text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
# Introduction
Let \(X\) be a real Banach space and \(Y\subset X\) be a finite-dimensional subspace. Let \(\mathcal{P}(X,Y)\) denote the set of all linear and continuous projections from \(X\) onto \(Y\), recalling that an operator \(P \colon X \rightarrow Y\) is called a *projection* onto \(Y\) if \(P|_Y={\rm Id}_Y.\) We define the *relative projection constant* of \(Y\) by \[\lambda(Y,X) :=\inf\lbrace\|P\|:\;P\in\mathcal{P}(X,Y)\rbrace\] and the *absolute projection constant* of \(Y\) by \[\label{DefMAPC} \lambda(Y) :=\sup\lbrace\lambda(Y,X):Y\subset X\rbrace.\] The literature also deals with the *maximal absolute projection constant*, which is defined by \[\lambda(m) :=\sup \lbrace\lambda(Y):\; \dim(Y)=m \rbrace.\] By the Kadec--Snobar theorem (see ), we have \(\lambda(m)\leq \sqrt{m}\). Moreover, it has been shown in that this estimate is asymptotically the best possible. However, the determination of the constant \(\lambda(m)\) seems to be difficult: apart from \(\lambda(1)=1\), the only known value of \(\lambda(m)\) is \(\lambda(2)=4/3\)---this is *Grünbaum conjecture*, formulated in and proved in. Numerical computations presented in indicate that \(\lambda(3)\) should equal \((1+\sqrt{5})/2\)---this was stated, with an erroneous proof, in. Other numerical experiments conducted by B. L. Chalmers (and unfortunately unpublished) suggest that \(\lambda(5)\approx 2.06919\). In this article, we show that \[\lambda(5)\geq 5(11+6\sqrt{5})/59\approx 2.06919.\] Viewed in isolation, this could seem anecdotal. However, several sources of evidence hint that this is the actual value of \(\lambda(5)\). This comes as a surprise, because it was growingly believed that obtaining exact formulas for \(\lambda(m)\) was an unreasonable quest. Now there is hope that this quest could be realized after all. To establish the announced lower bound, we make a detour via maximal relative projection constants. Recent results concerning maximal relative and absolute projection constants can be found in. Here, we only give the definition of the *maximal relative projection constant* for \(n \ge m\) as \[\lambda(m,n):=\sup\lbrace \lambda(Y, l_\infty^{(n)}):\; \dim(Y)=m \textrm{ and } Y\subset l_\infty^{(n)}\rbrace.\] This is motivated by the fact that, in the expression [\[DefMAPC\]](#DefMAPC){reference-type="eqref" reference="DefMAPC"} of \(\lambda(m)\), it suffices to take the supremum over finite-dimensional \(l_\infty\) superspaces (see e.g. ), so that the nondecreasing sequence \(( \lambda(m, n) )_{n \ge m}\) converges to \(\lambda(m)\). In reality, there even is an \(N\in\mathbb{N}\) such that \(\lambda(m,n) = \lambda(m)\) for all \(n\geq N\) (see ). Our estimation of \(\lambda(m,n)\) will rely on the following result proved in.
Although this theorem provides an essential tool for estimating the maximal relative projection constants, computing their exact values remains a challenging problem, carried out in just a few cases (see e.g. ). One particular situation where an explicit formula is available involves equiangular tight frames. Let us recall that a system of unit (i.e., \(l_2\)-normalized) vectors \((v_1,\dots, v_n)\) in \(\mathbb{R}^m\) is called *equiangular* if there is a constant \(c\geq 0\) such that \[|\langle v_i,v_j \rangle|=c \qquad \textrm{ for all } i,j\in \{1,\dots, n\},\;i\neq j.\] It is called a *tight frame* if \[VV^\top=\frac{n}{m}{\rm I}_m,\] where \(V\) is the matrix with columns \(v_1,\dots, v_n\). The system \((v_1,\dots, v_n)\) of unit vectors is called an *equiangular tight frame* if it is both equiangular and a tight frame. For an equiangular tight frame of \(n\) unit vectors in \(\mathbb{R}^m\), it is well known (see e.g. ) that \[|\langle v_i,v_j \rangle|=\sqrt{\frac{n-m}{m(n-1)}} \qquad \textrm{ for all } i,j\in \{1,\dots, n\},\;i\neq j.\] The above-mentioned explicit formula is presented as part of the result below. Built from Theorems 1 and 2 of, it appeared in a slightly different form as Theorem 5 in. A new self-contained proof is included later as an appendix.
In the rest of this paper, we present new explicit lower bounds for \(\lambda(m,n)\) under the condition that certain mutually unbiased equiangular tight frames for \(\mathbb{R}^m\) exist (see Theorem [\[amazing\]](#amazing){reference-type="ref" reference="amazing"}). We then provide a construction of an infinite family of such mutually unbiased equiangular tight frames (see Theorem [\[muetfFamily\]](#muetfFamily){reference-type="ref" reference="muetfFamily"}). Finally, combining these two ingredients, we highlight the resulting estimation of \(\lambda(5,16)\) to arrive at the promised lower bound for \(\lambda(5)\), conjectured to be its true value.
# The Lower Bound
Before stating the main result, we start with an observation about mutually unbiased equiangular tight frames, formally defined below.
This definition generalizes a concept introduced in so as to permit the case \(k \neq l\). We point out that the scalar \(c\) is uniquely determined, as also noted in .
Let \((v_1,\dots,v_k)\) and \((w_1,\dots,w_l)\) be mutually unbiased equiangular tight frames for \(\mathbb{R}^m\) and let \(V \in \mathbb{R}^{m \times k}\) be the matrix with columns \(v_1,\dots,v_k\). For any \(j\in\{1,\dots,l\}\), because the two frames are mutually unbiased, we have \[\|V^\top w_j\|_2^2 =\sum_{i=1}^k |\langle v_i,w_j \rangle|^2 =\sum_{i=1}^k c^2 =kc^2.\] Since \((v_1,\dots,v_k)\) is a tight frame for \(\mathbb{R}^m\), we also have \(VV^\top = (k/m){\rm I}_m\), and so \[\|V^\top w_j\|_2^2 =\langle V^\top w_j, V^{\top}w_j \rangle =\langle w_j, VV^{\top}w_j \rangle =\Big\langle w_j, \dfrac{k}{m}w_j \Big\rangle =\dfrac{k}{m}\| w_j\|_2^2 =\dfrac{k}{m}.\] It follows that \(kc^2=k/m\), and hence \(c=1/{\sqrt{m}}\), as claimed.0◻.1in We now present the main theorem of this section, whose statement involves the quantity \(\delta_{m,n}\) introduced in Theorem [\[ETFs\]](#ETFs){reference-type="ref" reference="ETFs"}.
Let \(V \in \mathbb{R}^{m \times k}\) be the matrix with columns \(v_1,\dots,v_k\) and \(W \in \mathbb{R}^{m \times l}\) the matrix with columns \(w_1,\dots,w_l\). For any \(\theta\in [0,\pi/2]\), let us consider the vector \(t_\theta\in \mathbb{R}^{k+l}\) and the matrix \(U_\theta \in \mathbb{R}^{m \times (k+l)}\) defined, in block notation, by \[\label{DeftU} t_\theta:= \begin{bmatrix} \cos\theta \dfrac{1}{\sqrt{k}} \mathbb{1}_k \\ \hline \sin \theta \dfrac{1}{\sqrt{l}}\mathbb{1}_l \end{bmatrix} \qquad \mbox{and} \qquad U_\theta:= \begin{bmatrix} \; \cos\theta \sqrt{\dfrac{m}{k}} V & \vline & \sin \theta \sqrt{\dfrac{m}{l}} W \; \end{bmatrix},\] where \(\mathbb{1}_n\) denotes the \(n\)-dimensional vector with all entries equal to \(1\). We observe that \(\|t_\theta\|_2=1\), that \[U_\theta {U_\theta}^\top=\cos^2\theta \frac{m}{k}VV^\top + \sin^2 \theta \frac{m}{k} W W^\top=\cos^2\theta \, {\rm I}_m+\sin^2\theta \, {\rm I}_m = {\rm I}_m,\] and that \[\label{UTU} {U_\theta}^\top U_\theta = \begin{bmatrix} \cos^2 \theta \dfrac{m}{k} V^\top V & \vline & \cos \theta \sin \theta \dfrac{m}{\sqrt{k l}} V^\top W\\ \hline \cos \theta \sin \theta \dfrac{m}{\sqrt{k l}} W^\top V & \vline & \sin ^2 \theta \dfrac{m}{l} W^\top W \end{bmatrix}.\] Therefore, according to the expression of \(\lambda(m,n)\) from Theorem [\[lammbda\]](#lammbda){reference-type="ref" reference="lammbda"}, we can make use of the tight frame and unbiasedness properties of \(U\) and \(V\) to obtain, with the shorthand notation \(\phi_{m,n} := \sqrt{(n-m)/(m(n-1))}\), \[\begin{aligned}
\lambda(m,k+l)& \geq \sum_{i,j=1}^{k+l} (t_\theta)_i (t_\theta)_j | {U_\theta}^\top U_\theta|_{i,j} \\ & = \cos^2 \theta \frac{1}{k} \times \cos^2\theta \frac{m}{k} \times k + \cos^2 \theta \frac{1}{k} \times \cos^2\theta \frac{m}{k} \phi_{m,k} \times k(k-1)\\ & + \sin^2 \theta \frac{1}{l} \times \sin^2\theta \frac{m}{l} \times l + \sin^2 \theta \frac{1}{l} \times \sin^2\theta \frac{m}{l} \phi_{m,l} \times l(l-1)\\ & + 2 \times \cos \theta \sin \theta \frac{1}{\sqrt{k l}} \times \cos \theta \sin \theta \frac{m}{\sqrt{k l}} \frac{1}{\sqrt{m}} \times k l \\ & = \cos^4 \theta \bigg( \frac{m}{k} + \frac{m}{k} (k-1) \phi_{m,k} \bigg)+ \sin^4 \theta \bigg( \frac{m}{l} + \frac{m}{l} (l-1) \phi_{m,l} \bigg)\\ & + 2 \cos^2 \theta \sin^2 \theta \sqrt{m}\\ & = \bigg(\frac{1+\cos(2\theta)}{2}\bigg)^2 \delta_{m,k} + \bigg(\frac{1-\cos(2\theta)}{2}\bigg)^2 \delta_{m,l} + \big( \sin(2\theta) \big)^2 \frac{\sqrt{m}}{2}. \end{aligned}\] Since this is valid for any \(\theta \in [0,\pi/2]\), after setting \(x:=\cos(2\theta)\), we arrive at \[\begin{aligned}
\nonumber \lambda(m,k+l)&\geq \max_{x\in[-1,1]}\left( \frac{\delta_{m,k}(1+2x+x^2)}{4}+ \frac{\delta_{m,l}(1-2x+x^2)}{4}+\frac{\sqrt{m}}{2}(1-x^2) \right)\\ &=\frac{1}{4}\max_{x\in [-1,1]}\left( ax^2+2 b x + c \right), \end{aligned}\] where \(a := \delta_{m,k}+\delta_{m,l}-2\sqrt{m}\), \(b := \delta_{m,k}-\delta_{m,l}\), and \(c:= \delta_{m,k}+\delta_{m,l}+2\sqrt{m}\). Taking momentarily for granted that \(a < 0\) and that \(x_* :=-b / a \in [-1,1]\), we deduce that \[\begin{aligned}
\lambda(m,k+l) & \ge \frac{1}{4} \left( ax_*^2+2 b x_* + c \right) = \frac{1}{4}\left(-\frac{b^2}{a} + c \right) = \frac{1}{4} \frac{b^2-a c}{-a}\\ & = \frac{1}{4} \frac{(\delta_{m,k}-\delta_{m,l})^2 + (2\sqrt{m}-\delta_{m,k}-\delta_{m,l}) (2\sqrt{m} + \delta_{m,k}+\delta_{m,l}) }{2\sqrt{m}-\delta_{m,k}-\delta_{m,l}}\\ & = \frac{1}{4} \frac{4m-4\delta_{m,k}\delta_{m,l}}{2\sqrt{m}-\delta_{m,k}-\delta_{m,l}}, \end{aligned}\] which is the announced lower bound. It now remains to notice that \(a < 0\) and that \(-b / a \in [-1,1]\), but both follow from the general observation that \(\delta_{m,n} < \sqrt{m}\) for \(n \ge m > 1\), see Remark [\[RemGam\]](#RemGam){reference-type="ref" reference="RemGam"}. 0◻Before uncovering a family of mutually unbiased equiangular tight frames in the next section, we emphasize here two noteworthy properties relating the vector \(t_\theta\) and the matrix \(U_\theta\) that appeared in the above proof.
When establishing both [\[SP1\]](#SP1){reference-type="eqref" reference="SP1"} and [\[SP2\]](#SP2){reference-type="eqref" reference="SP2"}, it will be useful to keep in mind that \(\delta_{m,n}\) is tied to \(\phi_{m,n} = \sqrt{(n-m)/(m(n-1))}\) via \[\delta_{m,n} = \frac{m}{n} \bigg( 1 + (n-1) \phi_{m,n} \bigg) = \frac{m}{n} \bigg( 1 + \frac{n-m}{m} \frac{1}{\phi_{m,n}} \bigg).\] Starting with the justification of [\[SP1\]](#SP1){reference-type="eqref" reference="SP1"}, we notice that, since the matrix \(V^\top V\) has diagonal entries equal to \(1\) and off-diagonal entries equal to \(\phi_{m,k}\) in absolute value, we have \[|V^\top V| = (1-\phi_{m,k}) {\rm I}_k + \phi_{m,k} \mathbb{1}_{k,k},\] where \(\mathbb{1}_{n,n'}\) denotes the \(n \times n'\) matrix with all entries equal to \(1\). It follows that \[|V^\top V| \mathbb{1}_{k} = (1-\phi_{m,k}) \mathbb{1}_{k} + k \phi_{m,k} \mathbb{1}_{k} = (1+(k-1) \phi_{m,k}) \mathbb{1}_{k} = \frac{k}{m} \delta_{m,k} \mathbb{1}_{k}.\] Likewise, we can obtain \[|W^\top W| \mathbb{1}_{l} = \frac{l}{m} \delta_{m,l} \mathbb{1}_{l}.\] Moreover, since the matrices \(V^\top W\) and \(W^\top V\) have entries all equal to \(1/\sqrt{m}\) in absolute value, we have \(| V^\top W | = (1 / \sqrt{m}) \mathbb{1}_{k,l}\) and \(| W^\top V | = (1 / \sqrt{m}) \mathbb{1}_{l,k}\), so that \[| V^\top W | \mathbb{1}_{l} = \frac{l}{\sqrt{m}} \mathbb{1}_{k} \qquad \mbox{and} \qquad | W^\top V | \mathbb{1}_{k} = \frac{k}{\sqrt{m}} \mathbb{1}_{l}.\] Therefore, according to the block-forms of \(t_\theta\) and \(U_\theta^\top U_\theta\) (see [\[DeftU\]](#DeftU){reference-type="eqref" reference="DeftU"} and [\[UTU\]](#UTU){reference-type="eqref" reference="UTU"}), we observe that \[\begin{aligned}
\nonumber |U_\theta^\top U_\theta | \, t_\theta & = \begin{bmatrix} \cos^2 \theta \dfrac{m}{k} \cos \theta \dfrac{1}{\sqrt{k}} \dfrac{k}{m} \delta_{m,k} \mathbb{1}_{k} + \cos \theta \sin \theta \dfrac{m}{\sqrt{kl}} \sin \theta \dfrac{1}{\sqrt{l}} \dfrac{l}{\sqrt{m}} \mathbb{1}_{k} \\ \hline \cos \theta \sin \theta \dfrac{m}{\sqrt{kl}} \cos \theta \dfrac{1}{\sqrt{k}} \dfrac{k}{\sqrt{m}} \mathbb{1}_{l} + \sin^2 \theta \dfrac{m}{l} \sin \theta \dfrac{1}{\sqrt{l}} \dfrac{l}{m} \delta_{m,l} \mathbb{1}_{l} \end{bmatrix} \\ \label{AlmostThereSP1} & = \begin{bmatrix} \cos \theta \dfrac{1}{\sqrt{k}} \left( \cos^2 \theta \delta_{m,k} + \sin^2 \theta \sqrt{m} \right) \mathbb{1}_{k}\\ \hline \sin \theta \dfrac{1}{\sqrt{l}} \left( \cos^2 \theta \sqrt{m} + \sin^2 \theta \delta_{m,l} \right) \mathbb{1}_{l} \end{bmatrix}. \end{aligned}\] Next, in view of \[\begin{aligned}
\cos^2 \theta & = \frac{1+\cos(2 \theta)}{2} = \frac{\sqrt{m}-\delta_{m,l} }{2 \sqrt{m}-\delta_{m,k}-\delta_{m,l}},\\ \sin^2 \theta & = \frac{1-\cos(2 \theta)}{2} = \frac{\sqrt{m}-\delta_{m,k} }{2 \sqrt{m}-\delta_{m,k}-\delta_{m,l}}, \end{aligned}\] we easily derive that \[\label{TrigEq} \cos^2 \theta \delta_{m,k} + \sin^2 \theta \sqrt{m} = \cos^2 \theta \sqrt{m} + \sin^2 \theta \delta_{m,l} = \gamma_{m,k,l}.\] When substituting the latter into [\[AlmostThereSP1\]](#AlmostThereSP1){reference-type="eqref" reference="AlmostThereSP1"}, the identity [\[SP1\]](#SP1){reference-type="eqref" reference="SP1"} immediately follows. Turning now to the justification of [\[SP2\]](#SP2){reference-type="eqref" reference="SP2"}, recalling that the matrix \(V^\top V\) has diagonal entries equal to \(1\) and off-diagonal entries equal to \(\phi_{m,k}\) in absolute value, the diagonal entries of the matrix \({\rm sgn}(V^\top V)\) are equal to \(1\) and its off-diagonal entries are equal to those of \(V^\top V\) divided by \(\phi_{m,k}\). In short, we see that \({\rm sgn}(V^\top V) = (1-1/\phi_{m,k}){\rm I}_k + (1/\phi_{m,k}) V^\top V\) holds, and a similar identity holds for \({\rm sgn}(W^\top W)\). Moreover, we also have \({\rm sgn}(V^\top W) = \sqrt{m} \, V^\top W\) and \({\rm sgn}(W^\top V) = \sqrt{m} \, W^\top V\), as a consequence of all the entries of \(W^\top V\) and \(W^\top V\) being equal to \(1/\sqrt{m}\) in absolute value. All in all, according to the block-form [\[UTU\]](#UTU){reference-type="eqref" reference="UTU"} of \(U_\theta^\top U_\theta\), we obtain \[{\small {\rm sgn}(U_\theta^\top U_\theta) = \begin{bmatrix} \left( 1-\dfrac{1}{\phi_{m,k}} \right) {\rm I}_k + \dfrac{1}{\phi_{m,k}} V^\top V & \vline & \sqrt{m} \, V^\top W \\ \hline \sqrt{m} \, W^\top V & \vline & \left( 1-\dfrac{1}{\phi_{m,l}} \right) {\rm I}_l + \dfrac{1}{\phi_{m,l}} W^\top W \end{bmatrix}. }\] In turn, using the block-form of \(T_\theta = {\rm diag}[t_\theta]\), we derive that \(T_\theta {\rm sgn}(U_\theta^\top U_\theta) T_\theta\) takes the form \[{\footnotesize \begin{bmatrix} \cos^2 \theta \dfrac{1}{k} \left( \left( 1-\dfrac{1}{\phi_{m,k}} \right) {\rm I}_k + \dfrac{1}{\phi_{m,k}} V^\top V \right) & \vline & \cos \theta \sin \theta \dfrac{1}{\sqrt{kl}} \sqrt{m} \, V^\top W \\ \hline \cos \theta \sin \theta \dfrac{1}{\sqrt{kl}} \sqrt{m} \, W^\top V & \vline & \sin^2 \theta \dfrac{1}{l} \left( \left( 1-\dfrac{1}{\phi_{m,l}} \right) {\rm I}_l + \dfrac{1}{\phi_{m,l}} W^\top W \right) \end{bmatrix}. }\] Multiplying on the right by the transpose of \(U_\theta= {\small \begin{bmatrix} \; \cos\theta \sqrt{\dfrac{m}{k}} V & \vline & \sin \theta \sqrt{\dfrac{m}{l}} W \; \end{bmatrix}}\) and making use of the facts that \(V V^\top = (k/m) {\rm I}_m\) and \(W W^\top = (l/m) {\rm I}_m\), the matrix \(T_\theta {\rm sgn}(U_\theta^\top U_\theta) T_\theta \, U_\theta^\top\) becomes \[\begin{aligned}
& {\footnotesize \begin{bmatrix} \cos^2 \theta \dfrac{1}{k} \cos \theta \sqrt{\dfrac{m}{k}} \left( \left( 1-\dfrac{1}{\phi_{m,k}} \right) + \dfrac{k}{m} \dfrac{1}{\phi_{m,k}} \right) V^\top + \cos \theta \sin \theta \dfrac{1}{\sqrt{k l}} \sqrt{m} \sin \theta \sqrt{\dfrac{m}{l}} \dfrac{l}{m} V^\top\\ \hline \cos \theta \sin \theta \dfrac{1}{\sqrt{kl}} \sqrt{m} \cos\theta \sqrt{\dfrac{m}{k}} \dfrac{k}{m} W^\top + \sin^2 \theta \dfrac{1}{l} \sin \theta \sqrt{\dfrac{m}{l}} \left( \left( 1-\dfrac{1}{\phi_{m,l}} \right) + \dfrac{l}{m} \dfrac{1}{\phi_{m,l}} \right) W^\top \end{bmatrix} }\\ & = \begin{bmatrix} \cos \theta \sqrt{\dfrac{m}{k}} \left( \dfrac{\cos^2 \theta}{k} \left( 1 + \dfrac{k-m}{m} \dfrac{1}{\phi_{m,k}} \right) + \sin^2 \theta \dfrac{1}{\sqrt{m}} \right) V^\top \\ \hline \sin \theta \sqrt{\dfrac{m}{l}} \left( \cos^2 \theta \dfrac{1}{\sqrt{m}} + \dfrac{\sin^2 \theta}{l} \left( 1 + \dfrac{l-m}{m} \dfrac{1}{\phi_{m,l}}\right) \right) W^\top \end{bmatrix}\\ & = \begin{bmatrix} \cos \theta \sqrt{\dfrac{m}{k}} \left( \dfrac{\cos^2 \theta}{m} \delta_{m,k} + \dfrac{\sin^2 \theta}{\sqrt{m}} \right) V^\top \\ \hline \sin \theta \sqrt{\dfrac{m}{l}} \left( \dfrac{\cos^2 \theta}{\sqrt{m}} + \dfrac{\sin^2 \theta}{m} \delta_{m,l} \right) W^\top \end{bmatrix}. \end{aligned}\] Similarly to [\[SP1\]](#SP1){reference-type="eqref" reference="SP1"}, the identity [\[SP2\]](#SP2){reference-type="eqref" reference="SP2"} now simply follows by exploiting [\[TrigEq\]](#TrigEq){reference-type="eqref" reference="TrigEq"} again. 0◻
# Construction of Mutually Unbiased Equiangular Tight Frames
To apply the result of Theorem [\[amazing\]](#amazing){reference-type="ref" reference="amazing"} in practical situations, we evidently need to uncover specific integers \(k\), \(l\), and \(m\) allowing mutually unbiased equiangular tight frames to exist. As a simple example, one can take \(k=l=m\) and consider \((v_1,\ldots,v_k)\) to be the canonical basis for \(\mathbb{R}^m\) and \((w_1,\ldots,w_l)\) to be the columns of an \(m \times m\) Hadamard matrix---recall that \(m \times m\) Hadamard matrices are conjectured to exist when and only when \(m\) is a multiple of \(4\) (the 'only when' part being acquired, of course). This would yield the lower bound \(\lambda(m) \ge (1+\sqrt{m})/2\), \(m \in 4 \mathbb{N}\), which is inferior to the lower bounds reported in for \(m=4\) and \(m=8\). As a slightly more elaborate example, one can take \(k=m\) and \((v_1,\ldots,v_k)\) to be the canonical basis of \(\mathbb{R}^m\), together with \(l > m\) and \((w_1,\ldots,w_l)\) to be a real equiangular tight frame for \(\mathbb{R}^m\) that is *flat*, in the sense that every entry of each vector \(w_j\) is either \(1/\sqrt{m}\) or \(-1/\sqrt{m}\). Real flat equiangular tight frames are equivalent to binary codes achieving equality in the Grey--Rankin bound and infinite families are known (see ). This would yield the lower bound \(\lambda(m,m+l) \geq (m-\gamma_{m,l})/(2\sqrt{m}-1-\gamma_{m,l})\). With \(m=6\) and \(l=16\), this provides the lower bound \(\lambda(6)\gtrsim 2.2741\), which is superior to the lower bounds reported in but inferior to the numerical evaluation \(\lambda(6) \approx 2.2857\) performed by B. L. Chalmers and corroborated by our own computations. In order to apply Theorem [\[amazing\]](#amazing){reference-type="ref" reference="amazing"} more effectively, we need further examples of mutually unbiased equiangular tight frames. To this end, we now relate such frames to a type of generalized Hadamard matrices.
Firstly, let us assume that there are mutually unbiased equiangular tight frames \((v_1,\dots,v_k)\) and \((w_1,\dots,w_l)\) for \(\mathbb{R}^m\). With \(V \in \mathbb{R}^{m \times k}\) and \(W \in \mathbb{R}^{m \times l}\) denoting the matrices with columns \(v_1,\dots,v_k\) and \(w_1,\dots,w_l\), respectively, we set \[X=\sqrt{m}\,V^\top W\in\mathbb{R}^{k\times l}.\] By Lemma [\[beta\]](#beta){reference-type="ref" reference="beta"}, we have \(|V^\top W|_{i,j} = |\langle v_i, w_j \rangle | = 1/\sqrt{m}\) for all \(i\in\{1,\dots,k\}\) and \(j\in \{1,\dots,l\}\), so Property [\[P1\]](#P1){reference-type="eqref" reference="P1"} is immediate. In view of \(VV^\top = (k/m)\,{\rm I}_m\) and of \(WW^\top=(l/m)\,{\rm I}_m\), it is also straightforward to see that \[\label{XX'} XX^\top = l \, V^\top V \qquad \mbox{and} \qquad X^\top X =k \, W^\top W.\] From here, using the fact that \(VV^\top = (k/m)\,{\rm I}_m\) one more time, we obtain that \(XX^\top X = (l \, V^\top V) (\sqrt{m}\, V^\top W)= (kl/m)\sqrt{m}\, V^\top W\), i.e., \(XX^\top X = a X\) with \(a = kl/m\), so Property [\[P2\]](#P2){reference-type="eqref" reference="P2"} is satisfied. Properties [\[P3\]](#P3){reference-type="eqref" reference="P3"} and [\[P4\]](#P4){reference-type="eqref" reference="P4"}, too, are consequences of [\[XX\'\]](#XX'){reference-type="eqref" reference="XX'"}, since e.g. the off-diagonal entries of \(X X^\top\) are constant in absolute value because those of \(V^\top V\) are. Finally, Property [\[P5\]](#P5){reference-type="eqref" reference="P5"} is also implied by [\[XX\'\]](#XX'){reference-type="eqref" reference="XX'"} via \({\rm rank}(X) = {\rm rank}(X X^\top) = {\rm rank}(V^\top V) = {\rm rank}(V V^\top) = {\rm rank}({\rm I}_m)~=~m\). Conversely, let us assume that Properties [\[P1\]](#P1){reference-type="eqref" reference="P1"}--[\[P2\]](#P2){reference-type="eqref" reference="P2"} are fulfilled by some matrix \(X \in \mathbb{R}^{k \times l}\). Consider the singular value decomposition of this matrix written as \(X = P \Sigma Q^\top\), where the diagonal matrix \(\Sigma \in \mathbb{R}^{m \times m}\) has positive entries (by [\[P5\]](#P5){reference-type="eqref" reference="P5"}) and where the matrices \(P \in \mathbb{R}^{k \times m}\) and \(Q \in \mathbb{R}^{l \times m}\) have orthonormal columns, i.e., \(P^\top P = {\rm I}_m\) and \(Q^\top Q = {\rm I}_m\). Property [\[P2\]](#P2){reference-type="eqref" reference="P2"} easily yields \(\Sigma^3 = a \, \Sigma\) and hence \(\Sigma = \sqrt{a} \, {\rm I}_m\). Then, looking at the squared Frobenius norm of \(X = \sqrt{a} \, P Q^\top\), we derive from [\[P1\]](#P1){reference-type="eqref" reference="P1"} that \(k l = a m\), i.e., that \(a = kl/m\). We now set \[V = \sqrt{\frac{k}{m}}P^\top \in \mathbb{R}^{m \times k} \qquad \mbox{and} \qquad W = \sqrt{\frac{l}{m}} Q^\top \in \mathbb{R}^{m \times l}\] and we claim that the columns \(v_1,\ldots,v_k\) of \(V\) and \(w_1,\ldots,w_l\) of \(W\) are mutually unbiased equiangular tight frames for \(\mathbb{R}^m\). Indeed, using \(V^\top V = (k/m) P P^\top\) and \(X X^\top = a P P^\top\), we see that \(V^\top V = (1/l) X X^\top\), so that the equiangularity of the system \((v_1,\ldots,v_k)\) is clear from [\[P3\]](#P3){reference-type="eqref" reference="P3"}. Note that each \(v_i\) is a unit vector, since \(\|v_i\|_2^2 = (V^\top V)_{i,i} = (1/l) (XX^\top)_{i,i} = (1/l) \sum_{j=1}^l X_{i,j}^2 = 1\) by [\[P1\]](#P1){reference-type="eqref" reference="P1"}. The fact that these vectors form a tight frame is seen from \(V V^\top = (k/m) \, P^\top P = (k/m) \, {\rm I}_m\). Similar arguments (using [\[P4\]](#P4){reference-type="eqref" reference="P4"}) would reveal that the system \((w_1,\ldots,w_l)\) is also an equiangular tight frame. At last, to see that these systems are mutually unbiased, it suffices to notice that \(V^\top W = (\sqrt{kl}/m) \, P Q^\top = (1/\sqrt{m}) \, X\) and to invoke [\[P1\]](#P1){reference-type="eqref" reference="P1"} once again. It finally remains to establish that the three quantities in [\[integrality\]](#integrality){reference-type="eqref" reference="integrality"} are integers. For the first one, we have seen (in the proofs of both implications) that \(a = kl/m\) and [\[P1\]](#P1){reference-type="eqref" reference="P1"}-[\[P2\]](#P2){reference-type="eqref" reference="P2"} show that \(a\) is an integer: any entry of \(XX^\top X=a X\) is on the one hand an integer and on the other hand equal to \(\pm a\). For the third one, say, looking e.g. at [\[XX\'\]](#XX'){reference-type="eqref" reference="XX'"}, any off-diagonal entry of \(X X^\top = l \, V^\top V\) is on the one hand an integer and on the other hand equal to \(l\) times the common absolute inner product in a \(k\)-vector equiangular tight frame for \(\mathbb{R}^m\), i.e., to \(l \sqrt{(k-m)/(m(k-1))}\). 0◻.1in Although conditions [\[P1\]](#P1){reference-type="eqref" reference="P1"}--[\[P5\]](#P5){reference-type="eqref" reference="P5"} are restrictive, there are matrices \(X\) satisfying them with \(m < k < l\). For instance, the \(6\times 10\) matrix \[\setlength{\arraycolsep}{1pt} X=\begin{small}\left[\begin{array}{rrrrrrrrrr} 1& 1& 1& 1& 1& 1& 1& 1& 1& 1\\ 1& 1&-1& 1&-1&-1& 1&-1&-1&-1\\ 1&-1& 1&-1& 1&-1&-1& 1&-1&-1\\-1& 1& 1&-1&-1& 1&-1&-1& 1&-1\\-1&-1&-1& 1& 1& 1&-1&-1&-1& 1\\-1&-1&-1&-1&-1&-1& 1& 1& 1& 1 \end{array}\right]\end{small}\] is one such matrix[^1]: it has \(\pm 1\) entries, the identity \(X X^\top X = a X\) is easily verified (at least computationally), and it was already observed in that both its rows and its columns form equiangular tight frames for their \(5\)-dimensional spans. Therefore, since \(X\) fulfills the conditions of Proposition [\[muetfCharacterization\]](#muetfCharacterization){reference-type="ref" reference="muetfCharacterization"} with \(m=5\), \(k=6\), and \(l=10\), we are guaranteed the existence of mutually unbiased equiangular tight frames \((v_1,\dots,v_6)\) and \((w_1,\dots,w_{10})\) for \(\mathbb{R}^5\). Remarkably, this example is but the first member of the infinite family presented below.
For any such \(s\), \(k\), \(l\) and \(m\), the requisite matrix \(X\) of Proposition [\[muetfCharacterization\]](#muetfCharacterization){reference-type="ref" reference="muetfCharacterization"} is produced in the recent paper , albeit nonobviously so. In brief, let \(Q\) and \(B\) be the canonical hyperbolic-quadratic and symplectic forms on the binary vector space \(\mathbb{F}_2^{2s}\), respectively: \[\begin{aligned}
Q(x)=Q(x_1,\dotsc,x_{2s}) &:=\sum_{r=1}^s x_{2r-1}x_{2r},\\ B(x,y)=B((x_1,\dotsc,x_{2s}),(y_1,\dotsc,y_{2s})) &:=\sum_{r=1}^s (x_{2r-1}y_{2r}+x_{2r}y_{2r-1}). \end{aligned}\] Let \(\Gamma\) be the corresponding character table of \(\mathbb{F}_2^{2s}\), defined by \(\Gamma(x,y)=(-1)^{B(x,y)}\) for all \(x,y\in\mathbb{F}_2^{2s}\). Any submatrix of \(\Gamma\) obviously satisfies [\[P1\]](#P1){reference-type="eqref" reference="P1"} from Proposition [\[muetfCharacterization\]](#muetfCharacterization){reference-type="ref" reference="muetfCharacterization"}. Let \(X\) be the specific submatrix of \(\Gamma\) whose rows and columns are indexed by \(\{x\in\mathbb{F}_2^{2s}: Q(x)=1\}\) and \(\{x\in\mathbb{F}_2^{2s}: Q(x)=0\}\), respectively. By Lemma 4.2 of , these two subsets of \(\mathbb{F}_2^{2s}\) are *difference sets* for \(\mathbb{F}_2^{2s}\) of cardinality \(k\) and \(l\), respectively. As detailed in , this means that the rows and columns of \(X\) are equiangular, namely that [\[P3\]](#P3){reference-type="eqref" reference="P3"} and [\[P4\]](#P4){reference-type="eqref" reference="P4"} hold. Theorem 4.4 of moreover gives that these two difference sets are *paired*, meaning that the columns of \(X\) form a tight frame for their span, so that [\[P2\]](#P2){reference-type="eqref" reference="P2"} holds. Theorem 3.3 of then implies that the rank of \(X\) is indeed \(m\), so that [\[P5\]](#P5){reference-type="eqref" reference="P5"} holds. 0◻.1in We close this section by highlighting that real mutually unbiased equiangular tight frames are rare objects. Precisely, we have obtained rather stringent necessary conditions for their existence (not included here because too detached from our main focus). For instance, these conditions imply that mutually unbiased \(k\)-vector and \(l\)-vector equiangular tight frames for \(\mathbb{R}^m\) can only exist for at most thirteen triples of integers \((m,k,l)\) with \(l > k > m+1\) when \(m \le 1000\), and that they cannot exist when \(l=k>m\), in contrast with the complex setting.
# Epilogue: the fifth maximal projection constant
By combining the main results derived in the two previous sections, namely Theorems [\[amazing\]](#amazing){reference-type="ref" reference="amazing"} and [\[muetfFamily\]](#muetfFamily){reference-type="ref" reference="muetfFamily"}, and after some tedious algebraic manipulation, we can state that the maximal relative projection constant at any \(m\) of the form \(m = (2^{2s}-1)/3\) for some integer \(s \ge 2\) is bounded below as \[\label{LBFam} \lambda ( m ,4^s) \ge \frac{2^{2s}-1}{2^{3s}-3\, 2^{s-1} + 1} \left( \frac{2^{2s-1} + 2^s-1}{3} + 2^{s-1} \sqrt{m} \right).\] If this was to be an equality, then the vector \(t_\theta \in \mathbb{R}_+^{n}\), \(n = 4^s\), and the matrix \(U_\theta \in \mathbb{R}^{m \times n}\), \(m = (2^{2s}-1)/3\), appearing in the proof of Theorem [\[amazing\]](#amazing){reference-type="ref" reference="amazing"} should be maximizers of the expression for \(\lambda(m,n)\) from Theorem [\[lammbda\]](#lammbda){reference-type="ref" reference="lammbda"}. For genuine maximizers \(\bar{t} \in \mathbb{R}_+^{n}\) and \(\bar{U} \in \mathbb{R}^{m \times n}\), we emphasize the following two necessary conditions:
(a) [\[NecCond1\]]{#NecCond1 label="NecCond1"} \(\bar{t}\) is a maximizer of \(\sum_{i,j} t_i t_j |\bar{U}^\top \bar{U} | _{i,j}\) subject to \(\|t\|_2 = 1\), so is characterized by the fact that \(\bar{t}\) is an eigenvector (in fact, the leading eigenvector) of \(|U^\top U|\)---this is indeed satisfied by \(t_\theta\) and \(U_\theta\), according to [\[SP1\]](#SP1){reference-type="eqref" reference="SP1"};
(b) [\[NecCond2\]]{#NecCond2 label="NecCond2"} \(\bar{U}\) is a maximizer of \(\sum_{i,j} \bar{t}_i \bar{t}_j {\rm sgn}(\bar{U}^\top \bar{U})_{i,j} (U^\top U) _{i,j} = \textrm{tr} (\bar{T} {\rm sgn}(\bar{U}^\top \bar{U}) \, \bar{T} U^\top U)\), \(\bar{T} := {\rm diag}[\bar{t}]\), subject to \(U U^\top = {\rm I}_m\), so is characterized by the fact that the rows of \(\bar{U}\) are eigenvectors corresponding to the \(m\) largest eigenvalues of \(\bar{T} {\rm sgn}(\bar{U}^\top \bar{U}) \bar{T}\)--- this is indeed satisfied by \(t_\theta\) and \(U_\theta\), according to [\[SP2\]](#SP2){reference-type="eqref" reference="SP2"}.
The observation that \(t_\theta\) and \(U_\theta\) do satisfy conditions [\[NecCond1\]](#NecCond1){reference-type="eqref" reference="NecCond1"}-[\[NecCond2\]](#NecCond2){reference-type="eqref" reference="NecCond2"} supports the belief that [\[LBFam\]](#LBFam){reference-type="eqref" reference="LBFam"} could be an equality. To the question of whether the right-hand side of [\[LBFam\]](#LBFam){reference-type="eqref" reference="LBFam"} also coincides with the value of the maximal absolute projection constant \(\lambda ( m )\), \(m = (2^{2s}-1)/3\), the answer is in general *no*. Indeed, for \(s=3\), hence for \(m=21\), \(k=28\), and \(l=36\), we have \(\gamma_{21,28,36} \approx 3.9397\), while a real equiangular tight frame for \(\mathbb{R}^{21}\) made of \(126\) vectors is known to exist (see e.g. ), so Theorem [\[ETFs\]](#ETFs){reference-type="ref" reference="ETFs"} yields \(\lambda(21) \ge \lambda(21,126) \gtrsim 4.3333\). However, for \(s=2\), hence for \(m=5\), \(k=6\), and \(l=10\), there are convincing reasons to believe that \(\gamma_{5,6,10} \approx 2.06919\) coincide with the value of \(\lambda(5)\). These reasons are the extensive numerical investigations carried out B. L. Chalmers, as well as our own computations (some of which can be found in a matlab reproducible available on the authors' webpages). All these clues prompt us to conclude with the following assertion.
.5in
# Appendix {#appendix .unnumbered}
As bonus material, we present here a new proof of Theorem [\[ETFs\]](#ETFs){reference-type="ref" reference="ETFs"} as a immediate consequence of the technical result below coupled with Theorem [\[lammbda\]](#lammbda){reference-type="ref" reference="lammbda"}.
For \(t \in \mathbb{R}^n\) satisfying \(\|t\|_2 = 1\) and \(U \in \mathbb{R}^{m \times n}\) satisfying \(U U^\top = {\rm I}_m\), we use the nonnegativity of \((U^\top U)_{i,i}\) (as the inner product of the \(i\)th column of \(U\) with itself) and Cauchy--Schwarz inequality to write \[\begin{aligned}
\Sigma & := \sum_{i,j=1}^n t_i t_j |U^\top U|_{i,j} = \sum_{i=1}^n t_i^2 |U^\top U|_{i,i} + \sum_{\substack{i,j=1 \\ i \not= j}}^n t_i t_j |U^\top U|_{i,j} \\ & \le \sum_{i=1}^n t_i^2 (U^\top U)_{i,i} + \sqrt{ \sum_{\substack{i,j=1 \\ i \not= j}}^n t_i^2 t_j^2 } \sqrt{ \sum_{\substack{i,j=1 \\ i \not= j}}^n (U^\top U)_{i,j}^2 } \\ & = \sum_{i=1}^n t_i^2 (U^\top U)_{i,i} + \sqrt{ \sum_{i,j=1}^n t_i^2 t_j^2-\sum_{i=1}^n t_i^4 } \sqrt{ \sum_{i,j=1}^n (U^\top U)_{i,j}^2-\sum_{i=1}^n (U^\top U)_{i,i}^2 }\\ & = \sum_{i=1}^n \alpha_i \beta_i + \sqrt{ A-\sum_{i=1}^n \alpha_i^2} \sqrt{ B-\sum_{i=1}^n \beta_i^2 }, \end{aligned}\] where we have set \(\alpha_i = t_i^2\), \(\beta_i = (U^\top U)_{i,i}\), \(A = \big(\sum_i t_i^2 \big) \big(\sum_j t_j^2 \big) = \|t\|_2^4 = 1\), and \(B = \sum_{i,j} (U^\top U)_{i,j}^2 = \|U^\top U\|_F^2 = \textrm{tr}(U^\top U U^\top U) = \textrm{tr}(U U^\top U U^\top) = m\). Setting also \(a = \|t\|_2^2 =1\), \(b = \textrm{tr}(U^\top U) = \textrm{tr}(U U^\top) = m\), as well as \[x_i := \frac{\alpha_i-a/n}{\sqrt{A-a^2/n}} \qquad \mbox{and} \qquad y_i := \frac{\beta_i-b/n}{\sqrt{B-b^2/n}},\] we notice that \(\sum_{i=1}^n x_i =0\) and \(\sum_{i=1}^n y_i =0\). We exploit these identities a few times to derive \[\begin{aligned}
\Sigma & \le \sum_{i=1}^n \bigg( \frac{a}{n} + \sqrt{A-\frac{a^2}{n}} x_i \bigg) \bigg( \frac{b}{n} + \sqrt{B-\frac{b^2}{n}} y_i \bigg)\\ & + \sqrt{A-\sum_{i=1}^n \bigg( \frac{a}{n} + \sqrt{A-\frac{a^2}{n}} x_i \bigg)^2} \sqrt{ B-\sum_{i=1}^n \bigg( \frac{b}{n} + \sqrt{B-\frac{b^2}{n}} y_i \bigg)^2}\\ & = \frac{ab}{n} + \sqrt{A-\frac{a^2}{n}} \sqrt{B-\frac{b^2}{n}} \sum_{i=1}^n x_i y_i\\ & + \sqrt{A-\frac{a^2}{n}-\Big( A-\frac{a^2}{n} \Big) \sum_{i=1}^n x_i ^2} + \sqrt{B-\frac{b^2}{n}-\Big( B-\frac{b^2}{n} \Big) \sum_{i=1}^n y_i ^2} \\ & = \frac{ab}{n} + \sqrt{A-\frac{a^2}{n}} \sqrt{B-\frac{b^2}{n}} \left[ \sum_{i=1}^n x_i y_i + \sqrt{1-\sum_{i=1}^n x_i^2} \sqrt{1-\sum_{i=1}^n y_i^2} \right]. \end{aligned}\] The latter term in square brackets is nothing but the inner product of the unit vectors \(\tilde{x} := \begin{bmatrix} x, \sqrt{1-\|x\|_2^2} \end{bmatrix}\) and \(\tilde{y} := \begin{bmatrix} y, \sqrt{1-\|y\|_2^2} \end{bmatrix}\), so it is bounded by one. Thus, keeping the values of \(a=1\), \(b=m\), \(A=1\), and \(B=m\) in mind, we arrive at \[\sum_{i,j=1}^n t_i t_j |U^\top U|_{i,j} \le \frac{m}{n} + \sqrt{1-\frac{1}{n}} \sqrt{m-\frac{m^2}{n}}.\] Taking the supremum over \(t\) and \(U\) leads to the desired inequality [\[UB_proved\]](#UB_proved){reference-type="eqref" reference="UB_proved"} after some algebraic manipulation. This inequality turns into an equality if the matrix \(U \in \mathbb{R}^{m \times n}\) with \(U U^\top = {\rm I}_m\) satisfies \((U^\top U)_{i,i} = m/n\) for all \(i \in \{1,\ldots, n\}\) and \(|U^\top U|_{i,j} = \sqrt{(n-m)m/(n-1)}/n\) for all \(i \not= j \in \{1,\ldots, n\}\), simply by choosing \(t \in \mathbb{R}^{n}\) with entries \(t_i = 1/\sqrt{n}\) for all \(i \in \{ 1, \ldots,n\}\). Conversely, let us assume that [\[UB_proved\]](#UB_proved){reference-type="eqref" reference="UB_proved"} is an equality. Our goal is now to prove that \((U^\top U)_{i,i} = m/n\) for all \(i \in \{1,\ldots, n\}\) and \(|U^\top U|_{i,j} = \sqrt{(n-m)m/(n-1)}/n\) for all \(i \not= j \in \{1,\ldots, n\}\), where \(U \in \mathbb{R}^{m \times n}\) satisfying \(U U^\top = {\rm I}_m\) achieves the maximum, together with \(t \in \mathbb{R}^n\) satisfying \(\|t\|_2 = 1\). We start by taking into account that equality must hold throughout the first part of the argument. Equality in Cauchy--Schwarz inequality implies the existence of \(c \in \mathbb{R}\) such that \[t_i t_j = c \, |U^\top U|_{i,j} \qquad \mbox{for all }i \not= j \in \{1,\ldots, n\}\] and equality in \(\langle \tilde{x},\tilde{y} \rangle \le 1\) yields \(x=y\), i.e., \[\label{EqIn...} (U^\top U)_{i,i}-\frac{m}{n} = \frac{\sqrt{m-m^2/n}}{\sqrt{1-1/n}} \bigg( t_i^2-\frac{1}{n} \bigg) \qquad \mbox{for all }i \in \{1,\ldots, n\}.\] Since the matrix \(T {\rm sgn}(U^\top U) T\) has diagonal entries \((T {\rm sgn}(U^\top U) T)_{i,i} = t_i^2\) and off-diagonal entries \[(T {\rm sgn}(U^\top U) T)_{i,j} = t_i t_j {\rm sgn}(U^\top U)_{i,j} = c \, |U^\top U|_{i,j} {\rm sgn}(U^\top U)_{i,j} = c \, (U^\top U)_{i,j},\] the necessary condition [\[NecCond2\]](#NecCond2){reference-type="eqref" reference="NecCond2"}, written for all \(i \in \{1,\ldots,n\}\) and \(h \in \{1,\ldots,m\}\) as \[\sum_{j=1}^n (T {\rm sgn}(U^\top U) T)_{i,j} U^\top_{j,h} = \mu_h U^\top_{i,h},\] where \(\mu_1 \ge \cdots \ge \mu_m\) are the \(m\) leading eigenvalues of \(T ({\rm sgn}(U^\top U) T\), becomes \[t_i^2 U^\top_{i,h} + \sum_{\substack{j=1\\j\not= i}}^n c \, (U^\top U)_{i,j} U^\top_{j,h} = \mu_h U^\top_{i,h}.\] In other words, for all \(i \in \{1,\ldots,n\}\) and \(h \in \{1,\ldots,m\}\), we have \[t_i^2 U^\top_{i,h} + c \, (U^\top U U^\top)_{i,h}-c \, (U^\top U)_{i,i} U^\top_{i,h} = \mu_h U^\top_{i,h},\] or equivalently, in view of \(U U^\top = {\rm I}_m\), \[\label{SomeEq} \left( t_i^2 + c-c \, (U^\top U)_{i,i} \right) U^\top_{i,h} = \mu_h U^\top_{i,h}.\] This actually shows that \(\mu_h\) is independent of \(h \in \{1,\ldots,m\}\) and---thanks to the alternate expression \(\lambda(m,n) = \mu_1+\ldots+\mu_m\) (see e.g. )---one must have \(\mu_h = \lambda(m,n)/m\). Now [\[SomeEq\]](#SomeEq){reference-type="eqref" reference="SomeEq"} reduces (say, by multiplying by \(U^\top_{i,h}\), summing over \(h\), and simplifying) to \(t_i^2 + c-c \, (U^\top U)_{i,i} = \lambda(m,n)/m\). Summing over \(i \in \{1,\ldots,n\}\)) yields \[1 + c \, (n-m) = \frac{n}{m} \lambda(m,n) = 1 + \sqrt{\frac{(n-1)(n-m)}{m}},\] which shows that \[c= \sqrt{\frac{n-1}{m(n-m)}}.\] Invoking Remark [\[RkCombine2NedCond\]](#RkCombine2NedCond){reference-type="ref" reference="RkCombine2NedCond"}, we notice that \((U^\top U)_{i,i} = m t_i^2\) for all \(i \in \{1,\ldots,n\}\), and therefore [\[EqIn\...\]](#EqIn...){reference-type="eqref" reference="EqIn..."} becomes \(m(t_i^2-1/n) = \sqrt{m(n-m)/(n-1)}(t_i^2-1/n)\). Given that \(m~\not=~\sqrt{m(n-m)/(n-1)}\) when \(m>1\), we consequently obtain \(t_i^2 = 1/n\) for all \(i~\in~\{1,\ldots,n\}\). In turn, we deduce from \((U^\top U)_{i,i} = m t_i^2\) that \((U^\top U)_{i,i} = m/n\) for all \(i \in \{1,\ldots,n\}\) and from \(c |U^\top U|_{i,j} = t_i t_j\) that \(|U^\top U|_{i,j} =\sqrt{m(n-m)/(n-1)}/n\) for all \(i \not= j \in \{1,\ldots,n\}\). The proof is now complete. 0◻
[^1]: As pointed out to us by Josiah Park, this same \(6\times 10\) matrix appeared in a recent investigation of spherical half-designs (see ). | {'timestamp': '2022-06-06T02:15:17', 'yymm': '2206', 'arxiv_id': '2206.01596', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01596'} |
null | null |
null | null |
null | null |
# Introduction
The investigation of turbulent flows involves description of a very complicated nonlinear dynamics of small scale fluctuations, hence it is extremely difficult and requires sophisticated mathematical tools. In particular the emergence of large-scale coherent structures is a topic of interest, i.e. the transfer of energy from small scales to the large scales termed the inverse energy cascade. To simplify the problem various assumptions have been put forward and in particular a common simplification in the theoretical approaches is the assumption of the so-called weak turbulence, which corresponds to weak amplitude of turbulent pulsations and linearization of their evolution, cf.. Such an approach lacks the crucial effect of nonlinear dynamics of the fluctuations. As argued in or in some cases the regime of weak turbulence can be sustained for long times, nevertheless, it is much more common for natural systems to develop into the strong turbulence regime, where the evolution of turbulent fluctuations becomes non-linear. In order to treat the fully nonlinear regime and reliably estimate turbulent diffusion the renormalization technique has been developed and applied to strongly turbulent flows. This is a statistical closure approximation which is based on systematic, subsequent (iterative) elimination of thin wave-number bands from the Fourier spectrum of rapidly evolving variables (cf. ). Notable contributions come from who have published comprehensive works on renormalization of hydrodynamic equations. Another powerful method which allows to relate the turbulent diffusion to the turbulent energy tensor is the so-called two-scale direct-interaction approximation (TSIDA) dating back to Kraichnan (1959, 1965). It involves introduction of a tensorial response function to an infinitesimal impulse-force and application of a two-scale approach in space and time related by the same parameter of expansion. Despite its limitations it allows to describe the turbulent viscosity in strong turbulence once the statistical properties of the underlying small-scale chaotic flow are known, see for a review. Recent investigations of involved applications of the renormalization group method to study the effect of non-stationarity and anisotropy on the magnetohydrodynamic turbulence in what could be called an 'intermediate regime of turbulence' or 'weakly nonlinear turbulence', in contrast to simple, linear, weak turbulence regime. Due to high complexity of the mathematical approach in the case of non-stationary and non-isotropic turbulence the effect of nonlinear evolution of the fluctuations has been included only at leading order at each step of the renormalization procedure. As a result, although reliable estimates of the turbulent electromotive force could be made the wave number dependence of all the turbulent coefficients likewise of the energy and helicity spectra was not fully resolved. Because the problem of turbulent convection is also anisotropic due to action of vertical buoyancy, full renoramalization of the Boussinesq equations turns out unattainable thus the latter approach corresponding to 'intermediate turbulence' is adopted here in order to study the physical mechanism of formation of the large-scale convection cells from small-scale energy input. The nonlinear evolution of turbulent pulsations is thus included through calculation of leading order expressions for the effective turbulent viscosity at each step of the iterative renormalization procedure, which corresponds to expansion of the full, renormalized coefficients up to first order in the Reynolds number. The turbulence is assumed to be driven by a small-scale, statistically random (Gaussian) heat source and the Prandtl number, that is the ratio of the viscosity to thermal diffusivity is assumed small, so that the temperature dynamics is dominated by rapid diffusion and the heat source. It is worth mentioning, that random heat sources are considered e.g. in the dynamics of dusty media as stochastically heated dust grains play an important role in transport of radiation (cf. ) and in the problem of stochastic heat engines (cf. ). Renormalized dynamical equations for the mean flow are obtained which contain turbulent coefficients describing the net nonlinear effect of short-wavelength fluctuations, such as the turbulent viscosity and the turbulent coefficient describing the effective 'motive force' at large scales, which takes the form of negative vertical diffusion in the studied regime. The results correspond to a somewhat initial stage of formation of the large-scale convective flow, as the 'intermediate turbulence' regime is necessarily eventually destroyed and strong turbulence must emerge. The dynamics of the turbulent Boussinesq convection involves formation of large-scale circulation (LSC) or the so-called 'wind of turbulence' (cf. ). In cylindrical geometry with comparable vertical and horizontal size, the LSC is believed to result from a quasi-two-dimensional, coupled horizontal sloshing and torsional (ST) oscillatory mode. Roche (2020) studied the physics of transition to the 'ultimate state' of convection at very high Rayleigh number (a measure of relative strength of the buoyancy force with respect to diffusion) and developed a model based on boundary layer stability. Zwirner et al. (2020) suggested, that such transitions could occur through the development of elliptical instabilities and showed that states with smaller amount of large-scale rolls built on top of each other transport heat more efficiently than states with more complex roll-structure. Vasiliev et al. (2019) discovered for the first time spontaneous formation of large-scale azimuthal flow. Here, we analyse the system driven thermally by a random heat source and in that way we avoid the problem of boundary conditions and thus the effect of boundary layers; we do not study the LSC structure, but simply study the physical mechanism of LSC formation, i.e. derive formula for the effective 'motive force' driving LSC, which is shown to take the form of negative vertical diffusion. This exact form is valid only as long as the regime of weak and 'intermediate' turbulence persists, that is at the initial and intermediate stages of evolution and turbulence development, but once the turbulence becomes strong the structure of the 'motive' force is expected to change. This is a fundamental study and the result sheds light on the physics of the process of energy transfer to large scales in thermal convection. The introduced simplification can be viewed as an advantage in the sense, that the problem of inverse turbulent energy cascade in convection is extracted and studied in isolation from the influence of velocity boundary conditions.
# Dynamical equations and mathematical formulation
To study the thermally driven turbulence in an incompressible fluid we consider a fluid layer between two flat, parallel boundaries distant \(L\) apart, with gravity \(\mathbf{g}=-g\hat{\mathbf{e}}_{z}\) pointing downwards and volume heat sources delivering heat to the system at a rate \(\mathcal{Q}(\mathbf{x},t)\). Such a system is governed by the following dynamical equations describing the evolution of the velocity field of the fluid flow \(\mathbf{u}(\mathbf{x},t)\) and the temperature field \(T(\mathbf{x}.t)\) under the Boussinesq approximation (cf. )
where \(\Pi=p/\bar{\rho}\) is the pressure divided by the mean density \(\bar{\rho}\), \(\kappa=k/\bar{\rho}\bar{c}_{p}=\mathrm{const.}\) is the thermal diffusivity of the fluid, uniform by assumption (\(k\) is the fluid's thermal conductivity) and \(\bar{T}\) is the mean temperature of the system which within the Boussinesq approximation is much greater than any temperature variations; \(\bar{c}_{p}\), \(\bar{\alpha}\) are the mean values of the specific heat at constant pressure and the coefficient of thermal expansion, respectively. The kinematic viscosity of the fluid is denoted by \(\nu\) and \(Q=\mathcal{Q}/\bar{\rho}\bar{c}_{p}\) is the heat source in \(K/s\). The solenoidal constraint for the velocity field ([\[eq:DIVS\]](#eq:DIVS){reference-type="ref" reference="eq:DIVS"}) simply expresses the law of mass conservation. As typically done in the case of Boussinesq convection we have also assumed, that the adiabatic temperature gradient \(g\bar{\alpha}\bar{T}/\bar{c}_{p}\ll\left\Vert \nabla T\right\Vert\) is much smaller than the typical temperature gradients in the fluid flow (typically about a thousand times smaller in laboratory flows, cf. e.g. ). We concentrate on fluids with low Prandtl numbers \[Pr=\frac{\nu}{\kappa}\ll1,\label{eq:Pr_small}\] such as e.g. liquid gallium and write down the equations in the following non-dimensional form
where the Reynolds number \((Re=)\epsilon\) and the P\(\acute{\textrm{e}}\)clet number \(Pe\) are defined in a standard form \[\label{RePePr} \epsilon=\frac{UL}{\nu},\qquad Pe=\epsilon Pr=\frac{UL}{\kappa},\] and we have chosen \(L^{2}/\nu\) for the time scale, \(L\) for the spatial scale, \(\nu U/g\bar{\alpha}L^{2}\) for the temperature scale, \(\kappa\nu U/g\bar{\alpha}L^{4}\) for the heat source scale and finally pressure was scaled with \(\bar{\rho}\nu U/L\). We will seek for the form of the large-scale flow equations in the limit of 'intermediate turbulence' (described in general terms in the introduction and in detail below and in ), through expansions in the Reynolds number \(\epsilon\); since \(Pr\ll1\) it follows, that the P\(\acute{\textrm{e}}\)clet number must also be small, \(Pe\ll1\). Hence the final set of equations describing convection at low Prandtl number takes the form
Furthermore, we introduce the Fourier transforms defined in the following way
where according to standard renormalization approach we have introduced the upper cut-off for the Fourier spectra \(\Lambda\), which in natural systems appears due to enhanced energy dissipation at very small scales of turbulence. The aim of this analysis is to study the large-scale flows in turbulent convection at low \(Pr\). In order to model developed turbulence we assume a stochastic heat source, Gaussian with zero mean \[\left\langle Q\right\rangle =0,\label{eq:zero_mean_HS}\] statistically homogeneous and stationary, fully defined by the following correlation function \[\left\langle \hat{Q}(\mathbf{k},\omega)\hat{Q}(\mathbf{k}',\omega')\right\rangle =\Xi\left(k\right)\delta(\mathbf{k}+\mathbf{k}')\delta(\omega+\omega'),\label{eq:force_correlations}\] where the function \(\Xi\left(k\right)\) will be specified later and angular brackets denote the ensemble mean, \[\left\langle \cdot\right\rangle-\textrm{ensemble mean}.\] Note, that so-induced turbulence will be *anisotropic* because of action of vertical gravity (buoyancy force). We can calculate a positive definite quantity \[\int k^{2}\mathrm{d}\mathring{\varOmega}_{\mathbf{k}}\int\mathrm{d}^{4}q'\left\langle \hat{Q}(\mathbf{k},\omega)\hat{Q}(\mathbf{k}',\omega')\right\rangle =4\pi k^{2}\Xi\left(k\right)>0,\label{eq:F2}\] where \(\mathring{\varOmega}_{\mathbf{k}}\) denotes a solid angle associated with the vector \(\mathbf{k}\), which implies \(\Xi\left(k\right)>0\). The approach will be based on the renormalization group technique, which is an iterative procedure of successive elimination of thin wave-number bands from the Fourier spectrum of fluctuating fields. In this way the effect of thin bands of modes with shortest wavelengths on the remaining modes is calculated at each step of the procedure. The final aim of this approach is to obtain recursion equations for coefficients describing the effective mean Reynolds stress \(\left\langle \mathbf{u}\mathbf{u}\right\rangle\) as a function of the wave number at each step of the procedure. The Reynolds stresses are responsible for creation of the turbulent viscosity and what can be called a 'motive force' for the large-scale flows. The recursion equations (provided in ([\[eq:xi-2-1-1\]](#eq:xi-2-1-1){reference-type="ref" reference="eq:xi-2-1-1"}-c)) are then solved for \(k\rightarrow0\) in order to obtain the final forms of the large-scale viscosity and the motive force which appear in the mean-field equations and include the effect of nonlinear evolution of turbulent fluctuations.
# The iterative weakly nonlinear renormalization procedure[\[sec:Renormalization-procedure\]]{#sec:Renormalization-procedure label="sec:Renormalization-procedure"}
Introducing new shorter four-component vector notation \[\mathbf{q}=(\mathbf{k},\omega),\qquad\int_{0}^{\Lambda}d^{3}k\int_{-\infty}^{\infty}d\omega(\cdot)=\int d^{4}q(\cdot),\label{eq:short_notation}\] so that e.g. \[u_{i}(\mathbf{x},t)=\int d^{4}q\hat{u}_{i}(\mathbf{q})\mathrm{e}^{\mathrm{i}(\mathbf{k}\cdot\mathbf{x}-\omega t)},\label{eq:short_notation_1}\] the equations take the form
In the above we have also defined the following convolution integral \[\mathbb{I}_{ij}^{(u)}=\int d^{4}q'\hat{u}_{i}(\mathbf{q}-\mathbf{q}')\hat{u}_{j}(\mathbf{q}'),\label{eq:Integrals_def}\] which possesses the symmetry property \[\mathbb{I}_{ij}^{(u)}=\mathbb{I}_{ji}^{(u)}.\label{eq:integral_symmetries}\] It should be noted at this stage, that the convolution integrals, which represent the nonlinear interactions between fluctuating turbulent fields are *not* neglected in the evolutional equations for fluctuations ([\[eq:u_Fourier\]](#eq:u_Fourier){reference-type="ref" reference="eq:u_Fourier"},b), and their effect will be included within the scope of the 'intermediate turbulence regime', based on the iterative renormalization procedure introduced in. Thus we go beyond the weak turbulence regime and quantitatively express the effect of this nonlinearity on the dynamics of the large-scale flow. However, contrary to the standard renormalization approach at each step of the iterative procedure based on a step-by-step elimination of thin wave number bands from the Fourier spectrum, only the terms of the leading order in \(\epsilon\) will be retained; in such a way the Taylor series in \(\epsilon\) will not be finally contracted as in the renormalization approaches, but instead we will obtain a weakly nonlinear expressions on effective coefficients appearing in the final 'renormalized' equation for the large-scale flow, i.e. up to the first order in \(\epsilon\). In order to eliminate pressure we apply the projection operator \[P_{ij}(\mathbf{k})=\delta_{ij}-\frac{k_{i}k_{j}}{k^{2}},\label{eq:projection_operator}\] to both sides of the Navier-Stokes equation ([\[eq:u_Fourier\]](#eq:u_Fourier){reference-type="ref" reference="eq:u_Fourier"})
where we have already introduced \(\hat{T}(\mathbf{q})=\hat{Q}(\mathbf{q})/k^{2}\) from ([\[eq:T_Fourier\]](#eq:T_Fourier){reference-type="ref" reference="eq:T_Fourier"}) and \[P_{imn}(\mathbf{k})=k_{m}P_{in}(\mathbf{k})+k_{n}P_{im}(\mathbf{k}),\label{Pimn_def}\] \[\gamma=-\mathrm{i}\omega+k^{2}.\label{eq:gammas_def}\] The smallness of the amplitude of turbulence \(\epsilon\ll1\) allows for proper mathematical formulation of the problem, since the velocity field is expressed at leading order by the driving \(\hat{Q}\) and the nonlinearity, which is of the order \(\mathcal{O}(\epsilon)\), can be treated in perturbational sense. The iterational procedure of renormalization is then applicable. The final recursion differential relations for the coefficients describing the Reynolds stresses can be solved analytically, thus in particular the turbulent viscosity and the motive force for the large-scale flow can be determined. We now start the iterative procedure of taking successive little bites off the Fourier spectrum from the short-wavelength side in order to obtain the final nonlinear effect of the fluctuations on the means. At the first step of the procedure we introduce a parameter \(\lambda_{1}\), which satisfies \[\delta\lambda=\Lambda-\lambda_{1}\ll1,\label{eq:lambda_bite}\] and divide the Fourier spectrum into two parts
and same way for \(\hat{Q}\). The equation for the field \(\hat{u}_{i}^{<}(\mathbf{k},\omega)\) is obtained by averaging ([\[eq:PijNS\]](#eq:PijNS){reference-type="ref" reference="eq:PijNS"}) over the first shell (\(\lambda_{1}<k<\Lambda\)) \[\begin{aligned}
\left(-\mathrm{i}\omega+k^{2}\right)\hat{u}_{i}^{<}(\mathbf{q})= & \frac{P_{i3}}{k^{2}}\hat{Q}^{<}(\mathbf{q})-\frac{1}{2}\mathrm{i}\epsilon P_{imn}\mathbb{I}_{mn}^{(u^{<})}\left(\mathbf{q}\right)\nonumber \\ &-\frac{1}{2}\mathrm{i}\epsilon P_{imn}(\mathbf{k})\int d^{4}q'\left\langle \hat{u}_{m}^{>}(\mathbf{q}')\hat{u}_{n}^{>}(\mathbf{q}-\mathbf{q}')\right\rangle _{c}.\label{eq:u<} \end{aligned}\] To get the equation for \(\hat{u}_{i}^{>}(\mathbf{k},\omega)\) we utilize ([\[eq:PijNS\]](#eq:PijNS){reference-type="ref" reference="eq:PijNS"}) again \[\begin{aligned}
\hat{u}_{i}^{>}(\mathbf{q})= & \frac{P_{i3}}{k^{2}\gamma}\hat{Q}^{>}\left(\mathbf{q}\right)-\mathrm{i}\epsilon\frac{P_{imn}}{\gamma}\mathbb{J}_{mn}^{(u)}\left(\mathbf{q}\right)\nonumber \\ &-\frac{1}{2}\mathrm{i}\epsilon\frac{P_{imn}}{\gamma}\mathbb{I}_{mn}^{(u^{<})}\left(\mathbf{q}\right)+R_{i},\label{eq:u_gtr} \end{aligned}\] where \(\left\langle \cdot\right\rangle _{c}\) denotes conditional average over the first shell (\(\lambda_{1}<k<\Lambda\)) statistical subensemble, described at the beginning of the Appendix (cf. ); furthermore, we have defined \[\mathbb{J}_{mn}^{(u)}\left(\mathbf{q}\right)=\int d^{4}q'\hat{u}_{m}^{<}(\mathbf{q}')\hat{u}_{n}^{>}(\mathbf{q}-\mathbf{q}'),\label{eq:Ju}\] and the rest in ([\[eq:u_gtr\]](#eq:u_gtr){reference-type="ref" reference="eq:u_gtr"}) is given by \[R_{i}^{(u)}=-\frac{1}{2}\mathrm{i}\epsilon\frac{P_{imn}}{\gamma}\mathbb{I}_{mn}^{\left(u^{>}\right)}.\label{eq:Ru}\] The rests will be neglected on the basis of generating triple order statistical correlations since they involve only terms second order in \(\mathbf{u}^{>}\) or because of the kept order of accuracy in the asymptotic limit \(\epsilon\ll1\), which will allow to neglect terms of the order \(\mathcal{O}(\epsilon^{3})\); for details the reader is referred to the Appendix. In what follows we provide a short description of the asymptotic iterative procedure, described in detail in the Appendix. First we introduce ([\[eq:u_gtr\]](#eq:u_gtr){reference-type="ref" reference="eq:u_gtr"}) into ([\[eq:u\<\]](#eq:u<){reference-type="ref" reference="eq:u<"}) and calculate the dynamical effect of short-wavelength components \(\hat{u}_{i}^{>}(\mathbf{k},\omega)\) on the evolution of \(\hat{u}_{i}^{<}(\mathbf{k},\omega)\) (long wavelength modes). This results in corrections to some of the terms in equation ([\[eq:u\<\]](#eq:u<){reference-type="ref" reference="eq:u<"}), but also generates terms with a new structure. Therefore a next step is necessary, involving calculation of the effect of the next shell \(\lambda_{2}=\lambda_{1}-\delta\lambda<k<\lambda_{1}\) (new short-wavelength modes) on the modes with \(k<\lambda_{1}-\delta\lambda\) (new long wavelength modes); this is continued until invariance of the equations for long-wavelength modes is achieved, i.e. the equations do not change from one iterational step to the next one. We can then take the limit of infinitesimally narrow wave number bands \(\delta\lambda\rightarrow0\), which leads to differential recursion relations for all the coupling constants introduced into the equations for long wavelength modes by couplings of the short wavelength ones. In the isothermal, fully isotropic case Yakhot and Orszag (1986) calculated the correction from short wavelength modes in the Navier-Stokes equation which was proportional to \(k^{2}\hat{u}_{i}^{<}\delta\lambda\) thus creating viscosity correction; the turbulent viscosity was then obtained from an equation of the form \(\mathrm{d}\nu_{turb}/\mathrm{d}\lambda=f(\lambda)\) with an 'initial' condition \(\nu_{eff}(\lambda=\Lambda)=\nu\). The case at hand is *anisotropic* because of the vertical gravity (buoyancy force); explicit calculation of two initial steps of the renormalization procedure is enough to derive the final differential recursion relations with satisfactory accuracy. The details of the procedure are provided in the Appendix.
