi}\) belongs to a particular class and consists of a well defined set of \(\ket{m}\) basis states, its time reversed counterpart \(\T\ket{\Psi}\) consists of \(\ket{-m}\) basis states and thus belongs to a class with a reversed sign of \(\phi\) (which might be the same class). On the class diagram (Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}a), these two classes are symmetric about the \(x\) axis. If the two classes are different, \(\ket{\Psi}\) and \(\T\ket{\Psi}\) are automatically orthogonal, and thus \(\T\ket{\Psi}\) must be the other state of the doublet up to a phase factor. We can write for two states \(\ket{\Psi^+}\) and \(\ket{\Psi^-}\) belonging to the same doublet \(\T \ket{\Psi^+} = e^{i\chi}\ket{\Psi^-}\). The phase \(\chi\) may be different in every doublet.
# Magnetization tunneling
In an isolated system, any eigenstate is indefinitely stable. For atoms adsorbed on the surface, however, the magnetic atom becomes an open quantum system via the interaction with the substrate. The magnetization can become unstable if we consider the interaction between the local 4\(f\) moment and the bath electrons. The scattering of these electrons on the atom within the \(J\) multiplet can be described by the operator \(\mathcal{V}\) . \[\mathcal{V} = \vec{J} \cdot \vec{\sigma} = \J{z}\sigma_z + \frac{1}{2} \left( \J{+}\sigma_-+ \J{-}\sigma_+ \right)\,,\] acting in the product space of the atom and the scattering electron (\(\vec{\sigma}\) is the spin of the latter). We note that the standard exchange coupling would include the magnetic atom's spin instead of the total angular momentum, but the Wigner-Eckart theorem ensures that \(\vec{S}\) can be written as \(\vec{J}\) up to a proportionality factor. This interaction couples different eigenstates of the atom. A system prepared in one of the two ground states can undergo a complete magnetization reversal, i.e. electron-assisted magnetization tunneling to the other ground state. We can characterize the stability of the system by the value of the matrix element \(V_n = \bra{\Psi_g^-}\mathcal{V}^n\ket{\Psi_g^+}\), where \(\Psi_g^\pm\) are the two ground states of the system (remember that we only consider cases where the ground states belong to a doublet). Namely, if \(V_1\) is nonzero, a single electron scattering event may reverse the magnetization of the atom. The energy cost of this process is zero, so that the switching will take place at arbitrary low temperatures and would eventually lead to Kondo screening . If \(V_1\) is zero, one can consider higher order matrix elements, corresponding to interactions with several electrons. Since \(\mathcal{V}\) operates in a product space of the atom and the interacting electrons, \(V_n\) are operators in the electron spin space. The equality \(V_n = 0\) is to be understood as every component of it in the \(J\) space being equal to zero. Thus, we will consider matrix elements of the form \(\matx{\Psi_g^-}{\J{i_n}\dots\J{i_1}}{\Psi_g^+}\), with \(i_k =x,y,z\) or, equivalently \(i_k = +,-, z\). If all such matrix elements are zero, \(V_n\) is a zero operator. Note that such matrix elements describe coherent multi-electron interactions, where the intermediate states are virtual states. This is analogous to co-tunneling . Thus, this analysis describes zero-energy transitions of the magnetic state of the atom. At finite temperature we additionally expect incoherent multi-electron interactions, where intermediate states are thermally populated, i.e. there is energy transfer between the atom and the bath. This is in analogy to sequential tunneling. In that case the relevant matrix elements are of the form \(\matx{\Psi_g^-}{\J{i_n}}{\Psi_{i_{n-1}}}\dots\matx{\Psi_{i_2}}{\J{i_2}}{\Psi_{i_1}}\matx{\Psi_{i_1}}{\J{i_1}}{\Psi_g^+}\) and a full solution requires solving the complete master equation. As will be shown, the symmetries of the system lead to \(V_1\) being zero for particular classes, thus protecting the ground state doublet from magnetization tunneling. **Protection by rotational symmetry:** Let us consider a ground state \(\ket{\Psi_g^+}\) from class \(2\pi g/q\) and the state \(\mathcal{V}\ket{\Psi_g^+}\). The three components of \(\mathcal{V}\), namely , and affect the class of the state differently (see arrows in Figure [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}a and [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}b). \(\mathcal{V}\ket{\Psi_g^+}\) is a linear combination of eigenstates with classes \(2\pi g/q\) (from \(\J{z}\ket{\Psi_g^+}\)), \(2\pi (g+1)/q\) (from \(\J{+}\ket{\Psi_g^+}\)) and \(2\pi (g-1)/q\) (from \(\J{-}\ket{\Psi_g^+}\)). The other ground state \(\ket{\Psi_g^-}\) is from class \(-2\pi g/q\). It is evident that \(V_1 = \matx{\Psi_g^-}{\mathcal{V}}{\Psi_g^+}\) will be exactly zero unless \(-g\) is equal to either \(g \text{ or } g\pm1\). This means that the matrix element will vanish when the start and end classes are not neighbors on the circle in the complex plane, or equivalently a single scattering event with a conduction electron can only transfer \(\pm1\hbar\). For example consider a doublet \(\pm3\pi/5\) from Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}c. \(\mathcal{V}\ket{\Psi^{\frac35\pi}}\) will only have components from \(\pi/5, 3\pi/5 \text{ and } \pi\), and so \(\matx{\Psi^{-\frac35\pi}}{\mathcal{V}}{\Psi^{\frac35\pi}} = 0\), as indicated in the table in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}d by a zero on a blue background. This is the protection mechanism reported in XMCD studies for Ho atoms on MgO and for Dy atoms on graphene . For higher-order electron scattering processes, that can transfer more than one \(\hbar\), the number of classes in \(\mathcal{V}^n\ket{\Psi_g^+}\) increases and at a certain order, depending on the separation between the doublet classes, \(V_n\) ceases to be zero and zero-energy transitions involving \(n\) electrons become allowed. A prominent example of this is the \(\pm4\pi/5\) doublet in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}a for \(J=2\). In general, for \(2J+1\leq q\) there are at least as many classes as basis states and the states become eigenstates of . Thus for \(J=2\) there is no direct path between \(4\pi/5\) and \(-4\pi/5\) (i.e. \(\J{+}\ket{2} = 0\)). The shortest path in this case contains four transitions between classes, and so \(V_4\) is the first matrix element that is non-zero. A similar situation is realized in the \(\pi\) doublet in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}b for \(J=5/2\), as again the two ground states are just \(\ket{5/2}\) and \(\ket{-5/2}\), and at least five transitions are needed to switch between the two states. Note that the two states in this doublet remain eigenstates of also for higher values of \(J\), up to \(\frac{11}2\). This means that every intermediate state \(\J{i_1}\dots \J{i_n}\ket{\Psi_g^+}\) is also an eigenstate of , and thus the first non-vanishing matrix element is of the fifth order. We remind the reader that this higher-order protection only holds for coherent processes. For incoherent processes, the first interaction excites the atom from the ground state \(\ket{\frac52}\) into any eigenstate containing \(\ket{\frac32}\), \(\ket{\frac52}\) or \(\ket{\frac72}\). From there, a direct path to the other ground state, \(\ket{-\frac52}\), might exist depending on the class. We have marked the cases where this distinction matters by an asterisk at the appropriate \(J\) values in figures [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"} and [\[fig_c\_all\]](#fig_c_all){reference-type="ref" reference="fig_c_all"}. **Protection by time-inversion symmetry:** Consider again the two ground states of an atom with integer \(J\). As shown above, the two states are connected by time reversal \(\T\ket{\Psi_g^+} = e^{i\chi_g}\ket{\Psi_g^-}\). This also implies \(\T\ket{\Psi_g^-} = \T\,e^{-i\chi_g}\,\T \ket{\Psi_g^+} = e^{i\chi_g}\T^2\ket{\Psi_g^+} = e^{i\chi_g}\ket{\Psi_g^+}\). We now consider one component of \(V_1\), \(\matx{\Psi_g^-}{\J{i}}{\Psi_g^+}\) where \(i=x, y \text{ or } z\). Transforming such a matrix element we can write \[\begin{aligned}
\begin{split} \matx{\Psi_g^-}{\J{i}}{\Psi_g^+} &= \matx{\T\Psi_g^+}{e^{i\chi_g}\J{i}\,e^{-i\chi_g}\T}{\Psi_g^-}\\ &\qquad\qquad\Big.{\scriptstyle \J{i}\T =-\T\J{i} } \\ &=-\matx{\T\Psi_g^+}{\T\J{i}}{\Psi_g^-}\\ &\qquad\qquad\Big.{\scriptstyle \bracket{\T\alpha}{\T\beta} = \bracket{\beta}{\alpha}} \\ &=-\bracket{\J{i}\Psi_g^-}{\Psi_g^+} \\ &\qquad\qquad\Big.{\scriptstyle \bracket{O_H\alpha}{\beta} = \bracket{\alpha}{O_H\beta}} \\ &=-\matx{\Psi_g^-}{\J{i}}{\Psi_g^+}\. \end{split}\label{timerev} \end{aligned}\] Thus, for integer \(J\) any matrix element of the form \(\matx{\Psi_g^-}{\J{i}}{\Psi_g^+}\) is zero. E.g. the \(\pm4\pi/5\) doublet in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}a for \(J > 2\) is protected from first-order transitions. This is the protection mechanism reported by Miyamachi et al. for Ho atoms on Pt(111) . For half-integer \(J\) the argument above does not apply. Instead, time-reversal protection manifests itself in systems where two ground states belong to the same class. These remain magnetic and do not form singlets, since Kramers theorem protects the doublet from mixing and splitting. For those doublets \(\J{+}\ket{\Psi_g^+}\) and \(\J{-}\ket{\Psi_g^+}\) both belong to a different class than \(\ket{\Psi_g^+}\), so only \(\matx{\Psi_g^-}{\J{z}}{\Psi_g^+}\) might be non-zero. We have however chosen eigenstates that maximize \(\expect\J{z}\), setting these matrix elements to zero. Thus also in this case, \(V_1\) vanishes. An example of such doublet is the \(\pi\) class in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}c, as indicated in the table Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}d by a yellow background. Note that protection by time reversal only holds for first order electron scattering for both integer and half-integer \(J\) systems. In total, we have identified two different mechanisms that can protect the ground state from zero-energy magnetization flips up to a specific order of electron scattering. First, the states can be protected by time reversal symmetry (yellow background in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}c & [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}d). Second, the states can be protected by rotational symmetry (blue background in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}c & [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}d). The two cases can coexist for certain combinations (green background in Fig. [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}c & [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"}d).
# Influence of magnetic field
Let us now consider how a magnetic field \(\vec{B}\) applied to the system affects the protection mechanisms. Strictly speaking, such a field lifts the degeneracy of the doublets, and so transitions within the doublet cannot be excited elastically any more. For finite temperatures, however, the substrate electrons have enough thermal energy to flip the magnetization when \(JB \lesssim k_\mathrm{B}T\). Note that he magnetic field may induce other zero-energy transitions, when two states become degenerate due to the Zeeman energy. This is reflected by distinct steps in the hysteresis loop due to Zener tunneling and has been described in detail elsewhere. In the following, we consider two different situations. First, the external magnetic field is applied along the \(z\) direction. Adding a term \(\Hs{Z} = g_J\J{z} B_z\) to the Hamiltonian does not change the rotational symmetry but breaks time reversal symmetry. Thus we expect no changes to the classes of the eigenstates but they will be no longer connected by . For small fields we can treat the Zeeman term as a perturbation. The resulting perturbed ground states \(\ket{\Phi_g^\pm}\) can then be written as \[\begin{aligned}
\ket{\Phi_g^\pm} &= \ket{\Psi_g^\pm} + \sum_{\phi^\pm_k = \phi_g^\pm} \frac{\matx{\Psi_k^\pm}{\Hs{Z}}{\Psi^\pm_g}}{E_g-E_k}\ket{\Psi^\pm_k} + O(B_z^2)\,, \end{aligned}\] where \(\phi^\pm_k = \phi^\pm_g\) reflects the fact that the matrix element for the Zeeman term is only nonzero if \(\Psi_k^\pm\) is in the same class as \(\Psi_g^\pm\) [^4]. As expected, the perturbed states are in the same class as the unperturbed. Thus we only need to consider the breaking of time inversion symmetry, lifting that particular protection. Indeed, if we consider the components of \(V_1\) for the new states: \[\begin{aligned}
\begin{split} \matx{\Phi_g^-}{\J{i}}{\Phi_g^+} &= \cancelto{\,=0}{\matx{\Psi_g^-}{\J{i}}{\Psi_g^+}}\\ &- \sum_{\phi^-_k = \phi_g^-} \frac{\matx{\Psi_g^-}{\Hs{Z}}{\Psi^-_k}}{E_l} \matx{\Psi^-_k}{\J{i}}{\Psi_g^+} \\ &-\sum_{\phi^+_k = \phi_g^+} \frac{\matx{\Psi^+_k}{\Hs{Z}}{\Psi_g^+}}{E_k}\matx{\Psi_g^-}{\J{i}}{\Psi^+_k}\\ &+ O(B_z^2)\. \end{split} \end{aligned}\] Since \(\Psi^+_k\) and \(\Psi^-_k\) are connected by time reversal, we can use the same reasoning as in Eqn. [\[timerev\]](#timerev){reference-type="ref" reference="timerev"}, to show that \[\begin{aligned}
\matx{\Psi_a^\pm}{\Hs{Z}}{\Psi_b^\pm} &=-e^{i(\chi_a-\chi_b)}\matx{\Psi_b^\mp}{\Hs{Z}}{\Psi_a^\mp}\\ \matx{\Psi_a^\pm}{\Hs{Z}}{\Psi_b^\mp} &=-(-1)^{2J}e^{i(\chi_a-\chi_b)}\matx{\Psi_b^\pm}{\Hs{Z}}{\Psi_a^\mp} \end{aligned}\] and to rewrite the first sum in terms of \(\Psi^+_k\): \[\begin{split} \sum_{\phi^-_k = \phi_g^-} \frac{\matx{\Psi_g^-}{\Hs{Z}}{\Psi^-_k}}{E_k} \matx{\Psi^-_k}{\J{i}}{\Psi_g^+} = \\ (-1)^{2J}\sum_{\phi^+_k = \phi_g^+} \frac{\matx{\Psi^+_k}{\Hs{Z}}{\Psi_g^+}}{E_k}\matx{\Psi_g^-}{\J{i}}{\Psi^+_k}\. \end{split}\] The resulting first order correction to \(V_1\) is thus \[\left[1 + (-1)^{2J}\right]\sum_{\phi^+_k = \phi_g^+} \frac{\matx{\Psi^+_k}{\Hs{Z}}{\Psi_g^+}}{E_k}\matx{\Psi_g^-}{\J{i}}{\Psi^+_k}\] and vanishes for half-integer \(J\). Note that the second-order terms are a product of three matrix elements, and the corresponding term contains \(\left[1-(-1)^{2J}\right]\), which does not vanish for half-integer \(J\) in the general case. Therefore, \(V_1\) is proportional to \(B_z\) for integer and to \(B_z^2\) for half-integer \(J\), as noted in yellow boxes in Figures [\[fig_c5\]](#fig_c5){reference-type="ref" reference="fig_c5"} & [\[fig_c\_all\]](#fig_c_all){reference-type="ref" reference="fig_c_all"}. Thus, we expect an increase of flipping rate of the magnetic moment proportional to \(B_z^2\) and \(B_z^4\) for integer and half-integer spins, respectively. This effect was reported in an STM study for Ho atoms on Pt(111) . Note that we can expect the same dependency for any perturbation proportional to . This includes e.g. the diagonal part of the exchange interaction between two nearby atoms 1 and 2 proportional to \(\J{z}^{(1)}\J{z}^{(2)}\). Thus, the lifetime of the magnetization is expected to drop for Ho on Pt(111) when increasing the coverage of Ho adatoms. This is also in agreement with numerical calculations by Hübner et al. , where the magnetic field dependency of the matrix element was computed for the case \(\expect{\J{z}}=8, q=3\). For integer \(J\) where the crystal field leads to the formation of pairs of singlet states, the magnetic field plays the opposite role. The Zeeman energy reverses the symmetric/antisymmetric pairing due to the crystal field and partially restores magnetic states. \(V_1\) is then inversely proportional to \(B_z\). For example in the case of Ho with \(J=8\) and an \(m=\pm8\) ground state in four-fold symmetry, an application of magnetic field results in magnetic states with long lifetimes . This reduces to the situation mentioned above, in which Zener tunneling would induce a step in the hysteresis curve near zero magnetic field. Again, the same effect would be observed for any perturbation proportional to . As shown elsewhere also the coupling to the electron bath can undo singlet formation leading to a doublet of ground states with short lifetimes. In the second case the field is applied perpendicular to \(z\) (in-plane). Such a field breaks both symmetries and quenches the magnetic moment along \(z\). Only in the case that this field is applied along a different high-symmetry direction (e.g. an in-plane two-fold rotation axis in \(D_3\)) can we expect some remnant magnetization. For the general case of a tilted magnetic field, no simple statement can be made on the power of the leading order term and effects like cross-over may arise.
# Practical symmetries
We can now apply the above analysis to systems of practical value, i.e. in 2-, 3-, 4-and 6-fold symmetric environments. The results, which are presented in detail in Fig. [\[fig_c\_all\]](#fig_c_all){reference-type="ref" reference="fig_c_all"}, can be summarized as follows:
- For integer \(J\)
- if the two ground states belong to the same class, they are non-magnetic singlets,
- if the two ground states belong to different classes, at least **two** electrons are needed for zero-energy magnetization switch.
- For half-integer \(J\)
- in 3-fold (6-fold) symmetric environments for the ground state containing \(\ket{\pm3/2}\) at least **two** (**three**) electrons are needed for zero-energy magnetization switch,
- otherwise **one** electron provides for zero-energy magnetization switch.
For systems with small \(J\), higher-order stability can be expected. See Fig. [\[fig_c\_all\]](#fig_c_all){reference-type="ref" reference="fig_c_all"} for details. Finally, for the simple case of rare earth atoms in the oxidation state 3+ and a giving ground states of maximal \(\expect{\J{z}}\), i.e. the situation best suited for magnetic data storage, we give a simple table that summarizes the findings (Table [\[fig_re\]](#fig_re){reference-type="ref" reference="fig_re"}). We note that the best candidates for magnetically stable systems are Sm and Ce in 6-fold symmetric adsorption sites.
# Higher symmetries
We now analyze the stability of degenerate doublets in other axial symmetries with a \(q\)-fold rotation axis. We show that these symmetries behave either like \(C_q\) or like \(C_{2q}\). In the latter case the stability is greatly enhanced---e.g. in a \(D_{5d}\)-symmetric environment there are 10 classes, so that ground states in a magnetic atom with integer \(J\) are protected up to second order in most of the cases.
## \(C_{qv}\) and \(D_q\)
\(C_{qv}\) adds a vertical mirror plane \(\sigma_v\) to \(C_q\). The \(\sigma_v\) operator anticommutes with and does not commute with : \(\sigma_v\R = \R[-1]\sigma_v\). Thus, the eigenstates we chose will not be eigenstates of \(\sigma_v\), and their composition will be the same as for \(C_q\). Marciani et al. have shown that the presence of the mirror plane introduces only a weak constraint on the switching process . \(D_q\) adds a 2-fold rotation axis perpendicular to the main axis to \(C_q\). The operator behaves exactly like \(\sigma_v\) with respect to and the components of \(\vec{J}\). Thus, again, the stability in \(D_q\) will follow the same rules as in \(C_q\).
## \(C_{qh}\) and \(D_{qh}\)
The rules change in the environments with a horizontal mirror plane \(\sigma_h\). The \(\sigma_h\) operator commutes with and anticommutes with , and . Thus \(\ket{m}\) are eigenstates of \(\sigma_h\) with eigenvalue sign alternating between neighboring \(m\)'s. We can write \(\sigma_h\ket{m} = (-1)^{J+m}\ket{m}\). \(\sigma_h\) also commutes with , and thus the eigenstates of can be chosen to be eigenstates of both and \(\sigma_h\). In fact, the eigenstates we chose in section [2](#symmetries){reference-type="ref" reference="symmetries"} are already eigenstates of both operators. This is easy to see in the case of doublets where the two states belong to different classes, as there the choice is unique. For same-class doublets we note that they only exist for half-integer \(J\) and odd \(q\) and are composed of kets with \(m = q(n + \frac12)\). If such a state is an eigenstate of \(\sigma_h\), then for every \(n\) it can contain either \(\ket{m}\) or \(\ket{-m}\), but not both, since \(\matx{-m}{\sigma_h}{-m} = (-1)^{2m}\matx{m}{\sigma_h}{m} = (-1)^q\matx{m}{\sigma_h}{m} =-\matx{m}{\sigma_h}{m}\). It follows that for every \(n\) one of the doublet's states will contain \(\ket{m}\), while the other \(\ket{-m}\), and so the matrix element of between these states will be zero, consistent with out choice from section [2](#symmetries){reference-type="ref" reference="symmetries"}. Now every state is characterized by the rotational phase \(\phi\) and the eigenvalue of \(\sigma_h\) (\(\pm1\)). Note that for even \(q\) all the states with the same \(\phi\) gain the same sign after reflection. Thus for even \(q\) the number of classes does not change, and the same analysis applies. For odd \(q\), however, half of the \(\ket{m}\) kets gain one sign and the other half the other, splitting every class in two. In this case we can reclassify the states according to a new phase \(\phi_{2q} = \frac{\pi m}q\), that would normally arise in \(C_{2qv}\) symmetry. The new classes combine states with a same eigenvalues of and \(\sigma_h\) and have the same properties we have used before, namely:
1. \(\J{\pm}\) operators switch the class one step forward or backward along the unit circle,
2. operator does not affect the class,
3. connects classes with opposite sign of \(\phi_{2q}\).
Thus the stability rules in \(C_{qh}\) are the same as the stability rules in \(C_q\) for even \(q\) and \(C_{2q}\) for odd \(q\). The \(D_{qh}\) behaves the same way, since the additional symmetry does not have a strong effect on stability.
## \(D_{qd}\) and \(S_{2q}\)
The \(S_{2q}\) point group is generated by the rotation-reflection operator \(\Sq = \Rhalf\sigma_h\) with \(\Sq\Sq = \R\). commutes with and anticommutes with , and . \(\Sq\ket{m} = (-1)^{J+m}e^{i\pi m/q}\ket{m}\). The choice of eigenstates as eigenstates of both \(\mathcal{H}\) and \(\Sq\) again maximizes \(\expect{Jz}\) for every doublet. We can thus use the phase \(\phi_s = \pi m (1 + 1/q)\) to classify the states. In the case of even \(q\) we obtain \(2q\) classes, while in the case of odd \(q\) there are only \(q\) classes. Based on this, we can reclassify the states again based either on \(\phi_{2q}\) or \(\phi\), depending on the parity of \(q\). As in the previous case, this preserves the composition of classes and the action of , , and with respect to the unit circle of classes. It follows that the stability rules in \(S_{2q}\) are the same as the stability rules in \(C_q\) for odd \(q\) and \(C_{2q}\) for even \(q\). The \(D_{qd}\) has additionally a vertical mirror plane, that does not alter stability rules. Table [1](#similar){reference-type="ref" reference="similar"} summarizes the found relations of higher order symmetries to the behavior of the simpler \(C_q\) symmetry.
# Magnetic stability of Ho atoms
There have been several reports of stable and unstable single rare earth atom systems, mostly based on Holmium single atoms. Here we review their results in the light of our stability criteria. Ho atoms on three-fold symmetric adsorption sites on Pt(111) have been investigated by STM and XMCD . Donati et al. report a ground state with \(J=8, m=\pm6\) (class \(0\)), obtained by fitting XMCD spectra based on multiplet calculations. In this configuration there are two singlet states, that will mix in a high magnetic field, leading to a hysteresis-free magnetization curve, as measured in the experiment. Ab initio calculations from the paper by Miyamachi et al., on the other hand, predict a ground state with \(J=8, m=\pm8\) (classes \(\pm2\pi/3\)). Such ground states are protected by time reversal symmetry, so we expect long lifetimes, unless a magnetic field is applied. The STM experiments showed long lifetimes at zero field that become shorter in small magnetic fields. The energy of an inelastic excitation observed in the experiment, also agrees with the theoretically predicted \(J\) multiplet structure. Interestingly, an ensemble of Ho atoms with time-reversal-protected ground state doublet would also show no magnetic hysteresis under the experimental conditions of the XMCD measurements (\(T=\SI{2.5}{\kelvin}\), \(\approx\SI{10}{\second}\) per point). At zero field the average magnetization of Ho atoms is zero, as the two stable magnetization orientations are equally probable. The application of a magnetic field along \(z\) splits the doublet, introducing a preferential orientation, while at the same time increasing the switching rate quadratically, leading to an overall increase in average magnetization. At high fields the two states split so far that switching rate becomes negligible and the magnetization saturates. As the magnetic field is reduced, the switching rate increases and then decreases again for vanishing magnetic field. Since the energy difference between the two states also goes down to zero, we again expect equal population of the ground states at zero field and no hysteresis. The stability at vanishing magnetic fields would be also influenced by the distance between Ho atoms due to exchange or dipolar interactions. As has been shown above, any operator linear in \(J\) acts like a magnetic field. Note that the STM experiments have been carried out at extremely low coverage , while the XMCD data was taken at a coverage of 0.04 % . Thus the STM and the XMCD results agree on the absence of hysteresis, but disagree on the multiplet structure. The question about the configuration of the crystal field and the ground state of this system remains open. Shick et al. have proposed that the crystal field parameters and thus \(\expect{\J{z}}\) of the ground state changes with magnetic field, due to an interaction between \(5d\) and \(4f\) orbitals, such that in low magnetic fields one expects \(\expect{\J{z}} \approx 8\), and in high magnetic fields \(\expect{\J{z}} \approx 6\). Similar to Ho on Pt(111), STM experiments and XMCD experiments on Ho atoms in 4-fold symmetric adsorption sites on MgO fundamentally disagree on the class of the ground state. The ground state deduced from XMCD spectra is the class \(\pm\pi/2\) doublet (\(\expect\J{z}=4.66\)). It is protected from switching by rotation symmetry and so the magnetization curves show a hysteresis. The estimated crystal field parameters predict the first excited state 4.5 meV higher than the ground state. At the same time, the STM experiments report the first excited state at 73 meV above the ground state. Natterer et al. estimate \(m=\pm8\) (class \(0\)) from the Ho magnetic moment. This ground state should be non-magnetic, but as the experiments are done in a mostly in-plane magnetic field, the two singlet states would mix and acquire a magnetic moment. Transitions between such mixed states are, in general, allowed, but the actual transition probability depends on the strength of the crystal field. In the STM setup the magnetic state was stable for several hours, likely indicating weak transversal anisotropy and, as expected, weak interaction with the electron bath. We would expect little or no hysteresis in this system, as at zero field the magnetic states should lose their magnetization. As in the previous case, the origin of the disagreement between STM and XMCD data on the energy spectrum in Ho on MgO remains open and needs further investigation. Possible explanations such as different MgO thickness and averaging over several adsorption sites in the XMCD experiment can only partially explain the discrepancies.
# Comparison to other theoretical results
## M. Marciani et al. (2017)
The paper by Marciani et al. uses an approach very similar to ours. However, we disagree with several points of the analysis. First, in the case of \(q=3\) Marciani et al. write that for \(J=9/2\) the ground state is not stable, based on their analysis of the structure of the Hamiltonian. This is not confirmed by a direct construction of an appropriate Hamiltonian and calculation of the transition rates. The relevant block of the Hamiltonian in the basis \(\ket{-\frac92}, \ket{-\frac32}, \ket{\frac32}, \ket{\frac92}\) has a general form \[\begin{bmatrix} 2d & a & b & 0\\ \bar{a} & 0 & 0 & b\\ \bar{b} & 0 & 0 &-a\\ 0 & \bar{b} &-\bar{a} & 2d \end{bmatrix},\] plus a constant term, and can be diagonalized analytically. The eigenstates with maximal \(\expect\J{z}\) within every doublet are not connected by either , or . Thus first-order electron transitions are forbidden between these states. Second, Marciani et al. assume large uniaxial anisotropy, such that \(\expect\J{z}\) of the ground state is very close to \(J\). For the case where the ground state has a smaller \(\expect\J{z}\), they propose using a smaller \(J_\mathrm{eff}\approx\expect\J{z}\) for analysis of the ground state stability. This, however, might lead to wrong conclusions if \(2J \geq q\) while \(2J_\mathrm{eff} < q\), as the latter suggests stable states. Consider, for example, \(J=\frac72\) in \(C_6\), and two doublets from the \(\phi=\pm\frac56\pi\) class pair, one with \(\expect{\J{z}} \approx \frac72\) and the other with \(\expect{\J{z}} \approx \frac52\). Both are unstable, as they contain states from neighboring classes. However, if the second doublet has lower energy, Marciani et al. suggest using \(J_\mathrm{eff} = \frac52\) to determine the stability. But in a system with \(J=\frac52\) the eigenstates of are eigenstates of , and the doublet \(\expect{\J{z}}=\frac52\) is stable up to fourth order. Using this conclusion for the \(\expect{\J{z}} \approx \frac52\) doublet in \(J=\frac72\) system is definitely unwarranted. To avoid such mistakes, one could amend the definition of \(J_\mathrm{eff}\) to include the class of the doublet in question instead of its \(\expect{\J{z}}\). That is, one should choose \(J_\mathrm{eff}\) equal to the highest \(m\) that belongs to the same class. For the example above that would be \(m = \frac72\) for both doublets, and thus the same conclusion on stability will be reached. One should keep in mind, however, that such a definition would only work for single-electron processes, as higher-order stability also depends on the neighboring classes. As an example, consider the \(\pm\frac23\pi\) doublet for \(J=2\) and \(3\) in \(C_6\). \(J_\mathrm{eff}\) is equal to 2 for both cases, but for \(J=3\) second-order transitions are possible at non-zero temperature, while for \(J=2\) they are forbidden.
## M. Prada (2017)
The paper by M. Prada classifies single atom systems by introducing a geometric phase, as a phase difference between two equivalent paths connecting the two ground states. However, this phase was introduced based on the equality \(\T\R = \R[-1]\T\), which, due to commutativity of and only holds for \(q=2\). The operators \(\T^+\) and \(\T^-\) are thus not equivalent to , as the paper claims, but to and , and the proposed phase is essentially double the eigenvalue of for the state. This has reprecussions throughout the paper, as for example, \(\T^+\) and \(\T^-\) do not anticommute with or , and thus Eqn. 15 in Ref. derived under the assumption \(\T^+\vec{J} =-\vec{J}\T^+\) becomes invalid. Through these mistakes the final classification scheme gives wrong results in almost one half of the cases: for half-integer \(J\) all doublets in 3-, 4-and 6-fold symmetric environments are listed as protected. A direct calculation shows that this is untrue. For example, let us consider the case of \(J=\frac72\) in \(C_{3v}\) and ground states from the \(\phi=\pm\frac\pi3\) class pair. The two relevant blocks of the Hamiltonian in \(\ket{-\frac72}, \ket{-\frac12}, \ket{\frac52}\) and \(\ket{-\frac52}, \ket{\frac12}, \ket{\frac72}\) bases have a general form \[\begin{bmatrix} e_1 & a & c\\ \bar{a} & e_2 & b\\ \bar{c} & \bar{b} & e_3 \end{bmatrix} \text{ and } \begin{bmatrix} e_3 &-b & c\\-\bar{b} & e_2 &-a\\ \bar{c} &-\bar{a} & e_1 \end{bmatrix}.\] We could now calculate the \(V_1\) analytically, but since it is enough to show that \(V_1\) is not zero for specific values of the constants, we choose to calculate matrix elements \(\matx{\Phi_g^-}{\J{\pm}}{\Phi_g^+}\) numerically for \(e_1=-1\), \(e_2=0\), \(e_3=-0.5\), \(a=b=0.01\), \(c=0.0001\). These values satisfy the criteria of \(\expect{\J{z}} \approx J\), out-of-plane easy axis and small transversal terms. We get a ground state energy of approximately-1.0001 and \(\matx{\Phi_g^-}{\J{-}}{\Phi_g^+} = 0.0004\), \(\matx{\Phi_g^-}{\J{+}}{\Phi_g^+} = 0\). Based on our analysis of the problem, we believe that it is impossible to classify all the possible systems based on a single number. However, if we ignore the complications arising at small \(J\) and in magnetic field, and focus on first-order interactions, it becomes possible. First, we note that for integer \(J\) everything except single states is protected, and single states gain a phase \(\phi_g = 0\text{ or }\pi\) under rotation by . Thus we write that the doublet is stable unless \(\cos^2{\phi_g} = 1\). For half-integer \(J\) the doublet is stable, unless the two states belong to neighbouring classes. This we write as \(\cos^2{\phi_g} = \cos^2\frac{\pi}q\). We combine the two cases and introduce a number \(\eta = \cos^2{\phi_g}-\cos^2{(\pi J)}-\cos^2\frac{\pi}q\sin^2{(\pi J)}\). Now given an eigenstate that belongs to a rotation class with phase \(\phi_g\), we conclude that this state belongs to a magnetic doublet, protected in first order against magnetization tunneling, unless \(\eta = 0\). To additionally check the correctness of our results, numerical calculations have been performed for \(q=2,3,4,5\text{ and }6\) and \(J\) up to 12 using Hamiltonians based on Stevens' operators . The corresponding [python]{.sans-serif} code is made available on Github .
# Conclusions
We have analyzed the stability of magnetization in arbitrary doublets in 2-, 3-, 4-and 6-fold symmetric adsorption sites. We show that depending on the symmetry, the value of \(J\) and the ground state doublet, different order processes are needed for a zero-energy magnetization switch. The most stable doublet requires a 5th order scattering process to switch (\(C_6: J=5/2, m=5/2\)). In real life systems the protection will not be absolute. Presence of magnetic fields (particularly in-plane) and Jahn-Teller distortion can break the symmetries that protect the magnetization. Further, at \(T > \SI{0}{\kelvin}\) also phonons can induce transitions, as they distort the lattice, temporarily inducing mixing of the states. This breaks both protection mechanisms. Similarly, the hyperfine interaction may lead to additional constraints on magnetization tunneling that depend on magnetic field. However, in the limit of low fields and a thermal population of the nuclear spin states, the nuclear spins may be treated as a simple heat bath in analogy to the electron heat bath. When taking the lowest order interaction of the electronic total angular momentum \(J\) with the nuclear spin \(I\) of the form \(A\vec{I}\vec{J}\), the selection rules are identical to electron scattering not qualitatively changing magnetic stability. Only, if simultaneous spin flips of a conduction electron and a nuclear spin are considered, the rules for elastic magnetization reversal change and stability can only be sustained for protection of at least second order. The difference between a theoretically stable and theoretically unstable system can however make a decisive difference in the lifetime of the magnetic state. The considerations above would limit the lifetime of any system. Magnetic stability cannot be achieved over arbitrary times. For applications of single atoms on surfaces for quantum information processing, additionally the decoherence time \(T_2\) is of importance. A functioning device must operate on the quantum states faster than the decoherence time and the magnetization lifetime. We note that maximally stable states (long lifetimes) are also maximally classical. If one prepares a superposition of the two ground states, e.g. \(\frac1{\sqrt2}\left(\ket{\Psi_g^+} + \ket{\Psi_g^-}\right)\), it only takes the elastic interaction with a single electron of the bath to completely dephase the superposition. | {'timestamp': '2017-11-13T02:03:22', 'yymm': '1706', 'arxiv_id': '1706.01760', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01760'} |
null | null |
null | null |
null | null |
# Introduction
In recent years numerous experiments have successfully trapped, or formed a beam of, antihydrogen atoms. The general aim of these experiments is to perform spectroscopic measurements in order to precisely compare the properties of the antihydrogen atom to that of the hydrogen atom, thus enabling a direct test of the CPT symmetry. Other experiments aim at testing the effects of gravity using antimatter. Production of antihydrogen is performed either with antiproton and positron plasmas injected into cryogenic multi-ring electrode traps where three-body recombination takes place (\(\bar{\mathrm{p}} + 2e^{+} \rightarrow \bar{\mathrm{H}} + e^{+}\)), or by the charge exchange processes between positronium, antiproton, and antihydrogen (\(\mathrm{Ps}^{(*)} + \bar{\mathrm{p}} \rightarrow \bar{\mathrm{H}} + e^{-}\) and \(\bar{\mathrm{H}} + \mathrm{Ps} \rightarrow \bar{\mathrm{H}}^{+} + e^{-}\)). In most of the cases, an antihydrogen event is detected by its annihilation signature whereby several charged pions are emitted from the annihilation vertex. These are distinguished from cosmic or instrumental background events using the hit multiplicity and the infered positions of the tracks and annihilation vertex. Previous experiments have detected trapped antimatter annihilation events using tracking and vertex-finding approaches, predominantly relying on algorithms adopted from High-Energy Physics (HEP). However, antimatter trap experiments are not designed for high-resolution and high-efficiency tracking performance; the necessary layers of vacuum chamber walls, multi-ring electrodes, and thermal isolation limits the resolution of the detector. Therefore, the vertex-reconstruction approach may not be optimal for detecting antihydrogen event signals. In this work, we present a machine learning approach for analyzing data from antimatter trap experiments that does not require explicit vertex reconstruction. In particular, we propose the use of deep learning, essentially large neural networks, to identify antihydrogen annihilation signatures in the low-level detector data. This approach has the advantage of being able to exploit subtle and previously-unrecognized characteristics in the data, as demonstrated in Higgs-search and Higgs-decay studies from HEP. We train our system using Monte Carlo simulated data of antihydrogen-like signal and antiproton background annihilation events, and compare the performance to a traditional track and vertex reconstruction method. Although the results presented in this paper are obtained using a specific antimatter trap and detector environment, the approach is general and can be applied to other environments.
# [\[sec:Experiment\]]{#sec:Experiment label="sec:Experiment"}Monte Carlo simulation and annihilation event reconstruction
Signal and background annihilation events were simulated with the ASACUSA experimental setup. The primary antiproton annihilation process, the trap environment, and the Asacusa Micromegas Tracker (AMT) detector were simulated with the Geant4 toolkit. The Monte Carlo data includes the full material budget of the cryogenic vacuum trap to account for the dominant effect of multiple Coulomb scattering of pions while emerging from the annihilation events and subsequently traversing the surrounding materials. Hits produced by the simulated tracks from annihilation events are recorded and saved for analysis according to the AMT detector details. Briefly, the AMT detector consists of two, half-cylinder, curved micro-strip pattern gaseous detector layers using Micromegas technology , and a single, full-cylinder layer of plastic trigger scintillator bars sandwiched in between the Micromegas layers, segmented into eight bars. The Micromegas single hit resolution was simulated with Garfield++, Magboltz, and Heed programs for the response of charged pions of \(p \simeq 100\) MeV/c momentum and was found to be much better (\(\sigma_{\mathrm{hit}} \simeq 250\) \(\mu\mathrm{m}\)) than the amount of smearing of the pion track positions (\(\sigma_{\mathrm{Coul.}} \simeq 2\) mm) due to Coulomb scattering when the pions reach the AMT detector layers. Using the detected hits, the reconstructed vertex position resolution was estimated to be \(\sigma_{vx} \simeq 1\) cm. The vertex position resolution may be used as a design parameter to discriminate antihydrogen signal events from antiproton or other background events. However, in numerous antimatter experiments aiming at trapping or forming a beam of antihydrogen the material budget is not optimized for annihilation vertex resolution but for antiproton and positron trapping performance. Antihydrogen experiments trap charged particles (antiprotons and positrons) on the multi-ring electrode trap axis. The continuous annihilation of trapped antiprotons on the residual gas atoms produces background annihilation events. When a neutral antihydrogen atom is produced as a result of recombination of an antiproton with a positron, it may escape the electric trapping potentials and annihilate on the inner wall of the multi-ring electrodes. However, the difference in the annihilation signature of antiprotons or antihydrogen is only due to the annihilation vertex position. Therefore, in order to simulate antihydrogen signal events, we generated antiprotons to annihilate on the multi-ring electrode material, and to simulate antiproton background events annihilation process on residual gas atoms was used to generate events at the central axis of the trap. This scenario reflects realistic conditions of antihydrogen production observed in experiments. In a typical track and vertex reconstruction algorithm, the recorded hit data is processed as follows: the single detector channels are searched for entries above threshold, and neighboring channels are iteratively clustered to form hits. The detected hits are then used to form and fit tracks, using a Kalman-filtering algorithm. After fitting all hit pair combinations in an event, the track candidates are filtered by their compatibility with the known antiproton cloud position. Finally, the selected track candidates are paired and their three-dimensional point of closest approach position is assigned as the vertex position. The result of such a reconstruction approach are illustrated in Fig [\[fig:XYPos\]](#fig:XYPos){reference-type="ref" reference="fig:XYPos"} for the case of antihydrogen-like signals (vertices distributed on the round multi-ring electrode walls) and antiproton background events (vertices confined to the center of the trap). In the ASACUSA experiment, tracking detector layers are only present in the upper half of the trap, which is reflected in the distribution of reconstructed vertices. The shape of the distribution also reflects the smearing of the position by 1 cm due to the Coulomb scattering of the charged pions in the trap material. Event classification is based on the radial position of the reconstructed vertex, \(R=\sqrt{X^2+Y^2}\), but the classes are not perfectly separable because low resolution results in heavy overlap in the distributions of \(R\) for signal and background. Thus, a probabilistic model is used to classify the events. In this work, we use a simple neural network model with a single input \(R\), one hidden layer of 100 hyperbolic tangent units, and a logistic output unit that returns a probability of an event being an antihydrogen annihilation. The cluster finding algorithm, track and vertex reconstruction algorithms, and neural network are all tuned from simulations and experimental test data.
# [\[sec:dl\]]{#sec:dl label="sec:dl"}Deep Learning Approach
We propose a deep learning approach in which a deep neural network model is trained to classify annihilations directly from raw detector data. This automated, end-to-end learning strategy can potentially identify discriminative information in the raw data that is typically discarded in the vertex-reconstruction algorithm. Two design choices were made regarding the neural network architecture in order to address the geometry of the detector: (1) data from micro-strips situated along the azimuth (\(\phi\)) and axial (\(Z\)) dimensions are processed along separate pathways that are then merged higher in the architecture; and (2) 1-D convolution and max-pooling layers account for translational invariances in these dimensions. The raw data consists of 1,430 binary values from the two detector layers and two micro-strip orientations: 246 inner and 290 outer azimuthal strips, and 447 inner and 447 outer axial strips. The inner and outer detector layers are treated as two associated input 'channels' in the convolutional neural network, where the outer azimuthal layer is downsampled with linear interpolation to match the dimensionality of the inner layer. A diagram of this architecture is shown in Figure [\[fig:deeparch\]](#fig:deeparch){reference-type="ref" reference="fig:deeparch"}.
# [\[sec:results\]]{#sec:results label="sec:results"}Results
One million annihilation events of each type were simulated and randomly divided into training (60%), validation (20%), and test subsets (20%). Because the particles emitted by annihilations often escape undetected, the vertex finding (VF) algorithm failed on 75% of these events: 20% did not result in any detector hits; 7% had a hit in only one detector; and 48% had hits in both detectors, but a vertex could not be reconstructed. Vertex reconstruction failed because either (1) there were not enough distinct hits to infer the presence of at least two tracks, or (2) the point of closest approach of the reconstructed tracks was greater than the threshold value of 1 cm. Thus, a direct performance comparison was only possible on the 25% of the test set events for which the VF algorithm succeeded, even though the deep learning approach provides a prediction for every event. Due to the complexity of the non-neutral plasma physics and atomic scattering processes, there is no physics model to our knowledge that would be able to predict the true axial distribution of the antihydrogen atoms in realistic experimental conditions. Therefore, the classifier should be invariant to translations in the axial dimension. This was achieved by augmenting the training data, in which simulated antihydrogen annihilations occured at \(Z = 100\) cm (see Figure [\[fig:heatmap\]](#fig:heatmap){reference-type="ref" reference="fig:heatmap"}), by randomly translating each event during training by \(t\) 'pixels' in the \(Z\) dimension, where \(t\) is sampled from the uniform distribution \(t \sim U(-100,347)\). The same augmentation was used on the validation and test sets to evaluate performance, but the performance was very similar on non-augmented data indicating that the network learned to focus on \(Z\)-invariant features. Various neural network architectures were trained in order to optimize generalization performance on the validation set. The best architecture consists of five 1-D convolutional layers with kernel sizes 7-3-3-3-3 (the size of the receptive fields for neurons in each layer), channel sizes 8-16-32-64-128 (the number of distinct feature detectors in each layer), and rectified linear activation . In order to account for translational invariance, each convolution layer is followed by a max-pooling layer with pool size 2 and stride length 2. The flattened representations from the two pathways are then concatenated and followed by two fully-connected layers of 50 and 25 rectified linear units, then a single logistic output unit with relative entropy error so that the output can be interpreted as a probability. During training, \(50\%\) dropout was used in the top two dense layers to reduce overfitting . The model weights were initialized from a scaled normal distribution as suggested by He et al. , then trained using the Adam optimizer (\(\beta_1=0.9, \beta_2=0.999, \epsilon=1e-08\)) with mini-batch updates of size 100 and a learning rate that was initialized to \(0.0001\) and decayed by 1% at the end of each epoch. Training was stopped when the validation objective did not improve within a window of three epochs. The models were implemented in Keras and Theano , and trained on a cluster of Nvidia Titan Black processors. Performance comparisons were made by plotting the Receiver Operating Characteristic (ROC) curve and calculating the Area Under the Curve (AUC). Figure [\[fig:roc\]](#fig:roc){reference-type="ref" reference="fig:roc"} shows that the deep neural network classifier outperforms the VF approach by a large margin on the test events for which a vertex could be reconstructed (87% vs. 76% AUC). Remarkably, the deep neural network achieves 78% AUC on a disjoint set of events for which a vertex could *not* be reconstructed, and 82% AUC on the union set containing all events for which both detector layers were hit (not shown). This effectively triples the event coverage---to 73% of all events---while *simultaneously* improving the AUC by more than 5%. These results clearly demonstrate that useful information contained in the raw data is being discarded by the VF algorithm, and that the deep learning approach is able to use this information to improve detection efficiency. Additional experiments evaluated the advantage of deep learning compared to 'shallow' machine learning with boosted decision trees and neural networks with only a single hidden layer. These experiments were performed on the non-augmented data, where the convolution and pooling layers provide *less* of an advantage for the DNN. Figure [\[fig:roc_depth\]](#fig:roc_depth){reference-type="ref" reference="fig:roc_depth"} shows that XGBoost and shallow neural networks both perform better than the VF algorithm using the raw detector data (89% and 90% AUC on the same test events), but not as well as the DNN (92% AUC). The shallow network had 2000 hidden rectified linear units and was trained with 70% dropout in the hidden layer and the ADAM optimizer (\(\alpha=0.0001, \beta_1=0.9, \beta_2=0.999, \epsilon=1e-8\)); the hidden layer size, dropout probability, and learning rate were optimized based on validation performance. The XGBoost classifier had 100 trees with a maximum depth of 100 and minimum child weight of 10, and was trained with \(\eta=0.1\), L2 regularization factor of 0.1, and feature sampling rate 0.5.
# [\[sec:Conclusion\]]{#sec:Conclusion label="sec:Conclusion"}Conclusion
In summary, we report the first application of deep learning for the identification of antihydrogen annihilation events using realistic Monte Carlo simulations of both antihydrogen-like signals and antiproton background events. Moreover, we address the scenario in which classification must be translationally-invariant in the axial dimension to address systematic uncertainty in the experiment. The results demonstrate significant performance improvements compared to traditional approaches that are based on track and vertex reconstruction. Vertex reconstruction necessarily discards statistical information when fitting tracks, and the approach fails completely on all but 25% of events. In the deep learning approach, the end-to-end neural network model extracts additional statistical information from the raw data, improving AUC from 76% to 87% on the same event subset. It also has far better coverage, achieving 82% AUC on the 73% of events in which both detector layers were hit. While this study was performed for the ASACUSA experiment, the deep learning approach can easily be deployed to other ongoing antimatter experiments with different instruments. Furthermore, the ability to detect complex signals in high-dimensional data without relying on explicit track reconstruction potentially offers a new direction in the design of future detectors. Finally, the data used in this paper, comprising two million events, are publicly available from the machine learning in physics web portal: <http://mlphysics.ics.uci.edu>.\ | {'timestamp': '2017-06-07T02:08:05', 'yymm': '1706', 'arxiv_id': '1706.01826', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01826'} |
null | null |
# Introduction
The original motivation of this paper comes from a partnership of the authors with Eurotunnel, the company operating the tunnel under the Channel. Eurotunnel is currently facing an increasing congestion due to the trucks waiting in the terminal before being loaded in the shuttles. A way to address this issue consists in scheduling the shuttles so that the trucks do not wait too long in the terminal. In railway transportation, a traditional point of view considers that the demand can be smoothed by offering sufficiently enough departures over a day. Timetabling is then guided by other considerations, such as robustness, maintainability, or rolling stock. For instance, Swiss, Dutch and German companies usually design periodic timetables, which present many advantages . The way to optimize this kind of timetables has been the topic of many researches, initiated by Serafini and Ukovich and by Voorhoeve explicitly in the railway context, see for further works. In the context of periodic timetables, a way to adapt the schedules to a demand with strong variations consists in inserting new departures at peak-hours and deleting departures when the demand is low. Since the trip of the trucks in the tunnel is a small part of their whole journey, it is a reasonable approximation to assume that they cannot choose their arrival time in the terminal. Moreover, increasing the size of the fleet is not always doable in practice (the shuttles are expensive and the tunnel is used by other vehicles, which limits the maximal number of shuttle trips over a day). We face thus a different problem than the one addressed in the aforementioned literature: the demand is assumed to be fixed and nonelastic to the departures times, and the number of shuttles cannot be adjusted to the demand. Given a fleet of shuttles and a demand of transportation known in advance, the problem consists in designing a schedule for the shuttles that minimizes the waiting time of the users. There are timetabling problems with similar features, see for instance, but these articles are at a more macroscopic level than what we require to solve our problem. Moreover, in the present work, the schedules have to be designed in an offline manner. In a transportation context, and especially for Eurotunnel, computing the schedule in advance is mandatory. We study several versions of the problem, mainly according to two features. The first feature is whether the shuttles are allowed to come back at the terminal after having realized the trip. The second feature is the objective function. We consider in turn the following two quantities to be minimized: the maximum waiting time and the average waiting time. The first objective is perhaps a fairer one regarding the users, while the second one is relevant for the global efficiency. It seems that the question we address in the present paper is new. Moreover, it may be relevant for any situation where a demand, known in advance, has to be processed by batches and for which we want to minimize the processing time. This kind of situation is often met in chemical industry. An example whose motivation is very close to ours considers a test to be performed on samples, which arrive continuously . The test takes a certain amount of time and can be performed on several samples simultaneously. The question is then to determine the test schedule that minimizes the processing time. We propose efficient algorithms for the different versions. "Efficient" here means "theoretically efficient", according to their time complexity and the performance guarantee. It also means "practically efficient" for almost all cases, as shown by numerical experiments conducted on real-world instances. It might also be worth noting that, depending on the version considered, the proof techniques rely on various fields of optimization (convexity, Karush-Kuhn-Tucker conditions, binary search, domination results, shortest paths in finite graphs\...).
# Model
## The problems
We are given a fleet of \(S\) shuttles, for which departure times have to be determined. All shuttles have a capacity \(C\geq 0\) and are situated in the same loading terminal at the beginning of the day. The users are infinitesimal and arrive continuously in the terminal over a finite period of time, modeled as the interval \([0,T]\), following the cumulative nondecreasing function \(D:[0,T]\rightarrow\mathbb{R}_+\), where \(D(t)\) is the total number of users arrived in the terminal during the interval \([0,t]\). We assume throughout the paper that \(D(T)>0\). The shuttles have to carry the users over a fixed trip. When the users arrive in the terminal, they enter a queue. This queue closes when all the users who will leave with the next shuttle have arrived in the queue and users can enter a new queue only if the previous one is closed (this is how it works at Eurotunnel). When a queue is closed, the users in that queue can start boarding the shuttle. The process in illustrated on Figure [\[fig:loading\]](#fig:loading){reference-type="ref" reference="fig:loading"}. Loading a shuttle with a total of \(x\) users takes a time \(\nu x\). Note that setting \(\nu\) to zero allows to model the case where the users do not have to wait for the last user before boarding. Even if no users arrive strictly after time \(T\), loading and departures are allowed after that instant. Two possibilities are considered regarding the shuttles. Either return is not allowed: once the shuttles leave, they never come back to the terminal; or it is allowed: once the shuttles leave, they come back to the terminal after a time equal to \(\pi\geq 0\). Two objective functions to be minimized are considered: the maximum waiting time and the average waiting time. We have thus four problems:
- [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, which consists of not allowing return and minimizing the maximum waiting time.
- [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, which consists of not allowing return and minimizing the average waiting time.
- [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright}, which consists of allowing return and minimizing the maximum waiting time.
- [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}, which consists of allowing return and minimizing the average waiting time.
Practical constraints impose that overtake is not possible and thus, when return is allowed, the departure orders of the shuttles remain the same over the whole period. It is nevertheless possible to have simultaneous trips. This is an approximation in the case of Eurotunnel (we neglect the security distance and the length of the shuttles). For other situations, as the chemical application mentioned in the introduction, it may match what is met in practice.
## The demand
Throughout the paper, we assume that \(D(\makebox[1ex]\cdot)\) is upper semicontinuous. It allows to model discontinuity in the arrival process (batch of users arriving simultaneously). Yet, a weaker requirement could lead to mathematical difficulties, e.g., nonexistence of optimal solutions even for very simple cases. The *pseudo-inverses* of \(D(\makebox[1ex]\cdot)\), defined by \[\tau\colon y\in[0,D(T)]\longmapsto \left\{\begin{array}{ll}\inf\left\{t\in[0,T]\colon D(t)>y\right\}\in[0,T]&\mbox{if}\;y<D(T)\\mathsf{T}&\mbox{otherwise}\end{array}\right.\] and \[\bar\tau\colon y\in[0,D(T)] \longmapsto \inf\left\{t\in[0,T]\colon D(t)\geq y\right\}\in[0,T],\] play an important role in the paper. Note that they are nondecreasing functions and that \(\tau(y)\geq\bar\tau(y)\) for all \(y\in[0,D(T)]\). Times \(\tau(y)\) and \(\bar\tau(y)\) are respectively interpreted as the arrival times of the first user after the \(y\) firsts and of the last user of these \(y\) firsts. Since \(D(\makebox[1ex]\cdot)\) is upper semicontinuous, we have the following properties, with proofs for sake of completeness.
## Mathematical model
For the four problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright}, and [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}, a feasible solution is characterized by two nondecreasing sequences of nonnegative real numbers \(\boldsymbol{d}=d_1,d_2,\ldots\) and \(\boldsymbol{y}=y_1,y_2,\ldots\). The \(d_j\)'s are the successive departure times of the shuttles, and the \(y_j\)'s are their successive cumulative loads: the \(j\)th departure occurs at time \(d_j\) with a load of \(y_j-y_{j-1}\) users, where we set \(y_0=0\). Denote by \(g^{\max}(\boldsymbol{d},\boldsymbol{y})\) the value of the maximum waiting time and by \(g^{\ave}(\boldsymbol{d},\boldsymbol{y})\) the value of the average waiting time. There are explicit expressions of these objective functions. Note that \(\tau(y_{j})\) can be interpreted as the first arrival time of a user leaving with the "\((j+1)\)th shuttle". \[\begin{aligned}
g^{\max}(\boldsymbol{d},\boldsymbol{y})&=&\max_{j\colon y_j>y_{j-1}}\big(d_j-\tau(y_{j-1})\big),\\ g^{\ave}(\boldsymbol{d},\boldsymbol{y})&=&\frac 1 {D(T)} \sum_j\int_{y_{j-1}}^{y_j}(d_j-\bar\tau(y))dy, \end{aligned}\] where the indices \(j\) range over all departures. Problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright} can be written under the following form, \[\label{Pnoreturn}\tag{P\(_{\mbox{\textup{\tiny no return}}}\)} \begin{array}{rl@{\hspace{1cm}}rr} \operatorname{Min} & g(\boldsymbol{d},\boldsymbol{y}) & \\ \mbox{s.t.} & y_j-y_{j-1}\leq C & j=1,\ldots,S & \textup{(i)}\\ & y_{j-1}\leq y_j & j=1,\ldots,S & \textup{(ii)} \\ & d_{j-1}\leq d_j & j=2,\ldots,S& \textup{(iii)}\\ & y_S=D(T) & & \textup{(iv)}\\ & \bar\tau(y_j)+\nu(y_j-y_{j-1})\leq d_j & j=1,\ldots,S & \textup{(v)}\\ & y_0=0, & & \end{array}\] where \(g(\makebox[1ex]\cdot)\) is either \(g^{\max}(\makebox[1ex]\cdot)\) or \(g^{\ave}(\makebox[1ex]\cdot)\). Constraint (i) ensures that the total amount of users in any shuttle does not exceed the shuttle capacity. Constraint (ii) ensures that the indices of the \(y_j\) variables are consistent. Constraint (iii) ensures that the shuttles do not overtake. Constraint (iv) ensures that every user eventually leaves the terminal in a shuttle. Constraint (v) ensures that the departure time of a shuttle occurs once the last user of this shuttle has arrived and the loading is over. Problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright} always admit optimal solutions when they are feasible, i.e., when \(CS\geq D(T)\). Indeed, \(\bar\tau(y_j)+\nu(y_j-y_{j-1})\) is upper-bounded by \(T+\nu C\) and adding a constraint \(d_j\leq T+\nu C\) for all \(j\) does not change the optimal value; since \(\bar\tau(\makebox[1ex]\cdot)\) is lower semicontinuous (Lemma [\[lem:semicont\]](#lem:semicont){reference-type="ref" reference="lem:semicont"}), the set of feasible solutions of the optimization problem obtained with this new contraint is compact; its objective function is lower semicontinuous (and even continuous in the case of [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}). The following properties for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright} will be useful in some proofs.
Problems [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} can be written almost identically under the following form. We use infinitely many variables since there is no *a priori* reason to have a bounded number of departures, and there are indeed special cases for which there is no optimal solution with a finite number of departures. However, if \(\pi>0\), we prove that any optimal solution of [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} requires a finite number of departures, see Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"}. The case of [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} remains open. \[\label{Preturn}\tag{P\(_{\mbox{\textup{\tiny return}}}\)} \begin{array}{rl@{\hspace{1cm}}rr} \operatorname{Min} & g(\boldsymbol{d},\boldsymbol{y}) & \\ \mbox{s.t.} & y_j-y_{j-1}\leq C & j=1,\ldots,+\infty & \textup{(i)}\\ & y_{j-1}\leq y_j & j=1,\ldots,+\infty & \textup{(ii)} \\ & d_{j-1}\leq d_j & j=2,\ldots,S & \textup{(iii)}\\ & \displaystyle{\lim_{j\rightarrow+\infty}y_j=D(T)} & & \textup{(iv)}\\ & \bar\tau(y_j)+\nu(y_j-y_{j-1})\leq d_j & j=1,\ldots,+\infty & \textup{(v)}\\ & d_j+\pi+\nu(y_{j+S}-y_{j+S-1})\leq d_{j+S}& j=1,\ldots,+\infty & \textup{(vi)} \\ & y_0=0, & & \end{array}\] where \(g(\makebox[1ex]\cdot)\) is either \(g^{\max}(\makebox[1ex]\cdot)\) or \(g^{\ave}(\makebox[1ex]\cdot)\). Constraints (i), (ii), (iii), (iv), and (v) have the same meaning as for the previous problems. Constraint (vi) ensures that the time between two consecutive departures of a same shuttle is not smaller than the time required for a full trip plus the time needed to load the users. In the model [\[Preturn\]](#Preturn){reference-type="eqref" reference="Preturn"}, the shuttles are not identified. Note however that their schedules can be easily be recovered: the departure times of a shuttle \(s\) is of the form \[d_s,d_{s+S},d_{s+2S},\ldots\] and the time at which the loading starts for a shuttle with departure time \(d_j\) can be chosen to be \(d_j-\nu(y_j-y_{j-1})\) (the loading starts as late as possible). While it can be shown that problem [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} always admits an optimal solution when it is feasible (see Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"}), we were not able to settle the case of problem [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}.
## Computational model
We assume that the following operations take constant time:
- Evaluation of \(D(t)\) for any \(t\in[0,T]\).
- Integration of \(D(\makebox[1ex]\cdot)\) between two values.
- Evaluation of \(\tau(y)\) and \(\bar\tau(y)\) for any \(y\in\mathbb{R}_+\).
- Evaluation of \(\sup\{y\colon\bar\tau(y)+\nu y\leq \alpha\}\) for any \(\alpha\in\mathbb{R}_+\).
Note that if \(D(\makebox[1ex]\cdot)\) is piecewise affine with a natural description, as it is usually the case in practice, these assumptions are easily matched. Moreover, we set as constants of the computational model the capacity \(C\), the length of the period \(T\), the cumulative demand \(D(\makebox[1ex]\cdot)\), the loading rate \(\nu\), and the return time \(\pi\). The complexity functions will be expressed in terms of \(S\) and the accuracy of the computed solution.
# Main results {#sec:mainresults}
In the present section, we present our main findings. Many results state the existence of algorithms with a guarantee that the returned solution has a value close to the optimal value \(OPT\) of the considered problem. Except for two easy results--Corollary [\[cor:approx\]](#cor:approx){reference-type="ref" reference="cor:approx"} and Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"}--all proofs are postponed to other sections. We organize the results presented in that section in three subsections. The first subsection--Section [3.1](#subsec:one){reference-type="ref" reference="subsec:one"}--deals with the special case where \(D(\makebox[1ex]\cdot)\) is a constant function, i.e., when all users are in the loading terminal from the beginning of the period, and with returns allowed. It seems to us that these results are also interesting in themselves, because they form a natural situation for which there is a very efficient algorithm. The second subsection--Section [3.2](#subsec:noret){reference-type="ref" reference="subsec:noret"}--deals with the general case where the shuttles are not allowed to come back, i.e., with the case covered by the problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. The case where the shuttles are allowed to come back, i.e., when we deal with the problems [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}, is discussed in Section [3.3](#subsec:ret){reference-type="ref" reference="subsec:ret"}.
## All users in the terminal from the beginning {#subsec:one}
In this subsection, we present results regarding the four problems when \(D(t)=D(T)\) for all \(t\in[0,T]\) (all users are from the beginning in the terminal). For the problems for which return is not allowed ([P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}), an obvious optimal solution is given by \(y_j^*=jD(T)/S\) and \(d_j^*=\nu D(T)/S\) for \(j\in\left\{1,\ldots,S\right\}\) and the optimal value is \(\nu D(T)/S\) for both problems, provided that \(D(T)\leq CS\) (otherwise, there is no feasible solution at all): the shuttles take all the same amount of users, start immediately the loading process, and have the same departure time. The rest of the section is devoted to the results regarding the problems [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}. For the first one, there are closed-from expressions for the optimal value and an optimal solution.
In the proof, we actually provide a closed-form expression for an optimal solution. For [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} however, there does not seem to be a closed-form expression for an optimal solution, and not even for the optimal value. There is nevertheless an efficient algorithm.
If \(\pi=0\), the optimal value is \(\frac {\nu D(T)}{2S}\), and it is not too difficult to see that there is no optimal solution. In a transportation context, \(\pi=0\) looks unrealistic. However, the chemical application mentioned in the introduction could be a situation where this equality could be met: as soon as the test is over for a batch, we can start the test for a new one.
## When return is not allowed {#subsec:noret}
We have the existence of an efficient approximation algorithm for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. The algorithm is actually an easy binary search (Section [5.1.1](#subsubsec:algo){reference-type="ref" reference="subsubsec:algo"}).
With an additional assumption on \(D(\makebox[1ex]\cdot)\), this theorem provides actually an approximation scheme.
A schedule for the shuttles requires to specify \(S\) real numbers. Taking an output sensitive point of view, this corollary states thus the existence of a polynomial approximation scheme in this particular case.
For [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, there exists an efficient approximation algorithm too. The algorithm is also described later (Section [5.2.1](#subsubsec:algo_ave){reference-type="ref" reference="subsubsec:algo_ave"}). We already outline that this algorithm is not a binary search as in the former case, but consists in building a quite simple weighted "approximative" graph, in which a shortest path is computed.
As for Corollary [\[cor:approx\]](#cor:approx){reference-type="ref" reference="cor:approx"} above, this theorem could be interpreted as a polynomial approximation scheme by using the fact that \(D(\makebox[1ex]\cdot)\) is increasing.
## When return is allowed {#subsec:ret}
The following proposition implies that when \(\pi\) is larger than \(0\), any optimal solution of [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} requires a finite number of nonempty departures.
The next theorem states that there exists an algorithm computing arbitrarily good feasible solutions for [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} within reasonable computational times when \(S\) is small. As for Section [3.2](#subsec:noret){reference-type="ref" reference="subsec:noret"}, this algorithm is described later in the paper (Section [6](#sec:ret){reference-type="ref" reference="sec:ret"}). It is based on the computation of a shortest path in an "approximative" graph, as for Theorem [\[thm:pave\]](#thm:pave){reference-type="ref" reference="thm:pave"}. It also uses Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"} in a crucial way (actually a slight variation of it: Lemma [\[lem:t+\]](#lem:t+){reference-type="ref" reference="lem:t+"}).
As above, the theorem actually ensures that the algorithm is an approximation scheme since we can bound from below \(OPT\) using only the input values. If \(S\) is considered as constant, this becomes even a polynomial approximation scheme.\
We do not know whether there is a counterpart to Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"} for problem [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}. If such a counterpart would exist, then almost the same technique as the one used in Section [6](#sec:ret){reference-type="ref" reference="sec:ret"} would lead to a theorem similar to Theorem [\[thm:pmaxret\]](#thm:pmaxret){reference-type="ref" reference="thm:pmaxret"} for [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}. The existence of such a theorem remains thus open.
# All users in the terminal from the beginning {#all-users-in-the-terminal-from-the-beginning}
Consider the case where all users are in the loading terminal from the beginning. To ease the reading, and for the present section only, we use \(D\) to denote the quantity \(D(T)\). We treat first the case of problem [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright}.
The rest of this section is devoted to the proof of Proposition [\[prop:S1\]](#prop:S1){reference-type="ref" reference="prop:S1"}, which ensures the existence of an efficient algorithm solving problem [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} when \(D(t)=D\) for all \(t\in[0,T]\). We start by considering the special case of problem [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} when \(S=1\). In such a case, it is always beneficial to define \(d_j=(j-1)\pi+\nu y_j\). Assuming that the \(y_j\)'s are given, it provides a feasible solution since \(\bar\tau(y)=0\) for all \(y\in[0,D]\). The objective function of [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} becomes thus \[\frac 1 D\sum_{j=1}^{+\infty}\left((j-1)\pi+\nu \sum_{i=1}^{j}x_i\right)x_j=\frac 1 D\left(\sum_{j=1}^{+\infty}(j-1)\pi x_j+\frac 1 2\nu\sum_{j=1}^{+\infty}x_j^2\right)+\frac {\nu D} 2 ,\] where \(x_j=y_j-y_{j-1}\). Solving [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright} when \(S=1\) reduces thus to solving \[\label{S1averet}\tag*{\(P(D)\)} \begin{array}{rlr} \operatorname{Min} & \displaystyle{\sum_{j=1}^{+\infty}(j-1)\pi x_j+\frac 1 2\nu\sum_{j=1}^{+\infty}x_j^2} \\ \mbox{s.t.} & \displaystyle{\sum_{j=1}^{+\infty}x_j=D} \\ & 0\leq x_j\leq C & j=1,\ldots,+\infty, \end{array}\] which is a convex program (with infinitely many variables). We will show that there is always an optimal solution of [\[S1averet\]](#S1averet){reference-type="ref" reference="S1averet"} with a finite support. Then, we will solve \(P_0(D)\), defined as the program [\[S1averet\]](#S1averet){reference-type="ref" reference="S1averet"} with the additional constraint \(|\{j\colon x_j\neq 0\}|<+\infty\), with the help of the Karush-Kuhn-Tucker conditions (that do not apply otherwise).
Let \(v(D)\) and \(v_0(D)\) be the optimal values of respectively [\[S1averet\]](#S1averet){reference-type="ref" reference="S1averet"} and \(P_0(D)\). Note that \(v(D)\leq v_0(D)\).
# When return is not allowed {#sec:noret}
## Minimizing the maximum waiting time {#subsec:pmax}
### The algorithm {#subsubsec:algo}
If \(CS<D(T)\), there is no feasible solution. We can thus assume that \(CS\geq D(T)\). The algorithm is a binary search starting with the values \(h^+=T+\nu D(T)\) and \(h^-=0\) which are respectively upper and lower bounds of the optimal value. While the gap \(h^+-h^-\) is larger than \(\rho\), we consider the tentative value \(h=\frac{h^++h^-} 2\) and the system \[\label{Smaxnoreth}\tag{S\(_h\)} \left\{ \begin{array}{l@{\hspace{1cm}}r} y_j = \sup \mathcal{S}_j^h & j=1,\ldots,S\\ y_S=D(T) &\\ y_0 = 0\\ d_j = h+\tau(y_{j-1}) & j=1,\ldots,S, \end{array}\right.\] where \(\mathcal{S}_j^h=\left\{y\in\mathbb{R}_+ \colon y\leq C+y_{j-1}, \bar{\tau}(y)+\nu(y-y_{j-1})-\tau(y_{j-1})\leq h,y\leq D(T)\right\}\). Each iteration of the binary search consists in deciding whether [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"} has a feasible solution or not, and it can be done in \(O(S)\) by computing the values of the \(y_j\)'s and the \(d_j\)'s iteratively (here we use in particular the computational assumptions on \(D\)). As we are going to prove, [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"} has a feasible solution if and only if the problem has a feasible solution with a value of the objective function at most \(h\). If [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"} has a feasible solution, we update thus the value of \(h^+\) with the current value of \(h\), otherwise, we update \(h^-\) with \(h\). When \(h^+-h^-\leq\rho\), the solution of program \((\text{S}_{h^+})\) is feasible for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and its value \(h^+\) is at most at \(\rho\) from the optimal value.
### Proof of Theorem [\[thm:pmax\]](#thm:pmax){reference-type="ref" reference="thm:pmax"}
For any fixed \(h\), [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} has a feasible solution with a value of the objective function at most \(h\) if and only if the following system has a feasible solution. \[\label{Qmaxnoreth}\tag{Q\(_h\)} \left\{ \begin{array}{l@{\hspace{1cm}}rr} d_j-\tau(y_{j-1})\leq h & j=1,\ldots,S & \textup{(Qi)}\\ y_j-y_{j-1}\leq C & j=1,\ldots,S & \textup{(Qii)}\\ y_{j-1}\leq y_j & j=1,\ldots,S & \textup{(Qiii)}\\ d_{j-1}\leq d_j & j=2,\ldots,S& \textup{(Qiv)}\\ y_S=D(T) & &\textup{(Qv)}\\ \bar\tau(y_j)+\nu(y_j-y_{j-1})\leq d_j& j=1,\ldots,S & \textup{(Qvi)}\\ y_0=0. & \end{array}\right.\] We claim that [\[Qmaxnoreth\]](#Qmaxnoreth){reference-type="eqref" reference="Qmaxnoreth"} has a feasible solution if and only if [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"} has one. Once this equivalence is established, the correctness of the binary search described above is almost immediate using Claim [\[claim:change_obj\]](#claim:change_obj){reference-type="ref" reference="claim:change_obj"}. Let \((\boldsymbol{d},\boldsymbol{y})\) be a feasible solution of [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"}. We use without further mention that \(\mathcal{S}_j^h\) is closed. It satisfies the constraints \(\textup{(Qi)}, \textup{(Qii)}\), and \(\textup{(Qv)}\). We have \(y_{j-1}\leq C+y_{j-1}\) and \(y_{j-1}\leq D(T)\). Since \(\bar{\tau}(y)\leq \tau(y)\) for all \(y\), we also have \(\bar{\tau}(y_{j-1})+\nu(y_{j-1}-y_{j-1})-\tau(y_{j-1})\leq h\). It means that \(y_{j-1}\) belongs to \(\mathcal{S}_j^h\), and thus \(y_{j-1}\leq y_j\). Hence, \((\boldsymbol{d},\boldsymbol{y})\) satisfies also constraint \(\textup{(Qiii)}\). Since \(\bar{\tau}(y_j)+\nu(y_j-y_{j-1})-\tau(y_{j-1})\leq h\), the solution also satisfies constraint \(\textup{(Qvi)}\) and since \(\tau(\makebox[1ex]\cdot)\) is nondecreasing, it satisfies constraint \(\textup{(Qiv)}\). Therefore, it is a feasible solution of [\[Qmaxnoreth\]](#Qmaxnoreth){reference-type="eqref" reference="Qmaxnoreth"} and the existence of a feasible solution of [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"} implies the existence of a feasible solution of [\[Qmaxnoreth\]](#Qmaxnoreth){reference-type="eqref" reference="Qmaxnoreth"}. For the converse implication, suppose that [\[Qmaxnoreth\]](#Qmaxnoreth){reference-type="eqref" reference="Qmaxnoreth"} admits a feasible solution, and consider the optimization problem consisting in maximizing \(\sum_{j=1}^Sy_j\) over its feasible solutions. These feasible solutions form a compact set of \(\mathbb{R}_+^S\) since it is obviously bounded and since the semicontinuities of \(\tau(\makebox[1ex]\cdot)\) and \(\bar\tau(\makebox[1ex]\cdot)\) imply that it is closed. There is thus an optimal solution \((\boldsymbol{d}^*,\boldsymbol{y}^*)\) to that optimization problem. Suppose for a contradiction that there is a \(j\) such that \(y_j^*< \sup \mathcal{S}_j^h\). Denote \(j_0\) the largest such index. Let us slightly increase \(y_{j_0}^*\), while letting the other \(y_j^*\) untouched. Redefine \(d_j^*\) to be \(h+\tau(y_{j-1}^*)\) for all \(j\geq j_0\). The pair \((\boldsymbol{d}^*,\boldsymbol{y}^*)\) remains feasible for [\[Qmaxnoreth\]](#Qmaxnoreth){reference-type="eqref" reference="Qmaxnoreth"} (we use here the fact that \(\sup\mathcal{S}_j^h\) is nondecreasing with \(j\)), while increasing the quantity \(\sum_{j=1}^Sy_j^*\), which is a contradiction with the optimality assumption. Thus, we have \(y_j^*=\sup \mathcal{S}_j^h\) for all \(j\) and \(d_j^*:=h+\tau(y_{j-1}^*)\) for all \(j\) provides a feasible solution for [\[Smaxnoreth\]](#Smaxnoreth){reference-type="eqref" reference="Smaxnoreth"}. 0◻
## Minimizing the average waiting time {#subsec:pave}
### The algorithm {#subsubsec:algo_ave}
The following map will be useful in the description of the algorithm. \[f^{\operatorname{ave}}:(d,y,y')\longmapsto\int_y^{y'}(d-\bar\tau(u))du.\] Define the directed graph \(\mathcal{G}=(\mathcal{V},\mathcal{A})\) by \[\begin{array}{rcl} \mathcal{V} & = & \{(0,0)\}\cup\{\eta,2\eta,\ldots,M\eta\}\times\{\eta,2\eta,\ldots,R\eta\}\smallskip\\ \mathcal{A} & = & \left\{\big((z,r),(z',r')\big)\in\mathcal{V}^2\colon r+z'=r'\;\mbox{and}\;\bar\tau(r')-\bar\tau(r)+\nu(z'-z)+\frac 1 2 \gamma\eta\geq 0\right\}, \end{array}\] where we use the following notations: \[\alpha=\inf_{t\in[0,T)}D'_+(t),\qquad R=\left\lfloor\frac{D(T)M}C\right\rfloor,\qquad\eta=\frac C M,\qquad\mbox{and}\qquad\gamma=2\left(\frac 1\alpha+2\nu\right).\] Set for each arc \(a=\big((z,r),(z',r')\big)\) a weight \(w(a)=f^{\operatorname{ave}}\big(\bar\tau(r')+\nu(z'-\eta),r+\eta,r'\big)\).\
If \(CS<D(T)\), there is no feasible solution. We can thus assume that \(CS\geq D(T)\). The algorithm consists first in computing a path \(\tilde p\) minimizing \(\sum_{a\in\mathcal{A}(p)}w(a)\), among all paths \(p\) with at most \(S\) arcs starting at \((0,0)\in\mathcal{V}\) and ending at a vertex \((z,r)\) with \(r=R\eta\). Such paths exist, see Lemma [\[lem:kopt\]](#lem:kopt){reference-type="ref" reference="lem:kopt"} below. The computation of \(\tilde p\) can be done in \(O(S|\mathcal{A}|)\) via dynamic programming. Let the vertex sequence of \(\tilde p\) be \(\big((z_0,r_0),(z_1,r_1),\ldots,(z_n,r_n)\big)\). The algorithm consists then in defining recursively \[\tilde y_j=\left\{\begin{array}{ll} 0 & \mbox{for \(j=0\)} \\ \min\big(r_j+\eta,\tilde y_{j-1}+C,D(T)\big) & \mbox{for \(j=1,\ldots,n\)} \smallskip\\ D(T) & \mbox{for \(j=n+1,\ldots,S\)} \end{array}\right.\] and \[\tilde d_j=\left\{\begin{array}{ll} \bar\tau(\tilde y_j)+\nu(\tilde y_j-\tilde y_{j-1})+j\gamma\eta & \mbox{for \(j=1,\ldots,n\)} \smallskip\\ \max(\tilde d_n,T+\nu(\tilde y_{n+1}-\tilde y_n)) &\mbox{for \(j=n+1,\ldots,S\)} \end{array}\right.\] and outputting the pair \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\). The construction of the graph is sketched on Figure [\[fig:algo\]](#fig:algo){reference-type="ref" reference="fig:algo"}.\
As it will be shown below, this \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\) is a feasible solution of [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright} providing a value to the objective function within a \(O\left(\frac {S^2} M\right)\) gap to the optimal value.
### Proof of Theorem [\[thm:pave\]](#thm:pave){reference-type="ref" reference="thm:pave"}
We provide three lemmas, which are proved in a separate section at the very end of the paper to ease the reading. Theorem [\[thm:pave\]](#thm:pave){reference-type="ref" reference="thm:pave"} results immediately from their combination. In the proofs of the lemmas, we assume that \(M\) is large enough so that \(\eta<D(T)\). Since in Theorem [\[thm:pave\]](#thm:pave){reference-type="ref" reference="thm:pave"}, \(M\) appears in 'big O' formulas, it is a valid assumption. Anyway it is what is sought in practice: the larger \(M\), the larger the accuracy of the solution. An \(\eta\) of same order of magnitude of \(D(T)\) would be useless.
## When the demand function is a step function
Better complexity results can be obtained when the demand is a step function. A *step function* is a function that can be written as a finite linear combination of indicator functions of intervals. The assumption on \(D(\makebox[1ex]\cdot)\) being a step function means that the users arrive only on a finite number of instants. As it has already been noted, the assumption \(\nu=0\) is equivalent to the assumption that every user boards a shuttle as soon as he arrives in the terminal.
It turns out that when \(C\) and the values taken by \(D(\makebox[1ex]\cdot)\) are integer, the loads of the shuttles in the optimal solution returned by the algorithm are also integer. We cover thus the case where the users are atoms.
### The algorithm {#the-algorithm .unnumbered}
We provide only the algorithm for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, the other case can be dealt similarly. Let \(t_1<\cdots<t_K\) be the \(K\) discontinuities. Define the directed graph \(\mathcal{G}=(\mathcal{V},\mathcal{A})\) by \[\begin{array}{rcl} \mathcal{V} & = & \{0\}\cup\left\{D(t_k)+Cq\colon k\in\left\{1,\ldots,K\right\}, q\in\{0,1,\ldots,Q\}\right\} \smallskip\\ \mathcal{A} & = & \{(y,y')\in\mathcal{V}^2\colon 0\leq y'-y\leq C\}, \end{array}\] where \(Q=\lfloor D(T)/C\rfloor\). Note that the vertex set is a finite subset of \(\mathbb{R}_+\). Set for each arc \(a=(y,y')\) a weight \(w(a)=\bar\tau(y')-\tau(y)\). We consider the two vertices \(0\) and \(D(T)\) (obtained with \(k=K\) and \(q=0\)).\
If \(CS<D(T)\), there is no feasible solution. We can thus assume that \(CS\geq D(T)\). The algorithm consists first in computing a \(0\)-\(D(T)\) path \(\tilde p\) with \(S\) arcs minimizing \(\max_{a\in \mathcal{A}(\tilde p)}w(a)\). Within the proof of Proposition [\[prop:pmax\]](#prop:pmax){reference-type="ref" reference="prop:pmax"} below, we show that from any feasible solution we can build a \(0\)-\(D(T)\) path with \(S\) arcs in \(\mathcal{G}\). Thus, when the problem is feasible, such paths exist in \(\mathcal{G}\). The computation of \(\tilde p\) can be done in \(O(S|\mathcal{A}|)\) via dynamic programming. Let the vertex sequence of \(\tilde p\) be \((\tilde y_0,\tilde y_1,\ldots,\tilde y_S)\). The end of the algorithm consists in defining \(\tilde d_j=\bar\tau(\tilde y_j)\) for all \(j\in\{1,\ldots,S\}\) and outputting the pair \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\). As it will be shown below, this \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\) is an optimal solution of [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}.
### Proof of Proposition [\[prop:pmax\]](#prop:pmax){reference-type="ref" reference="prop:pmax"} {#proof-of-proposition-proppmax .unnumbered}
According to Claim [\[claim:change_obj\]](#claim:change_obj){reference-type="ref" reference="claim:change_obj"}, we replace the objective function of [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} by \(\max_{j\in\left\{1,\ldots,S\right\}}(d_j-\tau(y_{j-1}))\). It can easily be checked that \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\) is feasible for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. It provides a value \(\max_{j\in\left\{1,\ldots,S\right\}}(\bar\tau(\tilde y_j)-\tau(\tilde y_{j-1}))\) for [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} (with the alternative objective function), and this value coincides with \(\max_{a\in \mathcal{A}(\tilde p)}w(a)\). The path \(\tilde p\) describes therefore a solution of [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} with a value equal to \(\max_{a\in \mathcal{A}(\tilde p)}w(a)\). Conversely, let \((\boldsymbol{d},\boldsymbol{y})\) be any feasible solution of [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. Let \(\bar\boldsymbol{y}\) be the sequence defined by \(\bar y_j=\min\{y\in\mathcal{V}\colon y\geq y_j\}\). On the one hand, we have \(\bar y_{j-1}\leq \bar y_j\) because \(y_{j-1}\leq y_j\). On the other hand, we have \(\bar y_{j-1}+C\geq y_{j-1}+C\geq y_j\). If \(\bar y_{j-1}+C\in\mathcal{V}\), we have \(\bar y_{j-1}+C\geq\bar y_j\) by definition of \(\bar y_j\). If \(\bar y_{j-1}+C\notin\mathcal{V}\), then \(\bar y_{j-1}+C> D(T)\geq\bar y_j\) since \(D(T)\in\mathcal{V}\). Thus, \((\bar y_{j-1},\bar y_j)\in\mathcal{A}\) for all \(j\in\{1,\ldots,S\}\). We have \(\bar y_0=0\) and \(\bar y_S=D(T)\) and the sequence \(\bar \boldsymbol{y}\) is a \(0\)-\(D(T)\) path \(p\) with \(S\) arcs. Second, we prove that \(\bar\tau(\bar y_j)-\tau(\bar y_{j-1})\leq d_j-\tau( y_{j-1})\) as follows. There exists a unique \(k\) such that \(D(t_k)<y_j\leq D(t_{k+1})\). By definition of \(D(\makebox[1ex]\cdot)\), we have \(D(t)=D(t_k)\) for all \(t\in[t_k,t_{k+1})\), and thus \(\bar\tau(y_j)=t_{k+1}\). Since \(D(t_{k+1})\in\mathcal{V}\), we have \(\bar y_j\leq D(t_{k+1})\) by definition of \(\bar y_j\), and hence \(\bar\tau(\bar y_j)\leq t_{k+1}\) (directly by definition of \(\bar\tau(\makebox[1ex]\cdot)\)). Therefore, \(\bar\tau(\bar y_j)-\tau(\bar y_{j-1})\leq\bar\tau(y_j)-\tau(y_{j-1})\leq d_j-\tau(y_{j-1})\) (where we use the fact that \(\tau(\makebox[1ex]\cdot)\) is nondecreasing). Finally, we have \(\max_{a\in\mathcal{A}(p)}w(a)\leq \max_{j\in\{1,\ldots,S\}}\big(d_j-\tau( y_{j-1})\big)\). As the path \(\tilde p\) is optimal, \(\max_{a\in\mathcal{A}(\tilde p)}w(a)\) is a lower bound on the value taken by the (alternative) objective function on \((\boldsymbol{d},\boldsymbol{y})\). 0◻
# When return is allowed {#sec:ret}
### The algorithm {#the-algorithm-1 .unnumbered}
The following map will be useful: \[f^{\max}\colon(\ell,y,y')\longmapsto\left\{\begin{array}{ll} \max(\ell,\bar\tau(y'))+\nu(y'-y)-\tau(y) & \mbox{if \(y'\geq y\)} \\ 0 & \mbox{if \(y'=y\).}\end{array}\right.\] We introduce the following two sets \[\begin{array}{rcl} \mathcal{Q} & = & \{0,\eta,\ldots,(\left\lfloor T^+/\eta\right\rfloor+1)\eta\}^S \smallskip\\ \mathcal{R} & = & \left\{\boldsymbol{r}\in\{0,\eta,\ldots,R\eta\}^S\colon 0\leq r_k-r_{k-1}\leq M\eta\mbox{ for \(k=2,\ldots,S\)}\right\}, \end{array}\] where \[\eta=\frac{C}{M},\qquad R=\left\lfloor\frac{D(T)M}C\right\rfloor,\qquad\mbox{and}\qquad T^+=T+\frac{\nu D(T)}S+\left(\left\lceil\frac{D(T)}{CS}\right\rceil-1\right)\pi.\] Define the directed graph \(\mathcal{G}=(\mathcal{V},\mathcal{A})\) by \[\begin{array}{rcl} \mathcal{V} & = &\left\{(z,\boldsymbol{q},\boldsymbol{r})\in\{0,\eta,\ldots,M\eta\}\times\mathcal{Q}\times\mathcal{R}\colon r_k\leq D(q_k)\;\mbox{for \(k=1,\ldots,S\)}\right\}\medskip\\ \mathcal{A} & = & \left\{\big((z,\boldsymbol{q},\boldsymbol{r}),(z',\boldsymbol{q}',\boldsymbol{r}')\big)\in \mathcal{V}^2\;\mbox{satisfying \((\star)\)}\right\}, \end{array}\] where \[(\star)\quad r_S+z'=r_1'\quad\mbox{and}\quad q_k'-q_k-\nu(r_k-r_{k-1})-\pi+(1+\nu)\eta\geq 0\mbox{ for \(k=1,\ldots, S\)}.\] We adopt the convention \(D(t)=D(T)\) when \(t\geq T\) and we define \(r_0=r_1-z\). Set for each arc \(a=\big((z,\boldsymbol{q},\boldsymbol{r}),(z',\boldsymbol{q}',\boldsymbol{r}')\big)\) a weight \(w(a)=\max_{k\in\left\{1,\ldots,S\right\}}f^{\max}(q_k'-\eta,r_{k-1}'+\eta,r'_k)\) where \(r'_0=r'_1-z'\).\
The algorithm consists first in computing a path \(\tilde p\) minimizing \(\max_{a\in A(p)}w(a)\) among all paths \(p\) starting at \((0,\boldsymbol{0},\boldsymbol{0})\in\mathcal{V}\) (the 'all zero' vector) and ending at a vertex \((z,\boldsymbol{q},\boldsymbol{r})\) with \(r_S=R\eta\). Such paths exist, see Lemma [\[lem:loptmax\]](#lem:loptmax){reference-type="ref" reference="lem:loptmax"} below. It can be done in \(O(|\mathcal{V}||\mathcal{A}|)\) via dynamic programming. Let the vertex sequence of \(\tilde p\) be \(\big((0,\boldsymbol{0},\boldsymbol{0}),(z_0,\boldsymbol{q}^{0},\boldsymbol{r}^{0}),\ldots,(z_n,\boldsymbol{q}^n,\boldsymbol{r}^{n})\big)\). The vector \(\boldsymbol{r}^i\) models the cumulative loads of the \(S\) shuttles when they perform their \(i\)th departure. The vector \(\boldsymbol{q}^{i}\) models the times at which the loading of the \(S\) shuttles starts when they perform their \(i\)th departure. These quantities are computed only approximatively (with an accuracy \(\eta\)). The algorithm consists then in defining recursively for all \(j=iS+k\) with \(i=0,\ldots,n\) and \(k=1,\ldots,S\) \[\begin{array}{rcl} \tilde y_{j}&=&\left\{\begin{array}{ll}\min\big(r_{k}^i+\eta,y_{j-1}+C,D(T)\big)&\mbox{if } r^i_k>r^i_{k-1}\\\tilde y_{j-1}&\mbox{otherwise}\end{array}\right.\\ \tilde d_{j}&=&\max(q_{k}^i,\bar\tau(\tilde y_{j}))+j\tilde\gamma\eta+\nu(\tilde y_{j}-\tilde y_{j-1}) \end{array}\] where \(\tilde y_0=0\), \(r^i_0=r^{i-1}_S\), \(r^0_0=0\), and \(\tilde\gamma=(1+2\nu+1/\alpha)\). For \(j=(n+1)S+1,\ldots,N\) \[\begin{array}{rcl} \tilde y_{j}&=&D(T)\\ \tilde d_{j}&=&\max\big(\tilde d_{j-S}+\pi,T\big)+\nu(\tilde y_j-\tilde y_{j-1}). \end{array}\] and outputting the pair \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\).\
As it will be stated below, this \((\tilde\boldsymbol{d},\tilde\boldsymbol{y})\) is a feasible solution of [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} providing a value to the objective function \(g^{\max}(\makebox[1ex]\cdot)\) within a \(O\left(\frac {S^2} M\right)\) gap to the optimal value.
### Proof of Theorem [\[thm:pmaxret\]](#thm:pmaxret){reference-type="ref" reference="thm:pmaxret"} {#proof-of-theorem-thmpmaxret .unnumbered}
We provide four lemmas. The proof of Lemma [\[lem:t+\]](#lem:t+){reference-type="ref" reference="lem:t+"} is almost the one of Proposition [\[prop:finite_max\]](#prop:finite_max){reference-type="ref" reference="prop:finite_max"} and the proofs of the three others follow the same scheme as the ones of Lemmas [\[lem:kopt\]](#lem:kopt){reference-type="ref" reference="lem:kopt"}, [\[lem:feas\]](#lem:feas){reference-type="ref" reference="lem:feas"}, and [\[lem:byy\]](#lem:byy){reference-type="ref" reference="lem:byy"}. They are thus omitted. Theorem [\[thm:pmaxret\]](#thm:pmaxret){reference-type="ref" reference="thm:pmaxret"} results immediately from their combination.
# Experimental results {#sec:experiments}
In this section, we test the performance of the algorithms described in previous sections for problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, and [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright}. As explained in Section [3](#sec:mainresults){reference-type="ref" reference="sec:mainresults"}, we do not have such an algorithm for problem [P\(^{\ave}_{\mbox{\textup{\tiny return}}}\)]{.upright}.
## Data
Our experiments are based on real data provided by our partner Eurotunnel. They are related to the transportation of freight trucks between France and Great Britain. Some parameters are fixed as constants of the problems and do not vary from an instance to another. For the constants \(C, T, \nu ,\pi\) of our problem, we take the real values used in practice by the company: \[C=32,\quad T=1440\min \mbox{ (one day)}, \quad \nu=0.625\min,\quad \pi=34\min.\] (The value taken for \(\pi\) is actually the duration of a trip going from France to Great-Britain, and not of the round trip, which lasts approximatively twice this quantity.) Two functions \(D(\makebox[1ex]\cdot)\) are used. The first one ("1P") is a piecewise affine map obtained by averaging the real demand over several days. It turns out that this function has a peak period in the morning. The second function ("2P"), also piecewise affine, is obtained from the first by artificially adding a second peak period in the evening. In both cases, \(D(\makebox[1ex]\cdot)\) is increasing and \(D(T)=2016\). For problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} and [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, we consider \(S\in[100,250]\) since the number of shuttle trips in every direction is within this range for a typical day. The numerical experiments are performed on a Macbook Pro of 2014 with four 2.2 Ghz processors and 16 Gb of ram.
## Results
The problems [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}, and [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} are solved with algorithms described in this article. The results are summarized in the following tables. Table [1](#tab:Pmaxnoret){reference-type="ref" reference="tab:Pmaxnoret"} gives the numerical results for problem [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. The next column is the number of shuttles \(S\) in the fleet. The third column provides the parameter \(\varepsilon\) of the algorithm, which is an *a priori* upper bound on the optimality gap (Corollary [\[cor:approx\]](#cor:approx){reference-type="ref" reference="cor:approx"}). The two next columns give respectively the lower bound and the upper bound (value of the feasible solution returned by the algorithm), both expressed in minutes. The next column is the optimality gap. The last column provides the CPU time spent solving the problem.
Table [2](#tab:Pavenoret){reference-type="ref" reference="tab:Pavenoret"} gives the numerical results for problem [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright}. The columns are the same as for Table [1](#tab:Pmaxnoret){reference-type="ref" reference="tab:Pmaxnoret"} except the third one which provides here the parameter \(M\) of the algorithm. We know from Theorem [\[thm:pave\]](#thm:pave){reference-type="ref" reference="thm:pave"} that the gap between the upper bound and the lower bound converges asymptotically to 0 when \(M\) goes to infinity. We tried \(M=32\) and \(M=128\).
Table [3](#tab:Pmaxret){reference-type="ref" reference="tab:Pmaxret"} gives the numerical results for problem [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright}. The columns are the same as for Table [2](#tab:Pavenoret){reference-type="ref" reference="tab:Pavenoret"}. Since the computation time was prohibitively long as soon as \(S\geq 2\), we made experiments for \(S=1\). To get realistic waiting times for the users, we divided the demand functions by 3.5 leading to ("1P\(^*\)") and ("2P\(^*\)"). Again, we know from Theorem [\[thm:pmaxret\]](#thm:pmaxret){reference-type="ref" reference="thm:pmaxret"} that for large \(M\), we will be close to the optimal solution and we tried \(M=16\) and \(M=32\).
## Comments
In Table [1](#tab:Pmaxnoret){reference-type="ref" reference="tab:Pmaxnoret"}, the results for problem [P\(^{\max}_{\mbox{\textup{\tiny no return}}}\)]{.upright} are extremely convincing, the optimal solutions were found almost immediately. In Table [2](#tab:Pavenoret){reference-type="ref" reference="tab:Pavenoret"}, the algorithm for problem [P\(^{\ave}_{\mbox{\textup{\tiny no return}}}\)]{.upright} was able to find provable good solutions within reasonable computation times. We may note that increasing \(M\) after some threshold does not seem to improve the quality of the return solution. This was confirmed by other experiments not shown here. It may indicate that the algorithm could be used efficiently in practice. In Table [3](#tab:Pmaxret){reference-type="ref" reference="tab:Pmaxret"}, the same holds for [P\(^{\max}_{\mbox{\textup{\tiny return}}}\)]{.upright} once we have accepted to work with one shuttle. Finding an efficient algorithm with at least two shuttles seems to remain a challenging task.
# Proofs of Lemmas of Section [5.2](#subsec:pave){reference-type="ref" reference="subsec:pave"} {#proofs-of-lemmas-of-section-subsecpave} | {'timestamp': '2017-06-07T02:06:00', 'yymm': '1706', 'arxiv_id': '1706.01733', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01733'} |
null | null |
# Introduction
This paper is about the use of transactions as a common means to control concurrent access of programs to shared locations and to avoid that values stored at these locations are changed almost randomly. A *transaction controller* interacts with concurrently running programs (read: sequential components of an asynchronous system) to control whether access to a shared location can be granted or not, thus ensuring a certain form of consistency for these locations. A commonly accepted consistency criterion is that the joint behavior of all transactions (read: programs running under transactional control) with respect to the shared locations is equivalent to a serial execution of those programs. Serialisability guarantees that each transaction can be specified independently from the transaction controller, as if it had exclusive access to the shared locations. It is expensive and cumbersome to specify transactional behavior and prove its correctness again and again for components of the great number of concurrent systems. Our goal is to define once and for all an abstract (i.e. programming language independent) transaction controller \(\ASM{TaCtl}\) which can simply be "plugged in" to turn the behavior of concurrent programs (read: components \(M\) of any given asynchronous system \(\cal M\)) into a transactional one. This involves to also define an operator \(TA(M,\ASM{TaCtl})\) which forces the programs \(M\) to listen to the controller \(\ASM{TaCtl}\) when trying to access shared locations. For the sake of generality we define the operator and the controller in terms of Abstract State Machines (ASMs) which can be read and understood as pseudo-code so that \(\ASM{TaCtl}\) and the operator \(TA\) can be applied to code written in any programming language (to be precise: whose programs come with a notion of single step, the level where our controller imposes shared memory access constraints to guarantee transactional code behavior). On the other side, the precise semantics underlying ASMs (for which we refer the reader to ) allows us to mathematically prove the correctness of our controller and operator. We concentrate here on transaction controllers that employ locking strategies such as the common two-phase locking protocol (2PL). That is, each transaction first has to acquire a (read-or write-) lock for a shared location, before the access is granted. Locks are released after the transaction has successfully committed and no more access to the shared locations is necessary. There are of course other approaches to transaction handling, see e.g. and the extensive literature there covering classical transaction control for flat transactions, timestamp-based, optimistic and hybrid transaction control protocols, as well as non-flat transaction models such as sagas and multi-level transactions. We define \(\ASM{TaCtl}\) and the operator \(TA\) in Sect. [2](#sect:TAoperator){reference-type="ref" reference="sect:TAoperator"} and the \(\ASM{TaCtl}\) components in Sect. [3](#sect:TaCtl){reference-type="ref" reference="sect:TaCtl"}. In Sect. [4](#sect:Thm){reference-type="ref" reference="sect:Thm"} we prove the correctness of these definitions.
# The Transaction Operator \(TA(M,\ASM{TaCtl}\)) {#sect:TAoperator}
As explained above, a transaction controller performs the lock handling, the deadlock detection and handling, the recovery mechanism (for partial recovery) and the commit of single machines. Thus we define it as consisting of four components specified in Sect. [3](#sect:TaCtl){reference-type="ref" reference="sect:TaCtl"}.
The operator \(TA(M,\ASM{TaCtl})\) transforms the components \(M\) of any concurrent system (asynchronous ASM) \({\cal M} = (M_i)_{i \in I}\) into components of a concurrent system \(TA({\cal M},\ASM{TaCtl})\) where each \(TA(M_i,\ASM{TaCtl})\) runs as transaction under the control of \(\ASM{TaCtl}\):
\(\ASM{TaCtl}\) keeps a dynamic set \(TransAct\) of those machines \(M\) whose runs it currently has to supervise to perform in a transactional manner until \(M\) has \(Terminated\) its transactional behavior (so that it can \(\ASM{Commit}\) it).[^1] To turn the behavior of a machine \(M\) into a transactional one, first of all \(M\) has to register itself with the controller \(\ASM{TaCtl}\), read: to be inserted into the set of currently to be handled \(TransAct\)ions. To \(\ASM{Undo}\) as part of a recovery some steps \(M\) made already during the given transactional run segment of \(M\), a last-in first-out queue \(history(M)\) is needed which keeps track of the states the transactional run goes through; when \(M\) enters the set \(TransAct\) the \(history(M)\) has to be initialized (to the empty queue). The crucial transactional feature is that each non private (i.e. shared or monitored or output) location \(l\) a machine \(M\) needs to read or write for performing a step has to be \(LockedBy(M)\) for this purpose; \(M\) tries to obtain such locks by calling the \(\ASM{LockHandler}\). In case no \(newLocks\) are needed by \(M\) in its \(currState\) or the needed \(newLocks\) can be \(Granted\) by the \(\ASM{LockHandler}\), \(M\) performs its next step; in addition, for a possible future recovery, the machine has to \(\ASM{Record}\) in its \(history(M)\) the current values of those locations which are (possibly over-) written by this \(M\)-step together with the obtained \(newLocks\). Then \(M\) continues its transactional behavior until it is \(Terminated\). In case the needed \(newLocks\) are \(Refused\), namely because another machine \(N\) in \(TransAct\) for some needed \(l\) has \(W\mbox{-}Locked(l,N)\) or (in case \(M\) wants a W-(rite)Lock) has \(R\mbox{-}Locked(l,N)\), \(M\) has to \(Wait\) for \(N\); in fact it continues its transactional behavior by calling again the \(\ASM{LockHandler}\) for the needed \(newLocks\)---until the needed locked locations are unlocked when \(N\)'s transactional behavior is \(\ASM{Commit}\)ed, whereafter a new request for these locks this time may be \(Granted\) to \(M\).[^2] As a consequence deadlocks may occur, namely when a cycle occurs in the transitive closure \(Wait^*\) of the \(Wait\) relation. To resolve such deadlocks the \(\ASM{DeadlockHandler}\) component of \(\ASM{TaCtl}\) chooses some machines as \(Victim\)s for a recovery.[^3] After a victimized machine \(M\) is \(Recovered\) by the \(\ASM {Recovery}\) component of \(\ASM{TaCtl}\), so that \(M\) can exit its \(waitForRecovery\) state, it continues its transactional behavior. This explains the following definition of \(TA(M,\ASM{TaCtl})\) as a control state ASM, i.e. an ASM with a top level Finite State Machine control structure. We formulate it by the flowchart diagram of Fig. [\[fig:TA(M,C)\]](#fig:TA(M,C)){reference-type="ref" reference="fig:TA(M,C)"}, which has a precise control state ASM semantics (see the definition in ). The components for the recovery feature are highlighted in the flowchart by a colouring that differs from that of the other components. The macros which appear in Fig. [\[fig:TA(M,C)\]](#fig:TA(M,C)){reference-type="ref" reference="fig:TA(M,C)"} and the components of \(\ASM{TaCtl}\) are defined below. The predicate \(NewLocksNeededBy(M)\) holds if in the current state of \(M\) at least one of two cases happens:[^4] either \(M\) to perform its step in this state reads some shared or monitored location which is not yet \(LockedBy(M)\) or \(M\) writes some shared or output location which is not yet \(LockedBy(M)\) for writing. A location can be \(LockedBy(M)\) for reading (\(R\mbox{-}Locked(l,M)\)) or for writing (\(W\mbox{-}Locked(l,M)\)). Formally:
The \(overWrittenVal\)ues are the \(currState(M)\)-values (retrieved by the \(eval\)-function) of those shared or output locations \((f,args)\) which are written by \(M\) in its \(currState(M)\). To \(\ASM{Record}\) the set of these values together with the obtained \(newLocks\) means to append the pair of these two sets to the \(history\) queue of \(M\) from where upon recovery the values and the locks can be retrieved.
To \(\ASM{CallLockHandler}\) for the \(newLocks\) requested by \(M\) in its \(currState(M)\) means to \(\ASM{Insert}(M,newLocks)\) into the \(\ASM{LockHandler}\)'s set of to be handled \(LockRequest\)s. Similarly we let \(\ASM{CallCommit(M)}\) stand for insertion of \(M\) into a set \(CommitRequest\) of the \(\ASM{Commit}\) component.
# The Transaction Controller Components {#sect:TaCtl}
A \(\ASM{CallCommit(M)}\) by machine \(M\) enables the \(\ASM{Commit}\) component. Using the \(\mathrel{\mathbf{choose}}\) operator we leave the order in which the \(CommitRequest\)s are handled refinable by different instantiations of \(\ASM{TaCtl}\). \(\ASM{Commit}\)ing \(M\) means to \(\ASM{Unlock}\) all locations \(l\) that are \(LockedBy(M)\). Note that each lock obtained by \(M\) remains with \(M\) until the end of \(M\)'s transactional behavior. Since \(M\) performs a \(\ASM{CallCommit(M)}\) when it has \(Terminated\) its transactional computation, nothing more has to be done to \(\ASM{Commit}\) \(M\) besides deleting \(M\) from the sets of \(CommitRequest\)s and still to be handled \(TransAct\)ions.[^8] Note that the locations \(R\mbox{-}Locked(l,M)\) and \(W\mbox{-}Locked(l,M)\) are shared by the \(\ASM{Commit}\), \(\ASM{LockHandler}\) and \(\ASM{Recovery}\) components, but these components never have the same \(M\) simultaneously in their request resp. \(Victim\) set since when machine \(M\) has performed a \(\ASM{CallCommit(M)}\), it has \(Terminated\) its transactional computation and does not participate any more in any \((M,L) \in LockRequest\) or \(Victim\)ization.
As for \(\ASM{Commit}\) also for the \(\ASM{LockHandler}\) we use the \(\mathrel{\mathbf{choose}}\) operator to leave the order in which the \(LockRequest\)s are handled refinable by different instantiations of \(\ASM{TaCtl}\). The strategy we adopt for lock handling is to refuse all locks for locations requested by \(M\) if at least one of the following two cases happens:
- some of the requested locations is \(W\mbox{-}Locked\) by another transactional machine \(N \in TransAct\),
- some of the requested locations is a \(WriteLoc\)ation that is \(R\mbox{-}Locked\) by another transactional machine \(N \in TransAct\).
This definition implies that multiple transactions may simultaneoulsy have a \(R\mbox{-}Lock\) on some location. It is specified below by the predicate \(CannotBeGranted\). To \(\ASM{RefuseRequestedLocks}\) it suffices to set the communication interface \(Refused\) of \(TA(M,\ASM{TaCtl})\); this makes \(M\) \(Wait\) for each location \(l\) that is \(W\mbox{-}Locked(l,N)\) and for each \(WriteLoc\)ation that is \(R\mbox{-}Locked(l,N)\) by some other transactional component machine \(N \in TransAct\).
A \(Deadlock\) originates if two machines are in a \(Wait\) cycle, otherwise stated if for some (not yet \(Victim\)ized) machine \(M\) the pair \((M,M)\) is in the transitive (not reflexive) closure \(Wait^*\) of \(Wait\). In this case the \(\ASM{DeadlockHandler}\) selects for recovery a (typically minimal) subset of \(Deadlocked\) transactions \(toResolve\)---they are \(Victim\)ized to \(waitForRecovery\), in which mode (control state) they are backtracked until they become \(Recovered\). The selection criteria are intrinsically specific for particular transaction controllers, driving a usually rather complex selection algorithm in terms of number of conflict partners, priorities, waiting time, etc. In this paper we leave their specification for \(\ASM{TaCtl}\) abstract (read: refinable in different directions) by using the \(\mathrel{\mathbf{choose}}\) operator.
Also for the \(\ASM{Recovery}\) component we use the \(\mathrel{\mathbf{choose}}\) operator to leave the order in which the \(Victim\)s are chosen for recovery refinable by different instantiations of \(\ASM{TaCtl}\). To be \(Recovered\) a machine \(M\) is backtracked by \(\ASM{Undo}(M)\) steps until \(M\) is not \(Deadlocked\) any more, in which case it is deleted from the set of \(Victim\)s, so that be definition it is \(Recovered\). This happens at the latest when \(history(M)\) has become empty.
Note that in our description of the and the (partial) we deliberately left the strategy for victim seclection and \(\ASM{Undo}\) abstract leaving fairness considerations to be discussed elsewhere. It is clear that if always the same victim is selected for partial recovery, the same deadlocks may be created again and again. However, it is well known that fairness can be achieved by choosing an appropriate victim selection strategy.
# Correctness Theorem {#sect:Thm}
In this section we show the desired correctness property: if all monitored or shared locations of any \(M_i\) are output or controlled locations of some other \(M_j\) and all output locations of any \(M_i\) are monitored or shared locations of some other \(M_j\) (closed system assumption)[^9], each run of \(TA({\cal M},\ASM{TaCtl})\) is equivalent to a serialization of the terminating \(M_i\)-runs, namely the \(M_{i_{1}}\)-run followed by the \(M_{i_{2}}\)-run etc., where \(M_{i_{j}}\) is the \(j\)-th machine of \(\cal M\) which performs a commit in the \(TA({\cal M},\ASM{TaCtl})\) run. To simplify the exposition (i.e. the formulation of statement and proof of the theorem) we only consider machine steps which take place under the transaction control, in other words we abstract from any step \(M_i\) makes before being \(\ASM{Insert}\)ed into or after being \(\ASM{Delet}\)ed from the set \(TransAct\) of machines which currently run under the control of \(\ASM{TaCtl}\). First of all we have to make precise what a *serial* multi-agent ASM run is and what *equivalence* of \(TA({\cal M},\ASM{TaCtl})\) runs means in the general multi-agent ASM framework.
### Definition of run equivalence.
Let \(S_0, S_1, S_2, \dots\) be a (finite or infinite) run of the system \(TA({\cal M},\ASM{TaCtl})\). In general we may assume that runs forever, whereas each machine \(M \in \mathcal{M}\) running as transaction will be terminated at some time--at least after commit \(M\) will only change values of non-shared and non-output locations[^10]. For \(i = 0,1,2,\dots\) let \(\Delta_i\) denote the unique, consistent update set defining the transition from \(S_i\) to \(S_{i+1}\). By definition of \(TA({\cal M},\ASM{TaCtl})\) the update set is the union of the update sets of the agents executing \(M \in \mathcal{M}\) resp. \(\ASM{TaCtl}\): \[\Delta_i = \bigcup\limits_{M \in \mathcal{M}} \Delta_i(M) \cup \Delta_i(\ASM{TaCtl}).\] \(\Delta_i(M)\) contains the updates defined by the ASM \(TA(M,\ASM{TaCtl})\) in state \(S_i\)[^11] and \(\Delta_i(\ASM{TaCtl})\) contains the updates by the transaction controller in this state. The sequence of update sets \(\Delta_0(M)\), \(\Delta_1(M)\), \(\Delta_2(M)\), ...will be called the *schedule* of \(M\) (for the given transactional run). To generalise for transactional ASM runs the equivalence of transaction schedules known from database systems we now define two *cleansing operations* for ASM schedules. By the first one (i) we eliminate all (in particular unsuccessful-lock-request) computation segments which are without proper \(M\)-updates; by the second one (ii) we eliminate all \(M\)-steps which are related to a later \(\ASM{Undo}(M)\) step by the \(\ASM{Recovery}\) component:
1. Delete from the schedule of \(M\) each \(\Delta_i(M)\) where one of the following two properties holds:
- \(\Delta_i(M)=\emptyset\) (\(M\) contributes no update to \(S_i\)),
- \(\Delta_i(M)\) belongs to a step of an \(M\)-computation segment where \(M\) in its \(ctl\_state(M)=\) TA-\(ctl\) does \(\ASM{CallLockHandler}(M,newLocks)\) and in its next step moves from control-state \(waitForLocks\) back to control state TA\(-ctl\), because the \(\ASM{LockHandler}\) refused new locks by \(Refused(M,newLocks)\).[^12]
In such computation steps \(M\) makes no proper update.
2. Repeat choosing from the schedule of \(M\) a pair \(\Delta_j(M)\) with later \(\Delta_{j'}(M)\) (\(j<j'\)) which belong to the first resp. second of two consecutive \(M\)-Recovery steps defined as follows:
- a (say \(M\)-RecoveryEntry) step whereby \(M\) in state \(S_j\) moves from control-state TA-\(ctl\) to \(waitForRecovery\), because it became a \(Victim\),
- the next \(M\)-step (say \(M\)-RecoveryExit) whereby \(M\) in state \(S_{j'}\) moves back to control state TA-\(ctl\) because it has been \(Recovered\).
In these two \(M\)-Recovery steps \(M\) makes no proper update. Delete:
1. \(\Delta_j(M)\) and \(\Delta_{j'}(M)\),
2. the \(((Victim,M),true)\) update from the corresponding \(\Delta_t(\ASM{TaCtl})\) (\(t< j\)) which in state \(S_j\) triggered the \(M\)-RecoveryEntry,
3. \(\ASM{TryToRecover}(M)\)-updates in any update set \(\Delta_{i+k}(\ASM{TaCtl})\) between the considered \(M\)-RecoveryEntry and \(M\)-RecoveryExit step (\(i<j<i+k<j'\)),
4. each \(\Delta_{i'}(M)\) belonging to the \(M\)-computation segment from TA-\(ctl\) back to TA-\(ctl\) which contains the proper \(M\)-step in \(S_i\) that is \(\ASM{UNDO}\)ne in \(S_{i+k}\) by the considered \(\ASM{TryToRecover}(M)\) step; besides control state and \(\ASM{Record}\) updates these \(\Delta_{i'}(M)\) contain updates \((\ell,v)\) with \(\ell = (f,(val_{S_i}(t_1),\dots,val_{S_i}(t_n)))\) where the corresponding \(\ASM{Undo}\) updates are \((\ell,val_{S_i}(f(t_1,\dots,t_n))) \in \Delta_{i+k}(\ASM{TaCtl})\),
5. the \(\ASM{HandleLockRequest}(M,newLocks)\)-updates in \(\Delta_{l\prime}(\ASM{TaCtl})\) corresponding to \(M\)'s \(\ASM{CallLockHandler}\) step (if any: in case \(newLocks\) are needed for the proper \(M\)-step in \(S_i\)) in state \(S_l\) (\(l<l^\prime<i\)).
The sequence \(\Delta_{i_1}(M), \Delta_{i_2}(M), \dots\) with \(i_1 < i_2 < \dots\) resulting from the application of the two cleansing operations as long as possible--note that confluence is obvious, so the sequence is uniquely defined--will be called the *cleansed schedule* of \(M\) (for the given run). Before defining the equivalence of transactional ASM runs we remark that \(TA({\cal M},\ASM{TaCtl})\) has indeed several runs, even for the same initial state \(S_0\). This is due to the fact that a lot of non-determinism is involved in the definition of this ASM. First, the submachines of are non-deterministic:
- In case several machines \(M, M^{\prime} \in \mathcal{M}\) request conflicting locks at the same time, the can only grant the requested locks for one of these machines.
- Commit requests are executed in random order by the submachine.
- The submachine chooses a set of victims, and this selection has been deliberately left abstract.
- The submachine chooses in each step a victim \(M\), for which the last step will be undone by restoring previous values at updated locations and releasing corresponding locks.
Second, the specification of \(TA({\cal M},\ASM{TaCtl})\) leaves deliberately open, when a machine \(M \in \mathcal{M}\) will be started, i.e., register as a transaction in \(TransAct\) to be controlled by . This is in line with the common view that transactions \(M \in \mathcal{M}\) can register at any time to the transaction controller and will remain under its control until they commit.
That is, we consider runs to be equivalent, if all transactions \(M \in \mathcal{M}\) read the same locations and see there the same values and perform the same updates in the same order disregarding waiting times and updates that are undone.
### Definition of serializability.
Next we have to clarify our generalised notion of a serial run, for which we concentrate on committed transactions--transactions that have not yet committed can still undo their updates, so they must be left out of consideration[^13]. We need a definition of the read-and write-locations of \(M\) in a state \(S\), i.e. \(ReadLoc(M,S)\) and \(WriteLoc(M,S)\) as used in the definition of \(newLocks(M,S)\). The definition of \(Read/WriteLoc\) depends on the locking level, whether locks are provided for variables, pages, blocks, etc. To provide a definite definition, in this paper we give the definition at the level of abstraction of the locations of the underlying class \(\cal{M}\) of component machines (ASMs) \(M\). Refining this definition (and that of \(newLocks\)) appropriately for other locking levels does not innvalidate the main result of this paper. We define \(ReadLoc(M,S) = ReadLoc(r,S)\), where \(r\) is the defining rule of the ASM \(M\), and analogously \(WriteLoc(M,S)\) \(= WriteLoc(r,S)\). Then we use structural induction according to the definition of ASM rules in . As an auxiliary concept we need to define inductively the read and write locations of terms and formulae. The definitions use an interpretation \(I\) of free variables which we suppress notationally (unless otherwise stated) and assume to be given with (as environment of) the state \(S\). This allows us to write \(ReadLoc(M,S)\), \(WriteLoc(M,S)\) instead of \(ReadLoc(M,S,I)\), \(ReadLoc(M,S,I)\) respectively.
### Read/Write Locations of Terms and Formulae.
For state \(S\) let \(I\) be the given interpretation of the variables which may occur freely (in given terms or formulae). We write \(val_S(construct)\) for the evaluation of \(construct\) (a term or a formula) in state \(S\) (under the given interpretation \(I\) of free variables).
Note that logical variables are not locations: they cannot be written and their values are not stored in a location but in the given interpretation \(I\) from where they can be retrieved. We define \(WriteLoc(\alpha,S)=\emptyset\) for every formula \(\alpha\) because formulae are not locations one could write into. \(ReadLoc(\alpha,S)\) for atomic formulae \(P(t_1, \ldots, t_n )\) has to be defined as for terms with \(P\) playing the same role as a function symbol \(f\). For propositional formulae one reads the locations of their subformulae. In the inductive step for quantified formulae \(domain(S)\) denotes the superuniverse of \(S\) minus the Reserve set and \(I_{x}^{d}\) the extension (or modification) of \(I\) where \(x\) is interpreted by a domain element \(d\).
Note that the values of the logical variables are not read from a location but from the modified state environment function \(I_{x}^{d}\).
### Read/Write Locations of ASM Rules.
.WriteLoc( r_1 r_2,S) = {
.\
ReadLoc( x = t r,S,I)= ReadLoc(t,S,I) ReadLoc(r,S,I_x\^val_S(t))\
WriteLoc( x = t r,S,I)= WriteLoc(r,S,I_x\^val_S(t)) // call by value\
ReadLoc( x r,S,I)= ReadLoc(x ,S,I) \_a range(x,,S,I)ReadLoc(r,S,I_x\^a) range(x,,S,I)= {d domain(S) val\_S,I_x\^d() = true}by-1WriteLoc( x r,S,I)= \_a range(x,,S,I)WriteLoc(r,S,I_x\^a)\
.
We say that \(M\) has or is committed (in state \(S_i\), denoted \(Committed(M,S_i)\)) if step \(\ASM{Commit}(M)\) has been performed (in state \(S_i\)).
That is, in a serial run all committed transactions are executed in a total order and are followed by the updates of transactions that did not yet commit.
By induction hypothesis \(S_0^{\prime\prime}, S_1^{\prime\prime}, S_2^{\prime\prime}, \dots\) is serialisable, so \(S_0,S_1^\prime,S_2^\prime,\dots\) and thereby also \(S_0,S_1,S_2,\dots\) is serialisable with \(M_1 < M\) for all \(M \in \mathcal{M}-\{ M_1 \}\). ◻
# Conclusion
In this article we specified (in terms of Abstract State Machines) a transaction controller \(\ASM{TaCtl}\) and a transaction operator which turn the behaviour of a set of concurrent programs into a transactional one under the control of \(\ASM{TaCtl}\). In this way the locations shared by the programs are accessed in a well-defined manner. For this we proved that all concurrent transactional runs are serialisable. The relevance of the transaction operator is that it permits to concentrate on the specification of program behavior ignoring any problems resulting from the use of shared locations. That is, specifications can be written in a way that shared locations are treated as if they were exclusively used by a single program. This is valuable for numerous applications, as shared locations (in particular, locations in a database) are common, and random access to them is hardly ever permitted. Furthermore, by shifting transaction control into the rigorous framework of Abstract State Machines we made several extensions to transaction control as known from the area of databases. In the classical theory schedules are sequences containing read-and write-operations of the transactions plus the corresponding read-and write-lock and commit events, i.e., only one such operation or event is treated at a time. In our case we exploited the inherent parallelism in ASM runs, so we always considered an arbitrary update set with usually many updates at the same time. Under these circumstances we generalised the notion of schedule and serialisability in terms of the synchronous parallelism of ASMs. In this way we stimulate also more parallelism in transactional systems. Among further work we would like to be undertaken is to provide a (proven to be correct) implementation of our transaction controller and the \(TA\) operator, in particular as plug-in for the CoreASM or Asmeta simulation engines. We would also like to see refinements or adaptations of our transaction controller model for different approaches to serialisability , see also the ASM-based treatment of multi-level transaction control in . Last but not least we would like to see further detailings of our correctness proof to a mechanically verified one, e.g. using the ASM theories developed in KIV (see for an extensive list of relevant publications) and PVS or the (Event-)B theorem prover for an (Event-)B transformation of \(TA({\cal M},\ASM{TaCtl})\) (as suggested in ). | {'timestamp': '2017-06-07T02:06:44', 'yymm': '1706', 'arxiv_id': '1706.01762', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01762'} |
# Introduction
This paper is devoted to generalize the Thomas projective parameter and the Weyl projective tensor as invariants of a mapping of an affine connection space. These invariants are primary generalized in but it is obtained these invariants are not unique ones in this paper. Therefore, we are interested to obtain some invariants of geometric mappings different of the invariants from in this paper.
## Spaces of affine connection
An \(N\)-dimensional manifold \(\mathcal M^N\) equipped with an affine connection \(\nabla\) with torsion is called the non-symmetric affine connection space \(\mathbb{GA}_N\). The affine connection coefficients of the affine connection \(\nabla\) are \(L^i_{jk}\) and they are non-symmetric by indices \(j\) and \(k\). The geometrical object \[L^i_{\underline{jk}}=\frac12\big(L^i_{jk}+L^i_{kj}\big),\] satisfies the equation \[L^{i'}_{\underline{j'k'}}=x^{i'}_ix^j_{j'}x^k_{k'}L^i_{\underline{jk}}+ x^{i'}_{i}x^i_{j'k'},\] so it is the affine connection coefficient of a symmetric affine connection space \(\mathbb A_N\). The manifold \(\mathcal M^N\) equipped with an affine connection \(\widetilde\nabla\) whose coefficients are \(L^i_{\underline{jk}}\) is **the associated space \(\mathbb A_N\)** (of the space \(\mathbb{GA}_N\)). There is one kind of covariant derivation with regard to a symmetric affine connection: \[a^i_{j|k}=a^i_{j,k}+L^i_{\underline{\alpha k}}a^\alpha_j- L^\alpha_{\underline{jk}}a^i_\alpha, \label{eq:covderivativesim}\] for a tensor \(a^i_j\) of the type \((1,1)\) and the partial derivative \(\partial/\partial x^i\) denoted by comma. From the corresponding Ricci-type identity, it is obtained one curvature tensor of the associated space \(\mathbb A_N\): \[R^i_{jmn}=L^i_{\underline{jm},n}-L^i_{\underline{jn},m} + L^\alpha_{\underline{jm}}L^i_{\underline{\alpha n}}- L^\alpha_{\underline{jn}}L^i_{\underline{\alpha m}}. \label{eq:R}\] Special symmetric affine connection spaces are Riemannian spaces \(\mathbb R_N\) whose affine connection coefficients are the corresponding Christoffel symbols \(\Gamma^i_{\underline{jk}}\) of the second kind. Many authors have developed the theory of symmetric affine connection spaces and mappings between them. Some of them are J. Mikeš [@mik1; @mik10; @mik2; @mik3; @mik5; @mik6; @mik7; @mik8], N. S. Sinyukov , V. E. Berezovski, L. P. Eisenhart and many others. M. S. Stanković (see ) obtained an invariant of an almost geodesic mapping of a non-symmetric affine connection space from the corresponding transformation of the curvature tensor ([\[eq:R\]](#eq:R){reference-type="ref" reference="eq:R"}).
## Motivation
H. Weyl and T. Y. Thomas obtained invariants of geodesic mappings of a symmetric affine connection space. J. Mikeš with his research group [@mik1; @mik10; @mik2; @mik3; @mik5; @mik6; @mik7; @mik8], N. S. Sinyukov, and many other authors have continued the process of generalization of these invariants. The search for general formula for invariants of geometric mappings is started in. The basic invariants of geometric mappings are obtained in this paper. It is also studied a special case in that paper. In this special case, it is obtained basic invariants of a mapping but the author founded some other invariants different of the basic ones. We are interested to obtain some of these invariants which are not basic in this paper. The invariants which we are interested to develop in this paper are
- The generalized Thomas projective parameter : \[T^i_{jk}=L^i_{\underline{jk}}- \frac1{N+1}\big(\delta^i_kL^\alpha_{\underline{j\alpha}}+ \delta^i_jL^\alpha_{\underline{k\alpha}}\big). \label{eq:Thomasgeodesic}\]
- The Weyl projective tensor : \[W^i_{jmn}=R^i_{jmn}+\frac1{N+1}\delta^i_jR_{[mn]}+ \frac N{N^2-1}\delta^i_{[m}R_{jn]}+ \frac1{N^2-1}\delta^i_{[m}R_{n]j}. \label{eq:Weylgeodesic}\]
In a Riemannian space \(\mathbb{R}_N\), the Weyl projective tensor ([\[eq:Weylgeodesic\]](#eq:Weylgeodesic){reference-type="ref" reference="eq:Weylgeodesic"}) reduces to \[W^i_{jmn}=R^i_{jmn}+ \frac 1{N-1}\big(\delta^i_{m}R_{jn}-\delta^i_{n}R_{jm}\big). \label{eq:WeylgeodesicRN}\]
# Reminder on basic invariants
Let \(f:\mathbb{GA}_N\to\mathbb{G\overline A}_N\) be a mapping between non-symmetric affine connection spaces \(\mathbb{GA}_N\) and \(\mathbb{G\overline A}_N\). The deformation tensor \(P^i_{jk}\) of this mapping satisfies the corresponding equation \[P^i_{jk}=\overline L^i_{jk}-L^i_{jk}= \overline\omega{}^i_{jk}-\omega^i_{jk}+\overline\tau{}^i_{jk}- \tau^i_{jk}, \label{eq:P}\] for geometrical objects \(\omega^i_{jk},\overline\omega{}^i_{jk},\tau^i_{jk},\overline\tau{}^i_{jk}\) of the type \((1,2)\) such that \(\omega^i_{jk}=\omega^i_{kj},\) \(\overline\omega{}^i_{jk}=\overline\omega{}^i_{kj},\) \(\tau^i_{jk}=-\tau^i_{kj},\) \(\overline\tau{}^i_{jk}=-\overline\tau{}^i_{kj}\). After symmetrizing the equation ([\[eq:P\]](#eq:P){reference-type="ref" reference="eq:P"}) by indices \(j\) and \(k\), one gets \[\begin{aligned}
\overline L^i_{\underline{jk}}=L^i_{\underline{jk}}+ \overline\omega^i_{jk}-\omega^i_{jk}. \label{eq:Psim} \end{aligned}\] With regard to the last result, it is obtained three kinds of invariants of the mapping \(f\) in. In this paper, we are interested to generalize just invariants of the second kind. The basic invariants which we will generalize are: \[\begin{aligned}
&\aligned {\widetilde{\mathcal T}}{}^i_{jk}=L^i_{\underline{jk}}-\omega^i_{jk}, \endaligned\label{eq:Tsimantisimgeneral}\\\displaybreak[0] &\aligned {\widetilde{\mathcal W}}{}^i_{jmn}= R^i_{jmn}-\omega{}^i_{jm|n}+ \omega{}^i_{jn|m}+ \omega{}^\alpha_{jm}\omega{}^i_{\alpha n}-\omega{}^\alpha_{jn}\omega{}^i_{\alpha m}, \endaligned\label{eq:Wbasic} \end{aligned}\] for \(p^1_1,\ldots,p^2_3=1,2,\omega_{(1).jk}^i=L^i_{\underline{jk}}, \omega^i_{(2).jk}=\omega^i_{jk}\). The invariant ([\[eq:Tsimantisimgeneral\]](#eq:Tsimantisimgeneral){reference-type="ref" reference="eq:Tsimantisimgeneral"}) is the basic invariant of the mapping \(f\) of the Thomas type but the invariant ([\[eq:Wbasic\]](#eq:Wbasic){reference-type="ref" reference="eq:Wbasic"}) is the basic invariant of the mapping \(f\) of the Weyl type.
# Derived invariants
In general, the geometrical object \(\omega^i_{jk}\) from the equation ([\[eq:P\]](#eq:P){reference-type="ref" reference="eq:P"}) has the form \[\omega^i_{jk}=s_{(1)}\big(\delta^i_j\rho_k+\delta^i_k\rho_j\big)+ s_{(2)}\big(F^i_j\sigma_k+F^i_k\sigma_j\big)+s_{(3)}\sigma_{jk}\varphi^i, \label{eq:omegagen}\] for \(s_1,s_2,s_3\in\mathbb R\), \(1\)-forms \(\rho_j,\sigma_j\), an affinor \(F^i_j\), a covariant tensor \(\sigma_{jk}\) symmetric by indices \(j\) and \(k\) and a contra-variant vector \(\varphi^i\).
With regard to the equation ([\[eq:omegagen\]](#eq:omegagen){reference-type="ref" reference="eq:omegagen"}), we get \[\aligned \omega^\alpha_{jm}\omega^i_{\alpha n}&=\big(s_{(1)}\big)^2\delta^i_j\rho_m\rho_n+\big(s_{(1)}\big)^2 \delta^i_m\rho_j\rho_n\\&+ \delta^i_n\Big(2\big(s_{(1)}\big)^2\rho_j\rho_m+ s_{(1)}s_{(2)}\big(F^\alpha_m\rho_\alpha\sigma_j+F^\alpha_j\rho_\alpha\sigma_m\big) +s_{(1)}s_{(3)}\sigma_{jm}\rho_\alpha\varphi^\alpha\Big)\\&+ \big(s_{(2)}\big)^2\Big(F^i_n\big(F^\alpha_m\sigma_j+F^\alpha_j\sigma_m\big)\sigma_\alpha+ F^i_\alpha\big(F^\alpha_m\sigma_j+F^\alpha_j\sigma_m\big)\sigma_n\Big)+ \big(s_{(3)}\big)^2\sigma_{jm}\sigma_{\alpha n}\varphi^\alpha\varphi^i\\&+ s_{(1)}s_{(2)}\Big(F^i_n\big(\rho_j\sigma_m+\rho_m\sigma_j\big)+ F^i_m\big(\rho_j\sigma_n+\rho_n\sigma_j\big)+ F^i_j\big(\rho_m\sigma_n+\rho_n\sigma_m\big)\Big)\\&+ s_{(1)}s_{(3)}\big(\sigma_{mn}\rho_j+\sigma_{jn}\rho_m+\sigma_{jm}\rho_n\big)\varphi^i\\&+ s_{(2)}s_{(3)}\Big(\big(F^\alpha_m\sigma_j+F^\alpha_j\sigma_m\big) \sigma_{\alpha n}\varphi^i+\big(F^i_n\sigma_\alpha+ F^i_\alpha\sigma_n\big)\sigma_{jm}\varphi^\alpha\Big) \endaligned\label{eq:omegaomegagen}\] It holds the following theorem:
## Invariants in associated space
We will analyze the invariants ([\[eq:basicThomasGeneral\]](#eq:basicThomasGeneral){reference-type="ref" reference="eq:basicThomasGeneral"}, [\[eq:basicWeylGeneral\]](#eq:basicWeylGeneral){reference-type="ref" reference="eq:basicWeylGeneral"}) of a mapping \(f:\mathbb{GA}_N\to\mathbb{G\overline A}_N\) bellow. From this analyzing, we will obtain some other invariants of the mapping \(f\). After contracting the equality \(\widetilde{\overline{\mathcal T}}{}^i_{jk}- \widetilde{\mathcal T}{}^i_{jk}=0\) by indices \(i\) and \(k\), we get \[\aligned (N+1)s_{(1)}\big(\overline\rho_j-\rho_j\big)&=\overline L^\alpha_{\underline{j\alpha}}-s_{(2)}\big(\overline F^\alpha_j\overline\sigma_k+\overline F\overline\sigma_j\big)- s_{(3)}\overline\sigma_{j\alpha}\overline\varphi^\alpha\\&- L^\alpha_{\underline{j\alpha}}+s_{(2)}\big(F^\alpha_j\sigma_\alpha+F\sigma_j\big)+ s_{(3)}\sigma_{j\alpha}\varphi^\alpha, \endaligned\label{eq:rho-rho}\] for \(F=F^\alpha_\alpha,\overline F=\overline F^\alpha_\alpha\). After substituting this equation into the equality \(\widetilde{\overline{\mathcal T}}{}^{(s).i}_{jk}- \widetilde{\mathcal T}{}^{(s).i}_{jk}=0\), one obtains \[\widetilde{{\overline T}}{}^{(s).i}_{jk}= \widetilde{T}{}^{(s).i}_{jk},\] for \[\aligned \widetilde T{}^{(s).i}_{jk}&=L^i_{\underline{jk}}- \frac{s_{(1)}}{N+1}\delta^i_j\Big(L^\alpha_{\underline{k\alpha}}- s_{(2)}(F^\alpha_k\sigma_\alpha+F\sigma_k\big)-s_{(3)}\sigma_{k\alpha}\varphi^\alpha\Big)\\& - \frac{s_{(1)}}{N+1}\delta^i_k\Big(L^\alpha_{\underline{j\alpha}}- s_{(2)}(F^\alpha_j\sigma_\alpha+F\sigma_j\big)-s_{(3)}\sigma_{j\alpha}\varphi^\alpha\Big)\\&- s_{(2)}\big(F^i_j\sigma_k+F^i_k\sigma_j\big)-s_{(3)}\sigma_{jk}\varphi^i, \endaligned\label{eq:generalThomassymmetric}\] and the corresponding \(\widetilde{\overline T}{}^{(s).i}_{jk}\).
The geometrical object ([\[eq:generalThomassymmetric\]](#eq:generalThomassymmetric){reference-type="ref" reference="eq:generalThomassymmetric"}) is *the derived associated invariant* of the Thomas type. Furthermore, from the invariance of the geometrical object ([\[eq:basicWeylGeneral\]](#eq:basicWeylGeneral){reference-type="ref" reference="eq:basicWeylGeneral"}) we obtain \[\aligned \overline R^i_{jmn}&=R^i_{jmn}+\delta^i_j\big(\overline\zeta^{(s)}_{[mn]}- \zeta^{(s)}_{[mn]}\big)+\delta^i_m\big(\overline\zeta^{(s)}_{jn}- \zeta^{(s)}_{jn}\big)-\delta^i_n\big(\overline\zeta^{(s)}_{jm}- \zeta^{(s)}_{jm}\big)\\&- \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).i}_{jmn}+ \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).i}_{jnm}+ \widetilde{{\mathcal D}}{}^{(s_2).(s_3).i}_{jmn}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).i}_{jnm}. \endaligned\label{eq:R-Rgeneral}\] After contracting this equation by indices \(i\) and \(j\), we get \[\aligned (N+1)\big(\overline\zeta^{(s)}_{[mn]}-\zeta^{(s)}_{[mn]}\big)&= -\overline R_{[mn]}+\widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha mn}- \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha nm}\\&+R_{[mn]}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha mn}+ \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha nm}. \endaligned\label{eq:R-Rgenerali=j}\] From the equations ([\[eq:R-Rgeneral\]](#eq:R-Rgeneral){reference-type="ref" reference="eq:R-Rgeneral"}, [\[eq:R-Rgenerali=j\]](#eq:R-Rgenerali=j){reference-type="ref" reference="eq:R-Rgenerali=j"}), one obtains \[\aligned \overline R^i_{jmn}&=R^i_{jmn}+\frac1{N+1}\delta^i_j\big(R_{[mn]}-\overline R_{[mn]}\big)+\delta^i_m\big(\overline\zeta^{(s)}_{jn}- \zeta^{(s)}_{jn}\big)-\delta^i_n\big(\overline\zeta^{(s)}_{jm}- \zeta^{(s)}_{jm}\big)\\& +\frac1{N+1}\delta^i_j\big(\widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha mn}- \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha nm}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha mn}+ \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha nm}\big)\\&- \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).i}_{jmn}+ \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).i}_{jnm}+ \widetilde{{\mathcal D}}{}^{(s_2).(s_3).i}_{jmn}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).i}_{jnm}. \endaligned\label{eq:R-Rgenerali=j2}\] From the contraction of this result by indices \(i\) and \(n\), we get \[\aligned (N-1)\big(\overline\zeta^{(s)}_{jm}-\zeta^{(s)}_{jm}\big)&= \frac N{N+1}\big(R_{jm}-\overline R_{jm}\big)+ \frac1{N+1}\big(R_{mj}-\overline R_{mj}\big)\\&+ \frac1{N+1}\big( \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jm]}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jm]}\big)\\&- \widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[m\alpha]}+ \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[m\alpha]}. \endaligned\label{eq:R-Rgenerali=n}\] With regard to the equations ([\[eq:R-Rgenerali=j2\]](#eq:R-Rgenerali=j2){reference-type="ref" reference="eq:R-Rgenerali=j2"}, [\[eq:R-Rgenerali=n\]](#eq:R-Rgenerali=n){reference-type="ref" reference="eq:R-Rgenerali=n"}), we get \[{{\overline W}}{}^{(s).[1].i}_{jmn}= {{W}}{}^{(s).[1].i}_{jmn},\] for \[\aligned {{W}}{}^{(s).[1].i}_{jmn}&=R^i_{jmn}+ \frac1{N+1}\delta^i_jR_{[mn]}+ \frac N{N^2-1}\delta^i_{[m}R_{jn]}+ \frac1{N^2-1}\delta^i_{[m}R_{n]j}\\& +\widetilde{\mathcal D}{}^{(s_2).(s_3).i}_{j[mn]}- \frac1{N+1}\delta^i_j\widetilde{\mathcal D}{}^{(s_2).(s_3).\alpha}_{\alpha[mn]}\\&+ \frac1{N^2-1}\delta^i_m\big((N+1)\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[n\alpha]}-\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jn]}\big)\\&- \frac1{N^2-1}\delta^i_n\big((N+1)\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[m\alpha]}-\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jm]}\big), \endaligned\label{eq:generalWeylderived1} \tag{\(i\)}\] and the corresponding \({{\overline W}}{}^{(s).[1].i}_{jmn}\). Let us test are some summands in the invariant ([\[eq:generalWeylderived1\]](#eq:generalWeylderived1){reference-type="ref" reference="eq:generalWeylderived1"}) invariants of the mapping \(f\). After contract the equality \({{\overline W}}{}^{(s).[1].i}_{jmn}-{{W}}{}^{(s).[1].i}_{jmn}=0\) by indices \(i\) and \(n\), we get \[\widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jm]}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{\alpha[jm]}=0.\] Hence, the invariant ([\[eq:generalWeylderived1\]](#eq:generalWeylderived1){reference-type="ref" reference="eq:generalWeylderived1"}) reduces to \[\aligned {{W}}{}^{(s).[2].i}_{jmn}&=R^i_{jmn}+ \frac1{N+1}\delta^i_jR_{[mn]}+ \frac N{N^2-1}\delta^i_{[m}R_{jn]}+ \frac1{N^2-1}\delta^i_{[m}R_{n]j}\\& +\widetilde{\mathcal D}{}^{(s_2).(s_3).i}_{j[mn]}+ \frac1{N-1}\big(\delta^i_m\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[n\alpha]}- \delta^i_n\widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[m\alpha]}\big), \endaligned\label{eq:generalWeylderived1final} \tag{\(ii\)}\] and the corresponding \({\overline{W}}{}^{(s).[2].i}_{jmn}\). Let us check-out are there invariants of the mapping \(f\) into the invariant ([\[eq:generalWeylderived1final\]](#eq:generalWeylderived1final){reference-type="ref" reference="eq:generalWeylderived1final"}). After contracting the equality \({\overline{W}}{}^{(s).[2].i}_{jmn}- {{W}}{}^{(s).[2].i}_{jmn}=0\) by the indices \(i\) and \(j\), one obtains \[\widetilde{\overline{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[n\alpha]}- \widetilde{{\mathcal D}}{}^{(s_2).(s_3).\alpha}_{j[n\alpha]}=0.\] Hereof, the invariant ([\[eq:generalWeylderived1final\]](#eq:generalWeylderived1final){reference-type="ref" reference="eq:generalWeylderived1final"}) reduces to \[\aligned W{}^{(s).i}_{jmn}&=R^i_{jmn}+ \frac1{N+1}\delta^i_jR_{[mn]}+ \frac N{N^2-1}\delta^i_{[m}R_{jn]}+ \frac1{N^2-1}\delta^i_{[m}R_{n]j}\\& +\widetilde{\mathcal D}{}^{(s_2).(s_3).i}_{jmn}- \widetilde{\mathcal D}{}^{(s_2).(s_3).i}_{jnm}. \endaligned\label{eq:Weylgensimf}\] It holds the following theorem:
The geometrical object ([\[eq:Weylgensimf\]](#eq:Weylgensimf){reference-type="ref" reference="eq:Weylgensimf"}) is *the derived associated invariant* of the Weyl type.
## \(F\)-planar mappings
In this part of paper, we will apply the above obtained results. The theoretical part of this application will be search for invariants of conformal mappings. The practical example will be about a transformation of a special Riemannian space.\
**\(F\)-planar mappings.** A mapping \(f:\mathbb{A}_N\to\mathbb{\overline A}_N\) is called *the \(F\)-planar mapping* if it is determined with the equation \[\overline L^i_{\underline{jk}}= L^i_{\underline{jk}}+\delta^i_k\psi_j+\delta^i_j\psi_k+ F^i_k\sigma_j+F^i_j\sigma_k, \label{eq:F-planardfn}\] for \(1\)-forms \(\psi_j,\sigma_j\) and an affinor \(F^i_j\). The basic equation of the inverse mapping \(f^{-1}:\mathbb {A}_N\to\mathbb{\overline{A}}_N\) is \[L^i_{\underline{jk}}=\overline L^i_{\underline{jk}}- \delta^i_k\psi_j-\delta^i_j\psi_k-F^i_k\sigma_j- F^i_j\sigma_k. \label{eq:F-planar-1dfn}\] Hence, we get the mapping \(f^{-1}\) is an \(F\)-planar mapping for \[\begin{aligned}
\overline F^i_j=F^i_j,& \overline\sigma_j=-\sigma_j,& \overline\psi_j=-\psi_j. \label{eq:F-planarpsisigmaF-1} \end{aligned}\] Moreover, it holds \[\overline L^i_{\underline{jk}}= L^i_{\underline{jk}}+\delta^i_k\psi_j+\delta^i_j\psi_k- \frac12\overline F^i_k\overline\sigma_j- \frac12\overline F^i_j\overline\sigma_k+ \frac12F^i_k\sigma_j+\frac12F^i_j\sigma_k. \label{eq:F-planaromega-omega1}\] Therefore, the \(F\)-planar mapping \(f\) is the case of \(s_1=1,s_2=1/2,s_3=0\), in the equation ([\[eq:omegagen\]](#eq:omegagen){reference-type="ref" reference="eq:omegagen"}). After contracting the equation ([\[eq:F-planaromega-omega1\]](#eq:F-planaromega-omega1){reference-type="ref" reference="eq:F-planaromega-omega1"}) by indices \(i\) and \(k\), one gets \[\aligned \psi_j&=\frac1{N+1}\Big(\overline L^\alpha_{\underline{j\alpha}} +\frac12\overline F^i_k\overline\sigma_j+ \frac12\overline F^i_j\overline\sigma_k\Big)- \frac1{N+1}\Big(L^\alpha_{\underline{j\alpha}} +\frac12F^i_k\sigma_j+ \frac12F^i_j\sigma_k\Big). \endaligned\label{eq:F-planarpsij}\] With regard to the equations ([\[eq:omegagen\]](#eq:omegagen){reference-type="ref" reference="eq:omegagen"}, [\[eq:F-planaromega-omega1\]](#eq:F-planaromega-omega1){reference-type="ref" reference="eq:F-planaromega-omega1"}, [\[eq:F-planarpsij\]](#eq:F-planarpsij){reference-type="ref" reference="eq:F-planarpsij"}), we obtain \[\rho_j=\frac1{N+1}\Big(L^\alpha_{\underline{j\alpha}} +\frac12F\sigma_j+ \frac12F^\alpha_j\sigma_\alpha\Big), \label{eq:F-planarrhoj}\] for \(F=F^\alpha_\alpha\). From the equation ([\[eq:basicThomasGeneral\]](#eq:basicThomasGeneral){reference-type="ref" reference="eq:basicThomasGeneral"}), it holds that the associated invariant of the Thomas type of the mapping \(f\) is \[\begin{aligned}
&\aligned \widetilde{T}{}^i_{jk}&=L^i_{\underline{jk}}-\frac12\big(F^i_j\sigma_k+F^i_k\sigma_j\big)- \frac1{N+1}\delta^i_j\Big(L^\alpha_{\underline{k\alpha}}- \frac12\big(F^\alpha_k\sigma_\alpha+F\sigma_k\big)\Big) \\&- \frac1{N+1}\delta^i_k\Big(L^\alpha_{\underline{j\alpha}}- \frac12\big(F^\alpha_j\sigma_\alpha+F\sigma_j\big)\Big), \endaligned\label{eq:F-planarThomas}\\& \aligned \widetilde{T}{}^i_{jk}&=T^i_{{jk}}-\frac12\big(F^i_j\sigma_k+F^i_k\sigma_j\big)+ \frac1{2(N+1)}\delta^i_j \big(F^\alpha_k\sigma_\alpha+F\sigma_k\big) \\&+ \frac1{2(N+1)}\big(F^\alpha_j\sigma_\alpha+F\sigma_j\big), \endaligned\tag{\ref{eq:F-planarThomas}'}\label{eq:F-planarThomas'} \end{aligned}\] for the generalized Thomas projective parameter \(T^i_{jk}\) from the equation ([\[eq:Thomasgeodesic\]](#eq:Thomasgeodesic){reference-type="ref" reference="eq:Thomasgeodesic"}). Based on the equation ([\[eq:F-planarpsisigmaF-1\]](#eq:F-planarpsisigmaF-1){reference-type="ref" reference="eq:F-planarpsisigmaF-1"}), we obtain \[\overline F^i_j\overline F^m_n\overline\sigma_p\overline\sigma_q= F^i_jF^m_n\sigma_p\sigma_q,\] i.e. the geometrical objects ([\[eq:zeta(s)\]](#eq:zeta(s)){reference-type="ref" reference="eq:zeta(s)"}, [\[eq:D(s2)(s3)\]](#eq:D(s2)(s3)){reference-type="ref" reference="eq:D(s2)(s3)"}) reduce to \[\begin{aligned}
&\aligned\widetilde{\mathcal D}{}^i_{jmn}=-\frac12\big(F^i_j\sigma_m+F^i_m\sigma_j\big)_{|n}, \endaligned\label{eq:F-planarD(s2)(s3)}\\ &\aligned \zeta_{ij}&=\frac1{N+1}L^\alpha_{\underline{i\alpha}|j}+\frac1{(N+1)^2} L^\alpha_{\underline{i\alpha}}L^\beta_{\underline{j\beta}}+ \frac1{2(N+1)}L^\beta_{\underline{\alpha\beta}} \big(F^\alpha_i\sigma_j+F^\alpha_j\sigma_i\big)\\&+ \frac1{2(N+1)}\Big(\big(F\sigma_i+F^\alpha_i\sigma_\alpha\big)_{|j}+ L^\alpha_{\underline{i\alpha}}\big(F\sigma_j+F^\beta_j\sigma_\beta\big)+ L^\alpha_{\underline{j\alpha}}\big(F\sigma_i+F^\beta_i\sigma_\beta\big)\Big). \endaligned\label{eq:F-planarsigma} \end{aligned}\] Therefore, the invariant ([\[eq:basicWeylGeneral\]](#eq:basicWeylGeneral){reference-type="ref" reference="eq:basicWeylGeneral"}) of the mapping \(f\) is \[\aligned \widetilde{\mathcal W}{}^i_{jmn}&=R^i_{jmn}+\frac1{N+1}\delta^i_jR_{[mn]}-\frac12\big(F^i_j\sigma_m+F^i_m\sigma_j\big)_{|n}+ \frac12\big(F^i_j\sigma_n+F^i_n\sigma_j\big)_{|m}-\delta^i_{[m}\zeta_{jn]}, \endaligned\label{eq:F-planarWbasic}\] for \(\zeta_{ij}\) given in the equation ([\[eq:F-planarsigma\]](#eq:F-planarsigma){reference-type="ref" reference="eq:F-planarsigma"}). The invariant ([\[eq:Weylgensimf\]](#eq:Weylgensimf){reference-type="ref" reference="eq:Weylgensimf"}) of the \(F\)-planar mapping \(f\) is \[\widetilde W^i_{jmn}=W^i_{jmn}-\frac12\big(F^i_j\sigma_m+F^i_m\sigma_j\big)_{|n}+ \frac12\big(F^i_j\sigma_n+F^i_n\sigma_j\big)_{|m}, \label{eq:F-planarWderived}\] for the Weyl projective tensor \(W^i_{jmn}\).\ | {'timestamp': '2018-11-05T02:08:56', 'yymm': '1706', 'arxiv_id': '1706.01784', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01784'} |
# Introduction
Determining a function (or a class of functions) of a certain regularity fitting to a prescribed set of data is one of the most challenging problems in modern mathematics. The origin of this problem is very old, since this general framework encompasses classical problems of applied analysis. Depending on the requested regularity, it goes from the Tietze extension theorem in normal topological spaces, where the required regularity is minimal (continuity), to results where the requested regularity is progressively increasing: McShane results on uniformly continuous, Hölder or Lipschitz extensions , Lipschitz extensions for vector-valued functions (Valentine ), differentiable and \(C^{k}\)-extensions (Whitney , Glaeser , and more recently Brudnyi-Shvartsman , Zobin , Fefferman ), monotone multivalued extensions (Bauschke-Wang ), definable (in some o-minimal structure) Lipschitz extensions (Aschenbrenner-Fischer ), etc. In this work we are interested in the Glaeser-Whitney \(C^{1,1}\)-extension problem, which we describe below.Let \(S\) be a nonempty subset of a Hilbert space \((\mathcal{H},\langle \cdot,\cdot\rangle, |\cdot|)\) and assume \(\alpha:S\rightarrow\mathbb{R}\) and \(v:S\rightarrow\mathcal{H}\) satisfy the so-called Glaeser-Whitney conditions: \[\begin{aligned}
\label{cond-GW} \left\{ \begin{array} [c]{l} \displaystyle \mathop{\rm sup}_{s_1,s_2\in S, s_1\not= s_2} \frac{|\alpha(s_{2})-\alpha(s_{1})-\langle v(s_{1}),s_{2}-s_{1}\rangle|}{|s_{1}-s_{2}|^{2}} := K_1 <+\infty, \bigskip \\ \displaystyle \mathop{\rm sup}_{s_1,s_2\in S, s_1\not= s_2} \frac{|v(s_{1})-v(s_{2})|}{|s_{1}-s_{2}|}:= K_2<+\infty. \end{array} \right. \end{aligned}\] In it has been shown that under the above conditions, in case \(\mathcal{H}=\mathbb{R}^{n},\) there exists a \(C^{1,1}\)-smooth function \(F:\mathbb{R}^{n}\rightarrow\mathbb{R}\) such that the prescribed \(1\)-Taylor field \((\alpha(s),v(s))\) coincides, at every \(s\in S\), with the \(1\)-Taylor field \((F(s),\nabla F(s))\) of \(F\). The above result has been extended to Hilbert spaces in Wells and Le Gruyer . In particular, in the following constant has been introduced: \[\begin{aligned}
\label{gamma1} && \Gamma^{1}(S,(\alpha,v)):= \mathop{\rm sup}_{s_1,s_2\in S, s_1\not= s_2} \left(\sqrt{A_{s_1 s_2}^2+B_{s_1 s_2}^2}+|A_{s_1 s_2}|\right), \end{aligned}\] where \[\begin{aligned}
A_{s_1 s_2}=\frac{2(\alpha(s_1)-\alpha(s_2))+\langle v(s_1)+v(s_2), s_2-s_1\rangle} {|s_1-s_2|^2}, \quad B_{s_1 s_2}=\frac{v(s_1)-v(s_2)}{|s_1-s_2|}. \end{aligned}\] It has been shown in that \(\Gamma^{1}(S,(\alpha,v))<+\infty\) if and only if conditions [\[cond-GW\]](#cond-GW){reference-type="eqref" reference="cond-GW"} hold. Moreover, in this case, the existence of a \(C^{1,1}\) function \(F:\mathcal{H}\rightarrow\mathbb{R}\) such that \(F|_{S}=\alpha,\) \(\nabla F|_{S}=v\) and \[\Gamma^{1}(\mathcal{H},(F,\nabla F))=\Gamma^{1}(S,(\alpha,v)),\label{1}\] has been established. Henceforth, every \(C^{1,1}\)-extension of \((\alpha,v)\) satisfying [\[1\]](#1){reference-type="eqref" reference="1"} will be called a minimal Glaeser-Whitney extension. The terminology is justified by the fact that, for every \(C^{1,1}\) function \(G:\mathcal{H}\rightarrow\mathbb{R},\) we have \(\Gamma^{1}(\mathcal{H},(G,\nabla G))=\mathrm{Lip}(\nabla G)\) (see ). Thus \(\mathrm{Lip}(\nabla F)\leq \mathrm{Lip}(\nabla G)\) for any \(C^{1,1}\)-extension \(G\) of the prescribed \(1\)-Taylor field \((\alpha(s),v(s))\). If for some universal constant \(K \geq 1\) (not depending on the data) we have \(\Gamma^{1}(\mathcal{H},(G,\nabla G)) \leq K \, \Gamma^{1}(S,(\alpha,v))\) then the extension \(G\) will be called almost minimal. Recently, several authors have been interested in extensions that are subject to additional constraints: extensions which preserve positivity or convexity . In, D. Azagra and C. Mudarra considered the problem of finding a *convex* \(C^{1,1}\)-smooth extention over a prescribed Taylor polynomial \((\alpha(s),v(s))_{s\in S}\) in a Hilbert space \(\mathcal{H}\) and established that the condition \[\alpha(s_{2})\geq\alpha(s_{1})+\langle v(s_{1}),s_{2}-s_{1}\rangle+\frac {1}{2M}|v(s_{1})-v(s_{2})|^{2},\quad\forall s_{1},s_{2}\in S,\label{form-azagra}\] is necessary and sufficient for the existence of such extension.Inspired by the recent work concerning \(C^{1,1}\)-convex extensions, we revisit the classical Glaeser-Whitney problem. We first provide an alternative shorter proof of the result of concerning \(C^{1,1}\)-convex extensions in Hilbert spaces by giving a simple explicit formula. This formula is heavily based on the regularization via sup-inf convolution in the spirit of Lasry-Lions and can be efficiently computed, see Remark [\[rmk-num\]](#rmk-num){reference-type="ref" reference="rmk-num"}. As an easy consequence, we obtain a direct proof for the classical \(C^{1,1}\)-Glaeser-Whitney problem in Hilbert spaces, which goes together with an explicit formula of the same type as for the convex extension problem. Let us mention that the previous proofs are quite involved both in finite dimension and Hilbert spaces . In the finite dimensional case, a construction of the extension is proposed in and some explicit formulae can be found in but both are not tractable (see however the work for concrete computations). Our approach also compares favorably to the result of , in which the existence of minimal extensions is established. On the other hand, the extension given by our explicit formula may fail to be minimal---though it is almost minimal up to a universal multiplicative factor. Before we proceed, we recall that a function \(f:\mathcal{H} \rightarrow\mathbb{R}\) is called \(C_{\ast}\)-semiconvex (resp., \(C^{\ast}\)-semiconcave) when, for all \(x,y\in H,\) \[f(y)-f(x)-\langle\nabla f(x),y-x\rangle\geq-\frac{C_{\ast}}{2}|x-y|^{2} \ \ \text{(resp., \(\displaystyle\leq\frac{C^{\ast}}{2}|x-y|^{2}\)).}\] This is equivalent to assert that \(f+\frac{C_{\ast}}{2}|x|^{2}\) is convex (respectively \(f-\frac{C^{\ast}}{2}|x|^{2}\) is concave). When \(f\) is both \(C\)-semiconvex and \(C\)-semiconcave, then \(f\) is \(C^{1,1}\) in \(\mathcal{H}\) with \(\mathrm{Lip}(\nabla f)\leq C\) (for a proof of this latter result in finite dimension, see and use the arguments of to extend the result to Hilbert spaces).
# Convex \(C^{1,1}\) extension of \(1\)-fields {#ss-2}
For any \(f:\mathcal{H}\rightarrow\mathbb{R}\) and \(\varepsilon>0,\) we define respectively the sup and the inf-convolution of \(f\) by \[f^{\varepsilon}(x)=\mathop{\rm sup}_{y\in\mathcal{H}}\left \{ f(y)-\frac{|y-x|^{2}}{2\varepsilon} \right \},\quad f_{\varepsilon}(x)=\mathop{\rm inf}_{y\in\mathcal{H} }\left \{f(y)+\frac{|y-x|^{2}}{2\varepsilon} \right \}.\]
# \(C^{1,1}\) extension of \(1\)-fields: explicit formulae
Let us now apply the previous result to obtain a general \(C^{1,1}\)-extension in the Glaeser-Whitney problem.
# Limitations of the sup-inf approach
The main result (Theorem [\[thm-main-2\]](#thm-main-2){reference-type="ref" reference="thm-main-2"}) is heavily based on the explicit construction of a \(C^{1,1}\)-convex extension of a 1-Taylor field \((\alpha,v)\) satisfying [\[form-azagra\]](#form-azagra){reference-type="eqref" reference="form-azagra"}, which in turn, is based on the sup-inf convolution approach. The reader might wonder whether our approach can be adapted to include cases where less regularity is required, as for instance \(C^{1,\theta}\)-extensions, that is, extensions to a \(C^{1}\)-function whose derivative has a Hölder modulus of continuity with exponent \(\theta\in(0,1)\). The existence of such *convex* extensions (and even \(C^{1,\omega}\) convex extensions with a general modulus of continuity \(\omega\)) was established in finite dimensions in Azagra-Mudarra by means of involved arguments. Indeed, it would be natural to endeavor an adaptation of formula [\[ext-C11-f\]](#ext-C11-f){reference-type="eqref" reference="ext-C11-f"} to treat the problem of \(C^{1,\theta}\)-convex extensions, for \(0<\theta<1\). According to , the adequate condition, analogous to [\[form-azagra\]](#form-azagra){reference-type="eqref" reference="form-azagra"}, is that the 1-Taylor field has to satisfy, for some \(M>0\), \[\begin{aligned}
\label{form-azagra-holder} \alpha(s_{2})\geq\alpha(s_{1})+\langle v(s_{1}),s_{2}-s_{1}\rangle +\frac{\theta}{(1+\theta)M^{1/\theta}}|v(s_1)-v(s_2)|^{1+\frac{1}{\theta}}. \end{aligned}\] Unfortunately, the technique developed in Section [2](#ss-2){reference-type="ref" reference="ss-2"} is specific to the \(C^{1,1}\)-regularity and cannot be easily adapted to this more general case. Let us briefly explain the reason. Considering the suitable sup and inf-convolutions \[\begin{aligned}
f^\varepsilon(x)= \mathop{\rm sup}_{y\in \mathcal{H}}\left \{f(y)-\frac{|y-x|^{1+\theta}}{(1+\theta)\varepsilon^\theta}\right \}, \quad f_\varepsilon(x)= \mathop{\rm inf}_{y\in \mathcal{H}}\left \{f(y)+\frac{|y-x|^{1+\theta}}{(1+\theta)\varepsilon^\theta}\right \}, \end{aligned}\] all of the arguments of the proof of Theorem [\[thm-main-1\]](#thm-main-1){reference-type="ref" reference="thm-main-1"} go through except [\[form-cruciale\]](#form-cruciale){reference-type="eqref" reference="form-cruciale"}, which fails to hold in this general case. More precisely, the convex extension \(f\) defined by [\[fcvx\]](#fcvx){reference-type="eqref" reference="fcvx"} satisfies \[\begin{aligned}
\label{f-line-holder} l(x)\leq f(x)\leq q(x) \qquad \text{for all \(x\in \mathcal{H}\) and \(s\in S,\)} \end{aligned}\] with equalities for \(x=s\), where \[\begin{aligned}
&& l(x):= \alpha(s)+ \langle v(s), x-s\rangle \label{linear-theta}\\ && q(x):= \alpha(s)+ \langle v(s), x-s\rangle \label{parab-theta} +\frac{M}{1+\theta}|x-s|^{1+\theta}. \end{aligned}\] Therefore for every \(\varepsilon>0\) such that \(M\varepsilon^{\theta} <1,\) we have \[\begin{aligned}
l(x)\leq (f^\varepsilon)_\varepsilon(x)\leq (q^\varepsilon)_\varepsilon(x)\. \end{aligned}\] Nonetheless, we may now possibly have \[\begin{aligned}
\label{contr-theta} (q^\varepsilon)_\varepsilon(s)\not= q(s), \end{aligned}\] yielding that \((f^\varepsilon)_\varepsilon\) is a \(C^{1,\theta}\)-convex function but may differ from \(f\) on \(S\), hence it is not an extension of the latter. Let us underline that the problem arises even in dimension 1 and even for small \(\varepsilon.\) In particular, the sup-convolution \(q^\varepsilon\) may develop singularities for arbitrary small \(\varepsilon\) so that \(q^\varepsilon\) is not anymore in the same class as \(q\), contrary to the quadratic case (see [\[sup-conv-quad\]](#sup-conv-quad){reference-type="eqref" reference="sup-conv-quad"}).
# Appendix
## Independence of the Glaeser-Whitney constants in [\[cond-GW\]](#cond-GW){reference-type="eqref" reference="cond-GW"}
It is worth-noticing that there is no link between \(K_1\) and \(K_2\) in [\[cond-GW\]](#cond-GW){reference-type="eqref" reference="cond-GW"}. Each of these constants can be 0 while the other one can be very large, as it is shown in the following examples.
## Inequality estimations between \(\Gamma^{1}(S,(\alpha,v))\), \(K_1\) and \(K_2\).
The following result has been used in the last part of the proof of Theorem [\[thm-main-2\]](#thm-main-2){reference-type="ref" reference="thm-main-2"}
**Acknowledgement.** This work was partially supported by the Centre Henri Lebesgue ANR-11-LABX-0020-01. Major part of this work has been done during a research visit of the first author to INSA Rennes. This author is indebted to his hosts for hospitality. Research of Aris Daniilidis was partially supported by BASAL PFB-03, FONDECYT grant 1171854 (Chile) and MTM2014-59179-C2-1-P grant of MINECO (Spain) and ERDF (EU). | {'timestamp': '2018-02-20T02:14:56', 'yymm': '1706', 'arxiv_id': '1706.01721', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01721'} |
# Introduction
The recovery of camera parameters and scene structure have been studied for over two decades since several applications, such as 3D vision from multiple views , are heavily dependent on the quality of the camera calibration. In particular, two major calibration types can be considered: aiming at the determination of the intrinsic and/or extrinsic parameters. The former ones include focal lengths, principal point, aspect ratio, and non-perspective distortion parameters, while the extrinsic parameters are the relative pose. Assuming two cameras with unknown extrinsic and *a priori* intrinsic parameters except a common focal length is called the *semi-calibrated case* . It leads to the *unknown focal-length problem*: estimation of the relative motion and common focal length, simultaneously. The semi-calibrated case is realistic since (1) the aspect ratio is determined by the shape of the pixels on the sensors, it is usually 1:1; (2) the principal point is close to the center of the image, thus it is a reasonable approximation and (3) the distortion can be omitted if narrow field-of-view lenses are applied. Considering solely the locations of point pairs makes the problem solvable using at least six point pairs . The objective of this paper is to *solve the problem exploiting only two local affine transformations*. In general, 3D vision approaches including state-of-the-art structure-from-motion pipelines apply a robust estimator, e.g. RANSAC , augmented with a minimal method, such as the five or six-point algorithm as an engine. Selecting a method exploiting as few point pairs as possible gains accuracy and drastically reduces the processing time. Benefiting from estimators which use less input data, the understanding of low-textured environment becomes significantly easier . Moreover, minimal methods are advantageous from theoretical point-of-view leading to deeper understanding. Local affine transformations represent the warp between the infinitely close vicinities of corresponding point pairs and have been investigated for a decade. Their application field includes homography and surface normal estimation; recovery of the epipoles ; triangulation of points in 3D ; camera pose estimation ; structure-from-motion . In practice, local affinities can be accurately retrieved using e.g. affine-covariant feature detectors, such Affine-SIFT and Hessian-Affine . To the best of our knowledge, no paper has dealt with the unknown focal length problem using local affine transformations. This paper proposes two novel linear constraints describing the relationship between local affinities and epipolar geometry. Forming a multivariate polynomial system and solving it by the *hidden-variable technique* , the proposed method is efficient and estimates the focal length and the relative motion using only two affinities. In order to eliminate invalid roots, a novel condition is introduced investigating the geometry of local affinities. To select the best candidate out of the remaining ones, we propose a root selection technique which is as accurate as the state-of-the-art for small noise and outperforms it for high-level noise.