# Dynamics of the large-scale flow[\[subsec:Dynamics-of-mean\]]{#subsec:Dynamics-of-mean label="subsec:Dynamics-of-mean"}
The mean-field equations are derived in the Appendix A and take the following form
where the coefficient \[\varsigma=\frac{2\pi^{2}}{735}\frac{g^{2}\bar{\alpha}^{2}L^{3}\mathbb{Q}^{2}}{\nu^{3}\kappa^{2}k_{\ell}^{7}},\label{eq:varsigma-1}\] includes the effect of the turbulent fluctuations on the means; in the above \(\mathbb{Q}\) is the magnitude of heat delivery rate (in \(K/s\)) and \(k_{\ell}\) is the wave number based on the length scale of most energetic turbulent eddies. The general differential recursion relations for the turbulent coefficients are solved in the Appendix A, see ([\[eq:xi-2-1-1\]](#eq:xi-2-1-1){reference-type="ref" reference="eq:xi-2-1-1"}-c). It is evident, that the term \(-9\varsigma\nabla^{2}\left\langle U\right\rangle _{z}\hat{\mathbf{e}}_{z}\) is the large-scale motive force responsible for energy transfer from small scales to large scales, i.e. the inverse energy cascade; it takes the form of negative diffusion in the vertical direction, which drives the large-scale flow. For the sake of a rough estimate we may take the Kolmogorov cut-off value \(\Lambda\sim(\sqrt{gL}/\nu)^{3/4}L^{-1/4}\), where the free-fall velocity \(\sqrt{gL}\) was used as the convective velocity scale, which yields \(\varsigma\sim\nu\mathcal{G}^{-13/4}\mathcal{H}^{2}\left(k_{\ell}/\Lambda\right)^{-7}\), with \(\mathcal{H}=g^{1/2}\bar{\alpha}L^{7/2}\mathbb{Q}/\kappa\nu\) and \(\mathcal{G}=\sqrt{gL^{3}/\nu^{2}}\). Since \(k_{\ell}/\Lambda\ll1\), this coefficient is expected to be much larger than the molecular viscosity, in particular in strongly driven turbulence \(\mathcal{H}\gg1\). It is, in fact a typical situation when the turbulent coefficients greatly exceed the molecular ones.
## Linear regime
Linearization of ([\[eq:MF_eq_WN-2\]](#eq:MF_eq_WN-2){reference-type="ref" reference="eq:MF_eq_WN-2"}) and substitution of normal modes in the form
leads to \[\sigma\left\langle \mathbf{U}\right\rangle =-\mathrm{i}\mathbf{K}\hat{\Pi}-\nabla\left\langle \Pi\right\rangle +9\varsigma K^{2}\left\langle U\right\rangle _{z}\hat{\mathbf{e}}_{z}-\left(\nu+6\varsigma\right)K_{h}^{2}\left\langle \mathbf{U}\right\rangle-\left(\nu+4\varsigma\right)K_{z}^{2}\left\langle \mathbf{U}\right\rangle ,\label{eq:MF_eq_WN-1-2}\] which under action of the projection operator \(\mathbf{1}-\mathbf{K}\mathbf{K}/K^{2}\) where \(\mathbf{1}\) is the unity matrix transforms into \[\sigma\left\langle U\right\rangle _{i}=9\varsigma K^{2}\left(\delta_{i3}-\frac{K_{i}K_{z}}{K^{2}}\right)\left\langle U\right\rangle _{z}-\left(\nu+6\varsigma\right)K_{h}^{2}\left\langle U\right\rangle _{i}-\left(\nu+4\varsigma\right)K_{z}^{2}\left\langle U\right\rangle _{i},\label{eq:MF_eq_WN-1-2-1}\] so that the pressure is eliminated. This yields for the growth rate of the mean flow \[\sigma=\left(3\varsigma-\nu\right)K_{h}^{2}-\left(\nu+4\varsigma\right)K_{z}^{2}.\label{eq:g_rate}\] It follows, that in turbulent convection driven by strong stochastic heat sources \(\mathcal{H}\gg1\) the growth rate takes the approximate form \[\sigma\approx3\varsigma K_{h}^{2}-4\varsigma K_{z}^{2},\label{eq:grate_approx}\] and thus turbulence excites large-scale modes with horizontal wavelengths shorter than vertical ones, \[\sigma>0\;\Leftrightarrow\;K_{z}^{2}<\frac{3}{4}K_{h}^{2}\;\Leftrightarrow\;L_{h}<\frac{\sqrt{3}}{2}L_{z}\approx0.87L_{z}.\label{eq:sigma_Ls}\] In other words in the studied problem the large scale flow is expected to form vertically elongated rolls. The growth rate increases unboundedly with \(K_{h}^{2}\), but since the equation ([\[eq:MF_eq_WN-2\]](#eq:MF_eq_WN-2){reference-type="ref" reference="eq:MF_eq_WN-2"}) describes the large-scale flow only, there is a natural upper bound on the horizontal wave number of the large-scale modes and thus on the growth rate. As argued in the Appendix B there also exists an additional term on the r.h.s. of the mean flow equation ([\[eq:MF_eq_WN-2\]](#eq:MF_eq_WN-2){reference-type="ref" reference="eq:MF_eq_WN-2"}) of the form \(\varrho\partial_{z}^{2}\left\langle U\right\rangle _{z}\hat{\mathbf{e}}_{z}\), which is of smaller (asymptotically negligible) magnitude than the other turbulent terms proportional to the coefficient \(\varsigma\), i.e. \(\varrho<\varsigma\). Thus the growth rate is modified to \(\sigma=3\varsigma K_{h}^{2}-4\varsigma K_{z}^{2}-\varrho\frac{K_{z}^{2}K_{h}^{2}}{K^{2}}\), but the sign of the turbulent coefficient \(\varrho\) remains undetermined, hence it is not clear whether it acts as additional diffusion (if \(\varrho>0\)) or additional motive force (if \(\varrho<0\)). The normal modes in the form ([\[eq:Fourier_lin\]](#eq:Fourier_lin){reference-type="ref" reference="eq:Fourier_lin"}) are individually also solutions of the nonlinear equation ([\[eq:MF_eq_WN-2\]](#eq:MF_eq_WN-2){reference-type="ref" reference="eq:MF_eq_WN-2"}). Of course in the problem at hand, the energy is transferred from the small scale fluctuations, where the flow is thermally driven, to the large-scale modes and thus in the limit of small \(K\) the dynamics naturally involves wave packets, rather than individual modes, which evolve nonlinearly. Still it is possible for the most unstable modes to dominate the dynamics, in which case the amplitude of convection grows unboundedly in time until the initial assumption of small Reynolds number ceases to be valid and saturation may occur. In other words the analysis of weakly nonlinear turbulence does not lead to saturation of large-scale modes, which is possible only beyond the scope of this approach, that is in fully developed, strong turbulence.
# Conclusions
The presented analysis was focused on derivation of the effective equation describing the dynamics of the large-scale flow (circulation) in turbulent convection driven by a random heat source at low \(Pr\). The applied technique was based on the renormalization approach of and (see also for a review of the method), which allowed to incorporate the effect of the nonlinear terms in the dynamical equations for small-scale turbulent fluctuations, and calculate the anisotropic turbulent viscosity and 'motive force' induced by the fluctuations and experienced by the large-scale flows. The renormalized mean-flow equation was derived and it was shown, that the 'motive force' acts in the form of negative vertical diffusion, \(\partial_{t}\left\langle \mathbf{U}\right\rangle =-9\varsigma\nabla^{2}\left\langle U\right\rangle _{z}\hat{\mathbf{e}}_{z}+\dots\), where \(\varsigma\) is given in ([\[eq:varsigma-1\]](#eq:varsigma-1){reference-type="ref" reference="eq:varsigma-1"}), leading to enhancement of the mean flow energy. The general recursion differential equations for all the turbulent coefficients are provided in ([\[eq:xi-2-1\]](#eq:xi-2-1){reference-type="ref" reference="eq:xi-2-1"}-c) for any form of the random heat-source function \(\Xi(k)\).\
\
**Data availability statement:** All data generated or analysed during this study are included in this published article | {'timestamp': '2022-06-06T02:13:08', 'yymm': '2206', 'arxiv_id': '2206.01539', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01539'} |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction {#intro}
This paper deals with mass transportation problems when couplings are finitely additive probabilities. To be more precise, we need to recall the standard framework where transportation problems are usually investigated. In the sequel, the abbreviation *f.a.p.* stands for *finitely additive probability* and a *probability measure* is a \(\sigma\)-*additive* f.a.p. Moreover, we use the notation \[P(f)=\int f\,dP\] whenever \(P\) is a f.a.p. and \(f\) a function such that \(\int f\,dP\) is well defined.
## The standard framework {#z34d8ng1q}
Let \(I=\{1,\ldots,n\}\) where \(n\) is a positive integer. For each \(i\in I\), let \((\Omega_i,\mathcal{A}_i)\) be a measurable space and \(\mu_i\) a probability measure on the \(\sigma\)-field \(\mathcal{A}_i\). Define \[\begin{gathered}
\Omega=\Omega_1\times\ldots\times\Omega_n\quad\text{and}\quad\mathcal{A}=\mathcal{A}_1\otimes\ldots\otimes\mathcal{A}_n \end{gathered}\] and denote by \(\pi_i:\Omega\rightarrow\Omega_i\) the \(i\)-th canonical projection, namely, \[\pi_i(\omega)=\omega_i\quad\quad\text{for all }i\in I\text{ and }\omega=(\omega_1,\ldots,\omega_n)\in\Omega.\] Moreover, let \(c:\Omega\rightarrow [-\infty,\infty]\) be a (cost) function. Various conditions on \(c\) can be taken into account. In this paper, \(c\) *is \(\mathcal{A}\)-measurable and takes values in \([0,\infty]\). In particular, it may be that \(c=\infty\)*. A *coupling* (or a *transport plan*) is a probability measure \(P\) on \(\mathcal{A}\) having \(\mu_1,\ldots,\mu_n\) as marginals, in the sense that \[P\circ \pi_i^{-1}=\mu_i\quad\text{for all }i\in I.\] The collection of all couplings, henceforth denoted by \(\Gamma\), plays a basic role. A few classical issues are:
- Give conditions for the existence of \(P\in\Gamma\) such that \[\begin{gathered}
\label{kant} P(c)=\inf_{Q\in\Gamma}Q(c); \end{gathered}\]
- Characterize those \(P\in\Gamma\) satisfying equation [\[kant\]](#kant){reference-type="eqref" reference="kant"} (provided they exist);
- Give conditions for the duality relation \[\begin{gathered}
\inf_{Q\in\Gamma}Q(c)=\sup_{f_1,\ldots,f_n}\,\sum_{i=1}^n\mu_i(f_i) \end{gathered}\] where \(\sup\) is over the \(n\)-tuple \((f_1,\ldots,f_n)\) such that \(f_i\in L_1(\mu_i)\) for all \(i\in I\) and \(\sum_{i=1}^nf_i\circ\pi_i\le c\).
Here, \(L_1(\mu_i)=L_1(\Omega_i,\mathcal{A}_i,\mu_i)\) is the class of \(\mathcal{A}_i\)-measurable functions \(f:\Omega_i\rightarrow\mathbb{R}\) such that \(\mu_i(\abs{f})=\int\abs{f}\,d\mu_i<\infty\) (without identifying maps which agree \(\mu_i\)-a.s.). A natural development is to fix a subset \(\Gamma_0\subset\Gamma\) and to investigate (i)-(ii)-(iii) (and possibly other problems) with \(\Gamma_0\) in the place of \(\Gamma\). Following, for instance, \(\Gamma_0\) could be \[\begin{gathered}
\Gamma_0=\bigl\{P\in\Gamma:P(\abs{f})<\infty\text{ and }P(f)=0\text{ for all }f\in F\bigr\} \end{gathered}\] where \(F\) is a given class of \(\mathcal{A}\)-measurable functions \(f:\Omega\rightarrow\mathbb{R}\). As a special case, when \((\Omega_i,\mathcal{A}_i)=(\mathbb{R},\mathcal{B}(\mathbb{R}))\) for each \(i\in I\), a suitable choice of \(F\) yields \[\begin{gathered}
\Gamma_0=\bigl\{P\in\Gamma:(\pi_1,\ldots,\pi_n)\text{ is a martingale under }P\bigr\}. \end{gathered}\] Such a \(\Gamma_0\), introduced in, corresponds to the so-called *martingale transport*. In addition to be theoretically intriguing, martingale transport has solid financial motivations; see e.g., , and references therein. A last remark is that, for any choice of \(\Gamma_0\), a preliminary question is whether \(\Gamma_0\ne\emptyset\). In martingale transport, for instance, as a consequence of some results by Strassen, one obtains \(\Gamma_0\ne\emptyset\) if and only if \[\int\abs{x}\,\mu_i(dx)<\infty\quad\text{and}\quad\mu_i(f)\le\mu_{i+1}(f)\] for all \(i\in I\) and all convex functions \(f:\mathbb{R}\rightarrow\mathbb{R}\).
## Content of this paper
To investigate mass transportation in a finitely additive setting is a quite natural idea. Accordingly, various hints in this direction are scattered throughout the literature; see e.g., , and references therein. Usually, however, f.a.p.'s are only instrumental. Typically, a result concerning f.a.p.'s is nothing but an intermediate step toward the corresponding \(\sigma\)-additive result. To the best of our knowledge, no systematic treatment of the finitely additive mass transportation is available to date. This paper aims to fill this gap in the special case where \(\mu_1,\ldots,\mu_n\) are probability measures. Let \[\mathbb{P}=\bigl\{\text{all f.a.p.'s on }\mathcal{A}\bigr\}\] and \[M=\bigl\{P\in\mathbb{P}:P\circ \pi_i^{-1}=\mu_i\text{ for all }i\in I\bigr\}.\] In this paper, problems (i)-(ii)-(iii) are investigated with \(M\) in the place of \(\Gamma\). Similarly, the subsets \(\Gamma_0\subset\Gamma\) are replaced by the corresponding subsets \(M_0\subset M\). Our main result is that, if \(\Gamma\) is replaced by \(M\), each of problems (i)-(ii)-(iii) admits a solution assuming only that \(c\) is non-negative and \(\mathcal{A}\)-measurable. On the contrary, to have a solution in the standard framework, further conditions on \(c\) and/or \(\mu_1,\ldots,\mu_n\) are needed (such conditions are recalled in Section [3](#gu4n77y){reference-type="ref" reference="gu4n77y"}). Special attention is devoted to martingale transport. To illustrate, suppose \((\Omega_i,\mathcal{A}_i)=(\mathbb{R},\mathcal{B}(\mathbb{R}))\) for each \(i\in I\) and define \[M_0=\bigl\{P\in M:(\pi_1,\ldots,\pi_n)\text{ is a }P\text{-martingale}\}.\] If \(M_0\ne\emptyset\), then \(P(c)=\inf_{Q\in M_0}Q(c)\) for some \(P\in M_0\). Similarly, fix a reference f.a.p. \(P^*\in\mathbb{P}\) and define \[M_1=\bigl\{P\in M:P\ll P^*\text{ and }(\pi_1,\ldots,\pi_n)\text{ is a }P\text{-martingale}\}.\] Once again, if \(M_1\ne\emptyset\), there is \(P\in M_1\) such that \(P(c)=\inf_{Q\in M_1}Q(c)\). In the standard martingale transport, to our knowledge, \(M_1\cap\Gamma\) is neglected. This is actually strange, since \(M_1\cap\Gamma\) is quite meaningful. In fact, the usual motivation for martingale transport is that martingale probability measures play a role in various financial problems. However, in most such problems, probability measures are also requested to be equivalent, or at least absolutely continuous, with respect to a reference probability measure \(P^*\). Hence, to focus on \(M_1\cap\Gamma\) is reasonable in the standard martingale transport. In turn, in the framework of this paper, it is natural to focus on \(M_1\). The result quoted above requires \(M_1\ne\emptyset\). This condition is investigated in the last part of the paper. Suppose \(P^*\) is a probability measure with compact support and define \[\mathcal{U}=\bigl\{Q\in\mathbb{P}:Q\ll P^*\text{ and }(\pi_1,\ldots,\pi_n)\text{ is a }Q\text{-martingale}\bigr\}.\] Then, \(M_1\ne\emptyset\) if and only if \[\begin{gathered}
\sum_{i=1}^n\mu_i(f_i)\ge\inf_{Q\in\mathcal{U}}Q\Bigl(\,\sum_{i=1}^nf_i\circ\pi_i\Bigr) \end{gathered}\] for all bounded Borel functions \(f_1,\ldots,f_n:\mathbb{R}\rightarrow\mathbb{R}\). This result admits a \(\sigma\)-additive version as well. In fact, if \(P^*\) has discrete marginals (except possibly one), the above condition implies the existence of \(P\in\Gamma\) such that \(P\ll P^*\) and \((\pi_1,\ldots,\pi_n)\) is a \(P\)-martingale.
# Preliminaries {#sect2}
For any measurable space \((S,\mathcal{E})\), we denote by \(B(S,\mathcal{E})\) the set of bounded \(\mathcal{E}\)-measurable functions \(f:S\rightarrow\mathbb{R}\). If \(P,\,Q\in\mathbb{P}\), we write \(P\ll Q\) to mean that \(P(A)=0\) whenever \(A\in\mathcal{A}\) and \(Q(A)=0\). Moreover, if \(f_i:\Omega_i\rightarrow\mathbb{R}\), the map \(\sum_if_i\circ\pi_i\) is denoted by \(\oplus_if_i\). Hence, \(\oplus_if_i\) is the function on \(\Omega\) defined by \[\bigl(\oplus_{i=1}^nf_i\bigr)(\omega)=\sum_{i=1}^nf_i(\omega_i)\quad\quad\text{for all }\omega=(\omega_1,\ldots,\omega_n)\in\Omega.\]
## Integrals with respect to f.a.p.'s {#e4n8h5}
Let \(P\in\mathbb{P}\) and \(f:\Omega\rightarrow\mathbb{R}\) a real-valued \(\mathcal{A}\)-measurable function. If \(f\) is simple, \(\int f\,dP\) is defined in the usual way. If \(f\) is bounded, \(f\) is the uniform limit of a sequence \(f_k\) of simple functions, and we let \(\int f\,dP=\lim_k\int f_k\,dP\). If \(f\ge 0\), then \(f\) is \(P\)-integrable if and only if \[\sup_k\int f\wedge k\,dP<\infty,\] and we let \[\begin{gathered}
\label{n6t4} \int f\,dP=\sup_k\int f\wedge k\,dP. \end{gathered}\] In general, \(f\) is \(P\)-integrable if and only if \(f^+\) and \(f^-\) are both \(P\)-integrable, or equivalently \(\int\abs{f}\,dP<\infty\), and in this case \(\int f\,dP=\int f^+\,dP-\int f^-\,dP\). Observe now that equation [\[n6t4\]](#n6t4){reference-type="eqref" reference="n6t4"} makes sense whenever \(0\le f\le\infty\), merely \(\int f\,dP=\infty\) if \(f\) is real-non-negative but not \(P\)-integrable or if \(P(f=\infty)>0\). Hence, \(\int f\,dP\) is always defined by [\[n6t4\]](#n6t4){reference-type="eqref" reference="n6t4"} whenever \(f\) is \(\mathcal{A}\)-measurable and takes values in \([0,\infty]\). In a nutshel, this is a concise summary on finitely additive integration (when the integrand \(f\) is measurable with respect to a \(\sigma\)-field). For more on this subject, we refer to. As already noted, we will use the notation \(P(f)=\int f\,dP\) whenever \(\int f\,dP\) is well defined.
## Finitely additive martingales {#nj887c4d6h}
Let \((\Omega_i,\mathcal{A}_i)=(\mathbb{R},\mathcal{B}(\mathbb{R}))\) for all \(i\in I\). Define a class \(H\) of real-valued functions on \(\Omega=\mathbb{R}^n\) as \[H=\{\pi_1\}\cup\bigl\{\bigl(\pi_{i+1}-\pi_i\bigr)\,g(\pi_1,\ldots,\pi_i):1\le i<n,\,\,g\in B(\mathbb{R}^i,\mathcal{B}(\mathbb{R}^i))\bigr\}.\] If \(P\) is a probability measure on \(\mathcal{A}\), then \((\pi_1,\ldots,\pi_n)\) is a \(P\)-martingale if
- \(P(\abs{\pi_i})<\infty\) for all \(i\in I\),
- \(P(\pi_1)=0\) and \[E_P\bigl(\pi_{i+1}\mid \pi_1,\ldots,\pi_i\bigr)=\pi_i,\,\,P\text{-a.s., for all }i<n.\]
However, under (a), condition (b) is equivalent to
- \(P(f)=0\) for each \(f\in H\).
Therefore, if \(P\in\mathbb{P}\) is a f.a.p. on \(\mathcal{A}\), then \((\pi_1,\ldots,\pi_n)\) is said to be a \(P\)-martingale if conditions (a) and (c) are satisfied; see -.
## The product topology on \([0,1]^\mathcal{A}\) {#b4rt6}
Let \([0,1]^\mathcal{A}\) be the collection of all functions \(P:\mathcal{A}\rightarrow [0,1]\). When equipped with the product topology, \([0,1]^\mathcal{A}\) is a compact Hausdorff space and convergence is setwise convergence. Precisely, a net \((P_\alpha)\) in \([0,1]^\mathcal{A}\) converges to \(P\in [0,1]^\mathcal{A}\) if and only if \[P(A)=\lim_\alpha P_\alpha(A)\text{ for all }A\in\mathcal{A}.\] In particular, \(\mathbb{P}\) and \(M\) are compact subsets of \([0,1]^\mathcal{A}\). Moreover, if \(f\in B(\Omega,\mathcal{A})\), the map \(P\mapsto P(f)\) is continuous on \(\mathbb{P}\). In the sequel, \([0,1]^\mathcal{A}\) *is given the product topology and all its subsets are equipped with the corresponding relative topology*.
# Finitely additive couplings {#gu4n77y}
In this section, \(c:\Omega\rightarrow [0,\infty]\) is an \(\mathcal{A}\)-measurable function (the value \(\infty\) is admissible for \(c\)). Recall also that \(\mathbb{P}\) denotes the set of all f.a.p.'s on \(\mathcal{A}\) and \(M\) the collection of those \(P\in\mathbb{P}\) with marginals \(\mu_1,\ldots,\mu_n\). We begin with a preliminary lemma. Fix a collection \(F\) of real-valued \(\mathcal{A}\)-measurable functions on \(\Omega\) and define \[\mathcal{K}(F)=\bigl\{P\in M:P(\abs{f})<\infty\text{ and }P(f)=0\text{ for all }f\in F\bigr\}.\]
Condition [\[b7t9\]](#b7t9){reference-type="eqref" reference="b7t9"} is a form of uniform integrability. Since \[P\bigl\{\abs{f}\,1(\abs{f}>k)\bigr\}\le k^{-\epsilon}\,P\bigl\{\abs{f}^{1+\epsilon}\bigr\},\] a sufficient condition for [\[b7t9\]](#b7t9){reference-type="eqref" reference="b7t9"} is that, for each \(f\in F\), there is \(\epsilon>0\) such that \(\sup_{P\in M}P\bigl\{\abs{f}^{1+\epsilon}\bigr\}<\infty\). In particular, condition [\[b7t9\]](#b7t9){reference-type="eqref" reference="b7t9"} holds whenever \(F\subset B(\Omega,\mathcal{A})\). In any case, since \(P\mapsto P(c)\) is lower semicontinuous, one obtains \[\begin{gathered}
\label{cb9u4} P(c)=\inf_{Q\in K}Q(c)\quad\quad\text{for some }P\in K \end{gathered}\] provided \(K\subset M\) is compact and non-empty. Thus, with reference to problem (i), the \(\inf\) is attained for *any* non-negative \(\mathcal{A}\)-measurable cost function \(c\) provided \(\Gamma\) is replaced by a non-empty compact \(K\subset M\). The next result highlights some meaningful special cases.
Whether or not the sets \(K_1\) and \(K_3\) are non-empty is investigated in Section [4](#h81a3a){reference-type="ref" reference="h81a3a"}. Let us turn now to problem (iii). Similarly to (i), everything goes smoothly if \(\Gamma\) is replaced by \(M\).
In the standard framework (i.e., with \(\Gamma\) in the place of \(M\)), to settle problems (i) and (iii), one needs some conditions. To fix ideas, suppose the \(\Omega_i\) are separable metric spaces and \(\mathcal{A}_i\) the corresponding Borel \(\sigma\)-fields. Then, for the \(\inf\) in problem (i) to be attained, \(c\) should be lower semicontinuous and each \(\mu_i\) a tight probability measure. As regards (iii), duality holds if \(0\le c<\infty\) and the \(\Omega_i\) are Polish spaces. Moreover, in case the \((\Omega_i,\mathcal{A}_i)\) are arbitrary measurable spaces, duality holds if \(c\le\oplus_{i=1}^nf_i\) for some \(f_i\in L_1(\mu_i)\) and all but one the \(\mu_i\) are perfect probability measures. See e.g., ,, ,, ,. In view of Theorems [\[v77h3s1z\]](#v77h3s1z){reference-type="ref" reference="v77h3s1z"}-[\[h9k33d\]](#h9k33d){reference-type="ref" reference="h9k33d"}, all these conditions are superfluous if the couplings are allowed to be finitely additive. The next example helps to understand why.
We close this section with problem (ii). Let \[L=\bigl\{\oplus_{i=1}^nf_i:f_i\in L_1(\mu_i),\,i\in I\bigr\}.\]
Once again, to have an analogous of Theorem [\[uf6fa1\]](#uf6fa1){reference-type="ref" reference="uf6fa1"} in the standard framework, one needs some conditions. Basically, for Theorem [\[uf6fa1\]](#uf6fa1){reference-type="ref" reference="uf6fa1"} to hold with \(\Gamma\) in the place of \(M\), all but one the \(\mu_i\) should be perfect probability measures; see. Note also that, since \(P\in M\) is not necessarily \(\sigma\)-additive, \(P(c>f+\epsilon)=0\) for all \(\epsilon>0\) does not imply \(P(c>f)=0\).
# Existence of f.a.p.'s and probability measures satisfying certain conditions {#h81a3a}
For Theorem [\[v77h3s1z\]](#v77h3s1z){reference-type="ref" reference="v77h3s1z"} to apply, certain collections \(K_j\) of f.a.p.'s should be non-empty. We now give conditions for \(K_j\ne\emptyset\). By the same argument, we also obtain conditions for the existence of certain probability measures. This section is based on ideas from which in turn was inspired by Strassen. Let \(\mathcal{U}\subset\mathbb{P}\) be any collection of f.a.p.'s on \(\mathcal{A}\). If \(P\in M\cap\mathcal{U}\) and \(f_i\in B(\Omega_i,\mathcal{A}_i)\) for each \(i\in I\), one trivially obtains \[\begin{gathered}
\sum_{i=1}^n\mu_i(f_i)=P\Bigl(\oplus_{i=1}^nf_i\Bigr)\ge\inf_{Q\in\mathcal{U}}Q\Bigl(\oplus_{i=1}^nf_i\Bigr) \end{gathered}\] where the equality is due to \(P\in M\) and the inequality to \(P\in\mathcal{U}\). Hence, \[\begin{gathered}
\label{j92x1ly6} \sum_{i=1}^n\mu_i(f_i)\ge\inf_{Q\in\mathcal{U}}Q\Bigl(\oplus_{i=1}^nf_i\Bigr)\quad\quad\text{whenever }f_i\in B(\Omega_i,\mathcal{A}_i)\text{ for all }i\in I. \end{gathered}\] Our next goal is proving that, for some choices of \(\mathcal{U}\), condition [\[j92x1ly6\]](#j92x1ly6){reference-type="eqref" reference="j92x1ly6"} is actually equivalent to \(M\cap\mathcal{U}\ne\emptyset\). In what follows, we adopt the usual convention \[\inf\emptyset=\infty.\] Therefore, condition [\[j92x1ly6\]](#j92x1ly6){reference-type="eqref" reference="j92x1ly6"} is trivially false if \(\mathcal{U}=\emptyset\).
Theorem [\[f6h99k33ws\]](#f6h99k33ws){reference-type="ref" reference="f6h99k33ws"} applies to martingale transport.
The same argument used for Theorem [\[f6h99k33ws\]](#f6h99k33ws){reference-type="ref" reference="f6h99k33ws"} and Corollary [\[gb9971s\]](#gb9971s){reference-type="ref" reference="gb9971s"} can be applied to prove the existence of an absolutely continuous, martingale, *probability measure* with given marginals. More precisely, our last result provides conditions for the existence of \(P\in\Gamma\) (so that \(P\) is a probability measure on \(\mathcal{A}\) with marginals \(\mu_1,\ldots,\mu_n\)) such that \(P\ll P^*\) and \((\pi_1,\ldots,\pi_n)\) is a \(P\)-martingale.
We conclude this paper with two remarks. Firstly, it would be nice to have an analogous of Theorem [\[thmsadd444\]](#thmsadd444){reference-type="ref" reference="thmsadd444"}, as well as of the other results of this paper, with \(P\sim P^*\) in the place of \(P\ll P^*\), where \(P\sim P^*\) means \(P\ll P^*\) and \(P^*\ll P\). To this end, however, the techniques of this paper seem not to work. A few partial results are in, ,. Secondly, the class \(\mathcal{U}\) involved in Corollary [\[gb9971s\]](#gb9971s){reference-type="ref" reference="gb9971s"} and Theorem [\[thmsadd444\]](#thmsadd444){reference-type="ref" reference="thmsadd444"} may be empty. Let us consider the case where \(P^*\) is a probability measure. Then, by, one obtains \(\mathcal{U}\ne\emptyset\) if \(P^*\) satisfies the *non-arbitrage* condition \[P^*(f\ge 0)=1\quad\Rightarrow\quad P^*(f=0)=1\quad\quad\text{for each }f\text{ in the linear span of }H.\] The non-arbitrage condition, however, is stronger than \(\mathcal{U}\ne\emptyset\). In fact, it implies the existence of a probability measure \(P\) on \(\mathcal{A}\) such that \(P\sim P^*\) and \((\pi_1,\ldots,\pi_n)\) is a \(P\)-martingale. Assuming \(P^*\) has compact support, a characterization of \(\mathcal{U}\ne\emptyset\) is provided by. According to the latter, \(\mathcal{U}\ne\emptyset\) if and only if \[P^*(f>a)>0\quad\text{ for all }a<0\text{ and all }f\text{ in the linear span of }H.\] | {'timestamp': '2022-06-06T02:16:58', 'yymm': '2206', 'arxiv_id': '2206.01654', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01654'} |
# Introduction
With the continuous advancement of machine learning methodologies and increasing availability of dedicated software and hardware, data-driven approaches have gained popularity across many different fields such as finance, medicine, science, and engineering. Across these disciplines, one important topic of interest is describing how a dynamical system changes through time to predict its future state. While some dynamical systems can be predicted through traditional physics-based modeling, there are many real-world examples of systems with unpredictable or chaotic behavior . There are many different approaches for predicting and describing chaotic systems such as physics-based modeling, deep learning, physics-informed machine learning, deep operator inference (DeepONets), and operator inference (OpInf). Physics-based models describing chaotic systems mainly rely on complex partial differential equations (PDEs) that require a tremendous amount of computational power running continuously on high-performance computing (HPC) infrastructure to be solved. Moreover, the systems' non-stationarities, nonlinearities, and intermittency make them intractable from a deterministic standpoint, rendering long-term forecasting unrealistic. Reduced-order models (ROMs) aim to solve the computational complexity issue by creating a low-rank representation of the original system that is capable of accurately approximating its full spatiotemporal evolution. This allows significant improvements in computational speed due to the lower number of variables at stake. ROMs are typically classified into two types: intrusive and non-intrusive. Intrusive ROMs project the high fidelity full order models (FOM) to a low-dimensional subspace through techniques such as proper orthogonal decomposition (POD). They require access to the differential operators of the FOM to compute the reduced operators. This means that the solution is exclusive to the equation studied, thus, nongeneralizable. Non-intrusive methods have no such drawback; they adopt a data-driven approach where snapshots, i.e. measurements of the states of the dynamical system, are used to learn the ROM. Therefore, non-intrusive approaches can be applied in settings where the FOM operators are unavailable such as in the case of proprietary simulation software. An alternative approach for physics-based ROMs is deep learning methods. These methods circumvent the massive online computational requirements for solving PDEs by leveraging the offline training capacity of artificial neural networks with many hardware architectures such as GPUs, TPUs, CPUs, and cloud services. However, one also expects challenges during the offline process such as long training time, availability of RAM, hyperparameter optimization, and scaling for large datasets. Parallelization schemes, such as the one implemented by, can mitigate the scaling problem by training multiple models simultaneously. These parallel schemes for ROMs incorporate two extra hyperparameters associated with the training subset sizes and the communication band between neighboring subsets which makes the convergence to an optimal ROM harder. Parallel models are prone to produce less representative ROMs since they strongly depend on additional hyperparameter tuning. Physics-informed neural networks (PINNs) aim to improve upon traditional neural networks by imposing physical constraints that must be satisfied. These networks are composed of two parts. The first one is a neural network that works as an approximation function that receives time and space coordinates as input providing an approximate solution \(\hat{u}\) of a given PDE. The second part takes the object \(\hat{u}\) and applies automatic differentiation (AD) over it to compute the residual format of the PDE which, together with its corresponding boundary (BC) and initial conditions (IC), are plugged in the neural network loss function. Thus a multi-task learning problem is set through a composite loss function that aims to fit any available data while minimizing the residuals of the PDE, BCs, and ICs. The PDE residual, BC, and IC terms in the loss function act as regularization terms discarding unrealistic solutions and consequently constraining the space of admissible solutions to those that adhere to the imposed physical laws. PINNs have many advantages compared to traditional deep learning: less data is required, a faster training process, and better extrapolation capacity. The downsides, however, are that its architecture is not generalizable due to the physics embedding being unique to the problem considered, together with convergence rate and computational cost issues. Recent advances such as the incorporation of Neural Tangent Kernel approach, tailored architectures and variational formulations represent great strides to overcome the stability and generalization issues of PINNs. Those techniques are key enabling features to make PINNs suitable for a wider range of real-world applications. Deep Operator Networks (DeepONets) are a new class of machine learning frameworks that can learn nonlinear operator mapping between infinite-dimensional Banach spaces. These frameworks are built upon the generalized universal approximation theorem for operators ensuring that nonlinear operators that map input entities to the corresponding latent space solutions of a given PDE can be learned. Although they provide a simple and intuitive model architecture, they require large amounts of high-fidelity data to be trained, whereas the learned operator may not be consistent with the underlying physical laws of the system. Since the output of DeepONets is differentiable with respect to the input coordinates, they also inherit the basic properties of PINNs. Both can be combined in a single physics-informed DeepONet framework. Such a combination can regularize the target output satisfying physical constraints through its loss function, which can lead to higher data efficiency and improved accuracy. The non-intrusive Operator inference (OpInf) technique proposed in is another particularly suitable approach within the broad class of operator learning methods since it offsets some of the disadvantages of intrusive ROMs and deep learning schemes. The non-intrusive nature of OpInf leads to a more general and explainable model compared to intrusive ROMs and standard deep learning architectures while having much fewer hyperparameters and a quite straightforward training procedure. Such features make the method pretty affordable and efficient since it doesn't rely on minimizing loss functions through gradient descent like DeepONets or other kinds of deep neural networks. The method learns the operators by solving a simple least-squares regression problem (more details in section [4](#sec_opinf){reference-type="ref" reference="sec_opinf"}). In a previous work, the authors applied echo state networks (ESN) alongside ROMs to capture the most relevant features of chaotic dynamical systems and analyzed their forecasting capability. The benchmark chaotic systems Lorenz 63/96 and the GEOS Composition Forecasting (GEOS-CF) dataset for atmospheric pollutants over the continental United States were used as examples. For the Lorenz 63 and 96, the ESN showed remarkable forecasting feature up to 11.8 and 22.8 Lyapunov time units respectively. For the GEOS-CF dataset, ESN showed fair results with pointwise errors of about 15% for a target 44-hour time horizon. However, that model was unable to perform long-term recursive predictions, being effective only as a single step ahead predictor. In this case, the ROM could reduce the number of dimensions from \(24,000\) to \(200\) while maintaining 99% of the system's total energy but showed a limited performance due to a lack of any physical driven mechanism, which turned the extrapolation process of the reduced time series quite hard. Such behavior demonstrated that even in reduced space with \(200\) dimensions, a pure data-driven machine learning approach based on ESN can fail in a long-term extrapolation regime. While machine learning approaches deal well with classical chaotic systems, they are still not quite generalizable to more realistic systems. Such empirical evidence clearly showed the need to further explore physics-driven machine learning schemes. Aiming at demonstrating the suitability of physics-inspired ROMs over pure machine learning-based ROMs to forecast canonical chaotic systems , this work focuses on demonstrating the effectiveness of applying OpInf to the classical Lorenz 96 and Kuramoto-Sivashinsky chaotic systems. The main contributions of this work are described below:
- The assessment of long-term forecasting features of Operator Inference (OpInf) technique for benchmark chaotic systems, namely, Lorenz 96 and Kuramoto-Sivashinsky equation.
- An improved parallelization scheme for decomposing the original data matrices used for solving the OpInf least-squares problem in batches.
# Lorenz 96 {#sec:Lorenz}
The Lorenz 96 equations are a 3-tier extension of the original Lorenz 63 model, resulting in three coupled nonlinear ODEs. Such system of equations are shown below: \[\begin{aligned}
\frac{d X_{k}}{d t} &=X_{k-1}\left(X_{k+1}-X_{k-2}\right)+ F-\frac{h c}{b} \, \Sigma_{j} Y_{j, k} \label{eq:1_Lorenz} \\ \frac{d Y_{j, k}}{d t} &=-c \, b \, Y_{j+1, k}\left(Y_{j+2, k}-Y_{j-1, k}\right)-c \, Y_{j, k} + \frac{h c}{b} \, X_{k}-\frac{h e}{d} \, \Sigma_{i} Z_{i, j, k} \label{eq:2_Lorenz} \\ \frac{d Z_{i, j, k}}{d t} &=e \, d \, Z_{i-1, j, k}\left(Z_{i+1, j, k}-Z_{i-2, j, k}\right)-g \, e \, Z_{i, j, k} + \frac{h e}{d} \, Y_{j, k} \label{eq:3_Lorenz} \end{aligned}\] This improved model encompasses additional features to its predecessor Lorenz 63 including a large-scale forcing \(F\) that makes the system highly chaotic and a set of constant coefficients \(b, c, d, e, g, h\) that can be tuned to produce appropriate spatiotemporal variability in all state variables \(X_{k},Y_{j, k}\) and \(Z_{i, j, k}\). Lorenz 96 was designed to model the large-scale behavior of the mid-latitude atmosphere. The numerical setup of this problem is outlined in section [6.1](#res:lor){reference-type="ref" reference="res:lor"}.
# Kuramoto--Sivashinsky equation {#sec:KS}
The Kuramoto--Sivashinsky (KS) equation was developed almost simultaneously by Yoshiki Kuramoto and Gregory Sivashinsky while studying the turbulent state in a chemical reaction system and the hydrodynamic instability in laminar flames. This equation can be written as: \[\frac{\partial u}{\partial t} + u\frac{\partial u}{\partial x} + a\frac{\partial^2 u}{\partial x^2} + b\frac{\partial^4 u}{\partial x^4} = 0 \label{eq_kuramoto}\] The parameters \(a\) and \(b\) in Equation [\[eq_kuramoto\]](#eq_kuramoto){reference-type="ref" reference="eq_kuramoto"} were both chosen as unitary. We considered a smooth initial condition for the KS system, which is shown in Equation [\[ks_boundary\]](#ks_boundary){reference-type="eqref" reference="ks_boundary"}, and periodic boundary conditions that match the periodicity of the initial condition. The boundary conditions are given by Equations [\[periodic1\]](#periodic1){reference-type="eqref" reference="periodic1"}, [\[periodic2\]](#periodic2){reference-type="eqref" reference="periodic2"}, [\[periodic3\]](#periodic3){reference-type="eqref" reference="periodic3"}. \[u(0,x) = cos\left(\frac{\pi x}{20}\right)\left[1+sin\left(\frac{\pi x}{20}\right)\right]\label{ks_boundary}\] \[u(t,0) = u(t,L) \label{periodic1}\] \[u_x(t,0) = u_x(t,L) \label{periodic2}\] \[u_{xx}(t,0) = u_{xx}(t,L) \label{periodic3}\] The numerical setup and parameters used in solving Kuramoto-Sivashinsky's equation are detailed in section [6.2](#res:ks){reference-type="ref" reference="res:ks"}.
# Operator inference {#sec_opinf}
Operator inference (OpInf) provides a non-intrusive way of creating ROMs by learning the physical operators of the FOM in latent space by solving a least-squares optimization problem. Given measurements of the full order model, an approximate output operator that describes the mapping from the full state to the outputs is found. Upon discretization, many important dynamical systems including those considered in this work can be cast into a linear-quadratic system of ODEs as follows: \[\begin{split} \frac{d}{dt}\mathbf{q}(t) =\mathbf{c}+\mathbf{Aq}(t)+\mathbf{H}(\mathbf{q}(t)\otimes \mathbf{q}(t)) + \mathbf{Bu}(t), \\ \mathbf{q}(t_0) = \mathbf{q_0}, t \in [t_0,t_f] \label{eq:1} \end{split}\] where \(\mathbf{q}(t) \in \mathbb{R}^n\) is the state vector at time \(t\) with dimension \(n\), \(\mathbf{q}_0 \in \ \mathbb{R}^n\) is the initial condition, \(\mathbf{A} \in \mathbb{R}^{n \times n}\) is the linear ODE operator, \(\mathbf{H} \in \mathbb{R}^{n \times n^2}\) is the quadratic ODE operator, \(\mathbf{c} \in \mathbb{R}^{n}\) are constant terms, \(\mathbf{u}(t) \in \mathbb{R}^{m}\) are input terms related to BCs or forcing terms, \(\mathbf{B} \in \mathbb{R}^{n \times m}\) is a linear operator in the forcing term \(\mathbf{u(t)}\), and \(\mathbf{q} \otimes \mathbf{q} = [q_1^2,q_1 q_2...,w_1,q_n,q_2q_1,q_2^2...,q_2q_n...,q_n^2] \in \mathbb{R}^{n^2}\) Note that Equation [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} has a polynomial structure. OpInf is especially suited for dynamical systems with polynomial nonlinearities such as Lorenz 96 and KS. For systems with non-polynomial nonlinear operators, one can use lifting maps to transform non-polynomial dynamics into higher-order polynomial dynamics. The OpInf method is built upon a sequence of steps starting by gathering a snapshot matrix \(\mathbf{Q}\) that contains a finite set of measurements of the FOM, as seen in the following: \[\mathbf{Q} = [\mathbf{q}(t_0),\mathbf{q}(t_1)...,\mathbf{q}(t_{f})] \ \in \ \mathbb{R}^{n \times k} \label{eq:2}\] where \(k\) is the total number of snapshots. In the next step, one chooses a low-dimensional basis \(\mathbf{V} \in \mathbb{R}^{n \times r}\), with \(r \ll n\), to represent each vector \(\mathbf{q}(t)\) of the matrix \(\mathbf{Q}\). In this case, \(\mathbf{q}(t) \approx \mathbf{V} \mathbf{q}_{r}(t)\), where \(\mathbf{q}_{r}(t) \in \mathbb{R}^{r}\) is the reduced state vector. This study applied Principal Component Analysis (PCA) to generate the basis \(\mathbf{V}\). A Galerkin projection on Equation [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"} using the basis \(\mathbf{V}\) results in a system of ODEs representing the original dynamics in the reduced order space. If the PDE operators of the FOM are known, it is straightforward to check that the Galerkin projection preserves the polynomial structure of the operators. This observation suggests that even if we do not have access to the FOM operators, we can assume that the reduced model operators have the same shape (i.e., the same polynomial structure) as those in Equation [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"}. Such an assumption is central in the OpInf method and constitutes the third building block of the OpInf ROM. Rewriting the projected snapshots \(\mathbf{q}_{r}(t)\) as \(\widehat{\mathbf{q}}(t)\), the OpInf seeks reduced operators \(\widehat{\mathbf{c}} \sim \mathbf{c}\), \(\widehat{\mathbf{A}} \sim \mathbf{A}\), \(\widehat{\mathbf{H}} \sim \mathbf{H}\), \(\widehat{\mathbf{B}} \sim \mathbf{B}\) that mimic the original full order operators in a non intrusive way, which are shown in Equation [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"} \[\begin{split} \frac{d}{dt}\mathbf{\widehat{q}}(t) =\mathbf{\widehat{c}}+\mathbf{\widehat{A}\widehat{q}}(t)+\mathbf{\widehat{H}}(\mathbf{\widehat{q}}(t)\otimes \mathbf{\widehat{q}}(t)) + \mathbf{\widehat{B}u}(t), \\ \mathbf{\widehat{q}}(t_0) = \mathbf{V}^{T}\mathbf{q_0}, t \in [t_0,t_f] \label{eq:3} \end{split}\] Given the projected state variables \(\mathbf{\widehat{q}}(t)\) and its time derivatives \(\mathbf{\dot{\widehat{q}}}(t)\), calculated with any numerical procedure such as finite differences, the best possible match of the reduced operators with respect to the reduced-order system of ODEs can be found by minimizing the residual of Equation [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}. Thus, the fourth and final building block of the OpInf is to solve a least-squares problem, where the objective function seeks to minimize the residual of the latent dynamics in the Euclidean norm and the objective variables are the reduced operators \(\widehat{\mathbf{c}}, \widehat{\mathbf{A}}, \widehat{\mathbf{H}}, \widehat{\mathbf{B}}\). This problem is ill-conditioned and prone to overfitting. To mitigate this issue, we can rewrite the least-squares problem in a regularized version as shown below [\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"}. \[\begin{split} \min _{\widehat{c}, \widehat{A}, \widehat{\mathbf{H}}, \widehat{\mathbf{B}}} \sum_{j=0}^{k-1}\left\|\widehat{\mathbf{c}}+\widehat{\mathbf{A}} \widehat{\mathbf{q}}_{j}+\widehat{\mathbf{H}}\left(\widehat{\mathbf{q}}_{j} \otimes \widehat{\mathbf{q}}_{j}\right)+\widehat{\mathbf{B}} \mathbf{u}_{j}-\dot{\widehat{\mathbf{q}}}\right\|_{2}^{2}+ \\ \lambda_{1}\|\widehat{\mathbf{c}}\|_{2}^{2}+\lambda_{2}\|\widehat{\mathbf{A}}\|_{F}^{2}+\lambda_{3}\|\widehat{\mathbf{H}}\|_{F}^{2}+\lambda_{4}\|\widehat{\mathbf{B}}\|_{F}^{2} \label{eq:5} \end{split}\] For the sake of conciseness, the regularized regression problem can be cast into a matrix form written as \[\min _{\mathbf{O}}\left\|\mathbf{D O}^{\top}-\mathbf{R}^{\top}\right\|_{F}^{2}+\left\|\mathbf{\Gamma O}^{\top}\right\|_{F}^{2} \label{eq:6}\] where \[\begin{aligned} &\mathbf{O}=\left[\begin{array}{llll} \widehat{\mathbf{c}} & \widehat{\mathbf{A}} & \widehat{\mathbf{H}} & \widehat{\mathbf{B}} \end{array}\right] \in \mathbb{R}^{r \times d(r, m)} \\ &\mathbf{D}=\left[\begin{array}{llll} \mathbf{1}_{k} & \widehat{\mathbf{Q}}^{\top} & (\widehat{\mathbf{Q}} \otimes \widehat{\mathbf{Q}})^{\top} & \mathbf{U}^{\top} \end{array}\right] \in \mathbb{R}^{k \times d(r, m)} \\ &\widehat{\mathbf{Q}}=\left[\begin{array}{llll} \widehat{\mathbf{q}}_{0} & \widehat{\mathbf{q}}_{1} & \cdots & \widehat{\mathbf{q}}_{k-1} \end{array}\right] \in \mathbb{R}^{r \times k} \\ &\mathbf{R}=\left[\begin{array}{llll} \dot{\widehat{\mathbf{q}}}_{0} & \dot{\widehat{\mathbf{q}}}_{1} & \cdots & \dot{\widehat{\mathbf{q}}}_{k-1} \end{array}\right] \in \mathbb{R}^{r \times k} \\ &\mathbf{U}=\left[\begin{array}{llll} \mathbf{u}_{0} & \mathbf{u}_{1} & \cdots & \mathbf{u}_{k-1} \end{array}\right] \in \mathbb{R}^{m \times k} \\ &\mathbf{\Gamma}= \lambda \mathbf{I} = \Lambda(\lambda_1,\lambda_2,\lambda_3,\lambda_4), \ \lambda > 0. \end{aligned}\] In this set of equations \(d(r,m) = 1 + r + \binom{r+1}{2} + m\) and \(\mathbf{1}_{k} \in \mathbb{R}^{k}\) is a column vector of length \(k\) with unity entries. We notice that in Equation [\[eq:6\]](#eq:6){reference-type="eqref" reference="eq:6"}, \(\mathbf{O}\) contain the unknown operators, \(\mathbf{D}\) are known data values, \(\widehat{\mathbf{Q}}\) are the projected state variables, \(\mathbf{R}\) are the time derivatives of the projected state variables, \(\mathbf{U}\) represents possible forcing terms, and \(\mathbf{\Gamma}\) is a diagonal regularizer. The minimizer of the regression problem in Equation [\[eq:6\]](#eq:6){reference-type="eqref" reference="eq:6"} satisfies the modified normal equations: \[\left(\mathbf{D}^{\top} \mathbf{D}+\mathbf{\Gamma}^{\top} \boldsymbol{\Gamma}\right) \mathbf{O}^{\top}=\mathbf{D}^{\top} \mathbf{R}^{\top} \label{eq:8}\] which constitutes an algebraic linear system that can be solved by a variety of efficient algorithms. Furthermore, the optimization problem can also be transformed into \(n\) independent least-squares problems and solved efficiently using standard solvers in parallel architectures. The most computationally expensive operation in the closed-form construction stage (cf. Equation [\[eq:8\]](#eq:8){reference-type="eqref" reference="eq:8"}) is the evaluation of \(\mathbf{D}^T\,\mathbf{D}\). Matrix \(\mathbf{D}\) is usually dense and, depending on the number of samples and degrees of freedom required to fit the model, it becomes challenging to allocate such matrix operations to a single computational node. By splitting the inner product operation into batch-wise operations along the samples axis and dispatching them in multiple MPI processes, we can rely on huge amounts of data for creating \(\mathbf{D}\) without suffering memory issues. Appendix [\[parallel_eval\]](#parallel_eval){reference-type="ref" reference="parallel_eval"} shows a pseudocode detailing this batch-wise approach. Alternatively, one can compute a solution for the algebraic linear system by using the Moore-Penrose pseudoinverse \(\mathbf{O}^{\top}=\mathbf{D}^\dagger\mathbf{R}^{\top}\) in which no regularization parameter is necessary. When OpInf learns the reduced operators, it indirectly considers information coming from the original PDE and, thus, learns the underlying Physics of the dynamical system without having access to the full order operators showed in Equation [\[eq:1\]](#eq:1){reference-type="eqref" reference="eq:1"}. The incorporation of physics principles into the machine learning pipeline is the main motivation for choosing OpInf as a chaotic dynamical system forecasting tool.
# Quantifying model performance
To quantify a system's chaoticity, it is common to study its sensitivity to initial conditions. Consider two initial conditions represented by the vectors \(x(t)\) and \(x(t)+\delta(t)\), where \(\delta(t)\) is an infinitesimally small displacement. At time \(t=0\), the distance between both trajectories is set as \(\delta(t=0)=\delta_0\). For chaotic systems, trajectories start diverging exponentially in time. Such divergence can be measured through the following mathematical relation: \[\|\delta(t)\| \approx \|\delta_0\| \ e^{\Lambda t} \label{1}\] The variable \(\Lambda\) in Equation [\[1\]](#1){reference-type="eqref" reference="1"} is referred to as the Lyapunov exponent and it characterizes the stability of the system. If the Lyapunov exponent is positive, it means the system is chaotic while negative values represent stability. Considering multi-dimensional systems, typically, there are as many Lyapunov exponents as there are dimensions. If at least one positive Lyapunov exponent exists, the system is chaotic, however, the rate of divergence can be qualitatively observed by the magnitude of these exponents. The value of \(\Lambda\) can also be different depending on the orientation of the displacement vector \(\delta(t)\), given by \(\delta(t)/|\delta(t)|\). Therefore, the possible values of \(\Lambda\) defines a spectrum \(\Lambda_1 \geq \Lambda_2 \geq \Lambda_3 \geq... \geq \Lambda_n\) with \(n\) being the dimensionality of the phase space. Note that when \(\Lambda\)'s with positive values are dominant, we observe faster divergence among trajectories. From Equation [\[1\]](#1){reference-type="eqref" reference="1"}, the largest positive value of \(\Lambda\) causes the largest divergence, which is called the Maximal Lyapunov exponent (MLE) denoted by \(\Lambda_1\). The MLE has two main purposes:
- To measure the level of unpredictability of the dynamical system;
- Provide a characteristic time scale to quantify the quality of predictions based on the forecasting error growth (Lyapunov time).
In this work, we scale the time axis of all plots in numerical results by the largest Lyapunov exponent \(\Lambda_1\) in order to characterize the chaoticity of the system. This is done by multiplying the time vector \(t\) by the MLE of the model (\(T^{\Lambda_1} = t \Lambda_1\)). Aiming at establishing a straightforward comparison with methods based on backpropagation and reservoir computing RNN architectures showed in, we considered the same setups and geometries used in that paper. Thus, we considered the Lyapunov exponent of the Kuramoto-Sivashinsky problem as \(\Lambda_1 = 0.094\) while for Lorenz-96 system we fixed \(\Lambda_1=1.68\) for \(F=8\) and \(\Lambda_1=2.27\) for \(F=10\). As suggested in, this work also considers the valid prediction time \(\mathrm{(VPT)}\) to quantify the model's predictive performance through a single metric. The VPT is computed in terms of the system's MLE (\(\Lambda_1\)) using the following definition \[\mathrm{VPT}=\frac{1}{\Lambda_{1}} \underset{t_{f}}{\operatorname{argmax}}\left\{t_{f} \mid \operatorname{NRMSE}\left(\boldsymbol{o}_{t}\right)<\epsilon, \forall t \leq t_{f}\right\}. \label{2}\] Textually, \(\mathrm{VPT}\) finds the largest time \(t_f\) that the model can forecast with a normalized root mean square error (NRMSE) of the observable time-series \(o_t\) smaller than a threshold \(\epsilon\) and normalizes it using \(\Lambda_1\). The value of \(\epsilon\) is set to 0.5 to allow comparisons with. \(\mathrm{NRMSE}\) is computed as follows \[\label{eq:nrmse} \mathrm{NRMSE}(\mathbf{\widehat{x}}) = \sqrt{\frac{\sum^{N-1}_{i=0} (\mathbf{x}_{i}-\mathbf{\widehat{x}}_{i})^2}{N\boldsymbol{\sigma}^2}},\] where \(\mathbf{x}_{i}\) and \(\mathbf{\widehat{x}}_{i}\) identify the reference and predicted time-series, respectively, \(N\) is the total number of time-series considered (possibly in latent space), and \(\boldsymbol{\sigma}\) is the array of standard deviations over time for all time-series. The division in Equation [\[eq:nrmse\]](#eq:nrmse){reference-type="ref" reference="eq:nrmse"} should be understood as an element-wise operation, that is, for each time-series, there is a single \(\sigma_{i}\).
# Numerical results
In this section, we assess the effectiveness of the OpInf method in creating a reduced-order model through two benchmark examples: the Lorenz 96 and the Kuramoto-Sivashinsky systems. The former is a suitable emulator for complex multi-scale atmospheric systems. The latter shows rich dynamical characteristics with many bifurcations depending on the size \(L\) of the spatial domain culminating in strong chaotic regimes.
## Lorenz 96 {#res:lor}
The Lorenz 96 numerical experiment is performed using \(N=40\) degrees of freedom for the interval \(t = [0, 2000]\,s\) with a timestep \(dt=0.01\,s\). The forcing term \(F\) is chosen as either \(F=8\) or \(F=10\). Assuming periodic boundary conditions and starting from a randomly generated initial condition, the problem is solved by employing SciPy's LSODA algorithm, a solver that automatically switches between a nonstiff Adams solver and a Backward Differentiation Formula method for stiff problems. The first \(1000\,s\) are discarded to get rid of the initial transient behavior, and the remaining \(1000\,s\) are split into two sets of equal size corresponding to the training and testing datasets. Since the Lorenz 96 system is randomly initialized, the simulation is repeated \(100\) times for each forcing term to analyze the sensitivity of the reduced model with respect to the randomized ICs. The OpInf reduced models are obtained by solving a least-squares problem via Moore-Penrose pseudoinverse for each random initialization and subsequently used for computing extrapolations. To evaluate the quality of each model, we employ the VPT criteria. For each simulation set, corresponding to a different forcing term \(F\), the VPT values are recorded and their minimum, maximum, mean, and standard deviation values are showed in Table [1](#tab:VPT){reference-type="ref" reference="tab:VPT"}. In this table \(\sigma^{*}\) is computed over all approximate solutions associated to each IC. The VPT threshold value was fixed as \(0.5\).
[\[tab:VPT\]]{#tab:VPT label="tab:VPT"}
The left-hand side of Figs. [\[contour_lorenz_96_F=8\]](#contour_lorenz_96_F=8){reference-type="ref" reference="contour_lorenz_96_F=8"} and [\[contour_lorenz_96_F=10\]](#contour_lorenz_96_F=10){reference-type="ref" reference="contour_lorenz_96_F=10"} compare the approximate and target contour plots of the spatiotemporal forecast of the Lorenz 96 chaotic system for the configurations \(F=8\) and \(F=10\), respectively. The right-hand side of the same Figs. show the normalized root square error (NRSE) associated to each one of the forcing configurations. The OpInf results correspond to the best approximation among all random ICs. Fig. [\[NRMSE_F=8_F=10\]](#NRMSE_F=8_F=10){reference-type="ref" reference="NRMSE_F=8_F=10"} shows the evolution of the mean \(NRMSE\) along time for both configurations of the forcing term, i.e., \(F=8\) (Fig. [\[NRMSE_F=8_F=10\]](#NRMSE_F=8_F=10){reference-type="ref" reference="NRMSE_F=8_F=10"}a) and \(F=10\) (Fig. [\[NRMSE_F=8_F=10\]](#NRMSE_F=8_F=10){reference-type="ref" reference="NRMSE_F=8_F=10"}b) considering all random ICs. The plots also depict the standard deviation envelope around the average \(NRMSE\) which is computed as \(\overline{NRMSE} \pm \sigma(NRMSE)\). These outcomes indicate a clear prediction improvement of the OpInf method over the backpropagation and reservoir computing RNN architectures considered in. For instance, comparing the average \(NRMSE\) for F=8 and F=10 reported in with the ones in Fig. [\[NRMSE_F=8_F=10\]](#NRMSE_F=8_F=10){reference-type="ref" reference="NRMSE_F=8_F=10"}, their best forecasting model reached VPT at \(0.79\) and \(0.83\) Lyapunov time units (cf. Table 1 in ) respectively, while OpInf did the same at \(4.42\) and \(2.32\) Lyapunov time units (Table [1](#tab:VPT){reference-type="ref" reference="tab:VPT"}). Note that if we consider the standard deviation envelope, the worst OpInf results for F=8 is still better than the best \(NRMSE\) in while for F=10, the worse OpInf results is better than the mean \(NRMSE\) in. Fig. [\[lorenz_96_modes_F=8,10\]](#lorenz_96_modes_F=8,10){reference-type="ref" reference="lorenz_96_modes_F=8,10"} shows the predictive power of OpInf for five different discrete variables of the Lorenz 96 model with \(F=8\) and \(F=10\), respectively. The plots compare the approximate integrated solution with the ground truth over time for the best approximate solution among all random ICs. We can observe from those plots that the OpInf performed remarkably for both \(F=8\) and \(F=10\) configurations matching the exact solutions for as long as \(6.66\) and \(3.19\) Lyapunov time units, respectively.
## Kuramoto-Sivashinsky {#res:ks}
The Kuramoto-Sivashinsky (KS) equation is defined over a domain \(\Omega = [0,L]\) with periodic boundary conditions. We chose the domain size \(L = 200 \,m\) and discretized it with \(512\) equally spaced grid points. The KS equation marches in time for a total \(T = 6 \cdot 10^4 \,s\) with a time step \(dt = 0.125\,s\) corresponding to \(48 \cdot 10^4\) steps. The reference solution of the KS equation is generated using the modified exponential time-differencing fourth-order Runge-Kutta scheme (ETDRK4) as described in. The ETDRK4 scheme combines a modified version of the exponential time-differencing (ETD) scheme with the fourth-order Runge-Kutta (RK4) time integrator. Time-stepping is performed with sequential FFT analysis where the number of waves considered is a function of the number of grid points. We employ the Principal Component Analysis (PCA) to reduce the dimensionality of the full order model from \(512\), which corresponds to the number of grid points, to \(r=160\) dimensions that match the number of reduced basis vectors in \(\mathbf{V}\). The final shape of the spatiotemporal dataset used to train and test the OpInf model is set to \((t,x) = (4,8\,\times 10^5, 160)\). The PCA decomposition is applied to \(90\%\) of the initial \((t,x)\) snapshots and the choice of \(k\) is guided by the projection error evaluated over the remaining \(10\%\) of the dataset. The projection error obtained in this region is approximately \(0.51 \%\) and the expected variance ratio higher than \(0.9999\) with respect to the \(90\%\) portion of the initial snapshots. Fig. [\[cumul_energy_PCA\]](#cumul_energy_PCA){reference-type="ref" reference="cumul_energy_PCA"} shows the cumulative explained variance ratio (or the energy content) preserved by the PCA decomposition as a function of the number of modes. The modeling dataset is split into three subsets with proportions (90%, 5%, 5%) corresponding to the training, validation and test datasets, respectively. The OpInf reduced model is generated considering a simple grid-search for the \(log_{10}\) of the Thikonov regularization penalties \((\lambda_2, \lambda_3)\) for the discrete domain \(\left\{(\lambda_2,\lambda_3);-3 \leq \lambda_2 \leq 3,-3 \leq \lambda_3 \leq 3\right\}\) with a \(0.5\) step for both \(\lambda_2\) and \(\lambda_3\). For this specific case we found the combination \((\lambda_2,\lambda_3) = (10^{-3}, 10^{-3})\) as the best option for minimizing the extrapolation error in the validation dataset. For this particular case, the regularizers \(\lambda_1\) and \(\lambda_4\) are set to zero since the constant polynomial term \(\mathbf{\widehat{c}}\) (cf. Equation [\[eq:3\]](#eq:3){reference-type="eqref" reference="eq:3"}) has minor influence over the final results and the KS system has no forcing term to fit the operator \(\mathbf{\widehat{B}}\). Once OpInf operators are obtained, they are used for extrapolating the state variable \(\widehat{\mathbf{q}}(t)\) over an unseen dataset. In order to evaluate the model sensibility to different initial conditions, we proceed by time-integrating the reduced model with the SciPy routine *odeint* using a set of 100 initial conditions randomly selected from the test dataset. We uses the same time horizon of 90 s (empirically chosen) to assess the model extrapolation for each IC. Table [2](#tab:VPT_ks){reference-type="ref" reference="tab:VPT_ks"} shows minimum, maximum, mean, and standard deviation VPT values for the KS system. In this table, \(\sigma^{*}\) is computed over all approximate solutions associated to each IC. The VPT threshold in this case was fixed as \(0.5\).
[\[tab:VPT_ks\]]{#tab:VPT_ks label="tab:VPT_ks"}
The extrapolated solutions are then reconstructed into the full dimensional space through the PCA basis \(\mathbf{V}\). The relative approximation error in \(L^2\)-norm achieved by the best reconstructed solution among all time-integrated solutions associated to each IC is \(25.8 \%\) for the time horizon of \(90 s\). Considering only the first \(75 s\) in the test dataset comprising the \(90 s\), the relative approximation error drops to \(1.2 \%\). The VPT metric associated to the best case gets the impressive value of \(794\). The NRMSE is evaluated for the \(160\) latent time-series \(\mathbf{\widehat{x}}_{i}\) referenced in Equation [\[eq:nrmse\]](#eq:nrmse){reference-type="ref" reference="eq:nrmse"}. A summary of the VPT results for this numerical experiment can be seen in Figure [\[NRMSE_KS\]](#NRMSE_KS){reference-type="ref" reference="NRMSE_KS"}. Comparing OpInf performance for the KS system with the results in, OpInf is way superior with a forecasting capability two orders of magnitude larger. We should also notice that the OpInf performed remarkably well with a reduced number of dimensions (i.e., using \(160\) eigenmodes). In contrast, all the network architectures analyzed in produced forecasts in the full order space. We still observe that OpInf also outperforms the Markov neural operators (MNO) as proposed in in terms of accuracy and stability. MNO keeps up with the exact trajectory for the KS system until \(t=50s\) while OpInf exceeds \(t=70s\). Fig. [\[KS_approx_vs_exact\]](#KS_approx_vs_exact){reference-type="ref" reference="KS_approx_vs_exact"} compares the contour plots of one of the best OpInf approximate solutions (considering all the solutions corresponding to each initial condition) with the ground truth solution of the KS chaotic system. The OpInf ROM can accurately predict the system dynamics for more than \(744\) Lyapunov time units (cf. black dashed lines) which is far beyond the results achieved in. Fig. [\[KS_error\]](#KS_error){reference-type="ref" reference="KS_error"} shows the spatiotemporal contour plots of the approximation errors of the OpInf ROM based on two distinct metrics: a) pointwise error and b) NRSE. The white dashed lines point out how far the reduced model can keep errors at an extremely low level. Such an outstanding performance highlights how physics inspired machine learning architectures can outperform deep learning state-of-the-art models such as backpropagation and reservoir computing RNN architectures in extrapolation regime. Fig. [\[ks_modes\]](#ks_modes){reference-type="ref" reference="ks_modes"} shows the comparison between approximate and reference temporal eigenmodes for six distinct integrated time-series in the KS system latent space. The plots evidence the capacity of OpInf ROM to match the reference eigenmodes during long-term forecasting even for the lowest energetic modes (e.g., eigenvectors 120 and 159). We notice in Fig. [\[ks_modes\]](#ks_modes){reference-type="ref" reference="ks_modes"} that the latent space representation of the KS data produces an increasing oscillatory pattern as the eigenmodes become less and less energetic. In contrast, the Lorenz 96 system with \(F=10\), for instance, displays dominant high frequency oscillations in every temporal eigenmode since it is represented in the full order space (see Fig. [\[lorenz_96_modes_F=8,10\]](#lorenz_96_modes_F=8,10){reference-type="ref" reference="lorenz_96_modes_F=8,10"}). We conjecture that the regularized representation of the PCA decomposition, specially for the most energetic modes, is likely the reason why Lorenz 96 forecasting results did not perform as remarkably as those for the KS system. As a final assessment, we check the physical consistency of the OpInf ROM. We select a point in the state-space of the ROM where the predicted solution and the ground truth already diverged (\(\approx\) \(750\) Lyapunov times in Fig. [\[KS_error\]](#KS_error){reference-type="ref" reference="KS_error"}) and use it as an initial condition for the discretized KS equation (i.e., the reference system). We then integrate the discretized equation in time for \(20 s.\) (\(\approx 212\) Lyapunov times) and compare the output with the OpInf solution. We call the time-integrated solution based on the reference system the restarted solution. Fig. [\[KS_restarted_vs_predicted\]](#KS_restarted_vs_predicted){reference-type="ref" reference="KS_restarted_vs_predicted"} shows the spatiotemporal representation of the restarted and OpInf solutions together with the absolute pointwise error between both solutions. We observe an excellent agreement between both KS systems' answers, with errors kept very small on most of the \((x,t)\) domain except for a few dark blue and red stripes. Fig. [\[tseries_restarted_vs_opinf\]](#tseries_restarted_vs_opinf){reference-type="ref" reference="tseries_restarted_vs_opinf"} compares the restarted and the OpInf ROM solutions at six regularly spaced grid points. These plots correspond to horizontal slices (time-series) of the contour plots in Fig. [\[KS_restarted_vs_predicted\]](#KS_restarted_vs_predicted){reference-type="ref" reference="KS_restarted_vs_predicted"}. Comparing the restarted and the OpInf solutions on these plots, we confirm a close similarity between them, especially at the boundaries of the KS domain \(u(0,t)\) and \(u(L,t)\) (cf. first and last plots from top to bottom in Fig. [\[tseries_restarted_vs_opinf\]](#tseries_restarted_vs_opinf){reference-type="ref" reference="tseries_restarted_vs_opinf"}). This result indicates that OpInf is physically meaningful since its outputs are consistent with the state-space generated by the reference system including its boundary conditions.
# Conclusion
In this paper we develop surrogate models based on the Non-intrusive Reduced Order Modeling via Operator Inference (OpInf) method to simulate benchmark chaotic systems, namely, Lorenz 96 and Kuramoto-Sivashinsky (KS) equation. We also propose a parallel scheme for decomposing the matrix product \(\mathbf{D}^T\,\mathbf{D}\) on the left hand side of Equation [\[eq:8\]](#eq:8){reference-type="eqref" reference="eq:8"} using a batch-wise approach as detailed in Appendix [\[parallel_eval\]](#parallel_eval){reference-type="ref" reference="parallel_eval"}. OpInf shows remarkable forecasting capabilities for both dynamical systems with two setup configurations for Lorenz 96 and one for KS system. We observe robust time-series predictions in latent space outperforming by far state-of-the-art machine learning techniques such as ESN-RC, Unitary cell, LSTM, and GRU analyzed in. OpInf performance also excels in the MNO method based on Fourier Neural Operators (FNO) being \(40\%\) more accurate in representing the KS system trajectory. It is worth noticing that for Lorenz 96 system with 40 degrees of freedom, OpInf reached VPT at \(4.42\) and \(2.32\) Lyapunov time units on average for forcing terms \(F=8\) and \(F=10\), respectively, while the best forecasting model in topped at \(0.79\) and \(0.83\) Lyapunov time units. Regarding the KS system, OpInf performs even better matching ground truth simulations for as long as \(520\) Lyapunov time units on average using \(160\) degrees of freedom. Referring to Fig. 12a in we notice that their best surrogate model defined in full space dimension (\(512\) degrees of freedom) reached \(5\) Lyapunov time units as VPT. Numerical experiments for the KS system also revealed that OpInf ROM produces physically meaningful solutions consistent with the state-space generated by the discretized KS equation, as depicted in Figs. [\[KS_restarted_vs_predicted\]](#KS_restarted_vs_predicted){reference-type="ref" reference="KS_restarted_vs_predicted"} and [\[tseries_restarted_vs_opinf\]](#tseries_restarted_vs_opinf){reference-type="ref" reference="tseries_restarted_vs_opinf"}. In this case, the OpInf dynamics show close correspondence with the solution generated by the reference system using an OpInf unpaired physical state as the initial condition. We confirm that OpInf has many attractive features. It is non-intrusive, generalizable (in the sense that it can approximate a large class of PDE operators with linear-quadratic structure), explainable, physics-informed, and computationally efficient. Moreover, it is based on a sound mathematical background (Koopman's theory), does not depend on gradient descent algorithm, uses standard linear algebra tools, and needs only a few regularizing parameters. The regularization strategy in the least-squares problem is a key ingredient for the OpInf success but it is quite tricky. The Tikhonov-based regularization method used in OpInf is obtained by means of a grid search over the space of the regularization parameters which makes the process of finding a global minimum pretty challenging. Keeping time integration errors bounded is another crucial point which we address by using SciPy's LSODA algorithm. As we are interested in tackling turbulent flows in the future, we plan to find ways to build closure mechanisms for the OpInf ROM in latent space aiming at capturing the energy of hidden scales smeared by the PCA decomposition. Closure mechanisms could be straightforwardly plugged into the OpInf least-squares problem through the \(\mathbf{\widehat{B}}\) forcing operator (cf. Equation [\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"}) producing a ROM closure. We also want to address the limitations of the OpInf grid search regularization technique by exploring convex and nonconvex optimization approaches that best fit the regularization parameters by means of approximations and exploration of equivalent formulations. | {'timestamp': '2022-06-07T02:31:17', 'yymm': '2206', 'arxiv_id': '2206.01604', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01604'} |
null | null |
# Introduction {#sec:intro}
Neural ordinary differential equations (NODEs) have opened a new perspective on continuous-time computation with neural networks (NNs) as a practical framework for machine learning based on differential equations. While the original approach---proposed as a continuous-depth version of deep feed-forward residual NNs---only covers autonomous ODEs entirely determined by the initial conditions, more recent extensions deal with sequential data (reviewed in Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}) in a way similar to what is typically done with standard recurrent NNs (RNNs) in the discrete-time scenario. This potential for continuous-time (CT) sequence processing (CTSP) is particularly interesting, since there are many applications where datapoints are observed at irregularly spaced time steps, and CT sequence models might better deal with such data than their discrete-time counterparts. However, the development of NODEs for CTSP is still at an early stage. For example, a popular approach of Neural Controlled Differential Equations (NCDEs; also reviewed in Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}) has in practice only one architectural variant corresponding to the "vanilla" RNN. Discrete-time processing, however, exploits many different RNN architectures as well as Transformers. While it is not straightforward to transform the standard Transformer into a CT sequence processor, we'll show that the closely related Fast Weight Programmers (FWPs) and linear Transformers (reviewed in Sec. [2.3](#sec:back_fwp){reference-type="ref" reference="sec:back_fwp"}) have direct CT counterparts. In FWPs, temporal processing of short-term memory (stored in fast weight matrices) uses learnable sequences of *learning rules*. Hence CT versions of FWPs will require differential equations to model the learning rules. This relates to an ancient trend of the 1980s/90s. Among many old connections between NNs and dynamical systems described by ODEs (e.g., ), the theoretical analysis of NN learning rules in the ODE framework has been particularly fruitful. Consider the famous example of Oja's rule (briefly reviewed in Sec. [2.2](#sec:back_lr){reference-type="ref" reference="sec:back_lr"}): many results on its stability, convergence, and connection to Principle Component Analysis were obtained using its ODE counterpart (e.g., ). Here we propose a novel combination of Neural ODEs and learning rules, to obtain a new class of sequence processing Neural ODEs which are continuous-time counterparts of Fast Weight Programmers and linear Transformers. To the best of our knowledge, there is no previous work on Neural ODE-based Transformer families, despite their dominance in important types of discrete time computations such as Natural Language Processing and beyond. We also show how our approach solves the fundamental limitation of existing Neural CDEs in terms of model size scalability. We conduct experiments on three standard time series classification tasks covering various scenarios (regularly sampled, irregularly sampled with missing values, and very long time series). We demonstrate that our novel models outperform existing Neural ODE-based sequence processors, in some cases by a large margin.
# Background {#sec:background}
We briefly review the main background concepts this work builds upon: NODEs for sequence processing (Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}), NN learning rules and their connection to ODEs (Sec. [2.2](#sec:back_lr){reference-type="ref" reference="sec:back_lr"}), and Fast Weight Programmers whose memory update is based on learning rules controlled by an NN (Sec. [2.3](#sec:back_fwp){reference-type="ref" reference="sec:back_fwp"}).
## Neural ODEs (NODEs) and Their Extensions for Sequence Processing {#sec:back_ode}
Here we review the core idea of NODEs. In what follows, let \(n\), \(N\), \(d\), \(d_\text{in}\) denote positive integers, \(T\) be a positive real number, and \(\theta\) denote an arbitrary set of real numbers. The first step is to consider a residual layer (say, the \(n\)-th layer with a dimension \(d\)) in an \(N\)-layer deep NN which transforms an input \({\bm{h}}_{n-1} \in \mathbb{R}^{d}\) to an output \({\bm{h}}_{n} \in \mathbb{R}^{d}\) with a parameterised function \({\bm{f}}_\theta: \mathbb{R}^{d} \rightarrow \mathbb{R}^{d}\) as follows: \[\begin{aligned}
\label{eq:resnet} {\bm{h}}_n = {\bm{h}}_{n-1} + {\bm{f}}_\theta({\bm{h}}_{n-1}) \end{aligned}\] This coincides with the following equation for \(\epsilon=1\) \[\begin{aligned}
{\bm{h}}({t_n}) = {\bm{h}}({t_{n-1}}) + \epsilon {\bm{f}}_\theta({\bm{h}}({t_{n-1}})) \end{aligned}\] where \({\bm{h}}: [0, T] \rightarrow \mathbb{R}^d\) is a function such that \({\bm{h}}(t_n) = {\bm{h}}_n\) holds for all \(n: 0\leq n \leq N\) and \(t_n \in [0, T]\) such that \(t_n-t_{n-1} = \epsilon > 0\) if \(n \geq 1\). This equation is a forward Euler discretisation of the ordinary differential equation defined for all \(t \in (t_0, T]\) as \[\begin{aligned}
{\bm{h}}'(t) = {\bm{f}}_\theta({\bm{h}}(t)) \quad \text{or} \quad {\bm{h}}(t) = {\bm{h}}(t_0) + \int_{s=t_0}^t {\bm{f}}_\theta({\bm{h}}(s))ds \end{aligned}\] where \({\bm{h}}'\) denotes the first order derivative. This establishes the connection between the ODE and the deep residual net with parameters \(\theta\) shared across layers[^1]: given the initial condition \({\bm{h}}(t_0)={\bm{h}}_0\), the solution to this equation evaluated at time \(T\), i.e., \({\bm{h}}(T)\), corresponds to the output of this deep residual NN, which can be computed by an ODE solver. We denote it as a function \(\mathrm{ODESolve}\) taking four variables: \({\bm{h}}(T) = \mathrm{ODESolve}({\bm{f}}_\theta, {\bm{h}}_0, t_0, T)\). During training, instead of backpropagating though the ODE solver's operations, the *adjoint sensitivity method* (which essentially solves another ODE but backward in time) can compute gradients with \(O(d)\) memory requirement, constant w.r.t. \(T\). A natural next step is to extend this formulation for RNNs, i.e., the index \(n\) now denotes the time step, and we assume an external input \({\bm{x}}_n \in \mathbb{R}^{d_\text{in}}\) at each step \(n\) to update the hidden state \({\bm{h}}_{n-1}\) to \({\bm{h}}_n\) as \[\begin{aligned}
\label{eq:rnn} {\bm{h}}_n = {\bm{f}}_\theta({\bm{h}}_{n-1}, {\bm{x}}_n) \end{aligned}\] Depending on the property of external inputs \(({\bm{x}}_n)_{n=1}^N = ({\bm{x}}_1..., {\bm{x}}_N)\), there are different ways of defining NODEs for sequence processing. We mainly distinguish three cases. **First**, when there is a possibility to construct a *differentiable* control signal \({\bm{x}}:t\mapsto {\bm{x}}(t) \in \mathbb{R}^{d_\text{in}}\) for \(t \in [t_0, T]\) from the inputs \(({\bm{x}}_n)_{n=1}^N\); an attractive approach by handles the corresponding dynamics in a *neural controlled differential equation* (NCDE): \[\begin{aligned}
\label{eq:cde} {\bm{h}}(t) = {\bm{h}}(t_0) + \int_{s=t_0}^t {\bm{F}}_\theta({\bm{h}}(s))d{\bm{x}}(s) = {\bm{h}}(t_0) + \int_{s=t_0}^t {\bm{F}}_\theta({\bm{h}}(s)){\bm{x}}'(s)ds \end{aligned}\] where \({\bm{F}}_\theta\) is a parameterised function (typically a few-layer NN) which maps a vector \({\bm{h}}(s) \in \mathbb{R}^d\) to a matrix \({\bm{F}}_\theta({\bm{h}}(s)) \in \mathbb{R}^{d \times d_{\text{in}}}\) (we'll discuss how this component already relates to Fast Weight Programmers) and thus, \({\bm{F}}_\theta({\bm{h}}(s))d{\bm{x}}(s)\) denotes a matrix-vector multiplication. The *control* \({\bm{x}}: [t_0, T] \rightarrow \mathbb{R}^{d_\text{in}}\) is typically constructed via natural cubic spline over all data points \(({\bm{x}}_n)_{n=1}^N\) such that its differentiability is guaranteed (which, however, makes it incompatible with auto-regressive processing ). Since the final equation is again an NODE with a vector field of form \({\bm{g}}_{\theta, {\bm{x}}'}(s, {\bm{h}}(s)) = {\bm{F}}_\theta({\bm{h}}(s)){\bm{x}}'(s)\), all methods described above are applicable: ODE solver for evaluation and adjoint method for memory efficient training. A notable extension of Neural CDEs is the use of *log-signatures* to sub-sample the input sequence. Resulting NCDEs are called *neural rough differential equations* (NRDEs), which are relevant for processing long sequences. On a side note, the NCDE is often referred to as the "continuous-time analogue" to RNNs, but this is a bit misleading: discrete-time RNN equations corresponding to the continuous-time Eq. [\[eq:cde\]](#eq:cde){reference-type="ref" reference="eq:cde"} do not reflect the standard RNN of Eq. [\[eq:rnn\]](#eq:rnn){reference-type="ref" reference="eq:rnn"} but: \[\begin{aligned}
\label{eq:cde_discrete_1} {\bm{h}}_n &= {\bm{h}}_{n-1} + {\bm{W}}_{n-1} ({\bm{x}}_n-{\bm{x}}_{n-1}) \\ \label{eq:cde_discrete_2} {\bm{W}}_{n} & = {\bm{F}}_\theta({\bm{h}}_{n}) \end{aligned}\] where one network (Eq. [\[eq:cde_discrete_1\]](#eq:cde_discrete_1){reference-type="ref" reference="eq:cde_discrete_1"}) learns to translate the variation of inputs \(({\bm{x}}_n-{\bm{x}}_{n-1})\) into a change in the state space using a weight matrix \({\bm{W}}_{n-1}\) which itself is generated by another network (\({\bm{F}}_\theta: \mathbb{R}^{d} \rightarrow \mathbb{R}^{d \times d_\text{in}}\); Eq. [\[eq:cde_discrete_2\]](#eq:cde_discrete_2){reference-type="ref" reference="eq:cde_discrete_2"}) on the fly from the hidden state. This model is thus a kind of Recurrent FWP. **Second**, even if \({\bm{x}}\) is not differentiable, having access to \({\bm{x}}\) defined and bounded over an interval of interest \([t_0, T]\) is enough to define a sequence processing NODE, by making it part of the vector field: \[\begin{aligned}
\label{eq:direct-ode} {\bm{h}}(t) = {\bm{h}}(t_0) + \int_{s=t_0}^t {\bm{f}}_\theta({\bm{h}}(s), {\bm{x}}(s))ds \end{aligned}\] where the vector field \({\bm{f}}_\theta({\bm{h}}(t), {\bm{x}}(t)) = {\bm{g}}_{\theta, {\bm{x}}}(t, {\bm{h}}(t))\) can effectively be evaluated at any time \(t \in [t_0, T]\). We refer to this second approach as a *direct NODE* method. While theoretically and empirically show that this approach is less expressive than the NCDEs above, we'll show how in our case of learning rules one can derive interesting models within this framework, which empirically perform on par with the CDE variants. **Finally**, when no control function is available (i.e., we only have access to \({\bm{x}}_n\) at discrete time steps \((t_n)_{n=1}^N\)), a mainstream approach consists in dissociating the continuous-time hidden state update via ODE for the time between two observations (e.g., Eq. [\[eq:ode-rnn-ode\]](#eq:ode-rnn-ode){reference-type="ref" reference="eq:ode-rnn-ode"} below) from integration of the new data (Eq. [\[eq:ode-rnn-rnn\]](#eq:ode-rnn-rnn){reference-type="ref" reference="eq:ode-rnn-rnn"} below). Notable examples of this category include ODE-RNNs which transform the hidden states \({\bm{h}}_{n-1}\) to \({\bm{h}}_n\) for each observation \({\bm{x}}_n\) available at time \(t_n\) as follows: \[\begin{aligned}
\label{eq:ode-rnn-ode} {\bm{u}}_n &=& \mathrm{ODESolve}({\bm{f}}_{\theta_1}, {\bm{h}}_{n-1}, t_{n-1}, t_n) \\ \label{eq:ode-rnn-rnn} {\bm{h}}_n &=& \bm{\phi}_{\theta_2}({\bm{x}}_{n}, {\bm{u}}_{n}) \end{aligned}\] where Eq. [\[eq:ode-rnn-ode\]](#eq:ode-rnn-ode){reference-type="ref" reference="eq:ode-rnn-ode"} autonomously updates the hidden state between two observations using a function \({\bm{f}}_{\theta_1}\) parameterised by \(\theta_1\), while in Eq. [\[eq:ode-rnn-rnn\]](#eq:ode-rnn-rnn){reference-type="ref" reference="eq:ode-rnn-rnn"}, function \(\bm{\phi}_{\theta_2}\) parameterised by \(\theta_2\) integrates the new input \({\bm{x}}_n\) into the hidden state. In Latent ODE-RNN, a popular extension of this approach to the variational setting, the initial recurrent state \({\bm{h}}_0\) is sampled from a prior (during training, an additional encoder is trained to map sequences of inputs to parameters of the prior). While this third case is not our focus, we'll also present how to use FWPs in this scenario in Sec. [3.3](#sec:ode_rfwp){reference-type="ref" reference="sec:ode_rfwp"} for the sake of completeness.
## Learning Rules and Their Connections to ODEs {#sec:back_lr}
Learning rules of artificial NNs describe the process which modifies their weights in response to some inputs. This includes the standard backpropagation rule (also known as the reverse mode of automatic differentiation) derived for the case of supervised learning, as well as rules inspired by Hebb's informal rule in "unsupervised" settings. Here we focus on the latter. Let \(n\), \(d_\text{in}\) \(d_\text{out}\) be positive integers. Given a linear layer with a weight matrix \({\bm{W}}_{n} \in \mathbb{R}^{d_\text{out} \times d_\text{in}}\) (the single output neuron case \(d_\text{out}=1\) is the focus of the classic works) at time \(n\) which transforms input \({\bm{x}}_{n} \in \mathbb{R}^{d_\text{in}}\) to output \({\bm{y}}_{n} \in \mathbb{R}^{d_\text{out}}\) as \[\begin{aligned}
\label{eq:hebb} {\bm{y}}_{n} &=& {\bm{W}}_{n-1} {\bm{x}}_n \end{aligned}\] the pure Hebb-style additive learning rule modifies the weights according to \[\begin{aligned}
\label{eq:hebb_update} {\bm{W}}_n &=& {\bm{W}}_{n-1} + \eta_n {\bm{y}}_{n} \otimes {\bm{x}}_n \end{aligned}\] where \(\otimes\) denotes outer product and \(\eta_n \in \mathbb{R}_{+}\) is a learning rate at time \(n\). proposed stability improvements to this rule through a decay term \[\begin{aligned}
\label{eq:oja_82} {\bm{W}}_n = {\bm{W}}_{n-1} + \eta_n {\bm{y}}_n \otimes ({\bm{x}}_n-{\bm{W}}_{n-1}^\top {\bm{y}}_n) \end{aligned}\] whose theoretical analysis has since the 1980s been a subject of many researchers covering stability, convergence, and relation to Principal Component Analysis. One key approach for such theoretical analysis is to view the equation above as a discretisation of the following ODE: \[\begin{aligned}
\label{eq:oja_ode_original} {\bm{W}}'(t) = \eta(t) {\bm{y}}(t) \otimes ({\bm{x}}(t)-{\bm{W}}(t-1)^\top {\bm{y}}(t)) \end{aligned}\] On a related note, the study of RNNs has also profited from ODEs.
## Fast Weight Programmers & Linear Transformers {#sec:back_fwp}
Fast Weight Programmers (FWP; ) are general-purpose (auto-regressive) sequence processing NNs. In general, an FWP is a system of two NNs: a *slow* NN, the *programmer*, rapidly generates during runtime weight changes of another neural network, the *fast* NN. The (slow) weights of the slow net are typically trained by gradient descent. Variants of FWPs whose weight generation is based on outer products between keys and values have been shown to be equivalent to Linear Transformers (using the mathematical equivalence known from perceptron/kernel machine duality ). These FWPs use sequences of learning rules to update short-term memory in form of a fast weight matrix. A practical example of such FWPs is the DeltaNet which transforms an input \({\bm{x}}_n \in \mathbb{R}^{d_\text{in}}\) into an output \({\bm{y}}_n \in \mathbb{R}^{d_\text{out}}\) at each time step \(n\) while updating its fast weight matrix \({\bm{W}}_{n-1} \in \mathbb{R}^{d_\text{out} \times d_\text{key} }\) as follows: \[\begin{aligned}
\label{eq:proj} \beta_n, {\bm{q}}_n, {\bm{k}}_n, {\bm{v}}_n &=& {\bm{W}}_\text{slow}{\bm{x}}_n\\ \label{eq:lr} {\bm{W}}_n &=& {\bm{W}}_{n-1} + \sigma(\beta_n)({\bm{v}}_n-{\bm{W}}_{n-1} \phi({\bm{k}}_n)) \otimes \phi({\bm{k}}_n) \\ {\bm{y}}_n &=& {\bm{W}}_n \phi({\bm{q}}_n) \label{eq:fw_get} \end{aligned}\] where the slow net (Eq. [\[eq:proj\]](#eq:proj){reference-type="ref" reference="eq:proj"}; with weights \({\bm{W}}_\text{slow} \in \mathbb{R}^{(1 + 2 *d_\text{key} + d_\text{out}) \times d_\text{in}}\)) generates vectors key \({\bm{k}}_n \in \mathbb{R}^{d_\text{key}}\) and value \({\bm{v}}_n \in \mathbb{R}^{d_\text{out}}\) as well as a scalar \(\beta_n \in \mathbb{R}\) to obtain a dynamic learning rate by applying a sigmoid function \(\sigma\), and \(\phi\) is an element-wise activation function whose output elements are positive and sum up to one (typically softmax). These fast dynamic variables generated by a slow NN are used in a learning rule (Eq. [\[eq:lr\]](#eq:lr){reference-type="ref" reference="eq:lr"}) akin to the classic delta rule to update the fast weight matrix. The output is finally produced by the forward computation of the fast NN, i.e., by *querying* the fast weight matrix by the generated query vector \({\bm{q}}_n \in \mathbb{R}^{d_\text{key}}\) (Eq. [\[eq:fw_get\]](#eq:fw_get){reference-type="ref" reference="eq:fw_get"}). An intuitive interpretation of the fast weight matrix is a key-value associative memory with write and read operations defined by Eq. [\[eq:lr\]](#eq:lr){reference-type="ref" reference="eq:lr"} and [\[eq:fw_get\]](#eq:fw_get){reference-type="ref" reference="eq:fw_get"}, respectively. This encourages intuitive thoughts about memory capacity (limited by the number of "keys" we can store without interference). For instance, if we replace the learning rule (i.e., memory writing operation) of Eq. [\[eq:lr\]](#eq:lr){reference-type="ref" reference="eq:lr"} by a pure additive Hebb-style rule (and a fixed learning rate of 1.0): \({\bm{W}}_n = {\bm{W}}_{n-1} + {\bm{v}}_n \otimes \phi({\bm{k}}_n)\), we obtain the Linear Transformer (we refer to prior work for further explanations of the omission of attention normalisation). Such a purely additive learning rule often suffers from long term dependencies, unlike the delta rule. We'll confirm this trend also in the CT models (using the EigenWorms dataset). For later convenience, we introduce a function \(\mathrm{FWP}\) which denotes generic FWP operations: \({\bm{y}}_n, {\bm{W}}_n = \mathrm{FWP}({\bm{x}}_n, {\bm{W}}_{n-1}; {\bm{W}}_\text{slow})\).
# Continuous-Time Fast Weight Programmers {#sec:ct_fwp}
We propose continuous-time counterparts of Fast Weight Programmers (Sec. [2.3](#sec:back_fwp){reference-type="ref" reference="sec:back_fwp"}) which naturally combine ODEs for learning rules (Sec. [2.2](#sec:back_lr){reference-type="ref" reference="sec:back_lr"}) and existing approaches for sequence processing with NODEs (Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}). We present three types of these CT FWP models in line with the categorisation of Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"} while the main focus of this work is on the two first cases.
## Direct NODE-based FWPs {#sec:sub_direct_ode}
In the *direct NODE* approach (reviewed in Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}), we assume a (bounded) control signal \({\bm{x}}: t\mapsto {\bm{x}}(t)\) defined over an interval \([t_0, T]\). We make it part of the vector field to define an ODE describing a *continuous-time learning rule* for a fast weight matrix \({\bm{W}}(t)\): \[\begin{aligned}
\label{eq:ode-lr} {\bm{W}}(t) = {\bm{W}}(t_0) + \int_{s=t_0}^t {\bm{F}}_\theta({\bm{W}}(s), {\bm{x}}(s))ds \end{aligned}\] where \({\bm{W}}: t\mapsto {\bm{W}}(t) \in \mathbb{R}^{d_\text{out} \times d_\text{key}}\) is a function defined on \([t_0, T]\), and \({\bm{F}}_\theta\) is an NN parameterised by \(\theta\) which maps onto \(\mathbb{R}^{d_\text{out} \times d_\text{key}}\). This is a neural differential equation for learning to train a neural net through continuous learning rules, that is, to train a fast weight matrix \({\bm{W}}(t)\) of a fast NN (Eq. [\[eq:querying_2\]](#eq:querying_2){reference-type="ref" reference="eq:querying_2"} below) for each sequential control \({\bm{x}}\). Like in the discrete-time FWPs (Sec. [2.3](#sec:back_fwp){reference-type="ref" reference="sec:back_fwp"}), the output \({\bm{y}}(T) \in \mathbb{R}^{d_\text{out}}\) is obtained by *querying* this fast weight matrix[^2] (e.g., at the last time step \(T\)): \[\begin{aligned}
\label{eq:querying_1} {\bm{q}}(T) &= {\bm{W}}_q {\bm{x}}(T) \\ \label{eq:querying_2} {\bm{y}}(T) &= {\bm{W}}(T) {\bm{q}}(T) \end{aligned}\] where \({\bm{W}}_q \in \mathbb{R}^{d_\text{key} \times d_\text{in}}\) is a slow weight matrix used to generate the query \({\bm{q}}(T) \in \mathbb{R}^{d_\text{key}}\) (Eq. [\[eq:querying_1\]](#eq:querying_1){reference-type="ref" reference="eq:querying_1"}). Now we need to specify \({\bm{F}}_\theta\) in Eq. [\[eq:ode-lr\]](#eq:ode-lr){reference-type="ref" reference="eq:ode-lr"} to fully define the learning rule. We focus on three variants: \[\begin{aligned}
\label{eq:direct_ode} {\bm{F}}_\theta({\bm{W}}(s), {\bm{x}}(s)) &= \sigma(\beta(s)) \begin{cases} {\bm{k}}(s) \otimes {\bm{v}}(s) & \text{Hebb-style}\\ {\bm{v}}(s) \otimes \big({\bm{k}}(s)-{\bm{W}}(s)^\top {\bm{v}}(s) \big) & \text{Oja-style}\\ \big({\bm{v}}(s)-{\bm{W}}(s){\bm{k}}(s)\big) \otimes {\bm{k}}(s)& \text{Delta-style} \end{cases} \end{aligned}\] where \([\beta(s), {\bm{k}}(s), {\bm{v}}(s)] = {\bm{W}}_\text{slow} {\bm{x}}(s)\) with a slow weight matrix \({\bm{W}}_\text{slow} \in \mathbb{R}^{(1 + d_\text{key} + d_\text{out}) \times d_\text{in}}\). As in the discrete-time FWP (Sec. [2.3](#sec:back_fwp){reference-type="ref" reference="sec:back_fwp"}), the slow NN generates \(\beta(s) \in \mathbb{R}\) (to which we apply the sigmoid function \(\sigma\) to obtain a learning rate), key \({\bm{k}}(s) \in \mathbb{R}^{d_\text{key}}\) and value \({\bm{v}}(s) \in \mathbb{R}^{d_\text{out}}\) vectors from input \({\bm{x}}(s)\). These variants are inspired by the respective classic learning rules of the same name, while they are crucially different from the classic ones in the sense that all variables involved (key, value, learning rate) are continually generated by the slow NN. In the experimental section, we'll comment on how some of these design choices can result in task-dependent performance gaps. In practice, we use the *multi-head* version of the operations above (i.e., by letting \(H \in \mathbb{N}\) denote the number of heads, query/key/value vectors are split into \(H\) sub-vectors and Eqs. [\[eq:querying_2\]](#eq:querying_2){reference-type="ref" reference="eq:querying_2"}-[\[eq:direct_ode\]](#eq:direct_ode){reference-type="ref" reference="eq:direct_ode"} are conducted independently for each head). The output is followed by the standard feed-forward block like in Transformers. Possible extensions for deeper models are discussed in Appendix [\[app:deeper\]](#app:deeper){reference-type="ref" reference="app:deeper"}.
## NCDE based FWPs {#sec:cde-lr}
Here we present models based on NCDEs (reviewed in Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}). We assume availability of a differentiable control signal \({\bm{x}}(t)\), whose first order derivative is denoted by \({\bm{x}}'(t)\). Given the NCDE formulation of Eq. [\[eq:cde\]](#eq:cde){reference-type="ref" reference="eq:cde"}, the most straight-forward approach to obtain a CT Fast Weight Programmer is to extend the dimensionality of the recurrent hidden state, i.e., we introduce a parameterised function \({\tens{F}}_\theta\) which maps a matrix \({\bm{W}}(t) \in \mathbb{R}^{d_\text{out} \times d_\text{key}}\) to a third-order tensor \({\tens{F}}_\theta({\bm{W}}(t)) \in \mathbb{R}^{d_\text{out} \times d_\text{key} \times d_\text{in}}\): \[\begin{aligned}
\label{eq:cde-lr} {\bm{W}}(t) = {\bm{W}}(t_0) + \int_{s=t_0}^t {\tens{F}}_\theta({\bm{W}}(s))d{\bm{x}}(s) = {\bm{W}}(t_0) + \int_{s=t_0}^t {\tens{F}}_\theta({\bm{W}}(s)){\bm{x}}'(s)ds \end{aligned}\] However, this approach is obviously not scalable since the input and output dimensions (\(d_\text{out} \times d_\text{key}\) and \(d_\text{out} \times d_\text{key} \times d_\text{in}\)) of \({\tens{F}}_\theta\) can be too large in practice. A more tractable CDE-based approach can be obtained by providing \({\bm{x}}\) and/or \({\bm{x}}'\) to the vector field: \[\begin{aligned}
\label{eq:cde-lr-x-dx} {\bm{W}}(t) = {\bm{W}}(t_0) + \int_{s=t_0}^t {\tens{F}}_\theta({\bm{W}}(s), {\bm{x}}(s), {\bm{x}}'(s)){\bm{x}}'(s)ds \end{aligned}\] While this equation still remains a CDE because of the multiplication from the right by \(d{\bm{x}}={\bm{x}}'(s)ds\), the additional inputs to the vector field offer a way of making use of various learning rules, as in the case of direct NODE approach above (Sec. [3.1](#sec:sub_direct_ode){reference-type="ref" reference="sec:sub_direct_ode"}). To be specific, either \({\bm{x}}\) and \({\bm{x}}'\) or only \({\bm{x}}'\) is required in the vector field to obtain these tractable CDEs. Here we present the version which uses both \({\bm{x}}\) and \({\bm{x}}'\) [^3]. The resulting vector fields for different cases are: \[\begin{aligned}
\label{eq:cde_lr} {\tens{F}}_\theta\big({\bm{W}}(s), {\bm{x}}(s), {\bm{x}}'(s)\big){\bm{x}}'(s) &= \sigma(\beta(s)) \begin{cases} {\bm{W}}_k {\bm{x}}(s) \otimes {\bm{W}}_v {\bm{x}}'(s) & \text{Hebb}\\ \big({\bm{W}}_k {\bm{x}}(s)-{\bm{W}}(s)^\top {\bm{W}}_v {\bm{x}}'(s) \big) \otimes {\bm{W}}_v {\bm{x}}'(s) & \text{Oja} \\ \big({\bm{W}}_v {\bm{x}}(s)-{\bm{W}}(s){\bm{W}}_k {\bm{x}}'(s)\big) \otimes {\bm{W}}_k {\bm{x}}'(s) & \text{Delta} \end{cases} \end{aligned}\] As can be seen above, the use of CDEs to describe a continuous fast weight learning rule thus naturally results in a key/value memory where \({\bm{x}}'\) is used to generate either key or value vectors. Because of the multiplication from the right by \({\bm{x}}'\), the role of \({\bm{x}}'\) changes depending on the choice of learning rule: \({\bm{x}}'\) is used to generate the key in the Delta case but the value vector in the case of Oja. In the case of Hebb, the choice made in Eq. [\[eq:cde_lr\]](#eq:cde_lr){reference-type="ref" reference="eq:cde_lr"} of using \({\bm{x}}\) for keys and \({\bm{x}}'\) for values is arbitrary since Eq. [\[eq:cde_lr\]](#eq:cde_lr){reference-type="ref" reference="eq:cde_lr"} is symmetric in terms of roles of keys and values (see an ablation study in Appendix [\[app:cde_variants\]](#app:cde_variants){reference-type="ref" reference="app:cde_variants"} for the other case where we use \({\bm{x}}'\) to generate the key and \({\bm{x}}\) for the value). The querying operation (analogous to Eqs. [\[eq:querying_1\]](#eq:querying_1){reference-type="ref" reference="eq:querying_1"}-[\[eq:querying_2\]](#eq:querying_2){reference-type="ref" reference="eq:querying_2"} for the direct NODE case) is also modified accordingly, depending on the choice of learning rule, such that the same input (\({\bm{x}}\) or \({\bm{x}}'\)) is used to generate both key and query: \[\begin{aligned}
\label{eq:cde_fwp_query} {\bm{y}}(T) &= \begin{cases} {\bm{W}}(T)^\top {\bm{W}}_q {\bm{x}}(T)& \text{Hebb and Oja}\\ {\bm{W}}(T) {\bm{W}}_q {\bm{x}}'(T) & \text{Delta} \end{cases} \end{aligned}\] Note that since the proposed vector field \({\tens{F}}_\theta\big({\bm{W}}(s), {\bm{x}}(s), {\bm{x}}'(s)\big){\bm{x}}'(s)\) is more general than the one used in the original NCDE \({\tens{F}}_\theta\big({\bm{W}}(s)\big){\bm{x}}'(s)\), any theoretical results on the CDE remain valid (which, however, does not tell us anything about the best choice for its exact parameterisation).
## ODE-RFWP and Latent ODE-RFWP {#sec:ode_rfwp}
The main focus of this work is the setting of where we assume the existence of some control signal \({\bm{x}}\). However, here we also show a way of using FWPs in the third case where no control \({\bm{x}}(t)\) is available, i.e., we only have access to discrete observations \(({\bm{x}}_n)_{n=0}^N\). Here we cannot directly define the vector field involving continuous transformations using the inputs. We follow the existing approaches (ODE-RNN or Latent ODE; Sec. [2.1](#sec:back_ode){reference-type="ref" reference="sec:back_ode"}) which use two separate update functions: A discrete recurrent state update is executed every time a new observation is available to the model, while a continuous update using an autonomous ODE is conducted in between observations. Unlike with standard recurrent state vectors, however, it is not practical to autonomously evolve high-dimensional fast weight matrices (which would require an expressive matrix to matrix transformation). We therefore opt for using a Recurrent FWP (RFWP) and combine it with an ODE: \[\begin{aligned}
\label{eq:ode-rfwp-ode} {\bm{u}}_n &=& \mathrm{ODESolve}({\bm{f}}_{\theta_1}, {\bm{h}}_{n-1}, t_{n-1}, t_n) \\ \label{eq:ode-rfwp-rnn} {\bm{h}}_n, {\bm{W}}_{n} &=& \mathrm{FWP}([{\bm{x}}_{n}, {\bm{u}}_{n}], {\bm{W}}_{n-1}; \theta_2) \end{aligned}\] where we keep the fast weight learning rule itself discrete (Eq. [\[eq:ode-rfwp-rnn\]](#eq:ode-rfwp-rnn){reference-type="ref" reference="eq:ode-rfwp-rnn"}), but evolve the recurrent state vector \({\bm{u}}_n\) using an ODE (Eq. [\[eq:ode-rfwp-ode\]](#eq:ode-rfwp-ode){reference-type="ref" reference="eq:ode-rfwp-ode"}) such that the information to be read/written to the fast weight matrix is controlled by a variable which is continuously updated between observations. We refer to this model as ODE-RFWP and its variational variant as Latent ODE-RFWP. While this case is not of central interest to the present work, as the learning rule remains discrete (Eq. [\[eq:ode-rfwp-rnn\]](#eq:ode-rfwp-rnn){reference-type="ref" reference="eq:ode-rfwp-rnn"}), we also provide some experimental results (for model-based reinforcement learning settings) in Appendix [\[app:rl\]](#app:rl){reference-type="ref" reference="app:rl"}.
# Experiments {#sec:exp}
We consider three datasets covering three types of time series which are regularly sampled (Speech Commands ), irregularly sampled with partially missing features (PhysioNet Sepsis ), or very long (EigenWorms ). We compare the proposed direct NODE and CDE based FWP models (Sec. [3.1](#sec:sub_direct_ode){reference-type="ref" reference="sec:sub_direct_ode"} & [3.2](#sec:cde-lr){reference-type="ref" reference="sec:cde-lr"}) to NODE baselines previously reported on the same datasets. Appendix [\[app:exp_detail\]](#app:exp_detail){reference-type="ref" reference="app:exp_detail"} provides further experimental details.
:::
#### EigenWorms.
The EigenWorms dataset (which is part of the UEA benchmark ) is a 5-way classification of roundworm types based on time series tracking their movements. To be more specific, motions of a worm are represented by six features corresponding to its projections to six template movement shapes, called "eigenworms". While this dataset contains only 259 examples, it is notable for its very long sequences (raw sequence lengths exceed 17 K) and long-span temporal dependencies. We use the same train/validation/test split ratio as the prior work which reports neural RDEs (NRDEs) as achieving the best NODE model performance on this dataset. The equations of our CT FWPs for the RDE case can be straightforwardly obtained by replacing the input \({\bm{x}}\) in Eqs. [\[eq:ode-lr\]](#eq:ode-lr){reference-type="ref" reference="eq:ode-lr"} and [\[eq:querying_1\]](#eq:querying_1){reference-type="ref" reference="eq:querying_1"} of the direct NODE formulation by the corresponding log-signatures. Table [2](#tab:eigenworms){reference-type="ref" reference="tab:eigenworms"} shows the results, where "Step" denotes the time sub-sampling rate which is fixed to 4 for which the prior work reports the best NRDE and NCDE performance. "Sig-Depth" denotes the depth of the log-signature (the deeper, the more log-signature terms we take into account, thus ending up with a larger input feature vector; we refer to the original paper for further details). We consider two values for this parameter: 1 and 2. When set to 1, the input feature contains only the first derivative \({\bm{x}}'(s)\) and thus the NRDE is reduced to a NCDE. We take the best NCDE performance from as the depth-1 baseline. report the best overall performance for the depth-2 NRDE (depth-2 baseline in our table). In both cases, we first note a large performance gap between models with different learning rules. While the naive Hebb and Oja based models struggle with this very long sequence processing (sequence length still exceeds 4 K with a down-sampling step size of 4), the Delta rule performs very well. This confirms the prior result in the discrete-time domain which motivated the Delta rule design by its potential for handling long sequences (we refer to prior work for further explanations). Since its performance on other tasks is comparable to the one of Hebb and Oja variants, the Delta rule is a natural default choice for parameterising the CT FWPs. In both the depth-1 and depth-2 cases, we obtain large improvements compared to the respective baselines. It is counter-intuitive to find certain depth-2 models underperforming their depth-1 counterparts, but this trend has also been observed in the original NRDEs. Our best overall performance is obtained in the depth-1 case: 91.8 % (3.4) exceeds the previous best NRDE based model performance of 83.8 % (3.0). This almost matches the state-of-the-art accuracy of 92.8 % (1.8) reported by (using an ODE-inspired discrete-time model). Our model's performance variance is high (best single seed performance is 97.4% while the standard deviation is 3.4). This calls for future investigations to improve stability of ODE-based models across random initialisations. The wall clock time is similar for our best model (last row in Table [2](#tab:eigenworms){reference-type="ref" reference="tab:eigenworms"}) and the NRDE baseline (36s/epoch on a GeForce RTX 2080 Ti) and their sizes are comparable (87 K vs. 65 K parameters respectively).
# Discussions {#sec:diss}
#### Scalability Advantage Compared to Standard NCDEs.
Apart from our primary motivation, our FWP approach also addresses an important limitation of existing NCDEs: their scalability in terms of model size. The vector field in standard NCDEs (Eq. [\[eq:cde\]](#eq:cde){reference-type="ref" reference="eq:cde"}) requires an NN \({\bm{F}}_\theta\) which takes a vector \({\bm{h}}(s) \in \mathbb{R}^d\) as an input to produce a matrix of size \(\mathbb{R}^{d \times d_{\text{in}}}\). This can be very challenging when \(d_{\text{in}}\) or/and \(d\) is large. Actually, the same bottleneck is present in the weight generation of FWPs. The use of outer products can remediate this issue in discrete FWPs as well as in CT FWPs: the computations in our FWP-based NODE/NCDEs only involve "first-order" dimensions (i.e., no multiplication between different dimensions, such as \(d \times d_{\text{in}}\)) for NN outputs. This can scale well with increased model size, making feasible larger scale tasks infeasible for existing NCDEs. On the other hand, report that using outer products (in their "Sec. 6.1 on limitations") in standard NCDEs does not perform well. Why do outer products work in FWPs but not in the original NCDEs? The answer may be simple. In the original NCDEs (Eq. [\[eq:cde\]](#eq:cde){reference-type="ref" reference="eq:cde"}), multiplications occur at each (infinitesimal) time step between the generated rank-one weight matrix \({\bm{F}}_\theta({\bm{h}}(s))\) and \({\bm{x}}'(s)\) before the sum. All these transformations are thus of rank one while we expect expressive transformations to be necessary to translate \({\bm{x}}'(s)\) into changes in the state space. In contrast, in FWPs, the ODE only parameterises the weight generation process of another net, and thus the rank-one matrices are never used in isolation: they are summed up over time (Eq. [\[eq:ode-lr\]](#eq:ode-lr){reference-type="ref" reference="eq:ode-lr"} or [\[eq:cde-lr-x-dx\]](#eq:cde-lr-x-dx){reference-type="ref" reference="eq:cde-lr-x-dx"}) to form an expressive weight matrix which is only then used for matrix multiplication (Eq. [\[eq:querying_2\]](#eq:querying_2){reference-type="ref" reference="eq:querying_2"} or [\[eq:cde_fwp_query\]](#eq:cde_fwp_query){reference-type="ref" reference="eq:cde_fwp_query"}. The proposed FWP-NODE/NCDEs thus offer scalable alternatives to existing NCDEs, also yielding good empirical performance (Sec. [4](#sec:exp){reference-type="ref" reference="sec:exp"}).
#### Memory Efficient Backpropagation for FWPs
Memory efficiency of adjoint backpropagation may be not so important for standard NCDEs of state size \(O(d)\), but is crucial for FWPs of state size \(O(d^2)\) which can quickly become prohibitive for long sequences, as naive backpropagation stores all states used in the forward pass. Prior works on discrete FWPs solve this problem by a custom memory-efficient implementation. Here, the adjoint method naturally addresses this problem.
#### Related Work on Parameter/Weight ODEs.
There are other works which use ODEs to parameterise time-evolving weights of some model. However, they are limited to autonomous ODEs (i.e., no external control \({\bm{x}}\) is involved). and study coupled ODEs where one ODE is used for temporal evolution of parameters of the main neural ODE. The scope of these two works is limited to autonomous ODEs corresponding to continuous-depth residual NNs with different parameters per depth. consider an ODE version of a gradient descent learning process for adaptation, but also formulated as an autonomous ODE. In contrast, our focus is really on *sequence processing* where the model continuously receives external controls \({\bm{x}}\) and translates them into weight changes of another network.
# Conclusion {#sec:ccl}
We introduced novel continuous-time sequence processing neural networks that learn to use sequences of ODE-based continuous learning rules as elementary programming instructions to manipulate short-term memory in rapidly changing synaptic connections of another network. The proposed models are continuous-time counterparts of Fast Weight Programmers and linear Transformers. Our new models experimentally outperform by a large margin existing Neural ODE based sequence processors on very long or irregularly sampled time series. Our Neural ODE/CDE based FWPs also address the fundamental scalability problem of the original Neural CDEs, which is highly promising for future applications of ODE based sequence processors to large scale problems. | {'timestamp': '2022-06-06T02:16:45', 'yymm': '2206', 'arxiv_id': '2206.01649', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01649'} |
# INTRODUCTION
Trust has been identified as a key factor for effective human-robot interaction. Consequently, substantial research efforts have been devoted to identifying factors that influence humans' trust in robots, developing computational models for trust estimation, and developing trust-aware robots. Previous computational models of trust require binary performance measures of the autonomy. However, such measures cannot be directly applied in scenarios wherein the robotic agent needs to perform complex trade off decisions to maximize the cumulative reward and hence the autonomy performance is more difficult to quantify. In addition, although previous studies have revealed the existence of different types of trust dynamics, no research has tried to associate personal characteristics with the type of trust dynamics. To fill these research gaps, we develop a novel reward-based performance metric to drive the trust estimation algorithm in a Markov Decision Process (MDP). The proposed performance metric is a binary surrogate of the autonomous agent's reward function, which allows our autonomous agent to utilize the trust update model developed in for trust-aware decision-making. Moreover, from the data collected through human-subject experiments, we analyze how human trust evolves with their earned reward over repeated interactions with the autonomy. We find three distinct types of trust dynamics through \(k-\)means clustering analysis and examine associations between personal characteristics and type of trust dynamics.
# RELATED WORK {#sec:relatedWork}
Extensive work has been done in identifying factors that affect trust in automation. However, most of these studies provide a snapshot view of trust, and failed to characterize the dynamics of trust. More recently, dynamic models of trust has been proposed, including the auto-regressive moving average vector (ARMAV) model, the Online Probabilistic Trust Inference Model (OPTIMo) model, and the experience-based model. These models enable a robot to estimate a human's trust in real time and has led to the development of trust-aware robots, allowing robots to reason upon human trust in their planning. Most existing research investigating individual differences aims to find associations between (snapshot) trust and individual characteristics, including propensity to trust automation, personality trait of neuroticism, and dispositional expectations of technology. It is possible that these individual differences will be differentially represented within trust profiles related to trust dynamics. Recent research reveal the existence of different types of trust dynamics, namely, rational decision maker, oscillator, and disbeliever. It is reasonable to investigate the association between a human agent's trust dynamics and factors that are shown to affect a human agent's (snapshot) trust.
# PROBLEM FORMULATION {#sec:model}
In this section, we propose a finite horizon Markov Decision Process (MDP) for modeling and incorporating trust in the decision-making system of a robotic agent. The robot provides recommendations to its human partner about the action that they should take, but the final decision lies with the human. The specific scenario we target is an 'Intelligence, Surveillance, and Reconnaissance' (ISR) mission in which a human soldier teams up with an *intelligent* drone to search through a town for threats. The drone guides the soldier on whether s/he should breach a site directly or deploy a Robotic Armored Rescue Vehicle (RARV). Using the RARV prevents any health loss to the soldier in the presence of a threat but takes additional time to the team; on the contrary, breaching a site directly is faster, but the soldier will be harmed if s/he encounters a threat. Here, two natural (but conflicting) goals that arise are to minimize any damage to the soldier and minimize the time to search through all the sites. The key components of the MDP model of the ISR mission are given below.
### **States**
We use the estimated trust of the human on the robot as the state. Specifically, we use the trust model in, wherein trust level at the \(i^\text{th}\) stage, \(t_i\), follows a Beta distribution, i.e., \(t_i \sim Beta(\alpha_i, \beta_i),\) where \(\alpha_i\) and \(\beta_i\) are the positive experience and negative experience at time \(i\). And, the state at stage \(i\) is specified by the tuple \((\alpha_i, \beta_i)\).
### **Actions**
The drone has two actions: Recommend to use or not use the RARV.
### **Human behavior model**
We assume that human will accept the recommendation given by the robotic agent with probability \(t_i\); and the human chooses the opposite action of the one that was recommended with probability \(1-t_i\).
### **Reward function**
We use the weighted sum of the health loss cost and the time cost to define the cost of choosing an action. Similar to, we add a trust-gaining term to incentivize the autonomous agent to gain trust. As a result, the reward \(R_i\) at the \(i^\text{th}\) stage is defined as \[\label{eq::reward-function} R_i(a) =-w_h h(a)-w_c c(a) + \gamma_i\cdot \mathbb{1}(A),\] where \(w_h\) and \(w_c\) are the weights for the health loss cost and time loss cost respectively. We define \(A\) as the event when trust increases, i.e., \(\mathbb{1}(A) = 1\) if the performance of the autonomous agent is a success and \(\mathbb{1}(A) = 0\) otherwise. \(\gamma_i (=w_t \sqrt{N-i})\) is a weight given to the trust gain reward that decreases with the stage number. The idea behind this is to support trust-gaining behavior near the current stage, and task performance optimizing behavior towards the later stages of planning.
### **Transition function** {#sec::TransitionFunction}
The state \((\alpha_i, \beta_i)\) is updated by \[\label{eq::alpha-update} ( \alpha _{i} ,\beta _{i}) =\begin{cases} \left( \alpha _{i-1} +w^{s} ,\beta _{i-1}\right) & \text{if } p_{i} =1,\\ \left( \alpha _{i-1} ,\beta _{i-1} +w^{f}\right) & \text{if } p_{i} =0. \end{cases}\] Here, \(p_{i}\) is the autonomous agent's performance at the \(i^\text{th}\) stage. Since the team's goal is to maximize earned reward, we define \(p_i=1\) if at the \(i^\text{th}\) site, the immediate reward for following the recommendation by the drone was greater than that for not following the recommendation and \(p_i=0\) otherwise. We fit the parameters \((\alpha_0, \beta_0, w^s, w^f)\) for each participant to model their trust dynamics. We use gradient descent to estimate the parameters in real time after each feedback from the participant. We use value iteration to solve the MDP.
# Experiment {#sec:case}
This section describes details of the human-subject experiment. The experiment complied with the American Psychological Association code of ethics and was approved by the Institutional Review Board at the University of Michigan.
## Participants
A total of 46 adults participated in the study. One participant's data was discarded as the participant failed the attention check. The remaining 45 participants consisted of 21 females and 24 males (Age: Mean = 22.8 years, *SD* = 3.6 years). Each participant was reimbursed with a base pay of \\(20 with a bonus of up to \\)10 based on their performance, which was measured by the time taken by the participants to complete the task and the final health level of the soldier.
## Testbed
We developed a 3D testbed in Unreal Engine. A screenshot of the testbed is shown in Fig. [\[fig::Background\]](#fig::Background){reference-type="ref" reference="fig::Background"}. Fig. [\[fig::Recommendation\]](#fig::Recommendation){reference-type="ref" reference="fig::Recommendation"} shows the recommendation dialog box where the participant was recommended to not use the RARV. After choosing an action, the four possibilities depending on the presence of threat and participant's selected actions are shown in Fig. [\[fig::Outcomes\]](#fig::Outcomes){reference-type="ref" reference="fig::Outcomes"}. The participants are told that if they encounter a threat without the RARV, they will lose \(5\) points of health. They were told that deploying the RARV takes about \(10\) seconds. They were told to choose an action based on their interaction history and the recommendation. Their goal was to minimize the time taken for the mission, while maintaining the soldier's health.
## Measures
### Personality
The big 5 factors of personality (Extraversion, Conscientiousness, Agreeableness, Neuroticism, and Imagination) were measured using the 20-item mini-IPIP scale. This 5-point Likert scale has widely been used in human-robot trust research.
### Propensity to Trust
Propensity to trust autonomous systems was assessed using a 6-item scale developed in. This was also a 5-point Likert scale
### Perfect Automation Schema
Perfect Automation Schema (PAS) was measured using the 7-item scale developed by. Of these, 4 items measure high expectations from the autonomy and the other 3 items measure All-or-none thinking. This was a 7-point Likert scale.
### Moment-to-Moment Trust
During the task, the participants were asked to rate their moment-to-moment trust on the drone by adjusting a slider on a 100-point scale.
### Post-experiment trust
After the experiment, we used two scales for assessing trust. The first one was a 9-item questionnaire with sliders while we used 6-items from the second scale all of which were 7-point Likert type questions.
### Workload
Workload was measured using the NASA Task Load Index. We only used 5 of the 6 items as there was no physical demand from the participants in our experiment. All the items had the participants rate their feelings on a slider with values ranging from Very low to Very high.
## Experimental procedure
Prior to the experiment, participants provided informed consent and completed the pre-experiment surveys. They were oriented to the steps of the experiment and walked through each of the screens they would see during the experiment. The two-fold objective of minimizing time and maximizing health was emphasized. They were informed about the performance-based bonus pay. They were told that the drone was imperfect, but they were not informed of the exact reliability level. They were also told that the robotic agent's recommendations would help them achieve the two-fold objective. Participants then proceeded to the experimental trials, wherein they had to search through 100 houses sequentially. After searching each house, the participants were asked to report their level of trust on the autonomous agent's recommendations. The participants took on average 51.5 minutes to complete the task. At the end of the experiment, participants completed the post-experiment surveys.
# RESULTS & DISCUSSION {#sec:result}
## Using Immediate Actual Reward as a performance metric
Since the participants are explicitly told to consider both the soldier's health and the time to complete the search as their objectives, we expect their trust to be correlated with the immediate reward that they receive upon choosing an action. We expect that a participant's trust would be likely to increase if following the recommendation by the drone gets them a higher reward than doing the opposite and vice versa. Fig. [\[fig::ImmediateRewardAndTrust\]](#fig::ImmediateRewardAndTrust){reference-type="ref" reference="fig::ImmediateRewardAndTrust"} shows a representative example. In the figure, a green triangle represents the site at which following the recommendation would result in a better immediate reward gain (\(p_i=1\)) and the red triangles represent the opposite. It is quite clear that a red triangle is often followed by a decrease in trust and a green triangle is followed by an increase in trust. Thus, our reward-based performance metric is able to capture moment-to-moment trust changes of the participant. Using this metric in our trust update model, we get a prediction root mean squared error of \(0.1266\) \((SD=0.078)\).
## Clustering of Trust Dynamics
We employ k-means clustering to group participants with similar trust dynamics. We use the root mean squared error (\(E_{RMS}\)) between the predicted trust and the trust feedback and the average logarithm of trust as the clustering features. For computing \(E_{RMS}\), we consider the feedback at the first \(20\) sites as a training set to fit the personalized parameters \((\alpha_0, \beta_0, w^s, w^f)\). Thereafter, we use the feedback after every 5 sites to update the parameters. Fig. [\[fig::Clustering-clusters\]](#fig::Clustering-clusters){reference-type="ref" reference="fig::Clustering-clusters"} shows the results. We choose \(k=3\) as the optimum number of clusters because it is a turning point of both the within-cluster variance and the silhouette score. We call the cluster with small \(E_{RMS}\) and generally higher values of trust *Bayesian Decision Makers*. The second cluster, called *Disbelievers*, consists of participants whose \(E_{RMS}\) values are small but whose trust is generally low. The third significant group is the *Oscillators*, whose trust changes rapidly, making it harder to predict.
## Personal Characteristics and Type of Trust Dynamics
We performed one-way ANOVA between the data of the clustered participants. Results showed significant difference between the three types of trust dynamics in Extraversion (\(F(2,42)=4.991, p=0.011\)), Agreeableness (\(F(2,42)=3.276, p=0.048)\), and the High Expectations facet of the Perfect Automation Schema \((F(2,42)=5.752, p=0.006)\). Further, propensity to trust automation \((F(2,42)=3.002, p=0.06)\) and intellect/imagination \((F(2,42)=2.687,p=0.08)\) had marginally significant differences. Post-hoc analysis with Bonferroni adjustment shows that the disbelievers are significantly less extroverted than the oscillators \((p=0.009)\). Disbelievers were only marginally less extroverted than Bayesian Decision Makers \((p=0.061)\). Disbelievers were marginally less agreeable than the Bayesian Decision Makers (\(p=0.068\)) and Oscillators (\(p=0.06\)). Disbelievers had significantly lower expectations from automation compared to both Bayesian Decision Makers \((p=0.005)\) and Oscillators \((p=0.023)\). One way ANOVA on the post-experiment measures show significant difference between the three clusters in their post-experiment trust reports (Trust questionnaire by Muir and Moray \(F(2,42)=22.167, p<0.001\), Trust questionnaire by Lyons and Guznov \(F(2,42)=15.183, p<0.001\)) and their frustration levels (\(F(2,42)=4.136, p=0.023)\)). Post-hoc analysis with Bonferroni adjustment shows that there are significant differences between each of the three groups' trust reports according to the trust questionnaire by Muir and Moray (\(p<0.001\) between Bayesian Decision Makers and Disbelievers, \(p=0.006\) between Bayesian Decision Makers and Oscillators, and \(p=0.009\) between Oscillators and Disbelievers with the highest trust for Bayesian Decision Makers and lowest for Disbelievers). Trust reported with the questionnaire by Lyons and Guznov only showed significant difference between Disbelievers and Bayesian Decision Makers \((p<0.001)\), and between Oscillators and Bayesian Decision Makers \((p=0.002)\). Oscillators were significantly more frustrated than Bayesian Decision Makers \((p=0.025)\).
# CONCLUSIONS {#sec:discussion}
In this study, we formulated the human-robot sequential decision-making problem as an MDP and proposed an innovative reward-based performance metric for trust-aware decision-making. Through a human-subject experiment, we found three distinct types of trust dynamics, namely, Bayesian Decision Makers, Oscillators, and Disbelievers. Our model achieved different accuracies on different types on trust dynamics, which points to a requirement to use different models for people in different categories. Analyses suggest that those individuals classified as Bayesian Decision makers evidenced high expectations of automation. When combined with other state measures such as trust and frustration, one might be able to identify a profile of the Bayesian Decision makers as the combination of individual differences and state measures could be used to parse the sample into the three categories of trust dynamics. Knowing that an individual might fall into one of the categories could influence whether or not a machine partner that is equipped with a dynamic trust model is a feasible solution for that individual. In future studies, besides the reverse-psychology model, different human trust-behavior models should be considered. In addition, techniques such as inverse reinforcement learning can be employed to learn the preference of the human teammate to improve the team performance. | {'timestamp': '2022-06-06T02:16:39', 'yymm': '2206', 'arxiv_id': '2206.01645', 'language': 'en', 'url': 'https://arxiv.org/abs/2206.01645'} |
null | null |
# Introduction: Arguing for Confinement in Non-Supersymmetric Theories Obtained by Deformation of Confining Supersymmetric Theories
Confinement and the existence of a mass gap are readily established features of many \(N=1\) supersymmetric gauge theories. This is the case, in particular, for gauge theories without matter fields and gauge theories with only massive matter. Exploiting tools such as holomorphy, it is possible to study many aspects of the problem, such as the role of monopole condensation. It is natural to ask whether these results can be extended to demonstrate confinement in non-supersymmetric theories. One could start, for example, by taking a supersymmetric gauge theory without matter, and then introducing a gaugino mass term. Due to the mass gap, for small gaugino mass both the mass gap and confinement persist. The question is: is there a deconfining phase transition as \(m_\lambda\) becomes large compared to the QCD scale? The analogous problem can be addressed in \(CP^N\) theories in two dimensions at large \(N\), where one can readily check that there is no phase transition as one passes from the supersymmetric to the non-supersymmetric theory. These theories are confining and gapped in both limits. Confinement is a phenomenon of the far infrared theory, and the infrared theory is the same in both limits, up to a rescaling of lengths. In four dimensional gauge theories, we don't have a simple (continuum) picture of confinement, so the problem is more challenging. Again, if we deform the supersymmetric theories with small soft breaking terms, confinement will survive, as readily follows from the existence of a mass gap. But for large soft breakings, i.e. to describe real QCD, more is required to establish confinement. In, arguments were put forward that if one considered theories with anomaly mediated supersymmetry breaking, there would be no phase transition. These arguments relied on working only to linear order in soft breakings, and considering the monopole condensate, but as explained in, non-linear effects (in the low energy dynamics) become important (and difficult to control) as we approach the QCD scale. So while the result is plausible, its foundations are shaky. But focussing on the heavy quark-antiquark potential (Wilson line), we will argued that a deconfining transition can be ruled out. Framed this way, confinement follows from rather mild assumptions about the theory in the far infrared. In large \(N\), these assumptions reduce to properties which are generally assumed to be true of the planar diagram expansion. In this paper, we give simple arguments for confinement from this perspective. Possible loopholes are enumerated, but are not very plausible. The rest of this note is organized as follows. We briefly review confinement in supersymmetric theories in the next section. In section [3](#cpn){reference-type="ref" reference="cpn"}, we then remark on confinement in the two-dimensional \(CP^N\) models at large \(N\), with and without supersymmetry, noting certain features which we expect to be relevant to four dimensions. Turning to four dimensions, we give naive arguments for confinement and then a more general argument for confinement for large \(m_\lambda\) in section [4](#generalargument){reference-type="ref" reference="generalargument"}. We explore the question of large \(N\) theories in four dimensions with small and large soft breakings in section [5](#largen){reference-type="ref" reference="largen"}. Limiting consideration to CP conserving values of parameters reduces the problem of confinement to standard assumptions about large \(N\). In section [6](#lightquarks){reference-type="ref" reference="lightquarks"}, we extend the analysis to theories with light quarks. In section [7](#etaprime){reference-type="ref" reference="etaprime"}, we compare these assumptions with those governing the \(\eta^\prime\) at large \(N\). In section [8](#thetadependence){reference-type="ref" reference="thetadependence"}, we remark on possible behaviors with CP violating parameters. Our conclusion, that confinement holds in the theories with large soft breakings, is presented in section [9](#conclusions){reference-type="ref" reference="conclusions"}.
# Confinement in Supersymmetric Gauge Theories {#susyconfinement}
Proving confinement in non-abelian gauge theories without supersymmetry remains one of the great challenges of theoretical physics. What is remarkable is that, with \(N=1\) supersymmetry, one can establish confinement in a range of gauge theories, in a variety of ways. These include:
1. Following Seiberg and Witten, one can prove confinement by exhibiting monopole condensation in theories which can be obtained by deformation of \(N=2\) theories.
2. Starting with \(SU(N)\) supersymmetric QCD with \(N-1\) massive flavors, one can perform a systematic computation, for small masses, of the spectrum and other features. Using holomorphy, one can study the theory in various regimes of quark mass, establishing, among other results, gaugino condensation and the existence of a mass gap in the \(N_F = 0\) theory.
In the case of the \(SU(N)\) \(N=2\) theory, before one adds a mass term for the adjoint field, \(\phi\), there are \(N\) points on on the moduli space of vacua where monopoles become massless: \[\langle u \rangle \equiv \langle {\rm Tr}~ \phi^2 \rangle = u_0 e^{2 \pi i \over N}. \label{seibergwittenvacua}\] These transform into one another under a \(Z_N\) symmetry. Adding a large mass for the adjoint, the theory at low energies is supersymmetric \(SU(N)\) QCD without matter, and these states correspond to the \(N\) possible values of the gaugino condensate, \(\langle \lambda \lambda \rangle \propto e^{2 \pi i \over N}\). With a gaugino mass, the degeneracy is lifted, and all but one of the states are unstable. For large enough gaugino mass, we would expect that all but the ground state are highly unstable. So confinement and the existence of a mass gap are common in supersymmetric field theories. Notably, they are features of certain large \(N\) theories, such as the \(SU(N)\) supersymmetric gauge theory without matter, or with number of flavors, \(N_f\), much less than \(N\), providing additional tools to study non-supersymmetric theories.
# Lessons from the \(CP^N\) Models in Two Dimensions {#cpn}
The \(CP^N\) models at large \(N\) in two dimensions are instructive for the case of four dimensional QCD. These theories are confining and gapped. While confinement itself is much simpler in two than in four dimensions, other aspects of the problem, particularly the scaling of terms in a four dimensional effective action, have close parallels. What is similar about the two theories is that confinement is a problem of the far infrared. As a result, it is natural to focus on a Wilsonian action with cutoff, \(\Lambda_w\), well below the scale of the theory, \(\Lambda\) (the dynamically generated scale in the two dimensional case, \(\Lambda_{QCD}\),the QCD scale in the four dimensional case). Confinement is a feature of this very low energy effective theory. For the non-supersymmetric \(CP^N\) models, the microscopic degrees of freedom are a set of \(N\) complex scalars on the \(CP^N\) manifold; more conveniently, they are a set of scalar fields along with a scalar lagrange multiplier and a vector field with no kinetic term at the classical level. The scalars gain mass of order \(\Lambda\), the dynamically generated scale of the theory. In the far infrared, the the scalar degrees of freedom are integrated out and the Wilsonian action is written in terms of a gauge field, \(A_\mu\), and its field strength \(F_{\mu \nu}\). In the non-supersymmetric case, the low scale effective action can be computed in the loop expansion, with successive loops suppressed by by powers of \(1/N\). At one loop, one has a series of operators of the form \(F_{\mu \nu}^{2n}\), with coefficients scaling as \(N \Lambda^{-4n+2}\). But setting aside the explicit one loop computation, we can understand this scaling by the requirement that if one perform a one loop computation with the effective action, it yields corrections to the effective action smaller than, or at least not parameterically larger than, the action itself. Here we can take the ultraviolet cutoff to be (at most) of order \(\Lambda\). So, for example, writing the coefficient of \(F_{\mu \nu}^4\) as \({1 \over M^6}\), the one loop correction to the \(F_{\mu \nu}^2\) coupling behaves as \[N^{0}{1 \over M^{6}}\int^\Lambda d^2 k {k^4 \over k^2} \sim N^0 {\Lambda^4 \over M^6}.\] This is (up to a factor of \(N\)) of order the leading term in the action, provided \(M \sim \Lambda\). We can place a sharper bound if we insist on a particular behavior of the \(1/N\) expansion; the result, needless to say, is in agreement with the explicit computation. The hierarchy of operators is also consistent with confinement. \(F^2\) is the most relevant operator of the low energy theory and the \(F^2\) term in the action makes the dominant contribution to the potential at large distances. For \(\Lambda_w \ll \Lambda\), the coefficient of \(F^2\) does not evolve with decreasing \(\Lambda_w\) It is crucial that the infrared theory is confining for any value of the coefficient of the \(F_{\mu \nu}^2\) term, \(1 \over e_0^2\). Changing \(e_0\), the coefficient of the most relevant operator in the infrared theory, just rescales lengths in the large distance theory. In four dimensions, we will argue for similar behavior. The theory exhibits non-trivial \(\theta\) dependence. We will consider \(\theta\)-dependence later in four dimensions, where our arguments for confinement will hold on the CP conserving subspace of couplings. The supersymmetric version of the theory microscopically includes fermionic partners of the scalar fields. It is convenient to work with the theory in the form described in, in which form it is relatively easy to explicitly break the supersymmetry. In terms of the scalars, \(n_i\) fermions \(\psi^i\), bosonic auxiliary fields \(\lambda, \sigma, \pi, A_\mu\) and fermionic auxiliariy fields \(\chi, \bar \chi\), the lagrangian is: \[{\cal L} = {N \over g^2} [n_i^* (\partial_\mu-i A_\mu-i \pi)^2n^i + \bar \psi_i (i \not{\hbox{\kern-4pt \(\partial\)}}-\not{\hbox{\kern-4pt \(A\)}}) \psi^i-\bar \psi_i m_\psi \psi^i-{1 \over 2}(\sigma^2 + \pi^2)-\alpha m_\psi \sigma\] \[~~~~-{1\over \sqrt{2}} \bar \psi^i(\sigma + i \pi \gamma_5) \psi^i + \lambda(n_i^* n^i-1) + \bar \chi n_i \psi^I + \bar \psi_i n^i \chi^* ].\] Here we have included explicit supersymmetry violating terms, proportional to \(m_\psi\). The term linear in \(\sigma\) is necessary for the renormalizability of the theory. Including \(m_\psi\), we can perform a one loop computation at large \(N\) (the leading effect) of the effective action. We can look for stationary points of this action. We can obtain, as in, the values of \(\sigma,\pi\), and \(\lambda\) at large \(N\) by computing the effective action at one loop. Setting \(\pi = A_\mu = 0\), and differentiating the action with respect to \(\lambda\) and \(\sigma\) yields two equations: \[N \int {d^2 k \over (2 \pi)^2} {1 \over k^2 + \lambda} = {N \over g^2}; \label{lambdaequation}\] \[N \int {d^2 k \over (2 \pi)^2} {\sqrt{2}({\sigma \over \sqrt{2}} + m_\psi)\over k^2 + ({\sigma \over \sqrt{2}} + m_\psi)^2} = N( {\sigma\over g^2}). \label{sigmaequation}\] Here we have noted that while in general there is a different renormalization of the fermion and boson kinetic terms, at large \(N\) they are the same, so the same factor of \(1/g^2\) appears in both [\[lambdaequation\]](#lambdaequation){reference-type="ref" reference="lambdaequation"} and [\[sigmaequation\]](#sigmaequation){reference-type="ref" reference="sigmaequation"}. The first equation is solved by \[\lambda \equiv \Lambda^2 =M^2 e^{-{4 \pi \over g^2}}.\] The second is solved, for small \(m_\psi\), by \(\sigma^2 = 2 \lambda\). We can think of the scale, \(\Lambda\), as a fixed physical scale, analogous to \(\Lambda_{QCD}\), measuring \(m_F = m_\psi + {\sigma \over \sqrt{2}}\) and other scales relative to this one. For large \(m_\psi\), \({\sigma \over \sqrt{2} } + m_\psi \sim m_\psi\). We can understand this if we think of integrating out the fermion, with mass much greater than \(\Lambda\), In this case, the effects on the dynamics at low energies are small. At scales below the scale \(\Lambda\), the system, in either limit, is described by the gauge field \(A_\mu\) This field acquires a non-trivial kinetic term at one loop. It does not create physical excitations, but it does give rise to a potential between static charged sources, which grows linearly at large distances; equivalently, it yields an area law for large Wilson loops. The vacuum polarization is given by: \[\pi_{\mu \nu}(q)= (\pi_{\mu \nu})_B(q)+ (\pi_{\mu \nu})_F(q)\] where \[(\pi_{\mu \nu})_B(q) = (g_{\mu \nu} q^2-q_\mu q_\nu) \pi_B(q^2); ~~\pi_B(q^2) = {1 \over 4 \pi} {1 \over m_B^2} ~~~~q^2 \ll m_B^2.\] \[(\pi_{\mu \nu })_F(q) = (g_{\mu \nu} q^2-q_\mu q_\nu) \pi_F(q^2);~~\pi_F(q^2) = {1 \over 4 \pi} {1 \over m_F^2} ~~~~q^2 \ll m_F^2.\] Each of these terms has the structure of a local interaction, \[-{1 \over 4 e_0^2} F_{\mu \nu}^2,\] with the bosonic term dominating for large \(m_F^2\). In either limit, the low energy theory is the \(U(1)\) gauge theory. We can think of a Wilsonian effective action, obtained by integrating out physics at scales above \(\Lambda_w\). \(\Lambda_w\) can be arbitrarily small; we still find a linear potential. This potential arises from the most relevant operator, \(F_{\mu \nu}^2\), with coefficient \(-{1 \over 4 e_0^2}\). This is a reflection of the fact that the linear potential is a phenomena of the far infrared in this theory--arbitrarily far. Less relevant operators yield contributions to the potential suppressed by powers of \((\Lambda R)^{-1}\). The coupling \(e_0^2\) is proportional to \(\Lambda=\Lambda(\Lambda_w)\), which we can now think of as a coupling in the Wilsonian theory. For small \(\Lambda_w\), examining equation [\[lambdaequation\]](#lambdaequation){reference-type="ref" reference="lambdaequation"}, we see that \(\Lambda\) goes to a constant as \(\Lambda_w \rightarrow 0\) as \(({\Lambda_w \over \Lambda})^2\). As a result, the coefficient of the separation \(R\) in the linear potential tends to a constant. We will argue for similar behavior in four dimensional gauge theory. We might have anticipated this result by noting, first, that the supersymmetric theory is confining. Because the theory is gapped, we expect this to remain true in the presence of a small \(m_\psi\). For small \(\Lambda_w \ll m_F,~\Lambda_w \ll \Lambda\), the theory is a \(U(1)\) gauge theory, characterized by the coefficient of the most relevant operator. As a result, the theory is confining for all \(\Lambda_w\), and therefore for all \(m_F\). The argument could have failed if there were two most relevant operators of the same dimension in the small \(\Lambda_w\) theory. Also, if the most relevant operator was marginal, it could conceivably have a fixed point for some value of the coupling. In four dimensions, we don't know the structure of the theory in the far infrared, but we can attempt to make an argument for confinement along these lines. It is worth noting that confinement occurs in a regime where there are no physical excitations of the system, so Greens functions of gauge-invariant local operators fall to zero exponentially rapidly with growing separation. Such correlators can be represented by contact terms in the low energy theory; the value of this contact term can be thought of as information which supplements the Wilsonian action. The Wilson loop is different; not being a correlation function of a local operator, it does not satisfy a simple spectral relation. Its large distance behavior can be computed from the Wilsonian action with arbitrarily small \(\Lambda_w\). There are lessons here for four dimensions. In particular, for small \(m_\lambda\), the theory with \(N_f=0\) is gapped, confining, and has non-trivial \(\theta\) dependence, implying a non-vanishing contact term in the \(F\tilde F\) two-point function. This is necessary to give the correct \(\theta\)-dependence, and should be thought of as a feature which supplements the Wilsonian action. Similarly, the large radius Wilson loop obeys an area law, which can be computed from the small \(\Lambda_w\) Wilsonian action. The question is: do these features change as we take \(m_\lambda\) large? What might change in this case as the theory flows to small \(\Lambda_w\)?