# Preliminaries and Notation
**Epipolar geometry.** Assume two perspective cameras with a common intrinsic camera matrix \(\mathbf{K}\) to be known. Fundamental and essential matrices are as follows: \[\mathbf{F} = \begin{bmatrix} f_1 & f_2 & f_3 \\ f_4 & f_5 & f_6 \\ f_7 & f_8 & f_9 \end{bmatrix}, \quad \mathbf{E} = \begin{bmatrix} e_1 & e_2 & e_3 \\ e_4 & e_5 & e_6 \\ e_7 & e_8 & e_9 \end{bmatrix}.\] If the cameras are calibrated (\(\mathbf{K}\) is known) matrix \(\mathbf{F}\) can be transformed to be an essential matrix \(\mathbf{E}\) as follows: \[\mathbf{E} = \textbf{K}^{T} \mathbf{F} \mathbf{K}.\] The epipolar relationship of corresponding point pair \(\mathbf{p}_1\) and \(\mathbf{p}_2\) are described by \(\mathbf{F}\) as \[\label{eq:epipolar_constraint} \mathbf{p}_2^T \mathbf{F} \mathbf{p}_1 = 0.\] A valid fundamental matrix must satisfy singularity constraint \(\det(\mathbf{F}) = 0\). Considering this cubic constraint and the fact that a fundamental matrix is defined up to an arbitrary scale, its degrees-of-freedom is reduced to seven. Thus seven point pairs are enough for the estimation. As the essential matrix encapsulates the full camera motion, the orientation and direction of the translation, it has five degrees-of-freedom. The two additional constraints are described by the well-known trace constraint as \[\label{eq:trace_constraint} 2 \mathbf{E} \mathbf{E}^T \mathbf{E}-\textrm{tr}(\mathbf{E} \mathbf{E}^T) \textbf{E} = 0.\] Even though Eq. [\[eq:trace_constraint\]](#eq:trace_constraint){reference-type="ref" reference="eq:trace_constraint"} yields nine polynomial equations for \(\mathbf{E}\), only two of them are algebraically independent. Semi-calibrated case is assumed in this paper as only the common focal-length \(f\) is considered to be unknown. Without loss of generality, the intrinsic camera matrix is = \^T= (f, f, 1), \(where\)f\(is the unknown focal-length. In order to replace\[with\]in Eq.~\ref{eq:trace_constraint} we define matrix\[as follows: \begin{equation} \mathbf{Q} = \text{diag} \left( 1,1,\tau \right), \quad \tau = f^{-2}. \end{equation} \noindent Due to the fact that\]is non-singular, and trace(\) \^T\() identifies a scalar value, Eq.~\ref{eq:trace_constraint} can be simplified by multiplying with\)\^-T\(and\)\^-1\(from the left and the right sides, respectively. Moreover, trace is invariant under cyclic permutations. As a consequence, Eq.~\ref{eq:trace_constraint} is written as~ \begin{equation} \label{eq:focal_trace_constraint} 2 \mathbf{F} \mathbf{Q} \mathbf{F}^T \mathbf{Q} \mathbf{F}-\textrm{tr}(\mathbf{F} \mathbf{Q} \mathbf{F}^T \mathbf{Q}) \mathbf{F} = 0. \end{equation} This relationship will help us to recover the focal length and the fundamental matrix using two affine correspondences. \noindent \textbf{An affine correspondence}\)(\_1, \_2, )\(consists of a corresponding point pair and the related local affinity\[transforming the vicinity of point\)\_1\(to that of\)\_2\(. In the rest of the paper,\]is considered as its left\)`<!-- -->`{=html}2 \(submatrix \begin{equation*} \mathbf{A} = \begin{bmatrix} a_1 & a_2 \\ a_3 & a_4 \end{bmatrix} \end{equation*} since the third column--the translation part--is determined by the point locations. We use the \textbf{hidden variable technique} in the proposed method. It is a resultant technique in algebraic geometry for the elimination of variables from a multivariate polynomial system~. Suppose that\)m\(polynomial equations in\)n\(variables are given. In brief, one can assume an unknown variable as a parameter and rewrite the equation system as\)**C**(y_1) **x** = 0\(, where\)**C**\(is a coefficient matrix depending on the unknown\)y_1\((hidden variable) and vector\)**x**\(is the vector of\)n-1\(unknowns. If the number of equations equals to that of the unknown monomials in\)**x**\(, i.e. matrix\)**C**\(is square, the non-trivial solution can be carried out as\)(**C**(y_1)) = 0\(. Solving the resultant equation for\)y_1\(and back-substituting it, the whole system is solved. \section{Focal-length using Two Correspondences} This section aims the recovery of the unknown focal length and fundamental matrix using two affine correspondences. First, the connection between the fundamental matrix and local affinity is introduced, then we discuss the estimation technique. \vspace{5px} \noindent \vspace{2px} \textbf{3.1. Exploiting a Local Affine Transformation} Suppose that an affine correspondence\)(, , )\(and fundamental matrix\[are known. It is trivial that every affine transformation preserves the direction of the lines going through points\]and\[on the first and second images. As a consequence, the link between directions\]and\[of epipolar lines can be described~ by affine transformation\]as \begin{equation} \label{eq:a_v1_par_v2} \mathbf{A} \mathbf{v_1} \parallel \mathbf{v_2}. \end{equation} Reformulating Eq.~\ref{eq:a_v1_par_v2} using the well-known fact from Computer Graphics~ leads to\)\^-T \^90 = \^90 \(, where matrix\)\^90\(is a 2D orthonormal (rotation) matrix rotating with\)`<!-- -->`{=html}90\(degrees and\[is an unknown scale. Vectors\)\^90 \(and\)\^90 \(are the line normals\]and\[as \begin{equation} \label{eq:a_n1_eq_n2} \mathbf{A}^{-T} \mathbf{n_1} = \beta \mathbf{n_2}. \end{equation} In Appendix~\ref{appendix:proof_affine_constraints}, it is proven that\]is equal to\)-1\(if\[and\]are calculated from the fundamental matrix using relationships\) \(and\)\^T \(and they are \textit{not normalized}. In brief, it is given as the distance ratio of neighboring epipolar lines on the two images. For the case when the normals are not normalized--the original scale has not been changed--,\[is only a scale inverting the directions. Normals are expressed from\]as the first two coordinates of the epipolar lines:\)\_1 = (\_1)\_(1:2) = (\^T \_2)\_(1:2)\(and\)\_2 = (\_2)\_(1:2) = ( \_1)\_(1:2)\(~, where the lower indices select a subvector. Therefore, Eq.~\ref{eq:a_n1_eq_n2} is written as \begin{equation} \mathbf{A}^{-T} (\mathbf{F}^T \mathbf{p}_2)_{(1:2)} =-(\mathbf{F} \mathbf{p}_1)_{(1:2)} \end{equation} and forms a system of linear equations consisting of two equations as follows: \begin{eqnarray} \label{eq:affine_constraint_1} (u_2 + a_1 u_1) f_1 + a_1 v_1 f_2 + a_1 f_3 + (v_2 + a_3 u_1) f_4 + \nonumber \\ a_3 v_1 f_5 + a_3 f_6 + f_7 = 0 \\ \label{eq:affine_constraint_2} a_2 u_1 f_1 + (u_2 + a_2 v_1) f_2 + a_2 f_3 + a_4 u_1 f_4 + \nonumber \\ (v_2 + a_4 v_1) f_5 + a_4 f_6 + f_8 = 0. \end{eqnarray} Thus each local affine transformation \textit{reduces the degrees-of-freedom by two}. \vspace{5px} \noindent \vspace{2px} \textbf{3.2. Two-point Solver} Suppose that two affine correspondences (\)\_1\^1\(,\)\_2\^1\(,\)\^1\() and (\)\_1\^2\(,\)\_2\^2\(,\)\^2\() are given. Coefficient matrix \begin{equation*} \resizebox{1.0\columnwidth}{!}{ \)\mathbf{C}^i = \begin{bmatrix} u_2 + a_1 u_1 & a_1 v_1 & a_1 & v_2 + a_3 u_1 & a_3 v_1 & a_3 & 1 & 0 & 0 \\ a_2 u_1 & u_2 + a_2 v_1 & a_2 & a_4 u_1 & v_2 + a_4 v_1 & a_4 & 0 & 1 & 0 \\ u_1 u_2 & v_1 u_2 & u_2 & u_1 v_2 & v_1 v_2 & v_2 & u_1 & v_1 & 1 \end{bmatrix}\( } \end{equation*} related to the\)i\(-th (\)i {1,2}\() correspondence is formed as the combination of Eqs.~\ref{eq:epipolar_constraint},~\ref{eq:affine_constraint_1},~\ref{eq:affine_constraint_2} and satisfies formula\)\^i = 0\(, where\) =
\^T\(is the vector of unknown elements of the fundamental matrix. We denote the concatenated coefficient matrix of both correspondences as follows: \begin{equation} \mathbf{C} = \begin{bmatrix} \mathbf{C}^1 \\ \mathbf{C}^2 \end{bmatrix}. \end{equation} It is of size\)`<!-- -->`{=html}6 \(, therefore, its left null space is three-dimensional. The solution is carried out as \begin{equation} \label{eq:null_space} \mathbf{x} = \alpha \mathbf{a} + \beta \mathbf{b} + \gamma \mathbf{c}, \end{equation} where\[,\]and\[are the singular vectors and\],\[,\]are unknown non-zero scalar values. Remember that only the common focal length is unknown from the intrinsic parameters, therefore, we are able to exploit the trace constraint. Eq.~\ref{eq:focal_trace_constraint} yields ten cubic equations for four unknowns\[,\],\[and\], where\)= f\^-2\(encapsulates the unknown focal length. We consider\[as the hidden variable and form coefficient matrix\)()\(w.r.t.\ the other three ones--thus the rows of\)()\(are univariate polynomials with variable\]. Even though\[,\]and\[are defined up to a common scale, we do not fix this scale in order to keep the homogenity of the system. The monomials of this polynomial system are as\) = \[ \^3 \^2 \^2 \^2 \^2 \^3 \^2 \^2 \^3 \]\^T \(. Table~\ref{tab:coefficient_table} demonstrates the coefficient matrix. Since the scale of monomial vector\]has not been fixed, the non-trivial solution of equation\)() = 0\(is when the determinant vanishes as \begin{equation} \label{eq:determinant_hidden} \det(\mathbf{C}(\tau)) = 0. \end{equation} Therefore, the hidden-variable resultant--a polynomial of the hidden variable--is\)(())\(. As the current problem is fairly similar to that of~, we adopt the proposed algorithm. It is proved that\)(())\(is actually a 15-th degree polynomial and it obtains the candidate values for\[. Then the solution for\],\[,\]and\[is given as\) = (())\(. Finally, fundamental matrix\]regarding to each obtained focal length can be directly estimated using Eq.~\ref{eq:null_space}. \begin{table}[h!] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{ | c | c c c c c c c c c c | } \hline \multirow{2}{*}{\)\mathbf{C}(\tau)\(} & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ & \)\alpha^3\( & \)\alpha^2 \beta\( & \)\alpha^2 \gamma\( & \)\alpha \beta^2\( & \)\alpha \beta \gamma\( & \)\alpha \gamma^2\( & \)\beta^3\( & \)\beta^2 \gamma\( & \)\beta \gamma^2\( & \)\gamma^3\( \\ \hline 1 & \)c_1\( & \)c_2\( & \)c_3\( & \)c_4\( & \)c_5\( & \)c_6\( & \)c_7\( & \)c_8\( & \)c_9\( & \)c_{10}\( \\ . &. &. &. &. &. &. &. &. &. &.\\ 10 & \)c_{91}\( & \)c_{92}\( & \)c_{93}\( & \)c_{94}\( & \)c_{95}\( & \)c_{96}\( & \)c_{97}\( & \)c_{98}\( & \)c_{99}\( & \)c_{100}\( \\ \hline \end{tabular} } \caption{ The coefficient matrix \)\mathbf{C}(\tau)\( related to the ten polynomial equations of the trace constraint.} \label{tab:coefficient_table} \end{table} \section{Elimination and Selection of Roots} In this section, a novel technique is proposed to omit roots on the basis of the underlying geometry. Then we show a heuristics considering the properties of digital cameras to remove invalid focal lengths. In the end, we introduce a root selection algorithm. \vspace{5px} \noindent \vspace{2px} \textbf{4.1. Elimination of Invalid Focal Lengths} A solution is proposed here based on the underlying geometry to eliminate invalid focal lengths. Suppose that a point pair\)(\_1, \_2)\(, the related local affinity\[, the fundamental matrix\], and an obtained focal length\)f\(are given. As the semi-calibrated case is assumed,\[and\)f\(exactly determines the projection matrices\)\_1\(and\)\_2\(of both cameras~. Denote the 3D coordinates and the surface normal induced by point pair\)(\_1, \_2)\(, local affinity\]and the projection matrices with\) = \[x y z\]\^T\(and\) = \[n_x n_y n_z\]\^T\(, respectively. According to our experiences, linear triangulation~ is a suitable and efficient choice to estimate\[. Surface normal\]is estimated exploiting affinity\[by the method proposed in~.\footnote{\url{http://web.eee.sztaki.hu/~dbarath/}} Without loss of generality, we assume that a point of a 3D surface cannot be observed from behind. As a consequence, the angle between vectors\)\_i-\(and\]must be smaller than\)`<!-- -->`{=html}90\^\(for both cameras, where\)\_i\(is the position of the\)i\(-th camera (\)i {1,2}\(). This can be interpreted as follows: each camera selects a half unit-sphere around the observed point\)**q**\(. Surface normal\)**n**\(must lie in the intersection of these half spheres. These half spheres are described by a rectangle in the spherical coordinate system as follows: \textrm{rect}_{i} = \begin{bmatrix} \theta_{i}-\frac{\pi}{2} & \sigma_{i}-\frac{\pi}{4} & \pi & \frac{\pi}{2} \end{bmatrix}\), where \(\theta_{i}\), \(\sigma_{i}\) are the corresponding spherical coordinates and \(\textrm{rect}_{i}\) is of format \(\begin{bmatrix} \textrm{corner}_\theta & \textrm{corner}_\sigma & \textrm{width} & \textrm{height} \end{bmatrix}\). The intersection area induced by the two cameras is as \[\textrm{rect}_\cap = \bigcap_{i \in [1,2]} \textrm{rect}_{i}.\] Point \(\textbf{q}\) is observable from both cameras *if and only if* surface normal \(\textbf{n}\), represented by spherical coordinates \(\Theta\) and \(\Sigma\), lies in the intersection area: \(\begin{bmatrix} \Theta & \Sigma \end{bmatrix} \in \textrm{rect}_\cap\). A setup, induced by focal length \(f\), not satisfying this criteria is an invalid one and can be omitted. Note that this constraint can be straightforwardly extended to the multi-view case making the intersection area more restrictive. **4.2. Physical Properties of Cameras** We introduce restrictions on the estimated roots considering the physical limits of the cameras. The focal length within camera matrix \(\mathbf{K}\) is not equivalent to the focal length of the lenses, since it is the ratio of the optical focal length and the pixel size . Particularly, the latter one is a few micrometers, while the optical focal length are within interval \([1 \dots 500]\) mm. Therefore, coarse lower and upper limits for a realistic camera are \(100\) and \(500.000\). Focal lengths out of this interval are automatically discarded. Note that these limits can be easily changed considering cameras with different properties. **4.3. Root Selection** To resolve the ambiguity of multiple roots and to minimize the effect of the noise, the classical way is to exploit multiple measurements eliminating the inconsistent ones. Since Eq. [\[eq:determinant_hidden\]](#eq:determinant_hidden){reference-type="ref" reference="eq:determinant_hidden"} is a high-degree polynomial it is sensitive to noise--small changes in the coordinates and affine elements cause significantly different coefficients. RANSAC is a successful technique for that problem, e.g. in the five-point relative-orientation one . Recent methods, i.e. Kernel Voting, exploit the property that the roots form a peak around the real solution . Kernel Voting maximizes a kernel density function like a maximum-likelihood-decision-maker. To our experiences, this technique works accurately if the noise in the coordinates does not exceed \(1-2\) pixels on average. Over that, the roots may form several strongly supported peaks and it is not guaranteed that the true solution is found. Thus we formulate the problem as a mode-seeking in a one dimensional domain: the real focal length appears as the most supported mode. Among several mode-seeking techniques the most robust one is the Median-Shift according to extensive experimentation. Median-Shift providing Tukey-medians as modes does not generate new elements in the domain it is applied to. In particular, there is no significant difference in the results of Tukey- and Weiszfeld-medians , however, the former one is slightly faster to compute. Finally, in order to overcome the discrete nature of Median-Shift--since it does not add new instances, only operates with the given ones--, we apply a gradient descent from the retrieved mode \(x_0\) maximizing function \[f(x) = \sum_{i = 1}^{n} \frac{\kappa(x_i-x)}{h},\] where \(n\) is the number of focal lengths, \(\kappa\) is a kernel function--we chose Gaussian-kernel--, \(x_i\) is the \(i\)-th focal length, and \(h\) is a bandwidth same as for the Median-Shift.
# Experimental Results
For the synthesized tests, we used the MATLAB code shown in Alg. [\[MatlabAlgorithm\]](#MatlabAlgorithm){reference-type="ref" reference="MatlabAlgorithm"}. For the real world tests, we used our C++ implementation[^1] which is a modification of the solver of Hartley et al. . **5.1. Synthesized tests** For synthesized testing, two perspective cameras are generated by their projection matrices \(\mathbf{P}_1\) and \(\mathbf{P}_2\). The first camera is at position \([0 \; 0 \; 1]^T\) looking towards the origin, and the distance of the second one from the first is \(0.15\) in a random direction. Five random planes passing over the origin are generated and each is sampled in fifty random locations. The obtained 3D points are projected onto the cameras. Zero-mean Gaussian-noise is added to the point coordinates. The local affine transformations are calculated by derivating the homographies induced by the tangent planes at the noisy point correspondences similarly to . Figure [\[fig:root_number\]](#fig:root_number){reference-type="ref" reference="fig:root_number"} reports the kernel density function with Gaussian-kernel width \(10\) plotted as the function of the relative error (in percentage). Candidate focal lengths are estimated as follows: **1.** Select two affine correspondences. **2.** Apply the proposed 2-point method. **3.** Repeat from Step 1. The iteration limit is chosen to \(100\). The blue horizontal line reports the result of Median-Shift, the green one is that of Kernel Voting. The \(\sigma\) value of the zero-mean Gaussian-noise added to the point locations and affinities is (a) \(0.01\) pixels, (b) \(0.1\) pixels, (c) \(1.0\) pixels, (d) \(3.0\) pixels, (e) \(3.0\) pixels and there are \(10\%\) outliers, (f) \(1.0\) pixels with some errors in the aspect ratio: the true one is \(1.00\) but \(0.95\) is used. The real focal length is \(600\). Confirming the validity of the proposed theory, the peak is over the ground truth focal length: \(0\%\) relative error. The proposed root selection is more robust than the Kernel Voting approach since the blue line is closer to the zero relative error even if the noise is high. Fig. [\[fig:fundamental_error\]](#fig:fundamental_error){reference-type="ref" reference="fig:fundamental_error"} reports the mean (top) and median (bottom) errors of the estimated fundamental matrices plotted as the function of the noise \(\sigma\) and compared with the results of Hartley et al. and Perdoch et al.. The error is the Frobenious norm of the estimated and ground truth fundamental matrices. \(100\) runs were performed on each noise level. It can be seen that the accuracy of the estimated fundamental matrices is similar to that of Hartley et al. . **5.2. Tests on Real Images[^2]** To test the proposed method on real world photos, \(104\) image pairs were downloaded[^3] each containing the ground truth focal length in the EXIF data (see Fig. [\[fig:example_pairs\]](#fig:example_pairs){reference-type="ref" reference="fig:example_pairs"} for examples). Affine correspondences are detected by ASIFT and the same procedure is applied as for the synthesized tests. Fig. [\[fig:real_focals\]](#fig:real_focals){reference-type="ref" reference="fig:real_focals"} reports the histogram of the relative errors (in percentage) in the focal length estimates on all the \(104\) pairs. It can be seen that in most of the cases the obtained results are accurate, the relative error is close to zero. Fig. [\[fig:real_image\]](#fig:real_image){reference-type="ref" reference="fig:real_image"} shows the first image of an example pair and the point correspondences. In Table [1](#tab:comparison1){reference-type="ref" reference="tab:comparison1"}, the proposed method is compared with the 6-point algorithm and the one creating point correspondences from two local affinities . The reported relative errors are computed as the ratio of the estimation error and the ground truth focal length as \(|f_{est}-f_{gt}|/f_{gt}\). It can be seen that the 2-point technique outperforms the other ones in terms of both mean and median accuracy and spread.
[\[tab:comparison1\]]{#tab:comparison1 label="tab:comparison1"}
**5.3. Time Demand** Augmenting RANSAC or other robust statistics with the proposed method significantly reduces the processing time. Table [2](#tab:ransac_iterations){reference-type="ref" reference="tab:ransac_iterations"} reports the required iteration number of RANSAC to converge using different minimal methods (columns) as engine. Rows show the ratio of the outliers.
[\[tab:ransac_iterations\]]{#tab:ransac_iterations label="tab:ransac_iterations"}
# Conclusion
A theory and an efficient method is proposed to estimate the unknown focal-length and the fundamental matrix using only two affine correspondences. The 2-point method is validated on both synthesized and real world data. Compared with the state-of-the-art methods, it obtained the most accurate focal lengths with fundamental matrices having similar quality as the recent algorithms. Combining the minimal solver with a robust statistics, e.g. RANSAC, allows significant reduction in computation. Particularly, its time demand is around a few milliseconds, thus it is much faster than affine-covariant detectors providing the input. The proposed algorithm can also be applied in reconstruction or multi-view pipelines, e.g. that of Bujnak et al. , if at least two images of the same camera with fixed focal length are available.
[^1]: <http://web.eee.sztaki.hu/~dbarath/>
[^2]: Test data are provided as supplemental material.
[^3]: <http://www2c.airnet.ne.jp/kawa/photo/ste-idxe.htm> | {'timestamp': '2017-06-07T02:04:27', 'yymm': '1706', 'arxiv_id': '1706.01649', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01649'} |
null | null |
null | null |
# Introduction
When designing machine learning formulations, choice of penalty plays a key role in the accuracy of the inferred model, and robustness of the learning procedure. Consider Figure [\[1D example\]](#1D example){reference-type="ref" reference="1D example"}, where data from a simple linear regression has been contaminated with asymmetric outliers. The data generating mechanism is shown in black. The linear regression model for the data \(\{y_i, a_i\}\) is given by \[\label{eq:LG} y = \ip{a_i,x} + \epsilon_i,\] with \(\epsilon_i\) assumed i.i.d. Gaussian variables. The maximum likelihood formulation is equivalent to the least square problem, \[\min_x \frac{1}{2}\|Ax-y\|^2.\] This assumption is violated in Figure [\[1D example\]](#1D example){reference-type="ref" reference="1D example"}; the data are corrupted with asymmetric errors, and contain outliers. The least squares fit, shown in blue dash, fails to detect the true data generating mechanism. To learn effectively in these cases, we consider a parameterized family of penalties \(\rho(x;\theta)\), where \(x\) are model parameters and \(\theta\) control the shape of the penalty. The family is rich enough to allow the kinds of errors in Figure [\[1D example\]](#1D example){reference-type="ref" reference="1D example"}, and we learn \(x\) and \(\theta\) simultaneously using an extended statistical model. Two immediate examples of \(\theta\) are the robustness threshold \(\kappa\) in the Huber penalty, and the slope \(\tau\) in the asymmetric quantile penalty, see Figure [\[fig:Qhub\]](#fig:Qhub){reference-type="ref" reference="fig:Qhub"}. Selecting the appropriate \(\theta\) is important. The quantile Huber case was considered by.
For example, the fit with the correctly set quantile penalty is shown in red dash in Figure [\[1D example\]](#1D example){reference-type="ref" reference="1D example"}. The value of \(\tau\) was obtained automatically from the data using a statistical model detailed in Section [2](#sec:model){reference-type="ref" reference="sec:model"}, and did not require cross-validation. The main focus of this paper is data-drive approaches for simultaneously selecting \(\theta\) and solving for \(x\), without cross-validation or prior/additional information.
## Related work
Meta-parameters are classically estimated using cross-validation or grid search. These methods typically require multiple solutions of any given learning problem, where a held-out dataset is used to evaluate each configuration. More recently, Bayesian optimization and random search have come to the forefront as two leading techniques that can be used to obtain meta-parameters in a very wide range of contexts. All of these techniques can also be used for the problem class we consider. However, applying these approaches is always more computationally expensive than solving a single instance of a learning problem; both random search and Bayesian optimization require many instance evaluations. In contrast, for the narrower context of shape parameter estimation, we solve a *single* extended problem to simultaneously fit the \(x\) and \(\theta\). The most relevant works related to this paper focus on the relation between the quantile penalty and the asymmetric Laplace distribution(ALD) . jointly estimate the model and the shape parameters for quantile penalty, and infer the joint posterior distribution of these parameters.
## Roadmap
The paper proceeds as follows. In Section [2](#sec:model){reference-type="ref" reference="sec:model"} we derive the maximum likelihood model for joint inference in \(x\) and \(\theta\) and characterize theoretical properties of the resulting objectives from an optimization perspective. In Section [3](#sec:algo){reference-type="ref" reference="sec:algo"}, we consider first-and second-order algorithms for the structured but generally nonconvex and nonsmooth objective [\[eq:mainFormulation\]](#eq:mainFormulation){reference-type="eqref" reference="eq:mainFormulation"}. Section [4](#sec:synthetic){reference-type="ref" reference="sec:synthetic"} illustrates the convergence rates of the methods, as well as behavior of the shape-tuned estimates, using synthetic data. In Section [5](#sec:real){reference-type="ref" reference="sec:real"}, we develop self-tuning RPCA approaches, and apply them to real data.
# Statistical Model and Properties of Joint Objective {#sec:model}
Penalties in learning formulations have underlying statistical assumptions. In this section we first review the relationship between penalties and corresponding residual distributions. We then use this relationship to develop a joint maximum likelihood approach for model and shape parameter inference, and characterize properties of the resulting objective function.
## Statistical view
Recall the quantile penalty in Figure [\[fig:Qhub\]](#fig:Qhub){reference-type="ref" reference="fig:Qhub"}. If we choose \(\tau\) to be close to 1, we penalize the negative errors a lot more than the positive. Equivalently, we assume that the distribution of the errors \(\epsilon_i\) is biased towards positive errors. The relationship between penalties and associated densities can be made precise. Given a penalty \(\rho(r; \theta)\), we assume \(\epsilon_i\) in [\[eq:LG\]](#eq:LG){reference-type="eqref" reference="eq:LG"} are i.i.d. samples from the distribution with density \[\label{eq:penalty-density} p(r;\theta) = \frac{1}{n_c(\theta)}\exp[-\rho(r;\theta)],\quad \text{where} \quad n_c(\theta) = \int_\mathbb{R} \exp[-\rho(r;\theta)]\,dr.\] The term \(n_c(\theta)\) is a normalization constant that ensures that \(\rho(r,\theta)\) can be interpreted as a density as in [\[eq:penalty-density\]](#eq:penalty-density){reference-type="eqref" reference="eq:penalty-density"}. We can now formulate the *joint* maximum likelihood problem in \((x,\theta)\), or equivalently minimize its negative log: \[\label{eq:obj} \min_{x,\theta\in\mathcal{D}} \sum_{i=1}^m \rho(y_i-\ip{a_i,x};\theta) + g(x) + m\log[n_c(\theta)].\] The parameter \(\theta\) may be restricted to a domain \(\mathcal{D}\); for example, the slope parameter \(\tau\) must be between \(0\) and \(1\) (see Figure [\[fig:Qhub\]](#fig:Qhub){reference-type="ref" reference="fig:Qhub"}). The term \(g(x)\) is an optional regularization function, e.g. \(\lambda \|x\|_1\) or indicator of \(x \in \mathcal{C}\) for some set \(\mathcal{C}\). The objective in the quantile example used to obtain the penalty-tuned fit in Figure [\[1D example\]](#1D example){reference-type="ref" reference="1D example"} is given by \[\label{eq:quantile} \min_{x, \tau \in [0,1]} q_{\tau} (Ax-b) + m\log\left(\frac{1}{\tau} + \frac{1}{1-\tau}\right),\] with \(q_\tau\) the asymmetric 1-norm, and \(m\) the length of the residual. In this special case, \(\log(n_c)\) is available in closed form, is smooth in the interior of its domain, and acts as a barrier function for the interval \([0,1]\) that favors \(\tau = 0.5\). It's also a strongly convex function, but has *no global quadratic upper bound*, violating a key assumption often required by optimization algorithms. In the remainder of this section, we characterize theoretical properties of the objective [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}.
## Theoretical properties
Smoothness, convexity, and quadratic upper bounds are at the center of algorithm design, and understanding these properties guide the choice of algorithms for [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}.
Under these assumptions, we can obtain formulas for the first and second derivatives of \(n_c(\theta)\).
The proof is straightforward, and included in the supplementary materials. The derivative formulas [\[eq:nc_form\]](#eq:nc_form){reference-type="eqref" reference="eq:nc_form"} are used for first-and second-order methods to infer \(x\) and \(\theta\). The parametrization conditions in \(\theta\) are satisfied by all commonly used piecewise linear quadratic (PLQ) examples , including Huber and quantile penalties in Figure [\[fig:Qhub\]](#fig:Qhub){reference-type="ref" reference="fig:Qhub"}. The theorem applies more generally to densities that are not log-concave. For example, the Student's \(t\) density and associated penalty satisfy Assumption [\[asp:smoothness\]](#asp:smoothness){reference-type="ref" reference="asp:smoothness"} and other assumptions of Theorem [\[th:smoothness\]](#th:smoothness){reference-type="ref" reference="th:smoothness"} for \(\nu > 1\). In the quantile case [\[eq:quantile\]](#eq:quantile){reference-type="eqref" reference="eq:quantile"}, the term \(\log[n_c(\theta)]\) is convex. We characterize sufficient conditions for convexity of \(\log[n_c(\theta)]\) for a general class of penalties \(\rho\).
This result follows from. Theorems [\[th:smoothness\]](#th:smoothness){reference-type="ref" reference="th:smoothness"} and [\[th:convexity\]](#th:convexity){reference-type="ref" reference="th:convexity"} tell an interesting story. The log-normalization constant \(\log[n_c(\theta)]\) is nearly always smooth; even when the loss \(\rho\) is nonsmooth in \(x\). The inference problem [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"} is *never jointly convex* in \((x,\theta)\); in particular looking for a jointly convex formulation \(\rho(x;\theta)\) guarantees \(\log[n_c(\theta)]\) will be *concave*. This is intuitive, as we are attempting to learn both the model and error structure at the same time. Objective [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"} is, in general, nonsmooth and nonconvex; but it has a fairly simple structure that is exploited to design first and second order methods in the next section. To understand how non-convex [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"} is, we apply partial minimization, and consider the function \[\varrho(\theta) = \min_{x} \sum_{i=1}^m \rho(y_i-\ip{a_i,x};\theta) + m\log[n_c(\theta)].\] This is the *value function* of the shape parameters, after \(x\) has been minimized. For simple examples, \(\theta\) may have dimension 1 or 2, we can plot either the graph or the level sets of this function. We generate the samples \(\epsilon_i\) from distribution defined by quantile Huber function with \(\kappa = 1\) and \(\tau = 0.05\), and plot \(\varrho(\theta)\) in Figure [\[fig:levelset\]](#fig:levelset){reference-type="ref" reference="fig:levelset"}. From Figure [\[fig:levelset\]](#fig:levelset){reference-type="ref" reference="fig:levelset"} (a), we can see that for the quantile penalty, the value function \(\varrho\) appears to be quasi-convex for this example; and we can expect to find the unique global minimum in \(x\) and \(\tau\), since computing the projection require solving a convex problem. When \(\theta\) comprises both \(\tau\) and \(\kappa\) in (b), the joint objective is clearly noncovex and may be more challenging for a local search (the level sets are stretched and bent). Nonetheless, there is a unique global minimum that is close to the true parameters; and this minimum was found by a local search.
# First-and Second-Order Algorithms. {#sec:algo}
In this section, we consider first-and second-order methods for problems of type [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}. When \(\rho\) is smooth in \(x\) and \(\theta\), we show how to apply the Proximal Alternating Linearized Minimization (PALM) algorithm. The development is straightforward, but the log normalization constant \(\log[n_c(\theta)]\) must be treated carefully, as its gradient does not have a global Lipschitz constant. Requiring smoothness in \(\rho\) is restrictive; and in particular eliminates the quantile example [\[eq:quantile\]](#eq:quantile){reference-type="eqref" reference="eq:quantile"}. The quantile penalty is not smooth, but it is piecewise linear. Interior point methods have been shown to be effective for convex problems of moderate scale (thousands of variables and data points) where \(\rho\) and \(g\) are nonsmooth piecewise linear-quadratic penalties . Examples include symmetric penalties (1-norm, quadratic) as well as asymmetric penalties, such as quantile and quantile Huber . All of these penalties are shown in detail in the supplementary materials. Our main algorithmic contribution is to extend this approach to the joint nonconvex inference problem [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}.
## PALM for inference and shape estimation {#sec:PALM}
The PALM algorithm can be used to minimize any problem of form \[\min_{x,\theta} H(x,\theta) + r_1(x) + r_2(\theta),\] where \(H\) is \(C^1\), with globally Lipschitz partial gradients, while the functions \(r_1\) and \(r_2\) are required to be only proper lower semicontinuous (in particular not necessarily convex, finite valued, or smooth). Even though \(\log[n_c(\theta)]\) is smooth (see Theorem [\[th:smoothness\]](#th:smoothness){reference-type="ref" reference="th:smoothness"}), it must be relegated to \(r_2(\theta)\), since otherwise it can easily violate the Lipschitz assumptions on \(H\). Therefore, to apply PALM to [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}, we take \[\label{eq:palm_detail} H(x,\theta) = \sum_{i=1}^m \rho(y_i-\ip{a_i,x};\theta), \quad r_1(x) = g(x), \quad r_2(\theta) = \delta_\mathcal{D}(\theta) + m\log[n_c(\theta)].\] Here \(\delta_{\mathcal{D}}\) is the indicator function for the set \(\mathcal{D}\), ensuring \(\theta \in \mathcal{D}\), and \(g\) is any 'prox-friendly' regularizer for \(x\). The PALM algorithm is detailed in Algorithm [\[alg:PALM\]](#alg:PALM){reference-type="ref" reference="alg:PALM"}. The steps \(c_k\) and \(d_k\) are obtained from Lipschitz constants of the (partial) gradients of \(H\).
The proof appears in the appendix.
# Synthetic Data Experiments {#sec:synthetic}
We illustrate the proposed approach using a linear regression example. We consider a data set contaminated by asymmetric errors and outliers, two features captured by the quantile Huber penalty (Figure [\[fig:QuantileQHub\]](#fig:QuantileQHub){reference-type="ref" reference="fig:QuantileQHub"}). The slope \(\tau\) controls for the asymmetry, while the threshold \(\kappa\) detects the point beyond which a residual might be classified as an 'outlier'. The goal of the experiment is to simultaneously learn the regression model parameters \(x\) as well as obtain the correct \(\tau\) and \(\kappa\). Simple residual analysis is not possible *a priori*, since the model parameters \(x\) are also unknown. When \(\kappa > 0\) in quantile Huber, \(\rho(x;\theta)\) is smooth, and we can use the PALM algorithm from Section [3.1](#sec:PALM){reference-type="ref" reference="sec:PALM"}. The quantile penalty is also PLQ, so we can also apply the proposed IP method from Section [3.2](#sec:IP){reference-type="ref" reference="sec:IP"}. We use both and compare their performance. The primal form for the quantile Huber penalty is \[\label{eq:quantileHuberPrimal} \small \rho(r;\theta) = \begin{cases}-\tau\kappa r-\frac{(\tau\kappa)^2}{2},& r <-\tau\kappa\\ \frac{1}{2}r^2,& r\in[-\tau\kappa,(1-\tau)\kappa]\\ (1-\tau)\kappa r-\frac{((1-\tau)\kappa)^2}{2},& r > (1-\tau)\kappa. \end{cases}\] We must choose a parametrization in terms of \(\theta\). One option would be to take \(\theta = [\tau,\kappa]^{\mathsf{T}}\). But this parametrization would violate assumptions of both the first-and second-order approaches in Section [3](#sec:algo){reference-type="ref" reference="sec:algo"}. Indeed, \(\nabla_\theta \rho(r;\theta)\) would not have a global Lipschitz constant, so we could not use PALM. Similarly, we could not write the conjugate representation [\[eq:PLQprimaldual\]](#eq:PLQprimaldual){reference-type="eqref" reference="eq:PLQprimaldual"} using affine functions of \(\theta\). Looking carefully at either [\[eq:quantileHuberPrimal\]](#eq:quantileHuberPrimal){reference-type="eqref" reference="eq:quantileHuberPrimal"} or [\[eq:PLQprimaldual\]](#eq:PLQprimaldual){reference-type="eqref" reference="eq:PLQprimaldual"}, we instead choose \(\theta_1 = \tau\kappa, \theta_2 = \tau(1-\kappa)\). The only requirement on these parameters is that they are each non-negative. The primal objective can be written as \[\min_{x,\theta\geq 0} \rho(Ax-y;\theta) + m\log[n_c(\theta)].\] where \(A\in\mathbb{R}^{m\times n}\) is the design matrix, \(x\in\mathbb{R}^n\) is the model parameter vector, and \(y\in\mathbb{R}^m\) is the observed data vector contaminated by outliers. From Theorem [\[th:smoothness\]](#th:smoothness){reference-type="ref" reference="th:smoothness"}, \(n_c(\theta)\) is \(\mathcal{C}^2\) smooth. From Theorem [\[th:convexity\]](#th:convexity){reference-type="ref" reference="th:convexity"}, the objective in \(\theta\) is the sum of a concave term \(\rho(Ax-y;\theta)\) and a convex term \(m\log(n_c(x))\). The joint problem in \((x,\theta)\) is nonconvex. Nonetheless, both first-and second-order methods from Section [3](#sec:algo){reference-type="ref" reference="sec:algo"} can be applied to solve the problem. We generate synthetic data with \(m = 1000\), \(n = 50\), and generate the elements of \(A\in \mathbb{R}^{m\times n}\) from a standard Gaussian random distribution. The measurement errors are sampled from quantile Huber distributions, to verify that the approach is able to recover 'ground truth' values for \((\tau, \kappa)\) parameters. We denote ground truth parameters as \(x_t\), \(\tau_t\), \(\kappa_t\), while \(x^*\), \(\tau^*\), \(\kappa^*\) are the solutions obtained by solving [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"}. We provide two reference solutions: \(x_{LS}\) is the least square solution, and \(x_M\) is the solution obtained by solving \(\|Ax-b\|_1\). For each \(\kappa\) and \(\tau\) setting, we run the simulation 10 times, and show the average of the results in Table [1](#tb:exp){reference-type="ref" reference="tb:exp"}. Results shown are obtained by the IP method.
In order to get the result in Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"} (b), we introduce a variance parameter \(\sigma\) for the Huber penalty to automatically estimate the right scale of the residual. The joint \((\kappa, \sigma)\) parametrization is given by \[\rho(r;[\kappa,\sigma]) = \begin{cases} \kappa|r|/\sigma-\kappa^2/2, & |r|>\kappa\sigma\\ r^2/(2\sigma^2), & |r|\le\kappa\sigma, \end{cases}\] with the resulting self-tuning RPCA formulation (solved by Algorithm [\[alg:PALM\]](#alg:PALM){reference-type="ref" reference="alg:PALM"}): \[\min_{U,V,\kappa>0,\sigma>0} \sum_{i,j}\rho(\ip{U_i,V_j}-Y_{i,j};[\kappa,\sigma]) + mn\log[n_c([\kappa,\sigma])].\] The result is shown in Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"}(b). As the optimization proceeds, \(\kappa, \sigma \rightarrow0^+\) with a fixed ratio \(\alpha = \kappa/\sigma\). The self-tuning Huber becomes the scaled 1-norm, recovering the original RPCA formulation . The result in Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"}(b) is an improvement over the result with \((\kappa, \sigma)\) fixed at the initial values in Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"}(a). The weakness of the Huber is that the \(\kappa\) has to work well for residuals near the origin as well as in the tail. The self-tuning approach makes it easy to create and adapt new penalties. To get additional flexibility, we introduce an inflection point; letting the 'slope' near the origin be different from slopes in the tails. The Huberized Student's t penalty is shown in Figure [\[fig:Tiber\]](#fig:Tiber){reference-type="ref" reference="fig:Tiber"}, and detailed below: \[\label{eq:tiber} \hspace{-.4cm} \rho(r;[\kappa,\sigma]) = \begin{cases} \frac{2\kappa}{\sigma(\kappa^2+1)}(|r|-\kappa\sigma) + \log(1+\kappa^2), & |r| > \kappa\sigma\\ \log(1+r^2/\sigma^2),&|r|\le\kappa\sigma \end{cases}\] Penalty [\[eq:tiber\]](#eq:tiber){reference-type="eqref" reference="eq:tiber"} is flexible, in the sense that behavior for large residuals and small ones is decoupled. When we self-tune this new penalty, the additional flexibility indeed improves on the self-tuned Huber, recovering the results shown in Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"}(d). It is clear the self-tuning approach succeeds, as the Huberized Student's \(t\) result at the initial \(\kappa, \sigma\) values is useless (Figure [\[RPCA\]](#RPCA){reference-type="ref" reference="RPCA"}(c)). In this data set, the advantage of self-tuned Huberized Student's \(t\) over self-tuned Huber may not be that obvious. We also apply our approach to Escalator data set[^3]. This data set is is less noisy than the first, but contains multiple moving objects. We select a time window with \(200\) pictures, and \(Y\in \mathbb{R}^{20800\times200}\), and apply the huber and Huberized Student's \(t\) to this data set, with results shown in Figure [\[fig:esc_result\]](#fig:esc_result){reference-type="ref" reference="fig:esc_result"}. We can see that there are a lot of artifacts in the self-tuned Huber penalty result, e.g., the escalator stairs and shadow of people. Instead the result for self-tuned Huberized Student's \(t\) is much cleaner and we isolate the moving people successfully. All of these results are achieved with fully automatic parameter tuning.
Here, we use self-tuning Huber and self-tuning Huberized Student's \(t\) penalties. Looking at Figure [\[fig:esc_result\]](#fig:esc_result){reference-type="ref" reference="fig:esc_result"}, the self-tuned Huber has stronger signal but includes more background; the picture looks better to the eye but a closer examination reveals parts of the escalator are present. The self-tuned Huberized Student's \(t\) has weaker \(S\) which is harder to see; but actually gives a much better result. To get more insights into the problem, we also look at the empirical distribution of the residual. In Figure [\[fig:dis\]](#fig:dis){reference-type="ref" reference="fig:dis"}, the light blue line denote the empirical CDF for the residual (\(R = Y-U^{\mathsf{T}} V\)), red dashed line is the the best Huberized Student's \(t\) fit, blue dashed line is the best \(\ell_2\) fit and green dashed line is the best \(\ell_1\) fit. We can see that Huberized Student's \(t\) has a perfect fit, and in particular is much better that \(\ell_1\) and \(\ell_2\).
# Conclusions
We introduced a method for self-tuning error penalties, incorporates the log of the normalization constant (a function of the shape parameters) into an extended joint inference problem. Results with synthetic data as well as real data show promise. Future research includes designing innovative 'flexible' penalties and automatically adapting them to different applications.
[^1]: If \(A\) has dimensions \(m\) and \(n\), interior point methods require \(O(n(m^2+n^2))\) arithmetic operations, where \(n\) is the smaller dimension. This limits practical applications for large-scale problems; to go beyond \(2000 \times 2000\) with modest compute, some sort of special structure or technique (sparsity, preconditioning) is typically needed.
[^2]: Publicly available at <http://vis-www.cs.umass.edu/~narayana/castanza/I2Rdataset/>
[^3]: Publicly available at <http://vis-www.cs.umass.edu/~narayana/castanza/I2Rdataset/> | {'timestamp': '2017-06-07T02:08:52', 'yymm': '1706', 'arxiv_id': '1706.01865', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01865'} |
# Introduction
Given an unweighted 2-edge connected graph \(G=(V, E)\), the *minimum size 2-edge connected spanning subgraph problem* (henceforth *2EC*) consists of finding a 2-edge connected spanning subgraph of \(G\) with the minimum number of edges. Multiple copies of an edge are not allowed, and are not necessary. The problem has important applications in the design of networks that can survive the failure of a link, but it is MAX SNP-hard even for special cases. Letting \(x_{e}\) be a binary variable with value 1 when edge \(e\in E\) is in a 2EC solution, 2EC can be formulated as an integer program as follows:\[\label{ilpTwoEC} \begin{aligned} \text{Minimize}\quad & \sum_{e\in E}x_e& \\ \text{Subject to}\quad &\sum(x_{ij}:i\in S, j\notin S) \geq 2 &\text{for all } \emptyset \subset S \subset V,\\ & 0 \leq x_e \leq 1\text{, and integer} &\text{for all } e \in E. \\ \end{aligned}\] The linear programming (LP) relaxation of 2EC, denoted by 2EC\(^\text{LP}\), is obtained by relaxing the integer requirement in ([\[ilpTwoEC\]](#ilpTwoEC){reference-type="ref" reference="ilpTwoEC"}). Let OPT\((G)\) and OPT\(_{\text{LP}}(G)\) denote the optimal value of 2EC for \(G\) and the value of its LP relaxation, respectively. In order to measure the quality of the LP relaxation as a lower bound, we define the *unit integrality gap* \(\alpha\)`<!-- -->`{=html}2EC as the supremum of the ratio of OPT\((G)\) over OPT\(_{\text{LP}}(G)\) over all input graphs \(G\). This gives a measure of the quality of the lower bound provided by 2EC\(^\text{LP}\), which is crucial for approximation algorithms and methods like branch and bound. Note that a polynomial-time constructive proof that the unit integrality gap equals \(k\in\mathbb{R}\) would provide a \(k\)-approximation algorithm for 2EC. While \(\alpha\)`<!-- -->`{=html}2EC has been intensely studied, little is known about it other than that \(\frac{9}{8}\leq\alpha\text{2EC}\leq\frac{4}{3}\). A natural next step is to focus on the simplest form of the problem known to remain NP-hard, when 2EC is restricted to graphs that are *cubic* (i.e., every vertex has degree three) and 3-edge connected; we thus hope to generalizing successful methods. In this paper, we demonstrate an improved upper bound for \(\alpha\)`<!-- -->`{=html}2EC when restricted to 3-edge connected cubic graphs.
# Notation and Background {#notation}
This section is dedicated to basic notions of graph theory, which will be used throughout the paper. Let \(G=(V, E)\) be a simple undirected graph with vertex set \(V\), edge set \(E\) and \(n=|V|\). A graph is *subcubic* if all vertices have degree 3 or less. A *subgraph* of \(G\) is a graph \(H=(V_{H}, E_{H})\) where \(V_{H}\subseteq V\) and \(E_{H}\subseteq E\), such that if edge \(uv\in E_{H}\), then \(u, v\in V_{H}\). Subgraph \(H\) is said to be *spanning* if \(V_{H}=V\). For any subgraph \(H\) of \(G\) we sometimes use the notation \(E(H)\) and \(V (H)\) to denote the edge set and the vertex set for \(H\), respectively, and we use \(\chi^{E(H)} \in\mathbb{R}^{E}\) to denote the incidence vector of subgraph \(H\) (i.e. \(\chi^{E(H)}_{e}\) is the number of copies of edge \(e\) in \(H\)). A *component* of \(G\) is a subgraph of \(G\) for which any two vertices are connected to each other by paths, and which is not connected to any other vertex in \(G\). For any subset \(A\subseteq V\), the *complement* of \(A\) is \(\overline{A}=V\setminus A\). For any two non-overlapping subsets \(A, B\subseteq V\), the edge set between \(A\) and \(B\) is denoted by \(E[A: B]\), i.e. \(E[A: B] = \{ij\in E\ |\ i\in A, j\in B\}\). Note that the subset \(F=\delta(A)\) of \(E\) for some \(A\subset V\) is called a *cut* \(\delta(A)=E[A:V\setminus A]\) as it disconnects \(G\). Graph \(G\) is considered \(k\)-edge connected if and only if all cuts of \(G\) have size greater or equal to \(k\), \(k\in\mathbb{N}\). An edge cut that contains \(k\) edges is a *\(k\)-edge cut*. An edge cut \(F\) of \(G\) is *essential* if \(G'=(V, E\setminus F)\) has at least two components each containing more than one vertex. Each of those components is called a *shore* of the cut. For a \(k\in\mathbb{N}\), \(G\) is said to be *essentially \(k\)-edge connected* if and only if \(G\) does not have an essential edge cut \(F\) with \(|F|<k\). If \(G\) is cubic, simple and essentially 4-edge connected, then it is also 3-edge connected. The function \(|\delta(\cdot)|\) is *symmetric submodular* for \(G\), i.e. for every two sets \(Y, Z\subseteq V\), the following two properties hold: \[\begin{aligned}
|\delta(Y)|+|\delta(Z)|\geq |\delta(Y\cup Z)| + |\delta(Y\cap Z)|\text{,}\\ |\delta(Y)|+|\delta(Z)|\geq |\delta(Y\setminus Z)| + |\delta(Z\setminus Y)|\text{.} \end{aligned}\] The concept of convex combination is used extensively: in the context of this paper, we say that a vector \(y\in\mathbb{R}^{E}\) is a *2EC convex combination* if there exist 2-edge connected spanning subgraphs \(H_{i}\) with multipliers \(\lambda_{i}\in{\mathbb R}_{\geq 0}, i=1, 2, \dotsc, j\) such that \(y=\sum_{i=1}^{j}\lambda_{i}\chi^{E(H_{i})}\) and \(\sum_{i=1}^{j}\lambda_{i} = 1\).
# Contribution
Our main result, which is proven in Section [6](#sectionProof){reference-type="ref" reference="sectionProof"}, is the following.
In other words, we show that the unit integrality gap for 2EC is bounded above by \(\frac{7}{6}\) for 3-edge connected cubic graphs, which improves upon Boyd, Iwata and Takazawa's upper bound of \(\frac{6}{5}\)Ṡince 2EC restricted to 3-edge connected cubic graph is the simplest form of the problem known to remain NP-hard, it also implies that successful methods can be generalized. Our methods are not polynomial and thus, do not result in an approximation algorithm. Nevertheless, they give hope that a \(\frac{7}{6}\)-approximation algorithm exists, which would improve on the existing methods which gives a \(\frac{6}{5}\)-approximation. It would also extend Takazawa's \(\frac{7}{6}\)-approximation for bipartite cubic 3-edge connected graphs to all cubic 3-edge connected graphs. In order to prove Corollary [\[mainCorollary\]](#mainCorollary){reference-type="ref" reference="mainCorollary"}, we prove the following stronger statement which uses the concept of convex combination.
In Section [6](#sectionProof){reference-type="ref" reference="sectionProof"}, we also prove the following intermediary result for essentially 4-edge connected cubic graphs.
Besides being useful in our proof for the main result, Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"} is also of independent interest because it is a new reduction operation that is very specific to essentially 4-edge connected cubic simple graphs.
# Literature Review on 2EC
General case-Khuller + Vishkin, 3/2-Cheriyan, Sebo, Szigeti, 17/12-Vempala, Vetta, 4/3-Sebo, Vygen, 4/3-Krysta, Kuma, 4/3-1/1344-Jothi, Raghavachari, Varadarajan, 5/4-Explain why not Frederickson, Jaja For special cases-subcubic bridgeless -Csaba, Karpinski, Krysta, 5/4 + epsilon -Boyd, Fu, Sun, 5/4 -cubic, 3-edge connected -Boyd, Iwata, Takazawa, 6/5 -Huh, 5/4 -bipartite: Takazawa: 7/6 Bounds-Sebo, Vygen, 9/8 \<= alpha2EC \<= 4/3 in general Boyd, Iwata,, Takazawa, alpha2EC \<= 6/5 for cubic 3-edge connected Yu Sun showed that for cubic bridgeless graphs, you can't do better than 7/6n (but that isn't a bound on the integrality gap). For subcubic, the integrality gap is at least 8/7. For cubic 3-edge connected, as far as I know, the lower bound stands at 11/10 (the Petersen graph).
# Literature Review
The first to improve the 2-approximation ratio for 2EC were Khuller and Vishkin, in 1994, with their \(\frac{3}{2}\)-approximation algorithm obtained using depth-first search trees and a method called "tree-carving". This was later refined in 1998 to \(\frac{17}{12}\) by Cheriyan, Sebő and Szigeti via ear decompositions, then to \(\frac{4}{3}\) by Sebő and Vygen, in 2014. Krysta and Kumar improved the ratio to \(\frac{4}{3}-\epsilon\) based on a charging scheme. Other results have been claimed, but have either been shown to be false, or are without conclusive proofs. In parallel, much development also occurred for special cases, especially cubic and subcubic graphs. Krysta and Kumar designed a \(\frac{21}{16}\)-approximation algorithm for cubic graphs. For subcubic graphs, in 2002, Csaba, Karpinski and Krysta attained the approximation ratio of \(\frac{5}{4} + \epsilon\), which was then improved to \(\frac{5}{4}\) by Boyd, Fu and Sun in 2014, using circulations. When 2EC is further restricted to 3-edge connected cubic graphs, we denote the work of Huh, and of Boyd, Iwata and Takazawa, who achieved ratios of \(\frac{5}{4}\), and \(\frac{6}{5}\), respectively. Very recently, Takazawa obtained a \(\frac{7}{6}\)-approximation algorithm for bipartite cubic 3-edge connected graphs, inspired from the work of Boyd, Iwata and Takazawa, by using 2-factors covering specified cuts. In the general case, Sebő and Vygen proved that the unit integrality gap \(\alpha\)`<!-- -->`{=html}2EC is bounded as follows: \(\frac{9}{8}\leq\alpha\text{2EC}\leq\frac{4}{3}\). The integrality gap 2EC for cubic 3-edge connected graphs was shown in 2013 by Boyd, Iwata and Takazawa to be bounded above by \(\frac{6}{5}\), and it is known that they are bounded below by the Petersen graph, at \(\frac{11}{10}\), as shown in Figure [\[Petersen\]](#Petersen){reference-type="ref" reference="Petersen"} (where bold lines represent edges in the subgraph and dotted edges stand for edges in the Petersen graph, but not in the subgraph). -Define ILP and LP.-Define integrality gap -Importance: indicator of the quality of the lower bound given by the LP relaxation (important for branch and bound and approximations that depend on lower bounds for success)-Conjecture -Where is that gap today? Define 2EC -finding a minimum size 2-edge connected spanning subgraph in a given 2-edge connected graph Why is 2EC important -survivable network design, train network -NP hard even for special cases (e.g. regular bipartite graphs)-Why do we care if it's not an approximation algorithm? Khuller and Vishkin gave a \(\frac{3}{2}\)-approximation for this problem using depth-first search trees and a notion called "tree carving" followed by a 17/12-approximation algorithm due to Cheriyan, Sebő and Szigeti \[6\]. They use ear decomposition. Garg et al. \[3\] claimed to have a 5/4-approximation algorithm for the minimum size 2-ECSS problem, but no proof was given. Improvements in the approximation ratio in several graph classes are made. For the minimum 2-edge connected spanning subgraph problem in 3-edge connected cubic graphs, Huh \[16\] gave a 5/4-approximation algorithm. A further improvement is given by Boyd, Iwata and Takazawa \[5\], who designed two algorithms with approximation ratio 6=5. Kenjiro came up with a 7/6 approximation scheme on bipartite graphs, using two-factors. Based on Boy'd method and exploiting the bipartiteness of graphs. Lower bounds: Yu Sun showed that for cubic bridgeless graphs, you can't do better than 7/6n (but that isn't a bound on the integrality gap). For subcubic, the integrality gap is at least 8/7. For cubic 3-edge connected, as far as I know, the lower bound stands at 11/10 (the Petersen graph).
# 2EC Convex Combinations for 3-edge Connected Graphs {#sectionProof}
In this section, we provide a proof for Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"}, i.e. we show that for any essentially 4-edge connected cubic graph \(G = (V, E)\) with \(|V|>6\) and edges \(au, uv, vc, vd \in E\), no two subsets \(S, S'\subset V\) different than \(\{u, v\}\) exist such that \(au, vd \in \delta(S)\), \(au, vc \in \delta(S')\), and \(|\delta(S)|=|\delta(S')|=4\). This result will be used to prove the main theorem.
## Proof of Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"}. {#proof-of-lemma-submodularlemma. .unnumbered}
Here we provide a proof of Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"}. Suppose that \(G = (V, E)\) with \(|V|>6\), edges \(au, uv, vc, vd \in E\) and subsets \(S, S'\subset V\) is the smallest counter-example to the lemma. Let edges incident to \(u\) and \(v\) be as illustrated in Figure [\[UvAndIncidentEdges\]](#UvAndIncidentEdges){reference-type="ref" reference="UvAndIncidentEdges"}, with the unlabelled vertex neighbour to \(u\) called \(b\). Some edges incident to the unlabelled vertices are not shown. \(G\) is simple and essentially 4-edge connected, and \(|V|>6\), which implies that labelled vertices in Figure [\[UvAndIncidentEdges\]](#UvAndIncidentEdges){reference-type="ref" reference="UvAndIncidentEdges"} are distinct, as there would otherwise exist an essential 3-edge cut. Without loss of generality, assume that \(u \in S\) and \(u \in S'\) (if not, take the complement of the set). The cuts \(\delta(S)\) and \(\delta(S')\) are minimal and \(G\) is cubic, which implies that two adjacent edges may not be in the same cut. Therefore \(v, b, c \in S\) and \(v, b, d \in S'\). We already know that \(a, d \notin S\) and \(a, c \notin S'\). Figure [\[CutsDisposition\]](#CutsDisposition){reference-type="ref" reference="CutsDisposition"} shows this disposition of the vertices in the cut. Throughout this proof, we exploit the symmetric submodularity property of the function \(|δ\delta(\cdot)|\), which states that \[\begin{aligned}
|\delta(S)|+|\delta(S')|\geq |\delta(S\cup S')| + |\delta(S\cap S')|\text{,}\nonumber\\ 8\geq |\delta(S\cup S')| + |\delta(S\cap S')|\text{.}\label{eqn1} \end{aligned}\] Because \(u, v \in S \cap S'\), then \(|S\cap S'| > 1\); similarly because \(a \notin S\cup S'\), then \(\overline{|S \cup S'|} > 0\). It follows that \(|\delta(S\cup S')| \geq 3\) and \(|\delta(S\cap S')| \geq 4\) (since \(G\) has no essential 3-edge cut). There exist two cases, when \(|\delta(S\cup S')| = 3\) and \(|\delta(S\cap S')|=4\) or \(5\) and where \(|\delta(S \cup S')| = 4\), because the cardinality of the cuts are restricted by ([\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}).
Graph \(G\) is simple and essentially 4-edge connected, which implies that one of the shores of \(\delta(S\cup S')\) must consist of a single vertex, as the cardinality of this cut is 3: vertices \(u\) and \(v\) belong to \(S\cup S'\), so \(\overline{S\cup S'}=\{a\}\). Any partition of a cubic graph has an even number of odd parts, because cubic graphs have an even number of vertices. We use this parity to show that \(|\delta(S\cap S')| = 5\). The partition \(\{S \setminus S', S' \setminus S, S \cap S', \overline{S \cup S'}\}\) (displayed in Figure [\[Partition\]](#Partition){reference-type="ref" reference="Partition"}) has \(|\overline{S \cup S'}|\) odd, which implies that at least one of \(S\setminus S'\), \(S'\setminus S\) and \(S\cap S'\) also has an odd number of vertices. The graph \(G\) is cubic, which implies that for any subset \(S\subseteq V\) of vertices, \(\delta(S)\) is odd if and only if \(|S|\) is odd. Given that \(|\delta(S)| = |\delta(S')| = 4\), then \(|S|\) and \(|S'|\) must be even. This implies that the parity of \(|S' \setminus S|\) and \(|S \cap S'|\) must be the same; similarly, the parity of \(|S \setminus S'|\) and \(|S \cap S'|\) must also be the same. Therefore, either \(|S \setminus S'|\), \(|S' \setminus S|\) and \(|S \cap S'|\) are all odd, or all even, to respect parity. But from the above, they cannot all be even. Thus, \(|S \cap S'|\) is odd, which means that \(|\delta(S \cap S')|\) must be odd as well. Therefore, \[|\delta(S \cap S')| = 5\text{.}\label{SandSEquals5}\] All edges at \(u\) and \(v\) are known: let us examine an edge \(bx\), for a vertex \(x \in V\). We already know of 3 edges in \(\delta(S \cap S')\), which means that there are exactly 2 others edges \(e\) and \(f\) in the cut. Assume that \(x \in S \cap S'\): this means that \(bu\), \(e\) and \(f\) are in an essential 3-edge cut, because such a cut contains \(b\) and \(x\). Such a cut contradicts the fact that \(G\) is essentially 4-edge connected. Therefore, \(x \notin S \cap S'\), which immediately implies that \(|S \cap S'| = 3\). We have that \[\begin{aligned}
4=|\delta(S)| = |\delta(S\setminus S')| + |\delta(S\cap S')|-2|E[S\setminus S': S\cap S'])|\text{,}\label{eqn5}\\ 4=|\delta(S')| = |\delta(S'\setminus S)| + |\delta(S\cap S')|-2|E[S'\setminus S: S\cap S']|\text{.}\label{eqn6} \end{aligned}\] We also have that \(ba \notin E\) or else there would be an essential 3-edge cut \(\delta(\{u, b, a\})\), which means that from the five edges in \(\delta(S \cap S')\), four link to either \(S \setminus S'\) or \(S' \setminus S\): \[\begin{aligned}
4 = |E[S'\setminus S: S\cap S']| + |E[S\setminus S': S\cap S']|\text{.}\label{eqn7} \end{aligned}\] From the premise \(|V | > 6\), it is implied that one of \(S \setminus S'\) and \(S' \setminus S\) contains more than one vertex. Without loss of generality, let it be \(S \setminus S'\). Because \(G\) has no essential 3-edge cut and \(|S \setminus S'|>1\), we infer that. \[\begin{aligned}
4 \leq |\delta(S \setminus S')| \text{,}\label{sNewBound}\\ 3 \leq |\delta(S' \setminus S)|\text{.}\label{sPrimeNewBound} \end{aligned}\] The symmetric submodularity property ensures that \[\begin{aligned}
|\delta(S)|+|\delta(S')|&\geq |\delta(S\setminus S')| + |\delta(S'\setminus S)|\nonumber\\ 8&\geq |\delta(S\setminus S')| + |\delta(S'\setminus S)|\text{.} \label{symMod2} \end{aligned}\] We will now use algebraic manipulations to show a contradiction. We subtract ([\[eqn6\]](#eqn6){reference-type="ref" reference="eqn6"}) from ([\[eqn5\]](#eqn5){reference-type="ref" reference="eqn5"}) and add ([\[eqn7\]](#eqn7){reference-type="ref" reference="eqn7"}) twice: \[\begin{aligned}
8 = |\delta(S\setminus S')|-|\delta(S'\setminus S)| + 4|E[S'\setminus S: S\cap S']|\text{.}\label{eqn8} \end{aligned}\] In equation ([\[eqn8\]](#eqn8){reference-type="ref" reference="eqn8"}), \(|\delta(S\setminus S')|-|\delta(S'\setminus S)|\) must be a multiple of 4. Because inequalities ([\[sNewBound\]](#sNewBound){reference-type="ref" reference="sNewBound"}), ([\[sPrimeNewBound\]](#sPrimeNewBound){reference-type="ref" reference="sPrimeNewBound"}) and ([\[symMod2\]](#symMod2){reference-type="ref" reference="symMod2"}) restrict the values of \(|\delta(S\setminus S')|\) and \(|\delta(S'\setminus S)|\), their difference must be zero, i.e. \[\begin{aligned}
|\delta(S\setminus S')|=4\text{.}\label{diffZero} \end{aligned}\] Equation ([\[eqn8\]](#eqn8){reference-type="ref" reference="eqn8"}) is simplified to \[\begin{aligned}
2 = |E[S\setminus S': S\cap S']|\text{.}\label{eqn9} \end{aligned}\] We now conclude by substituting ([\[SandSEquals5\]](#SandSEquals5){reference-type="ref" reference="SandSEquals5"}), ([\[diffZero\]](#diffZero){reference-type="ref" reference="diffZero"}) and ([\[eqn9\]](#eqn9){reference-type="ref" reference="eqn9"}) in ([\[eqn5\]](#eqn5){reference-type="ref" reference="eqn5"}): \[\begin{aligned}
& 4=|\delta(S\setminus S')| + |\delta(S\cup S')|-2|E[S\setminus S': S\cap S']|\nonumber\\ & \quad = 4 + 5-2 \times 2\nonumber\\ & \quad = 5\nonumber\text{,} \end{aligned}\] which gives a contradiction.
Since \(|\delta(S\cap S')|=4\), it follows that \(|S\cap S'|\) is even, because \(G\) is cubic. So there is at least one more vertex \(w\) in \(S\cap S'\). Since \(G\) is 3-edge connected, at least one edge incident with \(w\) is in \(\delta(S\cap S')\). However, \(vc, vd, au\in\delta(S\cap S')\) already, so we only have one other edge \(e\) in the cut (as shown in Figure [\[Case2\]](#Case2){reference-type="ref" reference="Case2"}). Therefore \(w\) is one end of \(e\). This means that \(\{ub, e\}\) is a 2-edge cut in \(G\), which is a contradiction.0◻
## Proof of Theorem [\[mainThm\]](#mainThm){reference-type="ref" reference="mainThm"}. {#proof-of-theorem-mainthm. .unnumbered}
Let \(G=(V, E)\) be the smallest counter-example for which the theorem does not hold. There are only three 3-edge connected essentially 4-edge connected cubic graphs with 6 vertices or less, on which the theorem can be shown to be true directly, as demonstrated in Figure [\[basecase\]](#basecase){reference-type="ref" reference="basecase"}, where bold lines represent edges in the subgraph, and dotted lines represent edges in \(G\) not in the subgraph. In the figure, for each such graph \(G\), the subgraphs \(H_{i}\) and the corresponding \(\lambda_{i}\) values for the required convex combination are shown. The smallest cubic 3-edge connected simple graph which is not essentially 4-edge connected has \(|V|=6\), which means that either
1. \(G\) is essentially 4-edge connected and \(|V| > 6\),
2. \(G\) has an essential 3-edge cut and \(|V| \geq 6\).
We pick an arbitrary edge \(uv\in E\), and we label the other adjacent vertices of \(u\) as \(a\) and \(b\), and the other adjacent vertices at \(v\) as \(c\) and \(d\). We will use this edge to create the vector \(z\in\mathbb{R}^{E}\) as a convex combination of incidence vectors of 2-edge connected spanning subgraphs, where for all \(e\in E\), \[z_{e} = \begin{cases} \hfill 1 \hfill & \text{ if \(e = uv\),} \\ \hfill \frac{1}{2} \hfill & \text{ if \(e \in \{ua, ub, vc, vd\}\),} \\ \hfill \frac{8}{9} \hfill & \text{ if \(e \neq uv\) and \(e\) adjacent to one of \(ua, ub, vc\) or \(vd\),} \\ \hfill 1 \hfill & \text{ if \(e \neq uv\) and \(e\) adjacent to two of \(ua, ub, vc\) or \(vd\),} \\ \hfill \frac{7}{9} \hfill & \text{ otherwise.} \\ \end{cases}\] Graph \(G\) has no essential 3-edge cut and \(|V|>6\), which implies that \(a\), \(b\), \(c\) and \(d\) are distinct. Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"} states that there does not exist two subsets \(A, B\subset V\) different from \(\{u, v\}\) such that \(au, vc\in \delta(A)\), \(au, vd\in \delta(B)\) and \(|\delta(A)|=|\delta(B)|=4\). Thus, assuming that \(\delta(B)\) is an essential 4-edge cut for some \(B\) automatically means that \(\delta(A)\) is not a 4-edge cut for any such \(A\), and vice-versa. Without loss of generality, assume that edges \(au\) and \(vc\) are not in an essential 4-edge cut together, other than \(\delta(\{u, v\})\), i.e. \(\delta(A)>4\). By Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"} again, there does not exist two subsets \(C, D\subset V\) different from \(\{u, v\}\) such that \(bu, vd\in \delta(C)\), \(bu, vc\in \delta(D)\) and \(\delta(C)=\delta(D)=4\). Knowing that \(|\delta(A)| > 4\) for all \(A\) such that \(au, vc\in\delta(A)\) and \(A\neq\{u,v\}\) implies that \(|\delta(C)|>4\) for all \(C\) such that \(bu, vd\in\delta(C)\) and \(C\neq\{u, v\}\), which means that edges \(bu\) and \(vd\) are not in an essential 4-edge cut together. Armed with these facts, we create graph \(G_{1}\) by removing edges \(au\) and \(vc\), and \(G_{2}\) by removing edges \(bu\) and \(vd\). In both graphs, any vertex \(s\) of degree 2 with incident edges \(st\) and \(sr\) is removed, and the edge \(sr\) is added. Because \(\delta(A)>4\), \(bd, ac\notin E\), no multi-edge are created. This situation is illustrated in the first part of Figure [\[UVPartialTransform\]](#UVPartialTransform){reference-type="ref" reference="UVPartialTransform"}; the second part of the figure illustrates the same situation, but when \(|\delta(A)| = 4\) and \(|\delta(B)| > 4\), for any \(B\) such that \(au, vd\in \delta(B)\) and \(B\neq\{u, v\}\). The astute reader may ask what happends if an edge links one of \(a\), \(b\), \(c\) or \(d\): assume that such an edge \(rs\in E\) exists, for \(\{r, s\}\subset\{a, b, c, d\}\). It is immediately apparent that \(r\) and \(s\) cannot both be neighbor to the same vertex, either \(u\) or \(v\), otherwise there would exist an essential 3-edge cut, e.g. \(\{rsu\}\). Without loss of generality, let \(r\) be adjacent to \(u\) and let \(s\) be adjacent to \(v\). For simplicity, let \(r=a\) and \(s=c\). Cut \(\delta(\{a, c\})\) is an essential 4-edge cut and \(ua, vc\in\delta(\{a, c\})\). By Lemma [\[submodularLemma\]](#submodularLemma){reference-type="ref" reference="submodularLemma"} and because \(|V|>6\), edges \(au\) and \(vd\) are not in an essential 4-edge cut together, and edges \(bu\) and \(vc\) are not in an essential 4-edge cut together, which means that the transformation is valid. Both \(G_{1}\) and \(G_{2}\) are 3-edge connected and cubic, because \(G\) was simple and essentially 4-edge connected and the two edges removed in \(G\) were not in an essential 4-edge cut together. The theorem holds for \(G_{1}\) and \(G_{2}\), as they are both smaller than \(G\), and gives two convex combinations of incidence vectors of 2-edge connected spanning subgraphs \(C_{1}\) and \(C_{2}\) for \(G_{1}\) and \(G_{2}\), respectively. Note that an edge in the convex combination is present \(\frac{7}{9}\) times overall in subgraphs. We will now modify the subgraphs in \(C_{1}\) and \(C_{2}\) to form sets \(C_{1}'\) and \(C_{2}'\) of subgraphs for \(G\) in the following way: assume that edges \(s\) and \(t\) were the two edges that we removed to transform \(G\) into \(G_{i}\), then any 2-edge connected spanning subgraph for \(G_{i}\) is 2-edge connected and spanning for \(G\) with the same edge selection, save for edges adjacent to \(s\) and \(t\) which we always select, and \(s\) and \(t\) which we always omit. This is illustrated on the left side of Figure [\[PatternsToMissingEdgeSevenNineth\]](#PatternsToMissingEdgeSevenNineth){reference-type="ref" reference="PatternsToMissingEdgeSevenNineth"}, where bold lines represent edges in the subgraph, dotted line represent edges omitted in the subgraph and dashed lines represent edges which may or may not be in the subgraph. In the figure, edges incident to \(a\), \(b\), \(c\) and \(d\), but not adjacent to \(uv\) are shown as distinct, for simplicity. We now take \(\frac{1}{2}C_{1}'+\frac{1}{2}C_{2}'\) to obtain a 2EC convex combination for \(G\). The occurrence of each edge in this convex combination is shown in the rightmost part of Figure [\[PatternsToMissingEdgeSevenNineth\]](#PatternsToMissingEdgeSevenNineth){reference-type="ref" reference="PatternsToMissingEdgeSevenNineth"}. Note that this is the occurrence for the case that no edges with both endpoints in the set \(\{a, b, c, d\}\) exist. In the event such edges exist, then these edges have an occurrence of \(1\), which does not affect the end result, as is shown later. The convex combination \(\frac{1}{2}C_{1}'+\frac{1}{2}C_{2}'\) is symmetrical (see the rightmost part of Figure [\[PatternsToMissingEdgeSevenNineth\]](#PatternsToMissingEdgeSevenNineth){reference-type="ref" reference="PatternsToMissingEdgeSevenNineth"}): therefore, repeating all the steps from Case [\[case1\]](#case1){reference-type="ref" reference="case1"}, taking every edge of \(G\) in turn as edge \(uv\) gives \(|E|\) convex combinations which we label \(F_{i}\), for \(i=1,\ldots,|E|\). We build a new convex combination by setting the lambda value \(\lambda_{F_{i}}=\frac{1}{|E|}\) for each \(F_{i}\), and conclude that it selects every edge \(\leq\frac{7}{9}\) times overall. We shall show this property in more detail here for an edge \(yz\in E\):
- edge \(yz\) is treated as edge \(uv\) in exactly one of the \(|E|\) convex combinations. It then has an occurrence of 1.
- there are no doubled edges, which means that \(yz\) is adjacent to \(uv\) in exactly 4 of the \(|E|\) convex combinations. It then has an occurrence of \(\frac{1}{2}\).
- Because \(G\) is cubic, there are exactly 8 ways to find an edge \(yz\) exactly one edge away from \(uv\), which are handled differently if \(yz\) and \(uv\) are in a 4-cycle together or not. Assume that \(yz\) and \(uv\) are not in a 4-cycle \(r\) times out of \(|E|\), where \(r \leq 8, r\in \mathbb{N}\) (the graph is cubic and simple, so there can be at most 8 such distinct edges for \(yz\)). It then has an occurrence of \(\frac{8}{9}\). If \(yz\) and \(uv\) are in a 4-cycle (i.e. \(yu, zu\in E\)), then \(yz\) is one edge away from \(uv\) both by \(yu\) and \(zu\). This happens \(t\) times over the \(|E|\) convex combinations and \(yz\) then has an occurrence of 1. Because there are 8 ways to be one edge away from \(uv\), and any 4-cycle with \(yz\) and \(uv\) uses two of those ways, \[2t + r = 8,\text{ for }t\in \mathbb{N}\text{.} \label{firstLine}\]
- The occurrence of \(yz\) in the other convex combinations is \(\frac{7}{9}\).
The average occurrence of an edge over the \(|E|\) convex combinations is \[\begin{aligned}
&\leq \frac{1}{|E|}\Big(1+t+4\times \frac{1}{2}+ r\times\frac{8}{9}+(|E|-r-t-5)\times\frac{7}{9}\Big)\nonumber\\ &=\frac{1}{|E|}\Big(\frac{2t}{9} + \frac{r}{9}-\frac{8}{9}+\frac{7|E|}{9}\Big)\nonumber\\ &= \frac{7}{9} + \frac{1}{9|E|}(2t+r-8)\text{.}\label{lastLine} \end{aligned}\] Since \(2t + r = 8\), by ([\[firstLine\]](#firstLine){reference-type="ref" reference="firstLine"}), it follows from ([\[lastLine\]](#lastLine){reference-type="ref" reference="lastLine"}) that the average edge occurrence over the \(|E|\) convex combinations is at most \(\frac{7}{9}\). The convex combination which results from giving each convex combination a weight of \(\frac{1}{|E|}\) selects every edge \(\frac{7}{9}\) times. If an edge is selected less than \(\frac{7}{9}\) of the time, we add it back arbitrarily to have it selected exactly \(\frac{7}{9}\) of the time. The theorem holds.
Notice that the ends of the three edges in \(C\) must be distinct because \(G\) is 3-edge connected. In this case we contract each shore of \(C\) to a single pseudo-vertex, to obtain graphs \(G_{1} = (V_{1},E_{1})\) with pseudo-vertex \(v_{1}\) and \(G_{2} = (V_{2},E_{2})\) with pseudo-vertex \(v_{2}\) (as shown in Figure [\[3ECSplittingSevenNinth\]](#3ECSplittingSevenNinth){reference-type="ref" reference="3ECSplittingSevenNinth"}). Both \(G_{1}\) and \(G_{2}\) are smaller than \(G\) so the theorem holds for \(G_{1}\) and \(G_{2}\). Moreover the patterns formed by the occurrence of the edges incident to \(v_{1}\) and \(v_{2}\) are unique and identical in the subgraphs in the corresponding convex combination. More specifically, exactly \(\frac{2}{9}\) of the time, one of the incident edges will not be in the subgraph, on both sides of the cut, and this is true for each of the three incident edges. The remaining subgraphs contain all three incident edges. These constant patterns allow us to "glue" (reconnect the edges as there were before the inductive step) the subgraphs for \(G_{1}\) and \(G_{2}\) together, in such a way that identical patterns at \(v_{1}\) and \(v_{2}\) are matched and the resulting subgraphs are 2-edge connected. Figure [\[3ECSplittingSevenNinth\]](#3ECSplittingSevenNinth){reference-type="ref" reference="3ECSplittingSevenNinth"} displays the patterns formed by the occurrence of the edges incident to \(v_{1}\) and \(v_{2}\) and their occurrence, on the left. The figure also illustrates the process of gluing the 2-edge connected spanning subgraph in the convex combinations for \(G_{1}\) and \(G_{2}\). This results in a convex combination that shows that the theorem holds for \(G\), which gives a contradiction.0◻
## Proof of Corollary [\[mainCorollary\]](#mainCorollary){reference-type="ref" reference="mainCorollary"}. {#proof-of-corollary-maincorollary. .unnumbered}
Theorem [\[mainThm\]](#mainThm){reference-type="ref" reference="mainThm"} implies that for graph \(G=(V, E)\), there exists a set of 2-edge connected spanning subgraphs \(H_{i}\) with \(\lambda_{i}\in\mathbb{R}_{\geq 0}\), \(i=1,\ldots,j\) such that \(\sum_{i = 1}^{j}\lambda_{i}=1\) and \(\sum_{i = 1}^{j}\lambda_{i}\chi^{E(H_{i})}_{e}=\frac{7}{9}\) for all \(e\in E\). This implies that for at least one of the \(H_{i}\), \(\sum_{e \in E}\chi^{E(H_{i})} \leq \frac{7|E|}{9}=\frac{7n}{6}\), since \(|E|=\frac{3n}{2}\) for cubic graphs.0◻ In other words, we show that the unit integrality gap for 2EC is bounded above by \(\frac{7}{6}\) for 3-edge connected cubic graphs, which improves upon Boyd, Iwata and Takazawa's upper bound of \(\frac{6}{5}\). Our methods are not polynomial and thus, do not result in an approximation algorithm. Nevertheless, they give hope that a \(\frac{7}{6}\)-approximation algorithm exists, which would improve upon the \(\frac{6n}{5}\)-approximation algorithm by Boyd, Iwata and Takazawa, and the \(\frac{7n}{6}\)-approximation algorithm when \(G\) is bipartite, by Takazawa. | {'timestamp': '2017-06-07T02:03:22', 'yymm': '1706', 'arxiv_id': '1706.01609', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01609'} |
null | null |
# Introduction
The Minimal Supersymmetric Standard Model (MSSM) is a leading candidate for physics beyond the Standard Model (SM). However, superpartners have remained stubbornly absent at the Large Hadron Collider (LHC). Moreover, bounds from direct detection experiments, including LUX, PandaX, and Xenon1T present an increasingly strong challenge to the WIMP (weakly interacting massive particle) paradigm both in the MSSM and more broadly. These experiments place the most pressure on models where dark matter--nucleon scattering is related to the cosmological history via a crossing symmetry. Crossing symmetry is spoiled if the dark matter coinhabits the thermal bath with another exotic state at the time of freeze-out. Then processes involving this co-annihilating partner can be important for the determination of the relic density, but are unrelated (at tree level) to direct detection. These co-annihilating scenarios are therefore among the WIMP models least constrained by direct detection bounds. The MSSM realizes this scenario when the superpartner of the top quark, the stop, is light, and neutralino co-annihilations with this state determine the relic density. In this work, we examine the stop co-annihilation scenario in terms of the low energy parameters most relevant for cosmology and direct detection. Many analyses of the stop co-annihilation parameters place an emphasis on simplified high-energy models, see e.g. Ref. However, because there are typically a small number of processes that dominate the cosmology involving only a handful of particles and couplings, it is illuminating to analyze these models in terms of the low energy parameters. Consistent with this approach, in previous work we considered the possibility where a single top partner (perhaps the superpartner of the right-handed top) was responsible for co-annihilation, see also Refs. for related work. In the context of the MSSM, however, a simplified model that includes a single co-annihilator, e.g. a \(\tilde{t}_R\), may be too simple to capture the physics of both cosmology and direct detection. Mixing between the light stop and the heavy stop can impact both of these processes. Indeed, the measurement of the Higgs boson mass suggests that there may be large mixing in the stop sector. In this case, it makes sense to include the full stop sector in the simplified model. This article is organized as follows. In Sec. [2](#s.orientation){reference-type="ref" reference="s.orientation"} we discuss the basics of the parameter space and discuss constraints unrelated to the dark matter story. Cosmology is analyzed in Sec. [3](#s.relic){reference-type="ref" reference="s.relic"} and we turn to implications for direct detection in Sec. [4](#s.DD){reference-type="ref" reference="s.DD"}. Finally, Sec. [5](#s.conc){reference-type="ref" reference="s.conc"} is reserved for our conclusions.
# Orientation {#s.orientation}
We consider the case of pure bino neutralino dark matter \(\chi\), with co-annihilations resulting from the presence of one or more colored states \(Y\) with masses not dissimilar to that of the neutralino. If equilibrium between the neutralino and the colored state is maintained (as is typically the case due to scatterings off the SM bath), then processes of the form \((\chi Y \rightarrow SM)\) or \((YY \rightarrow SM)\) are relevant for setting the dark matter abundance. A small admixture of Higgsino will not affect the cosmological history in detail. However, doping the bino with even a small Higgsino fraction can impact direct detection since the Higgs boson has a tree-level bino-Higgsino coupling. The Higgsino fraction of the neutralino is controlled by mixing suppressed by \(M_{Z}/\mu\), with \(\mu\) the Higgsino mass parameter. The tree-level direct detection cross section is well approximated by: \[\label{e:HDD} \sigma_{SI}^{tree-mixing} \approx 3 \times 10^{-48}\textrm{cm}^2 \left( \frac{2 \textrm{ TeV}}{\mu} \right)^4 \left( \frac{m_\chi}{600 \textrm{ GeV}} \right)^2 \left(1+\frac{\mu \, s_{2 \beta }}{m_{\chi }}\right)^2 \left(1-\frac{m_{\chi }^2}{\mu ^2}\right)^{-2}.\] Here, \(s_{2 \beta} \equiv {\sin 2 \beta}\) where \(\tan \beta\) is the ratio of vacuum expectation values of the two Higgs doublets. In this work, we will be interested in the case where this contribution to direct detection is subdominant. That is, we consider the case where \(\mu\) is (quite) large, and we ask the question: what is the direct detection cross section induced only by the presence of a stop co-annihilation cosmology? Because we will be interested in discussing the effects of reproducing the required Higgs boson mass, we will include the full stop sector, i.e both \(\tilde{t}_R\) and \(\tilde{t}_L\). This means that the left-handed sbottom \(\tilde{b}_L\) is necessarily in the spectrum as well. For simplicity, we assume the partner of the right-handed bottom is decoupled. The precise value of its mass has little effect on either the physical Higgs boson mass or any fine-tuning arguments.[^1] The stop mass matrix is given by: \[M^2_{\tilde{t}}=\left( \begin{array}{cc} m_{Q_3}^2 + m_t^2 + D_L & \quad m_t X_t \\ & \\ m_t X_t & \quad m_{u_3}^2 + m_t^2 + D_R\\ \end{array} \right),\] where \(X_t = (A_t-\mu/\tan\beta)\), \(m_{Q_3},m_{u_3}\) are the left and right-handed soft masses, and the \(D\)-terms are: \(D_L= M_Z^2 \cos 2 \beta (\frac{1}{2}-\frac{2}{3} s_W^2)\) and \(D_R=\frac{2}{3} M_Z^2 \cos 2\beta s_W^2\). Here, \(s_{W} = \sin \theta_W\) is the weak mixing angle. We choose sign conventions for the sign of the stop mixing angle consistent with `SuSpect_2.41` . We have: \[\sin 2\theta_{\tilde{t}} \equiv s_{2t}= \frac{-2 m_t X_t}{m_{\tilde{t}_2}^2-m_{\tilde{t}_1}^2 }\;,\] where we denote the stop mass eigenstates as \(\tilde{t}_1 = c_t \, \tilde{t}_L +s_t \, \tilde{t}_R\) and \(\tilde{t}_{2} =-s_t \, \tilde{t}_{L} +c_t \, \tilde{t}_R\) and use the shorthand \(s_{t}= \sin \theta_{\tilde{t}}\), etc. The stop mixing angle is a function of \(X_t\) rather than \(A_t\), \(\mu\) or \(\tan \beta\) directly. In the region of interest, \(X_t \sim A_t\), and the dependence on \(\tan \beta\) is minimal. For concreteness we fix \(\tan\beta=10\) but numerically verify that results are approximately independent of \(\tan\beta\), as long as the values are not too extreme. The heaviness of the right-handed sbottom ensures that the \(\tilde{b}_1\) state has mass given approximately by \(m_{Q_3}\). But for simplicity, we also set \(X_b=0\) so there is no mixing in the sbottom sector, regardless of \(\tan \beta\). This ensures that the \(\tilde{b}_1\) is completely left-handed, and its mass may be expressed in terms of the stop masses and mixing angles: \[m_{\tilde{b}_1}\equiv m_{\tilde{b}_L} = c_t^2 m_{\tilde{t}_1}^2+s_t^2 m_{\tilde{t}_2}^2-M_W^2 c_{2\beta} +m_b^2-m_t^2\;.\] For later reference, we record the couplings of the stops and the left-handed sbottom to the Higgs and the bino: \[\begin{aligned}
\label{ht1t1} g_{h\tilde{t}_1\tilde{t}_1}&=&-\frac{2}{v}\left[ M_Z^2 c_{2\beta}\left( \frac{1}{2} c_t^2-\frac{2}{3} s_w^2 c_{2 t}\right) +m_t \left(m_t+\frac{1}{2} s_{2 t} X_t\right) \right] \; ,\\ \label{ht2t2} g_{h\tilde{t}_2\tilde{t}_2}&=&-\frac{2}{v}\left[ M_Z^2 c_{2\beta}\left( \frac{1}{2} s_t^2 +\frac{2}{3} s_w^2 c_{2 t}\right) +m_t \left(m_t-\frac{1}{2} s_{2 t} X_t\right) \right] \;, \\ \label{ht1t2} g_{h\tilde{t}_1\tilde{t}_2}&=& \frac{2}{v}\left[ M_Z^2 c_{2\beta}s_{2 t} \left( \frac{1}{4}-\frac{2}{3} s_w^2 \right)-\frac{1}{2} m_t c_{2 t} X_t \right] \; ,\\ \label{t1xx} g_{\tilde{B}\tilde{t}_1{t}}&=&\frac{\sqrt{2}}{3} g_Y \left(-\frac{1}{2}c_t P_L + 2s_t P_R\right)\; ,\\ \label{t2xx} g_{\tilde{B}\tilde{t}_2{t}}&=& \frac{\sqrt{2}}{3} g_Y \left(\frac{1}{2}s_t P_L + 2c_t P_R\right)\; ,\\ \label{bLx} g_{\tilde{B}\tilde{b}_L{b}}&=&-\frac{\sqrt{2}}{6}g_Y \;, \end{aligned}\] where the Higgs vacuum expectation value \(v=246\) GeV, \(g_Y\) is the hypercharge gauge coupling and \(P_L\), \(P_R\) are the projection operators. We implemented this simplified model containing the third generation squark doublet, right-handed stop and a bino neutralino in `MicroOmegas` . As a cross-check we validated our numerics with the full MSSM implementation.
## Indirect Constraints {#s.Higgs}
Under the assumption that the MSSM is the underlying theory, important information about the superpartner spectrum can be gleaned from the Higgs boson mass. Indeed, the MSSM cannot reproduce a Higgs boson mass of \(125\) GeV without substantial radiative corrections from the stop sector. Because the Higgs mass constraint impacts the allowed properties of the stops, it is of interest to study whether it has phenomenological implications for the stop co-annihilation region. Nevertheless, one should bear in mind the possibility that the observed Higgs mass is generated by physics beyond the MSSM. We therefore prefer to begin by considering additional indirect consequences of the stop sector, potentially relevant even if there are additional contributions to the mass of the Higgs boson. In particular, before discussing the impact of the Higgs mass, we discuss three sets of constraints: electroweak precision observables (EWPO), Higgs boson production and decays, and the stability of the physical vacuum due to the presence of charge and color breaking vacuua. It is a model-dependent question as to whether the above would be affected by deviations from the MSSM. At minimum, however, these constraints give an indication of cancellations that would need to occur between the stop sector and any additional contributions. Once collider bounds are imposed, we find constraints from vacuum stability are generally the strongest of the three. Regarding EWPO, we require the stop sector contribution to \(\delta \rho\) not exceed the \(2\sigma\) bound from the measured \(\delta \rho = 3.7 \pm 2.3 \times 10^{-4}\) . Assuming negligible mixing in the sbottom sector, the third generation squarks yield: \[\delta \rh = \frac{3 G_{F}}{8 \sqrt{2} \pi^2} \left(-s_t^2 c_t^2 F_0[m_{\tilde{t}_1}^2,m_{\tilde{t}_2}^2] + c_t^2 F_0[m_{\tilde{t}_1}^2, m_{\tilde{b}_L}^2] + s_t^2 F_0[m_{\tilde{t}_2}^2, m_{\tilde{b}_L}^2]\right),\] with \(G_{F}\) the Fermi constant while the loop function \(F_0\) is given by \[F_0[x,y] \equiv x+ y-\frac{2 xy }{x-y} \log\left[\frac{x}{y}\right].\] In the limit of negligible mixing in the stop sector, the bound on \(\delta \rho\) implies \(m_{\tilde{t}_L} > 290\) GeV. In the case with hierarchical stops, where the lightest stop is still mostly left-handed, it can be shown that \(X_t\sim m_{\tilde{t}_2}\) minimizes the above contribution to \(\delta\rho\) . This is due to an approximate custodial symmetry between the sbottom and the lightest stop--these two states have an approximate mass degeneracy in this regime: \(m_{\tilde{t}_1}^2 \equiv m_{Q_3}^2 +m_t^2\left( 1-X_t^2/m_{u_3}^2\right) \sim m_{Q_3}^2\equiv m_{\tilde{b}_1}^2\). Deviating too far from this limit (i.e. very large \(X_{t}\)), causes the constraints to become significant. The excluded region is shown in Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"} (red region) in the \(\sin\theta_{\tilde{t}}\) vs. \(m_{\tilde{t}_2}\) plane for two different values of \(m_{\tilde{t}_1} = 600\) GeV (a) and 1.5 TeV (b). Higgs boson properties can also be influenced by the presence of a light stop. The primary impact is a potential contribution to the gluon fusion production cross section. The contribution to \(H\rightarrow \gamma \gamma\), while less pronounced, should also be considered. Using the Higgs low energy theorem, the contribution of \(\tilde{t}_{1,2}\) to the production amplitude of the Higgs boson in gluon fusion can be approximated as \[\label{e:gf} \mathcal{A}_{hgg} \simeq \mathcal{A}_{hgg}^{\rm SM} + \frac{m_t^2}{ m_{\tilde{t_1}}^ 2m_{\tilde{t_2}}^2}(m_{\tilde{t}_1}^2+m_{\tilde{t}_2}^2-X_t^2)\,,\] where \(A^{\rm{SM}}_{hgg} = 4\) is the SM amplitude. Analogously we can write the amplitude for \(H\rightarrow \gamma \gamma\) as \[\label{e:gg} \mathcal{A}_{h\gamma\gamma} \simeq \mathcal{A}_{h\gamma\gamma}^{\rm SM} +\frac{8}{9} \frac{m_t^2}{ m_{\tilde{t_1}}^ 2m_{\tilde{t_2}}^2}(m_{\tilde{t}_1}^2+m_{\tilde{t}_2}^2-X_t^2)\,,\] where \(\mathcal{A}_{h\gamma\gamma}^{\rm SM} =-13\) . Again, if \(X_{t} \gg m_{\tilde{t}_2}\), both gluon fusion and the diphoton rate will be impacted, but because \(| \mathcal{A}_{hgg}^{\rm SM} | \ll |\mathcal{A}_{h\gamma\gamma}^{\rm SM}|\), the stop contribution to gluon fusion provides a larger deviation from SM expectations. The combined analysis of the 8 TeV LHC run by ATLAS and CMS constrains the signal strength of the production in gluon fusion, normalized to the SM prediction, to be \(\mu_{ggF} = 1.03^{+0.17}_{-0.15}\) while \(\mu_{\gamma \gamma} = 1.16^{+0.20}_{-0.18}\) . Allowing for near future improvements in these measurements, we allow a deviation of up to 20% in these observables. With the increasingly powerful direct searches for the stop (which impose \(\tilde{t}_{1} \mathop{}_{\textstyle \sim}^{\textstyle >} 500\) GeV, more below), these constraints are only relevant for particularly large values of the \(X_{t}\) parameter. The region excluded by Higgs signal strength measurements is shown in orange in Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}. Charge and color breaking vacua may appear at large values of the trilinear coupling \(A_{t}\) and the physical vacuum can become metastable, possibly with a lifetime shorter than that of the observed universe. To ensure a metastable physical vacuum with a sufficiently long lifetime, following Ref. [^2] we impose the rule of thumb condition \[\label{eq:At} A_{t}^{2} \leq \left( \alpha + \beta \frac{|1-r|}{1+r}\right)(m_{\tilde{t}_1}^2 +m_{\tilde{t}_2}^2)-\gamma M_Z^2/2\;,\] with \(\beta =-0.5\), \(\alpha= 3.4\), \(\gamma =60\) and \(r \equiv m_{u_3}^2/m_{Q_3}^2\). We see that stability of the physical vacuum roughly requires \(|A_{t}| \lesssim m_{\tilde{t}_2}\). Because \(X_t \equiv \left(A_t-\mu/\tan\beta\right)\), the precise impact of Eq. ([\[eq:At\]](#eq:At){reference-type="ref" reference="eq:At"}) on the stop sector will depend on the choice of \(\mu\) and \(\tan\beta\). However, even for fairly large values of \(\mu\), as long as \(\tan\beta\) is larger than a few, \(X_t\sim A_t\) and the impact of the exact choice of \(\mu\) and \(\tan\beta\) is minimal. For concreteness, when imposing the vacuum stability constraint, Eq. ([\[eq:At\]](#eq:At){reference-type="ref" reference="eq:At"}), on the stop parameters, we fix \(\mu = 3~ m_{\tilde{t}_1} \sim 3~m_\chi\), a value large enough that the tree-level direct detection cross-section will be small, see Eq. ([\[e:HDD\]](#e:HDD){reference-type="ref" reference="e:HDD"}).[^3] In Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}, the purple region denotes where the vacuum becomes unstable according to this prescription. The slight asymmetry between positive and negative values of \(\sin \theta_{\tilde{t}}\) is because of the difference between \(X_t\) and \(A_t\) due to our choice of \(\mu/\tan\beta\). In the absence of Higgs mass information, the above constraints are responsible for limiting the maximal value of \(X_t\) allowed. Because \(X_{t}\) enters the Higgs boson couplings to the stops, this affects both cosmology and direct detection. As we will now see, the Higgs mass constraint gives comparable, and often stronger bounds on \(X_{t}\).
## Higgs Mass {#s.Mass}
At tree level, the MSSM predicts a Higgs boson with mass below \(M_{Z}\). However, loop effects give important contributions. In our analysis, we employ the Higgs mass calculation as implemented in `FeynHiggs-2.13.0` . In the stop sector, we scan over \(m_{Q_3}\), \(m_{u_3}\) and \(A_t\). For concreteness we fix \(M_3 = \pm 1.5\) TeV, \(X_b=0\), and all other soft masses and trilinear couplings to 5 TeV. To capture potential uncertainties due to even higher order corrections and -- more importantly -- uncertainties in inputs (such as the top quark mass or the Gluino mass, which can change the Higgs mass up to 4 GeV when the Gluino mass is varied from 1 to 3 TeV ) we allow a Higgs boson mass in a window from 122 to 128 GeV. The sign of \((A_t M_3)\) has an asymmetric effect on the Higgs mass primarily due to the running of the top Yukawa (see Ref. and references therein). However, this sign does not otherwise affect the dark matter phenomenology. Hence, we check both signs of the gluino mass, to see whether either realizes a valid Higgs mass. Consistency with the observed Higgs boson mass in the presence of a relatively light \(\tilde{t}_1 \mathop{}_{\textstyle \sim}^{\textstyle <}\) TeV requires a relatively heavy \(\tilde{t}_2\). Unless one wishes to allow for an extreme hierarchy between the stop masses, an appreciable mixing, \(|X_t| \sim m_{\tilde{t}_2}\) is also preferred . In Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}, the green swath denotes the region with an observationally consistent Higgs boson mass. The two bands that comprise the swath correspond to the two signs of \(M_3\). Also in Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}, we indicate where the sbottom mass is degenerate with \(m_{\tilde{t}_1}\) by the solid black line. The two dashed (dotted) contours show where the sbottom is 5 GeV lighter (heavier) than \(m_{\tilde{t}_1}\). The central region enclosed by the thick black lines is then where the sbottom is lighter than the lightest stop, and hence would be the next to lightest supersymmetric particle (NLSP). Close to these lines, we expect the sbottom and stop to act as co-NLSPs. This information will be relevant when we discuss the cosmology associated with this scenario. To summarize, the region compatible with the Higgs boson mass is mostly unaffected by the indirect constraints of the previous section. However, the largest values of \(X_{t}\) consistent with the Higgs mass can be in tension with the stability of the vacuum.
## LHC {#s.LHC}
Owing to its small production cross section, pure bino dark matter is difficult to probe at the LHC. The colored, nearly degenerate co-annihilation partner, on the other hand, possesses a relatively large production cross section. While small mass splittings between the states degrade the classic missing energy signature, a variety of searches targeting this region now exist. We review a few of the most relevant searches here. Other particles in the SUSY sector are also potentially accessible, i.e. \(\tilde{t}_2\) and \(\tilde{b}_1\). Since they will typically have a larger splitting with the lightest supersymmetric particle (LSP), the sensitivity to these states can be enhanced. Hence, in some cases it is possible that their limits can be important, despite the smaller production cross section. In the co-annihilation scenario the mass splitting between the \(\tilde{t}_1\) and the dark matter is always less than \(m_{t}\), so the two-body decay of the stop \((\tilde{t_1} \rightarrow \chi^0 t)\) is kinematically forbidden. Indeed, for all but the heaviest stops, the splitting is much less than \(M_W\), which also forbids the three-body decay \((\tilde{t_1} \rightarrow \chi^0 W^+ b)\), leaving only the four-body decay \((\tilde{t} \rightarrow b \chi^{0} f \bar{f}')\), or the flavor violating two-body decay \((\tilde{t} \rightarrow c \chi^{0})\). The relative importance of these last two decays depends on the flavor properties of the SUSY breaking sector . Recent searches by the CMS collaboration have focused on this compressed region, and have ruled out stop masses up to roughly 500 GeV . The exact mass excluded depends on the precise value of the mass splitting and the relative branching ratio of these two channels. Monojet searches, such as those undertaken at ATLAS can also be relevant, but typically are not as constraining. For \(m_{\tilde{b}} > m_{\tilde{t}_1}\) the sbottom decay pattern depends strongly on the mass splitting. If kinematically accessible and if there is any left-handed component in the lightest stop, (\(\tilde{b} \rightarrow \tilde{t} W\)) dominates, which is challenging to observe. However, a dedicated search may be possible. If the splitting is less than \(M_{W}\), \((\tilde{b} \rightarrow b \chi)\) robustly excludes \(m_{\tilde{b}} < 625\) GeV . A phenomenological projection suggests that a combined search for all the different decay modes of the sbottom could ultimately probe \(m_{\tilde{b}} \lesssim 900\) GeV with \(300 \; \mbox{fb}^{-1}\) of data. Finally, the decay width of the lightest stop is suppressed if the mass splitting is small, and for \(\Delta m \lesssim 10\) GeV the \(\tilde{t}_1\) has a proper decay length exceeding a meter. Current limits on such long-lived particles exclude stable stops with \(m_{\tilde{t}_1} < 1020\) GeV. As we shall see, a co-annihilation cosmology does not motivate such a small mass splitting for stops this light (see next section). Even with 300 fb\(^{-1}\) it will be challenging for these searches to probe the relevant region, where the stop mass exceeds 1.5 TeV. At present, stops with modest mass splittings from the neutralino and mass roughly greater than 500 GeV are allowed, so in what follows we will focus on this region. The LHC will continue to push these bounds upwards as additional data are taken.
# Relic Density {#s.relic}
In the model studied here, co-annihilation of \(\chi\) with \(\tilde{t}_1\) and \(\tilde{b}_1\) , as well as the annihilation of pairs of \(\tilde{t}_1\) and/or \(\tilde{b}_1\) are relevant for the precise determination of \(\Omega h^2 = 0.11805 \pm 0.0031\). For some spectra, even processes involving \(\tilde{t}_{2}\) can be relevant. Contributions from processes with particles other than \(\chi\) in the initial state must be appropriately weighted by their thermal abundance. The impact of these additional processes is accounted for via an effective annihilation cross section \[\begin{aligned}
\sigma_{eff} {\rm v} = \sum_{i,j} \frac{n_i^{eq} n_j^{eq}}{\left(\sum_k n_k^{eq}\right)^2}\sigma_{ij} {\rm v}\;, \end{aligned}\] where \(n_i^{eq} = g_i [m_i T/ (2 \pi)]^{3/2} e^{-m_i/T}\); \(m_i\) is the mass of the particle \(i\), and \(g_i\) counts the number of internal degrees of freedom. Since the abundance of a heavier state \(i\) is suppressed relative to that of the LSP by factors of \(e^{-\Delta m /T_F}\), where \(\Delta m = \left(m_{i}-m_{\chi}\right)\), the relic density is extremely sensitive to the mass splitting between \(\chi\) and the co-annihilator. For \(m_{i} \gtrsim 1.2\, m_{\chi}\) co-annihilations can safely be neglected. Contributions to the effective cross section from the annihilations of a pair of NLSPs are doubly exponentially suppressed compared to \(\chi\chi\) annihilations, and the annihilations of \(\chi\) with a single co-annihilator are singly exponentially suppressed. Note, strongly interacting particles \(\tilde{f}\) always possess the annihilation channel \((\tilde{f} \tilde{f^{\ast}} \rightarrow gg)\) with cross section \[\sigma {\rm v} \left(\tilde{f} \tilde{f}^*\rightarrow gg\right)= \frac{7 g_s^4}{216 \, \pi \, m_{\tilde{f}}^2} \,.\] Thus, co-annihilations impose a lower limit on \(\Delta m\) as a function of \(m_{\chi}\), since co-annihilations will be too effective if the mass splitting is too small.[^4] A complication arises from the Sommerfeld effect, which is known to have an impact on the annihilation rate of charged non-relativistic particles . While the Sommerfeld effect does not directly affect the annihilation of binos, it will have an impact on the annihilation rates of \(\tilde{t} \tilde{t}^{\ast}\) and \(\tilde{t} \tilde{t}\) (and potentially their analogs including the sbottom). These corrections can modify the relic density significantly . Qualitatively, the Sommerfeld effect can be understood as an enhancement (or suppression) of the leading order cross section due to the presence of an attractive (repulsive) potential between the initial state particles generated by the exchange of light force carriers. In the context of color charged particles such a potential is generated by gluon exchange and can be approximated by the leading term of the static QCD potential ,: \[\begin{aligned}
V(r)\approx C \frac{\alpha_S}{r} = \frac{\alpha_S}{2 r} \left(C_Q-C_R-C_{R'}\right)\;, \end{aligned}\] where \(C_R\) and \(C_R'\) are the quadratic Casimirs of the color representation of the incoming particles while \(C_Q\) is the quadratic Casimir of the final state. In this case the long and the short range contributions to the annihilation process factorize. The Sommerfeld corrected s-wave cross section is given by \(\sigma^S \equiv S_0 \;\sigma_{0}\) where \(\sigma_{0}\) is the short distance contribution, and \(S_0\) is an enhancement (suppression) factor given by \[\begin{aligned}
S_0=\frac{\pi \alpha/\beta}{ e^{\pi \alpha/\beta}-1}\;. \end{aligned}\] Here \(\alpha \equiv 1/2 \,\alpha_S \,(C_Q-C_R-C_{R'})\) is the strength of the potential, and \(\beta=v/2\) where \(v\) corresponds to the velocity of the particles in the initial state. The appropriate Sommerfeld factors for higher partial waves \(S_l\) can be constructed using a recursion relation starting from \(S_0\) . For a detailed description of our treatment of the Sommerfeld effect see Ref. . We have implemented the Sommerfeld effect in `MicrOmegas 3.3` which we use to solve the relevant Boltzman equations numerically[^5]. NLO calculations of the relic density have made significant progress in recent years and indicate that perturbative corrections can have a relevant impact on the annihilation rate. For stop coannihilation these corrections were found to change the relic density by \(\approx 20 \%\) . As expected the impact is even larger if Sommerfeld corrections are also included . This is certainly larger than the observational uncertainty on the relic density and should be included in detailed studies of the MSSM. In the scenario at hand this correction can be absorbed via a shift of \(\Delta m\) which does not have an appreciable effect of the qualitative features of the phenomenology. It has recently been emphasized that bound state formation may have an impact on dark matter freeze-out . For squarks the non-Abelian structure of QCD leads to a partial accidental cancellation in the matrix element for bound state formation which somewhat diminishes its importance. Once all electroweak annihilation channels are included in the freeze-out calculation, the impact of bound states was found to be at the \(10\%\) level for all viable neutralino masses . We do not consider their effects here, but note that while the formation of bound states would allow for somewhat heavier dark matter consistent with the relic density constraint, the qualitative picture would remain the same. In Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} we show the mass difference \(\Delta m\) between \(\chi\) and the NLSP where co-annihilations provide a viable mechanism for reproducing the dark matter density. We have shown separately the regions where the left-handed sbottom is the NLSP and where the lightest stop is the NLSP. To give guidance for the importance of various processes in the early universe we have shown two additional curves. The first is the \(\Delta m\) curve (solid black line) if only the process \((\tilde{f} \tilde{f}^* \to gg)\) were active. The true \(\Delta m\) is always larger, showing the importance of other annihilation channels. Another curve, labeled \(\tilde{t}_R\) (dot-dashed black line), shows the mass splitting if the NLSP were a single pure right-handed stop. In this case, there are no \(X_{t}\) enhanced couplings to the Higgs boson, but the pure right-handed stop has the largest coupling to the bino \[cf. Eqs. ([\[ht1t1\]](#ht1t1){reference-type="ref" reference="ht1t1"})-([\[bLx\]](#bLx){reference-type="ref" reference="bLx"})\]. In this case, the processes \(\chi \chi \to t \bar{t}\), \(\chi \tilde{t}_R \to t g/h\), \(\tilde{t}_R \tilde{t}_R^* \to t\bar{t}/gg/hh\) are relevant. The precise value of the mass splitting depends strongly on the mixing in the stop sector. This mixing controls both interactions with the Higgs boson and the bino, see Eqs. ([\[ht1t1\]](#ht1t1){reference-type="ref" reference="ht1t1"})-([\[t2xx\]](#t2xx){reference-type="ref" reference="t2xx"}). The variation in these couplings due to different mixings explains the width of the band of consistent \(\Delta m\) at each given value of the dark matter mass. In Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"}, the allowed mass splitting for stop NLSPs broadens for higher dark matter masses. Near \(m_{\chi} \sim 500\) GeV, it populates a band from \(\sim 35-60\) GeV, but in the multi-TeV range it can range from near degeneracy to mass splittings approaching 100 GeV. The largest mass splittings require contributions from channels such as \((\tilde{t}_1 \tilde{t}_1^* \to hh)\) to be large--realized by increasing \(X_t\) \[cf. Eq. ([\[ht1t1\]](#ht1t1){reference-type="ref" reference="ht1t1"})\]. The cut-off in the maximal values of \(\Delta m\) seen in the left panel results from the need for an \(X_t\) \((A_t)\) so large as to make the vacuum unstable. The smallest \(\Delta m\) obtained are somewhat smaller than the values consistent with a purely right handed stop NLSP. This is because perturbing away from the pure \(\tilde{t}_R\) case, an important effect is that the \(\tilde{t}_L\) admixture has a much smaller hypercharge and hence the coupling to \(\chi\) is reduced, see Eq. ([\[t1xx\]](#t1xx){reference-type="ref" reference="t1xx"}). This suppresses diagrams that rely on this coupling, necessitating a smaller \(\Delta m\) to maintain the thermal relic abundance. In contrast, sbottom NLSPs never have \(\Delta m\) exceeding 80 GeV, irrespective of the vacuum stability constraint. This is because for the sbottom to be the NLSP, \(X_t\) cannot be too large (lest the \(\tilde{t}_1\) mass be driven down by level repulsion). In the sbottom NLSP case, the lightest stop will always be nearby, and for some parameter choices the heavier stop is also allowed to be degenerate. In fact the largest mass splitting obtained for the sbottom NLSP is precisely where both the stops are nearly degenerate with the sbottom, and dominant annihilation channels may include \(\tilde{t}_{1,2}\) and \(\tilde{b}_1\); for example \((\tilde{f} \tilde{f}^* \to gg)\) or \((\tilde{t}_2/\tilde{t}_1 \tilde{b}_1 \to Wg)\). As the heavier stop decouples from the cosmology, the required mass splitting decreases and the smallest \(\Delta m\) is obtained when only \(\tilde{t}_1\) and \(\tilde{b}_1\) contribute to co-annihilations. The right panel of Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} shows the region that is carved out demanding consistency with the Higgs mass. Note that in particular the width of the sbottom NLSP region is significantly reduced. This can be understood by looking at the region in Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"} where the sbottom is the NLSP: Only values of \(s_t\) close to 0 and very large \(m_{\tilde{t}_2}\) can accommodate both a sbottom NLSP and a consistent Higgs mass. This results in a well-defined cosmology for this scenario, and hence a well-defined \(\Delta m\): values of the mass splitting are obtained close to the ones denoted by the bottom of the \(\tilde{b}_1\) NLSP band in the left panel of Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} (\(m_{\tilde{b}_L}\sim m_{\tilde{t}_1} \ll m_{\tilde{t}_2}\)). Regarding the stop NLSP region, some of the largest \(\Delta m\) are eliminated as they require an \(X_{t}\) so large that a large enough Higgs boson mass may not be achieved (see the right panel of Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}, particularly for \(s_t > 0\)). While the masses of the stop NLSPs shown in Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} are in broad agreement with the latest LHC limits, it should be noted that the very lightest masses for a \(\tilde{b}_1\) NLSP (approximately \(m_{\chi} \lesssim 625\) GeV) are excluded by LHC searches, see Sec. [2.3](#s.LHC){reference-type="ref" reference="s.LHC"}. The mass splittings shown in Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} represent important future targets for the LHC experiments.
# Direct Detection {#s.DD}
Having determined the regions of parameter space in which thermal freeze-out can account for the observed relic density, we now turn to direct searches for dark matter. The bino dark matter candidate considered here has a vanishing direct detection cross section at tree level. A process that can generate a direct detection cross section arises from the loop-induced coupling of the dark matter with the Higgs boson. Such an effective \(\chi\chi h\) coupling is generated by triangle diagrams with tops and stops in the loop, see Figs. [\[fig:DDdiagrams\]](#fig:DDdiagrams){reference-type="ref" reference="fig:DDdiagrams"}a and [\[fig:DDdiagrams\]](#fig:DDdiagrams){reference-type="ref" reference="fig:DDdiagrams"}b. This effective coupling has been calculated in Ref. and re-derived by us using the low energy Higgs theorem .[^6] In addition to the Higgs triangle-diagrams there are also box-diagrams with \(\tilde{t}_{1,2}/t\) and \(\tilde{b}_1/b\) in the loop which induce an effective coupling between the bino and the gluon content in the nucleus (see Fig. [\[fig:DDdiagrams\]](#fig:DDdiagrams){reference-type="ref" reference="fig:DDdiagrams"}c as an example diagram). The full loop result for this contribution is available in the literature and is already included in the MSSM implementation of `MicrOmegas`. We show the direct detection cross section as a function of the neutralino mass in Fig. [\[f:DD\]](#f:DD){reference-type="ref" reference="f:DD"} considering different constraints. First, in red, we show all points in our scan, which include \(X_{t}\) that would violate the vacuum stability constraint by as much as \(50\%\). Next, in blue, we show points consistent with the observed Higgs signal strength (gluon fusion rate), EWPO (\(\delta \rho\)), and a metastable vacuum with sufficiently long lifetime. Because all points shown have \(m_{\chi} > 500\) GeV, they are consistent with current stop searches. In addition, we ensure that the points are consistent with the recent sbottom searches, which eliminates some points with mass between 500 and 625 GeV where the sbottom was nearly degenerate with the LSP. Finally, in green, we show points that also yield a Higgs mass 122 GeV \(< m_{h} <\) 128 GeV. Once the Higgs mass is imposed, all points lie well below future experimental sensitivity, and indeed below the neutrino floor. We can understand the behavior of the direct detection in some detail by examining the relevant amplitudes. At the nucleon level the direct detection cross section is given by: \[\sigma_{SI}^n = \frac{4m_\chi^2m_n^4\, \mathcal{A}^2}{\pi (m_\chi+m_n)^2}\,,\] where \(n\) specifies the nucleon (i.e. the neutron or the proton) and \(\mathcal{A}\) is the amplitude which can be separated into a Higgs exchange contribution and the gluon box contribution \[\mathcal{A}=\mathcal{A}^h+\mathcal{A}^g.\] We discuss each of these contributions in turn, attempting to understand how large an amplitude can be realized for each. The gluon box diagram is suppressed by the relevant fermion mass and hence can be enhanced when the sbottom is the NLSP and the mass splitting is comparable to \(m_b\). In this case, \(\mathcal{A}^g\) is well approximated by \[\begin{aligned}
\label{e:glubox} \mathcal{A}^g &\simeq& \frac{g^2 t_w^2}{36}\frac{1}{m_\chi \left(m_b+\Delta m\right)^2}\left[ \frac{f_{T_g}^{n}}{270}+\frac{13}{240 \pi} G_\alpha^b\right] \nonumber \\ &\approx &1.5 \times 10^{-11} \, \mathrm{GeV}^{-3} \left(\frac{2 \, \mathrm{TeV}}{m_\chi}\right)\left( \frac{m_b+20~ \mathrm{ GeV} }{m_b+\Delta m}\right)^2, \end{aligned}\] where we use default values from `MicrOmegas` for \(f_{Tg^b}^n\simeq 0.98\) and \(G_\alpha^b\simeq 0.1\). The Higgs amplitude receives contributions from diagrams where the Higgs boson couples to the \(t\) as well as to both the stops \(\tilde{t}_{1,2}\). It can be decomposed as \[\begin{aligned}
\label{eqn:HiggsExchangeDD} \mathcal{A}^h&=&\frac{g^2t_w^2}{24 \pi^2}\frac{m_t}{v^2 m_h^2}~ \mathcal{A}^h_{t\tilde{t}}~\sum_q f_q^n , \qquad\quad f_q^n = \left \{\begin{array}{c c} f_{Tq}^n,& \quad \quad q=\{u,d,s\} \\ \frac{2}{27} f_{Tg}^n, & \,\;\;\quad q=\{c,b,t\} \end{array} \right. \, \nonumber\\ & \approx& 2.7 \times 10^{-11} \mathrm{GeV}^{-3} ~ \mathcal{A}^h_{t\tilde{t}}, \end{aligned}\] where \(\mathcal{A}^h_{t\tilde{t}} = \left( \mathcal{A}^h_t+\mathcal{A}^h_{\tilde{t}_{1,1}}+\mathcal{A}^h_{\tilde{t}_{1,2}}+\mathcal{A}^h_{\tilde{t}_{2,2}} \right)\), \(f_{Tg}^n = (1-\sum_{u,d,s}f_{Tq}^n)\) parametrizes the gluon contribution to the nucleon mass, \(f_{Tq}^n\) are the contributions from the light quark species \(u,d\) and \(s\), and \(v =246\) GeV is the electroweak vacuum expectation value. We use the default values in `MicrOmegas` which lead to \(\sum_q f_q^n \simeq 0.28\). As we showed in the previous section, the mass splitting necessary to reproduce the thermal relic density is much smaller than the top mass independently of whether the NLSP is the stop or the sbottom. In this case, taking advantage of the expressions for the couplings Eqs. ([\[ht1t1\]](#ht1t1){reference-type="ref" reference="ht1t1"})--([\[bLx\]](#bLx){reference-type="ref" reference="bLx"}), and using the relevant limit for the loop integrals, the various contributions to the amplitudes are well approximated by \[\begin{aligned}
\mathcal{A}_t^h&\simeq& \frac{s_{2t}}{2}\left(1-\frac{m_{\tilde{t}_1}^2}{m_{\tilde{t}_2}^2}\right)+\left(s_t^2+\frac{c_t^2}{16}\right) \frac{m_t}{m_{\tilde{t}_1}}+\left(c_t^2 +\frac{s_t^2}{16}\right)\frac{m_t m_{\tilde{t}_1}}{m_{\tilde{t}_2}^2} \;,\label{eq:t}\\ \mathcal{A}^h_{\tilde{t}_{1,1}} &\simeq & \frac{m_t}{m_{\tilde{t}_1}}\left( 1-\frac{X_t^2}{m_{\tilde{t}_2}^2-m_{\tilde{t}_1}^2} \right)\left[ 2 \left(s_t^2+\frac{c_t^2}{16}\right)-s_{2t} \frac{m_t}{m_{\tilde{t}_1}}\right] \;, \label{eq:t1}\\ \mathcal{A}^h_{\tilde{t}_{1,2}} &\simeq &-\frac{X_t}{m_{\tilde{t}_2}} c_{2t}\left(-\frac{15}{16}s_{2t}+c_{2t} \frac{m_t}{m_{\tilde{t}_1}} \right) \;, \label{eq:t1t2}\\ \mathcal{A}^h_{\tilde{t}_{2,2}} &\simeq & \frac{m_t m_{\tilde{t}_1}}{m_{\tilde{t}_2}^2}\left( 1+\frac{X_t^2}{m_{\tilde{t}_2}^2-m_{\tilde{t}_1}^2} \right)\left[2 \left(c_t^2 +\frac{s_t^2}{16}\right)+ s_{2t} \frac{m_t}{m_{\tilde{t}_1}}\right]\;.\label{eq:t2} \end{aligned}\] Unless \(X_t \gg m_{\tilde{t}_{1,2}}\), the above can be seen to be at most \(\mathcal{O}(1)\). Taking into account these potentially important contributions, we can write down how the direct detection cross section scales (\(m_\chi \sim m_{\tilde{t}_1}\sim m_{\tilde{b}_1}\)) \[\label{e:SIapprox} \sigma_{SI} \sim 4 \times 10^{-48} \mathrm{cm^2}\left[ 0.15\left( \frac{2~ \mathrm{TeV}}{m_\chi}\right)\left(\frac{m_b +20~ \mathrm{GeV}}{m_b+\Delta m}\right)^2 + \mathcal{A}^h_{t\tilde{t}} \right]^2,\] where \(\Delta m\) is the splitting with the sbottom. It should be noted that this expression will breakdown at smaller cross sections, in part, due to the omission of stop loop contributions to the gluon amplitude. We now discuss how \(\mathcal{A}^h_{t\tilde{t}}\) behaves in various limits to gain further insight into the expected size of the direct detection cross section. Naively, it appears that the cross section may be driven arbitrarily large for maximally mixed stops, \(s_{2t} \simeq1\) by taking \(X_t\) large (hence \(m_{\tilde{t}_2} \gg m_{\tilde{t}_1}\)), where the \(\tilde{t}_1\) contribution completely dominates \(\mathcal{A}^h_{t\tilde{t}}\) giving \(\mathcal{A}^h_{t\tilde{t}} \sim X_t/ 2m_{\tilde{t}_1}\). However, constraints both from vacuum stability and the Higgs mass prevent going too far into this regime. Once we impose the Higgs mass constraint it is possible to make detailed statements about the Higgs mediated amplitude. As can be seen from Fig. [\[fig:Higgs\]](#fig:Higgs){reference-type="ref" reference="fig:Higgs"}, for any given mixing angle, generally the largest values of \(X_t\) or \(m_{\tilde{t}_2}\) consistent with the Higgs boson mass are close to the largest values allowed by the stability of the vacuum. It is useful to consider different mixing angles separately. First, we examine the case of significantly mixed stops. Here, the proper Higgs boson mass is obtained for values of \(m_{\tilde{t}_2}\) not too much heavier than \(m_{\tilde{t}_1}\) and approximately \(|X_t| \sim 2~ m_{\tilde{t}_2}\). The contribution to the amplitude from the stops can then be approximated as \(\mathcal{A}^h_{t\tilde{t}} \sim m_t/ m_{\tilde{t}_1}\left[ 3-X_t/m_{\tilde{t}_2}\left(2+X_t/m_{\tilde{t}_2} \right) \right]\sim 0.26\) for positive \(s_t\) (\(X_t<0\)) and \(m_{\tilde{t}_1}\) = 2 TeV. For negative \(s_t\) (\(X_t >0\)), this contribution can be approximately a factor of 2 larger in magnitude than the \(s_t >0\) case, but with a negative sign. This negative sign yields destructive interference with the subdominant gluon contribution. We now turn to the case of small, but finite mixing, (\(s_t\sim1\) or 0). In this case, a consistent Higgs mass is obtained when \(m_{\tilde{t}_1} \ll m_{\tilde{t}_2}\) and \(|X_t| \sim m_{\tilde{t}_2}\). For a predominantly right-handed light stop, we find \(\mathcal{A}^h_{t\tilde{t}} \sim m_t/m_{\tilde{t}_1} \left[ 3-X_t/m_{\tilde{t}_2}\left(1+2X_t/m_{\tilde{t}_2} \right) +X_t m_{\tilde{t}_1}/m_{\tilde{t}_2}^2\left(-1+2X_t/m_{\tilde{t}_2} \right) \right]\sim 0.35\) for \(m_{\tilde{t}_1}=2\) TeV, \(m_{\tilde{t}_2}=3\) TeV and positive \(s_t~(X_t<0)\). As can be seen from Eqs. ([\[eq:t\]](#eq:t){reference-type="ref" reference="eq:t"})-([\[eq:t2\]](#eq:t2){reference-type="ref" reference="eq:t2"}), most of the contributions to the direct detection cross section for a purely left-handed stop, close to \(c_t\sim1\), are suppressed by factors of 1/16 compared to the right-handed approximation owing to the smaller hypercharge of the left-handed multiplet. However, there is an unsuppressed contribution from \(\mathcal{A}^h_{\tilde{t}_{1,2}}\), and as long as the mixing angle is not too small, it tends to dominate. It leads to an approximate amplitude \(\mathcal{A}^h_{t\tilde{t}} \sim-m_t X_t/m_{\tilde{t}_1} m_{\tilde{t}_2}\sim 0.09\) for negative \(X_t\) and \(m_{\tilde{t}_1}=\) 2 TeV. Given these considerations, and demanding consistency with the Higgs mass, it is clear that unless the mass splitting between \(\chi\) and \(\tilde{b}_L\) is \(\sim m_b\), which happens for \(m_\chi\gtrsim 2.5\) TeV, the direct detection cross section for a predominantly left-handed stop is much smaller than a right-handed stop. Due to the destructive interference of the stop contributions with the gluon contribution for \(s_t<0\), the maximal cross section, in the absence of very small mass-splitting with the sbottom, consistent with the Higgs mass is obtained for a dominantly right-handed light stop with \(s_t>0\). This maximal value can be roughly approximated by \(\sigma_{SI} \sim 10^{-49}\left( 1+ 3 \tilde{t}_1/2\tilde{t}_2\right)^2\left(2 ~\mathrm{TeV}/ \tilde{t}_1\right)^2\mathrm{cm}^2\), a value well below the neutrino floor. Finally, a comment regarding the direct detection prospects for a mixed \(\tilde{b}\) co-annihilation scenario is in order. We have set \(X_{b}\) identically zero throughout, in part for simplicity, and in part because a mixing in the sbottom sector is not motivated by the Higgs boson mass. Relaxing this assumption would allow the lightest sbottom to be an admixture of the right and left-handed states. In cases where the sbottom is the co-NLSP, this enhances the direct detection cross section with respect to Eq. ([\[e:SIapprox\]](#e:SIapprox){reference-type="ref" reference="e:SIapprox"}) due to the larger hypercharge of the \(\tilde{b}_R\). To get a feel for the maximum size of this effect, we consider the limiting case of a pure \(\tilde{b}_R\) NLSP. The \(\tilde{b}_L\) and the entire stop sector can be substantially heavier than the NLSP. In this case the thermal expectation for \(\Delta m\) found in Ref. is rather similar to the expectation for \((\tilde{f} \tilde{f}^* \rightarrow g g)\) shown in Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"}. Note that this is substantially smaller than the expectation for a purely left-handed sbottom NLSP and hence will lead to significant enhancement of the gluon box contribution, Eq. ([\[e:glubox\]](#e:glubox){reference-type="ref" reference="e:glubox"}). The corresponding direct detection cross section has recently been studied in Ref. , and for a thermal mass splitting, as read from our Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"}, their results indicate an expected scattering rate above the neutrino floor throughout the allowed parameter space. However, LZ is only expected to be sensitive to the very smallest mass splittings, corresponding to 1300 GeV \(\mathop{}_{\textstyle \sim}^{\textstyle <} m_{\chi} \mathop{}_{\textstyle \sim}^{\textstyle <}\) 1700 GeV.
# Conclusion {#s.conc}
We have revisited the stop co-annihilation scenario wherein dark matter freeze-out, stop-induced corrections to the Higgs boson mass as well as an irreducible direct detection cross section mediated by stop/sbottom loops are interrelated. After accounting for collider limits on light stops and sbottoms we find that the constraints on the stop sector from the stability of the vaccuum are more stringent than EWPOs and the Higgs signal strength measurements. The region of stop parameters consistent with the Higgs mass is generally unconstrained by any of the above listed indirect constraints, however there may be slight tension between the largest values of \(X_t\) consistent with the Higgs mass and the stability of the vacuum. We analyze the irreducible, stop/sbottom-loop induced direct detection cross section in detail taking into account the mass splittings motivated by cosmology. Unfortunately, we find that this minimal direct detection rate generally falls well below the neutrino floor, particularly once the Higgs mass is taken into account. It should be kept in mind, however, that even a small amount of bino-Higgsino mixing can induce substantial spin-independent neutralino-nucleus scattering at tree-level. In this case the rate could be large, even going up to the current limit, but the direct connection between the cosmological origin of dark matter in the early Universe and direct detection rates would be lost. In light of the potentially challenging direct detection situation, collider searches may be the best available probes for this scenario. The mass splitting between the \(\tilde{t}\) (or \(\tilde{b})\) and the neutralino is small in the co-annihilation region, and naively, absent a detailed model of supersymmetry breaking it is not clear why this should be so. However, given that it produces the proper thermal dark matter relic density consistent with increasingly strong direct detection bounds, this mass spectrum is extremely well motivated. Therefore the region in the \(\Delta m\) vs. \(m_{\chi}\) plane shown in Fig. [\[f:deltam\]](#f:deltam){reference-type="ref" reference="f:deltam"} represents an important target. Dedicated LHC searches for such compressed spectra will be crucial in testing the co-annihilation scenario. | {'timestamp': '2018-01-23T02:13:16', 'yymm': '1706', 'arxiv_id': '1706.01911', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01911'} |
null | null |
# Introduction
Our knowledge of how baryons were accreted by galaxies such as the Milky Way is limited however but unprecedented amounts of data are becoming available from large scale surveys. It is vital that we use these data to improve our understanding of the local group and an excellent starting point is to find the current distribution of matter in the Galaxy. We cannot observe directly the distribution of dark matter so we must use the observable components as tracers to build dynamical models from which we can discover the distribution of dark matter. If we assume the Galaxy is in statistical equilibrium, we can exploit Jeans theorem () and presume that the distribution function () \(f(x,v)\) is a function of integrals of motion \(I(x,v)\) only. There is an infinite choice of integrals because any function of the integrals is also an integral but the best choices are the actions \(J_i\), which can be uniquely complemented by canonically conjugate variables to make a complete set (\(\theta\),J) for phase-space coordinates.
# Modelling Process
The Milky Way can be built up from its components; the dark halo, thin and thick discs, stellar halo, bulge and gas disc. Each of these components is modelled either by a distribution function, \(f(\textbf{J})\) (first three components above), or a fixed potential. By making a sensible estimate for the parameters of these components we can make an estimate for the total potential. We then use the Stäckel Fudge ([@Binney2012; @Binney2014 Binney 2012,2014]) to find the actions and by integrating over velocity: \[\rho(x) = \int d^3v f(\textbf{J}(\textbf{x},\textbf{v}))\] we estimate the density. We then solve Poisson's equation to find the new potential and iterate this procedure until after a few iterations the total potential converges.
## Disc
The of the discs in our models is superposition of the "quasi-isothermal" components introduced by [@Binney2011 Binney & McMillan (2011)]. It has the form \[\label{eq:qi} f(J_r,J_z,J_\phi)=f_{\sigma_r}(J_r,J_\phi)f_{\sigma_z}(J_z,J_\phi),\] where \(f_{\sigma_r}\) and \(f_{\sigma_z}\) are \[\label{planeDF} f_{\sigma_r}(J_r,J_\phi)\equiv \frac{\Omega\Sigma}{\pi\sigma_r^2\kappa} [1+\tanh(J_\phi/L_0)]\e^{-\kappa J_r/\sigma_r^2}, f_{\sigma_z}(J_z,J_\phi)\equiv\frac{\nu}{2\pi\sigma_z^2}\,\e^{-\nu J_z/\sigma_z^2}.\] The thick disc is represented by a single quasi-isothermal , while the thin disc's is built up with a quasi-isothermal for each coeval cohort of stars. The velocity-dispersion parameters depend on \(J_\phi\) and the age of the cohort. The star-formation rate in the thin disc decreases exponentially with time, with characteristic time scale \(t_0=8\Gyr\). In addition a parameter \(F_{\rm thk}\) controls the fraction of mass contributed by the thick disc.
## Dark halo
Our dark halo is based on the form introduced by [@Posti2015 Posti et al. (2015)] which in isolation self-consistently generates a density distribution which has a NFW ([@Navarro1997 Navarro, Frenk & White 1997]) profile (equation [\[eq:Posti\]](#eq:Posti){reference-type="ref" reference="eq:Posti"}). This is a function of \(h(\textbf{J})\) which is an almost linear function (a homogeneous function of degree unity) of the actions \(J_i\). The haloes generated are isotropic centrally and mildly radial when \(r>r_s\). The anisotropy can be changed by varying the linear function of the \(J_i\). These haloes closely resemble the haloes formed in dark-matter-only simulations. Specifically \[\label{eq:Posti} \fP(\vJ)=\frac{N}{J_0^3}\frac{(1+J_0/h)^{5/3}}{(1+h/J_0)^{2.9}}\] The scale action \(J_0\) encodes the scale radius around which the slope of the radial density profile shifts from \(-1\) at small radii to \(-3\) far out. From equation ([\[eq:Posti\]](#eq:Posti){reference-type="ref" reference="eq:Posti"}) it follows that \(f_P(\vJ)\sim |\vJ|^{-5/3}\) as \(|\vJ|\rightarrow0\).
## Observational constraints
We use several sets of observations to constrain the parameters of the s at various stages of the modelling. The observations include the astrometry of H\(_2\)O and SiO masing stars ([@Reid2014 Reid et al. 2014]), the distribution of radio-frequency lines of and CO emission in the longitude-velocity plane, the stellar parameters and distance estimates in the fourth RAVE data release ([@Kordopatis2013 Kordopatis et al. 2013]) (see Section [4](#sec:kine){reference-type="ref" reference="sec:kine"}) and finally the vertical density profile from SDSS. We assume that the population from which the RAVE sample is drawn is identical to that studied by.
# Results
used RAVE data and SDSS Juric (2008) data to constrain the mass of DM within solar radius, R\(_0\). The dark halo was included as a potential not a . [@Binney2015 Binney & Piffl (2015)] used the in equation [\[eq:Posti\]](#eq:Posti){reference-type="ref" reference="eq:Posti"} for the dark halo in their self-consistent model of the Milky Way. In this model the halo was in self-consistent equilibrium with the other components so it had been adiabatically compressed by the baryons from its original NFW form. Hence it mimicked a scenario where baryons accumulated quiescently in the Galaxy's dark halo. An NFW halo becomes more centrally concentrated, and with so much dark matter at low radii (Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}) the matching disc contains too few stars to satisfy the microlensing data ([@Popowski2005 Popowski et al. 2005]). This implies that the infinite phase-space density of particles at \(J=0\) characteristic of an NFW does not survive the accretion of baryons. The baryons cannot have accumulated entirely adiabatically but the most tightly bound dark matter particles were upscattered. This scattering of DM particles by baryons reduces their phase-space density and had greatest impact near \(J=0\). In order to model this scattering process [@Cole2017 Cole & Binney (2017)] modified the NFW by setting \(f(\textbf{J})=g(h)f_{NFW}(h)\) with \(g\sim{h}^{5/3}\) for small \(h\) and \(g\sim1\) at large \(h\). This shifts particles from very low \(J\) to higher \(J\). The functional form for \(g\) is \[\label{eq:defsg} g(h)=\left[{h_0^2\over h^2}-\beta{h_0\over h}+1\right]^{-5/6},\] where \(h_0\) is an arbitrary constant with the dimensions of action that sets the scale of the almost constant-density core of the and \(\beta\) is determined by requiring the total mass of dark matter is conserved. The free parameters in \(\fDM\) are \(J_0\), which sets the NFW scale radius, \(h_0\), which sets the size of the dark halo's core, and the normalisation \(N\). found (see Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}) that with this modified dark halo their models of the Milky Way have a similarly good fit to the observations but the central regions are now dominated by baryons. The dark matter fraction is reduced and consistent with the results of surveys of microlensing events. In summary they found the local DM density \(\rho_{DM}{\gtrsim}0.012\) \(M_\odot\) pc\(^{-3}\), stellar disc scale radius \(R_d\sim2.9\) kpc and stellar disc mass M\(_d\gtrsim4\times10^{10}\) \(M_\odot\).
# Using RAVE/TGAS {#sec:kine}
Our current modelling uses RAVE data in 8 spatial bins at \(R_0\pm\)`<!-- -->`{=html}1 kpc and at 0, 0.3, 0.6, 1.0, 1.5 kpc in \(\left|z\right|\) to compute velocity distributions predicted by the at the mean positions. In order to take advantage of Gaia DR1 ([@Gaia2016 Gaia Collaboration 2016]) we are developing a method using the RAVE/TGAS observations (). In order to do this we will develop a full selection function \(S(s)\) for this sample. Deriving an a priori calculation of \(S(s)\) needs a full chemodynamical model of the MW disc \(S(s,τ,[Fe/H])\) such as used in and in addition we need to model the exact distribution of stars in age and metallicity in the solar neighbourhood. [@Schönrich2017 Schönrich and Aumer 2017] did this using the RAVE selection function of and found that at fixed metallicity \(S(s,\tau)\) falls off exponentially with scale 0.12 kpc at \(s>0.2\) kpc. The selection function for TGAS is biased towards younger stars which are more likely to be seen so the kinematics will appear cooler than they really are. Our models need to take this into account. They already have age but not metallicity and so we can add metallicity by use of a suitable metallicity . Then we can compute the likelihoods of our model based on the resulting selection function.
# Action based modelling software library
Our modelling is currently being rewritten using AGAMA (Action-based Galaxy Modelling Architecture) which is a library of low-level programs containing interfaces and generic routines required to create the functions described here. The main sets of functions include gravitational potential and density interfaces, action/angle interface, interface for creating gravitationally self-consistent multicomponent galaxy models etc. The code can be downloaded from https://github.com/GalacticDynamics-Oxford/Agama | {'timestamp': '2017-06-07T02:04:46', 'yymm': '1706', 'arxiv_id': '1706.01672', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01672'} |
# Introduction
<https://doi.org/10.1103/PhysRevB.95.174440> Understanding and manipulating magnetic anisotropy is important in both basic and applied physics research. For instance, anisotropic magnetism is essential for the development of permanent magnets, and it appears to be a fundamental component in many systems that display unconventional superconductivity. Recently, research on materials that are crystallographically layered (held together by van der Waals bonds) has gained prominence in condensed matter physics. This is largely driven by the now-realized prospect of building van der Waals heterostructures from materials with complementary properties. Two-dimensional and quasi-2D materials are of fundamental interest from a bulk perspective as well, in part because they exist in the limit of very anisotropic interactions (strong in-plane interactions and weak cross-plane interactions). For instance, the compounds CrSiTe\(_3\), MnPS\(_3\), and CrI\(_3\) have layers connected by van der Waals bonds, and demonstrate bulk magnetic ordering with anisotropic interactions manifesting themselves in anisotropic properties, suppressed 3D ordering temperatures, two-dimensional order, and persistent short-range correlations above \(T_{\rm C}\)İn this work, we probe the behavior of Mn\(_3\)Si\(_2\)Te\(_6\) , which can be considered as a three-dimensional analogue of CrSiTe\(_3\). While its crystal structure is interesting, very little research has been reported on Mn\(_3\)Si\(_2\)Te\(_6\)Ṫhe initial report of its existence and physical properties discussed the magnetism within the framework of an incorrect stoichiometry and structure, and thus the initial hypothesis of ferrimagnetism could not be fully evaluated. Despite containing nominally Mn\(^{2+}\) with no orbital moment (3\(d^5\) with S=5/2, L=0), the magnetism displays a large anistropy field on the order of 10 T at 5 K. The crystal structure was reported several years after the initial characterization, but the magnetic properties were not revisited. Mn\(_3\)Si\(_2\)Te\(_6\) has a trigonal crystal structure (space group No. 163) that is shown in Fig. [\[Structure\]](#Structure){reference-type="ref" reference="Structure"}. Fig. [\[Structure\]](#Structure){reference-type="ref" reference="Structure"} also contains the magnetic structure obtained from our neutron diffraction data. Mn\(_3\)Si\(_2\)Te\(_6\) is composed of MnTe\(_6\) octahedra that are edge-sharing within the \(ab\) plane (Mn1 site), and along with Si-Si dimers this creates layers of Mn\(_2\)Si\(_2\)Te\(_6\). The layered framework is analogous to that of CrSiTe\(_3\), which is hexagonal and has a van der Waals gap between the layers. In Mn\(_3\)Si\(_2\)Te\(_6\) , however, the layers are linked by the filling of one-third of the octahedral holes within the van der Waals gap by Mn atoms at the Mn2 site, yielding a composition of Mn\(_3\)Si\(_2\)Te\(_6\)İmportantly, the multiplicity of Mn1 is twice that of Mn2. In this work, Mn\(_3\)Si\(_2\)Te\(_6\) was characterized using magnetization, specific heat and electrical resistivity measurements, as well as via powder x-ray and single crystal neutron diffraction. We find Mn\(_3\)Si\(_2\)Te\(_6\) to be a ferrimagnet due to antiparallel alignment of moments on the Mn1 and Mn2 atomic positions, the different multiplicity of which yields a net magnetization with moments preferring to lie in the trigonal plane. The data, including diffuse neutron scattering, reveal the existence of strong spin correlations well-above \(T_{\rm C}\) , which may be associated with short range order or the persistence of correlated excitations in the paramagnetic region. The experimental probes were complemented by first principles calculations that revealed a competition between antiferromagnetic (AFM) exchange interactions that are frustrated with respect to each other. Interestingly, the longer-range, third-nearest neighbor coupling (Mn1-Mn2 pairs at 5.41Å) dominates over the in-plane coupling of second-nearest neighbors at 4.06Å (Mn1-Mn1 pairs). The net result is an antiferromagnetic configuration close in energy to the ferrimagnetic one, and this competition for the ground state leads to a relative suppression of the ordering temperature.
# Experimental Details
Mn\(_3\)Si\(_2\)Te\(_6\) single crystals were obtained by melting a stoichiometric mixture of the elements in a vacuum sealed quartz ampoule at \(T\)=1125 K. High-purity Si lump (Alfa Aesar 99.9999%), and Te shot (Alfa Aesar, 99.999%) were combined with Mn granules (99.98% Alfa Aesar) that were arc-melted prior to use. Single crystals of Mn\(_3\)Si\(_2\)Te\(_6\) were mechanically isolated from the as-grown ingot. Field-cooled magnetization measurements were performed in a Quantum Design Magnetic Property Measurement System. AC magnetic susceptibility, electrical transport, and specific heat capacity measurements were performed in a Quantum Design Physical Property Measurement System. Powder x-ray diffraction measurements were performed using a PANalytical X'Pert Pro MPD with a Cu K\(_{\alpha,1}\) incident-beam monochromator. An Oxford PheniX Cryostat was utilized to obtain data between 20 K and 300 K. Additional powder x-ray diffraction measurements were performed between 300 and 400 K on a PANalytical X'Pert diffractometer with Cu K\(_{\alpha}\) radiation and an XRK900 oven-type furnace. The furnace chamber was purged with He for 12 hours prior to initiating the high-temperature measurements and gas was continuously flowed throughout the measurements. The measurements were delayed at each temperature for 30 minutes to allow the sample to equilibrate. At 300 K, Rietveld refinement of data from the low \(T\) stage yielded \(a\)=7.0321(6)Å and \(c\)=14.249(1)Å, while refinement of the data from the high \(T\) stage yielded \(a\) =7.0343(2)Å and \(c\) =14.249(2)Å. Single crystal neutron diffraction data were measured using the HB3A diffractometer at the High Flux Isotope Reactor at Oak Ridge National Laboratory (ORNL). The neutron wavelength of 1.546 Å from the bent perfect Si-220 monochromator was used for the data collections, and 380 reflections were measured at 4, 100, and 380 K. The nuclear-only refinement at 380 K yielded occupancies equal to unity within the error bars, and thus the occupancies were fixed at unity for all \(T\). The magnetic refinement at 4 K required three equivalent magnetic domains. All refinements were performed using the program FullProf. Diffuse neutron scattering measurements were performed on the CORELLI diffuse scattering spectrometer at the Spallation Neutron Source located at ORNL. CORELLI is a time-of-flight spectrometer with a pseudo-statistical chopper, which generates both an elastic scattering signal (with an average energy resolution \(\approx\)`<!-- -->`{=html}1 meV) and a total scattering signal from a single measurement. An approximately 7 mg single crystal was attached to an Al plate in a manner to facilitate inspection of the HHL scattering plane. The measurements were taken at temperatures of 6, 120, and 350 K, where the sample was rotated in steps of 3\(^{\circ}\) over ranges of 120\(^{\circ}\) (6 K), 120\(^{\circ}\) (120 K) and 90\(^{\circ}\) (350 K). A significant amount of time was spent at 120 K measuring 30 degrees centered around the 002 Bragg reflection. Data around 002 were also collected on warming from 6 K to 120 K. Mantid was utilized to perform the Lorentz and spectrum corrections and merge the full volume of the scattering data. First principles calculations were performed using the linearized augmented plane-wave (LAPW) code WIEN2K, within the generalized gradient approximation. Spin-orbit coupling was only included during the calculation of magnetic anisotropy within the observed ground state. Sphere radii of 2.04, 2.5 and 2.5 Bohr were chosen for Si, Mn and Te, respectively, with an RK\(_{max}\) of 7.0 (here RK\(_{max}\) is the product of the smallest sphere radius and the largest planewave expansion wavevector). All calculations used the experimental, room-temperature lattice parameters of \(a\)=7.029Å and \(c\)=14.255Å taken from Ref. . For each magnetic configuration, the internal coordinates were relaxed within the magnetic ordering pattern until the forces were less than 2 mRyd/Bohr. Optimization within the magnetic state yields significantly different atomic coordinates than optimization within the non-magnetic state (an energy gain of some 250 meV/Mn results), suggesting a strong coupling of magnetism to the lattice. The structure theoretically optimized within the ferrimagnetic ground state is much closer to the experimental structure than that optimized within the non-magnetic state. For example, the value for Mn1 of \(\Delta\) z \(\equiv\) z\(_{Exp.}\)-z\(_{Calc}\) is-8.7 \(\times\) \(10^{-4}\) for the ferrimagnetic ground state but nearly 5 times as large in magnitude at 4.02 \(\times\) 10\(^{-3}\) for the non-magnetic state. Similarly large ratios apply for the Si z and Te y and z coordinates, with the Te x coordinate \(\Delta\)x roughly twice as large for the non-magnetic state as for the ferrimagnetic state. Additional details are provided in the Supplemental Information.
# Results and Discussion
## Magnetization, neutron diffraction and theoretical calculations
The magnetization \(M\) data for single crystalline Mn\(_3\)Si\(_2\)Te\(_6\) are shown in Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a. Mn\(_3\)Si\(_2\)Te\(_6\) is observed to have a Curie temperature of \(T_{\rm C}\) =78 K. The ordering temperature was also confirmed by AC magnetic susceptibility measurements, the results of which are shown in Fig. [\[ACMag\]](#ACMag){reference-type="ref" reference="ACMag"}. The anisotropy of \(M\) below \(T_{\rm C}\) suggests the ordered moments lie primarily within the \(ab\)-plane. This is demonstrated by a significantly larger \(M\) when the applied field \(H\) lies within the \(ab\)-plane as compared to when \(H || c\); note the different vertical axes in the main panel of Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a. Consistent with this, below \(T_{\rm C}\) the magnetization saturates rapidly for \(H || ab\) and reaches \(\approx\)`<!-- -->`{=html}1.6 \(\mu_B\)/Mn at \(T\)=5 K (inset Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a). The data imply an anisotropy field of approximately 9 T (90 kOe), and a small ferromagnetic component is observed for \(H || c\). These data are consistent with the initial report on Mn\(_3\)Si\(_2\)Te\(_6\)Ẇe note there is no remanant moment for either orientation, which is different from the prior report and, to some extent, suggests the present single crystals are of high quality. Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}b plots the magnetization data as \(H/M\), which is equivalent to \(1/\chi\) when the susceptibility can be defined as \(\chi=M/H\) (when \(M\) is linear in \(H\)). The data demonstrate Curie-Weiss behavior between approximately 350 and 750 K, the region where 1/\(\chi\) is linear in \(T\). The data above 400 K were fit to a simple Curie-Weiss law, \(\chi\)= C/(\(T-\Theta\)), where C is the Curie constant and \(\Theta\) the Weiss temperature. This fitting produced an effective moment of 5.6\(\mu_B\)/Mn and a Weiss temperature of-277 K. The effective moment is consistent with the presence of Mn\(^{2+}\) ions and the negative Weiss temperature indicates antiferromagnetic correlations. The saturation magnetization (inset, Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a) is about one-third of that expected on the basis of this effective moment, which suggests the presence of one uncompensated Mn\(^{2+}\) local moment per formula unit in the ordered phase. In Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}b, the low-temperature data were collected on a large mass of ground crystals, while high-temperature data were collected using six small crystals sealed under vacuum in a thin quartz tube (the crystals are free to rotate in an applied field). The AC susceptibility \(\chi_{AC}\) = d\(M\)/d\(H\) can be described by in-phase (real) \(\chi\)' and out-of-phase (imaginary) \(\chi\)" components, which are shown in Fig. [\[ACMag\]](#ACMag){reference-type="ref" reference="ACMag"}. The out-of-phase component \(\chi\)" relates to dissipative losses, for instance the movement of domain walls in ferromagnets; \(\chi\)" is typically zero for simple antiferromagnets and non-zero in ferromagnets and sometimes in metals. The AC data shown in Fig. [\[ACMag\]](#ACMag){reference-type="ref" reference="ACMag"} were collected in zero applied DC field, with a small amplitude \(A\)=2 Oe and a frequency \(f\)=997 Hz. These data reveal a sharp onset at the Curie temperature, with \(\chi\)' and \(\chi\)" increasing below \(T_{\rm C}\) for both orientations. In addition, the data reveal anisotropy above \(T_{\rm C}\) , which was also observed with the DC measurements. A comparison of \(\chi\)' and the DC analog \(M/H\) (\(H\)=20 Oe) is shown in the Supplemental Information. As shown in Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}b, the low-field magnetization data on ground crystals deviate from the high \(T\) Curie-Weiss behavior below \(\approx\) 330 K. Single crystal magnetization data reported in Fig. [\[Short2\]](#Short2){reference-type="ref" reference="Short2"} demonstrate that this behavior is associated with a relatively abrupt onset of magnetic anisotropy below \(\approx\) 330 K, which is observed in both the low-field DC data (Fig. [\[Short2\]](#Short2){reference-type="ref" reference="Short2"}a) and the AC data (Fig. [\[Short2\]](#Short2){reference-type="ref" reference="Short2"}b). The anisotropy is the same as that in the ferrimagnetic phase below \(T_{\rm C}\) (easy-plane anisotropy). Upon cooling below 330 K, \(\chi\)" becomes non-zero for in-plane data, and it reaches a maximum near 110 K before rising sharply at \(T_{\rm C}\)Ṫhis may suggest the anisotropy is associated with, or derived from, a ferromagnetic component (uncompensated moment) that is present for \(H \perp c\). Indeed, isothermal magnetization data reveal a small (soft) ferromagnetic contribution for \(H \perp c\) in the region \(T_{\rm C}\) \(< T <\)`<!-- -->`{=html}330 K, which increases upon cooling. The low-field \(M(H)\) data at \(T\)=110 K are shown in Fig. [\[MH110\]](#MH110){reference-type="ref" reference="MH110"}, and the corresponding AC data are shown in the Supplemental along with \(M(H)\) curves at different \(T\). Interestingly, \(\chi\)' plateaus near 85-110 K, which would suggest that this ferromagnetic component has saturated. That is, this plateau in \(\chi\)' seems to suggest that the anisotropy does not originate in precursory short-range order of the ferrimagnetic phase, which would be expected to continue increasing all the way down to \(T_{\rm C}\) as the correlation length begins to diverge. However, the ferromagnetic-like contribution suggests the anisotropy originates from some form of magnetic order, as opposed to a crystal-field induced anisotropy of a paramagnet. As discussed below, our diffraction measurements did not reveal any significant structural change near 330 K, but a structural contribution and/or origin cannot be ruled out at this time. The true nature of the transition near 330 K is thus unclear and warrants further investigation. As discussed in the Supplemental Information, this behavior seems to be intrinsic to our single crystals. Single crystal neutron diffraction and spectroscopy were utilized to examine the nuclear and magnetic structures of Mn\(_3\)Si\(_2\)Te\(_6\)Ȧs shown in Fig. [\[ND1\]](#ND1){reference-type="ref" reference="ND1"}, the intensity of the 002 Bragg reflection increases upon cooling through \(T_{\rm C}\)Ṫhis suggests a strong in-plane component of the moment, consistent with the magnetization data. The lack of additional scattering at the 110 reflections at low \(T\) is consistent with the moment being in the \(ab\)-plane, but this observation alone is not definitive given the presence of three magnetic domains. The square root of the integrated magnetic intensity is proportional to the ordered moment, and can thus be used as a magnetic order parameter. Order parameter data employing intensity from the 002 reflection were obtained using both HB3A and CORELLI, as shown in Fig. [\[ND1\]](#ND1){reference-type="ref" reference="ND1"}b where fits to power law behavior are included. The data in Fig. [\[ND1\]](#ND1){reference-type="ref" reference="ND1"}b from CORELLI are the integrated intensity of 002, while that from HB3A were constant Q data collections; neither data set are corrected for background. Additional temperature-dependent data from HB3A are shown in the Supplemental Information. The order parameter fits shown as solid curves in Fig. [\[ND1\]](#ND1){reference-type="ref" reference="ND1"}b yield a critical exponent \(\beta\) \(\approx\) 0.25 and \(T_{\rm C}\) values consistent with that obtained from magnetization measurements, though some minor discrepancies are observed because the data were collected while ramping temperature in different conditions. The fits employ a simple power law, \((I-I_0)^{0.5}\)=\(A\)(\(T_{\rm C}\) -\(T\))\(^{\beta}\), where \(I_0\) represents the non-magnetic intensity (background and nuclear, assumed temperature independent). Data from just below \(T_{\rm C}\) down to \(\approx\)`<!-- -->`{=html}40 K are fit, and the resulting curves are extended to lower \(T\) in Fig. [\[ND1\]](#ND1){reference-type="ref" reference="ND1"}b for comparison. For the CORELLI data we obtain \(\beta\)=0.26(1) and \(T_{\rm C}\) =81.7(2) K while for the HB3A data we obtain \(\beta\)=0.23(1) and \(T_{\rm C}\) =78.5(1) K. The standard deviations originate from the fitting procedures, but the absolute errors on critical exponents estimated in this way are expected to be larger. The inset contains these same results in log-log representation to emphasize the similar temperature-dependence for the two data sets, and to highlight the good quality of the fits. The critical exponent of \(\beta\) \(\approx\) 0.25 for Mn\(_3\)Si\(_2\)Te\(_6\) can be compared to those obtained for simple mean field models. For instance, a 2D Ising system is expected to have \(\beta\)=0.125, while the 3D Ising and Heisenberg models have \(\beta\)=0.326 and 0.367, respectively. In CrSiTe\(_3\), a critical exponent of \(\beta\)=0.12 was reported, suggesting 2D character of the magnetism in that van der Waals bonded ferromagnet. Mn\(_3\)Si\(_2\)Te\(_6\) is a crystallographically three-dimensional material. The deviation of \(\beta\) from the simple 3D models is not surprising, however, given the existence of multiple Mn sites and ferrimagnetic order. Refinement of the 4 K neutron diffraction data from HB3A yielded the magnetic structure shown in Fig. [\[Structure\]](#Structure){reference-type="ref" reference="Structure"} (\(R_f=4.6\) and \(\chi^2=3.6\)). This spin structure is defined by parallel alignment of the moments on Mn1 combined with anti-parallel alignment to the moments on Mn2. Within the limits of these data, the refined moments are the same on both Mn sites, consistent with the expectation of similar oxidation states (both positions are octahedrally-coordinated by Te). At 4 K, the refined moment is 4.2(1)\(\mu_B\)/Mn when the moments are constrained to be equal and to lie within the trigonal plane. When permitted, a small moment of 0.2(2)\(\mu_B\)/Mn along the \(c\) axis is observed, but the quality of the refinement does not improve. However, the isothermal magnetization data \(M(H)\) at 5 K suggest a soft ferromagnetic component for \(H || c\) (inset Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a), and the low-field \(M(T)\) data are anisotropic below \(T_{\rm C}\) (Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}a). It is possible that an applied field causes alignment of canted moments or domains, and the lack of an applied field during neutron diffraction leads to an average canting that is negligible. Note that in Fig. [\[Structure\]](#Structure){reference-type="ref" reference="Structure"}, the moments are shown to be along the \(a\) axis, but their specific orientation within the \(ab\)-plane cannot be determined due to symmetry. In trying to understand this magnetic configuration, we first consider the crystal structure and apparent exchange interactions. Mn1-Mn2 have an interatomic distance of 3.55 Å at room temperature, and are linked through face-sharing octahedra along the \(c\) axis (Fig.[\[Structure\]](#Structure){reference-type="ref" reference="Structure"}), which would lead to a direct exchange interaction that we expect to be antiferromagnetic (AFM) for these Mn\(^{2+}\) ions. The Mn1-Mn1 interaction is more complicated because it occurs via edge sharing octahedra, which results in a competition of direct interaction (AFM) and \(\approx\) 90\(^{\circ}\) Mn1-Te-Mn1 interactions that can be either ferromagnetic (FM) or AFM depending on the \(p\) and \(d\) orbitals involved. Since the ground state is ferrimagnetic with parallel alignment of Mn1 moments, it is tempting to suggest that FM superexchange is the dominant in-plane interaction. Such an explanation would seem consistent with the observation of ferromagnetism in the structurally-related material CrSiTe\(_3\), in which the analogous edge-sharing octahedra have ferromagnetic coupling and orovide the shortest exchange pathway (there is no inter-plane occupancy). Theoretical calculations were performed to get a better understanding of the coupling mechanisms. First principles calculations were performed to understand the ferrimagnetic ground state and the associated \(T_{\rm C}\)Ėnergies of the following six states were calculated: non-magnetic, ferromagnetic (FM), ferrimagnetic states FI1 and FI2, and antiferromagnetic states AF1 an AF2. FI1 is the ferrimagnetic ground state, as also observed experimentally, and its magnetic symmetry is shown in Fig. [\[Structure\]](#Structure){reference-type="ref" reference="Structure"}. The first competing state is AF1, in which all Mn1-Mn1 planar neighbors are anti-aligned, while all Mn1-Mn2 nearest neighbors are aligned. Additional details regarding the various magnetic configurations are in the Supplemental Information. The energies associated with the magnetic states studied fall more than an eV per Mn below the non-magnetic state, and the Mn moments can accurately be considered local moments.
:::
## Lattice behavior and electrical resistivity
As shown in Fig. [\[Lattice\]](#Lattice){reference-type="ref" reference="Lattice"}, the evolution of the \(a\) and \(c\) lattice parameters appears smooth across \(T_{\rm C}\) and near 330 K. The thermal expansion along the \(a\) axis is fairly typical for all \(T\). The \(c\) parameter, however, shows a change in temperature dependence across \(T_{\rm C}\) and an unusual curvature up to near room temperature. This seems to bolster the idea that an ordering transition occurs near 330 K and has some coupling to the lattice, though additional measurements would be useful. The neutron diffraction data at 100 and 380 K refined well within the published crystal structure. The specific heat capacity of Mn\(_3\)Si\(_2\)Te\(_6\) is shown in Fig. [\[Cp\]](#Cp){reference-type="ref" reference="Cp"}. An anomaly can be observed near \(T_{\rm C}\) , and as shown in the inset an applied field (\(H || c\)) causes a broadening of this feature to higher temperature. This behavior is consistent with ferromagnetic-like ordering, where an applied field suppresses fluctuations and broadens the specific heat anomaly by removing more of the magnetic entropy above \(T_{\rm C}\)Ẇhile we do not have a good lattice (phonon) standard for Mn\(_3\)Si\(_2\)Te\(_6\) , an overestimation of the entropy change that occurs near \(T_{\rm C}\) can be made by integrating \(C_P/T\) with a straight line for a background (see Supplemental Information). When fitting between 60 and 95 K, this procedure yields an entropy change of only 8% of that expected for ordering of S=5/2 local moments on Mn\(^{2+}\). The magnetic entropy per mole of Mn in the paramagnetic state is \(S_{mag}\) = RLn\[2S+1\] = R Ln(6) = 14.9 J/mol-Mn/K, and the integration from 60 to 95 K only yields \(\Delta S_{mag}\approx\)`<!-- -->`{=html}1.2 J/mol-Mn/K. Again, we emphasize that this is likely to be an overestimation of \(\Delta S_{mag}\) due to the use of a linear background; this approach is utilized to emphasize how much entropy is released away from \(T_{\rm C}\) (both above and below). A similarly small \(\Delta S_{mag}\) is observed in MnPS\(_3\), where two-dimensional order is observed above \(T_{\rm C}\) , as well as in CrSiTe\(_3\) where short range correlations are observed above \(T_{\rm C}\) and the magnetism is strongly coupled to the lattice. As shown in Fig. [\[Cp\]](#Cp){reference-type="ref" reference="Cp"}b, there is a strong contribution to the specific heat capacity at low temperature, and this decreases with increasing applied field. At the lowest temperatures, the data can be well-described by a Debye term (\(\beta T^3\)) plus a linear term (\(\gamma T\)), the latter of which is the typical way to include an electronic contribution. In this case, however, Mn\(_3\)Si\(_2\)Te\(_6\) is a semiconductor and the strong magnetic-field dependence also points to a magnetic origin; \(\gamma\) should not be associated with the typical electronic Sommerfeld coefficient. As shown in the Supplemental Information, the low \(T\) magnetic contribution in Mn\(_3\)Si\(_2\)Te\(_6\) (and CrSiTe\(_3\)) can also be described using a term proportional to \(T^{1.5}\). Data below 2 K would likely help in considering the nature of the low \(T\) magnetic contribution in Mn\(_3\)Si\(_2\)Te\(_6\)Ṫhe in-plane electrical resistivity \(\rho_{ab}\) of Mn\(_3\)Si\(_2\)Te\(_6\) is shown in Fig. [\[Resist\]](#Resist){reference-type="ref" reference="Resist"} for two crystals. The value at 300 K is consistent with that reported by Ref. , and similar to their data we observe activated-conduction and a strong anomaly at the ferrimagnetic ordering temperature. An increase in the carrier mobility is expected below \(T_{\rm C}\) due to the suppression of spin fluctuations, and this may explain the sharp dip in \(\rho_{ab}\) below \(T_{\rm C}\)Ṫhe feature at \(T_{\rm C}\) is somewhat dramatic, though, especially when considering the implications of a small \(\Delta S_{mag}\) near \(T_{\rm C}\)Ṫhe present data, which extends to higher temperatures than previously reported, also shows a broad feature near 330 K. This appears to demonstrate the bulk-nature of magnetic ordering associated with the susceptibility anomaly observed at the same temperature in Fig. [\[Mag\]](#Mag){reference-type="ref" reference="Mag"}, and the associated behavior in Fig. [\[Resist\]](#Resist){reference-type="ref" reference="Resist"} could be related to carrier scattering or electronic structure effects. To inspect the changes in \(\rho(T)\) above and below the magnetic transitions, we fit the data to \(\rho=\rho_0{\rm Exp[}E_A/k_B T{\rm]}\) where \(E_A\) and \(k_B\) are the activation energy and Boltzmann constant, respectively. The fitted curves are shown in red in Fig. [\[Resist\]](#Resist){reference-type="ref" reference="Resist"}. Firstly, we note that these fits suffer from inadequacy of the simple model (particularly at the lowest \(T\)), as well as the limited amount of data utilized. However, the trends and magnitudes are worth noting. The activation energy obtained by fitting \(\rho\) from 350-380 K is roughly 0.3 eV; this is the region where all of the material appears to be in a paramagnetic state. Below the magnetic anomaly observed at 330 K, the activation energy decreases substantially, \(E_A\)=0.04 eV for 160-300 K. Below \(T_{\rm C}\) , the simple Arrhenius behavior is not a good model for the data. For comparison sake, a small temperature range allows for a reasonable fit to be obtained, with \(E_A\)=0.007 eV for 25-40 K. Interestingly, these results show the decrease of the activation energy with increasing magnetic order, and if \(E_a\) is related to an energy gap then this is rather unusual behavior. Note that the reduction of the activation energy at low \(T\) could be due to the activation of low-level defects.
## Diffuse neutron scattering
Motivated by the bulk measurements that revealed anomalous behavior below 330 K, neutron scattering was performed on CORELLI to search for diffuse scattering. These measurements revealed diffuse scattering near the 002 reflection at 120 K (Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"}), and diffuse scattering along L in the (H-HL) plane at 6 K (Fig. [\[Diffuse6\]](#Diffuse6){reference-type="ref" reference="Diffuse6"}). The scattering data were binned into 3D bins with three orthogonal axes along \[H H 0\], \[H-H 0\] and \[0 0 L\] directions, respectively. The bin sizes are 1%, 2% and 5% of the reciprocal vectors \[1 1 0\], \[1-1 0\] and \[0 0 1\], respectively. Figure [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"}a is a 2D map of the elastic neutron scattering intensity for Mn\(_3\)Si\(_2\)Te\(_6\) at 120 K, showing diffuse scattering around the 002 reflection. The 2D map shows the integrated intensity from \[-0.035-0.035 0\] to \[0.035 0.035 0\], i.e., 7 bins, along the \[H H 0\] out-of-plane direction. The strong scattering intensity reaching out from 000 in Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"} was strongly angle dependent and did not appear consistent with the usual instrument background, but it was essentially independent of temperature. We thus speculate that this feature may have been caused by the glue utilized to orient the small crystal. The diffuse scattering is emphasized in Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"}c, which is a map of the intensity at 120 K with the intensity at 350 K subtracted. By contrast to the 002 reflection, the strong nuclear-only 004 reflection is sharp and is not surrounded by diffuse scattering. This diffuse cloud around 002 is not observed at 6 or 350 K, suggesting it is from short-range magnetic scattering. It is interesting to speculate whether this diffuse scattering may be related to the weak ferromagnetic component of the magnetization seen below 330 K. This hypothesis could be tested with additional temperature-dependent INS measurements. The absence of this diffuse scattering at 6 K suggests that it coalesces into the ferrimagnetic order. This is consistent with the observation of diffuse scattering around the 002 reflection, which has a strong magnetic contribution in the ferrimagnetic phase. The strong contribution from the long-range magnetic order to the intensity of the 002 reflection would also hinder detection of diffuse scattering at 6 K. In Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"}b, line scans along L are shown around the 002 and 004 reflections using different integration areas dH (resolution is inversely proportional to dH). The data for 004 demonstrate the behavior expected for a nuclear-only Bragg peak; the 004 reflection has a sharp peak with little dependence on the temperature or integration area. By contrast, the nuclear component of the 002 reflection is very weak (see Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"}d) and the data at 120 K demonstrate a very broad region of increased intensity centered around 002. The sharpening of the intensity at 002 with decreasing dH (higher resolution) suggests the presence of a small, sharper Bragg reflection underneath a diffuse cloud of scattering. The diffuse scattering in Fig. [\[Diffuse120\]](#Diffuse120){reference-type="ref" reference="Diffuse120"} demonstrate the existence of short-range spin correlations well-above \(T_{\rm C}\) in Mn\(_3\)Si\(_2\)Te\(_6\) , though the nature of these correlations is unclear. The scattering could be from short range order, potentially relating to the magnetization anomaly near 330 K. Alternatively, the diffuse scattering could be associated with excited states that persist above \(T_{\rm C}\) and cause inelastic scattering that cannot be entirely excluded due to the nature of the experiment. Above \(T_{\rm C}\) , the paramagnons could be broad, un-gapped and without a defined dispersion, which would promote isotropic, diffuse scattering in these nearly-elastic measurements. Both short-range order and excited states will exist above a second order magnetic transition, but in some cases these short-range correlations persist to much higher temperatures than is typically observed (alternatively, their correlations lengths are relatively longer above \(T_{\rm C}\) ). The existence of short-range correlations above \(T_{\rm C}\) is generally enhanced when \(T_{\rm C}\) is suppressed relative to the strongest exchange interaction. Geometric frustration, competing interactions, or strong anisotropy in the exchange constants can all lead to such behavior. In these cases, strong magnetic exchange interactions can promote short-range correlations (of the ground or excited states) well-above \(T_{\rm C}\) because thermal energy has less influence near \(T_{\rm C}\)Ȧs noted in Ref. , the relevant ratio to consider when thinking about short-range correlations near \(T_{\rm C}\) is not \(T/\[T_{\rm C}\) but rather \(k_BT\)/JS. In many quasi-2D systems, anisotropic interactions lead to a relative suppression of \(k_B\]T_{\rm C}\) , and a strong JS can enhance correlation lengths above \(T_{\rm C}\)Ḟor Mn\(_3\)Si\(_2\)Te\(_6\) , competing antiferromagnetic interactions appear to provide the underlying mechanism that ultimately promotes the persistent, short-range correlations above \(T_{\rm C}\)Ȧ counter example is MnF\(_2\), where the thermal energy is large relative to the exchange energy at the antiferromagnetic ordering \(T_{\rm N}\) , and as a result the majority of the magnetic entropy in MnF\(_2\) is released below \(T_{\rm N}\) \(\approx\) 67 K. At 6 K, diffuse scattering was detected along L in the (H-HL) plane, as shown in Fig. [\[Diffuse6\]](#Diffuse6){reference-type="ref" reference="Diffuse6"}. The anisotropy of this low-temperature diffuse scattering suggests it is associated with a short correlation length along the \(c\) axis. The 6 K diffuse scattering is centered around L=2, and is only observed around particular reflections (H-K = 3n+2) that are only contributed to by Mn scattering (both sites contribute). This enhances the relative contribution of the magnetic intensity in these reflections and facilitates detection of this diffuse scattering. Despite appearing to have a magnetic origin, there is perhaps a minor indication of this diffuse scattering at 350 K (though this is within the noise of the background scattering). Therefore, this feature may be associated with defects on the Mn positions. From a magnetic perspective, the diffuse scattering seems to suggest that there is a short correlation length of spin canting along the \(c\) axis. While this seems plausible and consistent with our other measurements, the current data cannot determine the origin of this diffuse scattering. An additional scattering experiment with an applied field would be the next step in determining the origin of this anisotropic diffuse scattering. We note that the tails extending toward L = 0 (the asymmetry of 004 for instance) are an artifact of the time of flight measurement technique and are not observed in the elastic data (see Supplemental Information).
# Summary
This work has characterized the magnetism of Mn\(_3\)Si\(_2\)Te\(_6\) and found it to be a ferrimagnet due to anti-parallel alignment of moments on different Mn atomic positions. If the potential canting of the moments along the \(c\) axis is ignored, then the magnetic structure is fairly simple. By contrast, the mechanisms that lead to this ground state are relatively complex. A competition between antiferromagnetic exchange interactions exists, and it is a third-nearest neighbor interaction that ultimately determines the ground state. This frustration suppresses \(T_{\rm C}\) and likely enhances the importance of fluctuations, leading to persistent short-range correlations above \(T_{\rm C}\)Ṁagnetization measurements also suggest that some additional ordering may exist below \(\approx\)`<!-- -->`{=html}330 K. Diffuse scattering was observed at 6 K that may indicate short-range correlations of spin canting along the \(c\) axis. Due to the competing antiferromagnetic interactions, the manipulation of bond distances or the targeted substitution of the Mn2 atoms would likely yield new ground states with potentially higher ordering temperatures. | {'timestamp': '2017-06-08T02:01:14', 'yymm': '1706', 'arxiv_id': '1706.01925', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01925'} |
# Introduction {#1}
A *quasigroup* \(\mathcal{L}\) is a set with a binary operation \(*: \mathcal{L} \times \mathcal{L} \rightarrow \mathcal{L}\), such that every equation \(x*a = b\) or \(a*x = b\) with \(a, b \in \mathcal{L}\) has a unique solution \(x\). In this case, for every \(a \in \mathcal{L}\), the *right multiplication* \(R_a: \mathcal{L} \rightarrow \mathcal{L}, x \mapsto x * a\) is a permutation of \(\mathcal{L}\) (and of course so is every *left multiplication*). A quasigroup is a *loop*, if it contains an identity element. Thus a group is just a loop, in which the operation is associative, and we will indeed view groups as loops. In the following, we will only consider finite loops. Let \(\mathcal{L}\) be a (finite) loop, whose identity element we denote by \(e\). The *right multiplication group of \(\mathcal{L}\)* is the group \(G := \langle R_a \mid a \in \mathcal{L} \rangle\), a subgroup of the symmetric group on \(\mathcal{L}\). Clearly, \(G\) acts faithfully and transitively on \(\mathcal{L}\) and \(R_e\) is the identity element of \(G\), which we denote by \(1\). Let \(H \leq G\) denote the stabiliser in \(G\) of \(e \in \mathcal{L}\), and let \(T := \{ R_a \mid a \in \mathcal{L} \}\). Then \(T\) is a transversal for \(H^g \backslash G := \{ H^g x \mid x \in G \}\) for every \(g \in G\), the identity element of \(G\) is contained in \(T\), and \(\langle T \rangle = G\). The triple \((G,H,T)\) is called the *envelope* of \(\mathcal{L}\), a group theoretic object.. Conversely, starting from group theory, one defines a *loop folder* to be a triple \((G,H,T)\) of a finite group \(G\), a subgroup \(H \leq G\) and a subset \(T \subseteq G\) with \(1 \in T\), such that \(T\) is a transversal for \(H^g \backslash G\) for every \(g \in G\). Given a loop folder \((G,H,T)\) one can construct a loop \((\mathcal{L},*)\) on the set \(H\backslash G\) of right cosets of \(H\) in \(G\). However, the envelope of \(\mathcal{L}\) need not be equal to \((G,H,T)\). In contrast to the right multiplication group of \(\mathcal{L}\), in general the group \(G\) will not act faithfully on \(\mathcal{L}\), and the transversal \(T\) will not generate \(G\). On the other hand, it is not difficult to construct the envelope of \(\mathcal{L}\) from \((G,H,T)\). These results, as well as the notion of *loop folder* and *envelope of a loop* are contained in. However, the connection between loops and their envelopes goes back to Baer . Let \(\mathcal{L}\) be a loop with envelop \((G,H,T)\). We say that \(\mathcal{L}\) is *right conjugacy closed*, or an \(\text{\rm RCC\ }\) loop, if \(T = \{ R_a \mid a \in \mathcal{L} \}\) is closed under conjugation by itself. Clearly, this is the case if and only if \(T\) is invariant under conjugation in \(G = \langle T \rangle\); in other words, if \(T\) is a union of conjugacy classes of \(G\). We shortly say that \(T\) is \(G\)-invariant in the following. Thus an \(\text{\rm RCC\ }\) loop gives rise to a \(G\)-invariant transversal of \(H\), the stabilizer of \(e\) in \(G\). (A \(G\)-invariant transversal of a subgroup \(H\) of a group \(G\) is sometimes called a *distinguished transversal* in the literature.) On the group theoretic side, this leads to the notion of an \(\text{\rm RCC\ }\) loop folder. This is a loop folder \((G,H,T)\), where \(T\) is \(G\)-invariant. More definitions regarding loop folders are given at the beginning of Section [\[3\]](#3){reference-type="ref" reference="3"}. It has been shown by Drápal that an \(\text{\rm RCC\ }\) loop of prime order is a group. In this paper we determine all \(\text{\rm RCC\ }\) loops of order \(2p\), where \(p\) is an odd prime. In order to achieve this, we first describe the possible envelopes \((G,H,T)\) of such loops. Our approach is group theoretic. In Sections [2](#2){reference-type="ref" reference="2"} and [\[3\]](#3){reference-type="ref" reference="3"} we show that if \((G,H,T)\) is an \(\text{\rm RCC\ }\) loop folder such that \(G\) acts faithfully on \(H\backslash G\) and the index \(|G\colon H|\) is the product of two distinct primes, then \(G\) acts imprimitively on \(H\backslash G\) (Theorem [\[Imprimitive\]](#Imprimitive){reference-type="ref" reference="Imprimitive"}). This result uses the classification of the finite simple groups and is based on the classification of finite primitive permutation groups of squarefree degree by Li and Seress, and on the determination of the minimal degrees of permutation representations of finite groups of Lie type by Patton, Cooperstein and Vasilyev. For the purpose of our further investigation, it would suffice to enumerate the primitive permutation groups of degree \(2p\) for odd primes \(p\); we are not aware of any result in this direction which does not rely on the classification of the finite simple groups. In Section [4](#4){reference-type="ref" reference="4"}, we continue with some basic results on permutation groups of degree \(p\) and give a new proof of Drápal's theorem on \(\text{\rm RCC\ }\) loops of prime order (Corollary [\[DrapalsTheorem\]](#DrapalsTheorem){reference-type="ref" reference="DrapalsTheorem"}). Let \((G,H,T)\) be the envelope of an \(\text{\rm RCC\ }\) loop of order \(2p\), where \(p\) is an odd prime. Using Theorem [\[Imprimitive\]](#Imprimitive){reference-type="ref" reference="Imprimitive"} mentioned above, we may now assume that there is a subgroup \(K \lneq G\) with \(H \lneq K\), and also that one of the indices \(|G\colon K|\) or \(|K\colon H|\) is equal to \(2\), and the other index is equal to \(p\). This configuration is analysed in Section [5](#5){reference-type="ref" reference="5"} with elementary group theoretical methods. It turns out that there are three possible types for \(G\). Firstly, \(G\) can be isomorphic to the wreath product \(C_p \wr C_2\), where \(C_p\) and \(C_2\) denote (cyclic) groups of order \(p\) and \(2\), respectively. Secondly, \(G\) can be isomorphic to a subgroup of \(\mbox{\rm Aff}(1,p)\), the affine group over \(\mathbb{F}_p\). Thirdly, \(G\) can be isomorphic to a group \(K \times \langle a \rangle\), where \(K\) is an odd order subgroup of \(\mbox{\rm Aff}(1,p)\) and \(a\) is an element of order \(2\) (Theorem [\[MainResult\]](#MainResult){reference-type="ref" reference="MainResult"}). In particular, \(G\) is soluble. Ultimately, our results rely on the classification of the finite simple groups. One could avoid this by assuming from the outset that \(G\) is soluble. This would lead to exactly the same list of \(\text{\rm RCC\ }\) loops of order \(2p\), but of course without the guarantee to have found them all. In Section [6](#6){reference-type="ref" reference="6"}, we determine the number of isomorphism classes of loops of order \(2p\) (Theorem [\[NumberOfAllLoops\]](#NumberOfAllLoops){reference-type="ref" reference="NumberOfAllLoops"}). Finally, Section 7 introduces a series of examples of \(\text{\rm RCC\ }\) loops of order \(q^2-1\) and multiplication groups \(\mbox{\rm GL}(2,q)\) (Proposition [\[GL2Examples\]](#GL2Examples){reference-type="ref" reference="GL2Examples"}). For \(q = 4\), we obtain a loop of order \(3 \cdot 5\), whose multiplication group is not soluble. These examples indicate that a generalisation of our results to \(\text{\rm RCC\ }\) loops of order \(pq\) for distinct primes \(p\) and \(q\) could be substantially more difficult. This is a good place to discuss some related results. In , Stein shows that if \(T\) is a conjugacy class in a finite group and at the same time a tranversal to a subgroup, then \(\langle T \rangle\) is soluble. This result uses the classification of the finite simple groups. Without the classification, but with the help of the Odd Order Theorem, Csőrgő and Niemenmaa in obtain the solubility of the full multiplication group of a loop under certain conditions on the stabilizer of a point. Their paper contains further references for results along this line. In, Csőrgő and Drápal characterise left conjugacy closed loops inside the class of nilpotent loops of nilpotency class two. In the same paper, these authors also determine the nilpotent left conjugacy closed loops of order \(p^2\) for primes \(p\). In, Kunen shows that for each odd prime \(p\) there is exactly one non-assciative conjugacy closed loop of order \(2p\), up to isomorphism (a loop is conjugacy closed, if it is both left and right conjugacy closed). Burn shows in that every Bol loop of order \(p^2\) or \(2p\) for a prime \(p\) is a group. Finally, in Daly and Vojtěchovský determine the number of nilpotent loops of order \(2p\), where again \(p\) is a prime, up to isomorphism. This paper builds upon the PhD thesis of the first author, written under the direction of the second author and Alice Niemeyer. Theorem [\[Imprimitive\]](#Imprimitive){reference-type="ref" reference="Imprimitive"} is contained in this thesis, but also a complete classification of all \(\text{\rm RCC\ }\) loops of order at most \(30\). These have been incorporated into the GAP package *Loops* of Nagy and Vojtěchovský . The classification of the \(\text{\rm RCC\ }\) loops of order \(2p\) is, to the best of our knowledge, new. The examples computed in were of considerable importance for confirming our theoretical results of Section 6. The example of an \(\text{\rm RCC\ }\) loop of order \(15\) and multiplication group \(\mbox{\rm GL}(2,4)\) contained in , gave rise to the series of examples constructed in Section 7. Our group theoretical notation is standard. For example, we write \(G'\) for the commutator subgroup of the group \(G\). We do recall the notion of an almost simple group and that of the core of a subgroup in the introductions to Section [2](#2){reference-type="ref" reference="2"} and Section [\[3\]](#3){reference-type="ref" reference="3"}, respectively. As already indicated above, a cyclic group of order \(n\) is denoted by \(C_n\), and the symmetric and alternating groups of degree \(n\) are denoted by \(S_n\) and \(A_n\), respectively.
# Primitive permutation groups of squarefree degree {#2}
We begin with a remark on the sizes of conjugacy classes in almost simple groups. Recall that a group \(G\) is *almost simple*, if there is a non-abelian finite simple group \(S\) such that \(S \leq G \leq \mbox{\rm Aut}(S)\) (where \(S\) is identified with the group of inner automorphisms of \(S\)). In this context, \(S\) is called the *socle* of \(G\).
The following theorem combines some major results by Li and Seress on finite primitive permutation groups of squarefree degree, and by Patton, Cooperstein and Vasilyev on the minimal degrees of permutation representations of finite groups of Lie type.
In the lemma below we are going to make use of Zsigmondy primes, also known as primitive prime divisors. Let \(r\) and \(d\) be integers greater than \(1\). We call a prime \(\ell\) a Zsigmondy prime for \(r^d-1\), if \(\ell\) divides \(r^d-1\), but not \(r^i-1\) for \(1 \leq i < d\). A Zsigmondy prime for \(r^d-1\) exists whenever \(d > 2\) and \((r,d) \neq (2,6)\) (see ).
# The action of the right multiplication groups of rcc loops of [\[3\]]{#3 label="3"} twice prime order
The results of the previous section are now applied to \(\text{\rm RCC\ }\) loops whose order is the product of two distinct primes. Recall the notion of the envelope of a loop as introduced in the second paragraph of Section [1](#1){reference-type="ref" reference="1"} (which follows ). Recall also that a loop folder is a triple \((G,H,T)\), where \(G\) is a finite group, \(H\) is a subgroup of \(G\) and \(T\) is a transversal, with \(1 \in T\), for all coset spaces \(H^g\backslash G\) with \(g \in G\) (see the third paragraph of Section [1](#1){reference-type="ref" reference="1"} and ). A loop folder \((G,H,T)\) is an \(\text{\rm RCC\ }\) loop folder, if \(T\) is invariant under conjugation by \(G\) (see also Section [1](#1){reference-type="ref" reference="1"}). We will now introduce further notation, although only needed in later sections. Let \((G,H,T)\) be a loop folder. By definition, the *order* of \((G,H,T)\) is the size of \(T\). We say that \((G,H,T)\) is *faithful*, if \(G\) acts faithfully on \(H\backslash G\). This is the case if and only if the core of \(H\) in \(G\) is trivial. Recall that the smallest normal subgroup \(C\) of \(G\) containd in \(H\) is called the *core of \(H\) in \(G\)*. Thus \(C\) is the intersection of all the \(G\)-conjugates of \(H\) in \(G\), i.e. \(C := \cap_{g \in G} H^g\). The core of \(H\) in \(G\) is equal to the kernel of the permutation representation of \(G\) on the (right or left) cosets of \(H\). Clearly, the envelope of a loop is a faithful loop folder. Here is the main result of this section. It is only used in the setup of Subsection [5.1](#Generalities){reference-type="ref" reference="Generalities"}, and nowhere else in this paper.
We end this section with two general results on loop folders with certain invariance properties. The first will be used in an extension of a theorem of Drápal .
# Right conjugacy closed loops of prime order {#4}
In this section we give a new proof of a theorem of Drápal which states that left conjugacy closed loops of prime order are groups. We prove the analogue for \(\text{\rm RCC\ }\) loops, but as the opposite loop of a left conjugacy closed loop is an loop, our version is equivalent to Drápal's result. Recall the notions related to loop folders summarized at the beginning of Section 3. We begin with an easy lemma.
We will also need the following generalization of Corollary [\[DrapalsTheorem\]](#DrapalsTheorem){reference-type="ref" reference="DrapalsTheorem"}.
Notice that the above proposition is a generalization of Drápal's theorem (see Corollary [\[DrapalsTheorem\]](#DrapalsTheorem){reference-type="ref" reference="DrapalsTheorem"}); indeed, if \((G,H,T)\) is the envelope of a loop, then \(G = \langle T \rangle\), and if, moreover, \(G\) is abelian, then \(H = 1\), as the core of \(H\) in \(G\) is trivial.
# The right multiplication groups of rcc loops of twice prime order {#5}
We refer the reader to the introduction of Section [\[3\]](#3){reference-type="ref" reference="3"} for the notions related to loop folders.
## Generalities {#Generalities}
Let \(p\) and \(q\) be distinct primes and let \((G,H,T)\) denote the envelope of an \(\text{\rm RCC\ }\) loop of order \(pq\). This implies in particular that \(G\) acts faithfully on \(H\backslash G\), i.e. the core of \(H\) in \(G\) is trivial. It is at this stage, and only here, where we impose an important consequence of Theorem [\[Imprimitive\]](#Imprimitive){reference-type="ref" reference="Imprimitive"}. This states that \(G\) acts imprimitively on \(H\backslash G\), and hence \(H\) is not a maximal subgroup of \(G\). We let \(K \lneq G\) such that \(H \lneq K\). We choose notation such that \(|G\colon K| = q\) and \(|K\colon H| = p\). Put \(T_1 := T \cap K\), \(K_1 := \langle T_1 \rangle \leq K\) and \(H_1 := H \cap K_1 \leq H\). We collect first properties.
## The case \(q = 2\). {#CaseQEqal2}
Let us assume throughout this subsection that \(q = 2\). Then \(K \unlhd G\). Moreover, \(K_1 \unlhd G\), as \(K_1^g = \langle (T \cap K)^g \rangle = \langle T \cap K \rangle = K_1\) for all \(g \in G\).
We record two consequences which will be used later on.
We now distinguish two cases.
### Case \(1\) {#Case1Section}
Assume that \(H \neq 1\) and that \([s, K_1] = 1\) for all \(s \in T \setminus K\).
### Case \(2\) {#Case2Section}
Assume that there is \(s \in T \setminus K\) with \([s,K_1] \neq 1\). In this case we put \(Z := K_1 \cap C_G(s)\). Also, we let \(C := \cap_{k \in K} H^k \unlhd K\) denote the kernel of the action of \(K\) on the cosets of \(H\) in \(K\).
The previous result implies in particular that \(G\) is soluble. Indeed, \(K_1\) is a normal subgroup of \(G\), as we have remarked at the beginning of [5.2.2](#Case2Section){reference-type="ref" reference="Case2Section"}. Now \(K_1\) is abelian by Corollary [\[StructureOfK1\]](#StructureOfK1){reference-type="ref" reference="StructureOfK1"}, and \(G/K_1 = C_G(s)K_1/K_1 \cong C_G(s)/Z\) by Part (b) of the lemma above and by the definition of \(Z\). By Part (c) of the lemma, \(C_G(s)/Z\) is cyclic, hence \(G\) is soluble.
## The case \(p = 2\).
Let us assume throughout this subsection that \(p = 2\). Then \(H \unlhd K\). Here, we put \(D := \cap_{g \in G} K^g\), the kernel of the action of \(G\) on the set of right cosets of \(K\). Then \(D\) is an elementary abelian \(2\)-group, as \(G\) acts faithfully on the set of right cosets of \(H\) and as \(H\) has index \(2\) in \(K\). We write \(\bar{\ }: G \rightarrow \bar{G} := G/D\) for the canonical epimorphism. Then \(\bar{G}\) is a faithful permutation group on \(q\) letters, i.e. \(\bar{G}\) is isomorphic to a subgroup of \(S_q\).
### Case 1 {#case-1}
Here, we consider the case that \(K\) is normal in \(G\). Let us keep the notation of Lemma [\[KIntersectT\]](#KIntersectT){reference-type="ref" reference="KIntersectT"} in the following.
### Case 2 {#case-2}
Here, we consider the case that \(K\) is not normal in \(G\). Again, we use the notation of Lemma [\[KIntersectT\]](#KIntersectT){reference-type="ref" reference="KIntersectT"}.
## The main result
We can now summarize our results for envelopes of \(\text{\rm RCC\ }\) loop folders of orders \(2p\) for odd primes \(p\).
# The rcc loops of twice prime order {#6}
Let \(p\) be an odd prime. In this section we determine the number of isomorphism classes of loops of order \(2p\). Let \(\mathcal{L}\) denote such a loop and let \((G,H,T)\) be its envelope. By numbering the elements of \(\mathcal{L}\) by the integers \(1, \ldots, 2p\), where \(1\) numbers the identity element of \(\mathcal{L}\), we may and will view \(G\) as a subgroup of \(S_{2p}\), and \(H\) as the stabilizer in \(G\) of \(1\). If \(\mathcal{L}_1\) and \((G_1,H_1,T_1)\) is another such configuration, then \(\mathcal{L}\) and \(\mathcal{L}_1\) are isomorphic as loops, if and only if there is an element of \(S_{2p}\), conjugating \((G,H,T)\) to \((G_1,H_1,T_1)\). The isomorphism types of the right multiplication groups arising in loops of order \(2p\) have been described in Theorem [\[MainResult\]](#MainResult){reference-type="ref" reference="MainResult"}. For each of these groups \(G\) we have to determine their embeddings into \(S_{2p}\) up to conjugation. This will yield the possible pairs \((G,H)\) to be considered. For each of these pairs we have to determine the normalizer \(N\) in \(S_{2p}\) of \(G\) and \(H\), and then find the distinct \(N\)-orbits of \(G\)-invariant transversals \(T\) for \(H\backslash G\) such that \(1 \in T\) and \(\langle T \rangle = G\). We will refer to the three different types of \(G\) in Theorem [\[MainResult\]](#MainResult){reference-type="ref" reference="MainResult"}(a), (b), and (c) as Case (a), (b) and (c), respectively. We begin with some preliminary results. As usual, the largest normal \(p\)-subgroup of a finite group \(U\) is denoted by \(O_p(U)\), and its largest normal subgroup of odd order by \(O_{2'}(U)\).
Let \(n, d\) be positive integers, and let \(\zeta \in S_n\) denote an \(n\)-cycle. Let us put \[\label{Involutions1} I_{n,d} := |\{ \tau \in S_n \mid \tau^2 = 1, \tau \zeta^d = \zeta^d \tau \}|\] and \[\label{Involutions2} I_n := I_{n,n}.\] Thus \(I_{n,d}\) is one more than the number of involutions in \(C_{S_n}( \zeta^d )\) and \(I_n\) is one more than the number of involutions in \(S_n\). Notice that the definition of \(I_{n,d}\) does not depend on the chosen \(n\)-cycle \(\zeta\), as all \(n\)-cycles are conjugate in \(S_n\). It is not difficult to derive a formula for \(I_{n,d}\), where the formula for \(I_n\) is certainly well known. In the following result, \(n \text{\rm\ mod\ } 2 \in \{ 0, 1 \}\) denotes the remainder of the division of \(n\) by \(2\).
We summarise our results in the following theorem.
The table below contains the numbers obtained by evaluating formula ([\[FormulaForNumberOfAllLoops\]](#FormulaForNumberOfAllLoops){reference-type="ref" reference="FormulaForNumberOfAllLoops"}) for small values of \(p\). These numbers have also been obtained for \(p \leq 13\) in the PhD-thesis of the first author by different methods. \[\begin{array}{c|cccccccc} p & 2 & 3 & 5 & 7 & 11 & 13 & 17 & 19 \\ \hline {\rm (\ref{FormulaForNumberOfAllLoops})} & 2 & 5 & 18 & 99 & 10\,489 & 151\,973 & 49\,096\,721 & 1\,052\,729\,657 \end{array}\] One of the referees has kindly pointed out that formula ([\[FormulaForNumberOfAllLoops\]](#FormulaForNumberOfAllLoops){reference-type="ref" reference="FormulaForNumberOfAllLoops"}) evaluates to an integer, even if \(p\) is not a prime (and larger than \(1\)). This follows from the fact that for general positive integers \(n, d\), the number \(I_{n,d}\) equals the number of fixed points of the element \(\zeta^d\) on the set \(\{ \tau \in S_n \mid \tau^2 = 1 \}\), where the \(n\)-cycle \(\zeta\) acts by conjugation. Thus, by the Burnside-Cauchy-Frobenius lemma, the number of orbits of \(\langle \zeta \rangle\) on \(\{ \tau \in S_n \mid \tau^2 = 1 \}\) equals \(1/n \sum_{d = 1}^{n} I_{n,d}\), so that this number is an integer.
# A series of examples {#7}
According to Theorem [\[MainResult\]](#MainResult){reference-type="ref" reference="MainResult"}, the right multiplication group of an \(\text{\rm RCC\ }\) loop of order \(2p\), where \(p\) is an odd prime, is soluble. This is no longer the case for right multiplication groups of \(\text{\rm RCC\ }\) loops of order \(pq\), where \(p\) and \(q\) are distinct primes. An example is given in of an \(\text{\rm RCC\ }\) loop of order \(15\) with right multiplication group isomorphic to \(\mbox{\rm GL}(2,4)\). This fits into an infinite series of examples. | {'timestamp': '2017-12-20T02:06:03', 'yymm': '1706', 'arxiv_id': '1706.01645', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01645'} |
# Introduction
Evidence suggests that cognitive performance decreases following periods of sustained mental effort and that taking breaks helps knowledge workers of all types from paid-per-task Mechanical Turk workers to full-time analysts of every domain. For instance, asserts that workers are refreshed after breaks and that workers who take breaks perform more work compared to those who do not. It has also been observed that participant's resting baseline behavior predicts subsequent performance declines. It may be possible to improve a worker's overall performance by improving his/her baseline between periods of task activity. But there is competition between the production rate of work from knowledge worker and the quality of their work. Highly motivated knowledge workers' performance may decrease in quality if they continue to work for long periods of time without some sort of break. Knowledge workers focused on completion of tasks may not be able to accurately monitor their own fatigue level and may miss subtle cues indicating that they would benefit from some sort of break. We explore the notion of cognitive depletion where knowledge workers' cognitive resources are depleted by continued work and their ability to accurately complete their task or provide meaningful analytical feedback becomes compromised. Cognitive depletion extends beyond simple sleep-deprivation related fatigue, and experts are susceptible to cognitive depletion because they may not feel (or admit to) the early signs of mental fatigue we often associate with exhaustion. It is important then to have a mechanism for identifying impending cognitive depletion without the direct input of an individual to describe their current levels of fatigue and depletion. This mechanism must be able to reason about a person's state based on overt, observable symptoms in an unbiased way if it is to benefit its users. Identifying such observable symptoms remains a challenge because fatigue, in general, is not well understood. In this work, we present the results of a user study designed to evaluate a set of potential symptoms of cognitive depletion from a total of 19 participants. We use a browser-based game to simulate the analysis of Nuclear Magnetic Resonance (NMR) spectrum and asked participants to play this game for up to 3 hours. During this session researchers observed the participants for candidate symptoms of cognitive depletion while collecting game event logs and eye tracking data. We use the results of this study to draw inferences about how cognitive depletion manifests when people are asked to perform demanding problem-solving tasks for long durations of time. Our work provides early insights into cognitive depletion and addresses the future research implications of the findings. The following are the contributions of this paper:
- We summarize fourteen symptoms observed over the course of three hours of the user study (19 participants).
- We describe characteristics of each symptom and its relationship with others through the triangulation of different datasets (e.g., symptom observations, event logging, and eye-tracking) as well as individual-and group-level analyses.
# Related Works
Our work understanding cognitive depletion builds on research in the related fields of attention, interruption and task resumption, multitasking, error and mistake frameworks, quality control, working memory, and cognitive overload.
## Attention, Interruptions, and Task Resumption
Research into attention processes often attempts to measure and predict how long individuals can attend to a specific task. One important concept from attention research is the often observed vigilance decrement: after periods of sustained effort on a vigilance task, individuals' begin to miss cues critical to their task or workflow . Task-unrelated thoughts (TUTs) are another important phenomenon from attentional research. Individuals required to focus on a task for long periods of time often report self-distracting thoughts that are unrelated to the task at hand . Suppressing TUTs is a current topic of research but for our purposes, we view TUTs as a potential symptom of cognitive depletion. Interruption and task resumption research seeks to understand workflows and predict the optimal time to interrupt an individual so that the interruption is as minimally detrimental as possible (see our references for just a few examples). Such understanding is important to the study of cognitive depletion because any coping or mitigation strategies must also be minimally intrusive to workflows.
## Quality Control, Error, and Mistake Frameworks
The study of Mechanical Turk-style economies has provided a wealth of techniques for assessing the quality of worker contributions and for detecting workers who are abusing task structures for their own gain . These techniques are beneficial to the study of cognitive depletion as they provide potential metrics that can be automatically collected without interrupting workers as well as providing insight into the working patterns of large groups of individuals. It also provides an easily accessible real-world example of an ideal scenario for cognitive depletion research: an economy based on constant completion of micro-tasks where workers are motivated to work beyond the point where their cognitively depleted state begins affecting the quality of their output. A tremendous amount of work has been done to understand and categorize human error in a variety of contexts. This work is particularly important in fields such as air traffic control where human error endangers hundreds of lives . Our work and model presented here do not attempt to replicate the in depth error frameworks completed by others. Rather, we use them as a source of observations and discrete observable phenomena which may be symptoms of cognitive depletion.
## Working Memory and Cognitive Overload
Research into working memory and cognitive overload are the closest parallels to cognitive depletion that we have found and numerous works have provided the foundation for this current work. From these works important concepts such as mental and cognitive fatigue have arisen . Such research makes important distinctions between the effects of acute workload and extended engagement. Particularly, a sudden acute workload may cause a sort of 'buckling stress' where an individual is unable to cope, while fatigue is the gradual loss of work capacity . defines mental fatigue as a change in psychophysiological state due to sustained performance. Both working memory and cognitive overload research study the immediate effects of task load in order to determine when a person is cognitively overloaded. We view cognitive overload as one mechanism which will, with extended engagement, result in a state of cognitive depletion. Highly overloaded persons will reach a state of cognitive depletion at a rate faster than those who are not cognitively overloaded. Mental fatigue as defined by is then a cue that an individual is cognitively depleted. One seminal work of this field is the frequently used NASA-TLX which provides a six-axis inventory for self-assessing the cognitive load of a task. Our work builds on this foundation to assess signs indicating that a person may be in a state of cognitive depletion.
# Symptoms of Cognitive Depletion
Building on the discussed related fields, we drafted a collection of potential symptoms of cognitive depletion based on the results of published experimental evaluations (see Table [1](#tab:summary_symptoms){reference-type="ref" reference="tab:summary_symptoms"}). This gave us a starting collection of possible symptoms that we can group into three broad categories as follows.
## Vigilance and Reactionary Symptoms (VR)
Vigilance and reactionary symptoms include phenomena such as habituation to status alerts, increased reaction time, distraction, and inattention. These symptoms are most likely to manifest in conditions require active monitoring of real-time events. For example, 'vigilance decrement' is a well known phenomenon where reaction time and attentiveness to a stimuli stagnates as function of time task . In this study, we focus on four potential symptoms: distraction (D), inattention (I), increased decision time (IDT), and task-unrelated thoughts (TUT).
## Physical and Emotional Deregulation (PED)
Physical and motor symptoms include drowsiness, exhaustion, clumsiness, etc. as well as unintentional mistakes providing input, data, or keystrokes involving fine-motor skills. For example, a person may begin to fidget or change their seated posture rapidly or begin to click the wrong buttons on an interface . This category also includes emotional effects such as increased negative affect where a person may begin to complain about the difficulty of a task . In this study, physical and emotional deregulation manifests as data/command errors (DCE), increased negative affect (INA), and physical effects (PE).
## Personal Judgment Symptoms (PJS)
Personal judgment symptoms include occasions where a person is not accurately accounting for the state or their task and is confused or rushing tasks. It also includes disruptive workflows and the failure to form or adjust strategies. The distinguishing characteristic of symptoms in this category is that the person intends the actions resulting from an inaccurate judgement. If asked, the person would not say that he/she had made a mistake in the same sense that a person who misspelled a word might. In the latter case, the person knew how to spell the word but physically hit the wrong key. In the former, a person may intentionally use the word \"then\" when the grammatically correct word would be \"than\". Another symptom belonging to this category would be a demonstrated fixedness or rigidity in strategy towards a task even when it repeatedly fails . Symptoms such as confusion (C), effort/risk under/over estimation (ERE), forgetting (F), task abandonment (TA), lack of advancement/progress (LAP), strategy inefficiency (STI), and task rushing (TR) belong to this category.
# Methodology
Here we describe the environment and protocol we used to investigate cognitive depletion. Our study involved participants playing a puzzle-solving game for several hours while we collected data on overt symptoms of cognitive depletion.
## Apto Game Environment
Participants were recruited to play a web-game, Apto (see Figure [\[aptoss\]](#aptoss){reference-type="ref" reference="aptoss"}), which simulated the Nuclear Magnetic Resonance (NMR) spectrum analysis process and was designed to be similar to the professional software packages such as The Chenomx NMR suite [^1] used by experts to analyze spectrum results. Each spectrum was randomly generated and consisted of 1 to 16 compounds depending on the difficulty level. The score for each puzzle was calculated based on a formula that awarded more points per compound to higher levels. During the game, these scores were abstracted to a star system based on the game level: participants earned one bronze star for each completed level 1 puzzle, 2 bronze stars for level 2, 3 bronze stars for level 3, 1 silver star for level 4, and 1 gold star for each level 5 puzzle. The total score earned by a participant was available to him/her in this star form through a slide-down window. Participants solved puzzles by clicking peaks of the spectrum to search for compounds with peaks near the search point. Participants selected candidate matching compounds from the search results of suggested compounds and then adjusted the fit of each compound to match as precisely as possible by clicking and dragging the compound's individual spectrum up or down relative to the puzzle spectrum. Compounds could have closely overlaping spectra, so three additional tools were available to help participants. A \"sum line\" showed the summation of the candidate compounds the participant currently had selected, so they could compare against the puzzle. A \"difference line\" drew a red line along the spectrum to highlight the difference between the sum line and the spectrum. An \"error line\" drew a bright-magenta line to draw attention to areas where the mean squared error between the sum and spectrum lines was great. These three tools operated independently and could be used in any combination by the participant. Participants had a \"restart puzzle\" button available to them which undid all their actions on a puzzle and returned it to its initial state. A \"skip\" button allowed participants to abandon the current puzzle in an unfinished state and be given a new puzzle to try. The restart and skip buttons did not affect the participant's score. When a participant completed a puzzle, they were given the option to change their level and could jump to any level from 1-5. Puzzles were considered solved when participants achieved 99% of the available points.
## Participants
We recruited 19 participants (8 female) from our organization through flyers, emails, and meeting announcements. 9 of these participants wore glasses. Participants ranged in age and job description but did not have prior experience conducting NMR analysis. Participants were compensated for their time away from their regular jobs. In addition, we offered a small prize to be awarded to the participant with the highest session score at the end of the study.
## Protocol
Our procedures were reviewed and approved by our IRB. Participants arrived and signed a consent form which explained what data would be collected and what their task for the session would be. Next, study facilitators attempted to set up the eye tracking device. Participants then completed an in-game tutorial which taught them how to query the system for suggested compounds and fit compounds into a spectrum to solve the puzzle. During this tutorial period, participants could ask for as much help as they needed from study facilitators to feel comfortable. When the tutorial finished, participants were given an additional tip-sheet which reminded them of how to use game controls and provided some basic strategies for solving complicated puzzles. Participants were then directed to play the game for the entire 3 hour session. Because we hypothesized that taking breaks related to cognitive depletion, we instructed them that they could take breaks whenever they chose, but we neither encouraged nor discouraged the users to take breaks. A study facilitator was present for the entire session and collected observational data by hand.
## Data Collection
The game participants played automatically collected event information as the participants interacted with it. From this data, we have information about how many puzzles participants solved, how many compounds were in each puzzle (which refer to difficulty levels), and click data representing participant interactions with controls. We also collected timing information about how long it took for each puzzle to be solved through this game log. A Tobii Glasses 2 Eye Tracker [^2] was used to collect eye movement, accelerometers, pupil dilation, and fixation data. We could not calibrate the eye tracker and collect eye-tracking data on 6 of the 9 participants who wore glasses. For these 6 participants, we rely on event log and observational notes to capture activity during the session. The eye tracker was turned off when participants took breaks so that they could leave to use the restroom and then re-calibrated upon their return from break. Study facilitators used a predetermined coding scheme to capture the symptoms observed during the session as well as general notes about the session such as errors encountered, break times, and any interrupting factors. To account for differences in timing devices and the reaction-time of study facilitators, the observed time of each symptom was recorded at the minute level. After the study, for data reliability, four people (including two study facilitators and two researchers) looked into the logs and made sure that all the symptoms were all correctly and consistently annotated based on the symptom descriptions.
# Results
## Study Engagement
A total of 19 participants engaged in the study differently with respect to average level of task difficulty, the number of trials and completions, the number of skips, total game score, and break (Table [2](#tab:summary_participant){reference-type="ref" reference="tab:summary_participant"}). Each participant showed different degrees of engagement but we found some correlations among the variables. Participants who worked at higher difficulty levels solved and skipped fewer puzzles (\(p<\).05) but also spent more time on these difficult puzzles and so attempted fewer. We consider a break as variable and found that participants who did not take a break completed more difficult tasks, higher scores, and had less skips than those who took a break (\(p<\).05). We were interested in which types of symptoms and how much of each appeared in general. In addition, as we found different activity levels between the non-break and break groups, we decided to look into symptoms at the group level. *We did not find any differences in symptoms between difficulty levels.* In the following subsections, we will first summarize 14 symptoms identified in the study and then report characteristics of the symptoms from all participants as a whole and at the group level.
## Observed Symptoms
Symptoms were collected through observations and cross-checked with inter-rater agreement, game event logs, and eye tracking data when possible (the eye tracker would not calibrate on 5 participants). Symptoms in our study were operationalized as follows:
- *Distraction* (D, appeared 17 times) was recorded when participants showed signs of attending to stimuli that did not come from the Apto game environment. This included briefly re-orienting towards hallway noise, diverting attention to their cell phone if it rang, etc.
- *Inattention* (I, N=30) was typically characterized as a participants lack of correction of errors: i.e. some participants picked poorly fitting compounds which caused them to lose points, however eye tracking data confirms they did not always look at their score to pay attention to the effects of their actions.
- *Increased decision time* (IDT, N=73) and *Lack of advancement/progress* (LAP) were both judged based on the time it took a participant to take an action and complete a puzzle relative to their performance during the rest of the session. Participants who took longer than usual to make a single decision were noted for IDT. Participants who took longer than usual to solve a puzzle or never attempted puzzles of higher levels were noted for LAP.
- *Data/Command errors* (DCE, N=34) manifest as actions which were immediately undone and followed by a nearly identical action with preferred outcome (frequently paired with a audible "Oops").
- *Increased Negative Affect* (INA, N=44) was recorded when participants verbal expressed emotional states such as frustration or impatience ("I'm still at 96%? This is so frustrating!").
- *Physical effects* (PE, N=251) were the easiest symptoms for observers to confirm as they included events such as fidgeting and posture changes. The study protocol did not request or encourage a "talk-aloud" protocol so participants who chose to speak were often expressing confusion ("I don't know why that happened"), forgetfulness ("How do I scroll again?"), distraction("The browser is running very slow"), effort/risk over/under estimation ("This is a tough one!"), negative affect ("This \[puzzle\] is taking forever"),or task-unrelated-thoughts ("I'm glad I participated today, this study is interesting!"). These verbalizations were not counted as PE.
- *Confusion* (C, N=27) was typically noted through verbalizations such as \"I don't understand\" or \"What did I just do?\".
- *Effort/Risk Over/Under Estimation* (ERE, N=6) manifested as level-increase actions followed by quick task abandonment and level decreases or verbalizations such as \"This is harder than I thought!\" and \"Level 5 is much harder than it looks.\"
- *Forgetting* (F, N=21) is indicated by participants asking study facilitators to repeat instructions covered in the tutorial or re-reading the tip sheet multiple times. In one case, a participant asked for scratch paper to use as a memory aid to write down the names of compounds she had looked at in solving a puzzle.
- *Strategy inefficiency* (STI, N=53) typically manifest as a lack of help-line use: some participants manually searched for incomplete puzzle pieces instead of using the error or difference lines explicitly described in the tip-sheet.
- *Task abandonment* (TA, N=62) was recorded as use of the game's 'skip puzzle' button.
- *Task Rushing* (TR, N=33) is similar to IDT in that it represents a deviation from some baseline-level of participant actions. In the case of TR, this is a pace beyond what would be expected from experience and naturally improved skill. TR will likely manifest with DCE and I: participants rushing through a task make mistakes but do not correct them right away. TR also manifests as trial-and-error attempts to improve a participant's score on a puzzle (as opposed to carefully searching for a matching compound or using the error/difference/sum line tools).
- *Task Unrelated Thoughts* (TUT, N=27) were noted when participants verbalized their thoughts which had no impact at all on their puzzle solving tasks such as whether or not the random spectrum matched real-world substances (\"Is this hydrogen or something?\") or even what researchers did in their free time (\"You mush all meet up for drinks when this is over!\")
Note that this list is not meant to be exhaustive. It is an overview of the symptoms we learned about from our study. Certain symptoms suggested by related work were not expected to be seen in this study. For instance, habituation occurs when people stop reacting to important information in the form of alerts. The Apto game environment has no such alerts and therefor habituation (among other real-time-response-based symptoms) would not be detected in this study.
### Overall Summary
From a total of 14 symptoms, physical effects (PE) were detected the most followed by increased decision time (IDT), increased task abandonment (ITA), etc. As shown in Table [3](#tab:summary_symptom_participant){reference-type="ref" reference="tab:summary_symptom_participant"}, more than 50% of the participants (10) showed more than 10 symptom types. We also found that the number of symptoms was positively associated with the number of clicks, indicating that the participants tended to click more when they presented symptoms (Figure [\[Symptoms_clicks\]](#Symptoms_clicks){reference-type="ref" reference="Symptoms_clicks"}). The number of symptoms generally increased over time (Figure [\[Symptoms_over_time\]](#Symptoms_over_time){reference-type="ref" reference="Symptoms_over_time"}). We saw the significant drop at the end of the study. When we saw the results at the group level, PED and VR were found to be increased and PJS was decreased over time. As the participants progressed the study, it seems that they showed less task-related judgement symptoms but showed more physical and emotional symptoms. As shown in Figure [\[symptoms_by_type_over_time_5\]](#symptoms_by_type_over_time_5){reference-type="ref" reference="symptoms_by_type_over_time_5"}, each symptom shows a different distribution over time. Linear regression results showed statistically significant increases in physical effects (PE) and inattention (I) and significant decreases in strategy inefficiency (STI) at \(p<\).05. Given the length of the user study (3 hours), an increase in fatigue or decrease in concentration could contribute to PE and I. Learning effects could contribute to a decrease in STI because it is more pertinent to task efficiency. Although not statistically significant, TA generally decreased over time, also possibly due to a learning effect. We also considered a possibility of two symptoms' appearing together or sequentially. Our observation logs indicated that there were lists of symptoms which together in the same one-minute window. We measured symptom pairs and their frequency. Out of 706 individual observations, a total of 456 symptoms in 62 groups of two symptoms (64%) were identified. Figure [\[Symptoms_sequence\]](#Symptoms_sequence){reference-type="ref" reference="Symptoms_sequence"} illustrates the results. As PE occurred the most during the study, it occurred together with many other symptoms. The symptom group with the top two symptoms showed the highest count. Figure [\[Symptoms_CoSymptoms\]](#Symptoms_CoSymptoms){reference-type="ref" reference="Symptoms_CoSymptoms"} presents another aspect of co-occurrence. We measured the percentage of co-occurrence from the total number of occurrences for each symptom. Although the raw number of co-occurrences for PE was the largest, when it comes to percentage, it showed lower than 50%. This indicates that PE also occurs independently and individually. This result applies to IDT (51%) as well. In addition, TA, INA, TR, and I exhibited very high percentages (especially TA, INA, and I showed 100%), indicating that these are the symptoms that could entail or be led by other symptoms.
### Symptoms by Participant and Break Group
We now switch our focus to the results by an individual level and a group level. We had two user groups (i.e., break and non-break groups) for the analysis. Table [3](#tab:summary_symptom_participant){reference-type="ref" reference="tab:summary_symptom_participant"} summarizes the number of symptoms, symptom diversity, and co-symptoms which appeared in one minute (Figure [\[Symptoms_CoSymptoms\]](#Symptoms_CoSymptoms){reference-type="ref" reference="Symptoms_CoSymptoms"}) by each participant. The number of symptoms is strongly correlated to symptom diversity (\(r\)=.75, \(p<\).001) and the number of co-symptoms (\(r\)=.80, \(p<\).0001). However there was no significant difference between break and non-break groups. Figure [\[symptoms_by_participant_over_time\]](#symptoms_by_participant_over_time){reference-type="ref" reference="symptoms_by_participant_over_time"} illustrates the number of symptoms from each participant (clustered into either a break or non-break group) over time (5-minute bin). Red arrows represent breaks. Note that the break was found to be the last event of each time slot. We confirmed that no symptom appeared after the break for all participants. The results indicate that for quite a few participants (i.e., P1, P3, P9, P10, P15, P19) in the break group, the number of symptoms appeared at the same time of the break or one time slot before the break was above the average. Figure [\[Symptoms_by_type_over_time_5\_break_non_break\]](#Symptoms_by_type_over_time_5_break_non_break){reference-type="ref" reference="Symptoms_by_type_over_time_5_break_non_break"} shows a variance of each symptom by the break and non-break groups. On the one hand, the break group showed the results that were quite consistent with the ones from all participants (see Figure [\[symptoms_by_type_over_time_5\]](#symptoms_by_type_over_time_5){reference-type="ref" reference="symptoms_by_type_over_time_5"}). F, I, IDT, and PE generally increased while C and STI decreased. On the other hand, for the non-break group three symptoms (i.e., I, PE, and TR) increased and other three symptoms (i.e., STI, TA, and TUT) decreased. Especially for TA and TUT, two groups showed quite opposite results. It appears that the participants in the break group tended to be less focused than those in the non-break group. Another interesting finding is that the non-break group showed more PE than the break group. We were interested in the effects of breaks and investigated the differences in the types and frequencies of symptoms during the 10 minutes before and the 10 minutes after breaks. Figure [\[Symptom_diff_before_after_break\]](#Symptom_diff_before_after_break){reference-type="ref" reference="Symptom_diff_before_after_break"} showed several interesting insights. First, PE, C, and ERE appeared more before the break. PE appeared 17 times more before breaks, which indicates it could be a sign of one's intention to take a break. Second, for some symptoms, we found differences in frequency between their overall trends (see Figure [\[Symptoms_by_type_over_time_5\_break_non_break\]](#Symptoms_by_type_over_time_5_break_non_break){reference-type="ref" reference="Symptoms_by_type_over_time_5_break_non_break"}, top) and break periods. PE and IDT generally increased over time, but when it comes to the time around breaks, they appeared more before the break. TA, INA, and C appeared more before the break. Participants showed more physical effects (PE), increased decision time (IDT), task abandonment (TA), negative effects (INA), and confusion (C) before they decided to take a break. This indicates that these symptoms could be considered as possible signs of cognitive depletion and factors that could indicate that someone is in need of a break. Our last investigation in this section was looking into the participants in the break group. In the previous section, we showed generally lower performance by the break group than the non-break group. However we realized that some participants actually showed good performance and engagement (e.g., P6, P8). Thus, we measured differences in symptoms between two groups of participants based on their performance. We used the median of the scores to have the high and low groups (7 participants were assigned to each group). Figure [\[Symptoms_by_low_and_high_performance_group\]](#Symptoms_by_low_and_high_performance_group){reference-type="ref" reference="Symptoms_by_low_and_high_performance_group"} shows the results. Both the low and high groups showed a decrease in PJS; however, showed different results in PED and VR. This low group still showed increases in PED and VR even they took breaks during the study.
### Symptoms with Eye-Tracking Data
Our last investigation was to map the eye-tracking data to the observed symptoms. A pupil dilation response points to a level of emotion, arousal, stress, pain, or cognitive load of a human. In this regard, we were especially interested in the relationship between pupil dilation and the number of symptoms both by an individual level and a group level. For the group level, Spearman's correlation results (which is a nonparametric measure of the strength and direction of association) show that both break and non-break groups did not show any significant difference in pupil dilation. However, when looking into individual differences, we found that two participants from the break group showed significantly positive relationships (P13: \(\rho\)=0.28, \(p<\)`<!-- -->`{=html}0.05; P19: \(\rho\)=0.23, \(p<\)`<!-- -->`{=html}0.05), meaning that more symptoms appeared as the average of the pupil dilation increased. We took a look at them more closely. It turned out that P13 and P19 were in the break group and also showed the lowest performance and least engagement during the study. As shown in Figure [\[P16_P23_and_Others\]](#P16_P23_and_Others){reference-type="ref" reference="P16_P23_and_Others"}, these two participants worked primarily on low level (easy) tasks, had significantly fewer task trials, more symptoms, and lower game scores compared to the rest of the participants. This demonstrates the effect of cognitive loads and loss of concentration well but it is unclear why these two participants were affected more than others.
# Discussion
We gained a number of insights from this study. Firstly, while the total number of symptoms tended to increase over time, individual symptoms varied. Physical effects (PE), forgetting (F), inattention (I), and task rushing (TR) generally increased, whereas strategy inefficiency (STI) and task abandonment (TA) decreased over time. At the higher symptom-group level, physical and emotional deregulation (PED), and vigilance and reactionary symptoms (VR) increased, whereas personal judgement symptoms (PJS) decreased over time. This result is likely due to learning effects: as participants gained familiarity over the course of the three hour study, they became more confident in completing puzzles and had developed the most efficient strategies for doing so. But even as participants learned, cognitive depletion produced other effects to make them appear restless and impatient. The number of symptoms and that of clicks are positively correlated, which indicates that a degree of clicks could be used as a signal of symptoms. Future work will verify this result and investigate how click behavior might be used as a metric of cognitive depletion. An increase in symptoms seems to be a predictor of a break. Six participants (P1, P3, P9, P10, P15, P19) in the break group showed an above average number of symptoms at the same time as a break or one time slot before a break. This gives us an opportunity to provide systematic support (e.g., encouraging users to take a short break) to individuals when they show more symptoms over time. Particularly, some symptoms (e.g., PE, IDT, TA, INA, C) appeared more before the break compared to after breaks. Although in general both the break and non-break groups showed similar patterns of symptoms over time, TA and TUT were found to be quite different between the groups. It appears that the participants in the break group tended to be less focused than those in the non-break group. We identified additional two sub-break groups based on their performance. It turned out that the low performance group still showed an increase in PED and VR, whereas the high performance group showed a decrease. Systematic supports or suggestions could be applied to people in the low performance group to lower cognitive depletion and recover concentration.
## Limitations and Future Work
Although our study gives us a number of interesting insights, there exist some limitations which we believe will be addressed in our future study. First, our findings may not be generalizable due to small sample size. Results could also be sensitive on other characteristics of individuals like their experience and expertise in using computer systems, concentration levels, work strategies, etc. We plan to conduct a more refined user study with more participants, and one of our focuses will be to see repeatability of the results. Second, although we were confident that symptoms were correctly detected during the study, we also acknowledge that human errors and bias may exist. Presumably, some symptoms were easy to detect (e.g., most people would easily identify yawning) while others might be more difficult. We believe that symptoms would be better detected and understood by combining system usage logs and human annotations, which will be a primary method in our future study. Lastly, expanding the second point, cognitive depletion may manifest in more ways than the 14 symptoms observed in our study. For example, in many real scenarios, people sometimes work on multiple tasks. A cognitive depleted person may attempt to unnecessarily complete multiple tasks at once, increase the number of simultaneous tasks they attempt, switch between concurrent tasks more frequently, or make mid-progress switches between tasks more often. Moreover, a cognitively depleted person's strategy for handling pushed information (e.g., emails, updates, text messages, phone calls, etc.) will fail and become unable to continue tasks. Our current protocol and study environment was not designed to produce these conditions or provide an opportunity for related symptoms to be observed.
# Conclusion
This paper presents the results of a three hour user study designed to induce cognitive depletion in 19 participants. Through the collection of different datasets including activity logs, observations, and eye-tracking data and through the individual and group-level analyses, we have gained insight into how cognitive depletion manifests in novice NMR analysts. Our work provides early insights into cognitive depletion in human computation and addresses future research implications of the findings. Future work will focus on repeatability of these results, building models that will predict cognitive status and depletion, and introducing interventions target to relieve cognitive depletion and improve overall task performance.
[^1]: http://www.chenomx.com/software/
[^2]: http://www.tobiipro.com/product-listing/tobii-pro-glasses-2/ | {'timestamp': '2017-06-08T02:01:05', 'yymm': '1706', 'arxiv_id': '1706.01919', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01919'} |
# introduction
In many areas of applied mathematics, various types of special functions become essential tools for scientists and engineers. The continuous development of mathematical physics, probability theory and other areas has led to new classes of special functions and their extensions and generalizations. The study of one-variable hypergeometric functions appear in the work of Euler, Gauss, Riemann, and Kummer and the integral representations were studied by Barnes and Mellin. The special properties of one variable hypergeometric function were studied by Schwarz and Goursat. For more details about the recent works in the field of dynamical systems theory, stochastic systems, non-equilibrium statistical mechanics and quantum mechanics, the readers may refer to the recent work of the researchers and the references cited therein. Throughout this paper, we denote by \(\mathbb{N}, \mathbb{Z}^-\) and \(\mathbb{C}\) the sets of positive integers, negative integers and complex numbers, respectively, and also \(\mathbb{N}_0:= \mathbb{N} \cup\{0\}\) and \(\mathbb{Z}_0^-:=\mathbb{Z}^-\cup\{0\}\). The extended generalized hypergeometric function is defined by Srivastava *et al.* : \[\begin{aligned}
\label{1} _{r}F_{s}\left[\begin{array}{c l} (a_{1},p),a_{2},\cdots, a_{r};\quad \quad \quad\\\left. \quad \quad \quad\quad \quad \quad\quad \quad \quad\quad z \right.\\\left. b_{1},\cdots, b_{s};\quad\quad\quad\,\right. \end{array}\right]=\sum_{n=0}^{\infty}\frac{(a_{1};p)_{n}(a_{2})_{n}\cdots (a_{r})_{n}}{(b_{1})_{n}\cdots (b_{s})_{n}}\frac{z^{n}}{n!} \end{aligned}\] where, in terms of generalized Pochammer symbol \((\mu;p)_{v}\): \[(\mu;p)_{v}= \begin{cases} \frac{\Gamma_{p}(\mu+v)}{\Gamma(v)},& (\Re(p)>0, \mu,v \in \mathbb{C})\\ (\mu)_{v}, & (p=0, \mu, v\in \mathbb{C}) \end{cases}\] Here \(\Gamma_{p}(z)\) is the generalized gamma function introduced by Chaudhry and Zubair as follows: \[{\Gamma}_{p}(z)= \begin{cases} \int_{0}^{\infty}{t^{z-1}exp\left(-p-\frac{p}{t}\right)dt},& (\Re(p)>0, z \in \mathbb{C})\\ {\Gamma}(z), & (p=0, \Re(z)>0) \end{cases}\] The corresponding extensions of Gauss's hypergeometric and confluent hypergeometric functions are as follows: \[\begin{aligned}
\label{2} _{2}F_{1}\left[\begin{array}{c l} (a,p),\beta;\quad \quad\\\left. \quad \quad\quad \quad \quad\quad z \right.\\\left. \gamma;\quad\,\right. \end{array}\right]=\sum_{n=0}^{\infty}\frac{(a,p)_{n}\beta_{n}z^n}{(\gamma_{n}n!}, \end{aligned}\] and \[\begin{aligned}
\label{3} _{1}F_{1}\left[\begin{array}{c l} (a,p);\quad \quad\\\left. \qquad \qquad\qquad z \right.\\\left. \gamma;\,\quad\right. \end{array}\right]=\sum_{n=0}^{\infty}\frac{(a,p)_{n}z^n}{(\beta_{1})_{n}(\gamma_{n}n!}, \end{aligned}\] where \(a,\gamma\in \mathbb{C}\). In this paper, we derive two new integral formulas involving the generalized hypergeometric function ([\[1\]](#1){reference-type="ref" reference="1"}). Further we give corollaries as special cases for the extended Gauss' hypergeometric and confluent hypergeometric functions. For the present investigation, we need the following result of Oberhettinger. \[\begin{aligned}
\label{5} \int_0^\infty z^{\alpha-1} (z+b+\sqrt{z^{2}+2bz})^{-\beta} dz=2\beta b^{-\beta}\left(\frac{b}{2}\right)^{\alpha}\frac{\Gamma(2\alpha) \Gamma(\beta-\alpha)}{\Gamma(1+\alpha+\beta)}. \end{aligned}\] where \(0<\Re(\alpha) <\Re(\beta)\). For various other investigations involving certain special functions, interested reader may be referred to several recent papers on the subject (see, for example, [@Baleanu; @Choi; @Choi1; @Srivastava1; @Srivastava3; @Srivastava4] and the references cited in each of these papers.
# Main Result
In this section, the generalized integral formulas involving the extended generalized hypergeometric function defined in ([\[1\]](#1){reference-type="ref" reference="1"} are established here by inserting with the suitable argument in the integrand of ([\[5\]](#5){reference-type="ref" reference="5"}) and we express the obtained result in terms of an extended Wright-type hypergeometric function.
# Special Cases
In this section, we present certain special cases of (2:1) and (2:2) as corollaries given below for extended Gauss' hypergeometric and confluent hypergeometric functions (1:4) and (1:5). | {'timestamp': '2017-06-08T02:01:21', 'yymm': '1706', 'arxiv_id': '1706.01931', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01931'} |
null | null |
# Introduction
Interferometers, the most precise metrological instruments constructed by humans, have played a major role in many break-through experiments. In the famous Michelson-Morely failed attempt to detect the ether, the device, now called the "Michelson interferometer", was used. The negative result of this experiment was relevant for the subsequent formulation of the special theory of relativity. Almost 130 years later, the LIGO team used a similar Michelson setup to detect the gravitational wave coming from a collision of two black holes, which took place over 1 bilion years ago. At the moment of the detection, the sensitivity of the LIGO instruments was high enough to observe the displacement \(\Delta L\) of the test masses \(\sim10^{-22}\) times smaller than the length \(L\) of the interferometric arms. With \(L\simeq4\) km, this gives a truly impressive \(\Delta L\simeq10^{-19}\) m, which is \(\sim10^3\) times smaller than the radius of a proton. Matter-wave interferometers at this moment can measure the gravitational acceleration \(g\) with the sensitivity of \(\Delta g\simeq10\frac{\rm nm}{\rm s}\). Once miniaturized, such device could serve as an ultra-precise geological instrument. An atomic interferometer calibrated through the measurement of the close to the surface Casimir-Polder forces could yield information about the gravitational constant \(G\) and the possible deviations from the Newton \(1/r^2\) scaling of the gravitational force at small distances. The gravity-field curvature has recently been observed using an atomic interferometer. Interference of the matter waves is also used in the ultra-precise measurements of the fine-structure constant \(\alpha\), which is of fundamental importance. At the current stage, these devices operate at best at the shot-noise level, i.e., the sensitivity of the estimation of the parameter \(\theta\) does not break the shot-noise limit (SNL), \(\Delta\theta\propto\frac1{\sqrt N}\). Here \(N\) is the number of probes---for instance atoms or photons---which carry the information about the interferometric phase. However, theoretical results and multiple proof-of-principle experiments show that the entanglement between these probes is a resource for the sub shot-noise (SSN) sensitivity. Those experiments follow different routes to create non-classical states of light or matter. For instance, the LIGO interferometer already displayed the SSN sensitivity when one of its input ports was fed with a squeezed state of light. For interferometers operating on matter--waves, the non-classicality is associated with the entanglement between the particles. This effect often manifests through the spin-squeezing of the sample . It is a phenomenon---quantified by the spin-squeeizng parameter---associated with the two-mode algebra. The use of this algebra is quite natural when discussing the interferometric problems---the two arms are identified with the two modes. Most of the interferometric arguments, such as that relating the entanglement to the SSN sensitivity, are also invoked within this two-mode description. However, in principle, the signal propagating through each arm can have a rich multi-mode structure. This might be a result of thermal excitations, as witnessed in, or the inherently multi-mode nature of the process generating the entangled sample, as in. Here we generalize the central theorem of quantum metrology---that relating the SSN sensitivity to the entanglement---to the case when each arm of the interferometer has a complicated multi-mode structure. We show that the particle entanglement remains the key resource for beating the SNL. Our proof is of complete generality---it does not make any assumption about the state and works for identical and distinguishable particles or combination of both. Also, it applies to systems with a fixed or fluctuating number of particles. We derive the sensitivity of the phase estimation from the measurement of the population imbalance between the two arms. Finally, we show how the method of detecting the particle entanglement, which works for the two-mode systems, can incorrectly indicate the presence of non-classical correlations in the multi-mode configurations. This work is organized as follows. In Section [2](#sec.ent){reference-type="ref" reference="sec.ent"} we derive the relation between the entanglement and the SSN sensitivity for identical (Section [2.1](#sec.ent.id){reference-type="ref" reference="sec.ent.id"}) and distinguishable particles (Section [2.2](#sec.ent.dist){reference-type="ref" reference="sec.ent.dist"}). Other configurations are discussed in Section [2.3](#sec.ent.other){reference-type="ref" reference="sec.ent.other"}. In Section [3](#sec.est){reference-type="ref" reference="sec.est"} we calculate the phase sensitivity for the estimation based on the knowledge of the average population imbalance between the two arms. In Section [\[sec.det\]](#sec.det){reference-type="ref" reference="sec.det"} we show that the entanglement witness often used for two-mode systems, can give false results in the multi-mode setups. The summary is contained in Section [4](#sec.sum){reference-type="ref" reference="sec.sum"}.
# Entanglement and SSN sensitivity in multi-mode systems {#sec.ent}
## Bosons {#sec.ent.id}
### Multi-mode interferometric transformations
Let \(a\) and \(b\) denote the two arms of an interferometer. Here, for illustration, we will assume that these arms are spatially separated, however it could be the momentum or other degree of freedom, such as the fine structure, that distinguish \(a\) from \(b\). In the standard two-mode case, with each arm a single operator \(\hat a\) and \(\hat b\) is associated. The interferometric transformations are generated by the angular-momentum operators
To account for the multi-mode structure of each arm, we introduce the bosonic field operator \(\hat\Psi(\mathbf{r})\) which consists of two parts, i.e., \[\label{field} \hat\Psi(\mathbf{r})=\hat\Psi_a(\mathbf{r})+\hat\Psi_b(\mathbf{r}).\] Our aim is to construct the interferometric transformation which will act on the multi-mode arms \(a\) and \(b\), rather than on two modes only. It is clear, that such operations can be constructed in many different ways and that the interferometer's performance depends on the structure of the regions as well as on our choice of transformations. To limit the number of possibilities, we will first assume that the mode structure of \(a\) and \(b\) is the same, and it is only the spatial separation that makes the distinction between them, see Fig. [\[fig.ladder\]](#fig.ladder){reference-type="ref" reference="fig.ladder"}. Therefore, each operator can be expanded into its corresponding basis
and the spatial wave-functions are shifted by \(\mathbf d\), i.e., \[\label{rel1} \psi^{(n)}_b(\mathbf{r}+\mathbf{d})=\psi^{(n)}_a(\mathbf{r}).\] Also, we assume \(|\mathbf d|\) to be much larger than the characteristic widths of the wave-packets, giving \[\label{rel2} \psi^{(n)}_{a}(\mathbf{r})\psi^{(m)}_{b}(\mathbf{r})=0\ \ \ \forall\ n,m.\] Once the spatial structure of the system is determined, we pick the interferometric transformations. To establish the analogy with the two-mode case, we consider two types of transformations---the mode mixing and the phase imprint. A natural extension of ([\[J_2m\]](#J_2m){reference-type="ref" reference="J_2m"}) is
These integrals can be calculated using relations ([\[rel1\]](#rel1){reference-type="ref" reference="rel1"}) and ([\[rel2\]](#rel2){reference-type="ref" reference="rel2"}), and the outcome is
Therefore, the mode mixing operators act on each pair of modes separately. This is a result of 1\(^\circ\) the assumption about the symmetry between the regions and 2\(^\circ\) the particular form of the coupling in Equations ([\[J_mm\]](#J_mm){reference-type="ref" reference="J_mm"}). While 1\(^\circ\) can be regarded as unnatural or highly idealistic, such symmetry is encountered for instance in twin-beam configurations formed by the scattering of atoms from a Bose-Einstein condensate. On the other hand, 2\(^\circ\) seems a natural choice: the interferometer should simply "copy" a particle from one arm to another. This second condition can be expressed in other words: the interferometric transformation does not use any knowledge about the internal structure of the arms. It treats each arm as a whole and does not penetrate the inner structure. The multi-mode interferometric transformations which are single-particle operations (i.e., do not entangle the resources) are usually considered in a form \[\label{trans} \hat U(\theta)=e^{-i\theta\hat J_n}.\] Here \(\hat J_n=\vec n\cdot\hat{\vec J}\) is a scalar product of a unit vector \(\vec n\) and a vector of the angular momentum operators [^1]. We will now demonstrate that, in analogy to the two-mode case, the particle entanglement is a necessary resource for beating the shot-noise limit of the phase estimation also in the multi-mode case.
### Role of the particle entanglement {#sec.qfi}
Let us begin with a two-mode separable (i.e., non-entangled) pure state of \(N\) bosons. It is the spin-coherent state \[\label{css_2m} \ket{z,\varphi;N}=\frac1{\sqrt{N!}}\left(\sqrt ze^{i\varphi}\hat a^\dagger+\sqrt{1-z}\hat b^\dagger\right)^N\ket0.\] Here, \(z\in[0,1]\) is the population imbalance between the two modes, while \(\varphi\in[0,2\pi]\) is the relative phase. This state is a basic building block of the density matrix of \(N\) non-entangled bosons, which reads \[\label{sep_2m} \hat\varrho_{\rm sep}=\int_0^1\!\!dz\int_0^{2\pi}\!\!\!d\varphi\,{\mathcal P}(z,\varphi)\ket{z,\varphi;N}\bra{z,\varphi;N},\] where \({\mathcal P}\) is a probability distribution of the variables \((z,\varphi)\). The expression ([\[sep_2m\]](#sep_2m){reference-type="ref" reference="sep_2m"}) is the fixed-\(N\) analog of the classical state of the electromagnetic field, according to the Glauber-Sudarshan criterion. The expressions ([\[css_2m\]](#css_2m){reference-type="ref" reference="css_2m"}) and ([\[sep_2m\]](#sep_2m){reference-type="ref" reference="sep_2m"}) can be easily generalized to the multi-mode setup. Namely, the former transforms to \[\label{css_mm} \ket{z,\varphi;N}\rightarrow\ket{\vec\alpha,\vec\beta;N}=\frac1{\sqrt{N!}}\left(\vec\alpha\,\hat{\vec a}^\dagger+\vec\beta\,\hat{\vec b}^\dagger\right)^N\ket0.\] The vectors of complex amplitudes \(\vec\alpha\) and \(\vec\beta\) are normalized, i.e, \(|\vec\alpha|^2+|\vec\beta|^2=1\), while \(\hat{\vec a}\) and \(\hat{\vec b}\) are vectors of mode operators introduced in Eq. ([\[ops\]](#ops){reference-type="ref" reference="ops"}). Similarly, the density matrix of unentangled bosons now reads \[\label{sep_mm} \hat\varrho_{\rm sep}=\iint\!\! d\vec\alpha d\vec\beta\,{\mathcal P}(\vec\alpha,\vec\beta)\,\ket{\vec\alpha,\vec\beta;N}\bra{\vec\alpha,\vec\beta;N}.\] We will show that for the separable states ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}) and the interferometric transformations ([\[trans\]](#trans){reference-type="ref" reference="trans"}), the phase estimation sensitivity is bounded by the shot-noise. The sensitivity of the phase estimation is limited by the Cramer-Rao lower bound \[\label{crlb} \Delta\theta\geqslant\frac1{\sqrt m}\frac1{\sqrt{F_q}}.\] Here, \(m\) is the number of the independent repetitions of the estimation experiment. The \(F_q\) is called the quantum Fisher information (QFI) and it quantifies the amount of information about \(\theta\), which can be extracted from \(m\) measurements using any estimation strategy. For pure states, the QFI is simple to calculate, \[\label{pure} F_q=4(\av{\hat J_n^2}-\av{\hat J_n}^2)\equiv4\av{(\Delta\hat J_n)^2},\] where the mean is calculated with the state undergoing the interferometric transformation. For mixed states, the QFI is much more complex, \[F_q=2\sum_{i,j}\frac{(p_i-p_j)^2}{p_i+p_j}\left|\bra i\hat J_n\ket j\right|^2,\] where \(\ket{i/j}\) are the eigen-vectors and \(p_{i/j}\) are the corresponding eigen-values of the density matrix. Therefore, to obtain the \(F_q\), one would need to diagonalize \(\hat\varrho_{\rm sep}\) from Eq. ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}), which is numerically feasible, but analytically very hard, since different \(\ket{\vec\alpha,\vec\beta;N}\)'s are not orthogonal, just as nonorthogonal are the coherent states of light. However, a useful feature of the QFI---its convexity---allows to lower-bound the sensitivity for mixed stated. Namely, for \(\hat\varrho=\sum_ip_i\hat\varrho_i\), i.e., a statistical mixture of density matrices, we have \[F_q\left[\sum_ip_i\hat\varrho_i\right]\leqslant\sum_ip_iF_q[\hat\varrho_i].\] This property, applied to Eq. ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}), gives \[\label{convex} F_q\leqslant\iint\!\! d\vec\alpha d\vec\beta\,{\mathcal P}(\vec\alpha,\vec\beta)F_q^{(\vec\alpha,\vec\beta)}.\] Here, \(F_q^{(\vec\alpha,\vec\beta)}\) is the QFI calculated with a pure state \(\ket{\vec\alpha,\vec\beta;N}\), thus it is given by Eq. ([\[pure\]](#pure){reference-type="ref" reference="pure"}). We calculate the QFI for \(\hat J_{\vec n}=\hat J_z\) and show that for separable states, \(F_q\leqslant N\). Any other direction \(\vec n'\) can be obtained by a series of rotations of \(\hat J_z\), generated by the operators ([\[mJ\]](#mJ){reference-type="ref" reference="mJ"}). They are single-particle objects, therefore, once applied to the separable state ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}), rather than to the evolution operator, they would transform one \(\hat\varrho_{\rm sep}\) into another, i.e., would only change \(\mathcal P(\vec\alpha,\vec\beta)\) into some \(\tilde{\mathcal P}(\vec\alpha,\vec\beta)\), but will not modify the structure of Eq. ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}). Thus it is enough to show that \(F_q\leqslant N\) for \(\hat J_z\) and a general probability distribution. First, we calculate \(F_q^{(\vec\alpha,\vec\beta)}\) with Eq. ([\[pure\]](#pure){reference-type="ref" reference="pure"}). The mean of \(\hat J_z\) is \[\label{msq} \av{\hat J_z}=\frac N2\sum_n\left(|\alpha_n|^2-|\beta_n|^2\right)\equiv\frac N2(|\vec\alpha|^2-|\vec\beta|^2),\] where we used the expression ([\[mJz\]](#mJz){reference-type="ref" reference="mJz"}) and the relation
The second moment is calculated in a similar way, i.e., \[\av{\hat J_z^2}=\sum_{n,m}\av{\hat J_z^{(n)}\hat J_z^{(m)}}=\sum_{n}\av{\left(\hat J_z^{(n)}\right)^2}+\sum_{n\neq m}\av{\hat J_z^{(n)}\hat J_z^{(m)}}.\] The two contributions must be treated separately \[\begin{aligned}
\sum_n\av{\left(\hat J_z^{(n)}\right)^2}&=&\frac N4\left[1+(N-1)\sum_n(|\alpha_n|^2-|\beta_n|^2)^2\right],\\ \sum_{n\neq m}\av{\hat J_z^{(n)}\hat J_z^{(m)}}&=&\frac{N(N-1)}4\times\\\ &\times&\sum_{n\neq m}(|\alpha_n|^2-|\beta_n|^2)(|\alpha_m|^2-|\beta_m|^2). \end{aligned}\] We add these two terms and obtain \[\av{\hat J_z^2}=\frac N4+\frac{N(N-1)}4(|\vec\alpha|^2-|\vec\beta|^2)^2.\] The substraction of the squared mean from Eq. ([\[msq\]](#msq){reference-type="ref" reference="msq"}) gives \[\label{bound} 0\leqslant F_q^{(\vec\alpha,\vec\beta)}=N\left(1-(|\vec\alpha|^2-|\vec\beta|^2)^2\right)\leqslant N.\] This result, combined with Eq. ([\[convex\]](#convex){reference-type="ref" reference="convex"}) yields \[\label{snl} F_q\leqslant N.\] Since the QFI can surpass the SNL, for instance with two-mode spin-squeezed states, we conclude that the particle entanglement is the resource for the SSN metrology also in multi-mode systems. We now extend the above formalism to distinguishable particles.
## Distinguishable particles {#sec.ent.dist}
### Multi-mode interferometric transformations
Note that the formalism of second quantization allows for a quick generalization of the above results to distinguishable particles. Namely, another index must be attributed to the operators \(\hat a_n\) and \(\hat b_n\), such that labels the species of the particle (and of the associated field). This way, for the particle of type \(j\), we obtain \(\hat a^{(j)}_n\) and \(\hat b^{(j)}_n\). The interferometric transformations cannot transmute a particle of one type into another---such process would violate the conservation laws and the related super-selection rules. Thus a particle of each type undergoes a separate transformation, which means that the operators ([\[J_mm\]](#J_mm){reference-type="ref" reference="J_mm"}) change into
Once the interferometric transformations are determined, we discuss the role of the particle entanglement for the SSN sensitivity.
### Role of the particle entanglement {#role-of-the-particle-entanglement}
The separable state of \(N\) distingishable particles is constructed from the one-body pure states. For the particle of type \(j\) distributed among \(a\) an \(b\) it reads \[\label{css_dist} \ket{\vec\alpha^{(j)},\vec\beta^{(j)};N}=\left(\vec\alpha^{(j)}\hat{\vec a}^{(j)\dagger}+\vec\beta^{(j)}\hat{\vec b}^{(j)\dagger}\right)\ket0.\] The parrallel with the coherent state from Eq. ([\[css_mm\]](#css_mm){reference-type="ref" reference="css_mm"}) is evident and it is even more pronounced when the separable state of distinguishable particles is introduced
The state of \(N\) distinguishable and non-entangled particles is the state ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}) but with each mode---now labeled with two numbers \(n\) and \(j\), rather than with \(n\) only---occupied with one particle. The deep analogy holds since the transformation ([\[op_dist\]](#op_dist){reference-type="ref" reference="op_dist"}) cannot transmute the particles. Once this close relation is noticed, the QFI can be bounded from above in a similar fashion to that presented in Section [2.1.2](#sec.qfi){reference-type="ref" reference="sec.qfi"} and the calculation is straightforward. We again pick the interferometric transformation to be generated by \(\hat J_z\) and using the convexity of the QFI declared in Eq. ([\[convex\]](#convex){reference-type="ref" reference="convex"}), we have
Here we used the fact that the operator ([\[Jz_mm\]](#Jz_mm){reference-type="ref" reference="Jz_mm"}) acts on each particle independently. Every \(F_q^{(\vec\alpha^{(j)},\vec\beta^{(j)})}\) is calculated with a single-particle state ([\[css_dist\]](#css_dist){reference-type="ref" reference="css_dist"}), therefore it is bounded as in Eq. ([\[bound\]](#bound){reference-type="ref" reference="bound"}) but with \(N=1\) (because there is only a single particle of a given type), namely \[0\leqslant F_q^{(\vec\alpha^{(j)},\vec\beta^{(j)})}=\left[1-(|\vec\alpha^{(j)}|^2-|\vec\beta^{(j)}|^2)^2\right]\leqslant 1.\] Therefore the sum from Eq. ([\[qfi_dist\]](#qfi_dist){reference-type="ref" reference="qfi_dist"}) is bounded as follows \[\sum_{j=1}^NF_q^{(\vec\alpha^{(j)},\vec\beta^{(j)})}\leqslant N.\] Since \(\mathcal{P}\) from Eq. ([\[qfi_dist\]](#qfi_dist){reference-type="ref" reference="qfi_dist"}) is normalized we obtain \[F_q\leqslant N\] for all separable states of distinguishable particles defined in Eq. ([\[sep_dist\]](#sep_dist){reference-type="ref" reference="sep_dist"}). On the other hand, take an entangled NOON state of \(N\) distinguishable particles---for instance each in the same spatial mode---in the form \[\ket\psi=\frac1{\sqrt2}\left(\prod_{j=1}^N\hat a^{(j)\dagger}+\prod_{j=1}^N\hat b^{(j)\dagger}\right)\ket0.\] This state will give \(F_q=N^2\) (the Heisenberg scaling), therefore the QFI is a witness of particle entanglement also for distinguishable particles.
## Other cases {#sec.ent.other}
For the story to be complete we must consider two other possible cases. One is when the system contains both distinguishable particles and bosons, all together forming a separable state. Again, we repeat that the interferometric transformations---to be consistent with the conservation laws and the super-selection rules---cannot transmute the particles from these two groups into each other. Therefore, since they form a separable state where the mode occupied by each particle can be addressed individually, the set of distinguishable particles and bosons can be treated separately. Thus to each of these sets the arguments from the above sections apply, so also in such configuration the particle-entanglement will be a necessary resource to beat the shot-noise limit. Another possibility is that the system does not contain a fixed number of particles, but rather its amount fluctuates from shot to shot, governed by the probability distribution \(P(N)\). The separable state is now \[\hat\varrho_{\rm sep}=\sum_{N=0}^\infty P(N)\hat\varrho^{(N)}_{\rm sep},\] where \(\hat\varrho^{(N)}_{\rm sep}\) contains \(N\) particles and is given either by Eq. ([\[sep_mm\]](#sep_mm){reference-type="ref" reference="sep_mm"}) or ([\[sep_dist\]](#sep_dist){reference-type="ref" reference="sep_dist"}) or by a mixture of those, as discussed in the above paragraph. Since the operators \(\hat J_i\) do not couple states with different number of particles, each fixed--\(N\) sector can be treated separately. Therefore, using again the convexity of the QFI, in all the cases we obtain for separable states \[F_q\leqslant\sum_{N=0}^\infty P(N)N\equiv\av{N},\] which defines the shot-noise limit. Finally, we note that these arguments do not apply to collections of fermions, for which the separable states do not exist due to the Pauli exclusion principle.
# Estimation from the mean population imbalance {#sec.est}
We now abandon the general considerations and switch to a particular phase estimation protocol. We derive the phase sensitivity for a multi-mode Mach-Zehdner interferometer (MZI). We take the most common estimation protocol, where the phase is deduced from the average population imbalance between the two regions. Although the derivation is done for bosons, according to the above arguments the results apply also to distinguishable particles and collections of both. In analogy to the two-mode case, the multi-mode MZI evolution operator is \[\label{evo} \hat U(\theta)=e^{-i\theta\hat J_y},\] with the generator given by Eq. ([\[y_mm\]](#y_mm){reference-type="ref" reference="y_mm"}). At the output, in the \(i\)-th repetition of the experiment, the number of particles in each arm is measured, i.e., \(n_a^{(i)}\) and \(n_b^{(i)}\). From this data, the population imbalance between the two sub-systems is calculated, \(n_i=\frac{n_a^{(i)}-n_b^{(i)}}2\), and the sequence is repeated \(m\) times and averaged to give \[\overline n_m=\frac1m\sum_{i=1}^mn_i.\] If \(m\gg1\), the central limit theorem tells that the probability for obtaining \(\overline n_m\) is a Gaussian function \[p(\overline n_m)\propto\exp\left[-m\frac{\left(\overline n_m-\av{\hat J_z(\theta)}\right)^2}{2\av{(\Delta\hat J_z(\theta))^2}}\right].\] When the phase is estimated from the maximum of the likelihood function for this probability, we obtain \[\label{sens_mzi} \Delta^2\theta=\frac1m\frac{\av{(\Delta\hat J_z(\theta))^2}}{\left(\frac{\partial\av{\hat J_z(\theta)}}{\partial\theta}\right)^2}.\] The \(\hat J_z(\theta)\) is obtained by evolving the \(\hat J_z\) with the operator ([\[evo\]](#evo){reference-type="ref" reference="evo"}), i.e, \[\label{Jz_evo} \hat J_z(\theta)=\hat U^\dagger(\theta)\hat J_z\hat U(\theta)=\hat J_z\cos\theta+\hat J_x\sin\theta,\] where the last equality was obtained using the commutation relations of operators ([\[J_mm\]](#J_mm){reference-type="ref" reference="J_mm"}) and the Baker--Campbell--Hausdorff formula. Substitution of the result ([\[Jz_evo\]](#Jz_evo){reference-type="ref" reference="Jz_evo"}) into Eq. ([\[sens_mzi\]](#sens_mzi){reference-type="ref" reference="sens_mzi"}) gives for \(\theta=0\) \[\Delta^2\theta=\frac1m\frac{\av{(\Delta\hat J_z)^2}}{\av{\hat J_x}^2}.\] This is a natural extensions of the spin-squeezing parameter (for balanced systems, where \(\av{\hat J_y}=0)\), to the multi-mode case. The combination of Equations ([\[crlb\]](#crlb){reference-type="ref" reference="crlb"}) and ([\[snl\]](#snl){reference-type="ref" reference="snl"}) means that also \[\label{ss_mm}\] is a witness of quantum correlations among the particles---when \(We now show that the measurement of this quantity in the multi-mode system is not as straightforward as in the two-mode case. \section{Detection of entanglement---fluctuations and visibility}\label{sec.det} First however we recall that for the two-mode systems, the spin-squeezing parameter is defined as \begin{equation} \end{equation} Here\)`<!-- -->`{=html}1,2,3\(are three orthogonal directions obtained by the combinations of\)J_x\(,\)J_y\(and\)J_z\(from Equations (\ref{J_mm}). When\) It it most common to take \((1,2,3)\) as \((x,y,z)\), giving \[\label{ss}\] The motivation for this particular choice lies in direct experimental accessibility both to the nominator and the denominator of Eq. ([\[ss\]](#ss){reference-type="ref" reference="ss"}). This can be particularly easily seen when the average number of atoms in each mode is high. In this case, the mode operators can be approximated as follows \[\label{mf} \hat a\simeq\sqrt{n_{a}}e^{i\frac\phi2},\ \ \ \hat b\simeq\sqrt{n_{b}}e^{-i\frac\phi2}.\] Here, \(n_{a/b}\) are the average mode occupations and \(\phi\) is the relative phase, which fluctuates from shot to shot. In this approximation, the average of \(\hat J_x\) and \(\hat J_y\) is \[\av{\hat J_x}=\sqrt{n_an_b}\av{\cos\phi},\ \ \ \av{\hat J_y}=\sqrt{n_an_b}\av{\sin\phi}.\] Similarly, the atom number fluctuations normalized to the shot-noise is \[\label{fluct}\] Therefore, the spin squeezing from Eq. ([\[ss\]](#ss){reference-type="ref" reference="ss"}) can be expressed as follows \[\label{eff}\] Note that for the spin-coherent state \[\^2+\^2=\^2+\^2\(, giving\) In the experiments, the atom number fluctuations are calculated by measuring in each shot the population of each mode. In another series of experiments, the two modes are let interfere and the relative phase of the pattern is recorded to give, after many runs, the denominator of Eq. ([\[eff\]](#eff){reference-type="ref" reference="eff"}). Note that there is an alternative method for measuring ([\[ss\]](#ss){reference-type="ref" reference="ss"}), which does not rely on the mean-field approximation ([\[mf\]](#mf){reference-type="ref" reference="mf"}). In the far-field, the spatial wave-packet of the modes \(a\) and \(b\) are plane-waves (or, more precisely broad functions with fast oscillations on top), i.e., \]\hat\Psi(x)=e^{ikx}\hat a+e^{-ikx}\hat b.\[ The density of this system is \]\begin{aligned}
\rho(x)&=&\av{\hat\Psi^\dagger(x)\hat\Psi(x)}=\av{\hat a^\dagger\hat a}+\av{\hat b^\dagger\hat b}+\nonumber\\ &+&2\av{\hat J_x}\cos(2kx)+2\av{\hat J_y}\sin(2kx). \end{aligned}\[ Since \(\av{\hat a^\dagger\hat a}+\av{\hat b^\dagger\hat b}=N\), the normalized density reads \]\begin{aligned}
\label{normalized} p(x)&=&\frac1N\rho(x)=1+\nu_1\cos(2kx)+\nu_2\sin(2kx)=\nonumber\\ &=&1+\nu\cos(2kx-\alpha), \end{aligned}\[ where \(\nu_{1/2}=\frac2N\av{\hat J_{x/y}}\) and \(\alpha=\arccos\left(\frac{\nu_1}\nu\right)\), while \(\nu=\sqrt{\nu_1^2+\nu_2^2}\) is the fringe visibility. Therefore, the spin-squeezing ([\[ss\]](#ss){reference-type="ref" reference="ss"}) can be expressed in terms of the atom number fluctuations \(\eta^2=\frac4N\av{(\Delta\hat J_z)^2}\) and the visibility \]\label{op_def}\[ We now demonstrate that the use of this operational definition of the spin-squeezing, without the *a priori* knowledge about the mode structure of each arm, can lead to false conclusions regarding the presence of the particle entanglement in the system. In the multi-mode case, the density calculated with the operator ([\[field\]](#field){reference-type="ref" reference="field"}) is \]\rho(x)=\rho_{aa}(x)+\rho_{bb}(x)+\rho_{ab}(x)+\rho_{ba}(x),\[ where \(\rho_{ij}(x)=\av{\hat\Psi^\dagger_i(x)\hat\Psi^\dagger_j(x)}\). Clearly the density contains multiple interference terms, not only resulting from the overlap of the wave-functions coming from the opposite regions but also from different modes residing initially in the same region. Therefore, the visibility of fringes cannot be linked to the denominator of Eq. ([\[ss_mm\]](#ss_mm){reference-type="ref" reference="ss_mm"}), which only quantifies the \(a/b\) coherence. For illustration, take a coherent spin state from Eq. ([\[css_2m\]](#css_2m){reference-type="ref" reference="css_2m"}) with \(\varphi=0\). Assume now that in the region \(a\), some process splits the mode \(a\) into the coherent superposition of \(a_1\) and \(a_2\), schematically shown in Fig. [\[fig.dw\]](#fig.dw){reference-type="ref" reference="fig.dw"} in a double-well setup. This way, a state \]\label{css2} \ket\psi=\frac1{\sqrt{N!}}\Big[\sqrt z\left(\sqrt{\zeta}\hat a_1^\dagger+\sqrt{1-\zeta}\hat a_2^\dagger\right)+\sqrt{1-z}\hat b_1^\dagger\Big]^N\ket0\[ is obtained. It is an example of a non-entangled state from Eq. ([\[css_mm\]](#css_mm){reference-type="ref" reference="css_mm"}) with \(\vec\alpha=\sqrt z(\sqrt\zeta,\sqrt{1-\zeta},0\ldots)^{\rm T}\) and \(\vec\beta=(\sqrt{1-z},0\ldots)^{\rm T}\). The splitting does not influence the nominator of Eq. ([\[ss_mm\]](#ss_mm){reference-type="ref" reference="ss_mm"}), but in the far-field the interference of the two modes \(a_1\) and \(a_2\) will have impact on the visibility of fringes. The field operator after the expansion will read \]\hat\Psi(x)=e^{i(k+\delta k)x}\hat a_1+e^{i(k-\delta k)x}\hat a_2+e^{-ikx}\hat b.\[ giving the normalized density
This density is plotted in Fig. ([\[fig.pattern\]](#fig.pattern){reference-type="ref" reference="fig.pattern"}) with \(z=0.91\), \(\zeta=0.5\), \(k=10\) and \(\delta k=0.5\). The fringe visibility, calculated as the ratio \(\nu=(p_{\rm max}-p_{\rm min})/(p_{\rm max}+p_{\rm min})\) is \(\nu^2=0.326\). This combined with the normalized population imbalance between the two regions, \(\eta^2=1-(2z-1)^2=0.321\), gives the ratio \(\eta^2/\nu^2=0.988\), suggesting the presence of the particle entanglement in the system. However, it is a false statement: \(\nu\) cannot be identified with the denominator of Eq. ([\[ss_mm\]](#ss_mm){reference-type="ref" reference="ss_mm"}) in this case. *Some remarks*---Naturally, the three-mode example invoked here is quite artificial. It is hard to imagine that a coherent physical process, which splits \(\hat a\) into \(\hat a_1\) and \(\hat a_2\) could be uncontrolled and unnoticed by the experimentalists. Also, any experimentalist would immediately notice two frequencies of oscillations in the interference pattern. Last but not least, usually the multi-mode structure of the two regions \(a\) and \(b\) comes from the thermal excitations, and thus reveals no coherence between the modes. Nevertheless, the example shows that the method of detecting the particle-entanglement through the fluctuations-to-visibility ratio can be safely used only when the structure of each sub-system is known.
# Summary {#sec.sum}
The main outcome of this work is the establishment of the SNL for the two-arm multi-mode interferometers. This is a general result, as it applies to any thinkable quantum system where the entangled/non-entangled dichotomy exists. It is valid for both fixed-and non-fixed-\(N\) systems, for as long as the coherence between states carrying different number of particles is absent. Note that our results are valid for a particular choice of the interferometric transformations, such that do not penetrate the inner structure of each arm but rather treat them as a whole. In principle, any other type of such transformation requires a dedicated calculation of the SNL. Otherwise, conclusions about the presence of metrologically useful particle entanglement in the system can be incorrect. We have also calculated the phase sensitivity for the standard estimation protocol based on the knowledge of the mean population imbalance between the two arms. Finally, we have shown that if the popular method of detecting the entanglement in two-mode systems is used without the *a priori* knowledge about the modal structure of the arms, a false conclusion from the number-fluctuations-to-visibility ratio can drawn.
[^1]: Note that in principle, an interferometer can be a series of linear transformations, rather than a single one Still, such composite operation remains linear and does not entangle the particles. | {'timestamp': '2017-06-07T02:05:23', 'yymm': '1706', 'arxiv_id': '1706.01700', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01700'} |
# Introduction {#sec:intro .unnumbered}
Single-photon states are of outstanding interest in modern quantum optics, as the basis of fundamental tests of quantum mechanics and of a variety of quantum technologies (see e.g. for recent surveys of single-photon techniques and applications). In order to generate truly single photons, one of the most efficient method is based on a conditional measurement, where a two-photon state is generated and the presence of a single photon is *heralded* by detection of its partner. To this end, parametric processes such as spontaneous parametric down-conversion (PDC) or four-wave mixing (FWM), where one or two photons belonging to a high energy pump laser are occasionally converted into pairs of photons, are routinely employed. The conservation laws ruling these microscopic processes originate a quantum correlation in the spatial and temporal degrees of freedom of the pair, which typically extends over broad spectral and angular bandwidths. Such a high dimensional entanglement may represent a resource for broadband quantum communication or quantum metrology, but is detrimental for the purity of heralded single photons, because detection of the trigger photon projects the state of its twin in a highly mixed state. This represents a limitation for quantum communication/information protocols where single photons are required to be in indistinguishable and capable of high-visibility interference.\
The recent development of waveguided PDC (see and reference therein) and of FWM in single-mode fibers opened the possibility to control the spatial degrees of freedom, and to generate the twin photons into a single of few spatial modes. In order to eliminate also the spectro-temporal correlation, a possibility is filtering hard enough that a single spectral mode is selected, but clearly the efficiency of the source is reduced. In order to achieve pure heralded photons with high fluxes, therefore, considerable effort has been devoted to find alternative techniques, which directly ruduce the degree of entanglement of the source in order to produce uncorrelated twin photons. In such a way a conditional measurement projects the field in a pure single-photon state rather than in a mixed state. In the standard co-propagating geometry (Fig.[\[fig1\]](#fig1){reference-type="ref" reference="fig1"}b) this task requires careful techniques of group velocity matching, which can be implemented only in some materials and tuning conditions.
In recent years, the counter-propagating configuration of PDC, where twin photons are emitted along opposite directions (Fig.[\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a), emerged as a promising source of heralded pure single photons without the need of group-velocity matching. Proposed by Harris in the sixties and implemented in 2007 by Canalias et al., this configuration presents the challenge of requiring a very short poling of the nonlinear material. On the other hand, counter-propagating twin photons and twin beams possess peculiar and attractive features, such as their narrowband character and the potentiality to generate in the high-gain regime a robust continuous-variable entanglement . In the spontaneus regime, twin-photons may be naturally generated in either spectrally entangled or decorrelated states by simply modifying the pump pulse duration (or crystal length), without the need of special tuning conditions. Similar features have recently been described in counterpropagating FWM.
In this work we focus on the PDC process, and on the temporal correlation of twin photons thereby generated, making a parallel analysis of the two co-propagating and counterpropagating geometries (Fig.[\[fig1\]](#fig1){reference-type="ref" reference="fig1"}) as sources of heralded single photons. By a systematic comparison of the two configurations, we aim at providing a deeper understanding of the physical mechanisms under which temporal correlation emerges and can eventually be eliminated, which may turn useful to optimize the existing configurations and to design new ones. Event though our analysis is restricted to the PDC process, some of our results may qualitatively hold also for FWM.
A large part of our analysis will be based on the correlation of twin-photons in the temporal domain. This approach is not standard in the literature, which mostly focused on their spectral correlation, and provides, in our opinion, a more direct physical view. For example, it will show that the temporal correlation between twin photons can be eliminated only creating conditions such that the timing provided by the pump pulse is more precise than that offered by detecting any of the twin photons. In this way, detection of one photon does not give any better information on the exit time of its twin than by not detecting it, and the two-photon state appears uncorrelated.\
From a more quantitative point of view, we will derive an approximated formula for the Schmidt number, to our knowledge not present in the former literature, which holds for both geometries and gives a simple means for evaluating the degree of purity of the heralded photon based only on two parameters, that describe the characteristic time scales of the source.\
Finally, the individual spectro-temporal properties of the heralded photons, which are of great importance in view of different applications will be investigated in parallel. In particular, we will see that either narrowband or broadband pure heralded photons are characteristic of the two geometries, and that different strategies of group-velocity matching lead to very different efficiencies of pair production.
The paper is organized as follows: After introducing in Sec.[1](#sec:phasematching){reference-type="ref" reference="sec:phasematching"} the two geometries and their different phase-matching conditions, the core of our results is contained in sections [2](#sec:temporal){reference-type="ref" reference="sec:temporal"}-[5](#sec:spectrum){reference-type="ref" reference="sec:spectrum"}. Sec.[2](#sec:temporal){reference-type="ref" reference="sec:temporal"} describes the temporal correlation of twin photons, in terms both of general formulas valid for any configuration of PDC, and of numerical calculations for specific examples. Sections [3](#sec:Schmidt){reference-type="ref" reference="sec:Schmidt"} calculates the Schmidt number of entanglement. General analytic results obtained within the Gaussian approximation for the biphoton amplitude are compared with more exact numerical ones. Sec [4](#sec:interpretation){reference-type="ref" reference="sec:interpretation"} offers a physical interpretation of the results obtained, while the final Sec.[5](#sec:spectrum){reference-type="ref" reference="sec:spectrum"} describes the spectro-temporal properties of heralded photons and compares the efficiency of heralding within the different strategies investigated.
# Counter-propagating and co-propagating geometries {#sec:phasematching}
As in most of the previous literature, our analysis is restricted to the purely temporal domain, assuming either a single-mode waveguided configuration, or that a small angular bandwidth is collected.
In the counter-propagating geometry shown in Fig.[\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a, a coherent pump pulse of central frequency \(\omega_p\) and temporal profile \(\alpha_p(t)\) impinges a nonlinear \(\chi^{(2)}\) crystal of length \(l_c\) from the left face, and generates counter-propagating photon pairs with, say, the idler photon backpropagating towards the laser source. For the counter-propagating three-wave interaction, momentum conservation requires quasi-phase matching in a periodically poled structure, with a short poling period \(\Lambda\), such that the momentum associated to the nonlinear grating of the nonlinearity \(k_G=\frac{2 \pi m}{\Lambda}\), \(m=\pm 1, \pm3\hdots\), approximately compensates the momentum of the pump photon. At first order, it requires \(\Lambda\) on the same order as the pump wavelength in the medium. The central frequencies of the emitted signal and idler fields, \(\omega_s\) and \(\omega_i=\omega_p-\omega_s\), are thus determined by the poling period \(\Lambda\) and the pump central frequency \(\omega_p\) according to \[\label{qpm} k_{s}-k_{i}=k_{p}-k_G\qquad \text{\small{{\bf (a)counter-propagating case }}}\] where \(k_{j}=\frac{\om_j}{c}n_j(\omega_j)\), \(j=s,i,p\) are the wave-numbers at the corresponding central frequencies \(\omega_j\).
In the co-propagating geometry (Fig.[\[fig1\]](#fig1){reference-type="ref" reference="fig1"}b) all the three fields propagate along the positive \(z\) direction. In this case momentum conservation requires \[\label{qpm2} k_{s}+k_{i}=k_{p}-k_G\qquad\text{\small{{\bf (b) co-propagating case}}}\] which describes both the case of a bulk crystal, in which \(k_G=0\), or quasi-phasematching in periodically poled structures. The efficiency of the down-conversion process is described by a dimensionless gain parameter \(g\), proportional to the peak amplitude of the pump laser, the crystal length and the nonlinear susceptibility. Considering the spontaneous regime where \(g \ll 1\), and retaining terms at most linear in \(g\ll1\), the signal-idler state at the crystal output takes the well-known form \[\begin{aligned}
&|\phi^{}\rangle = \left |0\right \rangle + |\phi^{}_2 \rangle\label{state}\\ &|\phi^{}_2\rangle= \int d\Om_s d\Om_i \psi^{\text{\tiny{(\(j\))}}} (\Om_s, \Om_i) \hat{a}_s^\dagger (\Om_s) \hat{a}_i^\dagger (\Om_i) \left |0\right \rangle \,, \label{statec} \end{aligned}\] corresponding to the superposition of the vacuum and the two-photon state \(|\phi^{}_2\rangle\) (see e.g. for a derivation from the more powerful quantum-field description). Here \(\hat{a}_s^\dagger (\Om_s)\) and \(\hat{a}_i^\dagger (\Om_i)\) are signal and idler photon creation operators in the frequency domain, \(\Om_j\) being the offset from the reference frequency \(\omega_j\), and \[\label{psi} \psi^{\text{\tiny{(\(j\))}}}(\Omega_s,\Omega_i) =\frac{g }{\sqrt{2\pi}} \tilde{\alpha}_p(\Omega_s+\Omega_i) \mathrm{sinc}\left[\frac{{\cal D}^{\hspace{-0pt}\text{\tiny{\(\pm\)}}}(\Omega_s,\Omega_i)l_c}{2}\right] e^{i\beta(\Omega_s,\Omega_i) }\] is the so-called spectral biphoton amplitude, giving the joint probability amplitude of detecting a signal photon at frequency \(\omega_s+\Om_s\) and an idler photon at frequency \(\omega_i+\Om_i\). In formula [\[psi\]](#psi){reference-type="eqref" reference="psi"} \[\label{alphap} \tilde{\alpha}_p(\Om)=\int\frac{dt}{\sqrt{2\pi}}e^{i\Om t} {\alpha_p(t)}\,;\] is the spectral amplitude of the pump pulse, where the temporal profile \({\alpha_p(t)}\) is normalized to its peak value; \({\cal D}^{\hspace{-0pt}\text{\tiny{\)\pm\(}}}(\Oms,\Omi)l_c\) is the phase-mismatch function, where here and in the following the index \(+\) and \(-\) refer to the counter-propagating and co-propagating geometries (a) and (b), respectively \[\begin{aligned}
{\cal D}^{\hspace{-0pt}\text{\tiny{\(\pm\)}}}(\Oms,\Omi) &&= \left\{ \begin{array}{lr} k_p(\Oms+\Omi) -k_s(\Oms) + k_i(\Omi)-k_G&\quad \text{ \; \bf (a)} \vspace{3pt}\\ k_p(\Oms+\Omi)-k_s(\Oms)-k_i(\Omi)-k_G&\quad \text{ \; \bf (b)} \end{array} \right. \label{mismatch} \end{aligned}\] Finally \[\beta (\Oms,\Omi)= \frac{l_c}{2} \left[ k_s(\Oms) +k_i (\Omi) + k_p (\Omega_s + \Omega_i) \right]\] is a global phase, acquired during the propagation along the crystal in the absence of any nonlinear effect by a pair of photons generated at the crystal center \(z=l_c/2\).\
As described in, the different sign in front of the idler wave-number \(k_i(\Om_i)\) in Eq. [\[mismatch\]](#mismatch){reference-type="eqref" reference="mismatch"} is at the origin of the radically different properties of the two geometries of PDC. This is best seen by expanding the phase mismatch ([\[mismatch\]](#mismatch){reference-type="ref" reference="mismatch"}) at first order around the reference frequencies (corresponding to \(\Omega_j=0\)), \[\begin{aligned}
\label{Dlin} {\cal D}^{\hspace{-0pt}\text{\tiny{\(\pm\)}}} (\Oms,\Omi) \frac{l_c}{2} &\approx&\frac{l_c}{2}[(k_p'\pm k_i')\Omi +(k_p'-k_s')\Oms ] \nonumber \\ &\equiv& {\tau_{i}^{\text{\tiny{\(\pm\)}}}}\Omi + {\tau_{s}} \Oms \;,\label{Dlin2} \end{aligned}\] where \(k_j'\equiv v_{gj}^{-1} =\left(\frac{d k_j}{d \Omega_j}\right)_{\Omega_j=0}\) is the inverse group velocity of wave \(j\) and we introduced the characteristic times \[\begin{aligned}
{\tau_{s}} = \frac{1}{2}\left(\frac{l_c}{v_{gp}}-\frac{l_c}{v_{gs}}\right) \label{taus}\\ {\tau_{i}^{\text{\tiny{\(\pm\)}}}}= \frac{1}{2}\left(\frac{l_c}{v_{gp}}\pm\frac{l_c}{v_{gi}}\right) \label{taui} \end{aligned}\] which involve either the difference or the sum of the group velocities of the pump and the down-converted field, depending whether the latter co-propagates or counterpropagates with respect to the former, or, in other words, they involve the *relative* group velocities of the down-converted photons with respect to the pump. These constants represent the characteristic temporal separations between the centers of the pump and down-converted wave-packets, and describe how much delayed from the pump a pair of twin photons generated at the crystal center arrive at their end faces. In the co-propagating case \({\tau_{s}}\) and \({\tau_{i}^{\text{\tiny{\)-\(}}}}\) are both determined by the group velocity mismatch (GVM) with respect to the pump, and are typically on the same order of magnitude, unless some particular strategy of group velocity matching is employed. By contrast, in the counter-propagating case, the time constant associated to the backward photon \({\tau_{i}^{\text{\tiny{\)+\(}}}}\) involves the group velocities sum (GVS) and is on the order of the photon transit time across the crystal, which exceeds the GVM time by one or two orders of magnitude.\
As will be clear in the following (see also ), the possibility to generate heralded photons with a high degree of purity depends on the relative sizes and signs of these time constants, and on how they compare to the pump duration \({\tau_\mathrm{p}}\). Therefore we introduce the ratio \[\eta^{}=\frac{ {\tau_{s}} }{{\tau_{i}^{\text{\tiny{\(\pm\)}}}}} \;, \label{eta}\] which can be positive or negative ( only \({\tau_{i}^{\text{\tiny{\)+\(}}}}\) is always positive, while \({\tau_{s}}\) and \({\tau_{i}^{\text{\tiny{\)-\(}}}}\) can be either positive or negative). Clearly, in the counter-propagating case \[|\eta| = \frac {|k'_p-k'_s|}{k'_p +k'_i} =\frac{| {\tau_{s}} | }{{\tau_{i}^{\text{\tiny{\(+\)}}}}} \ll 1\] for any choice of materials and phase matching conditions. Without loss of generality, we assume that in the co-propagating case the signal and idler are chosen so that \(| {\tau_{s}} |\leq|{\tau_{i}^{\text{\tiny{\)-\(}}}}|\), so that in both configurations \[-1\leq\eta^{\text{\tiny{(\(j\))}}}\leq 1 \label{etarange}\] An analogous linear approximation of the global phase appearing in Eq.[\[psi\]](#psi){reference-type="eqref" reference="psi"} gives: \[\begin{aligned}
\beta (\Oms,\Omi) &\approx \frac{l_c}{2} \left[ k_s +k_i + k_p + (k_s' + k'_p) \Oms + (k_i' + k'_p ) \Omega_i\right] \nonumber \\ &= \text{const.} + t_{As} \Oms + t_{Ai} \Omi \,, \label{betalin} \end{aligned}\] where
can be considered as the times at which two twin photons downconverted from the pump peak at \(z= l_c/2\)-pictured as wave-packets propagating without deformation-arrive at their respective end faces of the crystal. They represent the exit time of the centers of the signal and idler wavepackets, assuming that at \(t=0\) the center of the pump pulse enters the crystal (\(t_{Ap} = \frac{l_c}{v_{gp}}\) is then the exit time of the pump). We notice that the omission of second and higher order terms in Eqs. [\[Dlin2\]](#Dlin2){reference-type="eqref" reference="Dlin2"} and [\[betalin\]](#betalin){reference-type="eqref" reference="betalin"} is valid only for small bandwidths, and corresponds to neglecting the temporal dispersion. This is is well justified in the counter-propagating configuration, which involves narrow down-conversion spectra, while it is less justified in the co-propagating case, because of the larger bandwidths in play. In particular, it is not justified when \({\tau_{s}} \simeq{\tau_{i}^{\text{\tiny{\)-\(}}}}\) ( e.g. for type I PDC close to degeneracy).
# Time-domain view: the temporal correlation of twin photons {#sec:temporal}
An alternative and perhaps more intuitive view of the problem is offered by the joint amplitude in the temporal domain, which is obtained by back-Fourier transforming the joint spectral amplitude \[\label{phi} \phi(t_s,t_i)=\int\frac{d\Oms}{\sqrt{2\pi}} \int\frac{d\Omi}{\sqrt{2\pi}} e^{-i (\Oms t_s+\Omi t_i)} \psi(\Oms,\Omi)\] In terms of this function, the two-photon state [\[statec\]](#statec){reference-type="eqref" reference="statec"} becomes \[\begin{aligned}
&|\phi^{}_2 \rangle= \int d t_s d t_i \phi (t_s, t_i) \hat{a}_s^\dagger (t_s) \hat{a}_i^\dagger (t_i) \left |0\right \rangle \,, \label{statec2} \end{aligned}\] where \(\hat{a}_s^\dagger (t_s)\), \(\hat{a}_i^\dagger (t_i)\) are photon creation operators in the time domain, and for example \(\hat{a}_s^\dagger (t_s) |0\rangle \equiv |t_s\rangle\) represents the state with exactly one signal photon at time \(t_s\) at the crystal output face. Thus, \(\phi(t_s,t_i)\), which was extensively analysed in in the counter-propagating case, represents the joint probability amplitude that a signal and an idler photon exit the crystal slab at times \(t_s\) and \(t_i\), respectively, and describes the temporal correlation between the twin photons. The rate of coincidence counts at the crystal exit faces is then given by \[\begin{aligned}
G_{si}^{(2)} (t_s,t_i) \equiv \langle \hat{a}_s^{\dagger} (t_s) \hat{a}_s (t_s) \hat{a}_i^{\dagger} (t_i) \hat{a}_i (t_i) \rangle \approx \left| \phi (t_s,t_i)\right|^2. \label{G2} \end{aligned}\]
Several plots of this correlation function are shown in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}, where the three columns correspond to three examples chosen as representatives of the two geometries (see Table [\[table1\]](#table1){reference-type="ref" reference="table1"} for parameters, and Appendix [\[sec:examples\]](#sec:examples){reference-type="ref" reference="sec:examples"} for details), namely **(i)** a generic counter-propagating configuration, not specifically optimized for separability, where \(\eta\simeq 0.01\), and two co-propagating configurations optimized for separability, with **(ii)** \(\eta=0\), and **(iii)** \(\eta=-1\). Condition (ii) is referred to as *asymmetric group-velocity matching* and requires that the signal photon propagates with the same group velocity as the pump (\({\tau_{s}} =0\)). Condition (iii) corresponds to the *symmetric group-velocity matching*, and can be realized only in the co-propagating case, requiring \({\tau_{s}} =-{\tau_{i}^{\text{\tiny{\)+\(}}}}\). They are usually difficult to satisfy in the visible range, but can be achieved in some \(\chi^{(2)}\) material in the near infrared and at telecom wavelengths. Experimental evidence of frequency decorrelated photon pairs through this technique were first reported in.
The plots in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"} have been calculated from Eq.[\[phi\]](#phi){reference-type="eqref" reference="phi"} without resorting to any approximation, by using the complete Sellmeier dispersion formula in for evaluating the wave-numbers. If the linear approximations [\[Dlin2\]](#Dlin2){reference-type="eqref" reference="Dlin2"} and [\[betalin\]](#betalin){reference-type="eqref" reference="betalin"} are instead employed, the Fourier transform ([\[phi\]](#phi){reference-type="ref" reference="phi"}) can be explicitly calculated , obtaining \[\begin{aligned}
\phi(\bar{t}_s,\bar{t}_i) = \frac{g}{2| \tauipm-\taus|} {\alpha}_p\left(\frac{\bar{t}_s-\eta\bar{t}_i}{1-\eta}\right) {\rm Rect}\left(\frac{\bar{t}_s-\bar{t}_i}{|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |}\right) \label{psitemp} \end{aligned}\] where constant phase factors have been omitted. The barred arguments \[\begin{aligned}
\bar{t}_j&=&t_j-t_{Aj} \qquad j=s,i \,, \end{aligned}\] are the exit times of the twin photons measured relative to the central exit times \(t_{Aj}\) of their wavepackets, \[see Eq. [\[tA\]](#tA){reference-type="eqref" reference="tA"}\]. The temporal correlation in Eq.[\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} is the product of two factors: the first one is the pump temporal profile \({\alpha}_p(t)\), that we can e.g. take as a Gaussian, the second one is the box function \[{\rm Rect}\left(\frac{t}{\Delta \tau } \right)=\begin{cases} 1 \qquad &\text{for } \, t\epsilon \, [-\Delta \tau, \Delta \tau]\\ 0 \qquad &\text{elsewhere} \end{cases}\] of width \(2\Delta \tau\) and unitary height, where \[\Delta \tau:=|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |=\frac{l_c}{2}\left|\frac{1}{v_{gs}}\pm\frac{1}{v_{gi}}\right| \label{Dtau}\] represents the typical delay intercurring between the exit times of two twin photons.
The result [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} relies merely on linearization of the propagation phases, and as such fails to be valid for \({\tau_{s}} \to {\tau_{i}^{\text{\tiny{\)\pm\(}}}}\), when the first order of the Taylor espansion of \({\cal D}^{\hspace{-0pt}\text{\tiny{\)\pm\(}}}\) vanishes. It coincides with the expression analysed in , where it was shown that in the counter-propagating configuration the temporal correlation [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} reduces to a factorable function of \(\bar t_s, \bar t_i\) for a range of pump durations intermediate between the two characteristic time scales: \(|{\tau_{i}^{\text{\tiny{\)+\(}}}}| \gg {\tau_\mathrm{p}}\gg | {\tau_{s}} |\). The analysis in can then be generalized to a generic geometry.
From a geometric/mathematic viewpoint, we have the noticeable asymptotic behaviors of the temporal correlation in Eq.[\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"}:\
**1)**Limit of a long pump pulse \({{\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\ge | {\tau_{s}} |}\). When the pump pulse is much longer than both characteristic time scales, the pump profile has a slow variation with respect to the much narrower box function. Therefore we can put \(\bar t_s= \bar t_i\) inside the term \({\alpha}_p\left(\frac{\bar{t}_s-\eta\bar{t}_i}{1-\eta}\right)\), and \[\begin{aligned}
\phi(\bar{t}_s,\bar{t}_i) \underset{ {\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}|} {\longrightarrow} & \frac{g}{2| \tauipm-\taus|} \alpha_p \left(\bar{t}_s \right) {\rm Rect}\left( \frac{\bar{t}_s-\bar t_i }{|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |} \right]) \label{asym1} \end{aligned}\] This formula well reproduces the three upper plots in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}, which exhibit a a sharp maximum of the temporal correlation along the diagonal \(\bar t_s = \bar t_i\), espressing the fact that twin photons can be generated at any time along the pump pulse, but they tends to exit almost simultaneously from the crystal (a part from the fixed offset \(t_{As}-t_{Ai}\)) with a flat distribution of their mutual time delay in the interval \([-\Delta \tau, \Delta \tau]\). This situation corresponds to a joint temporal amplitude not factorable in its arguments, and the two-photon state [\[statec2\]](#statec2){reference-type="eqref" reference="statec2"} is entangled.\
**2)**Limit of an ultrashort pump pulse \({ {\tau_\mathrm{p}} \ll | {\tau_{s}} | \le |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}| }\). In the opposite limit, the narrow pump profile forces \(\bar{t}_s = \eta \bar{t}_i\) inside the box function, and \[\begin{aligned}
\phi(\bar{t}_s,\bar{t}_i) \underset{ {\tau_\mathrm{p}} \ll | {\tau_{s}} |} {\longrightarrow} & \frac{g}{2| \tauipm-\taus|} \alpha_p \left(\frac{\bar{t}_s-\eta\bar{t}_i}{1-\eta}\right) {\rm Rect}\left[\frac{\bar{t}_s}{ {\tau_{s}} } \right] \label{asym2} \end{aligned}\] Eq.[\[asym2\]](#asym2){reference-type="eqref" reference="asym2"} approximately describes the three lower plots of Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}, where the correlation function is peaked along the line \(\bar t_s = \eta \bar t_i\). In the example (ii) where \(\eta=0\), this line is parallel to the \(\bar{t}_i\) axis, and the joint temporal amplitude is approximately factorable (see Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}f), implying that the two-photon state [\[statec2\]](#statec2){reference-type="eqref" reference="statec2"} is separable. In the other two examples with \(\eta \ne 0\), the state is entangled. Notice the opposite signs of \(\eta\) in panels (c) and (j) of Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}, and that in terms of the original time arguments, by using Eq.[\[tA\]](#tA){reference-type="eqref" reference="tA"}: \(\bar t_s =\eta \bar{t}_i\) implies \((t_s-t_{Ap} )= \eta (t_i-t_{Ap})\). Thus, in the ultrashort pump regime, the exit times of twin photons are correlated for \(\eta>0\), while they are *anticorrelated* for \(\eta <0\), as a result of the fact that one photon is slower than the pump while the other is faster.\
**3)**Intermediate pump pulses \({|{\tau_{i}^{\text{\tiny{\)\pm\(}}}}| \gg {\tau_\mathrm{p}} \gg| {\tau_{s}} | }\). Notice that this limit is well defined in the counterpropagating case, where the two scales are naturally separated, but needs not to exist in the co-propagating case, where it basically requires that one of the two time constant vanishes, e.g. \({\tau_{s}} \to 0\) as in the example (ii). If this limit exists, then \[\begin{aligned}
\phi(\bar{t}_s,\bar{t}_i) \underset { |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}|\gg{\tau_\mathrm{p}}\gg | {\tau_{s}} | } {\longrightarrow} \frac{g}{2|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |} {\alpha}_p\left(\bar{t}_s\right)& {\rm Rect}\left(\frac{\bar{t}_i}{|{\tau_{i}^{\text{\tiny{\(+\)}}}}-{\tau_{s}} |}\right) \, \label{asym3} \end{aligned}\] and the biphoton correlation becomes separable in its arguments, as approximately shown in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}(b),(f), implying that two-photon state [\[statec2\]](#statec2){reference-type="eqref" reference="statec2"} is separable.
As a final for this section, we notice that despite the fact that the joint temporal probability as a function of \((t_i,t_s)\) changes completely its shape in different pump regimes, when considered only as a function of the time difference \(t_i-t_s\), it always retains a box-function shape. This means that if the coincidence count rate [\[G2\]](#G2){reference-type="eqref" reference="G2"} at the crystal exit faces is registered only as a function of the delay between the exit times of the twins, no information is gained about the entanglement or separability of the state. Indeed, if we rewrite the temporal correlation in Eq. [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} as a function of \(\delta t = t_i-t_s\), we have \[\begin{aligned}
\phi (t_s,t_s+ \delta t ) & \propto \alpha_p \left(\bar{t}_s-\frac{\eta}{1-\eta} \delta \bar t \right) \mathrm{Rect}\left(\frac{\delta \bar t }{\Delta \tau}\right) \nonumber \\ \phi(t_i-\delta t ,t_i) & \propto \alpha_p \left(\bar{t}_i+ \frac{1}{1-\eta} \delta \bar t \right) \mathrm{Rect}\left(\frac{\delta \bar t }{\Delta \tau}\right) \nonumber \end{aligned}\] Then, if coincidence counts are measured only as a function of the time delay \(t_i-t_s\) between twin photons, without detecting the absolute arrival time of the signal or of the idler, \[\begin{aligned}
\bar G_{si}^{(2)} ( \delta t ) & = \int dt_s | \phi (t_s, t_s + \delta t ) |^2 = \int dt_i | \phi (t_i-\delta t, t_i) |^2 \nonumber \\ &=\frac{ {\cal N} }{2 \Delta \tau} \mathrm{Rect}\left(\frac{\delta \bar t }{\Delta \tau}\right). \end{aligned}\] where \({\cal N} =g^2 \frac{\sqrt{\pi}{\tau_\mathrm{p}}} {2 \Delta \tau}\) is the total number of photon pairs (see Sec.[5](#sec:spectrum){reference-type="ref" reference="sec:spectrum"}). Therefore the coincidence counts as a function of \(t_i-t_s\) always reproduces the flat distribution of time delays characteristic of the spontaneous process, regardless whether the state is temporally entangled or separable.
Section [4](#sec:interpretation){reference-type="ref" reference="sec:interpretation"} will offer an interpretation of these different behaviours of the temporal correlation of twin photons. Before that, we make first a quantitative analysis of the degree of the entanglement of the state.
# Entanglement quantification {#sec:Schmidt}
The degree of entanglement of the state is here characterized by the Schmidt number, which also estimates the number of independent modes participating to the entangled state. It is defined as the inverse of the purity of the state of each separate subsystem \[\cappa= \frac{1}{ {\mathrm Tr} \{\hat \rho_s^2\} }= \frac{1}{ {\mathrm Tr} \{\hat \rho_i^2\} }\] where \(\hat \rho_s\), \(\hat \rho_i\) are the reduced density matrix of the signal and idler when the PDC state [\[state\]](#state){reference-type="eqref" reference="state"} is conditioned to a photon count. For example for the signal \(\hat \rho_s=\frac{1}{ \langle \phi _{2} |\phi _{2}\rangle } {\mathrm Tr}_i \{ |\phi _2\rangle \, \langle \phi _2| \}\). The inverse of the Schmidt number thus straighforwardly gives the degree of purity of the heralded photon state, via a single parameter that can be calculated without resorting to the esplicit Schmidt decomposition. Indeed, for a biphoton state of the form [\[statec\]](#statec){reference-type="eqref" reference="statec"}, the Schmidt number can be expressed through an integral formula \[\cappa = \frac{ {\cal N}^2} {B} \label{kintegral}\] where \[\begin{aligned}
{\cal N} = & \int d\Oms \int d\Omi \, \left| \psi(\Oms, \Omi) \right|^2 \label{enne} \end{aligned}\] is the mean number of signal or idler photons \(\langle \hat N_s \rangle = \langle \hat N_i \rangle\) generated by the pump pulse; \[\begin{aligned}
B= & \int d\Oms \int d\Oms' \int d\Omi \int d\Omi ' \left[ \psi(\Oms, \Omi ) \psi(\Oms', \Omi ') \right. \left. \psi^* (\Om_s,\Om_i') \psi^* (\Om_s',\Om_i) \right] \,, \label{B} \end{aligned}\] represents their normally ordered fluctuations(see ): \(B=\langle: \delta\hat N_s^2: \rangle=\langle: \delta\hat N_i^2: \rangle\).
In this work the Schmidt number \(\cappa\) is calculated in two ways: (i) \"exact \" results are obtained by numerical integration of Eqs.([\[kintegral\]](#kintegral){reference-type="ref" reference="kintegral"})-([\[B\]](#B){reference-type="ref" reference="B"}), where the complete Sellmeier dispersion formula in are used; (ii) approximated analytical results are derived within the \"Gaussian approximation\" of the spectral amplitude extensively used in former studies of co-propagating PDC. This consists in replacing the sinc function in Eq.([\[psi\]](#psi){reference-type="ref" reference="psi"}) by a Gaussian of its argument, and then using the linear approximations [\[Dlin\]](#Dlin){reference-type="eqref" reference="Dlin"} and [\[betalin\]](#betalin){reference-type="eqref" reference="betalin"}, obtaining thereby \[\begin{aligned}
{\rm sinc}\frac{{\cal D}^{\hspace{-0pt}\text{\tiny{\(\pm\)}}}(\Oms,\Omi)l_c}{2}&\to e^{-\gamma\left[\frac{{\cal D}^{\hspace{-0pt}\text{\tiny{\(\pm\)}}}(\Oms,\Omi)l_c}{2}\right]^2} \approx e^{-\gamma \left( {\tau_{s}} \Oms+{\tau_{i}^{\text{\tiny{\(\pm\)}}}}\Omi\right)^2} \label{sinc_apr} \end{aligned}\] where e.g. \(\gamma=1/6\) if one equates the leading order of the Taylor expansions of the \({\rm sinc}\) and the Gaussian, or \(\gamma=0.193\) if one requires that they have the same full width at half maximum. In addition, one has to consider a Gaussian pump pulse \(\alpha_p(t)= e^{-t^2/2{\tau_\mathrm{p}}^2}\), of duration \({\tau_\mathrm{p}}\) and spectral width \(\Delta \Om_p=1/{\tau_\mathrm{p}}\), so that \(\tilde{\alpha}_p(\Om)={\tau_\mathrm{p}} e^{-\frac{{\tau_\mathrm{p}}^2}{2} \Om_p^2},\) The spectral amplitude ([\[psi\]](#psi){reference-type="ref" reference="psi"}) takes then the Gaussian form: \[\label{psi_gauss} \psi^{\text{\tiny{(\(j\))}}}(\Oms,\Omi)\to \frac{g{\tau_\mathrm{p}}}{\sqrt{2\pi}} e^{i\left[t_{As} \Oms+ t_{Ai} \Omi\right]} e^{-\sum_{i,j=s,i} c_{ij} \Om_i \Om_j}\] where constant phase terms have been omitted and the real coefficients \(c_{ij}\) \[see also \] are \[\begin{aligned}
c_{ss}&=&\frac{{\tau_\mathrm{p}}^2}{2}+\gamma {\tau_{s}} ^2,\label{c11}\\ c_{ii}&=&\frac{{\tau_\mathrm{p}}^2}{2}+\gamma{\tau_{i}^{\text{\tiny{\(\pm\)}}}}^2,\label{c22}\\ c_{is}&=&\frac{{\tau_\mathrm{p}}^2}{2}+\gamma {\tau_{s}} {\tau_{i}^{\text{\tiny{\(\pm\)}}}}.\label{c12} \end{aligned}\] The approximation [\[psi_gauss\]](#psi_gauss){reference-type="eqref" reference="psi_gauss"} allows to extract analytical formula, providing a straightforward comparison between the two cases. Inserting the formula [\[psi_gauss\]](#psi_gauss){reference-type="eqref" reference="psi_gauss"} in the expression of \(\cappa\) in Eqs.[\[kintegral\]](#kintegral){reference-type="eqref" reference="kintegral"}-[\[B\]](#B){reference-type="eqref" reference="B"}, and performing the Gaussian integrals involved, we find a single formula that holds for any PDC configuration: \[\begin{aligned}
\cappaG &=\sqrt{\frac{c_{ss}c_{ii}}{c_{ss}c_{ii}-c_{si}^2}} = \left[ 1+\left( 1 + \frac { 2\gamma\, {\tau_{s}} {\tau_{i}^{\text{\tiny{\(\pm\)}}}} } {{\tau_\mathrm{p}}^2} \right)^2 \frac{{\tau_\mathrm{p}}^2}{2\gamma |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |^2} \right]^{1/2} \label{Kgauss} \end{aligned}\] and gives a simple way of estimating the degree of purity \({1} /{\cappa}\) of heralded single photons for any crystal length and dispersion relations in the medium. We notice that in this formula the various parameters that characterize the PDC process (crystal length, dispersion relations for the three waves) have been condensed in only two parameters, i.e. in the two time constants \({\tau_{s}}\) and \({\tau_{i}^{\text{\tiny{\)\pm\(}}}}\) that rule the relative propagation of the three waves in a crystal of length \(l_c\).\
The Gaussian formula [\[Kgauss\]](#Kgauss){reference-type="eqref" reference="Kgauss"}, seen as a function of the pump duration \({\tau_\mathrm{p}}\) (see Fig.[\[figKtaup\]](#figKtaup){reference-type="ref" reference="figKtaup"}), presents for long pump pulses a linear asymptote: \[\cappaG \to \frac{{\tau_\mathrm{p}}}{\sqrt{2\gamma} |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |} \qquad \text{ for } {\tau_\mathrm{p}} \gg \sqrt{ 2\gamma |{\tau_{i}^{\text{\tiny{\(\pm\)}}}} {\tau_{s}} |}. \label{KGauss_long}\] corresponding to an entangled state with \(\cappa \gg 1\); by shortening the pump it reaches a minimum, and then it stays close to the curve \[\cappaG \to \frac{ \sqrt{2\gamma} | {\tau_{s}} {\tau_{i}^{\text{\tiny{\(\pm\)}}}}|}{{\tau_\mathrm{p}} |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |} \qquad \text{ for }{\tau_\mathrm{p}} \ll \sqrt{ 2\gamma |{\tau_{i}^{\text{\tiny{\(\pm\)}}}} {\tau_{s}} |} \label{Kgauss_short}\] \(\cappaG\) takes its minimum (the purity takes its maximum) for \[\label{taupmin} {\tau_\mathrm{p}}^{min} =\sqrt{2\gamma| {\tau_{s}} {\tau_{i}^{\text{\tiny{\(\pm\)}}}}|}\] which is basically the geometrical mean between the two time constants \(|\tau_s|\) and \(|{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\). The minimum value of \(\cappa\) depends both on the sign and on the magnitude of \(\eta= {\tau_{s}} /{\tau_{i}^{\text{\tiny{\)\pm\(}}}}\) and is given by \[\cappaG^{min} = \begin{cases} 1&\text{for \( {\tau_{s}} {\tau_{i}^{\text{\tiny{\)\pm\(}}}} \leq 0 \rightarrow \eta \leq 0\)} \\ \frac{{\tau_{i}^{\text{\tiny{\(\pm\)}}}}+ {\tau_{s}} ^{\text{\tiny{(\(j\))}}}}{{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} ^{\text{\tiny{(\(j\))}}}} =\frac{1+\eta^{\text{\tiny{(\(j\))}}}}{1-\eta^{\text{\tiny{(\(j\))}}}} \; &\text{for \( {\tau_{s}} {\tau_{i}^{\text{\tiny{\)\pm\(}}}} > 0 \rightarrow \eta >0 \)} \end{cases} \label{Kmin}\] As well known, the Gaussian approximation ([\[psi_gauss\]](#psi_gauss){reference-type="ref" reference="psi_gauss"}) predicts complete separability \(\cappa=1\) for \(\eta \leq 0\), as can be immediately recognized by inspection of the mixed term \(c_{is}\) in Eq. [\[c12\]](#c12){reference-type="eqref" reference="c12"}. The condition \(\eta=0\), corresponds to the asymmetric group velocity matching of the example (ii), and the ideal value \(\cappa=1\) is reached only asymptotically for \({\tau_\mathrm{p}} \to {\tau_\mathrm{p}}^{min} =0\). The condition \(\eta <0\) requires that \({\tau_{s}}\) and \({\tau_{i}^{\text{\tiny{\)\pm\(}}}}\) have opposite signs (as in the example (II) for \(\eta=-1\)) and a separable state can be in principle obtained for a finite pump duration \({\tau_\mathrm{p}}= {\tau_\mathrm{p}}^{min}\). These conditions straightforwardly lead to a separable Gaussian spectral amplitude [\[psi_gauss\]](#psi_gauss){reference-type="eqref" reference="psi_gauss"}.\
Alternatively, for positive \(\eta\), the two-photon state can be made almost separable by choosing a configuration for which \(\eta\) is sufficiently small, because \[\cappaG^{min} = \frac{1+\eta^{\text{\tiny{(\(j\))}}}}{1-\eta^{\text{\tiny{(\(j\))}}}} \simeq 1+ 2\eta \qquad \text{for \( 0<\eta \ll 1\)}\] Notice that this last condition is naturally fulfilled in the counter-propagating case.
Figures [\[figKtaup\]](#figKtaup){reference-type="ref" reference="figKtaup"} and [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} plot the Schmidt number \(\cappa\) as a function of the pump duration, for the three examples described in Table [\[table1\]](#table1){reference-type="ref" reference="table1"}. Fig.[\[figKtaup\]](#figKtaup){reference-type="ref" reference="figKtaup"} superimpose the three results, calculated in the Gaussian approximation ([\[Kgauss\]](#Kgauss){reference-type="ref" reference="Kgauss"}), while Fig.[\[fig3\]](#fig3){reference-type="ref" reference="fig3"} (notice here the horizontal logarithmic scale) shows, for each example, a comparison between the Gaussian result and the exact one, obtained by numerical integration of Eqs.([\[kintegral\]](#kintegral){reference-type="ref" reference="kintegral"})-([\[B\]](#B){reference-type="ref" reference="B"}). The Schmidt number curves reproduce in a more quantitative way the behaviour of the temporal correlation analysed in Sec. [2](#sec:temporal){reference-type="ref" reference="sec:temporal"} (notice that the hollow dots in Fig.[\[fig3\]](#fig3){reference-type="ref" reference="fig3"} correspond to the parameters of the plots of in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}): for long pump pulses the state is entangled in all the examples, by shortening the pump pulse \(\cappa\) reaches a minimum close to the ideal value \(\cappa^{min}=1\) and then it grows again in the two examples with \(\eta \ne 0\), while it stays close to the minimum in the example (ii) with \(\eta=0\). However, evident from these figures are the different ranges of pump durations \({\tau_\mathrm{p}}\) in which separability can be achieved: the counter-propagating case displays a broad plateau with \(\cappa \simeq 1\) for \({\tau_\mathrm{p}}\) in the range between \(2\)ps and \(10\)ps, while in the co-propagating case separability requires supbicosecond pulses, of duration \({\tau_\mathrm{p}}\lesssim 100\)fs \(\ll{\tau_{i}^{\text{\tiny{\)-\(}}}}=720\)fs in the \(\eta=0\) case (ii), and \({\tau_\mathrm{p}}\simeq {\tau_\mathrm{p}}^{min}=147\)fs in the \(\eta=-1\) case (iii).\
Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} also shows some discrepancy between the Gaussian results and the exact ones, especially for short pump pulses. In particular, the minimum of \(\cappa\) is always slightly larger than the Gaussian result [\[Kmin\]](#Kmin){reference-type="eqref" reference="Kmin"}, and never reaches the ideal value \(\cappa=1\) even for \(\eta \le 0\). Clearly these discrepancies have to be ascribed both to the replacement of the sinc by a Gaussian and to the effect of dispersion, which becomes relevant only for short pump pulses, and/or long crystal. Because of this, the degree of purity achievable in the counter-propagating configuration with \(\eta=0.01\) is comparable to what obtained in the two cases with \(\eta\le0\), which require much shorter pump pulses and stringent phase matching conditions.
It is also interesting to see how these results reflect in the shape of the joint spectral probability \(|\psi|^2\) in the \((\Om_i, \Oms)\) plane, plotted in Fig.[\[fig4\]](#fig4){reference-type="ref" reference="fig4"} for the same parameters as in Fig. [\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}. The red ellipses here represent the curves \(c_{11}\Om_s^2+c_{22}\Om_i^2+2c_{12}\Om_s\Om_i= 1\) where according to the Gaussian formula ([\[psi_gauss\]](#psi_gauss){reference-type="ref" reference="psi_gauss"}) \(|\psi|^2\) reduces by \(1/e^2\), and permit to estimate visually the validity of the Gaussian approximation. The spectral correlation in this figure gives the complementary view with respect to the temporal correlation in Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}, but obviously displays the same amount of entanglement: For long pump pulses (top raw) the state is highly entangled in all the three examples, with the biphoton amplitude peaked along the diagonal \(\Oms=-\Omi\) where energy conservation takes place. For short enough pump pulses \[bottom panels (c) and (i)\] the state is again entangled, with the biphoton amplitude peaked along the line \(\Oms=-\eta \Omi\) where momentum conservation, i.e. phase matching, is realized. Panels (b) (f) and (h) correspond to the optimal conditions for separability, and we see that in these cases the biphoton probability is a nearly factorable function of its arguments, being approximately an ellipse with its principal axes aligned along \(\Omega_s\) and \(\Omega_i\).
# Interpretation and discussion {#sec:interpretation}
In order to understand these behaviors, and to be able to describe the peculiarity of the co-propagating geometry (case \(\eta \approx-1\) ), we resort to the very notion of temporal correlation between the members of a pair, which relies on the possibility of predicting the arrival time of one photon by detecting the arrival time of its twin, with a precision better than what would be obtained with an unconditional measurement.\
Referring for definiteness to the signal photon (the one that in any case co-propagates with the pump), we first wonder what is the uncertainty in its arrival time conditioned to detection of the idler. Then we will compare it with the uncertainty in its arrival time not conditioned on the detection of the idler.\
There are actually two distinct mechanisms by which the exit time of one photon can be ascertained by detection its twin: one is simply based on their simultaneity, the other relies on the possibility to gain information on the point where down-conversion took place.\
The first one prevails for long pump pulses \({ {\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\ge | {\tau_{s}} |}\). When the duration of the pump pulse is much longer than the delays of twin photons with respect to its center, then the idler and signal wavepackets propagate under the pump pulse, and detection of neither photon gives information on the point of the crystal where the pair was generated. Thus, our ability to infer the arrival time of one photon by detection of its twin is limited by the degree to which they arrive simultaneously at their end faces, i.e. by the spread of the distribution of their mutual time delays. This is described by the box function in Eqs. [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} and[\[asym1\]](#asym1){reference-type="eqref" reference="asym1"}, characteristic of the spontaneous PDC processes, where a photon pair can be generated at any point along the crystal with uniform probability. Then, depending on the point where down-conversion took place, the delay between the exit of the twins has a flat distribution in the interval \([-\Delta \tau, + \Delta \tau ]\) [^1] where \(\Delta \tau:=\frac{l_c}{2}\left|\frac{1}{v_{gs}}\pm\frac{1}{v_{gi}}\right|\). Notice that counter-propagating photons (\(+\) sign) can be delayed up to their transit time along the sample, since they appear on opposite faces of the crystal. Conversely, co-propagating twins (\(-\) sign) exit from the same face, and appear with a small delay ruled by their GVM, the extreme case being when they propagate exactly at the same velocity, and always exit simultaneously, in which case the width of the box function vanishes. According to this mechanism the exit time of the signal can be deduced from that of the idler as \[\begin{aligned}
\quad \bar t_s = \bar t_i &\qquad \text{within} \; {\Delta\tau } \simeq \Delta \tau= | {\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} |&\qquad \text{for }\; {\tau_\mathrm{p}} \gg|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}| \label{tcorr_long} \end{aligned}\] where \({\Delta\tau }\) indicates the spread of the distribution of the signal arrival time conditioned to detection of the idler, i.e the *correlation time*.
The second mechanism, described by the factor \(\alpha_p \left(\frac{\bar{t}_s-\eta\bar{t}_i}{1-\eta}\right)\) in Eq. [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"}, prevails for ultrashort pump pulses \({\tau_\mathrm{p}} \ll | {\tau_{s}} | \le |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\). In this case the twin photons tend to separate from the pump during propagation, and detection of any of them provides an indication of the point where the pair was generated. By using a rough picture of photons as wavepackets propagating without deformation with their group velocities, one can imagine that if a photon pair was generated at a point \(z_0\), and conversion occurred from a pump photon delayed by \(\delta t_p\) from the centre of the pump pulse, than the idler photon will arrive at its end face at time \[t_i =\begin{cases} \frac{z_0 }{v_{gp} } + \frac{z_0 }{v_{gi}} + \delta t_p \quad &\text{counter-prop. case } \\ \frac{\ z_0 }{v_{gp} } + \frac{ l_c-z_0}{v_{gi}} + \delta t_p \quad &\text{co-prop. case } \end{cases}\] i.e, for the barred time argument \(\bar t_i = t_i-\frac{1}{2}\left( \frac{l_c}{v_{gp} } + \frac{l_c}{v_{gi} } \right)\): \[\begin{aligned}
\bar t_i &= \left(z_0-\frac{l_c}{2} \right) \frac{2}{l_c} {\tau_{i}^{\text{\tiny{\(\pm\)}}}} + \delta t_p \label{ti} \end{aligned}\] The same argument gives the arrival time of its twin signal photon as \[\begin{aligned}
\bar t_s &= \left(z_0-\frac{l_c}{2} \right) \frac{2}{l_c} {\tau_{s}} + \delta t_p \label{ts} \end{aligned}\] where in the above formulas \(\delta t_p\) can be considered as a Gaussian stochastic variable with variance \({\tau_\mathrm{p}}/\sqrt{2}\), because a pump photon can be downconverted at any time along the pump pulse, with a probability proportional to the Gaussian pump intensity. Then, by comparing Eqs ([\[ti\]](#ti){reference-type="ref" reference="ti"}) and ([\[ts\]](#ts){reference-type="ref" reference="ts"}): \[\begin{aligned}
\bar t_s = \bar t_i \frac{ {\tau_{s}} }{{\tau_{i}^{\text{\tiny{\(\pm\)}}}}} + \delta t_p \left( 1-\frac{ {\tau_{s}} }{{\tau_{i}^{\text{\tiny{\(\pm\)}}}}} \right) = \eta \bar t_i + (1-\eta) \delta t_p \end{aligned}\] Thus, according to these arguments, the arrival time of the signal conditioned to detection of the idler is \[\quad \bar t_s = \eta \bar t_i \quad \text{within} \; {\Delta\tau } \simeq T_p (1-\eta) \quad ( {\tau_\mathrm{p}} \ll | {\tau_{s}} | )\]
These conditional uncertainties have to be compared with the unconditional uncertainty in the arrival time of the signal., i.e. the spread of the distribution of the exit times of the signal when the idler is not detected. Also here there are two sources of uncertainty, one is the width \({\tau_\mathrm{p}}\) of the Gaussian distribution of the pump, because the photon can be downconverted from any portion of the pump pulse, the other one is the point of the crystal where down-conversion took place. Depending on the latter, the delay of the signal photon with respect to the center of the pump pulse ranges in the interval \[0, \(2 | {\tau_{s}} |]\) with uniform probability. Clearly, the first mechanism dominates for \({\tau_\mathrm{p}} \gg | {\tau_{s}} |\), as shown by Eqs. [\[asym1\]](#asym1){reference-type="eqref" reference="asym1"} and [\[asym3\]](#asym3){reference-type="eqref" reference="asym3"}, the second mechanism for \({\tau_\mathrm{p}} \ll | {\tau_{s}} |\) \[see Eq. [\[asym2\]](#asym2){reference-type="eqref" reference="asym2"}\]. A more precise analysis, based on the temporal correlation[\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} gives \[\begin{aligned}
(\delta t_s )_{\mathrm uncond.} &= \sqrt{\frac{T_p^2}{2} + \frac{ {\tau_{s}} ^2}{3} }\to \begin{cases} \frac{T_p}{\sqrt{2}} & {\tau_\mathrm{p}} \gg |\tau_s|\vspace{4pt } \\ \frac{|\tau_s|}{\sqrt{3}} & {\tau_\mathrm{p}} \ll |\tau_s| \end{cases} \label{unconditional} \end{aligned}\] where \((\delta t_s )_{\mathrm uncond.}\) is the variance of the distribution of the arrival times of the signal when the idler is not detected.
Summarizing the results, we have the following situation:
\(\boldmath{ {\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\ge | {\tau_{s}} |}\), comparing the unconditional and conditional uncertainties in the arrival time of the signal, in Eqs. [\[unconditional\]](#unconditional){reference-type="eqref" reference="unconditional"} and [\[tcorr_long\]](#tcorr_long){reference-type="eqref" reference="tcorr_long"}, we see that for such long pumps \({\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\), the state is in general highly entangled because [^2]. \[\frac{ (\delta t_s )_{\mathrm uncond.}} { {\Delta\tau } } \simeq \frac{{\tau_\mathrm{p}}}{ |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} | } = \frac{{\tau_\mathrm{p}}}{ |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}| (1-\eta) }\gg 1 \label{ratio1}\] Euristically, the above ratio also gives an estimate of the number of entangled modes, well in accordance with the asymptotic behaviour of the Schmidt number in Eq.[\[KGauss_long\]](#KGauss_long){reference-type="eqref" reference="KGauss_long"} for long pump pulses.\
\(\boldmath {\tau_\mathrm{p}} \ll | {\tau_{s}} | \le |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\), the state is again highly entangled because \[\frac{ (\delta t_s )_{\mathrm uncond.}} { {\Delta\tau } } \simeq \frac{| {\tau_{s}} |}{ {\tau_\mathrm{p}}(1-\eta) } = \frac{| {\tau_{s}} {\tau_{i}^{\text{\tiny{\(+\)}}}}|}{ {\tau_\mathrm{p}}|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}-{\tau_{s}} | } \gg 1 \label{ratio2}\] for \({\tau_\mathrm{p}} \ll | {\tau_{s}} |\). Also in this case the ratio in Eq.[\[ratio2\]](#ratio2){reference-type="eqref" reference="ratio2"} can be considered an estimate of the number of entangled modes, and indeed reproduces the asymptotic behaviour of the Schmidt number in Eq.[\[Kgauss_short\]](#Kgauss_short){reference-type="eqref" reference="Kgauss_short"}.
Following the above arguments, in particular the results [\[ratio1\]](#ratio1){reference-type="eqref" reference="ratio1"} and [\[ratio2\]](#ratio2){reference-type="eqref" reference="ratio2"}, there is no chance that the state become separable when \({\tau_{s}} \simeq {\tau_{i}^{\text{\tiny{\)\pm\(}}}}\), i.e. \(\Delta \tau\simeq 0\). When the two photons propagate in the same direction at similar velocities, in fact, detection of one photon will always provide an extremely precise information about the arrival time of the other, basically because they exit the crystal almost simultaneously. This is the typical situation that occurs in the co-propagating case, in the absence of any velocity matching strategy, and explains why co-propagating photons in general display high temporal entanglement.\
Conversely, in the counterpropagating geometry the exit times of the twins lack simultaneity because they propagate in opposite directions, and this case is naturally characterized by a strong asymmetry between the twins, in particular they have very different delays \({\tau_{i}^{\text{\tiny{\)+\(}}}} \gg | {\tau_{s}} |\) from the pump. Because of that, the separable limit [\[asym3\]](#asym3){reference-type="eqref" reference="asym3"} can be always realized for intermediate pump durations \({\tau_{i}^{\text{\tiny{\)+\(}}}}\gg {\tau_\mathrm{p}} \gg | {\tau_{s}} |\) which need not to be ultrashort because \({\tau_{i}^{\text{\tiny{\)+\(}}}}\) is a long time scale. This limit physically corresponds to a situation where the forward photon (here the signal) always propagate below the pump pulse, but the pump pulse is short enough that the backward photon rapidly separate from it. In these conditions the temporal localization of the pump provides an absolute timing information on the exit time of the signal, as precise (\(\sim {\tau_\mathrm{p}}\)) as the information that can be gained by detecting the idler \(\sim {\tau_\mathrm{p}} (1-\eta) \approx {\tau_\mathrm{p}}\), and the exit times of the twins appear uncorrelated. An other way of looking at the situation is that the forward photon is locked to the pump so that it cannot provide any information on the point where down-conversion occurred, and by detecting the signal one does not gain any more precise information on the exit time of the backward idler than by not detecting it.
In the co-propagating configuration, the same limit can be reached only by creating a strong asymmetry between the propagation velocities of the two photons relative to the pump \(| {\tau_{s}} | \ll |{\tau_{i}^{\text{\tiny{\)-\(}}}}|\), which in practice requires that the signal is velocity matched to the pump as in the example (ii). However, in order that the idler separate from the pump pulse along a finite propagation length, the pump pulse must be in this case ultrashort (Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}f). As a consequence, as we shall see in the following, the generated twin photons are broadband.
An alternative strategy, which can be realized only in the co-propagating case, is the so-called symmetric velocity matching where \({\tau_{s}} =-{\tau_{i}^{\text{\tiny{\)-\(}}}}\) and \(\eta =-1\), as in the example (iii). This condition, which is rather difficult to achieve, requires that one photon propagate faster than the pump while the other is slower, so that they are symmetrically delayed with respect to the pump center. Negative values of \(\eta\) are in general favorable to separability, because in this case \(\Delta \tau =|{\tau_{i}^{\text{\tiny{\)\pm\(}}}} | + | {\tau_{s}} | >|{\tau_{i}^{\text{\tiny{\)\pm\(}}}} |, | {\tau_{s}} |\), and twin photons are more simultaneous with the pump than between themselves. In these conditions, a properly localized pump pulse may provide a timing of the exits of the twins more precise than the information that can be inferred from detection of any of them. Focusing e.g on the case \(\eta\)=-1, when the pump duration \({\tau_\mathrm{p}}\) approaches \(|{\tau_{i}^{\text{\tiny{\)-\(}}}}| =| {\tau_{s}} |\), both photons are still locked under the pump, so they do not provide sufficient information on the point where generation occurred, and on the other side they are no more simultaneous than the pump duration \(\Delta \tau =2|{\tau_{i}^{\text{\tiny{\)-\(}}}}| >{\tau_\mathrm{p}}\). [^3]\
This configuration produces rather pure photons (see Figs.[\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(iii) and [\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}h). However, their purity is rather far from the ideal result \(\cappa_{min} =1\) predicted by the Gaussian formula [\[Kmin\]](#Kmin){reference-type="eqref" reference="Kmin"}, which seems rather an artifact of replacing the sinc by a Gaussian. Actually, if one looks at Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}h, which corresponds to the minimum of \(\cappa\) for \(\eta=-1\), the joint temporal probability exhibits a diamond-like shape, which does not look really separable in its arguments, and indeed \(\cappa_{min}=1.18\) is slightly higher than in the other two cases. Notice that in contrast the biphoton probability in the spectral domain in Fig.[\[fig4\]](#fig4){reference-type="ref" reference="fig4"}h has a nearly circular shape which looks much more factorable (actually in this case deviations from factorability are indicated by the sidelobes of the sinc functions ) This demonstrates the usefulness of taking the two complementary views in the temporal and spectral domains. Such imperfect separability is a consequence of the sharp boundaries of the nonlinear medium, which are at the origin of the rectangular box function appearing in Eq.([\[psitemp\]](#psitemp){reference-type="ref" reference="psitemp"}). Taking the Gaussian approximation ([\[psi_gauss\]](#psi_gauss){reference-type="ref" reference="psi_gauss"}), this box function is replaced by a smooth Gaussian of the same width, \({\rm Rect}\left(\frac{\bar{t}_s-\bar{t}_i}{\Delta \tau}\right) \to \exp\left(-\frac{(\bar{t}_s-\bar{t}_i)^2}{4\gamma \Delta \tau^2}\right)\), and then the temporal correlation becomes factorable: \[\begin{aligned}
\phi(\bar{t}_s,\bar{t}_i) &\propto {\alpha}_p\left(\frac{\bar{t}_s+\bar{t}_i}{2}\right) {\rm Rect}\left(\frac{\bar{t}_s-\bar{t}_i}{\Delta \tau}\right) \to e^{-\frac{\bar{t}_s^2+\bar{t}_i^2}{4\gamma(\Delta \tau)^2}} & \text{for \({\tau_\mathrm{p}}= {\tau_\mathrm{p}}^{min}= \sqrt{2\gamma} |{\tau_{i}^{\text{\tiny{\)-\(}}}}|\)} \label{phi_fact} \end{aligned}\] (the same result is obtained by directly Fourier transforming the Gaussian approximation of \(\psi(\Om_s,\Om_i)\) in Eq.([\[psi_gauss\]](#psi_gauss){reference-type="ref" reference="psi_gauss"})). Notice that the sharp crystal boundaries have less impact on the separability when \(|\eta| \ll 1\), as in the examples i) and ii), because of the elongated shape of the temporal correlation (see Fig.[\[fig4_temp\]](#fig4_temp){reference-type="ref" reference="fig4_temp"}b and f). These effects might be eliminated and the purity of the heralded photons increased by engineering a Gaussian nonlinearity profile of a poled crystal, as demonstrated by Branczyk et al..
# Spectro-temporal properties of heralded photons {#sec:spectrum}
Once assessed the conditions under which pure single heralded photons can be obtained, it is clearly important to study their individual properties in view of their use in quantum communication/metrology protocols (coupling with atoms, interferometry etc.).
For spontaneous PDC, the marginal statistics of individual photons is described by their first order coherence functions. In the temporal domain, for example, these functions are: \[\begin{aligned}
G_s^{(1)}(t_s, t_s') &=&\langle \hat a_s^{\dagger} (t_s) \hat a_s (t_s') \rangle = \int d t_i \phi^*(t_s, t_i)\phi(t_s', t_i) \label{G1s}\\ G_i^{(1)}(t_i, t_i') &=&\langle \hat a_i^{\dagger} (t_i) \hat a_i (t_i') \rangle = \int d t_s \phi^*(t_s, t_i)\phi(t_s, t_i') \label{G1i} \end{aligned}\] with analogous definitions in the spectral domain. In terms of its temporal \(G^{(1)}\), the reduced state of the heralded photon takes the form \[\begin{aligned}
\hat \rho_j &= \frac{1}{\cal N} \int dt \int dt' G_j^{(1)}(t', t) \, \hat a_j^\dagger (t) |0 \rangle\, \langle 0| \hat a_j (t') \nonumber \\ &=\frac{1}{\cal N} \int dt \int dt' G_j^{(1)}(t', t) \, |t \rangle_j \phantom{i}_j\langle t'| \label{rhoj} \end{aligned}\] which in general represents a mixed state, unless the coherence function \(G_j^{(1)}(t, t')\) is factorable in its arguments.\
For equal times, \(G_j^{(1)}(t, t) = \langle \hat A_j^{\dagger\, \mathrm{out}} (t) \hat A_j^{\mathrm{out}} (t) \rangle = {\cal I}_j(t)\) gives the temporal intensity profile of the j-th wave, i.e. the probabilty distribution of detecting the photon at time \(t\) at the crystal exit face. The spectral distribution of the heralded photon can be instead obtained via \[\begin{aligned}
S_j(\Om) = \langle \hat A_j^{\dagger\, \mathrm{out}} (\Om) \hat A_j^{\mathrm{out}} (\Om) \rangle= \int \frac{dt \, dt'} {2\pi} e^{i\Om (t-t')} G_j^{(1)}(t, t') \end{aligned}\] For the sake of brevity we limit our analysis to the cases of A) long pump pulses, where the state is highly entangled, and B) conditions where nearly separable states can be realized.
## Long pump pulses, high entanglement
According to the previous results, when \({{\tau_\mathrm{p}} \gg |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\ge | {\tau_{s}} |}\) the two-photon state state is highly entangled in any configuration. We can for example use the asymptotic form [\[asym1\]](#asym1){reference-type="eqref" reference="asym1"} of the joint temporal amplitude[\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"}, and insert it into formulas ([\[G1s\]](#G1s){reference-type="ref" reference="G1s"},[\[G1i\]](#G1i){reference-type="ref" reference="G1i"}). The coherence functions of the signal and the idler obtained in this way are identical (a part from negligible offsets), and have the form \[\begin{aligned}
G_s^{(1)}(t, t')&= G_i^{(1)}(t, t') \to \frac{g^2}{2 \Delta \tau } \left| \alpha_p (t) \right|^2 \mathrm{T} \left( \frac{t'-t}{2 \Delta \tau } \right) \label{G1long} \end{aligned}\] where \(\Delta \tau=|{\tau_{i}^{\text{\tiny{\)\pm\(}}}}-{\tau_{s}} |\), and \[\mathrm{T} \left( \frac{\delta t }{2\Delta \tau } \right) = \begin{cases} 1-\frac{|\delta t |}{2 \Delta \tau } & \delta t \epsilon (-2 \Delta \tau, +2 \Delta \tau) \\ 0 & \text{elsewhere} \end{cases}\] is the triangular function, which has the shape of a triangle of base \(4\Delta \tau\) and unitary height. This result is the generalizazion of formula (46) in Ref.. Actually, it holds for any PDC configuration provided that the bandwidths in play are not too broad, because it makes only use of the linear approximations [\[Dlin\]](#Dlin){reference-type="eqref" reference="Dlin"} and [\[betalin\]](#betalin){reference-type="eqref" reference="betalin"}. It shows that when the pump pulse is much longer than the two characteristic time scales, the twin photons have identical properties. The width \(\Delta \tau\) of \(G^{(1)} (t,t')\) as a function of the time difference \(t'-t\) is their *coherence time*, which in this limit is equal to their mutual correlation time \[see Eq. [\[tcorr_long\]](#tcorr_long){reference-type="eqref" reference="tcorr_long"}\]. Conversely, the temporal distributions of twin photons follow the profile of the much longer pump pulse: \[{\cal I}_s(t) ={\cal I}_i(t) = \frac{g^2}{2 \Delta \tau } \left| \alpha_p (t) \right|^2\] Notice that from the point of view of classical statistics of light, this behaviour of the \(G^{(1)}\), with the peak at \(t=t'\) much narrower of the intensity distribution is typical of multimode incoherent light. On the other hand, in the quantum description, the state of the heralded photon in Eq. [\[rhoj\]](#rhoj){reference-type="eqref" reference="rhoj"} is in this case mixed, because \(G_j^{(1)}(t, t')\) in Eq.[\[G1long\]](#G1long){reference-type="eqref" reference="G1long"} is not factorable in its arguments.\
The spectra of the photons have the usual \({\rm sinc}^2\) shape characteristic of spontaneous processes \[\begin{aligned}
S_s(\Om) = S_i(\Om) = \frac{g^2}{\sqrt{4\pi}} {\tau_\mathrm{p}} {\rm sinc}^2 (\Om \Delta \tau), \label{S} \end{aligned}\] and their spectral bandwidths \(\Delta \Omega_j = \frac{1}{\Delta \tau}\) are the inverse of the correlation-coherence times. Clearly, this explains why counterpropagating twin-photons are narrowband (order 10 GHz), while co-propagating twin-photons are in general broadband (order Thz or more)\
## Nearly separable regime
We start from the case \(\eta \ll 1\), including both the counter-propagating and the co-propagating case with asymmetric group velocity matching. We remind that for a pump pulse intermediate between the two time scales \(|{\tau_{i}^{\text{\tiny{\)\pm\(}}}}| \gg {\tau_\mathrm{p}} \gg | {\tau_{s}} |\), one photon propagates locked under the pump while the other separates from it after being generated. For this reason the properties of the two photons are strongly asymmetric, as e.g. shown by the spectra in column (i) and (ii) of Fig.[\[fig6\]](#fig6){reference-type="ref" reference="fig6"}. By using the asymptotic limit[\[asym3\]](#asym3){reference-type="eqref" reference="asym3"} of the joint temporal amplitude, and inserting it into formulas ([\[G1s\]](#G1s){reference-type="ref" reference="G1s"},[\[G1i\]](#G1i){reference-type="ref" reference="G1i"}), we obtain for the signal photon: \[\begin{aligned}
G_s^{(1)}(\bar{t}_s, \bar{t}_s') &=\frac{g^2}{2 \Delta \tau} \alpha_p ^*(\bar{t}_s) \alpha_p (\bar{t}_s') \end{aligned}\] which implies that \[\begin{aligned}
&{\cal I}_s ( t ) = \frac{g^2}{2 \Delta \tau} \left|\alpha_p ( t-t_{As} ) \right|^2, \\ &{S}_s (\Om) = \frac{g^2}{2 \Delta \tau} |\tilde \alpha_p (\Om) |^2 \label{Ss} \end{aligned}\] i.e. the photon that propagates locked under the pump entirely inherits the spectro-temporal properties of the coherent pump laser.\
Conversely for the idler photon \[\begin{aligned}
G_i^{(1)}(t, t') &= \frac{g^2}{2 \Delta \tau} \frac{ \sqrt{\pi} {\tau_\mathrm{p}}} {2\Delta \tau} \mathrm{Rect }\left( \frac{\bar{t}_i}{\Delta \tau} \right) \mathrm{Rect }\left( \frac{\bar{t}_i^\prime}{\Delta \tau} \right) , \end{aligned}\] and its properties depend on the dispersion properties of the crystal and on its length, through the parameter \(\Delta \tau= |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}-{\tau_{s}} |\). In particular \[\begin{aligned}
&{\cal I}_i (t) = \frac{g^2}{2 \Delta \tau } \frac{ \sqrt{\pi} {\tau_\mathrm{p}}} {2\Delta \tau} \mathrm{Rect }\left( \frac{t-t_{Ai}}{\Delta \tau} \right), \\ &{S}_i (\Om) = \frac{g^2}{\sqrt{4\pi}} {\tau_\mathrm{p}} {\rm sinc}^2 (\Om \Delta \tau), \label{Si} \end{aligned}\] The temporal distribution of the idler photon is a rectangular pulse of duration \(2 \Delta \tau\), because for such a short pump the idler photon distribution reflect simply the fact that it can be generated anywhere in the crystal with uniform probability. Remarkably, since \(\Delta \tau\simeq |{\tau_{i}^{\text{\tiny{\)\pm\(}}}} | \gg {\tau_\mathrm{p}}\), the idler wavepacket has a longer duration than the pump itself. In particular, in the counterpropagating case, its duration roughly corresponds to the transit time of the pump along the crystal. The idler spectrum retains the same \({\rm sinc}^2\) shape as in the long pump limit, with a bandwidth \(\Delta \Omega_i = 1/ \Delta \tau \ll \Delta \Omega_p\). Hence, when the state is approximately separable, the heralded idler photon is not only pure, but also more monochromatic than the pump laser that drives the process. However, only in the case (a) this result really means that the generated idler photon is narrowband, as shown by figure [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}d, while in case b) it is anyway quite broadband (see Fig.[\[fig6\]](#fig6){reference-type="ref" reference="fig6"}e), because separability requires ultra-broadband pump pulses. Interestingly, in the counterpropagating case the strong asymmetry between the twin photons, and the fact that the idler is more monochromatic than the pump, reflect the unusual coherence properties of the classical signal and idler field generated above the MOPO threshold, described in.\
The forms of the coherence functions of the signal and idler, \(G^{(1)} (t,t') = \sqrt{ {\cal I} (t)} \sqrt{{\cal I} (t')}\), is typical of single-mode light, which possess classical temporal coherence: indeed it means that the average length of a temporal fluctuation (the coherence time) is equal to the duration of the wave-packet, implying that the statistics is a sort of frozen in time. In the quantum description the states of the heralded signal and idler photon are pure \(\hat \rho_j = |\psi_j\rangle\, \langle \psi_j |\), with \[\begin{aligned}
|\psi_s \rangle &=\frac{1}{(\sqrt{\pi} T_p)^{1/2} } \int dt_s \alpha_p \left(t_s-t_{A_s} \right) \, \hat a_s^\dagger (t_s) |0 \rangle \\ |\psi_i \rangle &= \frac{1}{ \sqrt{2\Delta \tau} } \int dt_i \mathrm{Rect }\left( \frac{t_i-t_{Ai}}{\Delta \tau} \right) \, \hat a_i^\dagger (t_i) |0 \rangle \end{aligned}\]
Finally we consider the alternative technique for reaching separability, i.e. the configuration with \(\tau_s=-\tau_i\) (\(\eta=-1\)), in which the two photons co-propagate symmetrically delayed with respect to the pump center. As can be intuitively understood, in this case the properties of the twin photons are completely symmetric. The mere linear approximation does not give particularly expressive results, so that we have to resort to the stronger Gaussian approximation. According to it, separability is reached at \({\tau_\mathrm{p}}= {\tau_\mathrm{p}}^{\rm min}= \sqrt{2\gamma} |{\tau_{i}^{\text{\tiny{\)-\(}}}}|\), where (calculations not reported here) the state of each heralded photon becomes pure, \(G_j^{(1)} (t,t') = \sqrt{ {\cal I}_j (t)} \sqrt{{\cal I}_j (t')}\), with the temporal and spectral distributions of the two waves given by \[\begin{aligned}
{\cal I}_s(\bar{t}_s ) &= {\cal I}_i(\bar{t}_i ) \propto e^{-\frac{(t-t_{Aj})^2}{2{\tau_\mathrm{p}}^2}} = | \alpha_p \left( \frac{\bar{t}_j}{\sqrt{2}} \right) |^2 \\ S_s(\Om) & =S_i (\Om) \propto | \alpha_p \left( \sqrt{2} \Om \right) |^2 \end{aligned}\] i.e. the two waves have similar properties as the pump, but with a slightly longer duration and a slightly narrower spectrum. As shown by panel (iii) of Fig.[\[fig6\]](#fig6){reference-type="ref" reference="fig6"}, these Gaussian results are rather close to the exact ones.