# The Essence of the Argument for Confinement in Four Dimensions {#generalargument}
Turning now to QCD, to address confinement, rather than attacking the question of the fate of the magnetic monopole condensate at large \(m_{\lambda}\), we focus on the heavy quark-antiquark potential. Taking \(R\) as the quark anti-quark separation (and \(q\) the corresponding momentum transfer), for confinement we can study arbitrarily large \(R\) (small \(q\)) for general \(m_\lambda\). In particular, we can take \(R \gg m_{\lambda}^{-1}\). In the broken supersymmetric gauge theory without matter fields, for sufficiently small \(m_\lambda\), the potential is linear. Because confinement is a feature of the far infrared, it makes sense to consider a Wilsonian effective action with scale \(\Lambda_w\). Typically in thinking about this action in the continuum, one takes \(\Lambda_w \gg \Lambda_{QCD}\), so that the action can be computed in a weak coupling approximation. This would be appropriate to the problem of large \(m_\lambda\), where integrating out the gaugino would yield a theory well described in this way. But the case of which we have knowledge is that of small \(m_\lambda\), so we'll be interested in considering a range of \(\Lambda_w\), including \(\Lambda_w\) well below the QCD scale. The problem is then not a weak coupling one. Consider, then, the softly broken \(N=1\) gauge theory without matter fields as a function of \(x={m_\lambda \over \Lambda_{QCD}}\). We will restrict, first, to \(\theta=0\). More precisely we will take the \(\theta\) parameter in the langrangian to be zero, \(m_\lambda\) to be real, and the phase of the gaugino condensate or of the field \(u\) of equation [\[seibergwittenvacua\]](#seibergwittenvacua){reference-type="ref" reference="seibergwittenvacua"} to be zero. For small values of \(x\), \(x < x_0\), for some \(x_0\), we know that the low energy theory confines and is gapped. Confinement is the result of some dynamics at scales well below the mass gap. The \(CP^N\) model illustrates this, as we have seen in section [3](#cpn){reference-type="ref" reference="cpn"}. There it is the gauge field in the far infrared which is responsible for confinement. This field, at low energies, does not create any of the particles of the theory, which are massive. Integrating out physics above a scale \(\Lambda_w\) slightly below the gluino mass, we are left with an effective theory with Wilsonian cutoff, \(\Lambda_w\) slightly less than \(m_\lambda\). The pure gauge theory is confining for some \({\Lambda_w \over \Lambda_{QCD}} < y_0\), where \(y_0\) is slightly less than \(x_0\). Now provided that the theory flows smoothly from larger \(y\) to smaller, confinement is established. For this to fail, the theory obtained by flow from large \(\Lambda_w\) to small \(\Lambda_w\), i.e. the theory obtained in the far infrared by integrating out a heavy gluino must be different than that obtained by integrating out a light gluino. This does not seem particularly plausible, and we will now argue that the flow is smooth. We first give two somewhat naive arguments for confinement for large \(m_\lambda\). Each relies on the fact of confinement for small gluino mass. The first starts by examining the diagrams of perturbation theory, and noting that a substantial enhancement of the individual contributions at large distances is required to obtain a linear potential. Such enhancements take the form of (powers of) logarithms of \(q\). For \(q \ll \Lambda_w\) and \(m_\lambda \sim \Lambda_w\), large logs only arise from diagrams with gluons.[^1] So only gauge bosons play a role in confinement. The confining potential, taken to be the sum of these diagrams, is then a feature of the pure gauge theory, with Wilsonian cutoff \(\Lambda_w\) and with a coupling \(g(\Lambda_w)\) in the effective action. The coefficient of the linear term in the potential, \(k\), depends only on the scale parameter of the effective theory with small \(\Lambda_w\), \(\tilde \Lambda_{QCD}\), \(k = a \tilde \Lambda_{QCD}^2\). Here \[\tilde \Lambda_{QCD} = \Lambda_w e^{-\int_{g_{ref}}^{g(\Lambda_w)}{dg^\prime \over \beta(g^\prime)}}. \label{tildelambda}\] In this equation, \(g_{ref}\) is some fixed reference value of the coupling. The form of this result will not change as we increase \(\Lambda_w\) in the bosonic theory, as the effect is just a change in the overall scale. So confinement survives at large \(m_\lambda\). One might question the reliability of the guidance provided by Feynman diagrams, but other known non-perturbative effects, such as instantons, are also dominated at large distances by the purely gluonic contributions. The usual large \(N\) picture, as we will discuss further below, assumes that features of QCD, at least at large \(N\), are reflected in the Feynman diagram expansion. As for \(\Lambda\) in the \(CP^N\) model, \(\tilde \Lambda\) tends to a constant as \(\Lambda_w \rightarrow 0\), as follows from the fact of confinement in the small \(m_\lambda\) theory. We can reframe this argument, referring to a naive understanding of the operators of QCD, an understanding which holds more rigorously at large \(N\). For this argument for confinement, we will make three assumptions:
1. We can write an effective action, at large distances, in terms of gauge fields \(A_\mu\) (\(F_{\mu \nu}\)). This is plausible, and, at large \(N\), a basic feature.
2. Operators have roughly their naive dimensions. In particular, there is a single marginal operator, \(F_{\mu \nu}^2\), and associated dimensionless coupling, \(g^2(\Lambda_w)\).
3. The theory with small \(m_\lambda\) and \(\Lambda_w < m_\lambda\), defines the infrared sector of actual QCD, for some values of the parameters of the effective lagrangian. This is the case for large \(m_\lambda\), where integrating out the gaugino generates a change in the marginal coupling and leaves over some irrelevant operators. In computing the potential at very large distances. With our assumption above the situation is similar, in that the contributions of irrelevant operators to the potential at large distances will be suppressed by powers of \((R \Lambda_w)^{-1}\).
4. There are no zeros of the beta function. As will see, and we can state more sharply in this formulation, we can actually rule out this possibility.
With these assumptions, the theory in the far infrared has only one parameter, a mass or inverse length scale, \(\tilde\Lambda_{QCD}(\Lambda_w)\). To define this, we proceed somewhat heuristically, writing an effective action which governs very low scale phenomena (\(\mu = m_\lambda,\Lambda_w\)) as \[{\cal L} =-{1 \over 4 g^2(\mu)} F_{\mu \nu}^2 + \rm \dots,\] and we have dropped higher dimension terms suppressed by powers of \(\mu\). Arguments of the type we gave for the \(CP^N\) models bound these higher dimension terms by powers of \(1/\Lambda_w\) or \({1 \over \tilde \Lambda_{QCD}}\). Indeed we can again define the scale \(\tilde \Lambda_{QCD}\) as in equation [\[tildelambda\]](#tildelambda){reference-type="ref" reference="tildelambda"}. We know that confinement occurs for small \(m_\lambda\). We know, as well, that the coefficient of the linear potential tends to a constant as \(\Lambda_w \rightarrow 0\). As a consequence, \(\tilde \Lambda_{QCD}\) also tends to a constant as \(\Lambda_w \rightarrow 0\). This is similar to the behavior of the \(CP^N\) theory. Confinement, then, depends only on the length parameter \(\tilde \Lambda_{QCD}\). As we increase \(m_{\lambda}\), \(\tilde \Lambda_{QCD}\) will change, but confinement will occur, simply associated with this different scale. We can argue against an obstruction as we vary \(g(\Lambda_w)\) by noting that a break in the flow would correspond to an infrared fixed point at some intermediate value of \(\Lambda_w\). If this were all there is, then the theory in this region, starting with some \(\Lambda_w\) would flow to this fixed point, which is inconsistent with the fact that it is gapped and confining. So there must be at least two such points (and in general an even number). But now the theory would have an ultraviolet fixed point starting at small \(\Lambda_w\), which is not consistent with the fact that the theory with small \(m_\lambda\) is asymptotically free. At the level of planar Feynman diagrams, we can make diagrammatic versions of these statements, and we will do that in the next section. The argument for confinement can be stated more generally. Confinement follows if:
1. In the low \(\Lambda_w\) theory, there is a single, most relevant operator. Correspondingly, for a given \(\Lambda_w\), there is a single parameter which characterizes the flow, \(g(\Lambda_w)\). \(g(\Lambda_w)\) flows with \(\Lambda_w\) according to the renormalization group.
2. There are no fixed points in \(g(\Lambda_w)\).
If these statements are true, then the system flows from large \(\Lambda_w\) to smaller \(\Lambda_w\) where the system is known to be confining. Equivalently, we can define a parameter \(\tilde \Lambda_{QCD}\), and again the confining parameter is independent of \(\Lambda_w\). If there were two most relevant operators, then as one increased \(m_\lambda\), it is conceivable that one could follows a trajectory which at some point becomes non-confining. For small values of \(g(\Lambda_w)\), there is only one such operator. Suppose *there are* two or more most relevant operators (degenerate in dimension) at larger \(g(\Lambda_w)\), i.e. for \(g(\Lambda_w)>g_0\). This would appear to imply that actual QCD is a multiparameter theory, both in the supersymmetric case and the non-supersymmetric one. Given all we know about the supersymmetric theory, the former possibility appears highly unlikely. Alternatively, two operators might cross in dimension at a particular value of the coupling. But in order that one theory track the higher point, one the lower point, the system would have to be finely tuned. In this more general framework, the possibility of a fixed point (or multiple fixed points) can again be ruled out by facts we know about the theory. In particular, if there were a single fixed point, then starting with small \(m_\lambda\) and \(\Lambda_w < m_\lambda\), the theory would flow at large distances to the fixed point. But this is inconsistent with the fact of confinement in the small \(m_\lambda\) theory. If there were a pair of fixed points (more generally an even number of fixed points), we would also encounter inconsistency with known facts of the small \(\Lambda_w\) theory.
# Confinement in the Large \(N\) Expansion With Broken Supersymmetry {#largen}
These arguments for confinement are particularly compelling in the large \(N\) limit. The main reason one might question the Feynman diagram argument is that Feynman diagrams possibly might not capture the essence of the theory, but at large \(N\), the planar diagrams *should* capture the theory's essential features. Also at large \(N\), the assumptions underlying the second argument are believed to hold: the system is described, in particular, in terms of gluons and gluinos, and the arguments used to bound the coefficients of operators in the effective action (and correspondingly the dimensions of these operators) should hold.
# Theories with \(N_f\) Light Quarks At Large and Finite \(N\) {#lightquarks}
We can extend these result to theories with \(N_f\) light quarks. Add to the action of this theory mass terms for the squarks and a mass for the gluino. For small values of the soft breaking parameters, one can find the ground state and calculate an effective action and the spectrum of the theory. The theory is gapped and confined. One can again ask what happens when the soft breakings are increased, eventually to values larger than the QCD scale. Once again, we can focus on the heavy quark-antiquark potential. At large distances, this is an infrared phenomenon, as before. Taking the fermion masses to be non-zero but small, then as we did for the light gluino, we can consider the far infrared theory, and argue that this theory is the one we have encountered in the theory with \(N_f=0\). Again this theory is described by a single parameter and is necessarily confining. At large \(N\) with \(N_f \ll N\), we can simply repeat the arguments of the \(N_f =0\) theory.
# The QCD Description of the \(\eta^\prime\) at Large \(N\) {#etaprime}
While often unstated, the assumption that the underlying degrees of freedom at large \(N\) in QCD in the far infrared are the gluon fields (and the dominance of planar diagrams) underlies the usual treatment of the \(\eta^\prime\) at large \(N\). The analysis relies heavily on the anomaly equation, \[\partial_\mu j^{\mu 5} = {2 N_f \over 16 \pi^2 } F \tilde F,\] or, in terms of the \(\eta^\prime\), \[f_\pi \partial^2 \eta^\prime = {2N_f \over 16 \pi^2}F \tilde F.\] From this, one can compute the \(\eta^\prime\) mass in terms of the leading contribution in \(1/N\) to the two point function of the \(\eta^\prime\) at distances of order a power of \(N\) times \(\Lambda_{QCD}^{-1}\) but smaller than a power of \(N\) times the QCD scale. The result is expressed in terms of the correlator \[U(k) = \langle (F(x) \tilde F(x) ~F(0)\tilde F(0) \rangle_k\] at zero (small, by a power of \(N\)) momentum. All of this is written explicitly in terms of the degrees of freedom arising from \(A^\mu\). Because the correlator is evaluated at zero momentum, and because of the observational fact that the theory is gapped, it is, as in the case of \(CP^N\), controlled by a contact term \[U(x) = K \delta(x).\] This contact term can be computed, for example, by computing the \(\theta^2\) term in the action for a given \(\Lambda_w\). Because of the infrared sensitivity of the computation and because the system is gapped, this term is invariant under changes of \(\Lambda_w\). Contributions to \(U(k)\) from operators of higher dimension than \(F_{\mu \nu}^2\) in the Wilsonian action are suppressed by powers of \({m_{\eta^\prime}^2 \over \Lambda^2} \sim N^{-1}\) This is all quite parallel to the questions which arise for confinement: one can study the deep infrared (small \(\Lambda_w\)) and work with the degrees of freedom of the original continuum action. The dominant term in the action is again the naive leading term.
# \(\theta\) Dependence {#thetadependence}
Once we allow a non-zero \(\theta\), or equivalently treat \(m_\lambda\) as complex, a richer phase structure is possible. In particular, if we take \(m_{\lambda}\) and \(\langle \lambda \lambda\rangle\) as real, we can introduce a \(\theta F \tilde F\) term, and our lagrangian now has two couplings of dimension four: \[{\cal L} =-{1 \over 4 g^2_{eff}(\Lambda_w)} F_{\mu \nu}^2 + {\theta \over 16 \pi^2} F \tilde F.\] Our discussion of the last section applies to the theory along the line \(\theta =0\). It established that there is confinement in this space. With non-zero \(\theta\), a more intricate phase structure arises, and there are trajectories which are sensitive to these phase changes. The supersymmetric gauge theory with small gaugino mass exhibits a branched structure, with phase transitions as one varies \(\theta\). Recall in particular that: \[m_\lambda \langle \lambda \lambda \rangle \propto e^{i (\theta + {{2 \pi k} \over N})}.\] As a result, the theory undergoes phase transitions as \(\theta\) varies. It is quite possible that there are regions in the space \((g_{eff}(\Lambda_w),\theta)\) with non-zero \(\theta\) which evolve under the renormalization group to unconfined regions.
# Conclusion: Even With Large Soft Breakings, SUSY QCD is Confining {#conclusions}
From these rather simple considerations, it appears that, by deforming supersymmetric QCD with large gaugino and squark masses, one obtains a theory which is gapped and confined. To summarize the basic argument:
1. The linear confining potential of the supersymmetric theories survives with small soft breakings, due to the presence of a mass gap.
2. The linear confining potential of the slightly broken supersymmetric theories can be computed at arbitrarily large distance.
3. One can work with a Wilsonian action with scale \(\Lambda_w\) small compared to the soft breakings and large compared to \(R^{-1}\).
4. For small \(\Lambda_w\) and large \(R\), only the gluon fields contribute to the potential.
5. The potential depends only on \(\tilde \Lambda_{QCD}\), a parameter which depends in the usual renormalization group fashion on the value of \(\Lambda_w\) and the most relevant coupling. Increasing \(m_\lambda\) just results in a rescaling of lengths.
The last two points are basic features of the perturbative expansion, but it is hard to see how they could fail to hold more generally--at long enough distances, the gaugino mass should be irrelevant in the technical sense. For large \(N\), the analysis is rigorous to the extent that large \(N\) is defined as a sum of an infinite set of planar diagrams, resummed in some way. For finite \(N\), a breakdown of the components of the basic argument would be quite bizarre Confinement in the theory with large breaking of supersymmetry, then, seems an inevitable consequence of the fact that the supersymmmetric theory is confining and that confinement is a feature of the (arbitrarily) far infrared. | {'timestamp': '2022-12-02T02:06:16', 'yymm': '2211', 'arxiv_id': '2211.17134', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17134'} |
# Introduction {#sec:intro}
-1.5ex With the advent of mobile devices that are equipped with microphones, wirelessly linking hearing aids to these devices has become increasingly popular. In such an acoustic sensor network, jointly processing all available microphones, i.e. the hearing aid microphones in conjunction with the external microphones, has been shown to be beneficial for, e.g., noise reduction as well as for direction of arrival (DOA) estimation. Large research effort has already been dedicated to DOA estimation, in particular also for binaural hearing aid applications. In we proposed a DOA estimation method for a binaural hearing aid setup, which exploits an external microphone at an unknown position. In this method all available microphone signals are used to estimate the so-called relative transfer function (RTF) vector between all microphones and a reference microphone on one of the hearing aids. The estimated RTF vector corresponding to only the hearing aid microphones is then compared against a database of anechoic prototype RTF vectors (e.g., obtained through measurement), by constructing a one-dimensional spatial spectrum for different directions. It should be noted that the element in the estimated RTF vector corresponding to the external microphone cannot be used in this comparison, since the position of the external microphone is generally unknown. The DOA of the target speaker relative to the binaural hearing aid setup is then estimated as the direction for which the spatial spectrum is maximized. Contrary to, in the external microphone does not need to be in the vicinity of the target speaker in order to capture a nearly clean speech reference, while contrary to no pre-trained representation of clean speech spectral envelopes is required for DOA estimation. Whereas the methods in exploit a single external microphone, in this paper we assume the availability of a calibrated array of external microphones. With calibrated array we mean that (similarly as for the binaural hearing aid) a database of anechoic prototype RTF vectors for several directions is available for this array. The prototype RTF vectors can be obtained, e.g., through measurement or when the geometry of the external microphone array is known (and assuming free-field sound propagation). It should be noted that the relative position of the external microphone array with respect to the binaural hearing aid setup is obviously unknown We propose a method where the calibrated array of external microphones assists in estimating the DOA of the target speaker. Similarly as in, the RTF vector between all available microphones and a reference microphone on one of the hearing aids is first estimated using the state-of-the-art covariance whitening method. Using the entire RTF vector and the prototype databases of both arrays, i.e. the binaural hearing aid as well as the external microphone array, we now define a two-dimensional spatial spectrum. The DOA of the target speaker is then estimated by determining the location of the main peak of this two-dimensional spatial spectrum. To further investigate the potential of the proposed method, we also consider a special case, by transforming the two-dimensional spatial spectrum into a one-dimensional spatial spectrum using a coordinate system transformation which requires prior information. The performance of the proposed binaural DOA estimation method is evaluated using recordings for a single static speaker in a reverberant acoustic scenario with diffuse-like noise. Experimental results show the benefit of exploiting a calibrated external microphone array compared to exploiting an uncalibrated external microphone array or using only the hearing aid microphones for DOA estimation.
# Signal model and notation {#sec:signalModel}
-1.5ex We consider an acoustic sensor network composed of two microphone arrays with a total of \(M = M_{\rm H}+M_{\rm E}\) microphones (see Fig. [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}): a binaural hearing aid setup consisting of \(M_{\rm H}\) head-mounted microphones (i.e., \(M_{\rm H}/2\) microphones on each hearing aid) and an array consisting of \(M_{\rm E}\) external microphones, which is assumed to be spatially separated from the binaural hearing aid setup. We consider a noisy and reverberant acoustic scenario with a single speaker that is located at DOA \(\theta\) relative to the local coordinate system of the binaural hearing setup and angle \(\theta_{\mathrm{E}}\) relative to the local coordinate system of the external microphone array (in the azimuthal plane). In the short-time Fourier transform (STFT) domain, the \(m\)-th microphone signal can be written as \[Y_{m}\left(k,l\right) = X_{m}\left(k,l\right) + N_{m}\left(k,l\right)\,, \quad m \in \left\{1,\dots,M\right\}\,, \label{eq:signalModel_micComponent}\] where \(X_{m}\left(k,l\right)\) and \(N_{m}\left(k,l\right)\) denote the speech and noise component in the \(m\)-th microphone signal, respectively, and \(k\in\left\{1,\dotsc,K\right\}\) and \(l\in\left\{1,\dotsc,L\right\}\) denote the frequency bin index and the frame index, respectively. Since all frequency bins are assumed to be independent and are hence processed independently, the index \(k\) will be omitted in the remainder of the paper where possible. Stacking all microphone signals into the vector \(\mathbf{y}\left(l\right) = \left[Y_{1}\left(l\right),\dotsc,Y_{M_{\rm H}}\left(l\right),\dotsc,Y_{M}\left(l\right)\right]^{T} \in \mathbb{C}^{M}\), where \(\left(\cdot\right)^{T}\) denotes transposition, the noisy microphone signals can be written as \(\mathbf{y}\left(l\right) = \mathbf{x}\left(l\right) + \mathbf{n}\left(l\right)\), where the speech vector \(\mathbf{x}\left(l\right)\) and the noise vector \(\mathbf{n}\left(l\right)\) are defined similarly as \(\mathbf{y}\left(l\right)\). Assuming that the speech vector can be split into a direct-path component \(\mathbf{x}^{\rm DP}\left(l\right)\) and a reverberant component \(\mathbf{x}^{\rm R}\left(l\right)\) and assuming that the multiplicative transfer function approximation holds for the direct-path component, the speech vector \(\mathbf{x}\left(l\right)\) can be written as \[\mathbf{x}\left(l\right) = \mathbf{x}^{\rm DP}\left(l\right) + \mathbf{x}^{\rm R}\left(l\right) = \begin{bmatrix} \mathbf{a}_{\mathrm{H}}\left(\theta\right)\\ \mathbf{a}_{\mathrm{E}}\left(\theta_{\mathrm{E}}\right)\\ \end{bmatrix}S\left(l\right) + \mathbf{x}^{\rm R}\left(l\right)\,,\label{eq:modelSpeechComponentATF}\] where the \(M_{\rm H}\)-dimensional vector \(\mathbf{a}_{\mathrm{H}}\left(\theta\right)\) and the \(M_{\rm E}\)-dimensional vector \(\mathbf{a}_{\mathrm{E}}\left(\theta_{\mathrm{E}}\right)\) denote the direct-path acoustic transfer function vectors between the speaker \(S\) and the hearing aid microphones and the external microphones, respectively. By introducing the direct-path RTF vectors \(\mathbf{g}_{\mathrm{H}}\left(\theta\right) = \mathbf{a}_{\mathrm{H}}\left(\theta\right)/A_{1}\left(\theta\right)\) and \(\mathbf{g}_{\mathrm{E}}\left(\theta_{\mathrm{E}}\right) = \mathbf{a}_{\mathrm{E}}\left(\theta_{\mathrm{E}}\right)/A_{\mathrm{E},1}\left(\theta_{\mathrm{E}}\right)\), where the first microphone of each array is chosen as reference microphone without loss of generality, we can rewrite the direct-path speech component in [\[eq:modelSpeechComponentATF\]](#eq:modelSpeechComponentATF){reference-type="eqref" reference="eq:modelSpeechComponentATF"} as \[\mathbf{x}^{\rm DP}\left(l\right) = \begin{bmatrix} A_{1}\left(\theta\right)\mathbf{g}_{\mathrm{H}}\left(\theta\right)\\ A_{\mathrm{E},1}\left(\theta_{\mathrm{E}}\right)\mathbf{g}_{\mathrm{E}}\left(\theta_{\mathrm{E}}\right)\\ \end{bmatrix}S\left(l\right)=\mathbf{g}X_{1}^{\rm DP}\left(l\right)\,,\label{eq:modelDPSpeechComponentRTF}\] where the \(M\)-dimensional vector \(\mathbf{g}\) denotes the direct-path RTF vector between all microphones and the reference microphone of the hearing aid and \(X_{1}^{\rm DP}\left(l\right)\) denotes the direct-path speech component in the reference microphone of the hearing aid. Condensing the noise and reverberation components into the undesired component \(\mathbf{u}\left(l\right) = \mathbf{n}\left(l\right) + \mathbf{x}^{\rm R}\left(l\right)\), the vector of noisy microphone signals can be written as \(\mathbf{y}\left(l\right) = \mathbf{x}^{\rm DP}\left(l\right) + \mathbf{u}\left(l\right)\). The RTF vectors \(\mathbf{g}_{\rm{H}}\left(\theta\right)\) and \(\mathbf{g}_{\rm{E}}\left(\theta_{\rm{E}}\right)\) can be both extracted from the RTF vector \(\mathbf{g}\) in [\[eq:modelDPSpeechComponentRTF\]](#eq:modelDPSpeechComponentRTF){reference-type="eqref" reference="eq:modelDPSpeechComponentRTF"} as \[\boxed{\mathbf{g}_{\rm{H}}\left(\theta\right) = \mathbf{E}_{\rm{H}}\mathbf{g}\,, \quad \mathbf{g}_{\rm{E}}\left(\theta_{\rm{E}}\right) = \frac{\mathbf{E}_{\rm{E}}\mathbf{g}}{\mathbf{e}_{1,\rm{E}}^{T}\mathbf{E}_{\rm{E}}\mathbf{g}}} \label{eq:extractRTFvec_theo}\] \[\mathbf{E}_{\rm{H}} = \left[\mathbf{I}_{M_{\rm H}\timesM_{\rm H}},\mathbf{0}_{M_{\rm H}\timesM_{\rm E}}\right]\,,\quad \mathbf{E}_{\rm{E}} = \left[\mathbf{0}_{M_{\rm E}\timesM_{\rm H}},\mathbf{I}_{M_{\rm E}\timesM_{\rm E}}\right]\,,\] where \(\mathbf{I}_{N\times N}\) denotes an \(N\times N\)-dimensional identity matrix, \(\mathbf{0}_{N\times N^{\prime}}\) denotes an \(N\times N^{\prime}\) matrix of zeros, and \(\mathbf{e}_{1,\rm{E}}=\left[1,0,\dotsc,0\right]^{T}\) denotes the \(M_{\rm E}\)-dimensional selection vector. Both for the binaural hearing aid setup as well as for the external microphone array, we assume that a database of anechoic prototype RTF vector is available (referred to as calibrated array). The database for the binaural hearing aid setup is denoted as \(\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right),~i=1...,I\), while the database for the external microphone array is denotes as \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right),~ j=1...,J\). Assuming that the direct-path speech component is uncorrelated with the undesired component, the \(M\times M\)-dimensional covariance matrix of the noisy microphone signals can be written as \[\boldsymbol{\Phi}_{\rm y}\left(l\right) = \myExpectation{\mathbf{y}\left(l\right)\mathbf{y}^{H}\left(l\right)} = \mathbf{g} \mathbf{g}^{H} \Phi_{\rm{x}}^{\rm{DP}}\left(l\right) + \boldsymbol{\Phi}_{\rm u}\left(l\right)\,, \label{eq:def_PhiY}\] where \(\left(\cdot\right)^{H}\) and \(\myExpectation{\cdot}\) denote complex transposition and expectation operators, respectively, \(\Phi_{\rm{x}}^{\rm{DP}}\left(l\right) = \myExpectation{|X_{1}^{\rm DP}\left(l\right)|^{2}}\) denotes the power spectral density of the direct-path speech component in the reference microphone, and \(\boldsymbol{\Phi}_{\rm u}\left(l\right)= \myExpectation{\mathbf{u}\left(l\right)\mathbf{u}^{H}\left(l\right)}\) denotes the covariance matrix of the undesired component. The \(M_{\rm H}\timesM_{\rm H}\)-dimensional covariance matrices \(\boldsymbol{\Phi}_{\rm{y,H}}\left(l\right)\) and \(\boldsymbol{\Phi}_{\rm{u,H}}\left(l\right)\) corresponding to the noisy microphone signals and the undesired components in the hearing aid microphones can be extracted from [\[eq:def_PhiY\]](#eq:def_PhiY){reference-type="eqref" reference="eq:def_PhiY"} as \[\boldsymbol{\Phi}_{\rm{y,H}}\left(l\right) = \mathbf{E}_{\rm{H}}\boldsymbol{\Phi}_{\rm y}\left(l\right)\mathbf{E}_{\rm{H}}^{T}\,, \quad \boldsymbol{\Phi}_{\rm{u,H}}\left(l\right) = \mathbf{E}_{\rm{H}}\boldsymbol{\Phi}_{\rm u}\left(l\right)\mathbf{E}_{\rm{H}}^{T}\.\]
# RTF-vector-based DOA estimation {#sec:DOAest}
-1.5ex To estimate the DOA \(\theta\) of the target speaker relative to the binaural hearing setup, in this section we propose an extension of the RTF-vector-based DOA estimation method from. In Section [3.1](#ssec:DOAest_noCoop){reference-type="ref" reference="ssec:DOAest_noCoop"} we review the existing method from, where either no external microphone array or an uncalibrated external microphone array is used. In Section [3.2](#ssec:DOAest_Coop){reference-type="ref" reference="ssec:DOAest_Coop"} we propose a method to jointly use both calibrated arrays, i.e. the binaural hearing aid setup and the external microphone array for DOA estimation.
## Baseline RTF-vector-based DOA estimation {#ssec:DOAest_noCoop}
-1ex To estimate the \(M_{\rm H}\)-dimensional RTF vector \(\mathbf{g}_{\rm{H}}\) corresponding to the binaural hearing aid setup, we use the state-of-the-art covariance whitening (CW) method in each time-frequency bin. This RTF vector can be estimated from only the hearing aid microphone signals as \[\begin{aligned}
\hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW}\right)}\left(l\right) &=f\left(\hat{\boldsymbol{\Phi}}_{\rm{y,H}}\left(l\right),\hat{\boldsymbol{\Phi}}_{\rm{u,H}}\left(l\right)\right)\label{eq:estRTF_CW}\,,\\ f\left(\tilde{\boldsymbol{\Phi}}_{\rm{y}},\tilde{\boldsymbol{\Phi}}_{\rm{u}}\right)&=\frac{\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{1/2}\myPrincipalEigenvec{\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{-1/2}\tilde{\boldsymbol{\Phi}}_{\rm{y}}\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{-H/2}}}{\tilde{\mathbf{e}}_{1}^{T}\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{1/2}\myPrincipalEigenvec{\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{-1/2}\tilde{\boldsymbol{\Phi}}_{\rm{y}}\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{-H/2}}}\,, \end{aligned}\] where \(\myPrincipalEigenvec{\cdot}\) denotes the principal eigenvector of a matrix, \(\tilde{\boldsymbol{\Phi}}_{\rm{u}}^{1/2}\) denotes a square-root decomposition (e.g., Cholesky decomposition) of \(\tilde{\boldsymbol{\Phi}}_{\rm{u}}\) and \(\tilde{\mathbf{e}}_{1}=\left[1,0,\dotsc,0\right]^{T}\) denotes a selection vector of same dimensionality as the matrix \(\tilde{\boldsymbol{\Phi}}_{\rm{u}}\). Alternatively, the RTF vector \(\mathbf{g}_{\rm{H}}\) can be estimated from all microphone signals, i.e. the hearing aid and external microphone signals, as \[\begin{aligned}
\hat{\mathbf{g}}^{\left(\rm{CW-E}\right)}\left(l\right) &=f\left(\hat{\boldsymbol{\Phi}}_{\rm y}\left(l\right),\hat{\boldsymbol{\Phi}}_{\rm u}\left(l\right)\right)\,,\label{eq:estRTF_CWE}\\ \hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW-E}\right)}\left(l\right) &= \mathbf{E}_{\rm{H}}\,\hat{\mathbf{g}}^{\left(\rm{CW-E}\right)}\left(l\right)\.\label{eq:estRTF_CWEHA} \end{aligned}\] Due to the involved eigenvalue and square-root decomposition, it can be easily shown that in general \(\hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW-E}\right)}\left(l\right)\neq\hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW}\right)}\left(l\right)\). In it was experimentally shown that the RTF vector \(\mathbf{g}_{\rm{H}}\) can be more accurately estimated using all available microphone signals than using only the hearing aid microphone signals. Based on the estimated RTF vector \(\hat{\mathbf{g}}_{\rm{H}}\left(k,l\right)\) corresponding to only the hearing aid microphone signals and the database of anechoic prototype RTF vectors \(\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\), a one-dimensional spatial spectrum is constructed as \[P\left(l,\theta_{i}\right) =-\sum_{k=2}^{K-1}d\left(\hat{\mathbf{g}}_{\rm{H}}\left(k,l\right),~\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\right)\.\label{eq:spatialSpectrum_FejginEUSIPCO2021}\] The DOA of the speaker is estimated as the location of the main peak of this spatial spectrum. It should be noted that the spatial spectrum in [\[eq:spatialSpectrum_FejginEUSIPCO2021\]](#eq:spatialSpectrum_FejginEUSIPCO2021){reference-type="eqref" reference="eq:spatialSpectrum_FejginEUSIPCO2021"} is obtained via frequency-averaging of narrowband spatial spectra in order to make the DOA estimation more robust against estimation errors in the RTF vector at individual frequencies. Each narrowband spatial spectrum is obtained by assessing the similarity between the estimated RTF vector and an anechoic prototype RTF vector. Inspired by, in this paper we consider a similarity measure based on the \(\ell_{2}\)-norm, i.e. \[d\left(\mathbf{a},\mathbf{b}\right) = \|\exp\left(i\angle\mathbf{a}\right)-\exp\left(i\angle\mathbf{b}\right)\|_{2}\,,\label{eq:def_PHATL2norm}\] where the operators \(\angle\cdot\) and \(\exp\left(\cdot\right)\) are applied element-wise to extract the phase and apply the exponential.
## Assisted RTF-vector-based DOA estimation {#ssec:DOAest_Coop}
-1ex When considering an uncalibrated external microphone array as in Section [3.1](#ssec:DOAest_noCoop){reference-type="ref" reference="ssec:DOAest_noCoop"}, it should be realized that only the database \(\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\) and thus only the estimated RTF vector \(\hat{\mathbf{g}}_{\rm{H}}\left(k,l\right)\) corresponding to the hearing aid microphones can be considered for the construction of a one-dimensional spatial spectrum as in [\[eq:spatialSpectrum_FejginEUSIPCO2021\]](#eq:spatialSpectrum_FejginEUSIPCO2021){reference-type="eqref" reference="eq:spatialSpectrum_FejginEUSIPCO2021"}. Since in this paper the external microphone array is assumed to be calibrated, i.e. a database \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right)\) of anechoic prototype RTF vectors is available, the entire estimated RTF vector \(\hat{\mathbf{g}}\left(k,l\right)\) and both prototype databases can-and should-be utilized. We first propose to construct a two-dimensional spatial spectrum that exploits both RTF vector databases \(\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\) and \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right)\) jointly with the entire RTF vector \(\hat{\mathbf{g}}^{\left(\rm{CW-E}\right)}\left(k,l\right)\) in order to exploit spatial correlations between microphone signals of both arrays for improving the DOA estimation. We then consider a special case, requiring prior information about relative position and orientation of both arrays. We define two \(M\)-dimensional concatenated RTF vectors, similarly to \(\mathbf{g}\) in [\[eq:modelDPSpeechComponentRTF\]](#eq:modelDPSpeechComponentRTF){reference-type="eqref" reference="eq:modelDPSpeechComponentRTF"}. The vector \[\bar{\mathbf{g}}^{\mathrm{joint}}\left(k,\theta_{i},\theta_{\mathrm{E},j}\right) = \begin{bmatrix} \bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\\ \bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right) \end{bmatrix}\.\label{eq:gProto2D}\] concatenates the anechoic prototype RTF vectors from both databases and is obtained for each pair of DOAs \(\theta_{i}\) and \(\theta_{\mathrm{E},j}\). Similarly, the vector \[\tilde{\mathbf{g}}\left(k,l\right) = \begin{bmatrix} \hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW-E}\right)}\left(k,l\right)\\ \hat{\mathbf{g}}_{\rm{E}}^{\left(\rm{CW-E}\right)}\left(k,l\right) \end{bmatrix} \label{eq:tildeRTF_stack}\] concatenates the estimated RTF vector corresponding to the hearing aid microphones \(\hat{\mathbf{g}}_{\rm{H}}^{\left(\rm{CW-E}\right)}\left(k,l\right)\) in [\[eq:estRTF_CWEHA\]](#eq:estRTF_CWEHA){reference-type="eqref" reference="eq:estRTF_CWEHA"} and the estimated RTF vector corresponding to the external microphones, which is equal to \[\hat{\mathbf{g}}_{\rm{E}}^{\left(\rm{CW-E}\right)}\left(l\right) =\frac{\mathbf{E}_{\rm{E}}\,\hat{\mathbf{g}}^{\left(\rm{CW-E}\right)}\left(l\right)}{\mathbf{e}_{1,\rm{E}}^{T}\mathbf{E}_{\rm{E}}\,\hat{\mathbf{g}}^{\left(\rm{CW-E}\right)}\left(l\right)} \label{eq:gEhat}\] similarly to [\[eq:extractRTFvec_theo\]](#eq:extractRTFvec_theo){reference-type="eqref" reference="eq:extractRTFvec_theo"}. A two-dimensional spatial spectrum is now constructed by assessing the similarity between the concatenated vectors in [\[eq:tildeRTF_stack\]](#eq:tildeRTF_stack){reference-type="eqref" reference="eq:tildeRTF_stack"} and [\[eq:gEhat\]](#eq:gEhat){reference-type="eqref" reference="eq:gEhat"}, i.e. \[\boxed{P^{\rm{joint}}\left(l,\theta_{i},\theta_{\mathrm{E},j}\right) =-\sum_{k=2}^{K-1}d\left(\tilde{\mathbf{g}}\left(k,l\right),~\bar{\mathbf{g}}^{\mathrm{joint}}\left(k,\theta_{i},\theta_{\mathrm{E},j}\right)\right)}\label{eq:spatialSpectrum_2D}\] with the similarity measure defined in [\[eq:def_PHATL2norm\]](#eq:def_PHATL2norm){reference-type="eqref" reference="eq:def_PHATL2norm"}. The DOA of the speaker is estimated as the location of the main peak of this two-dimensional spatial spectrum. Fig. [\[fig:2SspatialSpectrum\]](#fig:2SspatialSpectrum){reference-type="ref" reference="fig:2SspatialSpectrum"} depicts an exemplary spectrum for a static speaker in a reverberant environment with spatially diffuse babble noise. It should be realized that the main peak consists of the estimated DOA \(\hat{\theta}\left(l\right)\) as well as the estimated angle \(\hat{\theta}_{\rm{E}}\left(l\right)\) of the speaker as seen from both local coordinate systems (compare Fig. [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}). In this paper, however, we will evaluate the estimation accuracy of \(\hat{\theta}\left(l\right)\) only, as we are mainly interested in how the array of external microphones assists the binaural hearing aid setup in estimating the DOA \(\theta\). To further investigate the potential of the proposed method, we also consider a special case that transforms the two-dimensional spatial spectrum in [\[eq:spatialSpectrum_2D\]](#eq:spatialSpectrum_2D){reference-type="eqref" reference="eq:spatialSpectrum_2D"} into a one-dimensional spatial spectrum. Contrary to \(P^{\rm{joint}}\left(l,\theta_{i},\theta_{\mathrm{E},j}\right)\) that considers each pair of DOAs \(\theta_{i}\) and \(\theta_{\mathrm{E},j}\) in this special case only a subset of pairs is considered. In particular, only those pairs of DOAs \(\theta_{i}\) and \(\theta_{\mathrm{E},i}\) are considered where the respective prototype RTF vectors \(\bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{i}\right)\) and \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},i}\right)\) are spatially matched, i.e. steer towards the same candidate position. To obtain the respective DOA pairs prior information about the relative position and orientation of both arrays is required. Based on this information, a coordinate system transformation between the two local coordinate systems can be applied, i.e. \(\theta_{\mathrm{E},i} = g\left(\theta_{i},\theta_{\mathrm{E},j}\right)\). Enforcing this matching condition makes the resulting spatial spectrum one-dimensional, i.e. \[P^{\rm{match}}\left(l,\theta_{i}\right) =-\sum_{k=2}^{K-1}d\left(\tilde{\mathbf{g}}\left(k,l\right),\bar{\mathbf{g}}^{\rm{match}}\left(k,\theta_{i},\theta_{\mathrm{E},i}\right)\right)\.\label{eq:spatialSpectrum_match}\] Similarly to [\[eq:spatialSpectrum_2D\]](#eq:spatialSpectrum_2D){reference-type="eqref" reference="eq:spatialSpectrum_2D"}, the DOA of the speaker is estimated as the location of the main peak.
# Experimental results {#sec:experiments}
-1.5ex In this section we compare the DOA estimation accuracy when the exploited array of external microphone is calibrated or not. The experimental setup is described in Section [4.1](#ssec:detailsSetup){reference-type="ref" reference="ssec:detailsSetup"} and the considered algorithms and implementation details are described in Section [4.2](#ssec:algos){reference-type="ref" reference="ssec:algos"}. The results are presented and discussed Section [4.3](#ssec:results){reference-type="ref" reference="ssec:results"}.
## Experimental setup and implementation details {#ssec:detailsSetup}
-1ex For the experiments we used signals that were recorded in a laboratory at the University of Oldenburg with dimensions about \(7\times6\times2.7 \mathrm{m}^{3}\), where the reverberation time can be adjusted by means of absorber panels, which are mounted to the walls and the ceiling. The reverberation time is set to approximately \(T_{\rm 60}\approx 1250\mathrm{ms}\). A dummy head with a binaural hearing aid setup (\(M_{\rm H}\) = 4) is placed approximately in the center of the laboratory. For this hearing aid setup a database of anechoic prototype RTF vectors is obtained from measured anechoic binaural room impulse responses (BRIRs) with an angular resolution of \(5^{\circ}\) (\(I = 72\)). As an array of external microphones we consider a second dummy head with the same binaural hearing setup (\(M_{\rm E}\) = 4) that is located about \(1\mathrm{m}\) and about \(80^{\circ}\) to the right side from the first dummy head (see Fig. [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}). As the same binaural hearing aid setup is considered for the external microphone array, \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right) = \bar{\mathbf{g}}_{\rm{H}}\left(k,\theta_{\mathrm{E},j}\right)\) and \(J=I\). The target speaker was a male English speaker played back via a loudspeaker. Relative to the first dummy head 9 different speaker DOAs in the range \(\left[-160^{\circ},-120^{\circ},\dotsc,160^{\circ}\right]\) at about \(2\mathrm{m}\) distance are considered. The speech signal is constantly active and is approximately \(4\mathrm{s}\) long. Diffuse-like noise is generated with four loudspeakers facing the corners of the laboratory, playing back different multi-talker recordings. The speech and noise components are recorded separately and are mixed at \(0\mathrm{dB}\) signal-to-noise ratio (SNR) averaged over all microphones of the first dummy head. All microphone signals are assumed to be synchronized such that latency aspects can ne neglected.
## Algorithms and implementation details {#ssec:algos}
-1ex We compare the following algorithms to assess whether a calibrated external microphone array can assist the binaural hearing setup in estimating the speaker DOA \(\theta\). The notation X/Y means that the microphone array X is used to estimate an RTF vector and that the microphone array Y is used to construct a spatial spectrum, where either the binaural hearing aid setup (denoted as H) alone or jointly with the external microphone array (denoted as H+E) are considered as options.
- H/H: considering only the hearing aid microphones of the first dummy head, both to estimate the RTF vector \(\hat{\mathbf{g}}_{\rm{H}}\left(k,l\right)\) in [\[eq:estRTF_CW\]](#eq:estRTF_CW){reference-type="eqref" reference="eq:estRTF_CW"} as well as to construct the spatial spectrum \(P\left(l,\theta_{i}\right)\) in [\[eq:spatialSpectrum_FejginEUSIPCO2021\]](#eq:spatialSpectrum_FejginEUSIPCO2021){reference-type="eqref" reference="eq:spatialSpectrum_FejginEUSIPCO2021"}.
- H+E/H: considering all microphones to estimate the RTF vector \(\hat{\mathbf{g}}_{\rm{H}}\left(k,l\right)\) in [\[eq:estRTF_CWEHA\]](#eq:estRTF_CWEHA){reference-type="eqref" reference="eq:estRTF_CWEHA"} but considering only the hearing aid microphones of the first dummy head to construct the spatial spectrum \(P\left(l,\theta_{i}\right)\) in [\[eq:spatialSpectrum_FejginEUSIPCO2021\]](#eq:spatialSpectrum_FejginEUSIPCO2021){reference-type="eqref" reference="eq:spatialSpectrum_FejginEUSIPCO2021"}.
- H+E/H+E (2D): considering all microphones, both to estimate the concatenated RTF vector \(\tilde{\mathbf{g}}\left(k,l\right)\) in [\[eq:tildeRTF_stack\]](#eq:tildeRTF_stack){reference-type="eqref" reference="eq:tildeRTF_stack"} as well as to construct the two-dimensional spatial spectrum \(P^{\rm{joint}}\left(l,\theta_{i},\theta_{\mathrm{E},j}\right)\) in [\[eq:spatialSpectrum_2D\]](#eq:spatialSpectrum_2D){reference-type="eqref" reference="eq:spatialSpectrum_2D"}.
- H+E/H+E (match): as a special case of H+E/H+E (2D) also all microphones are considered, both to estimate the concatenated RTF vector \(\tilde{\mathbf{g}}\left(k,l\right)\) in [\[eq:tildeRTF_stack\]](#eq:tildeRTF_stack){reference-type="eqref" reference="eq:tildeRTF_stack"} as well as to construct the one-dimensional spatial spectrum \(P^{\rm{match}}\left(l,\theta_{i}\right)\) in [\[eq:spatialSpectrum_match\]](#eq:spatialSpectrum_match){reference-type="eqref" reference="eq:spatialSpectrum_match"} exploiting prior knowledge about the microphone configuration. Due to the enforced matching condition the spatial spectrum \(P^{\rm{match}}\left(l,\theta_{i}\right)\) consists of only 20 candidate speaker positions.
The microphone signals (sampling frequency \(f_{\rm s} = 16\mathrm{kHz}\)) are processed in the STFT domain using \(32\mathrm{ms}\) square-root Hann windows with \(50\%\) overlap. For each time-frequency (TF) bin the covariance matrix of the noisy microphone signals \(\boldsymbol{\Phi}_{\rm y}\left(k,l\right)\) and the covariance matrix of the undesired components \(\boldsymbol{\Phi}_{\rm u}\left(k,l\right)\) are estimated recursively during speech-and-noise TF bins and noise-only TF bins as \[\begin{aligned}
\hat{\boldsymbol{\Phi}}_{\rm y}\left(k,l\right) &= \alpha_{\rm y}\hat{\boldsymbol{\Phi}}_{\rm y}\left(k,l-1\right) + \left(1-\alpha_{\rm y}\right)\mathbf{y}\left(k,l\right)\mathbf{y}^{H}\left(k,l\right)\,,\label{eq:SOS_RyUpdate}\\ \hat{\boldsymbol{\Phi}}_{\rm u}\left(k,l\right) &= \alpha_{\rm u}\hat{\boldsymbol{\Phi}}_{\rm u}\left(k,l-1\right) + \left(1-\alpha_{\rm u}\right)\mathbf{y}\left(k,l\right)\mathbf{y}^{H}\left(k,l\right)\,,\label{eq:SOS_RuUpdate} \end{aligned}\] where the smoothing factors \(\alpha_{\rm y}\) and \(\alpha_{\rm u}\) correspond to time constants of \(250\mathrm{ms}\) and \(500\mathrm{ms}\), respectively. The speech-and-noise TF bins are discriminated from noise-only TF bins based on estimated speech presence probabilities in the microphones of the first dummy head, which are averaged and thresholded. To assess the DOA estimation performance, we average the localization accuracy over the considered \(9\) speaker DOAs, where the localization accuracy is defined as the percentage of frames that are correctly localized within a range of \(\pm5^{\circ}\).
## Results {#ssec:results}
-1ex Fig. [\[fig:res\]](#fig:res){reference-type="ref" reference="fig:res"} depicts the average localization accuracy for the four investigated algorithms. First, it can be observed that for the algorithm H/H using only the hearing aid microphones of the first dummy head and for the algorithm H+E/H, where the external microphones are exploited only to estimate the RTF vector, there is only a minor performance difference (average localization accuracy for both algorithms about \(80\%\)). This result is in line with the high-reverberation results reported in. Second, considering the H+E/H+E algorithms, which both exploit the external microphones to estimate the RTF vector as well as to construct a spatial spectrum, the results clearly show that both algorithms yield a larger average localization accuracy than the algorithms that do not consider external microphones to construct a spatial spectrum. In case of the H+E/H+E (2D) algorithm, where a two-dimensional spatial spectrum is constructed, this improved average localization accuracy is due to exploitation of the additional database of prototype RTF vectors \(\bar{\mathbf{g}}_{\rm{E}}\left(k,\theta_{\mathrm{E},j}\right)\). Third, in case of the H+E/H+E (match) algorithm, where prior knowledge about the microphone configuration is additionally exploited, the average localization accuracy can be significantly improved. Based on these results, the potential of assisted DOA estimation is clearly demonstrated.
# Conclusions {#sec:print}
In this paper we explored how calibrated external microphone arrays can be exploited to assist in estimating the DOA of a target speaker. We extended a recently proposed binaural RTF-vector-based DOA estimation method that considered only uncalibrated external microphone arrays to exploit calibrated external microphone arrays. We proposed to exploit the availability of two databases of anechoic prototype RTF vectors, i.e. the binaural hearing aid as well as the external microphone, in order to construct a two-dimensional spatial spectrum from which the DOA of the speaker can be estimated. We compared RTF-vector-based DOA estimation algorithm that either exploit a calibrated array of external microphones or not. Experimental results clearly demonstrate a benefit of exploiting a calibrated external microphone array compared to not exploiting a calibrated external microphone array or using only the hearing aid microphones for DOA estimation. Thus, the benefit of assisted DOA estimation is clearly demonstrated. | {'timestamp': '2022-12-01T02:20:24', 'yymm': '2211', 'arxiv_id': '2211.17202', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17202'} |
# Introduction and summary
Entanglement entropy is now a central topic in the study of QFT. Entanglement entropy itself is UV divergent, so many of the derivations/proofs of important results involving entanglement entropy must play a delicate game of UV regularization. We would like to study quantities insensitive to this regularization procedure. One approach is to study new quantities, different from entanglement entropy, that are well defined in the continuum limit. In favorable circumstances these UV finite quantities can be thought of as approaching a regularized version of entanglement entropy in certain limits. A prominent example is half the mutual information \(I(A:B)/2\) of two spatial regions \(A,B\) on a fixed Cauchy slice that are almost complement to each other, but leave a small finite corridor between the entangling surfaces. The appropriate limit sends the size of the corridor to zero. Another quantity, the focus of this paper, is half the reflected entropy \(S_R(A:B)/2\) which uses this same geometric setup for \(A,B\). An advantage of reflected entropy compared to the mutual information is that the reflected entropy is an actual von Neumann entropy of some density matrix-the reflected density matrix. Thus we can use the same regulator to study the reflected entanglement spectrum as a proxy for the regular entanglement spectrum. For a 2d (non-chiral) CFT the entanglement spectrum for \(A\) a single interval of length \(L\) was computed by Calabrese-Lafevre. The result reads: \[D(\lambda) =\delta(\lambda-\lambda_{\max}) + \frac{\ln \lambda_{\max}^{-1}}{\lambda} f( \ln \lambda_{\max}^{-1} \ln \lambda_{\max}/\lambda) \qquad f(t) = I_1(2t^{1/2}) t^{-1/2} \theta(t)\] where \(\lambda_{\max} = (L/\epsilon)^{-c/6}\) and \(I_1\) is the Bessel function. This reproduces the expected Renyi entropies: \[\exp(-S_n(A)/(n-1)) = \int d \lambda \lambda^{n} D(\lambda)\] computed in. There are several undesirable features to this formula. The spectrum depends on the UV cutoff \(\epsilon\) and so will depend on the regularization procedure. The spectrum is continuous (aside from the single delta function at the edge of the continuum). It is also universal, only depending on the central charge of the CFT and none of the other CFT data such as the operator spectrum and OPE coefficients. In contrast the reflected entanglement spectrum is UV insensitive, discrete and depends on the operator spectrum and OPE coefficients of the CFT. We now give a brief introduction to reflected entropy and the associated spectrum, see for further details. Given a density matrix \(\rho\) acting on a finite-dimensional Hilbert space \(\mathcal{H}\), one can form the *canonical purification* \(\ket{\sqrt{\rho}}\) by interpreting \(\sqrt{\rho}\) as a state in the doubled Hilbert space \[\label{canonical} \ket{\sqrt{\rho}} \in \text{End}(\mathcal{H}) = \mathcal{H}\otimes \mathcal{H}^*,\] where \(\mathcal{H}^*\) is the dual of \(\mathcal{H}\). This doubled Hilbert space is equipped with the inner product \(\Braket{\rho|\sigma} = \text{Tr}(\rho^\dag \sigma)\). In the case where \(\rho_{AB}\in \mathcal{H}_A\otimes \mathcal{H}_B\) is a bipartite density matrix, the canonical purification lives in the space \((\mathcal{H}_A\otimes \mathcal{H}_{A^*}) \otimes (\mathcal{H}_B \otimes \mathcal{H}_{B^*}) \equiv \mathcal{H}_{AA^*BB^*}\) and one defines the *reflected entropy* as the von Neumann entropy \[S_R(A:B) = S(AA^*)_{\ket{\sqrt{\rho_{AB}}}} = S_{\text{vN}}(\rho_{AA^*}),\] where \(\rho_{AA^*} = \tr_{BB^*}\ket{\sqrt{\rho_{AB}}}\bra{\sqrt{\rho_{AB}}}\) is the reduced density matrix obtained by tracing over \(\mathcal{H}_{BB^*}\). The reflected entanglement spectrum is simply the spectrum of \(\rho_{AA^\star}\) and we claim this is discrete even in the continuum limit. Taking the continuum limit proceeds as follows. As we send \(B \rightarrow A^c\), the complement region to \(A\), the reflected entropy reduces to twice the entanglement entropy which is now divergent. This divergent behavior can be understood as the non-existence of a tensor factorization \(\mathcal{H} \neq \mathcal{H}_A\otimes \mathcal{H}_{A^c}\) of the global Hilbert space without introducing a cutoff. This is an intrinsic property of type-III von Neumann algebras \(\mathcal{A}_A\) that govern the local bounded operators associated to region \(A\). Keeping a finite gap between \(A\) and \(B\) however allows reflected entropy to be used as a regulated version of entanglement entropy. In particular for two disjoint regions \(A\cup B\), the "split property" guarantees the existence of at least one type-I factor \(\mathcal{N}\) splitting of the local algebras \[\mathcal{A}_A \subset \mathcal{N} \subset \mathcal{A}_B'.\] The canonical purification introduced in [\[canonical\]](#canonical){reference-type="eqref" reference="canonical"}, in the algebraic language, corresponds to the state induced on a canonical type-I splitting factor, which can be written algebraically as \[\mathcal{N} = \mathcal{A}_A \vee J_{AB} \mathcal{A}_A J_{AB},\] where \(J_{AB}\) is the anti-unitary Tomita-Takesaki modular conjugation operator associated to \(\mathcal{A}_{AB}\) and the vacuum state. In this language the reflected entropy is defined as the von Neumann entropy of this type-I factor. In particular, the density matrix of \(\mathcal{N}\) is trace-class, with a well-defined and discrete spectrum, allowing one to make sense of the entropy and spectrum directly in the continuum. It was shown in that in the AdS/CFT setting, the reflected entropy is dual to the area of entanglement wedge cross-section. Since the entanglement wedge cross-section is typically a deep bulk probe of the emergent geometry, it is interesting to study the reflected entropy in a more general class of QFTs. Indeed, there has been many previous works computing the reflected entropy on different quantum systems, such as free fermions, free scalars, CFT in arbitrary dimensions, 3D Chern-Simons, holographic tensor networks and JT gravity with EOW branes. In this note we will focus on free fermions in \(1+1\) dimensions. Reflected entropy in free fermion systems was also already studied in. In this paper we will make a few new observations. In particular compared to we will work directly in the continuum theory bypassing the need of discretization. Some of our results will be numerical although we will also make some new analytic predictions in various limits. Consider two intervals \(A\) and \(B\) on some equal time slice in vacuum and define the cross-ratio of the end points of these intervals as \(x\). We will give analytic predictions for \(S_R\) in the limit where the \(A\) and \(B\) intervals are far separated \(x \rightarrow 0\) or nearly touching \(x \rightarrow 1\). In the former case the reflected entropy behaves as: \[S_R = \alpha (-x \ln x) + \beta + \ldots\] where we give a simple integral expression [\[eq:alpha_int\]](#eq:alpha_int){reference-type="eqref" reference="eq:alpha_int"} for \(\alpha\) and a numerical prediction [\[eq:beta_int\]](#eq:beta_int){reference-type="eqref" reference="eq:beta_int"} for \(\beta\). This is in agreement with our numerics as well as the numerics in and qualitative agreement with. In the later case we show that to the leading order \[S_R = \frac{1}{6}\left(-\ln (1-x) + \ln 4 \right) + \ldots\] in agreement with the universal behavior of 2D CFTs. In addition we give the next order correction to \(S_R\), agreeing well with our numerics. We summarize the analytical predictions and numerics in Figure [\[fig:Rferm\]](#fig:Rferm){reference-type="ref" reference="fig:Rferm"}. We also derive analytically the entanglement spectrum of \(\rho_{AA^\star}\) as \(x \rightarrow 1\). It takes the simple form of that of the spectrum of the thermal density matrix for a free chiral fermion on the circle in the NS sector with inverse temperature to circle length ratio \(\beta/L \propto -1/\ln(1-x)\). By studying perturbative corrections to this later spectrum we give evidence that the reflected density matrix approaches rapidly the thermal density matrix. This agrees well with the eigenvalues obtained from numerical method, see Figure [\[fig:eigenvalue\]](#fig:eigenvalue){reference-type="ref" reference="fig:eigenvalue"}. In this paper we do not make use of the replica trick, as in previous work, but we use the correlation matrix technique that is based on the fact that a many-body Gaussian state is entirely determined by it's two point correlation function. And so the many-body entropy can also be extracted from this correlation function. It is sufficient to know certain modular flowed two point functions in order to construct the relevant correlation matrix that computes the reflected entropy. These modular flow correlators were computed in the continuum in. Our analytic computations proceed as follows. We setup a systematic expansion for the eigenfunctions of the correlation matrix in the limit \(x \rightarrow 1\) via a certain matching procedure, between the endpoints of \(A\). This is similar to a QM scattering problem. In this limit one starts out with a continuum of Rindler eigenfunctions near each \(A\) endpoint and the discrete spectrum arises from a matching condition in-between. The discreteness of the spectrum is important for the finite-ness of the reflected entropy. The plan of this paper is as follows. We setup the singular integral equations in Section [2](#sec:setup){reference-type="ref" reference="sec:setup"}. In Section [3](#sec:zero){reference-type="ref" reference="sec:zero"} and Section [4](#sec:one){reference-type="ref" reference="sec:one"} we discuss the \(x \rightarrow 0\) limit and \(x \rightarrow 1\) limits respectively. In the later limit we make a more careful study of the spectrum, computing sub-leading corrections to the thermal spectrum in Section [4.3](#sec:corrections){reference-type="ref" reference="sec:corrections"}. We also introduce a \(s\)-modular flowed version of reflected entropy, that we call deflected entropy in Section [4.2](#sec:deflected){reference-type="ref" reference="sec:deflected"}-this is a natural one parameter generalization of reflected entropy. We compute this quantity for free fermions and also in AdS\(_3\)/CFT\(_2\) using the methods of-the dual roughly corresponds to a reflected geodesic that tracks the entanglement wedge cross section but picks up a boosts of rapidity \(s\) at the Ryu-Takayanagi surface. In Section [5](#sec:discussion){reference-type="ref" reference="sec:discussion"} we discuss our results, and make some conjectures about more general CFTs. Numerical results will be presented throughout the paper but are summarized in Figure [\[fig:Rferm\]](#fig:Rferm){reference-type="ref" reference="fig:Rferm"}.
# Setup {#sec:setup}
We would like to compute the reflected entropy for 2d free chiral fermions \(\psi\) and two intervals \(A,B\) on a circle. As discussed in, the reflected density matrix is still Gaussian (that is the Fermion fields satisfy Wick's theorem) so that the reduced density matrix is completely determined by the two point correlation function of fundamental fermions. We can then infer the entropy by solving the singular integral equation of this correlation kernel. The entire system consists of Fermions in \(AB\) and in the doubled system \((AB)^\star\). This later system is governed by Fermionic operators \(\widehat{\psi}(x) = i \tilde{J}_{AB} \psi(x)^\dagger \tilde{J}_{AB}^\dagger\) for \(x\in AB\). Since Fermions anti-commute, we need a generalization of the Tomita-Takesaki theory to a graded algebra. Denoting \(\Gamma = (-1)^F\) where \(F\) is the Fermion number operator, we can form the Klein operator \(Z = \frac{1-i\Gamma}{1-i}\), which is a unitary operator that acts as \(1\) on even states and \(i\) on odd states. The modular conjugation operator \(\tilde{J}_{AB}\) for the graded system is then related to the vacuum modular conjugation of Tomita-Takesaki theory (for the full fermionic algebra of operators in \(AB\)) \(J_{AB}\) by \(\tilde{J}_{AB} = Z J_{AB} = J_{AB} Z^\dagger\). The conjugated modes \(\widehat{\psi}(x)\) then satisfy the canonical anti-commutation relations amongst themselves as well as with the original \(AB\) system. We are interested in the entanglement entropy of \(AA^\star\) so we may restrict to correlation functions of \(\psi|_A\) and \(\widehat{\psi}|_A\). In order to understand the smooth \(B \rightarrow \emptyset\) limit we will choose to geometrize the \(A^\star\) reflected fermions by using instead the following description: \[\widetilde{\psi}(x) \equiv \tilde{J}_{AB} \tilde{J}_A \psi(x) \tilde{J}_A^\dagger \tilde{J}_{AB}^\dagger = J_{AB} J_A \psi(x) J_A J_{AB} \,, \qquad x \in A^c\] which is a fermion that now lives on \(A^c\) where \(A^c\) is the complementary region to \(A\). The \(\tilde{J}_A\) operator acts as CPT conjugation with a reflection across the \(A\) entangling surface, thus these still represent the same modes as \(\widehat{\psi}|_A\). The system \(AA^\star\) has now been mapped to the entire circle: \(S^1 = AA^c\). If we additionally define \(\widetilde{\psi}(x)|_{A} \equiv \psi(x)\) then \(\widetilde{\psi}\) is the new Fermion that lives on this circle: it satisfies the standard canonical anti-commutation relations. We thus, simply need to work out the state of the \(\widetilde{\psi}\) fermion on this circle. As discussed above this is determined by the correlation function of \(\widetilde{\psi}\). Note that we expect this to be an appropriately smooth state since near \(\partial A\) one can check that \(J_{AB}\) acts geometrically like the Rindler reflection or \(J_A\), so the two modular operators cancel out. Notice that the \(B \rightarrow \emptyset\) now reproduces the Fermion on a circle with the vacuum state. This is as expected and will give zero reflected entropy. The free fermion correlator on the cylinder in the NS vacuum \(\left| \Omega \right>\) is the following distribution: \[\left< \Omega \right| \psi^\dag(x)\psi(y) \left| \Omega \right> = \mathcal{P} \frac{z^{1/2}w^{1/2}}{\pi (z-w)} + \frac{1}{2}\delta(x-y) \, \qquad z= e^{ix}, \quad w = e^{iy}\] where \(0 < x,y \leq 2\pi\) on the cylinder. We have the canonical anti-commutation relations \(\{ \psi^\dag(x), \psi(y) \} = \delta(x-y)\) (with all other anti-commutators vanishing). Consider a basis of normalized functions on the circle that is anti-periodic: \(e^{ i (m+1/2)x } (2\pi)^{-1/2}\). We can consider the modes: \[\psi_m =\frac{1}{\sqrt{2\pi}} \oint \frac{dz}{ i z } z^{m+1/2} \psi(x)\, \qquad m \in \mathbb{Z}\] which gives \(\{ \psi_n^\dagger, \psi_m \} = \delta_{m,n}\) from the canonical anti-commutation relations. The Fourier modes of this correlator (\(m,n\in \mathbb{Z}\)) can then be calculated: \[\label{eq:theta_m} C_{mn} = \oint \frac{dz}{z i \sqrt{2\pi}} z^{m+1/2} \oint \frac{dw}{w i \sqrt{2\pi}} w^{-(n+1/2)}\left< \Omega \right| \psi^\dag(x)\psi(y) \left| \Omega \right>\\ = \Theta_m\delta_{mn}\] where \(\Theta_m = 1\) if \(m\ge0\) and \(0\) otherwise. Thus we find that the correlation function/matrix becomes a projector with eigenvalues \(0\) and \(1\). Let us now compute the correlation function when \(B\neq\emptyset\): \[\label{eq:correlator} {\rm Tr}_{AA^c} \left( \rho_{ AA^c} \widetilde{\psi}^\dagger (x) \widetilde{\psi}(y) \right) = \begin{cases} \left< \Omega \right| \psi^\dagger (x) \psi(y) \left| \Omega \right> & x,y \in A \quad {\rm or} \quad x, y \in A^c \\ \left( \frac{\partial y}{\partial y_J} \right)^{1/2} \left< \Omega \right| \psi^\dagger (x) \Delta_{AB}^{1/2} \psi(y_J) \left| \Omega \right> & x \in A\,,\quad y \in A^c \\ -\left( \frac{\partial x}{\partial x_J} \right)^{1/2} \left< \Omega \right| \psi (y) \Delta_{AB}^{1/2} \psi^\dagger(x_J) \left|\Omega \right> & y \in A\,,\quad x \in A^c \\ \end{cases}\] where \(x_J\) is given by the geometric action of \(\tilde{J}_A\) on \(x\): if \(x \in A\) then \(x_J \in A^c\). We have used Tomita-Takesaki theory to replace the modular conjugation operator \(J_{AB}\) with the modular operator \(\Delta_{AB}\) for the vacuum state. The phase of the square root can be fixed by demanding we reproduce the vacuum correlator as \(B \rightarrow \emptyset\). This correlation function along with the Gaussian nature of \(\rho_{AA^c}\) will allow us to compute its spectrum. The spectrum of \(\rho_{AA^c}\) is the same as the spectrum of the reflected density matrix \(\rho_{AA^\star}\) or equivilently the density matrix on the canonical type-I factor \(\mathcal{N}\). The correlation functions can be computed from the formula for modular flow for fermions on the plane: \[\begin{aligned}
\left< \psi (z) \Delta_{AB}^{is} \psi^\dagger(w) \right> &= \left< \psi^\dagger (z) \Delta_{AB}^{is} \psi(w) \right> = \frac{e^{-\pi s}}{2\pi i (z-w)} \left( 1- \frac{1-e^{-2\pi s}}{ Q-e^{-2\pi s}} \right)\, \\ \qquad Q &= \frac{(z-a_1)(z-a_2)(w-b_1)(w-b_2)}{ (z-b_1)(z-b_2)(w-a_1)(w-a_2)} \end{aligned}\] where \(A = [a_1, b_1]\) and \(B = [a_2,b_2]\) and we may consider the boundaries of these intervals to be complex coordinates in the plane. For Fermions on the circle we are free to fix \(a_1 = 1, b_1 =-1\) and \(a_2 =-i e^{-i \phi}, b_2 =-i e^{i \phi}\) and set \(z = e^{i x}\) and \(w = e^{i y}\). Here we may consider \(x\) as a holomorphic coordinate on the cylinder with \(0 \leq {\rm Re} \, x \leq 2\pi\) (not to be confused with the cross-ratio that we will introduce later) and in these coordinates \(A = [0,\pi]\) and \(B = [ 3\pi/2-\phi, 3\pi/2+\phi]\) (see figure [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}). We use: \[\left< \psi^\dagger (x) \Delta_{AB}^{is} \psi(y) \right>_\Omega \equiv \left< \psi^\dagger (x) \Delta_{AB}^{is} \ \psi(y) \right>_{\rm cyl} = \left( \frac{ \partial z}{\partial x} \right)^{1/2} \left( \frac{ \partial w}{\partial y} \right)^{1/2} \left< \psi^\dagger (x) \Delta_{AB}^{is} \ \psi(y) \right>\] and \(x_J =-x\). If we subtract the vacuum correlator we find: \[\begin{aligned}
\left< \widetilde{\psi}^\dagger (x) \widetilde{\psi}(y) \right>_{\rho_{AA^c}} &-\left< \psi^\dagger (x) \psi(y) \right>_\Omega \\ &= \frac{1}{4 \pi i \sin(\frac{x-y}{2})} \frac{ \theta_A(x) \theta_{A^c}(y) \sin(x) \sin(y) \sin(\phi) }{ ( \sin^2(\frac{x-y}{2})-\sin\phi \sin^2(\frac{x+y}{2}) + \cos\phi \sin(\frac{x-y}{2}) \cos( \frac{x + y}{2} ) )} \nonumber \\ & \qquad \qquad \qquad \qquad \quad-(x \leftrightarrow y) \nonumber \end{aligned}\] where \(\theta_A(x) ( \theta_{A^c}(x))\) is a unit step for \(x\in A (A^c)\). Define the Fourier transform of the correlation kernel \[\Delta C_{mn} = \int_0^{2\pi} \frac{d x}{\sqrt{2\pi}} e^{i x(m+1/2) } \int_0^{2\pi} \frac{d y}{\sqrt{2\pi}} e^{-i y(n+1/2) }\left( \left< \widetilde{\psi}^\dagger (x) \widetilde{\psi}(y) \right>_{\rho_{AA^\star}}-\left< \psi^\dagger (x) \psi(y) \right>_\Omega \right)\] So that we have \[\label{eq:C_mn} \Delta C_{mn} = \int_0^{\pi} dx \int_\pi^{2\pi} dy \frac{ 1}{4 \pi^2 \sin(\frac{x-y}{2})} \frac{ \sin( x(m+1/2)-y(n+1/2)) \sin(x) \sin(y) \sin(\phi) }{ ( \sin^2(\frac{x-y}{2})-\sin\phi \sin^2(\frac{x+y}{2}) + \cos\phi \sin(\frac{x-y}{2}) \cos( \frac{x + y}{2} ) )}\] from which it is clear that \(\Delta C_{mn} = \Delta C_{nm}\) (from sending \(x \leftrightarrow 2\pi-y\)). Furthermore \(x \rightarrow \pi-x\) and \(y \rightarrow 3\pi-y\) gives: \[\Delta C_{mn} = (-1)^{m-n} \Delta C_{mn}\] which implies that \(\Delta C_{mn}\) vanishes unless \(m-n\) is even. Also it is clear that \(\Delta C_{m,n} =-\Delta C_{-1-m,-1-n}\). The actual correlation matrix is: \[C_{mn} = \Delta C_{mn} + \delta_{mn} \Theta_m\] with \(\Theta_m\) defined in [\[eq:theta_m\]](#eq:theta_m){reference-type="eqref" reference="eq:theta_m"}. The second term comes from the vacuum correlator (this is the projector which gave zero entanglement when \(\Delta C_{mn} = 0\).) The reflected entropy can be computed via \[\begin{aligned}
S_R =-\tr (C_{mn}\ln C_{mn})-\tr((1-C_{mn})\ln(1-C_{mn})) \end{aligned}\] After truncation of the modes we can numerically compute \(\Delta C_{mn}\) and use this to extract the reflected entropy (Figure [\[fig:Rferm\]](#fig:Rferm){reference-type="ref" reference="fig:Rferm"},[\[fig:Rferm_small_x\]](#fig:Rferm_small_x){reference-type="ref" reference="fig:Rferm_small_x"},[\[fig:Rferm_large_x\]](#fig:Rferm_large_x){reference-type="ref" reference="fig:Rferm_large_x"},[\[fig:s_plots\]](#fig:s_plots){reference-type="ref" reference="fig:s_plots"}) and entanglement spectrum (Figure [\[fig:eigenvalue\]](#fig:eigenvalue){reference-type="ref" reference="fig:eigenvalue"},[\[fig:lambdashift\]](#fig:lambdashift){reference-type="ref" reference="fig:lambdashift"}). The relevant cross ratio is: \[x = \frac{ 2 \sin\phi}{1 + \sin\phi}\] We truncate the matrix at some finite value \(|m|,|n|\le L\) so that the resulted entropy converge to prescribed tolerance. [^1] Starting from the next section we will study the reflected entropy under various limits. Clearly the limit \(\phi \rightarrow 0\) (\(x\to 0\)) is most naturally studied on the cylinder; whereas the limit \(\phi \rightarrow \pi/2\) (\(x\to 1\)) is most appropriately studied using Rindler space, since this is the limit where we recover the usual (divergent) entanglement entropy.