## Mean number of down-converted pairs-Efficiency
An other important point in view of applications is the efficiency by which heralded pure photons can be generated. In the separable regime, where the state is approximately single-mode, this ultimately depends on the probability that a photon pair is generated by a pump pulse crossing the crystal [^4] For a given pump energy and crystal length, we shall see that the symmetric case \(\eta=-1\) is much more efficient than the asymmetric case \(|\eta| \ll 1\): this is quite natural because in the first case both photons propagate close to the pump, while in the latter case one photonic wavepacket rapidly separates from it, so that the effective interaction length is strongly reduced.\
The probability of generating at least a photon pair can be easily calculated from the state [\[psi\]](#psi){reference-type="eqref" reference="psi"} as \(\frac{\cal N}{1+{\cal N}} \approx {\cal N}\) for \({\cal N} \ll 1\), where \({\cal N}\) is the mean number of photon pairs per pulse. Within the linear approximation, making e.g. use of Eq. [\[psitemp\]](#psitemp){reference-type="eqref" reference="psitemp"} we obtain \[\begin{aligned}
{\cal N} &= \int d t_s \int d t_i \left |\phi(t_s,t_i )\right|^2 =g^2 \frac{\sqrt{\pi}{\tau_\mathrm{p}}} {2 \Delta \tau} \end{aligned}\] Notice that this formula is valid for any geometry and any value of \({\tau_\mathrm{p}}\) (of course, provided that \(\Delta \tau \ne 0\)). However, the number of down-converted pairs may be quite different in different regimes.\
In the long-pump regime, where the state is highly entangled, the number of down-converted photons is approximately \[{\cal N} \simeq \frac{g^2}{2} \cappa \gg g^2,\] where we used the asymptotic expression of the Gaussian Schmidt number in Eq.[\[KGauss_long\]](#KGauss_long){reference-type="eqref" reference="KGauss_long"}\].Roughly speaking, the number of photon pairs is the number of photons per mode that one would have in the strictly CW pump regime, multiplied by the number of entangled modes \(\cappa\).\
Let us then consider the separable regimes. In the symmetric case \(\eta =-1\), \[{\cal N} =\frac{ g^2 }{4} \frac{\sqrt{\pi} {\tau_\mathrm{p}}} {|{\tau_{i}^{\text{\tiny{\(+\)}}}} |} \approx \frac{ g^2 }{4} \sqrt{ 2\pi \gamma}\] where in the last expression we inserted the Gaussian value of the pump duration [\[taupmin\]](#taupmin){reference-type="eqref" reference="taupmin"} \({\tau_\mathrm{p}}^{min} =\sqrt{2\gamma| {\tau_{s}} {\tau_{i}^{\text{\tiny{\)\pm\(}}}}|}\) at which the best separability is achieved. Considering that \(\sqrt{ 2\pi \gamma} \approx 1\), then the mean number of down-converted photons is close to \(g^2/4\).\
In the asymmetric case \(\eta \ll 1\) (counterpropagating or co-propagating case with asymmetric group-velocity matching), instead \[{ \cal N} = \frac{g^2 }{2} \frac{\sqrt{\pi} {\tau_\mathrm{p}} }{(1-\eta)|{\tau_{i}^{\text{\tiny{\(\pm\)}}}}| } \simeq \frac{g^2 }{2} \frac{\sqrt{\pi} {\tau_\mathrm{p}}} { |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}| } \ll \frac{g^2 }{2}\] because separability is achieved for \({\tau_\mathrm{p}} \ll |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\). [^5] Notice that in this latter case the efficiency is strongly reduced, because of the lack of superposition between the pump and the idler wave. We remind that this configuration is characterized by the fact that the idler wavepacket rapidly separates from the short pump pulse \({\tau_\mathrm{p}} \ll |{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\). Since \(g \propto l_c\) it effectively works as if the crystal length were reduced to a shorter length \[l_c \longrightarrow l_c \sqrt{\frac{ {\tau_\mathrm{p}}} { |{\tau_{i}^{\text{\tiny{\(\pm\)}}}}| }}.\] At this point one is naturally lead to wonder whether filtering the spectral or temporal modes would not lead to the same or better efficiency. If a long enough pump pulse is used, the number of photon pairs goes as \(\simeq \cappa \frac{g^2}{2}\). Let us then suppose that it is possible to make a \"clever\" filtering, which projects the state onto a single Schmidt mode: even assuming that all the Schmidt modes have the same weight, a fraction \(1/\cappa\) of the total number of modes would then be transmitted, leading to \({\cal N} = \frac{g^2}{2}\), which is anyway better than what could obtained with asymmmetric group-velocity matching. However, practical considerations, which are outside the scope of this theoretical work, may then lead to conclude that is anyway better to directly generate a factorable state
# Summary and conclusions
In this work we compared different phase-matching configurations which are known to generate pure heralded single photons from spontaneous parametric down-conversion, namely the counter-propagating configuration, and the co-propagating configuration with asymmetric and symmetric group velocity matching. We provided a systematic analysis of the conditions under which uncorrelated twin photons can be generated, which basically hold for any PDC configuration, provided that the bandwidth in play are not too broad. In our description all the properties of the source are condensed in two time scales \({\tau_{s}}, {\tau_{i}^{\text{\tiny{\)\pm\(}}}}\) characteristic of the relative propagation of the three waves inside the medium.\
On the one side, we performed a sort of standard analysis in the spectral domain, where we derived a simple formula for the Schmidt number of entanglement, that permits to evaluate the degree of purity of the heralded photons via a single parameter. On the other side, the less standard analysis of the correlation of twin photons in the temporal domain clarifies in a more physical sense the role of the two characteristic time scales, and permits to understand the mechanism under which the temporal correlation between twin photons can be eliminated. In particular it shows that one way of eliminating this temporal correlation relies on creating a strong asymmetry between the velocities of the two down-converted waves relative to the pump wave, so that one photon propagates locked under the pump while the other, for a pump pulse sufficiently localized inside the crystal, separates from it. In this conditions, the timing provided by the pump pulse may be more precise than that offered by detecting any of the twin photons, and the arrival times of the two photons appear uncorrelated. Such an asymmetry in the relative propagation velocities is naturally present in the counter-propagating geometry, because of the natural separation of the GVM and the GVS time scales, but requires particular tuning conditions in the co-propagating geometry. Because of this unique feature, counter-propagating twin photons in a pure state can in principle be heralded at any frequency by choosing the required poling period. Moreover, the twin photons are naturally narrow-band, especially the one propagating opposite to the pump direction, and separability is achieved for a broad range of pump pulse durations. An other way of eliminating the correlation, which can be implemented only for co-propagating photons, requires that the twin photons propagate symmetrically delayed with respect to the pump pulse. Also in this case our analysis shows that in proper conditions the timing information which can be gained from a localized pump can be better than from any of the twins, but the required pump duration is ultrashort, and the generated twins are broadband. Although the purity of heralded photons generated in this way appears somehow lower, symmetric group velocity matching has the advantage of a higher efficiency of pair production, because both twin photons propagate close to the pump. Conversely, the asymmetric group-velocity matching and the counterpropagating geometry are characterized by a low efficiency, because purity requires that one photonic wavepacket separates from the pump pulse, so that the effective interaction length is reduced.
In conclusion we provided a systematic study of the generation of pure heralded single-photons through spontaneous PDC, which may turn useful for optimizing the existing configurations in view of different applications. For example when narrowband single-photons are required, the counter-propagating configuration is the natural choice. In addition, once the technical challenges for the fabrication of crystals with sub-micrometer poling periods are overcome, this strategy offers more flexibility since it can be virtually implemented for any wavelength, and generates pure single-photons for a broad range of pump durations. In contrast, twin photons emitted in the common co-propagating geometry are naturally broadband and can be generated in a separable state only for very short pulses, under particular tuning conditions. Thus this is the good choice when broadband photons and or high repetition rates are required. Finally, if efficiency is the main issue, the symmetric group velocity matching should represent the best choice.\
We also hope that our general analysis, by providing a deeper and more intuitive understanding of the mechanism through which the temporal entanglement of twin photons con be eliminated, may stimulate new strategies of heralded photon generation.
[^1]: For the original time arguments, \(t_s-t_i = \bar t_s-\bar t_i + (t_{As}-t_{Ai} )\). In the co-propagating case the box function is then shifted to the interval \([0,2\Delta \tau]\) while in the copropagating case the shift is negligible.
[^2]: Notice that \(0<(1-\eta) \le 2\), because of Eq.[\[etarange\]](#etarange){reference-type="eqref" reference="etarange"}
[^3]: Quantitatively, when the pump duration approaches \(|{\tau_{i}^{\text{\tiny{\)\pm\(}}}}|\) in Eq. [\[ratio1\]](#ratio1){reference-type="eqref" reference="ratio1"}, or \(| {\tau_{s}} |\) in Eq. [\[ratio1\]](#ratio1){reference-type="eqref" reference="ratio1"}, then the ratio \(\frac{ (\delta t_s )_{\mathrm uncond.}} { \left. (\delta t_s )\right|_{\bar t_i}} \to \frac{1}{ 1+ |\eta| } <1\).
[^4]: We are assuming an ideal unit efficiency of detection
[^5]: If we use the Gaussian result for \({\tau_\mathrm{p}}^{min}\), we also can write \(N \approx \frac{ g^2}{2} \sqrt{ |\eta| }\). | {'timestamp': '2017-06-07T02:07:14', 'yymm': '1706', 'arxiv_id': '1706.01790', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01790'} |
null | null |
null | null |
# Introduction
# Introduction
If one wants to theoretically search for high-temperature superconductors, which is one of the most challenging problems in condensed matter physics, a guiding principle is imperative. While the large energy scale \(t\) of electrons naively favors high-\(T_c\) in electron mechanisms of superconductivity, we usually end up with \(T_c\) as low as \(\sim t/100\), which is still high with the energy scale \(t \sim \mathcal{O}(1)\) eV, as in the cuprates, a typical class of unconventional superconductors. On the other hand, \(T_c\) of the conventional phonon-mediated superconductors is about several percent of the energy scale of phonons. Given this, a real question is how we can design electron-mechanism superconductivity where the "low\" values of \(T_c\) could be enhanced. A large potential may be expected, since we have various degrees of freedom from lattice structures, orbital characters, etc, for manipulating electronic pairing mechanisms. What would then be the main bottlenecks for enhancing \(T_c\) in the unconventional superconductors? One dilemma is the difficulty in simultaneously realizing a strong pairing interaction and a light renormalized electron mass: A strong electron correlation can mediate strong pairing interactions, but usually makes the electron mass heavy due to the strong quasi-particle renormalization. For instance, if we consider the spin-fluctuation-mediated pairing in the single-band Hubbard model on a square lattice, which is the simplest model for the high-\(T_c\) cuprates, we have indeed the strong pairing interaction simultaneously with the strong quasiparticle renormalization. The former works in favor of superconductivity, whereas the latter works against it. In a previous paper, one of the present authors and his coworkers proposed a possible way to circumvent this problem. Let us consider a system where narrow and wide bands coexist due to a lattice structure such as a ladder, and set the Fermi level close to, but not right within, the narrow band. Then the electrons in the wide band, which are not very strongly renormalized, can form Cooper pairs with a pairing interaction strongly mediated by the large number of interband scattering channels, where the narrow (or even flat) band acts as intermediate states while its large density of states does not cause a heavy mass for the electrons that form the Cooper pair. In fact, pairing interaction originating from a narrow band lying below, but not far away from, the Fermi level has attracted much attention recently in the context of the "incipient band\" in iron-based superconductors especially after the observation of missing hole Fermi pockets in some of the iron-based compounds. The idea for narrow bands in Ref. may be regarded as a precursor of the incipient band in iron-based superconductors. Superconductivity involving flat bands has also been discussed recently in the context of the flat-band superconductivity, where the virtual pair scattering between the flat and dispersive bands is also at work. The incipient-band-induced superconductivity exploits finite-energy spin fluctuations, whose importance has recently been noted in other models as well. In Ref., a two-leg Hubbard ladder was considered as a candidate model in which the above situation can be realized. In fact, the study of unconventional superconductivity in ladder-type materials has a long history, starting from the seminal proposal by Dagotto and Rice. Experimentally, superconductivity with a \(T_c\) exceeding 10 K was then observed in (Sr,Ca)\(_{14}\)Cu\(_{24}\)O\(_{41}\). The subsequent proposal in Ref. was that a much higher \(T_c\) may take place in two-leg cuprate ladder compounds, provided that \(\sim 30\) % of electron doping is achieved. Namely, in a single-orbital Hubbard model on a two-leg ladder with only the nearest-neighbor hopping, there are two (bonding and antibonding) bands with an identical shape separated in energy by \(2t_\mathrm{rung}\), where \(t_\mathrm{rung}\) is the hopping in the rung direction. When the second-neighbor (diagonal) hoppings are introduced, one of the bands becomes narrower, while the other becomes wider. For the sign of the diagonal hopping corresponding to the real material, a large electron doping is required to put the Fermi level just above the top of the narrow band. A snag, however, is that doping carriers into the cuprate ladders is known to be difficult; especially electron doping with such a large amount may be rather unrealistic. In the present paper, we propose an entirely different, and realistic, way to realize the above-mentioned situation where narrow and wide bands coexist with the Fermi level lying close to the edge of the narrow one. We propose to start with the Ruddlesden-Popper-type layered perovskites with two layers in a unit cell. These bilayer compounds may first seem to have nothing to do with ladders. However, if we consider a case in which the \(t_{2g}\) orbitals are relevant (i.e., forming the bands crossing \(E_F\)), an electron in, say, the \(d_{xz}\) orbital selectively hops in the \(x\) direction as well as in the \(z\) direction normal to the bilayer. This means that the \(d_{xz}\) orbitals form a ladder with the \(x\) and \(z\) directions being the leg and rung directions, respectively. Similarly, \(d_{yz}\) orbitals form a ladder in the \(y\) and \(z\) directions, as schematically shown in Fig.[\[lattice\]](#lattice){reference-type="ref" reference="lattice"}. Hence ladder-like electronic structures are, in fact, *hidden* in the apparently bilayer Ruddlesden-Popper materials. To be precise, the \(d_{xz}\) and \(d_{yz}\) orbitals are weakly hybridized with each other, and also with the \(d_{xy}\) orbital, so that the quasi-one dimensional ladder character is slightly degraded, but they should still essentially be ladder bands.
# Formulation
## Intuitive idea
As mentioned in the Introduction, the band filling is an essential factor if we want to have the narrow band as incipient in the ladder-like electronic structure to realize a high-\(T_c\) superconductivity. An electron doping of \(\sim 30\) % is required for the cuprate ladder as mentioned, where the \(d_{x^2-y^2}\) (i.e., an \(e_g\)) orbital is relevant. By contrast, if the \(d_{xz}\) and \(d_{yz}\) (i.e., \(t_{2g}\)) orbitals are relevant, the sign of the hopping integral is reversed from the case of \(d_{x^2-y^2}\), so that the ideal situation should be \(\sim 30\) % *hole* doping, which corresponds to \(1/3\) filling, namely, two electrons on the three \(t_{2g}\) orbitals. If we consider non-copper materials of the form Sr\(_3\)*TM*\(_2\)O\(_7\) (*TM* indicates a transition metal), the valence of the transition metal should be +4. This implies that if we want to have a \(d^2\) electron configuration, we can take, instead of the hole doping, stoichiometries Sr\(_3\)Mo\(_2\)O\(_7\) (with a 4\(d\) transition metal Mo) and Sr\(_3\)Cr\(_2\)O\(_7\) (with a 3\(d\) transition metal Cr) for possible candidates as the hidden ladder materials.
## Band calculation and many-body analysis
We start with a first-principles band calculation for these materials with the WIEN2k package, assuming the ideal case of no long-range orders and adopting the experimental lattice structures in Refs.. In the first-principles calculation, the Perdew-Burke-Ernzerhof parametrization of the generalized gradient approximation was used, with 1000 \(k\)-meshes and \(RK_{\rm max}=7\). To proceed to many-body analysis, we first construct a three-dimensional tight-binding model with six orbitals (three \(t_{2g}\) orbitals in each layer) for Sr\(_3\)Mo\(_2\)O\(_7\) constructed from the maximally-localized Wannier orbitals using the Wannier90 code and the wien2wannier interface. To take account of the electron correlation effect beyond the LDA/GGA level for this tight-binding model, we introduce multi-orbital Hubbard-type interactions on each atomic site with the Hamiltonian, \[\begin{aligned}
H_{\mathrm{int}}=\sum_{i}\left(\sum_{\mu}U_{\mu}n_{i\mu\uparrow}n_{i\mu\downarrow}+\sum_{\mu>\nu}\sum_{\sigma\sigma^\prime}U'_{\mu\nu}n_{i\mu\sigma}n_{i\mu\sigma^\prime} \right. \nonumber \\ \left. -\sum_{\mu\ne\nu}J_{\mu\nu}\bm{S}_{i\mu}\cdot\bm{S}_{i\nu}+\sum_{\mu\neq\nu}J^\prime_{\mu\nu}c_{i\uparrow}^{\mu\dag}c_{i\downarrow}^{\mu\dag}c_{i\downarrow}^{\nu}c_{i\uparrow}^{\nu}\right), \nonumber \end{aligned}\] where \(i\) labels the site and \(\mu,\nu\) are the orbitals in a notation adopted in Ref.. To treat the many-body effect, we apply the multi-orbital fluctuation exchange (FLEX) approximation. We set the on-site interactions as \(U = 3.0\) eV or \(U=2.5\) eV, along with \(U^\prime (= U-2J\) for preserving the orbital SU(2) symmetry) and \(J = J^\prime = U/10\). These are typical values evaluated with the constrained random phase approximation or the constrained LDA method for the Hubbard-type model for the \(d\)-orbitals of transition metals. The temperature is fixed at \(T = 0.01\) eV, and we take \(N=32\times 32\times4\) \(k\)-point meshes and 2048 Matsubara frequencies. In the FLEX approximation, the effective electron-electron interaction for obtaining the self-energy is calculated by collecting bubble-and ladder-type diagrams consisting of the renormalized Green's function \(G(k)\), namely, by summing up powers of the irreducible susceptibility, \[\begin{aligned}
\chi^{\rm irr}_{l_1l_2, l_3l_4}(q) \equiv-\frac{T}{N}\sum_{k} G_{l_3l_1}(k)G_{l_2l_4}(k+q). \nonumber \end{aligned}\] Here \(k\) is a shorthand for the wave number and the Matsubara frequency, and \(l_i\) denotes orbitals. The renormalized Green's function is obtained by solving the Dyson equation. We then obtain the singlet pairing interaction \(\Gamma(q)\) mediated mainly by spin fluctuations, which is plugged into the linearized Eliashberg equation for superconductivity, \[\begin{aligned}
\lambda \Delta_{ll^{\prime}}(k) &=&-\frac{T}{N}\sum_{k^{\prime}m_i}\Gamma_{lm_1 m_4l^{\prime}}(k-k^{\prime})G_{m_1m_2}(k^{\prime})\nonumber\\ &&\times \Delta_{m_2m_3}(k^{\prime}) G_{m_4m_3}(-k^{\prime}). \nonumber \end{aligned}\] Since the eigenvalue \(\lambda\) of the equation reaches unity at \(T=T_c\), here we adopt \(\lambda\), obtained at a fixed temperature, to measure how close the system is to superconductivity.
# Results
We first display the band structures of for Sr\(_3\)Mo\(_2\)O\(_7\) and Sr\(_3\)Cr\(_2\)O\(_7\) obtained from the first-principles band calculation. The result in Fig.2 indeed shows that the \(d_{xz}\) and \(d_{yz}\) orbitals form quasi-one-dimensional narrow (antibonding) and wide (bonding) bands in both materials, as identified from the band dispersion along P-N and N-\(\Gamma\). For comparison, we have also performed a band calculation for the well-known two-leg ladder cuprate SrCu\(_2\)O\(_3\), with the lattice structure given in Ref.. In the result displayed in bottom panels of Fig.[\[band\]](#band){reference-type="ref" reference="band"} we have reversed the sign of the energy to facilitate comparison, since the roles of electrons and holes are exchanged between the \(e_g\) (cuprate) and the \(t_{2g}\) (present) systems. We can see that the band structures of the \(d_{xz/yz}\) orbitals of the bilayer Ruddlesden-Popper compounds are very similar to that of the ladder compound. However, an important difference lies in the position of the Fermi level: In the Ruddlesden-Popper compounds \(E_F\) is situated very close to the bottom edge of the narrow bands, namely, the narrow bands can be regarded as incipient. This is in sharp contrast to the case of SrCu\(_2\)O\(_3\), where \(E_F\) sits far away from the narrow band edge. These results confirm the intuitive expectation presented above. In Fig.[\[wannier\]](#wannier){reference-type="ref" reference="wannier"}, we superpose the band structure of the six-orbital tight-binding model for Sr\(_3\)Mo\(_2\)O\(_7\) constructed from the maximally-localized Wannier orbitals on that of the first-principles calculation. We can see that the fit is quite accurate for all the bands intersecting the Fermi level. Using the six-orbital model, we now study the possibility of superconductivity in Sr\(_3\)Mo\(_2\)O\(_7\). Figure [\[lambda\]](#lambda){reference-type="ref" reference="lambda"} displays the FLEX result for the eigenvalue \(\lambda\) of the Eliashberg equation against the band filling. Here the filling \(n\) is defined as the average number of electrons per unit cell, so that \(n=4\) corresponds to the stoichiometric \(d^{2}\) electron configuration. As seen from the result, \(\lambda\) is qualitatively unchanged when we change \(U=2.5-3.0\) eV within a realistic range, except that for \(U=2.5\) eV, (i) the maximum value of \(\lambda\) is somewhat enhanced from the case of \(U=3\) eV, and (ii) a decrease of \(\lambda\) in the left side of the dome becomes slightly steeper. In both cases, we can see that \(\lambda\) attains a maximum \(\simeq 0.8\) near the stoichiometric band filling of \(n=4\), which can be regarded to signal a high-\(T_c\) in this material, since the value of \(\lambda\) is similar to that of the cuprate superconductor HgBa\(_2\)CuO\(_4\) having \(T_c \simeq 90\) K as obtained with FLEX for systematic models that qualitatively explains the material dependence of \(T_c\) among the cuprates. This is a key result in the present work. To be precise, \(\lambda\) in Fig.[\[lambda\]](#lambda){reference-type="ref" reference="lambda"} is peaked at \(n\simeq 3.8\), but decreases as we move away from that point, which can be interpreted as follows: For a too small band filling, superconductivity is suppressed because the narrow band lies too far from the Fermi level. When the band filling is too large, on the other hand, the Fermi level lies deep inside the narrow band, so that the system is put close to antiferromagnetic order, suppressing \(\lambda\). Hence we conclude that the superconductivity should occur around the stoichiometric point (\(n=4\)) with an optimized band filling at a slightly hole-doped \(n\simeq 3.8\). If we turn from the Mo compound to the 3\(d\) compound, Sr\(_3\)Cr\(_2\)O\(_7\), we may also expect high-\(T_c\) superconductivity, since its band structure is similar to that of Sr\(_3\)Mo\(_2\)O\(_7\) as seen in Fig.2.
# Discussion
The materials, Sr\(_3\)Mo\(_2\)O\(_7\) and Sr\(_3\)Cr\(_2\)O\(_7\), have in fact been synthesized in the past. Sr\(_3\)Cr\(_2\)O\(_7\) turns out to be an antiferromagnetic insulator, while Sr\(_3\)Mo\(_2\)O\(_7\) a Pauli-paramagnetic metal, where superconductivity has not been observed so far in these materials. Let us then explore the reasons for that. Sr\(_3\)Cr\(_2\)O\(_7\) has two \(d\) electrons within the three \(t_{2g}\) orbitals, so that we have a possibility of orbital ordering which will put the system in an insulating state. | {'timestamp': '2017-11-20T02:06:29', 'yymm': '1706', 'arxiv_id': '1706.01711', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01711'} |
null | null |
# Introduction
In spite of the many constraints now known on the possible values of Seshadri constants (see for example ), the longstanding question of whether Seshadri constants on surfaces (defined below) can ever be irrational remains open. In the case of a surface \(X\) obtained as the blow up \(\pi: X \to \mathbb{P}^2\) of the complex projective plane \(\mathbb{P}^2\) at very general points \(p_1,\cdots,p_s\in\mathbb{P}^2\), recent work of Dumnicki, Küronya, Maclean and Szemberg,, shows for \(s\geq 9\) that the SHGH Conjecture implies that certain ample divisors \(L\) on \(X\) have irrational Seshadri constants \(\varepsilon(X,L,x)\) when \(x\) is a very general point of \(X\). In this note we show that less is needed to obtain this conclusion, namely one merely has to assume that prime divisors \(C\) on the blow up \(Y\) of \(X\) at \(x\) with \(C^2<0\) satisfy \(C^2=C\cdot K_Y=-1\). This assumption is itself a consequence of the SHGH Conjecture but it is not known to be equivalent to the full SHGH Conjecture, and it leads to a conceptually simpler proof than the one obtained in. It also leads us to raise the question if an even weaker assumption, viz., Nagata's Conjecture, suffices to draw the same conclusion.
# Main result
We recall some standard facts. Given a point \(x\) on a smooth projective surface \(S\) and an ample divisor \(L\), the Seshadri constant \(\varepsilon(S,L,x)\) is defined to be \[\varepsilon(S,L,x)=\inf_C \frac{L\cdot C}{{\rm mult}_x(C)},\] where the infimum is taken over all curves \(C\) containing \(x\). Alternatively, let \(\pi: Y \to S\) be the blow up of \(S\) at \(x\) with exceptional curve \(E\). Then \(\varepsilon=\varepsilon(S,L,x)\) is the supremum of all real \(t\) such that \(\pi^*(L)-tE\) is nef and hence \((\pi^*(L)-\varepsilon E)^2\geq0\). It follows that \(\varepsilon(S,L,x)\leq \sqrt{L^2}\). If \(\varepsilon(S,L,x) < \sqrt{L^2}\), one says that \(\varepsilon(S,L,x)\) is submaximal, in which case it is well known that there exists a reduced and irreducible curve \(C\) on \(S\) passing through \(x\) such that \(\varepsilon = \varepsilon(S,L,x) = \frac{L\cdot C}{{\rm mult}_x(C)}\) (i.e., such that \((\pi^*(L)-\varepsilon E)\cdot \tilde{C}=0\), where \(\tilde{C}\subset Y\) is the strict transform of \(C\)). Such a curve \(C\) is called a *Seshadri curve* for \(L\) at \(x\). Since \(\varepsilon=\varepsilon(S,L,x) < \sqrt{L^2}\) implies \((\pi^*(L)-\varepsilon E)^2>0\), it follows by the Hodge index theorem that \(\tilde{C}^2<0\). We will also need to refer to multi-point Seshadri constants. Given distinct points \(p_1,\cdots,p_s\) on \(S\) and an ample divisor \(L\), the multi-point Seshadri constant \(\varepsilon(S,L,p_1,\cdots,p_s)\) is defined to be \[\varepsilon(S,L,p_1,\cdots,p_s)=\inf_C \frac{L\cdot C}{\sum_i{\rm mult}_{p_i}(C)},\] where the infimum is taken over all curves \(C\) containing at least one of the points \(p_i\). Alternatively, let \(\pi: Y \to S\) be the blow up of \(S\) at \(p_1,\cdots,p_s\) with \(E_i\) being the exceptional curve for \(p_i\). Then \(\varepsilon=\varepsilon(S,L,p_1,\cdots,p_s)\) is the supremum of all real \(t\) such that \(\pi^*(L)-t(E_1+\cdots+E_s)\) is nef and hence \((\pi^*(L)-\varepsilon (E_1+\cdots+E_s))^2\geq0\). If \(0<t<\varepsilon\), it is easy to see that \(\pi^*(L)-t(E_1+\cdots+E_s)\) is ample (since \(F=(t/\varepsilon)(\pi^*(L)-\varepsilon (E_1+\cdots+E_s))\) is nef and meets any nonnegative linear combination of the \(E_i\) positively, and \(\pi^*(L)-t(E_1+\cdots+E_s)=F+(1-(t/\varepsilon))\pi^*(L)\)). When the points \(p_i\) are very general, we will write \(\varepsilon=\varepsilon(S,L,s)\) for \(\varepsilon=\varepsilon(S,L,p_1,\cdots,p_s)\). Our focus will be on surfaces \(\pi:Y\to X\to\mathbb{P}^2\) where \(X\to\mathbb{P}^2\) is obtained by blowing up very general points \(p_1,\cdots,p_s\) on \(\mathbb{P}^2\) and \(Y\to X\) is the blow up of a very general point \(x\in X\) with exceptional divisor \(E\). So let \(H = \pi^*(\mathcal{O}_{\mathbb{P}^2}(1))\) and let \(E_i\) be the exceptional curve for each point \(p_i\). Every divisor on \(Y\) is linearly equivalent to a unique integer linear combination \(F=dH-mE-m_1E_1-\cdots-m_sE_s\). (Since \(Y\to X\) is an isomorphism away from \(x\), we can regard the divisors \(H\) and \(E_i\) as also being on \(X\). With this abuse of notation, every divisor on \(X\) is linearly equivalent to a unique integer linear combination \(dH-m_1E_1-\cdots-m_sE_s\).) Such a divisor \(F\) is in *standard* form if \(m\geq m_1\geq\cdots\geq m_s\geq 0\) and \(d\geq m+m_1+m_2\). An *exceptional* curve on \(X\) (or \(Y\)) is a reduced and irreducible rational curve \(C\) with \(C^2 =-1\) (and hence \(-K_X\cdot C=1\), or \(-K_Y\cdot C=1\) respectively). If \(F\) is in standard form, then \(F \cdot C \geq 0\) for all exceptional curves \(C\) on \(Y\). (To see this, let \(F=dH-mE-m_1E_1-\cdots-m_sE_s\) be divisor on \(Y\). If \(F\) is in standard form and if \(C\) is one of the exceptional curves \(E,E_1,\cdots,E_s\) then clearly \(F\cdot C \ge 0\). So suppose that \(C\) is different from \(E,E_1, \cdots, E_s\). Note that \(F\) is in standard form if and only if \(F\) is a nonnegative linear integer combination of \(H_0=H\), \(H_1=H-E\), \(H_2=2H-E-E_1\), \(H_3=3H-E-E_1-E_2\), \(\cdots\), \(H_{s+1}=3H-E-E_1-\cdots-E_s=-K_Y\). But \(H_i\) is nef for \(i=0,1,2\) and \(H_i\cdot C\geq-K_Y\cdot C=1\) for \(i\geq 3\).) The above definition of standard divisors also extends to divisors with coefficients in \(\mathbb Q\) or \(\mathbb R\). If \(F\) is a standard \(\mathbb Q\)-divisor, then for a suitable positive integer \(n\), the \(\mathbb Z\)-divisor \(nF\) is standard. It follows that \(F\cdot C\ge 0\) for all exceptional curves \(C\) on \(Y\). If \(F\) is a standard \(\mathbb R\)-divisor, then \(F\) is the limit of a sequence of standard \(\mathbb Q\)-divisors. So again \(F\cdot C\ge 0\) for all exceptional curves \(C\) on \(Y\).
Combining Remark [\[FewPtsRemark\]](#FewPtsRemark){reference-type="ref" reference="FewPtsRemark"}, Proposition [\[main1\]](#main1){reference-type="ref" reference="main1"} and Proposition [\[remaining-cases\]](#remaining-cases){reference-type="ref" reference="remaining-cases"}, we obtain our main theorem.
Our results depend only on assuming all negative curves are exceptional. A somewhat weaker result was conjectured by Nagata, namely for a blow up \(S\) of \(\mathbb{P}^2\) at \(s\geq 10\) very general points, if \(dH-(m_1E_1+\cdots+m_sE_s)\) is linearly equivalent to an effective divisor, then \(d\sqrt{s}\geq \sum_im_i\). This is equivalent to conjecturing that \(F_0=\sqrt{s}H-E_1-\cdots-E_s\) is nef. Note for arbitrarily small \(\delta>0\) that \(F_\delta=(\delta+\sqrt{s})H-E_1-\cdots-E_s\) is rational and semi-effective (meaning that a positive integer multiple is linearly equivalent to an effective divisor, which follows since \(F^2>0\)). Thus if \(F_0\) is not nef, then there is a prime divisor \(C\) with \(C^2<0\) and \(C\cdot F_0<0\). From this we see that the SHGH Conjecture implies Nagata's Conjecture. In fact, if \(C\) being a prime divisor with \(C^2<0\) implies \(C^2=C\cdot K_S=-1\), then already Nagata's Conjecture is true. This is because if \(C^2<0\) for a prime divisor \(C\), then \(C\cdot (\sqrt{s}H-E_1-\cdots-E_s)\geq C\cdot (3H-E_1-\cdots-E_s)\geq 1\). Thus Nagata's Conjecture is weaker than the assumption we used. Note further that the Nagata Conjecture exhibits irrational multi-point Seshadri constants on \(\mathbb{P}^2\), since it is equivalent to the statement that \(\varepsilon(\mathbb{P}^2,\mathcal{O}_{\mathbb{P}^2}(1),s) = 1/\sqrt{s}\) for every \(s \ge 10\). These remarks raise the following question.
**Acknowledgement:** We thank Tomasz Szemberg for reading this paper and giving useful suggestions. We also thank the referee for giving an alternate argument in the proof of Proposition [\[main1\]](#main1){reference-type="ref" reference="main1"} and numerous other suggestions which improved the exposition. | {'timestamp': '2017-12-18T02:08:38', 'yymm': '1706', 'arxiv_id': '1706.01648', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01648'} |
null | null |
# Introduction
In the recent years we have witnessed a growing pattern of experimental anomalies in flavour physics, which can be schematically summarized as follows:
1. Semi-leptonic \(B\)-decays in flavour changing neutral currents (FCNC) \(b \to s \ell \bar \ell\), suggesting a deficit of muons compared to electrons. The main observables are: \(i)\) the angular distributions of \(B \to K^* \mu \bar \mu\), \(ii)\) the rate of semi-leptonic decays such as \(B \to K^* \mu \bar \mu\) and \(B_s \to \phi \mu \bar \mu\) and \(iii)\) the lepton flavour universality (LFU) violating observables \(R_K\) and \(R_{K^*}\), which are defined by the ratios \(\mathcal{B}(B \to K^{(*)} \mu \bar \mu) / \mathcal{B}(B \to K^{(*)} e \bar e)\). We remark that the Standard Model (SM) theoretical uncertainty for \(R_{K^{(*)}}\) is very small (few percent due to QED radiative corrections ). Updated fits based on effective field theory (EFT) analyses, including the most recent \(R_K{^*}\) measurement, can be found in.
2. Semi-leptonic \(B\)-decays in flavour changing charged currents (FCCC) \(b \to c \ell \bar \nu_{\ell}\), suggesting an excess of taus compared to muons and electrons. The main observables are the LFU violating ratios \(R_{D^{(*)}}\), defined as \(\mathcal{B}(B \to D^{(*)} \tau \bar \nu) / \mathcal{B}(B \to D^{(*)} \ell \bar \nu)\), with \(\ell = e, \mu\). In this case it is non-trivial that three different experiments agree well among each other. A recent EFT fit of \(R_{D^{(*)}}\) can be found for instance in. Very recently, there has also been a new measurement of \(R_{D^*}\) by the LHCb collaboration, which is remarkably compatible with the previous ones. However, this single measurement does not affect much the global fit.
In both cases the statistical significance reaches the \(4\sigma\) level, while from a theoretical standpoint it is very intriguing that both the set of anomalies can be interpreted within a coherent framework. In particular, one can envisage two orthogonal structures: \(i)\) a vertical (gauge) one: global fits seem to prefer effective operators featuring only \(SU(2)_L\) doublets and \(ii)\) a horizontal (flavour) one: data hints to violation of LFU with a similar hierarchical pattern as in the SM, with new physics contributions negligible in electrons (basically no effects), sizeable in muons (observable only in \(b \to s \mu \bar \mu\)) and large in taus (effects in \(b \to c \tau \bar \nu_\tau\) and potentially in \(b \to s \nu_\tau \bar \nu_\tau\)). These facts motivated the community to speculate about the simultaneous explanation of these two sets of anomalies and their connection with the origin of the SM flavour. It is then maybe not too early to dream about new physics and ask "what is the scale of new physics behind the \(B\)-flavour anomalies?" Here, we address this question by using an old tool of theoretical physics, namely perturbative unitarity. Perhaps most famously, constraints imposed by perturbative unitarity in \(WW\) scattering have been used in the past to infer an upper bound on the Higgs boson mass or, alternatively, on the scale where the SM description of weak interactions needed to be completed in the ultraviolet (UV) in terms of some new strongly coupled dynamics. What we are going to consider here instead resembles in some sense the Fermi theory of weak interactions. In fact, already in the 1930's, from the low-energy measurement of \(G_F\) one could have inferred what was the scale of "new physics" behind the Fermi theory. By looking at \(2 \to 2\) scatterings via four-fermion effective operators in the Fermi theory one finds that unitarity is violated[^1] at energies of the order of \(\Lambda_U = 900\) GeV (see e.g. ). As is well known, the dynamical degrees of freedom of the SM turned out to be weakly coupled and hence much lighter than the unitarity bound, e.g. \(M_W \ll \Lambda_U\). In this paper, we do something similar to the unitarity analysis in the Fermi theory by considering the four-fermion operators of the \(d=6\) SM-invariant EFT (SMEFT) semi-leptonic basis, under the hypothesis of a short-distance new physics explanation of the experimental anomalies in semi-leptonic \(B\)-meson decays.[^2] Note that the analysis can be independently carried out for the \(b \to s \ell \bar \ell\) and \(b \to c \ell \bar \nu\) anomalies, and we do not necessarily rely on a common explanation of the two. In short, once the Wilson coefficient of an effective operator is fixed by the fit to the anomaly, we can use it in order to extract the scale of unitarity violation without the need of passing through the ambiguous separation of mass vs. coupling. The common lore is that on-shell new degrees of freedom should appear below the scale of unitarity violation (see however for exceptions), with interesting consequences for direct searches at LHC and future colliders. A simple message that we would like to emphasize is that scattering amplitudes employing SM invariant effective operators lead to scales of unitarity violation \(\Lambda_U\) which are typically smaller than the naive dimensional analysis (NDA) estimate of the strong coupling regime \(g_\star = 4 \pi\), i.e. \(M_\star = 4 \pi \Lambda_{\mathcal{O}}\), where \(\Lambda_{\mathcal{O}}\) denotes the scale of the SM invariant effective operator required to fit the anomaly normalized to unit Wilson coefficient. This enhancement is also in part due to the correlation of the scattering amplitudes in the gauge group space, which is important to take into account when thinking about the energy reach of LHC or future colliders. A related point is the flavour structure of the effective operators. If these are oriented along the third generation fermion families (as motivated in various flavour models), one typically predicts a strong enhancement of the unitarity bound which can even reach few TeV (cf. ). Similarly to the EFT analysis, unitarity arguments can also be used in order to set perturbativity constraints on the parameter space of simplified models explaining the flavour anomalies. Note, however, that in the latter case the scattering amplitudes do not grow with the energy but reach asymptotic values proportional to the Yukawa-like couplings of the new mediators. It is possible then to translate the unitarity bounds on the coupling into an upper bound on the mass of the new states (once the ratio coupling/mass is fixed in terms of the fit to the relevant anomaly). Remarkably, in some cases the upper bound on the new mediators' mass is so strong that the perturbative interpretation of the anomaly within a given simplified model can be ruled out, or soon tested at the LHC. The layout of the paper is the following: in we start by introducing and comparing different kind of scales in the EFT. After discussing in motivated flavour structures for the effective operators, we briefly introduce the partial-wave-unitarity tool in . We continue in where we derive the unitarity bounds respectively in the EFT and for simplified models addressing the \(B\)-flavour anomalies. We finally conclude in , where we also provide a summary of our results. In , as a paradigmatic example, we report the details of the unitarity bound calculation in the presence of an \(SU(2)_L\) triplet effective operator.
# A tale of scales {#tale}
In what follows we will focus for simplicity on purely left-handed operators, since they provide the best fit for both the anomalies in \(b \to s \mu \bar \mu\) and \(b \to c \tau \bar \nu\) transitions. The analysis can be easily generalized to scenarios including more operators by using the results given in . In order to start the discussion it is useful to identify and compare four (conceptually different) scales in the EFT:[^3]
1. \(\Lambda_{A}\): the "Fermi constant" of the process.\
This is the scale required to explain the anomaly, to be evaluated at the typical energy of the process which is fixed by the \(B\)-meson mass. The low-energy EFT description is based on \(SU(3)_C \times U(1)_{EM}\) invariant operators. The index \(A\) on \(\Lambda_A\) runs over the anomalies, schematically \(A=\{ R_{D^{(*)}},R_{K^{(*)}} \}\), and the EFT Lagrangian featuring purely left-handed operators reads \[\label{LeffRDRK} \mathcal{L}_{\rm eff} \supset-\frac{1}{\Lambda^2_{R_{D^{(*)}}}} 2 \, \overline{c}_L \gamma^{\mu} b_L \overline{\tau}_L \gamma_{\mu} \nu_L + \frac{1}{\Lambda^2_{R_{K^{(*)}}}}\overline{s}_L \gamma^{\mu} b_L \overline{\mu}_L \gamma_{\mu} \mu_L + \textrm{h.c.} \,,\] where we assumed alignment with the phases of the CKM elements that appear in the corresponding SM operators. Note that the fit of the \(R_{D^{(*)}}\) and \(R_{K^{(*)}}\) anomalies requires an opposite sign interference with the SM contribution. We also included an extra factor of \(2\) in the definition of the charged-current operator, so that the latter has the same normalization of the neutral-current operator when considering a SMEFT. The best fit values of the \(R_{D^{(*)}}\) and \(R_{K^{(*)}}\) anomalies yield respectively \[\begin{aligned}
\Lambda_{R_{D^{(*)}}} &= 3.4 \pm 0.4 \textrm{ TeV} \,, \\ \Lambda_{R_{K^{(*)}}} &= 31 \pm 4 \textrm{ TeV} \,, \end{aligned}\] where the errors are at 1\(\sigma\). In the following we will only consider central values.
2. \(\Lambda_{\mathcal{O}}\): the scale of the SMEFT operator.\
This is the scale required to explain the anomaly using an EFT at higher energies[^4] (\(SU(3)_C \times SU(2)_L \times U(1)_Y\) invariant), with Wilson coefficient normalized to one. The index \(\mathcal{O}\) on \(\Lambda_{\mathcal{O}}\) is associated with an operator of the SMEFT semi-leptonic basis and runs over all the possible Lorentz and flavour structures. For definiteness we will consider here an \(SU(2)_L\) triplet operator (\(Q\) and \(L\) denoting \(SU(2)_L\) doublets) \[\label{opQijL33} \mathcal{L}_{\rm SMEFT} \supset \frac{1}{\Lambda^2_{Q_{ij} L_{kl}}} \left( \bar Q_i \gamma^\mu \sigma^A Q_j \right) \left( \bar L_k \gamma_\mu \sigma^A L_l \right) + \textrm{h.c.} \,,\] and two reference flavour structures such that the operator is aligned in the direction of the flavour eigenstates responsible for the anomalies, namely \(\mathcal{O} = Q_{23} L_{33}\) (for \(b \to c \tau \bar \nu\) transitions) and \(\mathcal{O} = Q_{23} L_{22}\) (for \(b \to s \mu \bar \mu\) transitions). The matching with yields \[\begin{aligned}
\label{LamQ23L33} \abs{\Lambda_{Q_{23} L_{33}}} &= \Lambda_{R_{D^{(*)}}} = 3.4 \textrm{ TeV} \,, \\ \label{LamQ23L33} \abs{\Lambda_{Q_{23} L_{22}}} &= \Lambda_{R_{K^{(*)}}} = 31 \textrm{ TeV} \,. \end{aligned}\] As we will discuss in detail in , depending on the specific flavour ansatz, the scale \(\Lambda_{\mathcal{O}}\) can be effectively reduced with respect to the "Fermi constant" of the process. For example, the transition \(b \to c \tau \bar \nu\) could originate from the operator \(\mathcal{O} = Q_{33} L_{33}\), where the \(3 \to 2\) transition in the up sector is due to a CKM mixing (in the basis where \(Q_i = (V^\dag_{ij} u^j_L, d^i_L)^T\)), which yields \(\Lambda_{Q_{33} L_{33}} / \sqrt{\abs{V_{cb}}} = \Lambda_{R_{D^{(*)}}}\).
3. \(\Lambda_U\): the scale of unitarity violation.\
This is the scale where the EFT description breaks down. The important point is that it can be expressed in terms of the scale \(\Lambda_{\mathcal{O}}\), without passing through the ambiguous separation between coupling and mass. Using the results of (which are based on a non-trivial calculation of the scattering amplitude, including gauge group multiplicity factors) we obtain \[\Lambda_U = \sqrt{\frac{4 \pi}{\sqrt{3}}} \abs{\Lambda_{Q_{ij} L_{kl}}} \,,\] which yields \[\begin{aligned}
\label{LambdaU1} \Lambda_U &= 9.2 \textrm{ TeV} \qquad (\mathcal{O} = Q_{23} L_{33} \ \ \text{case}) \,, \\ \label{LambdaU2} \Lambda_U &= 84 \textrm{ TeV} \qquad \ (\mathcal{O} = Q_{23} L_{22} \ \ \text{case}) \,. \end{aligned}\] These are the most conservative bounds on the scale of new physics responsible for the anomalies in \(b\to c \tau \bar \nu\) and \(b \to s \mu \bar \mu\).
4. \(M_\star\): the NDA mass scale in the strongly coupled regime.\
This is the mass scale associated with the effective operator when saturating perturbativity. After reintroducing \(\hbar\) in the NDA (see e.g. ), one can formally distinguish among scales (\(\Lambda\)), masses (\(M\)) and couplings (\(g\)), and set \(M = g \Lambda\). By naively saturating perturbativity at \(\abs{g_\star}= 4 \pi\), we can write \[\frac{1}{\abs{\Lambda_{\mathcal{O}}}} = \frac{4 \pi}{M_{\star}} \,,\] which leads to \[\begin{aligned}
\label{Mstar1} M_\star &= 43 \textrm{ TeV} \qquad \ (\mathcal{O} = Q_{23} L_{33} \ \ \text{case}) \,, \\ \label{Mstar2} M_\star &= 390 \textrm{ TeV} \qquad (\mathcal{O} = Q_{23} L_{22} \ \ \text{case}) \,. \end{aligned}\] Note that \(M_\star\) is a factor \(5\) larger than the scale of unitarity violation in .
Our results for the EFT analysis are summarized in (cf. also for more details on the flavour structure of the effective operators and the unitarity bounds), where we report the values of the four different scales discussed above for the anomalies in either \(b \to c \tau \bar \nu\) or \(b \to s \mu \bar \mu\) transitions, and depending on the flavour structure of the operator \(\mathcal{O}\). The two main points to be observed are the following: \(i)\) \(\Lambda_U\) is sizably smaller than \(M_\star\) and \(ii)\) depending on the flavour structure of the operator \(\mathcal{O}\), the scale \(\Lambda_U\) approaches the energy reach of LHC. This motivates an interesting interplay of the flavour anomalies with direct searches, which is further explored in by employing simplified models.
:::
[^1]: We will sometimes improperly use the term "unitarity violation", by which we mean perturbative unitarity (cf. the discussion in ).
[^2]: Unitarity bounds for the EFT intrerpretaton of \(b \to s \ell \bar \ell\) anomalies were briefly mentioned in Ref. .
[^3]: Some of the results presented here will be derived in the following sections.
[^4]: QCD running effects on the Wilson coefficients are of the order of \(1 +\frac{\alpha_s}{4 \pi} \times \log \frac{\Lambda_{\mathcal{O}}}{m_b}\). For \(\Lambda_{\mathcal{O}}= 1\) TeV, this corresponds to an \(\mathcal{O}(5\%)\) correction that will be neglected in the following. | {'timestamp': '2017-10-10T02:08:16', 'yymm': '1706', 'arxiv_id': '1706.01868', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01868'} |
null | null |
# Introduction {#sec:Introduction}
In an *interaction system*, economic (or general physical) agents interact pairwise, but do not necessarily cooperate towards a common goal. Yet, this model arises as a natural generalization of the model of cooperative TU-games, for which already Owen introduced the co-value as an assessment of the pairwise interaction of two cooperating players[^1]. In fact, it turns out that the context of interaction systems provides an appropriate framework for the analysis of cooperative games (*cf.* Faigle and Grabisch ). It is therefore of interest to investigate interaction systems in their own right. A second motivation comes from strong arguments (put forth most prominently by Penrose ) that the human mind with its decisions and actions constitutes a physical quantum system and should therefore be modelled according to the laws of quantum mechanics. This idea has furthered a seemingly new and rapidly growing branch of game theory where in so-called *quantum games* the players are assumed to execute actions on quantum systems according to the pertinent laws of physics[^2]. The actions of the quantum players typically transform quantum bit vectors according to the same mechanisms attributed to so-called *quantum computers*. Many classical games of economic or behavioral interest have been studied in this setting. The Prisoners' Dilemma, for example, has been found to offer a Pareto optimal Nash equilibrium in the quantum model where no such equilibrium exists in the classical variant (Eisert *et al.* ). The quantum games discussed in the literature are generally non-cooperative. Although cooperative game theory has been applied to quantum theory[^3], a cooperative quantum game theory has not really developped[^4]. In general, it is felt that quantum game theory has a highly physical side to it[^5] and that there is a gap between physics and rational game theoretic behavior[^6]. Therefore, it may be surprising that the mathematical analysis of interaction systems exhibits *no* conceptual gap to exist between the "classical" and the "quantum" world in game theory. In particular, no special knowledge of quantum mechanics (or physics in general) is needed to understand game theoretic behavior mathematically. Both aspects are sides of the same coin. The quantum aspects arise from the choice of terminology in the mathematical structural analysis rather than from innately physical environments. Yet, much structural insight is gained from the study of classical interaction and decision systems in terms of quantum theoretical language. The key in the analysis is the symmetry decomposition of the matrix of interaction coefficients \(\alpha_{xy}\) associated with a state \(\alpha\) of an interaction system. The decomposition allows us to represent states by self-adjoint matrices. Spectral decomposition then shows that interaction states have real eigenvalues. Moreover, each state is represented by a linear combination of pure density matrices. The connection to physical quantum models is now immediate: The axioms of quantum theory stipulate that quantum states can be described as convex combinations of pure density matrices. It follows that interaction systems subsume (finite-dimensional) quantum systems as special cases. We develop the mathematical details in Section [2](#sec:interaction-systems){reference-type="ref" reference="sec:interaction-systems"}. This section treats also measurements on interaction systems as linear functionals on the space of interaction matrices and provides a probabilistic interpretation that is commensurate with the probabilistic interpretation of measurements in standard physical quantum theory. Section [3](#sec.decision-analysis){reference-type="ref" reference="sec.decision-analysis"} approaches a binary decision system with a single decision maker as a system with two "interacting" alternatives. Decision systems with \(n\) decision makers arise as tensor products of \(n\) single decision systems. Again, the relation with quantum models is immediate: The states of an \(n\)-decision system are described by complex \(n\)-dimensional vectors of unit length in exactly the same way as such vectors are assumed to describe the states of an \(n\)-dimensional quantum system in the Schrödinger picture of quantum theory. The latter yields a very intuitive interpretation of quantum games: The players select their strategies as to move some decision makers into accepting an alternative that offers financial rewards to the individual players, where the financial reward is given as a linear functional on the space to decision states (Section [3.4](#sec:decision-games){reference-type="ref" reference="sec:decision-games"}). The choice of a set-theoretic representation of the decision states shows how classical cooperative TU games may be viewed as particular decision states of a finite set of decision makers. The associated probabilistic interpretation exhibits in particular Aubin's fuzzy games as decision instances where the decision makers (or "players") take their decisions probabilistically but independently from one another. The model of quantum computing[^7] views computation as the application of certain linear operators on quantum states. We point out in Section [3.3](#sec:linear-transformations){reference-type="ref" reference="sec:linear-transformations"} how well-known linear transformations (*e.g.*, Möbius or Walsh transform) in cooperative games theory arise as tensor products of simple transformations on single decision systems. Moreover, we show that the classical Fourier transform is well-defined in the context of interaction systems, while it does not seem to exist when attention is restricted to the classical space of TU games. We finally present a linear model for the evolution of interaction systems and discuss its "Markovian" aspects. In addition to the application examples in the course of the treatment, the appendix contains a worked out example of an interaction system with two agents.
# Interaction systems {#sec:interaction-systems}
An *interaction system* is a pair \(\mathfrak X =(X,\cal A)\), where \(X\) is a set of *agents* and \(\cal A\) a set of *states* such that for any state \(\alpha\in \cA\) and agents \(x,y\in X\), there is a well-defined *interaction coefficient* \(\alpha_{xy}\in \R\). The corresponding matrix \(A=[\alpha_{xy}]\in \R^{X\times X}\) of interaction coefficients reflects the *interaction instance* relative to the state \(\alpha\). Interpreting the diagonal elements \(A_{xx}\) (interactions of the agents \(x\) with themselves) as the activity levels of the individual agents in state \(\alpha\), we will refer to \(A\) also as an *activity matrix*. \(A^T\) denotes the transpose of \(A\). We do not necessarily assume interaction symmetry \(A_{xy}=A_{yx}\) (*i.e.*, \(A^T \neq A\) may hold). Moreover, we allow "interaction" and "agent" to be taken in the most general sense. So agents can be physical or more abstract entities. However, we assume finiteness
- \(|X|<\infty\)
throughout this investigation. We give some examples of interaction systems.
Returning to a general interaction system \(\mathfrak X =(X,\mathcal A)\), recall that the space \(\R^{X\times X}\) of all possible interaction matrices is a \(|X|^2\)-dimensional euclidian vector space. Moreover, \(\R^{X\times X}\) is a real Hilbert space relative to the inner product \[(A|B\) = \sum_{x,y\in X} A_{xy}B_{xy} = \sum_{x\in X} (A^TB)_{xx} = \tr(A^TB),\] where \(\tr(C)\) denotes the trace of a matrix \(C\). The associated norm is the so-called *Frobenius norm* \[\|A\| = \sqrt{\(A|A\)} =\sqrt{\sum_{x,y\in X} |A_{xy}|^2}.\] We define \(\|\alpha\| = \|[\alpha_{xy}]\|\) as the *norm* of the state \(\alpha\in\cA\).
## Symmetry decomposition and hermitian representation {#sec:symmetry-decomposition}
Denote by \(\S_X^+\) the subspace of symmetric and by \(\S_X^-\) the subspace of skew-symmetric matrices in \(\R^{X\times X}\). For any \(C\in \S_X^+\) and \(B\in \S_X^-\), one has \[\(C|B\) = \(C^T|B^T\) =-\(C|B\) \quad\mbox{and hence}\quad \(C|B\) = 0.\] Setting \(A^+ = \frac12(A+A^T)\in\S_X^+\) and \(A^-= \frac12(A-A^T)\in\S_X^-\), one finds that \(A\in \R^{X\times X}\) is the superposition \[\label{eq.symmetry-decomposition} A = A^+ +A^-\] of a symmetric and a skew-symmetric matrix. It follows that \(\S_X^+\) and \(\S_X^-\) are orthogonal complements in \(\R^{X\times X}\) and that the symmetry decomposition ([\[eq.symmetry-decomposition\]](#eq.symmetry-decomposition){reference-type="ref" reference="eq.symmetry-decomposition"}) of \(A\) is uniquely determined and obeys the pythagorean law \[\|A\|^2 = \|A^+\|^2 + \|A^-\|^2.\] A convenient way of keeping track of the symmetry properties of \(A\) is its *hermitian representation* with the complex coefficients \(\hat{A}_{xy} = A_{xy}^+ +\im A^-_{xy}\) \[\label{eq.hermitian-representation} \hat{A} = A^+ +\im A^-\in \C^{X\times X},\] where \(\C=\R+\im \R\) denotes the field of complex numbers with the imaginary unit \(\im = \sqrt{-1}\). So \(\hat{A} = A\) holds if and only if \(A\) is symmetric. Moreover, the *hermitian space* \[\H_X = \{A^++\im A^-\mid A\in \R^{X\times X}\} = \S_X^+ +\im \S_X^- \subseteq \C^{X\times X}\] is isomorphic with \(\R^{X\times X}\) as a real Hilbert space. Recalling the *conjugate* of the complex matrix \(C = D+\im F\) with \(D,F\in \R^{X\times X}\), as the matrix \(\ov{C} = D-\im F\) and the *adjoint* as the matrix \(C^* = \ov{C}^T\), the inner product becomes \[\(A|B\) = \tr(A^TB) = \tr(\hat{A}^*\hat{B}) =\(\hat{A}|\hat{B}\) \quad(A,B\in \R^{X\times X}).\] A readily verified but important observation identifies the hermitian matrices as precisely the self-adjoint matrices: \[C \in \H_X \quad \Longleftrightarrow\quad C^* =C \quad(C\in \C^{X\times X}).\] The fact that \(\R^{X\times X}\) and \(\H_X\) are isomorphic Hilbert spaces allows us to view the interaction matrices \(A\) and their hermitian representations \(\hat{A}\) as equally valid representatives of interaction states. We denote by \[\hat{\alpha} = \hat{A} = A^+ +\im A^-\] the hermitian representation of the state \(\alpha\) with interaction matrix \(A=[\alpha_{xy}]\).
### Binary interaction {#sec:binary-interaction}
We illustrate the preceding concepts with the interaction of just two agents \(x,y\), *i.e.*, \(X=\{x,y\}\). A basis for the symmetric space \(\S_X^+\) is given by \[I=\begin{pmatrix} 1&0\\0&1\end{pmatrix}, \pi_1 =\begin{pmatrix} 1&0\\0&-1\end{pmatrix}, \pi_2 =\begin{pmatrix} 0&1\\1&0\end{pmatrix}.\] The skew-symmetric space \(\S_X^-\) is \(1\)-dimensional and generated by \[\pi_3 =\begin{pmatrix} 0&-1\\1&0\end{pmatrix}.\] Thinking of the interaction of an agent with itself as its activity level, one can interpret these matrices as follows:
1. \(I\): no proper interaction, the two agents have the same unit activity level.
2. \(\pi_1\): no proper interaction, opposite unit activity levels.
3. \(\pi_2\): no proper activity, symmetric interaction: there is a unit "interaction flow" from \(x\) to \(y\) and a unit flow from \(y\) to \(x\).
4. \(\pi_3\): no proper activity, skew-symmetric interaction: there is just a unit flow from \(x\) to \(y\) or, equivalently, a \((-1)\)-flow from \(y\) to \(x\).
The corresponding hermitian representations are \(\hat{I} = I\), \(\hat{\pi}_1 = \pi_1\), \(\hat{\pi}_2 = \pi_2\) and \[\hat{\pi}_3 = \im\pi_3 = \begin{pmatrix} 0&-\im\\ \im&0\end{pmatrix}.\]
[\[r.complex-numbers\]]{#r.complex-numbers label="r.complex-numbers"}
## Spectral theory {#sec:spectral-theory}
It is a well-known fact in linear algebra that a complex matrix \(C\in \C^{X\times X}\) is self-adjoint if and only if there is a diagonal matrix \(\Lambda\) with diagonal elements \(\lambda_x\in \R\) and a *unitary* matrix \(U\in \C^{X\times }\) (*i.e.*, \(U^* = U^{-1}\)) that *diagonalizes* \(C\) in the sense \[\label{eq.diagonalization} \Lambda = U^*CU.\] The real numbers \(\lambda_x\) are the eigenvalues and form the *spectrum* of \(C\in \H_X\). The corresponding column vectors \(U_x\) of \(U\) are eigenvectors of \(C\) and yield a basis for the complex vector space \(\C^X\). If \(\alpha\) is a state of \(\mathfrak X=(X,\mathcal A)\), we refer to the eigenvalues of its self-adjoint representation \(\hat{\alpha}\) simply as the *eigenvalues of \(\alpha\)*. A state thus has always real eigenvalues. If the interaction matrix \([\alpha_{xy}]\) is symmetric, then \(\hat{\alpha} = [\alpha_{xy}]\) holds and the eigenvalues of \([\alpha_{xy}]\) and \(\hat{\alpha}\) coincide. In general, however, an interaction matrix \([\alpha_{xy}]\) does not necessarily have real eigenvalues. The diagonalization ([\[eq.diagonalization\]](#eq.diagonalization){reference-type="ref" reference="eq.diagonalization"}) implies the *spectral representation* \[\label{eq.spectral-representation} C = \sum_{x\in X} \lambda_x U_xU^*_x\] with pairwise orthogonal vectors \(U_x\in \C^X\) of unit length \(\|U_x\|= \sqrt{U^*_xU_x} = 1\) and real numbers \(\lambda_x\). ([\[eq.spectral-representation\]](#eq.spectral-representation){reference-type="ref" reference="eq.spectral-representation"}) shows that the members of \(\H_X\) are precisely the real linear combinations of (self-adjoint) matrices of the form \[uu^* \quad\mbox{with}\quad u\in \C^X \quad\mbox{s.t.}\quad \|u\|^2 =u^*u = 1.\]
## Measurements {#sec:measurements}
By a *(linear) measurement* on the interaction system \(\mathfrak X =(X,\mathcal{A})\), we understand a linear functional \(f:\R^{X\times X}\to \R\) on the space of all possible interaction instances. \(f(A)\) is the value observed when the agents in \(X\) interact according to \(A\). So there is a matrix \(F\in \R^{X\times X}\) such that \[\label{eq:meas} f(A) = \(F|A\) = \(\hat{F}|\hat{A}\) \quad\mbox{for all \(A\in \R^{X\times X}\).}\]
We give some illustrating examples.
- **Cooperative games.** A *linear value* for a player in a cooperative game (*sensu* Ex. [\[ex:coga\]](#ex:coga){reference-type="ref" reference="ex:coga"}), is a linear functional on the collection of diagonal interß-action matrices \(V\). Clearly, any such functional extends linearly to all interaction matrices \(A\). So the Shapley value (or any linear value (probabilistic, Banzhaf, egalitarian *etc.*) can be seen as a measurement. Indeed, taking the example of the Shapley value, for a given player \(i\in N\), the quantity \[\phi_i(v) = \sum_{S\subseteq N\setminus i}\frac{(n-s-1)!s!}{n!}(v(S\cup i)-v(S))\] with \(s=|S|\), acts linearly on the diagonal matrices \(V\) representing the characteristic function \(v\). Similar conclusions apply to bicooperative games.
- **Communication Networks.** The literature on graphs and networks[^10] proposes various measures for the centrality (Bonacich centrality, betweenness *etc.*) or prestige (Katz prestige index) of a given node in the graph, taking into account its position, the number of paths going through it, *etc.* These measures are typically linear relative to the incidence matrix of the graph and thus represent measurements.
Further important examples arise from payoff evaluations in \(n\)-person games below (*cf.* Remark [\[r.linear-payoff\]](#r.linear-payoff){reference-type="ref" reference="r.linear-payoff"}).
### Probabilistic interpretation {#sec:prob-interpretation1}
If \(\hat{A}= \sum_{x\in X} \lambda_x U_xU^*_x\) is the spectral representation of the self-adjoint matrix \(\hat{A}\), then the measurement ([\[eq:meas\]](#eq:meas){reference-type="ref" reference="eq:meas"}) takes the form \[\label{eq.measurement} f(A) = \sum_{x\in X} \lambda_x \(\hat{F}|U_x U^*_x\) = \sum_{x\in X}\sum_{y\in X}\lambda_x \mu_y\(V_yV_y^*|U_xU_x^*\),\] where \(\hat{F}=\sum_{y\in X} \mu_y V_yV_y^*\) is the spectral representation of \(\hat{F}\). Formula ([\[eq.measurement\]](#eq.measurement){reference-type="ref" reference="eq.measurement"}) has an immediate intuitive probabilistic interpretation, well-known in quantum theory, when we set \[p_{xy} = \(V_yV_y^*|U_xU_x^*\) = \tr(V_yV_y^*U_xU_x^*\).\] Indeed, Lemma [\[l.measure-probabilities\]](#l.measure-probabilities){reference-type="ref" reference="l.measure-probabilities"} (below) implies \(p_{xy} \geq 0\). Moreover, since the \(V_y\) and \(U_x\) yield unitary bases of \(\C^X\), we have \[\sum_{x,y\in X}p_{xy} = \sum_{x\in X}\sum_{y\in X} \(V_yV_y^*|U_xU_x^*\) =\sum_{x\in X} \(I|U_xU_x^*\) = \(I|I\) = 1.\] *i.e.*, the \(p_{xy}\) constitute a probability distribution on the set \(X\times X\) of pairs of agents \((x,y)\) and \[f(A)=\sum_{x,y\in X} \lambda_x\mu_y p_{xy}\] is the expected value of the corresponding eigenvalue products \(\lambda_x\mu_y\).
Let \(z=\sum_{j} \ov{u}_jv_j\). Since \(V_{ij} = v_i\ov{v}_j\) and \(U_{ij} = u_i\ov{u}_j\), one finds \[\(V|U\) = \sum_{i=1}^k\sum_{j=1}^k \ov{v_i}v_j u_i\ov{u}_j = \sum_{i=1}^k \ov{v}_iu_i \cdot z = \ov{z}z = |z|^2 \geq 0.\] 0◻
# Decision analysis {#sec.decision-analysis}
Consider \(n\) decision makers, each of them facing a choice between two alternatives. Depending on the context, the alternatives can have different interpretations: 'no' or 'yes' in a voting context, for example, or 'inactive' or 'active' if the decision maker is a player in a game theoretic or economic context *etc.*. Moreover, the alternatives do not have to be the same for all decision makers. Nevertheless, it is convenient to denote the alternatives simply by the symbols '\(0\)' and '\(1\)'. This results in no loss of generality.
## The case \(n=1\) {#sec:sinage}
We start with a single decision maker with the binary alternatives '\(0\)' and '\(1\)'. Let \(d_{jk}\in \R\) be a measure for the tendency of the decision maker to consider the choice of alternative \(j\) but to possibly switch to alternative \(k\). The \(4\)-dimensional coefficient vector \(d=(d_{00},d_{10},d_{01},d_{11})\) represents this state of the decision maker. Moreover, in view of the isomorphism \[\R^4 \;\cong\; (\R+\im \R) \times (\R+\im \R),\] we can represent \(d\) by a pair of complex numbers \(\delta_j\): \[\hat{d} = (d_{00}+ \im d_{10}, d_{01}+ \im d_{11}) =(\delta_0,\delta_1).\] In the non-trivial case \(d\neq 0\), there is no loss of generality when we assume \(d\) (and hence \(\hat{d}\)) to be unit normalized \[\|d\| = \sqrt{d_{00}^2+d_{10}^2 + d_{01}^2 +d_{11}^2} = \|\hat{d}\| = 1.\] So we think of the unit vector set \[\cS = \{\delta =(\delta_0,\delta_1)^T \in \C^2\mid \|\delta\|^2 = |\delta_0|^2 +|\delta_1|^2 =1\}\] as the set of *(proper) states* of the decision maker.
### Decisions and interactions {#sec:decisions-interactions}
While a vector state \(\delta=(\delta_0,\delta_1)^T\in \cS\) describes a state of the decision maker, the decision framework implies a natural interaction system in the set \(X=\{0,1\}\) of alternatives. Indeed, the matrix \[\label{eq.decision-density} D = \delta \delta^* = \begin{pmatrix} \delta_0\ov{\delta}_0 & \delta_0\ov{\delta}_1 \\ \delta_1\ov{\delta}_0 &\delta_1\ov{\delta}_1\end{pmatrix} = D = \delta \delta^* = \begin{pmatrix} |\delta_0|^2 & \delta_0\ov{\delta}_1 \\ \delta_1\ov{\delta}_0 &|\delta_1|^2\end{pmatrix}\] associates with the decision state \(\delta\) a "quantum state" with density \(D\), which may be interpreted as the self-adjoint representation of an interaction state on \(X\). The latter exhibits the alternatives '0' and '1' as interacting agents in their own right in the decision process.
### Decision probabilities {#sec:decision-probabilities}
The decision state vector \(\hat{d} = (d_{00}+ \im d_{10}, d_{01}+ \im d_{11}) =(\delta_0,\delta_1)\in \cS\) defines a probability distribution \(p\) on \(X^2 =\{0,1\}^2\) with probabilities \(p^{(d)}_{jk} = d_{jk}^2\). Accordingly, the probability for the alternative '\(k\)' to be accepted is \[p_k^{(d)} = p^{(d)}_{0k} + p^{(d)}_{1k} = |\delta_k|^2 \quad(k=0,1).\] We arrive at the same probabilities when we apply suitable measurements in the sense of Section [2.3](#sec:measurements){reference-type="ref" reference="sec:measurements"} on the interaction system on \(X\) with the self-adjoint state representation \(D\) as in ([\[eq.decision-density\]](#eq.decision-density){reference-type="ref" reference="eq.decision-density"}). To see this, let us measure the activity level of \(k\) (*i.e.*, the interaction of \(k\) with itself). This measurement corresponds to the application of the matrix \(F_k = e_ke_k^T\), where \(e_k\) is the \(k\)th unit vector in \(\R^2\). The expected value of this measurement is \[\(F_k|D\) = |\delta_k|^2 = p_k^{(d)}.\] Let us now take another look at earlier examples:
1. Influence networks: While it may be unusual to speak of "influence" if there is only a single agent, the possible decisions of this agent are its opinions ('yes' or 'no'), and the state of the agent hesitating between 'yes' and 'no' is described by \((\delta_0,\delta_1)\in\C^2\) (with probability \(|\delta_0|^2\) to say 'no' and \(|\delta_1|^2\) to say 'yes').
2. Cooperative games: As before, "cooperation" may sound awkward in the case of a single player. However, the possible decisions of the player are relevant and amount to being either active or inactive in the given game. Here, \((\delta_0,\delta_1)\) represents a state of deliberation of the player that results in the probability \(p_k^{(d)}\) for being active (resp. inactive).
### Quantum bits {#sec:qbits}
Denote by \(|j\)\) the event that the decision maker selects alternative \(j\). Thinking of \(|0\)\) and \(|1\)\) as independent generators of a \(2\)-dimensional complex vector space, we can represent the decision states equally well as formal linear combinations \[\label{eq.q-bit} \delta = \delta_0|0\) + \delta_1|1\) \quad ( (\delta_0,\delta_1)^T\in \cS).\] The expression ([\[eq.q-bit\]](#eq.q-bit){reference-type="ref" reference="eq.q-bit"}) is the representation of the states of a \(1\)-dimensional quantum system in the so-called *Schrödinger picture* of quantum mechanics. Strong arguments have been put forward[^11] that the decision process of the human mind should be considered to be a quantum process and a binary decision state should thus be described as a quantum state. In the theory of quantum computing[^12], an expression of the form ([\[eq.q-bit\]](#eq.q-bit){reference-type="ref" reference="eq.q-bit"}) represents a *quantum bit*. \(|0\)\) and \(|1\)\) correspond to the boolean bits of classical computation.
### Non-binary alternatives {#sec:non-binary-alternatives1}
It is straightforward to generalize the decision model to \(k>1\) alternatives '\(0\)', '\(1\)', \(\ldots\), '\(k-1\)'. We let \(|j\)\) denote the event that '\(j\)' is selected. Interpreting '\(j\)' as representing an "activity" at the \(j\)th of \(k\) possible levels, the decision maker is essentially a player in a *multichoice game* in the sense of Hsiao and Raghavan . Decision states are now described by formal linear combinations \[\delta = \sum_{j=0}^{k-1} \delta_j|j\) \quad\mbox{with \(\delta_j\in \C\) s.t. \( \|\delta\|^2 =\D\sum_{j=0}^{k-1} |\delta_j|^2 = 1\)}\] and \(|\delta_j|^2\) being the probability for alternative \(j\) to be taken.
## The case \(n\geq 2\) {#sec:n2}
We now extend our framework to the case of several decision makers and consider first the binary case. We let \(\mathfrak D\) denote of the decision system for \(n=1\) that can be in any state as in Section [3.1](#sec:sinage){reference-type="ref" reference="sec:sinage"}. The system of \(n >1\) decision makers is then given by the \(n\)-fold tensor product: \[\mathfrak D^n = \mathfrak D \otimes \cdots \otimes \mathfrak D \quad(\mbox{\(n\) times}).\] To make this precise, recall the tensor product of a vector space \(V\) with a fixed basis \(B=\{b_1,\ldots,b_n\}\) and a vector space \(W\) with basis \(G =\{g_1,\ldots,g_m\}\) over the common real or complex scalar field \(\F\). We first define the set \(B\otimes G\) of all (formal) products \(b_i\otimes g_j\) and extend the products linearly to products of vectors in \(V\) and \(W\), *i.e.*, if \(v =\sum v_i b_i\) and \(w =\sum w_j g_j\), then \[\label{eq.tensor-product} v\otimes w =\big(\sum_{i=1}^n v_ib_i\big)\otimes \big(\sum_{j=1}^m w_j g_j\big) = \sum_{i=1}^n\sum_{j=1}^m (v_iw_j) b_i\otimes g_j.\] So a coordinate representation of \(v\otimes w\) is given by the (\(n\times m\))-matrix with coefficients \(v_iw_j\). Moreover, we note the multiplication rule for the norm: \[\|v\otimes w\|^2 = \|v\|^2 \|w\|^2.\] The tensor product \(V\otimes W\) is the \(nm\)-dimensional vector space with basis \(B\otimes G\). It is well-known that the tensor product of vectors is associative and distributive and linear relative to the scalar field \(\F\). We define \(\mathfrak D^n\) as the system with the state set \[\cS^n = \{\delta \in \C^2\otimes \cdots \otimes \C^2\mid \|\delta\| = 1\}.\] The construction becomes immediately clear when bit notation is employed. For any sequence \(j_1\ldots j_n\in \{0,1\}^n\), we define the *\(n\)-bit* \[\label{eq.n-bit} |j_1\ldots j_n\) = |j_1\)\otimes \cdots\otimes |j_n\).\] The states of \(\mathfrak D^n\) now take the form of linear combinations \[\label{eq.n-decision-state} \delta = \sum_{k\in \{0,1\}^n} \delta_k |k\) \quad\mbox{with \(\delta_k\in \C\) and \(\sum_k |\delta_k|^2 = 1\).}\] In state \(\delta\), the (joint) decision \(k\in \{0,1\}^n\) is reached with probability \(|\delta_k|^2\).
As an illustration, consider the case \(n=2\). The decision system \(\mathfrak D^2=\mathfrak D\otimes \mathfrak D\) relative to a pair of binary alternatives represents the states in the form \[\delta = \delta_{00}|00\) + \delta_{10}|10\) + \delta_{11}|11\) +\delta_{01}|01\)\quad(\delta_{ij}\in \C, \sum_{j,k\in \{0,1\}} |\delta_{jk}|^2 = 1).\] In particular, we have \[(\alpha_0 |0\) + \alpha_1|1\))\otimes(\beta_0 |0\) + \beta_1|1\)) =\sum_{i,j\in \{0,1\}} \alpha_i\beta_j |ij\).\] In general, it is often convenient to represent states in set theoretic notation. So we let \(N=\{1,\ldots, n\}\) and identify \(n\)-bits with subsets: \[k = j_1\ldots j_n \in \{0,1\}^n\quad \longleftrightarrow\quad K =\{\ell\in N\mid i_\ell = 1\},\] which yields the state representation \[\label{eq.set-representation} \delta = \sum_{K\subseteq N} \delta_K |K\) \quad(\delta_K\in \C, \sum_{K\subseteq N} |\delta_K|^2 = 1).\] The latter allows us to view a state simply as a valuation \(\alpha:2^N\to \C\) that assigns to each subset \(K\subseteq N\) a complex number \(\alpha_K\in \C\). This generalizes the classical model of cooperative games with \(n\) players to complex-valued cooperative games in a natural way. The probabilistic interpretation says that the decision \(n\)-tuple \(k=i_1\ldots i_n\) is realized in a non-trivial state \(\alpha\) of \(\mathfrak D_n\) with probability \[\label{eq:proba} p^\alpha_k = \frac{|\alpha_k|^2}{\|\alpha\|^2}\;, \quad\mbox{where}\quad \|\alpha\|^2 = \sum_{k\in \{0,1\}^n} |\alpha_k|^2.\] Coming back to our examples, a similar interpretation as for the single agent case can be given. Specifically, in the case of the influence network in state \(\alpha\), \(p^\alpha_k\) in ([\[eq:proba\]](#eq:proba){reference-type="ref" reference="eq:proba"}) yields the probability that a given coalition \(K\) of agents says 'yes' and the other agents say 'no', while in the case of cooperative games, \(p^\alpha_k\) is the probability for a given coalition \(K\) to be active while the remaining players are not.
### Entanglement and fuzzy systems {#sec:entanglement}
We say that a state \(\delta\) of \(\mathfrak D^n\) is *reducible* if there is a \(0 < m<n\) such that \[\alpha = \alpha\otimes \beta\quad\mbox{for some state \(\alpha\) of \(\mathfrak D^m\) and state \(\beta\) of \(\mathfrak D^{n-m}\)}.\] In this reducible case, the state \(\delta\) arises from lower dimensional states \(\alpha\) and \(\beta\) that are probabilistically independent. Indeed, we have \(\|\delta\| = \|\alpha\|\cdot\|\beta\|\), \(\delta_{k\ell} = \alpha_k\beta_\ell\) and hence \[p^\delta_{k\ell} = p^\alpha_k\cdot p^\beta_\ell\quad\mbox{for all \(k\in \{0,1\}^m, \ell\in\{0,1\}^{n-m}\).}\] Already for \(n=2\), it is easy to see that \(\mathfrak D^n\) admits irreducible states. Such states are said to be *entangled*. Aubin has introduced the model of *fuzzy cooperation* of a set \(N\) of \(n\) players as a system whose states are characterized by real vectors \[w = (w_1,\ldots,w_n) \quad\mbox{with}\quad 0\leq w_j\leq 1.\] In state \(w\), player \(j\) decides to be active in the cooperation effort with probability \(w_j\) and inactive to be with probability \(1-w_j\). A coalition \(S\subseteq N\) is assumed to be formed with probability \[w(S) = \prod_{s\in S}\prod_{t\in N\setminus S} w_i(1-w_t)\] (see also Marichal ). So the players act independently in each "fuzzy" state \(w\). Entangled states are thus excluded. In particular, one sees that our model of interactive decisions is strictly more general than the fuzzy model.
## Linear transformations {#sec:linear-transformations}
Let again \(B=\{b_1,\ldots,b_n\}\) be a basis of the vector space \(V\) and \(G =\{g_1,\ldots,g_m\}\) a basis of the vector space \(W\) over the common scalar field \(\F\) and consider linear operators \(L:V\to V\) and \(M:W\to W\). The assignment \[L\otimes M(b\otimes g) = (Lb)\otimes (Mg) \quad((b,g)\in B\times G)\] extends to a unique linear operator \(L\otimes M: V\otimes W\to V\otimes W\) since the elements \(b\otimes g\) form a basis of \(V\otimes W\). The operator \(L\otimes M\) is the *tensor product* of the operators \(L\) and \(M\). For illustration, let us apply this construction to various linear operators on \(\mathfrak D\) and derive several linear transformation of well-known importance in game theory.
### Möbius transform {#sec:Moebius-transform}
The *Möbius operator* \(Z\) on the state vector space of \(\mathfrak D\) is the linear operator such that \[Z|0\) = |0\)+|1\) \quad\mbox{and}\quad Z|1\) = |1\).\] \(Z\) admits a linear inverse \(M = Z^{-1}\), which is determined by \[M |1\) = |1\) \quad\mbox{and}\quad M|1\) = |1\)-|0\).\] The \(n\)-fold tensor product \(Z^n =Z\otimes \cdots \otimes Z\) acts on the states of \(\mathfrak D^n\) in the following way: \[Z^n|i_1\ldots i_n\) = Z|i_1\)\otimes \cdots \otimes Z|i_n\).\] Setting \(N=\{1,\ldots,n\}\) and \(S = \{j\in N\mid i_j = 1\}\), the set theoretic notation thus yields \[Z^n|S\) = \sum_{T\supseteq S} |T\)= \sum_{T\subseteq N} \zeta_{ST}|T\)\] with the *incidence coefficients* \[\zeta_{ST} = \left\{\begin{matrix} 1 &\mbox{if \(S\subseteq T\),}\\ 0 &\mbox{otherwise.}\end{matrix}\right.\] For the inverse \(M^n = M\otimes \cdots \otimes M\) of \(Z^n\), one has \[M^n|i_1\ldots i_n\) = M|i_1\)\otimes \cdots \otimes M|i_n\)\] and thus for any \(S\subseteq N\), \[M^n|S\) = \sum_{T\supseteq S} (-1)^{|T\setminus S|}|T\) = \sum_{T\subseteq N} \mu_{ST}|T\)\] with the *Möbius coefficients*[^13] \[\mu_{ST} = \left\{\begin{matrix} (-1)^{|T\setminus S|} &\mbox{if \(S\subseteq T\),}\\ 0 &\mbox{otherwise.}\end{matrix}\right.\] The Möbius operator \(Z^n\) has well-known game theoretic implications. With the characteristic function \(v:2^N\to\R\) of a cooperative game on \(N\), we associate the state vector \[\tilde{v} = \sum_{S\subseteq N} v(S)|S\).\] \(v\) is said to be a *unanimity game* if \(\tilde{v} = Z^n|S\)\) holds for some \(S\subseteq N\). In general, we may compute the Möbius transform \[\hat{v} = Z^n\tilde{v} =\sum_{S\subseteq N} v(S)Z^n|S\) =\sum_{S\subseteq N} \hat{v}(S)|S\)\] which expresses the Möbius transform \(\hat{v}\) of \(\tilde{v}\) as a linear combination of state vectors of unanimity games. Similarly, an application of the inverse \(M^n\) yields \[\tilde{v} = M^n \hat{v} = \sum_{S\subseteq N} \hat{v}(S)M^n|S\) =\sum_{S\subseteq N}\sum_{T\subseteq S} (-1)^{|S\setminus T|}\hat{v}(T)|S\),\] which yields the *Möbius inversion formula* \[v(S) = \sum_{T\subseteq S} (-1)^{|S\setminus T|}\hat{v}(T) \quad(S\subseteq N)\] The parameters \(h^v_{ST} = (-1)^{|S\setminus T|}\hat{v}(T)\) are known as the *Harsanyi coefficients* of \(v\) in cooperative game theory.
### Hadamard transform {#sec:Hadamard-transform}
The *Hadamard transform* \(H\) of \(\mathfrak D\) is the linear transformation with the property \[H|0\) = \frac{|0\) + |1\)}{\sqrt{2}} \quad\mbox{and}\quad H|1\) = \frac{|0\)-|1\)}{\sqrt{2}}.\] The normalizing factor \(\sqrt{2}\) is chosen as to make \(H\) norm preserving. Note that in both states \(H|0\)\) and \(H|1\)\) the alternatives \(0\) and \(1\) are attained with equal probability \[p^{H|j\)}_0 = \frac12 = p^{H|j\)}_1 \quad(j=0,1)).\] It is easy to check that \(H\) is self-inverse (*i.e.*, \(H^{-1} = H\)). If follows that also the \(n\)-fold tensor product \(H^n = H\otimes \cdots \otimes H\) is norm preserving and self-inverse on the vector state space of \(\mathfrak D^n\). In particular, the \(2^n\) state vectors \[H^n|i_1\ldots i_n\) = H|i_1\) \otimes \cdots \otimes H|i_n\) \quad( i_j\in \{0,1\})\] are linearly independent. In the set theoretic notation, one finds \[H^n|S\) =\frac{1}{2^{n/2}} \sum_{T\subseteq N} (-1)^{|S\cap T|} |T\).\] Note that in each of the \(2^n\) states \(H^n|S\)\) of \(\mathfrak D^n\), the \(2^n\) boolean states \(|T\)\) of \(\mathfrak D^n\) are equi-probable.
Lastly, we point out an interesting connection of the Hadamard transform with interaction transforms of Grabisch and Roubens, already alluded to in Example [\[ex:in2a\]](#ex:in2a){reference-type="ref" reference="ex:in2a"}. These have been proposed in the context of cooperative games, and are essentially of two types: the Shapley interaction transform, which extends the Shapley value, and the Banzhaf interaction transform, which extends the Banzhaf value. The latter is expressed as follows. The *Banzhaf transform* of the TU-game \(v:2^N\rightarrow \R\) is the pseudo-boolean function \(I_{\rm B}\) with the values \[I_{\rm B}^v(S) = \Big(\frac{1}{2}\Big)^{n-s}\sum_{T\subseteq N}(-1)^{|S\setminus T|}v(T) \qquad (S\in 2^N)\] representing the quantity of interaction existing among agents in \(S\). \(I_{\rm B}^v(S)>0\) indicates that the cooperation of all agents in \(S\) brings more than any combination of subgroups of \(S\), while \(I_{\rm B}^v(S)<0\) implies some redundancy/overlap in the cooperation of agents in \(S\). By the identitification TU-game '\(v\) \(\leftrightarrow\) state \(\alpha\)', the Hadamard transform becomes a transform \(H^v\) on TU-games with values \[H^v(S) =\frac{1}{2^{n/2}} \sum_{T\subseteq N} (-1)^{|S\cap T|}v(T).\] Now, it is easy to check that \[I_{\rm B}^v(S) = \frac{(-2)^{s}}{2^{n/2}}H^v(S)\] holds, which yields an interpretation of the Hadamard transform in terms of interaction in cooperative contexts.
### Fourier transformation {#sec:Fourier-transformation}
We briefly recall the classical discrete Fourier transform of the coordinate space \(\C^k\). Setting \(\omega = e^{2\pi\im/k}\in \C\), one defines the (unitary) *Fourier matrix* \[\Omega = \frac{1}{\sqrt{k}}\begin{pmatrix}\omega^1&\omega^2 &\ldots& \omega^{k}\\ \omega^{2}&\omega^4 &\ldots& \omega^{2k}\\ \vdots &\vdots &\cdots &\vdots \\ \omega^k &\omega^{2k} &\ldots&\omega^{k^2} \end{pmatrix} \in \C^{k\times k}.\] The *Fourier transform* of \(v\in \C^k\) is the vector \(\Omega v\in \C^k\). Applied to decision systems, the Fourier transform of any state vector will also be a state vector.
The Fourier transform extends naturally to interactions. Indeed, for any matrix \(M\in \C^{k\times k}\), the linear operator \[C \mapsto \mu(C) = M C\ M^* \quad(C\in \C^{k\times k})\] preserves self-adjointness and thus acts as a linear operator on the Hilbert space \(\H_k\) of all (\(k\times k\)) self-adjoint matrices. In particular, we have \[C =\sum_{i=1}^k \lambda_i u_iu_i^*\;\Longrightarrow \; \mu(C) = \sum_{i=1}^k M(\lambda_i u_iu_i^*)M^* = \sum_{i=1}^k \lambda_i (Mu_i)(Mu_i)^*.\] Hence, if \(M\) is unitary, we observe that the spectral decomposition of a self-adjoint matrix \(C\) with eigenvectors \(u_i\) yields a spectral decomposition of \(\mu_M(C)\) with eigenvectors \(Mu_i\) (and the same eigenvalues). The choice \(M = \Omega\) thus yields the Fourier transform for interaction instances, which preserves the symmetry values: \[\hat{A} \mapsto \mu_\Omega(\hat{A}) = \Omega \hat{A}\Omega^* \quad(A\in \R^{k\times k}).\]
## Decision and quantum games {#sec:decision-games}
An *\(n\)-person game* \(\Gamma\) involves \(n\) players \(j\), each of them having a set \(S_j\) of feasible strategies relative to some system \(\mathfrak S\), which is assumed to be in an initial state \(\alpha^{(0)}\). Each player \(j\) selects a strategy \(s_j\in S_j\) and the joint selection \(s=(s_1,\ldots,s_n)\) of strategies then moves the system into a final state \(\alpha^{(f)}\). The reward of player \(j\) is \(P^{(j)}(\alpha^{(f)})\), where \(P^{(j)}\) is a pre-specified real-valued functional on the state space of \(\mathfrak S\). The \(n\)-person game \(\Gamma\) is a *decision game* if it is played relative to a decision system \(\mathfrak D^m\) with binary alternatives for some \(m>0\). By their strategic choices, the \(n\) players influence the \(m\) decision makers towards a certain decision, *i.e.*, the joint strategy \(s\) moves the game to a decision state \[\delta^{(f)} = \sum_{k\in\{0,1\}^m} \delta_k^{(f)} |k\) \quad(\delta^{(f)}_k\in\C, \sum_{k\in\{0,1\}^m} |\delta_k^{(f)}|^2 = 1).\] In the state \(\delta^{(f)}\), the \(m\) decision makers will accept the alternative \(k\) with probability \(|\delta_k^{(f)}|^2\), in which case \(|k\)\) is considered to be the final state of the game and player \(j\) receives a pre-specified reward \(p_k^{(j)}\). Hence \(j\)'s *expected* payoff is \[\label{eq.q-game-payoff} P^{(j)}(\delta^{(f)}) =\sum_{k\in\{0,1\}^m} p^{(j)}_k |\delta_k^{(j)}|^2.\] Interpreting the decision states of \(\mathfrak D^m\) as the states of an \(m\)-dimensional quantum system and regarding \(\delta^{(f)}\) as the final state of the game, we arrive at the model of a *quantum game* with the payoff functionals \(P^{(j)}\).
In the quantum games discussed in the literature, the strategies of the individual players typically consist in the application of certain linear operations on the states[^16]. As an illustration, consider the generalization of the classical Prisoners' Dilemma of Eisert *et al.* : : There are \(2\) players relative to the system \(\mathfrak D^2 = \mathfrak D\otimes \mathfrak D\) with states given as \[\delta = \delta_{00}|00\)+\delta_{01}|01\) + \delta_{10}|10\) + \delta_{11}|11\) \quad(\delta_{ij}\in \C).\] The game is initialized by a unitary operator \(U\) and prepared into the state \(\alpha^{(0)}= U|00\)\). The players select unitary operators \(A\) and \(B\) on \(\mathfrak{D}\), whose tensor product \(A\otimes B\) is applied to \(\alpha^{(0)}\). A further application of \(U^*\) results in the final state \[\alpha^{(f)} = U^*(A\otimes B)\alpha^{(0)} = U^*(A\otimes B)U|00\).\] The payoff coefficients \(p_k^{(j)}\), with \(k\in \{00,01,10,11\}\), are the coefficients of the payoff matrix in the classical Prisoners' Dilemma. Strategic choices of operators associated with the Pauli matrices \(\pi_1\) and \(\pi_2\) (Section [2.1.1](#sec:binary-interaction){reference-type="ref" reference="sec:binary-interaction"}), for example, would act on the states in the following way: \[\begin{aligned}
\pi_1 |0\) &=& |0\) \quad\mbox{and}\quad \pi_1 |1\) =-|1\)\\ \pi_2 |0\) &=& |1\) \quad\mbox{and}\quad \pi_2 |1\) = |0\). \end{aligned}\] Eisert *et al.* show that the set \(\{I,\pi_1,\pi_2\}\) of admissible strategies for each player can guarantee the existence of a Pareto optimal Nash equilibrium in cases where the classical variant does not admit such a Nash equilibrium. Without going into further details, we mention that the Hadamard transform \(H\) often turns out to be a powerful strategic operator in quantum games (see, *e.g.*, the seminal penny flip game of Meyer ).
# Markov evolutions {#sec:Markov-evolutions}
Interaction among agents may be time dependent. For example, opinions form over time in mutual information exchange, and game theoretic coalitions form due to changing economic conditions *etc.* Generally, we understand by an *evolution* of an interaction system \(\mathfrak X = (X,\cal A)\) in discrete time \(t\) a sequence \(\epsilon= (\alpha_t)_{t\geq 0}\) of states \(\alpha_t\in \mathcal A\). The evolution \(\epsilon\) is *bounded* if there is some \(c\in \R\) such that \(\|\alpha_t\|< c\) holds for all \(t\). \(\epsilon\) is *mean ergodic* if the averages of the state representations \[\ov{\alpha}^{(t)} = \frac1t\sum_{m=1}^t \hat{\alpha}_m \in \H_X\] converge to a limit \(\ov{\alpha}^{(\infty)}\). We say that \(\epsilon\) is a *Markov evolution* if there is a linear ("evolution") operator \(\Phi: \H_X\to\H_X\) such that \[\hat{\alpha}_{t+1} = \Phi^t \hat{\alpha}_0 \;,\emph{i.e.}, \; \hat{\alpha}_{t+1} = \Phi \hat{\alpha}_t \quad\mbox{holds for all \(t\geq 0\).}\] A Markov evolution is mean ergodic if and only if it is bounded:
Theorem 2 in Faigle and Schönhuth . 0◻The importance of these notions lies in the fact that the mean ergodic evolutions guarantee the statistical convergence of arbitrary measurements (in the sense of Section [2.3](#sec:measurements){reference-type="ref" reference="sec:measurements"}) on the evolution. More precisely, we have
We illustrate this model with two well-known examples.
#### Markov chains.
Let \(\pi\in\R^k\) be a probability distribution and \(M\in \R^{k\times k}\) such that every "state" \(\pi^{(t)} = M^t\pi\) is a probability distribution and hence satisfies \(\|\pi^t\| \leq k\). It follows from Theorem [\[t.FS\]](#t.FS){reference-type="ref" reference="t.FS"} that \((\pi^{(t)})_{t\geq 0}\) is a mean ergodic Markov evolution. If \(M\) is a probabilistic transition matrix (*i.e.*, all columns of \(M\) are probability distributions), the Markov evolution \((\pi^{(t)})_{t\geq 0}\) is a classical *Markov chain*, in which case mean-ergodicity is well-known.
#### Schrödinger's wave equation.
Recall that the Schrödinger picture of quantum mechanics describes a *state evolution* of a finite-dimensional quantum system by a time dependent function \(\psi\) with values \(\psi(t)\in \C^k\) for some \(k<\infty\), which is supposed to satisfy the differential equation \[\label{eq.Schroedinger} \frac{\partial \psi(t)}{\partial t} = \frac{\im }{\hbar}H\psi(t),\] with the matrix \(H\) being the so-called *Hamiltonian* of the system and \(\hbar\) Planck's constant. Assuming \(H\) to be normal (*i.e.*, \(HH^* = H^*H\)), one obtains the solution of ([\[eq.Schroedinger\]](#eq.Schroedinger){reference-type="ref" reference="eq.Schroedinger"}) in the form \[\psi(t) = U_t \psi(0) \quad\mbox{(with \(U_t= e^{-\im H t/\hbar}\)).}\] Note that \(U_t\) is unitary. So \(\|\psi(t)\| = \|\psi(0)\|\) is bounded. If \(H\) moreover is constant in time, one finds \[\label{eq.unitary-evolution} \psi(t) = U^t\psi (0) \quad\mbox{with \(U= e^{-\im H /\hbar}\in\C^{k\times k}\) unitary,}\] which exhibits the (discrete) Schrödinger evolution \((\psi(t))_{t\geq 0}\) as mean-ergodic.
The examples show that traditional models for the evolution of interaction systems are quite restrictive and suggest to study evolutions in a wider context.
# Appendix: An example with two agents {#appendix-an-example-with-two-agents .unnumbered}
Let us illustrate the main notions and results introduced in the paper. For simplicity, we consider a set \(X\) two agents and interaction matrices of the form \[A=\begin{pmatrix} w_1 & 1-w_1 \\ 1-w_2 & w_2\end{pmatrix}\] with \(w_1,w_2\in[0,1]\). \(A\) is row-stochastic but not symmetric unless \(w_1=w_2\). \(A\)'s entries are generally interpreted as interaction/activity levels, so that \(w_1,w_2\) are the activity levels of agents 1 and 2, while \(1-w_1,1-w_2\) are their interaction levels. Following Example [\[ex:inne\]](#ex:inne){reference-type="ref" reference="ex:inne"}, we understand \(A\) as an influence matrix. In this case, agent \(i\) would listen to the opinion of the other agent with weight \(1-w_i\), and to himself with weight \(w_i\). We first apply the symmetry decomposition and find \[A^+=\begin{pmatrix} w_1 & 1-\frac{1}{2}W \\ 1-\frac{1}{2}W & w_2\end{pmatrix}, \quad A^-=\begin{pmatrix} 0 &-\frac{1}{2}\Delta W \\ \frac{1}{2}\Delta W & 0\end{pmatrix}\] with the shorthand \(W=w_1+w_2\) and \(\Delta W=w_1-w_2\). Therefore, the self-adjoint representation is \[\hat{A} = \begin{pmatrix} w_1 & 1-\frac{1}{2}W-\im \frac{1}{2}\Delta W\\ 1-\frac{1}{2}W +\im \frac{1}{2}\Delta W & w_2\end{pmatrix}\] with the eigenvalues \[\lambda = \frac{W\pm \sqrt{(W-2)^2+2\Delta W^2}}{2}.\] The corresponding eigenvectors are \[u_1 = \frac{1}{\sqrt{8+2W^2-8W+4\Delta W^2-2\Delta W\sqrt{D}}}\begin{pmatrix} \sqrt{4+W^2-4W + \Delta W^2} \\ \displaystyle\frac{(1-\frac{1}{2}W + \im \frac{\Delta W}{2})(-\Delta W+\sqrt{D})}{\sqrt{1+\frac{1}{4}W^2-W + \frac{\Delta W^2}{4}}}\end{pmatrix}\] \[u_2 = \frac{1}{\sqrt{8+2W^2-8W+4\Delta W^2+2\Delta W\sqrt{D}}}\begin{pmatrix} \sqrt{4+W^2-4W + \Delta W^2} \\ \displaystyle\frac{(1-\frac{1}{2}W + \im \frac{\Delta W}{2})(-\Delta W-\sqrt{D})}{\sqrt{1+\frac{1}{4}W^2-W + \frac{\Delta W^2}{4}}}\end{pmatrix}\] One can see that calculations can become rapidly quite complex. In order to clarify the results, we perform a quantum theoretically standard change of variables and set \[\begin{aligned}
E_0 &=&\frac{W}{2}\\ \Delta & =& \sqrt{\frac{1}{2}\Delta W^2 + 1 + \frac{1}{4}W^2-W}\\ \tan\theta & =& \frac{2\sqrt{1+\frac{1}{4}W^2-W+\frac{1}{4}\Delta W^2}}{\Delta W}\\ e^{-\im\varphi} & =& \frac{1-\frac{1}{2}W-\im\frac{1}{2}\Delta W}{\sqrt{1 + \frac{1}{4}W^2-W +\frac{1}{4}\Delta W^2}}. \end{aligned}\] Then the eigenvalues are \[\lambda_1=E_0+\Delta, \quad \lambda_2=E_0-\Delta,\] and the unit eigenvectors take the form \[\label{eq:u} u_1=\begin{pmatrix}\cos(\theta/2) \\ e^{\im\varphi}\sin(\theta/2)\end{pmatrix}, \quad u_2 = \begin{pmatrix}-\sin(\theta/2)\\ e^{\im\varphi}\cos(\theta/2)\end{pmatrix}.\] Let us compute the evolution of the system by applying Schrödinger's equation, taking the Hamiltonian to be the self-adjoint matrix \(\hat{A}\), and \(\psi(0)=e_1=(1 \ \ 0)^T\). \(\psi(t)\) represents the state of the agents at timer \(t\). Expressed in the \((u_1,u_2)\) basis, we find, with the (reduced) Planck constant \(\hbar = h/2\pi\), \[\psi(t) = \cos(\theta/2)e^{-\frac{\im}{\hbar}\lambda_1t}u_1-\sin(\theta/2)e^{-\frac{\im}{\hbar}\lambda_2t}u_2\] and, in the standard \((e_1,e_2)\) basis, using ([\[eq:u\]](#eq:u){reference-type="ref" reference="eq:u"}): \[\psi(t) = e^{-\frac{\im}{\hbar}E_0t}\begin{pmatrix}\cos(t\Delta/\hbar) -\im\cos\theta\sin(t\Delta/\hbar) \\-\im e^{\im\varphi}\sin(\theta)\sin(t\Delta/\hbar)\end{pmatrix}.\] The probability of transition from state \(e_1=(1\ \ 0)^T\) to \(e_2=(0\ \ 1)^T\) after time \(t\) is given by \[\mathrm{Pr} = \|\langle e_2|\psi(t)\rangle\|^2 = \sin^2\theta\sin^2(t\Delta/\hbar).\] This probability has period \(\pi \hbar/\Delta\) and maximal value \(\sin^2\theta\). In our case, we have \[\sin^2\theta = \frac{\tan^2\theta}{\tan^2\theta+1}= 1-\frac{\Delta W^2}{4 + W^2-4W + 2\Delta W^2},\] which is the maximal value of the probability of transition. Its period is \[\frac{\pi\hbar}{\sqrt{\frac{1}{2}\Delta W^2 + 1 + \frac{1}{4}W^2-W}}.\] Let us consider slightly different interaction matrices \(A',A''\): \[A'=\begin{pmatrix} w_1 & w_1-1 \\ w_2-1 & w_2\end{pmatrix}, \quad A''=\begin{pmatrix} w_1 & 1-w_1 \\ w_2-1 & w_2\end{pmatrix}\] with \(w_1,w_2\in[0,1]\). In \(A'\), the antidiagonal elements are nonpositive, which, when \(A'\) is interpreted as an influence matrix, corresponds to an anticonformist attitude, since the agent has a tendency to adopt an opinion opposite to the one of the other agent. \(A''\) represents a mixed situation where agent 1 is conformist and agent 2 is anticonformist. The corresponding self-adjoint matrices are \[\widehat{A'} = \begin{pmatrix} w_1 & \frac{1}{2}W-1 +\im \frac{1}{2}\Delta W\\ \frac{1}{2}W-1-\im \frac{1}{2}\Delta W & w_2\end{pmatrix}\] \[\widehat{A''} = \begin{pmatrix} w_1 & \frac{1}{2}\Delta W +\im (\frac{1}{2} W-1)\\ \frac{1}{2}\Delta W-\im (\frac{1}{2} W-1) & w_2\end{pmatrix}\] Observe that since in all cases the term \(A_{12}\) is the same, all parameters \(E_0,\Delta, \theta\) remain the same, only \(\varphi\) changes. Consequently, the eigenvalues and probability of transition are the same, only the evolution of \(\psi\) changes. In the particular situation \[A=\begin{pmatrix}0 & 1 \\ 1 & 0\end{pmatrix},\] the agents display no activity (or no self-confidence) and interact positively (or follow the other's opinion fully). The eigenvalues are \(\pm 1\), \(\theta=\frac{\pi}{2}\) and \(\Delta =1\), yielding a maximal amplitude for probability transition and a period equal to \(\pi\hbar\). \(e^{-\im\varphi}=1\) holds and the evolution is \[\psi(t) = \cos(t/\hbar)\begin{pmatrix}1\\0\end{pmatrix}-\im\sin(t/\hbar)\begin{pmatrix}0\\1\end{pmatrix}.\] Another extreme case is \[A=\begin{pmatrix}0 &-1 \\-1 & 0\end{pmatrix},\] where the two agents have negative interaction or are pure anticonformists. The eigenvalues are \(\pm 1\), \(\theta=\frac{\pi}{2}\) and \(\Delta =1\), yielding a maximal amplitude for probability transition and a period equal to \(\pi\hbar\). Again \(e^{-\im\varphi}=-1\) holds with evolution \[\psi(t) = \cos(t/\hbar)\begin{pmatrix}1\\0\end{pmatrix} + \im\sin(t/\hbar)\begin{pmatrix}0\\1\end{pmatrix}.\] Note that in both cases, the state is oscillating with maximal amplitude, which is well in line with the intuition that the opinion is very unstable in case every agent just copies the activity/opinion of the other or does exactly the opposite.
[^1]: see also Grabisch and Roubens for a general approach
[^2]: see, *e.g.*, the surveys Grabbe and Huo *et al.*
[^3]: *e.g.*, Vourdas
[^4]: Iqbal and Toor discuss a 3 player situation
[^5]: see, *e.g.*, Levy
[^6]: *cf.* Wolpert
[^7]: see, *e.g.*, Nielsen and Chang
[^8]: see, *e.g.*, Grabisch and Rusinowska
[^9]: see Bilbao *et al.*, Labreuche and Grabisch
[^10]: see, *e.g.*, Jackson
[^11]: see, *e.g.*, Penrose
[^12]: see, *e.g.*, Nielsen and Chuang
[^13]: see Rota for generalizations
[^14]: Hammer and Rudeanu
[^15]: see, *e.g.*, Kalai , O'Donnell
[^16]: see, *e.g.*, Grabbe or Guo *et al.*
[^17]: see, e.g., Faigle and Grabisch | {'timestamp': '2017-06-07T02:06:50', 'yymm': '1706', 'arxiv_id': '1706.01769', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01769'} |
null | null |
# Introductions
Barrier coverage and area coverage are two important problems in applications of wireless sensor networks. In both two problems, sensors are deployed in such a way that every point of the target region is monitored by at least one sensor. For area coverage, the target region is traditionally a bounded area in the plane; while in the barrier coverage problem arising from border surveillance for intrusion detection, the target region are the borders and the goal is to deploy sensors along the borders such that at least one sensor will detect if any intruder crosses the border. Unlike area coverage, barrier coverage requires only to cover every points of the borders, rather than every point of the area bounded by the border. So barrier coverage uses much fewer sensors, and hence is more cost-efficient, particularly in practical large-scale sensor deployment. To accomplish a barrier coverage, sensors are dispersed along the borders. However, there may exist gaps after the dispersal, so the border line might not be completely covered. One approach is to disperse the sensors in multiple rounds, and guarantee the probability of complete coverage by the dispersal density of the sensors. The other approach is to acquire some sensors with the ability of relocation (i.e. mobility), such that after dispersal, the sensors can move to monitor the gaps on the barrier. In this context, since the battery of a sensor is limited, a smart relocation scheme is required to maximize the lifetime of the sensors, and hence ensures a maximum lifetime of the barrier coverage.
## Problem Statement
This paper studies the two-dimensional (2-D) barrier coverage problem with mobile sensors, in which the barrier is modeled by a line segment, while the sensors are distributed in the plane initially. The problem is to compute the relocated positions of the sensors, such that the barrier will be completely covered while the maximum relocation distance among all the sensors is minimized. Formally, we are given a line barrier \([0,M]\) on \(x\)-axis and a set of sensors distributed in the Euclidean plane, say \(\Gamma=\{1,\,\dots,\,n\}\), within which sensor \(i\) is with a radii \(r_{i}\) and a position \((x_{i},\,y_{i})\). The two dimensional Minimum Maximum Sensor Movement problem (MMSM) is to compute the minimum \(D\in\mathbb{Z}_{0}^{+}\) and a new position \((x_{i}',\,0)\) for each sensor \(i\), such that \(\max_{i\in n}\sqrt{(x_{i}-x_{i}')^{2}+(y_{i}-0){}^{2}}\leq D\) and each point on the line barrier is covered by at least one sensor (i.e. for each point on the line barrier there exists at least a sensor \(i\) within distance \(r_{i}\)). The paper finds that, MMSM can be reduced to a discrete version called DMMSM. In DMMSM, we are given a graph \(G=(V,\,E)\), where \(V=\{v_{0},\,v_{1},\,\dots,\,v_{M}\}\) and \(E=\{e_{i}=(v_{i},\,v_{i+1})\vert i\in V\setminus\{M\}\}\). We say an edge \(e_{j}\) is covered by a set of sensors \(\Gamma'\subseteq\Gamma\) if and only if every point on the edge is in the monitoring area of at least one sensor of \(\Gamma'\). The goal of DMMSM is also to compute a minimum maximum movement \(D\in\mathbb{Z}_{0}^{+}\) and the new relocate position for each sensor, such that every edge of \(E\) is covered by the sensors. We propose several algorithms that are actually first to solve decision MMSM and decision DMMSM, which is to determine, for a given the relocation distance bound \(D\), whether the sensors can be relocated within \(D\) to cover the line barrier.
## Related Works
The MMSM problem in 2D setting was first studied in , and shown strongly \({\cal NP}\)-complete for sensors with general sensing radii via a reduction from the 3-partition problem which is known strongly \({\cal NP}\)-complete. Later, an algorithm with a time complexity of \(O(n^{3}\log n)\) has been developed in for the problem where sensors are with identical sensing radii. In the same paper, an approximation algorithm with ratio \(\frac{y_{max}}{y_{min}}\) for general radii has also been developed, where \(y_{max}\) and \(y_{min}\) are respectively the maximum and minimum perpendicular relocation distance from the sensors to the barrier. To the best of our knowledge, there is no any other non-trivial approximation algorithm for MMSM with general radii. Unlike 2-D MMSM, the MMSM problem has been extensively studied and well understood in 1-D setting, in which the barrier are assumed to be a line segment on the same line where the sensors are initially located. Paper presented an algorithm which optimally solves 1D-MMSM for uniform radius and runs in time \(O(n^{2})\), by observing the order preservation property. The time complexity was improved to \(O(n\log n)\) later in paper , which also gave an \(O(n^{2}\log n)\) time algorithm for general radii. Recently, an \(O(n^{2}\log n\log\log n)\) time algorithm has been presented in for weighted 1D-MMSM with uniform radii, in which each sensor has a weight, and the moving cost of a sensor is its movement times its weight. Moreover, circle/simple polygon barriers has been studied besides straight lines in , in which two algorithms has been developed for MMSM, with an \(O(n^{3.5}\log n)\) time against cycle barriers and an \(O(mn^{3.5}\log n)\) time against polygon barriers, where \(m\) is the number of the edges on the polygon. The later time complexity was then improved to \(O(n^{2.5}\log n)\) in. Other problems closely related to MMSM have also been well studied in previous literature. In 1-D setting, the Min-Sum relocation problem, to minimize the sum of the relocation distances of all the sensors, is shown \({\cal NP}\)-complete for general radii while solvable in time \(O(n^{2})\) for uniform radii. The Min-Num relocation problem of minimizing the number of sensors moved, is also proven \({\cal NP}\)-complete for general radii and polynomial solvable for uniform radii. Similar to MMSM, where a PTAS has been developed for the Min-Sum relocation problem against circle/simple polygon barriers , which was improved by later paper that gave an \(O(n^{4})\) time exact algorithm. Paper studied a more complicated problem of maximizing the coverage lifetime, in which each mobile sensor is equipped with limited battery power, and the coverage lifetime is the time to when the coverage no longer works because of the death of a sensor. The authors presented parametric search algorithms for the cases when the sensors have a predetermined order in the barrier or when sensors are initially located at barrier endpoints. On the other hand, the same authors present two FPTAS respectively for minimizing sumed and maximum energy consumption when the radii of the sensors can be adjusted. When the sensing radii is fixed, i.e. unadjustable, the same paper showed the min-sum problem can not be approximated within \(O(n^{c})\) for any constant \(c\) under the assumption of \({\cal P}\neq{\cal NP}\), while the min-max version is known strongly \({\cal NP}\)-complete, as it can be reduced to 2-D MMSM which is know strongly \({\cal NP}\)-complete. Before deployment of mobile sensors, barrier coverage was first considered deploying stationary sensors for covering a closed curve (i.e. a moat), and an elegant algorithm was proposed by transferring the Min-Sum cost barrier coverage problem to the shortest path problem. It has then been extensively studied for line based employment, for better local barrier coverage, and for using camera sensors. The most recent result studied line barrier coverage using sensors with adjustable sensing ranges. They show the problem is polynomial solvable when each sensor can only choose from a finite set of sensing ranges, and \({\cal NP}\)-complete if each sensor can choose any sensing ranges in a given interval.
## Our Results and Technique
In this paper, we present two approximation algorithms for the decision MMSM problem. The first is a simple greedy approach based on our proposed sufficient condition of determining whether there exists a feasible cover for the barrier under the relocation distance bound \(D\). If \(D<D^{*}\), the algorithm outputs "infeasible"; Otherwise, the algorithm computes new positions for the sensors, resulting a maximum relocation distance \(D+2r_{max}\), where \(r_{max}=\max_{i}\{r_{i}\}\). The algorithm is so efficient that it runs in time \(O(n\log n)\), where \(n=\vert\Gamma\vert\) is the number of the sensors. The second is generally an linear programming (LP) rounding based approach, which first transfers MMSM to the fractional cardinality matching problem and then solves the LP relaxation we propose for the latter problem. The algorithm approximately solves the decision MMSM problem according to a solution to the LP relaxation. Similar to the case for the first algorithm, we show that the algorithm always outputs "feasible" if \(D\geq D^{*}\). Further, for any instance our algorithm returns "feasible", we give a method to construct a real solution for MMSM, with a maximum relocation distance \(D+r_{max}\), by rounding up a fractional optimum solution to the LP relaxation. The algorithm has a runtime \(O(n^{7}L)\), which is exactly the time of solving the proposed LP relaxation by Karmakar's algorithm , where \(L\) is the length of the input. As a by-product, we give the third algorithm for decision MMSM with a maximum relocation distance \(d(OPT)+r_{max}\), and time \(O(R^{2}\sqrt{\frac{M}{\log R}})\), where \(R=\sum_{i=1}^{n}2r_{i}\) is sum of the radii of the sensors. Based on the three above algorithms for the decision problem, the paper proposes an unified algorithm framework to actually calculate a solution to MMSM without a given \(D\). The time complexity and the maximum relocation distance are respectively \(O(n\log n\log d_{\max})\) and \(D^{*}+2r_{max}\) if employing the greedy algorithm; \(O(n^{7}L\log d_{max})\) and \(D^{*}+r_{max}\) if employing the LP based algorithm, where \(d_{max}\) is the maximum distance between the sensors and the barriers. The runtime \(O(n^{7}L\log d_{max})\) can be improved to \(O(R^{2}\sqrt{\frac{M}{\log R}}\log d_{max})\) if \(M\) is not large, by using the third algorithm based on matching. Note that, although our algorithm could compute a near-optimal solution when \(D^{*}\gg r_{max}\), the performance guarantee is not as good when \(D^{*}<r_{max}\). So in addition we give a simple factor-2 approximation algorithm for MMSM, by extending the optimal algorithm for 1D-MMSM as in paper. Consequently, the ratio of our first three algorithms can be improved for the case \(D^{*}<r_{max}\), by combining the factor-2 approximation.
## Organization of the Paper
The remainder of the paper is organized as follows: For decision MMSM, Section 2 gives a greedy algorithm as well as the ratio proof; Section 3 gives an approximation algorithm with an improved maximum relocation distance \(d(OPT)+r_{max}\) using LP rounding technique; Section 4 gives another approximation algorithm with the same maximum relocation distance guarantee but a different runtime, by using maximum cardinality matching; Section 5 present the algorithm which actually solve MMSM, using the algorithm given in Section 2, 3 and 4; Section 6 extends previous results and develops a factor-2 approximation algorithm with provable performance guarantee; Section 7 concludes the paper.
# A Simple Greedy Algorithm For Decision MMSM
This section presents an approximation algorithm for any instance of decision MMSM wrt a given \(D\): if the algorithm returns "infeasible", then the instance is truly *infeasible* with respect to \(D\); Otherwise, the instance of MMSM is feasible under the maximum movement of \(D+2r_{max}\), where \(r_{max}=\max_{i}\{r_{i}\}\). To show the performance guarantee of the algorithm, we propose a sufficient condition for the feasibility of decision MMSM against given \(D\).
## An Approximation Algorithm
Let \([l_{i},\,g_{i}]\) be the possible coverage range for sensor \(i\), where \(l_{i}\) and \(g_{i}\) are respectively the leftmost and the rightmost points of the barrier, i.e. the leftmost and the rightmost points sensor \(i\) can cover by relocating within distance \(D\). The key idea of our algorithm is to cover the barrier from left to right, using the sensor with minimum \(g_{i}\) within the set of sensors which can cover the leftmost uncovered point with a maximum relocation distance \(D+2r_{max}\). More detailed, the algorithm is first to compute for each sensor \(i\) its possible coverage range \([l_{i},\,g_{i}]\). Let \(s\) be the leftmost point of the uncovered part of the line barrier. Then among the set of sensors \(\{i\vert l_{i}-2r_{max}\leq s\leq g_{i}\}\), the algorithm repeats selecting the sensor with minimum \(g_{i}\) to cover an uncovered segment of the line barrier starting at \(s\). Note that \(\{i\vert l_{i}-2r_{max}\leq s\leq g_{i}\}\) is exactly the set of sensors, which are with \(g_{i}\geq s\) and can monitor an uncovered segment starting at \(s\) by relocating at most \(D+2r_{max}\) distance. If there is a tie on \(g_{i}\), then randomly pick a sensor within the tie. The selection terminates once the line barrier is completely covered, or the instance is found infeasible, i.e. there exists no such \(i\) with \(l_{i}-2r_{max}\leq s\leq g_{i}\) while the coverage is not done. The algorithm is formally as in Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"}.
For briefness, we will simply say an instance (or the input) of MMSM instead for the input of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} in the following paragraphs. Note that Steps 2-3 take time \(O(n)\) to compute \(l_{i}\) and \(g_{i}\) for all the sensors, Steps 4-7 take \(O(n\log n)\) time to assign the sensors to cover the line barrier. Therefore, we have the time complexity of the algorithm:
## The Ratio of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"}
The performance guarantee of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} is as below:
According to Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"}, we never move a sensor out of the range \([l_{i}-2r_{max},\,g_{i}]\). It remains to show Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} will always return a feasible solution when \(D\geq D^{*}\). For this goal, we will give a sufficient condition for the feasibility of decision MMSM. Below are two notations needed for the tasks: \[\lambda(i,\,D,\,x,\,x')=\begin{cases} 0 & g_{i}\leq x\,\mbox{or}\,l_{i}\geq x'\\ \min\{2r_{i},\,\min\{x',\,g_{i}\}-\max\{x,\,l_{i}\}\} & Otherwise \end{cases}\] and \[\sigma(i,\,D,\,{\cal S})=\begin{cases} 0 & g_{i}\leq x\,\mbox{OR}\,l_{i}\geq x'\\ \min\{2r_{i},\,\sum_{[x,x']\in{\cal S}}\lambda(i,\,D^{*},\,x,\,x')\} & Otherwise \end{cases}\] where \({\cal S}\) is a set of *disjoint* segments of the line barrier. Intuitionally \(\lambda(i,\,D,\,x,\,x')\) is the maximum coverage which sensor \(i\) can provide for segment \([x,\,x']\), and \(\sigma(i,\,D,\,{\cal S})\) is the sum of the coverage that sensor \(i\) can provide for all the segments in \({\cal S}\). Then a simple necessary condition for the feasibility of an instance of MMSM of as below:
Intuitionally, the above proposition states that the sum of the sensor coverage length must be not less than the length of the barrier segments to cover. The correctness of the above lemma is obviously, since a feasible relocation assignment must satisfy the condition. However, this is not a sufficient condition for the feasibility of decision MMSM (A counter example is as depicted in Figure [\[fig:A-tight-exampleforLP\]](#fig:A-tight-exampleforLP){reference-type="ref" reference="fig:A-tight-exampleforLP"} (a): For \(D=r_{1}\), the necessary condition holds for the given instance while the instance is actually infeasible). However, if \(2r_{max}\) more relocation distance is allowed as in Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"}, we have the following lemma:
Now we will prove Theorem [\[thm:greedyfinalthr\]](#thm:greedyfinalthr){reference-type="ref" reference="thm:greedyfinalthr"}. If \(D\geq D^{*}\), then the decision MMSM is feasible, and hence following Proposition [\[prop:necessary\]](#prop:necessary){reference-type="ref" reference="prop:necessary"} \(\sum_{i}\sigma(i,\,D,\,{\cal S})\geq\sum_{[x_{j},\,x_{j}']\in{\cal S}}\vert x_{j}'-x_{j}\vert\) holds for every \({\cal S}\) at the beginning of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"}. Then from Lemma [\[lem:sufficientcoreforgreedy\]](#lem:sufficientcoreforgreedy){reference-type="ref" reference="lem:sufficientcoreforgreedy"}, we immediately have the instance is feasible under relocation distance bound \(D\), which completes the proof of Theorem [\[thm:greedyfinalthr\]](#thm:greedyfinalthr){reference-type="ref" reference="thm:greedyfinalthr"}.
As the example depicted in Figure [\[fig:A-tight-maingreedalgorithm\]](#fig:A-tight-maingreedalgorithm){reference-type="ref" reference="fig:A-tight-maingreedalgorithm"}, \(d(OPT)=2r_{1}\) while the output of the algorithm is with a maximum relocation distance \(2r_{2}\). So when \(r_{2}\gg r_{1}\), \(d(OPT)+2r_{max}\thickapprox2r_{1}+2r_{2}\), and hence the analysis of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} is nearly tight in Theorem [\[thm:greedyfinalthr\]](#thm:greedyfinalthr){reference-type="ref" reference="thm:greedyfinalthr"}.
# An LP-based Approximation for Decision MMSM
This section will give an LP-based approximation algorithm to determine whether an instance of the decision MMSM problem is feasible. The algorithm first transfers the instance to a corresponding instance of decision DMMSM, and then an instance of the fractional cardinality matching problem with a proposed LP relaxation. Our algorithm answers "feasible" or "infeasible" according to the computed optimum solution of the LP relaxation. We show that if our algorithm returns *feasible*, then a solution to MMSM can be constructed under the maximum relocation distance \(D+r_{max}\) by rounding up a fractional optimum solution to the relaxation.
## Transferring to an Instance of DMMSM
The key idea of the transfer is first to compute \(l_{i}\) and \(g_{i}\) for each \(i\) wrt to the given \(D\), and then add \((l_{i},\,0)\) and \((g_{i},\,0)\) to the barrier as two vertices on the line. That is, \(V=\{(l_{i},\,0)\vert i\in[n]^{+}\}\cup\{(g_{i},\,0)\vert i\in[n]^{+}\}\). W.l.o.g. assume that the vertices of \(V\) appear on the line barrier in the order of \(v_{0},\,\dots,\,v_{|V|}\), from left to right on the line barrier. Then the algorithm adds an edge between every pair of \(v_{i}\) and \(v_{i+1}\). So \(E=\{e_{i}(v_{i},\,v_{i+1})\vert i\in[\vert V\vert-1]^{+}\}\). Formally the transfer is as in Algorithm [\[alg:The-transfering-algorithmfromMMSMtoDMMSM\]](#alg:The-transfering-algorithmfromMMSMtoDMMSM){reference-type="ref" reference="alg:The-transfering-algorithmfromMMSMtoDMMSM"}.
For the time complexity and the size of the graph, we have:
According to the algorithm, \(\vert V\vert=O(n)\) and \(\vert E\vert=O(n)\) hold trivially. Algorithm [\[alg:The-transfering-algorithmfromMMSMtoDMMSM\]](#alg:The-transfering-algorithmfromMMSMtoDMMSM){reference-type="ref" reference="alg:The-transfering-algorithmfromMMSMtoDMMSM"} takes in \(O(n\log n)\) time to sort (number) the vertices of \(V\) in Step 5, since \(\vert V\vert=O(n)\). Other steps of the algorithm takes trivial time compared to the sorting. So the total runtime of Algorithm [\[alg:The-transfering-algorithmfromMMSMtoDMMSM\]](#alg:The-transfering-algorithmfromMMSMtoDMMSM){reference-type="ref" reference="alg:The-transfering-algorithmfromMMSMtoDMMSM"} is \(O(n\log n)\).
## Fractional Maximum Cardinality Matching wrt DMMSM
Let \(\vert e_{i}\vert=v_{i+1}-v_{i}\) be the length of edge \(e_{i}\). Assuming that \(v_{i_{1}}=l_{i}\) and \(v_{i_{2}}=g_{i}\), we set \(J_{i}=\{e_{j}\vert j=i_{1},\,i_{1}+1,\,\dots,\,i_{2}-1\}\). Then the linear programming relaxation (LP1) for DMMSM is as below: \[\max\sum_{i\in[n]^{+}}\sum_{j\in J_{i}}x_{i,\,j}\] subject to \[\begin{aligned}
\sum_{j\in J_{i}}x_{i,j} & \leq2r_{i} & \forall i\in\{1,\,\dots,\,n\}\label{eq:sensoratmostone}\\ \sum_{i\in[n]^{+}}x_{i,\,j} & \leq\vert e_{j}\vert & \forall j\in\{1,\,\dots,\,\vert V\vert\}\label{eq:coverrequirement}\\ 0\leq x_{i,\,j} & \leq\vert e_{j}\vert & \forall i\in\{1,\,\dots,\,n\},\,j\in J_{i}\label{eq:capacitybd} \end{aligned}\] where \(\sum_{j\in J_{i}}x_{i,j}\leq2r_{i}\) is because a sensor \(i\) can at most cover length \(2r_{i}\) of the barrier, and Inequality ([\[eq:coverrequirement\]](#eq:coverrequirement){reference-type="ref" reference="eq:coverrequirement"}) is because the covered length of each edge \(e_{i}\) (segment) is at most \(\vert e_{j}\vert\). Our algorithm determines decision DMMSM according to the computed optimum solution, say \(\mathbf{x}\), to LP1: the algorithm outputs "feasible" if \(\sum_{i\in[n]^{+}}\sum_{j\in J_{i}}x_{i,\,j}=M\), and outputs "infeasible" otherwise.
It is known that there exist polynomial-time algorithms for solving linear programs. In particular, using Karmakar's algorithm to solve LP1 will take \(O(n^{7}L)\) time, since there are \(O(n^{2})\) variables totally in LP1.
It is worth to note that the simplex algorithm has a much better practical performance than Karmakar's algorithm. So using the simplex method the algorithm would be faster than \(O(n^{7}L)\) in real world applications. The performance guarantee of Algorithm [\[alg:mainLProundDMMSM\]](#alg:mainLProundDMMSM){reference-type="ref" reference="alg:mainLProundDMMSM"} is as given in the following theorem, whose proof will be given in next subsection.
From Figure [\[fig:A-tight-exampleforLP\]](#fig:A-tight-exampleforLP){reference-type="ref" reference="fig:A-tight-exampleforLP"}, an optimal fractional solution to LP1 is with a maximum relocation distance \(r_{1}\), while a true optimal solution for MMSM must be with a maximum relocation distance \(r_{2}\). Thus, the integral gap for LP1 is \(r_{2}-r_{1}\approx r_{max}\) when \(r_{2}\gg r_{1}\). That is, for any fixed \(\epsilon>0\), the maximum relocation distance increment could be larger than \(r_{max}-\epsilon\) for rounding an optimum solution of LP1 to a true solution of MMSM. Therefore, the ratio is nearly-tight for Algorithm [\[alg:mainLProundDMMSM\]](#alg:mainLProundDMMSM){reference-type="ref" reference="alg:mainLProundDMMSM"}.
## Proof of Theorem [\[thm:LPbasedratio\]](#thm:LPbasedratio){reference-type="ref" reference="thm:LPbasedratio"}
This subsection will prove Theorem [\[thm:LPbasedratio\]](#thm:LPbasedratio){reference-type="ref" reference="thm:LPbasedratio"} by showing a fractional optimum solution to LP1 can be rounded up to an integral solution of DMMSM with a maximum relocation distance \(D+r_{max}\). Let \(\mathbf{x}\) be an optimum solution to LP1. Recall that \(x_{ij}\) is the (fractional) portion sensor \(i\) covering edge \(j\), the key idea of our algorithm is to aggregate the portions of sensor \(i\) covering different \(j\)s, such that the portions combine to a line segment which \(i\) can cover within movement \(D+r_{max}\). Our algorithm is composed by two parts. The first part is called *pre-aggregation* which rounds \(x_{ij}\) to \(1\) in a "pseudo" way. More precisely, assume that \(0<x_{i_{1},\,j},\,\dots,\,x_{i_{h},\,j}<0\) are the variables shares edge \(e_{j}\). Then the *pre-aggregation* divides \(e_{j}\) to a set of sub-edges \(\{e_{j,\,i_{1}},\,\dots,e_{j,\,i_{h}}\}\), in which \(\vert e_{j,\,i_{l}}\vert=\vert e_{j}\vert\cdot x_{i_{l},\,j}\). We set \(x_{i_{l},\,j,\,l}=1\), which is, edge \(e_{j,\,i_{l}}\) completely covered by sensor \(i_{l}\). Let \({\cal S}_{i}\) be the set of sub-edges covered by sensor \(i\) accordingly. The second part, which is called *aggregation,* aggregates the edges of \({\cal S}_{i}\) for each \(i\) such that the edges covered by an identical sensor will connect together. The *aggregation* starts from the following simple observation whose correctness is obviously:
The key observation of the *aggregation* is that if case (2) and (3) in the above proposition can be eliminated, then the set of edges in \({\cal S}_{i}\) are aggregated together that they can be truly monitored by sensor \(i\). So the *aggregation* is accordingly composed by two phases called the *swap phase* and the *exchange* step, which are to eliminate case (2) and (3), respectively. Formally, the algorithm is as in Algorithm [\[alg:rounding\]](#alg:rounding){reference-type="ref" reference="alg:rounding"} (An example of such rounding depicted as in Figure [\[fig:ori\]](#fig:ori){reference-type="ref" reference="fig:ori"} and [\[fig:exe\]](#fig:exe){reference-type="ref" reference="fig:exe"}).
Note that if two adjacent sub-edges belong to the same \({\cal S}_{i}\), say \(e_{j_{1}},\,e_{j_{2}}\in{\cal S}_{i}\) with \(v_{j_{1}+1}=v_{j_{2}}\), then we can combine the two sub-edges as one, since they are segments both covered by sensor \(i\). So Step 4 of Algorithm [\[alg:rounding\]](#alg:rounding){reference-type="ref" reference="alg:rounding"} is actually to set \({\cal S}_{i}:={\cal S}_{i}\setminus\{e_{j_{1}},\,e_{j_{2}}\}\cup e(v_{j_{1}},\,v_{j_{2}+1})\) for every such pair of adjacent edges of \({\cal S}_{i}\) for every \(i\). The *swap* phase, as in Step 3 of Algorithm [\[alg:rounding\]](#alg:rounding){reference-type="ref" reference="alg:rounding"}, will eliminate case (2) (i.e. \(j_{1}<j_{1}'<j_{2}<j_{2}'\)) by swapping the coverage sensors of the edges without causing any increment on the maximum relocation distance. The observation inspiring the swap is that for any \(j_{1}<j_{1}'<j_{2}<j_{2}'\), we can swap the two sensors covering \(j_{1}'\) and \(j_{2}\) without increasing the maximum relocation distance. More precisely, we cover a portion of \(\min\{\vert j_{1}'\vert,\,\vert j_{2}\vert\}\) of \(j_{1}'\) with \(i\) and cover a portion of \(\min\{\vert j_{1}'\vert,\,\vert j_{2}\vert\}\) of \(j_{2}\) with \(i'\). The formal layout of the *swap phase* is as in Algorithm [\[alg:swap\]](#alg:swap){reference-type="ref" reference="alg:swap"}.
Note that Steps 8-13 will add new vertices and edges to the graph, so \(\sum_{i=1}^{n}\vert{\cal S}_{i}\vert\) may increases. However, we can always guarantee \(\sum_{i=1}^{n}\vert{\cal S}_{i}\vert\leq n\vert E\vert\). Since otherwise, following the pigeonhole principle, there must exist two sub-edges, say \(e_{j_{1}}\) and \(e_{j_{2}}\) which are in an identical \({\cal S}_{i}\) and within the range of an identical edge of \(E\). Then, such two sub-edges \(e_{j_{1}},\,e_{j_{2}}\) can be combined as one, by setting \(x(v_{j_{1}+1})=x(v_{j_{1}+1})+\vert e_{j_{2}}\vert\) and move every sub-edge between \(e_{j_{1}}\) and \(e_{j_{2}}\) to right with an offset with length \(\vert e_{j_{2}}\vert\). Clearly, following the meaning of a sensor covering edges as in the definition of DMMSM, these movement does not cause any increment on the relocation distance of each sensor. In Algorithm [\[alg:swap\]](#alg:swap){reference-type="ref" reference="alg:swap"}, the *while*-loop iterates for \(O(n)\) times and the outer *for*-loop iterates for \(O(n)\) times. Since the inner *for*-loop iterates for at most \(O(\sum_{i=1}^{n}\vert{\cal S}_{i}\vert)\) times. Then from \(\sum_{i=1}^{n}\vert{\cal S}_{i}\vert\leq n\vert E\vert=O(n^{2})\), the *for-loops iterates at most \(O(n^{2})\) time.* So we have the total runtime of the *swap phase*:
For the correctness of Algorithm [\[alg:swap\]](#alg:swap){reference-type="ref" reference="alg:swap"}, we have the following lemma:
The *exchange* phase, invoked in Step 5 in Algorithm [\[alg:rounding\]](#alg:rounding){reference-type="ref" reference="alg:rounding"}, is to eliminate case (3) (i.e. \(j_{1}<j_{1}'<j_{2}'<j_{2}\)). The key idea of the exchange is to move \(j_{1}\) to the place exactly before \(j_{2}\) (or to move \(j_{2}\) to the place exactly after \(j_{1}\)), and then move the edges between \(j_{1}\) and \(j_{2}\) accordingly. The choosing of movements (move \(j_{1}\) to \(j_{2}\), or \(j_{2}\) to \(j_{1}\)) depends on the current *offset* of the sub-edges between \(j_{1}\) and \(j_{2}\), as well as the length of the edge \(j_{1}\) and the length sum of the other edges in \({\cal S}_{i}\setminus\{j_{1}\}\), where the *offset* of an sub-edge \(j\) is the distance from the current position of \(j\) to its original position. Formally, the *exchange* phase is as in Algorithm [\[alg:exchange\]](#alg:exchange){reference-type="ref" reference="alg:exchange"}.
In Algorithm [\[alg:exchange\]](#alg:exchange){reference-type="ref" reference="alg:exchange"}, the function Mover (\({\cal E}\), \({\cal O}\), \(i\), \(i+\delta\)) actually decides whether to move \(C_{i+\delta}\) or \(C_{i}\), according to which one of the two values \(\vert C_{i}\vert-o_{i}\) and \(\sum_{j:\,C_{j}\in{\cal S}_{j_{i}}\setminus\{C_{i}\}}\vert C_{j}\vert+o_{i}\) is larger. Intuitionally, without considering offsets, if we move \(C_{i}\) then the moving distance of \(C_{l}\) between \(C_{i}\) and \(C_{i+\delta}\) will be \(\vert C_{i}\vert\); if move \(C_{i+\delta}\), the moving distance will be \(\sum_{j:\,C_{j}\in{\cal S}_{j_{i}}\setminus\{C_{i}\}}\vert C_{j}\vert\) instead of \(\vert C_{i+\delta}\vert\), since not only \(C_{i+\delta}\) but every \(C_{j}\in{\cal S}_{j_{i}}\setminus\{C_{i}\}\) will be moved to adjacent to \(C_{i}\). Then considering the offsets, we have the criteria of moving \(C_{i}\) or \(C_{i+\delta}\). The moving algorithm is as in Algorithm [\[alg:mover\]](#alg:mover){reference-type="ref" reference="alg:mover"}.
# A Matching-Based Solution to Decision MMSM
This subsection gives a pseudo polynomial algorithm for decision MMSM. The key idea of the algorithm is to consider MMSM as DMMSM with uniform edge length, where the barrier to cover is composed by \(M\) edges of length one. Then our algorithm is similar to the case in Section 3, excepting that we use maximum cardinality matching instead of fractional cardinality matching to compute an initial solution. Using a similar algorithm as in Subsection 3.3, we can round the initial solution, i.e. the maximum cardinality matching, to a solution to MMSM. To model a given instance of MMSM as maximum cardinality matching, we will construct an equivalent bipartite graph \(H=(U,\,V,\,E)\) in which the vertex set \(V\) corresponds to the sensors, the vertex set \(U\) corresponds to the edge set \(\cup_{i}J_{i}\), where \(J_{i}\) contains exactly the edges that can be completely sensed by sensor \(i\) within maximum movement \(D\), and the edge set \(E_{H}\) corresponds to the coverage of the sensors to the edge of \(\cup_{i}J_{i}\). Then we check whether there exists a maximal cardinality matching with size \(M\) in \(H\). If no such matching exists, the instance of DMMSM is infeasible under maximum relocation distance \(D\). Otherwise, similarly as in Subsection 3.3, we can aggregate the vertices of \(U\) that are fractionally covered by an identical sensor, such that the sensors can relocate within distance \(D+r_{max}\) to cover all the edges. The formal layout of the algorithm is as in Algorithm [\[alg:mainMatchingbasedalgorithm.\]](#alg:mainMatchingbasedalgorithm.){reference-type="ref" reference="alg:mainMatchingbasedalgorithm."}.
Firstly and apparently, if the DMMSM is feasible wrt a given \(D\), there must exist a matching with size \(M\) in the corresponding graph \(H\). So if no such matching exists wrt \(D\), then the instance DMMSM must be infeasible wrt \(D\). Secondly, similar to Algorithm [\[alg:rounding\]](#alg:rounding){reference-type="ref" reference="alg:rounding"}, the computed matching \({\cal Y}\) can be rounded to a true solution to MMSM using *swap* and *exchange*.
# The Complete Algorithm for Solving MMSM
The key idea of computing approximately a minimum \(D\) is to use binary search and call Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} (Or equivalently Algorithm [\[alg:mainLProundDMMSM\]](#alg:mainLProundDMMSM){reference-type="ref" reference="alg:mainLProundDMMSM"} or Algorithm [\[alg:mainMatchingbasedalgorithm.\]](#alg:mainMatchingbasedalgorithm.){reference-type="ref" reference="alg:mainMatchingbasedalgorithm."}) as a subroutine. Let \(d_{i}\) be the minimum distance between sensor \(i\) and any point on the line barrier, and \(d_{max}\) be the maximum distance between the sensors and the barriers. Then clearly every sensor of \(\Gamma\) can cover any point of the barrier within movement distance \(d_{max}\). That is, within movement distance \(d_{max}\) the line barrier can be covered successfully by the sensors of \(\Gamma\); Or the sensors in \(\Gamma\) is not enough to cover the barrier. Then, to find the minimum \(D\) for MMSM, we need only to use binary search within the range from 0 to \(d_{max}\). Apparently, this takes at most \(O(\log(d_{max}))\) calls of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} (Or [\[alg:mainLProundDMMSM\]](#alg:mainLProundDMMSM){reference-type="ref" reference="alg:mainLProundDMMSM"}, [\[alg:mainMatchingbasedalgorithm.\]](#alg:mainMatchingbasedalgorithm.){reference-type="ref" reference="alg:mainMatchingbasedalgorithm."}) to find the min-max feasible relocation distance bounded \(D_{SOL}\). Formally, the complete algorithm is as in Algorithm [\[alg:greeedybasedThe-complete-algorithm1\]](#alg:greeedybasedThe-complete-algorithm1){reference-type="ref" reference="alg:greeedybasedThe-complete-algorithm1"}.
From Theorem [\[thm:greedyfinalthr\]](#thm:greedyfinalthr){reference-type="ref" reference="thm:greedyfinalthr"}, we immediately have the time complexity and ratio for the algorithm as follows:
Note that, if using Algorithm [\[alg:mainLProundDMMSM\]](#alg:mainLProundDMMSM){reference-type="ref" reference="alg:mainLProundDMMSM"} instead of Algorithm [\[alg:mainA-simple-greedy\]](#alg:mainA-simple-greedy){reference-type="ref" reference="alg:mainA-simple-greedy"} in Step 8, the runtime and the maximum relocation distance of Algorithm [\[alg:greeedybasedThe-complete-algorithm1\]](#alg:greeedybasedThe-complete-algorithm1){reference-type="ref" reference="alg:greeedybasedThe-complete-algorithm1"} will be \(O(n^{7}L)\) and \(D_{SOL}=D+r_{max}\), respectively.
# A Simple Factor-2 Approximation Algorithm for MMSM
Following paper, MMSM is solvable in time \(O(n^{2}\log n)\) if all the sensors are on the line containing the barrier. Then a natural idea to solve 2D-MMSM is firstly to perpendicularly move (some of) the sensors to the line barrier, and secondly solve the consequent 1D-MMSM by using the algorithm in paper. Let \(d_{p,\,i}\) be the perpendicular distance between sensor \(i\) and the line barrier. Without loss of generality we assume that \(0=d_{p,\,0}\leq d_{p,\,1}\leq\dots\leq d_{p,\,i}\leq\dots\leq d_{p,\,n}\), where sensor \(0\) is a virtual sensor with radii 0. Let \(S(d_{p,\,i})=\{j\vert d_{p,\,j}\leq d_{p,\,i}\}\) be the set of sensors whose perpendicular distance to the line barrier is not larger than \(d_{p,\,i}\). Let \(D_{h}(S(d_{p,\,i}))\) be the maximum *horizontal* relocation distance of the sensors in \(S(d_{p,\,i})\) covering the barrier. Our algorithm will first simply compute \(D_{h}(S(d_{p,\,i}))\) for every \(i\in[n]^{+}\), and then select \(\min\{d_{p,\,i}+D_{h}(S(d_{p,\,i}))\vert i\in[n]^{+}\}\) as the maximum relocation distance.
Clearly, the above naive algorithm has to run the 1D-MMSM algorithm for \(O(n)\) times to compute \(\min\{d_{p,\,i}+D_{h}(S(d_{p,\,i}))\vert i\in[n]^{+}\}\). Hence, it runs in time \(O(n^{3}\log n)\). Note that the binary search cannot be immediately applied here, since \(f(i)=d_{p,\,i}+D_{h}(S(d_{p,\,i}))\) is neither monotonously increasing nor monotonously decreasing on \(i\). Anyhow, we will give an improve algorithm in which the number of times of solving 1D-MMSM is improved to \(O(\log n)\). We will show the ratio of the improved algorithm remains two by giving further observations. The key idea of our improved algorithm is, instead of finding an \(i\) such that \(d_{p,\,i}+D_{h}(S(d_{p,\,i}))\) attends minimum, to find an \(i_{alt}\), such that \(D_{h}(S(d_{p,\,i_{alt}})\leq d_{p,\,i_{alt}}\) and \(D_{h}(S(d_{p,\,i_{alt}-1})\geq d_{p,\,i_{alt}-1}\) both hold. Note that such \(i_{alt}\) can be found with solving the 1D-MMSM problem only for \(O(\log n)\) times, via a binary search in which a set of values \([lwr,\,upp]\), \(lwr\leq i_{alt}\leq upp\), is maintained. The formal layout of the algorithm is as in Algorithm [\[alg:factor2apprx\]](#alg:factor2apprx){reference-type="ref" reference="alg:factor2apprx"}.
The basic observation of our algorithm is that \(f(i)=D_{h}(S(d_{p,\,i}))\) will not increase while \(i\) increases.
The correctness of the above proposition immediately follows from the fact that \(S(d_{p,\,i})\supseteq S(d_{p,\,i-1})\). Then the performance guarantee of the algorithm is as in the following lemma:
Since Algorithm [\[alg:factor2apprx\]](#alg:factor2apprx){reference-type="ref" reference="alg:factor2apprx"} iterates the *while-loop* for at most \(O(\log n)\) times, we have the following theorem:
# Conclusion
This paper developed three algorithms for MMSM via solving decision MMSM, which are respectively with runtime \(O(n\log n \log d_{max})\), \(O(n^{7}L\log d_{max})\) and \(O\left(R^{2}\sqrt{\frac{M}{\log R}}\log d_{max}\right)\), and maximum relocation distance \(d(OPT)+2r_{max}\), \(d(OPT)+r_{max}\) and \(d(OPT)+r_{max}\), where \(n\) is the number of sensors, \(L\) is the length of input, \(M\) is the length of the barrier, \(d(OPT)\) is the maximum relocation distance in an optimum solution to MMSM, \(d_{max}\) is the maximum distance between the sensors and the barriers, and \(r_{max}=\max_{i}\{r_{i}\}\) is the maximum sensing radii of the sensors. We proved the performance guarantee for the first algorithm by giving a sufficient condition to check the feasibility of an instance of decision MMSM, and for the second (and hence the third) algorithm by rounding up an optimum fractional solution against the according LP relaxation, to a real solution of DMMSM. To the best of our knowledge, our method of rounding up a fractional LP solution is the first to round an LP solution by aggregation, and has the potential to be applied to solve other problems. In addition, we developed a factor-2 approximation by extending a previous result in paper. Consequently, the performance of our first three algorithms can be improved when \(r_{max}>d(OPT)\). We note that our proposed algorithms can only work for MMSM with only one barrier, and are currently investigating approximation algorithms for MMSM with multiple barriers. | {'timestamp': '2017-06-07T02:03:39', 'yymm': '1706', 'arxiv_id': '1706.01623', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01623'} |
# Introduction {#s:intro}
Doping of organic semiconductors (OSC) by introduction in the host matrix of strong electron-or hole-donating molecules has been shown to increase their electrical conductivity by orders of magnitude, leading to enhanced performances in organic light-emitting devices and photovoltaic cells. However, in contrast to inorganic semiconductors where doping is understood to proceed via the formation of shallow dopant levels, the case of OSC remains controversial . Fundamental questions regarding the electronic structure of doped OSC and the evolution of the transport and optical properties with doping load are still open and it is presently unclear why very large dopant concentrations (a few percents) are needed to boost their electrical conductivity. Experimental observations by Koch and coworkers point to contrasting pictures: Full dopant-OSC charge transfer (CT) seems to be the paradigm for conjugated polymers, while spectroscopic evidences of partial CT, or dopant-OSC orbital hybridization, have been reported for different molecular systems. Focusing on the paradigmatic case of 2,3,5,6-tetrafluoro-7,7,8,8-tetracyanoquinodimethane (F4TCNQ) as a hole dopant in bulk pentacene (PEN, see Figure [\[fig0\]](#fig0){reference-type="ref" reference="fig0"}) , it was shown that the introduction of the molecular dopant does not lead to any ultraviolet photoelectron spectroscopy (UPS) evidence of singly-occupied levels in the pentacene gap, as expected according to the standard polaronic picture for polymers. Doping results instead in the emergence of two novel optical absorption lines at \(\sim\)`<!-- -->`{=html}1.2-1.4 eV, located below the 1.8 eV pentacene absorption onset. These features have been ascribed to CT transitions pertaining to pentacene-F4TCNQ complexes on the basis of gas-phase density functional theory (DFT) calculations performed on a cofacial dimer . The emerging picture is that of strongly interacting molecular host-dopant pairs resulting in only *partially ionized* dopants. On the other hand, the scanning tunneling microscopy (STM) images by Ha and Kahn instead show that isolated F4TCNQ in pentacene films are *fully ionized* at room temperature, an observation that has been rationalized on the basis of electrostatic modeling. In this paper, we revisit the case of F4TCNQ-doped pentacene in the low-doping regime and analyze its electronic and optical properties with a combination of many-body *ab initio* and model Hamiltonian electronic structure calculations, both explicitly accounting for electron-hole correlations. We show that despite very deep acceptor levels in the pentacene gap, electron-hole interactions result in thermally accessible states with fully ionized F4TCNQ dopants. The broader picture obtained from our correlated electron-hole model describes doping as a competition between neutral and ionized dopants, passing through a narrow window of fractional CT. Our electronic structure calculations locate the pentacene-F4TCNQ system across this boundary, with structural molecular relaxation (polaronic effects) collapsing the system towards the full ionization of the dopant. The optical absorption signatures of CT vs. ionized species in F4TCNQ-doped pentacene and other doped OSC are analyzed in the light of recent experiments. Our computational approach relies on state-of-the-art *ab initio* electronic structure calculations using Green's function many-body perturbation theories within the \(GW\) and Bethe-Salpeter equation (BSE) formalisms. Extensive benchmarks against reference quantum-chemical calculations have demonstrated the accuracy of these approaches for calculating quasiparticle energy levels and optical excitation energies, BSE properly accounting for the long-range electron-hole interactions crucial for CT excitations. The \(GW\) formalism is embedded in a recently developed hybrid quantum/classical (QM/MM) scheme where many-body effects in the QM system are combined with an accurate discrete polarizable model accounting for the dielectric screening by the MM environment, known to largely affect the energies of charged and CT excitations. This approach that proved to reproduce accurately the experimental photoemission gap and bulk (periodic) \(GW\) calculations in pristine pentacene is here extended to optical excited states within the BSE framework. The paper is organized as follows. We first present in Section [2](#s:methods){reference-type="ref" reference="s:methods"} the embedded *ab initio* many-body formalism used, and the complementary model Hamiltonian that allows us to explore finite size effects and polaronic coupling. Our results are presented in Section [3](#s:results){reference-type="ref" reference="s:results"}, followed by a discussion in Section [4](#s:disc){reference-type="ref" reference="s:disc"}. After the conclusions and perspectives (Section [5](#s:conclu){reference-type="ref" reference="s:conclu"}), we gather in an Appendix convergence and validation tests performed on a small but representative subsystem.
# Methods {#s:methods}
## Technical details of embedded many-body *ab initio* calculations {#ss:details}
Our analysis relies on first-principles many-body calculations within the framework of the hybrid QM/MM formalism presented in Ref. . Our \(GW\) and Bethe-Salpeter (BSE) calculations are performed with the Fiesta package that relies on Gaussian bases and Coulomb-fitting resolution-of-the-identity (RI-V) techniques. We adopt the triple-zeta plus polarization 6-311G(d) atomic basis set combined with the Weigend Coulomb fitting auxiliary basis set, ensuring convergence values of the photoemission gap and excitation energies within a few meV (see Appendix [\[s:app\]](#s:app){reference-type="ref" reference="s:app"}). Input Kohn-Sham orbitals are generated by the NWChem package modified so as to provide the DFT Kohn-Sham eigenstates and exchange-correlation matrix elements in the Gaussian atomic basis. Our \(GW\) calculations are performed at the partially self-consistent (ev\(GW\)) level with update of the eigenvalues. Such an approach leads to improved accuracy for charged and neutral excitation energies and cures the impact of the dramatic dependency of the starting Kohn-Sham HOMO-LUMO gap on the starting functional, as already documented for model donor-acceptor complexes. For sake of consistency, we start with an hybrid functional PBEh(\(\alpha\)=0.4), namely a functional with 40% of exact exchange, selected so as to match the starting Kohn-Sham gap with the \(GW\) one (see below). BSE calculations are performed at the Tamm-Dancoff (TDA) level that produces accurate energies for CT excitations. The MM region is described by the charge response (CR) model by Tsiper and Soos in its [MESCAL]{.smallcaps} code implementation. This approach describes the anisotropic molecular response to electric fields in terms of induced atomic charges and induced dipoles, providing an accurate description of the static dielectric tensor of molecular solids. The molecular polarizability tensor is computed at DFT level \[B3LYP functional, 6-311++G(d,p) basis set\] and atom-atom polarizabilities governing intramolecular charge flows are evaluated with semiempirical Hartree-Fock calculations (ZINDO parametrization). Full technical details concerning the stability of our results with the starting DFT functional, convergency tests and comparison between full BSE and TDA calculations, are discussed in Appendix [\[s:app\]](#s:app){reference-type="ref" reference="s:app"}.
## Model Hamiltonian for molecular doping {#s:model}
Our accurate many-body *ab initio* analysis is complemented with a generalized Mulliken model for intermolecular CT that allows us to describe systems of larger size and introduce structural relaxation (polaronic) effects. For a single dopant (DOP) in the lattice of a host OSC we can represent the Hamiltonian on the basis of the neutral state \(|\mathcal{N}\rangle\) and singlet-coupled full-CT states \(|i_m\rangle\) with the electron populating the dopant LUMO and the hole in the HOMO\(-m\) orbital (\(m=0,1\)) of the OSC site \(i\). The electronic Hamiltonian reads: \[\begin{aligned}
H&=& \sum_{i,m}^{i\in \mathrm{OSC}} \varepsilon^\mathrm{CT}_{im} |i_m\rangle \langle i_m | + \sum_{i,m}^{i\in \mathrm{OSC}} t_{im}^\mathrm{CT} \left( |\mathcal{N} \rangle \langle i_m | + |i_m \rangle \langle \mathcal{N} | \right ) \nonumber \\ &+& \sum_{i,j}^{i,j\in \mathrm{OSC}} \sum_{m,n} t^\mathrm{h}_{im,jn} \left( |i_m \rangle \langle j_n | + |j_n\rangle \langle i_m | \right ), \label{e:H0} \end{aligned}\] where \(\varepsilon^\mathrm{CT}_{im}\) are CT states energies, \(t_{im}^\mathrm{CT}\) is the DOP-OSC charge transfer integral and \(t^{h}_{ij}\) is the hole transfer integral between OSC sites. Triplet excitations are described by the same Hamiltonian restricted to the subspace of CT states \(|i_m\rangle\) with triplet spin pairing. Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} describes Coulombically-bonded and possibly delocalized electron-hole pairs and can be considered an extension of similar models successfully applied to the description of intra- and intermolecular CT. Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} applies to doping in molecular or polymer OSC and is here accurately parametrized from first principles for F4TCNQ-doped pentacene. Diabatic CT states energies can be written as: \(\varepsilon^\mathrm{CT}_{im}=E^\mathrm{PEN}_{\mathrm{HOMO}-m}+E^\mathrm{DOP}_{\mathrm{LUMO}}+ P^\pm_{im}\). Molecular orbital energies are calculated from gas-phase ev\(GW\) calculations that yield \(E^\mathrm{PEN}_{\mathrm{HOMO}-1}=7.84\) eV, \(E^\mathrm{PEN}_{\mathrm{HOMO}}=6.36\) eV and \(E^\mathrm{DOP}_{\mathrm{LUMO}}=3.76\) eV. The polarization energy \(P^\pm_{im}\) accounts for electrostatic and screening effects in the solid-state. \(P^\pm_{i0}\) is evaluated for each CT state with CR calculations, whose results are extrapolated in the infinite crystal limit. Polarization energies of CT states with the hole in the pentacene HOMO\(-1\) are set to \(P_{i1}^\pm=P_{i0}^\pm-0.3\) eV. Hole transfer couplings \(t^{\mathrm{h}}\) are computed at the DFT level (PBE0 functional, 6-31G(d) basis) with the projective method. The same approach proved to be strongly dependent on the functional for pentacene-F4TCNQ CT couplings \(t^\mathrm{CT}\), for which we instead applied a multi-state generalized Mulliken-Hush diabatization scheme based on a post-Hartree-Fock description \[SCS-CC2/def-SV(P)\] of the ground and excited states. The sign of CT couplings is usually not attainable with standard quantum chemistry approaches because of the arbitrariness of the phase of \(\pi\) molecular orbitals. We overcome this limitation by using a fictitious \(s\) orbital placed above the molecular plane to probe the phase of frontier orbitals and then impose consistent phase relationships among all the molecules in the sample. Figure [\[CTene\]](#CTene){reference-type="ref" reference="CTene"} provides a graphical summary of the parameters entering Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"}, i.e. the CT couplings and the energies of localized (diabatic) CT states annotated on the lattice of doped pentacene. Figure [\[xcb\]](#xcb){reference-type="ref" reference="xcb"} shows the electron-hole distance dependence of the exciton binding energy of diabatic CT states from CR calculations, which closely follows a screened Coulomb potential even at relatively short distance. Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} can be extended to account for intramolecular structural relaxation upon charging within the framework of the Mulliken-Holstein model. We hence introduce one effective mode per molecule with coordinate \(q_i\), here treated within the adiabatic (Born-Oppenheimer) approximation, linearly modulating the energies of frontier molecular orbitals. The electronic Hamiltonian in the presence of Holstein coupling is formally equivalent to Equation [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} with \(E^\mathrm{DOP}_{\mathrm{LUMO}}\rightarrow E^\mathrm{DOP}_{\mathrm{LUMO}} + q_1\) (\(i=1\) labels the DOP site) and \(E^\mathrm{PEN}_{\mathrm{HOMO}-m} \rightarrow E^\mathrm{PEN}_{\mathrm{HOMO}-m} + q_i\), with \(i=2,\dots N\), being \(N\) is the number of molecular sites. The total energy includes the harmonic elastic contribution from molecular deformation, \[E_{harm}=\frac{1}{4\lambda^-} q_1^2 + \frac{1}{4\lambda^+} \sum_i^{i \in OSC} q_i^2,\] where \(\lambda^+\) and \(\lambda^-\) are the polaron binding energies for hole and electrons on OSC and DOP sites, respectively. These quantities have been calculated for PEN (\(\lambda^+=52\) meV) and F4TCNQ (\(\lambda^-=140\) meV) at the DFT level (PBE0 functional, 6-311G(d) basis) using differences of total energy obtained at the molecular geometries fully relaxed in the neutral and charged state (\(\Delta\)SCF scheme).
# Results {#s:results}
## Embedded \(GW\) and BSE calculations of F4TCNQ-doped pentacene {#s:res1}
The model system investigated here considers F4TCNQ substitutional defects in the pentacene crystal lattice --see Figure [\[fig0\]](#fig0){reference-type="ref" reference="fig0"}. Within our hybrid formalism, we describe a supramolecular complex (CPX) formed by one F4TCNQ molecule surrounded by its first shell of six pentacene neighbors (1+6 CPX henceforth) at the \(GW\)/BSE level. This CPX is then embedded into the pentacene crystal described within the charge response (CR) model, which provides an accurate description of the anisotropic static dielectric response of molecular crystals. F4TCNQ adopts the same position and orientation as that of the replaced pentacene molecule and its geometry has been optimized in vacuum at the CCSD level. The pentacene structure is taken from X-ray diffraction data for the vapor-grown polymorph. We stress that such an approach goes significantly beyond previous DFT electronic structure calculations, in terms of method accuracy, QM system size and account of an atomistic polarizable embedding. The first outcome of the present calculations, namely the accurate determination of the 1+6 CPX HOMO-LUMO gap in a proper dielectric environment, illustrates the striking difference between the physics of OSC doping and that prevailing in inorganic semiconductors where shallow impurity levels are located within a few dozen of meVs from the band edges. As shown in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(b), the \(GW\) HOMO-LUMO gap of the CPX is indeed found to be 0.67 eV, dramatically larger than room temperature thermal energy, clearly evidencing that the standard theory accepted for inorganic semiconductors does not apply to OSC. The analysis of the frontier orbital isocontours in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a) reveals that while the 1+6 CPX LUMO can be approximately identified as the LUMO of the F4TCNQ molecule, the two highest occupied orbitals have instead a supramolecular character, mostly involving the HOMOs of a specific pair of equivalent pentacene molecules. To better understand the origin of the HOMO-LUMO energy gap, it is useful to re-evaluate it in the absence of hybridization, namely by successively considering either a single F4TCNQ or a single pentacene molecule in the QM region. The resulting \(GW\) gap between pentacene HOMO and F4TCNQ LUMO in absence of dopant-host interactions, but with the proper dielectric environment, is found to be 0.45 eV. Hybridization therefore contributes to the HOMO-LUMO gap of the CPX, yet the latter is mostly sourced by the energy mismatch between the pentacene ionization potential (IP) and the F4TCNQ electron affinity (EA). The partial hybridization within the 1+6 CPX manifests in the frontier molecular orbitals shown in Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(a) and in an appreciable CT in the DFT ground state, with a net charge on F4TCNQ \(Q_\mathrm{DOP}=-0.25e\) as estimated from electrostatic potential fitting (ESP) atomic charges. We now turn to optical excitations as obtained within the BSE formalism. The (screened) electron-hole interaction dramatically lowers the optical gap as compared to the HOMO-LUMO gap, as evidenced by comparing Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(b) and [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}(c). The resulting CPX optical absorption spectrum is shown in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(a). The most salient feature is that the lowest singlet excitation energy (S\(_1\)) is found to be extremely low in energy, namely 34 meV above the ground state. The analysis of the corresponding BSE electron-hole two-body eigenstate, represented in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(b), reveals that S\(_1\) is an excitation of CT character taking place between HOMO and LUMO of the CPX. Additional absorption peaks in the 0.3-0.6 eV range (S\(_3\), S\(_5\)) are also associated with transitions mostly from the HOMO levels of other pairs of equivalent pentacene molecules in the CPX. Such peaks fall in the mid-infrared spectral region that has so far not been investigated experimentally. The spectrometer employed in Ref. was unable to measure absorption below 0.5 eV. The BSE optical spectrum also closely reproduces the fundamental absorption of pristine pentacene at \(1.85\) eV and, most interestingly, displays further new features in the 1.3--1.6 eV energy range that may correspond to the sub-bandgap peaks observed experimentally. The analysis of the contributing levels reveals that these excitations correspond mainly to CT transitions from the manifold of pentacene HOMO\(-1\) orbitals to the F4TCNQ LUMO. We will come back to this point in Section [4](#s:disc){reference-type="ref" reference="s:disc"}.
## Model validation and size effects {#s:res2}
Even though based on an accurate *ab initio* many-body framework which is at the forefront of what can be achieved today in terms of system size and complexity, the present \(GW\)/BSE calculations on the 1+6 CPX miss to incorporate potentially important effects associated with the delocalization of the transferred hole over pentacene molecules beyond the first shell of neighbors and local structural relaxation or polaronic effects. To extend the reach of our analysis to larger system sizes, we resort to the generalized Mulliken model for intermolecular CT presented in Section [2.2](#s:model){reference-type="ref" reference="s:model"}. The diagonalization of Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} with parameters specific to F4TCNQ-doped pentacene, as described in Section [2.2](#s:model){reference-type="ref" reference="s:model"}, yields ground and excited states for systems large enough to converge the quantities of interest with respect to the number of pentacene molecules (1+\(N\) CPX henceforth). Our model calculations yield a ground state of moderate CT, consistently with the *ab initio* findings. The charge on the dopant mildly increases from \(-0.15e\) in the 1+6 CPX to \(-0.24e\) for the 1+\(N\) CPX. This trend reflects the reduction of the gap due to hole delocalization in pentacene. The absorption spectrum computed with Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} for the 1+6 CPX is compared to BSE results in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(a). Again, the low-energy region is characterized by three CT transitions to electronic excited states where the hole is equally shared on the HOMOs of pairs of symmetry-equivalent pentacene molecules. In the 1.2--1.8 eV energy span, we predict three other absorption peaks corresponding to CT excitations where the hole lies in the pentacene HOMO\(-1\) orbitals. The agreement with BSE is excellent for both the relative energy and intensity of the electronic transitions and the shape of the excited-state electron-hole maps, see Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(b-c). This provides an important validation step for our model. Finite-size effects are also addressed in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(a), where we compare the absorption spectra of the 1+6 and 1+\(N\) CPX calculated with Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"}. Upon extending the size of the system (1+\(N\) CPX), we observe a small shift to lower energies of the lowest two energy bands due to the delocalization-induced stabilization of S\(_1\) and S\(_3\) and many additional transitions in the 0.4-0.8 eV range, signaling the presence of a dense manifold of delocalized CT states. Size effects affect in a similar way CT states arising from pentacene HOMO\(-1\) in the 1.3-1.9 eV range. On a more general vein, Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"} describes a smooth transition from a neutral (\(Q_\mathrm{DOP}=0\)) to a fully ionized dopant (\(Q_\mathrm{DOP}=-e\)) as a function of the difference between the IP and EA of the two components (with all other parameters fixed), Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(a). The width of the intermediate-ionicity window primarily depends on the magnitude of the couplings between the pentacene HOMO and the F4TNCQ LUMO (\(t_{i0}^\mathrm{CT}\)), with large values favoring a mixed-valence ground state and zero or full CT in the opposite case. We also remark that the ground-state value of \(Q_\mathrm{DOP}\) (0 K) can substantially differ from its thermal average at 300 K in the crossover regime, mostly owing to triplet excitations which are full-CT states of threefold multiplicity.
## Structural relaxation and polaronic effects {#s:res3}
The picture emerging so far depicts F4TCNQ-doped pentacene as a system close to the neutral-ionized boundary, as testified by the presence of CT excitations below 0.1 eV. Such a small gap is commensurate with the polaron binding energies for F4TCNQ-pentacene, \(\lambda^+ + \lambda^-=0.2\) eV (see Section [2.2](#s:model){reference-type="ref" reference="s:model"}), calling for the inclusion of vibrational effects in the analysis. We hence extended the model to account for one effective Holstein mode per molecule linearly coupled to the site charge as described in Section [2.2](#s:model){reference-type="ref" reference="s:model"}. The optimization of the ground state energy with respect to the set of intra-molecular coordinates leads to a qualitatively different *symmetry-broken ground state of full-CT character*, shown in the inset of Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(c). Indeed, because of the two low-lying degenerate CT states at the undistorted molecular geometries, a structural (Jahn-Teller like) instability develops in the system, leading to the collapse of the hole on one of the nearest pentacene neighbors --see Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(b). Structural relaxation therefore favors full dopant ionization and suppresses states of fractional CT. This is shown by the \(Q_\mathrm{DOP}\) vs. \((\mathrm{IP_{OSC}-EA_{DOP}})\) curve in Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(a) computed at relaxed molecular geometries (green line), presenting an abrupt step-like transition whose turning point is shifted to higher \((\mathrm{IP_{OSC}-EA_{DOP}})\) with respect to the rigid-molecule case. Hole localization is also observed upon introducing Gaussian energetic disorder in the model, which lifts the equivalence of pentacene sites. Structural relaxation and disorder are expected to concur in localizing charges in real (experimental) systems. The optical absorption of the 1+\(N\) CPX at the relaxed and unrelaxed geometry, both shown in Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(c), are only apparently similar. For instance, the lowest energy transition at 76 meV now leads the relaxed ground state to another CT state with the hole localized on the HOMO of the pentacene located on the opposite side of the dopant. Though located at similar energy, it has a completely different nature compared to the optical transition S\(_1\) in Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(c), which was a CT transition transferring one electron from pentacene to F4TCNQ from an almost neutral ground state. The most relevant feature of the absorption of the vibrationally-relaxed CPX is, however, the very intense band at 1.25 eV, which falls very close in energy to the full-CT (pentacene HOMO\(-1\) \(\rightarrow\) F4TCNQ LUMO) transition discussed above for the unrelaxed system. This quasi-coincidence is a specific feature of the pentacene-F4TCNQ CPX and of its nearly vanishing optical gap. Indeed, this band corresponds to a transition from the CT ground state in the Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(c) (inset), with the hole in the pentacene HOMO, to another full-CT excited state where the hole is in the HOMO\(-1\) of the same molecule. This band owns its large absorption intensity to the off-diagonal element of the dipole moment operator associated with such an intramolecular reshuffling of the hole density. In the most common nomenclature, the band at 1.25 eV corresponds to a *polaronic transition* of the charged pentacene in the presence of a counter-electron on F4TCNQ.
# Discussion {#s:disc}
The experimental observation of sub-bandgad absorption bands in F4TCNQ-doped pentacene can be rationalized both in terms of CT bands in weakly interacting complexes, or as polaronic features of ionized pentacene. Our accurate analysis leads us to assign these sub-bandgap features to the pentacene radical cation. This assignment is further corroborated by spectroscopic measurements on solutions of TIPS-pentacene oxidized by the very strong electron acceptor FeCl\(_3\). The strikingly coincident double peak observed at 1.2--1.4 eV both in F4TCNQ- and FeCl\(_3\)-doped (TIPS-)pentacene is a very strong indication of the same origin of these features, namely two vibronic replica of the same polaronic transition. The experimental band at \(\sim\)`<!-- -->`{=html}1.4 eV may also have a contribution from a characteristic absorption of the F4TCNQ anion. Extensive spectroscopic (photoemission, optical and vibrational) investigations by Koch and co-workers allowed identifying two different scenarios for doped molecular crystals, characterized by strong orbital hybridization and partial CT, and conjugated polymers, mostly exhibiting full ionization of dopant impurities. Our accurate theoretical analysis, tightly connected to experimental evidence, concludes that F4TCNQ-doped pentacene represents an exception to this empirical rule, at least in the low-doping regime targeted by our calculations. We further note that at least one case of partial CT has also been reported in conjugated polymers. We emphasize that the absence of intragap features in the photoemission spectra of doped pentacene is not inconsistent with full dopant ionization and with the presence of pentacene cation sub-bandgap optical excitations. Indeed, the intramolecular charging energy, associated with photoemission, is expected to split the singly and doubly occupied levels of pentacene, bringing the UPS features of the cation (polaron) deep in the valence band, as very recently shown for C\(_{60}\) fullerenes and a narrow-band polymer . This charging energy is instead irrelevant for optical transitions, namely excitations localized on the pentacene cation that do not affect the molecular charge. Vibrational and optical data on other molecular systems, such as quaterthiophene (T4) or 2,7-didecyl\[1\]benzothieno-\[3,2-b\]\[1\]benzothiophene (BTBT), instead provide clear evidence in support of the weak CT CPX scenario. This is rationalized by the curve in Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(a), when we consider the \(\sim\)`<!-- -->`{=html}0.4 eV larger IP of T4 and BTBT with respect to pentacene (\(\mathrm{IP}_{PEN}=4.9\) eV). Further confirmations of weak dopant ionization come also from very recent correlated-electron calculations targeting the 1:1 F4TCNQ:BTBT mixed-stack crystal. A less efficient charge injection in the OSC matrix does also rationalize the 2-3 orders of magnitude lower conductivity in doped T4 and BTBT with respect to doped pentacene, despite the comparable or higher charge carrier mobility in the formers. Doped polymers can also be described by Hamiltonian [\[e:H0\]](#e:H0){reference-type="ref" reference="e:H0"}, although quantitative differences in its parameters should be expected with respect to molecular systems. On the basis of approximate empirical parametrization, we suggest that the full dopant ionization observed experimentally in the large majority of doped polymers should be ascribed to their generally lower IP resulting from large bandwidths (e.g. in poly(3-hexylthiophene) IP=4.6 eV, bandwidth \(\sim3.5\) eV ). Bandwidths larger than the charging energy should also grant validity of the single particle picture in most conjugated polymers, hence explaining the observation of intra-gap states in UPS arising from polaronic relaxation.