# The \(x\to 0\) limit {#sec:zero}
Let us first consider the case when \(\phi\to0\). We will apply degenerate perturbation theory to the correlation matrix, which allows us to extract the \(\phi\) and \(\phi\ln\phi\) terms of the reflected entropy. Taylor expand [\[eq:C_mn\]](#eq:C_mn){reference-type="eqref" reference="eq:C_mn"} to the first order in \(\phi\) we have: \[\begin{aligned}
\Delta C_{mn} & \approx \phi {\rm Re} \int_0^\pi \frac{d x}{\pi} \frac{e^{i x m } \sin(x)}{1+i e^{-i x}} \int_\pi^{2\pi} \frac{d y}{\pi} \frac{e^{-i y n } \sin(y)}{1+i e^{i y}} \frac{ 1}{2\sin^2(\frac{x-y}{2})} \\ & = (-1)^n \phi {\rm Re} \int_0^\pi \frac{dx}{\pi} \int_0^\pi \frac{dy}{\pi} \frac{f_m(x) f_n(y)^\star}{2\cos^2(\frac{x-y}{2})} \,, \quad f_m(x) = \frac{e^{i x m} \sin(x)}{ 1 + i e^{-ix}} \end{aligned}\] Since the eigenvalue is degenerate at zeroth order, we need to diagonalize the two matrices on these degenerate subspaces: \[A = \lim_{\phi \rightarrow 0} \frac{1}{\phi} \Theta \Delta C \Theta \, \qquad B = \lim_{\phi \rightarrow 0} \frac{1}{\phi} (1-\Theta) \Delta C (1-\Theta)\] where we defined the projector \(\Theta_{mn} = \delta_{mn} \Theta_m\). Define the eigenvalues of these matrices as \(-a_k, b_k\) respectively such that \(a_k, b_k \geq 0\). Then the eigenvalues of \(C_{mn}\) are approximately \((1-\phi a_k)\) and \(\phi b_k\) to first order in \(\phi\). The leading correction to the reflected entropy is: \[\begin{aligned}
\label{eq:asym_smallx} S_R &\approx \sum_k \left( \phi a_k (1-\ln\left( \phi a_k \right)) + \phi b_k (1-\ln\left( \phi b_k \right)) \right) \\ &=-\phi \ln \phi \sum_k ( a_k + b_k) + \phi \sum_k ( a_k + b_k-a_k \ln a_k-b_k\ln b_k ) \end{aligned}\] For the \(\phi \ln \phi\) term the sum of eigenvalues is just the trace of the respective matrices which we can compute explicitly: \[\label{eq:alpha_int} \sum_k ( a_k + b_k) = \frac{1}{4\pi^2} \int_0^\pi dx \int_0^\pi dy \frac{ \sin(x) \sin(y)}{ \cos^3(\frac{x-y}{2}) ( \cos(\frac{x-y}{2}) + \sin(\frac{x+y}{2}))} \approx.149\] It seems hard to find the coefficient of the term linear to \(\phi\) using analytical method. Instead we have from the direct eigenvalue decomposition of matrices \(A\) and \(B\) that \[\label{eq:beta_int} \sum_k ( a_k + b_k-a_k \ln a_k-b_k\ln b_k ) \approx.560\] These coefficients agree with our numerical results, see figure [\[fig:Rferm_small_x\]](#fig:Rferm_small_x){reference-type="ref" reference="fig:Rferm_small_x"}.
# The \(x\to 1\) limit {#sec:one}
We move on to studying the reflected entropy in the limit \(x\to 1\) (or \(\phi\to\pi/2\)) in this section. We will setup a systematic expansion for the eigenfunctions of the correlation matrix via a certain matching procedure between the endpoints of \(A\). Similar to QM scattering, solving for these matching conditions discretizes the eigenvalue spectrum and we obtain a finite reflected entropy.
## Reflected entropy {#sec:rindler}
We need to reformulate our calculations adapted to \(A\) being a half space cut. We pick \(a_1 = 0, b_1 = \infty\) and \(a_2 =-1/b, b_2 =-b\). The cross ratio of these points is: \[x = \frac{ (a_1-b_1)(a_2-b_2)}{(a_1-a_2)(b_1-b_2)} =1-b^2\] where we restrict to \(0 < b < 1\). We also have: \[\label{full} Q = \frac{z(z+1/b)(w_J+b)}{ w_J (z+b)(w_J+1/b)}\] where we now used directly \((z,w)\) (rather than \((x,y)\)) as coordinates for the reflected fermion. Also \(w_J =-w\). Taking the limit \(b \rightarrow 0\) we find the correlator for the reflected fermion approaches (note that \(Q \rightarrow 1\)): \[\lim_{b \rightarrow 0} \left< \widetilde{\psi}^\dagger (z) \widetilde{\psi}(w) \right>_{\rho_{AA^c}} = \frac{\Theta_A(z) \Theta_A(w)+\Theta_{A^c}(z) \Theta_{A^c}(w) }{2\pi i (z-w-i \epsilon)}\] The eigenfunctions of this operator are Rindler modes with a continuous spectrum. This continuum leads to a divergent entanglement (since the operator is no longer trace class). These eigenfunctions are: \[f_\nu^A(z) =\frac{1}{\sqrt{2\pi}} \theta(z) z^{-1/2-i \nu} \qquad f_\nu^{A^c}(z) = \frac{1}{\sqrt{2\pi}} \theta(-z)(-z)^{-1/2-i \nu}\] and these satisfy: \[\int_0^\infty dz \frac{1}{2\pi i (z-w-i \epsilon)} f_\nu^A(z) = \frac{1}{2}( 1-\tanh(\pi \nu)) f_\nu^A(w) \,, \qquad w \in A\] and \[\int_{-\infty}^0 dz \frac{1}{2\pi i (z-w-i \epsilon)} f_\nu^{A^c}(z) = \frac{1}{2}( 1+\tanh(\pi \nu)) f_\nu^{A^c}(w) \,, \qquad w \in A^c\] To resolve this continuum we need to take the limit \(b \rightarrow 0\) more carefully. We hold \(|z|/b, |w|/b\) fixed as we send \(b \rightarrow 0\) (with fixed ratio \(|z/w|\)). This zooms in on the region near the entangling surface. Inside this region the correlator is: \[\delta \left< \widetilde{\psi}^\dagger (z) \widetilde{\psi}(w) \right>_{\rho_{AA^c}} \approx \frac{1}{2\pi i (z-w-2 wz/b)} \Theta_{A}(z) \Theta_{A^c}(w)-(z \leftrightarrow w)\] Overall we can approximate the correlator as: \[C(z,w) \equiv \left< \widetilde{\psi}^\dagger (z) \widetilde{\psi}(w) \right>_{\rho_{AA^c}} \approx \frac{1}{2\pi i (z-w-2 wz/b (\theta(-w)\theta(z)-\theta( w)\theta(-z) )}\] We can write this correlator as: \[C(z,w) =\left( \frac{ \partial \widetilde{z}}{\partial z} \right)^{1/2} \left( \frac{ \partial \widetilde{w}}{\partial w} \right)^{1/2} \frac{1}{ 2\pi i(\widetilde{z}-\widetilde{w}-i \epsilon) }\] where \[\widetilde{z}(z) = z \theta(z) + \frac{z}{1-2z/b} \theta(-z)\,, \qquad \widetilde{w}(w) = w \theta(w) + \frac{w}{1-2w/b} \theta(-w)\] and the new coordinate satisfies \(-b/2 \leq \widetilde{z} \leq \infty\). So the negative axis gets compactified. On this half space we know the Rindler eigenfunctions: \[\widetilde{f}_\nu (\widetilde{z}) = \frac{1}{\sqrt{2\pi}} \theta(\widetilde{z} + b) (\widetilde{z}+b)^{-1/2-i \nu}\] and these are the eigenfunctions of the integral Kernel \(\propto (\widetilde{z}-\widetilde{w}-i \epsilon)^{-1}\). Taking into account the conformal scale factors the would be eigenfunctions of \(C(z,w)\) are: \[\label{exp1} f_\nu(z) \approx \widetilde{f}_\nu (\widetilde{z}) \left( \frac{ \partial \widetilde{z}}{\partial z} \right)^{1/2} = \frac{1}{\sqrt{2\pi}}\left( (z+b/2)^{-1/2-i \nu} \theta(z) + (b/2)^{-2 i \nu} (b/2-z)^{-1/2 + i \nu} \theta(-z) \right)\] with eigenvalue \(( 1-\tanh(\pi \nu))/2\). This eigenfunction behaves badly at \(z \rightarrow \infty\). This is to be expected since at this point (when \(z \sim 1/b\)) our scaling limit breaks down. Instead we must match onto a new solution here. If we again go back to the full correlator (determined by [\[full\]](#full){reference-type="eqref" reference="full"}) and hold fixed \(|z| b\) and \(|w| b\) as we send \(b \rightarrow 0\) then we find: \[\delta \left< \widetilde{\psi}^\dagger (z) \widetilde{\psi}(w) \right>_{\rho_{AA^c}} \approx \frac{1}{2\pi i (z-w + 2/b)} \Theta_{A}(z) \Theta_{A^c}(w)-(z \leftrightarrow w)\] Overall we can approximate the correlator here as: \[D(z,w) \equiv \left< \widetilde{\psi}^\dagger (z) \widetilde{\psi}(w) \right>_{\rho_{AA^c}} \approx \frac{1}{2\pi i (z-w + 2/b (\theta(-w)\theta(z)-\theta( w)\theta(-z) )}\] which we can write as: \[D(z,w) = \frac{1}{ 2\pi i(\widehat{z}-\widehat{w}-i \epsilon) }\] where \[\widehat{z}(z) = z \theta(z) + (z-2/b) \theta(-z)\,, \qquad \widehat{w}(w) = w \theta(w) + (w-2/b) \theta(-w)\] the new coordinate satisfies \(0 \leq \widehat{z} \leq \infty\) and \(-\infty \leq \widehat{z} \leq-2/b\) (which is a domain that wraps around \(\infty\).) This can be described via \(1/\widehat{z} \geq-b/2\). On this domain we use the Rindler eigenfunctions: \[\widehat{f}_\nu (\widehat{z}) = \frac{1}{\sqrt{2\pi}} \theta(1/\widehat{z} + b/2 ) (\widehat{z})^{-1} (1/\widehat{z} + b/2)^{-1/2+i \nu}\] which has the same eigenvalue, for the kernel \(\propto (\widehat{z}-\widehat{w} + i \epsilon)^{-1}\) as we used in the previous patch. The eigenfunction for the integral equation of interested, in this new scaling limit must then be: \[f_\nu(z) \approx \frac{1}{\sqrt{2\pi}} \left( \theta(z) z^{-1} (1/z+b/2)^{-1/2+ i \nu} + \theta(-z) z^{-1} (b/2)^{ 2 i \nu} (-1/z+b/2)^{-1/2-i \nu} \right)\] The small \(z\) limit of this expression should match onto the large \(z\) limit of the previous expression [\[exp1\]](#exp1){reference-type="eqref" reference="exp1"} up to an overall scaling: \[\begin{aligned}
\begin{split} \kappa &\left( \theta(z) z^{-1} (1/z)^{-1/2+ i \nu} + \theta(-z) z^{-1} (b/2)^{ 2 i \nu} (-1/z)^{-1/2-i \nu} \right) \\ & \approx \left( (z)^{-1/2-i \nu} \theta(z) + (b/2)^{-2 i \nu} (-z)^{-1/2 + i \nu} \theta(-z) \right) \end{split} \end{aligned}\] which gives \(\kappa = 1\) and \((b/2)^{4 i\nu} =-1\). This is our quantization condition. More explicitly the two limits of the eigenfunction given above match smoothly if this condition is satisfied. We have the following allowed values of \(\nu\): \[\nu_m = \frac{\pi (m-\frac{1}{2})}{2 (-\ln(b/2))}\] where \(m \in \mathbb{Z}\). These give rise to eigenvalues of \(C\) as: \[\label{eq:lambdaleading} C_m = \frac{1-\tanh(\pi \nu_m)}{2} = ( 1 + e^{2\pi \nu_m})^{-1}\] The reflected entropy is: \[\label{sum} S_R = \sum_{m=-\infty}^{\infty}(-C_m \ln C_m-(1-C_m) \ln (1-C_m))\] Since the spacing is small as \(b \rightarrow 0\), we can approximate this by an integral with density: \[2 \int_0^1 d C \left| \frac{d\nu}{dC} \frac{dm}{d\nu_m} \right| (-C \ln C) = \frac{2 (-\ln(b/2))}{\pi^2} \int_0^1 d C \frac{1}{ C(1-C)} (-C \ln C) = \frac{(-\ln(b/2))}{3}\] Writing this in terms of the cross-ratio we have: \[\label{eq:SR_leading} S_R \rightarrow \frac{1}{6} (-\ln (1-x) + \ln 4 ) + \ldots\] which was as predicted by the replica trick. In principle we can compute corrections to this as an expansion in \((1-x)\) and \(1/\ln(1-x)\). In fact we can give a more general expression that is leading in \((1-x)\) but one that re-sums all \(1/\ln(1-x)\) effects. Firstly we notice that [\[sum\]](#sum){reference-type="eqref" reference="sum"} is the entropy deriving from a partition function: \[\label{eq:thermal_Z} S_R =-\beta^2 \partial_\beta \left( \beta^{-1} \ln Z(\beta) \right) \qquad\, Z(\beta) = \prod_{m=0}^\infty (1 + q^{m+1/2})^2\] where \(q = e^{i 2\pi \tau}\) and the inverse temperature is: \[\label{temp} \tau = i \frac{\pi}{2(-\ln( (1-x)/4))}\] We learn that the spectrum of the reflected density matrix approaches that of a free fermion in the NS sector. So the reflected entropy is simply given by the appropriate Jacobi theta function. And indeed, as is often the case, the universal term in [\[eq:SR_leading\]](#eq:SR_leading){reference-type="eqref" reference="eq:SR_leading"} arises simply via the Cardy formula. We would speculate that this is a more general result: the reflected entanglement spectrum approaches that of the thermal partition function of the CFT under consideration, with temperature given by [\[temp\]](#temp){reference-type="eqref" reference="temp"}. There seems to be some connection between the relfected entropy and the computable cross norm negativity (at least for some Renyi generalization of reflected entropy) which could possibly be used to give a more general proof of this fact.
## Deflected entropy {#sec:deflected}
We now study a generalization of the reflected entropy that we call deflected entropy. It is given by applying a modular flow on the correlation function [\[eq:correlator\]](#eq:correlator){reference-type="eqref" reference="eq:correlator"} by an additional amount of \(s \in \mathbb{R}\): \[\begin{aligned}
\Braket{\tilde{\psi}^\dag(x)\tilde{\psi}(y)}_{\rho_{AA^c}} = \begin{cases} \Braket{\psi^\dag(x) \psi(y)}_\Omega, \quad &x,y \in A \,\, {\rm or } \,\, x,y \in A^c \\ \left( \frac{\partial y_J}{\partial y} \right)^{1/2}\Braket{\psi^\dag(x) \Delta^{1/2+is}_{AB} \psi(y_J)}_\Omega, &x\in A, y\in A^c \\ \left( \frac{\partial x_J}{\partial x}\right)^{1/2} \Braket{\psi^\dag(x_J) \Delta^{1/2-is}_{AB} \psi(y)}_\Omega, &y\in A, x\in A^c \end{cases} \end{aligned}\] When \(s\neq0\) this correlator is not continuous across the entangling surface. To fix this discontinuity we replace the geometric reflection map by another vacuum modular flow (with respect to region \(A\)) with an amount \(-s+i/2\). In our current coordinate settings this action is simply \[\begin{aligned}
x\to x_S =-e^{2\pi s}x \end{aligned}\] This extra vacuum modular flow for the A region leaves the entropy invariant since it is generated by an operator that acts solely in \(AA^\star\) (working in the original frame of \(AB A^\star B^\star\))-the easiest way to see this is by using the modular operator for the split state, which agrees with the vacuum flow inside A. When \(s=0\) we simply get back the geometric reflection; whereas when \(s\neq 0\) this additional flow smooths out the jump discontinuity across the entanglement surface. Algebraically one can think of this construction as performing a Connes cocycle flow defined with the \(AB\) algebra and for the vacuum and split state. See for example. The analysis of the previous subsection still carries over as long as we assume \(b < e^{-\pi s}\) when we scale \(b\to 0\). We can approximate the new correlation function as \[\begin{aligned}
C(x,y) \simeq \begin{cases} \dfrac{1}{2\pi i} \dfrac{1}{x-y-\frac{2xy}{b}e^{\pi s}\cosh(\pi s)(\theta(x)\theta(-y)-\theta(-x)\theta(y))},\quad |x|,|y|\ll e^{-\pi s}\\ \dfrac{1}{2\pi i} \dfrac{1}{x-y+\frac{2}{b}e^{-\pi s}\cosh(\pi s)(\theta(x)\theta(-y)-\theta(-x)\theta(y))},\quad |x|,|y|\gg e^{-\pi s}\\ \end{cases} \end{aligned}\] The effect of the Connes cocyle flow is to shift \(b\) in both regimes accordingly and the matching point from \(1\) to \(e^{-\pi s}\). One can show the spectrum is now determined by: \[\nu_m = \frac{\pi(m-1/2)}{2\ln (2\cosh (\pi s)/b)}\] such that the reflected entropy becomes: \[\label{tocompff} S_R \rightarrow \frac{1}{6}\left(-\ln (1-x) + \ln 4 + 2\ln (\cosh(\pi s )) \right) + \ldots\] This result also agrees well with our numerics, see Figure [\[fig:s_plots\]](#fig:s_plots){reference-type="ref" reference="fig:s_plots"}. In the Appendix [\[app:A\]](#app:A){reference-type="ref" reference="app:A"} we perform a different computation of the deflected entropy in AdS\(_3\)/CFT\(_2\). We use this to compare to the free fermion computations. We find: \[\begin{aligned}
S_{R,\rm holographic} = \frac{c}{6} \ln \left| \frac{1+\sqrt{w}}{1-\sqrt{w}} \right|+ O(c^0), \quad w = \frac{x(1+e^{2\pi s})^2}{(x+e^{2\pi s})(1+xe^{2\pi s})} \end{aligned}\] From which we get the following asymptotic when \(x\to 1\): \[\begin{aligned}
S_{R,\rm holographic} = \frac{c}{6} (-\ln(1-x)+\ln 4 + \ln(\cosh(\pi s))) + \cdots \end{aligned}\] The overall factor of \(c\) (where in holographic theories \(c\) is large) accounts for the extra degrees of freedom. Comparing to [\[tocompff\]](#tocompff){reference-type="eqref" reference="tocompff"} we see similar behavior aside from the factor of \(2\). There is of course no reason to expect agreement. This difference persists away from \(x \rightarrow 1\) as can be seen in Figure [\[fig:s_plots\]](#fig:s_plots){reference-type="ref" reference="fig:s_plots"}.
## Next order corrections {#sec:corrections}
In this subsection we systematically improve on our results for the spectrum as \(x \rightarrow 1\). This also serves to convince the reader that our results are under control. We firstly setup the leading order answer more carefully. We will construct approximate eigenfunctions of the integral equation, valid as \(b \rightarrow 0\) (that is \(x \rightarrow 1\)). This also fills in some holes in the original discussion in Section [4.1](#sec:rindler){reference-type="ref" reference="sec:rindler"}. This subsection is a bit technical so the reader might prefer to skip to the conclusions. Let us see that the functions we wrote down are really approximate eigenfunctions as \(b \rightarrow 0\). *Define* the function: \[\begin{aligned}
g_{\nu}(z) &\equiv \nonumber \left( (z+b/2)^{-1/2-i \nu} \theta(z) + (b/2)^{-2 i \nu} (b/2-z)^{-1/2 + i \nu} \theta(-z) \right) \theta(1-|z |) \\& + \left( \theta(z) z^{-1} (1/z+b/2)^{-1/2+ i \nu} + \theta(-z) z^{-1} (b/2)^{ 2 i \nu} (-1/z+b/2)^{-1/2-i \nu} \right) \theta(|z |-1) \end{aligned}\] where we have designated an arbitrary matching point of \(|z|=1\). This function has a non-uniform expansion as \(b \rightarrow 0\). For \(|z| \sim b\) then \(|g_\nu(z)| \sim b^{-1/2}\), for \(|z| \sim 1/b\) then \(|g_\nu(z)| \sim b^{1/2}\) and finally for \(|z| \sim 1\) then \(g_\nu(z) \sim 1\). Consider, for \(w>0\) the following integral: \[\begin{aligned}
\int_{-\infty}^{\infty} d z C(z,w) g_{\nu}(z) \end{aligned}\] We need to check this integral in the various regimes. We first consider \(w \sim b\). One can see that the integral is dominated by the regime \(-1/w < z< 1/w\) giving an answer \(\mathcal{O}(b^{-1/2})\): \[\begin{aligned}
\int_{-\infty}^{\infty} d z C(z,w) g_{\nu}(z) \approx \lambda_\nu g_{\nu}(w) \end{aligned}\] Similarly for \(w \sim 1/b\) the integral is dominated in the region \(z>1/w\) and \(z<-1/w\) giving an answer \(\mathcal{O}(b^{1/2})\) with the same form as above. Finally for \(w \sim 1\) we only get a contribution from away from the scaling regimes, and \(z>0\) (since the kernel is otherwise suppressed by \(\mathcal{O}(b)\) for \(z<0\) when \(w>0\)) where: \[\int_{-\infty}^{\infty} d z C(z,w) g_{\nu}(z) \approx \int_0^\infty dz \frac{1}{2\pi i(z-w-i\epsilon)} z^{-1/2-i\nu} = \lambda_\nu w^{-1/2-i\nu} \approx \lambda_\nu g_\nu(w)\] Thus, at least for \(w>0\) it looks like \(g_\nu\) is an approximate eigenfunction with eigenvalue \(\lambda_\nu\). A similar analysis applies to the case \(w<0\), where the \(\mathcal{O}(1)\) regime is now: \[\begin{aligned}
\nonumber &\int_{-\infty}^{\infty} d z C(z,w) g_{\nu}(z) \approx \int_{-\infty}^0 dz \frac{ (-z)^{-1/2 + i \nu} }{2\pi i(z-w-i\epsilon)} \left( (b/2)^{-2 i \nu}\theta(1+z) -(b/2)^{ 2 i \nu} \theta(-(z+1)) \right) \\ & =-\lambda_\nu (b/2)^{2 i \nu} (-w)^{-1/2 + i \nu} + ((b/2)^{-2 i \nu} + (b/2)^{2 i \nu} )\int_{-1}^0 dz \frac{ (-z)^{-1/2 + i \nu} }{2\pi i(z-w-i\epsilon)} \\ & \approx \lambda_\nu g_\nu(w) + [ (b/2)^{-2 i \nu} + (b/2)^{2 i \nu}] \left(-\lambda_\nu \theta(1 + w) (-w)^{-1/2 + i \nu}+ \int_{-1}^0 dz \frac{ (-z)^{-1/2 + i \nu} }{2\pi i(z-w-i\epsilon)} \right) \end{aligned}\] So all together: \[\int_{-\infty}^\infty dz (C(z,w)-\lambda_\nu \delta(z-w)) g_\nu(z) \approx [ (b/2)^{-2 i \nu} + (b/2)^{2 i \nu}] G_\nu(w)\] where \(G_\nu(w)\) is \(\theta(-w)\) times the function above and satisfies \(G_\nu(w) \sim \mathcal{O}(1)\) if \(|w| \sim 1\) and \(\sim \mathcal{O}(1)\) for \(|w| \sim b\) and \(\sim \mathcal{O}(b^{1})\) if \(|w| \sim 1/b\). Thus we find that \(g_\nu\) is only an approximate eigenfunction upon imposing the quantization condition discussed in the previous subsection. We can now prove approximate orthogonality for distinct eigenvalues: \[0 \approx \int_{-\infty}^{\infty} dw \int_{-\infty}^\infty dz g_{\nu_m}^\star(w) (C(z,w)-\lambda_{\nu_n} \delta(z-w)) g_{\nu_n}(z) \approx ( \lambda_{\nu_m}-\lambda_{\nu_n}) \int_{-\infty}^{\infty} dw g_{\nu_m}^\star(w) g_{\nu_n}(w)\] and we can compute the normalization by moving slightly away from the quantization condition for one of the \(\nu\)'s above: \[( \lambda_{\nu_m}-\lambda_{\nu}) \int_{-\infty}^{\infty} dw g_{\nu_m}^\star(w) g_{\nu}(w) \approx [ (b/2)^{-2 i \nu} + (b/2)^{2 i \nu}] \int_{-\infty}^0 dw g_{\nu_m}^\star(w) G_\nu(w)\] One can see that the later integral \(I = \int_{-\infty}^0 dw g_{\nu_m}^\star(w) G_\nu(w)\) is dominated in the order one regime where it evaluates to: \[\begin{aligned}
I = e^{i\pi(m-1/2)} \int_{-\infty}^0 dw \left(-\lambda_{\nu} \theta(1 + w) (-w)^{-1/2 + i \nu}+ \int_{-1}^0 dz \frac{ (-z)^{-1/2 + i \nu} }{2\pi i(z-w-i\epsilon)} \right) (-w)^{-1/2-i \nu_m} \end{aligned}\] This integral is most efficiently evaluated by taking \(\nu\) complex and so that the \(w\) integral can be done first (the integral is well defined as stated above) \[\begin{aligned}
I = e^{i\pi(m-1/2)}\left( \frac{\lambda_{\nu}-\lambda_{\nu_m}} {i( \nu-\nu_m)} \right) \end{aligned}\] Thus: \[\int_{-\infty}^{\infty} dw g_{\nu_m}^\star(w) g_{\nu_n}(w) \approx \delta_{m,n} 4 (-\ln(b/2) )\] This gives (at least formally) the completeness relation: \[\label{complete} \sum_{m=-\infty}^{\infty} g_{\nu_m}^\star(w) g_{\nu_m}(w') \approx \frac{1}{4 (-\ln(b/2) )} \delta(w-w')\] We can now attempt to find corrections to these eigenfunctions and eigenvalues using perturbation theory. We write: \[h_m(z) = g_m(z) + \sum_{n=-\infty}^{\infty} \delta \mu_{mn} g_n(z) + \ldots\] where \(\delta \mu_{mn} = \mathcal{O}(b)\). Plugging this into the eigenfunction equation: \[\int_{-\infty}^{\infty} dz C(z,w) h_m(z) = (\lambda_m + \delta \lambda_m) h_m(w)\] To first order in \(b\) we have: \[\begin{aligned}
\label{tolead} \begin{split} 4 (-\ln(b/2)) \delta \lambda_m &= \int_{-\infty}^{\infty} dw\left( -\lambda_m \delta(g_m(w))+ \delta \int_{-\infty}^{\infty} dz C(z,w) g_m(z) \right)g_m^\star(w) \\ &= \int_{-\infty}^{\infty} dw \int_{-\infty}^{\infty} dz\delta( C(z,w) ) g_m(z) g_m^\star(w) \end{split} \end{aligned}\] which does not depend on \(d_{mn}\) or \(\delta \nu_m\). One sees the usual simplicity of first order perturbation theory. We now seek the leading term in [\[tolead\]](#tolead){reference-type="eqref" reference="tolead"}. We expect an expansion of the form: \[\delta\lambda_m = \sum_{k=1}^\infty b^{k} \ell_k( \ln b)\] where \(\ell_k\) are functions of \(\ln b\). We will aim for the order \(k=1\) term. Since the integrand in [\[tolead\]](#tolead){reference-type="eqref" reference="tolead"} has a non-uniform expansion in \(b\) we must expand it in various regions \(z,w \sim b, 1, b^{-1}\). We expand only in \(C(z,w)\), keeping the leading terms in the wave-functions and the measure \(dz dw\) (which gets \(b\) dependence after scaling into these regions.) This procedure goes under the name of matched asymptotic expansion. One often finds power law divergences after isolating the various regions using scaling arguments. The general rule is that, since the total integral is well defined, then any power law divergences will cancel amongst the various regions, that is after a re-arrangement of the order in the \(b\) expansion (these divergences typically lead to enhancements in the \(b\) expansion that mix between the orders.) In some instances we can proceed by using dimensional regularization. We give \(\nu\) a small imaginary part and this removes any power law diverges (log's would show up as poles, but we do not find any.) The diagonal regions \((z\sim 1,w \sim 1)\), \((z\sim b,w \sim b)\) and \((z\sim 1/b,w \sim 1/b)\) all give rise to \(\mathcal{O}(b^2)\) corrections. Since the pole/singularity in the kernel lives in the diagonal regions, we have the symmetry \(\delta (C(z,w)) =-\delta ( C(w,z)) = \delta ( C(w,z))^\star\) which allows us to restrict the integral to \(z > w\): \[\begin{aligned}
4 (-\ln(b/2)) \delta \lambda_m &= \int_{z>w} dw dz\delta( C(z,w) ) g_m(z) g_m^\star(w) + \int_{z<w} dw dz\delta( C(z,w) ) g_m(z) g_m^\star(w) \nonumber\\ & = 2 {\rm Re} \int_{z>w} dw dz\delta( C(z,w) ) g_m(z) g_m^\star(w) \end{aligned}\] after re-labeling \(w \leftrightarrow z\) in the second term. The non-diagonal regions with one \(z\sim 1\) or \(w \sim 1\) naively give \(b^{1/2}\) contribution, but they evaluate to simple power laws that do not contribute after following the above rules. We are left with the crossed regions \((z\sim b, w\sim 1/b)\) and \((z\sim 1/b, w\sim b)\). Note that the correlator satisfies \(C(z,w) = C(z^{-1},w^{-1})/(zw)\) under inversion. Also \(g_m(1/z) = g^*_m(z)/z\), one finds that the contribution from the two crossed regions are equal. [^2] Therefore we only need to study one crossed region, say \((z\sim 1/b, w\sim b)\): \[\begin{aligned}
\begin{split} \delta C(z,w) &= \frac{1}{2 \pi i z}\left( \theta(zw) + \frac{ (w/b+1)( z b-1) + w/b}{ (w/b+1)( z b-1)-w/b} \theta(-z) \theta(w) \right. \\ & + \left. \frac{ (w/b-1)( z b+1)-z/b}{ (w/b-1)( z b+1) + z/b} \theta(-z) \theta(w) \right) \end{split} \end{aligned}\] The wave function in this region is \[\begin{aligned}
g_{m}(z) g_{m}^\star(w) &\equiv \nonumber i (-)^m \left( (1+2w/b)^{-1/2+ i \nu_m}\theta(w) + (1-2w/b)^{-1/2-i \nu_m} \theta(-w) \right) \\& \times \left( (2/(zb)) (1+2/(zb))^{-1/2+i \nu_m} \theta(z) + (2/(zb)) (1-2/(zb))^{-1/2-i \nu_m} \theta(-z)\right) \end{aligned}\] It is convenient to use the scaled coordinates (related to \(\hat{z}, \widetilde{z}\) defined above): \[x = (1+2/(bz)) \,, \qquad y = (1+2w/b) \theta(w) + (1-2w/b))^{-1} \theta(-w)\] where we have assumed \(z > 0\). In the scaled region we cover \(0 < y< \infty\) and \(1 < x < \infty\). After rescaling to \(x,y\) coordinates we find: \[\begin{aligned}
\begin{split} 4 (-\ln(b/2)) \delta \lambda_m = \frac{ b (-)^m}{\pi} {\rm Re} & \left( \int_{1}^{\infty} d x x^{-1/2+i \nu_m} \int_0^1 d y y^{-1/2+i \nu_m} \frac{ 1+ x y}{y (x+y)} \right. \\ & \qquad + \left. \int_{1}^{\infty} d x x^{-1/2+i \nu_m} \int_1^\infty d y y^{-1/2+i \nu_m} \right) \end{split} \end{aligned}\] These later integrals converge if we analytically continue the various appearances of \(\nu_m\) differently. While this is rather crude, it does the job of extract the non-power law divergences. The final result is: \[\label{eq:lambdashift} \delta \lambda_m = \frac{b}{2 (-\ln b/2)} \frac{(-)^m \sech(\pi \nu_m)}{(1+ 4\nu_m^2)} + \mathcal{O}(b^2)\] This leads to a shift in the inferred spectrum: \[\delta m = \frac{b}{2} \frac{(-)^m \cosh(\pi \nu_m)}{\pi^2(1/4+ \nu_m^2)} + \ldots\] Which competes with the leading order answer when \(m \sim \frac{2}{\pi^2} (-\ln b/2)^2\). While we don't expect perturbation theory to break down here (since \(\delta \lambda_m\) remains small), it will become difficult to extract the spectrum at this order of \(m\), since the density matrix has an exponentially small depends on these energy shifts. These shifts that we found analytically agrees surprisingly well with our numerics, see Figure [\[fig:lambdashift\]](#fig:lambdashift){reference-type="ref" reference="fig:lambdashift"}: With the eigenvalue corrections at hand, one can check that the correlation matrix (hence also the density matrix) converges to the thermal answer [\[eq:lambdaleading\]](#eq:lambdaleading){reference-type="eqref" reference="eq:lambdaleading"} by examining the distance between two matrices. Given spectrum of two operators one can define the *spectral distance* by minimizing over the sum of the differences between two set of eigenvalues. The spectral distance is known to be equivalent to the distance between unitary orbits of two operators. In our setting this is simply \[\sum_{m} |\delta\lambda_m| = \sum_m\frac{b}{2(-\ln b/2)} \frac{\sech (\pi \nu_m)}{(1+4\nu^2_m)}\] It is easy to see that the sum is upper bounded by \[\begin{aligned}
\begin{split} \sum_{m} |\delta\lambda_m| \le \frac{b}{(-\ln b/2)} \left(\int^\infty_{1/2} \sech(\pi \nu_m)dm + \sech(\pi\nu_1)\right) = 2\pi b + \mathcal{O} (b/\ln(b)) \end{split} \end{aligned}\] which approaches zero as \(b\to 0\). Repeating the same analysis for deflected entropy correlator \(C^s(z,w)\) with non-zero \(s\), one gets (For details see Appendix [\[app:B\]](#app:B){reference-type="ref" reference="app:B"}): \[\begin{aligned}
\label{eq:s_lambdashift} \delta\lambda^s_m = \frac{b^s}{2(-\ln b^s/2)}\frac{(-)^m\sech(\pi \nu^s_m)}{(1+4\nu^{s}_m)^2} \left( \cosh(\pi s)\cos(2\pi s \nu^s_m)-\frac{\sinh(\pi s)\sin(2\pi s\nu^s_m)}{2\nu^s_m} \right) + \ldots \end{aligned}\] where \(b^s = b/\cosh(\pi s)\) and \(\nu^s_m = \frac{\pi (m-1/2)}{-2\ln(b^s/2)}\) is the effective spectrum parameters for \(s\ne0\). We see that it reduces to the unflowed case when \(s=0\). It gives a prediction of deflected entropy that agrees well with our numerics (Figure [\[fig:s_plots\]](#fig:s_plots){reference-type="ref" reference="fig:s_plots"}).
# Discussion {#sec:discussion}
In this note we calculated the reflected entropy of a chiral free fermion both numerically and analytically, and we give asymptotic formulas for \(S_R\) in both the limits \(x\to 0\) and \(x\to1\). In particular we have shown that in the latter case, the reflected spectrum of a free chiral fermion in the NS sector approaches a thermal spectrum with inverse temperature to length ratio \(\beta/L \propto-1/\ln(1-x)\). Indeed this temperature coincides with the expected answer one would have obtained in AdS/CFT. This thermal spectrum is theory dependent, as opposed to the universal spectrum of entanglement entropy in 2d CFTs which one can regard as being in the infinite temperature limit. As a practical matter one can then use the reflected entropy to extract the spectrum of operator dimensions directly from the vacuum wave-function. In this regard, the reflected entropy can be used as a diagnostic of different topological orders. However, due to the perturbative corrections and Boltzman suppression, the weights extracted in this way are only perturbatively reliable for a window of operator weights satisfying: \[\label{mod} h \lesssim \frac{1}{2\pi^2} ( \ln (1-x) )^2\] It is natural to conjecture that the above statement are more general statements about the reflected entanglement spectrum for any 2d CFT, with likely some modification of [\[mod\]](#mod){reference-type="eqref" reference="mod"} to include a varying central charge. This work is partially supported by the Air Force Office of Scientific Research under award number FA9550-19-1-0360 and the Department of Energy under award number DE-SC0019183.
[^1]: The appropriate value of \(L\) where we see convergence is highly dependent on the cross-ratio \(x\) as well as desired precision. In general we need higher \(L\) when \(x\to 1\). For the plots in this note the value of \(L\) ranges from \(20\) to \(500\).
[^2]: This follows from the scaling behavior of the 2D free Fermions. | {'timestamp': '2022-12-01T02:21:30', 'yymm': '2211', 'arxiv_id': '2211.17255', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17255'} |
null | null |
null | null |
null | null |
# Introduction {#sec:intro}
Electronic health records (EHR) are collected as part of routine medical care, and include demographic information, disease diagnoses, laboratory results, medication prescriptions, etc., providing a patient's clinical state over time. Recent machine learning techniques have been used to exploit the richness of EHR data at scale for diagnosis, prognosis, treatment and understanding of disease. Many medical terminologies are used across clinical data sets, and the standard practice involves mapping clinical terms across different resources or onto common data models. Here, we propose sEHR-CE (language modelling of **S**tructured **EHR** data for patient **C**ohort **E**xpansion), a novel framework based on transformers to enable the integrated analysis of multiple clinical resources without relying on any manual curation and mapping. Using text descriptions of concepts as input, our method generalises across data modalities and terminologies, i.e. text and structured EHR. This enables us to leverage a plethora of pre-trained language models like PubMedBERT to encode each patient's medical record. We ask the model to learn representations of clinical histories from diagnosed patients (cases) to predict phenotypes (such as diseases). In the absence of a directly comparable model, we evaluate the performance of our model against that of , as the state-of-the-art approach for learning clinical term embeddings for future disease prediction. See Appendix [\[sec:related_work\]](#sec:related_work){reference-type="ref" reference="sec:related_work"} for a more complete survey. In our validation experiments, we will show that the model can then identify individuals with missing diagnoses (controls) that share a similar clinical history with cases, indicating they might have the disease or be at risk of developing it. The contributions of our paper are **(i)** the presentation of a cohort expansion method that provides phenotype predictions outperforming non-text and single terminology approaches, and **(ii)** an in-depth qualitative evaluation demonstrating that positively predicted controls share similar clinical representations with cases, providing a high degree of evidence that these may be previously undiagnosed or misdiagnosed individuals. Finally, we demonstrate our method's potential for patient stratification by disease severity.
# Methods {#sec:method}
**Input Generation**. For each EHR source and associated ontology \(a \in \mathcal{A}\), we denote the set of concepts (e.g. clinical terms) as \(\Theta_a\), and the set of text descriptions as \(\Xi_a\). The total vocabulary of concepts and text descriptions across all ontologies is denoted by \(\Theta = \bigcup\limits_{a \in \mathcal{A}} \Theta_a\) and \(\Xi = \bigcup\limits_{a \in \mathcal{A}} \Xi_a\), respectively. For each patient, we define their full clinical history through time and across sources by the concatenation of sequences of clinical descriptions \(( \textbf{Label Generation}. Let\)=(d_1, ..., d_D)\(denote an ordered set of clinical outcomes or events, in our case disease phenotypes\)d_i\(. For each phenotype\)d_i\(, we let\)\_d_i\(be an indicator function that assigns a binary label to individuals according to the presence or absence of\)d_i\(. To define\)\_d_i\(, we rely on external oracles or phenotype definitions, such as CALIBER (Appendix \ref{labelling}). Figure \ref{figure:labelling} shows a schematic of input and label generation for a given patient. \footnotetext{Any patient data shown is simulated and does not represent data of real patients.} \textbf{Model Design}. Let\)X\^(p) = (x\^(p)\_1, ...,x\^(p)\_n)\(denote the tokenized input sequence of individual\)p\(. It forms the input to an encoding function\)\^(p)\_1, ..., \^(p)\_n\(= \textit{Encoder}\)(X\^(p))\(, where each\)\_i\(is a fixed-length vector representation of each input token\)x_i\(. Let\)\^(p)=(y\^(p)\_1, ..., y\^(p)\_D)\(,\)y\^(p)\_i{0,1}\(, be labels denoting presence or absence of phenotypes\)d_1,..., d_D\(. Given a learned representation over inputs, we decode over\)\^(p)\(under the predictive model\)(\^(p)\|X\^(p))\(. We calculate the probability of each phenotype\)d_i\(given the input sequence encoding\)(y\^(p)\_i\|**x**\^(p)\_1,..., **x**\^(p)\_n)\(via a decoder module. Specifically, we decode the representation into logits per phenotype\)z\^(p)\_1, ..., z\^(p)\_D =\(\textit{Decoder}\)(**x**\^(p)\_1,..., **x**\^(p)\_n)\(and calculate the probability per phenotype as\)(y\^(p)\_d\|**x**\^(p)\_1,..., **x**\^(p)\_n) = (z\^(p)\_d)\(, where\[denotes the sigmoid function (Figure \ref{figure:model_diagram_small}). We will omit the superscript\)(p)\(denoting the sample index in the remainder of the text. \subsection{Data Augmentation}\label{data_augmentation} \textbf{Clinical Masking}. We mask input descriptions\) 10% probability, and **replace** \(Figure \ref{figure:augment} displays a worked example. During testing, these term-description pairs are fully removed from the input sequence. \textbf{Comorbidities}. This masking approach is straightforward if an individual has only one positive label, but many people have comorbidities, e.g. co-occurring conditions. To allow for a sample with multiple positive labels\)d\_i_1, ..., d\_i_n\(, we create\)n\(input samples by replicating both the input sequence and target vector of phenotype labels. Consider the\)j\(th replicated input sequence, and let\)d = d\_i_j\(. We mask this replicated input sequence with the masking strategy for phenotype\)d\(described in the previous section; e.g.\)
## Defining a Loss Function in the Presence of Comorbidities and varying Prevalences {#loss_weights}
**Loss weights**. The data augmentation approach for comorbidities may lead to our model overfitting when an input sequence contains the descriptions associated with an existing phenotype \(d\) that is not masked. To account for the contribution of the target vector \(\mathbf{y}\) to the loss function in these scenario, we define a binary masking vector \(\mathbf{\gamma}^{j}=(\gamma^{j}_1, \ldots, \gamma^{j}_D)\) where \(\gamma^{j}_j=1\) and \(\gamma^{j}_k=0\), \(k=1,\ldots,D\), \(k\not=j\). Individuals with no positive phenotype labels are assigned an all-zero masking vector (Figure [\[figure:masking_comorbidities\]](#figure:masking_comorbidities){reference-type="ref" reference="figure:masking_comorbidities"}). Then, for a given input text sequence \(X\), target label vector \(\mathbf{y}=(y_1, \ldots, y_D)\) and masking vector \(\mathbf{\gamma}=(\gamma_{1}, \ldots, \gamma_{D})\), we define the loss weights by setting \(\omega_{d} = 0\) if \(y_{d} = 1\) and \(\gamma_d = 0\), and \(\omega_{d} = 1\) if \(y_{d} = 0\) or \(y_{d} = 1\) and \(\gamma_d = 1\). **Positive weights**. Because some diseases are very rare, whereas others are very common, our prediction classes are expected to be highly imbalanced in the practical setting. For cohort expansion, we want to increase recall while balancing a decline in precision. For \(d\in\Delta\), let \(TN_d\) and \(TP_d\) denote the total number of negative and positive examples of \(d\) in the training set. We define the positive weight \(\rho_d\) as \(\rho_d=TN_d/TP_d\), for all d \(\in\Delta.\) Our loss function is defined as a mean-reduced binary cross-entropy loss function over phenotypes, where disease prevalence and comorbidities are handled with loss and positive weights: \]\label{lossfunction} \mathcal{L}^{(p)}=-\frac{1}{|\Delta|} \sum_{d \in \Delta} \omega_{d}^{(p)} (\rho_dy_{d}^{(p)} \log \sigma({z_{d}^{(p)}}) \\ + (1-y_{d}^{(p)}) \log \sigma(1-z_{d}^{(p)})),\(\) where \(\sigma\) denotes the sigmoid function, \(\omega_d\) the comorbidity-derived loss weight, \(\rho_d\) the positive weight, and \(z_{d}^{(p)}\) the predicted probability for phenotype \(d\in \Delta\) for sample (e.g. individual) \(p\).
# Experiments and Results {#sec:results}
**Data.** This research has been conducted using the UK Biobank (UKBB) Resource under Application Number 43138, a large-scale research study of around \(500\)k individuals. We restrict the data set to those that have entries in both hospital and GP records, reducing our cohort to \(155\)k. To assess the quality of our model predictions, we choose four diseases that differ in terms of prevalence and clinical characteristics (Appendix [\[apd:model\]](#apd:model){reference-type="ref" reference="apd:model"}): Type 2 diabetes mellitus (T2DM), Heart failure (HF), Breast Cancer, and Prostate Cancer. We use validated phenotype definitions from CALIBER to label patients with each of the diseases (Appendix [\[apd:data_proc\]](#apd:data_proc){reference-type="ref" reference="apd:data_proc"}). We test the performance of our model on its ability to diagnose known cases, compare it to other methods, and evaluate associations of clinical features with predictions on T2DM. **sEHR-CE.** We use the pre-trained language model PubMedBERT as the encoder of the tokenised input sequences of clinical term descriptions. Since our input systematically differs from the general scientific text on which PubMedBERT was trained, we fine-tune on the masked-language modeling (MLM) task using the full UKBB cohort. The proposed model sEHR-CE uses the fine-tuned encoder and a fully connected linear layer as the decoder. To train on the multi-label classification task of outcome prediction, we split our data set into five equally sampled folds *\(f_0\)\...\(f_4\)* containing unique patients, and mask the data according to our strategy (Section [\[data_augmentation\]](#data_augmentation){reference-type="ref" reference="data_augmentation"}, Figure [\[figure:input_diagram\]](#figure:input_diagram){reference-type="ref" reference="figure:input_diagram"}). We train a total of five models on three folds, holding back folds \(f_i\) for validation and \(f_{(i+1)\bmod{5}}\) for testing for model \(i\), \(i=1,\ldots,5\) (Figure [\[figure:folds_diagram\]](#figure:folds_diagram){reference-type="ref" reference="figure:folds_diagram"}). All results presented are predictions on the independent test set. For masking and training details, see Appendix [\[apd:training\]](#apd:training){reference-type="ref" reference="apd:training"}. **Baseline Models**. We compare the performance of our model sEHR-CE to BEHRT, which takes a tokenised sequence of clinical terms, age and position embeddings as input. Ontologies from hospital and GP records are mapped to CALIBER definitions, removing unmapped terms (more details in Appendix [\[apd:eval\]](#apd:eval){reference-type="ref" reference="apd:eval"}). We train five such BEHRT models to predict an individual developing the four phenotypes on the same five data splits as sEHR-CE. Similarly, we train five sEHR-CE models restricted to CALIBER tokens (denoted sEHR-CE-codes) for comparison. Figure [\[figure:disease_dist_all\]](#figure:disease_dist_all){reference-type="ref" reference="figure:disease_dist_all"} shows the distributions of predicted probabilites for all phenotypes across all methods. sEHR-CE shows the best performance across all four phenotypes in terms of recall at \(0.5\) and AUC on the test set (Table [\[table:results_SOTA\]](#table:results_SOTA){reference-type="ref" reference="table:results_SOTA"}, Figure [\[figure:auprc\]](#figure:auprc){reference-type="ref" reference="figure:auprc"}). BEHRT performs slightly better than sEHR-CE-codes, indicating a benefit of adding visit position and age. Performance varies across phenotypes, presumably due to different clinical characteristics making some diseases easier to predict than others.
## Clinical evaluation on Type 2 Diabetes Mellitus
T2DM lends itself as a use case to qualitatively evaluate the predictions of missed cases, as it is a well-studied, slowly developing disease with varying disease severity. We used thresholds based on the 98\(^{th}\), 90\(^{th}\) and 12\(^{th}\) percentiles of sEHR-CE's predicted probabilities (Figure [\[figure:disease_dist_all\]](#figure:disease_dist_all){reference-type="ref" reference="figure:disease_dist_all"}) to define five different groups (Table [\[table:patient_groups\]](#table:patient_groups){reference-type="ref" reference="table:patient_groups"}): regular cases, cases predicted with high probability, cases predicted with low probability, regular controls and controls predicted with high probability (missed cases). **Measures of Disease Severity**. Haemoglobin A1c (HbA1c) is a blood biomarker used to diagnose and monitor diabetes. A level of 48mmol/mol or higher is considered diabetes; while a value between 42 and 48 mmol/mol is considered pre-diabetes. The input data did not include such biomarker data, so we use it here for evaluation. We aggregated HbA1c measurements taken in primary care (GP) of each individual by taking the \(95\)-th percentile value. Figure [\[figure:hba1c\]](#figure:hba1c){reference-type="ref" reference="figure:hba1c"} shows that cases predicted with high probability had the highest HbA1c mean levels. Cases identified with low probability were in the prediabetic range of HbA1c levels, indicating a less severe state. Missed cases had elevated HbA1c levels close to the prediabetic stage, representing individuals at risk of developing T2DM. We further investigated the association of the T2DM predicted probabilities with other measures of disease severity, expanded in Appendix [\[apd:t2dm_eval\]](#apd:t2dm_eval){reference-type="ref" reference="apd:t2dm_eval"}. Taken together, our results demonstrate that sEHR-CE's predicted probabilities of being diagnosed with T2DM are associated with disease severity. **Polygenic risk scores**. Genetic risk for complex diseases like T2DM arises from many genetic changes that, when taken together, can increase an individual's risk of developing the disease, which can be defined by polygenic risk scores (PRS). developed PRS based on the UK Biobank participants for a set of diseases, including T2DM. We computed and standardised T2DM PRS across all individuals in our cohort. Figure [\[figure:prs\]](#figure:prs){reference-type="ref" reference="figure:prs"} demonstrates that higher predicted probability of T2DM was associated with a higher genetic risk.
# Conclusion {#sec:conclusion}
We presented a data-driven method for cohort expansion based on language modelling. Our approach fuses primary and secondary care data via text, and we propose a data augmentation approach to allow for comorbidities in a patient's history. Our method predicts disease phenotype labels more accurately than non-text and single terminology approaches. We presented a high degree of evidence that our model identifies previously undiagnosed individuals that can extend the original cohort for downstream analysis. Future work will consider methods that are less restrictive on sequence length and allow for irregular time steps and age, as well as adding more data sources (e.g. medications). | {'timestamp': '2022-12-01T02:18:23', 'yymm': '2211', 'arxiv_id': '2211.17121', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17121'} |
null | null |
# Methods
## Device fabrication
The WSe\(_2\)/MoSe\(_2\)-heterobilayer devices are fabricated from exfoliated flakes with a dry-transfer method, protected by hBN from both sides and contacted via graphite electrodes (see schematic in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a; detailed information about device fabrication can be found in Refs. ). For STA device \#1 and LTA devices \#1 and \#3 we furthermore incorporated a graphite gate, whereas for STA device \#2 we use the Si\(^{++}\)/SiO\(_2\)() wafer for gating. An exception from this general fabrication procedure is LTA device \#2 that is used for the ARPES measurements and therefore cannot be protected by hBN. For this device we directly placed a MoSe\(_2\)/WSe\(_2\)/graphite-stack onto an n-doped silicon wafer.
## Measurements
Detailed technical information about the time-resolved Kerr rotation and photoluminesence measurement techniques and the used fitting procedures can be found in the Supporting Information sections of Refs. . We note that it is important to account for photo-induced screening effects of the gate electric field to unveil the true lifetimes and amplitudes of the Kerr signal that can be attributed to free charge carriers. In the Supporting Information of Ref. we show how the standard optical measurement techniques can be modified to account for the photo-induced gate screening and we demonstrate that disregarding this procedure can lead to erroneous conclusions drawn from gate-dependent measurements. **Time-resolved Kerr rotation:** Two mode-locked Ti:sapphire lasers are used to independently tune the energies of both pump and probe pulses. An electronic delay between both pulses covers the full laser repetition interval of . The pulse widths are on the order of . The laser beams are focused onto the device by a focal length aspheric lens (numerical aperture of 0.66 NA). Typical spot diameters are \(\unit[6-8]{\mu m}\) measured as a full width at half maximum (FWHM) value. The laser power was around \(\unit[500]{\mu W}\) for both pump and probe beams. **Photoluminescence:** A microscope objective lens (numerical aperture of 0.5 NA) is used for the photoluminescence measurement. The resulting spot size is around \(\unit[1-2]{\mu m}\) FWHM. A continuous wave laser with an energy of and a power between \(\unit[10-50]{\mu W}\) is used for these measurements. **Angle-resolved photoemission spectroscopy:** The ARPES measurements have been performed at the NanoESCA beamline of Elettra, the Italian synchrotron radiation facility, using a FOCUS NanoESCA photoemission electron microscope (PEEM) in the k-space mapping mode operation . The PEEM is operating at a background pressure \(p < \unit[5\times 10^{-11}]{mbar}\) and the photoelectron signal is collected from a spot size of about \(\unit[5-10]{\mu m}\). Before the experiment the sample was outgassed in UHV at \(T=\unit[180]{^\circ C}\) for . The measurements were conducted with a photon energy of and an overall energy resolution of using p-and s-polarized synchrotron radiation, while keeping the sample at .
# Second Harmonic Generation
The twist angles in the heterobilayer devices were determined by measurements of the polarization dependence of the second harmonic (SH) generation intensity. In monolayer samples, this dependence immediately indicates the orientation of the crystallographic axes . In a heterobilayer, the SH field is given by a vector superposition of the individual SH fields of both monolayers, which allows to distinguish natural stacking with almost complete destructive interference from the unnatural stacking with constructive interference . We measured the SH intensity on parts of isolated monolayers of both materials as well as on the bilayer part. Fig. [\[S1\]](#S1){reference-type="ref" reference="S1"} shows a representative measurement of the intensities of the SH emission as a function of the azimuthal angle of the linearly polarized incident laser in case of LTA device \#1 of the main manuscript. The data of the monolayers was fitted with the function \(I_{\Theta} = I_0 \cdot \sin^2(3\Theta)\). The difference of the numerically optimized angles \(\Theta\) between the different areas is the relative twist angle given in the main manuscript. In the case of small angles between the crystallographic axes (STA devices \#1 and \#2), the SH intensity of the bilayer was compared to the intensity of both monolayers, which identifies the twist angle of STA device \#1 as \(7^\circ\) and of STA device \#2 as \(56^\circ\), which is a small twist angle of \(60^\circ-56^\circ=4^\circ\) to an H-type stacking order. Due to quite different sizes of the two TMD flakes in LTA device \#3 and the use of the monolayer parts for electrical contacts, we were not able to obtain the necessary SHG measurements on the respective monolayer areas in this device. Therefore, we cannot give a twist angle for this device. However, we note that we intentionally misaligned the edges of the TMD flakes during the stacking process to achieve a large twist angle. The PL and TRKR measurements indicate that the large twist angle was indeed achieved. The laser used in the experiments had a pulse duration of approximately , a repetition rate of and a central wavelength of . The average power used was between and .
# Sensitivity of Kerr signal to free charge carriers
As we have demonstrated in Ref. , time-resolved Kerr rotation is not only sensitive to valley-polarized excitons but can also detect a valley polarization of free charge carriers, if the energy of the probe pulse is tuned to the trion (charged exciton) energy of the corresponding TMD. This is due to the fact that both the helicity of the incoming light and its energy determines what type of trion (singlet or triplet) is formed and in which valley the trion is created . As the formation of trions necessitates the presence of additional charge carriers , the free charge carrier density inside the valley directly impacts the probability with which photons with matching helicity and energy interact with the corresponding valley. The sensitivity of the linearly polarized probe pulse on a valley polarization can therefore be explained by the fact that the probe pulse is a superposition of left and right circularly polarized light, which have different probabilities to interact with the TMD in the presence of a valley polarization of free charge carriers. Of course, this is a simplified explanation as the interaction between a laser pulse and a very thin material on a reflective substrate is highly complex .
# Band structure calculations
We performed the first principles calculations of the R-stacked (0° twist) and large twist angles (21.8°, 38.2°, 32.2° and 27.8°) MoSe\(_2\)/WSe\(_2\) heterostructures based on the density functional theory using the all-electron full-potential linearized augmented plane-wave method implemented within the Wien2k package . We used the Perdew--Burke--Ernzerhof exchange-correlation functional with van der Waals interactions accounted via the D3 correction . The wavefunction expansion into atomic spheres takes into account orbital quantum numbers up to 10 and the plane-wave cut-off multiplied with the smallest atomic radii is set to 7. Spin--orbit coupling was included fully relativistically for core electrons, while valence electrons were treated within a second-variational procedure with the scalar-relativistic wavefunctions calculated in an energy window up to 2 Ry. Self-consistency was achieved using a two-dimensional Monkhorst--Pack k-grid with 12x12 (6x6) points for the R-stacked (twisted) structure and the convergence criteria of \(10^{-5}\) e for the charge and \(10^{-5}\) Ry for the energy was used. We built commensurate heterostructures with a common in-plane lattice parameter of 3.2855 \(\AA\) for MoSe\(_2\) and WSe\(_2\) layers (as an average value of their individual lattice parameters ). This average lattice parameter leads to an effective strain of \(\sim 0.1 \%\) to the individual layers, which is sufficiently small to affect the energy-and spin-dependent properties . A vacuum region of 16 \(\textrm{\AA}\) was assumed in all cases. We minimized the energy to find the equilibrium interlayer distances, obtaining 3.724 \(\textrm{\AA}\) for the Rhh stacking, 3.074 \(\textrm{\AA}\) for the RhM stacking, 3.084 \(\textrm{\AA}\) for the RhX stacking, 3.337 \(\textrm{\AA}\) for the 21.8° and 38.2° twisted structures and 3.365 \(\textrm{\AA}\) for the 32.2° and 27.8° twisted structures. In Fig. [\[S2\]](#S2){reference-type="ref" reference="S2"} we present the top view of each crystal structure considered in our study, namely, the Rhh, RhM and RhX stackings (0° twist angle) and the large twist angle cases 21.8°, 38.2°, 32.2° and 27.8°.
## Small twist angles
Fig. [\[S3\]](#S3){reference-type="ref" reference="S3"} shows the calculated band structures for the three different R-type stackings at \(\unit[0]{^\circ}\) twist-angle (see Fig. [\[S2\]](#S2){reference-type="ref" reference="S2"} for a description of these stacking orders). Depending on the type of stacking, the positions of the \(\Gamma\)-and Q-valleys can shift quite significantly with respect to the K-valleys. Furthermore, together with external parameters like interlayer distance (which might be impacted by fabrication-induced contamination between the TMD layers) or strain , the difference in stacking order might be the reason that STA device \#2 shows a less pronounced hybridization than STA device \#1 despite its smaller twist-angle between the crystallographic axes. The color-code in Fig. [\[S3\]](#S3){reference-type="ref" reference="S3"} depicts the spin-texture in the out-of-plane direction. Whereas the conduction band minimum at the Q-valley is showing a small spin-splitting, the valence band maximum at the \(\Gamma\)-valley is spin-degenerated, therefore likely enhancing spin-flip scattering in the valence band as described in the main manuscript.
## Large twist angles
Fig. [\[S4\]](#S4){reference-type="ref" reference="S4"} depicts the calculated band structures for the several large twist angles analyzed. We note that all of our calculations resulted in a type II band alignment, consistent with previous theoretical studies, but at odds with our experimental data. So far, theoretical studies only show a transition from a type II to a type I band alignment in case of large displacement fields . However, the necessary strength of the displacement field lies far beyond our applicable gate voltage range. Furthermore, in our devices we do not have the combination of both a bottom and a top gate, which is necessary to create a displacement field without simultaneously inducing charge carriers into the heterostructure . On the other hand, there is one experimental study that argues to observe a transition from a type II to a type I band alignment in case of large charge carrier densities . In this respect, we note that to our knowledge all previous band structure calculations of TMD heterostructures (including our own calculations) are done under the assumption of charge neutrality. This limitation is due to an increase in complexity of the models and especially a significant increase in required computational power if doping is also taken into account. Nonetheless, there has been theoretical efforts along this line but still restricted to monolayers . As most of our devices show a slight n-doping, we therefore propose that the unexpected band alignment, which is derived from our experimental data, might be caused by a combination of the large twist angle together with the presence of donor states in both TMD layers. It has to be the scope of further studies to investigate this idea further. Our calculations can also shed some light on the observation that most WSe\(_2\)/MoSe\(_2\)-heterobilayers (especially the ones with small twist angle) show a stronger quenching of the WSe\(_2\) intralayer exciton compared to the MoSe\(_2\) intralayer exciton, as mentioned in the main manuscript. Table [\[table-S1\]](#table-S1){reference-type="ref" reference="table-S1"} contains calculated transition rates for neutral excitons at the K-point for the different R-stackings and large twist angles analyzed. Whereas the transition rate for the interlayer exciton significantly drops towards larger twist angles (explaining the vanishing of the interlayer exciton in PL measurements of devices with large twist-angle), the intralayer transitions are barely effected by the twist angle and are similar to the ones in the respective monolayer case. Most importantly, the two bright exciton transitions of WSe\(_2\) \[VB+(W)\(\rightarrow\)CB+(W)\] and MoSe\(_2\) \[VB+(Mo)\(\rightarrow\)CB-(Mo)\] have very comparable transition rates. To understand the often missing WSe\(_2\) exciton emission in heterobilayers, we now only have to take into consideration that the intensity of a PL signal depends on both the transition rate and the occupation number in the respective states. As the transition rates are very comparable to each other, our calculations therefore suggest that photo-excited charge carriers must scatter quite fast away from either the valence band maximum or the upper, spin-split conduction band minimum of WSe\(_2\). We assume that the upper, spin-split conduction band minimum is actually responsible for this effect. To support our assumption, we present in Fig. [\[S5\]](#S5){reference-type="ref" reference="S5"} the band structures of the large twist angle cases highlight the contributions of the spin-up states of each layer, denoted by the molybdenum and tungsten atoms. Spin-down states are just the time-reversal partners of spin-up states are are not show for simplicity. We observe a magnitude of zone-folded bands near the conduction band minima of WSe\(_2\). Because of these bands, we assume that photo-excited electrons quite efficiently scatter away from the conduction band K-valleys of WSe\(_2\) and therefore are not available for the optical recombination process. On the other hand, at the K-valley both the valence and conduction bands of MoSe\(_2\) are quite well-separated from other bands. Therefore, scattering might take longer, which in turn increases the chance of optical recombination, which makes the corresponding PL signal to a quite prominent one in our measurements. Whereas these zone-folded bands can explain the significantly quenched PL emission from WSe\(_2\) in heterostructures that exhibit the widely assumed type II band alignment, the quite strong PL emission from WSe\(_2\) in both LTA devices \#1 and \#3 demonstrate that these devices do not possess these pronounced zone-folded bands near the conduction band minima of WSe\(_2\). It must be the scope of further studies, if an alignment of the conduction band minima of WSe\(_2\) and MoSe\(_2\) caused by donor states (as derived in the main manuscript) has this effect. | {'timestamp': '2022-12-01T02:20:30', 'yymm': '2211', 'arxiv_id': '2211.17210', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17210'} |
# Introduction
Planar group-III arsenide heterostructures enable optoelectronic devices such as vertical cavity surface emitting lasers in the near infrared spectral range. However, the monolithic integration of these GaAs-based devices on the Si platform remains an issue that has resisted a solution for decades despite considerable international research efforts. The high lattice mismatch between GaAs and Si results in threading dislocation densities in excess of \(10^7\,\)cm\(^{-2}\), and the high thermal mismatch leads to wafer bowing and cracks. Threading dislocations act as nonradiative defects, and their presence seriously degrades the internal quantum efficiency of any GaAs-based light emitter on Si. Both of these apparently insurmountable obstacles cease to be issues for GaAs-based nanowires (NWs) on Si\((111)\). Due to the limited footprint of these nanostructures, wafer bowing and cracks cannot occur. Furthermore, threading dislocation arms bend toward the NW surfaces close to the heterointerface, resulting in the complete absence of threading dislocations and the associated nonradiative recombination in GaAs NWs. Additionally, the high geometric aspect ratio of NWs offers a potentially more efficient absorption and extraction of light when compared to planar structures. Consequently, the monolithic integration of III-V NW devices on Si substrates has attracted considerable attention, and has been successfully demonstrated by several groups. GaAs NWs can be synthesized on Si(111) by both metal-organic vapor phase epitaxy (MOVPE) and molecular beam epitaxy (MBE). Most frequently, the vapor-liquid-solid (VLS) growth mode is employed, using either Au or Ga as the metal particles providing supersaturation, but growth in the vapor-soild (VS) growth mode has also been demonstrated. In any case, the NWs grow preferentially along the \([111]\)B direction, and are, regardless of growth technique and growth mode, prone to polytypism, i e., the stacking sequence along the length of the nanowire may change spontaneously between that of the zincblende or the wurtzite crystal structure. Crystallographically, polytypism manifests itself in stacking defects, namely, twin boundaries due to rotational twinning in the zincblende structure, and one of various types of stacking faults in the wurtzite structure. These defects may occur in an irregular sequence, and also longer segments of either crystal structure are possible. Electronically, these defects matter since the band alignment between wurtzite and zincblende GaAs is of type II. Thus, in the presence of stacking defects, spatially indirect, sub-bandgap transitions can occur between electrons and holes at the interfaces of the polytypes. If the density of stacking defects is sufficiently high, the emission spectra of the NWs are dominated by these parasitic transitions, completely masking the intrinsic optical properties of the GaAs NWs. Several studies attempted to establish a correlation between microstructure and the transitions observed in single GaAs NWs.. In some cases, single NWs were found exhibiting only a few stacking defects, thus displaying spectra dominated by the transitions typically observed in GaAs\((001)\) layers, particularly, the free exciton at higher excitation densities. Much more rare, but mandatory for any applications are ordered GaAs NW arrays with this desirable property, since any of these applications demand a deterministic and repeatable emission wavelength. reported photoluminescence (PL) spectra dominated by the free exciton with a full-width at half-maximum of \(7\) meV for random GaAs/(Al,Ga)As core/shell NW ensembles synthesized by MOVPE on GaAs\((111)\)B substrates. This group used the exceptionally high phase purity of their GaAs NW ensembles for detailed optical studies in subsequent work, deriving, for example, a value for the Mott density in GaAs NWs. Also using GaAs\((111)\)B substrates, obtained even narrower exciton lines (\(3\)--\(5\) meV depending on Al content) for site-controlled GaAs/(Al,Ga)As nanomembranes (NMs) fabricated by MBE. The authors of the latter work demonstrated the high uniformity of their arrays by comparing PL spectra of single NMs with those of the entire array, which were found to be essentially identical. An even more important figure-of-merit for optoelectronic applications is the minority carrier lifetime \(\tau\) or the internal quantum efficiency \(\eta_\text{int}=\tau/\tau_r\), where \(\tau_r\) is the radiative lifetime. Note that the stacking defects occurring in polytypic GaAs NWs induce spatially indirect radiative transitions of axially localized electron-hole pairs, and thus do not decrease, but rather tend to increase both \(\tau\) and \(\eta_\text{int}\). In contrast, the growth mode may play a crucial role, as shown by in their comparison of Au-and Ga-assisted GaAs/(Al,Ga)As NWs on Si\((111)\). Absolutely essential for obtaining high values for \(\tau\) and \(\eta_\text{int}\) is an efficient surface passivation because of the high surface recombination velocity of the bare GaAs\((110)\) sidewall surfaces. The most convenient and effective passivation technique consists in the radial overgrowth of the GaAs core with (Al,Ga)As. Still, even with an (Al,Ga)As shell, a finite interface recombination velocity remains, and the configuration of the passivating shell has been shown to affect \(\tau\) in various ways. Except for the few facts summarized above, we actually know very little about the mechanisms determining \(\tau\) and \(\eta_\text{int}\) in GaAs/(Al,Ga)As NWs. This statement applies to both radiative and nonradiative contributions, and in particular their dependence on excitation level. A thorough understanding of the carrier recombination dynamics in phase-pure GaAs/(Al,Ga)As NWs is essential for judging their eligibility for optoelectronic applications. In the present work, we use µ-PL spectroscopy to study the carrier recombination dynamics in ordered arrays of zincblende GaAs/(Al,Ga)As core/shell NWs grown on Si\((111)\) by MBE. The PL spectra of these NW arrays resemble those of the state-of-the-art NW and NM ensembles on GaAs\((111)\)B and are essentially free of transitions induced by stacking defects down to the lowest excitation levels. Combining complementary spectroscopic experiments and simulations, we extract the internal quantum efficiency at \(10\) K for excitation levels spanning over more than five orders of magnitude, estimate the extraction efficiency, and measure the thermal quenching of the internal quantum efficiency for selected excitation levels. Furthermore, we observe an abrupt Mott transition from exciton to electron-hole-plasma recombination, allowing us to determine the Mott density.
# Results and discussion {#sec:results}
## NW morphology and PL spectra
Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} depicts bird's eye (\(15^{\circ}\)) view secondary electron micrographs as well as low-temperature (nominally \(10\,\)K, actually \(\approx 25\,\)K--see the Supporting Information for details) continuous wave (cw) µ-PL spectra of different microfields containing GaAs/Al\(_{0.33}\)Ga\(_{0.67}\)As core/shell NW arrays (sample A). The selected microfields have the same hole size of \(90\) nm, but different pitches of (a,e) \(200\), (b,f) \(700\) and (c,g) \(5000\) nm (see Sec. for the definition of hole size and pitch). The secondary electron micrograph displayed in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a reveals a dense array of NWs with different lengths. The majority of NWs does not exceed a length of \(2\) m, but some approach the intended length of \(4\) m. This observation is likely an effect of shadowing of the Ga flux in the dense NW array, as investigated both experimentally and theoretically for Au-assisted InAs and Ga-assisted GaAs NWs. The bending of the long NWs arises from electrostatic effects during the exposure of the NWs to the electron beam in the scanning electron microscope. Figures [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}b,c show that the vertical yield improves drastically for fields with a larger NW pitch. In both cases, the micrographs show NWs of about \(4\) m length and \(160\) nm diameter (the GaAs core is \(100\,\)nm-thick) in an ordered hexagonal arrangement. The vertical yield exceeds \(90\)% and parasitic growth is almost absent. The µ-PL map displayed in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}d, which corresponds to the microfield of panel c, shows that only completely developed NWs exhibit a notable PL intensity, while the rare parasitic growth or NW stumps do not emit any detectable PL. Among the long vertical NWs, 11 out of the 12 of those visible in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}c emit with essentially equal intensity. The PL spectra in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}e are representative for NW arrays with a pitch of \(200\)--\(400\) nm, and are characterized by a broad band exhibiting a pronounced finestructure at low excitation density and a continuous blueshift with increasing excitation density. These features are typical of GaAs/(Al,Ga)As NWs with a high degree of polytypism. The random stacking sequences in the NWs result in a band of spatially indirect transitions with an almost continuous energy variation. Note that the peak energy of this band does not reach the band edge of GaAs even for the highest excitation density. In contrast, the spectra of NW arrays with a pitch \(\geq 700\) nm (Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}f) are dominated by two well-defined lines. The spectral position (accounting for strain) and relative intensity as a function of excitation density \(I_L\) of these lines, as well as their uniform appearance from NW to NW and along the length of single NWs (see below), identifies them as originating from the transitions prevalent in high-quality epitaxial GaAs layers (see the spectra in the Supporting Information for comparison). The dominance of these two lines even at the lowest excitation densities is a strong indication that the NWs possess the zincblende structure with a very low density of stacking defects such as twin boundaries or thicker wurtzite segments with stacking faults. Specifically, the line at \(1.483\)--\(1.491\) eV that prevails for low excitation densities stems from transitions related to the unintentional incorporation of the shallow acceptor C into the NWs, namely, the donor-acceptor-pair \[(\(D^0\),C\(^0\))\] transition at the lowest and the conduction-band-to-acceptor \[(\(e\),C\(^0\))\] transition at higher excitation density. The individual contributions are not spectrally resolved since the width of the band encompasses both of the underlying transitions. However, the blueshift of the band with increasing \(I_L\) is partly due to the actual characteristic blueshift of the (\(D^0\),C\(^0\)) transition, and partly an apparent blueshift due to the (\(e\),C\(^0\)) transition overtaking the (\(D^0\),C\(^0\)) transition at higher excitation density (\(I_L > 1\) W cm\(^{-2}\)). In the following, we will therefore refer to this band as (\(e/D^0\),C\(^0\)). The weak signals at about \(1.450\) and \(1.414\) eV are due to the first and second longitudinal optical (LO) phonon replicas of the (\(e/D^0\),C\(^0\)) transitions. For \(I_L > 20\) W cm\(^{-2}\), the (\(e/D^0\),C\(^0\)) band saturates altogether and a line at \(1.511\)--\(1.514\) eV starts to dominate the spectra. The spectral position of this line identifies it to be of excitonic origin. The slight blueshift with increasing excitation density suggests a participation of acceptor-bound \[(C\(^0\),\(X_{\text{hh}}\))\] and free excitons (\(X_{\text{hh}}\)), although their individual contributions are not resolved in the spectra shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}f. Because of the uniaxial tensile strain exerted by the (Al,Ga)As shell (see Supporting Information), these excitons are of heavy-hole character. The full width at half maximum (FWHM) of the \(X_{\text{hh}}\) line in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}f amounts to about \(7\)--\(8\) meV. While being remarkably narrow compared to the values commonly observed for GaAs NWs with the exception of Ref. , it is still rather broad compared to the \(X_{\text{hh}}\) line of planar GaAs for which we observe an FWHM of \((1.6 \pm 0.4)\) meV (see Supporting Information). With a pitch of \(700\) nm and a laser spot size of \((2 \pm 0.4)\) m, the spectra collected from this NW array originate from about \(20\) individual NWs. To examine whether the broadening is simply an ensemble effect, we record spectra from single NWs in arrays with a pitch of \(5\) m or larger such as depicted in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}c. Figures [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}g,h show spectra of two of such NWs. The spectra of NW1 in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}g are representative for most of the NWs and are essentially indistinguishably from the ensemble spectra depicted in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}f. This result suggests a high NW-to-NW homogeneity due to the virtual absence of polytypism, which would result in spectra that drastically differ from NW to NW. However, in some NWs we do detect transitions associated with stacking defects in spectra recorded at low excitation densities, such as in those of NW2 shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}h (labeled with an asterisk). Still, even in this case, the spectra are dominated by the (\(e/D^0\),C\(^0\)) and \(X_{\text{hh}}\) lines that are not resolvable in spectra of NWs with a high degree of polytypism as those displayed in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}e. The inset of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}g shows a cathodoluminescence (CL) hyperspectral linescan of another single NW. Evidently, both the C-assisted and the excitonic emission intensities are uniform along the entire length of the NW, in striking contrast to linescans taken from NWs with a high degree of polytypism. For the present NWs, we were unable to detect any transitions associated to stacking defects, presumably because the excitation density in CL is high enough to easily saturate transitions from isolated stacking defects (cf. spectrum in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}h taken at \(140\) W cm\(^{-2}\)). For all single NWs, the FWHM of the \(X_{\text{hh}}\) line amounts to about \(5\)--\(6\) meV and is thus only insignificantly narrower than the corresponding transition in the ensemble spectra. This finding shows that the broadening of the \(X_{\text{hh}}\) line is not an ensemble effect. Possible mechanisms for the broadening of the transitions from single NWs will be discussed below. The results above are representative for two categories of microfields. From those with a pitch \(<700\) nm we invariably record PL spectra characteristic of pronounced polytypism. In contrast, for all fields with a pitch \(>700\) nm (up to \(10\) µm) we obtain spectra that resemble those of bulk GaAs except for the larger spectral width of the individual lines, which we will discuss later in Sec. in detail. We observe this behavior for three samples containing either GaAs/Al\(_{0.33}\)Ga\(_{0.67}\)As (sample A) or GaAs/Al\(_{0.10}\)Ga\(_{0.90}\)As (samples B and C) core/shell NW arrays. The very different degrees of polytypism are observed in both samples for microfields in immediate vicinity (\(<1\,\)mm) and are thus definitely caused by the different pitch of the fields. In fact, it has been reported that the preferred polytype depends on the NW density due to the shadowing of the Ga flux, with zincblende being favored for larger interwire separation. Shadowing is indeed evident from the short NWs in the field with \(200\) nm pitch (cf. Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}a). We also note that shadowing, in conjunction with the continuous rotation of the sample during growth, leads to a periodical change of the effective V/III ratio at the tip of the NW, which may induce an instability at the triple-phase boundary (between the liquid Ga droplet, the GaAs NW, and the vacuum) with respect to the nucleation of zincblende or wurtzite monolayers. In any case, the PL spectra of NW arrays with a pitch equal or larger than \(700\) nm are close to those expected for phase-pure zincblende GaAs, and thus offer the opportunity to study the intrinsic optical properties of GaAs NWs in the sense that polytypism does not significantly affect these properties. In the remainder of this work, we consequently focus on the analysis of the PL spectra of NW fields with this desirable property, specifically, fields with a pitch of \(700\) and a hole diameter of \(90\) nm from samples A and B (for brevity, we refer to these fields in the following as \(700/90\) NW arrays). In particular, we are interested in the internal quantum efficiency of the NWs in these fields.
## Excitation-density dependence
To get access to the internal quantum efficiency of the phase-pure NWs, we extend the range of excitation densities to which they are subjected. Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a shows \(10\) K cw-µ-PL spectra of the GaAs/Al\(_{0.33}\)Ga\(_{0.67}\)As core/shell NW array depicted in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}b for low to high excitation densities. As discussed before, the spectra are dominated by the (\(e/D^0\),C\(^0\)) band at low, and by the \(X_{\text{hh}}\) line at intermediate excitation densitites. Upon even higher excitation densities, the \(X_{\text{hh}}\) line abruptly blueshifts and broadens at its high-energy side. As discussed in detail below, this blueshift and broadening signifies a transition from excitonic to electron-hole-plasma (EHP) recombination accompanied by a progressive band filling. These changes in the spectra are summarized quantitatively in Figs. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}b--d depicting the spectral position \(E_{\text{hh}}\), the FWHM, and the integrated intensity \(I_{PL}\) of the \(X_{\text{hh}}\) (full symbols) and the (\(e/D^0\),C\(^0\)) (open symbols) transitions, respectively, versus the excitation density \(I_L\). These quantities were extracted by line-shape fits of spectra measured on various microfields on samples A (circles) and B (squares). Regardless of sample, hole diameter, and pitch, the data are seen to follow the same trends.
### Emission energy: Band bending, strain, and Mott transition
We next discuss the dependence of the \(X_{\text{hh}}\) transition energy on excitation density and the composition of the passivating (Al,Ga)As shell. The spectral position of the \(X_{\text{hh}}\) line shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}b is seen to be constant for values of \(I_L\) between \(0.03\) and \(300\,\)W cm\(^{-2}\). In this range of excitation densities, the \(X_{\text{hh}}\) line is redshifted with respect to the \(X_{\text{hh}}\) energy in unstrained GaAs (indicated by the horizontal line) by, on average, (\(1.2 \pm 0.3\)) and \((0.3 \pm 0.1)\) meV for samples A and B, respectively. Such a redshift with increasing Al content in the shell has been observed by other groups as well. attributed this shift to strain exerted on the GaAs core by the lattice-mismatched (Al,Ga)As shell, and substantiated this interpretation by calculations of the strain state. In contrast, disregarded strain as the origin based on an absence of a shift in Raman spectra. Instead, these authors attributed the redshift to spatially indirect transitions between electrons and holes induced by a radial band bending in the NW core due to interface states formed by unintentionally incorporated O. and also observed an additional red-shift of the \(X_{\text{hh}}\) transition energy for decreasing excitation densities. The former authors attributed this shift to the gradual screening of axial piezoelectric fields in conjunction with compositional fluctuations in the shell, while the latter ones invoked the filling of interfacial trap states and a corresponding modification of a band bending. Finally, observed an increasing redshift with increasing diameter, and assigned this finding to the presence of radial electric fields induced by residual acceptors with a concentration of \(10^{15}\)--\(10^{16}\) cm\(^{-3}\). In the case of a redshift induced by an electric field, recombination has to occur between spatially separated electron-hole pairs with a transition energy shifted due to a radial Stark effect. The radiative lifetime of exciton-polaritons in high-quality GaAs/(Al,Ga)As double heterostructures is on the order of a few ns at low temperatures. The exciton decay times for Ga-assisted GaAs/(Al,Ga)As NWs, regardless of whether being fabricated by MBE or MOVPE, are close to this value. An exciton in an electric field ionizes when the tunneling time of the electron in the Coulomb potential of the hole becomes significantly shorter than the recombination time. Full ionization occurs at a field corresponding to a potential drop of one Rydberg across the Bohr radius of the exciton. For GaAs, this field amounts to \(3\) kV/cm. For fields of this magnitude, recombination can only occur between electron-hole pairs spatially separated by the radial fields, thus being shifted to lower energies (analogously to the Franz-Keldysh effect in bulk crystals). In our case, the energy of the \(X_{\text{hh}}\) line does not depend on \(I_L\) for low and intermediate excitation densities, i. e., recombination remains excitonic down to the lowest excitation density of \(30\) mW cm\(^{-2}\) (which is well within the range of excitation densities used, for example, in Refs. ). Hence, the radial electric field in the GaAs core of our NWs is too small to dissociate excitons prior to their recombination. As shown in the Supporting information, this situation changes for NWs of larger diameter (and thus larger fields), in which recombination is no longer excitonic. To obtain quantitative information of the band bending, the shell-induced strain and the associated radial piezoelectric fields, we have performed two-dimensional simulations of our GaAs/(Al,Ga)As core/shell NWs using the Poisson solver of [nextnano]{.smallcaps} and the strain, piezopolarization, and eight-band \(\mathbf{k \cdot p}\) modules of [SPHInX]{.smallcaps}. The results are presented in the Supporting Information, with the main insights summarized as follows. First of all, radial electric fields are induced intrinsically by the lattice-mismatched shell, but are too small (\(\ll 1\) kV/cm) to dissociate excitons for the present Al content and thickness of the shell. Second, the radial electric fields caused by the residual background doping stay below the strength required for dissociating excitons up to a residual doping density of \(1 \times 10^{16}\) cm\(^{-3}\). Third, the \(\mathbf{k \cdot p}\) calculations show that the redshift with increasing Al content can be safely assigned to strain. In fact, and as also discussed in the Supporting Information, values close to the experimental ones are obtained already with a crude approximation, taking into account solely the uniaxial strain induced by the lattice mismatch between core and shell along the NW axis and the respective cross-sectional areas, while neglecting the different elastic properties of core and shell. The constant energy of the \(X_{\text{hh}}\) line observed for both samples A and B for \(I_L < 300\,\)W cm\(^{-2}\) is expected for free exciton transitions because of the charge neutrality of the exciton. For higher excitation density, the continuum redshifts (band-gap renormalization) and eventually crosses the energy of the exciton and thus induces a Mott transition from excitonic to EHP recombination. In our case, we observe the exact opposite behavior: at \(I_L \approx 300\,\)W cm\(^{-2}\), the transition energy abruptly blueshifts, and the \(X_{\text{hh}}\) line broadens primarily at its high-energy side (cf. Figs. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a--c). The high-energy broadening is indicative of a progressive band filling due to the quasi-Fermi level for electrons entering the conduction band, creating a degenerate electron gas. The blueshift of the high-energy half-maximum thus appears to be analogous to the dynamic Burstein-Moss shift observed in pump-and probe absorption spectroscopy, and has indeed the same origin, but is *not* equivalent to it. In the absence of disorder as induced, for example, by heavy doping, the optical transitions in both absorption and emission obey strict **k** conservation. In absorption, this requirement of vertical transitions directly results in the blueshift of the absorption onset known as the dynamic Burstein-Moss shift. In emission, however, photoexcited holes rapidly relax toward the top of the valence band, and recombination can in principle only occur at \(\mathbf{k} = 0\). Transitions at higher energies require the presence of unoccupied (acceptor-like) states whose spread in **k** space allows recombination with electrons at \(\mathbf{k} > 0\), and these states can be provided intentionally or unintentionally in GaAs by incorporating shallow acceptors such as Be. For our nominally undoped NW arrays, the unintentionally incorporated C acts as shallow acceptor, and the pronounced high-energy broadening observed in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a suggests a comparatively high C concentration. In fact, for epitaxial GaAs layers with an extremely low background doping (\(p<10^{13}\) cm\(^{-3}\)), we observe the theoretically expected redshift of the transition (band-gap renormalization) with increasing excitation density (see the Supporting Information for details). Note that the band-gap renormalization is also visible for the NWs by the redshifting low-energy half maximum of the EHP band evident in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a. For the bulk-like layers (see Supporting Information), the Mott transition appears to be continuous, with excitons and the EHP coexisting over a certain range of excitation densities, but this effect may simply arise from the spatial nonuniformity of the carrier density induced by the finite excitation depth and carrier diffusion. In striking contrast, for the core/shell NWs under investigation, the sudden disappearance of the \(X_{\text{hh}}\) line and the associated abrupt shift of the transition energy signify a discontinuous (at least in comparison to the bulk-like GaAs layers as discussed in the Supporting Information) Mott transition from excitonic to EHP recombination. In fact, the coupling of light to a NW array is modified such as that for sufficiently large interwire separations (\(>600\) nm), the NWs are excited uniformly along their entire length. The NW array under investigation with a pitch of \(700\) nm is thus a realization of an electronically bulk-like GaAs sample with nearly uniform excitation density, and is thus ideally suited for studying the Mott transition in bulk GaAs. The excitation density of \(300\,\)W cm\(^{-2}\) at which the Mott transition occurs in our GaAs/(Al,Ga)As core/shell NWs is comparable to that observed for the Mott transition in state-of-the-art GaAs/(Al,Ga)As double-heterostructures as shown in the Supporting Information. At the first glance, this seems to imply that the coupled exciton-free carrier lifetimes are comparable as well, but we need to take into account the fact that the absorptance of the NW array is generally higher than that of a planar layer. In later sections, we will estimate the actual carrier density corresponding to this excitation density, determine the internal quantum efficiency, and measure the actual exciton and EHP lifetimes.
### Emission linewidth: Random dopant fluctuations and band filling {#Sec:linewidth}
Next, we discuss the physical origins of the spectral width of the \(X_{\text{hh}}\) line. For \(I_L<2\) W cm\(^{-2}\), the FWHM of the \(X_{\text{hh}}\) line is constant and amounts regardless of the Al content to values of \(5\)--\(8\,\)meV (see Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}c). In the inset of Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}c, we show that measurements on three different single NWs (NW 1 and 2 defined in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} and an additional single NW from the same sample) yield values in the same range, indicating that the width of the transition is not governed by ensemble effects, as already indicated in the context of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. For comparable temperatures and excitation densities, the narrowest FWHM of GaAs/(Al,Ga)As \(3\)D nanostructure arrays were reported by for GaAs nanomembranes. These authors measured an FWHM of \(2\) (\(4\))\(\,\)meV for membranes passivated with Al\(_{0.10}\)Ga\(_{0.90}\)As (Al\(_{0.33}\)Ga\(_{0.67}\)As) shells. In contrast to our results, the FWHM thus exhibited a notable dependence on the Al content in the shell. This fact was not commented upon by the authors, but a natural explanation follows from their previous interpretation of the transition energies. Due to the band bending, a comparatively large FWHM could result from a superposition of transitions with different energies, and as the band bending is supposed to increase with Al content, so would the FWHM of the transitions. However, this interpretation is not consistent with the experimental data obtained for the NWs investigated in the present work. As discussed above, the dependence of the transition energies on Al content and excitation density we have displayed in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}b is only compatible with free exciton recombination shifted by strain exerted on the core by the shell, but not with spatially indirect transitions between radially separated electrons and holes as proposed in Refs. and . The same applies to the FWHM of this transition. If the FWHM were to arise from a superposition of spatially indirect transitions in a radial electric field, it would be expected to monotonically decrease with increasing excitation density due to screening. Furthermore, the FWHM should depend on NW diameter similarly to the transition energy as discussed in Ref. . However, Fig, [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}c shows that the FWHM *increases* monotonically with excitation density, contrary to the expected behavior for spatially separated electron-hole pairs in radial electric fields. Moreover, we do not observe any change in the FWHM of the \(X_{\text{hh}}\) line for NWs of different diameter, as documented in the Supporting Information. An alternative explanation of the comparatively large FWHM of the transitions in our NWs is based on the fact that the residual C acceptors in the NW core are ionized in equilibrium even at cryogenic temperatures due to charge transfer to the surface. Solving the Poisson equation for our NW geometry (see the Supporting Information for details) reveals that this effect also applies to the core/shell NWs under investigation: the acceptors are fully ionized, but for the core diameters of the present NWs, the core remains fully depleted. The charge fluctuations created by the randomly positioned acceptor ions induce potential fluctuations that lead to a broadening of near band-edge transitions. In the present case, these fluctuations are particularly large due to the absence of free carriers, resembling the case of a perfectly compensated semiconductor as considered in Ref. . The FWHM of \(5\)--\(8\) meV observed for our samples then corresponds to acceptor concentrations on the order of \(10^{15}\) cm\(^{-3}\), at the lower end of the values estimated in Ref. , and still consistent with our above estimate of the maximum concentration compatible with free exciton recombination. Above an excitation density of \(300\) W cm\(^{-2}\), the FWHM is seen to steeply increase from \(10\) to more than \(40\) meV. As discussed above, the low-energy broadening represents the redshifting continuum due to bandgap renormalization, while the high-energy broadening reflects the progressive filling of the conduction band and the recombination of electrons at \(\mathbf{k} > 0\) with neutral acceptors. Note that the presence of neutral acceptors at high excitation densities is not in contradiction with the above assignment of the broadening to ionized acceptors at low excitation densities. It is well known that impurities ionized by compensating defects (which are here introduced by surface states) can be neutralized upon photoexcitation by capture of the respective photogenerated carrier species. Obviously, the fraction of photoneutralized states depends on the ratio of the impurity and the photogenerated carrier density, with the majority of the impurities being in an ionized state at low excitation density, and in a neutral state at high excitation density. Correspondingly, the FWHM of the band-edge transition may be dominated by random dopant fluctuations in the fomer, and by band filling in the latter case.
### Emission intensity: competition of radiative and nonradiative transitions and internal quantum efficiency
Figure [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}d shows the dependence of the integrated PL intensity \(I_{PL}\) of the \(X_{\text{hh}}\) (full symbols) as well as the (\(e/D^0\),C\(^0\)) (open symbols) lines on excitation density \(I_{L}\). Clearly, \(I_{PL}\) does not increase linearly with \(I_L\) over the whole range of excitation densities for either of these lines, but exhibits a more complex dependence on \(I_{L}\). As we will discuss in the following, this dependence reflects a competition between three different recombination channels, the two radiative transitions giving rise to the \(X_{\text{hh}}\) and the (\(e/D^0\),C\(^0\)) lines, and nonradiative Shockley-Read-Hall (SRH) recombination. In excitation regime *(i)*, the PL spectrum is dominated by the lowest-energy transition, the (\(e/D^0\),C\(^0\)) line. The dependence of this line on \(I_L\) is slightly sublinear, while the intensity of the \(X_{\text{hh}}\) line increases slightly superlinearly. In excitation regime *(ii)*, the \(X_{\text{hh}}\) line progressively overtakes the (\(e/D^0\),C\(^0\)) line, which starts to saturate in intensity. In excitation regime *(iii)*, the \(X_{\text{hh}}\) exhibits a strong superlinear increase in intensity. This increase cannot arise from the competition with the (\(e/D^0\),C\(^0\)) line, since it persists also for the total intensity of both lines. Rather, this increase reflects the presence of a nonradiative channel, which is gradually overtaken by radiative recombination in this excitation regime, resulting in an increase of the internal quantum efficiency \(\eta_\text{int}\). Finally, in excitation regime *(iv)*, the spectra are dominated by the EHP band, the intensity of which enters a linear dependence on \(I_L\), reflecting that radiative recombination starts to dominate. Consequently, \(\eta_\text{int}\) approaches unity in this regime. The behavior sketched above applies to both samples A and B despite the different Al contents in their shells. This finding suggests that the nonradiative process is not related to surface recombination, since the barrier height for sample A is drastically higher than for sample B. We would, in principle, also expect a dependence of the interface recombination velocity on the Al content, but it is certainly possible that such a dependence only sets in at higher Al content. Alternatively, point defects in the NW core may act as nonradiative SRH centers independent of the Al content of the shell. In any case, an increase of \(\eta_\text{int}\) as observed in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}d can originate either from a speedup of radiative recombination, a slowdown of the nonradiative one, or both effects acting together. For quantitatively analyzing these data, we integrate over both emission lines to obtain the total radiative intensity \(I_r\) as well as the overall internal quantum efficiency \(\eta_\text{int}\) by dividing the total intensity by excitation density. Models suitable for fitting such data are invariably based on the quasi-equilibria established between the individual populations at a given temperature and excitation density (such as the Saha equation for excitons and free carriers) in the spirit of Refs. , thus reducing a system of coupled rate equations to a single one that fully accounts for the variation of the spectrally integrated PL intensity with excitation density. For the present case of transitions involving excitons, free carriers, a shallow impurity, and a nonradiative center, we arrive, when summing over all radiative contributions, at an equation that is formally identical to either the simple model developed in Refs. (hereafter referred to as model I) or, when taking into account the SRH kinetics of the nonradiative process, at the model discussed in Ref. (hereafter referred to as model II). In both cases, the contribution of excitons is effectively taken into account by an additional term to the radiative recombination coefficient derived by the principle of detailed balance. Likewise, the contribution of the C-assisted transition replaces the usual monomolecular term proportional to the background carrier density, which is zero in the present case of NWs being depleted due to charge transfer of carriers to surface states. Explicit expressions for models I and II are given in the Supporting Information. Within the frame of model I, the increase of \(\eta_\text{int}\) as observed in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}d originates from a speedup of radiative recombination due to a transition from mono-to bimolecular recombination. In the present case of predominantly excitonic recombination, such a transition naturally takes place for excitation densities close to the Mott transition, i. e., in regime *(iii)*. Indeed, when taking into account the decrease of the exciton binding energy \(E_x\) with increasing excitation density, the Saha equation \[\frac{n_x}{np} \propto \exp\frac{E_x(n)}{k_B T}\] actually predicts that the fraction of free carriers over that of excitons increases with excitation density. Here, \(n\), \(p\), \(n_x\), \(k_B\), and \(T\) denote, respectively, the electron, hole, and exciton concentration, the Boltzmann constant, and the temperature. Hence, recombination close to the Mott transition may take an increasingly bimolecular character even at low temperatures. Within the frame of model II, an additional mechanism for an increase of \(\eta_\text{int}\) emerges from the potential slowdown of nonradiative recombination that may occur for recombination via an SRH center that preferentially captures minority over majority carriers. In this case, the recombination rate may decrease by a factor up to \(\tau_p/\tau_n\) from small to large signal excitation, where \(\tau_n\) and \(\tau_p\) are the capture times for minority and majority carriers, respectively. In the present case, however, this mechanism is effectively blocked by the depletion of the NWs and the corresponding vanishing background carrier density, resulting in a SRH recombination rate close to the large-signal case regardless of excitation density. The two considerations above are based on the assumption that the radiative and nonradiative recombination coefficients are constant and do not depend on the carrier density. This assumption is violated for degenerate semiconductors, for which the recombination coefficients decrease with increasing density and eventually enter a \(1/n\) dependence, thus effectively reentering a monomolecular recombination regime at high excitation densities. Likewise, nonradiative recombination slows down in the degenerate regime. However, the radiative and nonradiative rates in the degenerate regime can no longer be represented in terms of population densities multiplied by rate constants or coefficients, with the consequence that the concept of detailed balance to account for the coupling of these populations cannot be applied. The resulting equation systems are only accessible numerically, and less suitable for fits to experimental data. For simplicity, we treat both the radiative and the nonradiative coefficients as constants, which means that we will overestimate them in the degenerate regime. Since both of them decrease with increasing density in the degenerate regime, their ratio (determining \(\eta_\text{int}\) as well as \(I_{PL}\)) is expected to stay approximately constant, and we will see below that this is indeed the case when comparing these values with the actual lifetimes. Figures [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}a,b show the dependence of the overall \(\eta_\text{int}\) (as obtained by the spectrally integrated PL intensity divided by excitation density, and scaled to provide the best fit to the experimental data) and the total PL intensity on the generation rate \(G\), respectively. The sketch in the inset of Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}a illustrates the transitions relevant in the context of models I and II. To actually compare the experiment to simulations, both the \(y\) and \(x\) axes are rescaled to the units used in the simulation. For the carrier generation rate \(G = I_{L}\,\alpha\,\hbar^{-1}\,\omega^{-1}\), we arrive at \(G=10^{24}\,I_L\) by adopting an effective absorption coefficient of \(2.5 \times 10^{5}\,\)cm\(^{-1}\) as computed by for GaAs NWs with a diameter equal to the one of the NWs under investigation in the present work. Note that this value is about \(10\,\times\) higher than the one of bulk GaAs. The PL intensity is measured in arbitrary units, and our model thus includes a scaling factor to convert the PL efficiency (the PL intensity divided by the excitation density) to the internal quantum efficiency provided by the models. The best agreement is obtained when the point taken at highest excitation density is scaled to a value of \(\eta_\text{int} = 0.65 \pm 0.1\). The attraction of model I lies in its simplicity and the fact that it only has two free parameters, namely, the small-signal monomolecular radiative and nonradiative lifetimes \(\tau_r^s\) and \(\tau_{nr}\), respectively. In the present case, the former is given by the lifetime of the C-assisted transition \((b_A N_A)^{-1}\) with the recombination coefficient \(b_A\) and the concentration of C acceptors \(N_A\). The latter is the lifetime obtained from the full SRH expression for \(p_0 \approx 0\) and \(\Delta n \approx \Delta p\), i. e., \(\tau_n + \tau_p\). Thanks to having only two free parameters, we obtain a unique fit as shown in Figs. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}a,b returning values of \(\tau_r^s = (19 \pm 2)\) ns and \(\tau_{nr} = (33\pm 5)\) ps. Model II has three free parameters, with which we found it difficult to achieve convergence and to obtain a unique parameter set. To facilitate convergence, we run the fit with the radiative lifetime fixed to certain preset values. The best fit is obtained with essentially identical parameters compared to model I: \(\tau_r^s = 19\) ns, \(\tau_{n} = (12.5\pm 1)\) ps, and \(\tau_{p} = (20\pm 2)\) ps. The fit also shows that \(\Delta n \approx \Delta p\) except for the lowest generation rates. Hence, the total nonradiative lifetime varies only slightly from \(37\) to \(33\) ps from low to high excitation density, and the fit is essentially indistinguishable from the one of model I, which is based on the assumption \(\Delta n = \Delta p\) from the outset. Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}c shows the dependence of \(\Delta n\) on the generation rate as derived from the best fit of model II to the experimental data. As expected, \(\Delta n \propto G\) for small and \(\Delta n \propto \sqrt{G}\) for large signal excitation. According to the fit, the excitation density of \(300\,\)W cm\(^{-2}\), at which we observe the Mott transition to occur corresponds to a carrier density of \(9 \times 10^{15}\) cm\(^{-3}\). This value is close to the theoretically expected one as well as the few available values for bulk GaAs. In the only previous study on GaAs NWs, the Mott transition was reported to be gradual and to take place between \(2-4 \times 10^{16}\) cm\(^{-3}\). Note that the nature of the Mott transition (i. e., continuous vs discontinuous) is not clear even for quantum wells with their inherently uniform excitation. It is worth noting that degeneracy of the conduction band of GaAs sets in at \((10 \pm 3)\times10^{15}\) cm\(^{-3}\) at a temperature of \((25 \pm 5)\) K (see Supporting Information), i. e., band filling occurs indeed in parallel to the Mott transition.
### External quantum efficiency
The maximum internal quantum efficiency derived in the previous section is \(0.65\). Within the framework of the two models employed, we estimate the margin of error in this value to be not overly large, certainly less than \(\pm 20\)%. However, the models themselves are based on many assumptions and simplifications, and it is thus prudent to ask whether the order of magnitude of \(\eta_\text{int}\) is reasonable or not. To answer this question, we adopt the method used by to determine an absolute value of the external quantum efficiency \(\eta_\text{ext}\) from the sample under investigation. Specifically, we perform a side-by-side comparison of the PL intensity of the \(700/90\) NW array from sample C with the laser intensity reflected by a [Zenith Polymer]{.smallcaps} diffuse reflector, possessing a nearly ideal Lambertian reflectance characteristics. The \(\eta_\text{ext}\) of our NW array is then obtained by normalizing the spectrally integrated intensity of the PL from the NW array to that of the exciting laser, after properly correcting for the known reflectance of the [Zenith Polymer]{.smallcaps} and the system response to account for the different wavelengths of the PL signal and the laser. Note that a confocal setup, as used for all previous experiments, does not easily allow such a comparison, for which we have conducted these experiments in a separate non-confocal setup as described in the Methods section. The diamonds in Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}b represent, by definition, an internal and external quantum efficiency of unity, as they are equal to the measured intensities on the [Zenith Polymer]{.smallcaps} reflector, corrected for both the system response and the reflectivity of \(84\%\) for the [Zenith Polymer]{.smallcaps} thickness of \(500\) m used for this experiment. The squares display the ratio of the spectrally integrated PL intensity of our sample and the laser intensity measured side-by-side, i. e., the external quantum efficiency of this sample. Clearly, these data follow the same trend as those acquired with the confocal setup, but their absolute value is lower because the extraction efficiency is not unity even for deep-subwavelength NWs. In the present case, the extraction efficiency as given by the ratio of these two different measurements amounts to \(0.5 \pm 0.15\) on average. This value is in good agreement with theoretical work considering the effective diameter/wavelength ratio of \(\omega d/c \approx 1.6\) for the present NWs.
## Time-resolved PL experiments
The comparison between the internal and external quantum efficiencies, derived by completely different means, shows that the ratio of the lifetimes returned by the fit to our experimental data (see Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}a,b) is quite accurate. However, in deriving the model, we have made several assumptions and performed many simplifications. Hence, the absolute values of these lifetimes are most likely not correct. To obtain these absolute values, and to confirm the interpretation of the cw-PL experiments, we investigate the \(700/90\) NW array from sample A by time-resolved PL (TRPL) spectroscopy. Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}a shows a representative streak camera image taken at \(10\,\)K and a fluence of \(400\,\)pJ cm\(^{-2}\). At this intermediate fluence, the \(X_{\text{hh}}\) and the (\(e\),C\(^0\)) lines are having comparable intensities. While the \(X_{\text{hh}}\) line is seen to decay relatively fast, the intensity of the (\(e\),C\(^0\)) line stays almost constant within the \(2\) ns time windows, reflecting the long lifetime typical for impurity-assisted transitions. Assuming an exponential decay as expected for the (\(e\),C\(^0\)) transition, we obtain a decay time of \(60\) ns, corresponding (for a lattice temperature of \(25\) K) to an acceptor concentration of at most \(1 \times 10^{16}\) cm\(^{-3}\) if this time corresponds to the actual radiative lifetime of this transition. Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}b depicts PL spectra extracted from streak camera images recorded in focus mode at \(10\) K and with different fluences as indicated in the figure. The evolution of the spectra is identical to the one observed in the cw-PL spectra shown in Figs. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}f and [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a: at the lowest fluences, the spectra are dominated by the (\(e/D^0\),C\(^0\)) line, which saturates at intermediate fluences, with the \(X_{\text{hh}}\) line overtaking. At high fluences, the Mott transition takes place and band filling sets in, resulting in the blueshifted and asymmetrically broadened EHP emission. Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}c depicts spectrally integrated PL intensity transients of the \(X_{\text{hh}}\) line at \(10\,\)K and different fluences as indicated in the figure. With increasing fluence, the effective lifetime \(\tau_{\text{eff}}\) of the \(X_{\text{hh}}\) line is observed to increase by one order of magnitude from about \(0.11\) to \(1\) ns, reflecting the increase in \(\eta_{\text{int}}(G) = \tau_{\text{eff}}(G)/\tau_r(G)\) with increasing generation rate extracted from the cw-PL measurements. The value of \(1\,\)ns is reached for fluences \(\geq 9\) nJ cm\(^{-2}\), i. e., still well in the excitonic regime. This value is close to those reported for GaAs/(Al,Ga)As NWs previously. The smooth transition from the excitonic to the degenerate EHP regime evidences the strong coupling of excitons and free carriers already at \(10\) K, resulting in a common lifetime for these states. Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}d displays the temperature dependence of spectrally integrated PL intensity transients measured at a fluence of \(1.6\,\)µJ cm\(^{-2}\). At this fluence, the instantaneous carrier density after the pulse amounts to \(1.5\times10^{17}\,\)cm\(^{-3}\), corresponding to the highest carrier density created in the cw-PL experiments (cf. Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}c) and thus well above both the Mott transition and the crossover from a nondegenerate to a degenerate electron gas. The emission monitored thus stems from a degenerate EHP. Note that the transients are not scaled to allow a direct comparison of their intensities. For temperatures between \(10\) and \(50\) K, the transients hardly change in both peak intensity and decay time, and thus integrated intensity. For higher temperatures, all of these quantities start to decrease. The inset of Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}d displays the results of a quantitative analysis of the PL transients as function of temperature. Depicted are the effective lifetime \(\tau_{\text{eff}}\) obtained from exponential fits to the transients, the large-signal radiative lifetime \(\tau_r^\ell\) given (in arbitrary units) by the inverse of the transients maximum intensity, and the integrated intensity \(I_{PL}\). Clearly, for temperatures between \(10\) and \(50\) K, all of these quantities are essentially constant, implying that \(\tau_r^\ell \approx \tau_\text{eff}\) corresponding to an internal quantum efficiency \(\eta_\text{int}\) close to unity. Note that \(\tau_r^\ell\) is expected to be independent of temperature in a degenerate semiconductor, and to enter a \(T^{3/2}\) dependence in the nongenerate regime For the present conditions, the crossover between these regimes occurs at around \(80\) K. Finally, we can compare the small-signal lifetimes obtained from the fit of our simple recombination models to the cw-PL data (Figs. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}a,b) and the decay times obtained by TRPL. As already stated above, the fit to the data returned values of \(\tau_r^s = (19 \pm 2)\) ns and \(\tau_{nr} = (33\pm 5)\) ps. The TRPL transients of the (\(e/D^0\),C\(^0\)) and \(X_{\text{hh}}\) lines taken at low fluences yield effective lifetimes of \(60\) ns and \(110\) ps. The former of these values is likely to correspond to the actual radiative lifetime of this transition, while the latter should be close to the actual nonradiative lifetime. The absolute value of these lifetimes is about a factor of three larger than those extracted from the fit, but their ratio---which determines the internal quantum efficiency---is essentially the same. In view of the assumptions and simplifications underlying the model, this result is very satisfactory.
## Thermal quenching of the PL intensity
In the following, we analyze the temperature dependence of the cw-µ-PL data with a main focus on the effects of the Al content in the shell. Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}a depicts cw-µ-PL spectra of the \(700/90\) nm NW array from sample A recorded at temperatures between \(10\) and \(300\) K at an excitation density \(I_L = 10\) W cm\(^{-2}\), i. e, in the monomolecular regime (excitation regime *(ii)* in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}d). For this set of measurements, we resolve the acceptor-bound exciton (\(A^0\),\(X_{\text{hh}}\)) in addition to the (\(e/D^0\),C\(^0\)) and \(X_{\text{hh}}\) transitions at \(10\) K. The evolution of the spectra with temperature is identical to that observed for bulk GaAs. The individual transitions quench according to their binding energy, and at temperatures \(>200\) K, only the band-to-band transition participates in radiative recombination. Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}b displays a double-logarithmic Arrhenius representation of \(\eta_\text{int}(T)\) measured for the \(700/90\) nm microfields of samples A and B. The data from both samples are fit by the familiar expression for a thermally activated quenching of the PL intensity or internal quantum efficiency \[\eta_\text{int}(T) = \left[1+\sum_i a_i T^{3/2} \exp \left(-E_i/k_B T\right)\right]^{-1} \label{eq:arrhenius}\] with the activation energies \(E_i\) of the participating nonradiative channels, the prefactors \(a_i\), the temperature \(T\), and the Boltzmann constant \(k_B\). The increase of the radiative lifetime with temperature in the bulk-like GaAs core is explicitly taken into account by the factor \(T^{3/2}\). The data obtained from sample A are fit well with a single activation energy amounting to \((6.6 \pm 0.5)\) meV. This value is close to, but considering the error bar clearly larger than the exciton binding energy \(E_X = 4.2\) meV. In fact, the value represents an effective activation energy that subsumes over both the progressive dissociation of free excitons into free carriers and the thermal ionization of neutral C acceptors (with an energy of \(E_X = 26\) meV), resulting in the subsequent quenching of the (\(e/D^0\),C\(^0\)) line. For a temperature up to \(160\) K, sample B exhibits the same behavior as sample A governed by an essentially identical activation energy. For temperatures higher than \(160\) K, however, \(\eta_\text{int}\) of sample B declines sharply with a second activation energy of \((140 \pm 10)\,\)meV. This value is very close to the band gap energy difference of Al\(_{0.10}\)Ga\(_{0.90}\)As and GaAs, indicating that this second nonradiative channel is related to the bipolar thermionic emission of carriers above the bandgap of the Al\(_{0.10}\)Ga\(_{0.90}\)As shell. recently arrived at the same conclusion for the thermal quenching of the PL intensity from GaAs/(In,Ga)As core/shell NWs. This additional process leads to a severe drop of \(\eta_\text{int}\) at elevated temperatures. Extrapolating the fits to \(300\) K, we obtain values of \(2 \times 10^{-5}\) and \(6 \times 10^{-7}\) for samples A and B, respectively. This result underlines the critical role of the shell for carrier recombination as already reported in previous studies.
# Summary and Conclusions {#sec:summary}
In summary, we have presented a detailed spectroscopic investigation of ordered GaAs/(Al,Ga)As NW arrays on Si\((111)\). For a NW pitch equal or larger than \(700\) nm, we have found these arrays to exhibit a high phase purity and NW-to-NW uniformity, allowing us to quantitatively analyze systematic power-dependent, temperature-and time-dependent photoluminescence measurements. We have extracted several quantities of interest, such as the background acceptor concentration (\(<10^{16}\,\)cm\(^{-3}\)), the Mott density (\(9\times 10^{15}\,\)cm\(^{-3}\)), the extraction efficiency (\(0.5\)), and the internal quantum efficiency as a function of both carrier density and temperature (see below). To conclude, we put the values determined for the internal quantum efficiency at room temperature in broader context, and discuss the implications of our findings for applications of GaAs/(Al,Ga)As core/shell NWs in optoelectronics. At the first glance, the internal quantum efficiency of \(2 \times 10^{-5}\) obtained for small-signal excitation at \(300\) K seems disappointingly low. However, our measurements truly represent a worst case scenario for (intentionally undoped) GaAs/(Al,Ga)As NWs. First of all, the NWs are fully depleted due to charge transfer to surface states (see the Supporting Information), i. e., the radiative rate at \(300\) K, where neither impurity-assisted nor excitonic transitions are active anymore, scales directly with the photogenerated carrier density, and has no monomolecular contribution. For the present excitation density of \(I_L=10\) W cm\(^{-2}\), we have estimated an excess carrier density of \(\Delta n \approx 3 \times 10^{14}\) cm\(^{-3}\) at \(10\) K (see Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}c), but this value will (due to thermally activated nonradiative channels) be certainly lower at \(300\) K. Consequently, the radiative lifetime \(1/(B \Delta n)\) (with \(B \approx 2 \times 10^{-10}\) cm\(^3\)/s at \(300\) K) will be \(\gg 10\) s. Second, even in the total absence of nonradiative recombination in the bulk, the interface recombination velocity at GaAs/(Al,Ga)As interfaces is finite. For a NW diameter of \(100\) nm, the data compilation in Fig. 2 of Ref. shows that even with the interface recombination velocity of state-of-the-art GaAs/(Al,Ga)As\((001)\) double heterostructures, the resulting maximum lifetime would be on the order of \(5\) ns. Hence, even under the most optimistic assumptions, the internal quantum efficiency of GaAs/(Al,Ga)As NWs at this low excitation density is not expected to exceed \(5 \times 10^{-4}\). However, because of the purely bimolecular free-carrier recombination at room temperature, the radiative efficiency of GaAs NWs monotonically improves at least linearly with increasing excitation density. In fact, our measurements performed for large-signal excitation suggest an (extrapolated) internal quantum efficiency of \(0.1\) at \(300\) K. Still, the evolution of the effective lifetime with temperature clearly shows that \(\tau_{\text{nr}}\) is eventually approaching values close to \(\tau_{\text{r}}\) at higher temperatures, and thus values lower than those obtained for the corresponding interface for \((001)\)-oriented planar structures. In terms of applications, the GaAs/(Al,Ga)As NWs under investigation would obviously not be eligible for solar energy harvesting, taking place at the lower end of excitation densities used in the present work. However, they do seem to hold promise for applications such as light emitting or laser diodes, corresponding to the higher end of excitation densities. What is particularly attractive for light emitting diodes is the high extraction efficiency of the as-grown NW array without requiring any kind of postprocessing. Finally, for further improving the suitability of GaAs/(Al,Ga)As core/shell NWs for applications as light emitting devices, and to make them eligible for use as solar energy harvesters, it is crucial to identify the origin of the nonradiative recombination channel in GaAs/(Al,Ga)As core/shell NWs. In the present work, we have found that the internal quantum efficiency at low temperature does not depend on the Al content in the shell, contrary to what would be intuitively expected for interface recombination. However, further dedicated studies are required to decide whether the defect responsible is located at the interface or in the core, and to develop growth protocols to minimize the density of this defect.
# Methods {#sec:Exp}
## Synthesis
The samples were fabricated using an [Oxford/VG V\(80\)]{.smallcaps} molecular beam epitaxy system equipped with solid sources for Ga, Al, and As. The NW growth took place on \(p\)-doped Si\((111)\) substrates covered by a patterned thermal oxide. Ordered arrays of GaAs NWs were selectively grown within \(100 \times 100\) µm\(^2\) microfields, where holes of different sizes and distances (pitch) were previously patterned in the oxide layer by electron beam lithography. The substrate preparation prior to growth was carried out according to Ref. . Sample A was realized by first growing \(4\)-µm-long and \(100\)-nm-thick GaAs NWs utilizing the self-assisted vapor-liquid-solid mechanism and an As/Ga flux ratio of \(3.9\) (see Ref. for details on the core growth). The NWs were then passivated with \(30\)-nm-thick Al\(_{0.33}\)Ga\(_{0.67}\)As shells grown at \(500\,^{\circ}\)C and an As/III flux ratio of \(15\). Sample B and C were grown under the same conditions as sample A, but with a shell consisting of \(30\)-nm-thick Al\(_{0.10}\)Ga\(_{0.90}\)As. The NWs are unintentionally \(p\)-type doped with C as the predominant acceptor.
## Microscopy
The secondary electron micrographs shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} were acquired in a [Hitachi S-\(4800\)]{.smallcaps} field-emission scanning electron microscope using an acceleration voltage of \(5\) kV and magnifications of \(3500 \times\) (single NWs) and \(20000 \times\) (NW arrays).
## Spectroscopy
The NWs were investigated with three different µ-PL setups. Independently of the setup, the samples were mounted onto the coldfinger of a liquid He µ-PL cryostat allowing continuous control of the sample temperature between \(10\) (nominally, see Supporting Information) and \(300\,\)K, and the excitation densities were controlled over six orders of magnitude by neutral density filters placed in the excitation path. The carrier temperature \(T_C\) was determined by fitting the high-energy slope of PL spectra of both our planar GaAs reference samples and GaAs NWs by a single exponential. For low excitation density, the value of \(T_C\) was found to saturate, yielding the actual lattice temperature of \((25 \pm 5)\) K. The offset between the nominal and the actual temperature is due to the ambient \(300\) K radiation impinging onto the sample, which is located only a few mm beneath the cryostat window. The cw-PL data displayed in Figs. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}d--h, [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} (the circles), and [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}, were recorded with a [Horiba-Jobin Yvon Labram HR Evolution]{.smallcaps} Raman microscope. This setup features a confocal configuration (with a confocal hole size set to \(100\,\)µm for all measurements presented here) and a micro-positioning system for performing \(xy\) maps as shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}d. The samples were excited using a [Melles-Griot]{.smallcaps} \(25\) [LHP]{.smallcaps} \(925\) He-Ne Laser (\(632.8\,\)nm) focused onto the samples with an [Olympus]{.smallcaps} [MS Plan]{.smallcaps} \(50\times\) objective having a numerical aperture of \(0.55\). The spot size of the laser (or more accurately, the FWHM of the Gaussian beam) was directly measured by linescans of the PL intensity across single NWs as depicted in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}d and amounts to \((2 \pm 0.4)\) m. The FWHM of the detection spot (as defined by the objective's magnification and the confocal hole diameter) was the same, ensuring a uniform excitation density. The PL signal was collected by the same objective, passed though a notch filter suppressing laser stray light, and entered a spectrograph with a focal length of \(80\,\)cm equipped with \(600\,\)lines/mm grating and an LN\(2\)-cooled [Horiba Symphony II]{.smallcaps} CCD. The cw-PL data represented by the squares and diamonds in Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} were acquired with a custom-build -PL setup in a non-confocal configuration. The samples were again excited by a [Melles Griot]{.smallcaps} \(25\) [LHR]{.smallcaps} \(925\) He-Ne laser (\(632.8\,\)nm) focused onto the samples with an [Olympus Mplan]{.smallcaps} \(\times10\) microscope objective. The PL signal was collected by the same microscope objective and imaged onto the entrance slit of a \(75\,\)cm monochromator, passing an edge filter suppressing laser stray light. The signal was then dispersed with a \(750\,\)lines/mm grating and detected with an LN2-cooled [Acton]{.smallcaps} (In,Ga)As array. For determining the external quantum efficiency, a \(500\) m thick piece of [Zenith Polymer]{.smallcaps} was mounted in the cryostat next to the sample. The diffusely reflected laser intensity was measured for three different laser power levels with the edge filter removed from the detection path, and a wide open entrance slit. The PL intensity was measured subsequently under exactly the same conditions for seven different laser power levels. Due to the wide open slit, the detection area covers almost the entire NW microfield, and the PL spectra therefore contain a spurious (\(e/D^0\),C\(^0\)) line regardless of excitation density. For determining the PL intensity, we thus considered only the \(X_{\text{hh}}\) line. In addition, the measured intensities were corrected for the spectral dependence of the system response function which was determined using a calibrated halogen lamp. The hyperspectral cathodoluminescence linescan in the inset of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}g was recorded at a beam current of \(0.2\,\)nA by a [Gatan MonoCL\(4\)]{.smallcaps} system mounted to a [Zeiss Ultra\(55\)]{.smallcaps} analytical field-emission scanning electron microscope. An acceleration voltage of \(5\) kV was used for excitation, and the luminescence was dispersed by a \(30\) cm spectrograph equipped with a \(600\) lines/mm grating and an LN\(2\)-cooled [Princeton Instruments]{.smallcaps} Si CCD. The TRPL data shown in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} were acquired using a custom-build setup consisting of a [Coherent Mira]{.smallcaps} Ti:Sapphire femtosecond oscillator optically pumped by a [Coherent Verdi V\(10\)]{.smallcaps} Nd:YAG laser and a [Hamamatsu C\(5680\)]{.smallcaps} streak camera. The ultra-short laser pulses with a duration of \(200\,\)fs and a wavelength of \(706\,\)nm used for exciting the samples were focused onto the sample by a 75-mm-lens. The PL signal was collected by the same lens, dispersed by a [Jobin Yvon Spex]{.smallcaps} \(1681\) \(22\,\)cm spectrometer and detected by the streak camera.
## Investigated samples
Table [\[SI-table1\]](#SI-table1){reference-type="ref" reference="SI-table1"} gives an overview of all samples investigated in our study. For the continuous wave-photoluminescence (cw-PL) measurements, samples A, B, C, R\(1\) and R\(2\) were investigated in a confocal setup and samples C and S in a home-build non-confocal setup. Time-resolved PL measurements were performed only on sample A. More details can be found in the Methods section of the main text.
### Comparison of optical transitions, carrier and lattice temperature {#sec:comp}
Figure [\[SI-figure1\]](#SI-figure1){reference-type="ref" reference="SI-figure1"}a depicts µ-cw-PL spectra of nanowire (NW) ensembles from sample A recorded at a nominal temperature of \(10\,\)K and different excitation densities. Figure [\[SI-figure1\]](#SI-figure1){reference-type="ref" reference="SI-figure1"}b shows µ-cw-PL spectra of the reference sample R\(1\) taken at comparable excitation densities. Each excitation density corresponds to a specific color of the spectrum in Figs. [\[SI-figure1\]](#SI-figure1){reference-type="ref" reference="SI-figure1"}a,b. The lines observed for sample R\(1\) are considerably narrower than the ones recorded from sample A, and we can resolve several individual transitions as shown in the inset of Fig. [\[SI-figure1\]](#SI-figure1){reference-type="ref" reference="SI-figure1"}b. Table [1](#SI-table2){reference-type="ref" reference="SI-table2"} summarizes the energy and origin of these transitions. The inset also shows the exponential high-energy tail of band-to-band recombination, from which we can deduce the carrier temperature \(T_C\). For this and similar samples and for low excitation densities, \(T_C\) was found to saturate at about \((30 \pm 10)\) K, which we thus identify as the actual lattice temperature at a nominal temperature of \(10\) K. This temperature is also consistent with the comparatively high intensity of the band-to-band transition. A value of \((25\pm5)\,\)K has been determined for sample A.
### Mott transition in planar reference samples
Figure [\[SI-figure1\]](#SI-figure1){reference-type="ref" reference="SI-figure1"}c depicts µ-cw-PL spectra from sample R\(1\) recorded in a spectral range close to the free exciton energy for a wider range of excitation densities. The spectra are observed to gradually redshift above a certain excitation density. This redshift is expected to occur after the Mott transition from excitonic to electron-hole-plasma recombination due to bandgap renormalization. A high-energy broadening indicating band-filling is not observed even for the highest excitation densities, as expected for an ideal semiconductor with a negligible density of acceptor states.. However, the excitonic transitions persist even for the highest excitation densities, which is the inevitable consequence of the nonuniform excitation of bulk-like layers.
[\[SI-table2\]]{#SI-table2 label="SI-table2"}
The spectral position and the width of the free exciton transition is obtained by line-shape fits integrating, for simplicity, over both lines *iv* and *v*. Figure [\[SI-figure2\]](#SI-figure2){reference-type="ref" reference="SI-figure2"} shows (a) the peak position and (b) the full-width at half-maximum (FWHM) of the \(X_{\text{hh}}\) transition of samples R\(1\) (red symbols) and R\(2\) (violet symbols) over the excitation density \(I_L\). For both samples and for \(I_L\) lower than \(100\,\)W cm\(^{-2}\), the peak position is essentially constant, as expected for free exciton transitions, and close to the X\(_{\text{hh}}\) energy in bulk GaAs (indicated by the horizontal line) with the slight redshift being due to the inclusion of the donor-bound exciton in the fit. At excitation densities higher than \(100\,\)W cm\(^{-2}\), the \(X_{\text{hh}}\) position is seen to linearly redshift with increasing \(I_L\). The FWHM of the \(X_{\text{hh}}\) transition is obtained in the same way and amounts to \((1.6 \pm 0.4)\) meV for excitation densities below \(100\,\)W cm\(^{-2}\), and monotonically increases with increasing \(I_L\) for excitation densities larger than \(720\,\)W cm\(^{-2}\). As a consequence of these two observations, we place the Mott transition between \(100\) and \(720\,\)W cm\(^{-2}\). Contrary to the sharp Mott transition observed for the NWs (see discussions concerning Fig. 2 in the main text), the one observed here is more gradual. This gradual character of the Mott transition is primarily related to the fact that the excitation in the bulk is, in general, nonuniform.
## Reproducibility of samples and measurements
Figures [\[SI-figure3\]](#SI-figure3){reference-type="ref" reference="SI-figure3"}a--c depict µ-cw-PL spectra of NW ensembles with a hole size of \(90\,\)nm and a pitch of \(700\,\)nm from sample A, B, and C, respectively, recorded at different excitation densities in the confocal setup. These spectra show a very low degree of polytypism, and are comparable to each other, demonstrating a high degree of reproducibility of the samples. Figure [\[SI-figure3\]](#SI-figure3){reference-type="ref" reference="SI-figure3"}d depicts the dependence of the spectrally integrated emission intensity \(I_\text{PL}\) from samples A, B, C, R\(1\) and R\(2\) on excitation density \(I_L\). The samples were measured side-by-side in the confocal PL setup, and samples A, B, and C are found to exhibit very similar absolute values of \(I_\text{PL}\), and a very similar dependence of \(I_\text{PL}\) on \(I_L\). Clearly, our growth protocol is suitable for the reproducible fabrication of phase-pure GaAs/(Al,Ga)As core/shell NWs. Analogously, samples R\(1\) and R\(2\) are comparable as well. The PL intensity of these samples is about two orders of magnitude lower than the PL intensity emitted by the NW samples, which is due to the unpassivated surfaces of the layers, as well as to the enhanced light absorption and extraction of NW arrays.
## Simulations {#sec:simulations}
### Strain components and piezoelectric potential {#sec:strain}
In the following, we present an overview of the results obtained from two-dimensional (\(2\)D) simulations with the \(8\)-band \(\mathbf{k \cdot p}\) formalism as implemented within the [SPHInX]{.smallcaps} suite with parameters taken from Ref. . Figure [\[SI-figure4\]](#SI-figure4){reference-type="ref" reference="SI-figure4"} depicts the strain components of an (a) GaAs/Al\(_{0.1}\)Ga\(_{0.9}\)As and (b) GaAs/Al\(_{0.3}\)Ga\(_{0.7}\)As core/shell NW heterostructure. The results are consistent with previous simulations of zincblende core/shell NW heterostructures. In particular, the shear strain components \(\epsilon_{xz}\) and \(\epsilon_{yz}\) are found to be essentially negligible, while the shear strain component \(\epsilon_{xy}\) is on the same order of magnitude as the diagonal strain components. The strain along the \(z\) axis is uniform in shell and core, and the sum \((\epsilon_{xx}+\epsilon_{yy})/2\) exhibits a sixfold symmetry as reported previously. Finally, the symmetry of each strain component does not depend on the Al content, but the magnitude changes linearly. Figure [\[SI-figure6\]](#SI-figure6){reference-type="ref" reference="SI-figure6"} depicts the calculated piezoelectric potential in the plane perpendicular to the NW axis of a (a) GaAs/Al\(_{0.1}\)Ga\(_{0.9}\)As and a (b) GaAs/Al\(_{0.3}\)Ga\(_{0.7}\)As core/shell NW heterostructure. For both structures, the calculated potential profiles have periodic maxima and minima with a threefold symmetry, in exact agreement with previous calculations for strained zincblende core/shell heterostructures. While the maximum field induced by this potential remains well below \(1\) kV/cm for the Al content used in the present work, these radial piezoelectric fields must not be neglected in general. Since the GaAs core is under tensile strain, the heavy-hole valence band resides at higher energy than the light-hole one. Hence, the experimental transition energies have to be compared to the calculated electron-to-heavy-hole transition energies. Figure [\[SI-figure8\]](#SI-figure8){reference-type="ref" reference="SI-figure8"} depicts the measured energy shift with respect to the bulk exciton energy (blue triangles) and the calculated energy shift \(\Delta E (x) = E_1 (x)-E_0 (x)\) (red squares) with respect to the bulk band gap versus the Al content. Here, \(E_1\) is the single-particle transition energy calculated with [SPHInX]{.smallcaps} by taking into account the contribution from strain and the piezopotential, and \(E_0\) is the bulk band gap. The experimentally obtained values for \(\Delta E\) were obtained from the averaged position of the \(X_{\text{hh}}\) line measured at excitation densities smaller than \(300\,\)W cm\(^{-2}\) on different microfields of samples A and B. The calculated energy shift is notably larger than the experimental one. Since the calculated shift includes contributions due to both strain and the piezoelectric potential, we estimate the former using a simple analytical approach. Assuming that both core (\(C\)) and shell (\(S\)) share a coherent interface and thus an average (strained) lattice constant \(\bar{a}\) in \(z\) direction, and neglecting the difference in elastic constants between core and shell, we have: \[\bar{a} = \frac{A_{C} a_{C} + A_{S} a_{S}}{ A_{C} + A_{S}}.\] Here, \(a_{C}\) (\(a_{S}\)) is the unstrained lattice constant for the core (shell), and \(A_{C}\) (\(A_{S}\)) is the cross-sectional area of the core (shell). The uniaxial strain components in the NW core \(\epsilon_{zz,C}\) and in the shell \(\epsilon_{zz,S}\) read: \[\epsilon_{zz,C} = \frac{\bar{a}-a_{C}}{a_{C}} = \frac{A_S}{A_C+A_S}\frac{a_S-a_C}{a_C} \label{eq:core_strain}\] \[\epsilon_{zz,S} = \frac{\bar{a}-a_{S}}{a_{S}} = \frac{A_C}{A_C+A_S}\frac{a_C-a_S}{a_S} \label{eq:shell_strain}\] The expression deduced for the strain in the core is an approximation of the exact solution derived by by minimizing the strain energy in the structure. The two expressions differ by less than \(0.3\)% for an Al content of \(0.33\). The approximate solution was previously also used by for GaAs/(Al,Ga)As core/shell NWs. To obtain the values of \(\epsilon_{zz,C}\) and \(\epsilon_{zz,S}\) for the present NWs, we take the lattice constants of \(a_C = a_\text{GaAs}=5.65325\,\)Å, \(a_\text{AlAs}=5.66164\,\)Å, \(a_S = a_{\text{Al}_{x}\text{Ga}_{1-x}\text{As}} = x a_\text{AlAs} + (1-x) a_\text{GaAs}\), where \(x\) is the Al content in the shell. For simplicity, we assume a cylindrical shape of the NWs, so that the cross sectional area of the core (shell) amounts to \(A_{C}=\pi r_{C}^2\) \[\(A_{S}=\pi (r_{NW}^2-r_{C}^2)\)\]. Here, \(r_{C}=50\,\)nm and \(r_{NW}=80\,\)nm are the radii of core and the entire NW, respectively.
[\[SI-table3\]]{#SI-table3 label="SI-table3"}
Table [2](#SI-table3){reference-type="ref" reference="SI-table3"} compares the values for \(\epsilon_{zz,C}\) and \(\epsilon_{zz,S}\) obtained by [SPHInX]{.smallcaps} and the simple estimate outlined above. Despite the simplifying assumptions, the estimated values deviate by less than \(10\)% from the calculated ones. For a purely uniaxial tensile strain along the \([111]\) direction, the band gap shrinks linearly with strain at a rate of \(56\) meV/%. The dashed line in Fig. [\[SI-figure8\]](#SI-figure8){reference-type="ref" reference="SI-figure8"} shows the corresponding dependence versus the Al content. Despite the slightly higher strain values compared to the simulation (cf. Table [2](#SI-table3){reference-type="ref" reference="SI-table3"}), the trend predicted by our simple estimate is actually closer to the experimental data than the \(2\)D simulation. The reason for the larger redshift in the simulation is the fact that the single-particle states are affected by the radial piezoelectric potential depicted in Fig. [\[SI-figure6\]](#SI-figure6){reference-type="ref" reference="SI-figure6"}, while the exciton dominating the experimental spectra is not. Hence, the redshift of the \(X_{\text{hh}}\) line can be fully attributed to the tensile strain exerted on the GaAs core by the respective (Al,Ga)As shell.
### Radial electric fields due to surface band bending
Next, we address the magnitude of radial electric fields in the GaAs core arising from surface band bending induced by charge transfer from bulk acceptor to surface states, and their consequences on carrier recombination. Figure [\[SI-figure9\]](#SI-figure9){reference-type="ref" reference="SI-figure9"} depicts the magnitude of the electric field in the GaAs core close to the core/shell interface as a function of acceptor concentration, calculated with the \(2\)D Poisson solver in [nextnano]{.smallcaps} of a GaAs/Al\(_{0.3}\)Ga\(_{0.7}\)As core/shell NW at \(10\,\)K. The strength of the internal radial electric field stays below the exciton ionization threshold for doping levels up to \(10^{16}\,\)cm\(^{-3}\). Additionally, these simulations show that the GaAs NW core is fully depleted for doping levels up to \(10^{17}\,\)cm\(^{-3}\). The magnitude of the electric fields associated with surface band bending depends not only on acceptor concentration, but also on the diameter of the NWs. reported that the variation of the internal electric fields with diameter can be extracted from cw-PL spectra as a systematic shift of the emission energy. The band bending would also lead to a superposition of transitions with different energies and thus potentially influence the FWHM of the PL emission. Figure [\[SI-figure10\]](#SI-figure10){reference-type="ref" reference="SI-figure10"} depicts both the transition energy (a) and the width (b) of the highest energy line in the PL spectra of NW arrays from three samples, namely, the arrays with a hole size of \(90\,\)nm and a pitch of \(700\,\)nm from sample A and B, and an array with a hole size of \(40\,\)nm and a pitch of \(1000\,\)nm from sample D. The latter sample was grown under the same conditions as sample A, with the only difference that the core diameter has been increased by depositing an additional GaAs shell with a thickness of \(25\,\)nm (resulting in an increased core diameter by \(50\,\)nm) was grown prior to the growth of the \(30\,\)nm thick Al\(_{0.3}\)Ga\(_{0.7}\)As shell. The increased diameter of the NWs in sample A and D can be seen in secondary electron micrographs of representative single NWs shown in the inset of Fig. [\[SI-figure10\]](#SI-figure10){reference-type="ref" reference="SI-figure10"}b. While samples A and B exhibit a constant transition energy only slightly redshifted by strain, sample D shows a very different behavior: the redshift is much stronger, particularly at low \(I_L\), and decreases gradually with increasing \(I_L\). Since the strain in the core is reduced for sample D compared to samples A and B, this finding suggests that the radial electric field in sample D is of sufficient strength to dissociate excitons, resulting in recombination of spatially separated electron-hole pairs with an energy redshifted by a radial Stark effect.
## Recombination model for fitting \(\boldsymbol{I_{PL}(I_L)}\) and \(\boldsymbol{\eta_\text{int}(I_L)}\)
We start from a coupled system of rate equations for the densities of excitons, electrons and holes, shallow acceptors, and nonradiative Shockley-Read-Hall centers. Following Refs. , we can summarize this equation system into one effective equation reading \[G-B \Delta n \left( \hat{p}_0+ \Delta p \right)-\frac{\Delta n \Delta p}{\tau_p \Delta n + \tau_n \Delta p} = 0, \label{eq_ModelI}\] with the first term accounting for steady-state generation with rate \(G\), the second for radiative, and the third for nonradiative recombination. This expression is mathematically equivalent to Eq. 5 in Ref. , but with \(\hat{p}_0\) denoting not the background carrier concentration (which is essentially zero in our fully depleted NWs), but representing the monomolecular recombination due to shallow acceptors. In all what follows, symbols not explicitly defined have their usual meaning. In the second term of Eq. [\[eq_ModelI\]](#eq_ModelI){reference-type="ref" reference="eq_ModelI"}, \(B=b_r + \sigma_x \gamma_x\) is the effective radiative recombination coefficient including the contributions of free carriers and excitons. The former is given by \[b_r=\frac{\Gamma}{N_x}\] with the dipole transition rate \[\Gamma = \frac{n E_G e^2 \mid p_{cv} \mid^2}{3 \pi m_0^2 \hbar^2 c^3 \epsilon_0}\] and the total effective density of states \[N_{T}=\frac{1}{\sqrt{2}} \left( \frac{(m_e+m_{hh}) k_B T}{\pi{} \hbar^2} \right) ^{3/2}.\] In the contribution of excitons, \[\sigma_x = \frac{2}{N_{cv}} e^{E_x/k_B T}\] is the scattering volume for free carriers into the exciton state with the reduced density of states \[N_{cv}=\frac{1}{\sqrt{2}} \left( \frac{\mu k_B T}{\pi{} \hbar^2} \right) ^{3/2}\] and \(\gamma_x\) is the exciton's radiative rate. The monomolecular radiative term is proportional to \[\hat{p}_0=\frac{b_A}{B} (N_A-\Delta n + \Delta p ) \frac{\Delta p}{\Delta n}\] with the recombination coefficient for band-to-acceptor recombination \[b_A = \frac{2^5 \sqrt{2} \hbar n E_G e^2 \mid p_{CV} \mid ^2 }{4 \pi m_0^2 c^3 \epsilon_0 (m_h E_A)^{3/2}} \mathcal{I_A},\] where \[\mathcal{I_A} = 2 \pi \left( \frac{\beta}{\pi} \right)^{3/2} \int_{0}^{\infty} \frac{\sqrt{x} e^{-\beta x}}{(1 + x)^4} dx\] with \[\beta = \frac{m_h E_A}{m_e k_B T}.\] In the third term, the electron and hole capture times are given by \[\tau_n = \frac{1}{b_n N_C}~\text{and}~\tau_p = \frac{1}{b_p N_C},\] respectively, where \(b_n\) and \(b_p\) are the respective capture coefficients and \(N_C\) is the density of Shockley-Read-Hall centers. In addition to Eq. [\[eq_ModelI\]](#eq_ModelI){reference-type="ref" reference="eq_ModelI"}, we have to explicitly consider charge neutrality: \[\Delta n + N_A^-= \Delta p + N_C^+ \label{eq_chargeneutrality}\] where \[N_A = N_A^-+ N_A^0~\text{and}~N_C= N_C^+ + N_C^0.\] Note that at \(10\) K, the quasi-equilibrium between neutral (occupied) and ionized (empty) acceptor states implies that \[N_A^-= \frac{N_A}{1 + \sigma_A \Delta p} \approx 0,\] with \[\sigma_A = \frac{2}{N_V} e^{E_A/k_BT},\] where \[N_V = \frac{1}{\sqrt{2}} \left( \frac{m_h k_B T}{\pi \hbar^2} \right)^{3/2}\] is the valence band effective density of states. Equations [\[eq_ModelI\]](#eq_ModelI){reference-type="ref" reference="eq_ModelI"} and [\[eq_chargeneutrality\]](#eq_chargeneutrality){reference-type="ref" reference="eq_chargeneutrality"} define model II as investigated in detail in Ref. . When assuming \(\Delta n = \Delta p\), we obtain model I as \[G-B \Delta n \left( \frac{b_A}{B} N_A + \Delta n \right)-\frac{\Delta n}{\tau_n + \tau_p} = 0, \label{eq_ModelII}\] which is mathematically equivalent to the situation considered in Refs. . All temperature-dependent quantities have been calculated assuming a lattice temperature of \(25\) K (see Sec. ). | {'timestamp': '2022-12-01T02:19:30', 'yymm': '2211', 'arxiv_id': '2211.17167', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17167'} |
null | null |
# [\[sec: introduction\]]{#sec: introduction label="sec: introduction"} Introduction
Translating non-linear spectroscopy methods that are well established in the longer wavelength ranges to the X-ray regime is particularly attractive, since the strong localization of transitions involving core electrons as well as element-specific absorption edges promise additional selectivity to these already potent analytical tools. A prerequisite for non-linear X-ray spectroscopy is a high density of X-ray photons, such as those found in the tightly focused X-ray radiation from a Free-Electron Laser (FEL). The absorption of such pulses drastically modifies the electronic structure of any investigated material, even on the timescale of femtosecond-short pulses. Absorption around material resonances directly probes transitions between the core-levels and states around the Fermi level and is thus sensitive to the electron dynamics in the valence band of materials. Electronic structure changes in this region affect the degree of absorption experienced by later parts of the same pulse and the fluence-dependent absorption can be used to derive information on the excited state and its evolution . Several models have been put forward to describe aspects of this interplay between photon absorption and electronic system. Away from material resonances, fluence-dependent X-ray absorption has been successfully modeled using rate models in three-level systems modeling a ground, core-excited and intermediate valence excited state. When probing the valence bands around material resonances however, three population parameters representing the state of the material at a given point in time become insufficient to represent the non-thermal electron energy distribution around the Fermi level in an extended solid that is relevant for the rates of excitation and relaxation. This makes rate models using only three levels unsuitable to model non-linear absorption near resonances. In this work, we present an expanded rate model to describe the evolution of the electronic system in terms of an energy-resolved population of the valence band within a constant Density of States (DoS). The valence electrons are heated through energy transfer from the scattering of free electrons from Auger decays and non-resonant absorption. The model uses material parameters known or calculated for the ground-state and scales these parameters in accordance with changes in electronic populations. Only the time constants for the valence band thermalization and the scattering cascades of free electrons are treated as free parameters. We present calculations matching measurements of X-ray absorption spectra recorded with monochromatic X-rays in transmission through metallic nickel foils around the nickel 2\(p_{3/2}\) (\(L_3\)) edge; our model reproduces the main fluence-dependent changes in the measured spectra over more than three orders of magnitude. While the measurements are discussed in detail in a separate publication, this paper illustrates the framework of the rate model and strives to provide an intuitive understanding of the mechanisms that drive non-linear changes. The following text is structured as follows. In section ([\[sec: model\]](#sec: model){reference-type="ref" reference="sec: model"}), we give a qualitative overview of the rate model: Section [2.1](#sec: algorithm){reference-type="ref" reference="sec: algorithm"} describes the algorithm and approximations made in the interest of computational viability, Section [2.2](#sec: processes){reference-type="ref" reference="sec: processes"} outlines the mathematical formulation for the rates of all relevant processes, which are then assembled into differentials of the photon and electron populations in section [2.3](#sec: differentials){reference-type="ref" reference="sec: differentials"}. Section [2.4](#sec: parameters){reference-type="ref" reference="sec: parameters"} elaborates on relevant input parameters. Finally, we discuss the use and implications of our model in section [\[sec: discussion\]](#sec: discussion){reference-type="ref" reference="sec: discussion"}
# [\[sec: model\]]{#sec: model label="sec: model"} The rate model
We describe the propagation of X-ray photons through the sample as well as the dynamics of electron populations within the sample using a set of ordinary differential equations. These rate equations describe the evolution of photon and electron densities and are assembled from terms that each describe a specific physical process. The rate of each process is based on ground-state rates, scaled with the appropriate fractional occupation (the number of electrons divided by the number of states). Each process rate is described in detail in section [2.2](#sec: processes){reference-type="ref" reference="sec: processes"}. Such scaling inherently prevents any state from exceeding its physically meaningful population (between zero and the number of available states) and also enforces that the number of electrons in the sample is conserved over the simulated time. The model allows for an arbitrary number of incident resonant photon energies \(E_i\), for each of which a Gaussian temporal profile of incident intensity is assumed. For the presented calculations, only a single resonant photon energy was used, representing measurements with monochromatic X-rays. Incident photons are the only source of energy flow into the system, and all energy eventually contributes to the thermal energy of the valence system. Figure [\[fig: Variables\]](#fig: Variables){reference-type="ref" reference="fig: Variables"} schematically summarizes the modeling of the electronic system: The electron populations \(R_{C}\) and \(R_{V}\) describe the total number of electrons bound in the core and valence system, respectively, for a single atom in the sample. Their values are limited by the number of available states, \(M_{C}\) and \(M_{V}\). In the presented example on the nickel L\(_3\)-edge spectra, the ground-state populations are \(R_C = 4\), representing the 2p\(_{3/2}\)-electrons and \(R_{V}=10\), representing electrons from the 3\(d\) and 4\(s\) states. Electrons in intermediate shells (in the given example the 3\(p\) and 3\(s\) electrons) contribute to the non-resonant absorption but are not explicitly simulated. We describe the electronic population of the valence system in an energy-resolved manner, splitting it up into a discrete number of densities \(\rho_j\), where \(j\) represents the index along the valence energy axis. The number of available states \(m_j\) for each energy bin in the valence system is derived from the calculated ground-state DoS up to 30 eV above the Fermi level \(E_F\). Beyond this value, the free electron gas model is used to calculate the density of states. We set the energy cutoff of the energy-resolved valence band at 800 eV above the Fermi level. All electrons with even higher energies, such as photo-electrons created via non-resonant absorption and Auger-electrons from the decay of core-holes, are described in a separate pool of electrons \(R_{free}\) without energy resolution, although the total energy in this pool is tracked by the parameter \(E_{free}\). Within our model, the DoS does not change over time since the DoS is dominated by the atomic lattice potentials and thus mainly affected by nuclear motion. Since the coupling of electronic excitations to nuclear motion is much slower than the purely electronic dynamics, observations that are temporally confined to the duration of a single FEL pulse (on the order of tens of femtoseconds) are usually dominated by electronic processes.