# Conclusion and perspectives {#s:conclu}
In summary, we have developed a multifaceted approach combining state-of-the-art electronic structure methods and applied this formalism to gain a comprehensive picture of the electronic and optical properties of organic semiconductors in the low doping regime. Our analysis is based on novel many-body *ab initio* techniques that allow the accurate determination of charged and optical excitations in heterogeneous molecular systems, fully accounting for the complex non-local physics of interacting electrons and for the dielectric screening provided by the molecular environment. First principles calculations are also employed to parametrize a general model Hamiltonian for doped organic semiconductors covering the entire spectrum of possibilities between neutral and fully ionized dopants. By explicitly considering the prototypical case of F4TNQ-doped pentacene, our calculations show that even for acceptor dopant levels lying fairly deep into the gap, full ionization may still be possible thanks to the concurrent effects of electron-hole interaction, spin statistics and polaronic relaxation, yet leading to charge carriers that are strongly bound to the parent dopants. Our analysis confirms the common belief that the difference between the IP of the semiconductor and the dopant EA is an important parameter but certainly not the only one. Indeed, other quantities such as electron-hole interaction, dopant-host CT integrals and the coupling to vibrations are found to be key for dopant ionization. All these quantities can be strongly dependent on the material and on the morphology, calling for a detailed analysis of structure-property relationships. The proposed approach provides a robust framework to such a scope. On more general grounds, this study puts the accent on the central role of electron-hole interaction and polaronic effects in favoring dopant ionization, two effects that are missed in the well-established theories for electrical doping in inorganic semiconductors, whose applicability to the case of organic systems is here called into question. In contrast to the picture prevailing for doped inorganic semiconductors, the explanation for the high conductivity in heavily doped organics should thus be sought beyond independent-electron theories. | {'timestamp': '2017-06-30T02:03:35', 'yymm': '1706', 'arxiv_id': '1706.01853', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01853'} |
null | null |
null | null |
null | null |
# Introduction
Recent work has established the stability of three-dimensional spin liquid phases described by higher rank tensor gauge theories, such as the \(U(1)\) higher spin gauge theories and the "generalized lattice gauge theories\" of Vijay, Haah, and Fu, which are the natural discrete analogue. The most exotic feature of these higher-rank phases is the fact that the emergent tensor gauge field must be coupled to excitations with subdimensional behavior. These particles are restricted by gauge invariance to exist only in lower-dimensional subspaces of the full three-dimensional system. For example, some particles can be restricted to motion only along a one-dimensional line, or within a two-dimensional plane. As an extreme case of this phenomenon, some types of particles are restricted to a 0-dimensional subspace, \(i.e.\) a point, and become totally immobile. Such an immobile excitation is called a fracton, or a 0-dimensional particle. Theoretical evidence for these immobile particles has been seen for more than a decade, but interest in the topic has particularly surged just in the last few years, establishing the subject as a new subfield of condensed matter physics. In this work, our focus will be on a higher rank \(U(1)\) phase with fracton excitations. Previous work on this topic has already established many of the properties of these fractons, such as their conservation laws and their interactions with the tensor gauge field. This earlier work focused on the zero-temperature behavior of these phases, analyzing the physics of isolated particles. But what happens to these phases at finite temperature, when we have a soup of thermal excitations? In particular, is there a smooth crossover to a trivial phase, as in the conventional \(U(1)\) spin liquid, or could there be a robust finite-temperature phase? And to what extent are the finite-temperature properties controlled by the physics of the zero-temperature quantum spin liquid? Work on the finite-temperature behavior of the discrete fracton models has already yielded interesting physics, such as long thermalization times and glassy dynamics. Are there similar stories to be found in the \(U(1)\) fracton models? In the conventional (rank 1) \(U(1)\) spin liquid in three dimensions, the finite-temperature behavior is fairly easy to obtain, following from the familiar properties of Maxwell electromagnetism. At finite temperature, there will be a thermal distribution of the emergent charges which can self-consistently screen out their associated long-range electric fields. Very close to a charged particle, one will still see the Coulomb field of a point charge, decaying as \(1/r^2\). But at distances longer than some screening length \(\lambda\), the electric field will decay exponentially as \(e^{-r/\lambda}\), making the particle look effectively neutral (\(i.e.\) trivial) at the longest length scales. Since all excitations are neutralized at long distances, this allows for a smooth crossover at finite temperature between the \(U(1)\) spin liquid phase and the trivial phase, where all excitations are neutral. In the thermodynamic limit, the \(U(1)\) spin liquid is therefore only strictly distinct from the trivial phase at zero temperature. Technically, at any finite temperature, emergent \(U(1)\) charges are screened out if one looks at sufficiently long length scales. But, as we will review, the screening length grows exponentially as the temperature is decreased, \(\lambda \propto e^{m/T}\), where \(m\) is the mass scale of the charges. This rapid growth of the screening length allows for spin liquid physics to hold on macroscopic length scales at accessible temperatures. For example, suppose we are in a finite system of linear size \(L\). Once the temperature drops below \(T \sim m/\log L\), the screening length will be of the same order as the system size, so screening will no longer play any significant role in the system. Since this temperature only decreases logarithmically with system size, it is not unrealistically low, even for macroscopic systems. Therefore, at low (but accessible) temperatures, the emergent charges are well-defined over macroscopic length scales, and the zero-temperature physics of the \(U(1)\) spin liquid dominates the behavior of the system. In this work, we will generalize the notion of finite-temperature screening to a higher rank \(U(1)\) spin liquid phase, finding some other interesting finite-temperature physics along the way. We will put particular emphasis on a specific phase which hosts both fracton excitations and nontrivial mobile bound states. We will provide a review of the essential physics of such a phase of matter. The finite-temperature behavior of this system should be fairly representative of all \(U(1)\) phases which possess both fractons and mobile bound states. This is in contrast to two other types of higher rank \(U(1)\) phases: those without fractons, and those without nontrivial mobile bound states. Such phases have comparatively simpler finite-temperature physics, and we will comment on them towards the end. For the majority of this work, however, we specialize to the case of \(U(1)\) phases which have both fractons and mobile bound states. We will find that the fracton excitations in such a theory will indeed be screened at finite temperature. Therefore, this phase has a smooth finite-temperature crossover to the trivial phase, just like the conventional \(U(1)\) spin liquid. However, unlike the normal \(U(1)\) spin liquid, there are two distinct temporal regimes associated with screening of a charged disturbance in the system. At short timescales, the dominant screening mechanism is from thermally excited mobile bound states. In this regime, which we dub the weak-screening regime, the electric field of a fracton is weaker than the zero-temperature behavior, but still decays as a power law. At much longer timescales, a fracton can actually be screened by other fractons. In this strong-screening regime, the electric field of a point charge decays exponentially, and all correlations become short-ranged. In either case, the composite object of a fracton plus its screening cloud forms a mobile object which can move around the system, albeit at very slow speeds. In this sense, fracton immobility is technically absent at finite temperature. However, once again we will find that the screening length grows exponentially as the temperature is lowered. At sufficiently low temperatures, the zero-temperature physics will dominate on macroscopic length scales, and fractons will be effectively immobile. Furthermore, even above this temperature, the zero-temperature fracton physics will make its presence felt in the form of a large thermalization time. This obstruction to thermalization is particularly interesting since it occurs in a completely clean system, without the need for disorder, as in more conventional many-body localization situations. These unusual finite-temperature properties, such as slow thermalization in the clean limit and persistent power-law correlations, may serve as useful diagnostics in the experimental search for \(U(1)\) fracton phases.
# The Conventional \(U(1)\) Spin Liquid: A Review
We begin by briefly reviewing a simple calculation for screening in the conventional three-dimensional \(U(1)\) spin liquid, which closely follows the treatment of standard electromagnetic screening. The calculation will be useful in order to compare and contrast with the higher rank case. Note that the analysis will be almost purely classical. The particles have a finite energy gap, so at low temperatures the particles will be dilute, making the effects of particle statistics unimportant. A classical analysis will capture the correct qualitative physics throughout. We consider a \(U(1)\) spin liquid, which is described by an emergent \(U(1)\) gauge theory, at a finite temperature \(T\). We take the emergent charges of the \(U(1)\) gauge field, both positive and negative, to have some finite mass \(m\). At temperature \(T\), the equilibrium density of positive and negative charges will be: \[n_+ = n_-\equiv n_0(T) \propto e^{-m/T}\] (We will not consider bound states of higher charge, which are more energetically costly and contribute comparatively less to screening.) We now want to know how these thermal excitations respond to a "test charge\" introduced to the system. We imagine adding an extra point charge \(Q\) at a particular fixed location, then examining how the system responds to the resulting change in potential from the disturbance. The presence of a perturbing potential \(\phi(r)\) will shift the Boltzmann weight of the thermally excited particles, causing the density of positive (negative) charges to adjust to: \[n_\pm = n_0 e^{\mp\beta\phi} \approx n_0(1\mp\beta\phi)\] where we have defined \(\beta = T^{-1}\). The linearization is only valid for a small perturbation, breaking down in the immediate vicinity of the point charge, but captures the correct long-distance physics, which is our primary concern. Adopting a normalization such that the charge of the emergent particles is \(q=1\), the net charge density of the thermal excitations is then: \[\rho = n_+-n_-\approx-2n_0\beta\phi\] We recall that, in normal electrostatics, the potential of an arbitrary charge configuration can be written as: \[\phi(r) = \int d^3r' \frac{\rho(r')}{4\pi|r-r'|}\] where the integral is over three-dimensional space, and we have adopted units such that Gauss's law is \(\partial_i E^i = \rho\). Making use of this equation, the potential around the test charge will take the form: \[\phi(r) = \int d^3r' \frac{-n_0\beta\phi(r')}{2\pi|r-r'|} + \phi_{bare}(r)\] where \(\phi_{bare}\) corresponds to the bare potential of the unscreened test charge, \(Q/4\pi r\). Taking the Laplacian of both sides then gives us the Poisson equation: \[\partial^2\phi = 2 n_0\beta\phi-Q\delta^{(3)}(r)\] where the delta function represents the test charge. Away from the delta function (which sets boundary conditions at the origin), the potential obeys \(\partial^2\phi = 2n_0\beta\phi\), which has an exponentially decaying solution, \(\phi\propto e^{-r/\lambda}\), with a screening length given by: \[\lambda = (2 n_0\beta)^{-1/2} \propto e^{m/2T}\] where we use the fact that \(n_0\propto e^{-m/T}\). We therefore see that, at distances longer than \(\lambda\), any charge in the system will be screened, and there is little remnant of the emergent gauge structure. This allows for a smooth finite-temperature crossover to a trivial phase, which has only neutral excitations. However, we note that the screening length grows exponentially at low-temperatures, \(\lambda\propto e^{m/2T}\). For temperatures below \(T\sim m/\log L\), which is not unreasonably low, the screening length becomes larger than the system size, and the physics of the \(U(1)\) spin liquid will hold throughout the system.
# The \(U(1)\) Fracton Phase
## Review of the Model
We now wish to generalize the preceding treatment of screening to a higher rank \(U(1)\) phase possessing both fractons and also nontrivial mobile excitations. We begin by reviewing the simplest model of this type. We expect the same qualitative physics should hold for any \(U(1)\) model with these basic characteristics. The model we will work with has been studied in some detail in previous literature. We will review the most important features of this phase here, though we refer the reader to these previous works for more details. In particular, we will make use of some basic results of the generalized tensor electromagnetism, first established in Reference . We take the degrees of freedom of our model to be those of a symmetric rank 2 tensor field \(A_{ij}\), defined throughout three-dimensional space, which will play the role of the emergent gauge field. We call the canonically conjugate variable \(E_{ij}\), representing a symmetric tensor electric field. The theory can be defined by specifying a Gauss's law constraint on the electric field, which in turn specifies the gauge transformation. For the model we will focus on here, the generalized Gauss's law takes the form: \[\partial_i\partial_j E^{ij} = \rho\] for scalar charge density \(\rho\). We will take the charges of the system to generically have a finite energy gap. In this case, the low-energy sector is free of charges, \(\partial_i\partial_j E^{ij} = 0\), resulting in invariance under the following gauge transformation: \[A_{ij}\rightarrow A_{ij} + \partial_i\partial_j \alpha\] for scalar gauge parameter \(\alpha\) with arbitrary spatial dependence. This gauge invariance then dictates the form of the low-energy Hamiltonian governing the gapless gauge mode of the theory: \[H = \int d^3r\,\frac{1}{2}(E^{ij}E_{ij} + B^{ij}B_{ij})\] where the gauge-invariant magnetic tensor takes the form \(B_{ij} = \epsilon_{iab}\partial^aA^b_{\,\,\,j}\). This Hamiltonian leads to a linear dispersion for the gapless gauge mode, including both spin-2 and lower spin degrees of freedom. The gapless gauge mode is interesting in its own right, but here we will mostly be focusing on the physics of charges, as defined by the Gauss's law, \(\rho = \partial_i\partial_j E^{ij}\). These charges are most notable for their conservation laws. Like in the conventional \(U(1)\) spin liquid, this system will obey conservation of charge: \[\int d^3r\,\rho = \textrm{constant}\] \(i.e.\) the total charge in the system remains fixed. However, the charges of this higher rank \(U(1)\) gauge theory also obey a second conservation law: \[\int d^3r\,\rho\vec{r} = \textrm{constant}\] which reflects the conservation of dipole moment. This means that, not only must charge be conserved, but any motion of charges must proceed in such a way that the total dipole moment of the system remains fixed. The consequences of this new conservation law are severe. An isolated charge cannot move by itself, since this would change the total dipole moment. An isolated fundamental charge of this theory is locked in place and is therefore a fracton. While isolated charges of this theory are locked in place, charges can obtain mobility when they come together to form bound states. For example, a dipolar bound state of a positive and negative charge is free to move around the system in any direction, as long as its dipole moment remains fixed. Furthermore, such dipoles are nontrivial excitations of the system, since they cannot be directly created from or absorbed into the vacuum, due to the dipolar conservation law. There are many other types of \(U(1)\) fracton models, with various gauge structures. For example, most \(U(1)\) gauge theories of rank higher than 2 tend to have fracton excitations. But in all cases, the physics of fractons is essentially driven by the same basic principle: higher moment conservation laws which restrict the motion of charges.
## Weak-Screening Regime
Now that we have our model in hand, we want to know how the physics of screening carries over to this theory. One would like to once again say that thermally excited charges will rearrange themselves in such a way as to screen out long-range interactions. However, the fundamental charges of the theory are fractons, which are locked in place and cannot rearrange themselves into a screening pattern. (We will revisit and adjust this idea later.) But while the fractons cannot easily react to a potential, the dipoles of the system can. At nonzero temperature, the system will have a finite density of thermally excited mobile dipoles, which will react to the field of a test charge and can contribute to screening. However, since the dipoles are charge-neutral objects, we expect that they will not be able to screen as effectively as charged particles could. Indeed, we will find that dipoles can only partially screen charged disturbances. Just as we earlier focused on screening due to the minimal charges in the conventional \(U(1)\) spin liquid, we will here focus on screening due to the minimal dipoles. As discussed in earlier work, the dipole moment in this theory is quantized in a lattice-dependent way. For example, on a cubic lattice system, a dipole is labeled by three integers, \(\vec{p} = (n_x,n_y,n_z)\), in appropriate units. In this case, the minimal dipoles are \(\vec{p}= (\pm 1,0,0)\), \((0,\pm 1,0)\), and \((0,0,\pm 1)\). Such minimal dipoles will provide the dominant contribution to screening, since they are energetically cheapest and will therefore be thermally excited in the greatest numbers, with a density \(n_{0,d}(T)\propto e^{-m_d/T}\), where \(m_d\) is the mass of the minimal dipoles. Including the effects of higher strength dipoles should not affect the qualitative physics. We will now take advantage of some results regarding the generalized electromagnetism of this phase, established in Reference . By imposing the conditions of electrostatics, it is not hard to show that a static electric field can be described via a scalar potential formulation, \(E_{ij} = \partial_i\partial_j \phi\). For a point charge \(q\), this scalar potential takes the form: \[\phi(r) =-\frac{qr}{8\pi}\] and the corresponding electric field falls off as \(E_{ij}\sim 1/r\). In the presence of such a scalar potential, the potential energy of a dipole will be \(-p^j\partial_j\phi\), which shifts the Boltzmann weight. The density of thermally excited dipoles with dipole moment \(p^j\) will then become: \[n_p = n_{0,d}e^{\beta p^j\partial_j\phi}\approx n_{0,d}(1+\beta p^j\partial_j\phi)\] The scalar potential associated with a dipole takes the form: \[\phi_p(r) = \frac{(p\cdot r)}{8\pi r}\] Using this form of the scalar potential, the total potential around a test charge becomes: \[\phi(r) = \phi_{bare}(r) + \sum_{\{p^j\}} n_{0,d} \int dr' (1+\beta p^i\partial_i\phi(r')) \frac{p^j(r_j-r_j')}{8\pi|r-r'|}\] where the sum is over the possible orientations of the minimal dipoles. These orientations depend on the underlying lattice, but for any lattice, each orientation will also appear with its negative, so \(\sum p^j = 0\). This leads to the vanishing of the first term in the sum, leaving us with: \[\phi(r) = \phi_{bare}(r) + n_{0,d}\beta \int dr' \bigg(\sum_{\{p^j\}} p^ip^j\bigg) \frac{\partial_i\phi(r') (r_j-r_j')}{8\pi|r-r'|}\] For the cubic lattice, the sum in the integrand evaluates to \(2\delta^{ij}\). For other lattice choices, the sum will also be proportional to \(\delta^{ij}\), though the coefficient is non-universal, so we write \(\sum_p p^ip^j = \alpha\delta^{ij}\). Using this, our above equation becomes: \[\begin{split} \phi(r) = \,\,&\phi_{bare}(r) + \alpha n_{0,d}\beta \int dr' \frac{\partial^j\phi(r') (r_j-r_j')}{8\pi|r-r'|} = \\ &\phi_{bare}(r)-\alpha n_{0,d}\beta \int dr' \frac{\phi(r')}{4\pi |r-r'|} \end{split}\] where we have integrated by parts. We then take a Laplacian to yield: \[\partial^2\phi = \partial^2\phi_{bare} + \alpha n_{0,d}\beta\phi \label{dipolescreen}\] Let us now take \(\phi_{bare}\) to be that of a bare test charge, which has the form \(\phi_{bare} =-Qr/8\pi\). Taking the Laplacian yields \(\partial^2\phi_{bare} =-Q/4\pi r\), so Equation [\[dipolescreen\]](#dipolescreen){reference-type="ref" reference="dipolescreen"} becomes: \[\partial^2\phi =-\frac{Q}{4\pi r} + \alpha n_{0,d}\beta \phi\] Fourier transforming, we obtain: \[\begin{aligned}
-k^2\phi(k) =-\frac{Q}{k^2} + \alpha n_{0,d}\beta\phi(k) \\ \phi(k) = \frac{Q}{k^2(k^2+\alpha n_{0,d}\beta)} \end{aligned}\] At short distances close to the charge (corresponding to large \(k\)), we will have \(\phi(k) \sim k^{-4}\), which gives \(\phi(r)\sim r\). This is as expected, since at short distances we should see the field of a point charge. At long distances (small \(k\)), the scaling is adjusted to \(\phi(k)\sim k^{-2}\), so \(\phi(r) \sim 1/r\). The crossover occurs around a screening length of: \[\lambda = (\alpha n_{0,d}\beta)^{-1/2} \propto e^{m_d/2T}\] using the fact that \(n_{0,d}\propto e^{-m_d/T}\). At distances longer than this, the potential behaves (up to an additive constant) as: \[\phi(r) = \frac{Q\lambda^2}{4\pi r}\] In this regime, the electric field \(E_{ij} = \partial_i\partial_j\phi\) will scale as \(1/r^3\). We therefore see that, after accounting for finite-temperature dipole screening, the bare \(1/r\) electric field is screened to a \(1/r^3\) field. This reduction by two powers indicates that the first two terms of the generalized multipole expansion vanish. In other words, the composite object of the fracton plus its screening cloud is not only neutral, but also has zero dipole moment. Therefore, the screened fracton is carrying no nontrivial quantum numbers and is a trivial excitation. Since all finite-temperature excitations are effectively trivial at long length scales, this higher rank \(U(1)\) spin liquid can smoothly cross over to a trivial phase at finite temperature, just like its more conventional cousin. Nevertheless, the screening length still grows exponentially at low temperatures, \(\lambda\propto e^{m_d/2T}\). Once again, for \(T< m_d/\log L\), the screening length is cut off by the system size, and the physics of the higher rank spin liquid will hold throughout the system. Even above this temperature, the onset of this type of fracton mobility is quite slow. In order for the screened charge to move, it must drag along its screening cloud, of linear size \(\lambda\). As discussed in Reference , the effective mass of such a composite object grows exponentially in the separation of particles, which in this case is of order \(\lambda\) on average. In other words, the characteristic velocity of the screened excitations will be exponentially small in the screening length: \[v_{eff}\propto e^{-(\lambda/a)}\propto e^{-e^{m_d/2T}}\] where \(a\) is some lattice scale. At low temperatures, the mobility of a screened fracton decreases doubly exponentially, leading to negligible fracton motion via this mechanism. However, there is actually a different finite-temperature mechanism which can lead to slightly faster fracton motion, as we discuss next.
## Absorptive Diffusion
We have found that, once screening is taken into account, a fracton can move around the system, so long as it is willing to drag along its entire screening cloud. However, such a bulky process is not the most efficient type of fracton motion. Fractons can move more quickly through the direct absorption of thermally excited dipoles. As illustrated in Figure [\[fig:ab\]](#fig:ab){reference-type="ref" reference="fig:ab"}, a fracton can absorb a dipole and thereby hop in the direction of its dipole moment. In a certain sense, such a dipole is carrying a "quantum of position,\" which is transferred to the fracton. By absorbing randomly directed thermally excited dipoles, the fracton will thereby diffuse around the system, effectively performing a random walk. We refer to this process as "absorptive diffusion.\" Strictly speaking, in order to conserve momentum, the process depicted in Figure [\[fig:ab\]](#fig:ab){reference-type="ref" reference="fig:ab"} will have to involve the gapless gauge mode ("graviton\"). The initial fracton and the final fracton are both carrying zero momentum, while the dipole momentum is in general nonzero. The absorption of a dipole must therefore be supplemented by graviton emission in order to conserve momentum and energy. However, the gravitons are gapless and can be readily emitted and absorbed in a thermal system. The dominant factor setting the rate of dipole absorption will be the thermal density of dipoles, \(n_0\). The typical velocity of a fracton hopping via this mechanism therefore behaves as: \[v_{eff}\propto n_{0,d}\propto e^{-m_d/T}\] (The thermal distribution of gravitons, \(n_g\), can only provide a power-law prefactor to this exponential scaling.) This diffusive velocity is must faster than the typical velocity of a screening cloud at low temperatures, which we found in the previous section to behave doubly exponentially in \(T\). Fracton mobility is therefore dominated by this sort of process. Note, however, that the fracton does not maintain this velocity in one particular direction for a sustained period of time, but rather undergoes diffusive motion with a mean free path set by the lattice scale.
## Strong-Screening Regime
At the outset of our finite-temperature analysis, we assumed that the primary screening mechanism comes from thermally excited dipoles, since fractons are largely immobile, and therefore cannot easily rearrange themselves to create a screening cloud. However, we have now found that there actually are mechanisms by which a fracton can move at finite-temperature, albeit at very slow speeds. Therefore, we should expect fractons to contribute to screening on sufficiently long time scales. Before estimating the relevant time scale, let us examine the effects of such fracton-on-fracton screening. We now assume that the system has had sufficient time for the fracton sector to reach thermal equilibrium, such that we have a thermal distribution of fractons, \(n_{0,f}\propto e^{-m_f/T}\), where \(m_f\) is the mass scale of fractons. The scalar potential \(\phi\) plays the role of the potential energy per charge associated with fractons. The presence of such a potential will therefore shift the corresponding Boltzmann weights. In the presence of a small perturbing potential, the thermal density of fractons becomes: \[n_\pm = n_{0,f} e^{\mp\beta\phi}\approx n_{0,f}(1\mp\beta\phi)\] for positive and negative charges, respectively. The net charge density will then be: \[\rho = n_+-n_-=-2n_{0,f}\beta\phi\] As discussed earlier, the potential generated by a unit charge fracton takes the form \(\phi =-r/8\pi\). The net potential around a test charge will then take the form: \[\phi(r) = \int dr' \frac{n_{0,f}\beta}{4\pi}|r-r'|\phi(r') + \phi_{bare}(r)\] Taking a double Laplacian, we then find the generalized Poisson equation: \[\partial^4\phi(r) =-2n_{0,f}\beta\phi(r) + \partial^4\phi_{bare}(r)\] Taking the bare potential to be that of an isolated fracton, we obtain: \[\partial^4\phi(r) =-2n_{0,f}\beta\phi(r) + Q\delta^{(3)}(r)\] Taking a Fourier transform, we have: \[\begin{split} k^4\phi(k) &=-2n_{0,f}\beta\phi(k) + Q \\ \phi(k) &= \frac{Q}{k^4 + 2n_{0,f}\beta} \end{split}\] At large \(k\) (short distances), the \(k^{-4}\) behavior will lead to the same linear potential as the unscreened fracton, as expected. However, at small \(k\) we have that \(\phi(k)\) approaches a constant. This indicates that, in real space, \(\phi(r)\) decays exponentially at long distances, with a characteristic screening length given by: \[\lambda\propto (n_{0,f})^{-1/4} \propto e^{m_f/4T}\] We see that, at the longest timescales, potentials and electric fields decay exponentially, leaving only short-range correlations in the system. We refer to this as the strong-screening regime, in contrast with the power-law behavior of the weak-screening regime, brought about by dipole screening. While all potentials and fields will eventually become exponentially screened, the timescale for this to happen can be quite large. In order for fracton-on-fracton screening to become effective, we need fractons from the thermal bath to move distances of order the screening length \(\lambda \propto (n_{0,f})^{-1/4}\). As we identified in the previous section, the typical velocity of fractons will scale as \(n_{0,d}\), arising from the absorption of dipoles. The timescale for strong screening to come into effect then behaves as \(\lambda/v \sim (n_{0,f})^{-1/4}(n_{0,d})^{-1}\). (Note that the fracton is no longer undergoing a random walk, but rather being pulled in by the perturbing potential.) As a function of temperature, the strong-screening timescale behaves as: \[\tau_{ss}\propto (n_{0,f})^{-1/4}(n_{0,d})^{-1} \propto e^{(\frac{1}{4}m_f + m_d)/T}\] The important aspect of this formula is its exponential growth as \(T\) is lowered. At low temperatures, one would need to wait unreasonably long times for strong screening to come into effect, and the power-law behavior of weak screening will hold over experimentally relevant timescales. This is an interesting intermediate regime which has no analogue in the conventional \(U(1)\) spin liquid. For the rank 1 case, as soon as a finite temperature is turned on, exponential screening quickly comes into effect and all correlations of observables decay exponentially. For the fractonic \(U(1)\) spin liquid described here, however, a charged disturbance will leave residual power-law correlations in the system for exponentially long times at low temperature. These finite-temperature correlations will not be as long-ranged as the corresponding zero temperature ones. Nevertheless, the existence of these long-range correlations, and their eventual disappearance at long times, may serve as an interesting way to diagnose the presence of a \(U(1)\) fracton phase.
## Thermalization
Our analysis has indicated that, at low temperatures, the motion of \(U(1)\) fractons is still heavily suppressed, just as in the discrete fracton models. An interesting question which we can now ask is, starting from some fixed (non-thermal) configuration of fractons, how long does it take for these fractons to come to thermal equilibrium? For example, starting from a stationary finite density \(n\) of fractons, how long does it take for these fractons to relax to a thermal distribution? Let us first ask this question in the presence of a thermal bath of dipoles, in which case each fracton will effectively perform a random walk, where each step takes a time proportional to \(n_{0,d}^{-1}\). We expect that, in order to thermalize, each fracton must travel at least a distance of order the interfracton spacing, \(n^{-1/3}\), so that the fractons can mix together and smooth out the distribution. By random walk physics, this requires about \(n^{-2/3}\) steps. We can therefore bound that the thermalization time of the system should be at least: \[\tau_{therm} \sim n^{-2/3}n_{0,d}^{-1}\sim e^{m_d/T}\] which grows rapidly when the temperature becomes small. Of course, this behavior will not persist to arbitrarily low temperatures. When the temperature becomes low enough, dipoles will no longer be present in sufficient quantity to facilitate motion. At the lowest temperatures, fractons will move primarily via motion mediated by other fractons. As discussed in earlier work, a fracton will acquire finite mobility in the presence of a finite-density distribution of other fractons. In the presence of fracton density \(n\), the typical velocity of a fracton will behave as \(v\sim n\). Once again, we expect that thermalization will only occur after fractons have moved a distance on the order of the interparticle spacing, \(n^{-1/3}\). In this case, we can bound that the thermalization time should be at least: \[\tau_{therm}\sim \frac{n^{-1/3}}{n} \sim n^{-4/3}\] We therefore see that the thermalization time of an isolated fracton distribution (\(i.e.\) without a thermal dipole bath) can also grow to be large when the fracton density is small, though the growth is now merely algebraic, instead of exponential. But at low enough densities, such a state may still appear non-thermal on experimentally relevant timescales. One is reminded of the physics of many-body localization. Importantly, however, the resistance to thermalization of this system occurs even in the clean limit, without the need for any disorder. This is an interesting feature of fracton models, which has been noted in the past. Before concluding this section, we note that there is one more process which might alter the above analysis: fracton decay. We will discuss this issue in the Appendix, where we will find that fracton decay does not represent an important contribution to the thermalization of the system.
# Extensions to Other Higher Rank \(U(1)\) Spin Liquids
In the preceding sections, we have focused on a specific rank 2 \(U(1)\) spin liquid phase with both fractons and nontrivial mobile excitations. We found that this behavior leads to two distinct screening regimes. At short times, only the mobile excitations can contribute to the screening, leaving weak power-law correlations in the system. At much longer timescales, fractons will be able to fully screen each other, leaving only short-range correlations in the system. For other higher rank \(U(1)\) spin liquids with both fractons and mobile excitations, the story is likely much the same. Mobile excitations will play a dominant role in screening on short timescales, but they will not be able to fully screen the long-range interactions. On much longer timescales, however, we expect full screening to occur via the rearrangement of fractons. Some of the scaling estimates determined in previous sections will likely be different, depending on the details of the phase, but the qualitative physics should be unchanged. On the other hand, there are two other possible types of higher rank \(U(1)\) phases, which we briefly mention in turn.
## "Type 2\" \(U(1)\) Fracton Models
From earlier work on discrete fractons, we know that there are certain models, including Haah's code, in which all nontrivial excitations are fractons, whereas all mobile excitations are trivial (\(i.e.\) can decay directly to the vacuum). These models have been called "type 2\" fracton models. To the best of the author's present knowledge, no one has yet explicitly written a \(U(1)\) model which definitively has this property. But little systematic exploration has been done on \(U(1)\) models of rank higher than 2, and it seems likely that such "type 2\" models will be found somewhere in the higher spin hierarchy. Assuming that such models exist, we expect them to have slightly different finite-temperature behavior. Since there are no nontrivial mobile excitations, the system will have no screening ability at all on short timescales. We therefore expect the full long-range correlations of the ground state to survive until some long timescale at which fracton-on-fracton screening sets in and makes everything short-ranged. Such systems will therefore maintain a very strong "memory\" of their ground state behavior, even at finite temperature.
## Subdimensional Phases without Fractons
There are also higher rank spin liquid phases which possess subdimensional particle excitations, but not fractons. In a case like this, the fundamental charges have some degree of mobility, moving along a particular subspace, and can therefore rearrange their density to participate in screening. Furthermore, the different subspaces are all coupled via the gapless gauge mode, ensuring that all subspaces come to the same thermal equilibrium. We therefore expect phases like this will behave more like the conventional \(U(1)\) spin liquid and will be fairly normal in their screening and thermalization properties. As an example, let us consider a rank 2 \(U(1)\) spin liquid with 1-dimensional excitations. Our degrees of freedom will be the same as in the fracton model, a rank \(2\) symmetric tensor gauge field \(A_{ij}\). The Gauss's law for this theory is: \[\partial_i E^{ij} = \rho^j\] which defines a *vector* charge density. For the present purposes, the most important aspect of these charges is their conservation laws: \[\int d^3r\,\rho^j = \textrm{constant}\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\int d^3r\,(\vec{r}\times\vec{\rho})^j = \textrm{constant}\] These equations represent the conservation of charge and the conservation of the angular moment of charge, respectively. Unlike in the fracton model, the extra conservation law does not fully lock the charges in place, but rather constrains them to move only in the direction of their charge vector. Thus, the charges in this theory are 1-dimensional particles, forced to live their lives along a line determined by their internal charge vector. Let's work out precisely how screening works in this phase, again relying on some generalized electromagnetic results from Reference . We have a set of vector charges, \(\rho^j = \partial_i E^{ij}\), which have an electric field tensor \(E_{ij}\) falling off as \(1/r^2\), just as in conventional electromagnetism. Unlike previous theories, we no longer have a scalar potential for describing electrostatic situations. Instead, the potential formulation for this theory involves a vector-valued potential \(\phi_i\), which gives the electric field via \(E_{ij} =-\frac{1}{2}(\partial_i\phi_j + \partial_j\phi_i)\). This vector potential essentially represents the potential energy per charge in a given direction. The potential corresponding to a point charge \(p^j\) takes the form: \[\phi_p^j(r) = \frac{1}{8\pi}\bigg(\frac{(p\cdot r)r^j}{r^3}+3\frac{p^j}{r}\bigg)\] Let us now imagine a system at some finite temperature \(T\). We insert a test charge \(P^j\) into our system, then see how the thermal distribution responds to the extra potential. As in our earlier discussion of dipoles, the charge vectors of this theory are quantized in a lattice-dependent way. (For example, on the cubic lattice, the charge vector is labeled by three integers, \(\vec{p} = (n_x,n_y,n_z)\).) As before, we only account for screening by the minimal charges, which is the dominant effect. Let the thermally excited density of these minimal charges, in the absence of external perturbations, be denoted \(n_0(T)\propto e^{-m/T}\), where \(m\) is the mass of the minimal charges. After we introduce a perturbing potential \(\phi^j\), the density of species \(p^j\) will become \(n_0e^{-\beta p^j\phi_j}\approx n_0(1-\beta p^j\phi_j)\). The total potential near the test charge will be: \[\phi^j(r) = \phi^j_{bare}(r)-\sum_{p}\int dr'n_0\beta(p\cdot\phi(r'))\phi_p^j(r-r')\] where the sum is over the minimal charge vectors. Using the fact that \(-\frac{1}{2}(\partial^2\phi_p^j + \partial^j(\partial\cdot\phi_p)) = p^j\delta^{(3)}(r)\), we can then derive: \[\begin{aligned}
\begin{split}-\frac{1}{2}(\partial^2\phi^j + \partial^j(\partial\cdot\phi)) =-\frac{1}{2}&(\partial^2\phi_{bare}^j + \partial^j(\partial\cdot\phi_{bare})) \\ &-\sum_p n_0\beta(p\cdot\phi(r))p^j \end{split} \end{aligned}\] As in our previous discussion of dipole screening, we can write \(\sum_p p^ip^j = \alpha \delta^{ij}\) for some constant \(\alpha\). Also, assuming that the bare potential is that due to a point charge \(P^j\), the equation becomes: \[\partial^2\phi^j + \partial^j(\partial\cdot\phi) =-2P^j\delta^{(3)}(r) + 2 n_0\alpha\beta\phi^j(r)\] In momentum space, we then have: \[-k^2\phi^j(k)-k^j(k\cdot\phi(k)) =-2P^j + 2n_0\alpha\beta\phi^j(k)\] We can rewrite this equation in tensor language as: \[\bigg((k^2+2n_0\alpha\beta)\delta^{ij} + k^ik^j\bigg)\phi_i(k) = 2P^j\] Inverting for the potential, we obtain: \[\phi^i(k) = \frac{2}{(k^2 + 2n_0\alpha\beta)}\bigg( \delta^{ij}-\frac{k^ik^j}{2(k^2 + n_0\alpha\beta)}\bigg)P^j\] The transverse component \(\phi^j_\perp(k)\) (satisfying \(k\cdot\phi_\perp = 0\)) is given by: \[\phi_\perp^j(k) = \frac{2P^j}{2n_0\alpha\beta + k^2}\] while the longitudinal component \(\phi_\parallel\) (along the direction of \(k\)) is: \[\phi_\parallel^j(k) = \frac{P^j}{n_0\alpha\beta + k^2}\] In both cases, \(\phi(k)\) goes to a constant at \(k=0\), and the potential will fall off exponentially in real space. In either case, the screening length behaves as \(\lambda\propto n_0^{-1}\propto e^{m/T}\). Thus, at temperatures below \(T\sim m/\log L\), spin liquid physics will hold throughout the system. We see that the physics of screening in this phase closely mirrors that of the conventional \(U(1)\) spin liquid, without anything particularly exotic. Similar stories will hold for the other non-fractonic higher rank \(U(1)\) spin liquids. As long as the fundamental charges have some degree of mobility, they can quickly rearrange their density in such a way as to fully screen a test charge and kill all long-range interactions. Since there is only a strong-screening regime, the physics of these spin liquids is not quite as rich as their fractonic counterparts.
# Conclusion
In this paper, we have analyzed the finite-temperature screening behavior of higher rank \(U(1)\) spin liquids, with particular emphasis on a phase with both fractons and nontrivial mobile excitations. We have found that, as in their rank 1 counterparts, screening in these phases will allow for a smooth finite-temperature crossover to a trivial phase. Nevertheless, the exponentially large screening length at low temperatures allows for a realistic range of temperatures at which spin liquid physics still governs the system. Furthermore, for phases with fractons and mobile excitations, we have found that there are two distinct temporal regimes to the screening of charges. On short timescales, a charged disturbance (\(i.e.\) a fracton) is only weakly screened by the mobile excitations, leaving residual power-law correlations in the system. At much longer times, fractons can screen each other, and all correlations become short-ranged. The long timescales associated with screening and thermalization in these systems may be a useful tool for diagnosing the presence of such \(U(1)\) fracton phases.
# Appendix: Fracton Decay {#appendix-fracton-decay .unnumbered}
In the main text, we estimated the time it takes for a system of fractons to move around and reach thermal equilibrium, both with and without a thermal bath of dipoles. But how do we know that the fracton continues to exist at all over the relevant time scales? When there are other fractons around, a fracton can recombine with an appropriate set of neutralizing charges to annihilate into the vacuum. The energy gets carried away by the gapless gauge mode, which can freely propagate through the system, allowing for much faster thermalization. Let us estimate the timescale for a fracton decay process. First, we examine a state with a thermal distribution of dipoles. Each fracton in the system is screened by a cloud of such dipoles. As discussed in the main text, the composite object of a fracton plus its screening cloud is both charge-neutral and dipole-neutral, and can therefore decay directly into the vacuum. However, at low temperatures, this object is large and extended, of characteristic size \(\lambda\), the screening length. From previous studies of the lattice models for these phases, we know that the matrix element for such a state to decay to the vacuum behaves as \(e^{-(\lambda/a)^2}\), where \(a\) is the lattice spacing. The lifetime of the fractons then behaves as \(e^{(\lambda/a)^2}\propto e^{(a^2n_{0,d}\beta)^{-1}}\), where the density of thermally excited dipoles is \(n_{0,d}\propto e^{-m_d/T}\). Therefore, fracton decay is enormously slow at low temperatures: \[\tau_{decay} \sim e^{e^{m/T}}\] scaling as the exponential of an exponential. Compared to the other thermalization processes discussed in the text, we find that fracton decay is too slow to play a significant role in the equilibration of the system. Alternatively, let us start from the (non-thermal) state with fixed fracton density \(n\), where the typical spacing between fractons is roughly \(n^{-1/3}\). If all the fractons have the same sign of charge, then there is no fracton decay mechanism available. If there are a mix of positive and negative charges around, so that fracton decay can proceed, the lifetime of a fracton will be proportional to: \[\tau_{decay}\sim e^{\frac{1}{(an^{1/3})^{2}}}\] At low densities, this timescale is much larger than the thermalization time due to fracton motion, which we found behaves as a power law in \(n\). Once again, fracton decay is not an important process contributing to thermalization. | {'timestamp': '2017-09-11T02:03:52', 'yymm': '1706', 'arxiv_id': '1706.01899', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01899'} |
# Introduction
Electricity is non--storable, except marginally: any quantity which is consumed now must be produced now, and conversely. This means that the installed capacity must be sufficient to supply electricity when demand is maximal. As a consequence, part of this installed capacity will stay idle when demand is lower. This is the overcapacity problem, which is compounded by the fact that electricity consumption is far from stable, and little substitutable. People turn the heating on when it is cold, and the air conditioning on when it is warm, they turn the light on as soon as it gets dark and the home appliances when they come back from work. Unless they own a generator, there is no other source of energy they can turn to. So the collective demand for electricity must be satisfied, and satisfied right now. Since there are no stocks to dampen shocks and smooth discrepancy, adjusting supply to demand is a difficult task. One way to do this is to use prices. Very early on, power companies have hit upon the idea of making electricity more expensive in peak hours, so that consumers who are able to do so would switch their demand to off--peak periods. This falls naturally within the framework of Principal--Agent problems: the Principal (here the power company) offers a variety of contracts, and each Agent (consumer) picks the one which suits him best. It does not seem, however, that such an analysis is available at the present time, and the aim of the present paper is to fill this gap. We focus on the problem of finding an appropriate tariff: for a given installed capacity, or rather, for a given production function, how should a power company price electricity in order to maximise its profit? The company faces a variety of consumers, industrial users and domestic users, some of them are efficient, others less so. Some of them, for instance, live in insulated homes and need less electricity to achieve comfortable temperatures than others. The tariff the producer offers will be time--dependent and consumption--based. It will act in several ways:
- it will redirect part of the consumers to off-peak periods, by pricing properly the peak hours,
- it will avoid overly expensive production costs, by penalising higher consumptions,
- it will effectively exclude some of the users, who will find the proposed tariff too expensive, and who will look for better alternatives elsewhere.
The empirical effective response of individual agents to hourly pricing contracts is studied in or and is heterogeneous among the electricity consumers, as pointed out in or. Besides, more resilient measures of high electricity prices appear to be more efficient in practice for incentivising the reduction of electricity consumption, see or. In any case, the critical importance of the feedback information signals provided to the Agent has been highlighted in the empirical literature. In particular, the development of smart metering systems allows for the implementation of decentralised home automation systems, who partially regulate power consumption, as presented e.g. in or. | {'timestamp': '2018-06-20T02:11:18', 'yymm': '1706', 'arxiv_id': '1706.01934', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01934'} |
# Details of lattice model construction [\[app:model\]]{#app:model label="app:model"}
In this Appendix, we provided details of the construction of the lattice models realizing the decorated domain wall Floquet evolution described in the main text. We first construct a convenient lattice implementation of a \(2d\) CF phase, which forms the basis for the domain wall decoration in the \(3d\) lattice models of FTPMs.
## \(2d\) Chiral Floquet Model
To build chiral Floquet (CF) evolution \(H_{\text{CF},s,a}(t)\) utilized in the main text, we need to choose a particular implementation of the \(2d\) chiral Floquet unitary evolution of the \(p\)-state bosons. Previously constructed CF models based on applying a time-dependent sequence of boson SWAP operations have inconvenient properties for arranging onto arbitrary \(2d\) planar domains, and we will find it convenient to design an alternative (though topologically equivalent) implementation, whose edge acts as a uniform chiral edge-translation by one site, and which can be easily applied to a \(2d\) domain of arbitrary geometry. A key building block in this construction is an operator: \(C_{a,s}\) that cyclically permutes the p-state bosons around a square plaquette, \(a\), in either a right (\(s=+1\)) or left (\(s=-1\)) handed sense. For example, for a four site plaquette, \(a\), with sites labeled \(_A^D \square_B^C\): \[\begin{aligned}
C_{a,+} = \sum_{j_{A\dots D}=1}^p |j_D,j_A,j_B,j_C\>\<j_A,j_B,j_C,j_D| \equiv e^{-iH_aT} \end{aligned}\] The cyclic permutations, \(C_{a,s}\) can always be generated by a local Hamiltonian, \(H_{a}\), i.e. \(C_{a,s} = e^{-isH_{a}T}\), acting only on the spins in plaquette \(a\). To implement the CF evolution on the spin DWs, we can then label all of the plaquettes of the boson lattice by a number between \(1\) and \(4\) (see Fig. [\[fig:2dcf\]](#fig:2dcf){reference-type="ref" reference="fig:2dcf"}), and sequentially apply \(C_{a,s}\) on plaquettes of type \(1\), \(2\), \(3\), and then \(4\). The result of this four-step sequence is shown in Fig. [\[fig:2dcf\]](#fig:2dcf){reference-type="ref" reference="fig:2dcf"}b, for a \(3\times 3\) square. There is no motion for the site at the center of the square. The states of the edge sites are moved either \(0\), \(1\), or \(2\) sites along the edge in a chiral fashion (dashed arrows in Fig. [\[fig:2dcf\]](#fig:2dcf){reference-type="ref" reference="fig:2dcf"}b), such that one boson state is transferred across each point along the edge. This evolution differs from an ideal chiral edge translation only by a finite-depth local unitary evolution. We can remove this superficial difference by applying a 5th stage of the evolution, in which the boson states are swapped between neighboring sites shown with a dark gray oval in Fig. [\[fig:2dcf\]](#fig:2dcf){reference-type="ref" reference="fig:2dcf"}a. The resulting 5-step evolution implements an idealized chiral edge translation unitary, in which the bosons at the boundary are shifted by precisely one site in the right-handed direction. This idealized CF evolution can be implemented on arbitrary \(2d\) domains made from arbitrary edge sharing configurations of this minimal \(3\times 3\) unit. The precise implementation of step \(5\) depends on the local geometry of the \(2d\) domain (illustrative examples are given in Fig. [\[fig:2dcfexamples\]](#fig:2dcfexamples){reference-type="ref" reference="fig:2dcfexamples"}).
## \(3d\) Decorated Domain Wall Model
With the \(2d\) CF implementation in hand, we can begin to assemble the \(3d\) decorated domain wall model of a FTPM. The lattice model is formed from two types of degrees of freedom:
1. Spins-1/2, \(\vec{\sigma}_i\) that transform under TRS as \(\mathcal{T}\vec{\sigma}_i\mathcal{T}^{-1} = \sigma^x_i\vec{\sigma}_i^*\sigma^x_i\), and
2. p-state bosons, with an onsite Hilbert space spanned by a basis: \(\{|1\>,|2\>\dots|p\>\}\) that transforms trivially under TRS.
We arrange the spin-1/2, \(\sigma\) degrees of freedom on a layered triangular lattice with each layer being a vertically shifted copy of the one below it (Fig. [\[fig:ddwmodel\]](#fig:ddwmodel){reference-type="ref" reference="fig:ddwmodel"}). We take each \(\sigma\)-spin to be surrounded by a cube with a \(5\times 5\) grid of \(p\)-state boson sites on each face. In each layer, the boson cubes form a brick lattice around the spins, which has the advantage that the spin domain walls projected onto the boson cube faces will always contain an integer number of the elementary \(4\)-square-plaquettes used in the \(2d\) CF implementation (Fig. [\[fig:2dcf\]](#fig:2dcf){reference-type="ref" reference="fig:2dcf"}). This arrangement also avoids issues with point-like intersections between domain walls. Our strategy will be to apply the \(2d\) CF evolutions described in the previous section, to the boson plaquettes sitting on spin DWs. We define an orientation of the DWs point from down to up spins, and will evolve with the CF phase of right-handed chirality (with respect to the DW orientation). We further label the minimal 4-square boson plaquettes based on the direction of their normal vector: \(\pm x\), \(\pm y\), or \(\pm z\) (with \(\pm\) sign given by the DW orientation). Complications arise in "folding\" the \(2d\) CF evolution onto a closed \(3d\) surface. Namely, it is impossible to evolve all of the \(x\), \(y\), and \(z\) plaquettes simultaneously according to the 5-step CF evolution, without acting with \(C_{a,+}\) on multiple overlapping plaquettes at the same time. Since \(C_{a/b}\) commute only for disjoint plaquettes \(a\) and \(b\), this would spoil the desired zero-correlation length property of the model (i.e. render it not exactly solvable). To avoid this problem, we divide the CF evolution into two stages, evolving the \(x\) and \(y\) plaquettes in the first stage, and then the \(z\) plaquettes in the second stage (Fig. [\[fig:xyz\]](#fig:xyz){reference-type="ref" reference="fig:xyz"}). This division enables us to apply a \(C_a\)'s to a disjoint set of boson plaquettes at every step. In order to make the overall Floquet evolution TRS, we will need to further sub-divide the first stage into three steps:
1. Evolve the \(+\hat{x}\) and \(+\hat{y}\) facing plaquettes with the CF evolution. Denote this unitary evolution as \(U_{xy1}\).
2. Apply an appropriate set of SWAP operations to such that step xy1, and the next step, xy3, result in ideal chiral translations at the boundaries between xy and z surfaces (see Fig. [\[fig:corner\]](#fig:corner){reference-type="ref" reference="fig:corner"})
3. Evolve with \(U_{xy3}= \(\mathcal{T}^{-1} U_{xy1}\mathcal{T}\)^{-1}\)
Step xy3 is effectively the same as applying the CF evolution to the \(-\hat{x}\) and \(-\hat{y}\) oriented plaquettes. To see this, note that the time-reversal operators flip the spin-projectors in \(U_{xy1}\) so that \(\mathcal{T}^{-1} U_{xy1}\mathcal{T}\) implements a left-handed CF evolution as if the spin domain orientation were reversed. Hence, this will act on the negative \(xy\)-oriented plaquettes that were left out of step xy1. Lastly, the overall inversion in \(U_{xy3} = \(\mathcal{T}^{-1} U_{xy1}\mathcal{T}\)^{-1}\) switches the CF evolution back to the original right-handed one (though still acting on the \(-x\) and \(-y\) oriented plaquettes). Together with an appropriate choice of the SWAP operations in step xy2 (see Fig. [\[fig:corner\]](#fig:corner){reference-type="ref" reference="fig:corner"}), xy3 undoes the chiral motion at the boundary of the \(+\hat{x}\) and \(+\hat{y}\) plaquettes, leaving only a chiral motion around the z-plaquettes (which subsequently be undone in the second, \(z\), stage of the evolution). The virtue of dividing the xy-evolution into these steps is that it ensures that the overall evolution for the xy-stage: \[\begin{aligned}
U_{xy} = U_{xy3}U_{xy2}U_{xy1} \end{aligned}\] is manifestly time-reversal symmetric. Specifically, since \(\mathcal{T} U_{xy1} \mathcal{T}^{-1} = \mathcal{T} \(\mathcal{T}^{-1} U_{xy1}\mathcal{T}\)^{-1} \mathcal{T}^{-1} = U_{xy1}^{-1}\), and since the SWAP operations used in \(U_{xy2}\) are manifestly TRS (\(\mathcal{T} U_{xy2} \mathcal{T}^{-1} = U_{xy2}^\dagger\)), we verify: \(\mathcal{T} U_{xy}\mathcal{T}^{-1} = U_{xy}^\dagger\). To complete the construction, we need to remove the remaining chiral motion at the edge between the xy-facing and z-facing DW plaquettes. This is done by applying right-handed CF evolutions to the z-plaquettes. Again, to ensure TRS, it is convenient to divide this z-stage into two steps:
1. Evolve the \(+\hat{z}\) facing plaquettes.
2. Evolve with \(U_{z2} = \(\mathcal{T}^{-1} U_{z1}\mathcal{T}\)^{-1}\), which does the CF evolution of the same orientation to the \(-\hat{z}\) facing plaquettes in a way that is manifestly the time-reverse of step z1.
Finally, in order to combine the \(xy\) and \(z\) stages together in a way that is overall TRS, we should "sandwich\" the z-steps around the the xy steps as: \[\begin{aligned}
U(T) = U_{z1} U_{xy} U_{z2} \end{aligned}\] This unitary evolution implements the decorated domain wall picture of the FTPM phase described in the main text, while preserving the overall TRS. We note, in passing, that \(U(T)\) is unitarily equivalent to applying all the steps in sequence as shown in Fig. [\[fig:xyz\]](#fig:xyz){reference-type="ref" reference="fig:xyz"}: \(U_{z2}U_{z1} U_{xy} = U_{z2} U(T) U_{z2}^\dagger\). Changing between these two orderings simply amounts to a shift in our definition of the period, and a corresponding shift to the "center of inversion\" for the time-reversal operation. We can readily verify that the construction of \(U(T)\) results in a TRS evolution: \[\begin{aligned}
\mathcal{T} U(T) \mathcal{T}^{-1} &= \mathcal{T} U_{z1}\mathcal{T}^{-1} U_{xy}^\dagger \mathcal{T} U_{z2} \mathcal{T}^{-1} = U_{z2}^{\dagger} U_{xy}^\dagger U_{z1}^\dagger \nonumber\\ &= U(T)^\dagger \end{aligned}\] Crucially, each of these steps can be implemented by a local Hamiltonian involving projectors onto spin configurations multiplied by the local boson terms corresponding to the terms in the appropriate \(2d\) CF implementation described in the previous section. The evolution is easiest to picture when projected onto a \(2d\) plane. Fig. [\[fig:surfacedw\]](#fig:surfacedw){reference-type="ref" reference="fig:surfacedw"} shows the result of the \(z\) plaquette evolution for a fixed arrangement of spins. One can readily verify that, overall, every bulk site on the DW returns to itself after one period of this evolution (even at the corners and edges of various \(3d\) domain shapes, Fig. [\[fig:corner\]](#fig:corner){reference-type="ref" reference="fig:corner"}), so that the Floquet evolution is equivalent to the identity in the bulk. However, the intersection of a DW with a spatial boundary exposes a chiral edge (Fig. [\[fig:surfacedw\]](#fig:surfacedw){reference-type="ref" reference="fig:surfacedw"}), as required for the FTPM phase.
# Possible fermionic analogs [\[app:fermions\]]{#app:fermions label="app:fermions"}
In this section we investigate the possible connections of the ideas in the main text to fermionic SPT phases protected by TRS, whose surface states are characterized by time-reversal breaking domain walls that exhibit the chiral edge dynamics of a \(2d\) fermionic CF phase. Our current understanding of such fermionic generalizations is incomplete at present, and this appendix aims to assemble our current partial knowledge, lay out possible scenarios, and highlight open issues.
## Decorated domain walls with spinless fermions
A seemingly simple extension of the models described in the main text is to replace the \(p\)-state boson sites with complex (spinless) fermions described by creation operators \(f^\dagger_r\) on site \(r\). Since the on-site Hilbert space of a fermion has the same number of states as a \(p=2\) boson, this procedure would naively also produce a phase with \(\eta = 2\). In the presence of microscopic fermions, though, this \(\eta=2\) fermion phase seemingly requires charge conservation symmetry in addition to time-reversal symmetry for stability, since in superconducting systems, a purely \(2d\) Majorana CF phase with \(\nu = \sqrt{2}\) is possible. This construction appears to yield a Floquet topological insulator (FTI) protected by \(U(1)\) charge conservation and spinless time-reversal symmetry, \(\mathcal{T}\), with \(\mathcal{T}^2=1\) (class AI). We will refer to this phase as a "spinless fermion FTI\". In the absence of interactions, FTIs are described by a non-interacting Floquet band structure, whose possible topological features have exhaustively classified. For this class, there are no non-trivial Floquet-band TI phases, i.e. the spinless fermion FTI does not exist without interactions. Hence, we are left with three possibilities:
1. The spinless fermion FTI is topologically equivalent to a purely bosonic \(\eta=2\) phase,
2. the spinless fermion FTI is inequivalent to any bosonic phase, and instead constitutes an intrinsically interacting fermion Floquet SPT phase, or
3. there is a subtle (i.e. currently unknown) way in which the spinless fermion FTI surface states are not topologically protected.
While we do not currently have a definitive understanding of which option is correct, let us weigh some circumstantial evidence regarding each of these possible scenarios. Two observations speak in favor of scenario 1 (spinless fermion FTI = bosonic FTPM). First, we can employ a gedanken experiment that is frequently useful in equilibrium SPTs of inserting a \(\pi\)-flux (vison) into the surface state. In equilibrium, a non-trivial fermionic SPT will react non-trivially to such a \(\pi\)-flux (e.g. the flux will acquire a symmetry protected degeneracy or fractional symmetry charge) --otherwise one could proliferate such \(\pi\)-flux excitations and gap out the fermions at the surface, showing that the topological properties arise from purely bosonic degrees of freedom. The spinless fermion FTI phase, on the other hand, does not have a topological response to a \(\pi\)-flux, suggesting a topological equivalence to a purely bosonic system. A second piece of circumstantial evidence for scenario 1, is that in the absence of \(U(1)\)-number conservation symmetry, the \(\nu = \log 2\) CF phase of bosons and fermions are topologically equivalent. However, there are two possible reasons to doubt these arguments (supporting scenario 2, spinless fermion FTI \(\neq\) bosonic FTPM). First, the application of the \(\pi\)-flux proliferation to "gap\" out the fermion degrees of freedom is subtle in the context of highly excited states of a Floquet MBL system where energy is not conserved, and the topological properties come from highly excited dynamics, potentially involving excitations with arbitrary quasi-energy. Second, the demonstration of equivalence between boson and fermion CF phases with \(\nu = \log 2\) in Ref. , hinged on the absence of \(U(1)\) charge conservation, to show that the boson and fermion on-site Hilbert spaces can be made equivalent by tacking on auxiliary degrees of freedom with trivial dynamics. One can readily convince themselves that this trick cannot be done in a charge conserving way, so long as one has a finite on-site Hilbert space. Namely, in a fermion system, all bosonic degrees of freedom have even charge, whereas all fermionic degrees of freedom have odd charge. Hence, the maximal charge state in a fermionic site can never be equivalent to that of a bosonic site. This raises the more subtle possibility, that the fermion and boson phases may only be equivalent in a system with an infinite on-site Hilbert space (e.g. a quantum rotor model), though such an unbounded on-site Hilbert space may be problematic for MBL. Lastly, while we presently see no concrete issue with the fermionic decorated domain wall (DDW) model, there is potential cause to worry that there is some hidden obstruction that we have yet to identify (scenario 3). For example, in equilibrium, one could try to create a DDW model of an ordinary electronic TI, by decorating TRS-breaking magnetic domains with integer quantum Hall states of spinless fermions with \(\sigma^{xy} = \frac{e^2}{h}\). This would seemingly result in a model in which surface magnetic domains have a single chiral mode equivalent to the quantum Hall edge--the hallmark of an electronic TI with electromagnetic theta angle \(\theta_e=\pi\). However, in that context, it is known that the surface state is not protected, and that only spinful (Kramers doublet, \(\mathcal{T}^2=-1\)) electrons can form a stable TI phase. By analogy, it is conceivable that the "spinless FTI\" is not a stable topological phase, but rather, its surface state is not SPT protected. Instead, a non trivial SPT order requires spin-1/2 fermions to form. Such a spinful fermion FTI (class AII) does exist in the absence of interactions, and is characterized by a non-trivial Floquet band invariant. This spinful fermion FTI cannot be many-body localized without breaking time-reversal symmetry due to local Kramers degeneracies, however, it may occur as a long-lived pre-thermal phenomena. At the present, we are unable to definitively decide among these three scenarios, and raise this task as a challenge for future work.
## Spinful fermion Floquet topological insulator
In this section, we briefly outline the topological properties of a (non-interacting or prethermal) Floquet topological insulator of spin-1/2 electrons protected by charge conservation and spinful time-reversal symmetry (class AII). We will call this phase the "Floquet band TI\". This phase is classified by a non-trivial Floquet band invariant. In general, Floquet band structures are classified by two copies of the equilibrium band invariants. The first copy can be intuitively viewed as an equilibrium phase that is realized in a Floquet context, and has the usual equilibrium topological surface states at quasi-energy \(0\). Similarly, the extra Floquet phases can be viewed as a second set of equilbrium invariants for topological surface states at quasi-energy \(\pi\). This leads to an intuitive picture of the Floquet band TI surface states, as consisting of one Dirac cone \(0\) quasi-energy and another at \(\pi\) quasi-energy (strictly speaking, since there is no particle-hole symmetry, it is the quasi-energy difference between the two surface Dirac cones that is fixed at \(\pi\)). Namely, as was shown in Ref. (see also ), for any phase in which time-evolution by two periods can be implemented by an ordinary local, symmetric, and time-independent Hamiltonian \(U(2T) = e^{2iH_\text{eff}T}\) one can formally define an emergent dynamical \(\mathbb{Z}_2\) symmetry generator: \(g = U(T)e^{iH_\text{eff}T}\) where \(U(2T)\). The two surface Dirac cones have opposite symmetry "charge\" under this emergent dynamical symmetry. We can further understand the properties of this phase, via a gedanken experiment in which we insert a minimal flux magnetic monopole into the Floquet band TI. The familiar topological index theorems for Dirac cones guarantee that each surface Dirac cone contributes a charge-1/2 fermionic bound state to the monopole, described by annihilation operators \(\psi_{0,\pi}\) respectively. Crucially, the quasienergy difference between these modes is precisely: \(\delta \varepsilon = \pi\). The charge neutral monopole has two possible configurations, where one of the \(\psi_{0,\pi}\) is occupied and the other empty. These two configurations differ in quasi-energy by \(\pi\), or equivalently, the two states of the neutral monopole are degenerate with respect to time evolution by two periods, \(U(2T)\). This dynamical "degeneracy\" is protected by emergent dynamical \(\mathbb{Z}_2\) symmetry that descends from the time-translation symmetry of the drive. We close by remarking that this monopole gedanken experiment distinguishes the spinful Floquet band TI, from the putative spinless fermion FTI explained in the previous section. The latter does not respond in any non-trivial way to \(U(1)\) fluxes, and cannot be topologically equivalent to the Floquet band TI.
## Fractionalized generalizations
In the main text and previous sections, we have focused on short-range entangled bulk phases without anyon excitations, we may also consider consider "fractional\" analogs of these FTPM phases which can be accessed via a related decorated domain wall construction in which TRS breaking domains are decorated with radical CF phases, which would result in intrinsic 3D bulk topological order. The surface of a putative fractional DTI would then have an effective fractional value of the 3D TRS topological invariant \(\eta \in \prod_i \(\sqrt{p_i}\)^{n_i~\text{mod}~2}\), and surface states with effective chiral index that is a quartic root of a rational number, \(r\), \(\nu_\text{surface} = \pm\log \sqrt[4] r\). While such phases should be stable as metastable pre-thermal "ground-states\", there is a potential complication for realizing an MBL state in disordered versions of these systems. Namely, the bulk 3D topological order would exhibit string-like gauge-flux excitations, which, in the idealized zero-correlation length limit, would result in an exact degeneracy growing exponentially with the number of intersections between the string excitations and fluctuating spin-DWs. Upon moving away from the fine-tuned integrable limit, in the related case of 2D radical CF phases, these degeneracies were lifted either by a spontaneous breaking of time-translation symmetry, or a breakdown of MBL. Whether simply breaking time-translation symmetry in the above construction is sufficient to produce a stable MBL phase remains an open question for future study.
# General classification of bosonic Floquet topological paramagnets [\[app:generalclassification\]]{#app:generalclassification label="app:generalclassification"}
Recall that the static equilibrium classification of 3d bosonic phases with time reversal symmetry is \(\mathbb{Z}_2 \times \mathbb{Z}_2\), with one \(\mathbb{Z}_2\) generated by the in-cohomology (\(eTmT\)) SPT state and the other by the beyond-cohomology (\(eFmF\)) SPT state . In this appendix we will argue that the in-cohomology state can be realized by a many-body localizable (MBL) Hamiltonian, whereas the beyond-cohomology one cannot. A general proof that all in-cohomology states are MBL was given in . Here, we present a related, complementary argument that also allows us to argue that the beyond-cohomology state cannot be localized. Thus, the proposed full classification of bosonic Floquet topological paramagnets will include the in-cohomology SPT state, together with the infinite family of models constructed in this paper: i.e. the new infinite family replaces the beyond-cohomology state in the Floquet classification They key property of any in-cohomology SPTs in spatial dimension \(d \geq 1\) is the fact that its ground state can be disentangled into a symmetric product state by a finite depth unitary \(V\) that commutes with all symmetry generators: \[\begin{aligned}
|\Psi_{g.s.}\rangle = V |\Psi_{prod}\rangle. \end{aligned}\] Here \(|\Psi_{g.s.}\rangle\) is the SPT ground state, \(V\) is a finite depth circuit of local unitary operators, and \[\begin{aligned}
|\Psi_{prod}\rangle = \otimes_j |\psi_j\rangle \end{aligned}\] is a tensor product state over the sites \(j\) of the system of symmetric states \(|\psi_j\rangle\). Indeed, the disentangling circuit \(V\) can be constructed directly for the zero correlation length models introduced in, and its existence is a universal property of the SPT phase. This ground-state construction was generalized to all excited states of an MBL system in. Note that only the entire circuit \(V\) is symmetric--the individual unit-depth unitary steps making up \(V\) will not, by themselves, be symmetric for a non-trivial SPT phase. If the symmetry group \(G\) is onsite and Abelian, then each site Hilbert space decomposes as a sum of orthogonal one dimensional representations \(\alpha\). Letting \(P_j^\alpha\) denote the projector onto the \(\alpha\) representation at site \(j\), we see that the ground state \(|\Psi_{g.s.}\rangle\) is the unique state annihilated by \(\{V P_j^0 V^\dag\}\), where \(0\) denotes the symmetric representation generated by \(|\psi_j\rangle\). Thus the Hamiltonian \[\begin{aligned}
H_{\text{MBL}} = \sum_{j,\alpha} J_{j,\alpha} V P_j^{\alpha} V^\dag \end{aligned}\] with suitably chosen (random or quasi-periodic) couplings \(J_{j,\alpha}\) is a Hamiltonian with a full set of local conserved quantities believed to be in or close to an MBL phase, realizing \(|\Psi_{g.s.}\rangle\) as an eigenstate. Conversely, if an SPT ground state can be realized as the eigenstate of a symmetric Hamiltonian with a full set of symmetric local conserved quantities, then we conjecture that such a symmetric disentangling circuit \(V\) must exist. Indeed, in this case one can define \(V\) via the unitary transformation mapping projectors onto on-site degrees of freedom to projectors onto the conserved quantities ("l-bits\") of the MBL system, which is a locality preserving, and quasi-local transformation. Indeed, showed that in one dimension such locality-preserving unitary operators are always finite depth quantum circuits up to a generalized translation. Assuming some version of this result holds in higher dimensions, and the generalized translation can be argued to act trivially (i.e. take product states to product states), this would imply the existence of a symmetric finite depth circuit disentangling the ground state. Although we cannot prove this at this time, we conjecture that this is true in the case of 3 spatial dimensions and time reversal symmetry. In fact, we can almost take the existence of such a \(V\) as the definition of symmetry preserving MBL. Certainly any counterexamples would have a very different structure than known MBL phases, and would likely require modifying several common definitions of MBL. In particular, the in-cohomology, \(eTmT\) state has a TRS disentangling circuit \(V\) and can be many-body localized in this fashion. On the other hand, we claim that for the beyond cohomology state, no such circuit \(V\) exists, which by the argument above strongly suggests that it cannot be MBL. We will now argue by contradiction, that no such circuit exists for the beyond cohomology, \(eFmF\), state. Suppose that a symmetric circuit, \(V\) did exist, which could disentangle the bulk of the \(eFmF\) state in the absence of boundaries. Then, consider a system with a boundary (e.g. a solid rectangular block with a surface), and as the surface state, take the time reversal symmetric \(eFmF\) state. This is a gapped surface topological order with an anomalous realization of time reversal symmetry: namely, any truly 2d realization of the \(eFmF\) state necessarily has a chiral central charge equal to \(4\) modulo \(8\). In this open geometry, we can define a truncation of the hypothetical \(V\) to the bulk, which disentangles the bulk but not the boundary (generically it is not possible to disentangle the boundary of a nontrivial SPT with a finite depth unitary). The putative \(V\) would be TRS in the bulk. Namely, writing the time-reversal operator as \(\mathcal{T} = U_{\mathcal{T}}K\) with \(U_{\mathcal{T}}\) being a product of on-site unitary operators and \(K\) being complex conjugation (in some basis), \(V\) and \(U_{\mathcal{T}}^\dag V^* U_{\mathcal{T}}\) have the same action on operators localized in the bulk of the system. Note that time reversal property of the disentangling circuit \(V\) is not the same as that of the time evolution operator, which requires an extra inverse. Thus we see that the operator \(V^{-1} U^\dag V^* U\) acts only on the spins localized near the surface, i.e. is a surface operator. The key point now is that \(V\) would break time reversal symmetry near the surface and maps an SPT eigenstate to a product state in the bulk tensored with a surface state \(|\Psi_s\rangle\). This surface state \(|\Psi_s\rangle\) is now a truly 2d realization of the \(eFmF\) state, and hence has a nonzero chiral central charge of \(c\) equal to \(4\) modulo \(8\). On the other hand, if we had disentangled the bulk of the original (TRS) state with the time-reversed partner \(U^\dag V^* U\) of \(V\), we would have obtained the time-reversed \(eFmF\) state, with chiral central charge \(-c\). Also, these two time-reversed incarnations of the \(eFmF\) surface state are mapped into each other by the 2d locality preserving surface operator \(V^{-1} U^\dag V^* U\). This leads to a contradiction, as follows. Let us denote by \(eFmF_+\) and \(eFmF_-\) the surface states with chiral central charge \(c\) and \(-c\) respectively, and let \(W=V^{-1} U^\dag V^* U\) be the 2d locality preserving operator that maps one to the other. Now stack each of \(eFmF_+\) and \(eFmF_-\) with another copy of \(eFmF_-\). Augmenting \(W\) by the identity on this second copy of \(eFmF_-\), we obtain an operator \(\tilde{W}\) that maps \(eFmF_+ \times eFmF_-\) to \(eFmF_-\times eFmF_-\). However, \(eFmF_+ \times eFmF_-\) is simply the quantum double of \(eFmF\), and necessarily has a parent Hamiltonian equal to a sum of local commuting projectors. Conjugating these local commuting projectors by the locality-preserving operator \(\tilde{W}\) we would obtain a local commuting projector Hamiltonian for \(eFmF_-\times eFmF_-\), which is impossible because \(eFmF_-\times eFmF_-\) has nonzero chiral central charge. This is the desired contradiction. | {'timestamp': '2017-06-08T02:00:24', 'yymm': '1706', 'arxiv_id': '1706.01888', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01888'} |
null | null |
# Introduction
In the recent past, devices such as desktop computers and laptops were our best means for staying connected to the Internet community and to have access to online services. However, with enhanced capabilities, low cost and user-friendly interfaces, smartphones have become people's first choice to stay connected to the Internet. This is equally credited to the publicly available Internet which facilitates users to access their device contents regardless of their location. According to a recent study , over 80% of cell phone users globally using smartphones furthermore 2.56 billion people are estimated to own smartphones by the end of 2018. With this rapid growth trend in the usage of smartphones, device control, and data security have become extremely important. Not only personal and professional contact information is stored such as email IDs, passwords, banking details i.e. credit and debit card information, photos and videos, but users also store their sensitive and critical information in their smartphones . If the device is stolen, the stored information can be used to create many troubles not only for the user but also for the individuals in their contacts. In order to secure this access, it is important to develop fast and accurate methods for legitimate user identification and block-out impostors. Ideally, these methods should detect an impostor from the moment a device is stolen with acceptable accuracy within a minimum time span. Current identification methods such as secret PIN number (SPN) or lock codes are not only risky-since tools to extract SPNs are easily available on the Internet, but also difficult to use-particularly for the elderly and physically impaired users who find entering PIN codes or screen patterns difficult. Another factor which restricts the security of SPNs-based mechanisms is that smartphones are mostly used in public places with many other people around. This increases the chances for the codes to be found-out by potential attackers. To overcome these issues, fast and secure methods are required to intelligently verify the legitimate user and to block-out impostors. In the literature hereby referred to, several solutions were proposed addressing implicit user identification without involving the user, such as keystroke-based user identification , touch screen biometrics , application set fingerprints , hybrid user identification methods-such as accelerometers and gyroscopes , and gait based user identification . However, these solutions only discuss one aspect of user identification, either software or hardware. The first thing attackers do after stealing a smartphone is to walk away from the actual user. Considering this fact, the gait-based legitimate user identification methods are most efficient, as they detect the impostor from their walking or running patterns at that same time and thus can trigger an alarm to keep the real user informed. This type of legitimate user identification method protects not only the smartphone itself, as well as the stored data when the device is stolen by swiftly triggering an alarm. However, due to several contributing factors, the implementation of such real-time identification is quite challenging. There are 2 key important features in a gait-based legitimate user identification system:
1. The user can perform the same ambulatory sub-activity differently at different times.
2. Various users require a distinctive set of features for user identification: male and female have different physical characteristics, therefore, will have explicit identification processes.
This study focuses on the idea of identifying a smartphone user by applying different walking patterns, hereby referred to as sub-activities. Furthermore, it is assumed that the phone is freely placed without any particular orientation inside any of the user's pants pockets. Data on walking patterns are recorded using an embedded triaxial, where we intentionally limited the scope to just walking, in order to understand how fast and accurate can ambulatory sub-activity-based legitimate user identification be. The aim of this research is to propose a semi-controlled environment system in which we overcome the limitations of users' age, gender, jeans style (loose or tight) and walking style (we intentionally asked users to walk differently in various times to investigate the ambulatory activity performed by each user). In this regard, the aim of our current work is to investigate several research questions relevant to building a walking-based legitimate user identification system in real life:
1. Does the Extended Sammon Projection (ESP), a non-linear unsupervised feature selection method improve the identification accuracy more than the other existing and well-studied unsupervised feature selection methods such as Principal Component Analysis (PCA)?
2. Does kernel-based Extreme Learning Machine (KELM) an effective classifier for the non-linear signal-based user identification method?
3. How to achieve real-time user identification in practice? Since our goal is to develop an algorithm which identifies the user in real time, thus computation complexity is extremely important. System performance measurements ought to be considered to balance the trade-off between accuracy and computational cost.
4. Does the data variation affect the performance of the legitimate user identification process?
Our major research obligations are towards answering the above questions using ESP and KELM methods. In this work, we introduce a novel two-fold one-subject-cross-validation based legitimate user identification system, thus addressing the above questions within a sub-activity recognition-based legitimate user identification method on smartphones, in a real-time environment. The remainder of this paper is organized as follows: Section 2 explains related works; Section 3 describes the system design; Section 4 presents the system validation and the discussion on results, and finally, section 5 concludes with possible future directions.
# Related Work
The objective of this work is to provide convenience in using smartphones differently from explicit identification, by using sensor data information. Therefore, hereby we present some key related works on this field which can be categorized into two groups: implicit user identification and multiple modality biometrics. Shi, et al. proposed a Sen-guard method for user identification, offering continuous and implicit user identification service for smartphone users. This method leverages the sensors available on smartphones, e.g. voice, multi-touch and location; these sensors are processed together in order to get the user's identification features implicitly. Explicit identification is performed only when there is an important evidence of change in the user activity, which is not real-time upto some extent. In recent years, several other implicit identification approaches have been proposed leveraging smartphone's sensor devices such as accelerometer , touch screen , GPS and microphone . T. Feng, et al. proposed to extract finger motion speed and acceleration of touch patterns as features. Luca, et al. suggested to directly compute the distance between pattern traces using the dynamic time warping algorithm. Sae-Bae, et al. present 22 special touch patterns for user identification, most of which involve all five fingers simultaneously. They then computed dynamic time warping distance and Freshet distance between multi-touch traces. Frank, et al. studied the correlation between 22 analytic features from touch traces and classified them using k nearest neighbors and support vector machines. Shahzad, et al. explained the use of touchscreen patterns as a secure unlocking mechanism at the login screen. Moreover, the idea behind the behavior-based model is that the person's habits are a set of its attributes; therefore, each event (activity) has a correlation between two fundamental attributes: space and time. In addition, the architecture proposed in , utilizes the resources found in the smartphone devices, such as user calls, user schedule, GPS, device battery level, user applications, and sensors. A similar methodology has also been adopted in . Clarke, et al. proposed smartphone's user perception of identification, in which results showed the system implicitly and continuously performing user identification in the background. Koreman, et al. recommended continuous multiple model-based approaches for user identification. Mantyjarvi, et al. used an accelerometer in television remote controls to identify individuals. Gafurov, et al. and Cuntoor. et al. suggested an experimented user identification using gait analysis and recognition. Jakobsson, et al. put forward another unique implicit user identification framework by using recorded phone call history and location for continuous user identification. However, these approaches present several propositions for legitimate user identification, but to some extent, all required some additional information and source. Casale, et al. proposed a user verification and authentication method using gait as a biometric unobtrusive pattern. A four-layer architecture was built around the geometric concept of a convex hull. This was a twofold method, in which first, a general activity identification based classifier is personalized for a specific user based on their walking patterns, and second, where the author verifies whether the user is authorized or not by using the one-class classification method. The proposed architecture is able to improve robustness to the outliers, account for temporal coherence information and most importantly for non-convex shapes. The most important fact about this system was not its non-user-friendliness, thus the user can only operate this system in specific and controlled environments. Mantyjarvi studied a user's identification utilizing portable devices from gait signals acquired with three-dimensional accelerometers. They originally proposed three approaches: data distribution statistics, correlation, and frequency domain to identify the subject while walking at different speeds: fast, normal and slow, where the accelerometer device was placed on the user's belt only at the back. The identification process by this method was novel but only limited to walking by the same users and with limited variations. Similar studies for user identification were proposed in using external accelerometers. To some extent, these studies are innovative but solely rely on external accelerometers with limited activity or sub-activities. Conclusively, all previous works require users to either perform predefined touch patterns or for the data to be collected under controlled experimental environments, which might not be a real representation of common user interactions. For this situation, the proposed gait-based legitimate user identification work is an interesting application alternative for the process of identifying explicit and continuous legitimate user or impostor in a semi-controlled environment; hence, overcomes the smartphone's limitations in power consumption and cost. Moreover, today's research is towards the smartphone's emergence of these kinds of identification mechanisms. We worked on combining different approaches to deliver a more reliable legitimate user identification model using a built-in accelerometer for different sub-activities in the semi-controlled environment. In short, several studies have experimented user identification using gait recognition as a possible identification method. Our proposed approach utilizes acceleration signals and detects users by their way of walking under different sub-activities in a semi-controlled environment, and for this, a motion-recording device is used in order to measure the acceleration according to the three axes outlined in .
# System Design
Nowadays, smartphones are equipped with a variety of motion sensors which are useful for monitoring the device movements such as, tilt, rotation, shaking, and swinging. Two of these are the accelerometer (ACC) and the linear acceleration (LACC) sensors. This work explores built-in sensors to validate the sensor's capacities for legitimate user identification processing. To this end, the first task of our proposed system is to collect data using two different enabled sensors, namely, ACC and LACC. The second task is the feature extraction and analysis based on both the time and frequency domain features. The third task is feature selection and the fourth task is the classification for legitimate user identification-based on performed sub-activities while walking. Finally, the fifth task is towards the evaluation of our proposed method on a publicly available dataset.
## Data Collection
In this work, we have used two different datasets, the first data is collected using an Android smartphone. In this regards, we have collected raw signals from two different sensors, as the user performed daily sub-activities. These sub-activities included walking while the smartphone was intentionally placed in one of the subject's pants pockets, e.g. back right pocket (BRP), back left pocket (BLP), front right pocket (FRP) and front left pocket (FLP). We recorded these sub-activities without any constraints on the smartphone's orientation inside of any of the user's pockets. The first dataset was gathered from four users each one performing individual sub-activities each day; each sub-activity was performed at least twice a day. Thus, the compiled data is from one same user, doing the same activity on different days, for an entire month. Initially, we stored these raw signals in a micro SD card, which we later transferred to a computer for further analysis. It is worth pointing out since different smartphone models have different sampling rates, therefore, in order to control the data collection process and better validation and generalization, we have used 50Hz sampling rate instead to use highest sampling rate within different smartphones . A sample raw signal representation of each sub-activity for both sensors are shown in Figure [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}. The second dataset was acquired from 16 healthy users between the ages of 19 to 48 years old. The dataset was gathered from all users by using a linear acceleration sensor (LACC) with a constant rate of 50Hz while they were wearing a smartphone on their waist. The dataset was pre-processed by applying noise filters, using 2.56 sec windows with a 50% overlap. The second experimental dataset is an extended version of the UCI human activity recognition process using a smartphone dataset and can be freely downloaded from the UCI website .
## Feature Extraction
The accelerometer sensor generates time series signals which are highly fluctuating and oscillatory in nature , thus making the legitimate user identification more difficult. Therefore, it is compulsory to gather the nontrivial signals from the raw data through the feature extraction process. Particularly, we divided the raw signal into several equal size windows to control the flow rate, hence passing fewer data to the system to be able to extract all meaningful information. Given a sampling rate of 50Hz, we initially chose a window size of 50 samples for both sensors as shown in Figure 1. The selected window size provides enough data to be able to extract the quality features while ensuring a fast response . Our research analyzed the individual's user data applying time series modeling techniques to understand the behavior in each of the user's physical patterns. Time series analysis reveals unusual observations as well as particular data patterns citeWei1991. There are three commonly used models to perform time series analysis: moving average , auto-regressive and a combination of both the moving average and the auto-regressive models . We explored with the use of auto-correlation (AC) and partial auto-correlation (PAC) coefficients to identify the model for our data. These coefficients revealed the pattern of each datum and indicated the best model for our data. Each data model is determined based on the characteristics of the theoretical AC and PAC. Samples of AC and PAC for both sensors are shown in Figure [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}. The fitting process for the time series model is calculated by estimating the parameter values based on a previously selected model. Moving average and auto-regressive parameter estimations require an iteration process , and for that, we follow the \"box Jenkins model estimation\" due to its flexibility for the inclusion of both the moving average and the auto-regressive models . The parameters and the model need to be verified to ensure that the estimated results are statistically significant . Detailed analysis revealed that the time domain features, including the coefficients from the time series model, provided the same accuracy as the frequency domain features . Therefore, in this work, we have extracted both the time and frequency domain features. The extracted features are: mean, median, variance, standard deviation, inter-quartile, auto-correlation, partial correlation, coefficients of the auto-regressive model, coefficients of moving the average model, coefficient of moving average auto-regressive model and wavelet coefficients. These features are extracted from each axis of the three-dimensional acceleration signals. In total 72 features were extracted from each window. Prior to the feature extraction, a moving average filter of order three was employed for noise reduction purposes.
## Feature Selection
The output of an embedded motion sensor depends upon the position of the smartphone while walking. This could result in a high within class variance . Therefore, it is at the same time desirable to improve both the discriminatory power of the features and achieve dimensionality reduction, by employing an optimum method. Hence, our third important step towards legitimate user identification is feature selection. The advantages of the feature selection process are to avoid the curse of dimensionality , as well as to reduce the abundant, irrelevance, misleading and noisy features, but above all, to be able to reduce the system's cost pertaining to run-time applications . In addition to the above, the main purpose is to increase the accuracy of the resulting model within a short time. In recent years, several feature selection algorithms have been proposed, including the filter based approach, the wrapper method, the principal components analysis (PCA), the linear discriminant analysis (LDA) , KLDA a kernelized version of the LDA and SWLDA a stepwise linear discriminant analysis method . Filter methods are based on discriminating criteria which are relatively independent of classification and use minimum redundancy and maximum relevance for feature selection. These methods are comparatively fast, scalable and provide good computational complexity, but ignore the interaction with the classifier. Alternatively, wrapper methods utilize the classifier as a black box to obtain a subset of features based on their predictive power. These methods interact with the classifier to optimize the feature's subset. The main disadvantages of these methods are their dependency on the classifier, who makes the classifier selection a key important process. In different prospects, the LDA seeks a linear combination of features and characterizes two or more types. Among the three different discriminant analysis approaches, the number of dimensions returned by both the LDA and KLDA depends on the number of classes, which limit the use of LDA and KLDA in user identification systems. However, in this work, we consider only unsupervised non-linear feature selection methods. In this regard, the most commonly used method is PCA. Through this work, we have introduced an unsupervised and fully automatic feature selection method named extended Sammon projection (ESP) for the first for smartphone-based activity recognition and user identification. The ESP is one of the most successful non-linear metric multidimensional scaling methods and it projects the high dimensional space into a lower dimensional space while preserving the structure of inter-point distances in a high dimensional space within a lower dimensional space. Consequently, let us assume \(d_{i,j}\) be the distance between two adjacent samples \(x_i\) and \(x_j, i \neq j\) from the original space and \(d_{i,j}^*\) be the distance between two samples \(x_i^*\) and \(x_j^*\) in the mapped space; thus, the Sammon stress measure, which is also known as the Sammon error or simple error function E can be defined as; \[\begin{aligned}
E = \frac{1}{\sum_{i=1}^{n-1} \sum_{j = i+1}^n d_{i,j}} \sum_{i=1}^{n-1} \sum_{j = i+1}^n \frac{d_{i,j}^*-d_{i,j}}{d_{i,j}^*} \end{aligned}\] where \(d_{i,j}\) is the Euclidean distance and this error measurement is minimized by using the steepest and gradient decent methods, respectively. \[\begin{aligned}
x_{ik}^* (t + 1) = x_{ik}^*(t)-\alpha \frac{\partial E(t) \setminus \partial x_{ik}^* (t)}{\mid \partial^2 E(t) \setminus \partial^2 x_{ik}^*(t) \mid} \end{aligned}\] \[\begin{aligned}
x_{ik}^* (t + 1) = x_{ik}^*(t)-\alpha \frac{\partial E(t)}{\partial x_{ik}^* (t)} \end{aligned}\] In both cases \(x_{ik}^*\) is the \(k^{th}\) coordinate of the position of point \(x_i^*\) in the lower dimensional space. Since the Steepest descent method has issues at the inflection points, where the second-order derivative appears to quite small , therefore, we set \(\alpha\) between 0.3 0.4 as the optimal value using a grid operation between \[0, 1\], but there is no reason to expect this given range to be optimal for all problems and datasets.
## Classifier
Neural networks have quite diverse real-life applications and among different neural network approaches, extreme learning machines (ELM) have better generalization capabilities and a fast learning speed . An ELM is a single hidden layer feed-forward neural network which randomly determines the initial parameters of weights input and hidden biases with simple activation functions . Among the factors influencing learning performance, the hidden neurons are very important to improve generalization capabilities. Moreover, ELMs with a tune-able activation function were proposed to handle the data dependency on hidden neurons. However, the selection of suitable combinations for activation functions is still a big question within the research community. Therefore, kernelized ELMs are known to improve the generalization capabilities, when the feature mapping function of hidden neurons is unknown. However, the parameters for the kernel function need to be selected carefully and in order to improve the generalization performance for real-time applications, such as smartphone-based legitimate user identification, the kernel parameters need to be tuned carefully. In our work, the said parameters are optimized through the swarm optimization based method . In ELMs, the initial parameters of the hidden layer need not be tuned and almost all nonlinear, piece-wise continuous functions can be used as hidden neurons. Therefore, for \(N\) distinct samples \((x_i, t_i) \mid x_i \in R^n, t_i \in R^m, i \in 1,2,3..., N\), and the output function with L hidden neurons can be expressed as; \[\begin{aligned}
f_L(x) = \sum_{i=1}^L \beta_i h_i(x) = h(x)\beta \end{aligned}\] where \(\beta = [\beta_1, \beta_2,..., \beta_L]\) be the output weights between the hidden layer and the output neurons; \(h(x) = [h_1(x), h_2(x),..., h_L(x)]\) be the output vector of the hidden layer that maps the input space to the feature space . The output weights and training error should be minimized to enhance the generalization capabilities. The least square solution to the minimization problem can be expressed as: \[\begin{aligned}
\beta = H^T \Big(\frac{1}{C} + HH^T \Big)^{-1}T \end{aligned}\] where \(H\) be the hidden layer output matrix, T is the expected output matrix, and \(C\) is the regularization coefficient. Thus, the training model output can be expressed as: \[\begin{aligned}
f(x) = h(x) H^T \Big(\frac{1}{C} + HH^T \Big)^{-1} T \end{aligned}\] For an unknown mapping \(h(x)\), the kernel and the output function \(f(x)\) can be defined as; \[\begin{aligned}
M = HH^T; m_{i,j} = h(x_i)h(x_j) = \kappa(x_i, x_j) \end{aligned}\] \[\begin{aligned}
f(x) = \Big[\kappa (x, x_1) ,..., \kappa (x, x_N)\Big]\Big(\frac{1}{C} +M\Big)^{-1} T \end{aligned}\] where \(\kappa(x,y)\) is the kernel function and can be expressed as in expression (9), where \(a\) and \(b\) are adjustable kernel parameters, which play an important role in generalization performance and are updated using the cross-validation process. \[\begin{aligned}
\kappa (x, y) = cos \bigg(\frac{\|x-y\|^2}{a}\bigg) exp \bigg(\frac{\|x-y\|^2}{b}\bigg) \end{aligned}\]
# Experimental Settings, Results, and Discussion
To validate our proposed user identification system, we conducted different experiments on two different datasets as explained in the previous section. Based on first (our-own) dataset, the first experiment aims to show the sub-activities-based legitimate user identification system, in which we conducted detailed comparison (with and without feature selection) on two different unsupervised feature selection methods namely PCA and ESP. Our second experiment explains the identification behavior with different numbers of windows within the best settings obtained from our first experiment. Finally, our third experiment shows the computational cost for our proposed legitimate user identification system. Based on second (publicly available smartphone-based physical activity recognition) dataset, our first experiment presents a user identification method using one-subject-leave-out-cross-validation process during a walking activity, with different numbers of windows, in which the features are processed through ESP method. The second experiment explains the confidence interval for the computational cost of our proposed legitimate user identification system.
## Experimental Settings
The values utilized for the different parameters pertaining to the ESP, PCA, and KELM methods were optimized using the cross-validation process. These optimized setting were used during each case of user identification considering each sub-activity, as they provided the best results. In all cases, the wavelet kernel is used and the kernel parameters updated according to the practical swarm method. For the feature selection process, and in order to keep both methods consistent, the same numbers of PCs are selected as the number of features selected by the ESP. The intention behind to select the same number of features was to make this model reliable and consistent. Prior to the classification, all obtained features were normalized and bounded within \[0, 1\].
## Experimental Results and Discussion
In the introduction part we have stated four important research questions for legitimate user identification on smartphone and based on our findings; yes, it is possible to identify a smartphone's legitimate user by analyzing their walking patterns when the device is freely placed in any of their pants pocket. Furthermore, the unsupervised ESP feature selection method significantly boosts identification performance. The KELM method uses 72 features for user identification when the case is without feature selection; both feature selection methods are thus able to reduce this number down to a different number of features (5, 10, 15, 20, 25, 30, 35, and 40) keeping always the most informative features. Based on our findings we observe that the ESP together with the KELM classifier is more robust and accurate than the existing well studied unsupervised feature selection method such as PCA. Our first experiment details the process of our proposed scheme for analyzing the behavior of different numbers of samples per window (i.e. 25, 50, 75, 100, 125, 150, 175, and 200), on a legitimate user identification process within each sub-activity without feature selection method. Figure [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"} shows the user identification accuracy of a legitimate user being identified with a different number of samples per window, for both sensors' data. From these results, we observe that the BR pocket produced better results with 25 samples per window (80% to 88% confidence interval with the LACC and ACC sensors respectively). On average, we can see that, the ACC sensor outperformed than the LACC for legitimate user identification within each sub-activity without feature selection. In our second experiment, we investigated the average identification for all users and each sub-activity within a fixed size window (50 samples per window), for both feature selection methods with a different number of features. Table [\[Tab1\]](#Tab1){reference-type="ref" reference="Tab1"} lists the results (user identification accuracy across all users and each sub-activity with a 99% confidence interval) for the case of the PCA-KELM and ESP-KELM. According to the results obtained in Table [\[Tab1\]](#Tab1){reference-type="ref" reference="Tab1"}, the best performance was obtained by the KELM when the normalized features were extracted by both sensors and processed with the ESP method. The PCA performed slightly better than without the feature selection process, but on average, there was no such difference with and without the feature selection process for the PCA case. All these results were obtained by using 50 samples per window during the feature extraction process.
Figure [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"} shows the average accuracy obtained through ESP using a different number of windows as explained earlier and 30 number of selected features as 30 number of features provide best average results in our previous experiments for all pockets data. For all subsequent experiments, we have fixed 30 number of features selected by the ESP method. From Figure [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}, we can observe that the performance has significantly improved for the legitimate user identification for both sensors, each with a different number of samples. The average accuracy increased from 70% to 98% for 50 samples per window, which is a significant improvement for any legitimate user identification system. From these results, we conclude that the BLP and FRP have some variations according to the number of samples per window; however, this variation is not enough to exclude these sub-activities from our experimental setup, except the ones with 75 and 100 numbers of samples per window. This degradation happens due to the sudden change in the users walking patterns. In future research, we will further investigate legitimate user behaviors, while performing the same sub-activity to minimize the ambiguity to the legitimate user identification method. In our third experiment, we discussed the computational cost in terms of time comparison, pertaining to our previous experiments presented in Figures [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"} and [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}. Figures [\[Fig5\]](#Fig5){reference-type="ref" reference="Fig5"} and [\[Fig6\]](#Fig6){reference-type="ref" reference="Fig6"} show the computational time for KELM with and without using the ESP method; with different numbers of samples per window for each sub-activity and sensor individually. As shown in above results, the different number of samples per window (25 to 50 samples per window) for both sensors, the computational cost exhibits the huge difference, thus, indicating that identification has a strong influence on the computational time of KELM with the number of samples. However, when we increased the size of samples from 50 to 75 or even to 100, 125, 150, 175, and 200, both systems took almost the same time to complete the legitimate user identification process.
As shown in Figures [\[Fig7\]](#Fig7){reference-type="ref" reference="Fig7"} and [\[Fig8\]](#Fig8){reference-type="ref" reference="Fig8"}, the computational cost gradually increases as the size of the windows decrease. Therefore, dealing with such high computational time becomes an important issue. There are certain possibilities which can be used to overcome this problem. One way is to use a lightweight feature selection method such as PCA, although the problem with the PCA is its incompetency for statistical results. There is another immediate solution for such problems, and it is to further divide each window into other sub-windows. However, this presents another challenge, how to conduct the windowing process, so that it does not decrease the performance of the legitimate user identification process. We conducted our last experiment on a publicly available LACC sensor-based dataset, hereby this experiment, we investigated the one-subject-leave-out-cross-validation within the different size of windows. Table [\[Tab2\]](#Tab2){reference-type="ref" reference="Tab2"} lists the results of user identification and computational time to identify individual users where features are selected through the ESP method and the selected features are classified through the KELM method. The obtained results corroborate the effectiveness of our proposed methodology for real-time applications. From these results, we observe that the feature selection and classification methods significantly increases the accuracy for legitimate user identification within each sub-activity for both sensors; additionally, outperforming on publicly available smartphone-based physical activity recognition datasets. Figures [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"} and [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"} presents the 99% confidence intervals, pertaining to the average legitimate user identification by using the pairwise T-test between groups with and without feature selection data at the 99% confidence level. Looking at Figure [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}, significant statistical results are clearly seen, showing that the KELM method performs much better when selected features of data are used in all cases: an 80% and 88% to 99% performance increase. This leads us to prefer the use of the feature selection method in future applications. The KELM method provides acceptable performance and has numbers of other attractive features to its applicability. In term of the applicability within a smartphone system, the KELM has smaller confidence intervals, implying it has more reliability in training models. Since the structure of the network in the KELM is fixed, it has a better training and lower variance. This implies that systems using a connected network should keep a fixed connective structure, in order to increase its reliability. Holding a constant connective structure is a good feature for the network due to hardware constraints. Any potential manufacturer to capture this method into a chip, as a trainable network, needs such a constant size structure. These chips could take away from the phone's main CPU, thus increasing the speed of the legitimate user identification process. This would allow for greater device security, by not allowing for software-based attacks on this method, only hardware based manipulations. These hardware operations would require access to the smartphone, hence making such attacks subject to the device's defense.
# Conclusion and Future Work
This study substantiates the idea to be able to detect a legitimate smartphone user-based on their walking patterns through various sub-activities in a semi-controlled environment. Having used a commonly available mass-market consumer hardware, such as our experimental platform, we have demonstrated the global applicability of our proposed method with minimal accuracy variations. The KELM requires minimal battery consumption, and in order for us to run it as a practical application, we need to limit the number of required samples; and based on our results we observed that the KELM together with the ESP is the best methods in terms of accuracy and computational cost to a certain extent. However, the KELM is found to be the best examination method, in order to reduce battery consumption when performing checks to the sensor array. A key advantage of our proposed system is that the data for each user is collected on different days with different orientations and locations, different jeans styles which significantly helps to understand the characteristic behavior of individual user, and significantly improves the training and testing accuracy within a short time intervals, which is an important feature for any real-time legitimate user identification system. There are many ways to extend our current work. This work demonstrated that accelerometer data is more useful than linear accelerometer data, but it is possible that a fusion of accelerometer and linear accelerometer data will yield improved results. We have also been experimenting with more realistic features, which capture specific elements of a user's gait, and plan to investigate if these features can yield additional improvements. Another one of our goals for future work is to expand the evaluation of the proposed system so that it is applied to more real-life situations. Thus, we plan to expand our user base significantly, increase the diversity of the users (especially with respect to gender and age), and evaluate how the system operates when the training and test samples are collected over lower window size, which is an important future direction for our current work, in which we will split the size of the smallest windows (25 or 50 samples per window) in to even smaller windows within the classification process, but for this, there is another challenge: how to maintain performance, within each sub-window, and how to control the computational complexity during real-time deployment. One final goal is to incorporate this technology into a real-time system. | {'timestamp': '2018-12-03T02:23:13', 'yymm': '1706', 'arxiv_id': '1706.01739', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01739'} |
# Introduction
A Boolean function is called a *threshold function* (also known as *linearly separable* or a *halfspace*) if there exists a hyperplane separating true and false points of the function. Threshold functions play fundamental role in the theory of Boolean functions and they appear in a variety of applications such as electrical engineering, artificial neural networks, reliability theory, game theory etc. (see, for example, ). We study the problem of teaching threshold functions in the context of on-line learning with a helpful teacher. Speaking informally, *teaching* an unknown function \(f\) in a given class is the problem of producing its *teaching* (or *specifying*) set, i.e. a set of points in the domain which uniquely specifies \(f\). In the present paper, the universe is the set of threshold functions and a specifying set for \(f\) is a subset \(S\) of the points of the Boolean cube such that \(f\) is the only threshold function which is consistent with \(f\) on \(S\). It is not difficult to see that in the worst case the specifying set contains all the \(2^n\) points of the Boolean cube. However, in some cases, a threshold function \(f\) can be specified by a smaller set, for instance, when \(f\) depends on all its variables and is *positive* (or *increasing*), i.e. a function where an increase of a variable cannot lead to a decrease of the function. In this case, \(f\) can be specified by the set of its *extremal points*, i.e. its maximal false and minimal true points, of which there are at most \(\binom{n+1}{\left\lfloor \frac{n+1}{2}\right\rfloor}\). Moreover, this description can also be redundant, i.e. sometimes a positive threshold function \(f\) can be specified by a proper subset of its extremal points. The minimum cardinality of a teaching set of \(f\), i.e. the minimum number of points needed to specify \(f\), is the *specification number* of \(f\). The maximum specification number over all functions in a class is the *teaching dimension* of the class. showed that the specification number of a threshold function with \(n\) variables is at least \(n+1\). proved that this bound is attained for so-called nested functions by showing that positive nested functions contain precisely \(n+1\) extremal points. They also conjectured that for all other threshold functions with \(n\) variables the specification number is strictly greater than \(n+1\).
#### Our contribution {#our-contribution .unnumbered}
As our first result, we disprove the conjecture of by showing that for any \(n \geq 4\) there exist threshold functions with \(n\) variables which are non-nested and for which the specification number is \(n+1\). To state our second result, we observe that for positive nested functions the specifying set coincides with the set of extremal points. This is not the case in our counterexamples to the above conjecture. Therefore, our negative resolution of the conjecture leaves open the question on the number of extremal points: is it true that for any positive threshold function different from nested, the number of extremal points is strictly greater than \(n+1\)? In this paper, we answer this question positively. Moreover, we prove a slightly more general result dealing with so-called linear read-once functions, which is an extension of nested functions allowing irrelevant variables (see Section [2](#sec:pre){reference-type="ref" reference="sec:pre"} for precise definitions). More formally, we prove that a positive threshold function \(f\) with \(k \geq 0\) relevant variables has exactly \(k+1\) extremal points if and only if \(f\) is linear read-once. Our solution is based on revealing an underlying structure of the set of extremal points.
#### Related work {#related-work .unnumbered}
Upper and lower bounds and the average value for the specification number of a threshold Boolean function are obtained in. A number of papers are devoted to the teaching dimension for the class of threshold functions of \(k\)-valued logic, i.e. halfspaces defined on the domain \(\{0,1,\dots,k-1\}^n\). Upper bounds for the teaching dimension are obtained in. A tight lower bound is stated in. The special case \(n=2\) is considered in. The problem of teaching is closely related to the problem of learning . Learning threshold functions with membership or/and equivalence queries is studied by. Special case \(n = 2\) is considered in. Learning threshold Boolean functions with small weights is investigated in. Teaching or/and learning different classes of read-once (or repetition-free) functions are considered in. The importance of linear read-once functions in learning theory is evidenced, in particular, by their connection with special types of decision lists.
#### Organization of the paper {#organization-of-the-paper .unnumbered}
All preliminary information related to the paper can be found in Section [2](#sec:pre){reference-type="ref" reference="sec:pre"}. The refutation of the conjecture is presented in Section [3](#sec:cn){reference-type="ref" reference="sec:cn"}. Section [4](#sec:ex){reference-type="ref" reference="sec:ex"} contains the results about extremal points of a threshold function. Section [5](#sec:con){reference-type="ref" reference="sec:con"} concludes the paper with a number of open problems.
# Preliminaries {#sec:pre}
Let \(B = \{ 0, 1 \}\). For a *point* \(\vect{x} \in B^n\) we denote by \((\vect{x})_i\) the \(i\)-th coordinate of \(\vect{x}\), and by \(\notl{\vect{x}}\) the point in \(B^n\) with \((\notl{\vect{x}})_i = 1\) if and only if \((\vect{x})_i = 0\) for every \(i \in [n]\). Let \(f=f(x_1, \ldots, x_n)\) be a Boolean function on \(B^n\). For \(k \in [n]\) and \(\alpha_k \in \{ 0, 1\}\) we denote by \(f_{|x_k=\alpha_k}\) the Boolean function on \(B^{n-1}\) defined as follows: \[f_{|x_k=\alpha_k} (x_1, \ldots, x_{k-1}, x_{k+1}, \ldots, x_n) = f(x_1, \ldots, x_{k-1}, \alpha_k, x_{k+1}, \ldots, x_n).\] For \(i_1, \ldots, i_k \in [n]\) and \(\alpha_1, \ldots, \alpha_k \in \{ 0,1 \}\) we denote by \(f_{|x_{i_1} = \alpha_1, \ldots, x_{i_k} = \alpha_k}\) the function \((f_{|x_{i_1} = \alpha_1, \ldots, x_{i_{k-1}} = \alpha_{k-1}})_{|x_{i_k} = \alpha_k}\). We say that \(f_{|x_{i_1} = \alpha_1, \ldots, x_{i_k} = \alpha_k}\) is the *restriction* of \(f\) to \(x_{i_1} = \alpha_1, \ldots, x_{i_k} = \alpha_k\). We also say that a Boolean function \(g\) is a *restriction* of a Boolean function \(f\) if there exist \(i_1, \ldots, i_k \in [n]\) and \(\alpha_1, \ldots, \alpha_k \in \{ 0,1 \}\) such that \(g \equiv f_{|x_{i_1} = \alpha_1, \ldots, x_{i_k} = \alpha_k}\), i.e., \(g(\vect{x}) = f_{|x_{i_1} = \alpha_1, \ldots, x_{i_k} = \alpha_k}(\vect{x})\) for every \(\vect{x} \in B^{n-k}\).
Following the terminology of, we say that \(\vect{x} \in B^n\) is a *true point* of \(f\) if \(f(\vect{x}) = 1\) and that \(\vect{x} \in B^n\) is a *false point* of \(f\) if \(f(\vect{x}) = 0\).
## Positive functions and extremal points
By \(\below\) we denote a partial order over the set \(B^n\), induced by inclusion in the power set lattice of the \(n\)-set. In other words, \(\vect{x} \below \vect{y}\) if \((\vect{x})_i = 1\) implies \((\vect{y})_i=1\). In this case we will say that \(\vect{x}\) is *below* \(\vect{y}\). When \(\vect{x} \below \vect{y}\) and \(\vect{x} \neq \vect{y}\) we will sometimes write \(\vect{x} \sbelow \vect{y}\).
For a positive Boolean function \(f\), the set of its false points forms a down-set and the set of its true points forms an up-set of the partially ordered set \((B^n, \below)\). We denote by
- the set of maximal false points,
- the set of minimal true points.
We will refer to a point in \(Z^f\) as a *maximal zero of \(f\)* and to a point in \(U^f\) as a *minimal one of \(f\)*. A point will be called an *extremal point of \(f\)* if it is either a maximal zero or a minimal one of \(f\). We denote by
- the number of extremal points of \(f\).
## Threshold functions
The inequality \(w_1 x_1 + \ldots + w_n x_n \leq t\) is called *threshold inequality* representing function \(f\). It is not hard to see that there are uncountably many different threshold inequalities representing a given threshold function, and if there exists an inequality with non-negative weights, then \(f\) is a positive function. Let \(k \in \mathbb{N}, k \geq 2\). A Boolean function \(f\) on \(B^n\) is *\(k\)-summable* if, for some \(r \in \{ 2, \ldots, k \}\), there exist \(r\) (not necessarily distinct) false points \(\vect{x}_1, \ldots, \vect{x}_r\) and \(r\) (not necessarily distinct) true points \(\vect{y}_1, \ldots, \vect{y}_r\) such that \(\sum_{i=1}^r \vect{x}_i = \sum_{i=1}^r \vect{y}_i\) (where the summation is over \(\mathbb{R}^n\)). A function is *asummable* if it is not \(k\)-summable for all \(k \geq 2\).
## Linear read-once functions and nested functions
A Boolean function \(f\) is called *linear read-once* if it is either a constant function, or it can be represented by a *nested formula* defined recursively as follows:
1. both literals \(x\) and \(\notl{x}\) are nested formulas;
2. \(x \orl t\), \(x \andl t\), \(\notl{x} \orl t\), \(\notl{x} \andl t\) are nested formulas, where \(x\) is a variable and \(t\) is a nested formula that contains neither \(x\), nor \(\notl{x}\).
showed that the class of linear read-once functions is precisely the intersection of threshold and read-once functions. A linear read-once function is called *nested* if it depends on all its variables. For example, the function \((x_1 \orl x_2) x_3 x_5\) considered as a function of 5 variables \(x_1, \ldots, x_5\) is linear read-once, but not nested, since \(x_4\) is an irrelevant variable. If this function is considered as a function of 4 variables \(x_1, x_2, x_3, x_5\), then all its variables are relevant and therefore the function is also nested. It is not difficult to see that a linear read-once function \(f\) is positive if and only if a nested formula representing \(f\) does not contain negations.
## Specifying sets and specification number
Let \(\mathcal{F}\) be a class of Boolean functions of \(n\) variables, and let \(f \in \mathcal{F}\).
Let \(\mathcal{H}_n\) be the class of threshold Boolean functions of \(n\) variables. and later showed that the specification number of a threshold function of \(n\) variables is at least \(n+1\).
It was also shown in that the nested functions attain the lower bound.
## Essential points
In estimating the specification number of a threshold Boolean function \(f \in \mathcal{H}_n\) it is often useful to consider essential points of \(f\) defined as follows.
Clearly, any specifying set for \(f\) must contain all essential points for \(f\). It turns out that the essential points alone are sufficient to specify \(f\) in \(\mathcal{H}_n\). Therefore, we have the following well-known result.
## The number of essential points versus the number of extremal points
It was observed in that in the study of specification number of threshold functions, one can be restricted to positive functions. To prove Theorem [\[th:nested-spec\]](#th:nested-spec){reference-type="ref" reference="th:nested-spec"}, first showed that for a positive threshold function \(f\), which depends on all its variables, the set \(Z^f \cup U^f\) of extremal points specifies \(f\). Then they proved that for any positive nested function \(f\) of \(n\) variables \(|Z^f \cup U^f| = n+1\). In addition to proving Theorem [\[th:nested-spec\]](#th:nested-spec){reference-type="ref" reference="th:nested-spec"}, also conjectured that nested functions are the only functions with the specification number \(n+1\) in the class \(\mathcal{H}_n\).
In the present paper, we disprove Conjecture [\[con:conjecture\]](#con:conjecture){reference-type="ref" reference="con:conjecture"} by demonstrating for every \(n \geq 4\) a threshold non-nested function of \(n\) variables with the specification number \(n+1\). On the other hand, we show that the conjecture becomes a true statement if we replace 'specification number' by 'number of extremal points'. In fact, we prove a more general result saying that a positive threshold function \(f\) with \(k\) relevant variables is linear read-once if and only if it has exactly \(k+1\) extremal points. For this purpose, the following special type of functions appears to be technically useful.
In what follows, we will need the next two observations, which can be easily verified.
# Non-nested functions with small specification number {#sec:cn}
In this section, we disprove Conjecture [\[con:conjecture\]](#con:conjecture){reference-type="ref" reference="con:conjecture"}. To this end, we show in the following theorem that the minimum value of the specification number is attained in the class of threshold functions not only by nested functions.
# Extremal points of a threshold function {#sec:ex}
The main goal of this section is to prove the following theorem.
We will prove Theorem [\[th:extremal_main\]](#th:extremal_main){reference-type="ref" reference="th:extremal_main"} by induction on \(n\). The statement is easily verifiable for \(n=1\). Let \(n > 1\) and assume that the theorem is true for functions of at most \(n-1\) variables. In the rest of the section we prove the statement for \(n\)-variable functions. Our strategy consists of three major steps. First, we prove the statement for split functions in Section [4.2](#sec:split){reference-type="ref" reference="sec:split"}. This case includes linear read-once functions. Then, in Section [4.3](#sec:ns1){reference-type="ref" reference="sec:ns1"}, we prove the result for non-split functions \(f\) which have a variable \(x_i\) such that both restrictions \(f_{|x_i=0}\) and \(f_{|x_i=1}\) are split. Finally, in Section [4.4](#sec:ns2){reference-type="ref" reference="sec:ns2"}, we consider the case of non-split functions \(f\), where for every variable \(x_i\) of \(f\) at least one of the restrictions \(f_{|x_i=0}\) and \(f_{|x_i=1}\) is non-split. In this case, the proof is based on a structural characterization of the set of extremal points, which is of independent interest and which is presented in Section [4.1](#sec:str){reference-type="ref" reference="sec:str"}.
## The structure of the set of extremal points {#sec:str}
We say that a maximal zero (*resp.* minimal one) \(\vect{y}\) of \(f(x_1, \ldots, x_n)\) *corresponds to a variable* \(x_i\) if \((\vect{y})_i = 0\) (*resp.* \((\vect{y})_i = 1\)). A pair \((\vect{a}, \vect{b})\) of points in \(B^n\) is called *\(x_i\)-extremal for \(f\)* if
1. \(\vect{a}\) is a maximal zero of \(f\) corresponding to \(x_i\);
2. \(\vect{b}\) is a minimal one of \(f\) corresponding to \(x_i\); and
3. \((\vect{a})_j \geq (\vect{b})_j\) for every \(j \in [n] \setminus \{ i \}\).
Let \(g=g(y_1, \ldots, y_n)\) be a positive function, and let \(\{ y_{i_1}, \ldots, y_{i_k} \}\) be a subset of the relevant variables of \(g\). For every variable \(y_{i_j}\), \(j \in [k]\) we fix an \(y_{i_j}\)-extremal pair \((\vect{a}_{i_j}, \vect{b}_{i_j})\). Now we define a graph \(H(g, y_{i_1}, \ldots, y_{i_k})\) as an undirected graph with vertex set \(\{ \vect{a}_{i_j}, \vect{b}_{i_j}~|~j \in [k] \}\) and edge set \(\{ \{\vect{a}_{i_j}, \vect{b}_{i_j}\}~|~j \in [k] \}\). We call \(H(g, y_{i_1}, \ldots, y_{i_k})\) an extremal graph and observe that this graph is defined not uniquely.
## Split functions {#sec:split}
## Non-split functions with split restrictions {#sec:ns1}
## Non-split functions without split restrictions {#sec:ns2}
Due to Lemmas [\[lem:split\]](#lem:split){reference-type="ref" reference="lem:split"} and [\[lem:non-split-split-projection\]](#lem:non-split-split-projection){reference-type="ref" reference="lem:non-split-split-projection"} it remains to show the bound for a positive threshold non-split function \(f = f(x_1, \ldots, x_n)\) such that for every \(i \in [n]\) at least one of \(f_0 = f_{|x_i=0}\) and \(f_1 = f_{|x_i=1}\) is non-split. Assume without loss of generality that \(x_n\) is a relevant variable of \(f\), and let \(f_0 = f_{|x_n=0}\) and \(f_1 = f_{|x_n=1}\). We assume that \(f_0\) is non-split and prove that \(f\) has at least \(k+2\) extremal points, where \(k\) is the number of relevant variables of \(f\). The case when \(f_0\) is split, but \(f_1\) is non-split is proved similarly. Let us denote the number of relevant variables of \(f_0\) by \(m\). Clearly, \(1 \leq m \leq k-1\). Exactly \(k-1-m\) of \(k\) relevant variables of \(f\) became irrelevant for the function \(f_0\). Note that these \(k-1-m\) variables are necessary relevant for the function \(f_1\). By the induction hypothesis, the number \(r(f_0)\) of extremal points of \(f_0\) is at least \(m+2\). We introduce the following notation:
1. --the set of maximal zeros of \(f\) corresponding to \(x_n\);
2. --the set of all other maximal zeros of \(f\), i.e., \(P_0 = Z^f \setminus C_0\);
3. --the set of minimal ones of \(f\) corresponding to \(x_n\);
4. --the set of all other minimal ones of \(f\), i.e., \(P_1 = U^f \setminus C_1\).
For a set \(A \subseteq B^n\) we will denote by \(A^*\) the restriction of \(A\) into the first \(n-1\) coordinates, i.e., \(A^* = \{ (\alpha_1, \ldots, \alpha_{n-1})~|~(\alpha_1, \ldots, \alpha_{n-1},\alpha_{n}) \in A \text{ for some } \alpha_n \in \{0,1\} \}\). By definition, the number of extremal points of \(f\) is \[\label{eq:r_f} r(f) = |C_0| + |P_1| + |C_1| + |P_0| = |C_0^*| + |P_1^*| + |C_1^*| + |P_0^*|.\] We want to express \(r(f)\) in terms of the number of extremal points of \(f_0\) and \(f_1\). For this we need several observations. First, for every extremal point \((\alpha_1, \ldots, \alpha_{n-1}, \alpha_n)\) for \(f\) the point \((\alpha_1, \ldots, \alpha_{n-1})\) is extremal for \(f_{\alpha_n}\). Furthermore, we have the following straightforward claim.
In contrast to minimal ones of \(f_0\), the set of maximal zeros of \(f_0\) in addition to the points in \(C_0^*\) may contain extra points, which we denote by \(N_0^*\). In other words, \(Z^{f_0} = C_0^* \cup N_0^*\). Similarly, besides \(C_1^*\), the set of minimal ones of \(f_1\) may contain additional points, which we denote by \(N_1^*\). That is, \(U^{f_0} = C_1^* \cup N_1^*\).
From Claim [\[cl:p1_p0\]](#cl:p1_p0){reference-type="ref" reference="cl:p1_p0"} we have \(r(f_0) = |Z^{f_0} \cup U^{f_0}| = |C_0^*| + |N_0^*| + |P_1^*|\), which together with ([\[eq:r_f\]](#eq:r_f){reference-type="ref" reference="eq:r_f"}) and Claim [\[cl:n0_p0\]](#cl:n0_p0){reference-type="ref" reference="cl:n0_p0"} imply \[\label{eq:rf} \begin{split} r(f) & = |C_0^*| + |P_1^*| + |C_1^*| + |P_0^*| = |C_0^*| + |P_1^*| + |C_1^*| + |N_0^*| + |P_0^* \setminus N_0^*| \\ & = r(f_0) + |C_1^*| + |P_0^* \setminus N_0^*|. \end{split}\] Using the induction hypothesis we conclude that \(r(f) \geq m+2 + |C_1^*| + |P_0^* \setminus N_0^*|\). To derive the desired bound \(r(f) \geq k+2\), in the rest of this section we show that \(C_1^* \cup P_0^* \setminus N_0^*\) contains at least \(k-m\) points.
Recall that there are exactly \(s = k-1-m\) variables that are relevant for \(f_1\) and irrelevant for \(f_0\). We denote these variables by \(x_{i_1}, \ldots, x_{i_s}\). Let \(H\) be an extremal graph \(H(f_1, x_{i_1}, \ldots, x_{i_s})\) defined in such a way that all its vertices belong to \(C_1^* \cup P_0^* \setminus N_0^*\). Such a graph exists by Claim [\[cl:proper_pairs\]](#cl:proper_pairs){reference-type="ref" reference="cl:proper_pairs"}. By Lemma [\[lem:acyclic\]](#lem:acyclic){reference-type="ref" reference="lem:acyclic"} the graph \(H\) is acyclic, and hence it has at least \(s+1\) vertices. Therefore, the set \(C_1^* \cup P_0^* \setminus N_0^*\) has at least \(s+1 = k-m\) points. This conclusion establishes the main result of this section.
# Conclusion and open problems {#sec:con}
In this paper we studied the cardinality and structure of two sets related to teaching positive threshold Boolean functions: the specifying set and the set of their extremal points. First, we showed the existence of positive threshold Boolean functions of \(n\) variables, which are not linear read-once and for which the *specification number* is at its lowest bound, \(n + 1\) (Theorem [\[prop:1\]](#prop:1){reference-type="ref" reference="prop:1"}). An important open problem is to describe the set of all such functions. Second, we completely described the set of all positive threshold Boolean functions of \(n\) relevant variables, for which the number of *extremal points* is at its lowest bound, \(n + 1\). This is precisely the set of all positive linear read-once functions (Theorem [\[th:extremal_main\]](#th:extremal_main){reference-type="ref" reference="th:extremal_main"}). It would be interesting to find out whether this result is valid for all positive functions, not necessarily threshold. In other words, is it true that a positive Boolean function of \(n\) relevant variables has \(n + 1\) extremal points if and only if it is linear read-once? Finally, we ask whether the acyclic structure of the set of extremal points of a positive threshold function \(f\) can be helpful in determining the specification number of \(f\). | {'timestamp': '2017-06-07T02:06:34', 'yymm': '1706', 'arxiv_id': '1706.01747', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01747'} |
# Introduction
Coreference Resolution is the task of identifying groups of phrases in a text that refer to the same discourse entity. Such referring phrases are called mentions, a set of mentions that all refer to the same discourse entity is called a chain. Annotated corpora are important resources for developing and evaluating automatic coreference resolution methods. Turkish is an agglutinative language and Turkish coreference resolution poses several challenges different from many other languages, in particular the absence of grammatical gender, the possibility of null pronouns in subject and object position, possessive pronouns that can be expressed as suffixes, and ambiguities among possessive and number morphemes, e.g., can be analysed as or as , depending on context. No coreference resolution corpus exists for Turkish so far. We here describe the result of an effort to create such a corpus based on the METU-Sabanci Turkish Treebank (Say, Zeyrek, Oflazer, and Özge, 2004; Atalay, Oflazer, and Say, 2003; Oflazer, Say, Hakkani-Tür, and Tür, 2003) which is, to the best of our knowledge, the only publicly available Turkish Treebank. Our contributions are as follows.
We describe two stages of annotation: in , annotators created mentions and chains, which did not yield sufficient inter-annotator agreement. In , mentions were given to annotators who created only chains. We collected on average more than ten independent annotations per document for each document in the METU-Sabanci Turkish Treebank.
We describe annotator profiles and adjudication, which was done semi-automatically in and fully automatically in . We describe the principles of our automatic adjudication tool which uses a voting-like approach. Such an automatic approach is possible because we collected enough | {'timestamp': '2018-08-01T02:12:08', 'yymm': '1706', 'arxiv_id': '1706.01863', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01863'} |
# Introduction
alignment, also known as facial landmark localization, is an essential step in many computer vision methods such as face verification and facial motion capture. The majority of face alignment methods proposed in the last several years are based on Cascaded Shape Regression (CSR), which was first proposed in. CSR is usually initialized by placing the average face shape in the location provided by the face detector. Starting from this initialization, the face shape is refined in a fixed number of iterations. In each iteration features are extracted from the regions around each landmark of the face shape. The extracted features are then used in a regression method to estimate a correction to the current positions of the landmarks. In Ren et al. proposed Local Binary Features (LBF) where binary descriptors are created based on regression forests built on simple pixel difference features. The proposed features were incorporated into the CSR framework, which lead to a very fast method, running at the speed of 3000 frames per second on standard desktop hardware. Our face alignment method is based on LBF, but instead of standard regression forests with pixel features, we use novel K-cluster Regression Forests with Weighted Splitting (KRFWS) and Pyramid HOG (PHOG) features. PHOG features consist of HOG taken over the same area with progressively larger cell sizes, more details can be found in. K-cluster Regressions Forests (KRF) were originally proposed in for head pose estimation and car orientation estimation. In KRF, at each split, the target space is divided into K clusters, which locally minimize the loss function of the tree. Each K-fold split is performed using a separate k-class linear SVM classifier. The advantage of this type of splitting is that it allows to split on multidimensional features such as HOG, whereas standard regression forest splitting uses only scalar, one dimensional features. In this work we propose K-cluster Regression Forests with Weighted Splitting (KRFWS) by introducing weights on individual samples for classification at each split. We test KRFWS on the Pointing'04 dataset, for which KRF was originally designed, and show improved accuracy. Recently Li et al. proposed Affine-Transformation Parameter Regression for Face Alignment (APR), which estimates a 2D affine transform that refines the current face shape estimate. As opposed to the standard regression step in CSR, which refines the positions of the individual landmarks, APR applies a rigid transform. In it was shown that embedding such alignment before and between standard cascaded shape regression iterations improves face alignment accuracy. In this work we propose two major extensions to the APR method. First, we show that using KRFWS instead of linear regression improves the accuracy of the estimated transform. Second, we propose 3D-APR, where, we estimate a 3D transform of the current face shape estimate. The use of a 3D transform allows for the compensation of out-of-plane rotation of the head, which is not possible with standard APR.
# Methods
## K-cluster Regression Forests with Weighted Splitting
In this chapter we describe K-cluster Regression Forests with Weighted Splitting (KRFWS) which are an extension of K-cluster regression forests (KRF). Regression forests, like all other regression methods, map a point in the input space \(\pmb{X}\) onto a point in the target space \(\pmb{Y}\). KRF differ from other similar methods in the way splitting is performed and in the fact that in KRF each node has \(K \geq 2\) child nodes, whereas in most methods each node has 2 children. The aim of splitting is to find a rule that once applied to the input space \(\pmb{X}\), partitions the target space \(\pmb{Y}\) in a way that minimizes the loss function of the tree. KRF use the most common loss function, which is the sum of squared errors (SSE) of the child nodes. As explained in, the SSE of the child nodes of any given node is locally minimized by a partition defined by the result of k-means clustering on the target space of samples in that node. In KRF, the problem of splitting is defined as a classification problem. To preserve the partition defined by the k-means clustering, a classifier is trained at each split. This is achieved by applying a \(K\) class one-versus-rest linear SVM classifier on the input space of samples in each node. The advantage of this approach is that the splits can be performed on multidimensional features. This in turn facilitates the use of image features, for example PHOG, as the input space. In most cases the separating hyperplane produced by the SVM does not partition the input set perfectly. In practice this means that some of the samples will not be forwarded to the correct child node. The misclassified samples increase the value of the loss function and make the splitting problem more difficult in the child nodes. We propose to reduce the impact of this problem by giving more importance to the samples that, should they be misclassified, would have the largest impact on SSE. This is achieved by assigning each sample \(i\) in the current node \(T_j\) a weight \(w_i\in [0, 1]\), where 0 would signify that the sample is of no importance. The weights are then used as an input to a weighted SVM classifier included in the LIBLINEAR package. Below we describe our method for generating the sample weights for splits with \(K=2\) clusters. While the method does generalize to any number of clusters, the description of the generalized procedure is beyond the scope of this short article. Given a node \(T_j\) divided into two clusters \(C_1\) and \(C_2\) with their respective centroids \(c_1\) and \(c_2\), the weights are defined as follows: \[\label{eq:projection} v_i = \left(y_i-\frac{c_2 + c_1}{2}\right)^T \cdot \frac{c_2-c_1}{\lVert c_2-c_1 \rVert} ,\] \[\label{eq:normalization} w_i = \frac{\lvert v_i \rvert }{\max_{i\in T_j}[\lvert v_i \rvert]},\] where \(i \in T_j = C_1 \cup C_2\) and \(y_i\in\pmb{Y}\). Equation [\[eq:projection\]](#eq:projection){reference-type="eqref" reference="eq:projection"} transforms each point \(y_i\) of the target space into a distance \(v_i\) of that point from the hyperplane separating the two clusters. The distances are then normalized in equation [\[eq:normalization\]](#eq:normalization){reference-type="eqref" reference="eq:normalization"} so that \(w_i\in [0, 1]\). Effectively each weight is the normalized distance of a given sample from the hyperplane separating \(C_1\) and \(C_2\). The samples close to the boundary between the two clusters have a small weight as even if they are incorrectly classified, the SSE will not increase greatly. The samples that are far from the boundary on the other hand, have large weights as, should they be misclassified, the SSE would be heavily affected. The weighting scheme and its influence are illustrated in Figure [\[fig:drawing\]](#fig:drawing){reference-type="ref" reference="fig:drawing"}. KRFWS is not the only decision forest method that utilizes sample weights, one other method is boosting with decision trees. There are however several key differences between the two methods. In boosting with decision trees, the learning of each tree is dependent upon its predecessors. The samples are weighted based on their error in previously trained trees. In KRFWS each tree is trained independently using a different subset of the training data (bagging). The sample weights are calculated independently at each split, and reflect the influence a given samples would have on the loss if it was misclassified.
## Affine Pose Regression
Affine Pose Regression (APR) was recently proposed in as a method for improving the performance of face alignment methods. In contrast to Cascaded Shape Regression (CSR), APR estimates a rigid transforms of the entire face shape: \[\label{eq:transform} S' = \begin{bmatrix} a & b\\ c & d \end{bmatrix} S + \begin{bmatrix} t_x & \dots & t_x\\ t_y & \dots & t_y \end{bmatrix},\] where \(S\), a \(2\times n\) matrix, is the current estimate of the face shape, \(n\) is the number of landmarks in the face shape and \(a, b, c, d, t_x, t_y\) are the parameters of the transform. The parameters are estimated by linear regression based on HOG features extracted at the facial landmarks. APR can be applied before CSR or in between CSR iterations to efficiently compensate for inaccurate initialization of the face shape in scale, translation and in-plane rotation. In this work we propose to improve the original APR framework by using KRFWS instead of linear regression. We estimate all the transform parameters by creating separate KRFWS models for \(a,b,c,d\) and a joint model for \(t_x, t_y\). Instead of extracting features at individual landmarks we extract a single feature that covers the entire face. We show the effectiveness of our approach in experiments on the 300-W dataset in section [3](#sec:experiments){reference-type="ref" reference="sec:experiments"}.
## 3D Affine Pose Regression
As mentioned in the previous section, APR can be applied before CSR to compensate for inaccuracy in scale, translation and in-plane rotation of the face shape estimate. In this section we propose a method to extend APR by taking into account out-of-plane rotation of the head, namely: yaw and pitch. Our method, which we call 3D-APR, fits an average 3D face shape to the face in the image and uses the 2D projection of that shape as an initialization for face alignment. 3D Affine Pose Regression (3D-APR) consists of two steps: first we fit an average 3D face shape \(\bar{S}\) to the initial face shape estimate \(S\). The fitting is accomplished using a scaled orthographic projection: \[\label{eq:scaledorto} \bar{s} = k \cdot P \cdot \bar{S} + \begin{bmatrix} t_x & \dots & t_x\\ t_y & \dots & t_y \end{bmatrix},\] where \(\bar{s}\) is the projected shape, \(k\) is a scaling factor, \(P\) are the first two rows of a rotation matrix and \(t_x\), \(t_y\) are translation parameters. The values of the parameters \(\Gamma=\{k,P,t_x,t_y\}\) for any shape are obtained by solving the following optimization problem: \[\label{eq:gauss} \newcommand{\operatornamewithlimits{arg\min}}{\operatornamewithlimits{arg\min}} \Gamma = \operatornamewithlimits{arg\min}_{\Gamma} \lVert \bar{s}-S \rVert^2.\] The optimization is performed using the Gauss-Newton method as in. In the second step we estimate an update \(\Delta\) to \(\Gamma\) that refines the projected 3D shape \(\bar{s}\) so that it is closer to the true shape of the face in the image. As in APR the estimation is performed using KRFWS based on a PHOG descriptor extracted at the face region. Separate KRFWS models are trained to estimate \(k\) and \(P\), a joint model is used for \(t_x\) and \(t_y\). The projection matrix \(P\) is parametrized using euler rotations. In practice, the estimation of \(P\) is equivalent to head pose estimation. Learning is performed similarly to learning in APR and CSR. The training set consists of a set of images with corresponding ground truth landmark locations. For each image a number of initial shapes are generated from the ground truth shape. For each initial shape, the initial parameters \(\Gamma\) and the ground truth parameters \(\Gamma'\) are obtained using equation [\[eq:gauss\]](#eq:gauss){reference-type="eqref" reference="eq:gauss"} and the ground truth annotations. KRFWS learning is then applied to map the PHOG descriptor to the update \(\Delta = \Gamma'-\Gamma\).
# Experiments {#sec:experiments}
In this section we test the effectiveness of the proposed methods in affine pose regression, face alignment and head pose estimation. The parameters we use for our methods in APR and face alignment have been established through cross-validation, with the exception of the number of children \(K\). \(K\) was set following, where the authors have found \(K=2\) to be optimal for a target space similar to ours. We plan to investigate different values of \(K\) in future experiments..
## Affine pose regression
We test the effectiveness of APR and 3D-APR on the 300-W dataset, which consists of face images with corresponding ground truth annotations of 68 characteristic points and bounding boxes generated by a face detector. The images in 300-W are gathered from several other datasets: AFW, HELEN, IBUG and LFPW. For learning we use the AFW dataset and the training subsets of the HELEN and LFPW datasets, which together consist of 3148 face images. Our test dataset consists of two subsets: the challenging IBUG dataset (135 images) and the less challenging test subsets of the LFPW and HELEN datasets (554 images). Together the two datasets form what we refer to as the full set. This division of the 300-W dataset is a standard in face alignment testing, employed in many recent articles. Each method is initialized with the face detector bounding box provided in the 300-W dataset. Similarly to, we use the inter-pupil distance normalized landmark error, all errors are expressed as the % of the inter-pupil distance. The pupil locations are assumed to be the centroids of the landmarks located around each of the eyes. Five different configurations of APR and 3D-APR are tested: (1) Linear APR with feature extraction at landmarks, (2) Linear APR with a single feature extracted at the face center, (3) KRF APR with a single feature extracted at the face center, (4) KRFWS APR with a single feature extracted at the face center, (5) KRFWS APR followed by 3D-APR with a single feature extracted at the face center (Combined APR, CAPR). In all of the configurations the images are rescaled so that the face size is approximately \(64\times 64\) pixels. In all experiments APR is performed for two iterations, while 3D-APR is performed once. In the first configuration Pyramid HOG features covering \(32\times 32\) pixels are extracted at each landmark. The input descriptor for APR is formed by concatenating the descriptors from each of the landmarks. In configurations (2), (3), (4) and (5) a single PHOG is extracted at the face center. As the feature size is not a concern in this scenario (only one feature is extracted instead of 68) we use the extended version of the HOG feature described in. The descriptor covers an area of \(64\times 64\) pixels. The results of the experiments are shown in Table [1](#tab:APR){reference-type="ref" reference="tab:APR"}. KRFWS APR outperforms Linear APR on the challenging subset by 6%. CAPR shows the best accuracy of all tested methods, reducing the error of Linear APR by 35% on the full set.
## Face alignment
In face alignment we use the same training and evaluation data as in the APR experiments. In order to facilitate comparison with other methods we report the results of our full pipeline for both inter-pupil normalisation, and inter-ocular normalisation. Our face alignment method uses the Local Binary Feature framework, where instead of standard regression forests we use KRFWS, and instead of pixel difference features we use PHOG. The forest generated for each landmark consists of 5 trees with a maximum depth of 7, the PHOG extracted at landmarks cover an area of \(32\times 32\) pixels each, with a single block per pyramid level. Similarly to, our method is performed for 5 iterations. We test the accuracy of our face alignment method with four different initialization configurations: (1) face detector bounding boxes provided in the 300-W dataset, (2) Linear APR, (3) KRFWS APR, (4) KRFWS APR + 3D-APR (CAPR). The results of the experiments are shown in Table [2](#tab:alignment){reference-type="ref" reference="tab:alignment"} along with the results of several state-of-the-art methods, all trained on the same datasets and all initialized as in configuration (1). The methods we compare to perform face alignment using a variety of machine learning tools, including: linear regression, decision trees and deep learning. The experiments we have performed show the effectiveness of our face alignment combined with KRFWS APR and 3D-APR. In configuration (1) the proposed face alignment method outperforms the original LBF in all cases by between 2% and 8%. The addition of KRFWS APR and CAPR for initialization leads to an improvement of 18% and 20% over the original LBF on the challenging IBUG dataset. Our face alignment combined with CAPR initialization shows state-of-the-art results on all datasets. The addition of CAPR resulted in an error reduction of over 13% on the challenging subset in comparison to face detector initialization. Figure [\[fig:faces\]](#fig:faces){reference-type="ref" reference="fig:faces"} shows a diagram of our face alignment pipeline along with images at its consecutive stages.
## Head pose estimation {#sec:headpose}
In order to compare KRFWS to the original KRF we test our method on the head pose estimation task. We use the Pointing'04 dataset, which was also used to test the original KRF. The Pointing'04 dataset consists of images of 15 subjects, each photographed in two separate sessions. During both sessions each subject had 93 photographs taken with pitch and yaw of the head both varying from-90\(^{\circ}\) to +90\(^{\circ}\). Each image in the dataset is accompanied by a manually annotated bounding box containing the head. For fair comparison we use experimental settings identical to those proposed in. From each image we extract a single Pyramid HOG feature, set K in KRFWS to 2 and have each forest consist of 20 trees. The splitting stops when a node has less than 5 samples. We compare our method to the baseline KRF in two experiments, in both we use the Mean Absolute Error (MAE) measure. In the first experiment we perform 2-fold cross validation, where each fold consists only of images from a single session. In the second experiment we perform 5-fold cross validation, where the images in each fold are chosen at random. Our method shows an improvement of 4% over the baseline KRF in the first test and a 3.7% improvement in the second test. The results of both tests can be found in Table [3](#tab:pose){reference-type="ref" reference="tab:pose"} along with the results for KRF and Adaptive KRF (AKRF). In MATLAB the training time for a single tree in the 5-fold cross validation test is 6.65 sec for KRF and 6.45 sec for KRFWS. This shows that the training of KRFWS is actually faster than training of KRF, despite the additional weight calculation step. We believe that this is because in KRFWS less splits are required to reach the stopping criterion.
# Conclusion
In this article we have proposed a face alignment pipeline based on novel K-cluster Regression Forests with Weighted Splitting. Our pipeline consists of two separate stages: face shape initialization and face alignment. The first step performs APR and novel 3D-APR to improve the initial shape, provided by the face detector, in terms of translation, scale and in and out of plane rotation. The second step performs face alignment using an adapted version of the LBF framework. The proposed face alignment pipeline shows state-of-the-art results on the entire 300-W dataset. | {'timestamp': '2017-06-07T02:07:55', 'yymm': '1706', 'arxiv_id': '1706.01820', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01820'} |
null | null |
# Introduction
Black hole has attract people's attention since a long time ago. The pioneering work of Bekenstein, Hawking have suggested that black hole have temperature and entropy. So it is important to consider the quantum theory of black hole. But there are difficult problems with quantum black hole, such as information loss paradox, the firewall paradox and so on. Up to now, we still don't have a satisfying quantum theory of black hole. In the previous paper Ref., it was claimed that the black hole can be considered as a kind of topological insulator. For BTZ black hole in three dimensional \(AdS_3\) spacetime two evidences were given to support this claim: the first evidence comes from the black hole "membrane paradigm\", and the second evidence comes from the fact that the horizon of BTZ black hole can support two chiral massless scalar field with opposite chirality. Those are two key properties of 2D topological insulator. For higher dimensional black hole the first evidence is still valid. Since the general relativity can't be rewritten as Chern-Simons theory in higher dimensional spacetime, the second evidence fails. In this paper, starting from the boundary BF theory, which can be used to describe the boundary degrees of freedom of black hole in arbitrary dimension, it shown that the isolated horizon of \(3+1-\)D black hole can support massless scalar field and vector field. Those two field can be used to construct a massless Dirac field through the \(2+1-\)dimensional bosonization, which appeared on the boundary of \(3+1-\)D topological insulators. The article is organized as follows. In section 2 we consider the boundary modes on the horizon of BTZ black hole. In section 3 the results are generalized to \(3+1-\)D black hole. Section 4 is the conclusion.
# BTZ black hole
In this section, we study the simpler case: the BTZ black hole in three dimensional \(AdS_3\) spacetime. In the previous paper, it was shown that the horizon can have two chiral massless scalar fields with opposite chirality. The starting point is the reformulation of gravity theory in terms of Chern-Simons theory. But this method restrict to three dimension due to the property of Chern-Simons theory. On the other hand, it was shown that the degrees of freedom on the horizon can be described by BF theory with sources. This result hold for wider case, including general relativity in arbitrary dimension, Lovelock theory and scalar-tensor theory. To study the field theory on the horizon of BTZ black hole, the advanced Eddington coordinate \((v,r,\varphi)\) is more suitable. The metric of BTZ black hole is \[\label{1} ds^2=-N^2 dv^2+2 dv dr+r^2 (d\varphi+N^\varphi dv)^2.\] Choose the following Newman-Penrose null co-triads \[\label{2} l=-\frac{1}{2}N^2 dv+dr,\quad n=-dv,\quad m=r N^\varphi dv+r d\varphi,\] the corresponding spin connection is \(A_2=\alpha m-\kappa n\) with \(\alpha=N^\varphi,\kappa=r/L^2-r(N^\varphi)^2\). On the horizon \(r=r_+\), we choose \((x^0,x^1)=(v,\varphi)\). The boundary BF theory on the horizon \(\Delta\) is given by \[\label{3} S=\int_{\Delta}BF=\int_{\Delta}B {\rm d}A,\] with \[\label{4} {\rm d}B=\frac{1}{8\pi G}m,\quad A=\kappa dv+d\beta,\] where \(A\) is the non-rotating component of the connection \(A_2\), and \(\beta\) represent the SO\((1,1)\) symmetry on the horizon. Note that this form of BF theory also appeared in the BF theory description of topological insulator (the Eq.(42)). Up to a boundary term, the action can be rewritten as \[\label{5} S=\frac{1}{2}\int_{\Delta}B {\rm d}A-{\rm d}B\wedge A,\] The canonical form is \[\label{6}\begin{split} S=\frac{1}{2}\int_{\Delta}B(\partial_0 A_1-\partial_1 A_0)-( A_1\partial_0 B-A_0\partial_1 B )\\ =\frac{1}{2}\int_{\Delta}B\partial_0 A_1+A_0\partial_1 B-A_1\partial_0 B +A_0\partial_1 B. \end{split}\] Choose the gauge \(A_0=0\) to gives \[\label{6}\begin{split} S=\frac{1}{2}\int_{\Delta}B\partial_0 A_1-A_1\partial_0 B. \end{split}\] From ([\[4\]](#4){reference-type="ref" reference="4"}) it can be shown that the gauge \(A_0=0\) imply \[\label{7} \partial_0 \beta=-\kappa,\quad A_1=\partial_1 \beta.\] Insert into the action ([\[6\]](#6){reference-type="ref" reference="6"}) to give \[\label{8}\begin{split} S=\frac{1}{2}\int_{\Delta}B\partial_0 \partial_1 \beta-\partial_1 \beta\partial_0 B\\ =-\frac{1}{2}\int_{\Delta}\partial_1 B\partial_0 \beta+\partial_1 \beta \partial_0 B \end{split}\] which is the same as the Eq.(49) in Ref.. So we can get two chiral bosonic modes flowing in opposite directions. Those bosonic modes can combined into a one-dimensional fermionic system via standard \(1+1-\)dimensional bosonization. Now we have two fields \(B,\beta\). To construct a massless fermionic field those two fields should satisfy the dual relation \[\label{9} \partial_\mu \beta+\epsilon_{\mu \nu}\partial^\nu B=0.\] But due to the constraint ([\[4\]](#4){reference-type="ref" reference="4"}) and ([\[7\]](#7){reference-type="ref" reference="7"}), the dual relation don't hold. We must re-scale those fields \[\label{9a} \tilde{\beta}=\gamma_1 \beta, \tilde{B}=\gamma_2 B.\] We want the coefficient of BF action to be a dimensionless constant \(\frac{1}{C}\), then \[\label{10} \gamma_1=\sqrt{\frac{C}{8\pi G \kappa}},\quad \gamma_2=\pi \sqrt{8\pi C G \kappa}.\] The final action for BF theory on the horizon of BTZ black hole is \[\label{11} S=\frac{1}{C}\int_{\Delta}\tilde{B} {\rm d}\tilde{A},\] with \(\tilde{A}=\sqrt{\frac{C}{8\pi G \kappa}} \bar{A}\). When \(C=\pi\), one can get the BF theory for topological insulator.
# Four dimensional black hole
Unlike the Chern-Simons theory, the BF theory can be applied to 4D and higher dimensional black holes. Similar to the case in 3D, the boundary BF action on the isolated horizon is \[\label{12} S=\int_{\Delta}BF=\int_{\Delta}B \wedge {\rm d}A,\] with \[\label{14} {\rm d}B=\frac{\Sigma_{01}}{8\pi G}=\frac{1}{8\pi G} e^2\wedge e^3,\quad A=\kappa dv+d\beta.\] The action can also be written as \[\label{15} S=\frac{1}{2}\int_{\Delta}B\wedge {\rm d}A+{\rm d}B\wedge A,\] The canonical form is \[\label{16}\begin{split} S=\frac{1}{2}\int_{\Delta} \epsilon^{\mu\nu\rho}B_\mu F_{\nu\rho}+\epsilon^{\mu\nu\rho}\partial_\mu B_\nu A_\rho\\ =\frac{1}{2}\int_{\Delta}2B_0(\partial_1 A_2-\partial_2 A_1)+2A_0(\partial_1 B_2-\partial_2 B_1)-A_1\partial_0 B_2 +A_2\partial_0 B_1-B_1\partial_0 A_2 +B_2\partial_0 A_1. \end{split}\] The gauge \(A_0=0\) imply \[\label{17} \partial_0 \beta=-\kappa,\quad A_i=\partial_i \beta\Rightarrow \partial_1 A_2-\partial_2 A_1=0.\] Insert into the action ([\[16\]](#16){reference-type="ref" reference="16"}) to give \[\label{8}\begin{split} S=\frac{1}{2}\int_{\Delta}B_2\partial_0 \partial_1 \beta-\partial_1 \beta\partial_0 B_2-B_1\partial_0 \partial_2 \beta+\partial_2 \beta\partial_0 B_1\\ =-\frac{1}{2}\int_{\Delta}\partial_0 \beta\epsilon^{ij}\partial_i B_j+\partial_i \beta \epsilon^{ij} \partial_0 B_j, \end{split}\] which is the same as the Eq.(53) in Ref.. Now we have two fields \(\beta\) and \(B\). To construct a massless fermionic field those two fields should satisfy the dual condition \[\label{19} \partial_\mu \beta+\epsilon_{\mu \nu\rho}\partial^\nu B^\rho=0.\] First taking \(\mu=0\) to gives \[\label{20} \kappa=-\partial_0 \beta=\epsilon_{ij}\partial^i B^j=\frac{1}{8\pi G}.\] Due to the zero law of isolated horizon, the surface gravity \(\kappa\) is a constant on the horizon. Similar to the BTZ case, we can re-scale the two fields to satisfying the above condition and fixed coefficient. The result is \[\label{21} \tilde{\beta}=\sqrt{\frac{C}{8\pi G \kappa}} \beta, \quad \tilde{B}=\pi \sqrt{8\pi C G \kappa} B.\] Let's consider the canonical mass dimension of those fields. The spin connection has dimension \([A_\mu]=1\), the tetrad \([e_\mu]=0\), so one can get \([B_\mu]=1,[\beta]=0\). Since \([\kappa]=1,[G]=-2\), so \([\tilde{\beta}]=[\tilde{B}]=\frac{1}{2}\), which are the correct canonical mass dimension for scalar field and vector field in 3 dimension. Next taking \(u=i\) to gives \[\label{22} \partial_i \beta=\epsilon_{ij}\partial^0 B^j.\] Choose gauge \(B^0=0\), then the above equation gives the Coulomb gauge condition \(\partial_i B^i=0\). The final action for BF theory is \[\label{23} S=\frac{1}{C}\int_{\Delta}\tilde{B}\wedge {\rm d}\tilde{A},\] with \(\tilde{A}=\sqrt{\frac{C}{8\pi G \kappa}} A\).
# Conclusion
In this paper, we study the boundary field theory for the BTZ black hole and 4D black hole from the BF theory. For BTZ black hole, the boundary theory contain two chiral bosonic modes flowing in opposite direction. Through standard \(1+1-\)dimensional bosonization, those bosonic modes can combined into a one-dimensional fermionic system. For black hole in four dimensional spacetime, the BF theory transforms into a theory that contain a scalar field and a vector field, which can construct massless fermionic fields through \(2+1-\)dimensional bosonization. Those fermionic field also appears in the boundary of topological insulators. But we don't need to stop here. The BF theory can describe the degrees of freedom on the horizon for arbitrary dimensional black holes. Consider the \(n+1-\)dimensional spacetime. The horizon is \(n-\)dimension, and the BF theory reduces to a theory that contain a scalar field \(\beta\) and a \((n-2)-\)rank form field \(B_{\mu\nu\ldots \rho}\). This \(n-2-\)rank form field \(B\) is crucial for higher-dimensional bosonization. For example, in \(4+1-\)dimensional spacetime, those scalar-tensor fields can combined into a Weyl spinor in \(3+1-\)dimensional boundary. Due to the periodic table of topological insulators and superconductors, in every dimension, there are five types of non-trivial topological phases, two indexing by \(Z_2\) and three by \(Z\). The boundary modes of topological insulator in other dimension(not 2 and 3) may not be protected by time reversal symmetry, but by other symmetry. In black hole case, the boundary modes are protected not by symmetry, but by geometry. On the isolated horizon of black hole, due to the properties of isolated horizon, there exist the massless boundary modes. So the boundary modes are protected by the geometry of isolated horizon, and are difficult to destroy. Also the horizon of the black hole, which can conduct electricity, is intrinsic to black hole, so the black hole is more like the topological insulator rather than an insulator with a sheet of aluminum foil on it. The coefficient of BF theory is important, since for \(C=\pi\) it describe the integral topological insulator, and for other values it may describe the fractional topological insulators. We have not find a canonical way to fix the value of \(C\) from the black hole physics, so it is not clear that the black hole is integral or fractional topological insulators. This work is supported by the NSFC (Grant No.11647064) and Nanhu Scholars Program for Young Scholars of XYNU. | {'timestamp': '2017-12-21T02:03:46', 'yymm': '1706', 'arxiv_id': '1706.01630', 'language': 'en', 'url': 'https://arxiv.org/abs/1706.01630'} |
# Introduction {#sec:intro}
Determinism versus randomization in algorithm design is a fundamental concern in computer science and is the topic of a great many works in complexity theory. In "space-constrained" models such as communication complexity and data streaming, basic results show that derandomization can entail an exponential or worse blow-up in cost. For instance, in the two-party communication setting, the very basic \(n\)-bit [equality]{.smallcaps} problem admits a bounded-error randomized protocol with only \(O(1)\) communication (\(O(\log n)\) if restricted to private coins), whereas its deterministic communication complexity is as large as it gets, namely \(n+1\). In the data streaming setting, the similarly basic [distinct-elements]{.smallcaps} problem admits a one-pass bounded-error randomized algorithm that uses \(O(\log n)\) space to provide a \((1+\varepsilon)\)-approximation , whereas a deterministic algorithm would require \(\Omega(n)\) space, even if multiple passes and large approximation factors are allowed . In this work, we explore such a price-of-determinism phenomenon in the *query complexity* world, for a similarly basic search problem. The focus of our study is a search problem that we call [element-extraction]{.smallcaps} (henceforth, \(\elemx\)), where the input is a set \(Z \subseteq [n] := \{1,\ldots,n\}\), promised to be nonempty, and the goal is to extract any element from \(Z\). Formally, this is a total search problem given by the relation \(\elemx_n \subseteq 2^{[n]} \times [n]\), where \[\begin{aligned}
\label{eq:elemx-def-set} \elemx_n = \left\{ (Z,i): Z \subseteq [n],\, i \in [n],\, \text{and } |Z| > 0 \Rightarrow i \in Z \right\} \. \end{aligned}\] As is often the case, the natural correspondence between sets in \(2^{[n]}\) and vectors in \(\b^n\) will be useful. Indeed, we shall freely switch between these two viewpoints, using the notational convention that uppercase letters denote sets and their corresponding lowercase boldface variants denote characteristic vectors. Thus, we can also formalize \(\elemx\) as \[\begin{aligned}
\label{eq:elemx-def-vec} \elemx_n = \left\{ (\mathbf{z},i): \mathbf{z} = (z_1, \ldots, z_n) \in \b^n,\, i \in [n],\, \text{and } \mathbf{1}^\mathsf{T} \mathbf{z} > 0 \Rightarrow z_i = 1 \right\} \. \end{aligned}\] The goal of an algorithm solving \(\elemx\) is to produce an output \(i\) such that \((Z,i) \in \elemx\): with certainty in the deterministic setting, and with probability \(\ge 2/3\) (say) in the randomized setting. In other words, the algorithm must produce a *witness* of the nonemptiness of \(Z\). To do so, the algorithm may access \(Z\) (equivalently, \(\mathbf{z}\)) using *linear queries*, as we shall now explain. In a Boolean decision tree model, an algorithm may only access the input vector by querying its individual bits. In such a setting, there is not much to say about \(\elemx\): even randomized algorithms are easily seen to require \(\Omega(n)\) queries. But things get interesting if we allow more powerful queries: specifically, linear ones. Let us define a *linear query protocol over domain \(D\)* (a \(D\)-LQP, for short) to be a query protocol wherein each query is an evaluation of a linear form \(\sum_{i=1}^n a_i z_i\), where each \(a_i \in D\). The domain \(D\) should be thought of a "reasonable" subset of a ring containing \(\b\)---e.g., a finite field, or integers with bounded absolute value---and the linear functions will be evaluated in the underlying ring. The cost of an LQP is the number of linear form evaluations used.[^1] In this work we particularly care about the amount of *adaptivity* in an LQP, which quantifies the extent to which each query depends on the outcomes of previous queries. To set the stage, we recall the problem of \(\ell_0\)[-sampling]{.smallcaps} , from the world of sketching and streaming algorithms. The goal of \(\ell_0\)-sampling is to sample a pair \((i, x_i)\) from a nonzero input vector \(\mathbf{x} \in \mathbb{R}^n\) (say), so that \(x_i \ne 0\) and \(i\) is distributed nearly uniformly on the support of \(\mathbf{x}\). This is a fundamental primitive, used as a low-level subroutine in a wide range of applications in streaming and other "big data" algorithms. There are several solutions to this problem , most of which provide a linear *sketching* scheme, wherein one computes \(\mathbf{y} = S\mathbf{x}\) for a certain random \(d \times n\) matrix \(S\) and then runs a recovery algorithm on the low-dimensional vector \(\mathbf{y}\) to produce the desired sample. Notice that if the input is a vector \(\mathbf{z} \in \b^n\), such a scheme provides a randomized LQP for \(\elemx_n\) (allowing a small probability of error). In particular, using the optimal \(\ell_0\)-sampling sketch of Jowhari, Sağlam, and Tardos , we obtain a \(\mathbb{Z}\)-LQP that makes \(O(\log n)\) queries, using coefficients in \(\{0,1,\ldots,n\}\), and has the pleasing property of being *non-adaptive*. We can also obtain a \(\mathbb{Z}_q\)-LQP that makes \(O(\log^2 n/\log q)\) queries;[^2] details in . Turning to the deterministic setting---our main focus in this paper---it is easy to show that a non-adaptive \(\mathbb{Z}\)-LQP for \(\elemx_n\) must make \(\Omega(n/\log n)\) queries, for basic information-theoretic reasons. For completeness, we give the proof in . However, this heavy determinism penalty disappears upon moving to general deterministic LQPs, where we can use adaptivity. Indeed, a simple binary search strategy leads to a \(\mathbb{Z}\)-LQP that makes \(O(\log n)\) queries, using coefficients in \(\b\). We can refine this observation to trade off the query complexity for amount of adaptivity. This brings us to our central concept. Define a *\(k\)-round LQP* to be one where the queries are made in batches that we call *rounds*: the collection of linear forms defining the queries in round \(i\) depend only on the results of queries made in rounds \(1,\ldots,i-1\) (a formal definition appears in ). Then, a natural generalization of the binary search strategy provides a \(k\)-round \(\mathbb{Z}\)-LQP for \(\elemx\), using coefficients in \(\b\), making at most \(k (\lceil n^{1/k} \rceil-1)\) queries in total. When we are additionally promised that \(\mathbf{1}^\mathsf{T} \mathbf{z} \ne 0\), where addition is performed in the ring \(\mathbb{Z}_q\), then this algorithm also works as a \(\mathbb{Z}_q\)-LQP; details in . Notice that \(k\)-round LQPs naturally interpolate between linear sketches at one extreme (when \(k = 1\)) and linear decision trees at the other (when \(k = n\)). The most important message of this paper is that the above rounds-versus-queries tradeoff is asymptotically tight for deterministic linear query protocols for \(\elemx\), in several natural settings. We state our results informally for now, with formal statements given after the necessary definitions and preliminaries.
## Our Results and Techniques {#sec:results}
We shall study \(D\)-LQPs for the domains \(D = \mathbb{Z}_q\), the ring of integers modulo \(q\) (with \(q \ll n\)) as well as \(D = \mathbb{Z}\), but with coefficients of small magnitude (at most \(\poly(n)\), say). Such restrictions on the coefficients are necessary, because allowing arbitrary integer coefficients makes it possible to recover the entire input \(\mathbf{z}\) with the single query \(\sum_{i=1}^n 2^{i-1} z_i\). When \(D = \mathbb{Z}_q\), for small \(q\), solving \(\elemx\) without the promise that \(\mathbf{1}^\mathsf{T} \mathbf{z} \ne 0\) is hard, regardless of the number of rounds. Intuitively, there is no cheap way to deterministically verify that a subset \(I \subseteq [n]\) indeed contains an index \(i \in I\) where \(z_i \ne 0\). Defining the "cost" of an LQP to be the number of queries it makes in the worst case (formally defined in ), we obtain the following not-too-hard results.
As noted earlier, adding the promise that \(\mathbf{1}^\mathsf{T} \mathbf{z} \ne 0\) permits a more efficient \(k\)-round deterministic algorithm. For each integer \(q \ge 2\), define \(\elemxq_n\) to be the version of \(\elemx_n\) where we are given the stronger promise that \(\mathbf{1}^\mathsf{T} \mathbf{z} \ne 0\) under arithmetic in \(\mathbb{Z}_q\). Equivalently, using set notation, we are promised that \(|Z| \not\equiv 0 \pmod q\). We prove the following results, using similar round-elimination arguments.
Although subsumes in the asymptotic sense, we find it useful to present the former result in full, first, to lay the groundwork for our subsequent lower bound proofs. As we shall see, the fact that \(\mathbb{Z}_2\) is a field leads to an especially clean execution of the round elimination strategy. Note also that a weaker form of follows from existing work on formula size-depth tradeoffs (see ); however, the resulting proof, once fully unrolled, is considerably more complex than our direct argument. At a high level, a lower bound proof based on round elimination works as follows. We consider a hypothetical \(k\)-round protocol for \(n_k\)-dimensional instances of some problem \(P\) that does not incur much cost in its first round. Based on this low cost, we extract a \((k-1)\)-round protocol for \(n_{k-1}\)-dimensional instances of \(P\) by "lifting" these smaller instances to special \(n_k\)-dimensional instances on which the \(k\)-round protocol essentially "wastes" its first round. If we can carry out this argument while ensuring that the shrinkage from \(n_k\) to \(n_{k-1}\) is not too drastic, then a too-cheap \(k\)-round protocol will eventually give us a \(0\)-round protocol for a nontrivial instance dimension, leading to a contradiction. In the proofs of the above two theorems, this strategy is executed by identifying a large collection of pairwise disjoint sets that are treated identically in the protocol's first round. Viewing these sets as blocks of indices within \([n]\), we consider block-structured instances of \(\elemx_n\) and proceed to lift general instances of \(\elemx_{n'}\) into these block-structured ones. In , these blocks arise from elementary linear algebraic considerations. In , the fact that inputs are in \(\b^n\) instead of \(\mathbb{Z}_q^n\) necessitates a brief excursion into additive combinatorics. Finally, we consider LQPs over \(\mathbb{Z}\), the ring of all integers, but with bounds on the magnitude of coefficients (which, as we noted earlier, is necessary in order to have nontrivial results). To be precise, we consider domains of the form \(\mathbb{Z}_{[b,c]} := \{a \in \mathbb{Z}:\, b \le a \le c\}\). While we are unable to prove a full tradeoff lower bound in this case, we do obtain a near-optimal result for \(k=2\) rounds.
The former result is straightforward, based on the simple observation that such an LQP can extract the *entire* input \(\mathbf{z}\) followed by basic information theoretic considerations. Incidentally, the problem of extracting all of \(\mathbf{z}\) using \(\mathbb{Z}_{[0,1]}\)-LQPs has a long history as the *coin weighing* problem, for which a 1-round \(O(n / \log n)\) algorithm exists; see The significant result here is the latter. It again uses a round elimination strategy and, as before, the bird's-eye view is that we identify disjoint blocks of indices to engineer a suitable lifting. This time, the blocks arise out of extremal combinatorics considerations, specifically the sunflower lemma, in its recently strengthened form . Furthermore, upon carrying out this round elimination, we are left with a \(1\)-round LQP that solves \(\elemx\) only under a cardinality constraint on the input set. To finish the proof, we must demonstrate hardness even for this special case. This is not as straightforward as : our argument to handle this hinges on the Frankl--Wilson theorem on set systems with forbidden intersection sizes. Attempts to extend the above proof outline to handle more than two rounds runs into technical issues of integer divisibility. We suspect that this is an artifact of our proof machinery and not inherent to the problem. We conjecture that every deterministic \(k\)-round \(\mathbb{Z}_{[-B,B]}\)-LQP requires cost \(\widetilde{\Omega}(n^{1/k})\), suppressing polylogarithmic factors. Indeed, we believe that much more is true, and that a communication complexity analogue of such a tradeoff also holds. We shall take this up after a discussion of related work.
## Related Work and Connections {#sec:related}
Our work touches upon several themes with long histories of study in computer science: determinism versus randomization, adaptivity versus non-adaptivity, sublinear algorithms, and input access through structured queries. With these connections in mind, we recall a small number of works that are either close in spirit to ours or shed light on some aspect of this work. The most basic query model is the Boolean decision tree. In this setting, deterministic and randomized complexities are polynomially related for total Boolean functions , whereas arbitrarily large gaps are possible for search problems . Parity decision trees---equivalent to our \(\mathbb{Z}_2\)-LQPs---have been studied in several works (e.g., and the references therein), usually for Boolean functions and focusing on connections with communication complexity of XOR-composed functions. Beyond the Boolean---or indeed the discrete---setting lie *linear decision trees*, where the input is a *real* vector and one can query the *sign* of a linear form . All such "decision tree" models are fully adaptive and the vast majority of works using them do not focus on amount of adaptivity as a resource. At the other extreme is the (nonadaptive) linear sketching model, where a high-dimensional input is accessed through one batch of linear queries (equivalently, through a low-dimensional sketch of it produced by a sketching matrix). This paradigm is ubiquitous in data streaming algorithms and compressed sensing and has connections to dimension reduction and metric embeddings. Some recent work carries the message that linear sketching might be a *complete* paradigm for a large class of data streaming algorithms and certain communication protocols . Most work on linear sketching considers *randomized* sketches, since determinism often precludes sublinear cost. Turning to determinism, the well-studied *coin weighing* problem, put in our terms, asks for a \(\mathbb{Z}_{[0,1]}\)-LQP that retrieves the entire input \(\mathbf{z} \in \b^n\). It has long been known that \((2\pm o(1))n/\log n\) nonadaptive queries are necessary and sufficient. Special cases and variants of this problem have been studied over the years; see for some early history and for recent history. While some of these works consider adaptive LQPs, there is no strong rounds-vs-queries tradeoff for this problem, which is harder than \(\elemx\). The body of work on round complexity under linear queries is much smaller. There is one recent work very close to ours: Assadi, Chakrabarty, and Khanna studied a problem very similar to \(\elemx\) that they called [single-element-recovery]{.smallcaps}, where the input is a vector \(\mathbf{x} \in \mathbb{R}_{\ge0}^n\), and by applying \(\mathbb{R}\)-linear queries one wishes to recovery an arbitrary element from the support of \(\mathbf{x}\). While their query model is much stronger than our \(\mathbb{Z}\)-linear or \(\mathbb{Z}_q\)-linear queries, it is balanced by the \(\mathbb{R}_{\ge 0}\)-valued inputs that prevent tricks to recover the entire input in one query. Their main theorem implies that the deterministic \(k\)-round search algorithm making roughly \(k (n^{1/k}-1)\) queries in total---very similar to ---has cost exactly matching the lower bound. Linear queries and adaptivity are also featured together in some work on *sparse recovery* problems. One such problem is to find an approximately closest \(s\)-sparse vector \(\mathbf{x}^\star\) to an input \(\mathbf{x}\), using \(\mathbb{R}\)-linear queries to the input and \(r\) rounds of adaptivity. For this, have proven near optimal lower bounds of \(\Omega(r (\log n)^{1/r})\) when \(s=1\) and have extended them to small \(s\), proving \(\Omega(\frac{1}{r} s (\log n)^{1/r})\) queries are needed when \(\log s < (\log n)^{1/r}\). A number of works consider rounds of adaptivity in query models beyond linear queries. Recent examples include works on maximizing submodular functions through adaptive oracle queries ; on adaptivity hierarchy theorems in property testing ; on identifying biased coins through pairwise comparisons or in multi-armed bandit settings ; and on finding approximately maximum bipartite matchings through demand queries and OR-queries . Other works have studied adaptivity in the massive parallel communication/computation (MPC) model and in various graph query models . A rich body of work on cost/adaptivity tradeoffs is found in *communication complexity*, where adaptivity manifests as rounds of interaction. An early work gave exponential separations between \(k\) and \(k+1\) rounds for all \(k\) and introduced a round elimination paradigm that remains ubiquitous to this day. This work also explains how an earlier result connecting circuit and communication complexities can be used to relate bounded-round communication complexity for a specific problem to the size of bounded-depth, unbounded fan-in formulas. More work has been spurred by applications of bounded-round communication lower bounds in data structures, where they provide lower bounds in the cell-probe model ; and in streaming algorithms, where they translate naturally to tradeoffs between the number of passes made over the input stream and the working memory required . In much of this body of work, round elimination is performed using *information theoretic* arguments that naturally provide lower bounds against randomized algorithms. In contrast, it is rare to see deterministic tradeoffs where corresponding randomized ones do not hold because randomization makes the problem "too easy." This is exactly the situation with \(\elemx\), as shown by this work in the context of the randomized upper bounds () via \(\ell_0\)-sampling . In light of the preceding discussion, our instantiations of round elimination must use techniques beyond Shannon-style information theory. They indeed do. Our techniques therefore have the potential for further use in separating determinism from randomization in this fine-grained (round aware) sense. Our query complexity results on \(\elemx\) suggest a tantalizing communication complexity analogue. Let \(\textsc{ur}^\subset Clearly, a\)k\(-round query protocol for\[making\)q\(queries, with each answer lying in a set of size\)M\(, provides a\)k\(-round communication protocol for\)[ur]{.smallcaps}\^
We find the above conjecture compelling because it would demonstrate a new phenomenon in communication complexity, where a problem is easy for one-round randomized *and* for interactive deterministic protocols, but exhibits a nontrivial tradeoff for bounded-round deterministic ones. In passing, we note that the \(\textsc{ur}^\subset There are also two problems similar to\)[ur]{.smallcaps}\^
# Preliminaries {#sec:prelim}
Throughout the paper, we shall freely switch between the equivalent viewpoints of sets in \(2^{[n]}\) and vectors in \(\b^n\), using the notational convention that when an uppercase letter (e.g., \(S,Z\)) denotes a set, the corresponding lowercase boldface letter (e.g., \(\mathbf{s},\mathbf{z}\)) denotes the characteristic vector of that set and vice versa.
## Various Definitions {#sec:defs}
The search problem \(\elemx_n\) was already formally defined in [\[eq:elemx-def-set\]](#eq:elemx-def-set){reference-type="ref" reference="eq:elemx-def-set"}. We shall also work with special cases of this problem, where the cardinality of the input set is further restricted in some way. These are formalized as follows: we define \]\begin{aligned}
\elemxq_n &= \left\{ (Z,i): Z \subseteq [n],\, i \in [n],\, \text{and } |Z| \not\equiv 0 \tpmod q \Rightarrow i \in Z \right\} \,; \label{eq:elemxq-def} \\ \elemxqh_n &= \left\{ (Z,i): Z \subseteq [n],\, i \in [n],\, \text{and } |Z| \equiv h \tpmod q \Rightarrow i \in Z \right\} \,; \label{eq:elemxqh-def} \\ \elemxquarter_n &= \left\{ (Z,i): Z \subseteq [n],\, i \in [n],\, \text{and } |Z| = n/4 \Rightarrow i \in Z \right\} \label{eq:elemxquarter-def} \. \end{aligned}\[
## Useful Results from Combinatorics {#sec:comb}
In the course of this paper, we will use several important theorems from combinatorics. For results on \(\mathbb{Z}_q\)-LQPs (proved in ), we use the following result of van Emde Boas and Kruyswijk on zero sumsets, slightly reworded to use modern notation.
A stronger result that \(s(G) = 1 + r (q-1)\) applies when \(G = \mathbb{Z}_{q}^r\) and \(q\) is a prime power ; it is conjectured that the prime-power constraint is unnecessary When working over \(\mathbb{Z}\) (in ), we use the well-known notion of a *sunflower* and the following recent result of Rao , which refines the noted result of Alweiss, Lovett, Wu, and Zhang that improved the classic sunflower lemma of Erdős and Rado . The note further improves Rao's bound by replacing the \(\log(pt)\) factor with \(\log t\), but this will not affect our proof. Tao gives an alternative presentation of Rao's result which may be simpler to follow. In a different part of our argument, we will need a well known theorem of Frankl and Wilson .
## Our Round Elimination Framework {#sec:framework}
We now describe a framework for our round elimination arguments. For this section, we shall work over a general ring (with unity), \(R\), and "LQP" will mean a \(D\)-LQP where \(D \subseteq R\). Fix this ring \(R\).
Suppose we have an LQP \(\Pi\) that operates on inputs in \(\b^n\) and produces outputs in \([n]\). Further, suppose \(S_1, \ldots, S_m \subseteq [n]\) is a collection of pairwise disjoint nonempty sets. We then define a certain LQP \(\Pi^{(S_1,\ldots,S_m)}\) operating on inputs in \(\b^m\) and producing outputs in \([m]\). To aid intuition, we describe the construction procedurally in .
# Linear Queries Modulo 2 {#sec:z2-linear-measurements}
We begin our study of the element-extraction problem by considering \(\mathbb{Z}_2\)-linear queries. As noted in , we shall later generalize the results to \(\mathbb{Z}_q\), but we feel it is worth seeing our framework in action in the especially clean setting of \(\mathbb{Z}_2\). We begin by showing that the additional promise of odd cardinality on the input set \(Z\) is crucial, or else there is no interesting rounds-vs-queries tradeoff to be had.
Accordingly, for the rest of this section, we focus on the problem \(\elemxodd_n\), as defined in [\[eq:elemxq-def\]](#eq:elemxq-def){reference-type="ref" reference="eq:elemxq-def"}. We shall prove using a round elimination technique. As discussed in , this round elimination will be enabled by identifying a certain \(A\)-uniform family. The next lemma, establishing a useful fact about matrices over \(\mathbb{Z}_2\), will provide us this family.
The next step of the proof is to repeatedly invoke the above round elimination lemma and carefully control parameters. To perform a sharp analysis, we introduce the following concept.
This brings us to the main result of this section: a rounds-vs-queries tradeoff.
# Linear Queries Modulo ***q*** {#sec:zq-linear-measurements}
First, we use to show that \(\elemx_n\) is hard for \(\mathbb{Z}_q\)-LQPs.
A similar strategy proves a lemma analogous to :
Compared to \(\elemxodd\), there is a slight weakening of the main round elimination lemma, which is a direct consequence of the weakened . Instead of directly lower bounding the cost of \(\elemxq_n\), we prove separate lower bounds for each \(\elemxqh_n\), for all \(h \in \{1,\ldots,q-1\}\), and take their maximum. The search problem \(\elemxqh_n\) is \(\elemx_n\) with the additional promise that the input set \(Z\) has size \(\equiv h \pmod q\).
This brings us to the main result of this section, which essentially generalizes the modulo-\(2\) result from the previous section.
# Linear Queries Over the Integers {#sec:int-linear-measurements}
For \(\mathbb{Z}\)-LQPs, our main result is a 2-round lower bound for \(\elemx_n\). We require a careful accounting of the query cost of a protocol, to adjust for the fact that the (bit) size of the query results depends on the maximum entry value in a given query matrix. This motivates the following definition and observation.
Recall the problem \(\elemxquarter_n\) defined in [\[eq:elemxquarter-def\]](#eq:elemxquarter-def){reference-type="ref" reference="eq:elemxquarter-def"}. For \(n\) divisible by \(4\), this is simply \(\elemx_n\) under the additional promise that \(|Z| = n/4\). We first prove a \(1\)-round lower bound for this problem, under a slight additional assumption on \(n\).
It should be noted that without the promise that \(|Z| = n/4\), a one-round lower bound would follow very easily. By a standard "decoding" argument, a one-round protocol for \(\elemx\) can be used to recover the entire unknown input \(\mathbf{z}\). For completeness, we give the easy proof below. The reason we needed the much more complicated argument in above is that the promise in \(\elemxquarter_n\) prevents us from performing such a decoding.
For our round elimination argument, we require the following claim, similar to and . Even though the claim looks similar, the round elimination argument will be subtly different from its \(\mathbb{Z}_2\) and \(\mathbb{Z}_q\) predecessors.
The preceding round elimination lemma has a key limitation: it requires a protocol for \(\elemx_n\) to create one for \(\elemxquarter_m\). Because of this, it is not possible to apply the lemma to its own output, and thereby obtain a \(k\)-round lower bound. Say we were to try, and \(A\) were the matrix at the root of the protocol \(\Pi\) for \(\elemxquarter_n\). Then if \(A\) contained an all-ones row, might produce an \(A\)-uniform family with all set sizes \(|S_i|\) equal to some constant \(b\) which is not a factor of \(n/4\). Then lifting inputs \(W\) of size \(m/4\) to inputs \(Z\) of size \(n/4\) would fail, because \(n/4 = |Z| = b |W|\) would imply that \(b\) divides \(n/4\), a contradiction. With that said, we now use our round elimination lemma in a one-shot fashion to obtain our main result for integer LQPs.
# Upper Bounds {#sec:upper-bounds}
For the sake of completeness, we provide details of the LQPs attaining various upper bounds referenced throughout the paper. For the most part, these upper bounds are simple observations or extensions of well-known existing results.
## Deterministic *k*-round LQP for
The following family of protocols works both when \(D = \mathbb{Z}_{[0,1]}\) on the problem \(\elemx_n\), and when \(D = \mathbb{Z}_q\) on the problem \(\elemxq_n\). The algorithm appears to be well known, and versions of it are described in Lemma 4.1 of and Section 2.2 of. Let \(d_1, \ldots, d_{k}\) be a division sequence (see ) for \(n\), which minimizes \(\sum_{i=1}^{k} d_i\). makes no more than \(d_r\) queries in each round \(r\).
Detecting whether a \(\{0,1\}^n\) vector is not \(s\)-sparse is also easier than in \(\mathbb{R}^n\). For \(\mathbb{Z}_{[-B,B]}\)-LQPs, querying with the vector \(\mathbf{1} \in \mathbb{Z}^{d}\) suffices. For \(\mathbb{Z}_q\), because ensures that if a vector \(\mathbf{z}\) is s-sparse, it can be recovered exactly, it is enough to query \(O(1)\) random vectors in \(\mathbb{Z}_q^n\). Let \(\mathbf{r}\) be such a random vector, and let \(\mathbf{w}\) be the \(s\)-sparse vector in \(\{0,1\}\) recovered using \(H\); if \(\mathbf{z}\) was s-sparse, then \(\mathbf{z} = \mathbf{w}\) and \(\mathbf{r}^\mathsf{T} \mathbf{z} = \mathbf{r}^\mathsf{T} \mathbf{w}\); otherwise, \(\mathbf{r}^\mathsf{T} \mathbf{z}\) does not equal \(\mathbf{r}^\mathsf{T} \mathbf{w}\) with probability \(1-1/q\).
# Connections Between \(\mathbb{Z}_2\)-LQPs and Circuit Complexity {#sec:kw-lb}
A weaker version of can be proven by combining existing results. As shown in the following lemma, a given \(k\)-round \(\mathbb{Z}_2\)-LQP \(\Pi\) for \(\elemxodd\) can be converted to a communication protocol \(\Upsilon\) for the Karchmer-Wigderson game on \(\textsc{parity}_n\), with the communication cost \(C\) of \(\Upsilon\) being \(\le 2 \cost(\Pi)\). By a slight adaptation of the proof of Theorem 5 in , we can convert \(\Upsilon\) into an unbounded fan-in boolean formula with depth \(k+1\) and no more than \(2^C-1\) AND/OR gates that computes \(\textsc{parity}_n\). Relatively tight lower bounds on the size of such a formula date back to, but we use a result of, which says that a depth-\((k+1)\) unbounded fan-in formula computing \(\textsc{parity}_n\) must have at least \(2^{\Omega(k (n^{1/k}-1)}\) AND/OR gates. Thus \(\cost(\Pi) \ge \frac{1}{2} C \ge \Omega(k (n^{1/k}-1))\).
# Appendix {#sec:appendix}
The following estimate was used in during calculations in the proof of our \(\mathbb{Z}_q\)-LQP lower bound.
:::
[^1]: Note that this is somewhat lower than the number of *bits* needed to encode the output of the queries.
[^2]: Throughout this paper, "\(\log\)" denotes the base-\(2\) logarithm.
[^3]: The exponent of a group is the least common multiple of the orders of its elements. | {'timestamp': '2021-07-14T02:08:11', 'yymm': '2107', 'arxiv_id': '2107.05810', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05810'} |
# Introduction {#sec:intro}
Robust out-of-distribution (OOD) generalization is one of the key open challenges in machine learning. This is particularly relevant for the deployment of ML models to the real world, where we need systems that generalize beyond the i.i.d. (independent and identically distributed) data setting. One instance of such models are agents that learn by interacting with a training environment and we would like them to generalize to other environments with different statistics . Consider the example of a robot with the task of moving a cube to a target position: Such an agent can easily fail as soon as some aspects of the environment differ from the training setup, e.g. the shape, color, and other object properties, or when transferring from simulation to real world. Humans do not suffer from these pitfalls when transferring learned skills beyond a narrow training domain, presumably because they represent visual sensory data in a concise and useful manner . Therefore, a particularly promising path is to base predictions and decisions on similar low-dimensional representations of our world . The learned representation should facilitate efficient downstream learning and exhibit better generalization . Learning such a representation from scratch for every downstream task and every new variation would be inefficient. If we learned to juggle three balls, we should be able to generalize to oranges or apples without learning again from scratch. We could even do it with cherimoyas, a fruit that we might have never seen before. We can effectively reuse our generic representation of the world. We thus consider deep learning agents trained from pretrained representations and ask the following questions: To what extent do they generalize under distribution shifts similar to those mentioned above? Do they generalize in different ways or to different degrees depending on the type of distribution shift, including sim-to-real? Can we predict the OOD generalization of downstream agents from properties of the pretrained representations? To answer the questions above, we need our experimental setting to be realistic, diverse, and challenging, but also controlled enough for the conclusions to be sound. We therefore base our study on the robot platform introduced by. The scene comprises a robot finger with three joints that can be controlled to manipulate a cube in a bowl-shaped stage. conveniently introduced a dataset of simulated and real-world images of this setup with ground-truth labels, which can be used to pretrain and evaluate representations. To train downstream agents, we adapted the simulated reinforcement learning benchmark CausalWorld from that was developed for this platform. Building upon these works, we design our experimental study as follows (see [\[fig:experimental_setup\]](#fig:experimental_setup){reference-type="ref" reference="fig:experimental_setup"}): First, we pretrain representations from static simulated images of the setup and evaluate a collection of representation metrics. Following prior work, we focus on autoencoder-based representations. Then, we train downstream agents from this fixed representation on a set of environments. Finally, we investigate the zero-shot generalization of these agents to new environments that are out of the training distribution, including the real robot. The goal of this work is to provide the first systematic and extensive account of the OOD generalization of downstream RL agents in a robotic setup, and how this is affected by characteristics of the upstream pretrained representations. We summarize our contributions as follows:
- We train 240 representations and 11,520 downstream policies,[^1] and systematically investigate their performance under a diverse range of distribution shifts.[^2]
- We extensively analyze the relationship between the generalization of our RL agents and a substantial set of representation metrics.
- Notably, we find that a specific representation metric that measures the generalization of a simple downstream proxy task reliably predicts the generalization of downstream RL agents under the broad spectrum of OOD settings considered here. This metric can thus be used to select pretrained representations that will lead to more robust downstream policies.
- In the most challenging of our OOD scenarios, we deploy a subset of the trained policies to the corresponding real-world robotic platform, and observe surprising zero-shot sim-to-real generalization without any fine-tuning or domain randomization.
# Background {#sec:background}
In this section, we provide relevant background on the methods for representation learning and reinforcement learning, and on the robotic setup to evaluate out-of-distribution generalization. **Variational autoencoders.** VAEs are a framework for optimizing a latent variable model \(p_{\theta}(\mathbf{x}) = \int_\mathbf{z} p_{\theta}(\mathbf{x} \,|\, \mathbf{z}) p(\mathbf{z}) d\mathbf{z}\) with parameters \(\theta\), typically with a fixed prior \(p(\mathbf{z})=\mathcal{N}(\mathbf{z}; \mathbf{0}, \mathbf{I})\), using amortized stochastic variational inference. A variational distribution \(q_{\phi}(\mathbf{z} \,|\, \mathbf{x})\) with parameters \(\phi\) approximates the intractable posterior \(p_{\theta}(\mathbf{z} \,|\, \mathbf{x})\). The approximate posterior and generative model, typically called encoder and decoder and parameterized by neural networks, are jointly optimized by maximizing a lower bound to the log likelihood (the ELBO): \[\begin{aligned}
\log p_{\theta}(\mathbf{x}) &\geq \mathbb{E}_{q_{\phi}(\mathbf{z} \,|\, \mathbf{x})} \left[ \log p_{\theta}(\mathbf{x} \,|\, \mathbf{z}) \right]-D_\mathrm{KL}\left( q_{\phi}(\mathbf{z} \,|\, \mathbf{x}) \| p(\mathbf{z})\right) = \mathcal{L}^{ELBO}_{\theta,\phi} (\mathbf{x}) \. \end{aligned}\] In \(\beta\)-VAEs, the KL term is modulated by a factor \(\beta\) to enforce a more structured latent space. While VAEs are typically trained without supervision, we also employ a form of weak supervision that encourages disentanglement. **Reinforcement learning.** A Reinforcement Learning (RL) problem is typically modeled as a Partially Observable Markov Decision Process (POMDP) defined as a tuple \((S, A, T, R, \Omega, O, \gamma, \rho_{0}, H)\) with states \(s \in S\), actions \(a \in A\) and observations \(o \in \Omega\) determined by the state and action of the environment \(O(o|s,a)\). \(T(s_{t+1}|s_{t},a_{t})\) is the transition probability distribution function, \(R(s_t,a_t)\) is the reward function, \(\gamma\) is the discount factor, \(\rho_{0}(s)\) is the initial state distribution at the beginning of each episode, and \(H\) is the time horizon per episode. The objective in RL is to learn a policy \(\pi: S \times A \rightarrow [0, 1]\), typically parameterized by a neural network, that maximizes the total discounted expected reward \(J(\pi)=\mathbb{E} \big[ \sum_{t=0}^{H}\gamma^t R(s_t, a_t) \big]\). There is a broad range of model-free learning algorithms to find \(\pi^*\) by policy gradient optimization or by learning value functions while trading off exploration and exploitation. Here, we optimize the objective above with *Soft Actor Critic* (SAC), an off-policy method that simultaneously maximizes the expected reward and the entropy \(H(\pi(\cdot|s_t))\), and is widely used in control tasks due to its sample efficiency. **A robotic setup to evaluate out-of-distribution generalization.** Our study is based on a real robot platform where a robotic finger with three joints manipulates a cube in a bowl-shaped stage. We pretrain representations on a labeled dataset introduced by which consists of simulated and real-world images of this setup. This dataset has 7 underlying factors of variation (FoV): angles of the three joints, and position (x and y), orientation, and color of the cube. Some of these factors are correlated, which may be problematic for representation learners, especially in the context of disentanglement. After training the representations, we train downstream agents and evaluate their generalization on an adapted version of the simulated CausalWorld benchmark that was developed for the same setup. Finally, we test sim-to-real generalization on the real robot. Our experimental setup, illustrated in [\[fig:experimental_setup\]](#fig:experimental_setup){reference-type="ref" reference="fig:experimental_setup"}, allows us to systematically investigate a broad range of out-of-distribution scenarios in a controlled way. We pretrain our representations from this simulated dataset that covers 8 distinct cube colors. We then train an agent from this fixed representation on a subset of the cube colors, and evaluate it (1) on the same colors (this is the typical scenario in RL), (2) on the held-out cube colors that are still known to the encoder, or (3) OOD w.r.t. the encoder's training distribution, e.g. on novel colors and shapes or on the real world. We closely follow the framework for measuring OOD generalization proposed by. In this framework, a representation is initially learned on a training set \(\mathcal{D}\), and a simple downstream model is trained on a subset \(\mathcal{D}_1 \subset \mathcal{D}\) to predict the ground-truth factors from the learned representation. Generalization is then evaluated by testing the downstream model on a set \(\mathcal{D}_2\) that differs distributionally from \(\mathcal{D}_1\), e.g. containing images corresponding to held-out values of a chosen factor of variation (FoV). consider two flavors of OOD generalization depending on the choice of \(\mathcal{D}_2\): First, the case when \(\mathcal{D}_2 \subset \mathcal{D}\), i.e. the OOD test set is a subset of the dataset for representation learning. This is denoted by **OOD1** and corresponds to the scenario (2) from the previous paragraph. In the other scenario, referred to as **OOD2**, \(\mathcal{D}\) and \(\mathcal{D}_2\) are disjoint and distributionally different. This even stronger OOD shift corresponds to case (3) above. The generalization score for \(\mathcal{D}_2\) is then measured by the (normalized) mean absolute prediction error across all FoVs except for the one that is OOD. Following, we use a simple 2-layer Multi-Layer Perceptron (MLP) for downstream factor prediction, we train one MLP for each FoV, and report the *negative* error. This simple and cheap generalization metric could serve as a convenient proxy for the generalization of more expensive downstream tasks. We refer to these generalization scores as GS-OOD1, GS-OOD2-sim, and GS-OOD2-real depending on the scenario. The focus of was to scale VAE-based approaches to more realistic scenarios and study the generalization of these simple downstream tasks, with a particular emphasis on disentanglement. Building upon their contributions, we can leverage the broader potential of this robotic setup with many more OOD2 scenarios to study our research questions: To what extent can agents generalize under distribution shift? Do they generalize in different ways depending on the type of shift (including sim-to-real)? Can we predict the OOD generalization of downstream agents from properties of the pretrained representations such as the GS metrics from ?
# Study design {#sec:study_design}
**Robotic setup.** Our setup is based on TriFinger and consists of a robotic finger with three joints that can be controlled to manipulate an object (e.g. a cube) in a bowl-shaped stage. The agent receives a camera observation consistent with the images in and outputs a three-dimensional action. During training, which always happens in simulation, the agent only observes a cube of four possible colors, randomly sampled at every episode (see [\[fig:experimental_setup\]](#fig:experimental_setup){reference-type="ref" reference="fig:experimental_setup"}, step 2). **Distribution shifts.** After training, we evaluate these agents in 7 environments: (1) the training environment, which is the typical setting in RL, (2) the OOD1 setting with cube colors that are OOD for the agent but still in-distribution for the encoder, (3) the more challenging OOD2-sim setting where the colors are also OOD for the encoder, (4-6) the OOD2 settings where the object colors are as in the 3 previous settings but the cube is replaced by a sphere (a previously unseen shape), (7) the OOD2-real setting, where we evaluate zero-shot sim-to-real transfer on the real robotic platform. **Tasks.** We begin our study with the *object reaching* downstream control task, where the agent has to reach an object placed at an *arbitrary* random position in the arena. This is significantly more challenging than directly predicting the ground-truth factors, as the agent has to learn to reach the cube by acting on the joints, with a scalar reward as the only learning signal. Consequently, the compute required to learn this task is about 1,000 times greater than in the simple factor prediction case. We additionally include in our study a *pushing* task which consists of pushing an object to a goal position that is sampled at each episode. Learning this task takes one order of magnitude more compute than *object reaching*, likely due to the complex rigid-body dynamics and object interactions. To the best of our knowledge, this is the most challenging manipulation task that is currently feasible on our setup. report solving a similar pushing task, but require the full ground-truth state to be observable. **Training the RL agents.** The inputs at time \(t\) are the camera observation \(o_t\) and a vector of observable variables \(x_t\) containing the joint angles and velocities, as well as the target object position in *pushing*. We then feed the camera observation \(o_t\) into an encoder \(e\) that was pretrained on the dataset in. The result is concatenated with \(x_t\), yielding a state vector \(s_t = [x_t, e(o_t)]\). We then use SAC to train the policy with \(s_t\) as input. The policy, value, and Q networks are implemented as MLPs with 2 hidden layers of size 256. When training the policies, we keep the encoder frozen. **Model sweep.** To shed light on the research questions outlined in the previous sections, we perform a large-scale study in which we train 240 representation models and 11,520 downstream policies, as described below. See [\[app:implementation_details\]](#app:implementation_details){reference-type="ref" reference="app:implementation_details"} for further implementation details.
- We train 120 \(\beta\)-VAEs and 120 Ada-GVAEs with a subset of the hyperparameter configurations and neural architecture from. Specifically, we consider \(\beta \in \{1, 2, 4\}\), \(\beta\) annealing over \(\{0, 50000\}\) steps, with and without input noise, and 10 random seeds per configuration. The latent space size is fixed to \(10\) following prior work.
- For *object reaching*, we train 20 downstream policies (varying random seed) for each of the 240 VAEs. The resulting 4,800 policies are trained for 400k steps (approximately 2,400 episodes).
- Since *pushing* takes substantially longer to train, we limit the number of policies trained on this task: We choose a subset of 96 VAEs corresponding to only 4 seeds, and then use 10 seeds per representation. The resulting 960 policies are trained for 3M steps (about 9,000 episodes).
- Finally, for both tasks we also investigate the role of regularization on the policy. More specifically, we repeat the two training sweeps from above (5,760 policies), with the difference that now the policies are trained with L1 regularization on the first layer.
**Limitations of our study.** Although we aim to provide a sound and extensive empirical study, such studies are inevitably computationally demanding. Thus, we found it necessary to make certain design choices. For each of these choices, we attempted to follow common practice, in order to maintain our study as relevant, general, and useful as possible. One such decision is that of focusing on autoencoder-based representations. To answer our questions on the effect of upstream representations on the generalization of downstream policies, we need a diverse range of representations. How these representations are obtained is not directly relevant to answer our research question. Following, we chose to focus on \(\beta\)-VAE and Ada-GVAE models, as they were shown to provide a broad set of representations, including fully disentangled ones. Although we conjecture that other classes of representation learning algorithms should generally reveal similar trends as those found in our study, this is undoubtedly an interesting extension. As for the RL algorithm used in this work, SAC is known to be a particularly sample-efficient model-free RL method that is a popular choice in robotics. Extensive results on pushing from ground-truth features on the same setup in indicate that methods like TD3 or PPO perform very similarly to SAC under the same reward structure and observation space. Thus, we expect the results of our study to hold beyond SAC. Another interesting direction is the study of additional regularization schemes on the policy network, an aspect that is often overlooked in RL. We expect the potential insights from extending the study along these axes to not justify the additional compute costs and corresponding carbon footprint. However, with improving efficiency and decreasing costs, we believe these could become worthwhile validation experiments in the future.
# Results {#sec:results}
We discuss our results in three parts: In [4.1](#subsec:results_training_and_indistrib_performance){reference-type="ref" reference="subsec:results_training_and_indistrib_performance"}, we present the training results of our large-scale sweep, and how policy regularization and different properties of the pretrained representations affect in-distribution reward. gives an extensive account of which metrics of the pretrained representations predict OOD generalization of the agents in simulated environments. Finally, in [4.3](#subsec:results_ood_generalization_real_robot){reference-type="ref" reference="subsec:results_ood_generalization_real_robot"} we perform a similar evaluation on the real robot, in a zero-shot sim-to-real scenario.
## Results in the training environment {#subsec:results_training_and_indistrib_performance}
[\[fig:results_on_training_env\]](#fig:results_on_training_env){reference-type="ref" reference="fig:results_on_training_env"} shows the training curves of all policies for *object reaching* and *pushing* in terms of the task-specific success metric. Here we use success metrics for interpretability, as their range is always \([0,1]\). In *object reaching*, the success metric indicates progress from the initial end effector position to the optimal distance from the center of the cube. It is 0 if the final distance is not smaller than the initial distance, and 1 if the end effector is touching the center of a face of the cube. In *pushing*, the success metric is defined as the volumetric overlap of the cube with the goal cube, and the task can be visually considered solved with a score around 80%. From the training curves we can conclude that both tasks can be consistently solved from pixels using pretrained representations. In particular, all policies on *object reaching* attain almost perfect scores. Unsurprisingly, the more complex *pushing* task requires significantly more training, and the variance across policies is larger. Nonetheless, almost all policies learn to solve the task satisfactorily. To investigate the effect of representations on the training reward, we now compute its Spearman rank correlations with various supervised and unsupervised metrics of the representations ([\[fig:results_on_training_env\]](#fig:results_on_training_env){reference-type="ref" reference="fig:results_on_training_env"} bottom). By training reward, we mean the average reward of a fully trained policy over 200 episodes in the training environment (see [\[app:implementation_details\]](#app:implementation_details){reference-type="ref" reference="app:implementation_details"}). On *object reaching*, the final reward correlates with the ELBO and the reconstruction loss. A simple supervised metric to evaluate a representation is how well a small downstream model can predict the ground-truth factors of variation. Following, we use the MLP10000 and GBT10000 metrics (simply MLP and GBT in the following), where MLPs and Gradient Boosted Trees (GBTs) are trained to predict the FoVs from 10,000 samples. The training reward correlates with these metrics as well, especially with the MLP accuracy. This is not entirely surprising: if an MLP can predict the FoVs from the representations, our policies using the same architecture could in principle retrieve the FoVs relevant for the task. Interestingly, the correlation with the overall MLP metric mostly stems from the cube pose FoVs, i.e. those that are not included in the ground-truth state \(x_t\). These results suggest that these metrics can be used to select good representations for downstream RL. On the more challenging task of *pushing*, the correlations are milder but most of them are still statistically significant. **Summary.** Both tasks can be consistently solved from pixels using pretrained representations. Unsupervised (ELBO, reconstruction loss) and supervised (ground-truth factor prediction) in-distribution metrics of the representations are correlated with reward in the training environment.
## Out-of-distribution generalization in simulation {#subsec:results_ood_generalization_simulation}
**In-and out-of-distribution rewards.** After training, the in-distribution reward correlates with OOD1 performance on both tasks (especially with regularization), but not with OOD2 performance (see [\[fig:OOD_correlations_reward_trasfer\]](#fig:OOD_correlations_reward_trasfer){reference-type="ref" reference="fig:OOD_correlations_reward_trasfer"}). Moreover, rewards in OOD1 and OOD2 environments are moderately correlated across tasks and regularization settings. **Unsupervised metrics and informativeness.** In [\[fig:sim_world_OOD_rank_correlations_reaching_no_reg\]](#fig:sim_world_OOD_rank_correlations_reaching_no_reg){reference-type="ref" reference="fig:sim_world_OOD_rank_correlations_reaching_no_reg"} (left) we assess the relation between OOD reward and in-distribution metrics (ELBO, reconstruction loss, MLP, and GBT). Both ELBO and reconstruction loss exhibit a correlation with OOD1 reward, but not with OOD2 reward. These unsupervised metrics can thus be useful for selecting representations that will lead to more robust downstream RL tasks, as long as the encoder is in-distribution. While the GBT score is not correlated with reward under distribution shift, we observe a significant correlation between OOD1 reward and the MLP score, which measures downstream factor prediction accuracy of an MLP with the same architecture as the one parameterizing the policies. As in [4.1](#subsec:results_training_and_indistrib_performance){reference-type="ref" reference="subsec:results_training_and_indistrib_performance"}, we further investigate the source of this correlation, and find it in the pose parameters of the cube. Correlations in the OOD2 setting are much weaker, thus we conclude that these metrics do not appear helpful for model selection in this case. Our results on *pushing* confirm these conclusions although correlations are generally weaker, presumably due to the more complicated nature of this task. An extensive discussion is provided in [\[app:additional_results_ood_simulation\]](#app:additional_results_ood_simulation){reference-type="ref" reference="app:additional_results_ood_simulation"}. **Correlations with generalization scores.** Here we analyze the link between generalization in RL and the generalization scores (GS) discussed in [2](#sec:background){reference-type="ref" reference="sec:background"}, which measure the generalization of downstream FoV predictors *out of distribution*, as opposed to the MLP and GBT metrics considered above. For both OOD scenarios, the distribution shifts underlying these GS scores are the same as the ones in the RL tasks in simulation. We summarize our findings in [\[fig:sim_world_OOD_rank_correlations_reaching_no_reg\]](#fig:sim_world_OOD_rank_correlations_reaching_no_reg){reference-type="ref" reference="fig:sim_world_OOD_rank_correlations_reaching_no_reg"} (right) on the *object reaching* task. Reward in the OOD1 setting is significantly correlated with the GS-OOD1 metric of the pretrained representation. We observe an even stronger correlation between the reward in the simulated OOD2 setting and the corresponding GS-OOD2-sim and GS-OOD2-real scores. On a per-factor level, we see that the source of the observed correlations primarily stems from the generalization scores w.r.t. the pose parameters of the cube. The OOD generalization metrics can therefore be used as proxies for the corresponding form of generalization in downstream RL tasks. This has practical implications for the training of RL downstream policies which are generally known to be brittle to distribution shifts, as we can measure a representation's generalization score from a few labeled images. This allows for selecting representations that yield more robust downstream policies. **Disentangled representations.** [\[subsubsec:disentanglement_and_RL_reward\]]{#subsubsec:disentanglement_and_RL_reward label="subsubsec:disentanglement_and_RL_reward"} Disentanglement has been shown to be helpful for downstream performance and OOD1 generalization even with MLPs. However, in *object reaching*, we only observe a weak correlation with some disentanglement metrics ([\[fig:disentanglement_and generalization_reaching\]](#fig:disentanglement_and generalization_reaching){reference-type="ref" reference="fig:disentanglement_and generalization_reaching"}). In agreement with, disentanglement does not correlate with OOD2 generalization. The same study observed that disentanglement correlates with the informativeness of a representation. To understand if these weak correlations originate from this common confounder, we investigate whether they persist after adjusting for MLP FoV prediction accuracy. Given two representations with similar MLP accuracy, does the more disentangled one exhibit better OOD1 generalization? To measure this we predict success from the MLP accuracy using kNN (k=5) and compute the residual reward by subtracting the amount of reward explained by the MLP metric. [\[fig:disentanglement_and generalization_reaching\]](#fig:disentanglement_and generalization_reaching){reference-type="ref" reference="fig:disentanglement_and generalization_reaching"} shows that this resolves the remaining correlations with disentanglement. Thus, for the RL downstream tasks considered here, disentanglement per se does not seem to be useful for OOD generalization. We present similar results on *pushing* in [\[app:additional_results_ood_simulation\]](#app:additional_results_ood_simulation){reference-type="ref" reference="app:additional_results_ood_simulation"}. **Policy regularization and observation noise.** It might seem unsurprising that disentanglement is not useful for generalization in RL, as MLP policies do not have any explicit inductive bias to exploit it. Thus, we attempt to introduce such inductive bias by repeating all experiments with L1 regularization on the first layer of the policy. Although regularization improves OOD1 and OOD2 generalization in general (see box plots in [\[fig:disentanglement_and generalization_reaching\]](#fig:disentanglement_and generalization_reaching){reference-type="ref" reference="fig:disentanglement_and generalization_reaching"}), we observe no clear link with disentanglement. Furthermore, in accordance with, we find that observation noise when training representations is beneficial for OOD2 generalization. See [\[app:additional_results_ood_simulation\]](#app:additional_results_ood_simulation){reference-type="ref" reference="app:additional_results_ood_simulation"} for a detailed discussion. **Stronger OOD shifts: evaluating on a novel shape.** On *object reaching*, we also test generalization w.r.t. a novel shape by replacing the cube with a sphere. This corresponds to a strong OOD2-type shift, since shape was never varied when training the representations. Surprisingly, the policies appear to be robust to the novel shape. In fact, when the sphere has the same colors that the cube had during policy training, *all* policies get closer than 5 cm to the sphere on average, with a mean success metric of 95%. On sphere colors from the OOD1 split, more than 98.5% move the finger closer than this threshold, and on the strongest distribution shift (OOD2-sim colors, and cube replaced by sphere) almost 70% surpass that threshold with an average success metric above 80%. **Summary.** (1) In-and out-of-distribution rewards are correlated, as long as the representation remains in its training distribution (OOD1). (2) Similarly, in-distribution representation metrics (both unsupervised and supervised) predict OOD1 reward, but are not reliable when the representation is OOD (OOD2). (3) Disentanglement does not correlate with generalization in our experiments, while (4) input noise when training representations is beneficial for OOD2 generalization. (5) Most notably, the GS metrics, which measure generalization under distribution shifts, are significantly correlated with RL performance under similar distribution shifts. We thus recommend using these convenient proxy metrics for selecting representations that will yield robust downstream policies.
## Deploying policies to the real world {#subsec:results_ood_generalization_real_robot}
We now evaluate a large subset of the agents on the real robot without fine-tuning, quantify their zero-shot sim-to-real generalization, and find metrics that correlate with real-world performance. **Reaching.** We choose 960 policies trained in simulation, based on 96 representations and 10 random seeds, and evaluate them on two (randomly chosen, but far apart) goal positions using a red cube. While a red cube was in the training distribution, we consider this to be OOD2 because real-world images represent a strong distribution shift for the encoder . Although sim-to-real in robotics is considered to be very challenging without domain randomization or fine-tuning, many of our policies obtain a high fractional success without resorting to these methods. In addition, in [\[fig:OOD2_evaluations_reaching\]](#fig:OOD2_evaluations_reaching){reference-type="ref" reference="fig:OOD2_evaluations_reaching"} (left) we observe significant correlations between zero-shot real-world performance and some of the previously discussed metrics. First, there is a positive correlation with the OOD2-sim reward: Policies that generalize to unseen cube colors in simulation also generalize to the real world. Second, representations with high GS-OOD2-sim and (especially) GS-OOD2-real scores are promising candidates for sim-to-real transfer. Third, if no labels are available, the weaker correlation with the reconstruction loss on the simulated images can be exploited for representation selection. Finally, as observed by
for simple downstream tasks, input noise while learning representations is beneficial for sim-to-real generalization ([\[fig:OOD2_evaluations_reaching\]](#fig:OOD2_evaluations_reaching){reference-type="ref" reference="fig:OOD2_evaluations_reaching"}, middle). Based on these findings, we select 50 policies with a high GS-OOD2-real score, and evaluate them on the real world with a green and a blue cube, which is an even stronger OOD2 distribution shift. In [\[fig:OOD2_evaluations_reaching\]](#fig:OOD2_evaluations_reaching){reference-type="ref" reference="fig:OOD2_evaluations_reaching"} (right), where metrics are averaged over 4 cube positions per policy, we observe that most policies can still solve the task: approximately 80% of them position the finger less than 5 cm from the cube. Lastly, we repeat the evaluations on the green sphere that we previously performed in simulation, and observe that many policies successfully reach this completely novel object. See [\[app:additional_results_ood_real_world\]](#app:additional_results_ood_real_world){reference-type="ref" reference="app:additional_results_ood_real_world"} and the project website for additional results and videos of deployed policies. **Pushing.** We now test whether our real-world findings on *object reaching* also hold for *pushing*. We again select policies with a high GS-OOD2-real score and encoders trained with input noise. We record episodes on diverse goal positions and cube colors to support our finding that pushing policies in simulation can generalize to the real robot. In [\[fig:pushing_real_robot_frames\]](#fig:pushing_real_robot_frames){reference-type="ref" reference="fig:pushing_real_robot_frames"}, we show three representative episodes with successful task completions and refer to the project site for video recordings and further episodes. **Summary.** Policies trained in simulation can solve the task on the real robot without domain randomization or fine-tuning. Reconstruction loss, encoder robustness, and OOD2 reward in simulation are all good predictors of real-world performance. For real-world applications, we recommend using GS-OOD2-sim or GS-OOD2-real for model selection, and training the encoder with noise.
# Other related work
A key unsolved challenge in RL is the brittleness of agents to distribution shifts in the environment, even if the underlying structure is largely unchanged . This is related to studies on representation learning and generalization in downstream tasks, as well as domain generalization (see for an overview). More specifically for RL, focus on domain adaptation and zero-shot transfer in DeepMind Lab and MuJoCo environments, and claim disentanglement improves robustness. To obtain better transfer capabilities, argue for discretizing the state space in continuous control domains by clustering states where the optimal policy is similar. propose geometric object representations by means of keypoints or image-space coordinates and investigate the effect of different representations on the learning and exploration of different robotics tasks. Transfer becomes especially challenging from the simulation to the real world, a phenomenon often referred to as the sim-to-real gap. This is particularly crucial in RL, as real-world training is expensive, requires sample-efficient methods, and is sometimes unfeasible if the reward structure requires accurate ground truth labels. This issue is typically tackled with large-scale domain randomization in simulation.
# Conclusion
Robust out-of-distribution (OOD) generalization is still one of the key open challenges in machine learning. We attempted to answer central questions on the generalization of reinforcement learning agents in a robotics context, and how this is affected by pretrained representations. We presented a large-scale empirical study in which we trained over 10,000 downstream agents given pretrained representations, and extensively tested them under a variety of distribution shifts, including sim-to-real. We observed agents that generalize OOD, and found that some properties of the pretrained representations can be useful to predict which agents will generalize better. We believe this work brings us one step closer to understanding the generalization abilities of learning systems, and we hope that it encourages many further important studies in this direction.
# Ethics statement {#ethics-statement .unnumbered}
Our study is based on synthetic and real data of a robotic setup where a robotic finger interacts with a cube. Our study does therefore not involve any human subjects leading to discrimination, bias or fairness concerns, or privacy and security issues. Representation learning and generalization is important across many disciplines and applications and could have harmful consequences without humans in the loop in safety-relevant settings. Having a sound understanding of the robustness of a given ML system based on such pretrained representations to distributions shifts is crucial to avoid harmful consequences in potential future high-stake applications to society, such as human-robot interaction (e.g. robotic surgery), autonomous driving, healthcare applications or other fairness-related settings. Here, we investigate a narrow aspect of this, that is, learning arguably harmless manipulation skills like reaching or pushing an object with a simple robotic finger. Our conclusions for OOD generalization are based on this setting and thus cannot be directly transferred to any given application setting of concern.
# Reproducibility statement {#reproducibility-statement .unnumbered}
To make sure our experiments are fully reproducible, we provided a full account of all required setup and implementation details in [3](#sec:study_design){reference-type="ref" reference="sec:study_design"} and [\[app:implementation_details\]](#app:implementation_details){reference-type="ref" reference="app:implementation_details"}. | {'timestamp': '2022-04-19T02:18:29', 'yymm': '2107', 'arxiv_id': '2107.05686', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05686'} |
null | null |
# Introduction {#intro}
In a previous work the extended time-dependent Ginzburg-Landau (ETDGL) theory has been formulated, which is a generalization of the Ginsburg-Landau (GL) theory for the nonstationary regimes: the damped eigen oscillations (including the relaxation) and the forced oscillations of the order parameter (OP) \(\Psi(\mathbf{r},t)\) under the action of an external field. In this theory an action with Lorentz invariant Lagrangian for the complex scalar field \(\Psi=|\Psi|e^{i\theta}\) and the gauge field \(A^{\mu}=(\varphi, \mathbf{A})\) in some 4D Minkowski space \(\{\upsilon t,\mathbf{r}\}\), where speed \(\upsilon\) is determined with the dynamical properties of the system, has been proposed. At the same time, the dynamics of conduction electrons remains non-relativistic. Accounting of movement of the normal component, which is accompanied by friction, makes the theory be not Lorentz covariant. The superconducting (SC) system has two types of collective excitations: with an energy gap (quasi-relativistic spectrum) \(E^{2}=\widetilde{m}^{2}\upsilon^{4}+p^{2}\upsilon^{2}\) (where \(\widetilde{m}\) is the mass of a Higgs boson, so that \(\widetilde{m}\upsilon^{2}=2|\Delta|\))-Higgs mode, and with an acoustic (ultrarelativistic) spectrum \(E=p\upsilon\)-Goldstone mode. The light speed \(\upsilon\) is determined from the dynamical properties of the system, and it is much less than the vacuum light speed: \(\upsilon=v_{F}/\sqrt{3}\ll c\) (\(v_{F}\) is Fermi velocity). The Higgs mode is oscillations of modulus of the OP \(|\Psi(t,\mathbf{r})|\), and it can be considered as sound in the gas of above-condensate quasiparticles (at \(T\rightarrow T_{c}\)). The propagation of a Higgs boson is not accompanied by charge transfer. It should be noted, that the free Higgs mode is unstable due to both strong damping of these oscillations at \(T\rightarrow T_{c}\), so that the aperiodic relaxation takes place with decay into above-condensate quasiparticles, since \(E(q)\geq 2|\Delta|\). The Goldstone mode is oscillations of the phase \(\theta(\mathbf{r},t)\), and it is absorbed into the gauge field \(A^{\mu}\) according to the Anderson-Higgs mechanism. The Goldstone oscillations cannot be accompanied by oscillations of the charge density, they generate the transverse field \(\textrm{div}\mathbf{A}=0\) only and they are eddy currents (i.e. \(\textrm{div}\mathbf{j}=0\)), as result of the boundary conditions. From the gauge invariance of the Lagrangian it follows, that the superconductor is equivalent to a dielectric (in some effective sense) with the permittivity \(\varepsilon=\frac{c^{2}}{\upsilon^{2}}\) only for the induced electric field \(\mathbf{E}=-\frac{1}{c}\frac{\partial\mathbf{A}}{\partial t}\) (with frequencies \(0<\hbar\omega<2|\Delta|\)). Thus, the speed \(\upsilon\) is the speed of light in SC medium, if there were no skin-effect and Meissner effect. At the same time, inside the superconductor the potential electric field is absent \(\mathbf{E}=-\nabla\varphi=0\) as consequence of the boundary conditions. For the electrostatic field \(\mathbf{E}=-\nabla\varphi\) the permittivity is \(\varepsilon(\omega=0,\mathbf{q}=0)=\infty\) as in metals. Two-band superconductors are a specific class of superconductors essentially differing in their properties from single-band superconductors. The main feature of these materials is the presence of two OP-\"wave functions\" \(\Psi_{1}\) and \(\Psi_{2}\) corresponding to the condensates of Cooper pairs in each band. In a bulk isotropic s-wave superconductor the GL free energy functional can be written as : \[\begin{aligned}
\label{1.1} F&=&\int d^{3}r[\frac{\hbar^{2}}{4m_{1}}\left|\nabla\Psi_{1}\right|^{2}+\frac{\hbar^{2}}{4m_{2}}\left|\nabla\Psi_{2}\right|^{2} +\frac{\hbar^{2}}{4}\eta\left(\nabla\Psi_{1}\nabla\Psi_{2}^{+}+\nabla\Psi_{1}^{+}\nabla\Psi_{2}\right)\nonumber\\ &+&a_{1}\left|\Psi_{1}\right|^{2}+a_{2}\left|\Psi_{2}\right|^{2}+\frac{b_{1}}{2}\left|\Psi_{1}\right|^{4}+\frac{b_{2}}{2}\left|\Psi_{2}\right|^{4} +\epsilon\left(\Psi_{1}^{+}\Psi_{2}+\Psi_{1}\Psi_{2}^{+}\right)], \end{aligned}\] where \(m_{1,2}\) denote the effective mass of carriers in the corresponding band, the coefficients \(a_{1,2}\) are given as \(a_{i}=\gamma_{i}(T-T_{ci})\) where \(\gamma_{i}\) are some constants, the coefficients \(b_{1,2}\) are independent of temperature, the quantities \(\epsilon\) and \(\eta\) describe interband mixing of the two OP (proximity effect) and their gradients (drag effect), respectively. If we switch off the interband interactions \(\epsilon=0\) and \(\eta=0\), then we will have two independent superconductors with different critical temperatures \(T_{c1}\) and \(T_{c2}\) because the intraband interactions can be different. Thus, a two-band superconductor is understood as two single-band superconductors with the corresponding condensates of Cooper pairs \(\Psi_{1}\) and \(\Psi_{2}\) (so that densities of SC electrons are \(n_{\mathrm{s}1}=2|\Psi_{1}|^{2}\) and \(n_{\mathrm{s}2}=2|\Psi_{2}|^{2}\) accordingly), but these two condensates are coupled by both the internal proximity effect \(\epsilon\left(\Psi_{1}^{+}\Psi_{2}+\Psi_{1}\Psi_{2}^{+}\right)\) and the \"drag\" effect \(\eta\left(\nabla\Psi_{1}\nabla\Psi_{2}^{+}+\nabla\Psi_{1}^{+}\nabla\Psi_{2}\right)\). In presence of a magnetic potential \(\textbf{A}\) the replacement \(\nabla\rightarrow\nabla-\frac{i2e}{c\hbar}\textbf{A}\) must be done in the free energy functional ([\[1.1\]](#1.1){reference-type="ref" reference="1.1"}) for the gauge invariance. The magnetic response (penetration, critical fields etc.) has been considered in Ref.. Minimization of the free energy functional with respect to the OP, if \(\nabla\Psi_{1,2}=0\), gives \[\label{1.2} \left\{\begin{array}{c} a_{1}\Psi_{1}+\epsilon\Psi_{2}+b_{1}\Psi_{1}^{3}=0 \\ a_{2}\Psi_{2}+\epsilon\Psi_{1}+b_{2}\Psi_{2}^{3}=0 \\ \end{array}\right\},\] where the equilibrium values \(\Psi_{1,2}\) are assumed to be real (i.e. the phases \(\theta_{1,2}\) are \(0\) or \(\pi\)) in absence of current and magnetic field in the case of a two-band superconductor (but not for three-band superconductors, where the equilibrium phase differences can be not only \(0\) or \(\pi\), but it can be \(2\pi/3\) or \(\pi/3,2\pi/3\) depending on the signs of the interband interactions \(\epsilon_{ik}\), and chiral ground states triggered by the interband interaction occur ). Near the critical temperature \(T_{c}\) we have \(\Psi_{1,2}^{3}\rightarrow 0\), hence, we can find the critical temperature equating to zero the determinant of the linearized system ([\[1.2\]](#1.2){reference-type="ref" reference="1.2"}): \[\label{1.3} a_{1}a_{2}-\epsilon^{2}=\gamma_{1}\gamma_{2}(T_{c}-T_{c1})(T_{c}-T_{c2})-\epsilon^{2}=0.\] Solving this equation, we find \(T_{c}>T_{c1},T_{c2}\), moreover, the solution does not depend on the sign of \(\epsilon\). The sign determines the equilibrium phase difference of the OP \(|\Psi_{1}|e^{i\theta_{1}}\) and \(|\Psi_{2}|e^{i\theta_{2}}\): \[\label{1.4} \begin{array}{cc} \cos(\theta_{1}-\theta_{2})=1 & \mathrm{if}\quad\epsilon<0 \\ \cos(\theta_{1}-\theta_{2})=-1 & \mathrm{if}\quad\epsilon>0 \\ \end{array},\] that follows from Eq.([\[1.2\]](#1.2){reference-type="ref" reference="1.2"}). Then, in the linear approximation at \(T\rightarrow T_{c}\) (\(T>T_{c1},T_{c2}\)) we have \(\Psi_{2}=-\frac{a_{1}}{\epsilon}\Psi_{1}=-\mathrm{sgn}(\epsilon)\frac{a_{1}}{|\epsilon|}\Psi_{1} =-\mathrm{sgn}(\epsilon)\sqrt{\frac{a_{1}}{a_{2}}}\Psi_{1}\). The case \(\epsilon<0\) corresponds to an attractive interband interaction (for example, in \(\mathrm{MgB}_{2}\), where \(s^{++}\) wave symmetry occurs), the case \(\epsilon>0\) corresponds to a repulsive interband interaction (for example, in iron-based superconductors, where \(s^{+-}\) wave symmetry occurs). Thus, in the absence of interband coupling we obtain two independent bands, in which the OP can exhibit any sign. In the converse situation, the signs of the OP are determined by the sign of the interband coupling, that has been demonstrated with numerical calculations in the minimal two-orbital model of iron based superconductors in. The solutions of Eq.([\[1.2\]](#1.2){reference-type="ref" reference="1.2"}) are illustrated in Fig.[\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"} for the case of strongly asymmetrical bands \(T_{c1}\ll T_{c2}\). We can see, that the effect of interband coupling \(\epsilon\neq 0\), even if the coupling is weak \(|\epsilon|\ll |a_{1}(0)|\), is nonperturbative for the smaller OP \(\Psi_{1}\): applying of the interband coupling washes out the smaller parameter up to the new critical temperature \(T_{c}\gg T_{c1}\). At the same time, the effect on the larger parameter \(\Psi_{2}\) is not so significant-applying of the interband coupling only slightly increases the critical temperature \(T_{c}\gtrsim T_{c2}\). This property corresponds to numerical solutions of self-consistent equations for superconducting gaps \(\Delta_{1}\) and \(\Delta_{2}\) in two-band systems with both \(s\)-wave and \(d\)-wave symmetries and have been most clearly observed for \(\mathrm{Mg}_{1-x}\mathrm{Al}_{x}\mathrm{B}_{2}\) and \(\mathrm{LaO}_{1-x}\mathrm{F}_{x}\mathrm{FeAs}\) in Ref. and \(\mathrm{Mg}\mathrm{B}_{2}\) in Ref.. It is noteworthy, that the lower critical temperature \(T_{c1}\) makes some mark in the temperature dependence of the specific heat \(C(T)\): convexity, curvature and possible non-monotonicity near the temperature \(T_{c1}\), which is demonstrated by numerical calculations in Ref. and has been observed experimentally for some iron-based superconductors \(\mathrm{KFe}_{2}\mathrm{As}_{2}\), \(\mathrm{Lu}_{2}\mathrm{Fe}_{3}\mathrm{Si}_{5}\) and for classical two-band superconductor \(\mathrm{MgB}_{2}\). In references it has been shown, that in a two-band superconductor there are two coherence lengths, which are not related to the concrete bands involved in the formation of SC state in a system with the interband interaction: one of the lengths diverges at the critical temperature \(In Ref. it has been shown, that the term of the drag effect\)(\_1\_2\^++\_1\^+\_2)\(in the free energy functional of two-band superconductor plays an important role and the restrictions for the coefficient\[exist. If the coefficient is\)\^2=\(and it’s sign is opposite to the sign of the coefficient in the term of the proximity effect\)(\_1\^+\_2+\_1\_2\^+)\(, that is\)\<0\(, then this leads to single coherence length\) As has been demonstrated in Ref., in two-band superconductors the Goldstone mode splits into two branches: common mode oscillations with acoustic spectrum, and the oscillations of the relative phase \(\theta_{1}-\theta_{2}\) between two SC condensates with the energy gap in the spectrum determined by the interband coupling-the Leggett mode (for symmetrical condensates we have the gap: \(\sqrt{8|\epsilon|m\upsilon^{2}}\)). The common mode oscillations are absorbed into the gauge field \(A_{\mu}\) as in single-band superconductors, at the same time, the Leggett mode \"survives\" due to these oscillations are not accompanied by current. However, in this work, firstly, Higgs oscillations in two-band superconductors have not been investigated, secondly, effect of the drag term \(\eta\left(\nabla\Psi_{1}\nabla\Psi_{2}^{+}+\nabla\Psi_{1}^{+}\nabla\Psi_{2}\right)\) has not been investigated either. At the same time, as we can see from Ref., the drag term can have fundamental importance. Proceeding from the aforesaid, we aim to obtain eigen oscillations of internal degrees of freedom (the Higgs mode and the Goldstone mode) of two-band superconductors using the ETDGL theory formulated in Ref. for the case of two coupled OP by both the internal proximity effect and the drag effect. Our paper is organized as follows. In Sect.[2](#Goldstone and Higgs){reference-type="ref" reference="Goldstone and Higgs"} we generalize the two-band free energy ([\[1.1\]](#1.1){reference-type="ref" reference="1.1"}) to some action describing non-stationary regimes. Then, we obtain the spectra of both the Higgs mode and the Goldstone mode, which split into two branches each: common mode oscillations and anti-phase oscillations. It is demonstrated, that the second branch of the Higgs mode has a nonphysical property and it, together with the Leggett mode, can be removed with special choice of the coefficient \(\eta\). Thus, we demonstrate, that the drag effect plays a fundamental role in the internal dynamics of two-band superconductors. In Sect.[\[Josephson\]](#Josephson){reference-type="ref" reference="Josephson"} we analyze experimental data about the Josephson effect between two-band superconductors. In particular, it is demonstrated, that the resonant enhancement of the DC current through a Josephson junction at a resonant bias voltage \(V_{\mathrm{res}}\), when the Josephson frequency matches the frequency of some internal oscillation mode in two-band superconductors (banks), can be explained with the coupling between AC Josephson current and Higgs oscillations in two-band superconductors. Thus, explanation of this effect does not need the Leggett mode.
# Goldstone and Higgs oscillations in two-band superconductors {#Goldstone and Higgs}
## Ginzburg-Landau Lagrangian for two-band superconductors
In the general case the OP \(\Psi_{1,2}\) are both spatially inhomogeneous and they can change over time: \(\Psi_{1,2}=\Psi_{1,2}(\textbf{r},t)\). The OP are complex scalar fields, which are equivalent to two real fields each: modulus \(\left|\Psi(\textbf{r},t)\right|\) and phase \(\theta(\textbf{r},t)\) (the modulus-phase representation): \]\label{1.5} \Psi_{1}(\textbf{r},t)=\left|\Psi_{1}(\textbf{r},t)\right|e^{i\theta_{1}(\textbf{r},t)}, \quad\Psi_{2}(\textbf{r},t)=\left|\Psi_{2}(\textbf{r},t)\right|e^{i\theta_{2}(\textbf{r},t)}.\[ For the stationary case \(\Psi_{1,2}=\Psi_{1,2}(\textbf{r})\) the steady configuration of the field \(\Psi_{1,2}(\textbf{r})\) minimizes the free energy functional ([\[1.1\]](#1.1){reference-type="ref" reference="1.1"}). However, for the nonstationary case \(\Psi_{1,2}(\textbf{r},t)\) the minimization procedure makes no sense. According to the method described in, the parameter \(t\)-the time can be turned into a coordinate \(t\rightarrow\upsilon t\) in some 4D Minkowski space \(\{\upsilon t,\textbf{r}\}\), where \(\upsilon\) is an parameter of dimension of speed (as the light speed), which must be determined from the dynamical properties of the system. At the same time, the dynamics of conduction electrons remains non-relativistic. Then the two-component scalar fields \(\Psi_{1,2}(\textbf{r},t)\) minimizes some action \(S\) (as in the relativistic field theory ) in the Minkowski space: \]\label{1.6} S=\frac{1}{\upsilon}\int\mathcal{L}(\Psi_{1},\Psi_{2},\Psi^{+}_{1},\Psi^{+}_{2})\upsilon dtd^{3}r.\[ The Lagrangian \(\mathcal{L}\) is built by generalizing the density of free energy in Eq.([\[1.1\]](#1.1){reference-type="ref" reference="1.1"}) to the \"relativistic\" invariant form by substitution of covariant and contravariant differential operators: \]\label{1.7} \widetilde{\partial}_{\mu}\equiv\left(\frac{1}{\upsilon}\frac{\partial}{\partial t},\nabla\right), \quad\widetilde{\partial}^{\mu}\equiv\left(\frac{1}{\upsilon}\frac{\partial}{\partial t},-\nabla\right),\[ instead the gradient operators: \(\nabla\Psi\rightarrow\widetilde{\partial}_{\mu}\Psi,\quad\nabla\Psi^{+}\rightarrow\widetilde{\partial}^{\mu}\Psi^{+}\). Then the required Lagrangian takes the form: \]\begin{aligned}
\label{1.8} \mathcal{L}&=&\frac{\hbar^{2}}{4m_{1}}\widetilde{\partial}_{\mu}\Psi_{1}\widetilde{\partial}^{\mu}\Psi_{1}^{+} +\frac{\hbar^{2}}{4m_{2}}\widetilde{\partial}_{\mu}\Psi_{2}\widetilde{\partial}^{\mu}\Psi_{2}^{+} +\frac{\hbar^{2}}{4}\eta\left(\widetilde{\partial}_{\mu}\Psi_{1}\widetilde{\partial}^{\mu}\Psi_{2}^{+} +\widetilde{\partial}^{\mu}\Psi_{1}^{+}\widetilde{\partial}_{\mu}\Psi_{2}\right)\nonumber\\ &-&a_{1}\left|\Psi_{1}\right|^{2}-\frac{b_{1}}{2}\left|\Psi_{1}\right|^{4}-a_{2}\left|\Psi_{2}\right|^{2}-\frac{b_{2}}{2}\left|\Psi_{2}\right|^{4} -\epsilon\left(\Psi_{1}^{+}\Psi_{2}+\Psi_{1}\Psi_{2}^{+}\right), \end{aligned}\[ where the same speed \(\upsilon\) is used for both \(\Psi_{1}\) and \(\Psi_{2}\) with the masses \(m_{1}\) and \(m_{2}\) accordingly (it should be noted the property \(\widetilde{\partial}_{\mu}\Psi_{1}\widetilde{\partial}^{\mu}\Psi_{2}^{+}=\widetilde{\partial}^{\mu}\Psi_{1}\widetilde{\partial}_{\mu}\Psi_{2}^{+}\), etc.). The speed \(\upsilon\sim\upsilon_{F1},\upsilon_{F2}\) (here, \(\upsilon_{F1,2}\) are Fermi velocities in each band accordingly) plays role of the light speed in SC medium, and it will be found below. In the presence of el.-mag. field \(A_{\mu}=(\varphi,-\textbf{A})\), the replacement \(\partial_{\mu}\rightarrow\left(\partial_{\mu}+\frac{i2e}{c\hbar}\widetilde{A}_{\mu}\right)\) must be done in Lagrangian ([\[1.8\]](#1.8){reference-type="ref" reference="1.8"}) for a gauge invariance (here, \(\widetilde{A}_{\mu}=(\frac{c}{\upsilon}\varphi,-\textbf{A})\)). The electromagnetic response has been considered in Ref.. In the present work we will consider only collective excitations, which are not accompanied by a current and a charge transfer (the Higgs mode, the Leggett mode etc.). Substituting representation ([\[1.5\]](#1.5){reference-type="ref" reference="1.5"}) in the Lagrangian ([\[1.8\]](#1.8){reference-type="ref" reference="1.8"}) we obtain: \]\begin{aligned}
\label{1.9} \mathcal{L}&=&\frac{\hbar^{2}}{4m_{1}}\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}|\Psi_{1}| +\frac{\hbar^{2}}{4m_{1}}|\Psi_{1}|^{2}\widetilde{\partial}_{\mu}\theta_{1}\widetilde{\partial}^{\mu}\theta_{1} +\frac{\hbar^{2}}{4m_{2}}\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}|\Psi_{2}| +\frac{\hbar^{2}}{4m_{2}}|\Psi_{2}|^{2}\widetilde{\partial}_{\mu}\theta_{2}\widetilde{\partial}^{\mu}\theta_{2}\nonumber\\ &+&\frac{\hbar^{2}}{4}\eta\left(\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}|\Psi_{2}| +\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}|\Psi_{1}|\right)\cos\left(\theta_{1}-\theta_{2}\right) +\frac{\hbar^{2}}{4}\eta\left(\widetilde{\partial}_{\mu}\theta_{1}\widetilde{\partial}^{\mu}\theta_{2} +\widetilde{\partial}_{\mu}\theta_{2}\widetilde{\partial}^{\mu}\theta_{1}\right)|\Psi_{1}||\Psi_{2}| \cos\left(\theta_{1}-\theta_{2}\right)\nonumber\\ &+&\frac{\hbar^{2}}{4}\eta\left(\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}\theta_{2} +\widetilde{\partial}^{\mu}|\Psi_{1}|\widetilde{\partial}_{\mu}\theta_{2}\right)|\Psi_{2}|\sin\left(\theta_{1}-\theta_{2}\right) -\frac{\hbar^{2}}{4}\eta\left(\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}\theta_{1} +\widetilde{\partial}^{\mu}|\Psi_{2}|\widetilde{\partial}_{\mu}\theta_{1}\right)|\Psi_{1}|\sin\left(\theta_{1}-\theta_{2}\right)\nonumber\\ &-&a_{1}\left|\Psi_{1}\right|^{2}-\frac{b_{1}}{2}\left|\Psi_{1}\right|^{4}-a_{2}\left|\Psi_{2}\right|^{2}-\frac{b_{2}}{2}\left|\Psi_{2}\right|^{4} -2|\Psi_{1}||\Psi_{2}|\epsilon\cos\left(\theta_{1}-\theta_{2}\right), \end{aligned}\[ where we have used the following properties, as \(\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}|\Psi_{2}|=\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}|\Psi_{1}|\) and \(\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}\theta_{2}=\widetilde{\partial}^{\mu}|\Psi_{1}|\widetilde{\partial}_{\mu}\theta_{2}\). We will see, that modulus and phase variables can be separated in the linear approximation. Thus, the field coordinates \(|\Psi_{1,2}(\mathbf{r},t)|\) and \(\theta_{1,2}(\mathbf{r},t)\) are the normal coordinates, and their small oscillations are the normal oscillations.
## Goldstone oscillations
Let us consider movement of the phases \(\theta_{1,2}\). The corresponding Lagrange equations are \]\begin{aligned}
\widetilde{\partial}_{\mu}\frac{\partial\mathcal{L}}{\partial(\widetilde{\partial}_{\mu}\theta_{1})} -\frac{\partial\mathcal{L}}{\partial\theta_{1}}=0\Rightarrow&& \frac{\hbar^{2}}{4m_{1}}|\Psi_{1}|^{2}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{1}+ \frac{\hbar^{2}}{4}|\Psi_{1}||\Psi_{2}|\eta\cos\left(\theta_{1}-\theta_{2}\right)\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{2}-\nonumber\\ &&\frac{\hbar^{2}}{4}\eta|\Psi_{1}|\sin\left(\theta_{1}-\theta_{2}\right)\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}|\Psi_{2}| -|\Psi_{1}||\Psi_{2}|\epsilon\sin\left(\theta_{1}-\theta_{2}\right) = 0\label{2.2a}\\ \widetilde{\partial}_{\mu}\frac{\partial\mathcal{L}}{\partial(\widetilde{\partial}_{\mu}\theta_{2})} -\frac{\partial\mathcal{L}}{\partial\theta_{2}}=0\Rightarrow&& \frac{\hbar^{2}}{4m_{1}}|\Psi_{2}|^{2}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{2}+ \frac{\hbar^{2}}{4}|\Psi_{1}||\Psi_{2}|\eta\cos\left(\theta_{1}-\theta_{2}\right)\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{1} +\nonumber\\ &&\frac{\hbar^{2}}{4}\eta|\Psi_{2}|\sin\left(\theta_{1}-\theta_{2}\right)\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}|\Psi_{1}| +|\Psi_{1}||\Psi_{2}|\epsilon\sin\left(\theta_{1}-\theta_{2}\right) = 0\label{2.2b} \end{aligned}\[ where we have omitted nonlinear terms, such as \(\widetilde{\partial}_{\mu}\theta\widetilde{\partial}^{\mu}\theta\), \(\widetilde{\partial}_{\mu}|\Psi|\widetilde{\partial}^{\mu}|\Psi|\), \(\widetilde{\partial}_{\mu}\theta\widetilde{\partial}^{\mu}|\Psi|\). The phases can be written in the form of harmonic oscillations: \]\label{2.3} \begin{array}{c} \theta_{1}=\theta_{1}^{0}+Ae^{i(\mathbf{qr}-\omega t)}\equiv\theta_{1}^{0}+Ae^{-iq_{\mu}x^{\mu}} \\ \theta_{2}=\theta_{2}^{0}+Be^{i(\mathbf{qr}-\omega t)}\equiv\theta_{2}^{0}+Be^{-iq_{\mu}x^{\mu}} \\ \end{array},\[ where \(q_{\mu}=\left(\frac{\omega}{\upsilon},-\mathbf{q}\right)\), \(x^{\mu}=\left(\upsilon t,\mathbf{r}\right)\), equilibrium phases \(\theta_{1,2}^{0}\) satisfy the relation ([\[1.4\]](#1.4){reference-type="ref" reference="1.4"}), so that \(\epsilon\cos\left(\theta_{1}^{0}-\theta_{2}^{0}\right)=-|\epsilon|\). Linearizing Eqs.([\[2.2a\]](#2.2a){reference-type="ref" reference="2.2a"},[\[2.2b\]](#2.2b){reference-type="ref" reference="2.2b"}) by using relations \(\cos\left(\theta_{1}-\theta_{2}\right)\approx-\epsilon/|\epsilon|\), \(\epsilon\sin(\theta_{1}-\theta_{2})\approx-|\epsilon|\left(\theta_{1}-\theta_{2}-\theta_{1}^{0}+\theta_{2}^{0}\right)\) and by omitting the nonlinear terms \(\sin\left(\theta_{1}-\theta_{2}\right)\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}|\Psi_{1,2}|\), we obtain the following linear equations: \]\begin{aligned}
\frac{\hbar^{2}}{4m_{1}}|\Psi_{1}|^{2}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{1} -\frac{\hbar^{2}}{4}|\Psi_{1}||\Psi_{2}|\frac{\eta\epsilon}{|\epsilon|}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{2} +|\Psi_{1}||\Psi_{2}|\epsilon|\left(\theta_{1}-\theta_{2}-\theta_{1}^{0}+\theta_{2}^{0}\right) &=& 0\label{2.4a}\\ \frac{\hbar^{2}}{4m_{2}}|\Psi_{2}|^{2}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{2} -\frac{\hbar^{2}}{4}|\Psi_{1}||\Psi_{2}|\frac{\eta\epsilon}{|\epsilon|}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\theta_{1} -|\Psi_{1}||\Psi_{2}||\epsilon|\left(\theta_{1}-\theta_{2}-\theta_{1}^{0}+\theta_{2}^{0}\right) &=& 0.\label{2.4b} \end{aligned}\[ We can see, that oscillations of the phases \(\theta_{1,2}\) are separated from movement of the modules \(|\Psi_{1,2}|\) in a linear approximation. Substituting the phases ([\[2.3\]](#2.3){reference-type="ref" reference="2.3"}) in Eqs.([\[2.4a\]](#2.4a){reference-type="ref" reference="2.4a"},[\[2.4b\]](#2.4b){reference-type="ref" reference="2.4b"}) we obtain equations for the amplitudes \(A\) and \(B\): \]\label{2.5} \begin{array}{c} A\left(|\epsilon|-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4m_{1}}\frac{|\Psi_{1}|}{|\Psi_{2}|}\right) +B\left(-|\epsilon|+q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\right)=0 \\ A\left(-|\epsilon|+q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\right) +B\left(|\epsilon|-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4m_{2}}\frac{|\Psi_{2}|}{|\Psi_{1}|}\right)=0 \end{array}.\[ Equating to zero the determinant of the system ([\[2.5\]](#2.5){reference-type="ref" reference="2.5"}), we find a dispersion equation: \]\label{2.6} \left(q_{\mu}q^{\mu}\right)^{2}\frac{\hbar^{2}}{4}\left[\frac{1}{m_{1}m_{2}}-\eta^{2}\right] =\left(q_{\mu}q^{\mu}\right)|\epsilon| \left[\frac{m_{1}|\Psi_{2}|^{2}+m_{2}|\Psi_{1}|^{2}}{m_{1}m_{2}|\Psi_{1}||\Psi_{2}|}-2\frac{\eta\epsilon}{|\epsilon|}\right].\[ From where we can see, that one of dispersion relations is \]\label{2.7} q_{\mu}q^{\mu}=0\Rightarrow\omega^{2}=q^{2}\upsilon^{2},\[ wherein \(A=B\), thus this mode is common mode oscillations, as the Goldstone mode in single-band superconductors. There is another oscillation mode with spectrum \]\label{2.8} q_{\mu}q^{\mu}=\frac{4|\epsilon|}{\hbar^{2}}\left[\frac{1}{m_{1}m_{2}}-\eta^{2}\right]^{-1} \left[\frac{m_{1}|\Psi_{2}|^{2}+m_{2}|\Psi_{1}|^{2}}{m_{1}m_{2}|\Psi_{1}||\Psi_{2}|}-2\frac{\eta\epsilon}{|\epsilon|}\right],\[ wherein \]\label{2.9} \frac{A}{B}=-\frac{m_{1}}{m_{2}}\frac{|\Psi_{2}|^{2}}{|\Psi_{1}|^{2}}\frac{1-m_{2}\frac{\eta\epsilon}{|\epsilon|}\frac{|\Psi_{1}|}{|\Psi_{2}|}} {1-m_{1}\frac{\eta\epsilon}{|\epsilon|}\frac{|\Psi_{2}|}{|\Psi_{1}|}},\[ at that for symmetrical bands \(m_{1}=m_{2},|\Psi_{1}|=|\Psi_{2}|\) we have \(\frac{A}{B}=-1\). If we suppose the drag effect is absent: \(\eta=0\), then \]\label{2.10} (\hbar\omega)^{2}=4|\epsilon|\frac{|\Psi_{1}|^{2}m_{2}+|\Psi_{2}|^{2}m_{1}}{|\Psi_{1}||\Psi_{2}|}\upsilon^{2}+(\hbar q)^{2}\upsilon^{2},\[ wherein \]\label{2.11} \frac{A}{B}=-\frac{m_{1}}{m_{2}}\frac{|\Psi_{2}|^{2}}{|\Psi_{1}|^{2}}.\[ For symmetrical bands \(m_{1}=m_{2}\equiv m\) and \(|\Psi_{1}|=|\Psi_{2}|\) we obtain \]\label{2.12} (\hbar\omega)^{2}=8|\epsilon|m\upsilon^{2}+(\hbar q)^{2}\upsilon^{2},\quad A=-B,\[ that corresponds to results of Ref.. Thus, in two-band superconductors the Goldstone mode splits into two branches: common mode oscillations, where \(\nabla\theta_{1}=\nabla\theta_{2}\), with acoustic spectrum-Eq.([\[2.7\]](#2.7){reference-type="ref" reference="2.7"}), and the oscillations of the relative phase \(\theta_{1}-\theta_{2}\) between two SC condensates (for symmetrical condensates we have \(\nabla\theta_{1}=-\nabla\theta_{2}\)) with the energy gap in spectrum determined by the interband coupling-Eqs.([\[2.8\]](#2.8){reference-type="ref" reference="2.8"},[\[2.10\]](#2.10){reference-type="ref" reference="2.10"},[\[2.12\]](#2.12){reference-type="ref" reference="2.12"}), which can be identified as a Leggett mode. It should be noted, as we can see from Eqs.([\[2.8\]](#2.8){reference-type="ref" reference="2.8"},[\[2.10\]](#2.10){reference-type="ref" reference="2.10"}), \(\omega\) is not proportional to \(|\Psi_{1}||\Psi_{2}|\propto|\Delta_{1}||\Delta_{2}|\). Moreover, \(q_{\mu}q^{\mu}(T=T_{c})\propto |\epsilon|m\upsilon^{2}\neq 0\). It should be noted, that at \(T>T_{c}\) the Goldstone excitations make no sense (as the light speed \(\upsilon\) and the permittivity \(c^{2}/\upsilon^{2}\) ), so far as the phase \(\theta\) of the OP makes no sense, since \(\langle\Psi\rangle=0\), although \(\langle|\Psi|^{2}\rangle\neq 0\) occurs due to fluctuations. *If we suppose \(\eta^{2}=\frac{1}{m_{1}m_{2}},\quad\eta\epsilon<0\), then from* Eq.([\[2.6\]](#2.6){reference-type="ref" reference="2.6"}) *we can see, that the Leggett mode is absent, and the common mode oscillations with spectrum* ([\[2.7\]](#2.7){reference-type="ref" reference="2.7"}) *remains only*. In a two band superconductor a current (flow) takes the following form : \]\begin{aligned}
\label{2.13} \mathbf{j}&=&e\hbar\left[\frac{|\Psi_{1}|^{2}}{m_{1}}\nabla\theta_{1}+\eta|\Psi_{1}||\Psi_{2}| \left(\nabla\theta_{1}+\nabla\theta_{2}\right)\cos(\theta_{1}-\theta_{2})+\frac{|\Psi_{2}|^{2}}{m_{2}}\nabla\theta_{2}\right]\nonumber\\ &=&ie^{i(\mathbf{qr}-\omega t)}e\hbar\left[\frac{|\Psi_{1}|^{2}}{m_{1}}\left(1-m_{1}\frac{\eta\epsilon}{|\epsilon|} \frac{|\Psi_{2}|}{|\Psi_{1}|}\right)A+\frac{|\Psi_{2}|^{2}}{m_{2}}\left(1-m_{2}\frac{\eta\epsilon}{|\epsilon|} \frac{|\Psi_{1}|}{|\Psi_{2}|}\right)B\right]\mathbf{q}, \end{aligned}\[ from where we can see, that the Goldstone mode ([\[2.7\]](#2.7){reference-type="ref" reference="2.7"}) (where \(A=B\)) is accompanied by the current, therefore the gauge field \(\widetilde{A}_{\mu}\) absorbs the Goldstone bosons \(\theta_{1,2}\), as in single-band superconductors, i.e. the Anderson-Higgs mechanism takes place. *For the Leggett mode (where \(A/B\) is determined with Eq.([\[2.9\]](#2.9){reference-type="ref" reference="2.9"})) we obtain \(\mathbf{j}=0\), therefore such oscillations \"survive\" and can be observed.*
## Higgs oscillations
In the previous subsection we could see, that oscillations of the phases \(\theta_{1,2}\) are separated from movement of the modules \(|\Psi_{1,2}|\) in the linear approximation. Therefore, let us consider only movement of the modules (that is, assuming \(\theta_{1}=\theta_{1}^{0}\) and \(\theta_{2}=\theta_{2}^{0}\)), then Lagrangian ([\[1.9\]](#1.9){reference-type="ref" reference="1.9"}) takes the form: \]\begin{aligned}
\label{3.1} \mathcal{L}&=&\frac{\hbar^{2}}{4m_{1}}\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}|\Psi_{1}| +\frac{\hbar^{2}}{4m_{2}}\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}|\Psi_{2}| -\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\left(\widetilde{\partial}_{\mu}|\Psi_{1}|\widetilde{\partial}^{\mu}|\Psi_{2}| +\widetilde{\partial}_{\mu}|\Psi_{2}|\widetilde{\partial}^{\mu}|\Psi_{1}|\right)\nonumber\\ &-&a_{1}\left|\Psi_{1}\right|^{2}-\frac{b_{1}}{2}\left|\Psi_{1}\right|^{4}-a_{2}\left|\Psi_{2}\right|^{2}-\frac{b_{2}}{2}\left|\Psi_{2}\right|^{4} +2|\Psi_{1}||\Psi_{2}||\epsilon|. \end{aligned}\[ At \(T<T_{c}\) we can consider small variations of modules of OP from their equilibrium values: \(|\Psi_{1,2}|=\Psi_{01,02}+\phi_{1,2}\), where \(|\phi_{1,2}|\ll\Psi_{01,02}\). Then, \(|\Psi|^{2}\approx\Psi_{0}^{2}+2\Psi_{0}\phi+\phi^{2}\), \(|\Psi|^{4}\approx\Psi_{0}^{4}+4\Psi_{0}^{3}\phi+6\Psi_{0}^{2}\phi^{2}\), \(|\Psi_{1}||\Psi_{2}|\approx\Psi_{01}\Psi_{02}+\Psi_{01}\phi_{2}+\Psi_{02}\phi_{1}+\phi_{1}\phi_{2}\), and Lagrangian ([\[3.1\]](#3.1){reference-type="ref" reference="3.1"}) takes the form: \]\begin{aligned}
\label{3.2} \mathcal{L}&=&\frac{\hbar^{2}}{4m_{1}}\widetilde{\partial}_{\mu}\phi_{1}\widetilde{\partial}^{\mu}\phi_{1} +\frac{\hbar^{2}}{4m_{2}}\widetilde{\partial}_{\mu}\phi_{2}\widetilde{\partial}^{\mu}\phi_{2} -\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\left(\widetilde{\partial}_{\mu}\phi_{1}\widetilde{\partial}^{\mu}\phi_{2} +\widetilde{\partial}_{\mu}\phi_{2}\widetilde{\partial}^{\mu}\phi_{1}\right)\nonumber\\ &-&\phi_{1}^{2}\left(a_{1}+3b_{1}\Psi_{01}^{2}\right)-\phi_{2}^{2}\left(a_{2}+3b_{2}\Psi_{02}^{2}\right)+2|\epsilon|\phi_{1}\phi_{2}\nonumber\\ &+&2\phi_{1}\left(|\epsilon|\Psi_{02}-a_{1}\Psi_{01}-b_{1}\Psi_{01}^{3}\right) +2\phi_{2}\left(|\epsilon|\Psi_{01}-a_{2}\Psi_{02}-b_{2}\Psi_{02}^{3}\right)\nonumber\\ &-&a_{1}\Psi_{01}^{2}-\frac{b_{1}}{2}\Psi_{01}^{4}-a_{2}\Psi_{02}^{2}-\frac{b_{2}}{2}\Psi_{02}^{4}+2\Psi_{01}\Psi_{02}|\epsilon|. \end{aligned}\[ The last five terms can be omitted as a constant. The terms at \(\phi_{1}\) and \(\phi_{2}\) should be zero: \]\label{3.3} \left\{\begin{array}{c} a_{1}\Psi_{01}-|\epsilon|\Psi_{02}+b_{1}\Psi_{01}^{3}=0 \\ a_{2}\Psi_{02}-|\epsilon|\Psi_{01}+b_{2}\Psi_{02}^{3}=0 \\ \end{array}\right\},\[ that corresponds to Eq.([\[1.2\]](#1.2){reference-type="ref" reference="1.2"}). At \(T>T_{c1},T_{c2}\) we have \(a_{1,2}>0\) and \(\epsilon^{2}-a_{1}(T_{c})a_{2}(T_{c})=0\), at \(T<T_{c1},T_{c2}\) we have \(a_{1,2}<0\). For the case of *the weak interband coupling* \(\epsilon^{2}\ll a_{1}a_{2}\), at \(T\ll T_{c1},T_{c2}\) it is not difficult to obtain from Eq.([\[3.3\]](#3.3){reference-type="ref" reference="3.3"}): \]\label{3.4a} \begin{array}{c} \Psi_{01}=\sqrt{\frac{|a_{1}|}{b_{1}}} \left(1+\frac{|\epsilon|}{\sqrt{|a_{1}||a_{2}|}}\sqrt{\frac{b_{1}}{b_{2}}}\frac{|a_{2}|}{|a_{1}|}\right)\approx\sqrt{\frac{|a_{1}|}{b_{1}}}\\ \Psi_{02}=\sqrt{\frac{|a_{2}|}{b_{2}}} \left(1+\frac{|\epsilon|}{\sqrt{|a_{2}||a_{1}|}}\sqrt{\frac{b_{2}}{b_{1}}}\frac{|a_{1}|}{|a_{2}|}\right)\approx\sqrt{\frac{|a_{2}|}{b_{2}}}\\ \end{array}.\[ That is, the effect of the weak interband coupling on the both OP \(\Psi_{1,2}\) at \(T=0\) is not significant, and it can be described as perturbation. Let, for example, \(T_{c1}\ll T_{c2}\), then from Eq.([\[3.3\]](#3.3){reference-type="ref" reference="3.3"}) we can obtain: \]\label{3.4e} \begin{array}{ccc} \Psi_{01}^{3}=|\epsilon|\sqrt{\frac{|a_{2}|}{b_{2}b_{1}^{2}}}& \mathrm{at} & T=T_{c1}\\ \Psi_{02}^{2}=\frac{\epsilon^{2}}{a_{1}b_{2}}& \mathrm{at} & T=T_{c2}\\ \end{array}.\[ For symmetrical bands, i.e. \(a_{1}=a_{2}\), \(b_{1}=b_{2}\equiv b\), \(\Psi_{01}=\Psi_{02}\equiv\Psi_{0}\), at \(T=T_{c1}=T_{c2}\) we have \]\label{3.4d} \Psi_{0}^{2}=\frac{|\epsilon|}{b}.\[ At \(T\rightarrow T_{c}\) we have \(\Psi_{01,02}\rightarrow 0\), then it is not difficult to obtain from Eq.([\[3.3\]](#3.3){reference-type="ref" reference="3.3"}): \]\label{3.4f} \Psi_{01}^{2}=\frac{\epsilon^{2}\left(\epsilon^{2}-a_{1}a_{2}\right)}{\epsilon^{2}a_{1}b_{2}+b_{1}a_{2}^{3}},\quad \Psi_{02}^{2}=\frac{\epsilon^{2}\left(\epsilon^{2}-a_{1}a_{2}\right)}{\epsilon^{2}a_{2}b_{1}+b_{2}a_{1}^{3}}.\[ Thus, at high temperatures \(T\gtrsim T_{c1},T_{c2}\), the values of \(\Psi_{01,02}\) are determined by the interband coupling \(\epsilon\), so that, if \(\epsilon=0\), then \(\Psi_{01,02}=0\). Let us introduce the following notes: \]\label{3.4b} \alpha_{1}\equiv a_{1}+3b_{1}\Psi_{01}^{2}, \quad \alpha_{2}\equiv a_{2}+3b_{2}\Psi_{02}^{2},\[ then, \]\label{3.4c} \begin{array}{ccc} \alpha_{1,2}= a_{1,2}>0 & \mathrm{at} & T=T_{c} \\ \alpha_{1,2}=-2a_{1,2}=2|a_{1,2}| & \mathrm{at} & T\ll T_{c1},T_{c2}\\ \end{array}.\[ The second formula is correct only if the weak interband coupling \(\epsilon^{2}\ll a_{1}a_{2}\) takes place. The Lagrange equations for the Lagrangian ([\[3.2\]](#3.2){reference-type="ref" reference="3.2"}) are: \]\begin{aligned}
\frac{\hbar^{2}}{4m_{1}}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\phi_{1} -\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\phi_{2} +\alpha_{1}\phi_{1}-|\epsilon|\phi_{2} &=& 0\label{3.5a}\\ \frac{\hbar^{2}}{4m_{2}}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\phi_{2} -\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\widetilde{\partial}_{\mu}\widetilde{\partial}^{\mu}\phi_{1} +\alpha_{2}\phi_{2}-|\epsilon|\phi_{1} &=& 0.\label{3.5b} \end{aligned}\[ The fields \(\phi_{1,2}\) can be written in the form of harmonic oscillations: \(\phi_{1}=Ae^{-iq_{\mu}x^{\mu}}\), \(\phi_{2}=Be^{-iq_{\mu}x^{\mu}}\), where \(q_{\mu}x^{\mu}=\omega t-\mathbf{qr}\). Substituting them in Eqs.([\[3.5a\]](#3.5a){reference-type="ref" reference="3.5a"},[\[3.5b\]](#3.5b){reference-type="ref" reference="3.5b"}) we obtain equations for the amplitudes \(A\) and \(B\): \]\begin{aligned}
\label{3.6} \begin{array}{c} A\left(\alpha_{1}-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4m_{1}}\right) +B\left(-|\epsilon|+q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\right)=0 \\ A\left(-|\epsilon|+q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}\right) +B\left(\alpha_{2}-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4m_{2}}\right)=0 \end{array}. \end{aligned}\[ Equating to zero the determinant of the system ([\[3.6\]](#3.6){reference-type="ref" reference="3.6"}), we find a dispersion equation: \]\label{3.7} \left(q_{\mu}q^{\mu}\right)^{2}\frac{\hbar^{4}}{16}\left[\frac{1}{m_{1}m_{2}}-\eta^{2}\right] -\left(q_{\mu}q^{\mu}\right)\frac{\hbar^{2}}{4} \left[\frac{\alpha_{1}}{m_{2}}+\frac{\alpha_{2}}{m_{1}}-2\eta\epsilon\right]+\alpha_{1}\alpha_{2}-\epsilon^{2}=0.\[ From where we obtain the following dispersion relations: \]\label{3.8} q_{\mu}q^{\mu}=\frac{2}{\hbar^{2}}\left[\frac{1}{m_{1}m_{2}}-\eta^{2}\right]^{-1} \left[\left(\frac{\alpha_{1}}{m_{2}}+\frac{\alpha_{2}}{m_{1}}-2\eta\epsilon\right)\pm\sqrt{\mathfrak{D}}\right],\[ where \]\label{3.9} \mathfrak{D}=\left(\frac{\alpha_{1}}{m_{2}}+\frac{\alpha_{2}}{m_{1}}-2\eta\epsilon\right)^{2} -4\left[\frac{1}{m_{1}m_{2}}-\eta^{2}\right]\left(\alpha_{1}\alpha_{2}-\epsilon^{2}\right).\[ Thus, in two-band superconductors the Higgs mode splits into two branches. Let us consider each mode at \(T=T_{c}\), that is \(a_{1}a_{2}=\epsilon^{2}\) (however, we must apply for \(\alpha_{1}\alpha_{2}\) more accurate expressions than Eq.([\[3.4c\]](#3.4c){reference-type="ref" reference="3.4c"}) using Eq.([\[3.4f\]](#3.4f){reference-type="ref" reference="3.4f"},[\[3.4b\]](#3.4b){reference-type="ref" reference="3.4b"}), so that \(\alpha_{1}\alpha_{2}\approx a_{1}a_{2}+3a_{1}b_{2}\Psi_{02}^{2}+3a_{2}b_{1}\Psi_{01}^{2}\)): \]\begin{aligned}
q_{\mu}q^{\mu} &=& \frac{4}{\hbar^{2}}\frac{f(T_{c})(\epsilon^{2}-a_{1}a_{2})}{\frac{a_{1}}{m_{2}}+\frac{a_{2}}{m_{1}}-2\eta\epsilon}=0 \label{3.10a}\\ q_{\mu}q^{\mu} &=& \frac{4}{\hbar^{2}}\frac{\frac{a_{1}}{m_{2}}+\frac{a_{2}}{m_{1}}-2\eta\epsilon}{\frac{1}{m_{1}m_{2}}-\eta^{2}}\neq 0, \label{3.10b} \end{aligned}\[ where \(f(T_{c})\) is some finite dimensionless value. Obviously, the coefficient \(\eta\) must be such, that \(\frac{a_{1}}{m_{2}}+\frac{a_{2}}{m_{1}}-2\eta\epsilon>0\) and \(\eta^{2}<\frac{1}{m_{1}m_{2}}\). We can see, that for the first mode ([\[3.10a\]](#3.10a){reference-type="ref" reference="3.10a"}) the energy gap (the mass of a Higgs boson) vanishes at the critical temperature, as in single-band superconductors. At the same time, the energy gap of the second mode ([\[3.10b\]](#3.10b){reference-type="ref" reference="3.10b"}) does not vanish at the critical temperature. So, in a case of symmetrical bands \(m_{1}=m_{2}\equiv m\), \(a_{1}=a_{2}\equiv a\) at \(T=T_{c}\) (then \(a(T_{c})=|\epsilon|\)) and supposing, that the drag effect is absent \(\eta=0\) we obtain: \]\label{3.10c} (\hbar\omega)^{2}=8|\epsilon|m\upsilon^{2}+(\hbar q)^{2}\upsilon^{2},\[ which coincides with energy of the Leggett mode ([\[2.12\]](#2.12){reference-type="ref" reference="2.12"}). If \(\eta^{2}=\frac{1}{m_{1}m_{2}},\quad\eta\epsilon<0\), then only the single mode takes place: \]\label{3.11} q_{\mu}q^{\mu}=\frac{4}{\hbar^{2}} \frac{\alpha_{1}\alpha_{2}-\epsilon^{2}}{\frac{\alpha_{1}}{m_{2}}+\frac{\alpha_{2}}{m_{1}}+\frac{2|\epsilon|}{\sqrt{m_{1}m_{2}}}},\[ whose energy gap vanishes at \(T=T_{c}\), as for the mode ([\[3.10a\]](#3.10a){reference-type="ref" reference="3.10a"}). In Ref. it has been demonstrated how the energy gap \(\hbar\omega_{0}\) is related to coherence length \(\begin{eqnarray} \end{eqnarray} We can see, that the first coherence length diverges at\)T=T\_c\(. On the contrary, the second length remains finite and it varies little with temperature. These length scales are not related to the concrete bands involved in the formation of the superconducting ordering in a system with interband interaction. This result corresponds to the results in Ref. obtained with a microscopic approach, however they suggest, that the intergradient interaction is absent (i.e.\]), and corresponds to the results in Ref. obtained with a phenomenological approach. Let us consider the ratio of amplitudes for both branches of the spectrum. From Eq.(\ref{3.6}) we have: \begin{equation}\label{3.12} \frac{A}{B}=\frac{|\epsilon|-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}} {\alpha_{1}-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4m_{1}}}\equiv \frac{\alpha_{2}-q_{\mu}q^{\mu}\frac{\hbar^{2}} {4m_{2}}}{|\epsilon|-q_{\mu}q^{\mu}\frac{\hbar^{2}}{4}\frac{\eta\epsilon}{|\epsilon|}}. \end{equation} At\)TT\_c\(, using Eqs.(\ref{3.10a},\ref{3.10b}), we obtain: \begin{eqnarray} q_{\mu}q^{\mu}\rightarrow 0 &\Rightarrow& \frac{A}{B}=\frac{|\epsilon|}{a_{1}} =\frac{a_{2}}{|\epsilon|}=\sqrt{\frac{a_{2}}{a_{1}}}>0\label{3.13a}\\ q_{\mu}q^{\mu}\rightarrow \frac{4}{\hbar^{2}}\frac{\frac{a_{1}}{m_{2}}+\frac{a_{2}}{m_{1}}}{\frac{1}{m_{1}m_{2}}} &\Rightarrow& \frac{A}{B}=-\frac{|\epsilon|}{a_{2}}=-\frac{a_{1}}{|\epsilon|}=-\sqrt{\frac{a_{1}}{a_{2}}}<0,\label{3.13b} \end{eqnarray} where in Eq.(\ref{3.13b}) we suppose\[for simplicity. Thus, for the mode (\ref{3.10a}) (or for the mode (\ref{3.11}), i.e. when\)\^2=,\<0\(), oscillations of\)\|\_1\|\(and\)\|\_2\|\(occur in phase. For the mode (\ref{3.10b}), oscillations of\)\|\_1\|\(and\)\|\_2\|\(occur in anti-phase. The reason for the difference in energies between the common mode and anti-phase Higgs oscillations is as follows. Let us rewrite Eq.(\ref{3.3}) for equilibrium values\)\_01\(and\)\_02\(in the form: \begin{equation}\label{3.13c} \left\{\begin{array}{c} \Psi_{02}=\Psi_{01}(a_{1}+b_{1}\Psi_{01}^{2})/|\epsilon| \\ \Psi_{01}=\Psi_{02}(a_{2}+b_{2}\Psi_{02}^{2})/|\epsilon| \\ \end{array}\right\}. \end{equation} From the first equation we can see, that increasing (decreasing) of\)\_01\(results in increasing (decreasing) of\)\_02\(, then from the second equation we can see, that the increased (decreased)\)\_02\(causes increasing (decreasing) of\)\_01\(. This increasing (decreasing) of\)\_01\(causes the increasing (decreasing) of\)\_02\(again, and so on. Thus, Eqs.(\ref{3.3}) describe a system with the positive feedback. Therefore, excitation of the common mode oscillations requires less energy, than if the condensates from different bands were independent. On the contrary, to excite the anti-phase oscillations (\)\_01\(increases, but\)\_02\(decreases and vice versa) we must do the work against the positive feedback, hence minimal energy of these oscillations is larger. Thus, Higgs modes are oscillations of SC densities\)n\_i=2\|\_i\|\^2\(, at the same time, the normal density must oscillate in anti-phase so, that\)n=n\_+n\_=\(and\)n\_\_s+n\_\_n=0\(. To change SC density and, hence, the normal density, as least one Cooper pair must be broken, that is the energy of order of\)`<!-- -->`{=html}2\|\|\(must be spent. So, in Ref. it has been demonstrated, that in single-band superconductors\)q_q\^=4\|\|\^2\(. \emph{Thus, excitation of any Higgs mode at \)T=T_{c}\( does not require the energy consumption (for \)\mathbf{q}=0\(), since \)|\Delta|=0\( (i.e. a Cooper pair has zero binding energy). However, we could see, that for the second branch}-Eqs.(\ref{3.10b},\ref{3.10c}) \emph{we have \)q_{\mu}q^{\mu}(T_{c})\neq 0\(, which is a nonphysical property}. It should be noted, that this argument is not correct in the case of superconductors with a pseudogap, where uncorrelated pairs can be formed at\)T\>T\_c\(due to strong electron–electron interaction, but the phase coherence is possible only at\)T\<T\_c\(. Thus, we must suppose \begin{equation}\label{3.14} \eta^{2}=\frac{1}{m_{1}m_{2}},\quad\eta\epsilon<0, \end{equation} then from Eq.(\ref{3.7}) we can see, that the anti-phase Higgs mode is absent, and the common mode oscillations with zero energy gap at\)T=T\_c\((\ref{3.11}) remain only. Analogously, from Eq.(\ref{2.6}) we can see, that the Leggett mode is absent, and the common mode oscillations with gapless spectrum (\ref{2.7}) only remains. \emph{Thus, in two-band superconductors one Goldstone mode and one Higgs mode exist only, as in single-band superconductors, and the Leggett mode and the anti-phase Higgs mode are absent}. In the same time, the Goldstone mode is accompanied by current-Eq.(\ref{2.13}), therefore the gauge field\)\_\(absorbs the Goldstone boson\], as in single-band superconductors, i.e. the Anderson-Higgs mechanism takes place. In addition, the coherence length (\ref{3.11a}) only remains, that prohibits type 1.5 superconductors and corresponds to the result of Ref.. Let us consider regime of almost independent condensates in each bands. This means: 1) the temperature must be low, i.e.\)TT\_c1,T\_c2\(, 2) the weak interband coupling\)\^2a\_1a\_2\(must take place. Using Eqs.(\ref{3.4a},\ref{3.4c}), the energy gap\)\_0\((\)=0\() of Higgs mode (\ref{3.11}) can be presented in the form: \begin{equation}\label{3.15} (\hbar\omega_{0})^{2}=4\upsilon^{2} \frac{\alpha_{1}\alpha_{2}-\epsilon^{2}}{\frac{\alpha_{1}}{m_{2}}+\frac{\alpha_{2}}{m_{1}}+\frac{2|\epsilon|}{\sqrt{m_{1}m_{2}}}} \approx 4\upsilon^{2}\frac{4\sqrt{|a_{1}||a_{2}|b_{1}b_{2}}}{\frac{2|a_{1}|}{m_{2}}+\frac{2|a_{2}|}{m_{1}}}\Psi_{01}\Psi_{02}. \end{equation} Then the multiplier before\)\_01\_02\(depends on temperature very weakly, and this energy is symmetrical with respect to the bands. Using relationship between the "wave function" of Cooper pairs\[and the energy gap\], which can be generalized for two-band superconductors in the form: \begin{equation}\label{3.16} \Psi_{1}=\frac{\left(14\zeta(3)n_{1}\right)^{1/2}}{4\pi T_{c1}}\Delta_{1},\quad\Psi_{2}=\frac{\left(14\zeta(3)n_{2}\right)^{1/2}}{4\pi T_{c2}}\Delta_{2}, \end{equation} where\)n\_1,2=\(are electron densities for each band. Then, we can see, that\)(\_0)\^2\|\_1\|\|\_2\|\(, and we can suppose: \begin{equation}\label{3.17} (\hbar\omega_{0})^{2}=\chi\Delta_{01}\Delta_{02}, \end{equation} where\)=\((dimensionless) is such, that in superconductor with symmetrical\)m\_1=m\_2\(,\)n\_1=n\_2\(,\)a\_1=a\_2\(,\)b\_1=b\_2\(,\)T\_c1=T\_c2\[\]\_1=\_2\(and almost independent bands (i.e.\)\^2a\_1a\_2\(at\)TT\_c1,T\_c2\(), we should have\)=\(, since in single-band superconductors we have\)=\(and we can determine the "dielectric permittivity" as\)==\(, then the "mixture" of two superconductors is equivalent to two parallel dielectrics (capacitors) with total permittivity\)=\_1+\_2=\(; hence, we obtain for the "mixture":\)=\(. The coefficients\)a\_1,2\(,\)b\_1,2\(are : \begin{equation}\label{3.18} a_{1,2}=\frac{6\pi^{2}T_{c1,2}}{7\zeta(3)\varepsilon_{F1,2}}\left(T-T_{c1,2}\right),\quad b_{1,2}=\frac{6\pi^{2}T_{c1,2}}{7\zeta(3)\varepsilon_{F1,2}}\frac{T_{c1,2}}{n_{1,2}}. \end{equation} Substituting Eqs.(\ref{3.15},\ref{3.16},\ref{3.18}) in Eq.(\ref{3.17}) we obtain: \begin{equation}\label{3.19} \upsilon^{2}=\frac{\chi}{12}\frac{\upsilon_{F1}^{2}T_{c2}(T_{c2}-T)+\upsilon_{F2}^{2}T_{c1}(T_{c1}-T)} {\sqrt{T_{c1}T_{c2}}\sqrt{(T_{c1}-T)(T_{c2}-T)}}. \end{equation} If we consider symmetrical bands, i.e.\)\_F1=\_F2\(and\)T\_c1=T\_c2\(, then we obtain\)\^2=\_F\^2\(. It is not difficult to notice, that the dependence of\[on\)T\(in Eq.(\ref{3.19}) is weak (at\)TT\_c1,T\_c2\(), hence we can suppose: \begin{equation}\label{3.20} \upsilon^{2}\approx\frac{1}{12}\left(\frac{T_{c2}}{T_{c1}}\upsilon_{F1}^{2}+\frac{T_{c1}}{T_{c2}}\upsilon_{F2}^{2}\right). \end{equation} Strictly speaking, the relation\]is not valid at low temperatures\)TT\_c\(both in pure superconductors (where\)n\_s(0)=n\() and in dirty superconductors (where\)n\_s(0)\|\|\() (here,\)n\_s=2\|\|\^2\(is density of SC electrons,\)n\(is total electron density). In the same time, at\)TT\_c1,T\_c2\(the condensates cannot be considered as independent, hence we cannot find the speed\[using the above consideration. However, we can extrapolate the relation (\ref{3.17}) to all temperatures for dimensional reasons, where the dimensionless coefficient\]should be considered as an adjustable parameter. So, in general case, it is obvious to assume\)\~\_F1,\_F2\(so as to\)(\_0)\^2=\_01\_02\((i.e.\[). In Ref. the measurements of the excitation energy squared\)\_0\^2\(vs. the product of the gaps\)\_\_\(have been done for polycrystalline samples\)\_1-x\_x\_2\(and\)\_2\(with different\)T\_c\(at the same temperature. The results of the measurements are shown in Fig.(\ref{Fig2}). We can see, that the dependence\)\_0\^2=\_\_\(takes place for both substances, that corresponds to our theoretical result\)(\_0)\^2=\_\_\(for energy of the Higgs mode-Eq.(\ref{3.17}) with\]. If we interpret this result as the Leggett mode, then in a formula from Ref.:\)\_0\^2=4\_\_\(, where\)\_ij\(is dimensionless interband and intraband coupling constants, we must suppose\)(\_12+\_21)/(\_11\_22-\_12\_21)=1/4\(for different materials with different\)T\_c\(, hence with various\)\_ij\(, that is very unlikely. \section{Effect of Higgs oscillations on Josephson current}\label{Josephson} In Ref. a resonant enhancement of the DC current through a Josephson junction (JJ) at a bias voltage\)V\_\(has been observed when the Josephson frequency\)\_J=V\(or its harmonics\)(m\_J)\(match the energy of some internal oscillation mode of two-band superconductor\)\_0\(or its harmonics\)n\_0\(: \begin{equation}\label{4.1} V_{\mathrm{res}}=\frac{n}{m}\frac{\hbar\omega_{0}}{2e}, \end{equation} where\)n\(and\)m\(are integer numbers. For\)\_2\(the value of the energy is observed as\)\_0\(and\)n/m=3/2,1/1,1/2,1/3\(. This phenomenon has been observed as the dips in the\)I/V\(characteristics at the voltages\)V\_\(. In general, peculiarities of the same type could also appear due to interaction of the AC Josephson current with phonons or electromagnetic waves. Nevertheless, the authors in Ref. believe, that the peculiarities observed in their investigation are related namely to the Leggett collective excitations, based on a work, where it has been shown, that the Leggett collective mode resonantly couples to the AC Josephson current of a junction between a two-band superconductor and a single-band superconductor, so that, when the voltage matches the energy of the Leggett mode, the resonant enhancement of the DC current takes place. In their experimental samples, there are no optical phonons with energy as low as\)`<!-- -->`{=html}4\(in\)\_2\(. The effective interaction between Josephson current and low-energy acoustic phonons as well as electromagnetic waves can only exist in the presence of a resonator system inside the junction. Then, the observed subgap structure could appear at voltages matching the energies of resonator eigenmodes. It is very unlikely, that all their break-junctions demonstrating the discussed subgap structure possess identical resonator systems. Thus, the resonant enhancement of the DC current takes place if an additional coherence is introduced into the JJ with AC Josephson current (resonance with phonons, el.-mag. waves, the Leggett mode etc.). In Sect.\ref{Goldstone and Higgs} we could see, that, as in single-band superconductors, in two-band superconductors one Goldstone mode (unobservable due to absorbtion into el.-mag. field) and only one Higgs mode exists, at the same time, the Leggett mode is prohibited. Let us consider a S–I–S junction, where the right and left banks are the same isotropic s-wave two-band superconductors with the energy gaps\)\_1,2\(. We assume the capacity\)C\(of the JJ is small, so that a McCumber parameter is\)=(2e/)I\_mCR\^2\(, that is displacement current is negligible\)I\_D=CI\_m\(, and I-V characteristic is\)V=R\(. Calculation of the Josephson current between two-band superconductors with the strong impurity intraband scattering rates (the dirty limit described with Usadel equations) and the weak interband scattering in Ref. leads to: \begin{eqnarray}\label{4.2} I=\frac{\pi|\Delta_{1}|}{2eR_{N1}}\cos\frac{\theta}{2}\mathrm{Arctanh}\sin{\frac{\theta}{2}}+ \frac{\pi|\Delta_{2}|}{2eR_{N2}}\cos\frac{\theta}{2}\mathrm{Arctanh}\sin{\frac{\theta}{2}}. \end{eqnarray} Here,\)R\_N1\(and\)R\_N2\(are the contributions of the normal resistance for each band,\)\_1\^R-\_1\^L=\_2\^R-\_2\^L\(is the phase difference between the banks R and L. The interband scattering is neglected in the present approximation, so that the anharmonicity in Eq.(\ref{4.2}) is caused with the strong impurity intraband scattering; accordingly, in the pure limit we will have the usual harmonic expression\)J=+\(. Using Eqs.(\ref{3.16}), Eq.(\ref{4.2}) can be rewritten as follows: \begin{eqnarray}\label{4.3} I=\frac{2\pi^{2}T_{c1}|\Psi_{1}|}{eR_{N1}\sqrt{14\zeta(3)n_{1}}}\cos\frac{\theta}{2}\mathrm{Arctanh}\sin{\frac{\theta}{2}}+ \frac{2\pi^{2}T_{c2}|\Psi_{2}|}{eR_{N2}\sqrt{14\zeta(3)n_{2}}}\cos\frac{\theta}{2}\mathrm{Arctanh}\sin{\frac{\theta}{2}}. \end{eqnarray} Since\[(for maximal current we can suppose\)/2\() we can expand\]in the series as: \begin{eqnarray}\label{4.4} \cos\frac{\theta}{2}\mathrm{Arctanh}\sin{\frac{\theta}{2}}&\approx& \cos\frac{\theta}{2}\left[\sin{\frac{\theta}{2}}+\frac{1}{2}\sin^{3}{\frac{\theta}{2}} +\frac{1}{5}\sin^{5}{\frac{\theta}{2}}+\ldots\right]\nonumber\\ &=&\frac{1}{2}\sin{\theta}-\frac{1}{16}\sin{2\theta}+\frac{1}{8}\sin{\theta}+ \frac{1}{5\cdot 32}\sin{3\theta}-\frac{1}{40}\sin{2\theta}+\frac{1}{32}\sin{\theta}+\ldots \end{eqnarray} Then, \begin{equation}\label{4.5} I\approx I_{m1}\sin{\theta}-I_{m2}\sin{2\theta}+I_{m3}\sin{3\theta}+\ldots, \end{equation} where \begin{equation}\label{4.6} I_{m1}=\frac{4\pi^{2}T_{c1}|\Psi_{1}|}{3eR_{N1}\sqrt{14\zeta(3)n_{1}}} +\frac{4\pi^{2}T_{c2}|\Psi_{2}|}{3eR_{N2}\sqrt{14\zeta(3)n_{2}}}\equiv I_{m1}^{(1)}+I_{m1}^{(2)}, \end{equation} at that\)I\_m3\~0.1I\_m2\~0.01I\_m1\(. Let us consider the problem with the given voltage\)V\(. Then, in order to ensure the gauge invariance, the phase\[must be changed as:\)-Vt-t\(, where\)V\(. Then, \begin{equation}\label{4.7} I=-I_{m1}\sin{(\omega t-\theta)}+I_{m2}\sin{(2\omega t-2\theta)}-I_{m3}\sin{(3\omega t-3\theta)}+\ldots \end{equation} As in Sect.\ref{Goldstone and Higgs}, we consider the small variations of modulus of the OP from its equilibrium value:\)\|\_1,2\|=\_01,02+\_1,2\(, where\)\|\_1,2\|\_01,02\(. Thus, \begin{eqnarray}\label{4.8} I&=&-I_{m1}\sin{(\omega t-\theta)}-\frac{I_{m1}^{(1)}}{\Psi_{01}}\phi_{1}\sin{(\omega t-\theta)} -\frac{I_{m1}^{(2)}}{\Psi_{02}}\phi_{2}\sin{(\omega t-\theta)}\nonumber\\ &+&I_{m2}\sin{(2\omega t-2\theta)}+\frac{I_{m2}^{(1)}}{\Psi_{01}}\phi_{1}\sin{(2\omega t -2\theta)}+\frac{I_{m2}^{(2)}}{\Psi_{02}}\phi_{2}\sin{(2\omega t-2\theta)}\nonumber\\ &-&I_{m3}\sin{(3\omega t-3\theta)}-\frac{I_{m3}^{(1)}}{\Psi_{01}}\phi_{1}\sin{(3\omega t-3\theta)} -\frac{I_{m3}^{(2)}}{\Psi_{02}}\phi_{2}\sin{(3\omega t-3\theta)}+\ldots \end{eqnarray} The AC current with frequencies\),2,3,...\(stipulates oscillations of SC density\)n\_\(in each band, and besides, according to the result of Sect.\ref{Goldstone and Higgs}, the oscillations in each band occur in phase (in the linear approximation): \begin{eqnarray} e\frac{\partial n_{\mathrm{s1}}}{\partial t}&=&2e\frac{\partial|\Psi_{1}|^{2}}{\partial t}=4e\Psi_{01}\frac{\partial\phi_{1}}{\partial t}= -I_{m1}^{(1)}\sin{(\omega t-\theta)}+I_{m2}^{(1)}\sin{(2\omega t-2\theta)}-I_{m3}^{(1)}\sin{(3\omega t-3\theta)}+\ldots\label{4.9a}\\ e\frac{\partial n_{\mathrm{s2}}}{\partial t}&=&2e\frac{\partial|\Psi_{2}|^{2}}{\partial t}=4e\Psi_{02}\frac{\partial\phi_{2}}{\partial t}= -I_{m1}^{(2)}\sin{(\omega t-\theta)}+I_{m2}^{(2)}\sin{(2\omega t-2\theta)}-I_{m3}^{(2)}\sin{(3\omega t-3\theta)}+\ldots\label{4.9b} \end{eqnarray} Then we can write the equations: \begin{eqnarray} \frac{\partial^{2}\phi_{1}}{\partial t^{2}}&=&-\frac{I_{m1}^{(1)}\omega}{4e\Psi_{01}}\cos{(\omega t-\theta)} +\frac{I_{m2}^{(1)}\omega}{2e\Psi_{01}}\cos{(2\omega t-2\theta)}-\frac{3I_{m3}^{(1)}\omega}{4e\Psi_{01}}\cos{(3\omega t-3\theta)} +\ldots\label{4.10a}\\ \frac{\partial^{2}\phi_{2}}{\partial t^{2}}&=&-\frac{I_{m1}^{(2)}\omega}{4e\Psi_{02}}\cos{(\omega t-\theta)} +\frac{I_{m2}^{(2)}\omega}{2e\Psi_{02}}\cos{(2\omega t-2\theta)}-\frac{3I_{m3}^{(2)}\omega}{4e\Psi_{02}}\cos{(3\omega t-3\theta)} +\ldots\label{4.10b} \end{eqnarray} Thus, if there were no eigen oscillations (Higgs oscillations with the frequency\)\_0\(-Eqs.(\ref{3.15},\ref{3.17})), then the densities of SC electrons\)n\_\(and\)n\_\(oscillate with the frequencies of AC Josephson current\),2,3,...\(. On the other hand, the densities\)n\_\(and\)n\_\(can oscillate with the eigen frequency\)\_0\(at some attenuation constant\)\_0\(, and the AC Josephson current plays a role of a driving force. In this system a resonance occurs, if the frequency of AC current\)=2eV/\(coincides with the eigen frequency\)\_0\(, then Higgs oscillations in the banks can be excited. In other words, a Cooper pair can tunnel through the JJ, exciting a quantum of Higgs oscillations. Moreover, the tunneling with excitation of\)n\(quanta with the total energy\)n\_0\(by a Cooper pair due to passing through the energy difference\)`<!-- -->`{=html}2eV\(can occur, that is equivalent to the resonant excitation of an oscillator with an eigen frequency\)n\_0\(by the external driving force with the frequency\)=V\(. Hence, equations for such oscillations take the following form: \begin{eqnarray} \frac{\partial^{2}\phi_{1}}{\partial t^{2}}+2\gamma\frac{\partial\phi_{1}}{\partial t}+[n\omega_{0}]^{2}\phi_{1} &=&-\frac{I_{m1}^{(1)}\omega}{4e\Psi_{01}}\cos{(\omega t-\theta)} +\frac{I_{m2}^{(1)}\omega}{2e\Psi_{01}}\cos{(2\omega t-2\theta)}-\frac{3I_{m3}^{(1)}\omega}{4e\Psi_{01}}\cos{(3\omega t-3\theta)} +\ldots\label{4.11a}\\ \frac{\partial^{2}\phi_{2}}{\partial t^{2}}+2\gamma\frac{\partial\phi_{2}}{\partial t}+[n\omega_{0}]^{2}\phi_{2} &=&-\frac{I_{m1}^{(2)}\omega}{4e\Psi_{02}}\cos{(\omega t-\theta)} +\frac{I_{m2}^{(2)}\omega}{2e\Psi_{02}}\cos{(2\omega t-2\theta)}-\frac{3I_{m3}^{(2)}\omega}{4e\Psi_{02}}\cos{(3\omega t-3\theta)} +\ldots\label{4.11b} \end{eqnarray} The particular solutions of nonhomogeneous differential equations are (the general solutions are attenuating, therefore they can be omitted): \begin{eqnarray}\label{4.12} &&\phi_{1,2}\nonumber\\ &&=\frac{I_{m1}^{(1,2)}\omega \cos\left(\omega t-\theta-\varphi_{1}\right)}{4e\Psi_{01,02}\sqrt{\left(\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+4\gamma^{2}\omega^{2}}} +\frac{I_{m2}^{(1,2)}\omega \cos\left(2\omega t-2\theta-\varphi_{2}\right)}{2e\Psi_{01,02}\sqrt{\left(4\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+16\gamma^{2}\omega^{2}}} +\frac{3I_{m3}^{(1,2)}\omega \cos\left(3\omega t-3\theta-\varphi_{3}\right)}{4e\Psi_{01,02}\sqrt{\left(9\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+36\gamma^{2}\omega^{2}}}\nonumber\\ &&+\ldots, \end{eqnarray} where \begin{equation}\label{4.13} \cos\varphi_{1}=\frac{\omega^{2}-[n\omega_{0}]^{2}}{\sqrt{\left(\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+4\gamma^{2}\omega^{2}}},\quad \cos\varphi_{2}=\frac{[n\omega_{0}]^{2}-4\omega^{2}}{\sqrt{\left(4\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+16\gamma^{2}\omega^{2}}},\quad \cos\varphi_{3}=\frac{9\omega^{2}-[n\omega_{0}]^{2}}{\sqrt{\left(9\omega^{2}-[n\omega_{0}]^{2}\right)^{2}+36\gamma^{2}\omega^{2}}}. \end{equation} We can see, that resonance occurs at frequencies\)=n\_0,n\_0/2,n\_0/3,...\((if\)\_0\(). In the resonant frequencies we have\)\_1,2,3=/2\(. Then the solution (\ref{4.12}) in the resonant frequencies is \begin{eqnarray}\label{4.14} \phi_{1,2}\left(n\omega_{0}\right) &=& \frac{I_{m1}^{(1,2)}}{8e\Psi_{01,02}\gamma}\sin\left(n\omega_{0} t-\theta\right) \nonumber\\ \phi_{1,2}\left(\frac{n\omega_{0}}{2}\right) &=& \frac{I_{m2}^{(1,2)}}{8e\Psi_{01,02}\gamma}\sin\left(n\omega_{0} t-2\theta\right) \\ \phi_{1,2}\left(\frac{n\omega_{0}}{3}\right) &=& \frac{I_{m3}^{(1,2)}}{8e\Psi_{01,02}\gamma}\sin\left(n\omega_{0} t-3\theta\right). \nonumber \end{eqnarray} Out of the resonances the amplitudes of\)\_1,2\(are negligible. Substituting the solutions (\ref{4.14}) in the current (\ref{4.8}) we obtain the currents at the corresponding resonant frequencies: \begin{eqnarray}\label{4.15} I(n\omega_{0})&=&-I_{m1}\sin{(n\omega_{0} t-\theta)}+I_{m2}\sin{(2n\omega_{0} t-2\theta)}-I_{m3}\sin{(3n\omega_{0} t-3\theta)}\nonumber\\ &+&\frac{\left[I_{m1}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}\cos{(2n\omega_{0} t-2\theta)} +\frac{\left[I_{m1}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}\cos{(2n\omega_{0} t-2\theta)} -\frac{\left[I_{m1}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}-\frac{\left[I_{m1}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}+\ldots\nonumber\\ I\left(\frac{n}{2}\omega_{0}\right)&=&-I_{m1}\sin{\left(\frac{n}{2}\omega_{0} t-\theta\right)}+I_{m2}\sin{\left(n\omega_{0} t-2\theta\right)}-I_{m3}\sin{\left(\frac{3n}{2}\omega_{0} t-3\theta\right)}\nonumber\\ &-&\frac{\left[I_{m2}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}\cos{(2n\omega_{0} t-4\theta)} -\frac{\left[I_{m2}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}\cos{(2n\omega_{0} t-4\theta)} +\frac{\left[I_{m2}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}+\frac{\left[I_{m2}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}+\ldots\nonumber\\ I\left(\frac{n}{3}\omega_{0}\right)&=&-I_{m1}\sin{\left(\frac{n}{3}\omega_{0} t-\theta\right)}+I_{m2}\sin{\left(\frac{2n}{3}\omega_{0} t-2\theta\right)}-I_{m3}\sin{\left(n\omega_{0} t-3\theta\right)}\nonumber\\ &+&\frac{\left[I_{m3}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}\cos{(2n\omega_{0} t-6\theta)} +\frac{\left[I_{m3}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}\cos{(2n\omega_{0} t-6\theta)} -\frac{\left[I_{m3}^{(1)}\right]^{2}}{16e\Psi_{01}^{2}\gamma}-\frac{\left[I_{m3}^{(2)}\right]^{2}}{16e\Psi_{02}^{2}\gamma}+\ldots \end{eqnarray} Out of the resonances the current has the form: \begin{equation}\label{4.16} I(\omega)=-I_{m1}\sin{(\omega t-\theta)}+I_{m2}\sin{(2\omega t-2\theta)}-I_{m3}\sin{(3\omega t-3\theta)}+\ldots \end{equation} We can see, that \emph{at the frequencies \)\omega=n\omega_{0},n\omega_{0}/2,n\omega_{0}/3,\ldots\( the DC current occurs, however, unlike DC Josephson current} (\ref{4.2},\ref{4.5})\emph{, it is not function of the phase difference \)\theta\(}. Thus, at the voltage \begin{equation}\label{4.17} V_{\mathrm{res}}=\frac{n}{m}\frac{\hbar\omega_{0}}{2e}, \end{equation} where\)m,n=1,2,3...\(, DC current through the JJ appears. Amplitudes of the resonant peaks is proportional to the square of the critical Josephson currents\)I\_m1\^2,I\_m2\^2,I\_m3\^2,...\(. As we have seen above:\)I\_m3\~0.1I\_m2\~0.01I\_m1\(, hence, DC resonant currents are such, that\)I(\_0)\~10\^-2I(\_0)\~10\^-4I(\_0)\(, as we can see from Eq.(\ref{4.15}). Thus, the main resonance at the voltage\)V\_=\((that is, when\)=\_0\() is the most pronounced. In I-V characteristic of\)\_2\(break junction (S-I-S type) a "step" has been observed in Ref. just at voltage\)V\_=\(-Fig.(\ref{Fig3}b). In the same time, the currents in the subharmonics\)=\_0/2,\_0/3\(are very weak. In the I-V characteristics of a JJ between two-bands superconductors shown in Fig.(\ref{Fig3}a) we should observe the picture, as Shapiro spikes, at given voltage and, as Shapiro steps, at given current. Thus, the resonant enhancement of the DC current through the JJ is result of the coupling between AC Josephson current and Higgs oscillations. It should be noted, that we obtain the result (\ref{4.15}) neglecting of the interband scattering, hence, the effect would have been observed also in single-band superconductors. However, in the single-band superconductors the minimal energy of the Higgs mode is\)\_0=2\|\|\(, that is this mode exists in the free quasiparticle continuum. Hence, Higgs oscillations decay to quasiparticles with energy\]each. In other words, excitation of the Higgs mode is accompanied by the breaking of Cooper pairs with transfer of their constituents in the free quasiparticle states. Thus, the Higgs mode is unstable in single-band superconductors. Therefore, the resonant enhancement is strongly suppressed. On the contrary, as shown in Sect.\ref{Goldstone and Higgs}, in two-band superconductors\)(\_0)\^2=\|\_1\|\|\_2\|\(takes place. If\)\_0\<2(\|\_1\|,\|\_2\|)\(, then the Higgs mode is stable, therefore, the corresponding resonances in AC Josephson effect can be observable. The subharmonics at higher frequencies\)=\_0\>\_0\(can be observed if\)V\_=\<2(\|\_1\|,\|\_2\|)/e\(; at\)V\_(\|\_1\|,\|\_2\|)/e\(the Josephson current\)I(V)\(quickly goes to zero, and the total current is the tunneling of quasiparticles through the contact. Using experimental data of we can verify this criterion, which has a form\)\_0\_\(in this case. From Fig.(\ref{Fig4}) we can see, that the criterion is approximately satisfied. \section{Results}\label{results} In this work we investigate the eigen oscillations of internal degrees of freedom of two-band superconductors using the action (\ref{1.6}) with the corresponding Lorentz-invariant Lagrangian (\ref{1.8}), which are generalization of the GL free energy functional for a two-band isotropic s-wave superconductor on the one hand, and of the ETDGL theory formulated in Ref. on the other hand. Our results are as follows: 1) Due to the internal proximity effect, the Goldstone mode splits into two branches: common mode oscillations with an acoustic spectrum-Eq.(\ref{2.7}), and the oscillations of the relative phase\)\_1-\_2\(between two SC condensates with an energy gap in spectrum determined by the interband coupling\[-Eqs.(\ref{2.10},\ref{2.12}), which can be associated with the Leggett mode. The common mode oscillations are absorbed into the gauge field\)A\_\(due to these oscillations are accompanied by current, as in single-band superconductors. At the same time, Leggett oscillations are not accompanied by the current, therefore they "survive". Higgs oscillations splits into two branches also: a massive one, whose energy gap (mass) vanishes at\)T\_c\(-Eq.(\ref{3.10a}), another massive one, whose mass, determined by the interband coupling\], does not vanish at\)T\_c\(-Eqs.(\ref{3.10b},\ref{3.10c}). For the first mode, oscillations of\)\|\_1\|\(and\)\|\_2\|\(occur in a phase, for the second mode, the oscillations occur in an anti-phase. The mass of the Higgs mode is related to the coherence length\) 2) To excite one quanta of Higgs oscillations, as least one Cooper pair must be broken, i.e. the energy of order of \(2|\Delta|\) must be spent. Thus, excitations of any Higgs mode at \(T=T_{c}\) does not require the energy consumption, since \(|\Delta|=0\). In two-band superconductors for the anti-phase Higgs mode-Eqs.([\[3.10b\]](#3.10b){reference-type="ref" reference="3.10b"},[\[3.10c\]](#3.10c){reference-type="ref" reference="3.10c"}) we have the nonphysical property \(q_{\mu}q^{\mu}(T_{c})\neq 0\). If \(\eta^{2}=\frac{1}{m_{1}m_{2}},\quad\eta\epsilon<0\), then from Eq.([\[3.7\]](#3.7){reference-type="ref" reference="3.7"}) we can see, that the anti-phase Higgs mode is absent, and the common mode oscillations with the zero energy gap at \(T=T_{c}\)-Eq.([\[3.11\]](#3.11){reference-type="ref" reference="3.11"}) remains only. As a consequence, only the coherence length ([\[3.11a\]](#3.11a){reference-type="ref" reference="3.11a"}) remains, that prohibits type 1.5 superconductors. Analogously, from Eq.([\[2.6\]](#2.6){reference-type="ref" reference="2.6"}) we can see, that the Leggett mode is absent, and only the common mode oscillations with the gapless spectrum ([\[2.7\]](#2.7){reference-type="ref" reference="2.7"}) remains. Therefore, we must suppose Eq.([\[3.14\]](#3.14){reference-type="ref" reference="3.14"}), that ensures the correct property of the Higgs mode. *Thus, as in single-band superconductors, in two-band superconductors only one Goldstone mode and one Higgs mode exist, and the Leggett mode and the anti-phase Higgs mode are absent*. However, as mentioned above, the Goldstone mode is accompanied by the current, therefore the gauge field \(\widetilde{A}_{\mu}\) absorbs the Goldstone boson \(\theta\), as in single-band superconductors. 3) The energy gap of the Higgs mode in two-band superconductors-Eq.([\[3.11\]](#3.11){reference-type="ref" reference="3.11"}) can be represent in the form: \(\hbar\omega_{0}=\sqrt{|\Delta_{1}||\Delta_{2}|}\), that differs from the mass of the Higgs mode in single-band superconductors \(\hbar\omega_{0}=2|\Delta|\). Thus, in single-band superconductors this mode exists in the free quasiparticle continuum, hence, it is unstable. On the contrary, in two-band superconductors it can be \(\sqrt{|\Delta_{1}||\Delta_{2}|}<2\min(|\Delta_{1}|,|\Delta_{2}|)\), then the Higgs mode becomes stable. The reason for the difference in energies between the Higgs modes in two-band and single-band SC systems is, that the positive feedback in Eqs.([\[3.3\]](#3.3){reference-type="ref" reference="3.3"}) for equilibrium values of the OP occurs. The \"light\" speed \(\upsilon\) is determined with Eqs.([\[3.19\]](#3.19){reference-type="ref" reference="3.19"},[\[3.20\]](#3.20){reference-type="ref" reference="3.20"}), i.e. it is of order of Fermi velocities in the bands: \(\upsilon\sim\upsilon_{F1},\upsilon_{F2}\). The result for spectrum of Higgs oscillations in two-band superconductors has been compared with measurements of the excitation energy \(\varepsilon_{0}\) for polycrystalline samples \(\mathrm{Mg}_{1-x}\mathrm{Al}_{x}\mathrm{B}_{2}\) and \(\mathrm{Mg}\mathrm{B}_{2}\) with different \(T_{c}\) at the same temperature-Fig.([\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}). We can see, that the dependence \(\varepsilon_{0}^{2}=\Delta_{\sigma}\Delta_{\pi}\) takes place for both substances, that corresponds to our theoretical result for the Higgs mode. At the same time, these experimental results are difficult to interpret as manifestation of the Leggett mode, based on the formula of Ref.. 4) It is demonstrated, that the resonant enhancement of the DC current through a JJ at the resonant bias voltage \(V_{\mathrm{res}}\), when the Josephson frequency or its harmonics match the frequency of some internal oscillation mode or its harmonics in two-band superconductors (banks)-Eq.([\[4.1\]](#4.1){reference-type="ref" reference="4.1"}), which was observed in the experiments, is a result of coupling between AC Josephson current and Higgs oscillations-Eqs.([\[4.15\]](#4.15){reference-type="ref" reference="4.15"},[\[4.17\]](#4.17){reference-type="ref" reference="4.17"}). Amplitudes of the resonant peaks are proportional to the square of the critical Josephson currents. The main resonance at the voltage \(V_{\mathrm{res}}=\frac{\hbar\omega_{0}}{2e}\) is the most pronounced. The currents in subharmonics are very weak. These subharmonics are caused with the anharmonicity in the current-phase relation ([\[4.2\]](#4.2){reference-type="ref" reference="4.2"}) and with excitation of several quanta of the Higs mode simultaneously by a Cooper pair due to the passing through the energy difference \(2eV\). In the I-V characteristics of the JJ between two-bands superconductors we should observe picture, as Shapiro spikes, at given voltage and, as Shapiro steps, at given current-Fig.([\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"}). This phenomenon would have been observed in single-band superconductors also. However, as mentioned before, the Higgs mode is unstable in single-band superconductors. Therefore, the resonant enhancement is strongly suppressed. On the contrary, in two-band superconductors, if \(\hbar\omega_{0}<2\min(|\Delta_{1}|,|\Delta_{2}|)\) occurs, then the Higgs mode is stable, therefore, the corresponding resonances in AC Josephson effect can be observable. Thus, *explanation of this effect does not need Leggett oscillations, hence the effect cannot be considered as experimental confirmation of these oscillations*. This research was supported by grant of National Research Foundation of Ukraine \"Models of nonequilibrium processes in colloidal systems\" 2020.02/0220 and by theme grant of Department of physics and astronomy of NAS of Ukraine: \"Noise-induced dynamics and correlations in nonequilibrium systems\" 0120U101347 | {'timestamp': '2022-03-16T01:06:10', 'yymm': '2107', 'arxiv_id': '2107.05695', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05695'} |
null | null |
# Introduction
Classical black holes, as solutions of Einstein equations, are characterized by just a few quantities (mass, charge, angular momentum). When quantum processes are incorporated into their study, one finds that such objects behave as thermodynamic systems and radiate like a black body. To compensate the energy that is lost in this process, its temperature increases with the decrease of its radius, in such a way as to lead to a catastrophic behavior in its final stage, the evaporation. However, at this stage it is expected that processes linked to quantum gravity must be considered. In the context of string theory, more precisely within the black holes/string correspondence, it has been proposed that we can take into account the quantization of spacetime through the idea that the coordinates in the target space become noncommutative. Based on this idea, Nicolini, Smailagic and Spallucci studied the evaporation of a Schwarzschild-like black hole when spacetime is noncommutative and found that, rather than a catastrophic behavior, the final stage of such a black hole is actually a finite-mass cold remnant. The noncommutativity of spacetime can be described via the commutator \[= \theta\, \epsilon^{\mu\nu}\,,\] where \(\epsilon^{\mu\nu}\) is an antisymmetric matrix and \(\theta\) is the parameter that will measure the noncommutativity of the coordinates. Nicolini et al proposed that this noncommutativity can be incorporated into gravitation without changing Einstein's equations, but modifying the energy-momentum tensor to include such effects directly in the source of the curvature. In this way, a collapsed object like a black hole would not be described by a singularity, but by a kind of anisotropic fluid, diffused over a certain space region proportional to the parameter \(\theta\). Originally, this distribution was treated in a Gaussian way, for simplicity's sake, but non-Gaussian black holes, based on the noncommutativity of spacetime, were proposed afterwards (see, for instance, ). The thermodynamics of such objects has already been studied, as well as other properties. In this paper, we will focus on the Joule-Thomson expansion of noncommutative black holes, based on what is called black hole chemistry, where the mass of the black hole is identified with its enthalpy. Black hole chemistry is based on the idea that black holes have a mechanical pressure and a thermodynamic volume among their state variables (for a review on the subject, see ). The main reason for this proposal is perhaps based on the fact that, considering the cosmological constant as proportional to the pressure, its conjugate variable is exactly the volume of a sphere whose radius is the horizon radius for the Schwarzschild AdS metric. Among the processes studied in the context of black hole chemistry, the Joule-Thomson expansion analyzes the black hole temperature variation by its pressure, keeping the enthalpy of the system constant. Thus, it is worth studying different metrics for black holes to look for some similarity with fluid/gas systems, as van der Waals ones. An interesting result is that Schwarzschild AdS black holes do not have temperature inversion curves unless some kind of electrical charge is included in the system, that is, for a charged black hole like the Reissner-Nordström type. This brings us to the question of whether there are black holes that have temperature inversion curves even when not charged. The main motivation of this work is to investigate whether uncharged noncommutative black holes satisfy this criterion, i.e., that they have or not these inversion temperatures. The result obtained here is that noncommutative black holes do have these inversion temperature curves, and we also present their behaviors throughout \(T-P\) phase diagrams. The paper is arranged as follows. In section [2](#JTBHNC){reference-type="ref" reference="JTBHNC"} we establish the noncommutative black hole scenario, describe the Joule-Thomson expansion which is related to the inversion temperature. In section [3](#num){reference-type="ref" reference="num"} we present and discuss our numerical results for the inversion temperature curves as well as the isenthalpics in \(T-P\) phase diagrams. In section [4](#conc){reference-type="ref" reference="conc"} we bring our conclusion and last comments.
# Joule-Thomson expansion in non-commutative black holes {#JTBHNC}
Let us start this section introducing the metric for noncommutative black hole as proposed by Nicolini et al as \[ds^2 =-f(r)\, dt^2 + \frac{ dr^2}{f(r)}+ r^2\, d\Omega^2,\] where \(d\Omega^2\) is the solid angle, \(f(r)\), the so-called horizon function, \[f(r)=1-\frac{4 M }{\sqrt{\pi }\, r}\,\gamma \left(\frac{3}{2},\frac{r^2}{4 \theta }\right) + \frac{r^2}{L^2}, \label{funcaocompleta}\] and \(\gamma(a,x)\) is the lower incomplete gamma function, defined by \[\gamma (a, x) = \int_0^{x} dt \, t^{a-1} \, e^{-t}\.\] As stated in the introduction, the parameter \(\theta\) measures the noncommutativity of spacetime. The parameter \(M\) is the mass of the black hole, and the other parameter of the above metric, \(L\), is the radius of the anti-de Sitter spacetime. To get a feeling of the role of the noncommutative paremeter \(\theta\) in the black hole metric let us expand it for small \(\theta\), as \[\begin{aligned}
f(r) &\approx& 1-\frac{2 M}{r}+\frac{2 M e^{-\frac{r^2}{4 \theta }}}{\sqrt{\pi } \sqrt{\theta }}+\frac{4 \sqrt{\theta } M e^{-\frac{r^2}{4 \theta }}}{\sqrt{\pi } r^2} + \frac{r^2}{L^2}. \label{funcaoaproximada} \end{aligned}\] and for large \(r^2/4\theta\), one has a Taylor expansion for \(e^{-x^2}\) around \(x_0=M/\sqrt{\theta}\) (large) which means \(r\approx 2M\) (classical Schwarzschild radius): \[\begin{aligned}
e^{-\frac{r^2}{4\theta}} = e^{-\frac{M^2}{\theta}}\left[ 1-\frac {Mr}{\theta} + \frac {2M^2}{\theta} + \dots \right] \end{aligned}\] so that \[\begin{aligned}
f(r) &\approx& 1-\frac{2 M}{r}+ \left[ \frac{2 M }{\sqrt{\pi } \sqrt{\theta }}+\frac{4 \sqrt{\theta } M }{\sqrt{\pi } r^2} \right]e^{-\frac{M^2}{\theta}}\left[ 1-\frac {Mr}{\theta} + \frac {2M^2}{\theta} + \dots \right]+ \frac{r^2}{L^2}. \end{aligned}\] As one can see, in the above expression there is a \(1/r^2\) term which resembles a charge-like term, but whose charge would be dependent on the black hole mass. This is an indication that the noncommutative AdS black hole share some properties with a Reissner-Nordström AdS black hole. From now on, we will only work with the exact expression for the event horizon function, equation ([\[funcaocompleta\]](#funcaocompleta){reference-type="ref" reference="funcaocompleta"}). Following, the mass \(M\) of the black hole is identified with its enthalpy, then the first law of thermodynamics can be written as \[\label{firstLaw} dM\equiv dH = T dS + V dP\,\] and the thermodynamic volume can be calculated by \[\label{vol} V = \left( \frac{\partial M}{\partial P} \right)_{S}\.\] The mechanical pressure is defined as \[P = \frac{3}{8 \pi} \frac{1}{L^2}\,,\] so that for a Reissner-Nordström AdS black hole the volume would be the canonical one, \(V = (4/3) \pi r_+^3\), where \(r_+\) is the horizon radius. Note, however, that the volume of the noncommutative black hole, given by [\[vol\]](#vol){reference-type="eqref" reference="vol"}, will differ from the canonical one. To see this, let us calculate the black hole mass \(M\), from \(f(r_+)=0\), \[M( P, r_+)=\frac{2 \pi ^{3/2} P \, r_+ }{3 \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta}\right)}\, \left(\frac{3}{8 \pi P}+r_+^2\right),\] so that the thermodynamic volume of the system is given by: \[V = \left( \frac{\partial M}{\partial P} \right)_{S} = \frac{2 \, r_+^3 \, \pi^{3/2}}{3 \, \gamma\left(\frac{3}{2},\frac{r^2}{4 \theta}\right)}.\] Our interest, in this paper, is to study the inversion temperature in the Joule-Thomson (JT) expansion. To do this, we start defining the JT coefficient, given by \[\mu = \left(\frac{\partial T}{\partial P} \right)_M = \frac{1}{C_P} \left[T \left( \frac{\partial V}{\partial T} \right)_P-V \right], \label{JouleThomson}\] which tells us how the temperature varies with increasing (or decreasing) pressure. The inversion temperature occurs when \(\mu = 0\), i.e., when the system stops heating and starts to cool (or vice versa). Banerjee, Majhi and Samanta, studying the thermodynamics of noncommutative black holes, concluded that the Bekenstein expression is valid in this context, i.e., \(S = A/4\), so we can calculate the Hawking temperature as \[T_H=\frac{\partial M}{\partial S},\] which gives us \[\label{thawk} T_H= \frac{{\theta ^{3/2}} \left( \pi P \, r_+^2 +\frac{1}{8} \right) \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right) -{\frac{1}{12} r_+^3 \left( \pi P\, r_+^2 + \frac{3}{8}\right) e^{-\frac{r_+^2}{4 \theta }}}}{ \sqrt{\pi }\,\, {\theta ^{3/2}}\, r_+ \, \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)^2}.\] Using the condition \(\mu=0\) in Eq. ([\[JouleThomson\]](#JouleThomson){reference-type="ref" reference="JouleThomson"}), we obtain the inversion temperature as \[\begin{aligned}
\label{tinv} \nonumber T_i &=& V\left(\frac{\partial T}{\partial V}\right)_P \\ &=& \frac{e^{-\frac{r_+^2}{4 \theta }} \left(24 \theta ^3 \left(8 \pi P r_+^2-1\right) e^{\frac{r_+^2}{2 \theta }} \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)^2+r_+^6 \left(8 \pi P r_+^2+3\right)\right)}{48 \sqrt{\pi } \theta ^{3/2} \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)^2 \left(12 \theta ^{3/2} r_+ e^{\frac{r_+^2}{4 \theta }} \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)-r_+^4\right)} \nonumber \\ &+&\frac{e^{-\frac{r_+^2}{4 \theta }} \left(+\sqrt{\theta } r_+^3 e^{\frac{r_+^2}{4 \theta }} \left(-18 \theta +8 \pi P r_+^4+r_+^2 (3-112 \pi \theta P)\right) \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)\right)}{48 \sqrt{\pi } \theta ^{3/2} \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)^2 \left(12 \theta ^{3/2} r_+ e^{\frac{r_+^2}{4 \theta }} \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)-r_+^4\right)} \end{aligned}\] Imposing that the inversion temperature of the above equation and the Hawking temperature, Eq. ([\[thawk\]](#thawk){reference-type="ref" reference="thawk"}), are the same when \(P = P_i\), we can calculate their difference as: \[\begin{aligned}
\Delta T = \frac {1}{12\sqrt{\pi }} \Bigg\{ &-& 24\, \theta ^{3/2} \left(4 \pi P r_+^2+1\right) e^{\frac{r_+^2}{4 \theta }} + \, r_+^6 \theta ^{-3/2} \left( \pi P r_+^2+ \frac{3}{8}\right) e^{-\frac{r_+^2}{4 \theta }} \, \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)^{-2} \cr &+& \frac{1}{2\,\theta} r_+^3 \left[\frac{3}{4} \,\theta + \pi P r_+^4+ r_+^2 (\frac{3}{8}-2 \pi \theta P)\right] \gamma \left(\frac{3}{2},\frac{\text{rh}^2}{4 \theta }\right)^{-1} \Bigg\} \cr\cr &\times &\left[{ \, 12\, \theta ^{3/2} r_+ \gamma \left(\frac{3}{2},\frac{r_+^2}{4 \theta }\right)-r_+^4}\right]^{-1} = 0\,, \end{aligned}\] which is an equation for \(r_+\) and \(P\). Solving it numerically for \(r_+\), we can obtain the inversion temperature using Eq. ([\[thawk\]](#thawk){reference-type="ref" reference="thawk"}). These results will be presented and discussed in the next section.
# Numerical analysis {#num}
In this section we present the numerical solutions for the inversion temperature as a function of the pressure for some values of \(\theta\). Besides, we also present insenthalpic curves (\(M\)= constant) in \(T-P\) diagrams. To study the influence of the parameter \(\theta\) on the inversion temperature against the pressure, we analyzed the range of \(\theta\) from \(10^{-3}\) to \(10^{3}\). In the left panel of Fig. [\[inversioncurves1\]](#inversioncurves1){reference-type="ref" reference="inversioncurves1"}, we plot \(\theta\)= 0.001, 0.1, 0.2, 0.3, 1 and 10 and see that increasing these values the slope of these curves also increase. This behavior also holds for higher values of \(\theta\) and we checked it until \(\theta\) reaches \(10^3\). The shapes of the inversion temperature curves are similar to that of charged black holes in commutative theory. Notice that the limit \(\theta \rightarrow 0\) should be taken with caution, since for \(\theta = 0\), the system itself has no solution. In particular, for very small \(\theta\) the slope is also very small, the intercept is high and the curve appears to become a straight line. This seems to be an indication that in the commutative limit \(\theta\to 0\), there is no inversion temperature, as expected in AdS-Schwarzschild black holes. In the right panel of Fig. \(\ref{inversioncurves1}\), one can see the behavior of the curves for small values of the inversion pressure. We are interested in finding the minimum inversion temperature (\(P = 0\)), which was one of the values investigated in the first paper on the subject. As one can note, as the parameter \(\theta\) increases, \(T_i^{\text{min}}\) decreases. This is also the same behaviour for the charged black hole in general relativity. We also plot some isenthalpics of the system, i.e., when the mass is constant. We made such a plot for different values for mass and the parameter \(\theta\), as is show in figure [\[figIsenthalpics\]](#figIsenthalpics){reference-type="ref" reference="figIsenthalpics"}. The solid line is the inversion temperature curve for the given parameter \(\theta\), which intersects the isenthalpics at their maxima. These curves also resemble the corresponding ones for charged black holes. This feature still remains the same for higher values of \(\theta\), but in this case one should also increase the values of the black hole mass.
# Conclusions {#conc}
In this paper, we studied the Joule-Thomson expansion for a noncommutative black hole. Such an expansion is a thermodynamic process where the temperature of the system, usually a gas, varies with pressure, and may heat or cool. For real gases, there is usually an inversion temperature (and inversion pressure) where the gas stops heating and starts to cool, or vice versa. For black holes, this inversion only occurs in the presence of some type of charge, such as an electrical charge, for example. In the case of a Schwarzschild black hole, this reversal from heating to cooling does not occur. We did this study by dealing with the question numerically, plotting the temperature inversion curves for the noncommutative black hole, for different values of the noncommutativity parameter. We also plot the isenthalpics of the system to capture the whole thermodynamic scenario. Finally, our aim this work was to verify what happens when the black hole is uncharged but not commutative. We found that not only does inversion temperature occur, but the role of the noncommutativity parameter is similar to that of a charge, i.e., the higher its value, the lower the minimum temperature of the system for which the inversion occurs. | {'timestamp': '2021-11-03T01:24:09', 'yymm': '2107', 'arxiv_id': '2107.05781', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05781'} |
null | null |
null | null |
# Introduction {#sec:Intro}
The relationship between amenities in or near residential, work or school--neighborhood environments and health is receiving increasing attention, given that these environments can influence health-related behaviors and subsequent outcomes. Where spatial proximity to supermarkets is associated with diet, so too are recreational facilities associated with physical activity and fast food restaurants near schools associated with child obesity. Work in this area has been limited by the lack of knowledge of what geographic units are most relevant for exposure assessment, i.e. the well known modifiable unit areal problem (MAUP). Additionally, there may also be measured or unmeasured person-level behaviors or characteristics that give rise to the "uncertain geographic context problem" (UGCP). Whereas the former establishes that using different spatial units or spatial scales to define exposure measures will yield different estimates of association, the latter acknowledges that the most relevant spatial unit may differ from place to place or subject to subject due to place or person characteristics such as predominant transport modes in a given area or vehicle ownership, among others.
Recent work has begun addressing these issues by foregoing the pre-specification of the spatial unit used to construct exposure metrics. Instead, the association between proximity to amenities of interest, broadly referred to as built environment features (BEFs), and subjects' outcomes is estimated as a continuous function of distance between subjects and amenities. Whereas typical models regress the outcome on a BEF metric that depends on a pre-defined scale, these new methods use all the pair-wise distances between subjects and BEFs as inputs to the model. Specifically, in order to address the MAUP, an idealized smooth function \(f(d)\) is used to represent the association between the health outcome of interest and a single BEF placed at distance \(d>0\) from the subject. Having \(f(d)\) as the objective of inference enables the visualization of whether and how the association between availability of amenities and outcomes dissipates with distance, as well as estimation of the spatial scale, defined as the distance at which the association is negligible, i.e. \(d: f(d) =0\).
The function \(f(d)\) has been modeled in different ways: estimated \(f(d)\) non-parametrically by first discretizing the distances into a grid, and using the count of distances within bins defined by the grid as predictors in a Distributed Lag Model (DLM), i.e., the count of distances within each bin are conceptualized as distributed lag predictors, indexed by the corresponding value of the grid. The coefficients corresponding to each distributed lag predictor are smoothed using splines, yielding estimates of \(f(d)\) at the values of \(d\) used to construct the grid. Alternatively, modeled \(f(d)\) parametrically, typically using exponential functions to enforce the substantive belief that the association between health outcomes and spatial availability of amenities monotonically decays across distance, e.g. \(f(d) \propto \exp(-\frac{d}{\theta})\). However, the estimation of \(f(d)\) at the population level, as the previous methods propose, fails to account for the concerns the UGCP raises regarding unmeasured person-level behaviors or place-level factors that may determine subject -- or location -- specific spatial association.
Building upon their work in DLMs, constructed a hierarchical DLM (HDLM) allowing for the estimated \(f(d)\) to vary between subjects and or locations, according to pre-specified groups (e.g., different \(f(d)\) by sex), as well as unexplained variation in the association (i.e., using the idea of random coefficients to estimate \(f(d)\) for individual subjects). However, the HDLM approach, has some disadvantages: (1) it uses discretized distances to estimate association across space, unnecessarily coarsening the exposure information; (2) it requires pre-specifying the groups where heterogeneity in the association may occur (covariates and or subjects); and (3) by enforcing that heterogeneity in the association estimates to occur at the subject-level through random effects, it loses possible gains in precision that could result from pooling subjects with similar levels of association.
Motivated by the desire to identify schools where pupils may be at greater risk of obesity related to the proximity of fast food restaurants (FFRs), we propose a model that clusters schools-specific association curves, \(f(d)\) according to the strength of association between the spatial proximity of nearby FFRs and child obesity. Clustering provides investigators and policymakers with a greater understanding of the kinds of relationships that exist between students and their environment as well as identifies schools where students may be at greater risk, as identifying risk groups may help prioritize population level interventions. The data for this motivating study consists of body weight status of children nested within schools across Los Angeles County during academic years 2001-2008. Distances between schools and FFRs are calculated from geocoded school addresses, supplied by the California Department of Education, and geocoded FFR business addresses from the National Establishment Time Series Database
Our method uses the Dirichlet Process Mixture (DPM) prior and a spline basis function expansion to non-parametrically estimate both the number of cluster-BEF effects, and the nonlinear association functions across space, respectively. We name our method the Spatial Aggregated Predictor-Dirichlet Process, to reflect this dual non parametric estimation, but refer to it more generally as STAP-DP, given its potential for also modeling temporal exposures. Our approach is inspired by the work of and on clustering functions using DPM family priors. We use the penalized spline approach developed by and further popularized by to construct the estimates of the association functions, and use the DPM to cluster the spline coefficients.
Section 2 describes the model that estimates homo-and heterogeneous BEF effects. Section 3 studies the performance of the STAP-DP model in a variety of simulated data settings and discusses how the results may inform normative practice. Section 4 describes the application of the STAP-DP model to the motivating study on child obesityin Los Angeles. We conclude our work with a discussion of the model and future directions to explore.
# Model {#sec:Model}
We now introduce the STAP-DP framework, describing how we incorporate the estimation of heterogeneous BEF effects into a regression framework. We limit our discussion to the estimation of only one BEF's effects in space, FFRs for example, as the extension to multiple BEFs is straightforward. We organize our discussion into four parts. First, we build intuition for our approach by defining the STAP estimated via spline basis functions at the population level, i.e., homogeneous effect. Then, we define how to extend the STAP model to estimate heterogeneous effects--at the latent cluster level-- for a univariate outcome. In the final two sections we generalize the clustering framework for repeated outcome measures and discuss estimation.
## The STAP Model {#sec:Model_Univariate}
Suppose a continuous outcome \(Y_i\) (\(i=1...N\)) and corresponding covariates \(\bm{X}_i \in \mathbb{R}^{n \times p}\) are observed for a sample of \(N\) subjects. Additionally, spatial data, \(\mathcal{D}_i\) which contains distances, \(d\), between subject \(i\) and all FFRs within some substantively determined radius \(R\), are also measured. The inferential objective is to estimate function \(f(d)\), which represents the expected difference in the outcome associated with placing a single FFR at distance \(d\) after adjusting for covariates \(\bm{X}_i\). Defining \(F(\mathcal{D}_i) := \sum_{d \in \mathcal{D}_i} f(d)\), as the aggregated FFR effect under the assumption of additivity, we complete the initial STAP model formulation: \[\begin{aligned}
Y_i &= \bm{X}_i^{T}\bm{\gamma} + F(\mathcal{D}_i) + \epsilon_i, \tag{1} \label{eqn:stap_general} \\ \epsilon_i &\stackrel{iid}{\sim} N(0,\sigma^2), \end{aligned}\] where \(\epsilon_i\) is the residual error, with variance \(\sigma^2\). As mentioned in Section [1](#sec:Intro){reference-type="ref" reference="sec:Intro"}, there are a number of approaches to model \(f(d)\). In this work, we propose to model \(f(d)\) as a linear combination of basis functions, \(\{\phi\}_{l=1}^L\), which allows us to rewrite \(F(\mathcal{D}_i)\) as follows: \[\begin{aligned}
F(\mathcal{D}_i)&= \sum_{d \in \mathcal{D}_i} f(d) = \sum_{d \in \mathcal{D}_i} \sum_{l=1}^{L} \beta_l\phi_l(d), \label{eqn:f_def} \tag{2} \end{aligned}\] where \(\phi_l(d)\) is the evaluation of the distance through the \(l\)th basis function and \(\beta_l\) is the corresponding regression coefficient. In this work we use \(L\) spline basis functions defined across a set of equally spaced knots, though other knot placements or basis functions could be used. In order to avoid over fitting when \(L\) is large, the regression coefficients are regularized through the use of a quadratic penalty on \(\bm{\beta}\) implemented through a smoothing matrix \(S\) and tuned by penalty parameter \(\tau\). We use the difference penalty matrices of, a widely used spline penalty formulation.
Within a Bayesian paradigm, this penalty is equivalent to specifying a multivariate normal prior with improper precision matrix \(\tau S\). We adopt a variant of this Bayesian approach and, to improve computational efficiency in our more complex model formulations discussed in the next subsection, we first transform the spline basis function expansion matrix, \(\Phi(d)\), such that the transformed coefficients can have independent normal priors. While centering constraints are often imposed on \(\Phi(d)\) to avoid collinearity with the intercept in \(\bm{X}\), this constraint is not needed in our model (see supplementary material). Given that \(r_S =\)rank(\(S)<L\), two precision parameters for the priors are used, one for the first \(r_S\) coefficients and a second for the last \(L-r_S\) coefficients: \[\begin{aligned}
\bm{\beta}_1 &\sim MVN_{r_S} \left (\bm{0},\sigma^2 \tau_1^{-1} \bm{I}_{r_S} \right ) \quad \bm{\beta}_2 \sim MVN_{L-r_S}(\bm{0},\sigma^2 \tau_2^{-1}\bm{I}_{L-r_S}) \tag{3} \label{eqn:prior}\\ \tau_z &\stackrel{iid}{\sim} \text{Gamma}(a_{\tau},b_{\tau}) \quad z = 1,2. \end{aligned}\] In \((\ref{eqn:prior})\) we denote \(\bm{\beta}_z\), \(z=1,2\), as the regression coefficients in the penalty range and null space, respectively. Correspondingly, \(\tau_1\) and \(\tau_2\) are the respective precisions for these separate subsets of \(\bm{\beta}\). For ease of further exposition we define \(\bm{\Lambda}\) as the diagonal covariance matrix which has \(\tau_1^{-1}\) as the first \(r_S\) diagonal elements and \(\tau_2^{-1}\) as the last \(L-r_S\) diagonal elements, so that the prior in ([\[eqn:prior\]](#eqn:prior){reference-type="ref" reference="eqn:prior"}) can be written simply as \(\bm{\beta} \sim MVN_L(\bm{0},\sigma^2\bm{\Lambda})\). We place independent conjugate Gamma priors on \(\bm{\tau}= (\tau_1,\tau_2)\) so that both \(\bm{\beta}\)'s and \(\bm{\tau}\)'s conditional posterior distributions are available in closed form.
## STAP-DP with Univariate Outcomes
In alignment with this work's goal to estimate heterogeneous effects, we replace \(F_i(\mathcal{D}_i)\) with \(F_i(\Dset)\) in ([\[eqn:stap_general\]](#eqn:stap_general){reference-type="ref" reference="eqn:stap_general"}) while allowing for clustering in the \(f_i(d)\). Given that \(f_i(d)\) is represented by the fixed spline functions and random coefficients \(\bm{\beta}\), we implement this clustering goal by placing a DP prior on the vector of regression coefficients, \(\bm{\beta}\), and associated penalty parameter, \(\bm{\tau}\): \[\begin{aligned}
(\bm{\beta},\bm{\tau}) &\sim P \label{eqn:stapdp1} \tag{4}\\ P &\sim DP(\alpha,P_0)\\ P_0 &\equiv MVN_{L} \left(\bm{0},\sigma^2 \bm{\Lambda} \right )\times \prod_{z=1}^{2} \text{Gamma}(a_{\tau},b_{\tau}). \end{aligned}\] In ([\[eqn:stapdp1\]](#eqn:stapdp1){reference-type="ref" reference="eqn:stapdp1"}), \(P\) is a random measure drawn from Dirichlet Process \(DP(\alpha,P_0)\), where \(\alpha>0\) is a concentration parameter reflecting the variability of distribution \(P\) around base measure \(P_0\). \(P_0\) is chosen to retain the prior previously discussed in ([\[eqn:prior\]](#eqn:prior){reference-type="ref" reference="eqn:prior"}).
By placing the DP prior on \((\bm{\beta},\bm{\tau})\), clustering is induced on the \(f_i(d)\) as can be seen from the stick breaking construction of the DP: \(P = \sum_{k=1}^{\infty} \pi_k \delta_{(\bm{\beta}^\star,\bm{\tau}^\star)}(\cdot)\). In this representation \(\pi_k\) represents the probability the \(i\)th observation is assigned to the \(k\)th exposure function and \(\delta(\cdot)\) is the dirac-delta function. Each \(\pi_k\), itself is composed of the "broken sticks" created from variables drawn from a Beta distribution: \(\pi_k = v_k \prod_{u<k} (1-v_u); v_k \sim \text{Beta}(1,\alpha)\).
Combining all these pieces together, our proposed STAP-DP model for univariate outcome \(Y_i\) takes the following form: \[\begin{aligned}
Y_i &= \bm{X}_i\bm{\gamma} + \sum_{d \in \mathcal{D}_i} \sum_{l=1}^{L} \beta_{il}\phi_l(d) + \epsilon_i \label{eqn:mod1} \tag{5}\\ \epsilon_i &\stackrel{iid}{\sim} N(0,\sigma^2)\\ (\bm{\beta},\bm{\tau}) &\sim P\\ P &\sim DP(\alpha,P_0) \\ P_0 &\equiv MVN_L\left(\bm{0}, \sigma^2 \bm{\Lambda} \right )\times \prod_{z=1}^{2}\text{Gamma}(a_{\tau},b_{\tau}). \end{aligned}\] A final comment is warranted regarding the choice of the number and placement of the \(L\) knots in constructing the splines. While our approach follows previous work in placing a sufficient number of knots equally across the domain of observed distances, deciding what number of knots is "sufficient" requires greater statistical judgement than in standard applications. Clusters may be more difficult to detect when the dimension on which clusters are formed (i.e., number of coefficients) is large and the between-cluster differences are small (low signal effects). Conversely, more clusters may be identified in a setting with a stronger signal and greater number of knots. Thus, \(L\) must be chosen to balance accuracy in both function estimation *and* cluster discrimination.
## STAP-DP with Repeated Measurements {#sec:Model_Multivariate}
Extending ([\[eqn:mod1\]](#eqn:mod1){reference-type="ref" reference="eqn:mod1"}) to correlated outcomes, we consider the setting in which subjects are measured repeatedly over time, for \(j=1...,n_i\) occasions. This results in outcome \(Y_{ij}\) (\(i=1...,N,j=1...,n_i)\) modeled as a function of covariates \(\bm{X}_{ij}\), and their corresponding coefficients \(\bm{\gamma}\). The distance set adopts the new visit-specific index as well, i.e., \(\mathcal{D}_{ij}\), indicating it may vary over time; for instance FFRs may open and close between measurement occasions. Finally, a subset of \(\bm{X}_{ij}\), \(\bm{Z}_{ij}\), is included in the model, along with subject-specific coefficients \(\bm{b}_i \sim MVN(0,\Sigma)\) to account for within subject variability in standard fashion. Augmenting ([\[eqn:stapdp1\]](#eqn:stapdp1){reference-type="ref" reference="eqn:stapdp1"}) accordingly, we arrive at our final model: \[\begin{aligned}
\tag{6} \label{eqn:long1} Y_{ij} &= \bm{X}_{ij}^{T}\bm{\gamma} + \sum_{d\in\mathcal{D}_{ij}}\sum_{l=1}^{L}\beta_{il}\phi_l(d) + \bm{Z}_{ij}^{T}\bm{b}_i + \epsilon_i\\ \bm{b}_i & \stackrel{iid}{\sim} N(\bm{0},\Sigma)\\ \epsilon_i &\stackrel{iid}{\sim} N(0,\sigma^2)\\ (\bm{\beta},\bm{\tau}) &\sim P\\ P &\sim DP(\alpha,P_0) \\ P_0 &\equiv MVN_L \left ( \bm{0}, \sigma^2 \bm{\Lambda} \right )\times \prod_{z=1}^{2}\text{Gamma}(a,b). \end{aligned}\]
## Estimation {#sec:Model_Estimation}
In order to fit models of the form described in ([\[eqn:stapdp1\]](#eqn:stapdp1){reference-type="ref" reference="eqn:stapdp1"}) and ([\[eqn:long1\]](#eqn:long1){reference-type="ref" reference="eqn:long1"}), we truncate the DP so that a blocked Gibbs sampler can be used to draw samples from the posterior. While this sampler is fairly straightforward, it bears mentioning that \(\Phi(d)\) has to be adjusted at each iteration of sampling so that any DP components associated with 0 or some small number of observations are not included in the usual matrix inversion used to estimate the mean of the conditional posterior distribution for the regression coefficients, \(\bm{\beta}^{*} = [\bm{\gamma},\bm{\beta}]^{T}\). Instead, coefficients for those low-member cluster components are sampled with draws from the prior. For example, if on the \(m\)th iteration, none of the \(N\) observations are assigned to the \(k\)th DP component, then the samples of the spline regression coefficients for that iteration, \(\bm{\beta}_k^{(m)}\), are drawn from a \(MVN_L(\bm{0},\sigma^2\bm{\Lambda}_k)\) prior, where \(\bm{\Lambda}_k\) is the cluster specific covariance matrix, and the columns of zeros that would otherwise be included in \(\Phi(d)\) are omitted.
We present the closed form conditional posteriors and associated algorithm in the Supplementary Material. Our algorithm is implemented in C++ which can be called from our R package `rstapDP`.
For both our simulations and California data we use `rstapDP` to fit the STAP-DP in `R` (v.4.0.2) on a MacOS Catalina operating system with a 2.8 GhZ Quad-Core Intel Core i7 processor.
# Simulations {#sec:Simulations}
## Simulation Design
For a given sample size, the ability of the STAP-DP model to correctly classify subjects depends on (a) the proportion of subjects belonging to that cluster, (b) the difference in the \(f_i(d)\) functional forms, and (c) the distribution of distances (i.e., exposure information) present within each cluster. As the first of these three principles follows straightforward sample-size intuitions, in this section we study the STAP-DP's ability to correctly recover cluster specific functions,\(f_i(d)\), and cluster partitions in the latter two settings. Using simulated data we vary: (i) cluster effect size and (ii) distance distributions in order to see how these may impact correct cluster classification. We focus on evaluating cluster classification accuracy as it is the upstream predictor of all remaining model components, like the estimation of the \(f_i(d)\), which are all standard Bayes estimators conditional on the correct cluster classification.
We evaluate our method's ability to correctly classify subjects using a partition loss function developed by and used regularly in DP and other mixture model applications where label-switching may be of concern. Our employment of the loss function equally weights correct and incorrect classification, using the subjects' true and estimated class indicators, \(\zeta_i,\hat{\zeta}_i\), respectively: \[\begin{aligned}
\tag{7} \mathcal{\psi}(\bm{\zeta},\hat{\bm{\zeta}}) = \sum_{(i,i'); i< i' < N} I(\zeta_i = \zeta_{i'},\hat{\zeta}_i \neq \hat{\zeta}_{i'}) + I( \zeta_i \neq \zeta_{i'},\hat{\zeta}_i = \hat{\zeta}_{i'}).\label{eqn:loss} \end{aligned}\] Conceptually, ([\[eqn:loss\]](#eqn:loss){reference-type="ref" reference="eqn:loss"}) tallies the number of times that observations \(i\) and \(i'\) are incorrectly assigned to different clusters, when they in fact belong in the same cluster, as well as tallies when they are incorrectly assigned to the same cluster.
In each simulation setting discussed below, we generate 25 datasets and then fit the STAP-DP model shown in ([\[eqn:stapdp1\]](#eqn:stapdp1){reference-type="ref" reference="eqn:stapdp1"}), truncating the DP at K= 50 and using weakly informative Gamma(1,1) priors on \(\sigma^{-2},\alpha, \tau_1\) and \(\tau_2\), respectively. We draw 2000 samples from the posterior distribution for inference via Gibbs Sampling using `rstapDP` after discarding 2000 initial samples for burn-in. Across all 25 simulations we evaluate the loss ([\[eqn:loss\]](#eqn:loss){reference-type="ref" reference="eqn:loss"}) across all \(M=2000\) iterations of the posterior samples drawn via Gibbs sampling. Given that the loss function does not have a standard range, we normalize the loss results by the maximum loss across all simulation settings, so as to make the results more interpretable relative to one another. We have organized the files used to run the simulations in the STAPDPSimulations R package available via [Github]( https://github.com/apeterson91/STAPDPSimulations).
## Cluster Effect Size
Our first simulation study focuses on model performance as a function of the difference between two clusters' \(f(d)\), defined in (8) below, with each observation having a 50% probability of being assigned to either of these clusters. The cluster function set-up is intended to mimic a hypothetical high and low risk population scenario, in which subjects with equivalent exposure to the same BEFs experience different effects according to which risk population they belong. For each subject we generate a random number of distances uniformly so that the average number of BEFs is 15. Conditional on the number of distances, the distances themselves are then generated according to the "Skew" distribution shown in Figure S1 in the Supplementary Information. This distribution was selected in order to test our model's performance under a "worst case scenario", given that with this distribution there is relatively less exposure information at the distances where the BEF effects are non-zero. Specifically, the generative model takes the following form: \[\begin{aligned}
Y_i &= 26 +.5Z_i + \sum_{d \in \mathcal{D}_i} f_{\zeta_i}(d) + \epsilon_i \tag{8} \label{eqn:sim}\\ \epsilon_i &\sim N(0,\sigma^2=1) \qquad i = 1...,200\\ f_1(d) &= \exp\left \{ \left (\frac{-d}{.5} \right)^5 \right\} \qquad \\ f_2(d) &= \nu \exp\left \{ \left (\frac{-d}{.5} \right)^5 \right\} \qquad \nu = (0.25.5.75)\\ P(\zeta_i=1) &= P(\zeta_i=2)=.5; \end{aligned}\] where \(Z_i\) is a covariate generated as a fair Bernoulli random variable, \(\zeta_i\) is the subject specific cluster label indicating the true BEF effect, \(f(d)\), for the \(i\)th observation, and \(\nu\) represents the varying effect size at \(d=0\).
The relative loss as a function of the effect size \(\nu\) is shown in Figure [\[fig:effect_loss\]](#fig:effect_loss){reference-type="ref" reference="fig:effect_loss"}. As expected, one can see a decrease in relative loss and consequently improved classification as the effect size increases.
## Distance Distributions
As our method non-parametrically estimates cluster functions \(f_i(\mathcal{D}_i)\) across continuously measured space using a basis function expansion, correct estimation of the function requires there to be BEFs observed at the relevant distances, \(d: f(d)\neq 0\), within the study area of interest. Of course these "relevant" distances are not known *a priori* and so it is to the benefit of the investigator to err on the side of caution in specifying a larger study area if possible. However, despite any preparatory work that may be done to ensure an adequate area is included at the level of the sample study, it is not clear how differing distributions of distances at the latent cluster level may impact inference. For example, will suburbanites' lower exposure to proximate FFRs impact the ability of the stapDP model to discern the impact of FFRs on their health relative to their more exposure rich urban counterparts? For this reason our second simulation study examines how exposure to different distance distributions may impact classification.