## Algorithm {#sec: algorithm}
In contrast to a Monte-Carlo approach, where interaction pathways of many individual particles are computed and averaged, our rate model describes processes in terms of non-quantized, average quantities (densities) in a finite-element analysis. The problem is formulated as an idealized X-ray pulse, Gaussian in time and space, traveling at the speed of light through a three-dimensional sample. The photon density interacts non-linearly with the sample. Each voxel of the sample is characterized by a complex electronic state as described above and in Figure [\[fig: Variables\]](#fig: Variables){reference-type="ref" reference="fig: Variables"}, which itself evolves in time after excitation. To make this problem tractable, we apply additional key approximations that allow for a drastic reduction of the computational effort that would come with a naive implementation of this problem in a four-dimensional (three spatial dimensions and time) finite elements algorithm: First, we describe electronic excitations as strictly localized and neglect their propagation; instead, only photons propagate through the sample, and only in forward direction[^1]. Considering that a sample of 20 nm thickness is traversed by light in less than 100 as, the propagation of photons is calculated as if it happened instantaneous in between the time-steps performed to depict the evolution of the electronic system. This de-coupling of dimensional dependence effectively simplifies the problem into two sets of separately solvable, one-dimensional initial value problems: the photon propagation in space and the electronic evolution in time. To solve these, we formulate the time-differential of all electronic states (see section [2.3](#sec: differentials){reference-type="ref" reference="sec: differentials"}), depending on the photon density incident at a certain time. The temporal evolution is solved using the fourth-order Runge-Kutta method with adaptive time-stepping based on a fifth-order local approximation. In the meantime, whenever a time-differential is computed, the current incident photon density is first propagated through the sample using the explicit fourth-order Runge-Kutta method (in space) to retrieve the photon density at each depth of the sample. This way, the time-dependent transmission for a Gaussian temporal pulse profile is calculated for a range of different overall pulse energies incident per sample area. This treatment yet neglects the transversal profile of the beam. Because the absorption is non-linearly dependent on the incident fluence, the transmission of a transversally inhomogeneous beam must be integrated over a specific spot profile with spatially varying transmission. Here, a two-dimensional Gaussian profile of the FEL spot is then accounted for by integrating the transmitted intensity based on previously calculated fluence-dependent transmission. The Figures [\[fig: transmission\]](#fig: transmission){reference-type="ref" reference="fig: transmission"}, [\[fig: fermi solver\]](#fig: fermi solver){reference-type="ref" reference="fig: fermi solver"}, [\[fig: dos development\]](#fig: dos development){reference-type="ref" reference="fig: dos development"} and [\[fig: energy conservation\]](#fig: energy conservation){reference-type="ref" reference="fig: energy conservation"} in the following section show intermediate results from simulations calculated for a specific fluence, as opposed to integrating over the transmission in a Gaussian spot-profile. With these simplifications, the overall computational complexity is drastically reduced.
## Processes {#sec: processes}
The following describes the rates at which the physical processes occur for each atom. For a better overview, we introduce the processes as individual terms and assemble them into differential equations in the next section. In principle, each process is described using an absorption length or lifetime which is known from ground-state measurements and then scaled linearly with the changing electron populations with respect to the ground state. The normalization is such that the ground-state rate is reproduced for an undisturbed electron system and the rate vanishes when the corresponding transition cannot happen due to a lack of electrons or holes. We use the indices \(j\) and \(i\) to refer to specific energies, where the index \(i\) is used for photon energies of X-rays and the index \(j\) for the energy of electronic states in the valence band.
### Resonant interaction
The resonant interaction describes both resonant absorption (core-valence transitions) and stimulated emission (valence-core transitions) as a single process. It is calculated for each energy \(E_j\) in the valence system that is resonant with a given photon energy \(E_i\). \[P^{res}_j = \left(\frac{R_C}{M_C}-\frac{\rho_j}{m_j} \right) \frac{N_i}{\lambda_i^{res}} \delta_{ij}\]
The first terms (in brackets) represent the difference in the occupation of core states \(R_C/M_C\) and resonant valence states \(\rho_j/m_j\). The dominance of absorption over stimulated emission or vice-versa is determined solely by this difference, as they represent an optically driven two-level system in the incoherent limit. If the core level population is smaller than the valence population, the resonant interaction process becomes negative, representing the dominance of stimulated emission. The second term on the right is the number of irradiated photons divided by the penetration length. The Kronecker delta ensures that only photons and electrons in corresponding energy bins interact.
### Non-Resonant absorption
The non-resonant absorption summarizes photon absorption from other electronic states than the resonant core-level, especially from the valence electrons. Photon densities \(N_i\) at all incident energies reduce each population \(\rho_j\). \[P^\mathrm{non-res}_{i,j} = \frac{\rho_j}{R_{V}^0} \frac{N_i}{\lambda^{non-res}}\]
The interaction is normalized by the total valence band population in the ground state \(R_{V}^0\), so that the sum of the first term over all \(j\) becomes unity if all \(\rho_j = \rho_j^0\) (since \(R_{V}^0 \equiv \sum_j \rho_j^0\)). The second term represents the non-resonant absorption in the ground state as can be experimentally determined sufficiently before the resonance in the spectrum. This treatment does not explicitly differentiate the non-resonant absorption from core energy levels other than the one treated by \(R_C\). In the given example with photons resonant to the nickel 2\(p\)-absorption, the 3\(s\) and 3\(p\) core electrons only contribute to a minority of the non-resonant absorption events. In this model, we choose for simplicity to scale this contribution together with the non-resonant scattering from the valence electrons. An exemplary incidence profile and the resulting transmission over time are shown in Figure [\[fig: transmission\]](#fig: transmission){reference-type="ref" reference="fig: transmission"}.
### Auger decay
The model explicitly treats Auger decays that involve one core-hole and two electrons from the valence band. The rate at which an electron in density \(\rho_j\) would decay via an Auger process is calculated as: \[P^\mathrm{Auger}_j = (M_C-R_C)\frac{\rho_j}{R_{V}^0}\frac{R_{V}}{R_{V}^0} \frac{1}{\tau_{C}}\]
The first factor (in brackets) is the number of unoccupied core states, i.e. core-holes. The second factor describes the relative population of electrons at the energy \(E_j\) and the third term is the relative population of the entire valence band to which the electron could transfer its energy. The latter two are normalized by the respective ground state population. The last term is the decay rate in the ground state, where \(\tau_{C}\) represents the ground-state lifetime of a single core-hole. Altogether, this describes Auger decays as interactions between two valence electrons, one emitted and one filling the \(2p_{3/2}\) core-hole. In reality, some fraction of Auger decays will emit electrons from the 3\(s\) or 3\(p\) core levels instead, followed by further Auger processes which emit electrons with the remaining energy of the original core-hole. These are not treated separately in our description, since the indirect decay is, on the one hand, a minority contribution and on the other hand, ultimately results in the same energy transfer to the valence band, albeit with a slightly longer time delay due to the intermediate steps.
### Free-electron scattering
Inspired by earlier approaches to a simplified solution of the Boltzmann equation, we approximate the scattering rates of electrons in terms of characteristic time constants \(\tau_{scatt}\) and \(\tau_{th}\) for the free electrons and valence electrons, respectively. The lifetime of free electrons \(\tau_{scatt}\) represents the inverse rate at which free electrons \(R_{Free}\) scatter and decay to the valence system. While this parameter is ultimately empirical, it depicts a cascade of individual scattering events between electrons. In such a cascade, each free electron eventually transfers all its kinetic energy to the valence system. \[P^\mathrm{scatt} = R_{Free} \frac{1}{\tau_{scatt}}\]
### Electron thermalization
Similarly, \(\tau_{th}\) characterizes the time with which the valence system approaches an internal thermal equilibrium. [\[sec: el therm\]]{#sec: el therm label="sec: el therm"} \[P^\mathrm{therm}_j = \left[ r_j(T, \mu)-\rho_j \right] \frac{1}{\tau_{th}}\]
To this end, the chemical potential and equivalent electronic temperature are calculated in each time-step based on the current internal energy \(U\) and number of valence electrons \(R_{V}\). The Fermi distribution for the calculated chemical potential and temperature then yields a momentary target electron distribution \(r_j(T,\mu)\), which is approached with the electron thermalization constant \(\tau_{th}\). \[\begin{aligned}
r_j(T, \mu) = m_j \frac{1}{e^{(E_j-\mu)/ k_B T}}\\ U =\sum_j \rho_j E_j / \sum_j \rho_j\\ R_{V} = \sum_j \rho_j \end{aligned}\]
While the calculation of \(r_j\) from \(T\) and \(\mu\) is straightforward, determining \(T\) and \(\mu\) from \(U\) and \(R_{V}\) is an inverse problem. This is solved by iterative optimization using the Levenberg-Marquardt method[^2].\
Note that, because by definition both electron densities: the momentary \(\rho_j\) and the thermalized goal distribution \(r_j\) hold the same amount of electrons and internal energy, the sum over the valence band of all electron-distributing thermalization rates is always zero and the change in overall energy is also zero, i.e. \(\sum_j P^\mathrm{therm}_j = 0\) and \(\sum_j E_j P^\mathrm{therm}_j = 0\). Figure [\[fig: fermi solver\]](#fig: fermi solver){reference-type="ref" reference="fig: fermi solver"} shows how the temperature (given as an energy \(k_B T\) in units of eV) and the chemical potential develop over time for a high fluence of 30 \(\mathrm{J/cm^2}\). Although reaching higher temperatures, these results are in agreement with studies treating the nickel valence system heated with optical lasers.
## Differentials {#sec: differentials}
From these process terms, we can assemble the time-differential of the various pools of electrons, photons, and energy. Because the simulation conserves the number of electrons in the sample, in the sum over all \(j\), every process term describing electronic transitions appears equally often with positive and negative sign, representing a transition of electrons from one state to another. For example, the term for Auger decay appears twice with a negative sign in the valence electron differential, and each once with a positive sign in the differential for core-and free electrons.
### Valence electrons
\[\begin{split} \frac{d \rho_{j} }{dt} &= P^{res}_{i=j} \\ &-\sum_i P^\mathrm{non-res}_{i,j}\\ &-P^\mathrm{Auger}_j -\frac{\sum_j \rho_j P^\mathrm{Auger}_j }{R_{V}} \\ &+ P^\mathrm{therm}_j \\ &+ \frac{h_j}{H_{V}} P^\mathrm{scatt} + P_{red} \left(\frac{-\rho_j }{R_{V}} + \frac{h_j}{H_{V}}\right) \end{split}\]
The valence system interacts via all modeled processes, therefore we comment on the equation above line by line. The resonant absorption rate \(P^{res}_j\) changes the valence electron densities \(\rho_j\) at all incident photon energies \(E_i\) where \(i=j\). Via non-resonant absorption, incident photon energies \(E_i\) can remove electrons from \(\rho_j\). The third line shows the primary and secondary Auger electrons. Note that in the sum over all energies \(E_j\), terms 3 and 4 each remove one electron from the valence band for each Auger process occurring. The thermalization term in line four drives electrons towards the thermal distribution based on the current internal energy and population of the valence band, without changing the total valence occupation, as discussed in section [\[sec: el therm\]](#sec: el therm){reference-type="ref" reference="sec: el therm"}. The last line describes the effect of electron scattering. The first term represents electrons from the free electron pool that are re-joining the valence system into a random unoccupied state \(h_j\). The second term describes the electron redistribution inside the valence system in order to take up the energy released by the re-joining electron. This redistribution is calculated as a function of the rate of electronic scattering \(P^{scatt}\) and represents the effect of electronic scattering cascades. In such a cascade, a number of electrons is moved from occupied states \(\rho_j\) to unoccupied states \(h_j\). The total rate of electrons redistributed in this time step through scattering \(P_{red}\) is given by the ratio of the energy that needs to be taken up by the valence system to the energy that the valence system can additionally accommodate. \[P_{red}=\frac{S_{scatt}-S_{joining}}{U_h-U_e}\] The denominator of \(P_{red}\) represents the energy that can maximally be redistributed to the valence system as the difference between the energy that could be contained in the unoccupied states \[U_h = \frac{\sum_j h_j E_j}{H_V}\] and the one already contained in the occupied states \[U_{e} = \frac{\sum_j \rho_j E_j}{R_{V}}.\] The numerator is given by the difference between the rates at which energy is released from the free electron energy pool \[S_{scatt} = P^\mathrm{scatt} \frac{ E_{free}}{R_{free}}\] and the rate at which energy is gained in the valence system due to the formerly free electrons occupying random unoccupied valence states: \[S_{joining} = \sum_j \frac{h_j}{H_{V}} P^\mathrm{scatt} E_j\] Figure [\[fig: dos development\]](#fig: dos development){reference-type="ref" reference="fig: dos development"} shows an example for the development of the valence system within a single voxel at the sample surface, exposed to a pulse of 30 J/cm\(^2\) fluence, resonant to states 7 eV above \(E_F\).
### Core electrons
\[\frac{d R_{C} }{dt} = -\sum_{i,j} P^{res}_{i,j} + \sum_j P^\mathrm{Auger}_j\] The population of core electrons is reduced (or increased, depending on the sign of \(P^{res}_{i,j}\)) by resonant transitions of all incident photon energies \(E_i\) to states at all energies \(\rho_j\) (although this contribution is only non-zero at \(i=j\)), and is increased by Auger decay from electrons of all energies \(j\) in the valence system. Note that the spontaneous emission channel is neglected in our model as it is designed for soft X-ray energies where Auger emission accounts for most core-hole decays (here specifically, 99.1% of the nickel L\(_3\) core-hole decays ). In another concession to the specific experiment simulated here, we further neglect fast electrons leaving the sample, since the electron mean free path is much shorter than the sample thickness (about 1.3 nm compared to a 20 nm thick sample). While a loss process for free electrons would be trivial to implement, the total number of electrons in the system being strictly constant over time is a valuable indicator for the self-consistency of the calculation.
### Free electrons
\[\frac{d R_{Free} }{dt} = \sum_{i,j} P^\mathrm{non-res}_{i,j} + \sum_j P^\mathrm{Auger}_j -P^\mathrm{scatt}\] Unbound or free electrons are generated by non-resonant absorption from all incident photon energies \(E_i\) as well as Auger-decays from all energies in the valence band. The population is reduced by the free electron scattering rate \(P^\mathrm{scatt}\).
### Photon absorption and emission
\[\frac{d N_{i} }{dz} =-P^{res}_{i=j} -\sum_{j} P^\mathrm{non-res}_{i,j}\] The number of photons is reduced or increased by resonant interaction and reduced by non-resonant absorption. Note that this is a purely spatial differential that depicts the instantaneous transmission of a certain number of photons through the entire sample in each time-step.
### Energy of free electrons
\[\begin{split} \frac{d E_{Free} }{dt} &= \sum_{i,j} P^\mathrm{non-res}_{i,j}(E_F+E_i-E_j)\\ &+ \left(\frac{\sum_j E_j (\rho_j-P^\mathrm{Auger}_j)}{\sum_j (\rho_j-P^\mathrm{Auger}_j)}+E_F\right)\sum_j P^\mathrm{Auger}_j \\ &-P^\mathrm{scatt} \frac{E_{Free}}{R_{Free}} \end{split}\]
Because photons can \"kick out\" electrons from arbitrary states in the valence system, it becomes necessary to track the total kinetic energy of the free electrons, even though the distribution of energy among these electrons is not tracked. The rate of energy transfer to the free electron bath is described as the sum of non-resonant absorption and Auger decays, each multiplied by their respective energies (first and second line, respectively). The energy of the secondary Auger electrons is calculated as the average energy of all electrons other than the primary Auger electrons, as those drop to the core level. Finally, each electron that leaves \(R_{Free}\) reduces the energy of the bath by the average energy, which is \(\frac{E_{Free}}{R_{Free}}\). Explicitly tracking this energy also enables us to demonstrate the conservation of energy within the simulation; since there is no channel that allows energy to leave the sample, the energy held in the electronic sub-systems matches that of the absorbed photons at all times. Figure [\[fig: energy conservation\]](#fig: energy conservation){reference-type="ref" reference="fig: energy conservation"} shows the internal energy of the electronic subsystems over time, integrated over an area of 1 nm\(^2\) and the full 20 nm thickness of the sample. The comparison of the energy in various sub-systems demonstrates how quickly the energy of the photon pulse is transferred to valence excitations. Furthermore, observing the energy conservation has proven to be an invaluable tool to select sufficiently fine binning in time, space and energy, as it is particularly sensitive to the accumulation of numerical errors. For example, the calculation for the homogeneous illumination with 30 \(\mathrm{J/cm^2}\) shows a cumulative error in energy of 0.35%. This is in contrast to the electron conservation, which is strictly kept to machine precision level due to the symmetric way the process terms are arranged to form the time differentials.
## Parameters {#sec: parameters}
There are four categories of parameters: First, resolution parameters for the number of steps in time, space, and energy are chosen as a compromise between calculation time and numerical error; second, experimental parameters such as the pulse duration, peak fluence, and bandwidth of the interacting photons reflect experimental conditions; third, ground-state properties such as the atomic density and number of states and electrons, as well as Auger-decay limited core-hole lifetime are drawn from published literature, while the resonant and non-resonant absorption lengths are derived from the ground-state spectrum as described below. Fourth, the model-inherent phenomenological parameters are the valence thermalization time \(\tau_{th}\) and electron scattering time \(\tau_{scatt}\), which are varied to achieve the best match to the experimental results. A list of the relevant parameters and the chosen values for the present calculations, selected to match the experiment presented in, is shown in the appendix Table [\[tab: parameters\]](#tab: parameters){reference-type="ref" reference="tab: parameters"}. The parameterization of several ground-state properties deserves further comment. Firstly, the DoS (motivated at the start of this section) was subdivided into bins \(m_j\) of varying size, favoring a fine resolution for the bound states. The size of the energy bin that is resonantly coupled to the core level by the incident photons is chosen such that it represents the interaction bandwidth of the photons. The interaction bandwidth is calculated as the convolution of the bandwidth of incident photons (i.e. the resolution of the experiment) and the natural line width of the core excitation. We further account for the final state broadening of excitations into less tightly bound states by enlarging the interaction bandwidth by 0.1 eV per eV above the Fermi level. Furthermore, the resonant and non-resonant absorption lengths are derived from the ground-state spectrum. We treat the non-resonant absorption length as constant, i.e. independent of photon energy, and derive it from the pre-edge absorption level. The transition matrix element of a core-valence transition exhibits a resonant enhancement close to the absorption edge, which translates into an energy dependence of the resonant absorption length. Above the Fermi level, where the DoS is unoccupied in the ground state, the resonant absorption length is encoded in the ground-state absorption spectrum. As the transition matrix element from the core level to states below the Fermi level is experimentally not straightforward to access, we use the approximation that the energy-dependence of the transition matrix element is symmetric around the Fermi energy. To derive the resonant absorption length, the non-resonant absorption level is subtracted from the spectrum and line broadening is accounted for by deconvolution with a pseudo-Voigt-profile of 50% Gaussian and Lorentzian share and a width of 640 meV Full-Width Half-Maximum (FWHM), representing 420 meV broadening from the experimental resolution and 480 meV from the core-hole lifetime. The deconvolved resonant absorption spectrum above the Fermi level is then mirrored around the Fermi energy and the discontinuity within 320 meV around it is reconstructed with cubic interpolation. This results in the mirrored resonant absorption spectrum shown in Figure [\[fig: input spectrum\]](#fig: input spectrum){reference-type="ref" reference="fig: input spectrum"}, which is used as the resonant absorption length parameter \(\lambda_i^{res}\). Note that the results of simulated spectra were finally re-convolved with the same pseudo-Voigt-profile to simulate the same experiment.
# [\[sec: discussion\]]{#sec: discussion label="sec: discussion"} Discussion
## Parameter Study
In Figure [\[fig: spectra\]](#fig: spectra){reference-type="ref" reference="fig: spectra"} we show the measured non-linear X-ray absorption spectra labeled (I) presented and discussed in together with sets of simulated spectra computed for different sets of parameters. The first set of simulated spectra (also presented in ) represents the best match with the experimental conditions and the parameters shown in table [\[tab: parameters\]](#tab: parameters){reference-type="ref" reference="tab: parameters"} and is labeled (II), while the consecutive sets, labeled (III) to (VI), demonstrate how the results change when individual parameters are modified. We present this set of simulations to showcase how our model may be used to understand the relation between the non-linear changes and various parameters. To fit the experimental results, however, only the parameters \(\tau_{th}\) and \(\tau_{scatt}\) are treated as unknowns, while all other parameters are known experimental or ground-state parameters. In the best matching simulation (II), the experimental observations of a red-shifted rising edge, increased pre-edge absorption, as well as reduced absorption at and beyond the absorption peak, are reproduced. However, the saturation effect at the resonance is over-estimated and the lack of measured saturation around 852 eV cannot be explained by our model. See for a more detailed discussion of these deviations. The next simulation (III) was performed without non-resonant absorption. While this eliminates the pre-edge absorption rise, some shift of the absorption onset is still predicted within the original peak, while the spectra above the resonance onset behave similarly to the best-matching simulation. With simulations (IV) and (V), we demonstrate the effect of prolonging the thermalization time or shortening the pulse duration by a factor of ten, respectively. Both have the similar effect of strongly reducing the peak shift and a moderate decrease of saturation beyond the resonance. Finally, we show non-linear spectra (VI) where the free electron scattering process was eliminated (\(\tau_{scatt}=\infty\)). This prevents the majority of the excitation energy from entering the valence system and thus drastically reduces valence heating. This causes the pre-edge absorption rise and rising edge shift to vanish nearly entirely, and absorption-decrease due to saturation at and above the resonance is also strongly reduced. These differences underscore the importance of electronic scattering cascades to these phenomena. While an extensive study of the correlations between the various model parameters and the observed effects is beyond the scope of this work, this brief parameter study allows an interpretation of how red-shift and pre-edge absorption rise occur: Let us consider a case where the incident photon energy is slightly below the absorption edge. Initially, only non-resonant absorption transfers energy to the sample, specifically by creating free electrons in form of photo-electrons. This energy is then transferred to the valence band due to electronic scattering cascades, where it causes many small, non-thermal excitations. These are homogeneously distributed over all valence energies and thus only lead to a small increase in pre-edge absorption (see the simulation with slow thermalization). However, this distribution of excitations develops towards the shape of a Fermi distribution over time, specifically with the thermalization time constant. If the pulses are long in comparison to the thermalization time, a significant number of states below the Fermi level becomes available for core-valence transitions during the pulse duration. This causes the observed increase in pre-edge absorption. Once the first empty valence states become available at the current photon energy, resonant absorption begins to occur in addition to the non-resonant absorption. The additional resonant absorption leads to more free electrons from Auger-decay, which in the same way as the photo-electrons further contribute to secondary electron scattering and thermalization. Because thermalization mostly creates free states just below the Fermi level, this process of self-enhancing rise of overall absorption manifests in a fluence-dependent shift of the rising edge to lower energies.
## Limitations and Potential
As demonstrated in the parameter study above, this kind of analysis can provide a straightforward interpretation of how non-linear changes to the absorption spectrum emerge, and what each change says about the development of the electronic populations. While the model already demonstrates good agreement with experimental results, below we discuss the limitations of our approach and possibilities for expanding or refining it. In the model as described here, the DoS is not only assumed constant but also does not differentiate between spin-up and spin-down states, although we are treating a magnetic material. Splitting up the DoS in spin-up and spin-down states would allow for the inclusion of angular momentum conservation and transfer in the various scattering rates. For thin samples where the electron mean free path becomes similar to the layer thickness, a loss process for free electrons should be introduced to account for electrons leaving the sample. Fluorescent decay should be accounted for when moving to harder X-rays. While introducing the decay channel itself would be simple, accurately accounting for reabsorption may be less straightforward, since the model only propagates light in one direction. Furthermore, the thermalization time \(\tau_{th}\) is used in this work as a global fitting parameter, although electron thermalization times have been suggested to depend on electronic temperature. Since the electronic temperature and target distribution are calculated every time-step, an arbitrary dependence could be easily introduced, albeit with the necessity of additional fitting parameters. The DoS is dominated by the crystal lattice, which is typically stable on the sub-100 fs timescale. However, recent Time-Dependent Density Functional Theory (TDDFT) calculations show that electronic processes (i.e. sub 100 fs) can also lead to modifications of the DoS via spin-orbit coupling modeled by introducing an onsite Hubbard correlation \(U\) to the mean-field Hamiltonian. Since the rate model approach is generally not suited to calculate the DoS, incorporating modifications to the DoS would require a close interplay of the rate model with (TD)DFT calculations, leading to an ultimately much more complex approach reminiscent of models developed for the study of radiation-induced damage mechanisms. Furthermore, we derive the interaction bandwidth (the valence energy range to which core states can be resonantly coupled by incident photons) as a convolution of instrumental resolution and the lifetime of the core excitation, i.e. the Auger lifetime. The final state lifetime broadening of excitations into continuum states is described as a continuous broadening of 0.1 eV per eV above the Fermi energy. It is however reasonable to expect that the final state lifetime is further shortened at higher fluences, due to increased rates of both electronic scattering and stimulated emission; the latter would be particularly relevant at the resonance peak. Such further broadening would cause more valence states to be available for resonant interaction and reduce the observed saturation effect while increasing the number of core-holes that may be created by high fluences[^3]. Accounting for a fluence-dependent broadening of the interaction bandwidth in a refined rate model might help to remedy the overestimation of the saturation effect at the absorption peak in the presented calculations. Another candidate for further refinement is an energy dependence of the electronic scattering rates. The presented model uses fixed rates for thermalization and scattering cascades, which both act on all valence states indiscriminately. This description would be especially inadequate when applied without modifications to a bandgap material. An advanced model could describe both the thermalization rate of the valence band and the energy of excitations from scattering cascades in an energy-resolved manner[^4]. While such refinements may seem attractive, a core strength of the rate model approach is its relative simplicity and computational tractability, as well as the use of known ground-state parameters, which supports a straightforward physical interpretation. It is ultimately a mostly classical, phenomenological model which offers a complementary approach to ab-initio calculations. Every added complexity should therefore be weighed against its relevance, as a simpler model facilitates a meaningful interpretation and avoids introducing ambiguity in the results due to correlations between redundant input parameters. Since the model operates on widely applicable principles, we expect that it may be applied to a wide range of materials with some predictive power, while the limitations described above apply. The results will especially deviate from observations wherever multi-particle effects, such as electron correlations or quasiparticles become relevant. As presented, the model enables an understanding of the electronic population history under strong X-ray fluences and characterization of the resulting non-linear absorption near a core resonance. Non-linear absorption studies like the one analyzed here allow one to characterize the excitation dynamics of the electronic system under study. Furthermore, consideration of these dynamics is, due to the extreme fluences required, particularly relevant for methods in the emerging field of non-linear X-ray spectroscopy.
# [\[sec: conclusion\]]{#sec: conclusion label="sec: conclusion"} Conclusion
In this work, we propose a model of differential rate equations describing the various excitation and decay processes that connect core-and valence electronic states to quantitatively describe the non-linear changes to X-ray absorption around a core resonance that occur when employing increasingly high FEL pulse energies. We present the framework of the model in detail, applying it to the case of recently measured non-linear absorption spectra of nickel films at the \(L_3\) edge, recorded with monochromatic X-rays. We demonstrate how the rate model reaches good agreement with the experimental results while disentangling the contributions of the relevant physical processes: resonant and non-resonant absorption, Auger decay, electron thermalization, and electronic scattering cascades. Our rate model contains numerous simplifications in order to approximate the result of complex interactions between many particles in terms of average transition rates. Nevertheless, the resulting description enables a quantitative understanding of the evolution of the system and the processes responsible for spectral changes. In the presented calculations, the most relevant effect is valence system heating due to secondary electron cascades from free electrons. Our results allow quantifying the absorbance changes caused by fundamental electron population dynamics, which are crucial to disentangle from collective quantum effects studied with currently evolving non-linear X-ray spectroscopy techniques.
[^1]: We neglect spontaneous X-ray emission, which would constitute up to 0.9 % of nickel \(L_3\) core-hole decays.
[^2]: First, the Levenberg-Marquardt root-finding algorithm is applied with a maximum of 400 iterations. If the required residual is still exceeded, the algorithm switches to least-squares optimization to refine the root-finding result. This combination proved a good compromise between computation speed and stability in regions with small gradients in the loss function.
[^3]: When considering only resonant absorption, the core level, and resonant valence states constitute a classic two-level system. In the extreme fluence limit, the relative populations of such a system are given by the state degeneracy ratio, i.e. the ratio between the number of involved core and valence states.
[^4]: It might be non-trivial to normalize this energy dependence such that the electronic scattering processes do not violate the conservation of energy. | {'timestamp': '2022-12-01T02:19:00', 'yymm': '2211', 'arxiv_id': '2211.17144', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17144'} |
# Introduction
Bursts of star formation in galaxies are usually concentrated in small regions know as Super Star Cluster (SSC), which are compact in size (\(\sim\)`<!-- -->`{=html}1 pc), massive (\(M_{\rm \star}\[\sim\)`<!-- -->`{=html}10\(^5\) \(M_{\sun}\)) and young (from 1 up to 100 Myr). This mode of star formation is thought to play a major role in the evolution of galaxies, specially since it seems to be more frequent in galaxies with high star formation rates resulting from merging events, as in starbursts (SB) and in luminous and ultraluminous systems. It is expected that in a very short time scales (few 10\(^{6-7}\) yr) the most massive stars in SSCs will explode as supernova ejecting hot shells that eventually will generate a powerful outflow. So far, it is unclear to what extent the evolution of the SSC will produce a sequential SSCs formation due to positive feedback like in the Antennae galaxies (@Herrera11) or whether it will quench star formation (@Wall20). The study of SSCs in nearby galaxies is key in order to understand the triggering mechanisms that lead to the starbursts, and to establish the effects of the feedback processes (positive versus negative) in the earliest stage of their evolution. However, the earliest phases of SSCs formation take place deeply embedded in the parental cloud, preventing their observation even at mid-IR. Recently, have identified and studied the formation and the earliest stages of evolution of the SSCs, deeply embedded in the parent clouds, in the starburst galaxy NGC 253 by means of the HC\(_3\)N vibrationally excited (HC\(_3\)N\(^*\)) emission. They found that the SSCs in NGC 253 seem to be formed by the overpressure generated by the hot expanding gas produced by the SN explosion(s) (timescale \(\sim\)`<!-- -->`{=html}1 Myr) from an earlier star formation episode in the galaxy center but unrelated to the outflow, thus suggesting sequential SSC formation. NGC 4945, the closest (D\(\sim\)`<!-- -->`{=html}3.8 Mpc) composite galaxy in which an AGN and SB co-exist, shares with NGC 253 similar properties and they are both located at similar distances. Both galaxies have a star formation rate (SFR) of about 3 \(M_{\sun}\) yr\(^{-1}\) and show the presence of an outflow both in X-rays, optical and in the molecular emissions (see @Schurch02 [@Mingozzi19; @Bellocchi20; @Bolatto21] for NGC 4945 and @Strickland02 [@Westmoquette11; @Bolatto13; @Krieger19] for NGC 253): their ionized outflows share similar momentum flux and mass outflow rate (@Heckman90). However, the molecular outflow in NGC 4945 seems less massive (@Krieger19) and slightly less powerful (@Bolatto21) than that in NGC 253. The main difference between both galaxies is the age of the SB: in NGC 253 it is relatively younger (\(\sim\)`<!-- -->`{=html}1 Myr; @Leroy18) than that in NGC 4945 (\(\sim\)`<!-- -->`{=html}5 Myr; @Emig20). Then the study of the formation and early evolution of the SSCs in NGC 4945, when compared to the one already carried out toward NGC 253 (@Rico20 [@Rico21; @Rico22]), has the potential to provide further information on the evolution of the starburst and to establish to what extent the sequential star formation (positive feedback) also dominates at later stages or whether the SSC formation has been quenched. In this work we use ALMA data to study the early evolutionary stage of the proto-SSCs using the HCN and HC\(_3\)N vibrationally excited emission in NGC 4945. From the location and the ages of the young SSCs we found that the proto-SSCs are older than in NGC 253 and that the last generation of SSCs in NGC 4945 have been triggered by the positive feedback of the molecular outflow (@Bolatto21) arising from the previous generations of SSCs.
:::
**Notes:** Column (2): Size derived as in Sect. [5.1](#size_par){reference-type="ref" reference="size_par"}. Column (3) gas mass estimates. The (\(^{\star}\)) symbol indicates the mass of gas derived as described in Sect. 5.6. Column (4): Protostellar luminosity of the SSCs (see Sect. [5.2](#lum_derive){reference-type="ref" reference="lum_derive"}). When the v\(_{71}\) emission is not detected, we used the the upper limit to the size and the upper limit to the \(T_{\rm vib}\) listed in Columns 2 and 5 in Table [2](#Input_data){reference-type="ref" reference="Input_data"}, respectively. Column (5): The (\(^{\star\star}\)) symbol indicates the we have estimated the \(M_{\rm ZAMS}\) using the 90 GHz continuum emission following. Column (7): Estimated age of SSCs: for \(L_{\rm p*}\)/\(L_{\rm ZAMS}\]\gtrsim\)`<!-- -->`{=html}0.05, \(t_{\rm age}\)=10\(^5\)/(1 + \(L_{\rm p*}\)/\(L_{\rm ZAMS}\)) while for \(L_{\rm p*}\)/\(L_{\rm ZAMS}\[<\)`<!-- -->`{=html}0.05, \(t_{\rm age}\]\gtrsim\)`<!-- -->`{=html}10\(^5\) yr (assuming the timescale of an HII region to be \(\sim\)`<!-- -->`{=html}10\(^5\) yr).
We have also identified a clear linear trend (\(\rho\[\sim\)`<!-- -->`{=html}0.7) between the rotational temperature derived from the HC\(_3\)N,v\(_0\) line with the evolutionary stage of the SSCs, as illustrated in Fig. [\[fig_pos_Trot_2\]](#fig_pos_Trot_2){reference-type="ref" reference="fig_pos_Trot_2"}. The proto-HC\(_3\)N show the highest \(T_{\rm rot}\) (\(\sim\)`<!-- -->`{=html}65 K), while the proto-HCN show moderate rotational temperatures (\(\sim\)`<!-- -->`{=html}45 K). Finally, the ZAMS-SSCs show rather low \(T_{\rm rot}\) (\(\sim\)`<!-- -->`{=html}25 K; see also Table [3](#Params_physical){reference-type="ref" reference="Params_physical"}). In the earliest phase of star formation of the SSCs, with \(t_{\rm age}\) of few 10\(^4\) yr (proto-HC\(_3\)N phase), the radiation of the protostars excites HC\(_3\)N\(^*\) together with HCN\(^*\) emission, which increases \(T_{\rm rot}\). This is followed by a more evolved phase (proto-HCN) with \(t_{\rm age}\) closer to 10\(^5\) yr with moderate \(T_{\rm rot}\), likely when the HII regions start to dominate the heating and the radiative feedback in the SSCs, photodissociating the most fragile molecules like HC\(_3\)N and HNCO (see also @Rico20 [@Rico21]). Finally, the ZAMS dominated phase is characterized by rather low \(T_{\rm rot}\) and \(t_{\rm age}\]\gtrsim\)`<!-- -->`{=html}10\(^5\) yr. ZAMS-12 and-13 depart from this trend, showing a\(T_{\rm rot}\) (\(\sim\)`<!-- -->`{=html}50 K) closer to that derived for proto-HCN. The higher \(T_{\rm rot}\) suggests that they could maybe going through a transition phase between proto-HCN and ZAMS phases. The low excitation temperature of the parent cloud when the SSC reaches the ZAMS stage indicates that the heating by its ultracompact HII regions is mainly restricted to their very close neighborhood. The inner warm gas heated in the proto-HCN is likely dissociated and only the remaining outer cold parental cloud is observed.
## Triggering the SSC formation. Positive feedback {#new_region}
An inspection of Figs. [\[fig_pos_Trot_2\]](#fig_pos_Trot_2){reference-type="ref" reference="fig_pos_Trot_2"} and [\[outflow_map\]](#outflow_map){reference-type="ref" reference="outflow_map"} shows that the youngest proto-SSCs with ages of 7\(\times\)`<!-- -->`{=html}10\(^4\) yr are concentrated at both sides with respect to the major axis of the galaxy at a projected distance of-1.6 and +1.9 from the nucleus. Adopting an inclination of \(\sim\)`<!-- -->`{=html}62\(^\circ\) for the inner disk (@Chou07), we derived a deprojected distance of \(\sim\)`<!-- -->`{=html}65 and \(\sim\)`<!-- -->`{=html}80 pc, respectively. proposed the presence of an inflow of gas in NGC 4945 through a bar from (deprojected) galactocentric radii of 300 pc to \(\sim\)`<!-- -->`{=html}100 pc toward the nuclear disk, while the central parts (\(\lesssim\)`<!-- -->`{=html}100 pc) are disturbed by the outflowing gas. Our SSCs fall well within the outflow region, which makes less feasible that the inflow could induce star-formation. Furthermore, it is remarkable that the proto-HC\(_3\)N are just bordering the outer edges of the molecular outflow observed in CO by, as clearly illustrated in Fig. [\[outflow_map\]](#outflow_map){reference-type="ref" reference="outflow_map"}. In the region dominated by the outflow, we also found two out of three proto-HCN (\#5 and \#7), which are probably located outside the molecular outflow but likely associated with it. investigated the stability of the clumps in the outflow using the virial parameter, \(\alpha_{vir}\) finding that most of the gas in the outflow is unlikely to form stars. From the projected location of basically all the young proto-HC\(_3\)N, just at the edges of the blue and redshifted lobes of the molecular outflow, we propose that their formation could have been triggered in the shocked gas produced by the expansion of the molecular outflow. derived a dynamical age for the ionized outflow in NGC 4945 to be in the range 3-10 Myr. Even for the molecular phase of the outflow, which shows a more compact distribution (\(\sim\)`<!-- -->`{=html}70 pc), a dynamical timescale of about 0.3 Myr is derived (@Bolatto21). Both dynamical times are larger than the age of the proto-SSCs, making it possible the triggering of the most recent episode of massive star formation in NGC 4945 as seen in HC\(_3\)N\(^*\) and HCN\(^*\) by the outflow.
## The evolution of the star formation in NGC 4945 and NGC 253. Sequential SSC formation.
While the typical age of the proto-SSCs in NGC 4945 are about \(>\)`<!-- -->`{=html}0.8\(\times\)`<!-- -->`{=html}10\(^5\) (\(\pm<\)`<!-- -->`{=html}0.2\(\times\)`<!-- -->`{=html}10\(^5\)) yr, those in NGC 253 are younger, with \(t_{\rm age}\[\sim\)`<!-- -->`{=html}0.2-0.8 \(\times\)`<!-- -->`{=html}10\(^5\) yr. The derived ages for the SSCs in NGC 4945 indicate that the burst of star formation in this galaxy seems to be more evolved than in NCG 253, in agreement with the age of the burst in each galaxy (i.e., \(\sim\)`<!-- -->`{=html}1 My for NGC 253 and \(\gtrsim\)`<!-- -->`{=html}5 Myr for NGC 4945; @Leroy18 [@Emig20]). Also the mean gas mass reservoir available for star formation in the proto-SSCs in NGC 253 (\(\sim\)`<!-- -->`{=html}1.7\(\times\)`<!-- -->`{=html}10\(^5\) \(M_{\sun}\)) is larger, by at least a factor of 2, than in the NGC 4945 (7.5\(\times\)`<!-- -->`{=html}10\(^4\) \(M_{\sun}\)), also consistent with the different ages of the starburst in both galaxies. In fact, the gas mass reservoir in NGC 4945 is quite similar in both (proto and ZAMS) phases (see Table [3](#Params_physical){reference-type="ref" reference="Params_physical"}), of about 7\(\times\)`<!-- -->`{=html}10\(^4\) \(M_{\sun}\), indicating that the burst of star formation in NGC 4945 is likely reaching its end, as in NGC 253, which also shares similar amount of dense gas (HNCO/CS) with NGC 4945 (@Martin09). In fact, the different origins of the youngest proto-SSCs in both galaxies can be understood in the context of the scenario of feedback and sequential SSC formation. While the most evolved SSCs in NGC 253 are mainly located closer to the center of the galaxy and the younger proto-SSCs in the outer parts ('inside-out' scenario), in NGC 4945 this trend is the opposite, with most of the youngest proto-SSCs found toward the center. The SSCs age and spatial distribution in NGC 253 seem to suggest that the SSCs formation have been triggered by the overpressure produced by hot gas generated by the SN explosions from an early star formation episode in the center of the galaxy (@Rico20). In the case of more evolved starburst as NGC 4945, the well developed and strong molecular outflow, from the previous episode of star formation, induced the new generation of SSCs by positive feedback. Recently, a face-on (deprojected) view has been proposed in NGC 253 by, which still confirms the sequential star formation scenario of SSCs. Furthermore, according to the model shown by, the results obtained considering the deprojection of NGC 4945 would be still valid in a face-on view scenario.
# Conclusions
We present ALMA imaging of vibrationally excited HC\(_3\)N and HCN emissions to study the earliest phase of the formation and evolution of SSCs in the composite SB/AGN galaxy NGC 4945. Our main results can be summarized as follows:
- We have identified 14 forming SSCs in the center of NGC 4945: 8 of them show HCN vibrational excited emission (HCN\(^*\)), 5 of which also present HC\(_3\)N vibrationally excited emission (HC\(_3\)N\(^*\)). The other 6 SSCs do not show any vibrationally excited emissions in spite of their large stellar masses. From the LTE analysis of the HC\(_3\)N\(^*\) lines, we derived both rotational (\(T_{\rm rot}\)) and vibrational (\(T_{\rm vib}\)) temperatures of the molecular gas associated with the SSCs;
- We found that the 8 SSCs detected in vibrationally excited emission are in the proto phase, characterized by a mean age of \(\sim\)`<!-- -->`{=html}7\(\times\)`<!-- -->`{=html}10\(^4\) yr, in contrast with the 6 more evolved SSCs, which are in the ZAMS phase with ages \(\gtrsim\)`<!-- -->`{=html}10\(^5\) yr;
- We identified a clear trend between \(T_{\rm rot}\) from HC\(_3\)N with the different stages of evolution of the SSCs. The earliest phase of SSC formation with \(t_{\rm age}\]\lesssim\)`<!-- -->`{=html}7\(\times\)`<!-- -->`{=html}10\(^4\) yr is traced by HC\(_3\)N\(^*\) and HCN\(^*\) vibrationally excited emissions and high \(T_{\rm rot}\), then followed by a more evolved phase traced by the detection of only HCN vibrationally excited emission, with moderate \(T_{\rm rot}\) and \(t_{\rm age}\) closer to 10\(^5\) yr, when HII regions start to dominate the heating and the radiative feedback (photodissociation). Finally, the ZAMS phase is characterized by rather low \(T_{\rm rot}\), and \(t_{\rm age}\)\(\gtrsim\)`<!-- -->`{=html}10\(^5\) yr, without detection of the vibrationally excited emissions;
- We found that the youngest proto-SSCs are located at the edges of the outflow (@Bolatto21), suggesting that the outflow has triggered the formation of the youngest proto-SSCs (positive feedback) in the nearby shocked regions;
- SSCs in NGC 4945 are found to be in a more evolved stage than those in NGC 253, consistent with the sequential SSC formation scenario. While in NGC 253 SSCs seem to be formed by the overpressure produced by the SN explosions of the previous generation of stars, the youngest generation of SSCs in NGC 4945 have been triggered in the shocked regions produced by the well developed molecular outflow. | {'timestamp': '2022-12-01T02:21:45', 'yymm': '2211', 'arxiv_id': '2211.17268', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17268'} |
# Introduction
The classical Seshadri constant of a nef invertible sheaf \(\mathcal L\) on a projective variety \(X\) at a closed point \(p\), introduced by Demailly, is defined to be \[\begin{aligned}
\varepsilon(\mathcal L,p)&=\sup\{t\in\mathbb R_{\geq 0}\mid\pi^*c_1(\mathcal L)-tE\ \text{is nef},\ \pi:\mathrm{Bl}_p(X)\rightarrow X,\ E=\mathrm{exc}(\pi)\}\\ &=\inf \{\frac{c_1(\mathcal L)\cdot C}{\mathrm{mult_p(C)}}\mid C\ \text{is an irreducible curve passing through}\ p\}\in \mathbb R_{\geq 0}. \end{aligned}\] The Seshadri constants of nef \(\mathbb Q\)-Cartier \(\mathbb Q\)-divisors are defined in the same way. The constants measure the local positivity of \(\mathcal L\) and have a closed relation with the separation of jets of \(\mathcal L\) and of the adjoint bundle \(\omega_X\otimes_{\mathcal O_X}\mathcal L\). More precisely, the Seshadri criterion asserts that a Cartier divisor \(D\) on a projective variety \(X\) is ample if and only if \(\inf_{p\in X}\{\varepsilon(D,p)\}>0\), and, if \(X\) is smooth projective, \(\omega_X\otimes_{\mathcal O_X}\mathcal L\) separates \(\ell\)-jets at \(p\) if \(\varepsilon(\mathcal L,p)\geq\dim X+\ell+1\). See or for more detailed introductions. Despite the fact that the classical Seshadri constants are fundamental and richly structured, there are few studies on the possible generalizations. The most notable works in this direction, among others, are Beltrametti--Schneider--Sommese and Hacon on the Seshadri constants of ample vector bundles. Recently, Fulger and Murayama defined in the Seshadri constants of not necessarily nef coherent sheaves on projective varieties. To state their definition, recall that for any coherent sheaf \(\mathcal G\) on a smooth projective curve \(C\), the minimal slope of \(\mathcal G\) is defined to be \[\mu_{\min}(\mathcal G):=\min\{\mu(\mathcal H)\mid \mathcal H\ \text{is a qoutient of}\ \mathcal G,\ \mu(\mathcal H):=\frac{\deg{\mathcal H}}{\mathrm{rk}\,\mathcal H},\ \deg(\mathcal H):=\chi(\mathcal H)-\mathrm{rk}\,\mathcal H\cdot\chi(\mathcal O_C)\}\] with the convention that the slope of a torsion sheaf is \(\infty\). Following, we define \[\overline{\mu}_{\min}(\mathcal G):=\left\{\begin{array}{ll} \mu_{\min}(\mathcal G) &,\, \text{if}\ \mathrm{char}(\mathbf k)=0\\ \lim\limits_{n\rightarrow\infty}\frac{\mu_{\min}((F^n)^*\mathcal G)}{p^n}&,\, \text{if}\ \mathrm{char}(\mathbf k)=p \end{array}\right.,\] where \(\mathbf k\) is the underlying field and \(F:C\rightarrow C\) is the absolute Frobenius morphism. The Seshadri constant of a coherent sheaf \(\mathcal F\) on a projective variety \(X\) at a closed point \(p\in X\) is defined to be () \[\begin{aligned}
\label{def}\tag{1.1} \varepsilon(X,\mathcal F;p)&:=\inf_{p\in C\subseteq X}\{\frac{\overline{\mu}_{\min}(\nu^*\mathcal F)}{\mathrm{mult}_p(C)}\mid C\ \text{is an irreducible curve,}\ \nu:\tilde C\rightarrow C\ \text{is the normalization}\}\\ &\in\mathbb R\cup\{\pm\infty\}. \end{aligned}\] We usually write \(\varepsilon(\mathcal F,p)\) if no confusion arises. These constants agree with the classical Seshadri constants when \(\mathcal F\) is a nef invertible sheaf, and it has been shown in that they behave similarly to their classical counterparts.
- \(\inf\{\varepsilon(\mathcal F,p)\mid p\in X\}>0\) if and only if \(\mathcal F\) is ample ().
- Suppose \(X\) is smooth of dimension \(n\) and \(\mathcal F\) is an ample locally free sheaf of rank \(r\). If \(\varepsilon(\mathcal F,p)>\frac{n+s}{\ell+r}\), then \(\omega_X\otimes S^\ell\mathcal F\otimes\det\mathcal F\) separates \(s\)-jets at \(p\) ().
Besides being useful in studying the local positivity of coherent sheaves, Seshadri constants can also be employed to characterize the projective spaces. Bauer and Szemberg showed in that a complex Fano manifold is isomorphic to the projective space if the Seshadri constant of \(-K_X\) is larger than \(\dim X\) at some point. Later, Liu and Zhuang generalized the theorem to complex \(\mathbb Q\)-Fano varieties. Fulger and Murayama asked a question in higher rank similar to these observations:
In other words, it is expected that the Seshadri constants of tangent bundles can not strictly lie between \(0\) and \(1\). The conjecture has been verified when \(X\) is a smooth projective surface or a Fano manifold. The purpose of this paper is to study Conjecture [\[conj\]](#conj){reference-type="ref" reference="conj"} when \(X\) is toric. Our main results are the following:
If \(X\) is a projective \(\mathbb Q\)-factorial toric variety which is possibly singular, we still have \(\ve(T_X,p)\geq\min_{\rho\in\Delta(1)}\{\ve(D_\rho,p)\}\) (see Lemma [\[lemma1\]](#lemma1){reference-type="ref" reference="lemma1"} and the proof of Proposition [\[prop2\]](#prop2){reference-type="ref" reference="prop2"}). The author does not know any such \(X\) violating Corollary [\[cor0\]](#cor0){reference-type="ref" reference="cor0"}. It is worth mentioning that an explicit formula is given in for the Seshadri constants of nef toric vector bundles at the torus invariant points. On the other hand, when \(p\) is not a torus invariant point, it's not easy to calculate the Seshadri constants even for torus invariant divisors. For example, if \(H\) is the ample generator of the Picard group of \(X=\mathbb P(a,b,c)\) and \(1\in X\) is the identity of its dense torus, then \(\varepsilon(H,1)=1/\sqrt{abc}\) is equivalent to Nagata's conjecture for \(abc\) points (). Somewhat, Corollary [\[cor0\]](#cor0){reference-type="ref" reference="cor0"} is only of theoretical interest. The proof of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} is divided into two parts: \(p\in T\) or \(p\in X\setminus T\). If \(p\in X\setminus T\), we first show that the normal sequence of a \(T\)-invariant prime divisor \(Y\) inside the smooth projective toric variety \(X\) splits, and then apply an induction argument. To deal with the case \(p\in T\) (may assume \(p=1\)), we establish a combinatorial criterion of when \(\varepsilon(X,T_X;1)>0\) (Theorem [\[combinatoric\]](#combinatoric){reference-type="ref" reference="combinatoric"}). Combining with on the existence of certain primitive collection, we show that the only smooth projective toric variety with \(\varepsilon(X,T_X;1)>0\) is the projective space. For this part, if \(\mathrm{char}\,\mathbf k=0\), we remark that one can simply apply to get the same result (see for details). The advantage of our argument is that it does not depend on the characteristic of \(\bf k\).
# A combinatorial criterion
In the remaining of the paper, all varieties are defined over an algebraic closed field \(\mathbf k\) of any characteristic. The main result of this section is Theorem [\[combinatoric\]](#combinatoric){reference-type="ref" reference="combinatoric"}. Instead of repeating, we just recall some of the lemmas that are necessary in our calculation.
We refer to and for basic knowledge of toric geometry.
If \(\pi:Y\rightarrow X\) is the blow-up along a smooth subvariety \(Z\subseteq X\setminus\mathrm{Sing}(X)\) and \(E\subseteq Y\) is the exceptional divisor, there is a naturally defined morphism \(\pi^* T_{X}(-E)\rightarrow T_{Y}\). Such a morphism can be used to estimate the Seshadri constants of \(T_Y\) in terms of that of \(T_X\) and \(E\). The following is the toric analog.
Note that (2) is stronger than merely assuming \(\varepsilon(D,1)=\varepsilon(T_X,1)=0\); in (2) we are assuming that the Seshadri constants are zero and they can be achieved by the same sequence of irreducible curves.
# Proof of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"}
The following example demonstrates that the conjecture fails when the variety is not smooth, even in the toric case:
# some formulas for Seshadri constants
Note that the corollary is true no matter \(T_X\) is nef or not. It would be interesting to see whether Corollary [\[cor5\]](#cor5){reference-type="ref" reference="cor5"} is true for every smooth projective variety \(X\).
Although the above example shows that we can actually have \(0>\varepsilon(T_X,p)>-\infty\), it is quite rare in the sense of the following proposition.
Note that \(\mathrm{dim}(\overline{\mathrm{orb}(p)})\leq 1\) does not guarantee \(\varepsilon(T_X,p)>-\infty\). For example, let \(X=\mathrm{Bl}_{p}\mathbb P^3\overset{\pi}{\rightarrow} \mathbb P^3\) and let \(E=\mathrm{exc}(\pi)\). We have \(\varepsilon(E,p)=-\infty\) for any \(p\in E\) since \(E\simeq\mathbb P^2\) and \(E|_E\) is anti-ample. Then \(\varepsilon(T_X,p)=-\infty\) for any \(p\in E\) by Proposition [\[prop3\]](#prop3){reference-type="ref" reference="prop3"}.
Finally, we raise a question on the necessity of the projectiveness assumption.
In view of Proposition [\[primitive\]](#primitive){reference-type="ref" reference="primitive"}, \(X(\Delta)\) must be non-projective. If such an example does exist, then it will serve as a counterexample of Conjecture [\[conj\]](#conj){reference-type="ref" reference="conj"} without assuming that \(X\) is projective. If the answer to Question [\[projectiveness\]](#projectiveness){reference-type="ref" reference="projectiveness"} is no, then the proof of Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} can be carried over and we can generalize the theorem to smooth complete toric varieties. | {'timestamp': '2022-12-01T02:19:39', 'yymm': '2211', 'arxiv_id': '2211.17172', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17172'} |
null | null |
null | null |
null | null |
{'timestamp': '2022-12-05T02:06:14', 'yymm': '2211', 'arxiv_id': '2211.17143', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17143'} |
|
# Introduction
The paradigm of learning universal large-scale Transformer-based models has been recently transferred to speech from natural language processing . Due to the high complexity of speech data, the preferred way of downstream task adaptation for such models is to keep pretrained weights frozen and fine-tune small task-specific heads , so it is important to use the model's hidden states in the most efficient way. A common approach is to use outputs of the last layer of the model, combining them via various pooling methods; but it has been shown that for some tasks, e.g. phoneme prediction in HuBERT or the analogy task in BERT, embeddings from lower and middle layers are more useful . *Topological data analysis* (TDA) is a recently proposed way to get more efficient data representations from frozen Transformer weights . TDA features prove to be better suited for many downstream tasks, including artificial text detection and linguistic acceptability judgement. In particular, for ungrammatical sentence detection conventional sentence embeddings yield classification quality no better than random, while TDA has led to meaningful results. Inspired by these results, in this work we apply TDA to the HuBERT model in order to build more powerful speech representations. TDA has already been applied to signals of various nature; previous attempts for speech used topological statistics of the audio signal waveform: persistent entropy for noise classification and emotion recognition , detection of a periodic signal in noisy data etc. However, TDA for Transformer-based models has so far been limited to NLP, processing attention maps for artificial text detection and linguistic acceptability and word embeddings for dialogue term extraction and constructing story trees . Evolution of inner representations in neural networks has been studied with persistent Betti numbers and recently introduced representation topology divergence (RTD) . In this work, we apply RTD in particular to intermediate embeddings and attention maps of a speech Transformer. Another interesting area of TDA applications is the interpretation of pretrained Transformer models: it is known in NLP that different heads are sensitive to different phenomena . With TDA techniques, we demonstrate the same effect for speech Transformers, finding heads that are best for solving specific downstream tasks, e.g. separating a given pair of emotions, a given pair of speakers, detecting speech generated by a specific TTS model, or representing spectral features of sound samples (bit rate, LPCC etc.). We also investigate the pattern structure of attention maps in HuBERT with pattern classification methods and demonstrate that our TDA-based approach is able to extract meaningful representation from heads with any pattern types.
# Methods {#sec:methods}
In this section, we describe our approach. We first introduce topological data analysis for weighted graphs and then define the features we propose to extract. **Graph homologies**. Given a set of points in a high-dimensional space, TDA proposes ways to restore the underlying lower-dimensional manifold and study its topological properties. If the set of points has a graph structure, the \(d\)-dimensional manifold can be obtained in the form of a *simplicial complex*. Informally speaking, it is the structure obtained by replacing \(k+1\)-cliques in the graph by filled-in \(k\)-dimensional simplices for all \(0\le k \le d\): \(0\)-dimensional simplices are vertices, \(1\)-dimensional are edges, for \(k=2\) we get triangles, for \(k=3\), tetrahedrons, and so on . *Homology groups* are key topological features for the classification of topological spaces . Dimensions of the sequence of homology groups \(H_k\), \(k \ge 0\), are between of the main topological invariants. In the case of graph-like structures, \(\dim H_0\) equals the number of connected components, \(\dim H_1\) is the number of "loops", and \(\dim H_k\) counts \(k\)-dimensional "holes". **Persistent homologies for weighted graphs and point clouds**. We would like to apply TDA to sets of vectors in \({\mathbb R}^d\) (embeddings). Such a set (point cloud) can be viewed as a complete graph \(G\) with edges weighted by any distance-like measure between vectors. However, homology groups are only defined for unweighted graphs, so we make \(G\) unweighted by thresholding, leaving only edges with weights lower than a given \(\epsilon\); we denote the resulting graph by \(G_\epsilon\). TDA can track the changes of topology across varying thresholds via *persistent homology*. In this case, we build a family of graphs called a *filtration*: an ordered set of graphs \(\Gamma_\epsilon\) for a sequence of increasing \(\epsilon\). For \(\epsilon\) below the minimal distance between vertices the graph has no edges; as \(\epsilon\) increases new edges are added, ending in the complete graph; during this process, gradual changes of graph topology can be expressed in terms of "birth" or "death" of basic features. We begin with \(|V|\) connected components (all of them are "born"), and as \(\epsilon\) increases, pairs of them are joined together (one component "dies"). "Birth" and "death" moments can be represented with a diagram called the *barcode* , whose horizontal axis is a sequence of thresholds \(\epsilon\), and each horizontal bar corresponds to a single feature. \(H_0\)-bars correspond to connected components; they start at the minimal distance between vertices and correspond to edges of the minimal spanning tree (MST). \(H_1\)-bars correspond to simple cycles with more than \(3\) vertices; an \(H_1\)-bar starts when a cycle appears and ends when it is broken into triangles by a new edge. We denote by \(H_k^m(G)\) the average length of \(H_k\)-bars for a weighted graph or point cloud \(G\). *Representation topology divergence* (RTD) measures the topological dissimilarity between two data representations, formalized as two weighted graphs \(G^a\) and \(G^b\) with a common set of vertices. Namely, we track how the topology of the graph \(G^a_\epsilon \cap G^b_\epsilon\) changes as the filtration threshold \(\epsilon\) increases. These changes are then summarized via a special kind of barcode, called the *cross-barcode*, which reflects how much graphs \(G^a\) and \(G^b\) have in common on different scales. Our RTD feature is equal to the sum of bar lengths in this barcode. For a more general and formal definition, see . | {'timestamp': '2022-12-05T02:13:10', 'yymm': '2211', 'arxiv_id': '2211.17223', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17223'} |
null | null |
null | null |
null | null |
# Introduction {#sec:intro}
Over the past decade, end-to-end (E2E) supervised ASR has achieved outstanding improvements. These achievements keep pushing the limit of ASR performance in terms of word error rate (WER). However, training the state-of-the-art model still heavily relies on a reasonable amount of annotated speech. Unfortunately, labeled data is quite limited for most of the 7000 languages worldwide. The fast development of self-supervised learning (SSL) could mitigate the issue to some extent by leveraging unlabeled data. This paradigm first learns the speech representations from raw audio and then fine-tunes the model on limited transcribed speech data , resulting in reducing the need for annotated speech. However, as there is still a need for transcribed data for downstream model training, it is difficult to directly apply the system to all languages, especially those endangered languages that are extremely difficult to obtain data. Unsupervised ASR could be one possible direction to solve the problem where the model can be trained with more accessible unpaired speech and text data. Wav2vec-U is the state-of-the-art UASR framework. It utilizes both SSL (i.e., wav2vec2.0) and adversarial training for the UASR task. The framework is shown working not only in English, where the wav2vec2.0 is trained on but also in several other mainstream languages , as well as low-resource languages . This finding reveals the possibility of utilizing unsupervised learning for more languages. The authors of wav2vec-U have released their code in FAIRSEQ, which greatly improves reproducibility. The implementation mainly consists of 3 steps: data preparation, generative adversarial training (GAN) , and iterative self-training followed by Kaldi LM-decoding . Along with this reproducibility direction, we develop an unsupervised ASR toolkit named ESPnet Unsupervised ASR Open-source toolkit (EURO). EURO complements the original FAIRSEQ implementation with more efficient multi-processing data preparation, flexible choices over different SSLs, and large numbers of ASR tasks through ESPnet . EURO also integrates a weighted finite-state transducers (WFST) decoder using the k2 toolkit for word-level recognition. K2 is the updated version of the popular ASR toolkit Kaldi . It seamlessly integrates WFST and neural models implemented in PyTorch by supporting automatic differentiation for finite state automaton (FSA) and finite state transducer (FST), which are commonly used in ASR as a natural representation of the model's architecture . In EURO, k2 provides a compact WFST structured and efficient algorithm for decoding. With these advantages, EURO can considerably benefit the UASR study for the speech community, together with the FAIRSEQ UASR implementation. This paper first introduces the toolkit and its features. Then, we conduct experiments that explore mainstream self-supervised models as speech feature extractors for UASR in different languages. Finally, we provide details of the hyperparameters of our experiments.
# Related works {#sec:related works}
This section briefly compares the framework of EURO to wav2vec-U. As summarized in Table [\[tab: comparison\]](#tab: comparison){reference-type="ref" reference="tab: comparison"}, EURO provides more flexible choices of SSL model as the acoustic feature extractor by integrating with the S3PRL toolkit . With the comprehensive pipeline in the template, EURO enjoys a fast adoption to various datasets with a limited data preparation effort (less than 20 lines of code for a minimum runnable solution). Meanwhile, all the data preparation stages in EURO are designed to enable computing in parallel, which greatly minimized the preprocessing time compared to wav2vec-U. Besides of WFST decoder introduced in Sec. [1](#sec:intro){reference-type="ref" reference="sec:intro"}, EURO provides a self-implemented decoder to eliminate external dependencies.
[\[tab: comparison\]]{#tab: comparison label="tab: comparison"}
# Functionalities of EURO {#sec:euro}
Fig. [\[fig:architecture\]](#fig:architecture){reference-type="ref" reference="fig:architecture"} shows the architecture of EURO. Similar to other ESPnet tasks, EURO includes two major components: a Python library of network training/inference and a collection of recipes for running complete experiments for a number of datasets. The library is built upon PyTorch, while the recipes offer all-in-one style scripts that follow the data format style in Kaldi and ESPnet. In addition, a WFST decoder is included to perform word-level recognition.
## Models {#ssec: models}
As discussed in Sec. [2](#sec:related works){reference-type="ref" reference="sec:related works"}, we follow previous works in using adversarial training to achieve UASR in EURO. To be specific, we extend the Wav2vec-U framework into our implementation. Given a spoken utterance \(\mathbf{X} \in \mathcal{D}_{\text{speech}}\), we first extract the speech representation by using a speech SSL model as the feature extractor \(f(\cdot)\), resulting in a sequence of hidden representations \(\mathbf{H}\). Then, the sequence \(\mathbf{H}\) is passed into a preprocessor \(m(\cdot)\) to form a segmented feature sequence \(\mathbf{S}\), which is used for the generative adversarial network (GAN). The preprocessor \(m(\cdot)\) includes three steps, including adjacent clustering pooling, principle component analysis (PCA) dimension reduction, and mean pooling. The adjacent clustering pooling utilizes the K-Means cluster IDs from the input feature \(\mathbf{H}\) as guidance to merge the adjacent feature frames. The network is mainly trained with a GAN-based loss and some auxiliary supporting losses. Given the segmented feature \(\mathbf{S}\) and an unpaired phonemicized text sequence \(\mathbf{Y}_{u} \in \mathcal{D}_{\text{text}}\), the framework includes a generator \(\mathcal{G}\) and a discriminator \(\mathcal{C}\) as the classic GAN framework. The generator \(\mathcal{G}\), which serves as the ASR model, transcribes \(\textbf{S}\) into a phoneme sequence \(\mathbf{P}\) and the discriminator \(\mathcal{C}\) tries to distinguish \(\mathbf{Y}_u\) from the generated phoneme sequence \(\mathbf{P}\). The GAN-based loss \(\mathcal{L}_{\text{GAN}}\) is as follows: \[\mathcal{L}_{\text{GAN}} = \min_{\mathcal{G}} \max_{\mathcal{C}} \mathbb{E}_{\mathbf{Y}_u} [\log \mathcal{C}(\mathbf{Y}_u)]-\mathbb{E}_{\mathbf{H}} [\log(1-\mathcal{C(\mathcal{G}(\mathbf{S}))})].\] To stabilize the training, three auxiliary losses are also proposed, including (a) a gradient penalty loss \(\mathcal{L}_{\text{gp}}\) to sample the mixing rate of real and fake input for different steps: \[\mathcal{L}_{\text{gp}} = \underset{\mathbf{S}, \mathbf{Y}_u, \alpha \sim U(0, 1)}{\mathbb{E}} [(||\nabla \mathcal{C}(\alpha \mathcal{G}(\mathbf{S}) + (1-\alpha) \mathbf{Y}_u)||-1)^2],\] where \(\alpha\) is the mixing weight sampled from a uniform distribution . (b) a smoothness penalty to penalize inconsistent phoneme prediction between adjacent segments: \[\mathcal{L}_{\text{sp}} = \underset{(p_n, p_{n+1}) \in \mathcal{G}(\mathbf{S})}{\sum} ||p_n-p_{n+1}||^2,\] where \(p_n \in \mathbf{P}\) is the generator output distribution at the \(n\)'s segment. (c) a phoneme diversity loss to prevent the generator \(\mathcal{G}\) from generating the same phoneme all the time: \[\mathcal{L}_{\text{pd}} =-\sum_{n} \mathrm{Entropy}_{\mathcal{G}}(\mathcal{G}(\mathbf{S})),\] that is defined by the entropy of the average generator output distribution over every frame. The final loss is defined as \[\mathcal{L} = \mathcal{L}_{\text{GAN}} + \lambda \mathcal{L}_{\text{gp}} + \gamma \mathcal{L}_{\text{sp}} + \eta \mathcal{L}_{\text{pd}},\] where \(\lambda, \gamma, \eta\) are the weights for each term.
## Frontend {#ssec:frontend}
One of the major benefits of EURO compared to wav2vec-U is its tight integration with S3PRL[^1], a toolkit for speech/audio self-supervised models, to avoid manually managing and switching between different SSL models. S3PRL supports various SSL models as a general toolkit. S3PRL has kept up-to-date with the latest SSL models in the speech and audio domain. Based on the integration with S3PRL, by simply changing **one line** of the configuration, EURO can utilize up to **27** speech and audio SSLs with more than **70** of their variants.[^2]
## Decoding {#ssec: decoding}
EURO offers two methods for decoding, including a self-implemented prefix beam search method and a graph-based search method using k2. The prefix beam search utilizes the same decoding process as the CTC prefix decoding, but without the blank symbols. Similar to other ESPnet tasks, the decoding can be integrated with phoneme-level language models (LM) from both n-gram LMs and neural LMs. As briefly introduced in Sec. [1](#sec:intro){reference-type="ref" reference="sec:intro"}, the graph-based search employs WFST for decoding. Different from the prefix beam search method, the graph-based search can utilize word-level LMs in the search graph, and can also be extended to word recognition. The search graph \(T\) is a composition of three functional graphs: an alignment graph \(H\), a lexicon graph \(L\), and a grammar graph \(G\): \[T = H \circ L \circ G,\] where \(\circ\) is WFST composition. Specifically, \(H\) merges duplicated adjacent phones; \(L\) maps sequences of phonemes to sequences of responding words; \(G\) is an n-gram word LM. Fig. [\[fig:h\]](#fig:h){reference-type="ref" reference="fig:h"} and Fig. [\[fig:lexicon\]](#fig:lexicon){reference-type="ref" reference="fig:lexicon"} show an example of \(H\) and \(L\) implemented in k2. The lattice is generated during decoding , which represents the set of most likely hypothesis transcripts structured in a directed graph and can be easily integrated with neural LMs by performing lattice scoring . The best hypothesis is obtained by searching the best path in the lattice.
## Recipes for reproducible experiments {#ssec: recipes}
### Directory structure
EURO follows the unified directory organization of ESPnet as shown in Fig. [\[fig:tree\]](#fig:tree){reference-type="ref" reference="fig:tree"}. Similar to other tasks (e.g., ASR, speech translation (ST)), the recipe `usar.sh` and its related bash scripts are stored under `egs2/TEMPLATE/uasr/`. The model `espnet_model.py` and task `uasr.py` are stored at `espnet2/uasr/` and\
`espnet2/tasks/`, respectively. Decoding scripts\
`uasr_inference.py` and `uasr_inference_k2.py` are placed under `espnet2/bin/`.
### Recipe flow
The recipe in EURO follows the ESPnet2 style of task, which provides the template `uasr.sh`. The stages are defined as follows: ***Stage 1-5**: Data preparation.* The initial data format starts from the Kaldi style, but the `text` for transcription is supposed to be unpaired with the speech data. Then, we offer two optional data preprocessing stages: speed perturbation and voice activity detection (VAD). The VAD results can be applied in silence removal as it is shown to be important for some speech corpus for wav2vec-U. After the preprocessing, all speech data is converted into a standard format, by resampling, segmentation, silence removal, and dumping from pipe-style formats.\
***Stage 6-7**: Text tokenization and token list generation.* Texts are converted to phoneme tokens using the graphemes to phonemes toolkit g2p-en. Tokens are collected from the training text and formed into a corresponding token list for modeling. For UASR, the unpaired text \(\mathbf{Y}_u\) is fed into training in a random fashion. For efficiency purposes, we initialize a randomized text loader with the tokenized text, which is especially useful for large text data.\
***Stage 8-11**: LM preparation.* These stages train and evaluate LMs based on the unpaired text \(\mathcal{D}_{\text{text}}\). The LMs include both neural-based LMs and N-gram LM.\
***Stage 12**: WFST graph construction.* This stage creates the WFST decoding graph for the k2 decoder.\
***Stage 13**: UASR statistics collection.* In this stage, EURO collects necessary input statistics for batching and mean-variance normalization (MVN). Optionally, the feature from frontends (i.e., S3PRL module in EURO) can be extracted at the stage to support efficient training in further stages.\
***Stage 14**: UASR feature preprocessing.* This stage applies the preprocessing steps discussed in Sec. [3.1](#ssec: models){reference-type="ref" reference="ssec: models"}, including adjacent clustering pooling, PCA, and mean pooling. The resulting segments are used for UASR training.\
***Stage 15**: UASR training.* This stage conducts the adversarial training as discussed in Sec. [3.1](#ssec: models){reference-type="ref" reference="ssec: models"}.\
***Stage 16**: UASR decoding.* EURO has two decoding schemes as introduced in Sec. [3.3](#ssec: decoding){reference-type="ref" reference="ssec: decoding"}. This stage supports both decoding methods.\
***Stage 17**: UASR evaluation.* The evaluation in this stage utilizes the NIST *sclite* toolkit to compute the phone error rate (PER) or the word error rate (if applicable).\
***Stage 18-20**: Model packing and uploading.* This stage automatically packs the trained model checkpoint for easier sharing of the pre-trained model. EURO also supports uploading models to Huggingface for model sharing.