We study this problem by considering three different generative distance distributions which we label "Uniform","CA" and "Skew". The first, straightforwardly, refers to the idealistic-but unrealistic-scenario in which there is equivalent information available at all distances within the study area. The second two cases refer to more realistic situations in which there are more likely to be a higher number of BEFs found further away from the subject than close by -- a consequence of area's quadratic growth as a function of distance. We create the first of these skewed distributions, "CA", by using maximum likelihood to fit a beta distribution to the distribution of distances in our motivating California data distance distribution and the second by altering a beta distribution to be a more extreme version of the first. We generate distances under each distribution for each cluster in order to examine how differing exposure patterns between clusters impact cluster classification. The densities of each of these distributions can be found in Figure S1 in the supplementary material. Since the exposure information depends both on the distribution of distances and the number of BEFs, we generate scenarios where the amount of information increases as a function of the number of BEFs within the same distribution. We simulate data under the same model as proposed in ([\[eqn:sim\]](#eqn:sim){reference-type="ref" reference="eqn:sim"}), with \(\nu=0.25\), to illustrate how a substantial, but not obvious, difference in cluster functions manifest across the varying distance distribution settings. Fitting our STAP-DP model under the priors and sampler settings previously described, we plot the results below in Figure [\[fig:ddist_loss\]](#fig:ddist_loss){reference-type="ref" reference="fig:ddist_loss"}. Figure [\[fig:ddist_loss\]](#fig:ddist_loss){reference-type="ref" reference="fig:ddist_loss"} shows a number of patterns worth highlighting. First, across all distance distributions we observe a decrease in loss as the number of built environment features increases. This is as expected -- more information or exposure results in a more easily detectable signal. Further, distance distribution combinations that include more information result in lower levels of relative loss compared to more skewed distance distributions. There are number of cases where this can be seen in Figure [\[fig:ddist_loss\]](#fig:ddist_loss){reference-type="ref" reference="fig:ddist_loss"}, the most obvious being the top-left diagonal panel where both clusters have Uniform distance distributions; this has the lowest loss values across all panels due to the relative abundance in exposure information. This pattern holds when comparing to the more skewed distributions: The uniform-CA combination has higher error than the uniform-uniform combination when there are a relatively small number of built environment features present.
# Fast food restaurants near schools and child obesity among public school students in Los Angeles {#sec:Results}
There is a pressing need to understand contextual determinants of child obesity, in order to implement population level strategies to reduce and prevent it. The food environment near schools has been proposed and studied as a contextual factor that influences children's diet, and thus obesity. We use data on body weight of children attending public schools in Los Angeles, CA, along with data on the locations of FFRs as a marker of the food environment near schools, and apply our proposed method to identify schools where children may be at higher risk of obesity, related to food environment exposures. Identifying these schools may help prioritize or tailor population-level interventions to address child obesity.
## Data Description {#sec:Data}
Every year public schools in the State of California collect data on the fitness status of pupils in 5th, 7th and 9th grade, as part of a state mandate, using the Fitnessgram battery of tests. Child-level data available for our analysis were collected during academic years 2001-2008 on 5th and 7th graders, and consists of children's weight (Kg) and height (m) and, and the following categorical covariates: sex, race-ethnicity, fitness status (unfit, fit, fit above standard), and grade level. Weight and height are transformed to body mass index (Kg/m\(^2\)), and standardized to BMIz scores according to age-and sex-specific growth curves published by the United States' Center for Disease Control. In contrast to adults, standardization is needed when analyzing data from children of different ages, given children's rapid growth. To aid in managing the large database, and given that all child-level covariates are categorical, the dataset is "collapsed" so that each row represents a group of children within each school defined by the cross-classification of categorical child-level characteristics described above. The average BMIz of children in the group is the outcome of interest. Given the categorical nature of the covariates and our use of weighting by the size of the group represented by each row (below), this approach yields exactly the same results as would be obtained if data had not been grouped, thus avoiding biases in an ecological analysis.
Data on school-level characteristics are also available from the CDE website (see Table [\[tbl:cluster\]](#tbl:cluster){reference-type="ref" reference="tbl:cluster"}), and, importantly, so is the geocode of the school. School geocodes were used for two purposes. First, the geocodes were used to link schools to census tract level covariates. Second, the school geocodes were used to calculate the distances between the school and the geocoded location of each FFR in the LA area. FFRs were identified from the National Establishment Time Series (NETS) database, using a published algorithm that classifies specific food establishments as FFRs. Only FFRs within five miles of schools were kept for this analysis. This distance was chosen to be a conservative as previous work estimated that the distance at which FFRs cease to have an effect on childhood obesity is approximately one mile.
## Los Angeles STAP-DP Model
We fit models estimating both the population-level and latent cluster-level effects -- STAP and STAP-DP models, respectively. Given the available data consisting of subgroups of children defined by the cross-classification of categorical covariates, we use the standardized average BMI within the subgroup as the outcome. The models adjust for the student group and school-level covariates listed in Table [\[tbl:cluster\]](#tbl:cluster){reference-type="ref" reference="tbl:cluster"}. Denoting these covariates as \(\bm{X}_{ijq}\) for student group \(q=1...,n_{ij}\), measured at year \(j=2001...,2008\), attending school \(i=1...,N\), and using notation as described in ([\[eqn:long1\]](#eqn:long1){reference-type="ref" reference="eqn:long1"}) our model for analyzing the Los Angeles data is: \[\begin{aligned}
\text{BMIz}_{ijq} &= \bm{X}_{ijq}^{T}\bm{\gamma} + F_i(\mathcal{D}_{ij}) + b_{i1} + b_{i2}\frac{\text{year}_{ij}}{10} + \epsilon_{ijq}, \tag{9} \label{eqn:LA}\\ \epsilon_{iqj} &\sim N\left(0,\frac{\sigma^2}{n_{iqj}}\right), \\ \bm{b}_i &\sim MVN_2(\bm{0},\Sigma), \end{aligned}\] where \(n_{ijq}\) represents the number of students in student group \(q\) during year \(j\) at school \(i\). Given that FFRs may open or close during the study period, \(\mathcal{D}_{ij}\) represents the distances between school \(i\) and FFRs available within 5 miles during year \(j\). Similar to our simulations, we place a weakly informative Gamma(1,1) prior on each of the penalty parameters in \(\bm{\tau}\), associated with each cluster regression coefficients, the residual precision \(\sigma^{-2}\), and the concentration parameter \(\alpha\). The Gamma(1,1) prior on the concentration parameter is a common prior setting in the DP literature, reflecting the *a priori* expectation that the concentration parameter is \(1\), so that fewer clusters are favored. Additionally, we place a non-informative Jeffrey's prior on the covariance matrix for the school specific \(\bm{b}_i\) vector: \(p(\Sigma^{-1}) \propto \mid \Sigma \mid^{\frac{3}{2}}\). Estimation is conducted through `rstapDP`, drawing 2000 samples from each of 2 independent MCMC chains after 8000 samples have been iterated as "burn-in" on each chain. We check convergence via \(\hat{R}\) diagnostic and visually inspecting traceplots. We use \(L=7\) coefficients in our spline basis function expansion, and similarly use this basis to estimate the \(f_i(d)\) on a grid of values, calculating the 95% point-wise credible interval at each distance grid point. We also calculate the posterior probability of co-clustering which can be arranged in a matrix \(\bm{P} \in \mathbb{R}^{N\times N}\) so that \(\bm{P}_{i,i'} = P(\) school \(i\) is co-clustered with school \(i'\) across post burn-in iterations\()\). School cluster characteristics are tabulated using the cluster mode school assignment calculated using ([\[eqn:loss\]](#eqn:loss){reference-type="ref" reference="eqn:loss"}) as implemented in the `rstapDP` package via the `assign_mode` function.
For comparative purposes, we fit a model similar to ([\[eqn:LA\]](#eqn:LA){reference-type="ref" reference="eqn:LA"}) in all ways save for restricting the \(f_i(d)\) to be estimated at the population level-\(f(d)\). We fit this model using Hamiltonian Monte Carlo via the `rsstap` R package, drawing 1000 samples after 1000 warm-up across 4 independent MCMC chains. Convergence is assessed via \(\hat{R}\) diagnostic and we calculate the analogous posterior estimate for \(f(d)\) across the same grid of distance values.
## Los Angeles Results
Figure [\[fig:f_estimates\]](#fig:f_estimates){reference-type="ref" reference="fig:f_estimates"} shows four functions corresponding to the 3 estimated cluster functions from the STAP-DP model, as well as the 1 homogeneous effect estimate. We name the three cluster effects "Majority", "High Risk" and "Low Risk". These names derive from the proportion of schools assigned to the cluster as well as the relative effect size associated with the function at and around distance 0 mi from the school: In the cluster labeled "High Risk", one additional FFR placed at distance 0 from a school is associated with an expected 0.46 higher BMIz among students attending those schools (95% CI: 0.36, 0.58), all else equal. In contrast, placing one FFR at distance 0 from the schools assigned to the "Low Risk" cluster is associated with lower BMIz score, by-0.15 (95% CI:-0.17,-0.12), all else equal. The analogous values for the "Majority" and homogeneous function estimates, are 0.01 (95% CI: 0.00,0.02) and 0.01 (95% CI: 0.01,0.0132), respectively. In all clusters, the estimated associations rapidly decay with increasing distance, with all association estimates effectively zero by 1 mile. We now turn our attention to the matrix of co-clustering probabilities \(\bm{P}\) which we visualize using a heat-map in Figure S2 in the Supplementary material, after applying 's hierarchical sorting algorithm to group schools with similar co-clustering probabilities together. The estimate of co-clustering probabilities shows approximately 550 schools (90%) are consistently co-clustered within one of the three clusters, reflecting a high degree of model certainty in cluster configurations for these schools. The remaining \(\approx\) 50 schools show a greater uncertainty between being classified in the "Majority" or "Low Risk" cluster. This uncertainty likely stems from an insufficient number of FFRs present within the relevant \(\approx\) 1 mile distance from schools, where the cluster effects are most discernibly different.
Further examination of the school characteristics associated with each cluster details several suggestive, though not conclusive between-cluster differences (Table [\[tbl:cluster\]](#tbl:cluster){reference-type="ref" reference="tbl:cluster"}). Differences between the three clusters of schools are fairly muted, with summary statistics across student-and school-level measures describing similar student populations and levels of median household income and education amongst the neighborhoods of schools in each cluster. The most noteworthy differences amongst the clusters are in the number of FFRs within \(\frac{1}{2}\) mile of the school -- lower in the high risk group as compared to the other two clusters -- and the total enrollment -- higher in the low and high risk groups as compared to the majority cluster.
The lack of substantial differences in measured characteristics between these two clusters is noteworthy, suggesting that none of the observed characteristics appear to modify the obesity risk associated with FFR exposure. Although prior research shows that area-level education and income modify the effects of child obesity interventions, for instance, the protective effects of socioeconomic factors do not appear to extend to children's risk of obesity as due to proximate FFR exposure -- at least within this population. This lack of difference in socioeconomic characteristics suggests that there are unmeasured variables that account for the heterogeneous FFR effects. One potential unmeasured factor could be the type of FFRs proximal to the schools, e.g. chains vs non-chain FFR's. Another possibility could be unmeasured student-level measures of wealth -- which may modify obesity risk -- but we are not able to account for in our analysis.
# Discussion {#sec:Discussion}
This work proposed a modeling approach to identify heterogeneity in distance-dependent BEF effects. By allowing flexibility both across space and identifying subgroups of subjects with different effects, this modeling framework addresses two problems raised in the built environment literature, namely the MAUP and the UGCP, respectively. The modeling approach was shown to work well in both simulated data, as well as the data that motivated this work, concerning children's BMI and proximity to FFRs near their schools. While spatial point pattern built environment data are the primary motivation for this methodology, it could be also be applied to temporal or spatio-temporal data, the latter which we discuss in greater detail below.
Similar to the HDLM proposed by, we seek to allow for differences across subjects, or other substantively defined groups like schools, in the BEF associations across space. In contrast to that work, we pool subjects with similar association effects through the DPM, allowing us to identify latent risk subject groups.
In simulations our model demonstrated classification robustness to differing distributions of distances and expected improvement in classification due to increased information through BEF exposure or effect size. Our analysis of Fitnessgram data illustrated how one can analyze these data in terms of the spatial effects estimated as well as the characteristics associated with each latent cluster. The software to fit this model and perform the necessary auxiliary functions is freely available through our R package `rstapDP`.
There are a number of future directions with which to take this work. One obvious direction would be to extend the modeling framework for more general exponential family error distributions, though this makes estimation more difficult, as the posterior distribution of \(\bm{\beta}\) is no longer available in closed form. Work by has used a Riemann Hamiltonian Monte Carlo sampler in this context for models similar to ours, without smooth functional terms. This could provide one avenue to pursue. Another direction to explore would be to incorporate temporally indexed BEF data to enable spatio-temporal function estimation via tensor product of the spline basis function expansion used here. This approach would allow for cluster estimates across space and time, increasing the dimensionality and consequently, relevancy, of this work to more precisely target and understand how environments shape health and health behaviors across both time and space.
Finally, while there have been numerous methods to identify associations between subjects and BEF exposure we believe this to be the first to utilize techniques in both the Bayesian and functional non-parametric literature to identify heterogeneous BEF effects across a population. | {'timestamp': '2021-07-14T02:07:46', 'yymm': '2107', 'arxiv_id': '2107.05805', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05805'} |
null | null |
null | null |
# Introduction {#intro}
Healthcare is an extremely promising area for applying machine learning solutions--clinical support systems can aid physicians with decision making, text-based solutions can help match patients to relevant clinical trials, and ML systems can save lives by identifying patterns indicative of sepsis, stroke, or cancer. However, the healthcare domain also presents immense legal and ethical considerations, and model mistakes can cost patient lives. Clinical stakeholders not only need to know that models have been faithfully developed, but they also desire transparency in the output of the model so that they can more easily diagnose model failures. A lack of explainability in healthcare, especially clinical decision support models, has been criticized, leading to additional innovation and progress in the development of explainable solutions. Historically, a trade-off between model performance and explainability has been described, and lower performing yet explainable models are often preferred to more sophisticated black box models due to lack of trust or the presence of strict transparency requirements (e.g. regulatory constraints). To address this, model-agnostic *post hoc* explainability methods such as LIME and SHAP were developed, where a second model is trained to explain the original predictions. However, there have been issues reported with these methods: they do not always faithfully describe the original model's behavior and can give different explanations for similar examples. Recent efforts, namely, call for the use of inherently interpretable models in situations where human understanding is necessary--especially when consequences from undue faith in model output can be dire. These methods generally produce sparse models which allow human stakeholders to gain understanding about the broader system in which the model operates, as well as understand precisely how the model produces each prediction. Natural language processing (NLP) presents a unique modeling and interpretability challenge. Many traditional methods of text preparation are transparent in nature, but as Transformer-based models have gained popularity, we continue to see the boundaries of black-box modeling pushed. This raises a pressing issue for data scientists and their stakeholders within the healthcare domain: how should we appropriately balance the need for well-performing models with the desire for human interpretability of models and their outputs? With such a wide variety of available tools, it should also be noted that the use of these techniques should be tailored to the audience--for example, data scientists may be interested in interactively exploring neuron activations, while a physician may be interested in the features present in a medical record that contributed to a clinical decision.
## Interpretability versus Explainability
In Rudin's work [-@article], she defines "interpretable\" machine learning as the use of models which themselves are fully interpretable by human users, as opposed to using "explainability\" techniques to understand the behavior of non-interpretable models. We follow these definitions throughout, although we acknowledge that there is some degree of gray area with certain attribution methods. Importantly, neither of these concepts are in any way causal--all of the methods in this paper should be used to gain insight into how a specific model works, rather than providing any understanding of underlying causal structures of the broader system in which the models operate.
# Experiment Setup {#setup}
## Data
Given the sensitive nature of healthcare data, few openly accessible datasets of clinical text exist for broad research use. In collaboration with Beth Israel Deaconess Medical Center, Johnson et al. introduced MIMIC-III [-@Johnson2016]--a database containing deidentified electronic medical record (EMR) output of Intensive Care Unit (ICU) stays, including structured data elements of medication and treatment as well as unstructured clinical text. Because MIMIC-III is the only publicly available data source of its type and scale, it frequently serves as the basis for peer reviewed studies which make use of EMR data. Here, we utilize the unstructured discharge summaries for our classification task.
## Classification Task
We selected one task in order to compare interpretable features across model types and explainability methods within the context of a practically relevant use-case. We use the single-label, binary classification of in-hospital mortality prediction from simulated admissions notes. This benchmark was established in Van Aken et al. [-@van-aken-etal-2021-clinical], and we follow their process of data preparation. Details about this task dataset are provided in supplementary material^[\[supfoot\]](#supfoot){reference-type="ref" reference="supfoot"}^.
# Methods
## Classifiers
We include several classifiers with varying levels of complexity and interpretability:
- Logistic Regression (LR)
- Random Forest (RF)
- Explainable Boosting Machine (EBM)
- DL8.5
- Boosted Rule Sets
- Bayesian Rule Lists
- Optimal Classification Trees
- CORELS
- BigBird
## Explainability
For our traditional models (LR, RF, EBM), we selected two popular *post hoc* explainability methods: Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). Since some of our models are also inherently interpretable (LR and EBM), we were also able to compare *post hoc* explanations to the true local explanations. Transformer models are a rapidly evolving area of active research, and efforts are ongoing to develop explainability techniques for better understanding their underlying processes. Input attribution methods show importance of input tokens, illustrating the degree to which each token contributed to a given prediction. Several algorithms exist for performing input attribution with deep learning models: saliency and integrated gradients are somewhat agnostic methods which use various backpropagation features to perform attribution, while other methods exist for more specific purposes (e.g. deconvolution and GradCAM for computer vision). In addition to these methods, SHAP and LIME are compatible for *post hoc* explainability with deep learning models--however, both have been shown to possess worse qualities than techniques with properties native to deep learning. Along with the growing trend of Transformer interpretability features, researchers have created a number of visualization techniques to highlight various behaviors (self-attention weights, activations, etc.). BertViz and Ecco[^1] are two libraries which provide useful front-ends for visualizing such behaviors within Transformer models, and LIT is a more general framework for practitioners to visualize and perform behavioral testing on various NLP models.
## Explanation Quality Metrics
With a variety of options available for input attribution and local explainability for "black-box\" models, a user will naturally wonder which method best represents the underlying mechanisms of the model. There are a handful of metrics which seek to quantify the quality of model explanations, however there is no single metric which will adequately cover all aspects of interpretability and apply to all contexts. For example, parsimony of the global explanations (i.e. the inherent human interpretability of a model's structure) may be more important when developing clinical guidelines, while fidelity of prediction explanations might adequately capture the needs of a human-in-the-loop ML system. For our case study, we searched for metrics which measure various components contributing to the quality of *post hoc* explanations provided through input attribution methods (e.g. SHAP, integrated gradients, etc.).
### Generalizing Local Lipschitz
Local Lipschitz is an explanation quality metric which measures the relative change in attributions introduced by relatively small perturbations in the model input. This metric is closely related to the notion of a particular explanation method's sensitivity when slight changes are introduced--if the input changes by some small amount, then we would presume that the explanations should not change drastically. The formula for local Lipschitz is given as follows: \[\hat{L}(x_i) = \argmax_{x_j \in B_\epsilon(x_i)} \frac{||f(x_i)-f(x_j)||_2}{||x_i-x_j||_2}\] One challenge came to light when applying this metric to text data: the concept of a *local neighborhood* is difficult to define within the context of raw text--we can measure the distance between vectorized documents through metrics such as \(\ell2\) distance or cosine similarity, but there is no consistent radius (\(\epsilon\) in the formula) which can be equivalently applied to models with different preparation types (e.g. TF-IDF document vectors and a sequence of word embeddings). We address this issue by introducing a text perturbation method based upon token replacement via nearest neighbor sampling. Pseudocode for our document perturbation method is given in Algorithm [\[alg:perturbation\]](#alg:perturbation){reference-type="ref" reference="alg:perturbation"}. Each token--a single word in a document to be vectorized, or a single token ID to be fed into a language model--has probability \(\pi\) of being selected for replacement, and each selected token is replaced by sampling from its \(k\) nearest neighbors in vector space.--in this description, `sample_neighbors` is simply shorthand for sampling a replacement token from its closest neighbors. For traditional methods we use a word2vec model trained on the mortality training set. For the BigBird model, we construct a distance matrix of tokens using pairwise \(\ell2\) distance between entries within the language model's word embedding layer.
:::
# Conclusions
In this work, we explored practical aspects of interpretable and explainable ML methods within a case study of clinical text classification. We demonstrated some of the interpretability tools available to healthcare NLP practitioners, discussed existing definitions for explainability and interpretability, and we introduced a framework which can evaluate the quality of explanations across text classification models, including the infidelity and relative sensitivity of the presented attributions. We showed that this evaluation framework can be used in conjunction with preferred model evaluation metrics to provide a Pareto frontier illustrating the trade-off between a model's test set performance and the quality of the explanation methods available to it. This can be a very useful tool for practitioners in many domains, including healthcare, in which the data scientist is tasked with building and identifying an optimal model based on the desired level of human interpretability, which varies from problem to problem. It will also be useful for developers of AutoML solutions--the introduction of quantitative metrics for explanation quality presents an opportunity to jointly optimize for such metrics in addition to raw model performance. We showed that the interpretable ML techniques which are not typically used for text classification perform suboptimally despite their interpretable properties. In some cases, a deterioration in model quality may be acceptable for the purpose of full model interpretability, although other methods for producing sparse models (e.g. LASSO regression) may provide better performance. Through direct comparison between LIME, SHAP, and true local explanations, we show that LIME falls short on such high-dimensional tasks, where sparse attributions do not necessarily reflect the true underlying behavior of the model. We also demonstrated current work in visualizing the underlying behavior of BERT-style models, which attempts to build progress toward a sense of interpretability for these black-box models.
## Future Work
Future work in both explainable and interpretable ML will undoubtedly be a crucial factor in the effort to accelerate adoption within the healthcare sector. These efforts should include additional development across the spectrum of interpretability in addition to the goal of ever-improving model performance. Progress in this area will be aided by the development and use of common evaluation metrics, which seek to quantify the quality of explanations provided for black-box models. The essence of this effort is to NLP practitioners in answering the question, "*How much* better or worse are the SHAP attributions for a random forest compared to saliency for a BERT classifier?\" We make progress in this direction by introducing and open-sourcing prototypical methods for preparation-agnostic document perturbation within local neighborhoods, but additional exploration and improvement will undoubtedly aid adoption. In addition to better understanding of explanation quality, future work in reducing the computational complexity of purely interpretable and provably optimal ML solutions will allow for more competitive use in large-scale text classification. | {'timestamp': '2021-07-14T02:02:39', 'yymm': '2107', 'arxiv_id': '2107.05693', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05693'} |
# Introduction
Plastic organic scintillators are durable with rapid rise and decay times, making them useful in a wide variety of applications in nuclear and particle physics, national security, and medical imaging. For fast neutron detection using these materials, the dominant mechanism for light generation is n-p elastic scattering, where energy is transferred to the scintillating medium via the recoiling proton. The proton light yield (PLY), or light emitted as a function of proton recoil energy, is a critical input in the determination of detection efficiencies for absolute calibration of neutron time-of-flight (TOF) spectra and for reconstruction in kinematic neutron imaging algorithms. This work provides a measurement of the PLY relation for the EJ-230, EJ-232, and EJ-232Q fast plastic organic scintillators developed by Eljen Technologies. These scintillators are composed of a polyvinyltoluene (PVT) solvent with one or more proprietary fluors, which give rise to the unique properties of each material listed in Table [\[tab:scintprops\]](#tab:scintprops){reference-type="ref" reference="tab:scintprops"}. For the EJ-232Q sample used in this work, 0.5% benzophenone was added to the medium as a quenching agent, resulting in an ultra-fast rise time and a sub-ns prompt decay constant. The temporal properties of these materials make them particularly attractive for use in prospective single-volume neutron scatter camera systems, where two n-p elastic scattering events are detected within a contiguous volume. In such an approach, the energy and angle of the incident neutron is determined by measuring the position, time, and energy deposition associated with the double-scatter event. Specifically, the energy of the incident neutron, \(E_n\), is determined via conservation of energy: \(E_n = E_n^{\prime}+E_p\). The scattered neutron energy, \(E_n^{\prime}\), is inferred from the position of and time-of-flight between interaction points in the contiguous volume, necessitating fast scintillators and excellent timing resolution. The proton recoil energy, \(E_p\), is obtained by converting the measured scintillation light into proton energy via the empirically-determined PLY relation. The scattering angle of the incident neutron is then determined from \(E_p\) and \(E_n^{\prime}\) via n-p elastic scattering kinematics. The PLY measurements obtained in this work facilitate the design and operation of single volume neutron scatter camera systems relying on EJ-230, EJ-232, or EJ-232Q as the detection medium. Indeed, recent work from Sweany et al. suggests EJ-230 as a top candidate for use in the construction of an optically segmented single-volume neutron scatter camera, based on both position and timing resolution performance in test scintillator bars with dual-ended readout.
This work also addresses a recent controversial claim in the literature that the relative PLY of a scintillator depends upon the size of the scintillating volume. Enqvist et al. measured the relative PLY of three EJ-309 organic scintillator cells of different sizes and asserted that the relative PLY of EJ-309 varied with scintillator size due to self-attenuation of the scintillation light. The relative PLY provides the light output for proton recoils relative to that for electron recoils and the attenuation of the scintillation light as it travels through a medium is the same regardless of the means by which the light was produced. Excluding edge effects, the impact of a change in scintillator size should not manifest as a change in the relative PLY, and rather should be accounted for in the light unit calibration. A size-dependent relative PLY would have significant consequences in any application of organic scintillators for neutron detection, as such an effect would require the PLY to be independently evaluated for a particular scintillator geometry. To investigate this issue, the relative proton light yield of two different sized cylindrical EJ-232 and EJ-232Q scintillators was measured. The experimental methods and setup are detailed in Sec. [2](#exp){reference-type="ref" reference="exp"}. Section [3](#analysis){reference-type="ref" reference="analysis"} describes the data analysis and uncertainty quantification procedures. In Sec. [4](#results){reference-type="ref" reference="results"}, the measured PLY relations of EJ-230, EJ-232, and EJ-232Q are reported and compared with literature measurements where available. The EJ-232 and EJ-232Q PLY relations for two different sized scintillators are reported and the implications for applications are addressed. Concluding remarks are provided in Sec. [5](#conc){reference-type="ref" reference="conc"}.
# Experimental Methods {#exp}
The double-time-of-flight (dTOF) method of Brown et al. was recently developed for characterizing the relative PLY of organic scintillators. In accordance with this approach, the scintillator of interest, i.e., the *target scintillator*, was placed in an intense neutron beam generated at the 88-Inch Cyclotron at Lawrence Berkeley National Laboratory. The cyclotron accelerated a 16 MeV deuteron beam onto a 3-mm-thick beryllium target, resulting in a high-flux and broad-spectrum neutron source. When neutrons elastically scatter off of protons in the organic scintillator, the recoiling protons excite the scintillation medium. The goal of this work is to associate the recoil energy of the proton with the light output in the target scintillator (i.e., the PLY). To accomplish this, the energy and angle of the scattered neutron are used to kinematically determine the recoiling proton energy. For this purpose, an array of eleven EJ-309 liquid organic scintillators, i.e., *scatter scintillators*, surrounded the target at fixed positions out-of-beam during the experiment. Using the pulse-shape discrimination (PSD) properties of the EJ-309 scintillators, neutron interactions were identified in the scatter scintillators on an event-by-event basis. The incoming neutron energy for each event was determined using the time-of-flight (TOF) between the cyclotron RF signal and the target scintillator. Similarly, the outgoing scattered neutron energy was calculated using the TOF between the target scintillator and the corresponding scatter scintillator. The proton recoil energy \(E_p\) was then deduced from the scattered-neutron energy \(E_n'\) and the known scattered-neutron angle relative to the neutron beam axis, \(\theta\), via n-p elastic scattering kinematics: \[\label{eq:out} E_p = E_n' \tan^{2}(\theta).\] By associating this recoil energy with the corresponding light response in the target scintillator for each event, a continuous measurement of the target scintillator PLY was obtained. Figure [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"} shows a diagram of the EJ-230 experimental setup, which is similar in layout to those employed in the EJ-232 and EJ-232Q experiments. The \(x\)-coordinate corresponded to the distance from the west wall of the experimental cave, the \(y\)-coordinate to the distance from beamline-center in the north direction, and the \(z\)-coordinate to the vertical distance from beamline-center. The detector positions in each experiment are detailed in Appendix [6](#app-exp){reference-type="ref" reference="app-exp"}. The target scintillators were right cylinders (5.08 cm dia.) with each end of each cylinder optically coupled to a photomultiplier tube (PMT) using BC-630 silicone grease. The EJ-230 scintillator was 5.08 cm in height, and cylinders of two different heights (2.54 cm and 5.08 cm) were measured for both the EJ-232 and EJ-232Q media. Each target scintillator was wrapped in at least 10 layers of Teflon tape to maximize the reflectivity of the detector housing. The scatter scintillators (5.08 cm dia. x 5.08 cm h.) were composed of right-cylindrical cells of thin aluminum housing filled with EJ-309 and coupled to a single PMT via a borosilicate glass window and BC-630 silicone grease. All PMTs used in these measurements were obtained from Hamamatsu Photonics (either Type No. 1949-50 or 1949-51), and were negatively biased using either a CAEN R1470ET or CAEN NDT1470 power supply. Further details on the target scintillator PMTs are provided in Table [\[pmtTable\]](#pmtTable){reference-type="ref" reference="pmtTable"} in Appendix [6](#app-exp){reference-type="ref" reference="app-exp"}. The PMT signals were connected to a CAEN V1730 500 MS/s digitizer. Full waveforms were recorded within a 670-ns acquisition window for all scintillator signals as well as for the cyclotron RF control signal. The data acquisition was triggered on a coincidence between both of the target scintillator PMT signals as well as at least one of the scatter scintillator signals within a 400 ns coincidence window. The CAEN digital constant fraction discrimination algorithm, with a 75% fraction and a 4 ns delay, was used to determine the arrival time of the scintillator signals. The timing pick-off for the cyclotron RF signal was accomplished using leading edge discrimination. All waveforms were written to disk (using the ROOT file format ) in list mode with global time stamps to enable event reconstruction
# Data Analysis {#analysis}
The raw PMT waveforms were processed using a custom C ROOT-based analysis framework. Integration of each individual waveform provided a measure of the light collected in a particular event. The integration lengths were chosen to collect at least 95% of the scintillation light in the 670 ns acquisition window. For the EJ-232Q scintillator, the integration length was 140 ns, while for both the EJ-230 and EJ-232 scintillators, the integration length was 180 ns. As in previous studies, a 300 ns integration length was used for the EJ-309 scatter cells. For the target scintillators, the pulse integrals for each of the target PMTs were combined using a geometric average to obtain an event-by-event light output measurement independent of interaction position. The linearity of the response of each target PMT was evaluated using a specialized pulsed-LED circuit based on the approach of Friend et al. (see Appendix [6](#app-exp){reference-type="ref" reference="app-exp"} for details). The light-output unit for the PLY relations presented in this work is defined relative to the light produced by a 477 keV recoil electron. Electron-equivalent light units, although common in the literature, implicitly assume electron light proportionality. Several recent studies have shown that this assumption is not valid for plastic organic scintillators in the energy range relevant for this measurement. The use of a light-output unit relative to the light produced by a 477 keV recoil electron ensures that the results presented here capture the behavior of the PLY without bias introduced by potential electron light yield nonlinearity Light-output calibrations for each target scintillator were performed with a \(^{137}\)Cs source using the Compton edge (477 keV) of the 662 keV \(\gamma\)-ray. This source was chosen for its strong mono-energetic line that lay within the dynamic range of this experiment. The Cs source was positioned at a distance of at least 5 cm from the target scintillator for all but the 2.54 cm EJ-232 sample, where it was taped directly to the face of the cylinder during calibration. The resulting pulse-integral spectra were fit using \(\chi^2\)-minimization to simulated deposited-energy spectra (convolved with a PMT-resolution function ) to determine the channel number corresponding to the Compton edge. An example light output calibration spectrum is shown for the 5.08 cm EJ-232Q scintillator in Fig. [\[fig:pulseint\]](#fig:pulseint){reference-type="ref" reference="fig:pulseint"}. The \(\chi^2\)-minimization was implemented using the SIMPLEX and MIGRAD algorithms from the ROOT MINUIT2 package. In both the light-output calibrations and the neutron-beam data, edge effects due to escaping recoil particles are negligible due to the small size of the recoil-particle ranges (approximately 350 \(\mu\)m for a 5 MeV proton) compared to the scintillator dimensions To obtain the proton recoil energy (as described in Eq. [\[eq:out\]](#eq:out){reference-type="ref" reference="eq:out"}), timing calibrations were performed for both the incoming and outgoing neutron TOF. The incoming neutron TOF, \(t_{\text{inc}}\), was calculated as follows: \[\label{eq:inccalc} t_{\text{inc}} = \frac{t_1 + t_2}{2}-t_{\text{RF}}-t_{\text{cal}} + (n \times T_{\text{cyc}}),\] where \(t_1\) and \(t_2\) are the signal arrival times in each of the two target scintillator PMTs, \(t_{\text{RF}}\) is the time of the cyclotron RF signal, \(t_{\text{cal}}\) is a time-calibration constant, \({T_\text{cyc}}\) is the cyclotron pulse period (158.5 ns), and \(n\) is an integer to account for the possibility that the incoming neutron was generated by different cyclotron pulses. The determination of \(n\) will be discussed later in this section. The determination of \(t_{\text{cal}}\) was accomplished using the photon flash produced by deuterons impinging on the beryllium target and the known flight-path between the beryllium target and the target scintillator. The signal from the prompt-\(\gamma\) flash was fit with a Gaussian distribution plus a linear background term. Given the speed of light and the known flight-path distance, the centroid of the fit was used to determine \(t_{\text{cal}}\). The width of the prompt-\(\gamma\) distribution varies between 2.5 and 7 ns depending on the quality of the beam tune and is a significant contributor to the incoming-neutron energy uncertainty Similarly, the outgoing neutron TOF was calibrated by examining time differences between \(\gamma\)-ray events in the target scintillator and each individual scatter cell. For scatter scintillator \(i\), the outgoing neutron TOF, \(t_{i,\text{out}}\), is given by: \[\label{eq:outcalc} t_{i,\text{out}} = t_{i}-\frac{t_1 + t_2}{2}-t_{i,\text{cal}},\] where \(t_{i}\) is the measured time of the signal from the \(i\)th scatter scintillator and \(t_{i,\text{cal}}\) is the corresponding time-calibration constant. After gating on \(\gamma-\gamma\) coincidences using PSD in each EJ-309 scatter cell, histograms were made of the coincident TOF for each target-scatter pair. Each time-calibration constant \(t_{i,\text{cal}}\) was determined from the centroid of the prominent peak in the corresponding histogram. As in the case of the incoming TOF, the width of the peak (on average approximately 400 ps) contributes to the scattered-neutron energy uncertainty As this setup overconstrains n-p elastic scattering kinematics in the target scintillator, the integer ambiguity, \(n\), in Eq. [\[eq:inccalc\]](#eq:inccalc){reference-type="ref" reference="eq:inccalc"} can be resolved. Using the outgoing neutron TOF and the scattering angle, the expected incoming neutron TOF was calculated on an event-by-event basis and compared to the measured incoming neutron TOF for different values of \(n\). All events in which the expected incoming TOF differed from the measured incoming TOF by more than 5% of the cyclotron period (7.925 ns) were discarded. This additional constraint removes contributions from multiple neutron interactions in the target scintillator, which represent a large potential contributor to systematic bias in edge characterization approaches to PLY measurement The proton recoil energy was determined for each event using Eq. [\[eq:out\]](#eq:out){reference-type="ref" reference="eq:out"}. Figure [\[fig:ly2d\]](#fig:ly2d){reference-type="ref" reference="fig:ly2d"} shows a histogram of the relationship between light output and proton recoil energy, with a clearly visible kinematic band. Events in this histogram were gated on the observation of neutrons in the scatter cells using PSD. Also, a maximum neutron energy limit was applied using the known incoming-beam energy and the detector configuration. Uncertainty in the proton recoil energy was calculated from a Geant4 simulation taking into account the uncertainty in the outgoing neutron TOF, the uncertainty in the flight-path length, and the finite size of the detectors (resulting in an angular variance of approximately 0.8 degrees). For each simulated event, the proton recoil energy was reconstructed with the simulated scattered neutron energy and Eq. [\[eq:out\]](#eq:out){reference-type="ref" reference="eq:out"}. This reconstructed proton recoil energy was then compared to the simulated proton recoil energy to calculate the effects of the aforementioned uncertainties on the energy determination. The resulting energy resolution is heteroskedastic, providing an uncertainty that varies as a function of proton recoil energy. The energy resolution is dominated by the angular spread of the detectors, as opposed to the TOF resolution. The binning for the \(x\)-axis was determined using the proton recoil energy resolution function. Each individual bin was projected onto the light-yield axis and fit using a binned maximum-likelihood estimation with a Gaussian distribution and a continuous piecewise power law. Figure [\[fig:slices\]](#fig:slices){reference-type="ref" reference="fig:slices"} shows several examples of these individual energy-bin fits. The uncertainty in the light output was calculated from the following sources: sensitivity to parameters in the light-yield analysis, the stability of the PMT gain over time, and uncertainty in the Compton-edge determination during the pulse-integral calibration. To determine the sensitivity of the light output to analysis parameters, Monte Carlo calculations were performed in which each parameter was varied by sampling from the probability distribution defined by the associated uncertainty. The data reduction was then repeated with these varied parameters to determine the influence on the extracted light-yield. The varied parameters were the incoming/outgoing neutron TOF calibration constants, the measured scatter-scintillator locations, the measured target-scintillator location, and the flight path from the beryllium target to the origin of the coordinate system. Although these parameters directly affect only the proton recoil energy calculation, the fixed binning structure leads to some events in each Monte Carlo trial migrating into adjacent bins as the parameters are varied. As such, the uncertainty in the proton recoil energy is translated to an uncertainty in the light yield as the event composition of each bin changes. To evaluate the PMT gain stability, subsets of the beam data were analyzed to evaluate the light output as a function of time and a variation of 1% was observed. To evaluate the uncertainty from the pulse-integral calibration, the systematic uncertainty in the Compton-edge determination was estimated to be 1% by varying the fitted region around the Compton edge, while the statisical uncertainty was negligibly small. In the case of the 2.54 cm EJ-232 sample for which the Cs source was taped directly to the face of the cylinder, an edge-determination uncertainty of 3% was obtained. This was informed by comparing two sets of calibration data for the 5.08 cm EJ-232 scintillator (one in which the source was taped to the face of the cylinder, and another in which the source was placed 5 cm away) and observing the shift in the Compton edge. Simulations using Geant4 were also performed to study the effect of nonuniform illumination of the cylindrical scintillator. The resulting uncertainties from these three sources were combined in quadrature to arrive at the final light-output uncertainty for each data point.
# Results and Discussion {#results}
For the 5.08 cm h. x 5.08 cm dia. cylindrical scintillator, the measured relative PLY relations of the EJ-230, EJ-232, and EJ-232Q scintillators are shown in Figure [\[fig:results\]](#fig:results){reference-type="ref" reference="fig:results"}, and the corresponding data are tabulated in Appendix [7](#LYapp){reference-type="ref" reference="LYapp"}. The three scintillating media have relative PLYs that are similar in shape and magnitude. The difference between any two of the three materials is within approximately two sigma from zero across the full range. As ionization quenching (i.e., quenching of the primary excitation by a high density of excited and ionized molecules) is a primary process occurring in the solvent or host material for solid organic systems, this consistency is expected given that the three materials all employ the same PVT polymer base. Figure [\[fig:pozziComp\]](#fig:pozziComp){reference-type="ref" reference="fig:pozziComp"} shows a comparison of the EJ-230 PLY obtained in this work to a measurement by Pozzi et al. of the PLY of BC-420 (an EJ-230 commercial equivalent). Pozzi et al. used a \(^{252}\)Cf source incident on several plastic scintillator detectors and obtained the PLY by quantifying the electron-equivalent light output corresponding to an inferred minimum neutron energy for a given detection threshold. There is a clear discrepancy between the two measurements, as the present work shows significantly higher relative light yield for several-MeV proton recoil energies and lower relative light yield below 1 MeV proton recoils. The dTOF technique used in this work enables a direct measurement of the relative PLY compared to the method used by Pozzi et al., which infers the proton energy from neutron efficiency curves without fully accounting for systematic uncertainties. The black triangles show a dTOF measurement by Laplace et al. of the PLY of EJ-204, a fast plastic organic scintillator from Eljen Technology that also employs a PVT solvent. The EJ-230 result obtained in this work is consistent with the PLY of the EJ-204 scintillator within the estimated uncertainties. A recent study by Enqvist et al. of the EJ-309 organic scintillator suggested that the relative PLY varies depending on the scintillator cell size due to attenuation of the scintillation light. To test this conjecture, PLY measurements of cylindrical EJ-232 and EJ-232Q cells with heights of both 2.54 cm and 5.08 cm were performed. If self-attenuation resulted in size-dependence of the relative PLY (as Enqvist et al. argue), then such an effect would be particularly noticeable in EJ-232 and EJ-232Q given the short optical attenuation lengths of the materials (\(\sim\)`<!-- -->`{=html}17 cm and \(\sim\)`<!-- -->`{=html}6 cm, respectively). Figures [\[fig:comp_232\]](#fig:comp_232){reference-type="ref" reference="fig:comp_232"} and [\[fig:comp_232Q\]](#fig:comp_232Q){reference-type="ref" reference="fig:comp_232Q"} show comparisons of the relative PLY results for the 2.54 cm and 5.08 cm scintillator sizes for EJ-232 and EJ-232Q, respectively. In both cases, the relative PLY measurements for the two cell sizes agree within the estimated uncertainties. The lack of a statistically significant difference between the data obtained from the two different cell sizes presented here is consistent with basic physical intuition of the scintillation process. Attenuation occurs for scintillation light produced by both electron and proton recoils, and therefore should not affect the *relative* light yield. Recently, relative PLY measurements obtained with respect to the crystalline axes of multiple-sized stilbene samples were performed using a coincident kinematic technique. These measurements also showed no size-dependence.
# Summary {#conc}
The relative PLYs of EJ-230, EJ-232, and EJ-232Q were measured using a dTOF technique for proton recoil energies from approximately 300 keV to 4 MeV. The three resulting PLY relations are similar to each other, as well as to recent measurements of EJ-200, EJ-204, and EJ-208. This is reasonable considering that ionization quenching is a primary process and these scintillators share the same PVT base material. The EJ-230 result presented here disagrees significantly from the only previous literature measurement (which was of the commercial equivalent BC-420). Two different cylindrical cell sizes were measured for the EJ-232 and EJ-232Q scintillators, and in both cases the PLY was found to be consistent within uncertainty between the two sizes. The proton light yield relations provided herein support continued development of kinematic neutron-imaging systems, in particular for applications that require excellent timing resolution.
# Experimental Details {#app-exp}
The detector positions for the measurements of EJ-230, EJ-232, and EJ-232Q are listed in Tables [\[locationsEJ230\]](#locationsEJ230){reference-type="ref" reference="locationsEJ230"}, [\[locationsEJ232\]](#locationsEJ232){reference-type="ref" reference="locationsEJ232"}, and [\[locationsEJ232Q\]](#locationsEJ232Q){reference-type="ref" reference="locationsEJ232Q"}, respectively. The origin of the coordinate system is defined at the beamline center along the west wall of the experimental area (see Fig. [\[fig:schematic\]](#fig:schematic){reference-type="ref" reference="fig:schematic"}). The \(x\)-coordinate was taken as the distance from the west wall, the \(y\)-coordinate as the distance north from the beamline center, and the \(z\)-coordinate as the distance above and below the beamline center. The flight path between the beryllium target and the origin of the coordinate system was \(646.9~\pm~0.5\) cm along the \(x\)-axis. The PMT model numbers and bias voltages used for the target scintillators are detailed in Table [\[pmtTable\]](#pmtTable){reference-type="ref" reference="pmtTable"}. Using the method of Friend et al., the linearity of the responses of the target PMTs were evaluated and shown to be linear with one exception. One of the PMTs coupled to the 2.54 cm EJ-232Q scintillator (Hamamatsu H1949-50 PMT biased at 2050 V) exhibited a nonlinear response and waveforms were corrected on a per sample basis. This PMT was several years older than the other H1949-50 PMTs, and the manufacturer confirmed that variation in linearity is to be expected across different PMTs of the same model.
# Proton Light Yield Data {#LYapp}
The relative PLY data for EJ-230, EJ-232, and EJ-232Q obtained using integration lengths of 180, 180, and 140 ns, respectively, are summarized in Tables [\[resultsTable230\]](#resultsTable230){reference-type="ref" reference="resultsTable230"}, [\[resultsTable232\]](#resultsTable232){reference-type="ref" reference="resultsTable232"}, and [\[resultsTable232Q\]](#resultsTable232Q){reference-type="ref" reference="resultsTable232Q"}. | {'timestamp': '2021-07-14T02:04:00', 'yymm': '2107', 'arxiv_id': '2107.05725', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05725'} |
null | null |
null | null |
# Introduction
The idea of an information bottleneck (IB) is to learn a compressed representation \(Z\) of an input \(X\) that is predictive of a target \(Y\). This leads to the following training objective involving two mutual information terms: \[\label{eq:1} \min_{Z}-I(Z, Y) + \beta I(Z, X).\] This objective favours a representation \(Z\) that retains the minimum amount of information about \(X\) while being maximally predictive of \(Y\). The hyper-parameter \(\beta \ge 0\) controls the trade-off between the two losses. To make this objective practical, used variational techniques to construct the upper bound for the expression in Eq. [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"}--also known as the Variational Information Bottleneck (VIB) loss: \[\label{eq:2} \min_{p(\bm z| \bm x)} \mathbb E_{p(\bm x,\bm y)p(\bm z|\bm x)} \big[-\log q(\bm y | \bm z) + \beta \log \frac{p(\bm z|\bm x)}{q(\bm z)}\big],\] where \(p(\bm z|\bm x)\) is a stochastic encoder distribution, \(q(\bm y|\bm z)\) is a variational approximation to \(p(\bm y|\bm z)\), and \(q(\bm z)\) is the variational approximation to the marginal \(p(\bm z)\). In a similar way to the Variational Auto-Encoder (VAE) setup , we can parameterize Gaussian densities \(p(\bm z|\bm x)\) and \(q(\bm y|\bm z)\) using neural networks, and fix \(q(\bm z)\) to be a \(K\)-dimensional Gaussian \(\mathcal N(\bm 0, \bm I)\), where \(K\) is the size of the bottleneck layer. We can then use the reparameterization trick to learn the parameters of the neural networks when optimizing the stochastic estimate of the objective in Eq. [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}. A tighter bound on the IB objective is given by the Conditional Entropy Bottleneck (CEB) : \[\label{eq:3} \min_{p(\bm z|\bm x)} \mathbb E_{p(\bm x,\bm y)p(\bm z|\bm x)} \big[-\log q(\bm y | \bm z) + e^{-\rho} \log \frac{p(\bm z|\bm x)}{q(\bm z|\bm y)}\big],\] where the second term uses a class-conditional variational marginal \(q(\bm z|\bm y)\), and \(\rho\) is a hyper-parameter with the same role as \(\beta\) in Eq.[\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}. CEB parameterizes \(q(\bm z|\bm y)\) by a linear mapping that takes a one-hot label \(\bm y\) as input and outputs a vector \(\bm \mu_y\) representing the mean of the Gaussian \(q(\bm z|\bm y) = \mathcal N(\bm \mu_y, \bm I)\). CEB uses an identity matrix for the variance of \(q(\bm z|\bm y)\) and \(p(\bm z| \bm x)\), which is unlike VIB, where the variance of the encoder distribution is not fixed. Multiple studies suggest that IBs can reduce overfitting and improve robustness to adversarial attacks . For example, showed that CEB models can outperform adversarially trained models under both \(l_{\infty}\) and \(l_2\) PGD attacks while also incurring no drop in standard accuracy. However, no clear explanation has been found as to how IB models become more robust to adversarial examples. Previous works also failed to investigate possible effects of gradient obfuscation which could lead to a false sense of security . In this paper, we continue the analysis into the behaviour of IB models in the context of adversarial robustness. Our experiments provide evidence of gradient obfuscation, which leads us to conclude that the adversarial robustness of IB models was previously overestimated.
# Adversarial robustness
Since the discovery of adversarial examples for neural networks , there has been a lot of interest in creating new attacks and defenses. In this section we briefly review methods for crafting norm-bounded adversarial examples. Later, we use these methods to assess the adversarial robustness of IB models. The Fast Gradient Sign (FGS) attack is an \(l_{\infty}\) bounded single-step attack that computes an adversarial example \(\bm x_{adv}\) as \(\bm x + \epsilon \text{sign}(\nabla_{\bm x} \mathcal L(\bm \theta, y ,\bm x))\), where \(\bm x\) is the original image, \(y\) is the true label, \(\mathcal L\) is the cross-entropy loss, and \(\epsilon\) is the perturbation size. Projected Gradient Descent (PGD) is the multi-step variant of FGS. The \(l_{\infty}\) PGD attack finds an adversarial example by following iterative updates \({\bm x^{t+1} = \text{Proj}_{\mathcal B(\bm x, \epsilon)} \big(\bm x^t + \alpha \text{sign}(\nabla_{\bm x} \mathcal L(\bm \theta, y ,\bm x))\big)}\) for some fixed number of steps \(T\). Here, \(\text{Proj}_{\mathcal B(\epsilon, \bm x)}\) is a projection operator onto \(\mathcal B(\epsilon, \bm x)\) --the \(l_{\infty}\) ball of radius \(\epsilon\) around the original image \(\bm x\). The attack starts from an initial \(\bm x^0\) sampled randomly within \(\mathcal B(\epsilon, \bm x)\). The reliability of PGD attacks often depends on the choice of parameters such as the step size \(\alpha\), or the type of loss \(\mathcal L\). Recent PGD variants are designed to be less sensitive to these choices, and it is common to run an ensemble of attacks with different parameters and properties. AutoAttack and MultiTargeted are examples of this strategy.
# Experiments
In this section, we experiment with VIB and CEB models on MNIST and CIFAR-10. We run a number of diagnostics, which indicate that gradient obfuscation is the main reason why IB models are seemingly robust. In trying to understand their failure modes, we also look at some toy problems. Our interpretation of the results is deferred to the next section. Hyperparameters of all our models and additional plots are included in the appendix.
## MNIST {#sec:mnist}
For VIB experiments on MNIST, we follow the setup of. Namely, for the encoder network, we use a 3-layer MLP with the last bottleneck layer of size \(K=256\). This bottleneck layer outputs the \(K\) means and \(K\) standard deviations (after a softplus transformation) of the Gaussian \(p(\bm z|\bm x)\). The decoder distribution \(q(\bm y|\bm z)\) over 10 classes is parameterized by a linear layer ending with a softmax. During training, we use the reparameterization trick with \(S=12\) samples from the encoder \(\bm z \sim p(\bm z | \bm x)\) when estimating the expectation over \(p(\bm z | \bm x)\) in Eq. [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}. At test time, we also collect \(S\) samples of \(\bm z\), and compute \(p(\bm y|\bm x)\) as \(\frac{1}{S}\sum_{s=1}^{S} q(\bm y|\bm z^s)\). We refer to this evaluation as . In the , we only use the mean of \(p(\bm z|\bm x)\) as an input to the decoder. Our deterministic baseline is an MLP of the same overall structure as the VIB model. We train it with a cross-entropy loss without any additional regularization. First, we evaluate our models using the FGS attack. Figure [\[fig:0\]](#fig:0){reference-type="ref" reference="fig:0"} shows the robust accuracy of VIB models with varying \(\beta\) under the FGS attack with different perturbation sizes \(\epsilon\). For the rest of the paper, we assume that input images are in the \([0,1]\) range. Our results slightly differ from those of. In particular, the performance of our VIB models peaks at \(\beta=0.01\) instead of \(\beta=0.1\) as reported previously, and the evaluation in the and does not lead to the same results. Despite these differences, we can still achieve large gains in robust accuracy under the FGS attack for VIB models in comparison to the baseline. One result that stands out is the unusually high robust accuracy under the attack with \(\epsilon=0.5\). Indeed, with this perturbation size, one can design an attack that makes all images solid gray and, as such, the classifier should not do better than random guessing . The obtained robust accuracy above 10% indicates that gradients of VIB models do not always direct us towards stronger adversarial examples. To check if the improvements in robust accuracy generalize to stronger attacks, we evaluate VIB models with \(\beta=0.01\) under the PGD attack with 40 steps, \(\alpha=0.01\), \(\epsilon=0.2\), and a different number of restarts. Figure [\[fig:1a\]](#fig:1a){reference-type="ref" reference="fig:1a"} shows that we can drive the robust accuracy to zero as we increase the number of restarts. It is an indication of gradient obfuscation, as the loss landscape cannot be efficiently explored by gradient-based methods .
## CIFAR-10 {#sec:cifar}
For CIFAR-10, as our encoder network we use a PreActivation-ResNet18 followed by an MLP with the same architecture as the MNIST experiments. We train this network end-to-end, and only use random crops and flips to augment the data. As previously, we construct an analogous deterministic model that we do not regularize in any way, thus it overfits. In Figure [\[fig:2a\]](#fig:2a){reference-type="ref" reference="fig:2a"}, we evaluate the adversarial robustness of CEB models on \(l_{\infty}\) PGD attack with 20 steps and \({\alpha=0.007}\) . It is surprising that some of our deterministic models can outperform an adversarially-trained ResNet from with a reported robust accuracy of 45.8%. This result alone suggests that PGD attacks should be used with caution when evaluating models that might obfuscate the gradients. As with MNIST, we can again significantly reduce the robust accuracy by increasing the number of restarts as shown in Figure [\[fig:2b\]](#fig:2b){reference-type="ref" reference="fig:2b"}. To get a better estimate of the robust accuracy in the presence of gradient obfuscation, we use a set of stronger attacks: a mixture of AutoAttack (AA) and MultiTargeted (MT) . We execute the following sequence of attacks: AutoPGD on the cross-entropy loss with 5 restarts and 100 steps, AutoPGD on the difference of logits ratio loss with 5 restarts and 100 steps, MultiTargeted on the margin loss with 10 restarts and 200 steps. From Figure [\[fig:2c\]](#fig:2c){reference-type="ref" reference="fig:2c"}, we see that deterministic models have zero robust accuracy, while the performance of CEB models varies across models with different random seeds. This dependence on the seed could be the consequence of sub-optimal network initialization and difficulties related to training IB models. Some part of the variance in the robust accuracy might still be attributed to having an imperfect attack due to the unreliable gradients. Finally, in Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} and in the appendix, we show typical loss landscapes produced by the CEB model with \(\rho=2\) that scored 15.8% accuracy under the AA+MT ensemble of attacks. These plots are strikingly different from typical smooth non-flat loss landscapes obtained from adversarially trained models . The flatness of the plotted landscapes explains why gradient-based attacks with cross-entropy loss are not as effective. Moreover, since IB losses do not explicitly penalize misclassification for perturbed inputs within a certain \(l_p\)-ball, the model is free to choose where to place decision boundaries. Figure [\[fig:5\]](#fig:5){reference-type="ref" reference="fig:5"} suggests that CEB models could be robust to much smaller perturbation radii.
## A toy problem {#sec:toy}
We established that gradient obfuscation makes it harder to understand the robustness properties of IB models on real datasets. Thus, analysing toy examples can be a useful alternative. A classification task from is one example that can motivate the use of IBs, where their ability to ignore irrelevant features becomes helpful. We study this problem in the appendix. Here, we consider another simple setup where labels \(y\) are sampled uniformly at random from \(\{-1,1\}\), and two features have the following conditional distributions: \[\resizebox{\hsize}{!}{\( p(x_1|y=1) = \mathcal U(0,10), \; p(x_2|y=1) = \begin{cases} \mathcal U(0,1) \text{ w.p. } 0.9\\ \mathcal U(-1,0)\text{ w.p. } 0.1 \end{cases} \)}\] \[\resizebox{\hsize}{!}{\( p(x_1|y=-1) = \mathcal U(-10, 0), \; p(x_2|y=-1) = \begin{cases} \mathcal U(-1,0) \text{ w.p. } 0.9\\ \mathcal U(0,1)\text{ w.p. } 0.1 \end{cases} \)}\] In this example, the label can be predicted from the sign of \(x_1\), so in the optimal IB case, we need to communicate 1 bit of information about the input. The first feature is also more robust since it requires a larger perturbation before its sign gets flipped. In practice, we found that a simple VIB classifier does not exclusively focus on \(x_1\), and so it becomes prone to a rather trivial attack that substracts or adds \(\epsilon=1\) to \(x_2\) depending on the label, as shown in Figure [\[fig:toy\]](#fig:toy){reference-type="ref" reference="fig:toy"}. This could be the consequence of SGD training, the approximate nature of the objective function, VIB's formulation as a combination of competing objectives or other reasons we do not yet understand.
# Discussion
By re-evaluating adversarial robustness of VIB and CEB models, we have shown that weak adversarial attacks are often unable to provide reliable robustness estimates as these models create highly non-smooth loss surfaces, which are harder to explore with gradients. Therefore, we believe that previous, as well as future results on the robustness of IB models should include basic checks for gradient obfuscation. This is especially important when comparing different types of models, e.g. IBs versus adversarial training. Our experiments were inconclusive as to whether IB models offer adversarial robustness gains relative to the undefended deterministic baseline. For MNIST, the results under the FGS attack seemed promising. However, looking at the performance under the PGD attack with multiple restarts and different perturbation sizes showed a different picture. For CIFAR-10, some of the CEB models were significantly better than the baseline under the strongest attack. However, we did not identify the exact cause for having excessive variance in the results of models with different random seeds. Thus, it would be interesting to find regimes where CEB can reliably converge to more robust models. In this paper, we only considered IB models in discriminative settings. A generative model related to VIB is \(\beta\)-VAE . For auto-encoders, the adversarial attack amounts to finding inputs that would cause the decoder to reconstruct a visually distinct image, e.g. an object from a different class. showed that \(\beta\)-VAE for larger values of \(\beta\) is more robust to adversarial attacks. However, used a different set of evaluation metrics to challenge this claim. attributes the lack of robustness of \(\beta\)-VAE models to the inability of their objective to control the behaviour of the encoder outside of the support of the empirical data distribution. Namely, without additionally forcing the encoder to be smooth, tuning \(\beta\) alone is not enough for learning robust representations. Together with our observations for VIB and CEB models, the disagreement about \(\beta\)-VAE's results corroborates the need for more nuanced evaluation before adversarial robustness claims can be made. Overall, we believe that using IBs in the context of adversarial robustness is an idea that deserves further exploration. In this paper, we focused on the empirical evaluation of IB models under standard robustness metrics and illustrating the caveats related to it. An interesting future research direction would be to understand the properties of IB models, especially in the stochastic regime, from both information-theoretic and adversarial robustness perspectives. Another promising direction would be to explore IBs with additional curvature regularization or in combination with adversarial training. | {'timestamp': '2021-07-14T02:03:33', 'yymm': '2107', 'arxiv_id': '2107.05712', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05712'} |
# Introduction {#Introduction}
If one fixes the length of each edge of a triangle, or more generally of an \(n\)-dimensional simplex, then the geometry of the triangle (simplex) is completely determined: no vertex can be moved without changing the length of some adjacent edge. Contrast this to a square, for example, where two adjacent vertices can be shifted in the same direction while not changing the lengths of any edges. Thus, if \(\tau\) is an \(n\)-simplex and \(g\) is a metric on \(\tau\) with constant curvature, then the geometry of \((\tau,g)\) is completely determined by just the edge lengths that \(g\) assigns to \(\tau\). What we mean in the title by "synthetic geometry\" is formulas/procedures to compute geometric quantities of \((\tau,g)\) using only the edge lengths of \(\tau\), and without needing to isometrically embed \((\tau,g)\) into the appropriate model space and use explicit coordinates. The study of *Euclidean simplices*, or simplices \((\tau,g)\) where \(g\) has constant curvature zero, has a long history. Heron's formula, which computes the area of a triangle using its edge lengths, is an example of synthetic geometry in a Euclidean 2-simplex which dates back thousands of years. The study of higher dimensional simplices goes back (at least) to Menger in and Cayley in. More recently, the second author in developed a more efficient technique to encode the geometry of a Euclidean \(n\)-simplex using the edge lengths of the simplex (this formula is also developed in ). The techniques and results in will be summarized in Section [2](#section:Euclidean){reference-type="ref" reference="section:Euclidean"}, as they will play a vital role in the work done in this paper. Hyperbolic simplices, or metric simplices \((\tau,g)\) where \(g\) has constant curvature-1, play an important role in the current mathematical zeitgeist. For example, distances in hyperbolic triangles considering only the edge lengths of that triangle are used to determine whether a geodesic metric space is CAT(-1) (or an Alexandrov space with curvature bounded below by \(-1\)). Hyperbolic structures are used by Charney and Davis in for their strict hyperbolization, which are then further used by Ontaneda in in his smooth Riemannian hyperbolization. The goal of this research is to develop synthetic geometric formulas for simplices with constant curvature. This, together with, establishes a foundation on which further research may be more easily performed due to the formulas' relative simplicity in contrast to previous work in the literature. We will focus primarily on *hyperbolic simplices* due to their mathematical importance, but the last Section will discuss analogous results for *spherical simplices* (constant curvature 1), and more generally for simplices of constant curvature \(\kappa\). The main results of this paper are as follows:
1. We develop a simple criterion which determines whether or not a set of positive edge lengths for an \(n\)-simplex determine a legitimate hyperbolic simplex. (Section [3](#section:realizability){reference-type="ref" reference="section:realizability"}, Theorem [\[theorem: hyperbolic-realizability\]](#theorem: hyperbolic-realizability){reference-type="ref" reference="theorem: hyperbolic-realizability"}).
2. Given a hyperbolic simplex \((\tau,g_{\mathbb{H}})\) and two points \(x, y \in \tau\), we determine an easy procedure to find \(d_{\mathbb{H}}(x,y)\) using only the edge lengths of \(\tau\) and the barycentric coordinates of \(x\) and \(y\). (Section [5](#section:distances){reference-type="ref" reference="section:distances"}, Theorem [\[thm:hyp dist\]](#thm:hyp dist){reference-type="ref" reference="thm:hyp dist"}).
3. Given a Euclidean \(n\)-simplex \((\tau,g_{\mathbb{E}})\) where \(\tau = \langle v_1, v_2, \hdots, v_n, v_{n+1} \rangle\), we develop a formula for \(\text{proj}_{\tau_{n+1}}(v_{n+1})\), the orthogonal projection of \(v_{n+1}\) onto the \((n-1)\)-face opposite of it (denoted by \(\tau_{n+1}\)). (Section [6](#section:Euclidean proj){reference-type="ref" reference="section:Euclidean proj"}, Theorem [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"}).
4. Given a hyperbolic \(n\)-simplex \((\tau,g_{\mathbb{H}})\) where \(\tau = \langle v_1, v_2, \hdots, v_n, v_{n+1} \rangle\), we develop a formula for \(\text{proj}_{\tau_{1}}^{\mathbb{H}}(v_{1})\). (Section [7](#section:hyperbolic proj){reference-type="ref" reference="section:hyperbolic proj"}, Theorem [\[theorem:hyperbolic-projection\]](#theorem:hyperbolic-projection){reference-type="ref" reference="theorem:hyperbolic-projection"}).
The goal of all of this work is to develop formulas which are simple to use and reasonably intuitive. Toward this, in Section [8](#section:example){reference-type="ref" reference="section:example"} we work out an example with a 3-simplex in which we use all of the formulas mentioned above (and some of the formulas in ). The hope is that this example will be useful for any researchers who wish to use these formulas in the future. Lastly, there are more difficult formulas in the literature for some of the quantities listed above. In, Karliğa, Savas, and Yakut give a formula for orthogonal projection in hyperbolic space (Theorem 3 in ). As one can see, our formula is considerably simpler, and in any case their formula uses outward normal vectors and is therefore not a truly "synthetic\" formula. Also, in, Karliğa provides necessary and sufficient conditions for when a collection of edge lengths yields a legitimate hyperbolic simplex. But again, one can see that our necessary and sufficient condition listed in Section [3](#section:realizability){reference-type="ref" reference="section:realizability"} is more natural, and simpler to use. Also, just before submitting this paper to the arXiv, Abrosimov and Vuong posted the article which gives a geometric version for our Theorem [\[theorem: hyperbolic-realizability\]](#theorem: hyperbolic-realizability){reference-type="ref" reference="theorem: hyperbolic-realizability"} when \(n=3\).
# Notation and Formulas in Euclidean Simplices {#section:Euclidean}
Let us first establish some notation for the remainder of the paper. Let \(\tau = \langle v_1, v_2, \hdots, v_n, v_{n+1} \rangle\) be an \(n\)-dimensional simplex, and let \(g\) be a Riemannian metric on \(\tau\) with constant curvature. The notation \(g_\mathbb{E}\) means that \(g\) has constant curvature \(0\), or is *Euclidean*; the notation \(g_\mathbb{H}\) implies that \(g\) has curvature \(-1\), or is *hyperbolic*; and the notation \(g_\mathbb{S}\) means that \(g\) has curvature \(1\), or is *spherical*. Let \(e_{ij}\) denote the edge of \(\tau\) adjacent to the vertices \(v_i\) and \(v_j\), and let \(\gamma_{ij}\) denote the length of \(e_{ij}\) under \(g\), also denoted by \(g(e_{ij})\). We make the convention that \(\gamma_{ii} = 0\) for all \(i\). Denote the determinant of a matrix \(M\) by \(|M|\), and let \(M_{ij}\) denote the \(ij\)-th minor of \(M\) (that is, \(M_{ij}\) equals the determinant of the matrix obtained by removing the \(i^{th}\) row and \(j^{th}\) column of \(M\)). The purpose of this Section is to quickly summarize and explain the main results from to be used in this paper. Suppose a Euclidean \(n\)-simplex \((\tau,g_{\mathbb{E}})\) with \(\tau = \langle v_1, \hdots, v_{n+1} \rangle\) is linearly isometrically embedded into \(\mathbb{R}^m\) (\(m \geq n\)) endowed with some symmetric bilinear form \(\langle, \rangle\). Translate the image of \(v_{n+1}\) so that it is mapped to the origin, and by abuse of notation we associate each vertex \(v_i\) with its image in \(\mathbb{R}^m\). Define the vectors \(w_i = v_i-v_{n+1}\) for \(1 \leq i \leq n\). The collection \(( w_1, \hdots, w_n )\) forms a basis for the smallest subspace of \(\mathbb{R}^m\) containing \(\tau\), and the form \(\langle, \rangle\) is completely determined on this subspace by the \(n \times n\) matrix \(Q\) whose \(ij^{th}\) entry is defined by \[q_{ij} = \langle w_i, w_j \rangle.\] Now, notice that \[\gamma_{ij}^2 = \langle w_i-w_j, w_i-w_j \rangle = \gamma_{i,n+1}^2 + \gamma_{j,n+1}^2-2 \langle w_i, w_j \rangle\] and so \[\label{eqn:Q} q_{ij} = \langle w_i, w_j \rangle = \frac{1}{2} \left( \gamma_{i,n+1}^2 + \gamma_{j,n+1}^2-\gamma_{ij}^2 \right).\] This allows one to construct the matrix \(Q\) using only the edge lengths of \((\tau,g_\mathbb{E})\). Then, theoretically, one should be able to calculate any geometric quantity of \(\tau\) using \(Q\) since it completely determines the geometry of \((\tau,g_\mathbb{E})\). The main results from which follow from the definition of \(Q\) above are:
1. **Realizability of \((\tau,g_{\mathbb{E}})\):** \(n(n+1)/2\) positive real numbers \(\{ \gamma_{ij} \}\) are the edge lengths of some Euclidean simplex \((\tau, g_{\mathbb{E}})\) if and only if the matrix \(Q\) is positive definite.
2. **Distances in \((\tau,g_{\mathbb{E}})\):** Let \(x, y \in \tau\) be such that \(x\) has barycentric coordinates \((\alpha_i)_{i=1}^{n+1}\) and \(y\) has barycentric coordinates \((\beta_i)_{i=1}^{n+1}\), where we have \(\sum_{i=1}^{n+1} \alpha_i = 1 = \sum_{i=1}^{n+1} \beta_i\). Then the squared Euclidean distance between \(x\) and \(y\) can be calculated by the formula \[\label{euclid-dist} d_\mathbb{E}^2(x,y) = [x-y]^T Q [x-y]\] where \([x-y]\) is the vector in \(\mathbb{R}^m\) whose \(i^{th}\) coordinate is \((\alpha_i-\beta_i)\), with \(1 \leq i \leq n\).
3. **Volume of \((\tau, g_{\mathbb{E}}):\)** The \(n\)-dimensional volume of \((\tau, g_{\mathbb{E}})\) is given by \[\text{Vol}(\tau) = \frac{\sqrt{\text{det}(Q)}}{n!}.\]
# Determining the realizability of hyperbolic simplices {#section:realizability}
## The hyperboloid model for hyperbolic space {#the-hyperboloid-model-for-hyperbolic-space .unnumbered}
The goal of this research is to develop geometric formulas independent of the embedding of our hyperbolic simplex \((\tau, g_\mathbb{H})\) into hyperbolic space. But we will need to isometrically embed our hyperbolic simplex into some model space for hyperbolic space in order to prove that our formulas are correct. We will always use the hyperboloid model for hyperbolic space, and so we establish our notation for this now. We will use the notation \(\mathbb{R}^{n,1}\) to denote the standard Minkowski space with signature \((n,1)\). That is, as a vector space, \(\mathbb{R}^{n,1}\) is just \(\mathbb{R}^{n+1}\) endowed with the symmetric bilinear form \[\langle x,y \rangle = x_1y_1 + x_2y_2 + \hdots + x_ny_n-x_{n+1}y_{n+1},\] where \(x = (x_1, \hdots, x_{n+1})\) and \(y = (y_1, \hdots, y_{n+1})\). The solution set to the equation \(\langle x, x \rangle =-1\) forms a two-sheeted hyperboloid, and the "upper\" sheet (the sheet with \(x_{n+1} > 0\)) is our model for \(n\)-dimensional hyperbolic space \(\mathbb{H}^n\). Given two points \(x, y \in \mathbb{H}^n\), the hyperbolic distance \(d_\mathbb{H}(x,y)\) is given by \[\label{eqn:hyperbolic distance} d_\mathbb{H}(x,y) = \text{arccosh}(-\langle x,y \rangle)\] and is equivalent to the induced path-metric on \(\mathbb{H}^n\).
## Induced flat simplices {#induced-flat-simplices .unnumbered}
Let \((\tau, g_\mathbb{H})\) be an \(n\)-dimensional hyperbolic simplex, and assume that it is isometrically embedded in \(\mathbb{H}^n\). Let \(\tau = \langle v_1, \hdots, v_{n+1} \rangle\), and by abuse of notation associate \(v_i\) with its image in \(\mathbb{H}^n\). Identifying \(\mathbb{H}^n\) with the upper-half plane model described above, we can consider the convex hull of the vertices \(v_1, \hdots, v_{n+1}\) in \(\mathbb{R}^{n,1}\). This yields an \(n\)-dimensional simplex which we will call \(\sigma\). Note that \(\sigma\), endowed with the quadratic form inherited from \(\mathbb{R}^{n,1}\), has curvature \(0\).
In many of the formulas and arguments later in this paper, we will care about the \((n+1)\)-dimensional simplex \(\Sigma\) defined as follows. Assume we have a hyperbolic simplex \((\tau, g_\mathbb{H})\) isometrically embedded in \(\mathbb{H}^n\), and define \(\sigma\) as above. Then \(\Sigma := \{ \vec{0} \} \vee \sigma\). That is, \(\Sigma = (v_0, v_1, \hdots, v_{n+1})\) where \(v_0 = \vec{0}\) and \(v_1, \hdots, v_{n+1}\) are the vertices of \(\sigma\). So \(\Sigma\) is just the \((n+1)\)-dimensional simplex in \(\mathbb{R}^{n,1}\) obtained by combining the origin with the vertices of \(\sigma\). Using similar notation to Section [2](#section:Euclidean){reference-type="ref" reference="section:Euclidean"}, we can compute a simple formula for the matrix \(Q_\Sigma\) associated to \(\Sigma\). Let \(w_i = v_i-v_0 = v_i\). The collection \(( w_i )\) forms a basis for \(\mathbb{R}^{n,1}\). With respect to this basis, the \(ij^{th}\) entry of \(Q_\Sigma\) is given by \[\label{eqn:Qij} q_{ij} = \langle w_i, w_j \rangle = \langle v_i, v_j \rangle =-\cosh(\gamma_{ij})\] where \(\gamma_{ij} = d_\mathbb{H}(v_i, v_j)\) is the edge length of \(\tau\) determined by \(g_\mathbb{H}\). Note that the last equality in equation [\[eqn:Qij\]](#eqn:Qij){reference-type="eqref" reference="eqn:Qij"} is obtained directly from equation [\[eqn:hyperbolic distance\]](#eqn:hyperbolic distance){reference-type="eqref" reference="eqn:hyperbolic distance"}, and that the diagonal entries of \(Q_\Sigma\) are all \(-1\). Lastly, observe that the matrix \(Q_\Sigma\) can be constructed from \((\tau,g_\mathbb{H})\) without ever needing to isometrically embed \(\tau\) in \(\mathbb{H}\).
## Determining the realizability of hyperbolic simplices {#determining-the-realizability-of-hyperbolic-simplices .unnumbered}
Let \(\{ \gamma_{ij} \}_{i,j=1}^{n+1}\); \(\gamma_{ij} = \gamma_{ji}\) be a set of positive real numbers. The purpose of this Subsection is to establish whether this set, when defined as the edge lengths of \((\tau,g_\mathbb{H})\), will form a legitimate simplex in hyperbolic space.
# Barycentric Coordinates in Simplices of Constant Curvature {#section:coordinates}
For this Section let \((\tau,g)\) be a non-degenerate simplex with some constant curvature \(\kappa\) possibly different from 0 or-1. Linearly isometrically embed \(\tau\) into the model space \(\mathbb{R}^{n,1}\), and by abuse of notation let \(v_i\) denote the image of \(v_i\) under this realization. As before, let \(\sigma\) be the \(n\)-dimensional Euclidean simplex determined by the convex hull of \((v_1, \cdots, v_{n+1})\) in \(\mathbb{R}^{n,1}\). Let \(p \in \sigma\) be given by the barycentric coordinates \((\alpha_1, \cdots, \alpha_{n+1})\), where \(\sum_{i=1}^{n+1} \alpha_i = 1\). The purpose of this Section is to define the correspond point \(\tilde{p}\) in \(\tau\). The immediate idea is to project \(p\) onto \(\tau\) from the origin. This point \(\tilde{p}\) will then naturally depend on how \(\tau\) was embedded in \(\mathbb{R}^{n,1}\). In what follows we will give a formula for how to compute \(\tilde{p}\) using only the barycentric coordinates \((\alpha_i)_{i=1}^{n+1}\), the edge lengths \((\gamma_{ij})_{i,j=1}^{n+1}\), and the vectors \((v_i)_{i=1}^{n+1}\), therefore proving that \(\tilde{p}\) is well-defined with respect to the location of the vertices \((v_i)_{i=1}^{n+1}\) in the model space. Define \(\tilde{p}\) by \[\label{eqn:curved-barycentric} \tilde{p} = \begin{cases} \frac{p}{\kappa^2 \cdot \sqrt{\langle p, p \rangle }} & \qquad \text{if } \kappa > 0 \\ p & \qquad \text{if } \kappa = 0 \\ \frac{p}{\kappa^2 \cdot \sqrt{-\langle p, p \rangle }} & \qquad \text{if } \kappa < 0 \end{cases}\] Note that \(\tilde{p}\) is in fact the projection (from the origin) of \(p\) onto the model space with constant curvature \(\kappa\). We see from the above definition of \(\tilde{p}\) that it depends on \(p\) and \(\langle p, p \rangle\). Of course, \(p\) is completely determined by its barycentric coordinates and the location of the vertices of \(\tau\). To see that the same is true of \(\langle p, p \rangle\), we compute \[\label{eqn:p-dot-p} \langle p, p \rangle = \left\langle \sum_{i=1}^{n+1} \alpha_i v_i, \sum_{j=1}^{n+1} \alpha_j v_j \right\rangle = \sum_{i, j=1}^{n+1} \alpha_i \alpha_j \langle v_i, v_j \rangle\] where \[\langle v_i, v_j \rangle = \begin{cases} \frac{1}{\sqrt{\kappa}} \cos( \gamma_{ij} ) & \qquad \text{if } \kappa > 0 \\ \frac{1}{2} \left( \gamma_{i,n+1}^2 + \gamma_{j,n+1}^2-\gamma_{ij}^2 \right) & \qquad \text{if } \kappa = 0 \\ \frac{-1}{\sqrt{-\kappa}} \cosh( \gamma_{ij} ) & \qquad \text{if } \kappa < 0 \end{cases}\] Thus, the value of \(\langle p, p \rangle\) depends only on the the barycentric coordinates of \(p\), as well as the edge lengths of \(\tau\) as defined by \(g\). Then since \(p = \sum_{i=1}^{n+1} \alpha_i v_i\) depends only on its barycentric coordinates and the location of the vertices \((v_i)_{i=1}^{n+1}\), we see that the definition of \(\tilde{p} \in \tau\) above is well-defined with respect to the location of the vertices of \(\tau\) and the metric \(g\). Note that when referring to a specific point in \(\tau\), it is generally simpler to instead refer to the point's corresponding point in \(\sigma\), so that will be the convention used in the coming sections, including the example.
# Distances in hyperbolic simplices {#section:distances}
Let \((\tau, g_\mathbb{H})\) be a hyperbolic simplex, and let \(x, y \in \tau\) with barycentric coordinates \(x = (x_1, \hdots, x_{n+1})\) and \(y = (y_1, \hdots, y_{n+1})\). The purpose of this Section is to give a simple algorithm to compute \(d_\mathbb{H}(x,y)\), the hyperbolic distance between the points \(x\) and \(y\), using only the edge lengths \((\gamma_{ij})\) associated to \(g\) and the barycentric coordinates of \(x\) and \(y\). First, define \[\tilde{x} = \frac{x}{\sqrt{-\langle x, x \rangle}} \qquad \tilde{y} = \frac{y}{\sqrt{-\langle y, y \rangle}}\] where \(\langle x, x \rangle\) and \(\langle y, y \rangle\) can be easily calculated using the matrix \(Q_{\Sigma}\) as described in equation [\[eqn:p-dot-p\]](#eqn:p-dot-p){reference-type="eqref" reference="eqn:p-dot-p"}. Note that, if \(\tau\) were linearly isometrically embedded in the hyperboloid model, then \(x\) and \(y\) would lie on the convex hull \(\sigma\) while \(\tilde{x}\) and \(\tilde{y}\) would denote the corresponding projections of \(x\) and \(y\) onto the hyperboloid (from the origin). From equation [\[eqn:hyperbolic distance\]](#eqn:hyperbolic distance){reference-type="eqref" reference="eqn:hyperbolic distance"} we know that the hyperbolic distance from \(\tilde{x}\) to \(\tilde{y}\) is \(\text{arccosh}(-\langle \tilde{x}, \tilde{y} \rangle )\). This quantity corresponds to \(d_{\mathbb{H}}(x,y)\). That is, \(d_{\mathbb{H}}(x,y) = \text{arccosh}(-\langle \tilde{x}, \tilde{y} \rangle )\). We formally state this in the following Theorem.
Finally, note that the inner products in equation [\[eqn: hyp dist formula\]](#eqn: hyp dist formula){reference-type="eqref" reference="eqn: hyp dist formula"} are very easy to calculate using the matrix \(Q_\Sigma\). If one defines \[\vec{x} = \begin{pmatrix} x_1 \\ x_2 \\ \vdots \\ x_{n+1} \end{pmatrix} \qquad \text{and} \qquad \vec{y} = \begin{pmatrix} y_1 \\ y_2 \\ \vdots \\ y_{n+1} \end{pmatrix}\] then \[\langle x, y \rangle = \vec{x}^T Q_{\Sigma} \vec{y} \qquad \langle x,x \rangle = \vec{x}^T Q_\Sigma \vec{x} \qquad \langle y, y \rangle = \vec{y}^T Q_{\Sigma} \vec{y}.\]
# Orthogonal projection in Euclidean Simplices {#section:Euclidean proj}
For the remainder of this paper, we denote the sub-simplex generated by removing \(v_i\) from \((\tau,g)\) by \((\tau_i,g)\). Likewise, the sub-simplex generated by removing \(v_i\) and \(v_j\) is denoted by \(\tau_{ij}\), etc. The natural question arises as to the barycentric coordinates of the orthogonal projections of some point \(p \in \tau\) onto one of these sub-simplices in \((\tau,g_\mathbb{E})\) and \((\tau,g_\mathbb{H})\). Define the matrix \(Q_{\tau_{n+1}}\) as the form from for \((\tau_{n+1},g_\mathbb{E})\) (defined by equation [\[eqn:Q\]](#eqn:Q){reference-type="eqref" reference="eqn:Q"}). This Section will be focusing on \((\tau,g_\mathbb{E})\), and the following Section will discuss \((\tau,g_\mathbb{H})\). We first note that in both cases we need only consider the projection of a vertex onto a sub-simplex, as for any other point \(x \in \tau\), we may subdivide \(\tau\) in a manner which makes \(x\) into a vertex opposite the face we are projecting onto. Similarly, for projection onto a sub-simplex with \(n-2\) or fewer vertices, we need only know how to project onto an \((n-1)\)-face, as we may define a new simplex by removing the vertices that are not being projected and are not in the sub-simplex. The question, then, becomes one of optimization: we must minimize the distance from our projective vertex to the \((n-1)\)-face. For notational purposes, we relabel our vertices so that we are always projecting \(v_{n+1}\) onto the face \(\tau_{n+1}\). Let \(p = \text{proj}_{\tau_{n+1}}(v_{n+1})\). To prove our formula in Theorem [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"} for the barycentric coordinates of \(p\), we first need the following Lemma. Recall our notation that for a square matrix \(Q\), we denote its determinant by \(|Q|\).
# Orthogonal Projection in Hyperbolic Simplices {#section:hyperbolic proj}
We now turn our attention to orthogonal projection in hyperbolic simplices. In this Section we are going to project the vertex \(v_1\) onto the face \(\tau_1\), as opposed to last Section where we projected \(v_{n+1}\) onto \(\tau_{n+1}\). This reason for this change in notation is purely to make labeling subscripts easier. A first thought may be to just project within the convex hull \(\sigma\) of the vertices of \(\tau\), and then poject this point onto the hyperboloid. But, in general, this does not work. The reason for this can be found in Remark [\[rmk:non positive definite\]](#rmk:non positive definite){reference-type="ref" reference="rmk:non positive definite"}. The quadratic form in \(\mathbb{R}^{n,1}\) when restricted to the hyperplane containing the vertices of \(\tau\) may not be positive definite. In that case, the procedure in Theorem [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"} may not work.
Symmetrically to the Euclidean Projection, in order to project onto an \(n-2\) or smaller sub-simplex, we redefine a simplex that removes the irrelevant vertices and perform this calculation. As one can see, this calculation proves to be far simpler than other methods, and it only relies on the edge lengths of the simplex.
# An example {#section:example}
The formulas in our Theorems, specifically Theorems [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"} and [\[theorem:hyperbolic-projection\]](#theorem:hyperbolic-projection){reference-type="ref" reference="theorem:hyperbolic-projection"}, look a lot more complicated to use than they are in practice. The purpose of this Section is to work out an example to demonstrate the efficiency of these formulas. Let \(\tau = \langle v_1,v_2,v_3,v_4\rangle\) be a 3-simplex with edge lengths given by:
\(\gamma_{ij}\) 1 2 3 4
--------------- --- --- --- ---
1 0 2 3 4
2 2 0 4 5
3 3 4 0 3
4 4 5 3 0
In this section, we use the methods developed in this paper to compute various quanities in \(\tau\). We first verify that \(\tau\) is a non-degenerate simplex when given both the Euclidean metric \(g_\mathbb{E}\) and the hyperbolic metric \(g_\mathbb{H}\). We then follow this by finding the distances between \(p = (\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4})\) and \(q = (\frac{1}{3},\frac{1}{3},\frac{1}{3},0)\) in the Euclidean and Hyperbolic metric, and finally to find the orthogonal projection of \(v_1\) onto \(\tau_1\) with \(\tau\) viewed as both Euclidean and Hyperbolic simplex.
## Calculations in \((\tau,g_\mathbb{E})\)
## Verifying that \((\tau, g_\mathbb{E})\) is a legitimate Euclidean simplex {#verifying-that-tau-g_mathbbe-is-a-legitimate-euclidean-simplex .unnumbered}
. We construct the matrix \(Q\) from [\[eqn:Q\]](#eqn:Q){reference-type="eqref" reference="eqn:Q"}. \[Q = \begin{pmatrix} 1/2(\gamma_{12}^2 +\gamma_{12}^2-\gamma_{22}^2) & 1/2(\gamma_{12}^2 +\gamma_{13}^2-\gamma_{23}^2) & 1/2(\gamma_{12}^2 +\gamma_{14}^2-\gamma_{24}^2) \\ \\ 1/2(\gamma_{12}^2 +\gamma_{13}^2-\gamma_{23}^2) & 1/2(\gamma_{13}^2 +\gamma_{13}^2-\gamma_{33}^2) & 1/2(\gamma_{13}^2 +\gamma_{14}^2-\gamma_{34}^2) \\ \\ 1/2(\gamma_{12}^2 +\gamma_{14}^2-\gamma_{24}^2) & 1/2(\gamma_{13}^2 +\gamma_{14}^2-\gamma_{34}^2) & 1/2(\gamma_{14}^2 +\gamma_{14}^2-\gamma_{44}^2) \end{pmatrix} = \begin{pmatrix} 4 &-3/2 &-5/2 \\ \\-3/2 & 9 & 8\\ \\-5/2 & 8 & 16 \end{pmatrix}\] One can check that the eigenvalues of \(Q\) are approximately 21.7, 3.81, and 3.48. Then, since \(Q\) is positive-definite, \(\tau\) is a legitimate Euclidean simplex.
## Calculating distances in \((\tau, g_\mathbb{E})\) {#calculating-distances-in-tau-g_mathbbe .unnumbered}
Let \(p = (\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4})\) and \(q=(\frac{1}{3},\frac{1}{3},\frac{1}{3},0)\). We wish to calculate \(d_\mathbb{E}(p,q)\). The squared Euclidean distance between them, by Equation [\[euclid-dist\]](#euclid-dist){reference-type="ref" reference="euclid-dist"}, is \[d_\mathbb{E}^2(p,q) = [p-q]^TQ_E[p-q] = \begin{pmatrix} \frac{1}{12} & \frac{1}{12} & \frac{-1}{4} \end{pmatrix}\begin{pmatrix} 4 &-3/2 &-5/2 \\-3/2 & 9 & 8\\-5/2 & 8 & 16 \end{pmatrix}\begin{pmatrix} \frac{1}{12}\\\frac{1}{12}\\\frac{-1}{4} \end{pmatrix} = \frac{121}{144},\] so \(d_E(p,q)=11/12\).
## Projecting \(v_1\) onto \(\tau_1\) in \((\tau, g_\mathbb{E})\) {#projecting-v_1-onto-tau_1-in-tau-g_mathbbe .unnumbered}
Now, we find the projection of \(v_1\) onto \(\tau_1\). Firstly, we have \[Q_{\tau_1} = \begin{pmatrix} 1/2(\gamma_{23}^2 + \gamma_{23}^2-\gamma_{33}^2) & 1/2(\gamma_{23}^2 + \gamma_{24}^2-\gamma_{34}^2)\\ 1/2(\gamma_{23}^2 + \gamma_{24}^2-\gamma_{34}^2) & 1/2(\gamma_{24}^2 + \gamma_{24}^2-\gamma_{44}^2) \end{pmatrix} = \begin{pmatrix} 16 & 16 \\ 16 & 25 \end{pmatrix}\] So \(|Q_{\tau_1}|=144\). We now compute the minors of \(Q\): \[Q_{11}= \begin{vmatrix} 9 & 8 \\ 8 & 16 \end{vmatrix}=80 \qquad Q_{12}= \begin{vmatrix}-3/2 & 8 \\-5/2 & 16 \end{vmatrix}=-4 \qquad Q_{13}= \begin{vmatrix}-3/2 & 9 \\-5/2 & 8 \end{vmatrix}=21/2\]\[Q_{22}= \begin{vmatrix} 4 &-5/2 \\-5/2 & 16 \end{vmatrix}=231/4 \qquad Q_{23}= \begin{vmatrix} 4 &-3/2 \\-5/2 & 8 \end{vmatrix}=113/4 \qquad Q_{33}= \begin{vmatrix} 4 &-3/2 \\-3/2 & 9 \end{vmatrix}=135/4\] Thus, we can then find each \(\alpha_i\), noting that \(Q_{ij} = Q_{ji}\) due to the symmetry of \(Q\): \[\alpha_2 = \frac{Q_{11}-Q_{12} + Q_{13}}{|Q_{\tau_1}|} = \frac{80-(-4) + 21/2}{144} \approx 0.65625\]\[\alpha_3 = \frac{-Q_{21} + Q_{22}-Q_{23}}{|Q_{\tau_1}|} = \frac{-(-4) + 231/4-113/4}{144} \approx 0.23264\]\[\alpha_4 = \frac{Q_{31}-Q_{32}+Q_{33}}{|Q_{\tau_1}|} = \frac{21/2-113/4+135/4}{144} \approx 0.11111\] Let us quickly remark that the subscripts of the \(\alpha_i's\) are off by 1 from Theorem [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"} since we are projecting \(v_1\) as opposed to \(v_4\) as in the Theorem. Note that \(\alpha_2+\alpha_3+\alpha_4=1\). So the orthogonal projection of \(v_1\) onto \(\tau_1\) has barycentric coordinates \((0,0.6525, 0.23264, 0.11111)\). It is worth noting that each coordinate is positive, and so the projection \(p\) lies inside the triangle \(\tau_1\). Our formula gives an easy way to check if a vertex projects inside or outside of the opposite face.
## Calculating the "height\" of \((\tau, g_\mathbb{E})\) {#calculating-the-height-of-tau-g_mathbbe .unnumbered}
Consider the altitude of \(v_1\) over \(p\): \[d_\mathbb{E}(v_1,p) = \sqrt{p^TQp} \approx 1.4136.\] We will use this result to compare to the analogous result in the hyperbolic example.
## Calculations in \((\tau,g_\mathbb{H})\)
## Verifying that \((\tau, g_\mathbb{H})\) is a legitimate hyperbolic simplex {#verifying-that-tau-g_mathbbh-is-a-legitimate-hyperbolic-simplex .unnumbered}
Via Theorem [\[theorem: hyperbolic-realizability\]](#theorem: hyperbolic-realizability){reference-type="ref" reference="theorem: hyperbolic-realizability"}, we need to calculate \(Q_{\Sigma}\) using equation [\[eqn:Qij\]](#eqn:Qij){reference-type="eqref" reference="eqn:Qij"}. We have that \(Q_\Sigma\) = \[\begin{pmatrix}-1 &-\cosh(\gamma_{12}) &-\cosh(\gamma_{13}) &-\cosh(\gamma_{14})\\ \\-\cosh(\gamma_{12}) &-1 &-\cosh(\gamma_{23}) &-\cosh(\gamma_{24})\\ \\-\cosh(\gamma_{13}) &-\cosh(\gamma_{23}) &-1 &-\cosh(\gamma_{34})\\ \\-\cosh(\gamma_{14}) &-\cosh(\gamma_{24}) &-\cosh(\gamma_{34}) &-1 \end{pmatrix} = \begin{pmatrix}-1 &-\cosh(2) &-\cosh(3) &-\cosh(4)\\ \\-\cosh(2) &-1 &-\cosh(4) &-\cosh(5)\\ \\-\cosh(3) &-\cosh(4) &-1 &-\cosh(3)\\ \\-\cosh(4) &-\cosh(5) &-\cosh(3) &-1 \end{pmatrix}\] The eigenvalues of \(Q_\Sigma\) are approximately-90.1, 79.2, 5.5, and 1.4. Since \(Q_\Sigma\) has signature \((3,1)\), by Theorem 1 we know that \(\tau\) is a legitimate hyperbolic simplex.
## Calculating distances in \((\tau, g_\mathbb{H})\) {#calculating-distances-in-tau-g_mathbbh .unnumbered}
We now wish to calculate \(d_\mathbb{H}(p,q)\), where \(p = (\frac{1}{4},\frac{1}{4},\frac{1}{4},\frac{1}{4})\) and \(q=(\frac{1}{3},\frac{1}{3},\frac{1}{3},0)\). By Theorem [\[thm:hyp dist\]](#thm:hyp dist){reference-type="ref" reference="thm:hyp dist"}, we have \[d_\mathbb{H}(p,q) = \text{arccosh}\left(\frac{-\langle p,q \rangle}{\sqrt{\langle p,p \rangle \cdot \langle q,q \rangle}}\right),\] with: \[\langle p,q \rangle = p^t(Q_\Sigma) q \approx-16.40517\;, \quad \langle p,p \rangle = p^t(Q_\Sigma) p \approx-19.34049, \text{ and}\quad \langle q,q \rangle = q^t(Q_\Sigma) q \approx-9.47513.\] Thus, we have that \[d_\mathbb{H}(p,q) = \text{arccosh} \left(\frac{16.40516}{\sqrt{(-19.34049)(-9.47513)}}\right) = 0.63997.\] Notice here that \(d_\mathbb{H}(p,q) < d_\mathbb{E}(p,q)\), as expected.
## Projecting \(v_1\) onto \(\tau_1\) in \((\tau, g_\mathbb{H})\) {#projecting-v_1-onto-tau_1-in-tau-g_mathbbh .unnumbered}
We now consider the projection \(\tilde{p}\) of \(v_1\) onto \(\tau_1\). First, we find the barycentric coordinates for the corresponding point \(p = (0,\alpha_2,\alpha_3,\alpha_4)\) on the convex hull \(\sigma\), before finding the coordinates of \(\tilde{p}\). First, the relevant minors of \(Q_\Sigma\) are: \[\qquad Q_{12}^\Sigma = \begin{vmatrix}-\cosh(2) &-\cosh(4) &-\cosh(5)\\ \\-\cosh(3) &-1 &-\cosh(3)\\ \\-\cosh(4) &-\cosh(3) &-1 \end{vmatrix}\approx-12350.57\] \[Q_{13}^\Sigma = \begin{vmatrix}-\cosh(2) &-1 &-\cosh(5)\\ \\-\cosh(3) &-\cosh(4) &-\cosh(3)\\ \\-\cosh(4) &-\cosh(5) &-1 \end{vmatrix}\approx 2340.72 \quad Q_{14}^\Sigma = \begin{vmatrix}-\cosh(2) &-1 &-\cosh(4)\\ \\-\cosh(3) &-\cosh(4) &-1\\ \\-\cosh(4) &-\cosh(5) &-\cosh(3) \end{vmatrix}\approx-718.81\] From Theorem [\[theorem:hyperbolic-projection\]](#theorem:hyperbolic-projection){reference-type="ref" reference="theorem:hyperbolic-projection"}, we know that \[\alpha_2 = \frac{-Q_{12}}{-Q_{12}+Q_{13}-Q_{14}} \qquad \alpha_3 = \frac{Q_{13}}{-Q_{12}+Q_{13}-Q_{14}} \qquad \alpha_4 = \frac{-Q_{14}}{-Q_{12}+Q_{13}-Q_{14}}.\] Plugging in the values for the minors and calculating, we get that \(p = (0, 0.80146, 0.15190, 0.04665)\). Note that \(\alpha_2+\alpha_3+\alpha_4 = 1\), as expected. Also by Theorem [\[theorem:hyperbolic-projection\]](#theorem:hyperbolic-projection){reference-type="ref" reference="theorem:hyperbolic-projection"}, we have \[\tilde{p} = (0,\tilde{\alpha}_2,\tilde{\alpha}_3,\tilde{\alpha}_4); \quad \text{for } \tilde{\alpha}_i = \frac{(-1)^{i+1}Q^\Sigma_{1i}}{\sqrt{\sum_{i,j=2}^{n+1}(-1)^{i+j}Q^\Sigma_{1i}Q^\Sigma_{1j}(\cosh(\gamma_{ij}))}}\]And thus, since \[\sqrt{\sum_{i,j=2}^{n+1}(-1)^{i+j}Q^\Sigma_{1i}Q^\Sigma_{1j}(\cosh(\gamma_{ij}))} \approx 55578.499 \text{, we have } \tilde{p} = \left(0,0.22222,0.04212, 0.01293 \right).\] Alternatively, one could just calculate \(\langle p, p \rangle\), and then use that \(\tilde{p} = \frac{p}{\sqrt{-\langle p, p \rangle}}\).
## Calculating the "height\" of \((\tau, g_\mathbb{H})\) {#calculating-the-height-of-tau-g_mathbbh .unnumbered}
Let us find the altitude of \(v_1\) over \(\tilde{p}\): \[d_\mathbb{E}^2(v_1,\tilde{p}) = \begin{pmatrix}-1 & 0.222 & 0.042 & 0.0129 \end{pmatrix}Q_\Sigma \begin{pmatrix}-1 \\ 0.222 \\ 0.042 \\ 0.0129 \end{pmatrix}\approx 1.22644\]\[d_\mathbb{H}(v_1,\tilde{p}) = \text{arccosh}\left(\frac{2+1.22644}{2}\right) \approx 1.0575.\] Note that \(d_\mathbb{H}(v_1,p) < d_\mathbb{E}(v_1,p)\), as we would expect.
# Analogous formulas for spherical simplices {#section:spherical}
Our model space for \(\mathbb{S}^n\) is the unit sphere in \(\mathbb{R}^{n+1}\). Similar to equation [\[eqn:hyperbolic distance\]](#eqn:hyperbolic distance){reference-type="eqref" reference="eqn:hyperbolic distance"} we have \[\label{eqn:spherical distance} d_\mathbb{S}(x,y) = \arccos(\langle x,y \rangle).\] Also, in the same way as we did in Section [3](#section:realizability){reference-type="ref" reference="section:realizability"}, we can consider the \((n+1)\)-simplex \(\Sigma\) which is the convex hull of the vertices of a simplex and the origin. The Gram matrix \(Q_\Sigma\) is calculated by the formula \[\label{eqn:spherical Qij} q_{ij} = \langle w_i, w_j \rangle = \langle v_i, v_j \rangle = \cos(\gamma_{ij}).\]
Distances in spherical simplices are computed in the analogous way as they are in hyperbolic simplices: one projects the points onto the sphere using the techniques from Section [4](#section:coordinates){reference-type="ref" reference="section:coordinates"}, and then calculates the distance using equation [\[eqn:spherical distance\]](#eqn:spherical distance){reference-type="eqref" reference="eqn:spherical distance"} and \(Q_{\Sigma}\). For orthogonal projection in spherical simplices, you just project within the convex hull of the points and then project that point onto the sphere. More precisely, let \((\tau, g_\mathbb{S})\) be a spherical simplex. Linearly isometrically embed \(\tau\) into \(\mathbb{S}^n\) in some way, and identify the vertices \(v_i\) with their image in \(\mathbb{S}^n\). Let \(\sigma\) be the \(n\)-simplex formed by the convex hull of \((v_1, \dots, v_{n+1})\). Then, to calculate \(\text{proj}_{\tau_1}(v_1)\), you calculate \(\text{proj}_{\sigma_1}(v_1)\) using Theorem [\[thm:Euclid Projection\]](#thm:Euclid Projection){reference-type="ref" reference="thm:Euclid Projection"} and then project this point onto the sphere from the origin. This process works for spherical simplices but not for hyperbolic simplices because the quadratic form restricted to the hyperplane containing \(\sigma\) is always positive-definite for spherical simplices. Finally, the distance and projection formulas in this paper can be extended to simplices with constant curvature \(\kappa\) by adjusting equations [\[eqn:hyperbolic distance\]](#eqn:hyperbolic distance){reference-type="eqref" reference="eqn:hyperbolic distance"}, [\[eqn:Qij\]](#eqn:Qij){reference-type="eqref" reference="eqn:Qij"}, [\[eqn:spherical distance\]](#eqn:spherical distance){reference-type="eqref" reference="eqn:spherical distance"}, and [\[eqn:spherical Qij\]](#eqn:spherical Qij){reference-type="eqref" reference="eqn:spherical Qij"} accordingly. | {'timestamp': '2021-07-14T02:03:15', 'yymm': '2107', 'arxiv_id': '2107.05706', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05706'} |
null | null |
null | null |
null | null |
null | null |
# Introduction
Floating-point arithmetic in scientific, financial, and engineering applications suffers from rounding error: the results of a floating-point expression can differ starkly from those of the analogous mathematical expression | {'timestamp': '2021-07-14T02:06:53', 'yymm': '2107', 'arxiv_id': '2107.05784', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05784'} |
# Introduction
Natural language is an ubiquitous communication medium between human interlocutors, and is shaped by the desire to efficiently cooperate and achieve communicative goals. Because of this, there has been interest in creating artificial agents that mimic this communication process, with a wide variety of works examining communication between agents via either completely artificial *emergent* language, or through *natural* language such as English. In general, these methods model interaction between a pair of agents, a *speaker* and a *listener* that attempt to jointly achieve a goal where the language use is learned to optimize success with respect to a collaborative task (for example, in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"} the speaker instructs an embodied agent to perform a task in an environment). However, in contrast to this setup, human speakers interact with not a single listener, but many different conversational partners. In doing so, they also adapt to each other's language within short conversations. One of the representative phenomena is entrainment, in which interlocutors align their language on both acoustic-prosodic and lexical dimensions during communication. These issues remain mostly unattested in previous work on multi-agent communication---it remains an open question how to train an agent which can adapt to communicating with novel partners quickly. We define this adaptive communication problem as *few-shot language coordination*. As shown in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}, within a few rounds of the game, the speaker must adapt its language based on the responses from the listener to achieve the *communicative goal* of helping the listener perform the correct actions as many times as possible. This few-shot coordination setting provides the agents with the pressure to adapt on-the-fly--something that current models generally cannot achieve, even those that model pragmatics, e.g. the rational speech act model. Developmental psychology argues for the importance of Theory of Mind (ToM), the understanding of others' mental states, and also the ability of interlocutors to act upon others mental states to achieve desired effects. In this paper, we study the importance of modeling the beliefs of one's conversational partner to achieve few-shot language coordination. In particular, we train a model that quickly adapts to predict the actions of the listeners in real time. At each time step, the speaker predicts the listener's likely next action given all possible instructions and the listener's previous actions. This is essentially a few-shot learning problem, which we attack with model-agnostic meta-learning (MAML). In order to achieve the communicative goal, the speaker chooses the best instruction to give so that the probability of the listener performing the correct actions is maximised. We expect the resulting agent to not only mimic humans' ability to model a listener's mental state but also to leverage this estimate to choose better instructions. Through empirical evaluation, we aim to answer the question:
> *Can an agent, equipped with a model of theory of mind, quickly adapt to a listener's language in a few-shot language coordination game?*
Our experiments answer in the affirmative in the both a referential game and a vision-language navigation setting.
# Few-shot Language Coordination
Consider again the example depicted in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}. A speaker model observes the goal of the current task ("pick and cool wine bottle, place it on the dining table"), and sends messages to a listener model, in an attempt to help it finish the task. The listener model chooses actions to take given the latest and previous instructions. If the listener makes a mistake, such as going to the microwave instead of the refrigerator, the speaker should realize that the listener misunderstood some aspect of the high-level instructions in the first message. This misstep informs the speaker's model of the listener's ability, leading them to give lower-level instructions to help correct the trajectory. Through *several games*, the speaker gathers enough data to provide customized messages for individual listeners. This is only possible if a proper prior is provided to the speaker. The simplest prior can be hand-coded rules, e.g. if listener cannot understand abstract instructions, try simpler ones. However, to pursue a more general and powerful communication model (e.g. knowing when to simplify vs. rephrase), we study whether this kind of few-shot language coordination can be *learned* by playing with a *population* of listeners. This section proposes a method to construct meaningful populations.
## Asymmetric Speaker-Listener Games
Following previous work on communicative agents, we use goal-oriented language games as the test bed for few-shot language coordination. A general goal-oriented language game provides an environment where the participants uses language to communicate with each other to achieve the given goal. We consider the most basic setting of a simplified two-player shared-goal multi-round setup:
Environment:
: The environment is defined by \(\mathcal{O}\)bservation space, \(\mathcal{A}\)ction space, \(\mathcal{G}\)oal space and transition function \(E:\mathcal{O}\times \mathcal{A}\rightarrow\mathcal{O}\times \mathcal{G}\). At the start of each game, the environment provides the speaker with a goal and both participants with observations after each action is taken by the listener. A new game starts after the previous one succeeds or reaches a maximum number of steps.
Participants:
: The participants consist of a speaker and a listener sending and receiving natural language messages. After observing the goal, the speaker gives an instruction to the listener, and the listener performs an action in the environment. If the game is sequential, the speaker can also give an instruction after each action until the game is solved or the maximum number of steps is reached. The speaker is a message-and-action producing model defined by the *vocabulary* \(\Sigma\); the space of *observations* \(\mathcal{O}\); the space of *actions* \(\mathcal{A}\); and a *model* \(f:\mathcal{O}\times\mathcal{G}\rightarrow\Sigma^*\times \mathcal{A}\). The listener is an instruction-follower defined by the same vocabulary \(\Sigma\), observation space \(\mathcal{O}\), and space of *actions* \(\mathcal{A}\) as the speaker; and a *model* \(g: \Sigma^* \times \mathcal{O} \rightarrow \mathcal{A}\).
Multi-round Games:
: The pair of participants will play a *session* of \(N\) rounds of games, which are sampled independently. Different from single-round games (\(N\!=\!1\)) used in most previous work, the participants keep the memory of past games in the same session. Multi-round games are not only more general than single-round games, but are essential to few-shot language coordination, because participants have the opportunity to adapt to the interlocutors by learning from feedback during previous rounds.
Note that the listeners in this setting cannot directly observe the goal, so the speakers need to use instructions to inform the listeners about the final or intermediate goals of each game. Within \(N\) rounds, the speaker needs to adapt to the listener's knowledge to provide the most effective instructions.
## Population
coined the notion of "machine ToM", which is a model for tracking agents' behaviors. To train and evaluate its ability to adapt to different agents, they create populations of subject agents by using different neural architectures and random seeds for parameter initialization. Our design of populations draws inspiration from their work. However, to make the similarities and differences between agents controllable, we consider a population as a distribution over parameters neural listeners with the same architecture which have been trained on different datasets. A neural listener \(f_{\theta}^L: \mathcal{O}\times\mathcal{I}\rightarrow\mathcal{A}\) is a mapping from \(\mathcal{O}\)bservations and \(\mathcal{I}\)nstructions to \(\mathcal{A}\)ctions with parameter \(\theta\) of the neural networks. The parameters trained on dataset \(\mathcal{D}\) are: \[\theta_{\mathcal{D}} = \arg\min_{\theta} \mathcal{L}(f_L, \mathcal{D})\] In this way, the variation over listeners is mainly determined by the features of the dataset. By constructing datasets with different feature distributions, we control the listeners' language abilities. In the example of Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}, the observations in language games are the items in the visual field, and the agent may perform any number of actions (e.g. "`go to cabinet`"). The speaker may provide natural language instructions, which can range from high-level (e.g. "cool the winebottle") to low-level (e.g. "get the wine bottle", "take it to the refrigerator", "put it in the refrigerator"). A listener that has never been trained on a particular variety of (usually high-level) instruction would have trouble performing the appropriate actions. This leads to an exponential population of listeners that are trained on datasets containing, or not containing, particular relevant instructions. Because of this, in order to effectively and concisely communicate, an effective speaker will have to judge the language abilities of its various partners in the population and adjust appropriately; we explain how we do so in the following section.
## Theory-of-mind Model
Theory-of-mind, the ability to build a model of one's conversational partners, is deemed to be crucial in socio-pragmatics theory. Drawing inspiration from this, we build a theory-of-mind model to learn to mimic the behavior of the listener within a short time window.
#### Mental State
Modeling mental states is the central concept in building a theory-of-mind. We define the mental state of the listener as the parameters of a neural model, the *ToM model*, that produces the same output for the same inputs as the listener: \(\forall x\in \Sigma^*, o \in \mathcal{O}, g_{\text{ToM}}(x, o;\theta_{\text{mind}}) \approx g(x, o;\theta)\). It should be noted that in the general case, particularly when different model architectures are used to represent the model itself and the ToM model, the mental state representations may not be unique or even exist. In other words, for any model \(\theta\) there may be more than one parameter setting \(\theta_{\text{mind}}\) that satisfies this condition, or there may be no \(\theta_{\text{mind}}\) that produces the exact same output
#### Building a Theory-of-mind
Learning a ToM model is reduced to inferring the mental state of the listener. For a given listener \(g\) with parameters \(\theta\) and ToM model \(g_{\text{ToM}}\), we seek a mental state representation \(\theta_{\text{mind}}\). In practice, we use identical neural architectures for both the listener and ToM Model. However, inferring the exact mental state is infeasible within few interactions. Therefore, we estimate \(g_{\text{ToM}}\) such that \[\theta_{\text{mind}} = \arg\min_{\theta'} \mathbb{E}_{o, m}\mathcal{L} (g_{\text{ToM}}(o, m;\theta'), g(o, m;\theta))\] It is straightforward to apply this definition of mental state in the psychological context for which it was originally proposed. The mental state \(\theta_{\text{mind}}\) is the representation of the listener's language abilities, which are not directly observable, and which are ultimately used for predicting the belief and behavior of the speaker. For example, in our first set of experiments we focus on referential game where the speaker describes the target in order to let the listener pick it out from distractors. We construct a population in which neural listeners with LSTMs and word embeddings have different language comprehension abilities for different languages. One of the possible representations controls the word embeddings in different languages: the mental state of a good language listener should have more meaningful word embeddings, while the one which cannot understand the language should have more random ones. Given that the speaker can acquire an accurate mental state for the listener, it can be used for predicting the probability of listener choosing the correct image when hearing descriptions in different languages. By choosing the one that yields the correct image with the highest probability, the speaker generates the descriptions which improve the referential game. On the other hand, high quality descriptions help the speaker better narrow down the language abilities of the listener. This is similar to the two-way interrelation between language and ToM in humans. Following this direction, we present a dynamic view of ToM by putting the observer inside the conversation, instead of the static view of, which uses ToM for tracking the behavior of the agent without interfering in the games. Our training procedure is presented in Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"}. We aggregate a dataset \(D_{\theta_{\text{mind}}}\) at each epoch, and update the parameters by optimizing the ToM model on the dataset. To aggregate the dataset for each training listener, we randomly sample from the posteriors of the ToM model and uniform distributions over the candidates, which keeps a certain degree of exploration, modulated by distribution coefficient \(\sigma\) (through the paper, we use \(\sigma = 0.5\)). In practice, parameters are updated with stochastic gradient descent by sampling listeners and using the history of each listener at each time step as a support set for predicting the next actions of the listener. Following the literature on speech acts, e.g. , we also add exponential cost penalty \(\exp(-\kappa C(m))\) as a prior to penalize long instructions. (We have not explored the space of penalty functions in this paper, but the exponential function is widely used in the pragmatics literature, e.g. ,.) In Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"} (a&b), although "go to fridge" yields the highest probability of gold action, no instruction is given in order to express the goal concisely. Similarly to the imitation learning algorithm DAgger, the dataset is collected using expert actions. However, there is a major difference between Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"} and DAgger---we optimize the prediction of actions conditioned on the observations and instructions instead of the instruction probability directly. The following theorem shows that our model will improve the instruction generation quality:
#### Discussion
The conditions of Theorem [\[th:1\]](#th:1){reference-type="ref" reference="th:1"} mean that the speaker model \(S\) must be a well-trained model to produce good enough candidates pools. In practice, this condition is not hard to meet: for instance, in our language navigation experiment the listeners can at least understand the lowest-level instructions, and the speaker generates four levels of instructions by rule-based experts. Therefore, the practical implication of this theorem is helpful--our method reduces to DAgger without expert instructions. Different from DAgger, our training method doesn't directly optimize the instruction distribution against expert's instructions, but optimizes the action prediction loss instead, which upper-bounds the instruction loss.
## Meta-learning ToM Model
To acquire an estimate of the mental state from very few interactions, the ToM model needs to quickly learn from a small support set. In theory, any model for parameterizing \(\mathcal{P}_{\text{ToM}}(a \mid o, m, \mathcal{D}_{\text{supp}};\theta_{\text{mind}})\) could work in our framework. As a general method applicable to all ToM models, we apply model-agnostic meta-learning (MAML; ), a method that explicitly trains models to be easily adaptable in few-shot settings. Given support dataset \(\mathcal{D}_{\text{supp}}\), the inner loop updates parameters for \(N_{\text{inner}}\) steps: \[\begin{aligned} \theta^{(0)} &= \theta_{\text{mind}} \\ \theta^{(i+1)} &= \theta^{(i)}-\eta \nabla_{\theta} \mathcal{L}(\mathcal{D}_{\text{supp}}; \theta)|_{\theta=\theta^{(i)}} \end{aligned}\] where \(\eta\) is the inner loop learning rate, which, in practice, is not share across modules following. The loss function on the support set is the negative log-likelihood of listener's action given observations \(o\) and the instructions \(m\) \[\mathcal{L}(\mathcal{D}_{\text{supp}}; \theta) =-\mathbb{E}_{(o, m, a)\sim\mathcal{D}_{\text{supp}}} \log p_{\theta}(a\mid o, m)\] After \(N_{\text{inner}}\) steps, we get the prediction on the target observation \(o\) and instruction \(m\) \[\mathcal{P}_{\text{ToM}}(a \mid o, m, \mathcal{D}_{\text{supp}};\theta_{\text{mind}}) = p_{\theta^{N_{\text{inner}}}}(a\mid o, m)\] Outer loop optimize \(\theta_{\text{mind}}\) by mini-batch stochastic gradient descent \[\theta_{\text{mind}} \leftarrow \theta_{\text{mind}}-\eta_{\text{outer}} \nabla_{\theta} \mathcal{L}^{\text{pred}} (\mathcal{D}_{\theta_{\text{mind}}})\] The outer loop also runs for a given \(N_{\text{outer}}\) epochs.
## Deploying the ToM Model
Similar to Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"}, we evaluate ToM by using it to measure the probability of the gold action given the instructions. However, here we choose the best one instead of sampling from the posterior. Alg. [\[proc:test_tom\]](#proc:test_tom){reference-type="ref" reference="proc:test_tom"} shows the evaluation procedure.
The gold-standard speaker denotes the success rate of using the testing listener in place of the ToM listener. The score of over \(90\%\) indicates that the candidate pool is of high quality, so a speaker with a well-modeled ToM listener has ample room for achieving high accuracy. The non-ToM speaker uses the instruction with the highest probability in the speaker model; the RSA speaker uses the listener for training the speaker in place of the ToM listener. Our model achieves a significantly higher success rate, demonstrating that the ToM model could help produce better instructions for this referential game. However, does ToM model truly learn to adapt to individual listeners? We compute the accuracy of predicting the listener's behavior during the same session. Fig. [\[fig:acc\]](#fig:acc){reference-type="ref" reference="fig:acc"} shows that the prediction accuracy of listener's actions is significantly improved within sessions, which shows ToM indeed learns to adapt to individual test listeners.
# ALFWorld Task Navigation
In the previous section, we have shown that ToM could help games with simple dynamics, and learn to adapt to listeners. This section will show its application in a more complex game: language navigation. We use the Alfworld platform, which creates a natural language command action space upon the Alfred environment for vision-language navigation. In each game, a household task is given by the environment. We generate expert trajectories for all tasks in Alfworld, and manually create four levels of instructions, from task-level to action-level, which are denoted as \(\mathcal{I}_i, i=1, 2, 3, 4\). The differences between these four levels are the levels of abstraction. An action-level instruction corresponds to a single action in the Alfworld, while a task-level instruction corresponds to a whole trajectory which consists of more than eight commands. The two other levels are in between. The candidate pool at each time step consists of four instructions from each level (\(N_M=4\)). To create each listener, we draw an instruction distribution from \(Dir(0.6, 0.4, 0.3, 0.2)\) for each of the six types of task in the environment. Listeners are of the same neural model as in. While training the listeners, the instructions are randomly drawn from the instruction distributions according to the task type. This procedure produces 50 listeners. The average success rate of listeners is 83.6%. These listeners are randomly divided into training, validation, and testing listeners (30/10/10). We define the cost as the average total length of the set of instructions, i.e. repetitive instructions are only calculated once. The cost function is defined as \(C(m) = 2^i \ \text{if } m \in \mathcal{I}_i\). Within one session, the maximum number of interactions between speaker and listener is \(K=100\), and maximum number of interactions in a game is 20. Listeners' hyper-parameters are the same as the ones in, while MAML hyper-parameters are the same as referential game. In Fig. [\[fig:language_navigation\]](#fig:language_navigation){reference-type="ref" reference="fig:language_navigation"}, we compare ToM-assisted speakers and random speakers. We didn't compare with an RSA speaker, because differently from the multilingual referential games, the speaker is rule-based and no listener is used for training the speaker. A random speaker draws an instruction distribution from \(Dir(0.7, 0.7, 0.7, 0.7)\), and sample instructions from the candidate pools using the instruction distribution. The ToM speaker with \(\kappa=0\) predominantly uses action-level instructions, while the ToM speaker with \(\kappa=10\) uses task-level instructions most of the time. Comparing ToM speakers and random speakers, we find that for \(\kappa = \{0, 1, 2\}\), ToM speakers achieve higher game points and lower cost than random ones; for \(\kappa=10\), the ToM speaker does not have significant improvement over a random one, since only the listeners that have been trained on sufficient action-level instructions can succeed.
# Related Work
## Language Games
Language games have been the proving ground for various linguistics theories since their conception by. Recently, the most widely used language game is the referential game, in which the speaker observes the target and distractors and uses language to instruct the listener on how to pick out the target. **Emergent Communication** Without natural language annotations, this pressure for the speaker and listener enables language emergence. first uses the same recurrent neural networks as the speaker and the listener to conduct emergent communication in referential game. Following this lead, study how emergent languages are grounded to the input images. studies multi-turn communication via negotiation. study the compositionally and systematicity of emergent languages. **Learning Common Languages** By using natural language annotations, agents learn a common language so that agents that are never trained together can be expected to communicate. studies using MSCOCO annotations as gold labels for both speakers and listeners. found that alternating between self-playing and supervised learning benefits communication performance. show that humans have the ability to adapt to the machine's capability in language games. found that when language follows Zipf law, zero-shot communication is possible. **Language Games with Community** learns representations by training with a community of encoders and decoders. The difference between our work and theirs is that our MAML listener learns to adapt to different listeners in the population in a few games. The performance of their model should be equivalent to our model's result at time step 1. considers the adaptation problem, which is definitely relevant. However, adapting their model to our settings is non-trivial, which is beyond the scope of this paper. This referential game setting used in most previous work can be seen as a special case of our few-shot coordination formulation, where the number of games is one and the partners are mostly the same ones as in the training phase. These two differences prevent the previous models from learning to adapt due to the lack of pressure to do so.
## Machine Theory of Mind
Computational and neural models of theory-of-mind have been studied for decades. are the first to present a successful modeling of the mental state of various species of agent. While we also train a ToM model with meta learning, we put the ToM model into use. The predictions provided by the ToM model serve as reranker in the speaker's model. The variety of models is also more diverse than the species used in this paper. Importantly, find that neural models for question answering fail to keep track of inconsistent states of the world. extends machine ToM to neural recursive belief states. We expect improvement over our current model by modeling higher-order recursive belief, which is left for future work. explicitly trains belief state prediction with supervised learning, while our model's belief state is latent.
## Similar Topics in Reinforcement Learning
**Model-based Reinforcement Learning** Model-based reinforcement learning focuses on building a model of the environment to improve data efficiency, which could be applied to the zero-shot coordination problem by treating the listener as a part of the environment. Recently, neural networks have been used widely for model-based RL. We should point out that despite their similarities to our model, we focus on modeling different and unseen agents in the population within a few interactions. **Alternatives to Self-play** Zero-shot coordination has attracted much attention recently. propose other-play which maximizes the expected reward working with random partners. In contrast to their approach, we explicitly model the partner's ToM and focus on language coordination, which required more complicated modeling than the environments in their experiments.
# Implications and Future Work
We have introduced few-shot language coordination task and proposed ToM model for tracking the listener's mental state. Different from previous work using single-round games and self-play training, we consider more general multi-round games and playing with novel listeners. ToM model shows its ability to adapt to novel listeners and assist speakers in choosing the best instructions in both multilingual referential games and the language navigation task. We attribute the success of ToM model to modeling socio-pragmatics process in an explicit way. Many interesting questions about modeling ToM in few-shot language coordination remain open. The most immediate is how to model ToM from the listener's perspective, leading to a dialog agent that can acquire new knowledge through conversation. One step further, similar to RSA, ToM can also be modeled in a recursive manner, which may further improve language games.
# Introduction
Natural language is an ubiquitous communication medium between human interlocutors, and is shaped by the desire to efficiently cooperate and achieve communicative goals. Because of this, there has been interest in creating artificial agents that mimic this communication process, with a wide variety of works examining communication between agents via either completely artificial *emergent* language, or through *natural* language such as English. In general, these methods model interaction between a pair of agents, a *speaker* and a *listener* that attempt to jointly achieve a goal where the language use is learned to optimize success with respect to a collaborative task (for example, in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"} the speaker instructs an embodied agent to perform a task in an environment). However, in contrast to this setup, human speakers interact with not a single listener, but many different conversational partners. In doing so, they also adapt to each other's language within short conversations. One of the representative phenomena is entrainment, in which interlocutors align their language on both acoustic-prosodic and lexical dimensions during communication. These issues remain mostly unattested in previous work on multi-agent communication---it remains an open question how to train an agent which can adapt to communicating with novel partners quickly. We define this adaptive communication problem as *few-shot language coordination*. As shown in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}, within a few rounds of the game, the speaker must adapt its language based on the responses from the listener to achieve the *communicative goal* of helping the listener perform the correct actions as many times as possible. This few-shot coordination setting provides the agents with the pressure to adapt on-the-fly--something that current models generally cannot achieve, even those that model pragmatics, e.g. the rational speech act model. Developmental psychology argues for the importance of Theory of Mind (ToM), the understanding of others' mental states, and also the ability of interlocutors to act upon others mental states to achieve desired effects. In this paper, we study the importance of modeling the beliefs of one's conversational partner to achieve few-shot language coordination. In particular, we train a model that quickly adapts to predict the actions of the listeners in real time. At each time step, the speaker predicts the listener's likely next action given all possible instructions and the listener's previous actions. This is essentially a few-shot learning problem, which we attack with model-agnostic meta-learning (MAML). In order to achieve the communicative goal, the speaker chooses the best instruction to give so that the probability of the listener performing the correct actions is maximised. We expect the resulting agent to not only mimic humans' ability to model a listener's mental state but also to leverage this estimate to choose better instructions. Through empirical evaluation, we aim to answer the question:
> *Can an agent, equipped with a model of theory of mind, quickly adapt to a listener's language in a few-shot language coordination game?*
Our experiments answer in the affirmative in the both a referential game and a vision-language navigation setting.
# Few-shot Language Coordination
Consider again the example depicted in Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}. A speaker model observes the goal of the current task ("pick and cool wine bottle, place it on the dining table"), and sends messages to a listener model, in an attempt to help it finish the task. The listener model chooses actions to take given the latest and previous instructions. If the listener makes a mistake, such as going to the microwave instead of the refrigerator, the speaker should realize that the listener misunderstood some aspect of the high-level instructions in the first message. This misstep informs the speaker's model of the listener's ability, leading them to give lower-level instructions to help correct the trajectory. Through *several games*, the speaker gathers enough data to provide customized messages for individual listeners. This is only possible if a proper prior is provided to the speaker. The simplest prior can be hand-coded rules, e.g. if listener cannot understand abstract instructions, try simpler ones. However, to pursue a more general and powerful communication model (e.g. knowing when to simplify vs. rephrase), we study whether this kind of few-shot language coordination can be *learned* by playing with a *population* of listeners. This section proposes a method to construct meaningful populations.
## Asymmetric Speaker-Listener Games
Following previous work on communicative agents, we use goal-oriented language games as the test bed for few-shot language coordination. A general goal-oriented language game provides an environment where the participants uses language to communicate with each other to achieve the given goal. We consider the most basic setting of a simplified two-player shared-goal multi-round setup:
Environment:
: The environment is defined by \(\mathcal{O}\)bservation space, \(\mathcal{A}\)ction space, \(\mathcal{G}\)oal space and transition function \(E:\mathcal{O}\times \mathcal{A}\rightarrow\mathcal{O}\times \mathcal{G}\). At the start of each game, the environment provides the speaker with a goal and both participants with observations after each action is taken by the listener. A new game starts after the previous one succeeds or reaches a maximum number of steps.
Participants:
: The participants consist of a speaker and a listener sending and receiving natural language messages. After observing the goal, the speaker gives an instruction to the listener, and the listener performs an action in the environment. If the game is sequential, the speaker can also give an instruction after each action until the game is solved or the maximum number of steps is reached. The speaker is a message-and-action producing model defined by the *vocabulary* \(\Sigma\); the space of *observations* \(\mathcal{O}\); the space of *actions* \(\mathcal{A}\); and a *model* \(f:\mathcal{O}\times\mathcal{G}\rightarrow\Sigma^*\times \mathcal{A}\). The listener is an instruction-follower defined by the same vocabulary \(\Sigma\), observation space \(\mathcal{O}\), and space of *actions* \(\mathcal{A}\) as the speaker; and a *model* \(g: \Sigma^* \times \mathcal{O} \rightarrow \mathcal{A}\).
Multi-round Games:
: The pair of participants will play a *session* of \(N\) rounds of games, which are sampled independently. Different from single-round games (\(N\!=\!1\)) used in most previous work, the participants keep the memory of past games in the same session. Multi-round games are not only more general than single-round games, but are essential to few-shot language coordination, because participants have the opportunity to adapt to the interlocutors by learning from feedback during previous rounds.
Note that the listeners in this setting cannot directly observe the goal, so the speakers need to use instructions to inform the listeners about the final or intermediate goals of each game. Within \(N\) rounds, the speaker needs to adapt to the listener's knowledge to provide the most effective instructions.
## Population
coined the notion of "machine ToM", which is a model for tracking agents' behaviors. To train and evaluate its ability to adapt to different agents, they create populations of subject agents by using different neural architectures and random seeds for parameter initialization. Our design of populations draws inspiration from their work. However, to make the similarities and differences between agents controllable, we consider a population as a distribution over parameters neural listeners with the same architecture which have been trained on different datasets. A neural listener \(f_{\theta}^L: \mathcal{O}\times\mathcal{I}\rightarrow\mathcal{A}\) is a mapping from \(\mathcal{O}\)bservations and \(\mathcal{I}\)nstructions to \(\mathcal{A}\)ctions with parameter \(\theta\) of the neural networks. The parameters trained on dataset \(\mathcal{D}\) are: \[\theta_{\mathcal{D}} = \arg\min_{\theta} \mathcal{L}(f_L, \mathcal{D})\] In this way, the variation over listeners is mainly determined by the features of the dataset. By constructing datasets with different feature distributions, we control the listeners' language abilities. In the example of Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"}, the observations in language games are the items in the visual field, and the agent may perform any number of actions (e.g. "`go to cabinet`"). The speaker may provide natural language instructions, which can range from high-level (e.g. "cool the winebottle") to low-level (e.g. "get the wine bottle", "take it to the refrigerator", "put it in the refrigerator"). A listener that has never been trained on a particular variety of (usually high-level) instruction would have trouble performing the appropriate actions. This leads to an exponential population of listeners that are trained on datasets containing, or not containing, particular relevant instructions. Because of this, in order to effectively and concisely communicate, an effective speaker will have to judge the language abilities of its various partners in the population and adjust appropriately; we explain how we do so in the following section.
## Theory-of-mind Model
Theory-of-mind, the ability to build a model of one's conversational partners, is deemed to be crucial in socio-pragmatics theory. Drawing inspiration from this, we build a theory-of-mind model to learn to mimic the behavior of the listener within a short time window.
#### Mental State
Modeling mental states is the central concept in building a theory-of-mind. We define the mental state of the listener as the parameters of a neural model, the *ToM model*, that produces the same output for the same inputs as the listener: \(\forall x\in \Sigma^*, o \in \mathcal{O}, g_{\text{ToM}}(x, o;\theta_{\text{mind}}) \approx g(x, o;\theta)\). It should be noted that in the general case, particularly when different model architectures are used to represent the model itself and the ToM model, the mental state representations may not be unique or even exist. In other words, for any model \(\theta\) there may be more than one parameter setting \(\theta_{\text{mind}}\) that satisfies this condition, or there may be no \(\theta_{\text{mind}}\) that produces the exact same output
#### Building a Theory-of-mind
Learning a ToM model is reduced to inferring the mental state of the listener. For a given listener \(g\) with parameters \(\theta\) and ToM model \(g_{\text{ToM}}\), we seek a mental state representation \(\theta_{\text{mind}}\). In practice, we use identical neural architectures for both the listener and ToM Model. However, inferring the exact mental state is infeasible within few interactions. Therefore, we estimate \(g_{\text{ToM}}\) such that \[\theta_{\text{mind}} = \arg\min_{\theta'} \mathbb{E}_{o, m}\mathcal{L} (g_{\text{ToM}}(o, m;\theta'), g(o, m;\theta))\] It is straightforward to apply this definition of mental state in the psychological context for which it was originally proposed. The mental state \(\theta_{\text{mind}}\) is the representation of the listener's language abilities, which are not directly observable, and which are ultimately used for predicting the belief and behavior of the speaker. For example, in our first set of experiments we focus on referential game where the speaker describes the target in order to let the listener pick it out from distractors. We construct a population in which neural listeners with LSTMs and word embeddings have different language comprehension abilities for different languages. One of the possible representations controls the word embeddings in different languages: the mental state of a good language listener should have more meaningful word embeddings, while the one which cannot understand the language should have more random ones. Given that the speaker can acquire an accurate mental state for the listener, it can be used for predicting the probability of listener choosing the correct image when hearing descriptions in different languages. By choosing the one that yields the correct image with the highest probability, the speaker generates the descriptions which improve the referential game. On the other hand, high quality descriptions help the speaker better narrow down the language abilities of the listener. This is similar to the two-way interrelation between language and ToM in humans. Following this direction, we present a dynamic view of ToM by putting the observer inside the conversation, instead of the static view of, which uses ToM for tracking the behavior of the agent without interfering in the games. Our training procedure is presented in Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"}. We aggregate a dataset \(D_{\theta_{\text{mind}}}\) at each epoch, and update the parameters by optimizing the ToM model on the dataset. To aggregate the dataset for each training listener, we randomly sample from the posteriors of the ToM model and uniform distributions over the candidates, which keeps a certain degree of exploration, modulated by distribution coefficient \(\sigma\) (through the paper, we use \(\sigma = 0.5\)). In practice, parameters are updated with stochastic gradient descent by sampling listeners and using the history of each listener at each time step as a support set for predicting the next actions of the listener. Following the literature on speech acts, e.g. , we also add exponential cost penalty \(\exp(-\kappa C(m))\) as a prior to penalize long instructions. (We have not explored the space of penalty functions in this paper, but the exponential function is widely used in the pragmatics literature, e.g. ,.) In Fig. [\[fig:demo\]](#fig:demo){reference-type="ref" reference="fig:demo"} (a&b), although "go to fridge" yields the highest probability of gold action, no instruction is given in order to express the goal concisely. Similarly to the imitation learning algorithm DAgger, the dataset is collected using expert actions. However, there is a major difference between Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"} and DAgger---we optimize the prediction of actions conditioned on the observations and instructions instead of the instruction probability directly. The following theorem shows that our model will improve the instruction generation quality:
#### Discussion
The conditions of Theorem [\[th:1\]](#th:1){reference-type="ref" reference="th:1"} mean that the speaker model \(S\) must be a well-trained model to produce good enough candidates pools. In practice, this condition is not hard to meet: for instance, in our language navigation experiment the listeners can at least understand the lowest-level instructions, and the speaker generates four levels of instructions by rule-based experts. Therefore, the practical implication of this theorem is helpful--our method reduces to DAgger without expert instructions. Different from DAgger, our training method doesn't directly optimize the instruction distribution against expert's instructions, but optimizes the action prediction loss instead, which upper-bounds the instruction loss.
## Meta-learning ToM Model
To acquire an estimate of the mental state from very few interactions, the ToM model needs to quickly learn from a small support set. In theory, any model for parameterizing \(\mathcal{P}_{\text{ToM}}(a \mid o, m, \mathcal{D}_{\text{supp}};\theta_{\text{mind}})\) could work in our framework. As a general method applicable to all ToM models, we apply model-agnostic meta-learning (MAML; ), a method that explicitly trains models to be easily adaptable in few-shot settings. Given support dataset \(\mathcal{D}_{\text{supp}}\), the inner loop updates parameters for \(N_{\text{inner}}\) steps: \[\begin{aligned} \theta^{(0)} &= \theta_{\text{mind}} \\ \theta^{(i+1)} &= \theta^{(i)}-\eta \nabla_{\theta} \mathcal{L}(\mathcal{D}_{\text{supp}}; \theta)|_{\theta=\theta^{(i)}} \end{aligned}\] where \(\eta\) is the inner loop learning rate, which, in practice, is not share across modules following. The loss function on the support set is the negative log-likelihood of listener's action given observations \(o\) and the instructions \(m\) \[\mathcal{L}(\mathcal{D}_{\text{supp}}; \theta) =-\mathbb{E}_{(o, m, a)\sim\mathcal{D}_{\text{supp}}} \log p_{\theta}(a\mid o, m)\] After \(N_{\text{inner}}\) steps, we get the prediction on the target observation \(o\) and instruction \(m\) \[\mathcal{P}_{\text{ToM}}(a \mid o, m, \mathcal{D}_{\text{supp}};\theta_{\text{mind}}) = p_{\theta^{N_{\text{inner}}}}(a\mid o, m)\] Outer loop optimize \(\theta_{\text{mind}}\) by mini-batch stochastic gradient descent \[\theta_{\text{mind}} \leftarrow \theta_{\text{mind}}-\eta_{\text{outer}} \nabla_{\theta} \mathcal{L}^{\text{pred}} (\mathcal{D}_{\theta_{\text{mind}}})\] The outer loop also runs for a given \(N_{\text{outer}}\) epochs.
## Deploying the ToM Model
Similar to Proc. [\[proc:training_tom\]](#proc:training_tom){reference-type="ref" reference="proc:training_tom"}, we evaluate ToM by using it to measure the probability of the gold action given the instructions. However, here we choose the best one instead of sampling from the posterior. Alg. [\[proc:test_tom\]](#proc:test_tom){reference-type="ref" reference="proc:test_tom"} shows the evaluation procedure.
The gold-standard speaker denotes the success rate of using the testing listener in place of the ToM listener. The score of over \(90\%\) indicates that the candidate pool is of high quality, so a speaker with a well-modeled ToM listener has ample room for achieving high accuracy. The non-ToM speaker uses the instruction with the highest probability in the speaker model; the RSA speaker uses the listener for training the speaker in place of the ToM listener. Our model achieves a significantly higher success rate, demonstrating that the ToM model could help produce better instructions for this referential game. However, does ToM model truly learn to adapt to individual listeners? We compute the accuracy of predicting the listener's behavior during the same session. Fig. [\[fig:acc\]](#fig:acc){reference-type="ref" reference="fig:acc"} shows that the prediction accuracy of listener's actions is significantly improved within sessions, which shows ToM indeed learns to adapt to individual test listeners.
# ALFWorld Task Navigation
In the previous section, we have shown that ToM could help games with simple dynamics, and learn to adapt to listeners. This section will show its application in a more complex game: language navigation. We use the Alfworld platform, which creates a natural language command action space upon the Alfred environment for vision-language navigation. In each game, a household task is given by the environment. We generate expert trajectories for all tasks in Alfworld, and manually create four levels of instructions, from task-level to action-level, which are denoted as \(\mathcal{I}_i, i=1, 2, 3, 4\). The differences between these four levels are the levels of abstraction. An action-level instruction corresponds to a single action in the Alfworld, while a task-level instruction corresponds to a whole trajectory which consists of more than eight commands. The two other levels are in between. The candidate pool at each time step consists of four instructions from each level (\(N_M=4\)). To create each listener, we draw an instruction distribution from \(Dir(0.6, 0.4, 0.3, 0.2)\) for each of the six types of task in the environment. Listeners are of the same neural model as in. While training the listeners, the instructions are randomly drawn from the instruction distributions according to the task type. This procedure produces 50 listeners. The average success rate of listeners is 83.6%. These listeners are randomly divided into training, validation, and testing listeners (30/10/10). We define the cost as the average total length of the set of instructions, i.e. repetitive instructions are only calculated once. The cost function is defined as \(C(m) = 2^i \ \text{if } m \in \mathcal{I}_i\). Within one session, the maximum number of interactions between speaker and listener is \(K=100\), and maximum number of interactions in a game is 20. Listeners' hyper-parameters are the same as the ones in, while MAML hyper-parameters are the same as referential game. In Fig. [\[fig:language_navigation\]](#fig:language_navigation){reference-type="ref" reference="fig:language_navigation"}, we compare ToM-assisted speakers and random speakers. We didn't compare with an RSA speaker, because differently from the multilingual referential games, the speaker is rule-based and no listener is used for training the speaker. A random speaker draws an instruction distribution from \(Dir(0.7, 0.7, 0.7, 0.7)\), and sample instructions from the candidate pools using the instruction distribution. The ToM speaker with \(\kappa=0\) predominantly uses action-level instructions, while the ToM speaker with \(\kappa=10\) uses task-level instructions most of the time. Comparing ToM speakers and random speakers, we find that for \(\kappa = \{0, 1, 2\}\), ToM speakers achieve higher game points and lower cost than random ones; for \(\kappa=10\), the ToM speaker does not have significant improvement over a random one, since only the listeners that have been trained on sufficient action-level instructions can succeed.
# Related Work
## Language Games
Language games have been the proving ground for various linguistics theories since their conception by. Recently, the most widely used language game is the referential game, in which the speaker observes the target and distractors and uses language to instruct the listener on how to pick out the target. **Emergent Communication** Without natural language annotations, this pressure for the speaker and listener enables language emergence. first uses the same recurrent neural networks as the speaker and the listener to conduct emergent communication in referential game. Following this lead, study how emergent languages are grounded to the input images. studies multi-turn communication via negotiation. study the compositionally and systematicity of emergent languages. **Learning Common Languages** By using natural language annotations, agents learn a common language so that agents that are never trained together can be expected to communicate. studies using MSCOCO annotations as gold labels for both speakers and listeners. found that alternating between self-playing and supervised learning benefits communication performance. show that humans have the ability to adapt to the machine's capability in language games. found that when language follows Zipf law, zero-shot communication is possible. **Language Games with Community** learns representations by training with a community of encoders and decoders. The difference between our work and theirs is that our MAML listener learns to adapt to different listeners in the population in a few games. The performance of their model should be equivalent to our model's result at time step 1. considers the adaptation problem, which is definitely relevant. However, adapting their model to our settings is non-trivial, which is beyond the scope of this paper. This referential game setting used in most previous work can be seen as a special case of our few-shot coordination formulation, where the number of games is one and the partners are mostly the same ones as in the training phase. These two differences prevent the previous models from learning to adapt due to the lack of pressure to do so.
## Machine Theory of Mind
Computational and neural models of theory-of-mind have been studied for decades. are the first to present a successful modeling of the mental state of various species of agent. While we also train a ToM model with meta learning, we put the ToM model into use. The predictions provided by the ToM model serve as reranker in the speaker's model. The variety of models is also more diverse than the species used in this paper. Importantly, find that neural models for question answering fail to keep track of inconsistent states of the world. extends machine ToM to neural recursive belief states. We expect improvement over our current model by modeling higher-order recursive belief, which is left for future work. explicitly trains belief state prediction with supervised learning, while our model's belief state is latent.
## Similar Topics in Reinforcement Learning
**Model-based Reinforcement Learning** Model-based reinforcement learning focuses on building a model of the environment to improve data efficiency, which could be applied to the zero-shot coordination problem by treating the listener as a part of the environment. Recently, neural networks have been used widely for model-based RL. We should point out that despite their similarities to our model, we focus on modeling different and unseen agents in the population within a few interactions. **Alternatives to Self-play** Zero-shot coordination has attracted much attention recently. propose other-play which maximizes the expected reward working with random partners. In contrast to their approach, we explicitly model the partner's ToM and focus on language coordination, which required more complicated modeling than the environments in their experiments.
# Implications and Future Work
We have introduced few-shot language coordination task and proposed ToM model for tracking the listener's mental state. Different from previous work using single-round games and self-play training, we consider more general multi-round games and playing with novel listeners. ToM model shows its ability to adapt to novel listeners and assist speakers in choosing the best instructions in both multilingual referential games and the language navigation task. We attribute the success of ToM model to modeling socio-pragmatics process in an explicit way. Many interesting questions about modeling ToM in few-shot language coordination remain open. The most immediate is how to model ToM from the listener's perspective, leading to a dialog agent that can acquire new knowledge through conversation. One step further, similar to RSA, ToM can also be modeled in a recursive manner, which may further improve language games. | {'timestamp': '2021-07-14T02:03:03', 'yymm': '2107', 'arxiv_id': '2107.05697', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05697'} |
# Ethical Considerations {#ethical-considerations .unnumbered}
Research on active learning improves the labeling of data, by efficiently supporting the learning algorithm with targeted information, so that overall less data has to be labeled. This could contribute to creating machine learning models, which would otherwise be infeasible, either due to limited budget, or time. Active learning can be used for good or bad, and our contributions would---in both cases--show how to make this process more efficient. Moreover, we use pre-trained models, which can contain one or more types of bias. Bias, however, affects all approaches based on fine-tuning pre-trained language models, but therefore this has to be kept in mind and mitigated all the more.
# Introduction
Collecting labeled data for machine learning can be costly and time-consuming. A key technique to minimize labeling costs has been active learning, where an oracle (e.g., a human expert) is queried to label problem instances selected that are deemed to be most informative to the learning algorithm's next iteration according to a query strategy. Active learning is characterized by the real-world machine learning scenario in which large amounts of training data are unavailable, which may explain why comparably little research has investigated deep learning in this context. The recent widely successful transformer-based language models can circumvent the limitations imposed by small training datasets. Pre-trained on large amounts of unlabeled text, they can be fine-tuned to a given task using far less training data than when trained from scratch. However, their high number of model parameters renders them computationally highly expensive, for query strategies that are targeted at neural networks or text classification, resulting in prohibitive turnaround times between labeling steps. In this paper, we systematically investigate uncertainty-based query strategies as a computationally inexpensive alternative. Despite their relative disadvantages in traditional active learning, when paired with transformers, they are highly effective as well as efficient. Our extensive experiments assess a multitude of combinations including state-of-the-art transformer models BERT and DistilRoBERTa, five well-known sentence classification benchmarks, and five query strategies. | {'timestamp': '2022-03-22T01:33:13', 'yymm': '2107', 'arxiv_id': '2107.05687', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05687'} |
# Introduction
Understanding how the brain of mammals, including humans, represents, processes and stores information is one of the main challenges of contemporary science. In addition to the obvious direct interest in such an ambitious goal, any progress made towards elucidating the brain working principles would also help developing a new generation of artificial intelligence devises. Reversely, advances in computer science help shedding light on the analogies and differences between our present operational knowledge on \"artificial intelligence· and \"natural intelligence\". This two-sided dialogue is hoped to guide exciting breakthroughs in the next coming years in both fields. A popular idea, coming from the world of artificial neural networks and then exported to the realm of biological systems (see and refs. therein), is that information-processing complex systems, composed of many individual interacting units, are best suited to encode, respond, process, and store information if they operate in a dynamical regime nearby the critical point of a phase transition, i.e. at the edge between \"order\" and \"disorder\". In a nutshell, one can say that \"ordered phases\" encode information in a robust or stable way, but they are not flexible enough as to accommodate for or respond to input changes; on the other hand, \"disordered phases\" are dominated by noise, thus hindering information storage and making retrieval exceedingly difficult. Therefore, there needs to be some kind of trade-off between order and disorder that can be formulated in a number of different ways, e.g., between \"stability and responsiveness\" or between \"robustness and flexibility\". The criticality hypothesis poses that such a trade-off is best resolved near criticality or \"at the edge of chaos\", where combined advantages from the two alternative phases can be obtained. Furthermore, at critical points there is a concomitant scale invariance ---with its characteristic power-law distributions and scaling--- entailing the existence of broadly different time and length scales, which seem much convenient for the representation of multiscale complex inputs. Let us remark, that the terms \"criticality\" and \"edge of chaos\" are sometimes used indistinctly, though the last one applies to deterministic systems, in which a transition occurs between ordered and chaotic states, but as recently emphasized, they can describe two sides of the same coin (we refer to for an illuminating recent discussion). Empirical evidence that actual neural networks might operate close to criticality has kept accumulating in recent years. Most of this evidence (though not all) relies on the concept of neuronal avalanches which are empirically observed to be scale invariant across species, brain regions, resolution levels and experimental techniques. However, as of today, smoking-gun evidence is still needed to validate or dismiss this fascinating conjecture, and it remains controversial; thus, novel theoretical and data-oriented analyses are much needed. In a seemimgly-unrelated remarkable work, Stringer *e*t al. have recently made a step forward in understanding how neuronal networks actually represent complex inputs. In particular, these authors proved mathematically that the statistics of spiking neurons representing external sensory inputs (such as natural images represented in the mouse visual cortex) need to obey certain constraints for the input representation (or \"neural code\") to be \"continuous and differentiable\". These abstract mathematical properties are the formal counterpart of a much-desired property of neural networks: i.e., robustness of the representation against small perturbations of the inputs. Such a robustness is well-known to be often violated in artificial neural networks (ANNs); in particular, so-called, *adversarial attacks*, consisting in tiny variations in the input or their statistics, can fool the network, leading to wrong predictions and mis-classifications. We refer to Stringer *et al.* for an in-depth explanation and justification of these important ideas, as well as to for an application of them onto multi-layer ANNs. In any case, the conclusion of Stringer *et al.* is that, in order to achieve robust input representations, the covariance matrix of neuronal activities measured across time when the network is exposed to a sequential series of inputs, must obey the following spectral property: its rank-ordered eigenvalues should decay as a power law of their rank, with an exponent \(\alpha\) strictly larger than \(1+2/d\), where \(d\) is the embedding dimension of the input. Thus, \(\alpha=1\) sets a lower bound for the possible values of the eigenspectrum decay-exponent for complex, high-dimensional inputs. Rather remarkably, these theoretical predictions are verified to be fulfilled in experimental recordings of more than \(10000\) individual neurons in the mouse visual cortex exposed to a very large sequence of natural images. This confirms that information encoding occurs in as mathematically predicted, i.e. in a continuous and differentiable manifold. The main question we pose here is: are the internal representations of ANNs trained to classify images similar to those of the mouse visual cortex? More specifically, is the spectrum of eigenvalues of the associated covariance matrix a power law of the rank? Is the exponent in all cases larger than (and close to) \(1\)? If so, do the exponent values change with the images dimensionality in the way predicted by Stringer et al.? Here, in order to tackle all these questions within the simplest possible scenario, we analyze the neural encoding of inputs with different dimensions in a paradigmatic example of ANN: the *Echo state network* (ESN) . This type of networks, together with *liquid state machines*, constitute the prototype of *reservoir computing* (RC) approaches, a paradigm of computation that seems particularly well suited for exploiting the putative advantages of operating at the \"edge of chaos\".
# Materials and Methods
The Echo State Network, in its original formulation, was devised by Jaeger as a flexible and easy-trainable recurrent neural network (RNN) for time-series prediction tasks. More specifically, the architecture of ESNs is described as consisting of:
- An input layer, which scales a number \(L_{1}\) of inputs at each time step before they arrive in the reservoir, according to some random weights \(W^{in}\in\mathbb{R}^{N\times L_{1}}\).
- A reservoir consisting of \(N\) internal units connected with random weights \(W^{res}\in\mathbb{R}^{N\times N}\), whose states usually evolve according to a non-linear, time-discrete dynamical equation under the influence of a time-dependent input. In this way, the reservoir maps the external input into a high-dimensional space.
- An output layer, that converts the information contained in the high-dimensional states of the neurons (which serve as an internal representation of the inputs) to generate the final output.
Thus, unlike in other ANNs, the internal weights or \"synaptic connections\" in ESNs do not need to be updated during the learning process, and training is achieved by just modifying the layer of output weights that readout the network internal states. In order to adapt this architecture---usually employed in time-series analyses---for image classification tasks, we used black and white images with \(L_{1}\times L_{2}\) pixels (each of them characterized by a value in the \([0,1]\) interval, representing a normalized gray-scale) and converted them into multivariate time series by considering their vertical dimension as a vector of \(L_{1}\) elements or features, that \"evolve\" along \(T=L_{2}\) discrete \"time\" steps. One can then define a standard training protocol in which, as illustrated in Fig.[\[Figure_0\]](#Figure_0){reference-type="ref" reference="Figure_0"}, at each time \(t \in[0,T]\), vectors \(\mathbf{u}(t) \in[0,1]^{L_{1}}\) corresponding to columns of the image are fed as inputs to the ESN. In this way, the network dynamics for the reservoir states is given by the following non-linear activation function: \[\mathbf{x}(t)=\tanh(\varepsilon W^{in}\mathbf{u}(t)+W^{res}\mathbf{x}(t-1))\label{eq:States_Update}\] where \(\varepsilon\) is an input scaling factor. Using a supervised learning scheme, the goal of the ESN is to generate an output label \(\mathbf{y}\in\mathbb{\mathbb{N}}^{F}\) that correctly classifies each image in the test set as belonging to one of the \(F\) existing categories or classes. This label consists of a vector in which every element is zero, except for a value of one at the position corresponding to the assigned class (i.e., \"one-hot-encoded\" in the machine learning jargon). Several readout methods have been proposed in the literature to transform the information contained in the reservoir dynamics into the expected target output \(\mathbf{y}^{target}\in\mathbb{\mathbb{N}}^{F}\), ranging from linear regressions methods over the reservoir states, to the use of \"support vector machines\" or \"multilayer perceptrons as decoders. Here, we use a simple Ridge regression (see Appendix I for a detailed explanation of the algorithm) over the *\"reservoir model space\"*, a method that has been recently proposed for the classification of multivariate-time series. The \"reservoir model space\" is a set of parameters \(\theta_{x}\) that encodes the reservoir *dynamical* state for a given input (image). Such parameters are obtained from a Ridge regression to predict the next reservoir state from the past one at discrete time steps, \[\mathbf{x}(t+1)=W_{x}\mathbf{x}(t)+\mathbf{w}_{x},\] in such a way that \(\theta_{x}=\left[\mathrm{vec}(W_{x});\mathbf{w}_{x}\right]\in\mathbb{R}^{N\left(N+1\right)}\) provides a characterization of the internal reservoir dynamical state during the presentation of a given input, where \(\mathrm{vec}(\cdot)\) denotes reshaping to a one-column vector and \(";"\) vertical concatenation. Then, for each image a readout module or decoder can transform this internal representation into an output label: \[\mathbf{y}=W_{out}\theta_{x}+\mathbf{w}_{out}.\label{Eq_Readout}\] The parameters \(\theta_{out}=\left[\mathrm{vec}\left(W_{out}\right);\mathbf{w}_{out}\right]\)---where \(W_{out}\in\mathbb{\mathbb{R}}^{F\times N(N+1)}\) and \(\mathbf{w}_{out}\in\mathbb{\mathbb{R}}^{F}\) are defined as output weights and biases, respectively---are determined again through Ridge regression, minimizing the error between the produced and target label for all the presented images in the training set. Let us remark that the presented framework can be naturally extended to include, for instance, leakage and noise terms in Eq.[\[eq:States_Update\]](#eq:States_Update){reference-type="ref" reference="eq:States_Update"}, feedback connections from the output to the reservoir, or plastic rules that modify the reservoir weights according to the inputs , among other possible extensions. However and since our aim here is not to reach state-of-the-art classification accuracy---but rather highlight the link between optimal input representation and the internal dynamical state---for the sake of parsimony, we refrain from adding further features to our model for the time being.
# Results
## Non-trivial scaling at the edge of chaos.
Although relatively simple, our proposed ESN model has several hyperparameters than can be tuned, affecting its performance. More specifically, the *spectral radius* \(\rho\) of the reservoir internal weight matrix and the *scaling factor* \(\varepsilon\) of the input weights are two variables that usually determine the dynamical regime within the reservoir. The spectral radius---or largest eigenvalue of the reservoir weight matrix---controls the dynamical stability inside the reservoir when no input is fed into the network. Thus, a spectral radius exceeding unity has been often regarded as a source of instability in ESNs due to the loss of the so-called "*echo state property*", a mathematical condition ensuring that the effect of initial conditions on the reservoir states fades away asymptotically in time. Nevertheless, later studies have shown that the echo state property can be actually maintained over a unitary spectral radius, and different sufficient conditions have been proposed (see in particular, where the authors analyze the problem from the lens of non-autonomous dynamical systems, deriving a sufficient condition for the echo state property with regard to a given input). On the other hand, increasing the value of \(\varepsilon\) can convert an initially expanding mapping into a contracting dynamics, as stronger inputs tend to push the activities of the reservoir units towards the tails of the non-linearity. In what follows, we analyze the input representation that the reservoir codifies in terms of the trade-off between \(\rho\) and \(\varepsilon\), which *together* determine the dynamical operating regime of the ESN and the presence or absence of the echo state property. For the rest of parameters, the number of units in the reservoir is kept fixed to \(N=2000\) and the density of the reservoir-weight-matrix elements (i.e., the percentage of non-zero connections) to \(10\%\), while both reservoir and input weights are extracted at random from a uniform distribution in the interval \([-1,1]\). Following the same methodology as in Stringer *e*t al., the ESN was first presented with a large set of high-dimensional, natural images, and the activity of the internal units in the reservoir was stored for each step of the training. Then, principal component analysis (PCA) was performed directly over the full set of neuron activities \(X\in\mathbb{R}^{N\times( T\times M)}\), where \(T=90\) is the number of pixels in the horizontal dimension of the images and \(M=2800\) is the total number of images. In this way, we obtained the variance along each principal component or eigenvector of the covariance matrix, that serves as a basis for the activity inside the reservoir (see for a very gentle but rigorous introduction to PCA). Notably, we found that the spectrum of eigenvalues as a function of their rank (i.e., the variance associated to the n-th principal component, when ordered from the largest to the smallest) can be well fitted to a power-law \(n^{-\alpha}\) (see insets in Fig.[\[Figure_1\]](#Figure_1){reference-type="ref" reference="Figure_1"}), whose associated exponent \(\alpha\) decays (flatter spectrum) with the spectral radius \(\rho\); while, the exponent increases (faster decay) with the input scaling factor \(\varepsilon\) for most of the parameter space (see Fig. 2). In it was found that the exponent of this power-law relation is close to \(1\) when natural, high-dimensional images were shown to the mouse as an input. As discussed above, the authors proved mathematically that \(\alpha>1+2/d\) is a necessary condition for the neural manifold that emerges from the representation of a \(d\)-dimensional input to be continuous and differentiable. For natural images, \(d\) is very large and one can approximate the critical exponent by \(\alpha_{c}\approx1\) (this condition is marked by the purple plane in Fig.[\[Figure_1\]](#Figure_1){reference-type="ref" reference="Figure_1"}). As a following step, one might naturally wonder if there is some aspect of our model that characterizes such a regime of robust representations in the parameter space \((\rho,\varepsilon)\), for which an exponent \(\alpha\) close to unity is found. In other words: is the dynamics of the system inherently different in the regions for which the input representation manifold is found to be non-analytical (i.e. below the purple plane in Fig.[\[Figure_1\]](#Figure_1){reference-type="ref" reference="Figure_1"})? As it turns out, rather remarkably, the exponent \(\alpha\) characterizing the decay of the eigenspectrum approaches unity for choices of \(\rho\) and \(\varepsilon\) that drive the network dynamics towards the so-called \"edge- of instability\" or \"edge of chaos\", this is, near a transition point between an ordered and a chaotic regime. Traditionally, chaotic regimes are characterized by their average sensitivity to perturbations in the initial conditions; to quantify this effect, one usually measures the rate of divergence of two trajectories with a very small difference in their initial conditions: \[\lambda=\lim_{k\to\infty}\dfrac{1}{k}log\left(\dfrac{\gamma_{k}}{\gamma_{0}}\right)\]where \(\lambda\) is termed the maximum Lyapunov exponent (MLE), \(\gamma_{0}\) is the initial distance between the perturbed and unperturbed trajectories, and \(\gamma_{k}\) is the distance between the trajectories at step \(k\) (we refer the reader to and for a detailed explanation of the algorithm used to compute the MLE). Thus, chaotic dynamics is typically associated with a positive MLE, while the system is said to be stable to local perturbations provided \(\lambda<0\). It can be clearly seen from Fig.[\[Figure_1\]](#Figure_1){reference-type="ref" reference="Figure_1"} that the region in which one finds non-analytical representations of the input (below the purple plane) matches almost perfectly with the region (colored in green) in which a positive MLE is found. The transition order-to-chaos can be also visualized looking directly at the activities inside the reservoir (see Fig.[\[Figure_5\]](#Figure_5){reference-type="ref" reference="Figure_5"}). Observe that, when the network is in an \"ordered\" state, with \(\lambda<0\), the responses of the neurons are quite heterogeneous when compared among them, but they are highly localized within each neuron, i.e., individual neurons experience a limited response to stimuli. On the other hand, dynamical states characterized by \(\lambda>0\) have neurons whose response extends across the full range of the non-linearity (with higher probability along the tails, reflecting a saturated behavior), but it is this same \"phase space expansion\" that makes units almost indistinguishable from each other. It is only around the critical point or f-chaos, that we find a trade-off between dynamical richness in individual units and variability across units. Coming back to the results in Stringer *et al.*, one may also wonder whether the continuity and differentiability condition \(\alpha>1+2/d\) holds also for low-dimensional inputs, for which the expected bound \(\alpha_{c}=1+2/d\) deviates considerably from unity. To this purpose, Fig.[\[Figure_2\]](#Figure_2){reference-type="ref" reference="Figure_2"} shows the measured eigenspectrum of the reservoir activity covariance matrix (i.e. eigenvalues as a function of their rank) when images of different dimensionality (the same ones used by Stringer *et al.* in their experiments) are presented as inputs, and the reservoir is tuned to operate at the onset of a chaotic regime, i.e., for values in the parameter space \(\left(\rho,\varepsilon\right)\) for which \(\lambda\) was near zero but still negative. Remarkably, we find in all cases that *the exponents observed in the mouse visual-cortex activity are best reproduced when the reservoir dynamics is tuned close to the \"edge of chaos\"*. This finding suggests that one can set the network parameters in such a way that the neural activity manifold in which the input is represented is almost as high-dimensional as possible without loosing its \"smoothness\", and that such optimal solution is found at the edge of chaos. At this point, it is pertinent and timely to dig a bit deeper on the similarities and differences between the results presented in for real, V1-cortex neurons in the mouse, and the power-law exponents obtained through our reservoir computing model. (i) First of all, as in the case of real neurons, the observed correlations between the internal units are not just a byproduct emerging from scale-free features of natural images (see second column in Fig.[\[Figure_2\]](#Figure_2){reference-type="ref" reference="Figure_2"}). In particular, one can see that the power-law decay of the eigenspectrum persists even in response to low-dimensional inputs whose embedding vector space can be spanned with just a few principal components (i.e. without a power-law decaying intrinsic spectrum). (ii) In our model, images are processed sequentially in time along their horizontal dimension, so that for each image one can measure the activity of the \(N\) internal units over \(T=L_{2}\) time steps. In contrast, activity of V1 neurons in is scanned at a relatively low rate, so that for each image the neural representation in characterized by just one amplitude value in each neuron. (iii) To avoid confusion, let us remark that the variance observed by Stringer et al. is not directly measured over the raw activity of the neurons. Instead, the authors first project out the network spontaneous activity from the data, and then perform a cross-validated PCA (cvPCA) that allows them to filter out the trial-to-trial variability or \"noise\". The cvPCA method is thus able to estimate the stimulus-related variance confined in an n-dimensional manifold by first computing the eigenvectors spanning this manifold from a first repeat of the full training set, and then measuring the amount of a second repeat's variance that is confined to this plane (we refer to for a detailed explanation and derivation of the cvPCA method). However, since our model is completely deterministic for a given initialization of an ESN, the stimulus-related variance computed through cvPCA trivially matches that of a standard PCA. A natural question then arises from this last point: what happens when a noise term is included in Eq.[\[eq:States_Update\]](#eq:States_Update){reference-type="ref" reference="eq:States_Update"}, so that the dynamics becomes stochastic? Are the power-law exponents robust to the introduction of noise? To answer these questions, we considered stochastic versions of the ESNs ---including an independent small additive noise term in their inputs---and presented them with two repeats of the same input training set. The internal states of the noisy reservoirs were collected at each time step. We then performed the same type of cvPCA analyses proposed in to estimate the signal variance in our reservoirs (see fourth and fifth columns in Fig.[\[Figure_2\]](#Figure_2){reference-type="ref" reference="Figure_2"}). Just as in the case of real V1 neurons, the exponents measured over the raw, noisy activity are lower and below the critical threshold for continuity and differentiability of the neural manifold. Nevertheless, a cvPCA over the internal states retrieves the expected exponents after noise has been filtered out. We will further comment on the possible implications of this finding in the Discussion section, but for now, let us wrap up our findings tackling what we believe is a fundamental question from the perspective of machine learning: does working at the edge of chaos (or, equivalently, having maximal, continuous and differentiable neural manifolds) provide any functional advantage?
## Solving a bench-mark classification task.
The advantages of working at the so-called edge of chaos were first pointed in general dynamical systems and cellular automata, and only later analyzed in reservoir computer models with binary and analog internal units. In particular, in the authors showed that ESNs presented maximal information storage and transfer, as well as enhanced memory capacity right at the edge of chaos. However, while ESNs and other RC approaches have been previously applied to classification tasks with very good results, to the best of our knowledge an analysis of the influence of the dynamical regime on the performance of RC architectures for classification tasks is still missing. To this end, we measure the performance of ESNs in a classification task over the canonical MNIST dataset, which includes \(60,000\) handwritten instances of the first \(10\) digits as training (although only a third of them was used) and \(10,000\) for testing, following the training and testing procedure as described in Materials and methods. The results, showed in Fig.[\[Figure_4\]](#Figure_4){reference-type="ref" reference="Figure_4"}, highlight the fact that optimal performance (\(\sim 2.2 \%\) error rate) is found just below the onset of chaos, when \(\lambda \lesssim 0\). Most notably, the plot also evinces that the decay in performance is not only preceded by a positive MLE, but coincides too with exponents \(\alpha\) for the fit of the covariance-matrix eigenpectrum that are below the limiting value \(\alpha_{c}\approx1\), indicating the loss of continuity and differentiability of the neural representation manifold for high-dimensional images. Let us remark that, the slower the decay (i.e. the larger the exponent) the more weight is given to fine details of the input, but if the decay is too slow (smaller than the lower bound above), an excessive importance is given to such fine details at the cost of hampering the existence of a \"smooth manifold\" representation. Thus, operating near the edge of instability could provide the network with an optimal trade-off between representing as much details as possible and constructing operative, smooth representations. We finally remark that the results shown here were obtained with a reservoir consisting only of \(500\) internal units and using only one-third of the training set, with no pre-processing of the images. In contrast, the current best performance in MNIST digit recognition (\(0.81\%\) error rate) using reservoir computing networks has been achieved with a two-layer architecture, each consisting of \(16,000\) units, which amounts to a total \(880,000\) trainable parameters. In this sense, a simple ESN with readouts over the reservoir model space, when tuned near the edge of chaos, is able to outperform ESNs with a greater number of units and much more complicated dynamics (including feedback connections and leakage terms), trained over the full MNIST dataset.
# Discussion
Stringer *et al.* observed in that neural coding of different inputs in the mouse V1-cortex is close to optimal for each type of input, constrained by requirements of continuity and differentiability of the neural response manifold. In this paper, we open the door to the possibility that optimal, continuous and differentiable response manifolds emerge for neuron dynamics laying close to an edge of chaos type of critical point. Indeed, we have shown that a simple non-linear model of randomly-connected neurons, when subject to an external input, is able to reproduce power-law exponents similar to those found in mouse V1-cortex for the decay of the covariance matrix eigenspectrum. We find nevertheless important to clarify that the term edge of chaos---and the concept of chaos itself---should be taken with caution as it is not devoid of criticism in this context. As pointed out in, ESNs are an example of nonautonomous dynamical systems, for which typical concepts based in the theory of autonomous systems (e.g., \"sensitivity to initial conditions\", \"attractor\" and \"deterministic chaos\") do not directly apply. In fact, the authors of claim that local perturbation experiments cannot represent an ultimate evidence of chaotic dynamics in non-autonomous systems, since it might well be the case that the input drives the system towards and expanding dynamics for a certain time span, while the system shows on average a contracting, non-chaotic dynamics. Despite these caveats, at the light of the presented results it appears like there is indeed an actual dynamical phase transition occurring as the maximum Lyapunov exponent crosses zero. Thus, in any case, it seems a sensible choice to use such a quantity as a control parameter when analyzing the underlying neural representation of external inputs. On the other hand, adding stochasticity in the form of small-amplitude white noise naturally leads to flatter eigenspectra, much like those found when PCA is performed over raw experimental data. Nevertheless, one can use the same cvPCA technique introduced also in to extract the input-related variance of the activity, thus obtaining similar exponents to the fully deterministic case. This remarkable result therefore suggests that the role of spontaneous activity and trial-to-trial variability on the representation of external inputs can be easily accounted for in our simple echo-state-network model. Finally, results obtained on a benchmark classification task suggest that input-representation manifolds that are critically high-dimensional (from the point of view of their analytic properties) may serve a bigger purpose than just being a mathematical curiosity, as ESNs show a better performance when poised near such a critical point, while the accuracy falls rapidly as soon as the representation manifold becomes fractal. Therefore, the presented results open the path for very exciting research avenues at the boundary of biology and machine learning, calling for theoretical formulations that can shed light into the fascinating properties of these input-representation neural manifolds and their relation with the criticality hypothesis.
# Appendix A: Ridge regression {#appendix-a-ridge-regression .unnumbered}
Readouts in ESNs are typically linear, as it can be seen from Eq.[\[Eq_Readout\]](#Eq_Readout){reference-type="ref" reference="Eq_Readout"}. To simplify the forthcoming derivations, let us rewrite Eq.[\[Eq_Readout\]](#Eq_Readout){reference-type="ref" reference="Eq_Readout"} as: \[\mathbf{y}=\tilde{W}_{out}\left[\theta_{x};1\right]\] where \(\tilde{W}_{out}\in\mathbb{\mathbb{R}}^{F\times\left(N(N+1)+1\right)}\) and \"; \" indicates vertical vector concatenation. Thus, for a given input image a reservoir representation of the states \(\theta_{x}\) is constructed, and one can use the above equation to generate the corresponding output (which in our case will be a one-hot-encoded label \(\mathbf{y}\in\mathbb{\mathbb{N}}^{F}\) classifying the image). We can trivially generalize the above equation to apply to the full set of \(M\) images in the training set \(Y\in\mathbb{\mathbb{N}}^{F\times M}\): \[Y=\tilde{W}_{out}\Theta_{x}\] where \(\Theta_{x}\in\mathbb{R}^{\left(N\left(N+1\right)+1\right)\times M}\) contains as columns the vectors \(\left[\theta_{x};1\right]\) generated for each of the \(M\) input images. Finding the optimal weights \(\tilde{W}_{out}\) that minimize the squared error between the produced and target labels, \(\mathbf{y}\) and \(\mathbf{y}^{target}\), is then reduced to a standard linear regression problem, which is the greatest strength of the reservoir-computing approach: \[Y^{target}=\tilde{W}_{out}\Theta_{x}.\] Owing to the fact that large output weights are commonly associated to over-fitting of the training data, it is a common practice to add a regularization term to the error in the target reconstruction, usually defined in terms of the root-mean-squared error (RMSE). Although several methods have been proposed to achieve this regularization, one of the most efficient and stable algorithms is Ridge regression, which aims to solve: \[\tilde{W}_{out}=\underset{\left\{ \tilde{W}_{out}\right\} }{\arg\min}\dfrac{1}{M}\sum_{n=1}^{M}\sum_{i=1}^{F}\left(y_{i}[n]-y_{i}^{target}[n]\right)^{2}+\beta\left\Vert \tilde{w}_{i}^{out}\right\Vert ^{2}=Y^{target}\Theta_{x}^{T}\left(\Theta_{x}\Theta_{x}^{T}+\beta I\right)^{-1},\label{eq:MSE_Error}\] where \(\left\Vert \text{\ensuremath{\cdot}}\right\Vert\) stands for the Euclidian norm, \(I\) is the identity matrix and \(\beta\) is the regularization coefficient. Notice that choosing \(\beta=0\) removes the regularization, turning the Ridge regression into a standard generalized linear regression problem (we used \(\beta = 1\) across all simulations in the paper). We finally remark that, in order to obtain the reservoir model space parameters \(\Theta_{x}\), the same Ridge regression is also implemented to solve Eq. 2 for each of the input images.
# Appendix B: Phase space of reservoir units {#appendix-b-phase-space-of-reservoir-units .unnumbered}
In Fig.5, we reproduce the phase space of \(4\) different internal units of a reservoir operating (from top to bottom lines) in a sub-critical, critical and super-critical regimes, respectively. Each plot represents the activity of the corresponding neuron at each time step against the total input (sum of external input plus reverberating activity in Eq.[\[eq:States_Update\]](#eq:States_Update){reference-type="ref" reference="eq:States_Update"}) arriving to the neuron at the previous time. Since each image is first transformed into a multivariate time series of \(T=L_{2}=90\) time steps---and we plot the activity along the first three images only---each panel in Fig.[\[Figure_5\]](#Figure_5){reference-type="ref" reference="Figure_5"} contains 270 points. | {'timestamp': '2021-07-14T02:03:27', 'yymm': '2107', 'arxiv_id': '2107.05709', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05709'} |
null | null |
null | null |
null | null |
# Introduction
Frames which are an expansion of bases in Hilbert spaces, were first proposed by Duffin and Schaeffer to deal with nonharmonic Fourier series in. Indeed, frames are sets of vectors in a Hilbert space that yield one representation for each vector in the space, but which may have many different representations for a given vector. They play a fundamental role in mathematics, filter bank theory, coding and communications, signal processing, system modeling and model sensor networks (e. g. ) and they have some generalizations as fusion frames, g-frames, c-frames, \(K\)-frames, controlled frame, woven and etc. The notation of fusion frames have been introduced by Casazza and Kutyniok in, provide a means to improve robustness or develop feasible reconstruction algorithms. They were able to construct robustness of these frames under erasures in by using the methods suggested by Bodman for Parseval frames under the term weighted projective resolution of the identity for optimal transmission of quantum states and for packet encoding. Recently, controlled frames have been presented by Balazs and et al. in to improve the numerical efficiency of interactive algorithms for inverting the frame operator on Hilbert spaces. After, controlled frames have been studied for another kind of frames in. This manuscript is organized as follows: In Section 2, we study a new identity for the eigenvalues of the controlled frame operator and review the notation of controlled fusion frames with some results about these operators. In Section 3, we focus on the study of robustness of controlled fusion frames under erasures. Finally, we will present an interesting result about the composition of the operators of two controlled fusion frames. Throughout this paper, \(H\) and \(K\) are separable Hilbert spaces, \(\mathcal{B}(H,K)\) is the family of the all bounded linear operators on \(H\) into \(K\). When \(H=K\) we use \(\mathcal{B}(H)\) instead of \(\mathcal{B}(H,H)\). We denote \(\mathcal{GL}(H)\) as the set of all bounded linear operators which have bounded inverses and \(\mathcal{GL}^{+}(H)\) is the set of all positive operators in \(\mathcal{GL}(H)\). Also, we define \(\pi_W\) as the orthogonal projection from \(H\) onto a closed subspace \(V\subseteq H\).
# Review of Controlled Frames
In this section, we review the notations of controlled frames, fusion frames and present a useful identity about the eigenvalues of the controlled frame operator. Throughout this paper, \(C, C'\in\mathcal{GL}(H)\) and \(\Bbb I\subset\Bbb Z\).
In particular, we say \(\lbrace f_{i}\rbrace_ {i\in\Bbb I}\) a \(C\)-controlled frame for \(H\) if there exist constants \(0< A \leq B< \infty\) such that for each \(f \in H\), \[\begin{aligned}
A \Vert f \Vert^{2} \leq \sum _{i\in\Bbb I}\langle f,f_i\rangle\langle Cf_i, f\rangle\leq B \Vert f\Vert^{2}. \end{aligned}\] When \(A=B=1\), we say \(\lbrace f_{i}\rbrace_ {i\in\Bbb I}\) a Parseval \((C,C')\)-controlled frame and if the right hand of [\[con1\]](#con1){reference-type="eqref" reference="con1"} holds, \(\lbrace f_{i}\rbrace_ {i\in\Bbb I}\) is called a \((C,C')\)-controlled Bessel sequence. In this case, the controlled frame operator is defined by \[S_{CC'}f=\sum_{i\in\Bbb I}\langle C'f, f_i\rangle C^*f_i,\qquad f\in H.\] The next result gives a new identity about the eigenvalues of \(S_{CC'}\).
Next, we focus on controlled fusion frames for a Hilbert space which have been introduced by Khosravi and Musazadeh in.
Throughout this paper, \(W=\lbrace (W_{i},v_{i})\rbrace _ {i \in I}\) unless otherwise stated. We call \(W\) a tight \((C,C')\)-controlled fusion frame, if \(A=B\) and a Parseval \((C,C')\)-controlled fusion frame provided \(A=B=1\). We call \(W\) is a \(C^2\)-controlled fusion frame if \(C=C'\). If only the second inequality is required, We call \((C,C')\)-controlled Bessel fusion sequence with bound \(B\). If \(W\) is a \((C,C')\)-controlled fusion frame and \(C^*\pi_{W_i}C'\) is a positive operator for each \(i\in\Bbb I\), then \(C^*\pi_{W_i}C'=C'^*\pi_{W_i}C\) and we have \[\begin{aligned}
A \Vert f \Vert^{2} \leq \sum _{i\in\Bbb I} v_{i}^{2} \Vert(C^* \pi_{W_{i}} C')^{\frac{1}{2}}f\Vert^2 \leq B \Vert f\Vert^{2}. \end{aligned}\] We define the controlled analysis operator by (for more details, we refer to ) \[\begin{aligned}
T_{W}&:H \rightarrow \mathcal{K}_{2,W} \\ T_{W}f=&\lbrace v_{i}(C^{*}\pi_{W_{i}}C')^{\frac{1}{2}}f\rbrace_{i\in\Bbb I}, \end{aligned}\] where \[\begin{aligned}
\mathcal{K}_{2,W}:=\big\lbrace \lbrace v_i(C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f\rbrace_{i\in\Bbb I} \: \ f\in H\big\rbrace \subset (\bigoplus_{i \in\Bbb I} H)_{l^{2}}. \end{aligned}\] It is easy to see that \(\mathcal{K}_{2,W}\) is closed and \(T_{W}\) is well defined. Moreover \(T_{W}\) is a bounded linear operator with the adjoint operator \(T ^{*}_{W}\) (the controlled synthesis operator) defined by \[\begin{aligned}
T^{*}_{W}&:\mathcal{K}_{2,W} \rightarrow H \\ T ^{*}_{W}\lbrace v_i(C^{*}&\pi_{W_{i}} C')^{\frac{1}{2}}f\rbrace_{i\in\Bbb I}=\sum _{i\in\Bbb I} v_{i}^{2}C^{*}\pi_{W_{i}} C'f. \end{aligned}\] Therefore, we define the controlled fusion frame operator \(S_{W}\) on \(H\) by \[\begin{aligned}
S_{W}f=T ^{*}_{W}T_{W}(f)=\sum _{i\in\Bbb I} v_{i}^{2}C^{*}\pi_{W_{i}} C'f. \end{aligned}\] It is easy to check that \(S_W\) is a well defined and \[AId_H\leq S_W\leq B Id_H,\] hence, \(S_W\) is a bounded, invertible, self-adjoint and positive linear operator. The next result provides some conditions to get a controlled fusion frame.
# Erasures of subspaces and the approximation operator
First, we provide sufficient conditions on the weights for a subspace to be deleted yet still leave a controlled fusion frame with the same method in. Let \(\Bbb J\subset\Bbb I\) and for any \(i\in\Bbb J\) we define \[\mathcal{M}_i=\{f\in H: \ (C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f=f\}.\] It is easy to check that \(\mathcal{M}_i\) is a closed subspace of \(H\) for each \(i\in\Bbb J\). Indeed, if \(\{f_n\}\) is a sequence in \(\mathcal{M}_i\) such that \(f_n\rightarrow f\) and \(\varepsilon>0\), then \[\Vert(C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f-f\Vert\leq\Vert(C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f-(C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f_n+f_n-f\Vert<\varepsilon.\]
The following corollary yields immediately when one single subspace can be deleted.
Next, we make the reconstruction error for controlled fusion frames. The optimality under erasures has been presented for frames in and for Parseval fusion frames have been studied in. Suppose that \(\Bbb I=\{1,2,\cdots, m\}\) is finite and \(W\) is a Parseval \((C,C')\)-controlled fusion frame for \(H\) where \(\dim H<\infty\). For every \(i_0\in\Bbb I\), we consider the following operator: \[\begin{aligned}
\mathcal{D}&_{i_0}:\mathcal{K}_{2,W}\longrightarrow \mathcal{K}_{2,W},\\ \mathcal{D}_{i_0}\lbrace &v_i(C^{*}\pi_{W_{i}} C')^{\frac{1}{2}}f\rbrace_{i\in\Bbb I}=\delta_{i,i_0}v_{i_0}(C^{*}\pi_{W_{i_0}} C')^{\frac{1}{2}}f. \end{aligned}\] We define the associated 1-erasure reconstruction error \(\mathcal{E}_1(W)\) to be \[\mathcal{E}_1(W)=\max_{i\in\Bbb I}\Vert T^*_{W}\mathcal{D}_i T_{W}\Vert.\] Since \[\Vert T^*_{W}\mathcal{D}_i T_{W}\Vert=\sup_{\Vert f\Vert=1}\Vert T^*_{W}\mathcal{D}_i T_{W}f\Vert=v_i^2\sup_{\Vert f\Vert=1}\Vert C^*\pi_{W_i}C'f\Vert\leq v_i^2\Vert C\Vert\Vert C'\Vert,\] therefore, \[\mathcal{E}_1(W)=\max_{i\in\Bbb I}v_i^2\Vert C\Vert\Vert C'\Vert.\]
Finally, we study the composition of the controlled synthesis and analysis operators of two controlled fusion frames. This topic has been studied in for continuous frames and for g-fusion frames have been discussed by the author in.
Suppose that \(W=\{W_i\}_{i\in\Bbb I}\) and \(Z=\{Z_i\}_{i\in\Bbb I}\) are two sequences of closed subspaces of \(H\) and \(\{v_i\}_{i\in\Bbb I}\) is a set of weights. If \(C^*\pi_{W_i}C'\) and \(C^*\pi_{Z_i}C'\) are positive operators for each \(i\in\Bbb I\), we define the \(CC'\)-approximation operator \[\begin{aligned}
\Phi&: H\longrightarrow H,\\ \Phi f=&\sum_{i\in\Bbb I}v_i(C^*\pi_{Z_i}C')^{\frac{1}{2}}\big(v_i(C^*\pi_{W_i}C')^{\frac{1}{2}}f\big). \end{aligned}\] | {'timestamp': '2021-07-14T02:03:29', 'yymm': '2107', 'arxiv_id': '2107.05711', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05711'} |
# Analysis of Social Tie Evolution
## Structural Balance with \(W(t)\) {#subsec:structural}
## Solution to the General Riccati Equation {#subsec:general-Riccati}
We first focus on solving a general form of the matrix Riccati equation as stated below. \[\label{eqn:Evolution} \left\{ \begin{array}{rl} W' & = W^2 + C \\ W(0) & = B. \end{array} \right.\] The equation in our co-evolution model, i.e., Equation ([\[eqn:Riccati\]](#eqn:Riccati){reference-type="ref" reference="eqn:Riccati"}), satisfies \(C+B^2=V(0)V(0)^T\). Notice that the right-hand side \(V(0)V(0)^T\) is an \(n \times n\) matrix with rank one, which is a special condition. The analysis in this subsection applies for general matrices \(B, C\). By using a result in Reid , we can turn the matrix Riccati equation to a linear ODE system.
From Lemma [\[lemma:recatti\]](#lemma:recatti){reference-type="ref" reference="lemma:recatti"}, we can focus on solving the linear Ordinary Differential Equation (ODE) in Equation ([\[eqn:Rccati\]](#eqn:Rccati){reference-type="ref" reference="eqn:Rccati"}), which can be written in a matrix form. The analysis below is new. \[\label{eqn:matrix} \left[ \begin{array}{c} Y \\ Z \end{array} \right]' = \left[ \begin{array}{cc} 0 & I \\-C & 0 \end{array} \right] \left[ \begin{array}{c} Y\\ Z \end{array} \right],\] where we define \(A = \left[ \begin{array}{cc} 0 & I \\-C & 0 \end{array} \right]\in \mathbb{R}^{2n \times 2n}\) in block form and \(X = \left[ \begin{array}{c} Y\\ Z \end{array} \right] \in \mathbb{R}^{2n \times n}\). Now, let us solve the evolution equation \(X' = AX\), where \(X(0) = \left[ \begin{array}{c} I\\-B \end{array} \right]\). It is well known that the solution is, \[\label{eqn:X(t)} X(t) = (\sum_{n=0}^{\infty}\frac{t^n A^n}{n!}) X(0).\] Let us compute \(A^n\) using the block multiplication of matrices . \[A^2 = \left[ \begin{array}{cc} 0 & I \\-C & 0 \end{array} \right] \left[ \begin{array}{cc} 0 & I \\-C & 0 \end{array} \right] = \left[ \begin{array}{cc}-C & 0 \\ 0 &-C \end{array} \right].\] It implies: \[A^{2n} = \left[ \begin{array}{cc} (-1)^n C^n & 0 \\ 0 & (-1)^n C^n \end{array}, \right]\] \[A^{2n+1} = A^{2n}\cdot A = \left[ \begin{array}{cc} (-1)^n C^n & 0 \\ 0 & (-1)^n C^n \end{array} \right] \cdot \left[ \begin{array}{cc} 0 & I \\-C & 0 \end{array} \right] = \left[ \begin{array}{cc} 0 & (-1)^n C^{n} \\ (-1)^{n+1}C^{n+1} & 0 \end{array} \right].\] Recall that \(W =-ZY^{-1}\). Now we are ready to solve for \(W\).
## When \(B, C\) are both symmetric {#subsec:socialtie}
In our model, we assume that the initial tie matrix \(W(0)\) is symmetric. Thus the social tie evolution follows \(W' = W^2 + C\), where \(W(0) = B\), both \(B\) and \(C\) are symmetric. We are able to derive more detailed closed form solutions for the matrix Riccati equation in this setting. Since \(C\) is symmetric, there exists an orthogonal matrix \(U\) such that \(U^TCU\) is a diagonal matrix: \(U^TCU=\diag\{a^2_1, \cdots, a^2_k,-d^2_1, \cdots,-d^2_l, 0, \cdots, 0\}\), where \(a_i>0, d_j>0\). Furthermore, if \(BC=CB\), by the simultaneous diagonalization theorem, we may choose \(U\) such that both \(U^TCU\) and \(U^TBU\) are diagonal. By Lemma [\[conjugation\]](#conjugation){reference-type="ref" reference="conjugation"}(1), without loss of generality, we are going to solve the equation with the initial opinion vector \(U^TV(0)\) and initial weight matrix \(U^TW(0)U\). This leads to a system as below \[\label{eqn:evolution-diagonalized} \left\{ \begin{array}{rl} (U^TWU)' & = (U^TWU)^2 + U^T C U \\ U^T W(0)U & = U^TB U. \end{array} \right.\] The solution of this system can be easily transformed back to the solution to the original system by conjugation. For simplicity, the \((i,j)\)-th entry of a matrix \(M\) will be denoted by \(M_{ij}\). Define \(\beta_{ii}\) the \(i\)th diagonal element of \(U^TBU\), i.e., \(\beta_{ii}=(U^TBU)_{ii}\). Let us now work out explicitly the matrices \(Y\) and \(Z\) in Theorem [\[mainth\]](#mainth){reference-type="ref" reference="mainth"}.
1. For the positive eigenvalue \(a_i^2\) of \(C\), \[\begin{split} &\sum_{n=0}^{\infty}\frac{(-1)^n t^{2n} a_i^{2n}}{(2n)!} = \cos{(a_i t)} \\ &\sum_{n=0}^{\infty}\frac{(-1)^{n+1} t^{2n+1} a_i^{2n}}{(2n+1)!} =-\frac{1}{a_i}\sin{(a_i t)}. \end{split}\] Thus, \((U^TYU)_{ii} = \cos{(a_i t)}-\frac{1}{a_i} \sin{(a_i t)}\cdot \beta_{ii}\). Similarly, we have \[\begin{split} & \sum_{n=0}^{\infty}\frac{(-1)^{n+1} t^{2n} a_i^{2n}}{(2n)!} =-\cos{(a_i t)} \\ & \sum_{n=0}^{\infty}\frac{(-1)^{n+1} t^{2n+1} a_i^{2n+2}}{(2n+1)!} =-a_i \sin{(a_i t)}. \end{split}\] So, \((U^TZU)_{ii} =-\cos{(a_i t)}\beta_{ii}-a_i \sin{(a_i t)}\).
2. For the zero eigenvalues of \(C\), i.e., \(c_i = 0\), we have \((U^TYU)_{ii} = 1-t\cdot \beta_{ii}\) and \((U^TZU)_{ii} =-\beta_{ii}\).
3. For the negative eigenvalue \(c_i =-d_i^2\), we have \[\begin{split} (U^TYU)_{ii} = \sum_{n=0}^{\infty} \frac{t^{2n}d_i^{2n}}{(2n)!} + (-1)\sum_{n=0}^{\infty}\frac{t^{2n+1} d_i^{2n}}{(2n+1)!}\beta_{ii} = \cosh{(d_i t)}-\frac{1}{d_i}\sinh{(d_i t)}\beta_{ii} \\ (U^TZU)_{ii} = \sum_{n=0}^{\infty} (-1)\frac{t^{2n} d_i^{2n}}{(2n)!}\beta_{ii} + \sum_{n=0}^{\infty} \frac{t^{2n+1}d_i^{2n+2}}{(2n+1)!} =-\cosh{(d_i t)} \beta_{ii} + d_i \sinh{(d_i t)}. \end{split}\]
So we can summarize the above formulas as: \(U^TYU = D_1-D_2 U^T BU\) and \(U^TZU = D_3-D_1 U^TBU\), where \[\begin{aligned}
D_1&=&\diag\{\cos{(a_1 t)}, \cdots,\cos{(a_k t)}, \cosh{(d_1 t)}, \cdots, \cosh{(d_l t)}, 1, \cdots, 1 \}, \\ D_2&=&\diag\{\sin(a_1 t)/a_1, \cdots,\sin(a_kt)/a_k, \sinh(d_1 t)/d_1, \cdots, \sinh(d_l t)/d_l, 1, \cdots, 1 \},\\ D_3&=&\diag\{-a_1\sin(a_1 t), \cdots,-a_k\sin(a_kt), d_1\sinh(d_1 t), \cdots, d_l\sinh(d_l t), 0, \cdots, 0 \}, \end{aligned}\]
## When both \(B, C\) are symmetric and \(BC=CB\) {#subsec:BCequalCB}
Now, we consider the special case that \(BC=CB\). Using a basic fact that two commuting symmetric matrices can be simultaneously orthogonally diagonalized . Let \(U\) be an orthogonal matrix such that \[\begin{split} U^TCU & = \diag(a_1^2, \cdots, a_k^2,-d_1^2, \cdots,-d_l^2, 0, \cdots, 0)\\ U^TBU & = \diag(\lambda_1, \cdots, \lambda_k, \mu_1, \cdots, \mu_l, \delta_1, \cdots, \delta_h). \end{split}\] where \(a_i, d_j > 0\). In this case we can further simplify the solution. \[\label{eqn:WWW} \begin{aligned} W = &-Z \cdot Y^{-1} \\ = & U\diag(\frac{a_1 \sin{(a_1 t)} + \cos{(a_1 t)}\lambda_1}{\cos{(a_1 t)}-\frac{1}{a_1}\sin{(a_1 t)}\lambda_1}, \cdots, \frac{a_k \sin{(a_k t)}) + \cos{(a_k t)}\lambda_k}{\cos{(a_k t)}-\frac{1}{a_k}\sin{(a_k t)}\lambda_k},\\ &-\frac{d_1 \sinh{(d_1 t)}-\cosh{(d_1 t)}\mu_1}{\cosh{(d_1 t)}-\frac{1}{d_1}\sinh{(d_1 t)}\mu_1}, \cdots,-\frac{d_l \sinh{(d_l t)}-\cosh{(d_l t)}\mu_l}{\cosh{(d_l t)}-\frac{1}{d_l}\sinh{(d_l t)}\mu_l}, \\ & \frac{\delta_1}{1-t\delta_1}, \cdots, \frac{\delta_h}{1-t\delta_h})U^T. \end{aligned}\] Note that above equation for \(W(t)\) implies that \(V(t)\) is an eigenvector of \(W(t)\) for all time \(t\). Now we are ready to analyze the behavior of \(W\) over time in the case of a symmetric initial condition matrix \(W(0)\). We start with a technical lemma.
By Equation ([\[eqn:W\]](#eqn:W){reference-type="ref" reference="eqn:W"}) and Lemma [\[lemma:sin\]](#lemma:sin){reference-type="ref" reference="lemma:sin"}, we know that some diagonal entry of \(W\) converges to the infinity at a finite time \(T\) under some appropriate conditions.
## Structural Balance when \(BC=CB\) {#subsec:structural-commute}
In the following, we prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} which characterizes the conditions for structural balance, when \(BC=CB\).
Now we are ready to discuss our co-evolution model (Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"})). First, the condition \(BC=CB\) when \(B=W(0)\) and \(C=V(0)V(0)^T-W(0)W(0)^T\) means that \(V(0)\) is an eigenvector of \(W\). We finish the proof here.
For Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"}), we have an additional condition \(B^2+C = V(0)V(0)^T\) which has rank one. In the discussion below, we need the following fact about the eigenvalues and eigenvectors of rank-1 symmetric matrices \(H=uu^T\) where \(u \in \mathbb {R}^n-\{0\}\). The matrix \(H\) has eigenvalues \(||u||^2\) and \(0\) such that the associated eigenvectors are \(u\) and \(z\)'s which are perpendicular to \(u\), i.e., \(u^Tz=0\). Therefore, the unique largest positive eigenvalue is \(||u||^2\) with the associated eigenvector \(u\).
## Extensions and Conjectures {#subsec:general}
**General Graphs** The discussion so far has assumed a complete graph. For a connected graph \(G\), with our evolution model, our conjecture is that all the weights and opinions will go to extreme values. This appears to be the case for all simulations we have run. The reason is that it is hard for the weights to converge to finite values. One exception is that all the opinions are \(0\). But this is not a stable state. Any small perturbation on the opinion will break this stable state. At convergence, however, there might be more than \(2\) communities. The reason is that the edges with negative weights may not be in any triangle of the graph. Thus, multiple communities can be separated by the negative edges and it does not break the structure balance requirement. **High Dimensional Opinions** When each node has multiple opinions on different issues, its opinion can be represented as a \(m\)-dimensional vector, where \(m\) is the number of opinions in one node. Each entry of the weight matrix is a \(m\times m\) matrix instead of a real number. For the high dimensional setting, our equation for a general graph is: \[\label{eqn:derivation} \left\{ \begin{array}{l} V_i' = \sum_{j \sim i}W_{ij}V_j \\ W_{ij}' = V_i V_j^T, i \sim j \end{array} \right.\] where \(i\sim j\) means that there exists an edge between nodes \(i\) and \(j\). Then, we have the following theorem:
Based on our evolution equation and previous properties in the \(1\)-dimension opinion case, we make the following conjecture: For a complete graph with self-loop edges, all the opinion vectors and the weight matrix will converge to extreme values. Any two adjacent nodes \(V_i(t)\) and \(V_j(t)\) have the same opinion or the complete opposite opinion, i.e., \(V_i(t) = V_j(t)\) or \(V_i(t) =-V_j(t)\) as time approaches \(T_0\). Based on the above conjecture, each entry of \(\lim_{t \to T_0} W_{ij}(t)\) should have the same sign as that of \(\lim_{t \to T_0} V_i(t)V_j(t)^T\).
# Introduction
We live in a continuously changing world in which social interactions dynamically shape who we are, how we view the world and what decisions to make. Various social processes, naturally intertwined, operate on both the properties of individuals and social ties among them. Social influence, for example, describes how people's behaviors, habits or opinions are shifted by those of their neighbors. Social influence leads to homophily ( similarity of node attributes between friends) and leaves traces in the network structure such as high clustering coefficient and triadic closure (there are likely social ties among one's friends). Social influence and homophily, however, do not fully interpret the global network structure. One of the widely observed structural properties in social networks is the community structure. Nodes within the same community are densely connected and nodes from different communities are sparsely connected. Community detection is an important topic in social network analysis and has been investigated extensively . But why does community structure emerge and become a persistent feature? Are there social processes that encourage or maintain the community structure? In the literature, there have been a lot of studies of opinion dynamics, where people's behaviors, habits or opinions are influenced by those of their neighbors. Network models that capture social influence, such as French-DeGroot model , Friedkin-Johnsen model [@Friedkin1990-wl], Ku\(\l\)akowski et al. model naturally converge to global consensus. Actually it seems to be non-trivial to create non-homogeneous outcomes in these models, while in reality social groups often fail to reach consensus and exhibit clustering of opinions and other irregular behaviors. Getting a model that may produce community cleavage or diversity often requires specifically engineered or planted elements in a rather explicit manner. For example, bounded confidence models limit social influence only within pairs with opinions sufficiently close. Other models introduce stubborn nodes whose opinions remain unchanged throughout the process. The extreme case of community structure is dictated by structural balance , a common phenomenon observed in many social relationships. The notion was first introduced in a seminal paper by Heider in the 1940s in social psychology. It describes the stability of human relations among three individuals when there are only two types of social ties: positive ties describe friendship (or sharing common opinions) and negative ties describe hostility (or having opposite opinions). Heider's axioms state that among three individuals, only two kinds of triangles are stable: the triangles where all three ties are positive, indicating all three individuals are mutual friends; and the triangle with two negative edges and one positive edge, describing the folklore that "the enemy of your enemy is your friend." The other two types of triangles (e.g., a triangle of one negative tie and two positive ties, or three mutually hostile individuals) incur emotional stress or are not strategically optimal. See Figure [\[fig:structural-balance\]](#fig:structural-balance){reference-type="ref" reference="fig:structural-balance"}. Thus they are not socially stable--over time, they break and change to the stable ones. In fact, the structural balance theory not only describes the local property in a signed complete network, but also predicts the global network behavior--the only type of network in which all triangles are stable must have the nodes partitioned into two camps, inside each of which the edges are all positive and between them the edges are negative (one of the camps can be empty). The structural balance theory only describes the equilibrium state and does not provide any model on evolution or dynamics--what happens when a network has unstable triangles? Follow up work on structural balance dynamic models update the sign/weight of an edge \((i, j)\) towards a more balanced triangle by considering the sign/weights of neighboring edges \((i, k)\) and \((k, j)\) in a triangle \(\triangle i j k\). These two threads of research, opinion dynamics and structural balance dynamics, are currently orthogonal to each other. In opinion dynamics, opinions on vertices are influenced by each other through edges but researchers struggle to maneuver the model to create community cleavage. In structural balance dynamics, edge weights are dynamically updated to meet structural balance, which is explicitly coded as the projected outcome and optimization objective. In this paper, we propose co-evolution models for both dynamics of opinions (people's view on a variety of topics) and dynamics of social appraisals (the approval or disapproval towards each other). We show that by using two simple rules, node opinions evolve into opposing communities and structural balance naturally emerges.
## Our Contribution
In this paper, we consider the co-evolution of opinions via social influence and tie strength/appraisal/sign updates by discrepancies of node opinions. We assume a set of \(n\) individuals where individual \(i\) has an opinion \(v_i \in \mathbb{R}\), and an appraisal matrix \(W=[w_{ij}]\), where \(w_{ij}\) is interpreted as the influence from individual \(j\) on individual \(i\). Here \(w_{ij}\) does not need to be non-negative and takes values in \(\mathbb{R}\). We consider two update rules:
- *Opinion dynamics:* The opinion of \(i\) is updated as a linear combination of its current opinion and the weighted sum of neighbors' opinions with coefficients in matrix \(W\).
- *Appraisal dynamics:* The appraisal \(w_{ij}\) is updated as a linear combination of its current value and the agreement of the opinion of \(i\) and \(j\). The agreement of opinion \(v_i\) and \(v_j\) is taken as the dot product \(v_i \cdot v_j\).
The opinion dynamics model is similar to classical social influence models such as the DeGroot model and Friedkin-Johnsen model [@Friedkin1990-wl], except that the edge weight matrix \(W\) is dynamic as well. The model for appraisal dynamics is motivated by tie dynamics that can be traced back to Schelling's model of residential segregation . In modern society, tie changes on Facebook and Twitter can be easily triggered by disparities on their opinions , especially among the users who are most politically engaged. Our goal is to analyze the evolution and in particular, the conditions that lead to *sign stability* and *structural balance*--for any triangle with individual \(i, j, k\), \(w_{ij}w_{jk}w_{ki}\geq 0\). We call a network to reach strictly structural balance if \(w_{ij}w_{jk}w_{ki}> 0\), \(\forall i, j, k\). A structurally balanced network has two possible states: *harmony*, when all edges are positive; and *polarization*, when there are two communities with only positive ties within each community and negative ties across the two communities. We show that with our dynamics model, \(W(t)\) evolves by the following matrix Riccati Equation \(W' = WW^T + C\), where \(C= V(0)V(0)^T-W(0)W(0)^T\) is a symmetric constant \(n \times n\) matrix if \(W(0)\) is. Further the opinion vector \(V(0)\) evolves by the differential equation \(V''(t)=2|V|^2 \cdot V-C\cdot V\). Our main result is to analyze the asymptotic behavior of \(V(t)\) and \(W(t)\) and prove structural balance at the limit. Our results can be summarized in the following:
1. By analyzing the evolving equation for \(V(t)\), we show that either the network reaches strict structural balance or \(|V(t)|\rightarrow 0\). To prove this limit behavior, one crucial observation is that the length of the opinion vector \(|V(t)|^2\) is strictly convex, unless \(V(t)\equiv 0\).
2. We show how to solve the general matrix Riccati equation \(W'=W^2+C\), \(W(0)=B\), for any parameter \(B, C\). In particular, the eigenvector corresponding to the largest eigenvalue of \(W(t)\) encodes the two communities formed in the network; those with a positive value in the eigenvector versus those with a negative value in the eigenvector. As a byproduct of the analysis of \(W(t)\), we also show that when \(V(0)\) is an eigenvector of the initial matrix \(W(0)\), \(V(t)\) remains to be an eigenvector of \(W(t)\) for and structural balance must occur in finite time. In this case we can write down exact evaluations on the blowup time and the rate convergence.
The evolving of \(W(t)\) by \(W' = WW^T + C\) is strictly a generalization of the dynamic structural balance model by Marvel et al. . Their model captures the dynamics of edge appraisals by \(W' = W^2\) and does not consider user opinions. The behavior of our model becomes much more complex, as the initial user opinions are factored into the system dynamics through the matrix \(C\). We also performed extensive simulations with different initial conditions and graph topology. We examined the network evolution on the final convergence state (harmony v.s., polarization) and convergence rate. We observed that a higher network density or a higher initial opinion magnitude, empirically, speeds up the convergence rate. We tested our dynamic model on two real world graphs (Karate club graph and a political blog graph ). Both networks are known to have two communities with opposing opinions. A small number of seeds, randomly selected, are assigned with ground truth opinions and all other nodes start neural. The network evolution can successfully detect the final community structure and recover the ground truth with good accuracy. Apart from being a transparent and explainable label propagation algorithm, the model sheds lights on why community structure emerges and becomes a widely observed, sustainable property in complex networks.
# Related Work
## Opinion Dynamics and Social Influence
Opinions in a sociological viewpoint capture the cognitive orientation towards issues, events or other subjects, and mathematically represent signed attitudes or certainties of belief. Opinion dynamics is an extensively studied topic about how opinions change in a network setting with social influence from neighbors. One of the first models of opinion dynamics, French-DeGroot model , considers a discrete time process of opinion \(\{v_1, v_2, \cdots, v_n\}\) for a group of \(n\) individuals. An edge \((i, j)\) carries a non-negative weight \(w_{ij} \geq 0\). The opinion of node \(i\) at time \(t+1\) is updated by \[v_i(t+1)=\sum_{j}w_{ij}v_j(t).\] The weight matrix \(W=[w_{ij}]\) is taken as a stochastic matrix. The dynamics can be written as \(V(t+1)=WV(t)\). The continuous-time counterpart is called the Abelson's model where the dynamics is defined by \[\label{ref:eqn-abelson} \frac{dV(t)}{dt}=-LV(t),\] where \(V(t)=(v_1(t), v_2(t), \cdots v_n(t))^T\) and \(L\) is the Laplacian matrix \(L=I-W\). Opinions following the French-DeGroot model or the Abelson's model typically converge unless the network is disconnected or there are stubborn nodes (with \(w_{ii}=1\)). The most popular opinion dynamics model is probably the Friedkin-Johnsen model . It takes a stochastic matrix \(W\) as the influence model, and a diagonal matrix \(\Lambda=\diag(\lambda_1, \cdots, \lambda_n)\) where \(\lambda_i\in [0, 1]\) is the susceptibility of individual \(i\) to social influence. The opinions of the individuals are updated by the following process \[V(t+1)=\Lambda W V(t)+(I-\Lambda)u,\] where \(u\) is a constant vector of the individuals' prejudices and is often taken as the initial opinion \(V(0)\). When \(\Lambda=I\) the model turns to French-DeGroot model. Most of the literature on these two models assume a fixed weight matrix \(W\) and prove asymptotic convergence under favorable assumptions . There have been extensions when \(W\) is a time-varying matrix, but \(W(t)\) is still independent of \(V(t)\) (e.g., ). A significant deviation from the above family considers a time-varying matrix \(W\), by incorporating the principle of homophily, that similar individuals interact more than dissimilar ones. This is called the bounded confidence model . A few such models (Hegselmann-Krause (HK) model , Deffuant and Weisbuch ) introduced a fixed range of confidence \(d>0\): individual \(i\) is insensitive to opinions that fall outside its confidence set \(I_i=[v_i-d, v_i+d]\), and the opinion \(v_i\) is only updated by the average opinion of those opinions within \(I_i\). In other words, the matrix \(W\) is derived from the set of opinions at time \(t\) and thus co-evolves with the opinions. This model generates situations when the individuals converge to a set of different opinions, and has been extended to the multi-dimensional setting . All models above have only considered the case of positive influence, that the interactions of individuals change their opinions *towards* each other. It has been argued in both social settings and many physical systems that there is negative or repulsive influence (repulsive interactions in biological systems or collision avoidance in robot swarm formation ). Abelson argued that any attempt to persuade a person may sometimes shift his or her opinion away from the persuader's opinion, called the boomerang effect . Bhawalkar et al. presented game-theoretic models of opinion formation in social networks by maximizing agreement with friends weighted by the strength of the relationships. Thus interactions between individuals with similar opinions move their opinions closer; interactions between individuals with opinions that are very different shift their opinions away from each other. Here the edge weights are fixed. Many models have included negative ties but they are still awaiting rigorous analysis . The most notable work in this direction is by Altafini . The model starts to be similar to Abelson's model in Equation ([\[ref:eqn-abelson\]](#ref:eqn-abelson){reference-type="ref" reference="ref:eqn-abelson"}) with a *fixed* weight matrix \(W\) except that the weights in \(W\) do not need to be non-negative. The system is shown to be Lyapunov stable and studies have focused on the initial conditions of \(W\) for the system to converge to harmony or polarization. The matrix \(W\) is assumed to be either static or, in very recent studies , time-varying (but independent of the opinion changes). The negative influence is closely related to signed networks and structural balance theory, which will be discussed next.
## Structural Balance and Signed Networks
Notice that the structural balance theory only describes the equilibrium state and does not provide any model on evolution or dynamics--what happens when a network has unstable triangles? Follow up work proposed a few models, that can be categorized by discrete models or continuous models--depending on whether the appraisal on a social tie takes binary values \(\{+1,-1\}\) or a real number. Antal et al. considered the discrete model where the sign of an edge is flipped if this produces more balanced triangles than unbalanced ones. The balanced graph is clearly a stable state but the dynamics also has many local optima called jammed states . A similar model is used in Andreida et al. . Samin et al. considered removing the minimum number of edges to make the graph balanced, which is an NP-hard problem. In the continuous setting, the influence-based model describes an influence process on a complete graph, in which an individual \(i\) updates her appraisal of individual \(j\) based on what others positively or negatively think of \(j\). In other words, let us use \(w_{ij}\) to describe the type of the social tie between two individuals \(i, j\). \(w_{ij}>0\) if \(i, j\) are friends and \(<0\) if they feel negative about each other. The absolute value of \(w_{ij}\) describes the magnitude of the appraisal. The update rule says that the update to \(w_{ij}\) will take value \[\label{eq:dynamic-w} \frac{d w_{ij}}{dt}=\sum_{k}w_{ik}\cdot w_{kj}.\] Specifically, when \(w_{ik}\) and \(w_{kj}\) have the same sign, the value of \(w_{ij}\) is guided to the positive direction; when \(w_{ik}\) and \(w_{kj}\) have opposite signs, the value of \(w_{ij}\) is guided to the negative direction. Both cases try to enforce a balanced triangle on \(\{i, j, k\}\). Empirically, it has been observed that for essentially any initial value of \(W\), as the matrix where the \((i, j)\) element is \(w_{ij}\), the system reached a balanced pattern in finite time. In , Marvel et al. proved that for a random initial matrix \(W\) the system reaches a balanced matrix in finite time with probability converging to \(1\) as \(n \rightarrow \infty\). They also characterized the converged value and its relationship to the initial value. In a recent paper , Cisneros-Velarde et al. considered a *pure-influence* model, where the self-appraisal (such as \(w_{ii}\)) is taken out of Equation ([\[eq:dynamic-w\]](#eq:dynamic-w){reference-type="ref" reference="eq:dynamic-w"}) to be a more faithful interpretation of Heider's structural balance. They proved that when \(W\) is symmetric their continuous-time dynamic model is exactly the gradient flow of an energy function called *dissonance* , defined as \[-\sum_{[ij],[jk],[ki] \in E} w_{ij} \cdot w_{jk} \cdot w_{ki}.\] Dissonance characterizes the degree of violation to Heider's structural balance axioms in the current network. The global minimum of this energy function corresponds to signed networks that satisfy structural balance in the case of real-values appraisals. When the initial matrix \(W\) is symmetric the authors also provided characterizations of the critical points of the dissonance function (aka the equilibrium states of the dynamic model). The discussions of opinion dynamics and dynamics with structural balance, so far, have focused on node opinion changes or link appraisal changes, separately. Holme and Newman presented a simple model combing the two but did not present theoretical analysis. There is little work on combining both dynamics into a co-evolving model, which is the focus of this paper.
# Co-Evolution Model
Suppose there are \(n\) individuals, each one with its own opinion \(v_i \in \mathbb{R}\). Define the opinion vector \(V = (v_1, v_2, \dots, v_n )^{T} \in \mathbb{R}^{n}\). The influence model among the \(n\) individuals is characterized as an \(n\times n\) matrix \(W = [w_{ij}]\) with entries taking real values. A positive value of \(w_{ij}\) indicates a positive social influence between \(i, j\), where the opinions under the influence become similar. A negative value of \(w_{ij}\) means a negative influence and their opinions under influence become dissimilar. In our theoretical study, we consider the case of a complete graph. The evolution model that we introduce works for any network. In our simulations we also evaluate networks and opinion co-evolution on a general graph. Both the opinions of individuals and the influence matrix are dynamically evolving. Assume that the initial opinion vector is \(V(0)\) and the initial influence matrix is \(W(0)\). In this paper we assume the initial weight matrix \(W(0)\) is symmetric, i.e., \(w_{ij}(0)=w_{ji}(0)\), \(\forall i, j\). Define the opinion vector and influence matrix at time \(t\), in a discrete-time model, as \(V(t)\) and \(W(t)\) respectively. We propose the dynamic system governing the evolution of the relationship over integer time: \[\label{eq:discrete-model} \left\{ \begin{array}{l} V(t+1) = V(t) + W(t)V(t) \\ W(t+1) = W(t) + V(t) V(t)^T. \end{array} \right.\] In the first equation, the opinion of an individual \(i\) is shifted by the weighted sum of its neighbors' opinions, with coefficients in the influence matrix \(W\). In the second equation, the appraisal value \(w_{ij}\) between two individuals \(i, j\) is updated by the differences of opinions \(v_i, v_j\). If \(v_i, v_j\) generally agree (with a positive dot product), \(w_{ij}\) moves in the positive direction; otherwise moves in the negative direction. In a continuous-time model, the dynamics are driven by the following ODE: \[\label{eqn:continuous-model} \left\{ \begin{array}{l} V' = WV \\ W' = VV^T. \end{array} \right.\] where \(V'\) and \(W'\) is the coordinate-wise time derivative of \(V\) and \(W\). From this point on, we focus on solving the continuous time model. First we present a couple of basic properties of Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"}). This means we can focus on solving the system defined by Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"}) without losing generality. The detailed proof is provided in Appendix [\[subsec:model\]](#subsec:model){reference-type="ref" reference="subsec:model"}.
The main objective of this paper is to analyze how this system evolves. In particular, we care about system evolution to reach *sign stability* for \(w_{ij}\), \(\forall i, j\), and \(v_i\), \(\forall i\), as well as *structural balance*--\]\lim_{ t \to T} w_{ij}w_{jk}w_{ki} \geq 0,\[ for all indices \(i,j,k\) where \([0, T)\) is the maximum interval on which the solution \(W(t)\) exists. Notice that in the classical structural balance theory, the two types of stable triangles--with edge signs as either all positive (\(+1\)) or have two negative (\(-1\)) and one positive--satisfy this property. The evolution of \(W(t)\) and \(V(t)\) is described in the following two lemmas.
Remark that matrix \(C\) in our setting is a special symmetric matrix. Specifically, \(C+W(0)W(0)^T=V(0)V(0)^T\) has rank one. This property turns out to be useful for characterization of the system behavior.
# Analysis of the Opinion and Social Tie Evolution
Our analysis has two parts. First we focus on the opinion evolution model (Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"})). Here we provide analysis of the asymptotic behavior for \(V(t)\). Then we study the social tie evolution (Equation ([\[eqn:Riccati\]](#eqn:Riccati){reference-type="ref" reference="eqn:Riccati"})) for \(W(t)\). By solving Riccati equation explicitly for \(W(t)\) we are able to provide more detailed characterization of the evolving behavior.
## Analysis of Opinion Evolution
By analyzing the opinion evolution (Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"})), our main result is the following.
The theorem says that structural balance is always achieved, unless the opinions converge to a zero vector, in which case the entire network becomes neutral. We can consider the second case as a boundary case of structural balance. The rest of the subsection will focus on proving this theorem. An important observation is that the norm of \(V\), \(|V(t)|^2\), is a convex function. The detailed proof is in Appendix [15](#app:opinion){reference-type="ref" reference="app:opinion"}.
Now, let us understand Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"}) using coordinates. Since the matrix \(C\) is symmetric, by the orthogonal diagonalization theorem, there exists an orthogonal matrix \]U = [\beta_1, \cdots, \beta_n]=[u_{ij}]_{ n \times n}\[ such that \(C\beta_i = a_i \beta_i, i=1, 2, \cdots, n,\) where \(a_1, \cdots, a_n\) are eigenvalues of \(C\). By our assumption that \(C = V(0)V(0)^T-W(0)W(0)^T\), the eigenvalues \(a_1, \cdots, a_n\) are non-positive except for one. So we may assume \(a_1 \geq 0\) and \(a_2, \cdots, a_n \leq 0\). Because \(\beta_1, \cdots, \beta_n\) form an orthonormal basis of \(\mathbb{R}^n\), we can write \]V(t) = \sum_{i=1}^n \lambda_i(t)\beta_i = U\lambda, \mbox{ where } \lambda = [\lambda_1, \cdots, \lambda_n]^T.\[ This implies that \(V'(t) = \sum_{i=1}^n \lambda_i'(t)\beta_i\), \(V''(t) = \sum_{i=1}^n\lambda_i''(t) \beta_i\), \(|V(t)|^2 = \sum_{j=1}^n \lambda_j^2(t)\) and \(C\cdot V = \sum_{i=1}^n a_i \lambda_i \beta_i\). Therefore, Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"}) becomes \]\sum_{i=1}^{n} \lambda_i''(t) \cdot \beta_i = \sum_{i=1}^n (2(\sum_{k=1}^n \lambda_k^2)-a_i) \lambda_i \cdot \beta_i\[ Since \(\beta_1, \cdots, \beta_n\) are independent, we obtain the system of ODE with \(\lambda = [\lambda_1, \cdots, \lambda_n]^T\) in the form \]\label{eqn:lambda} \lambda_i''(t) = (2\sum_{k=1}^n \lambda_k^2-a_i) \lambda_i(t), i=1, 2, \cdots, n.\[ Denote \(W = [w_{ij}(t)]\). Then \(W' = V\cdot V^T\) implies \]W' = U\cdot \lambda \cdot \lambda^T \cdot U^T\[ Therefore, \(w'_{ij}(t)=\sum_{k,l=1}^n u_{ik}u_{jl}\lambda_{k}\lambda_l\) and \(w_{ij}(t) = \sum_{k,l=1}^n (\int_0^t \lambda_k(s)\cdot \lambda_l(s) ds) u_{ik}u_{jl} + w_{ij}(0)\). Our next goal is to show the following proposition,
Assuming the proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"}, Theorem [\[thm:main_theorem\]](#thm:main_theorem){reference-type="ref" reference="thm:main_theorem"} follows. Indeed, the leading term in \(w_{ij}(t)\) is \(u_{ih}u_{jh}\psi_{hh}\) as \(t \rightarrow T\). Therefore the sign of \(w_{ij}\) is the same as the sign of \(u_{ih}u_{jh}\psi_{hh}\). The leading term of \(w_{ij}w_{jk}w_{ki}\) as \(t \to T\) is \]u_{ih}^2 u_{jh}^2 u_{kh}^2 \psi_{hh}^2 \geq 0.\[ This shows that structural balance occurs eventually for generic initial values. Here the generic condition is used to ensure that all entries \(u_{ij}\) of the orthogonal matrix \([u_{ij}]\) are not zero and \(\psi_{hh}\) is the unique term with the maximum growth rate. Finally, if \(T=\infty\) and \(\lim_{t \to \infty} |V(t)|=0\), then by Corollary [\[cor:v\'\]](#cor:v'){reference-type="ref" reference="cor:v'"}, \(\lim_{t \to \infty} V'(t)=0\). The proof for proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"} is fairly technical and can be found in Appendix [15](#app:opinion){reference-type="ref" reference="app:opinion"}.
## Analysis of Social Tie Evolution
The analysis on the evolution of \(V\) in the previous section shows convergence. To further understand the community formed at the limit of convergence, we need to study the evolution of \(W\). The following theorem explains the reason behind the appearance of structure balance when at least one eigenvalue of \(W(t)\) tends to infinity. This was proved in . We include the statement here and the proof in Appendix [1.1](#subsec:structural){reference-type="ref" reference="subsec:structural"} for completeness.
Furthermore, the two antagonistic communities are given by \(U^+=\{i \in V|u_i>0\}\) and \(U^-=\{i \in V| u_i<0\}\) where \(u=[u_1,..., u_n]^T\) is a \(\beta_1(t)\) eigenvector. All edges connecting vertices within the same community are positive while edges connecting two vertices in different communities are negative. When one of \(U^+, U^-\) is empty, the network has only one community. The model in is the Riccati equation \(W'=W^2\), whose solution is \(W(t)=W(0)(1-W(0)t)^{-1}\). As a consequence, Marvel et. al. showed that structure balance occurs in the Riccati equation \(W'=W^2\) for generic initial parameter \(W(0)\) with a positive eigenvalue at finite time. Our model strictly generalizes the previous model and consider how the initial opinions may influence the system evolution. In the following we show how to solve the general Riccati equation and also when an eigenvalue goes to infinity. With most of the details in the Appendix, we carry out rigorous analysis of the general form of the matrix Riccati equation as stated below. \]\label{eqn:Evolution} \left\{ \begin{array}{rl} W' & = W^2 + C \\ W(0) & = B. \end{array} \right.\[ For general matrices \(B, C\) we can solve for \(W(t)\) as shown in the following theorem. The proof details can be found in Appendix [1.2](#subsec:general-Riccati){reference-type="ref" reference="subsec:general-Riccati"}.
In our co-evolution model, \(B=W(0)\) is assumed to be symmetric. Thus \(C\) is also symmetric. This allows us to simplify the solution further, as shown in Appendix [1.3](#subsec:socialtie){reference-type="ref" reference="subsec:socialtie"}. Then we analyze a special case when \(BC=CB\). In this case we can characterize the conditions when structural balance is guaranteed to occur, with details in Appendix [1.1](#subsec:structural){reference-type="ref" reference="subsec:structural"}, [1.4](#subsec:BCequalCB){reference-type="ref" reference="subsec:BCequalCB"}, and [1.5](#subsec:structural-commute){reference-type="ref" reference="subsec:structural-commute"}. Specifically, using a basic fact that two commuting symmetric matrices can be simultaneously orthogonally diagonalized , we can get the following theorem where the conditions of the eigenvalues of \(B, C\) for structural balance are characterized:
In our co-evolution model, \(BC=CB\) happens when \(V(0)\) is a eigenvector of \(W(0)\), an interesting initial condition. To see that, recall \(B=W(0)\), \(C= V(0)V(0)^T-W(0)W(0)^T\), and \(W(0)=W(0)^T\). To check if \(BC=CB\), we just need to check if \(W(0)V(0)V(0)^T=V(0)V(0)^TW(0)\) and apply the following Lemma (Proof in Section [1.5](#subsec:structural-commute){reference-type="ref" reference="subsec:structural-commute"}).
Further, the equation in our co-evolution model, i.e., Equation ([\[eqn:Riccati\]](#eqn:Riccati){reference-type="ref" reference="eqn:Riccati"}), satisfies \(C+B^2=V(0)V(0)^T\). Notice that the right-hand side \(V(0)V(0)^T\) is an \(n \times n\) matrix with rank one. This property actually ensures that the conditions characterized in Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} are met and thus structural balance is guaranteed. At the same time, the convergence rate is \(O(\frac{1}{|T-t|})\), which is proved by Lemma [\[lemma:sin\]](#lemma:sin){reference-type="ref" reference="lemma:sin"} in Appendix [1.4](#subsec:BCequalCB){reference-type="ref" reference="subsec:BCequalCB"}.
If \(V(0)=0\), the system stays at the fixed point with \(V\) remaining zero and the weight matrix unchanged. The case when \(V(0)\) is an eigenvector of \(W(0)\) includes a few interesting cases in practice. When \(W(0)=0\) or \(W(0)=I\), this models a group of individuals that start as complete strangers with uniform self-appraisals. Their non-homogeneous initial opinions \(V(0)\) may drive the network to be segmented over time. Finally the fact that \(V(t)\) remains an eigenvector of \(W(t)\) for all \(t\) follows from Equation ([\[eqn:WWW\]](#eqn:WWW){reference-type="ref" reference="eqn:WWW"}) in Appendix C. We conjecture that even in the general case (when \(V(0)\) is not necessarily an eigenvector of \(W(0)\)) the limit vector \(\lim_{t \to T} V(t)/|V(t)|\) is an eigenvector of the limit tie relation matrix \(\lim_{t \to T} W(t)/|V(t)|\). This is supported by our numerical evidences and Corollary [\[cor:wv\]](#cor:wv){reference-type="ref" reference="cor:wv"} which says that \(\lim_{t \to T} V(t)/|V(t)|\) is an eigenvector of \(\lim_{t \to T} (W(t)/|V(t)|)^2\).
# Simulations
In this section, we provide simulation results to accompany our theoretical analysis of the co-evolution model. We also present simulation results on general graphs and real world data sets to understand the behavior of network evolution. Here is a brief summary of observations from simulations. The details can be found in the Appendix.
- Except a few carefully crafted cases[^1] that make the opinion vector to be zero, strict structural balance is always reached, regardless of whether the graph is complete or general, whether \(B, C\) commute or not, or whether \(W(0)\) is symmetric or as a general matrix. An example is shown in Figure [\[fig:heatmap\]](#fig:heatmap){reference-type="ref" reference="fig:heatmap"}.
- In general, we observed through simulations that the number of iterations to convergence is inversely proportional to the magnitude of initial opinions and the edge density in the graph.
- In two real world data sets (the Karate club and a political blog data set), a few initially planted polarized opinions can successfully predict the ground truth community structure with high accuracy. See Figure [\[fig:application\]](#fig:application){reference-type="ref" reference="fig:application"}.
# Conclusion and Future Work
In this paper, we have provided a co-evolution model for both opinion dynamics and appraisal dynamics. We provided solutions to the system and rigorously characterized how the stable states depend on the initial parameters. There are a few follow-up problems that remain open, for example, when the social ties are directional/asymmetric, when the network is not a complete graph, and when each agent has an \(m\)-dimensional opinion vector. We include some discussion and conjectures on these cases in Appendix [1.6](#subsec:general){reference-type="ref" reference="subsec:general"} and consider this as interesting future work.
# Introduction
We live in a continuously changing world in which social interactions dynamically shape who we are, how we view the world and what decisions to make. Various social processes, naturally intertwined, operate on both the properties of individuals and social ties among them. Social influence, for example, describes how people's behaviors, habits or opinions are shifted by those of their neighbors. Social influence leads to homophily ( similarity of node attributes between friends) and leaves traces in the network structure such as high clustering coefficient and triadic closure (there are likely social ties among one's friends). Social influence and homophily, however, do not fully interpret the global network structure. One of the widely observed structural properties in social networks is the community structure. Nodes within the same community are densely connected and nodes from different communities are sparsely connected. Community detection is an important topic in social network analysis and has been investigated extensively . But why does community structure emerge and become a persistent feature? Are there social processes that encourage or maintain the community structure? In the literature, there have been a lot of studies of opinion dynamics, where people's behaviors, habits or opinions are influenced by those of their neighbors. Network models that capture social influence, such as French-DeGroot model , Friedkin-Johnsen model [@Friedkin1990-wl], Ku\(\l\)akowski et al. model naturally converge to global consensus. Actually it seems to be non-trivial to create non-homogeneous outcomes in these models, while in reality social groups often fail to reach consensus and exhibit clustering of opinions and other irregular behaviors. Getting a model that may produce community cleavage or diversity often requires specifically engineered or planted elements in a rather explicit manner. For example, bounded confidence models limit social influence only within pairs with opinions sufficiently close. Other models introduce stubborn nodes whose opinions remain unchanged throughout the process. The extreme case of community structure is dictated by structural balance , a common phenomenon observed in many social relationships. The notion was first introduced in a seminal paper by Heider in the 1940s in social psychology. It describes the stability of human relations among three individuals when there are only two types of social ties: positive ties describe friendship (or sharing common opinions) and negative ties describe hostility (or having opposite opinions). Heider's axioms state that among three individuals, only two kinds of triangles are stable: the triangles where all three ties are positive, indicating all three individuals are mutual friends; and the triangle with two negative edges and one positive edge, describing the folklore that "the enemy of your enemy is your friend." The other two types of triangles (e.g., a triangle of one negative tie and two positive ties, or three mutually hostile individuals) incur emotional stress or are not strategically optimal. See Figure [\[fig:structural-balance\]](#fig:structural-balance){reference-type="ref" reference="fig:structural-balance"}. Thus they are not socially stable--over time, they break and change to the stable ones. In fact, the structural balance theory not only describes the local property in a signed complete network, but also predicts the global network behavior--the only type of network in which all triangles are stable must have the nodes partitioned into two camps, inside each of which the edges are all positive and between them the edges are negative (one of the camps can be empty). The structural balance theory only describes the equilibrium state and does not provide any model on evolution or dynamics--what happens when a network has unstable triangles? Follow up work on structural balance dynamic models update the sign/weight of an edge \((i, j)\) towards a more balanced triangle by considering the sign/weights of neighboring edges \((i, k)\) and \((k, j)\) in a triangle \(\triangle i j k\). These two threads of research, opinion dynamics and structural balance dynamics, are currently orthogonal to each other. In opinion dynamics, opinions on vertices are influenced by each other through edges but researchers struggle to maneuver the model to create community cleavage. In structural balance dynamics, edge weights are dynamically updated to meet structural balance, which is explicitly coded as the projected outcome and optimization objective. Holme and Newman presented a simple model of this combination without any theoretical analysis. In this paper, we propose co-evolution models for both dynamics of opinions (people's view on a variety of topics) and dynamics of social appraisals (the approval or disapproval towards each other). We show that by using two simple rules, node opinions evolve into opposing communities and structural balance naturally emerges.
## Our Contribution
In this paper, we consider the co-evolution of opinions via social influence and tie strength/appraisal/sign updates by discrepancies of node opinions. We assume a set of \(n\) individuals where individual \(i\) has an opinion \(v_i \in \mathbb{R}\), and an appraisal matrix \(W=[w_{ij}]\), where \(w_{ij}\) is interpreted as the influence from individual \(j\) on individual \(i\). Here \(w_{ij}\) does not need to be non-negative and takes values in \(\mathbb{R}\). We consider two update rules:
- *Opinion dynamics:* The opinion of \(i\) is updated as a linear combination of its current opinion and the weighted sum of neighbors' opinions with coefficients in matrix \(W\).
- *Appraisal dynamics:* The appraisal \(w_{ij}\) is updated as a linear combination of its current value and the agreement of the opinion of \(i\) and \(j\). The agreement of opinion \(v_i\) and \(v_j\) is taken as the dot product \(v_i \cdot v_j\).
The opinion dynamics model is similar to classical social influence models such as the DeGroot model and Friedkin-Johnsen model , except that the edge weight matrix \(W\) is dynamic as well. The model for appraisal dynamics is motivated by tie dynamics that can be traced back to Schelling's model of residential segregation . In modern society, tie changes on Facebook and Twitter can be easily triggered by disparities on their opinions , especially among the users who are most politically engaged. Our goal is to analyze the evolution and in particular, the conditions that lead to *sign stability* and *structural balance*--for any triangle with individual \(i, j, k\), \(w_{ij}w_{jk}w_{ki}\geq 0\). We call a network to reach strictly structural balance if \(w_{ij}w_{jk}w_{ki}> 0\), \(\forall i, j, k\). A structurally balanced network has two possible states: *harmony*, when all edges are positive; and *polarization*, when there are two communities with only positive ties within each community and negative ties across the two communities. We show that with our dynamics model, \(W(t)\) evolves by the following matrix Riccati Equation \(W' = WW^T + C\), where \(C= V(0)V(0)^T-W(0)W(0)^T\) is a symmetric constant \(n \times n\) matrix if \(W(0)\) is. Further the opinion vector \(V(0)\) evolves by the differential equation \(V''(t)=2|V|^2 \cdot V-C\cdot V\). Our main result is to analyze the asymptotic behavior of \(V(t)\) and \(W(t)\) and prove structural balance at the limit. Our results can be summarized in the following:
1. By analyzing the evolving equation for \(V(t)\), we show that either the network reaches strict structural balance or \(|V(t)|\rightarrow 0\). To prove this limit behavior, one crucial observation is that the length of the opinion vector \(|V(t)|^2\) is strictly convex, unless \(V(t)\equiv 0\).
2. We show how to solve the general matrix Riccati equation \(W'=W^2+C\), \(W(0)=B\), for any parameter \(B, C\). In particular, the eigenvector corresponding to the largest eigenvalue of \(W(t)\) encodes the two communities formed in the network; those with a positive value in the eigenvector versus those with a negative value in the eigenvector. As a byproduct of the analysis of \(W(t)\), we also show that when \(V(0)\) is an eigenvector of the initial matrix \(W(0)\), \(V(t)\) remains to be an eigenvector of \(W(t)\) for and structural balance must occur in finite time. In this case we can write down exact evaluations on the blowup time and the rate convergence.
The evolving of \(W(t)\) by \(W' = WW^T + C\) is strictly a generalization of the dynamic structural balance model by Marvel et al. . Their model captures the dynamics of edge appraisals by \(W' = W^2\) and does not consider user opinions. The behavior of our model becomes much more complex, as the initial user opinions are factored into the system dynamics through the matrix \(C\). We also performed extensive simulations with different initial conditions and graph topology. We examined the network evolution on the final convergence state (harmony v.s., polarization) and convergence rate. We observed that a higher network density or a higher initial opinion magnitude, empirically, speeds up the convergence rate. We tested our dynamic model on two real world graphs (Karate club graph and a political blog graph ). Both networks are known to have two communities with opposing opinions. A small number of seeds, randomly selected, are assigned with ground truth opinions and all other nodes start neural. The network evolution can successfully detect the final community structure and recover the ground truth with good accuracy. Apart from being a transparent and explainable label propagation algorithm, the model sheds lights on why community structure emerges and becomes a widely observed, sustainable property in complex networks.
# Related Work
## Opinion Dynamics and Social Influence
Opinions in a sociological viewpoint capture the cognitive orientation towards issues, events or other subjects, and mathematically represent signed attitudes or certainties of belief. Opinion dynamics is an extensively studied topic about how opinions change in a network setting with social influence from neighbors. One of the first models of opinion dynamics, French-DeGroot model , considers a discrete time process of opinion \(\{v_1, v_2, \cdots, v_n\}\) for a group of \(n\) individuals. An edge \((i, j)\) carries a non-negative weight \(w_{ij} \geq 0\). The opinion of node \(i\) at time \(t+1\) is updated by \]v_i(t+1)=\sum_{j}w_{ij}v_j(t).\[ The weight matrix \(W=[w_{ij}]\) is taken as a stochastic matrix. The dynamics can be written as \(V(t+1)=WV(t)\). The continuous-time counterpart is called the Abelson's model where the dynamics is defined by \]\label{ref:eqn-abelson} \frac{dV(t)}{dt}=-LV(t),\[ where \(V(t)=(v_1(t), v_2(t), \cdots v_n(t))^T\) and \(L\) is the Laplacian matrix \(L=I-W\). Opinions following the French-DeGroot model or the Abelson's model typically converge unless the network is disconnected or there are stubborn nodes (with \(w_{ii}=1\)). The most popular opinion dynamics model is probably the Friedkin-Johnsen model . It takes a stochastic matrix \(W\) as the influence model, and a diagonal matrix \(\Lambda=\diag(\lambda_1, \cdots, \lambda_n)\) where \(\lambda_i\in [0, 1]\) is the susceptibility of individual \(i\) to social influence. The opinions of the individuals are updated by the following process \]V(t+1)=\Lambda W V(t)+(I-\Lambda)u,\[ where \(u\) is a constant vector of the individuals' prejudices and is often taken as the initial opinion \(V(0)\). When \(\Lambda=I\) the model turns to French-DeGroot model. Most of the literature on these two models assume a fixed weight matrix \(W\) and prove asymptotic convergence under favorable assumptions . There have been extensions when \(W\) is a time-varying matrix, but \(W(t)\) is still independent of \(V(t)\) (e.g., ). A significant deviation from the above family considers a time-varying matrix \(W\), by incorporating the principle of homophily, that similar individuals interact more than dissimilar ones. This is called the bounded confidence model . A few such models (Hegselmann-Krause (HK) model , Deffuant and Weisbuch ) introduced a fixed range of confidence \(d>0\): individual \(i\) is insensitive to opinions that fall outside its confidence set \(I_i=[v_i-d, v_i+d]\), and the opinion \(v_i\) is only updated by the average opinion of those opinions within \(I_i\). In other words, the matrix \(W\) is derived from the set of opinions at time \(t\) and thus co-evolves with the opinions. This model generates situations when the individuals converge to a set of different opinions, and has been extended to the multi-dimensional setting . All models above have only considered the case of positive influence, that the interactions of individuals change their opinions *towards* each other. It has been argued in both social settings and many physical systems that there is negative or repulsive influence (repulsive interactions in biological systems or collision avoidance in robot swarm formation ). Abelson argued that any attempt to persuade a person may sometimes shift his or her opinion away from the persuader's opinion, called the boomerang effect . Bhawalkar et al. presented game-theoretic models of opinion formation in social networks by maximizing agreement with friends weighted by the strength of the relationships. Thus interactions between individuals with similar opinions move their opinions closer; interactions between individuals with opinions that are very different shift their opinions away from each other. Here the edge weights are fixed. Many models have included negative ties but they are still awaiting rigorous analysis . The most notable work in this direction is by Altafini . The model starts to be similar to Abelson's model in Equation ([\[ref:eqn-abelson\]](#ref:eqn-abelson){reference-type="ref" reference="ref:eqn-abelson"}) with a *fixed* weight matrix \(W\) except that the weights in \(W\) do not need to be non-negative. The system is shown to be Lyapunov stable and studies have focused on the initial conditions of \(W\) for the system to converge to harmony or polarization. The matrix \(W\) is assumed to be either static or, in very recent studies , time-varying (but independent of the opinion changes). The negative influence is closely related to signed networks and structural balance theory, which will be discussed next.
## Structural Balance and Signed Networks
Notice that the structural balance theory only describes the equilibrium state and does not provide any model on evolution or dynamics--what happens when a network has unstable triangles? Follow up work proposed a few models, that can be categorized by discrete models or continuous models--depending on whether the appraisal on a social tie takes binary values \(\{+1,-1\}\) or a real number. Antal et al. considered the discrete model where the sign of an edge is flipped if this produces more balanced triangles than unbalanced ones. The balanced graph is clearly a stable state but the dynamics also has many local optimals called jammed states . Andreida et al. determine the sign of an edge according to the sign of the other two edges in the triangles to make more triangles balanced. Samin et al. try to remove the minimum edges to make the graph balanced which is NP-hard problem. In the continuous setting, the influence-based model describes an influence process on a complete graph, in which an individual \(i\) updates her appraisal of individual \(j\) based on what others positively or negatively think of \(j\). In other words, let us use \(w_{ij}\) to describe the type of the social tie between two individuals \(i, j\). \(w_{ij}>0\) if \(i, j\) are friends and \(<0\) if they feel negative about each other. The absolute value of \(w_{ij}\) describes the magnitude of the appraisal. The update rule says that the update to \(w_{ij}\) will take value \]\label{eq:dynamic-w} \frac{d w_{ij}}{dt}=\sum_{k}w_{ik}\cdot w_{kj}.\[ Specifically, when \(w_{ik}\) and \(w_{kj}\) have the same sign, the value of \(w_{ij}\) is guided to the positive direction; when \(w_{ik}\) and \(w_{kj}\) have opposite signs, the value of \(w_{ij}\) is guided to the negative direction. Both cases try to enforce a balanced triangle on \(\{i, j, k\}\). Empirically, it has been observed that for essentially any initial value of \(W\), as the matrix where the \((i, j)\) element is \(w_{ij}\), the system reached a balanced pattern in finite time. In , Marvel et al. proved that for a random initial matrix \(W\) the system reaches a balanced matrix in finite time with probability converging to \(1\) as \(n \rightarrow \infty\). They also characterized the converged value and its relationship to the initial value. In a recent paper , Cisneros-Velarde et al. considered a *pure-influence* model, where the self-appraisal (such as \(w_{ii}\)) is taken out of Equation ([\[eq:dynamic-w\]](#eq:dynamic-w){reference-type="ref" reference="eq:dynamic-w"}) to be a more faithful interpretation of Heider's structural balance. They proved that when \(W\) is symmetric their continuous-time dynamic model is exactly the gradient flow of an energy function called *dissonance* , defined as \]-\sum_{[ij],[jk],[ki] \in E} w_{ij} \cdot w_{jk} \cdot w_{ki}.\[ Dissonance characterizes the degree of violation to Heider's structural balance axioms in the current network. The global minimum of this energy function corresponds to signed networks that satisfy structural balance in the case of real-values appraisals. When the initial matrix \(W\) is symmetric the authors also provided characterizations of the critical points of the dissonance function (aka the equilibrium states of the dynamic model). The discussions of opinion dynamics and dynamics with structural balance, so far, have focused on node opinion changes or link appraisal changes, separately. There is little work on combining both dynamics into a co-evolving model, which is the focus of this paper.
# Co-Evolution Model
Suppose there are \(n\) individuals, each one with its own opinion \(v_i \in \mathbb{R}\). Define the opinion vector \(V = (v_1, v_2, \dots, v_n )^{T} \in \mathbb{R}^{n}\). The influence model among the \(n\) individuals is characterized as an \(n\times n\) matrix \(W = [w_{ij}]\) with entries taking real values. A positive value of \(w_{ij}\) indicates a positive social influence between \(i, j\), where the opinions under the influence become similar. A negative value of \(w_{ij}\) means a negative influence and their opinions under influence become dissimilar. In our theoretical study, we consider the case of a complete graph. The evolution model that we introduce works for any network. In our simulations we also evaluate networks and opinion co-evolution on a general graph. Both the opinions of individuals and the influence matrix are dynamically evolving. Assume that the initial opinion vector is \(V(0)\) and the initial influence matrix is \(W(0)\). In this paper we assume the initial weight matrix \(W(0)\) is symmetric, i.e., \(w_{ij}(0)=w_{ji}(0)\), \(\forall i, j\). Define the opinion vector and influence matrix at time \(t\), in a discrete-time model, as \(V(t)\) and \(W(t)\) respectively. We propose the dynamic system governing the evolution of the relationship over integer time: \]\label{eq:discrete-model} \left\{ \begin{array}{l} V(t+1) = V(t) + W(t)V(t) \\ W(t+1) = W(t) + V(t) V(t)^T. \end{array} \right.\[ In the first equation, the opinion of an individual \(i\) is shifted by the weighted sum of its neighbors' opinions, with coefficients in the influence matrix \(W\). In the second equation, the appraisal value \(w_{ij}\) between two individuals \(i, j\) is updated by the differences of opinions \(v_i, v_j\). If \(v_i, v_j\) generally agree (with a positive dot product), \(w_{ij}\) moves in the positive direction; otherwise moves in the negative direction. In a continuous-time model, the dynamics are driven by the following ODE: \]\label{eqn:continuous-model} \left\{ \begin{array}{l} V' = WV \\ W' = VV^T. \end{array} \right.\[ where \(V'\) and \(W'\) is the coordinate-wise time derivative of \(V\) and \(W\). From this point on, we focus on solving the continuous time model. First we present a couple of basic properties of Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"}). This means we can focus on solving the system defined by Equation ([\[eqn:continuous-model\]](#eqn:continuous-model){reference-type="ref" reference="eqn:continuous-model"}) without losing generality. The detailed proof is provided in Appendix [\[subsec:model\]](#subsec:model){reference-type="ref" reference="subsec:model"}.
The main objective of this paper is to analyze how this system evolves. In particular, we care about system evolution to reach *sign stability* for \(w_{ij}\), \(\forall i, j\), and \(v_i\), \(\forall i\), as well as *structural balance*--\[\lim_{ t \to T} w_{ij}w_{jk}w_{ki} \geq 0,\] for all indices \(i,j,k\) where \([0, T)\) is the maximum interval on which the solution \(W(t)\) exists. Notice that in the classical structural balance theory, the two types of stable triangles--with edge signs as either all positive (\(+1\)) or have two negative (\(-1\)) and one positive--satisfy this property. The evolution of \(W(t)\) and \(V(t)\) is described in the following two lemmas.
Remark that matrix \(C\) in our setting is a special symmetric matrix. Specifically, \(C+W(0)W(0)^T=V(0)V(0)^T\) has rank one. This property turns out to be useful for characterization of the system behavior.
# Analysis of the Opinion and Social Tie Evolution
Our analysis has two parts. First we focus on the opinion evolution model (Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"})). Here we provide analysis of the asymptotic behavior for \(V(t)\). Then we study the social tie evolution (Equation ([\[eqn:Riccati\]](#eqn:Riccati){reference-type="ref" reference="eqn:Riccati"})) for \(W(t)\). By solving Riccati equation explicitly for \(W(t)\) we are able to provide more detailed characterization of the evolving behavior.
## Analysis of Opinion Evolution
By analyzing the opinion evolution (Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"})), our main result is the following.
The theorem says that structural balance is always achieved, unless the opinions converge to a zero vector, in which case the entire network becomes neutral. We can consider the second case as a boundary case of structural balance. The rest of the subsection will focus on proving this theorem. An important observation is that the norm of \(V\), \(|V(t)|^2\), is a convex function. The detailed proof is in Appendix [15](#app:opinion){reference-type="ref" reference="app:opinion"}.
Now, let us understand Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"}) using coordinates. Since the matrix \(C\) is symmetric, by the orthogonal diagonalization theorem, there exists an orthogonal matrix \[U = [\beta_1, \cdots, \beta_n]=[u_{ij}]_{ n \times n}\] such that \(C\beta_i = a_i \beta_i, i=1, 2, \cdots, n,\) where \(a_1, \cdots, a_n\) are eigenvalues of \(C\). By our assumption that \(C = V(0)V(0)^T-W(0)W(0)^T\), the eigenvalues \(a_1, \cdots, a_n\) are non-positive except for one. So we may assume \(a_1 \geq 0\) and \(a_2, \cdots, a_n \leq 0\). Because \(\beta_1, \cdots, \beta_n\) form an orthonormal basis of \(\mathbb{R}^n\), we can write \[V(t) = \sum_{i=1}^n \lambda_i(t)\beta_i = U\lambda, \mbox{ where } \lambda = [\lambda_1, \cdots, \lambda_n]^T.\] This implies that \(V'(t) = \sum_{i=1}^n \lambda_i'(t)\beta_i\), \(V''(t) = \sum_{i=1}^n\lambda_i''(t) \beta_i\), \(|V(t)|^2 = \sum_{j=1}^n \lambda_j^2(t)\) and \(C\cdot V = \sum_{i=1}^n a_i \lambda_i \beta_i\). Therefore, Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"}) becomes \[\sum_{i=1}^{n} \lambda_i''(t) \cdot \beta_i = \sum_{i=1}^n (2(\sum_{k=1}^n \lambda_k^2)-a_i) \lambda_i \cdot \beta_i\] Since \(\beta_1, \cdots, \beta_n\) are independent, we obtain the system of ODE with \(\lambda = [\lambda_1, \cdots, \lambda_n]^T\) in the form \[\label{eqn:lambda} \lambda_i''(t) = (2\sum_{k=1}^n \lambda_k^2-a_i) \lambda_i(t), i=1, 2, \cdots, n.\] Denote \(W = [w_{ij}(t)]\). Then \(W' = V\cdot V^T\) implies \[W' = U\cdot \lambda \cdot \lambda^T \cdot U^T\] Therefore, \(w'_{ij}(t)=\sum_{k,l=1}^n u_{ik}u_{jl}\lambda_{k}\lambda_l\) and \(w_{ij}(t) = \sum_{k,l=1}^n (\int_0^t \lambda_k(s)\cdot \lambda_l(s) ds) u_{ik}u_{jl} + w_{ij}(0)\). Our next goal is to show the following proposition,
Assuming the proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"}, Theorem [\[thm:main_theorem\]](#thm:main_theorem){reference-type="ref" reference="thm:main_theorem"} follows. Indeed, the leading term in \(w_{ij}(t)\) is \(u_{ih}u_{jh}\psi_{hh}\) as \(t \rightarrow T\). Therefore the sign of \(w_{ij}\) is the same as the sign of \(u_{ih}u_{jh}\psi_{hh}\). The leading term of \(w_{ij}w_{jk}w_{ki}\) as \(t \to T\) is \[u_{ih}^2 u_{jh}^2 u_{kh}^2 \psi_{hh}^2 \geq 0.\] This shows that structural balance occurs eventually for generic initial values. Here the generic condition is used to ensure that all entries \(u_{ij}\) of the orthogonal matrix \([u_{ij}]\) are not zero and \(\psi_{hh}\) is the unique term with the maximum growth rate. Finally, if \(T=\infty\) and \(\lim_{t \to \infty} |V(t)|=0\), then by Corollary [\[cor:v\'\]](#cor:v'){reference-type="ref" reference="cor:v'"}, \(\lim_{t \to \infty} V'(t)=0\). The proof for proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"} is fairly technical and can be found in Appendix [15](#app:opinion){reference-type="ref" reference="app:opinion"}.
## Analysis of Social Tie Evolution
The analysis on the evolution of \(V\) in the previous section shows convergence. To further understand the community formed at the limit of convergence, we need to study the evolution of \(W\). The following theorem explains the reason behind the appearance of structure balance when at least one eigenvalue of \(W(t)\) tends to infinity. This was proved in . We include the statement here and the proof in Appendix [1.1](#subsec:structural){reference-type="ref" reference="subsec:structural"} for completeness.
Furthermore, the two antagonistic communities are given by \(U^+=\{i \in V|u_i>0\}\) and \(U^-=\{i \in V| u_i<0\}\) where \(u=[u_1,..., u_n]^T\) is a \(\beta_1(t)\) eigenvector. All edges connecting vertices within the same community are positive while edges connecting two vertices in different communities are negative. When one of \(U^+, U^-\) is empty, the network has only one community. The model in is the Riccati equation \(W'=W^2\), whose solution is \(W(t)=W(0)(1-W(0)t)^{-1}\). As a consequence, Marvel et. al. showed that structure balance occurs in the Riccati equation \(W'=W^2\) for generic initial parameter \(W(0)\) with a positive eigenvalue at finite time. Our model strictly generalizes the previous model and consider how the initial opinions may influence the system evolution. In the following we show how to solve the general Riccati equation and also when an eigenvalue goes to infinity. With most of the details in the Appendix, we carry out rigorous analysis of the general form of the matrix Riccati equation as stated below. \[\label{eqn:Evolution} \left\{ \begin{array}{rl} W' & = W^2 + C \\ W(0) & = B. \end{array} \right.\] For general matrices \(B, C\) we can solve for \(W(t)\) as shown in the following theorem. The proof details can be found in Appendix [1.2](#subsec:general-Riccati){reference-type="ref" reference="subsec:general-Riccati"}.
In our co-evolution model, \(B=W(0)\) is assumed to be symmetric. Thus \(C\) is also symmetric. This allows us to simplify the solution further, as shown in Appendix [1.3](#subsec:socialtie){reference-type="ref" reference="subsec:socialtie"}. Then we analyze a special case when \(BC=CB\). In this case we can characterize the conditions when structural balance is guaranteed to occur, with details in Appendix [1.1](#subsec:structural){reference-type="ref" reference="subsec:structural"}, [1.4](#subsec:BCequalCB){reference-type="ref" reference="subsec:BCequalCB"}, and [1.5](#subsec:structural-commute){reference-type="ref" reference="subsec:structural-commute"}. Specifically, using a basic fact that two commuting symmetric matrices can be simultaneously orthogonally diagonalized , we can get the following theorem where the conditions of the eigenvalues of \(B, C\) for structural balance are characterized:
In our co-evolution model, \(BC=CB\) happens when \(V(0)\) is a eigenvector of \(W(0)\), an interesting initial condition. To see that, recall \(B=W(0)\), \(C= V(0)V(0)^T-W(0)W(0)^T\), and \(W(0)=W(0)^T\). To check if \(BC=CB\), we just need to check if \(W(0)V(0)V(0)^T=V(0)V(0)^TW(0)\) and apply the following Lemma (Proof in Section [1.5](#subsec:structural-commute){reference-type="ref" reference="subsec:structural-commute"}).
Further, the equation in our co-evolution model, i.e., Equation ([\[eqn:Riccati\]](#eqn:Riccati){reference-type="ref" reference="eqn:Riccati"}), satisfies \(C+B^2=V(0)V(0)^T\). Notice that the right-hand side \(V(0)V(0)^T\) is an \(n \times n\) matrix with rank one. This property actually ensures that the conditions characterized in Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} are met and thus structural balance is guaranteed. At the same time, the convergence rate is \(O(\frac{1}{|T-t|})\), which is proved by Lemma [\[lemma:sin\]](#lemma:sin){reference-type="ref" reference="lemma:sin"} in Appendix [1.4](#subsec:BCequalCB){reference-type="ref" reference="subsec:BCequalCB"}.
If \(V(0)=0\), the system stays at the fixed point with \(V\) remaining zero and the weight matrix unchanged. The case when \(V(0)\) is an eigenvector of \(W(0)\) includes a few interesting cases in practice. When \(W(0)=0\) or \(W(0)=I\), this models a group of individuals that start as complete strangers with uniform self-appraisals. Their non-homogeneous initial opinions \(V(0)\) may drive the network to be segmented over time. Finally the fact that \(V(t)\) remains an eigenvector of \(W(t)\) for all \(t\) follows from Equation ([\[eqn:WWW\]](#eqn:WWW){reference-type="ref" reference="eqn:WWW"}) in Appendix C. We conjecture that even in the general case (when \(V(0)\) is not necessarily an eigenvector of \(W(0)\)) the limit vector \(\lim_{t \to T} V(t)/|V(t)|\) is an eigenvector of the limit tie relation matrix \(\lim_{t \to T} W(t)/|V(t)|\). This is supported by our numerical evidences and Corollary [\[cor:wv\]](#cor:wv){reference-type="ref" reference="cor:wv"} which says that \(\lim_{t \to T} V(t)/|V(t)|\) is an eigenvector of \(\lim_{t \to T} (W(t)/|V(t)|)^2\).
# Simulations
In this section, we provide simulation results to accompany our theoretical analysis of the co-evolution model. We also present simulation results on general graphs and real world data sets to understand the behavior of network evolution. Here is a brief summary of observations from simulations. The details can be found in the Appendix.
- Except a few carefully crafted cases[^2] that make the opinion vector to be zero, strict structural balance is always reached, regardless of whether the graph is complete or general, whether \(B, C\) commute or not, or whether \(W(0)\) is symmetric or as a general matrix. An example is shown in Figure [\[fig:heatmap\]](#fig:heatmap){reference-type="ref" reference="fig:heatmap"}.
- In general, we observed through simulations that the number of iterations to convergence is inversely proportional to the magnitude of initial opinions and the edge density in the graph.
- In two real world data sets (the Karate club and a political blog data set), a few initially planted polarized opinions can successfully predict the ground truth community structure with high accuracy. See Figure [\[fig:application\]](#fig:application){reference-type="ref" reference="fig:application"}.
# Conclusion and Future Work
In this paper, we have provided a co-evolution model for both opinion dynamics and appraisal dynamics. We provided solutions to the system and rigorously characterized how the stable states depend on the initial parameters. There are a few follow-up problems that remain open, for example, when the social ties are directional/asymmetric, when the network is not a complete graph, and when each agent has an \(m\)-dimensional opinion vector. We include some discussion and conjectures on these cases in Appendix [1.6](#subsec:general){reference-type="ref" reference="subsec:general"} and consider this as interesting future work.
# Model
[\[subsec:model\]]{#subsec:model label="subsec:model"}
# Evolution of Opinion Dynamics {#app:opinion}
Indeed, by \((|V(t)|^2)'=2V^TV'\) and Equation ([\[eqn:opinion_evolution\]](#eqn:opinion_evolution){reference-type="ref" reference="eqn:opinion_evolution"}) that \(V''=2|V|^2V-CV\), we have \((|V(t)|^2)''=2 ( V')^T V' + 2V^T V'' = 2|V'(t)|^2 + 2V^T(2|V|^2V-CV) = 2|V'(t)|^2 +4|V|^4-2V^TCV.\) Comparing it with Equation ([\[eqn:length\]](#eqn:length){reference-type="ref" reference="eqn:length"}), we see the corollary holds. The last statement of the corollary follows from Equation ([\[v\'\]](#v'){reference-type="ref" reference="v'"}). We now prove Proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"} using several lemmas. For simplicity, if \(f(t)\) is a function defined on an open internal \((a, b)\) (here \(b\) may be \(+\infty\)), we say \(f\) has property \(P\) (e.g., positive, non-negative, monotonic, convex etc) near \(b\) if there exists \(\epsilon > 0\) such that the restriction of \(f\) on the interval \((b-\epsilon, b)\) (if \(b < +\infty\)) or \((\frac{1}{\epsilon}, \infty)\) (if \(b=+\infty\)) has property \(P\). For example, \(t^4 + 3t^3-9 t^2 +t-5\) is positive and convex near \(+\infty\). The notation \(C^{k}(a,b)\) stands for all functions \(f(t)\) for which \(f, f', \cdots, f^{(k)}\) are continuous on the interval \((a, b)\).
Now, let us prove Proposition [\[prop:growth\]](#prop:growth){reference-type="ref" reference="prop:growth"}.
Now, for \(T < +\infty\), Theorem [\[thm:differential\]](#thm:differential){reference-type="ref" reference="thm:differential"} implies \(\lim_{t\rightarrow T} |V(t)| = +\infty\). Indeed, if \(\lim_{t\rightarrow T} |V(t)| < + \infty\), then \(V(t)\) is bounded on \([0, T)\). This implies \(W'(t) = V(t)\cdot V(t)^T\) is bounded. but \(W(t) = W(0) + \int_0^t W'(s) ds\). Therefore W(t) is bounded. This implies the solution \((V(t), W(t))\) for \(t\in [0, T)\) lies in a bounded set in \(\mathbb{R}^n \times \mathbb{R}^{n\times n}\) which contradicts Theorem [\[thm:differential\]](#thm:differential){reference-type="ref" reference="thm:differential"}. Now \(\lim_{t\rightarrow T} |V(t)|^2 = \lim_{t \rightarrow T} \sum_{j=1}^n \lambda_j^2(t) = + \infty\) implies, by Corollary [\[corollary:sol\]](#corollary:sol){reference-type="ref" reference="corollary:sol"}(2), there exists \(i\) for which \(\lim_{t \rightarrow T} \lambda_i^2(t) = + \infty\). Furthermore, by Corollary [\[corollary:sol\]](#corollary:sol){reference-type="ref" reference="corollary:sol"}(3), there exists an index \(h\) for which \(\lambda_h^2(t) \geq \lambda_j^2(t)\) near \(T\) for all \(j\). Thus \(\lambda_h^2(t)\) has the largest growth rate tending \(+ \infty\) as \(t\rightarrow T\). For generic initial value \(V(0)\) and \(W(0)\), \(\lambda_{h}^2(t)\) is the unique term of maximum growth rate. Therefore, we see that part (1) of Therem [\[thm:main_theorem\]](#thm:main_theorem){reference-type="ref" reference="thm:main_theorem"} holds. **Case 2.** If \(T = + \infty\), let us assume that the limit \(\lim_{t \rightarrow T} |V(t)| = L > 0\) and show that structural balance occurs eventually. By Corollary [\[corollary:sol\]](#corollary:sol){reference-type="ref" reference="corollary:sol"}, we may assume that \(\lambda_h^2(t) \geq \lambda_j^2(t)\) for \(t\) near \(\infty\) for all \(j\) (if \(\lim_{t \to T} |V(t)|^2=\infty\)) or \(\lim_{t \rightarrow T} \lambda_h^2(t) =\max\{ \lim_{ t \to T} \lambda^2(t) | i=1..., n\} < \infty\). Let \(L' =\lim_{t \rightarrow T} \lambda_h^2(t)\). Then \(L'>0\) since \(L>0\). In the case of \(\lim_{t \to T} |V(t)|^2=\infty < \infty\), for generic initial value, we may assume that \(\lim_{t \rightarrow T} \lambda_h^2(t)\) is the unique maximum value among all \(\lim_{t \rightarrow T} \lambda_i^2(t)\), \(i=1,..., n\). Then we see that \[\int_0^t |\lambda_h^2(s)| ds \geq (t-t_0) L'' + c_0\] for some constants \(L''>0\) and \(c_0\). It tends to \(+ \infty\) as \(t \rightarrow \infty\). Furthermore, by the Cauchy inequality, \[(\int_{0}^t \lambda_i(s) \lambda_j(s) ds )^2 \leq \int_{0}^{t} \lambda_i^2(s) ds \cdot \int_0^{t} \lambda_j^2(s) ds\] and \[\int_0^{t} \lambda_i^2(s) ds \leq \int_0^t \lambda_h^2(s) ds + c_1\] for t large, we see the growth rate of \(\int_{0}^t \lambda_i(s) \lambda_j(s) ds\) is at most that of \(\int_0^t \lambda_h^2(s) ds\) as \(t \to \infty\). This shows, by the same argument, that the growth rate of \(w_{ij}(t)\) is dominated by \(u_{ih} u_{jh}\int_0^t \lambda_h^2(s) ds\). Therefore, structural balance occurs again for generic initial values. Finally, we prove that in Case 1 or in Case 2 that \(T=\infty\) such that \(\lim_{t \to \infty} |V(t)| =L >0\), the limit \(\lim_{t \to \infty} V(t)/|V(t)|\) exists. By corollary [\[cor:v\'\]](#cor:v'){reference-type="ref" reference="cor:v'"} (2), \(\lim_{t \to T} \lambda_i\) exists in \([-\infty, \infty]\). Therefore, if \(\lim_{t \to \infty} |V(t)| =L\) is a finite positive number, then \(\lim_{t \to \infty} V(t)\) exists in \(\mathbb{R}^n-\{0\}\). Hence \(\lim_{t \to \infty} V(t)/|V(t)|\) exists. In the remaining cases, we have \(\lim_{ t \to T} |V(t)|=\infty\). In this case, by the argument above, we see that \(\lambda_i(t)\) has the same sign near \(T\). We claim that the function \(\lambda_i/\lambda_j\) is monotonic near \(T\). Indeed, by the quotient rule for derivative, the sign of derivative of \(\lambda_i/\lambda_j\) is the same as that of \(h(t)=\lambda_i'\lambda_j-\lambda_j'\lambda_i\). Now \(h'(t) = \lambda_i''\lambda_j-\lambda_j''\lambda_i= (2|V|^2-a_i)\lambda_i\lambda_j-(2|V|^2-a_j)\lambda_j\lambda_i=(a_j-a_i)\lambda_i\lambda_j\). Therefore, either \(h'(t)\) has the same sign for \(t\) near \(T\) (when \(a_i \neq a_j)\) or \(h'(t)=0\) (when \(a_i =a_j)\). If \(h'(t)=0\), then \(\lambda_i/\lambda_j\) is a constant near \(T\) and the claim follows. If \(h'(t)\) has the same sign near \(T\), then \(h(t)\) is strictly monotonic near \(T\). Therefore, \(h(t)\) has the same sign for \(t\) near \(T\). As a consequence we see that \(\lambda_i/\lambda_j\) is a monotonic near \(T\). In particular, the limit \(\lim_{ t \to T} \lambda_i(t)/\lambda_j(t)\) exists. Since \(\lim_{t \to T} \sum_{k=1}^n \lambda_k(t)^2=\infty\), this implies the limit \(\lim_{t \to T}\frac{\lambda_i(t)}{\sqrt{\sum_{j=1}^n \lambda_j(t)^2}}=\lim_{t \to T} \frac{\pm 1}{\sqrt{\sum_{j=1}^n \lambda_j(t)^2/\lambda_i(t)^2}}\) exists for any index \(i\). The last statement is the same as that \(\lim_{t \to T} V(t)/|V(t)|\) exists. This ends the proof of Theorem [\[thm:main_theorem\]](#thm:main_theorem){reference-type="ref" reference="thm:main_theorem"}. ◻
To see this, let \(v\) be \(\lim_{t \to T} V/|V|\). From \(W^2=VV^T-C\) and \(\lim_{t \to T} |V(t)| =\infty\), we see that \(W^2/|V|^2 =(V/|V|) (V/|V|)^T-C/|V|^2\) implies \(\lim_{ t \to T} W^2/|V|^2 =vv^T\). Since \(vv^Tv=v\) due to \(|v|^2=1\), the result follows. We end the appendix by making several remarks and a conjecture. The 1-dimensional case of equation \(V''(t) = (2|V(t)|^2-C)\cdot V(t)\), \(V(t) \in \mathbb{R}^n\) is \(y''(t) = 2y^3(t)-cy(t)\). The function \(f(t) = \frac{a}{\sinh{(at + b)}}\) solves \(f'' = 2f^3 + a^2f\) and \(g(t) = \frac{a}{\sin(at+b)}\) solves \(g'' = 2g^3-a^2g\). We may assume \(a>0\). Therefore, if \(b<0\), then both \(f\) and \(g\) exist only on a finite maximum interval (it is \([0,-\frac{b}{a})\) for \(f(t)\)), i.e., \(T < + \infty\). If \(b > 0\), then the function \(f(t)\) exists on \([0, + \infty)\) but the function \(g(t)\) exists only on a finite interval \([0, T)\). It indicates that if \(C\) has a positive eigenvalue \(a_1>0\), then the solution \(\lambda_1(t)\) may exist only on a finite interval. This prompts us to conjecture that
If the conjecture holds, by Theorem [\[thm:main_theorem\]](#thm:main_theorem){reference-type="ref" reference="thm:main_theorem"}, we see structural balance must occur eventually for generic initial value \(C\) which has a positive eigenvalue. Therefore, it also justifies our experimental observation that structural balance occurs almost all the time.
# Simulation
We use the discrete time model as described in Equation ([\[eq:discrete-model\]](#eq:discrete-model){reference-type="ref" reference="eq:discrete-model"}). The convergence of the dynamic process defined by Equation ([\[eq:discrete-model\]](#eq:discrete-model){reference-type="ref" reference="eq:discrete-model"}) is very fast. For visualization purposes, in the simulation, we set our evolution model using Equation ([\[modii\]](#modii){reference-type="ref" reference="modii"}) with \(a = b = 0.01\). When a graph is said to satisfy structural balance, the multiplication of weights along all triangles is non-negative. In addition, either all nodes have the same opinion (i.e., harmony) or the graph is partitioned into two antagonistic groups (i.e., polarization).
## Harmony vs Polarization
In this section, we work with a complete graph and examine when structural balance and/or polarization appears. **\(B\) and \(C\) commute.** As mentioned in Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, when the matrices \(B = W(0)\) and \(C = V(0)V(0)^T-W(0)^2\) are symmetric with \(BC=CB\), structural balance occurs in finite time for \(W(t)\) if \(W(t)\) has an unique largest eigenvalue. We take three special cases mentioned in the previous section, namely when \(W(0) = I\), \(W(0) = 0\), or \(W(0) = V(0)V(0)^T\). The evolution process of these three cases are shown in Figure [\[fig:commute_evolution\]](#fig:commute_evolution){reference-type="ref" reference="fig:commute_evolution"}. The initial opinions are selected uniformly at random in \([-1, 1]\). The dash curves represent the evolution of opinions and the solid ones shows the edge weights. The color of the curves represent the final sign of opinions and weights at convergence. In all three cases the network reaches polarization where the opinions of some vertices go to positive infinity and the opinion of the others go to negative infinity. In these cases, the node opinions do not change signs, because \(v_iw_{ij}v_j \geq 0\) holds all the time. The gradient direction of opinions and weights are same with their signs. Figure [\[fig:case0_C\]](#fig:case0_C){reference-type="ref" reference="fig:case0_C"} shows a case when both opinions and edge weights converge to positive values. All the opinions are assigned initially as a positive value. The initial weight matrix is a diagonal matrix with the same diagonal entry. All final opinions and weights are positive. **\(B\) and \(C\) do not commute.** Next, we consider the case when the matrices \(B\) and \(C\) do not commute. We take different random initial cases and show the evolution process in Figure [\[fig:non_commute_evolution\]](#fig:non_commute_evolution){reference-type="ref" reference="fig:non_commute_evolution"}. In the first two plots [\[fig:case0_NC\]](#fig:case0_NC){reference-type="ref" reference="fig:case0_NC"} and [\[fig:case1_NC\]](#fig:case1_NC){reference-type="ref" reference="fig:case1_NC"}, we select a symmetric random matrix as the initial weight matrix \(W(0)\). The opinions are assigned random initial values. In simulation we observed both cases of harmony and polarization as the final state. Some edges and vertices change signs in the process. We also show an example of the evolution process in Figure [\[fig:heatmap\]](#fig:heatmap){reference-type="ref" reference="fig:heatmap"} which shows how two communities emerge. **\(W(0)\) is not symmetric.** In Figure [\[fig:case2_NC\]](#fig:case2_NC){reference-type="ref" reference="fig:case2_NC"}, the initial weight matrix is not symmetric. Both vertex opinions and edge weights go to infinity after several iterations. Figure [\[fig:case3_NC\]](#fig:case3_NC){reference-type="ref" reference="fig:case3_NC"} shows when all the entries in the matrix and opinions are initially negative. Some opinions and weights change signs in the evolution process. For all cases we have tested when \(B\) and \(C\) do not commute, structural balance is always satisfied in the limit.
## Convergence Rate
In section, we check the convergence rate in different settings. This helps us understand intuitively the factors that influence the convergence rate. **Magnitude of Initial Opinions.** Figure [\[fig:convergence_1\]](#fig:convergence_1){reference-type="ref" reference="fig:convergence_1"} and [\[fig:convergence_2\]](#fig:convergence_2){reference-type="ref" reference="fig:convergence_2"} show two cases \(W(0) = I\) and \(W(0) = 0\) respectively. In both cases \(BC=CB\). The initial opinions are randomly selected from \((-1, 1)\). We check the number of iterations until all entries in the opinion vector and weight matrix have absolute value larger than \(10^{20}\). From the analysis in the previous section, the largest eigenvalue of \(W(t)\) determines the convergence rate. Given that the initial weights are determined, we check the positive eigenvalue of \(V(0)V(0)^T\). We can see that the number of iteration until convergence is inversely proportional to the magnitude of the positive eigenvalue of \(V(0)V(0)^T\). The more extreme the initial opinions are, the faster the network reaches convergence. We also tested networks of different sizes, which seems to be generally oblivious to the convergence rate. **\(W(0)\) as a Random Matrix.** In Figure [\[fig:convergence_3\]](#fig:convergence_3){reference-type="ref" reference="fig:convergence_3"}, we compare the case of \(W(0) = V(0)V(0)^T\) with \(W(0)\) being a random symmetric matrix (\(W(0)\) and \(C\) are generally not commutative). The orange dots show the number of iterations till convergence when entries in \(W(0)\) are selected uniformly in \((-1, 1)\). The green dots show the number of iterations until convergence when entries in \(W(0)\) are selected uniformly in \((-5, 5)\). There are a few observations. 1) When \(W(0)\) take random values, it requires more iterations to convergence compared to the case when \(W(0) = V(0)V(0)^T\). Specifically, during the evolution when \(W(0) = V(0)V(0)^T\) both opinions and weights do not change signs. 2) When the initial weights take greater absolute values in general, the system converges faster. Again the more extreme the opinions/weights are, the faster the system reaches structural balance. **On a General Graph.** We also test on graphs generated by social network models. In Figure [\[fig:convergence_4\]](#fig:convergence_4){reference-type="ref" reference="fig:convergence_4"}, we show the convergence results on graphs generated by the Erdős-Renyi Model \(G(n, p)\), where \(n\) is the number of nodes and \(p\) is the probability of each pair of nodes connected by an edge. As \(p\) increases from \(0\) to \(1\), the graph becomes denser and it requires fewer iterations to converge. In Figure [\[fig:convergence_5\]](#fig:convergence_5){reference-type="ref" reference="fig:convergence_5"} and [\[fig:convergence_6\]](#fig:convergence_6){reference-type="ref" reference="fig:convergence_6"}, we tested on Watts Strogatz model \(G(n, k, p)\), where \(n\) is the number of nodes, \(k\) is the number of neighbors of each nodes and \(p\) is the rewiring probability. The network starts as a regular ring lattice, where each node connects to \(k\) nearest neighbors on the ring. Each neighbor has a probability \(p\) to be 'rewired' to another non-neighbor node. Thus, with a larger value of \(k\), there are more edges in the graph. With a larger rewiring probability \(p\), there is more randomness in the graph. In these simulations, the initial weight matrices are set as a zero matrix, i.e., \(W(0) = 0\). Similarly, with the same number of edges in the graph, the relationship between the positive eigenvalue and the number of iterations follows the similar trend. When there are more edges in the graph, the convergence speed increases. In all cases, the network reaches structural balance.
## Emergence of Community Structure
In this section, we ran experiments on real-world network data sets. We are particularly interested in the following question. Can a few planted seeds with opposite opinions influence the other nodes and drive the network into structural balance and polarization? Does the final state coincide with the community structure in the network? Our first experiment is based on a study by Zachary who witnessed the breakup of a karate club into two small clubs. As shown in Figure [\[fig:karate_1\]](#fig:karate_1){reference-type="ref" reference="fig:karate_1"}, the networks captures \(34\) members, documenting links between pairs of members who interacted outside the club. During the study, a conflict arose between the administrator (label \(0\)) and the instructor (label \(33\)), which led to the split. The red and green nodes represent the choice of each individual in the end. In this experiment, we assign the administrator and the instructor opposite opinions as \(1\) and \(-1\). The other members start with opinion \(0\). Given that the links represent interaction and positive friendship between members, each edge is assigned a small positive value in the initial matrix \(W(0)\). We then run our co-evolution dynamics till convergence. The edges with negative weight are removed. The graph is separated into two communities, shown in Figure [\[fig:karate_2\]](#fig:karate_2){reference-type="ref" reference="fig:karate_2"}. It nearly predicted the same division as in the ground truth except for two members (\(\#8\) and \(\#19\)) which are somewhat ambiguous. The second experiment is based on the political blogs network. It is a directed network of hyperlinks between weblogs on US politics, recorded in 2005 by Adamic and Glance . There are \(1,490\) nodes and \(19,025\) directed edges in the graph. Each node has its political preference (\(-1\) as liberal, \(1\) as conservative) shown in Figure [\[fig:blog_1\]](#fig:blog_1){reference-type="ref" reference="fig:blog_1"}. We randomly select \(20\%\) nodes and assign initial opinions according to their ground truth values. All edges are assigned initial weights as a small positive value. When the graph reaches convergence, two big communities appear, as indicated by their final opinions and the sign of edges. Figure [\[fig:blog_2\]](#fig:blog_2){reference-type="ref" reference="fig:blog_2"} shows the detected communities after negative edges are removed. Compared with the ground truth, the predicted opinions by our dynamical model has an accuracy of \(97.21\%\), averaged by \(200\) simulation runs. If only \(3\%\) nodes are assigned ground truth opinions in the initial state, the prediction accuracy for the final opinions of all nodes, on average, is as high as \(82.12\%\). Our dynamic model, as shown by these experiments, explains why community structures appear. It can also be understood as an algorithm for label propagation or node classification. Compared with other methods for the same task that generally use data-driven machine learning approaches, our dynamic model has better transparency and interpretability.
[^1]: For example, if \(W(0) = B = \diag(-\frac{1}{a},-\frac{1}{a}, \cdots,-\frac{1}{a})\) and we use the dynamic update rule \(v_i(1) = v_i(0) + a\cdot w_{ii}(0)v_i(0) = 0\), for all \(1 \leq i \leq n\). After one iteration, all node opinions become \(0\). After that, the opinions and weights do not change anymore.
[^2]: For example, if \(W(0) = B = \diag(-\frac{1}{a},-\frac{1}{a}, \cdots,-\frac{1}{a})\) and we use the dynamic update rule \(v_i(1) = v_i(0) + a\cdot w_{ii}(0)v_i(0) = 0\), for all \(1 \leq i \leq n\). After one iteration, all node opinions become \(0\). After that, the opinions and weights do not change anymore. | {'timestamp': '2021-11-05T01:06:31', 'yymm': '2107', 'arxiv_id': '2107.05796', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05796'} |
null | null |
# Introduction {#sec:1}
In the most widely accepted picture of structure formation, the initial seeds of gravitational instability are inherited from the adiabatic density fluctuations in the Big Bang, which can be well measured from the cosmic microwave background. After the recombination era, the competition between the gravitational force and cosmic expansion leads to the formation of dark matter halos at small scales in the regions where the linear density contrast surpasses the threshold. On large scales, both observations of galaxy surveys and Sloan Digital Sky Survey and numerical simulations have revealed that the matter distribution can be characterized as a cosmic web (e.g. voids, walls, filaments, and clusters). As dark matter halos are residing in the cosmic web, one would expect that properties of dark matter halos are correlated with their surrounding cosmic environments or large-scale structures (hereafter LSS). The correlation between halo shape/spin and LSS has been extensively investigated in detail (see table 1 and table 2 in, also see table 1 in ) in the last two decades in both observations and theoretical work. On the theory side, many studies that used N-body and hydrodynamical simulations have reached almost converged conclusion that the major axes of halos/galaxies tend to be parallel with the slowest collapse direction of LSS, and the vector of angular momentum (spin) of halos/galaxies is preferentially either aligned or perpendicular to the LSS, depending on the halo/galaxy mass. Along with the theoretical efforts, observational studies have focused on testing the results from simulations about the correlations between shape and spin of galaxy with the large scale structure. Using galaxy shape as a proxy of galaxy spin, several studies claimed that galaxy spin of low-mass, spiral galaxies tend to align with the LSS, while high-mass, elliptical galaxies are preferentially perpendicular to the LSS, consistent with the simulation results. Some investigations also found that red central galaxies show the strongest alignment signal between galaxies major axes with the host filaments or walls. Generally speaking, the spin of a dark matter halo is the composition of the linear collapse of cosmological density field during the early stages, and the nonlinear collapse and mergers with other halos during the late stages. The former part can be well explained by the linear theory of the large scale tidal field, i.e., tidal torque theory. However, the later nonlinear part, which can only be studied with the help of N-body simulations, could be more dominant in the final angular momentum acquisition. For a long time, the correlations between halo properties and the LSS environment were usually explored in a statistical way using halo population as whole. If we focus on the merger history of every single halo, the individual differences can blur any conclusion that can be drawn. On the other hand, if we want to study the effect of the LSS direction or the environment, another halo sample with different merger histories has to be chosen from the simulation. It is thus difficult to study the effect of halo and environment separately. The degeneracy of LSS environment and nonlinear factors originating from small scales weaken our understanding of these halo-LSS conclusions mentioned above. Consequently, it's necessary to set up controlled N-body simulation in which we can somehow manipulate the large-scale environment of the halo and see how properties of given halos correspond to the changes of their LSS environments. In this work, we implement twin N-body simulations (explained in detail in Sec. [6](#sec:2){reference-type="ref" reference="sec:2"}) to show how the spin (and major-axis) of matching halos evolves when the \(\textbf{e}_3\) vector (i.e. the least compressed direction) changes its direction. This paper is organized as follows. In Sec. [6](#sec:2){reference-type="ref" reference="sec:2"} we present the method to produce twin universes using N-body simulation data and how to identify identical or matching dark matter halos. In Sec. [7](#sec:3){reference-type="ref" reference="sec:3"} we present the main results of the work. As a first step, we show the halo mass function and the correlations between halo spin/major-axis and \(\textbf{e}_3\) vector of the LSS to check if the twin simulations are equal statistically. We then show how the spin/major axis of halo changes correspond to the change of the LSS environments. Lastly, the influence of LSS on halo shape parameters (\(S\), \(Q\) and \(T\), see their definition in Sec. [7](#sec:3){reference-type="ref" reference="sec:3"}) will be explored. Conclusions and discussions are presented in Sec. [8](#sec:4){reference-type="ref" reference="sec:4"}.
# Methods and Simulations {#sec:2}
According to linear theory, the power spectrum of cosmic density perturbations and the two-point correlation function are Fourier transform pair, which can be expressed as follow: \[\label{equ:equ1}\] where \(P\left(k\right)=V_u\left\langle\vert\delta_{\textbf{k}}\vert^2\right\rangle\) is the power spectrum, \(x=\vert\textbf{x}\vert\) is the distance, \(k\) is the wave number, \(V_u\) is the volume we consider. This relation often holds true for \(V_u\rightarrow\infty\), or equivalently for \(k\rightarrow0\). This indicates that the information of LSS in the late stages is totally contained in the small \(k\) of the initial over-density field \(\delta\left(k\right)\), just like fingerprints. In the most favoured frame for the generation of \(\Lambda\)CDM initial conditions, the power spectrum and the phase information are encapsulated as a Gaussian white noise field, which is particularly convenient to work with numerically. One can expect that the whole LSS pattern will change severely if we replace the random seeds for the phase of those \(k\) mode smaller than a given scale, while the formation of dark matter halos on small scales will not be significantly affected. Besides, with the help of octree basis function, the LSS controlled initial conditions can also be generated in real-space. This kind of technique paves a new way to manipulate the LSS environment of dark matter halos by a couple of N-body simulations. In this work, we produce twin N-body simulations L500N1024, L500N1024a, L500N1024b, L500N1024c (hereafter N1024, N1024a, N1024b and N1024c). The letters "a/b/c\" here mark the three simulations for which the random seeds for small \(k\) in the initial conditions have been changed, i.e., we replace the random seeds for the phase of \(k\) mode such that: \[\label{equ:equ2} \vert\textbf{k}\vert<\frac{2\pi}{L_0}\;.\] Note that \(L_{0}\) is an arbitrarily selected scale (discussed later) and should not be confused with the box-size of the simulation. In addition, we also run an independent simulation L500N1024r (hereafter N1024r) as a reference, for which the random seed for the phase of all \(k\) modes is completely different from that used in the twin simulations. This reference simulation is only used to test the stochastic background of our method. For the N-body simulations, the initial conditions are set at redshift \(z=127\). At this initial time, the dark matter particles which later settle into a typical halo (with mass around \(10^{13}M_{\odot}\) at \(z=0\)**)** occupied an initial volume about several cubic megaparsec. After some simple tests, we choose \(L_0=20 h^{-1}\mathrm{Mpc}\). This scale is selected to produce significant changes of large-scale environments for most halos, but ensuring halo identities (more explicit definition in Sec. [6.1](#sec:2.1){reference-type="ref" reference="sec:2.1"}) are not severely affected. Here we remind the reader that, in terms of statistical properties, all twin simulations are identical. Unless otherwise specified, all figures in this paper are plotted in a symmetric way, i.e., the conclusions do not change if we switch the N-body simulations. That is why we call them twin simulations. Our simulations were run using the P-Gadget-3 code, a TreePM code based on the publicly available code Gadget-2, and they all contain \(1024^3\) dark matter particles in a periodic box of \(500h^{-1}\mathrm{Mpc}\). The cosmological parameters are chosen from the WMAP7 data, namely: \(\Omega_{\Lambda}=0.7274\), \(\Omega_m=0.2726\), \(h=0.704\) and \(\sigma_8=0.821\). The mass of single particle in our simulation is \(8.8\times10^9\;h^{-1}\mathrm{M}_{\odot}\), and \(136\) snapshots are stored from \(z=127\) to \(z=0\). We identify dark matter halos at each snapshot using the standard FoF (friend-of-friend) algorithm with a linking length that is 0.2 times the mean inter-particle separation. To ensure relatively robust results of halo spin and shape, and to avoid the spurious matches to the same halo purely by chance, the main halos containing at least 100 particles are chosen for further analysis. For each FoF halo, we identify subhalos using the SUBFIND algorithm and construct merger trees for N1024 and N1024c, respectively. Around \(\sim568~000\) main halo candidates were identified from each simulation (see Tab. [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"}).
## Match ratio and matching halos {#sec:2.1}
There is no unique way to match halos across N-body simulations. For example, use the halo's \(50\) most bound particles and a mass criterion to select matching halos. Here we simply define the match ratio between halos from the twin N-body simulations as Eq. [\[equ:equ3\]](#equ:equ3){reference-type="ref" reference="equ:equ3"}. Supposing main halo \(A\) from N1024 contains \(a\) dark matter particles, main halo \(B\) from N1024a contains \(b\) particles, and they have \(c\) common particles according to the particle ids, then the match ratio is defined as: \[\label{equ:equ3} R\left(A,B\right)=\frac{c^2}{ab}\;.\] Generally speaking, when the random seeds for the phase of small \(k\) are changed, a dark matter halo in N1024 may go through different evolutionary paths, such as being split into a few FoF groups, merge with other halos, or simply disappear. In other words main halo \(A\) may be linked to multiple corresponding halos in N1024a. Each of them has a match ratio with main halo \(A\) according to the Eq. [\[equ:equ3\]](#equ:equ3){reference-type="ref" reference="equ:equ3"} and the one with the largest ratio will be chosen as the matching halo of \(A\) in N1024a (and vice versa). To ensure that there are enough matched halos for a statistical analysis, we set an arbitrary threshold to \(R\left(A,B\right)\geq0.4\) for matching halo pairs, which roughly corresponds to each halo containing more than \(65\%\) of its matching halo. After this selection, \(157~748\) matching halo pairs were identified from N1024 and N1024a simulation at \(z=0\). More information about the number of identified matching halo pairs are listed in Tab. [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"} and Tab. [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}. Moreover, the number of main halos containing at least 100 particles at \(z=2\) and \(z=1\) is also listed.
As mentioned in Sec. [5](#sec:1){reference-type="ref" reference="sec:1"}, once the sample of matching halos between the twin N-body simulations is established, we can study the halo-LSS correlations. We employ the Hessian matrix method used in many previous articles to define the matrix as: \[\label{equ:euq4} H_{ij}=\frac{\partial^2\rho_s\left(\textbf{x}\right)}{\partial x_j\partial x_i}\;.\] This method is based on the smoothed density field \(\rho_s\left(\textbf{x}\right)\) at the halo position (based on a more accurate and improved algorithm by ), which can be given by the Cloud-in-Cell (CIC) technique. The smoothing length \(R_s\) is the only parameter of CIC, which can be regarded as the typical scale of halo LSS environment identified by Hessian matrix method. Many previous works used a constant smoothing length. To determine which Rs value should be chosen, we test some LSS properties (e.g., environment, eigenvectors) of matching halos for three fixed Rs: \(2.5,\;5,\;10h^{-1}\mathrm{Mpc}\). We find that the halo-LSS correlation, as well as other main conclusions we make below, become stronger as Rs decreases, which is a reasonable result according to our previous discussions. Consequently \(R_s=2.5h^{-1}\mathrm{Mpc}\) is chosen hereafter. The three eigenvalues of the Hessian matrix are marked as \(\lambda_1\), \(\lambda_2\), \(\lambda_3\), with corresponding eigenvectors \(\textbf{e}_1\), \(\textbf{e}_2\) and \(\textbf{e}_3\). Eigenvalues \(\lambda_i\) can be used to define the LSS environment of dark matter halos according to the number of positive eigenvalues, i.e.,
- void: \(0<\lambda_1<\lambda_2<\lambda_3\)
- wall: \(\lambda_1<0<\lambda_2<\lambda_3\)
- filament: \(\lambda_1<\lambda_2<0<\lambda_3\)
- cluster: \(\lambda_1<\lambda_2<\lambda_3<0\)
and the eigenvectors \(\textbf{e}_i\) stand for the three compressed directions of the smoothed density field. \(\textbf{e}_3\) vector indicates the least compressed direction, which is a robust and universal definition of the LSS. In this work, we will focus on the alignment of halo spin and shape with the \(\textbf{e}_3\) vector, i.e., \(\cos\theta_3=\textbf{a}\cdot\textbf{e}_3\), where \(\textbf{a}\) is the halo spin or major-axis vector. Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} displays a \(100\times100\times20h^{-1}\mathrm{Mpc}\) slice from the twin simulations N1024 and N1024a. It presents a global view of the twin simulations. Over two hundred matching halo pairs are identified in this displayed slice, but here we only mark nine matching halos with hollow circles, with size proportional to their virial radii. We marked their main progenitors in upper panels if and only if they are still matched at \(z=1\). We can see that the LSS environment of these samples really changes tremendously after we replaced the random seeds for the small \(k\) of the initial over-density field.
# Results {#sec:3}
Our main results, mostly regarding the influence of LSS environment transformation on the properties of dark matter halos, will be shown in this section. Before that, we show some statistical results, including the halo mass function and correlations between halo properties (spin and major-axis) and the LSS environment, to test whether the twin simulations are statistically equal. Then we will focus on the mass dependence of the match ratio \(R\) in different LSS environment. Lastly, the influence of LSS on the total spin, major-axis, and shape parameters of the matching halo will be explored. Specifically, the shape parameters of dark matter halo, \(S\), \(Q\) and \(T\), are defined by: \[\label{equ:equ5} S=\frac{c}{a}, \quad Q=\frac{b}{a}, \quad T=\frac{a^2-b^2}{a^2-c^2},\] where \(a\) is the major-axis of modeled ellipsoid (by calculating the inertia tensor) of FoF halo, \(b\) is the intermediate-axis, and \(c\) is the minor-axis. The parameters \(S\) and \(Q\) give the ellipticity of dark matter halos, and \(T\) quantifies the triaxiality. Purely prolate halos have \(T=1\) while purely oblate halos have \(T=0\).
## Statistical tests {#sec:3.1}
Firstly, we show the halo mass function as an important statistical test of our twin simulations. Since the amplitude of the initial power spectrum of the twin simulations is exactly the same, one would expect that the difference in their halo mass function is very small. On the other hand, each N-body simulation has a finite box size, hence the cosmic variance between different simulations will produce different mass functions. Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} shows the halo mass function of N1024, N1024a, N1024b and N1024c. As we can see, all curves are nearly identical except at the high mass end, which can be explained by the effect of cosmic variance (well within the range of grey shadows). And because of the same reason, any similar figure regarding the statistical properties of the twin simulations gives nearly the same result. So we will not distinguish them hereafter. In Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} we show another statistical test regarding the mass dependence of correlations between halo spin or halo major-axis with the \(\textbf{e}_3\) vector for matching halo pairs in the twin simulations. The upper two panels are for halo spin, while the lower two panels are for halo major-axis. This figure shows that the correlations between the twin simulations agree with each other within \(1\sigma\) scatter, showing again that the twin simulations are statistically equivalent. In the left panels the halo spin and major axis are calculated using all particles within the virial radius of the halo. The trend of mass dependence of spin-\(\textbf{e}_3\) correlation agrees well with previous works. Note that here we only use halos with more than 100 particles (\(\sim10^{12}\mathrm{M}_{\odot}/h\)), roughly corresponding to the flip mass of the spin-LSS correlation found by previous works. In N-body simulations, the direction of halo spin is often slowly twisted along the halo radius, i.e., there is a misalignment between the inner spin and the total spin. Motivated by the fact that the baryons experience the same tidal forces as the dark matter, the spin of the galaxies embedded in halos usually follows the inner spin of the host halo. To ensure a relatively reliable result, we calculate the spin and major-axis vectors at \(0.3\) times the virial radius to describe the inner properties of matching halo hereafter. The right two panels of Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} show the mass dependence of inner properties (i.e. spin and major-axis at \(0.3\) times the virial radius). It is seen that, comparing to the alignment of total spin and \(\textbf{e}_3\) vector, the correlation of inner spin and \(\textbf{e}_3\) is nearly independent of the mass. On the other hand, as mentioned by, the halo spin-\(\textbf{e}_3\) correlation is stronger at high redshift. Considering that the inner spin is inherited from the total spin at high redshift, the alignment signal of inner spin and \(\textbf{e}_3\) vector is partly lost due to the nonlinear process in the center of high mass halos at later time.
## Match ratio and mergers {#sec:3.2}
In this section we focus on the matching halo pairs between the twin simulations. Firstly in Fig. [\[fig:fig4\]](#fig:fig4){reference-type="ref" reference="fig:fig4"}, we show the mass dependence of match ratio \(R\) in \(6\) twin simulation pairs (i.e., matched halos between N1024 and N1024a, N1024 and N1024b, N1024 and N1024c, N1024a and N1024b, N1024a and N1024c, N1024b and N1024c) at \(z=0\) and \(z=2\). The shaded region marks the maximum and minimum median match ratio, which roughly illustrate the systematic errors introduced by cosmic variance. For each twin simulation pair, the mass dependence of match ratio \(R\) is for the first simulation. If the mass bins are sorted by the second one, as we mentioned in Sec. [6](#sec:2){reference-type="ref" reference="sec:2"}, the conclusions are still the same. The lower limit of the y-axis is set to \(0.4\), which corresponds the arbitrary threshold imposed to select matching halo pairs. It can be seen that the match ratio of low mass halos is relatively higher, and comparing the same line of both panels, we also find that halos at high redshift usually have higher match ratios. For massive halos at lower redshift, the frequent mergers or mass accretion along the large-scale structure decrease the match ratio between the twin simulations. The yellow and red lines are for matching halos located in the different environments in both simulations. The median match ratio \(R\) of halos in filaments decreases much faster compared to those in clusters, indicating that halos in filaments are more influenced by mass accretion (along the filament) than in clusters. Here we shortly explain how the value of match ratio is reduced in minor merger events. Suppose that halo \(A\) merges with another halo \(b\) in N1024, while the matching halo \(A'\) and \(b'\) in N1024a are not involved in any merger event. At the end of the minor merger, the dark matter particles of \(A\) and \(b\) mix together. The information about the matching halo pair \(b\) and \(b'\) in the later snapshots is completely lost, and the match ratio of halo \(A\) and \(A'\) will be reduced because of the external particles from halo \(b\).
## The evolution of spin with LSS {#sec:3.3}
As we mentioned in Sec. [6.1](#sec:2.1){reference-type="ref" reference="sec:2.1"}, the \(\textbf{e}_{3}\) obtained from the hessian matrix indicates the direction of the LSS around a halo. For matching halo pairs in the twin simulation, we calculate the rotational angle of the corresponding eigenvectors between the twin simulation as, \[\label{equ:equ6} \cos\left(\beta\right)=\frac{\textbf{e}_3\cdot \textbf{e}_3'}{\vert\textbf{e}_3\vert\vert\textbf{e}_3'\vert}\;,\] here \(\textbf{e}_3\) and \(\textbf{e}_3'\) are the corresponding eigenvectors of matching halo pairs. In Fig. [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"} the red histogram shows the distribution of the rotation angle between the \(\textbf{e}_{3}\) vector in the twin simulation pair, N1024 and N1024a. It is found that there is more probability to have halo pairs for which the rotation angle of \(\textbf{e}_{3}\) is not significant, which suggests that although the power spectrum on large scales (\(>20\) Mpc/h) is totally different in the twin simulations, the nearby local environment smoothed on 2.5 Mpc/h is more or less maintained by the matched halo pairs. The black histogram shows a comparison with a larger smoothing length with \(R_s=5.0\) Mpc/h. In this case the rotation angle of large-scale environment becomes larger, but still maintaining some level of parallel signal. In this work we are mainly interested in how halo properties, spin or shape, correspond to the change of the large-scale environment, in particular, the direction of the LSS given by \(\textbf{e}_{3}\). To clearly select matching halo pairs with significant change of their large-scale environments, we select halo pairs from the twin simulations that have: \[\label{equ:equ7} \vert\cos\left(\beta\right)\vert\leq 0.707\quad\mathrm{or}\quad \beta\geq45^{\circ}\;.\] In this way we are looking at halo pairs with significant rotation of their \(\textbf{e}_{3}\) vectors. Of course, we can select halo pairs with larger rotation angles to increase the signal, but that will reduce the sample size. Based on Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}, halos in twin simulations show the same spin-LSS correlation. We can then ask what will happen if the original spin (of halo in, for example, N1024) does not change, can it still maintain a similar correlation with the new \(\textbf{e}_3\) from other twin simulation? If not, it means that the spin needs to be changed systematically to match the correlations (as shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}). This can prove that the correlation between spin and LSS is mainly affected by large-scale structures. In Fig. [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"} we plot the distribution of the spin-LSS correlation at three different redshifts for halo pairs with larger rotation angles given by Eq. [\[equ:equ7\]](#equ:equ7){reference-type="ref" reference="equ:equ7"}. The red shaded areas are the PDFs of halos whose spin and \(\textbf{e}_3\) are from the same simulation N1024a. The grey histograms are the cross correlations of the halo spin-\(\textbf{e}_3\) correlation, where halo spin is from N1024, but the \(\textbf{e}_3'\) vector is from N1024a. For matching halo pairs, the spin-LSS correlation in the same simulation, N1024a, is similar to what is shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} and this is not surprising as the selected halos are a subsample of those shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}. It is interesting to note that the cross correlations (black histograms) are reversed, showing a different trend with the red histograms. This result shows that the rotation of \(\textbf{e}_3\) vectors will lead to proper rotation of the halo spin so to maintain the original halo spin-\(\textbf{e}_3\) correlation seen in one simulation. It indicates that the halo spin is mainly affected by the orbital angular momentum of accreted mass from the LSS along the \(\textbf{e}_3\) vector.
## Matching halos and \(\textbf{e}_3\) rotation angle {#sec:3.4}
From Sec. [7.3](#sec:3.3){reference-type="ref" reference="sec:3.3"} we learn that the spin/major-axis will adjust its direction appropriately when the \(\textbf{e}_3\) vector changes across twin simulations. In this subsection we will explore this reaction in more detail. To maintain the spin-LSS correlation, one may intuitively expect that the rotation angle of halo spin/major-axis will depend strongly on the rotation angle \(\beta\) (see Sec. [7.3](#sec:3.3){reference-type="ref" reference="sec:3.3"}, Eq. [\[equ:equ6\]](#equ:equ6){reference-type="ref" reference="equ:equ6"}). More specifically, we expect that the rotation angle of spin is larger for halos with larger \(\beta\) so as to maintain the spin-\(\textbf{e}_{3}\) relation. However, if we check the correlation mentioned above, we will find that the rotation of spin/major-axis is nearly independent of angle \(\beta\). This is a bit surprising and it is related to the 3D space configuration of halo spin and LSS. Even in 2D space, it is easy to configure a case where the \(\textbf{e}_3\) vector changes significantly while the halo spin/major-axis does not need any adjustment to maintain the same spin-LSS correlation. The situation is more complicated in the 3D case. To better show how this could happen, we introduce the cross-product between \(\textbf{e}_3\) vector and spin/major-axis as the normal vector of halo-LSS system: \[\label{equ:equ8} \textbf{n}=\textbf{e}_3\times\textbf{a}\;,\] where \(\textbf{a}\) is the halo spin or major-axis vector. Under this definition, the rotation of the \(\textbf{e}_3\) vector will lead to the corresponding rotation of the **n** vector in most cases. In Fig. [\[fig:fig8\]](#fig:fig8){reference-type="ref" reference="fig:fig8"} we show the relationship between the rotation angle of \(\textbf{e}_{3}\) and the normal vector defined above. Here we can see that a strong positive proportional relationship between the rotation angle \(\beta\) and **n** vectors for all redshifts, i.e., the change of \(\textbf{e}_3\) vector will lead to corresponding rotation of the whole halo-LSS system, in which the correlation between spin/major-axis and \(\textbf{e}_3\) vector always holds. This gives us a picture of how halo spin/major-axis reacts according to the change of LSS. On the other hand, the proportional relationship in both two panels are nearly independent of the redshift (only the **n** vector from major-axis shown in the right panel has a weak evolution from \(z=2\) to the present-day). This implies an universality for the picture described above.
## Halo shape parameters with LSS {#sec:3.5}
Some previous works have revealed the correlations between halo shape and the LSS environment. In this work we also explore the impact of LSS environment on ellipticity and triaxiality of halos (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)). In Fig. [\[fig:fig9\]](#fig:fig9){reference-type="ref" reference="fig:fig9"} the mass dependence of three shape parameters \(S\), \(Q\) and \(T\) at \(z=0\) are shown in three panels, respectively. We note that the results are nearly independent of redshift, therefore, we only show the \(z=0\) results. The red and green lines are for halos in different twin simulations. The solid lines are for halos in specific LSS environment (\(0\) stands for cluster and \(1\) stands for filament). In addition, we also choose the matching halo pairs which go through an environment transformation from cluster to filament. These halos are denoted by the subscript \(0\rightarrow1\). Firstly it can be seen that, statistically, the shapes of halos in different LSS environment have some systematic differences, i.e., compared to cluster halos, halos in filament (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)) are much more spherical and more oblate, which agrees with previous results. On the other hand, lines in the same color are quite consistent for all shape parameters (reminding that there is a symmetry between twin simulations in terms of individual measurements), which proves that the current LSS environment really has dominant impact on the ellipticity and triaxiality of halos, regardless of their environments in the other symmetric simulation. Combined with results presented in the previous section on the spin-LSS correlation, our results suggest that current environment dominates the correlation between halo spin/major axis with the LSS.
# Conclusions and Discussion {#sec:4}
In this paper, we investigate the effect of LSS on halo pairs matched between twin N-body simulations. The only difference between the twin simulations is that in one run, the random seeds for the phase of small \(k\) where replaced when the initial conditions are generated. Once the sample of matching halos at different redshifts is established, one can use them to separate the entangled factors of large and small scales in the formation of dark matter halos. In this work we mainly focus on the correlations between the halo properties (spin, major-axis, shape parameters) and the \(\textbf{e}_3\) vector of the Hessian matrix, which is thought to be an universal definition of the LSS. Our main results are summarized as follows,
- A few statistical tests, including the halo mass function and the halo spin-\(\textbf{e}_3\) correlation, show that the twin simulations are statistically equal, ensuring that our technique for producing controlled twin simulation is robust and reliable.
- The ratio of matching halo pairs in the twin N-body simulations, based on a merit function, decreases with increasing halo mass. This is due to the later stronger evolution for massive halo by merger or mass accretion along the large-scale structure which is different in the twin simulations.
- The halo spin/major-axis will adjust its direction appropriately when LSS environment changes across twin simulations, i.e., the change of \(\textbf{e}_3\) vector will lead to corresponding rotation of the whole halo-LSS system, to ensure the correlation between spin/major-axis and \(\textbf{e}_3\) vector identical in the twin simulations.
- Statistically, halos in filament (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)) are much more spherical and more oblate than cluster halos, and the LSS environment really has dominant impact on ellipticity and triaxiality of halos.
Since our technique of running twin simulations and establishing the matching halo pairs between them are verified to be robust and reliable, in the future, this technique can be implemented in some larger N-body simulations, or even hydrodynamical simulations, to identify matching galaxies embedded in the center of matching halo pairs. Moreover, multi-scale initial conditions for cosmological simulations, e.g. the zoom-in techniques, can also be employed to study how LSS environment influences galaxy formation and evolution in more details. On the other hand, one can degenerate the entangled factors on different scales in the halo and galaxy formation by choosing different \(L_0\) in Eq. [\[equ:equ2\]](#equ:equ2){reference-type="ref" reference="equ:equ2"}. We conclude that twin simulations with controlled initial condition on different scales could be useful to entangle the formation of halo/galaxy and the large-scale environment. We thank the anonymous reviewer for constructive and insightful suggestions that improved this paper. We acknowledge support from the joint Sino-German DFG research Project "The Cosmic Web and its impact on galaxy formation and alignment\" (NSFC No.11861131006, DFG-LI 2015/5-1 )， and the funds of cosmology simulation database (CSD) in the National Basic Science Data Center (NBSDC). HZC and XK acknowledge support from the the NSFC (No. 11825303, 11333008), the 973 program (No. 2015CB857003), and the China Manned Space Program (No. CMS-CSST-2021-A03, CMS-CSST-2021-B01). NIL acknowledges financial support of the Project IDEXLYON at the University of Lyon under the Investments for the Future Program (ANR-16-IDEX-0005).
# Introduction {#sec:1}
In the most widely accepted picture of structure formation, the initial seeds of gravitational instability are inherited from the adiabatic density fluctuations in the Big Bang, which can be well measured from the cosmic microwave background. After the recombination era, the competition between the gravitational force and cosmic expansion leads to the formation of dark matter halos at small scales in the regions where the linear density contrast surpasses the threshold. On large scales, both observations of galaxy surveys and Sloan Digital Sky Survey and numerical simulations have revealed that the matter distribution can be characterized as a cosmic web (e.g. voids, walls, filaments, and clusters). As dark matter halos are residing in the cosmic web, one would expect that properties of dark matter halos are correlated with their surrounding cosmic environments or large-scale structures (hereafter LSS). The correlation between halo shape/spin and LSS has been extensively investigated in detail (see table 1 and table 2 in, also see table 1 in ) in the last two decades in both observations and theoretical work. On the theory side, many studies that used N-body and hydrodynamical simulations have reached almost converged conclusion that the major axes of halos/galaxies tend to be parallel with the slowest collapse direction of LSS, and the vector of angular momentum (spin) of halos/galaxies is preferentially either aligned or perpendicular to the LSS, depending on the halo/galaxy mass. Along with the theoretical efforts, observational studies have focused on testing the results from simulations about the correlations between shape and spin of galaxy with the large scale structure. Using galaxy shape as a proxy of galaxy spin, several studies claimed that galaxy spin of low-mass, spiral galaxies tend to align with the LSS, while high-mass, elliptical galaxies are preferentially perpendicular to the LSS, consistent with the simulation results. Some investigations also found that red central galaxies show the strongest alignment signal between galaxies major axes with the host filaments or walls. Generally speaking, the spin of a dark matter halo is the composition of the linear collapse of cosmological density field during the early stages, and the nonlinear collapse and mergers with other halos during the late stages. The former part can be well explained by the linear theory of the large scale tidal field, i.e., tidal torque theory. However, the later nonlinear part, which can only be studied with the help of N-body simulations, could be more dominant in the final angular momentum acquisition. For a long time, the correlations between halo properties and the LSS environment were usually explored in a statistical way using halo population as whole. If we focus on the merger history of every single halo, the individual differences can blur any conclusion that can be drawn. On the other hand, if we want to study the effect of the LSS direction or the environment, another halo sample with different merger histories has to be chosen from the simulation. It is thus difficult to study the effect of halo and environment separately. The degeneracy of LSS environment and nonlinear factors originating from small scales weaken our understanding of these halo-LSS conclusions mentioned above. Consequently, it's necessary to set up controlled N-body simulation in which we can somehow manipulate the large-scale environment of the halo and see how properties of given halos correspond to the changes of their LSS environments. In this work, we implement twin N-body simulations (explained in detail in Sec. [6](#sec:2){reference-type="ref" reference="sec:2"}) to show how the spin (and major-axis) of matching halos evolves when the \(\textbf{e}_3\) vector (i.e. the least compressed direction) changes its direction. This paper is organized as follows. In Sec. [6](#sec:2){reference-type="ref" reference="sec:2"} we present the method to produce twin universes using N-body simulation data and how to identify identical or matching dark matter halos. In Sec. [7](#sec:3){reference-type="ref" reference="sec:3"} we present the main results of the work. As a first step, we show the halo mass function and the correlations between halo spin/major-axis and \(\textbf{e}_3\) vector of the LSS to check if the twin simulations are equal statistically. We then show how the spin/major axis of halo changes correspond to the change of the LSS environments. Lastly, the influence of LSS on halo shape parameters (\(S\), \(Q\) and \(T\), see their definition in Sec. [7](#sec:3){reference-type="ref" reference="sec:3"}) will be explored. Conclusions and discussions are presented in Sec. [8](#sec:4){reference-type="ref" reference="sec:4"}.
# Methods and Simulations {#sec:2}
According to linear theory, the power spectrum of cosmic density perturbations and the two-point correlation function are Fourier transform pair, which can be expressed as follow: \[\label{equ:equ1}\] where \(P\left(k\right)=V_u\left\langle\vert\delta_{\textbf{k}}\vert^2\right\rangle\) is the power spectrum, \(x=\vert\textbf{x}\vert\) is the distance, \(k\) is the wave number, \(V_u\) is the volume we consider. This relation often holds true for \(V_u\rightarrow\infty\), or equivalently for \(k\rightarrow0\). This indicates that the information of LSS in the late stages is totally contained in the small \(k\) of the initial over-density field \(\delta\left(k\right)\), just like fingerprints. In the most favoured frame for the generation of \(\Lambda\)CDM initial conditions, the power spectrum and the phase information are encapsulated as a Gaussian white noise field, which is particularly convenient to work with numerically. One can expect that the whole LSS pattern will change severely if we replace the random seeds for the phase of those \(k\) mode smaller than a given scale, while the formation of dark matter halos on small scales will not be significantly affected. Besides, with the help of octree basis function, the LSS controlled initial conditions can also be generated in real-space. This kind of technique paves a new way to manipulate the LSS environment of dark matter halos by a couple of N-body simulations. In this work, we produce twin N-body simulations L500N1024, L500N1024a, L500N1024b, L500N1024c (hereafter N1024, N1024a, N1024b and N1024c). The letters "a/b/c\" here mark the three simulations for which the random seeds for small \(k\) in the initial conditions have been changed, i.e., we replace the random seeds for the phase of \(k\) mode such that: \[\label{equ:equ2} \vert\textbf{k}\vert<\frac{2\pi}{L_0}\;.\] Note that \(L_{0}\) is an arbitrarily selected scale (discussed later) and should not be confused with the box-size of the simulation. In addition, we also run an independent simulation L500N1024r (hereafter N1024r) as a reference, for which the random seed for the phase of all \(k\) modes is completely different from that used in the twin simulations. This reference simulation is only used to test the stochastic background of our method. For the N-body simulations, the initial conditions are set at redshift \(z=127\). At this initial time, the dark matter particles which later settle into a typical halo (with mass around \(10^{13}M_{\odot}\) at \(z=0\)**)** occupied an initial volume about several cubic megaparsec. After some simple tests, we choose \(L_0=20 h^{-1}\mathrm{Mpc}\). This scale is selected to produce significant changes of large-scale environments for most halos, but ensuring halo identities (more explicit definition in Sec. [6.1](#sec:2.1){reference-type="ref" reference="sec:2.1"}) are not severely affected. Here we remind the reader that, in terms of statistical properties, all twin simulations are identical. Unless otherwise specified, all figures in this paper are plotted in a symmetric way, i.e., the conclusions do not change if we switch the N-body simulations. That is why we call them twin simulations. Our simulations were run using the P-Gadget-3 code, a TreePM code based on the publicly available code Gadget-2, and they all contain \(1024^3\) dark matter particles in a periodic box of \(500h^{-1}\mathrm{Mpc}\). The cosmological parameters are chosen from the WMAP7 data, namely: \(\Omega_{\Lambda}=0.7274\), \(\Omega_m=0.2726\), \(h=0.704\) and \(\sigma_8=0.821\). The mass of single particle in our simulation is \(8.8\times10^9\;h^{-1}\mathrm{M}_{\odot}\), and \(136\) snapshots are stored from \(z=127\) to \(z=0\). We identify dark matter halos at each snapshot using the standard FoF (friend-of-friend) algorithm with a linking length that is 0.2 times the mean inter-particle separation. To ensure relatively robust results of halo spin and shape, and to avoid the spurious matches to the same halo purely by chance, the main halos containing at least 100 particles are chosen for further analysis. For each FoF halo, we identify subhalos using the SUBFIND algorithm and construct merger trees for N1024 and N1024c, respectively. Around \(\sim568~000\) main halo candidates were identified from each simulation (see Tab. [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"}).
## Match ratio and matching halos {#sec:2.1}
There is no unique way to match halos across N-body simulations. For example, use the halo's \(50\) most bound particles and a mass criterion to select matching halos. Here we simply define the match ratio between halos from the twin N-body simulations as Eq. [\[equ:equ3\]](#equ:equ3){reference-type="ref" reference="equ:equ3"}. Supposing main halo \(A\) from N1024 contains \(a\) dark matter particles, main halo \(B\) from N1024a contains \(b\) particles, and they have \(c\) common particles according to the particle ids, then the match ratio is defined as: \[\label{equ:equ3} R\left(A,B\right)=\frac{c^2}{ab}\;.\] Generally speaking, when the random seeds for the phase of small \(k\) are changed, a dark matter halo in N1024 may go through different evolutionary paths, such as being split into a few FoF groups, merge with other halos, or simply disappear. In other words main halo \(A\) may be linked to multiple corresponding halos in N1024a. Each of them has a match ratio with main halo \(A\) according to the Eq. [\[equ:equ3\]](#equ:equ3){reference-type="ref" reference="equ:equ3"} and the one with the largest ratio will be chosen as the matching halo of \(A\) in N1024a (and vice versa). To ensure that there are enough matched halos for a statistical analysis, we set an arbitrary threshold to \(R\left(A,B\right)\geq0.4\) for matching halo pairs, which roughly corresponds to each halo containing more than \(65\%\) of its matching halo. After this selection, \(157~748\) matching halo pairs were identified from N1024 and N1024a simulation at \(z=0\). More information about the number of identified matching halo pairs are listed in Tab. [\[tab:1\]](#tab:1){reference-type="ref" reference="tab:1"} and Tab. [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}. Moreover, the number of main halos containing at least 100 particles at \(z=2\) and \(z=1\) is also listed.
As mentioned in Sec. [5](#sec:1){reference-type="ref" reference="sec:1"}, once the sample of matching halos between the twin N-body simulations is established, we can study the halo-LSS correlations. We employ the Hessian matrix method used in many previous articles to define the matrix as: \[\label{equ:euq4} H_{ij}=\frac{\partial^2\rho_s\left(\textbf{x}\right)}{\partial x_j\partial x_i}\;.\] This method is based on the smoothed density field \(\rho_s\left(\textbf{x}\right)\) at the halo position (based on a more accurate and improved algorithm by ), which can be given by the Cloud-in-Cell (CIC) technique. The smoothing length \(R_s\) is the only parameter of CIC, which can be regarded as the typical scale of halo LSS environment identified by Hessian matrix method. Many previous works used a constant smoothing length. To determine which Rs value should be chosen, we test some LSS properties (e.g., environment, eigenvectors) of matching halos for three fixed Rs: \(2.5,\;5,\;10h^{-1}\mathrm{Mpc}\). We find that the halo-LSS correlation, as well as other main conclusions we make below, become stronger as Rs decreases, which is a reasonable result according to our previous discussions. Consequently \(R_s=2.5h^{-1}\mathrm{Mpc}\) is chosen hereafter. The three eigenvalues of the Hessian matrix are marked as \(\lambda_1\), \(\lambda_2\), \(\lambda_3\), with corresponding eigenvectors \(\textbf{e}_1\), \(\textbf{e}_2\) and \(\textbf{e}_3\). Eigenvalues \(\lambda_i\) can be used to define the LSS environment of dark matter halos according to the number of positive eigenvalues, i.e.,
- void: \(0<\lambda_1<\lambda_2<\lambda_3\)
- wall: \(\lambda_1<0<\lambda_2<\lambda_3\)
- filament: \(\lambda_1<\lambda_2<0<\lambda_3\)
- cluster: \(\lambda_1<\lambda_2<\lambda_3<0\)
and the eigenvectors \(\textbf{e}_i\) stand for the three compressed directions of the smoothed density field. \(\textbf{e}_3\) vector indicates the least compressed direction, which is a robust and universal definition of the LSS. In this work, we will focus on the alignment of halo spin and shape with the \(\textbf{e}_3\) vector, i.e., \(\cos\theta_3=\textbf{a}\cdot\textbf{e}_3\), where \(\textbf{a}\) is the halo spin or major-axis vector. Fig. [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"} displays a \(100\times100\times20h^{-1}\mathrm{Mpc}\) slice from the twin simulations N1024 and N1024a. It presents a global view of the twin simulations. Over two hundred matching halo pairs are identified in this displayed slice, but here we only mark nine matching halos with hollow circles, with size proportional to their virial radii. We marked their main progenitors in upper panels if and only if they are still matched at \(z=1\). We can see that the LSS environment of these samples really changes tremendously after we replaced the random seeds for the small \(k\) of the initial over-density field.
# Results {#sec:3}
Our main results, mostly regarding the influence of LSS environment transformation on the properties of dark matter halos, will be shown in this section. Before that, we show some statistical results, including the halo mass function and correlations between halo properties (spin and major-axis) and the LSS environment, to test whether the twin simulations are statistically equal. Then we will focus on the mass dependence of the match ratio \(R\) in different LSS environment. Lastly, the influence of LSS on the total spin, major-axis, and shape parameters of the matching halo will be explored. Specifically, the shape parameters of dark matter halo, \(S\), \(Q\) and \(T\), are defined by: \[\label{equ:equ5} S=\frac{c}{a}, \quad Q=\frac{b}{a}, \quad T=\frac{a^2-b^2}{a^2-c^2},\] where \(a\) is the major-axis of modeled ellipsoid (by calculating the inertia tensor) of FoF halo, \(b\) is the intermediate-axis, and \(c\) is the minor-axis. The parameters \(S\) and \(Q\) give the ellipticity of dark matter halos, and \(T\) quantifies the triaxiality. Purely prolate halos have \(T=1\) while purely oblate halos have \(T=0\).
## Statistical tests {#sec:3.1}
Firstly, we show the halo mass function as an important statistical test of our twin simulations. Since the amplitude of the initial power spectrum of the twin simulations is exactly the same, one would expect that the difference in their halo mass function is very small. On the other hand, each N-body simulation has a finite box size, hence the cosmic variance between different simulations will produce different mass functions. Fig. [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} shows the halo mass function of N1024, N1024a, N1024b and N1024c. As we can see, all curves are nearly identical except at the high mass end, which can be explained by the effect of cosmic variance (well within the range of grey shadows). And because of the same reason, any similar figure regarding the statistical properties of the twin simulations gives nearly the same result. So we will not distinguish them hereafter. In Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} we show another statistical test regarding the mass dependence of correlations between halo spin or halo major-axis with the \(\textbf{e}_3\) vector for matching halo pairs in the twin simulations. The upper two panels are for halo spin, while the lower two panels are for halo major-axis. This figure shows that the correlations between the twin simulations agree with each other within \(1\sigma\) scatter, showing again that the twin simulations are statistically equivalent. In the left panels the halo spin and major axis are calculated using all particles within the virial radius of the halo. The trend of mass dependence of spin-\(\textbf{e}_3\) correlation agrees well with previous works. Note that here we only use halos with more than 100 particles (\(\sim10^{12}\mathrm{M}_{\odot}/h\)), roughly corresponding to the flip mass of the spin-LSS correlation found by previous works. In N-body simulations, the direction of halo spin is often slowly twisted along the halo radius, i.e., there is a misalignment between the inner spin and the total spin. Motivated by the fact that the baryons experience the same tidal forces as the dark matter, the spin of the galaxies embedded in halos usually follows the inner spin of the host halo. To ensure a relatively reliable result, we calculate the spin and major-axis vectors at \(0.3\) times the virial radius to describe the inner properties of matching halo hereafter. The right two panels of Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} show the mass dependence of inner properties (i.e. spin and major-axis at \(0.3\) times the virial radius). It is seen that, comparing to the alignment of total spin and \(\textbf{e}_3\) vector, the correlation of inner spin and \(\textbf{e}_3\) is nearly independent of the mass. On the other hand, as mentioned by, the halo spin-\(\textbf{e}_3\) correlation is stronger at high redshift. Considering that the inner spin is inherited from the total spin at high redshift, the alignment signal of inner spin and \(\textbf{e}_3\) vector is partly lost due to the nonlinear process in the center of high mass halos at later time.
## Match ratio and mergers {#sec:3.2}
In this section we focus on the matching halo pairs between the twin simulations. Firstly in Fig. [\[fig:fig4\]](#fig:fig4){reference-type="ref" reference="fig:fig4"}, we show the mass dependence of match ratio \(R\) in \(6\) twin simulation pairs (i.e., matched halos between N1024 and N1024a, N1024 and N1024b, N1024 and N1024c, N1024a and N1024b, N1024a and N1024c, N1024b and N1024c) at \(z=0\) and \(z=2\). The shaded region marks the maximum and minimum median match ratio, which roughly illustrate the systematic errors introduced by cosmic variance. For each twin simulation pair, the mass dependence of match ratio \(R\) is for the first simulation. If the mass bins are sorted by the second one, as we mentioned in Sec. [6](#sec:2){reference-type="ref" reference="sec:2"}, the conclusions are still the same. The lower limit of the y-axis is set to \(0.4\), which corresponds the arbitrary threshold imposed to select matching halo pairs. It can be seen that the match ratio of low mass halos is relatively higher, and comparing the same line of both panels, we also find that halos at high redshift usually have higher match ratios. For massive halos at lower redshift, the frequent mergers or mass accretion along the large-scale structure decrease the match ratio between the twin simulations. The yellow and red lines are for matching halos located in the different environments in both simulations. The median match ratio \(R\) of halos in filaments decreases much faster compared to those in clusters, indicating that halos in filaments are more influenced by mass accretion (along the filament) than in clusters. Here we shortly explain how the value of match ratio is reduced in minor merger events. Suppose that halo \(A\) merges with another halo \(b\) in N1024, while the matching halo \(A'\) and \(b'\) in N1024a are not involved in any merger event. At the end of the minor merger, the dark matter particles of \(A\) and \(b\) mix together. The information about the matching halo pair \(b\) and \(b'\) in the later snapshots is completely lost, and the match ratio of halo \(A\) and \(A'\) will be reduced because of the external particles from halo \(b\).
## The evolution of spin with LSS {#sec:3.3}
As we mentioned in Sec. [6.1](#sec:2.1){reference-type="ref" reference="sec:2.1"}, the \(\textbf{e}_{3}\) obtained from the hessian matrix indicates the direction of the LSS around a halo. For matching halo pairs in the twin simulation, we calculate the rotational angle of the corresponding eigenvectors between the twin simulation as, \[\label{equ:equ6} \cos\left(\beta\right)=\frac{\textbf{e}_3\cdot \textbf{e}_3'}{\vert\textbf{e}_3\vert\vert\textbf{e}_3'\vert}\;,\] here \(\textbf{e}_3\) and \(\textbf{e}_3'\) are the corresponding eigenvectors of matching halo pairs. In Fig. [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"} the red histogram shows the distribution of the rotation angle between the \(\textbf{e}_{3}\) vector in the twin simulation pair, N1024 and N1024a. It is found that there is more probability to have halo pairs for which the rotation angle of \(\textbf{e}_{3}\) is not significant, which suggests that although the power spectrum on large scales (\(>20\) Mpc/h) is totally different in the twin simulations, the nearby local environment smoothed on 2.5 Mpc/h is more or less maintained by the matched halo pairs. The black histogram shows a comparison with a larger smoothing length with \(R_s=5.0\) Mpc/h. In this case the rotation angle of large-scale environment becomes larger, but still maintaining some level of parallel signal. In this work we are mainly interested in how halo properties, spin or shape, correspond to the change of the large-scale environment, in particular, the direction of the LSS given by \(\textbf{e}_{3}\). To clearly select matching halo pairs with significant change of their large-scale environments, we select halo pairs from the twin simulations that have: \[\label{equ:equ7} \vert\cos\left(\beta\right)\vert\leq 0.707\quad\mathrm{or}\quad \beta\geq45^{\circ}\;.\] In this way we are looking at halo pairs with significant rotation of their \(\textbf{e}_{3}\) vectors. Of course, we can select halo pairs with larger rotation angles to increase the signal, but that will reduce the sample size. Based on Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}, halos in twin simulations show the same spin-LSS correlation. We can then ask what will happen if the original spin (of halo in, for example, N1024) does not change, can it still maintain a similar correlation with the new \(\textbf{e}_3\) from other twin simulation? If not, it means that the spin needs to be changed systematically to match the correlations (as shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}). This can prove that the correlation between spin and LSS is mainly affected by large-scale structures. In Fig. [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"} we plot the distribution of the spin-LSS correlation at three different redshifts for halo pairs with larger rotation angles given by Eq. [\[equ:equ7\]](#equ:equ7){reference-type="ref" reference="equ:equ7"}. The red shaded areas are the PDFs of halos whose spin and \(\textbf{e}_3\) are from the same simulation N1024a. The grey histograms are the cross correlations of the halo spin-\(\textbf{e}_3\) correlation, where halo spin is from N1024, but the \(\textbf{e}_3'\) vector is from N1024a. For matching halo pairs, the spin-LSS correlation in the same simulation, N1024a, is similar to what is shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"} and this is not surprising as the selected halos are a subsample of those shown in Fig. [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}. It is interesting to note that the cross correlations (black histograms) are reversed, showing a different trend with the red histograms. This result shows that the rotation of \(\textbf{e}_3\) vectors will lead to proper rotation of the halo spin so to maintain the original halo spin-\(\textbf{e}_3\) correlation seen in one simulation. It indicates that the halo spin is mainly affected by the orbital angular momentum of accreted mass from the LSS along the \(\textbf{e}_3\) vector.
## Matching halos and \(\textbf{e}_3\) rotation angle {#sec:3.4}
From Sec. [7.3](#sec:3.3){reference-type="ref" reference="sec:3.3"} we learn that the spin/major-axis will adjust its direction appropriately when the \(\textbf{e}_3\) vector changes across twin simulations. In this subsection we will explore this reaction in more detail. To maintain the spin-LSS correlation, one may intuitively expect that the rotation angle of halo spin/major-axis will depend strongly on the rotation angle \(\beta\) (see Sec. [7.3](#sec:3.3){reference-type="ref" reference="sec:3.3"}, Eq. [\[equ:equ6\]](#equ:equ6){reference-type="ref" reference="equ:equ6"}). More specifically, we expect that the rotation angle of spin is larger for halos with larger \(\beta\) so as to maintain the spin-\(\textbf{e}_{3}\) relation. However, if we check the correlation mentioned above, we will find that the rotation of spin/major-axis is nearly independent of angle \(\beta\). This is a bit surprising and it is related to the 3D space configuration of halo spin and LSS. Even in 2D space, it is easy to configure a case where the \(\textbf{e}_3\) vector changes significantly while the halo spin/major-axis does not need any adjustment to maintain the same spin-LSS correlation. The situation is more complicated in the 3D case. To better show how this could happen, we introduce the cross-product between \(\textbf{e}_3\) vector and spin/major-axis as the normal vector of halo-LSS system: \[\label{equ:equ8} \textbf{n}=\textbf{e}_3\times\textbf{a}\;,\] where \(\textbf{a}\) is the halo spin or major-axis vector. Under this definition, the rotation of the \(\textbf{e}_3\) vector will lead to the corresponding rotation of the **n** vector in most cases. In Fig. [\[fig:fig8\]](#fig:fig8){reference-type="ref" reference="fig:fig8"} we show the relationship between the rotation angle of \(\textbf{e}_{3}\) and the normal vector defined above. Here we can see that a strong positive proportional relationship between the rotation angle \(\beta\) and **n** vectors for all redshifts, i.e., the change of \(\textbf{e}_3\) vector will lead to corresponding rotation of the whole halo-LSS system, in which the correlation between spin/major-axis and \(\textbf{e}_3\) vector always holds. This gives us a picture of how halo spin/major-axis reacts according to the change of LSS. On the other hand, the proportional relationship in both two panels are nearly independent of the redshift (only the **n** vector from major-axis shown in the right panel has a weak evolution from \(z=2\) to the present-day). This implies an universality for the picture described above.
## Halo shape parameters with LSS {#sec:3.5}
Some previous works have revealed the correlations between halo shape and the LSS environment. In this work we also explore the impact of LSS environment on ellipticity and triaxiality of halos (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)). In Fig. [\[fig:fig9\]](#fig:fig9){reference-type="ref" reference="fig:fig9"} the mass dependence of three shape parameters \(S\), \(Q\) and \(T\) at \(z=0\) are shown in three panels, respectively. We note that the results are nearly independent of redshift, therefore, we only show the \(z=0\) results. The red and green lines are for halos in different twin simulations. The solid lines are for halos in specific LSS environment (\(0\) stands for cluster and \(1\) stands for filament). In addition, we also choose the matching halo pairs which go through an environment transformation from cluster to filament. These halos are denoted by the subscript \(0\rightarrow1\). Firstly it can be seen that, statistically, the shapes of halos in different LSS environment have some systematic differences, i.e., compared to cluster halos, halos in filament (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)) are much more spherical and more oblate, which agrees with previous results. On the other hand, lines in the same color are quite consistent for all shape parameters (reminding that there is a symmetry between twin simulations in terms of individual measurements), which proves that the current LSS environment really has dominant impact on the ellipticity and triaxiality of halos, regardless of their environments in the other symmetric simulation. Combined with results presented in the previous section on the spin-LSS correlation, our results suggest that current environment dominates the correlation between halo spin/major axis with the LSS.
# Conclusions and Discussion {#sec:4}
In this paper, we investigate the effect of LSS on halo pairs matched between twin N-body simulations. The only difference between the twin simulations is that in one run, the random seeds for the phase of small \(k\) where replaced when the initial conditions are generated. Once the sample of matching halos at different redshifts is established, one can use them to separate the entangled factors of large and small scales in the formation of dark matter halos. In this work we mainly focus on the correlations between the halo properties (spin, major-axis, shape parameters) and the \(\textbf{e}_3\) vector of the Hessian matrix, which is thought to be an universal definition of the LSS. Our main results are summarized as follows,
- A few statistical tests, including the halo mass function and the halo spin-\(\textbf{e}_3\) correlation, show that the twin simulations are statistically equal, ensuring that our technique for producing controlled twin simulation is robust and reliable.
- The ratio of matching halo pairs in the twin N-body simulations, based on a merit function, decreases with increasing halo mass. This is due to the later stronger evolution for massive halo by merger or mass accretion along the large-scale structure which is different in the twin simulations.
- The halo spin/major-axis will adjust its direction appropriately when LSS environment changes across twin simulations, i.e., the change of \(\textbf{e}_3\) vector will lead to corresponding rotation of the whole halo-LSS system, to ensure the correlation between spin/major-axis and \(\textbf{e}_3\) vector identical in the twin simulations.
- Statistically, halos in filament (\(M_h>10^{12}h^{-1}\mathrm{M}_{\odot}\)) are much more spherical and more oblate than cluster halos, and the LSS environment really has dominant impact on ellipticity and triaxiality of halos.
Since our technique of running twin simulations and establishing the matching halo pairs between them are verified to be robust and reliable, in the future, this technique can be implemented in some larger N-body simulations, or even hydrodynamical simulations, to identify matching galaxies embedded in the center of matching halo pairs. Moreover, multi-scale initial conditions for cosmological simulations, e.g. the zoom-in techniques, can also be employed to study how LSS environment influences galaxy formation and evolution in more details. On the other hand, one can degenerate the entangled factors on different scales in the halo and galaxy formation by choosing different \(L_0\) in Eq. [\[equ:equ2\]](#equ:equ2){reference-type="ref" reference="equ:equ2"}. We conclude that twin simulations with controlled initial condition on different scales could be useful to entangle the formation of halo/galaxy and the large-scale environment. We thank the anonymous reviewer for constructive and insightful suggestions that improved this paper. We acknowledge support from the joint Sino-German DFG research Project "The Cosmic Web and its impact on galaxy formation and alignment\" (NSFC No.11861131006, DFG-LI 2015/5-1 )， and the funds of cosmology simulation database (CSD) in the National Basic Science Data Center (NBSDC). HZC and XK acknowledge support from the the NSFC (No. 11825303, 11333008), the 973 program (No. 2015CB857003), and the China Manned Space Program (No. CMS-CSST-2021-A03, CMS-CSST-2021-B01). NIL acknowledges financial support of the Project IDEXLYON at the University of Lyon under the Investments for the Future Program (ANR-16-IDEX-0005). | {'timestamp': '2021-07-14T02:02:25', 'yymm': '2107', 'arxiv_id': '2107.05689', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05689'} |
null | null |
null | null |
# Introduction
Biological systems are inherently mechanochemical from the molecular to the tissue scale. Morphogenetic processes during embryo development involve coordinated cell and tissue shape changes driven by mechanical forces actively generated in the cell's cytoskeleton. These are in turn patterned by spatial gradients in concentration of molecules including myosin molecular motors. The chemical signals, such as molecular motors (or signaling factors upstream of motors that influence their contractile activity), are in principle also affected by mechanical forces, resulting in mechanochemical pattern formation during tissue development. Myosins constitute a superfamily of molecular motors that undergo cyclic enzymatic reactions via ATP hydrolysis and generate sliding motion of actin filaments in the cell cytoskeleton. They thus transduce chemical free energy into active mechanical forces, and form the fundamental force generation machinery of animal cells that is responsible for a variety of cell functions ranging from muscle contraction, cell motility and cell division to coordinated cell and tissue shape changes during morphogenesis. Myosin, like many other molecules involved in cellular force generation and transmission, is "mechanosensitive" in that its favored conformations can be modified by the exertion of mechanical force. Single molecule experiments have shown that the actin-myosin bond has catch bond behavior, that is the attachment time of the bound myosin increases on application of forces. A response of biomolecules to mechanical forces can change the generation and transmission of forces by the cells. This can lead to mechanochemical feedback loops that play a key role in many cellular, including morphogenetic, processes. The integration of such molecular signaling and subcellular processes with larger scale cell and tissue mechanics therefore remains an ongoing area of research in mechanobiology. Given the number and complexity of mechanosensitive biochemical signaling pathways and their inter--dependent nature, theoretical models built on general principles that predict pattern formation based on simple couplings of chemical kinetics with continuum mechanics are desirable and can open the path to future experimental tests. Inspired by the shape changes of epithelial tissue sheets as well as of sheets of reconstituted cytoskeletal gel by spatial gradients of myosin-generated contractile forces, we consider the coupling of elastic deformations and chemical gradients in slender bodies. Thin elastic plates and shells constitute a fundamental class of soft matter that exhibit sensitive response to stimuli changes because of the geometric nonlinearity of their mechanical properties. The resultant shapes of such material systems in response to a chemical signal or internal active forces (such as those generated by myosin molecular motors ) constitute a framework of ideas useful and important to material science and bioengineering. In the following, we describe a numerical procedure that captures shape changes coupled with chemical kinetics and illustrate the efficacy of the procedure by calculating the evolution of patterns in shape and chemical concentration in two distinct systems: one motivated by the coupling of morphogen gradients to tissue curvature, and the other based on the mechanical force-dependent binding of myosin motors to actin fibers. Our numerical procedure enables calculation of large elastic deformations in conjunction with chemical concentration evolution. The latter typically entails the solution of chemical rate kinetics in the form of differential equations on a curved surface. Our method can also describe separate time scales in the elastic relaxation and chemical kinetics. We illustrate this numerical scheme by applying it two different biologically relevant models.
# Models
We aim to describe the active deformations of elastic thin sheets that constitute a commonly studied class of active and living matter. Examples range from freely suspended epithelial cell monolayers to reconstituted actomyosin gels. A model sheet is generically considered to be a continuum elastic medium with isotropic mechanical properties throughout its thickness, which is assumed to be much smaller compared to the system size in the other two dimensions. This assumption allows us to neglect any dependence of the energy on strains transverse to the sheet. The problem can then be reduced to two dimensions by representing the sheet as a 2D surface embedded in the 3D space. The elastic free energy of a deformed shell is given by contributions from the stretching energy \(\mathcal{U}_s\) that is proportional to the shell thickness, \(h\), and arises due to in-plane extension and compression in the shell, and the bending energy contribution \(\mathcal{U}_b\), which accounts for curvature and scales as \(h^3\). Specifically, the elastic energy of a deformed elastic shell is given by | {'timestamp': '2021-07-14T02:04:01', 'yymm': '2107', 'arxiv_id': '2107.05726', 'language': 'en', 'url': 'https://arxiv.org/abs/2107.05726'} |
Subsets and Splits