# Experiments {#sec:experiments}
## Datasets
**TIMIT**: The TIMIT dataset is a popular benchmark for the UASR task. It contains 6300 sentences (5.4 hours) of reading speech. All sentences are manually transcribed to phonemes with time alignment. We use the standard split of the train (3696 sentences), dev (400 sentences), and test (192 sentences) sets for our experiments.\
**Librispeech**: The LibriSpeech dataset is a common benchmark for the ASR task . It contains 960 hours of reading speech automatically derived from the audiobooks of LibriVox. This corpus is split into 3 training sets (100 and 360 hours of clean speech, and 500 hours of other speech), 2 dev sets (each has 5 hours), and test sets (each has 5 hours).
## SSL models
We explore three SSL models for UASR in EURO, wav2vec 2.0 (wav2vec2-large-ll60k), HuBERT (hubert-large-ll60k), and WavLM (wavlm-large).[^3] The three models have the same architecture that consists of 24 layers of Transformer encoders with a similar number of parameters. Among them, wav2vec 2.0 and HuBERT are pre-trained on 60,000 hours of Libri-Light. In addition to Libri-Light , Wavlm uses 10,000 hours of Gigaspeech and 24,000 hours of VoxPopuli for pre-training.
## Comparison of SSL models
**TIMIT**: We test EURO on TIMIT to confirm that the toolkit works properly. We use the text from the same training set for unsupervised training. To make it comparable with wav2vec-U, we adopt the same setup for EURO wav2vec 2.0. More specifically, we use the model wav2vec2-large-ll60k and the features are extracted from the layer of the model. For Hubert and WavLM, we explore the performance of features from different layers and report the best PER results. Table [\[tab: timit\]](#tab: timit){reference-type="ref" reference="tab: timit"} shows the results on TIMIT. Wav2vec-U serves as the baseline and achieves 17.0% and 17.8% PER on dev and test sets, respectively. EURO with the same setup gets 18.5% and 19.8% PER on the dev and test set which is slightly worse than wav2vec-U. While our model is not heavily tuned under this setup, the results are comparable with wav2vec-U. For Hubert (hubert-large-ll60k), EURO performs best on features extracted from the layer. The PERs on the dev set and test set are 14.9% and 16.4%. Compared with the wav2vec-U baseline and EURO wav2vec 2.0, it provides a relative improvement of 10% and 20%, respectively. WavLM (wavlm-large) provides further improvement. The model trained using features extracted from layer of WavLM achieves 17% relative improvement compared with baseline and 25% relative improvement compared with EURO wav2vec 2.0 in terms of PER. It reduces the PER to 14.3% on the dev set and 14.6% on the test set.\
\
**Librispeech**: For Librispeech, because of the limitation of computing resources, we use 100 hours of clean speech for UASR training. Unlike TIMIT, we use the text from the whole training set (960 hours) excluding the overlap part with the training speech. The text is phonemicized using G2P phonemizer . In total, around 25m sentences are used for UASR training. We measure both the PER and WER of UASR systems using different SSL models for this dataset. Wav2vec-U uses a sophisticated decoder to convert phoneme sequences to word sequences and it may not be easily applied to other datasets. To make a fair comparison, we train the wav2vec-U model using the same data and load the model into EURO. We use the same prefix beam search decoder for PER and k2 WFST decoder for WER. Table [\[tab: librispeech\]](#tab: librispeech){reference-type="ref" reference="tab: librispeech"} show the results of LibriSpeech's standard dev and test sets. All EURO models outperform the baseline wav2vec-U. For phone recognition, the HuBERT model performs best on clean sets. It achieves PER 15.2 on the dev clean set and PER 15.1 on the test clean set. Wav2vec 2.0 performs best on more difficult sets. It achieves PER 19.3 and 19.8 on dev other and test other sets, respectively. For word recognition, Hubert gets the best WER of 29.3, 22.8, and 29.8 on dev other, test clean and test other sets. WavLM performs best on dev clean and achieves 23.0 WER which is slightly better than HuBERT.
## Experimental details {#ssec: exp details}
For training, we set \(\lambda=1.5\), \(\gamma=0.5\), and \(\eta=2.0\) for TIMIT and \(\lambda=2.0, \gamma=1.0, \eta=4.0\) for LibriSpeech datasets. For (phone-level) prefix beam search decoding, we set \(\text{beam size}=2\) and tunes the weight n-gram language model in \(\left[ 0,0.9 \right]\). For (word-level) graph-based WFST decoding, we set \(\text{search beam size}=30\), \(\text{output beam size}=15\), \(\text{min active states}=14,000\), and \(\text{max active stats}=56,000\). More details can be found in the configuration file under `egs2/TEMPLATE/uasr/conf/`.
# Conclusions {#sec:conclusions}
This work introduces a new toolkit for unsupervised ASR, namely EURO. The toolkit is developed as an open platform for the research field of unsupervised ASR. The current architecture of EURO is based on the Wav2vec-U framework but greatly improves the reproducibility with flexible frontends of almost 30 SSL models and a faster preparation/inference compared to its original implementation in FAIRSEQ. By integrating with k2, EURO provides a WFST decoder for word recognition. Our experiments on TIMIT and LibriSpeech show that we could get comparable performances with wav2vec-U in FAIRSEQ but even better results with Hubert and WavLM as new frontends. | {'timestamp': '2022-12-02T02:15:19', 'yymm': '2211', 'arxiv_id': '2211.17196', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17196'} |
# Introduction
The quantum Hall effect occurs when a two-dimensional electron gas (2DEG) is cooled to low temperature and placed in a strong magnetic field. Fractional quantum Hall states exhibit fractionally quantized Hall conductance, and have excitations that behave as fractionally charged quasiparticles which obey anyonic braiding statistics. The concept of particle-hole symmetry plays a central role in the hierarchical construction of daughter states from the \(\nu=1/m\) Laughlin series. If Landau-level mixing is neglected, particle-hole symmetry is expected for each Landau level; a similar approximate particle-hole symmetry exists for composite fermions. The \(\nu = 2/3\) fractional quantum Hall state can be considered to be the hole-conjugate state to the \(\nu = 1/3\) state. In this picture, the fully filled \(\nu = 1\) state is considered to be the vacuum, and holes form a \(\nu = 1/3\) state on top of this vacuum, leading to an overall 2/3 filling factor. The 2/3 state has been investigated in numerous theoretical works. The edge mode structure in hierarchical states may possess several branches. Moreover, multiple edge structures have been proposed for the 2/3 state. The edge structure which most straightforwardly follows from the hole-conjugate nature of the state was proposed by MacDonald. In this picture, the edge consists of an outer downstream integer edge mode (from the underlying \(\nu = 1\) state) and an inner, counterpropogating-1/3 edge mode. In this context, counterpropagating (or upstream) edge modes travel in the direction opposite to the conventional quantum Hall chiral edge current (edge modes following the conventional direction may be referred to as downstream). Equilibration between the two edge modes leads to a total conductance of \(\frac{2}{3} \frac{e^2}{h}\). A different edge structure was proposed by Kane and collaborators in which disorder dominates and drives an edge phase transition to a single downstream charge mode with conductance \(\frac{2}{3}\frac{e^2}{h}\) and a counterpropagating neutral mode. A model developed by Meir and coworkers considered the impact of soft confining potential in addition to disorder and predicted a reconstructed edge involving the formation of a strip with filling factor \(\nu = 1/3\) at the edge leading to two downstream charge modes with conductance \(\frac{1}{3} \frac{e^2}{h}\) along with counterpropagating neutral modes; other theoretical works have also examined the possibility of incompressible strips with fractional fillings when the confining potential is soft. Experimental studies of the \(\nu = 2/3\) state in quantum point contacts have usually shown an intermediate conductance plateau with \(G = \frac{1}{3} \frac{e^2}{h}\) when the QPC gates are negatively biased to bring the edges close together, which supports the existence of the Meir edge structure in those experiments. Experiments in which this \(G = \frac{1}{3} \frac{e^2}{h}\) intermediate plateau was observed also detected upstream noise attributed to counterpropagating neutral modes. These neutral modes are likely to be detrimental to interference experiments because their entanglement with the charge modes leads to dephasing. The device investigated in our experiments consists of two quantum point contacts (QPCs) separated by a pair of side gates, and utilizes a GaAs/AlGaAs heterostructure with auxiliary screening wells above and below the main quantum well (shown in Supp. Fig. 1 and described in Supp. Section 1). This geometry can be used to define a Fabry-Perot interferometer, but in this work we probe the QPCs independently without inducing interference. A false color SEM image is shown in Fig. [\[Integer_data\]](#Integer_data){reference-type="ref" reference="Integer_data"}a. The QPCs have a separation of 300 nm, and the dimensions of the interferometer are 800 nm \(\times\) 800 nm. We probe the QPCs individually, measuring the conductance of one QPC while tuning the gate voltage on the other to full transmission (in this mode of operation no interference occurs since backscattering only occurs in the QPC being probed). We apply a voltage on the source contact \(V_{SD}\) which is a combination of an AC voltage and a DC voltage (the application of DC bias is described where relevant). The AC bias typically used is \(10\) \(\mu\)V. Using another pair of contacts, the voltage drop across the device is measured, as illustrated in Fig. [\[Integer_data\]](#Integer_data){reference-type="ref" reference="Integer_data"}a, which makes it possible to calculate the conductance as \(G = \frac{I}{V}\). We stress that the semiconductor heterostructure employed here has been designed to produce sharp confining potentials consistent with our recent interferometry experiments. Details of the interferometric measurements at \(\nu=2/3\) will be detailed in a future publication.
# Integer Regime
In Fig. [\[Integer_data\]](#Integer_data){reference-type="ref" reference="Integer_data"}b we show the conductance of the individual QPCs, QPC 1 and QPC 2, versus gate voltage at zero magnetic field. The QPCs exhibit quantized conductance plateaus corresponding to fully transmitted conductance modes. In Fig. [\[Integer_data\]](#Integer_data){reference-type="ref" reference="Integer_data"}c we show conductance versus gate voltage at the \(\nu = 1\) integer quantum Hall state. Unlike at zero field, where the conductance gradually decreased in steps as the applied gate voltage reduces the effective width of the QPC, at \(\nu = 1\) the QPCs exhibit a single wide plateau (corresponding to full transmission of the \(\nu = 1\) edge state) and a very abrupt drop in conductance when the edge state is pinched off in the QPC. This abrupt pinchoff suggests that the confining potential is sharp, as expected for devices utilizing the screening well structure, since minimal backscattering occurs until the edges are brought very close together. This behavior contrasts with the behavior generally observed in QPCs on standard GaAs heterostructures without screening wells at \(\nu = 1\), which often show gradual decreases in conductance, resonances, and sometimes intermediate fractionally quantized plateaus due to edge reconstruction.
# QPC conductance at \(\nu = 2/3\)
Next we operate the device at bulk filling factor \(\nu = 2/3\). Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}a shows the bulk Hall resistance \(R_{xy}\) and the diagonal resistance \(R_D\) measured across the device with the 2DEG under the gates depleted, but with small enough voltages on the gates that minimal backscattering occurs in the QPCs, \(V_{QPC 1}=V_{QPC 2} =-0.3\) V. A wide plateau is visible in \(R_{xy}\) with resistance \(\frac{3}{2} \frac{h}{e^2}\approx 38.7\) k\(\Omega\), corresponding to the \(\nu = 2/3\) state, and \(R_D\) exhibits a nearly quantized plateau over a somewhat narrower region. Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}b shows conductance versus gate voltage for both QPCs at \(B\) = 5.9 T, near the center of the plateau. When there is zero applied DC bias (dashed lines), the QPCs exhibit exhibit a somewhat less sharp pinchoff than at \(\nu = 1\). Between 0 and-2 V both QPCs show conductance close to \(\frac{2}{3} \frac{e^2}{h}\), indicating close to full transmission of all edge states (note that the curve for QPC 2 is shifted by 0.5 V to the right for clarity). Past the point where conductance starts to drop more rapidly, both QPCs exhibit a nearly flat region before pinchoff. Unlike the experiments in Ref., there is no intermediate quantized conductance plateau at \(G=\frac{1}{3} \frac{e^2}{h}\). The behavior changes when a finite source-drain bias of 50 \(\mu\)V is applied, shown in the solid lines; in this situation the applied source-drain bias is a combination of a 10 \(\mu\)V AC bias which is used to probe the differential conductance and the 50 \(\mu\)V DC bias. With the finite bias both QPCs exhibit an intermediate conductance plateau; however, the value of the intermediate conductance plateaus is 0.5\(\frac{e^2}{h}\) rather than the \(\frac{1}{3} \frac{e^2}{h}\) seen in previous experiments (the quantization is not exact, but the conductance is within 2% of 0.5 \(\frac{e^2}{h}\) for both QPCs). This observation suggests that a different edge structure may be present in our system, although it is not obvious why the intermediate plateau only occurs when finite bias is applied. Intermediate conductance plateaus are generally interpreted as occurring when one or more inner edge states are fully reflected while the outer edge states are fully transmitted, leading to some range of gate voltage before the next outer edge state starts to be backscattered when there is an incompressible region in the middle of the QPC and the conductance is constant. In Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}c we plot the differential conductance for QPC 1 as a function of source-drain bias and QPC voltage; dashed lines indicate regions where \(G \approx 0.5 \frac{e^2}{h}\). It appears that the nearly-flat region in conductance observed at zero applied bias develops into the 0.5 \(\frac{e^2}{h}\) plateau as finite bias is applied, and the effect is symmetric around zero bias (there is a small offset of a few \(\mu\)V due to an offset in the amplifier). The value of the intermediate plateau does not change significantly when \(V_{SD}\) is increased past \(\approx 25\) \(\mu\)V. In Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}d we plot conductance of QPC 1 as a function of magnetic field \(B\) and QPC voltage \(V_{QPC 1}\) at \(V_{SD}\) = 50 \(\mu\)V. Both the primary plateau at \(G = \frac{2}{3} \frac{e^2}{h}\) and the intermediate \(G = 0.5 \frac{e^2}{h}\) plateau persist over a significant range of magnetic field from approximately 5.77 T to 6.05 T. Outside of this range the conductance drops and the plateaus are no longer quantized, as expected from inspection of Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}a. Line cuts at several values of magnetic field are shown in Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}e. Despite the requirement of a small finite bias, the \(G = 0.5 \frac{e^2}{h}\) plateau appears to be a robust feature, since it appears in both QPCs, spans a significant range of magnetic field, and does not have significant \(V_{SD}\) dependence past \(25\) \(\mu\)V. We have observed this \(G = 0.5 \frac{e^2}{h}\) conductance plateau in another set of QPCs on a different chip fabricated on the same wafer (Supp. Fig. 2).
# Edge model at \(\nu = 2/3\)
The 0.5 \(\frac{e^2}{h}\) conductance plateau is not straightforward to interpret, since no proposed edge structure contains an outer edge state with \(\Delta \nu = 0.5\) which would carry 0.5 \(\frac{e^2}{h}\) on its own (unlike the \(\nu = 5/2\) fractional state, \(\nu = 1/2\) is expected to be a compressible sea of composite fermions that does not form a quantum Hall state ). However, a possible explanation for the 0.5 \(\frac{e^2}{h}\) plateau could come from the counterpropagating charge mode as proposed in the Macdonald edge structure. Following the model of the \(\nu = 2/3\) state as a 1/3 state of holes on top of a vacuum of a fully filled \(\nu = 1\) state, the filling factor must increase from \(\nu = 2/3\) in the bulk to \(\nu = 1\) at the edge before decreasing to zero, shown schematically in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}a. The re-entrant density profile is enabled by the combination of the confining potential with the electrostatic interaction between electrons in the vacuum \(\nu = 1\) state and holes that form the inner-1/3 state. This results in an inner counterpropagating edge state and outer downstream edge state. In this case, the inner upstream edge state and outer downstream edge state on one edge of a Hall bar can have different chemical potentials because they are emitted from different Ohmic contacts. The situation is shown schematically in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}b. Assuming the edge modes have a small enough separation to allow charge transfer between them, the counterflowing edge modes will equilibrate and reach the same chemical potential. Based on the well-established experimental properties of quantum Hall states in bulk samples, the equilibrium chemical potential of each edge should be the potential of the contact from which the downstream, higher-conductance edge state emanated from. This leads to a potential difference \(V_{Hall} = \frac{\mu _{S}-\mu_{D}}{e}\), and a two-terminal conductance \(G = \frac{2}{3} \frac{e^2}{h}\), consistent with the observed bulk conductance properties of the 2/3 state. A quantum point contact which causes backscattering will take the edge states out of equilibrium and reduce the conductance of the device. Due to the somewhat more complicated nature of equilibration between counterflowing edge states compared to coflowing ones, this situation is less straightforward than the case of a QPC with all-downstream edge states. We analyze the situation of a bulk Hall bar and of a QPC when a counter-propagating edge mode is present. The conductance of a pair of counterflowing edge modes can be treated in a simple model where the current carried by each is equal to the mode's conductance \(G_i\) times its local chemical potential \(\mu_i (x)\) at its position along the edge (similar models were discussed in Refs. ), \(I_1 (x)=\mu_1(x)G_1\) and \(I_2(x)=\mu_2(x)G_2\). Equilibration (which has been studied theoretically ) is taken into account by allowing charge to be exchanged between the two modes and assuming a tunneling current per unit length between mode 1 and mode 2 \(dI_{12}/dx\) proportional to the chemical potential difference: \[\label{dI} d I_{12}=\frac{e^2}{h}\frac{\mu_2-\mu_1}{l_0}dx\] Here \(l_0\) is the characteristic equilibration length between the two edge modes. Conservation of current requires \(I = \mu_1 (x) G_1-\mu_2 (x) G_2 =\) constant. Eqn. [\[dI\]](#dI){reference-type="ref" reference="dI"} can be used to obtain differential equations for \(\mu_1\) and \(\mu_2\): \[\label{dmu1} \frac{d \mu_1}{dx}=\frac{(\mu_2-\mu_1)G_0}{G_1 l_0}\] \[\label{dmu2} \frac{d \mu_2}{dx}=\frac{(\mu_1-\mu_2)G_0}{G_2 l_0}\] \(G_0 \equiv \frac{e^2}{h}\) is the conductance quantum. Considering the bottom edge of the Hall bar in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}b where the inner edge emanates from the source contact, \(\mu_1 (0)=\mu_S\), and the outer edge from the drain contact, \(\mu_2 (L)=\mu_D\), the differential equations are straightforward to solve. They yield: \[\label{mu1x} \mu_1(x) = \frac{I}{G_1-G_2}+\left(\mu_S-\frac{I}{G_1-G_2} \right)e^{-\alpha x}\] \[\label{mu2x} \mu_2(x) = \frac{I}{G_1-G_2}+\frac{G_1}{G_2}\left(\mu_S-\frac{I}{G_1-G_2} \right)e^{-\alpha x}\] \[\label{Ix} I = \frac{(\mu_D-\mu_S \frac{G_1}{G_2}e^{-\alpha L})(G_1-G_2)}{1-\frac{G_1}{G_2}e^{-\alpha L}}\] Here \(\alpha\equiv \frac{G_0(G_2-G_1)}{l_0G_1G_2}\), and \(L\) is the length of the edge of the Hall bar. For large \(L\) (much larger than the characteristic equilibration length \(l_0\)) \(I\approx \mu_D (G_1-G_2)\), reflecting the fact that the edge states equilibrate at the chemical potential of the contact that the higher conductance outer edge state emanates from. The chemical potentials of the inner and outer edge states, \(\mu_1\) and \(\mu_2\), are qualitatively sketched as a function of length along the Hall bar in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}c. An equivalent relationship will hold on the top edge of the Hall bar so that the total current \(I_{total} = (\mu_S-\mu_D)(G_2-G_1) = \frac{2}{3}\frac{e^2}{h}(\mu_S-\mu_D)\), thus recovering the two-terminal conductance \(\frac{2}{3}\frac{e^2}{h}\) observed in macroscopic quantum Hall samples at \(\nu = 2/3\). Furthermore, in this limit \(\mu_2 (0) = \mu_D + \frac{G_1}{G_2}(\mu_S-\mu_D)\) and \(\mu_1 (L)=\mu_D\). This solution can be applied to the case of a QPC that fully transmits the outer mode and fully reflects the inner mode, as illustrated in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}d. The solutions to the differential equations can be applied to each quadrant of the Hall bar, allowing the chemical potential \(\mu_{QPC}^{BL}\) on the bottom left edge of the QPC and \(\mu_{QPC}^{UR}\) at the upper right edge of the QPC to be solved (note that on the top left and bottom right edges the inner and outer modes will not be at equilibrium due to the backscattering of the QPC, so there will be additional equilibration at those positions, as indicated by the white dashed lines). Applying the solutions to the differential equations in the limit \(L>>l_0\) gives the equations: \[\label{muBL} \mu_{QPC}^{BL}=\mu_{QPC}^{UR}\frac{G_1}{G_2}+\mu_D \left(1-\frac{G_1}{G_2}\right)\] \[\label{muUR} \mu_{QPC}^{UR}=\mu_{QPC}^{BL}\frac{G_1}{G_2}+\mu_S \left(1-\frac{G_1}{G_2} \right)\] Solving yields \(\mu_{QPC}^{BL}=\frac{\mu_D G_2+\mu_S G_1}{G_2+G_1}\) and \(\mu_{QPC}^{UR}=\frac{\mu_S G_2+\mu_D G_1}{G_2+G_1}\). This results in a total current: \[\label{Iqpc} I_{QPC}=(\mu_S-\mu_D)\frac{G_2-G_1}{1+\frac{G_1}{G_2}}\] For the case of \(G_1 = \frac{1}{3} \frac{e^2}{h}\) and \(G_2=\frac{e^2}{h}\) (as expected for the MacDonald edge structure), the result is \(G_{QPC} = \frac{1}{2} \frac{e^2}{h}\), matching the value of the intermediate plateaus observed in Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}b for \(V_{SD}\) = 50 \(\mu\)V. The chemical potential on the bottom edge in this situation is sketched in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}e. An abrupt discontinuity occurs for \(\mu_1\) in the middle of the Hallbar at the position of the QPC since at that point the inner mode is fully reflected. Although inner edge modes are usually assumed to be reflected before outer ones as more negative gate voltages are applied, at \(\nu = 2/3\) achieving full reflection of the inner edge state while maintaining full transmission of the outer one would seem to require an unusual situation where the filling factor in the middle of the QPC is \(\nu = 1\), higher than the bulk density of \(\nu = 2/3\). The Macdonald model arrives at a re-entrant density profile in which the density at the edge increases from the bulk filling factor value of 2/3 to \(\nu = 1\) due to the combined electrostatic potential of the vacuum state of the fully-filled N = 0 Landau level and the external confining potential as shown in Fig. [\[nu=2_3\_cartoons\]](#nu=2_3_cartoons){reference-type="ref" reference="nu=2_3_cartoons"}a. Full reflection of the inner mode would be achieved by moving the edges together via more negative gate voltage until the outer strips of \(\nu = 1\) fluid merge in the middle of the QPC so that the interface between \(\nu = 1\) and \(\nu = 2/3\) is absent inside the QPC; thus the inner mode cannot propagate through. This would be a somewhat unusual situation since QPCs are usually considered to be regions of lower density than the bulk. The lack of a fully quantized plateau at zero bias might indicate that the electrostatics in the QPC do not enable a region reaching full \(\nu = 1\) filling inside the QPC region once the inner mode is fully reflected, but rather the filling is somewhat lower, resulting in a small amount of simultaneous backscattering of the outer mode and \(G< 0.5 \frac{e^2}{h}\) as observed in Fig. [\[nu=2_3\_QPCs\]](#nu=2_3_QPCs){reference-type="ref" reference="nu=2_3_QPCs"}b after the initial sharp drop in conductance. Disorder-mediated tunneling may also be a contributing factor. Finite bias tends to reduce backscattering in Luttinger liquids, so the emergence of the quantized intermediate plateau at \(V_{SD}= 50\) \(\mu\)V could be due to the suppression of backscattering for the outer mode in this regime. A specific prediction of Luttinger liquid theory is that when \(V_{SD}\) is increased, transmission of the edge state through a barrier will increase until it reaches full transmission, which agrees well with our observation that after a small \(V_{SD}\) is applied the conductance increases to the half-integer quantized value indicating full transmission of the outer edge state, and then remains nearly constant. Although integer quantum Hall edge states are generally expected to exhibit Fermi liquid rather than Luttinger liquid behavior based on theory, when an integer edge state lies parallel to a co-propagating or counter-propagating edge state this may induce Luttinger liquid behavior. Additionally, non-linear conductance suggestive of Luttinger liquid behavior has been observed for the \(\nu =1\) integer quantum Hall state in QPCs, suggesting that in real systems even isolated integer edge states may have an interaction parameter that is not exactly 1 and may thus diverge from ideal Fermi liquid behavior. An assumption for applying this model to our device is that the equilibration length \(l_0\) is much smaller than the length from the QPC to the source and drain Ohmic contacts. In our devices, this length is \(\approx 500\) \(\mu\)m. Measurements in which both QPCs are tuned to backscatter the inner edge indicate that the edges partially, but not fully, equilibrate on the 800 nm length separating the two QPCs (see Supp. Section 3 and Supp. Fig 3), which makes it very likely that on the much longer length scale from the QPC to the Ohmic contacts, the edges will indeed fully equilibrate. Equilibration at \(\nu = 2/3\) has been investigated in Hall bar geometries in GaAs. Lin *et al*. found an equilibration length of \(\approx 8\) \(\mu\)m at \(B = 6.5\) T. Maiti *et al*. found an equilibration length which depended sharply on magnetic field, varying from \(\approx 50\) \(\mu\)m at \(B=8.8\) T to \(\approx 800\) \(\mu\)m at \(B = 10.8\) T; since our magnetic field of 5.9 T is much smaller, it is reasonable that the equilibration length should be smaller. Notably, however, is not clear what the edge structure was in these previous experiments given that they used different types of GaAs/AlGaAs heterostructures than the one used in our experiment, so the equlibration length scales may not be directly comparable. Measurements of the \(\nu = 2/3\) state in graphene have shown evidence for upstream transport, but charge equilibration appears to occur on short length scales. To our knowledge half integer quantized conductance in a QPC at \(\nu = 2/3\) has not previously been reported; instead, an intermediate plateau at \(G = \frac{1}{3} \frac{e^2}{h}\) is usually observed. This may reflect the fact that modulation doped GaAs/AlGaAs heterostructures used for quantum Hall experiments usually have shallow confining potentials due to the large setback of the doping layers from the quantum well and large distance of the surface gates from the quantum well (typically on the order of 100 nm). Such shallow confining potentials may be likely to place devices in the regime described by Meir in which there are two downstream edge modes with conductance \(\frac{1}{3} \frac{e^2}{h}\). On the other hand, in a few experiments a conductance plateau with \(G = 1.5 \frac{e^2}{h}\) has been observed in QPCs at \(\nu = 5/3\), which is the equivalent of the \(\nu = 2/3\) state with opposite spin. It is noteworthy that these experiments used techniques likely to increase the density of the 2DEG in the QPC. In Ref. and Ref., doping well heterostructures were used, and gate biases were applied at an elevated temperature of 4 K and 6 K, allowing electrons in the X-band of the AlAs layers next to the doping wells to equilibrate with the 2DEG and partially screen the gate voltage. The \(G = 1.5 \frac{e^2}{h}\) plateau was then observed when the gate voltage was increased above the annealed-in value. This unusual decrease in conductance from \(\frac{5}{3} \frac{e^2}{h}\) to \(\frac{3}{2} \frac{e^2}{h}\) suggests that the half-integer plateau occurred when the density in the center of the QPC was increased above the bulk density due to increasing of the gate voltage after annealing. In Ref., an additional gate in the middle of the device had a positive bias applied, which would be expected to yield higher electron density in the middle of the QPC. A possible explanation for the \(G = \frac{3}{2} \frac{e^2}{h}\) plateau is that the techniques of Refs. forced the density in the center of the QPC to be higher than the bulk by their external gate voltages, thus yielding \(\nu = 2\) in the QPC, which with full equilibration of the additional edge modes in the QPC results in quantized conductance \(G = \frac{3}{2} \frac{e^2}{h}\). Similar to the QPC gate voltage dependence in Ref., Ref. shows a non-monotonic dependence on the center gate voltage, supporting the possibility that the 2DEG density in the middle of the QPC is being forced above the bulk value by the external gate potential. A notable difference in our experiment compared to these previous works is that in our device, the conductance increases nearly monotonically with gate voltage, with the primary \(G = \frac{2}{3} \frac{e^2}{h}\) plateau occurring at higher gate voltage compared the \(G = 0.5 \frac{e^2}{h}\) plateau, suggesting that the lower-conductance intermediate plateau does not occur because the externally applied gate voltages have forced the density in the QPC to be higher, but because the intrinsic interactions of the electrons favor the formation of a \(\nu = 1\) strip and re-entrant density profile at the edge. Similar to the previous experiments, we also observe an intermediate plateau at \(G = 1.5 \frac{e^2}{h}\) at \(\nu = 5/3\) in both QPCs (Supp. Fig. 4a and Supp. Section 4), which suggests that an edge structure occurs at \(\nu =5/3\) similar to the one at \(\nu = 2/3\) in our device. On the other hand, at the particle-like fractional quantum Hall states \(\nu = 4/3\) and \(\nu = 2/5\), we do not observe anomalous intermediate plateaus that would imply a re-entrant density profile, which is consistent with the expectation that the \(\nu=1\) strip and re-entrant density profile occur only for hole-like fractional quantum Hall states. This further distinguishes our work from Ref., in which anomalous plateaus were observed even for particle-like states such as \(\nu = 4/3\). It is also noteworthy that evidence for counter-propagating charge transport at short distances has been observed at \(\nu = 2/3\), but only in the spin-unpolarized state. Due to the fact that our 2DEG has relatively high density and uses a wide, symmetric quantum well, the 2/3 state is very likely polarized in our system. Notably, numerical studies in the composite fermion picture predict counterpropagating modes for both the spin polarized and unpolarized \(\nu = 2/3\) states. The interplay of spin polarization and confining potential at \(\nu = 2/3\) requires more investigation.
# Comparison to a QPC on a non-screening well structure
In order to confirm that the use of the screening well heterostructure enables the observation of the \(G = \frac{1}{2} \frac{e^2}{h}\) plateau, we have also measured a QPC fabricated on a standard non-screening well heterostructure. The standard heterostructure uses a single GaAs/AlGaAs interface rather than a quantum well to confine the 2DEG. Fig. [\[Non_SW\]](#Non_SW){reference-type="ref" reference="Non_SW"}a shows the standard structure. Fig. [\[Non_SW\]](#Non_SW){reference-type="ref" reference="Non_SW"}b shows conductance versus gate voltage at \(\nu = 1\) (note that the range of gate voltage is smaller because the gate potential is not screened by the screening layers). In contrast to the wide plateau and very sharp pinchoff in the screening well QPCs, this standard structure QPC shows a narrow primary plateau with \(G = \frac{e^2}{h}\) and a much more gradual decrease in conductance. At \(\nu = 1\) there also appears to be an intermediate quasi-plateau at \(G \approx \frac{1}{3} \frac{e^2}{h}\), consistent with the phenomenon of edge reconstruction reported in. Fig. [\[Non_SW\]](#Non_SW){reference-type="ref" reference="Non_SW"}c shows conductance versus gate voltage at \(\nu = 2/3\), where a clear intermediate plateau at \(G = \frac{1}{3} \frac{e^2}{h}\) is present, consistent with previous results using non-screening well heterostructures. These results support the theoretical picture that a soft confining potential (as usually realized in standard GaAs/AlGaAs heterostructures not utilizing screening layers) results in a reconstructed edge at \(\nu = 2/3\) with an outer strip of filling \(\nu = \frac{1}{3}\) and two downstream \(\frac{1}{3}\frac{e^2}{h}\) charge modes, as described by the Meir edge structure. The absence of this \(G = \frac{1}{3} \frac{e^2}{h}\) plateau in the QPCs with sharp confinement indicates that the outer \(\nu = \frac{1}{3}\) strip does not form, and thus a different edge structure is present. This demonstrates the profound effect that confining potential can have on edge structure in the quantum Hall regime.
# Conclusions
We have measured conductance across QPCs at the \(\nu = 2/3\) fractional quantum Hall state. When a small finite source-drain bias is applied, both QPCs fabricated on the screening well heterostructure exhibit an intermediate conductance plateau with \(G\approx 0.5 \frac{e^2}{h}\); this plateau is robust over a moderate range of magnetic field, gate voltage, and source drain bias. A simple model taking into account equilibration between counterflowing edge modes suggests that this plateau may be explained in terms of the Macdonald edge structure as being due to full transmission of the outer downstream integer edge mode and full reflection of the inner fractional edge mode. Measurements of a QPC fabricated on a standard, non-screening well heterostructure reveal instead an intermediate conductance plateau with \(G = \frac{1}{3}\frac{e^2}{h}\), which suggests that in this type of structure there is an outer strip with filling factor \(\nu = 1/3\) when the bulk is at filling factor \(\nu = 2/3\), as occurs in the Meir edge structure. This suggests that, as predicted by theory, a transition from the Macdonald edge structure to a reconstructed edge given by the Meir edge structure occurs when the confining potential changes from sharp to soft. Our experiments clearly show the importance of confining potential on edge structure in the quantum Hall regime. The lack of edge reconstruction may explain why it has been possible to observe Aharonov-Bohm interference interference at quantum Hall states such as \(\nu = 1\) and \(\nu = 1/3\) in devices using the screening well heterostructure, but not in devices using standard non-SW heterostructures.
# Supplementary Section 1: Screening Well Heterostructure
The main device used in this work uses the screening well heterostructure shown in Fig. [\[SW\]](#SW){reference-type="ref" reference="SW"}. The structure is grown by molecular beam epitaxy. Similar structures were used in previous works. The device includes additional gates on the surface and on the back side of the chip which go around the Ohmic contacts, and these gates are biased to just deplete the electrons from the screening wells so that the Ohmic contacts are disconnected from the screening wells and no parallel conduction occurs in the magnetotransport.
# Supplementary Section 2: Replication in a second set of QPCs
We have measured QPC conductance in a second set of QPCs fabricated on a separate chip on the same screening well wafer at \(\nu = 2/3\). Similar to the QPCs discussed in the main text, these QPCs both exhibit a conductance plateau at \(G = 0.5 \frac{e^2}{h}\) as shown in Supp. Fig. [\[repeat\]](#repeat){reference-type="ref" reference="repeat"}.
# Supplementary Section 3: Equilibration in double QPC measurement
The fact that the bulk Hall conductance in our experiment matches the equilibrated value of \(\frac{2}{3} \frac{e^2}{h}\) suggests that the characteristic equilibration length \(l_0\) is much smaller than the separation of Ohmic contacts on our Hall bar, which is approximately 500 \(\mu\)m. An important consideration for the operation of interferometers is whether scattering between different edge modes can occur. If there is no scattering between edge states inside the interferometer, then when the inner edge is fully reflected by the QPC 1 while the outer edge is fully transmitted the overall conductance across the device should not change as the voltage on the QPC 2 is varied until it reaches the point of backscattering the outer edge state. On the other hand, if there is significant scattering between the edge states such that they fully equilibrate on the length between the QPCs, then the total transmission across the device will be the product of the individual QPC transmissions. With both QPCs tuned to the \(G = 0.5 \frac{e^2}{h}\) plateaus (transmission \(T = \frac{0.5}{2/3}=0.75\)), as sketched in Fig. [\[nu=2_3\_double\]](#nu=2_3_double){reference-type="ref" reference="nu=2_3_double"}a, the resulting overall conductance would be \(0.5 \frac{e^2}{h}\) in the case of no inter-edge scattering and \(0.375 \frac{e^2}{h}\) in the case of full equilibration. We evaluate inter-edge equilibration inside the device by measuring conductance across the device when both QPCs are tuned to backscatter the edge states. Fig. [\[nu=2_3\_double\]](#nu=2_3_double){reference-type="ref" reference="nu=2_3_double"}b shows conductance at zero applied bias as a function of the voltage on both QPCs. Here, the side gate are set to-1.5 V, so the 2DEG undernath them is depleted and there is an approximately 800 nm path length between the two QPCs. As in the individual QPC sweeps, a quantized intermediate plateau is absent, although in the region where the conductance bump occurs (\(V_{QPC 1}\approx\)-2.2 V and \(V_{QPC 2}\approx\)-2.15 V), the conductance is weakly dependent on gate voltage, forming a quasi-plateau. Fig. [\[nu=2_3\_double\]](#nu=2_3_double){reference-type="ref" reference="nu=2_3_double"}c shows differential conductance versus QPC voltages with 50 \(\mu\)V source-drain bias. A clear intermediate 0.5 \(\frac{e^2}{h}\) plateau is observed in both QPCs, as expected based on the individual QPC sweeps in the main text Fig. 2. In Fig. [\[nu=2_3\_double\]](#nu=2_3_double){reference-type="ref" reference="nu=2_3_double"}d we show line cuts from the colormap plot in Fig. [\[nu=2_3\_double\]](#nu=2_3_double){reference-type="ref" reference="nu=2_3_double"}c at different values of the \(V_{QPC 2}\). At \(V_{QPC 2} =-1.4\) V, there is full transmission through QPC 2, and the conductance vs. \(V_{QPC 1}\) shows a plateau at \(G = \frac{2}{3} \frac{e^2}{h}\) and an intermediate plateau at \(G = 0.5 \frac{e^2}{h}\). At \(V_{QPC 2}\) =-2.13 V and-2.19 V, QPC 2 is in the region of the \(G = 0.5 \frac{e^2}{h}\) plateau. In both of these line cuts the conductance exhibits a plateau at \(G\approx 0.5 \frac{e^2}{h}\) for \(V_{QPC 1} >-2.1\) V (although the conductance is slightly lower for-2.19 V, reflecting the fact that the intermediate plateau in QPC 2 is not completely flat). This plateau persists until the region of the \(0.5 \frac{e^2}{h}\) in QPC 1 (highlighted yellow region), at which point the conductance drops. The value of the conductance in this range is intermediate between the \(G = 0.5 \frac{e^2}{h}\) limit of zero inter-edge scattering and the \(G = 0.375 \frac{e^2}{h}\) limit of full equilibration, suggesting that the edge partially, but not fully, equilibrate along the length of the interferometer. This situation may complicate interference experiments which attempt to interfere either the outer or inner edge states. A line cut at \(V_{QPC 2} =-2.28\) V is also shown; this cut is along the region where the QPC begins fully pinching off, reflecting the backscattering of the outer edge mode, and it shows a gradual decrease in conductance as a function of \(V_{QPC 1}\) before fully pinching off. This may be due to a slight cross-coupling between the QPCs, since in this region the conductance of QPC 2 is very sensitive to changes in potential. In Ref, an equilibration length of approximately 2 \(\mu\)m was found for the counter-propagating edge mode in the spin-unpolarized state. In Ref., the edge structure appeared to consist of two downstream charge modes with counterpropagating neutral modes, and equilibration between the charge modes occurred over lengths of a few \(\mu\)m. Although our edge structure may be different from these previous experiments due to the possibility of a counterpropagating inner mode in the spin-polarized state, it appears that the effective equilibration length is comparable since the edge modes only partially equilibrate over the edge length of \(\approx 800\) nm in our device. In another experiment conducted at much higher magnetic fields of \(B\approx 14\) T and in a sample appearing to support the Meir edge structure of two downstream charge modes, a much larger equilibration length of \(l_0 = 104 \mu\)m for these two co-propagating edge modes was measured. This may suggest that the short magnetic lengths and large Coulomb energy scales at high magnetic field work to reduce inter-edge scattering.
# Supplementary Section 4: QPC sweeps at other filling factors
Similar to the previous experiments, we also observe an intermediate plateau at \(G = 1.5 \frac{e^2}{h}\) at \(\nu = 5/3\) in both QPCs (Fig. [\[Other_QPC\]](#Other_QPC){reference-type="ref" reference="Other_QPC"}a), which suggests that a similar edge structure occurs at \(\nu =5/3\) to the one at \(\nu = 2/3\) in our device. Notably, this plateau does not require finite DC bias to be observed, and instead occurs at \(V_{SD}= 0\). Surprisingly, at \(\nu = 5/3\) there is not an intermediate plateau at \(G = \frac{e^2}{h}\) which would be expected to occur when only the outermost edge mode is fully transmitted, even though there are \(G = \frac{e^2}{h}\) intermediate plateaus at the nearby states \(\nu = 4/3\) and \(\nu = 5/3\), shown in Fig. [\[Other_QPC\]](#Other_QPC){reference-type="ref" reference="Other_QPC"}b and c. While there does appear to be a rough second intermediate plateau at \(\nu = 5/3\), it occurs at conductance somewhat less than \(\frac{e^2}{h}\), and the value is somewhat different for QPC 1 and QPC 2. There may non-trivial effects of spin leading to greater scattering between inner and outer edge states than would be naively expected. In Fig. [\[Other_QPC\]](#Other_QPC){reference-type="ref" reference="Other_QPC"}b, c, and d we show QPC sweeps at \(\nu = 4/3\), \(\nu = 2\), and \(\nu = 2/5\). Notably, at \(\nu = 4/3\) there is a primary plateau at \(G = \frac{4}{3} \frac{e^2}{h}\) and an intermediate plateau at \(G = \frac{e^2}{h}\), in line with expectations of an inner downstream 1/3 edge mode and outer integer mode. At \(\nu = 2/5\), there is a primary plateau at \(G = \frac{2}{5} \frac{e^2}{h}\) and intermediate plateau at \(G = \frac{1}{3} \frac{e^2}{h}\), consistent with expectations of an inner downstream 1/15 mode (consisting of \(e/5\) quasiparticles) and an outer 1/3 mode. The fact that these particle-like fractional quantum Hall states exhibit conductance plateaus quantized to conventional fractions (in contrast to the anomalous \(G = \frac{1}{2} \frac{e^2}{h}\) plateau observed at the hole-like \(\nu=2/3\) state) is consistent with the expectation that only hole-like states have the re-entrant density profile with a strip of \(\nu = 1\) filling at the outer edge. | {'timestamp': '2022-12-01T02:21:17', 'yymm': '2211', 'arxiv_id': '2211.17240', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17240'} |
null | null |
null | null |
# Introduction
Quantum theory reveals interesting and counter-intuitive phenomena in even the simplest physical systems. Paradigmatic examples are Bell nonlocality and Einstein-Podolsky-Rosen (EPR) steering . These nonlocal phenomena appear as strong correlations between the outcomes of spatially separated measurements performed by independent observers. These correlations enable us to distinguish the classical and quantum origins of the experiments. Recently, a similar split between classical and quantum features was found in a setup closely related to quantum communication tasks, the so-called prepare-and-measure (PM) scenario . This scenario can be viewed as a communication game between two parties, Alice (the sender) and Bob (the receiver), where the dimension of the classical (versus quantum) system communicated from Alice to Bob is bounded from above. The PM game is described as follows (see panel (a) of Fig. [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"}). Upon receiving an input \(x=(1,\ldots,n)\), a preparation device (controlled by Alice) emits a physical system in a quantum state \(\rho_x\). We assume \(\rho_x\in\mathcal L(\mathbb{C}^d)\) for a given \(d\ge 2\). In the following, however, we will focus explicitly on \(d=2\), that is, we assume that two-dimensional quantum systems (qubits) or classical systems (bits) are transmitted from Alice to Bob. The state \(\rho_x\) is passed to a measurement device which, upon receiving an input \(y=(1,\ldots,m)\) performs a measurement and obtains an outcome \(b=(1,\ldots,o)\). In this paper we will focus on the smallest, nontrivial case of \(o=2\), i.e., measurements with two outcomes, in which case we denote the outcomes by \(b=\pm 1\). Our goal in this scenario is to compare the power of qubits with classical bits. This scenario has been discussed in some extent for a small number of preparations \(n\) and measurements \(m\) (see e.g. Refs. . However, in this paper we would like to turn our attention to the case of large \(n\) and \(m\) (i.e. in the range of 70). We will see that the main bottleneck of the study is the computation of the relevant quantities associated with the classical bit case for which we develop large scale numerical tools in this paper. We first concentrate on the qubit case, and then we will elaborate on the classical bit case.
*Qubit case.*---In the qubit binary outcome (\(o=2\)) case, the measurement is described by two positive operators \(\{\Pi_{b|y}\}\), \(b=\pm 1\) acting on \(\mathbb{C}^2\) which sum to the identity \(\Pi_{b=+1|y}+\Pi_{b=-1|y}=\leavevmode\hbox{\small1\normalsize\kern-.33em1}\) for each \(y\), where \(\leavevmode\hbox{\small1\normalsize\kern-.33em1}\) denotes the \(2\times 2\) identity matrix. The statistics of the experiment are then given by the formula \[P(b|x,y)=\operatorname{Tr}(\rho_x \Pi_{b|y}). \label{born}\] It is important to note that both the preparations and the measurements are unknown to the observer, up to the fact that the dimension of the transmitted system is two. Since we have binary outcomes \(b=\{+1,-1\}\) it becomes convenient to use expectation values \[E_{x,y}=P(b=+1|x,y)-P(b=-1|x,y). \label{Exy}\] Note that \(E_{x,y}\) can take up the values in \(\left[-1,+1\right]\) for all \(x,y\). However, if the Hilbert space dimension of the communicated particle is bounded, then in general not all expectation values \(E_{x,y}\) in \(\left[-1,+1\right]\) become possible. The simplest scenario that shows this effect appears already for \(n=3\), \(m=2\) and \(o=2\) (see for an example). With respect to the measurement operators \(M_{b|y}\), one case, namely the set of projective rank-1 measurements, is of particular interest to us. In this case, we have \[\Pi_{b|y}=\frac{\leavevmode\hbox{\small1\normalsize\kern-.33em1}_2+b\vec b_y\cdot\vec\sigma}{2},\] where \(\vec b_y\in S^2\), \(b=\pm 1\) and \(\vec\sigma=(\sigma_x,\sigma_y,\sigma_z)\) is the vector of Hermitian \(2\times 2\) Pauli matrices. On the other hand, let us set \[\rho_x = \frac{\leavevmode\hbox{\small1\normalsize\kern-.33em1}+\vec a_x\cdot\vec\sigma}{2},\] where \(\vec a_x\in S^2\). This density matrix corresponds to a pure state with Bloch vector \(\vec a_x\). Note that in this case, the above equations give us \[\label{Exydot} E_{x,y}=\vec a_x\cdot\vec b_y,\] where \(\vec a_x,\vec b_y\in S^2\). Limits on the set of possible distributions in dimension two can be captured by the following expression \[W=\sum_{x=1}^{n}\sum_{y=1}^{m} M_{x,y}E_{x,y}, \label{dimwit}\] where \(M_{x,y}\) are coefficients of a real witness matrix \(M\) of dimension \(n\times m\). Let us then define the quantity \[Q(M)=\max\sum_{x=1}^{n}\sum_{y=1}^{m}M_{x,y}E_{x,y}, \label{QM}\] where \(E_{x,y}\) is of the form ([\[Exy\]](#Exy){reference-type="ref" reference="Exy"}), and where we maximize the expression over Bob's measurements \(\{M_{b|y}\}\) and the qubit state \(\rho_x\) in Eq. ([\[born\]](#born){reference-type="ref" reference="born"}). Thus, \(Q(M)\) is the value that is achievable with the most general two-dimensional quantum resources in our PM setup. We further define the quantity \[q(M)=\max\sum_{x=1}^{n}\sum_{y=1}^{m}M_{x,y}E_{x,y}, \label{qM}\] where \(E_{x,y}=\vec a_x\cdot\vec b_y\) and we maximize over the unit vectors \(\vec a_x\) and \(\vec b_y\) in the three-dimensional Euclidean space. It turns out that \(Q(M)\) can be obtained with pure qubit states and projective measurements . However, the optimal projective measurements are in general not of rank-1, they can be of rank-2 as well. Indeed, there are example matrices \(M\) (even in the simple \(n=m=3\), \(o=2\) case) for which \(Q(M)>q(M)\). Note that \(q(M)\) corresponds to projective qubit measurements of rank 1, in which case \(E_{x,y}=\vec a_x\cdot\vec b_y\) (see equation ([\[Exydot\]](#Exydot){reference-type="ref" reference="Exydot"})). Yet, as we will see, the set \(\{E_{x,y}\}_{x,y}\) obtained by rank-1 projective measurements is a significant subset of the set \(\{E_{x,y}\}_{x,y}\) corresponding to the most general qubit measurements. The tools for computing the value \(Q(M)\) can be found in Refs. . Importantly, the value of \(Q(M)\) can serve as a dimension witness in the prepare-and-measure scenario . Indeed, if \(W>Q(M)\) for some \(M\) (where the witness \(W\) is defined by equation ([\[dimwit\]](#dimwit){reference-type="ref" reference="dimwit"})), this implies that the set of states \(\{\rho_x\}_{x=1}^{n}\) transmitted to Bob must have contained at least one state \(\rho_{x=x'}\) of at least three dimensions (that is qutrit). *Classical bit versus qubit case.*---It turns out that the witness \(W\) can also serve as a quantumness witness. To this end, let us discuss the classical bit case. That is, we want to bound the expression ([\[dimwit\]](#dimwit){reference-type="ref" reference="dimwit"}) if Alice can only prepare classical two-dimensional systems (i.e. bits). Let us denote the bound on ([\[dimwit\]](#dimwit){reference-type="ref" reference="dimwit"}) by \(L_2(M)\), which corresponds to this situation. If \(W>L_2(M)\), this certifies that some of the measurements performed by Bob are true (incompatible) quantum measurements acting on true qubit states . Mathematically, the classical bit case is equivalent to the qubit case discussed above, with the restriction that all qubits are sent in the same basis, and all measurements of Bob are carried out in the very same basis. That is, if we want to maximize ([\[dimwit\]](#dimwit){reference-type="ref" reference="dimwit"}) for correlations \(E_{x,y}\) arising from classical two-dimensional systems, the maximum can be attained with pure states \[\rho_x = \frac{\leavevmode\hbox{\small1\normalsize\kern-.33em1}_2+a_x\cdot\sigma_z}{2}, \label{rhoxbit}\] where \(a_x=\pm 1\), and observables \(B_y=\Pi_{0|y}-\Pi_{1|y}\) which have the form \[B_y=b_y^+\ket{0}\bra{0}+b_y^-\ket{1}\bra{1}, \label{Bybit}\] where \(\sigma_z\) is the standard Pauli matrix \[\begin{aligned}
\sigma_z=\begin{pmatrix*}[r] 1 & 0 \\ 0&-1 \end{pmatrix*} \end{aligned}\] and both \(b_y^+\), \(b_y^-\) are \(\pm 1\) variables. Inserting these values into ([\[Exy\]](#Exy){reference-type="ref" reference="Exy"}) we obtain \[E_{x,y} = \frac{(1+a_x)b_y^++(1-a_x)b_y^-}{2}. \label{Exybit}\] Since we have binary variables \(a_x=\pm 1\), they translate to \(E_{x,y}=b_y^+\) if \(a_x=1\) and \(E_{x,y}=b_y^-\) if \(a_x=-1\). Then the classical one-bit bound \(L_2(M)\) is given by \[L_2(M)=\max\sum_{x=1}^{n}\sum_{y=1}^{m}M_{x,y}E_{x,y}, \label{L2M}\] where \(E_{x,y}\) is defined by ([\[Exybit\]](#Exybit){reference-type="ref" reference="Exybit"}) and we maximize over all binary variables \(a_x, b^+_y,b^-_y\in\{-1,+1\}\). In words, the expression ([\[Exybit\]](#Exybit){reference-type="ref" reference="Exybit"}) corresponds to the following deterministic protocol. Alice, depending on \(x\), prepares a bit \(a_x=\pm 1\), which she sends to Bob, who outputs \(b=\pm 1\) depending on the value of \(a_x\) and the measurement setting \(y\). That is, Bob's output is a deterministic function \(b=f(a_x,y)\), where the output assumes \(b=\pm 1\). We can write \[\begin{aligned}
&L_2(M)\nonumber\\ &=\max \left(\sum_{x: a_x=+1}\sum_{y=1}^{m} M_{xy}b^+_y+\sum_{x: a_x=-1}\sum_{y=1}^{m} M_{xy}b^-_y\right), \label{L2Mv2} \end{aligned}\] where the maximum is taken over all binary \(a_x\), \(b_y^+\) and \(b_y^-\) variables \(\pm 1\). We can eliminate the variables \(b^+_y\) and \(b^-_y\) from the above expression and get the following formula for \(L_2(M)\): \[L_2(M) = \max_{a_x=\pm 1} \left(\left\|\sum_{x: a_x=+1} M_x\right\|_1+\left\|\sum_{x: a_x=-1} M_x\right\|_1\right), \label{L2Mv3}\] which only consists of maximization over the binary variables \(a_x=\pm 1\). In the above formula, \(M_x\) denotes the \(x\)th row of the real \(n\times m\) matrix \(M\), where \(\|v\|_1\) denotes the Manhattan norm of the real vector \(v\), i.e., \(\|v\|_1=\sum_x |v_x|\). We prove several interesting properties of \(L_2(M)\) in the Methods section [3.1](#sec:L2proofs){reference-type="ref" reference="sec:L2proofs"}. In particular, \(L_2\) is proven to be a matrix norm. Let us recall that \(L_2(M)\) is a key quantity in our study, as it enables witnessing both quantumness of preparations and quantumness of measurements. Indeed, \(W>L_2(M)\), where \(W\) is defined in equation ([\[dimwit\]](#dimwit){reference-type="ref" reference="dimwit"}), certifies incompatible quantum measurements acting on true qubit states. That is, not all the performed measurements and not all prepared states originate from the same basis . In section [3.1](#sec:L2proofs){reference-type="ref" reference="sec:L2proofs"} we do not restrict our study to the properties of the \(L_2\) norm but generalize \(L_2(M)\) to \(L_k(M)\) for any \(k>2\) and prove that \(L_k\) is a norm as well, moreover \(L_k(M)\) is a monotonic increasing function of \(k\). Furthermore, in section [3.2](#sec:L2tips){reference-type="ref" reference="sec:L2tips"} we give tips for an efficient implementation of the branch-and-bound algorithm for computing the \(L_k(M)\) bound for \(k=2\) and for \(k>2\) as well. *Introducing the constants \(K_{\text{PM}}\) and \(K_{\text{D}}\).*---We define two quantities \(K_{\text{PM}}\) and \(K_{\text{D}}\) which are related to \(L_2(M)\) and \(q(M)\), and are defined as follows. Let us first introduce \(K_{\text{PM}}\), in which case we ask for the maximum ratio between \(q(M)\) and \(L_2(M)\). That is, we are interested in the value \[K_{\text{PM}}=\max_M \frac{q(M)}{L_2(M)}, \label{KPM}\] where the maximization is taken over all possible real \(n\times m\) matrices \(M\), where \(q(M)\) is defined by ([\[qM\]](#qM){reference-type="ref" reference="qM"}) and \(L_2(M)\) is defined by ([\[L2M\]](#L2M){reference-type="ref" reference="L2M"}). Let us now recall the Grothendieck constant of order 3 , which is given by \[K_G(3)=\max_M \frac{q(M)}{L(M)}, \label{KG3}\] where the maximization is taken over real matrices \(M\) of arbitrary dimensions \(n\times m\), \(q(M)\) is defined by ([\[qM\]](#qM){reference-type="ref" reference="qM"}) and \(L(M)\) is defined as follows \[L(M)=\max \sum_{x=1}^{n}\sum_{y=1}^{m} M_{x,y}a_xb_y, \label{LMdef}\] where the maximum is taken over all \(a_x, b_y\in\{-1,+1\}\). According to the literature, the value of \(K_G(3)\) in ([\[KG3\]](#KG3){reference-type="ref" reference="KG3"}) is bounded by \[\label{KG3bounds} 1.4359870\le K_G(3)\le 1.4643558,\] where the lower bound is due to Ref. and the upper bound is due to Ref. . Note that the upper bound is an improved version of the bound in Ref. . We prove that \(K_{\text{PM}}=K_G(3)\), which will be given in the Results section [2.1](#sec:KPMeqKG3){reference-type="ref" reference="sec:KPMeqKG3"}. We are interested in \(K_{\text{D}}\) as well, a quantity similar to \(K_{\text{PM}}\). We define this quantity as follows \[K_{\text{D}}=\max_M \frac{q(M)-S(M)}{L_2(M)-S(M)}, \label{KD}\] where \[S(M)=\sum_{x=1}^n\sum_{y=1}^m M_{x,y}. \label{SM}\] Note the relation \[\frac{q(M)-S(M)}{L_2(M)-S(M)}\ge \frac{q(M)}{L_2(M)},\] whenever \(L_2(M)>S(M)\) (also note that \(q(M)\ge L_2(M)\)), therefore we have \(K_{\text{D}}\ge K_{\text{PM}}=K_G(3)\). From this we immediately obtain the lower bound \(K_{\text{D}}\ge 1.4359\). In this paper, we give efficient large-scale numerical methods to obtain even better lower bounds on the above quantity. Namely, we prove the lower bound \(K_{\text{D}}\ge 1.5682\). We also prove an upper bound of 2 on this quantity, so putting all together we have the following interval \[1.5682\le K_{\text{D}}\le 2\] for the constant \(K_{\text{D}}\). It is an open problem to close or at least reduce the gap between the lower and upper limits. We next present the Results section, which contains our main findings in three subsections.
# Results {#sec:results}
## Proof of the relation \(K_{\text{PM}}=K_G(3)\) {#sec:KPMeqKG3}
To prove our claim, we relate \(L(M')\) to \(L_2(M')\), where \(M'\) is given by the following matrix (see also ([\[Mv\]](#Mv){reference-type="ref" reference="Mv"})) \[\begin{aligned}
M'=\begin{pmatrix*} &M \\ -&M \end{pmatrix*}, \label{Mvv} \end{aligned}\] where \(M\) is a real \(n\times m\) matrix. Denote by \(M_x\) the \(x\)-th row of the matrix \(M\). Note that according to the above definition \(M'\) has size \(2n\times m\) and \(M'\) has rows such that \(M'_x=M_x\) and \(M'_{x+n}=-M_{x}\) for all \(x=1,\ldots,n\). Then the following lemma holds.
The proof of this lemma is given in Methods section [3.3](#sec:LL2Mp){reference-type="ref" reference="sec:LL2Mp"}. Then we need to prove the following lemma.
For an arbitrary matrix \(M\), we have \(L_2(M)\ge L(M)\). This has been proved in Methods section [3.1](#sec:L2proofs){reference-type="ref" reference="sec:L2proofs"}. Then the lemma follows from the definitions ([\[KG3\]](#KG3){reference-type="ref" reference="KG3"}) and ([\[KPM\]](#KPM){reference-type="ref" reference="KPM"}). Our next lemma reads
*Proof.* To prove this, it suffices to show that for an arbitrary real matrix \(M\), there exists the matrix \(M'\) defined by ([\[Mvv\]](#Mvv){reference-type="ref" reference="Mvv"}) such that \(q(M')=2q(M)\) and \(L_2(M')=2L(M)\). The first relation follows from the special structure of \(M'\). The second relation has been shown in Lemma [\[L2L\]](#L2L){reference-type="ref" reference="L2L"}. Therefore, \(K_{\text{PM}}\) cannot be less than \(K_G(3)\), which proves our claim. \(\qed\)
Hence we have the same bounds \(1.4359\le K_{\text{PM}}\le 1.4644\) as for \(K_G(3)\) (see ([\[KG3bounds\]](#KG3bounds){reference-type="ref" reference="KG3bounds"})). From the corollary above, we have a matrix \(M'\) of size \(48\times 24\) with \(q(M')/L_2(M')>\sqrt 2\). Indeed, the construction is based on a matrix \(M\) of size \(24\times 24\), which provides \(q(M)/L(M)>\sqrt 2\) . To the best of our knowledge, this is the smallest \(M\) matrix that has the property \(q(M)/L(M)>\sqrt 2\). Then the \(48\)-by-\(24\) matrix \(M'\) follows from ([\[Mvv\]](#Mvv){reference-type="ref" reference="Mvv"}). On the other hand, \(q(M)/L(M)=\sqrt 2\) is already attained with a \(2\times 2\) matrix \(M\) in the CHSH-form : \[\begin{aligned}
M=\begin{pmatrix*}[r] 1 & 1 \\ 1&-1 \end{pmatrix*}. \end{aligned}\] It remains an open question to show that \(K_{\text{PM}}>\sqrt 2\) with a matrix size smaller than \(48\times 24\), which might use a different construction than the one above.
## Proof of the bounds \(1.5682\le K_{\text{D}}\le 2\) {#sec:uplow}
*Upper bound.*---We first prove the upper bound. Translating the Gisin-Gisin model from the Bell nonlocality to the PM scenario , we find that the following statistics can be obtained in the PM scenario with 1 bit of classical communication: \[\label{Egg} E(\vec a,\vec b)=\frac{\vec a\cdot\vec b + 1}{2},\] where \(\vec a\in S^2\) denotes the preparation vector and \(\vec b\in S^2\) denotes the measurement Bloch vector. We give the proof of this formula in the Methods section [3.4](#sec:gisinmodel){reference-type="ref" reference="sec:gisinmodel"} and we show panel (b) in figure [\[fig:setup\]](#fig:setup){reference-type="ref" reference="fig:setup"} for the description of the classical one-bit model. On one hand, due to the above Gisin-Gisin one-bit model, we have for an arbitrary \(n\times m\) matrix \(M\): \[\max\sum_{x=1}^{n}\sum_{y=1}^{m}M_{x,y}E_{x,y}\le L_2(M), \label{giga1}\] where \(E_{x,y}\) has the form ([\[Egg\]](#Egg){reference-type="ref" reference="Egg"}) and we maximized over the unit vectors \(\vec a_x\) and \(\vec b_y\) in the three-dimensional Euclidean space. On the other hand, substituting \(E_{x,y}:=E(\vec a_x,\vec b_y)\) in the formula ([\[Egg\]](#Egg){reference-type="ref" reference="Egg"}) into ([\[giga1\]](#giga1){reference-type="ref" reference="giga1"}) we find \[\begin{aligned}
\max\sum_{x=1}^{n}\sum_{y=1}^{m}M_{x,y}E_{x,y} &= \frac{\max\sum_{xy} M_{xy}\left(\vec a_x\cdot\vec b_y + 1\right)}{2}\nonumber\\ &= \frac{q(M) +S(M)}{2}, \label{giga2} \end{aligned}\] where maximization is over the unit vectors \(\vec a_x\) and \(\vec b_y\) in the three-dimensional Euclidean space, and we also used the definition of \(q(M)\) in ([\[qM\]](#qM){reference-type="ref" reference="qM"}) and the definition of \(S(M)\) in ([\[SM\]](#SM){reference-type="ref" reference="SM"}). Comparing the right-hand side of ([\[giga1\]](#giga1){reference-type="ref" reference="giga1"}) with ([\[giga2\]](#giga2){reference-type="ref" reference="giga2"}), we have \[\frac{q(M)-S(M)}{L_2(M)-S(M)}\le 2,\] where the left-hand side of ([\[KD\]](#KD){reference-type="ref" reference="KD"}) is just \(K_{\text{D}}\), which proves the upper bound \(K_{\text{D}}\le 2\). 0◻*Lower bound.*---In the following, we prove the lower bound using large-scale numerical tools. Note, however, that the resulting bound is rigorous and in particular the final result is due to exact computations. The steps are as follows. Given a fixed setup with Alice's Bloch vectors \(\vec a_x\), \(x=(1,\ldots,n)\) and Bob's Bloch vectors \(\vec b_y\), \(y=(1,\ldots,m)\) the method is the following. We define the \((n\times m)\)-dimensional one-parameter family of matrices \(E_{xy}(\eta)\) with entries \[E_{x,y}(\eta)=\eta E_{x,y}+(1-\eta), \label{Exyeta}\] where \(E_{x,y}=\vec a_x\cdot\vec b_y\). We wish to show that for some \(\eta\in\left[0,1\right]\), the distribution ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}) in the PM scenario cannot be simulated with one bit of classical communication. In fact, due to the expectation value ([\[Egg\]](#Egg){reference-type="ref" reference="Egg"}) of the Gisin-Gisin model, it is enough to consider the interval \(\eta\in\left[1/2,1\right]\). To show quantumness, we therefore need to find a matrix \(M\) of certain size \(n\times m\) and a given \(\eta\in\left[1/2,1\right]\) such that \[\label{hyper} \sum_{x=1}^n\sum_{y=1}^m M_{xy}E_{xy}(\eta)>L_2(M)\] for \(E_{xy}(\eta)\) defined by ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}), and \(L_2(M)\) is defined by ([\[L2M\]](#L2M){reference-type="ref" reference="L2M"}). The above problem, i.e., finding a suitable \(M\) with the smallest possible \(\eta\) in ([\[hyper\]](#hyper){reference-type="ref" reference="hyper"}), can be solved by a modified version of the original Gilbert algorithm , a popular collision detection method used, for example, in the video game industry. The algorithm is iterative, and the procedure adapted to our problem is given in Sec. [3.5](#sec:gibi){reference-type="ref" reference="sec:gibi"}. Indeed, using the algorithm of Gilbert, we find the value \[\eta^*=0.6377\] and a corresponding \(70\times 70\) matrix \(M\) and \(E_{xy}(\eta^*)\) in the form ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}) which satisfies inequality ([\[hyper\]](#hyper){reference-type="ref" reference="hyper"}). We will give more technical details of the input parameters and the implementation of the algorithm in Section [3.6](#sec:techgibi){reference-type="ref" reference="sec:techgibi"}. Then, rearranging ([\[hyper\]](#hyper){reference-type="ref" reference="hyper"}) and making use of equation ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}), we find the bound \[\frac{\sum_{xy}M_{xy}E_{xy}-S(M)}{L_2(M)-S(M)}>\frac{1}{\eta^*},\] where due to the definitions ([\[qM\]](#qM){reference-type="ref" reference="qM"},[\[KD\]](#KD){reference-type="ref" reference="KD"}) the lower bound \[K_{\text{D}}>(1/\eta^*)\simeq1.5682\] on \(K_{\text{D}}\) follows.
## Physical meaning of the constants \(K_{\text{PM}}\) and \(K_{\text{D}}\)
*The role of \(K_{\text{PM}}\) in the PM scenario.*---The value of \(K_{\text{PM}}\) is interesting from a physical point of view as well, since it is related to the critical noise resistance of the experimental setup if the transmitted \(\rho_x\) goes through a noisy, fully depolarizing channel. That is, \(1-p_{\text{crit}}=1-(1/K_{\text{PM}})\) gives the amount \((1-p_{\text{crit}})\) of critical white noise \(\leavevmode\hbox{\small1\normalsize\kern-.33em1}/2\) that the PM experiment with rank-1 projective qubit measurements can maximally tolerate while still being able to detect quantumness. Namely, for a fully depolarizing channel with visibility parameter \(p\) the qubits \(\rho_x\) emitted by Alice's turn into \(p\rho_x+(1-p)\leavevmode\hbox{\small1\normalsize\kern-.33em1}/2\), and the expectation value ([\[Exydot\]](#Exydot){reference-type="ref" reference="Exydot"}) becomes \[E_{xy}=p\vec a_x\cdot\vec b_y, \label{Exyp}\] where \(\{\vec a_x\}_x\) are the Bloch vectors of Alice's qubits, whereas \(\{\vec b_y\}_y\) are the Bloch vectors of Bob's measurements. To witness quantumness, there must exist expectation values \(E_{xy}\) in ([\[Exyp\]](#Exyp){reference-type="ref" reference="Exyp"}) and a matrix \(M\) of arbitrary size such that \[\sum_{xy}M_{xy}E_{xy}>L_2(M). \label{Witp}\] Inserting ([\[Exyp\]](#Exyp){reference-type="ref" reference="Exyp"}) into ([\[Witp\]](#Witp){reference-type="ref" reference="Witp"}) and making use of ([\[qM\]](#qM){reference-type="ref" reference="qM"}), we obtain \[\label{pcrit} p_{\text{crit}}=\min_M\frac{L_2(M)}{q(M)}=\frac{1}{K_{\text{PM}}}=\frac{1}{K_G(3)}.\] for the critical noise tolerance. In fact, the value of \(K_G(3)\) appears in the studies of the Bell nonlocality of two-qubit Werner states . Note that a recent approach in Ref. , based on the simulability of Werner states with local models, yields the same relation ([\[pcrit\]](#pcrit){reference-type="ref" reference="pcrit"}) between \(p_{\text{crit}}\) and \(1/K_G(3)\). From the upper and lower bounds on \(K_{\text{PM}}\), the following bounds on the amount \((1-p_{\text{crit}})\) of critical white noise follow: \[0.3036\le (1-p_{\text{crit}})\le 0.3172.\] *The role of \(K_{\text{D}}\) in the PM scenario.*---In section [2.2](#sec:uplow){reference-type="ref" reference="sec:uplow"} we proved the lower bound of \(K_{\text{D}}\ge 1.5682\). Below we prove that this bound is related to the finite detection efficiency threshold of Bob's measurements. To this end, we assume that Bob's detectors are not perfect and only fire with probability \(\eta\). Assume that when the measurement \(y\) fails to detect, Bob outputs \(b_y=1\) (due to possible relabelings there is no loss of generality). Assume further that the probability of detection \(\eta\) is the same for all \(y\). This is the problem of symmetric detection efficiency. A review of this problem in the Bell scenario can be found in Ref. . On the other hand, the same problem in the PM scenario has been elaborated in Refs. and the upper bound of \(1/\sqrt 2\) on the critical value of the symmetric detection efficiency was found. Since \(\eta\) does not depend on \(y\), the expectation value \(E_{x,y}\) becomes \(E_{x,y}(\eta)=\eta E_{x,y} + (1-\eta)\) for all \(x\) and \(y\). Hence, the witness matrix \(M\) detects quantumness with finite detection efficiency \(\eta\) (assuming optimal preparation states and measurements) whenever we have \[\eta q(M) + (1-\eta)\sum_{x,y}M_{x,y}>L_2(M). \label{etaineq}\] Recalling \(S(M)=\sum_{x,y}M_{x,y}\), solving the above relation for \(\eta\), and optimizing over all \(M\) witness matrices, we obtain the critical detection efficiency \(\eta_{\text{crit}}\): \[\eta_{\text{crit}}=\min_M{\frac{L_2(M)-S(M)}{q(M)-S(M)}}=\frac{1}{K_{\text{D}}}, \label{etacrit}\] where \(K_{\text{D}}\) is defined by ([\[KD\]](#KD){reference-type="ref" reference="KD"}). In particular using the lower bound value \(K_{\text{D}}\ge 1.5682\), we obtain the improved upper bound \(0.6377\) on \(\eta_{\text{crit}}\). It should be noted, however, that the above is not the most general detection efficiency model. Rather than outputting \(b_y=1\), Bob can output a third result, which could potentially give a lower detection efficiency threshold. An open problem is whether this third outcome can lower the detection efficiency threshold. In the above, we also assumed that Bob's qubit measurements are rank-1 projectors that can achieve \(q(M)\). However, it is known that the true qubit maximum \(Q(M)\) (in ([\[QM\]](#QM){reference-type="ref" reference="QM"})) can be larger than \(q(M)\) (in ([\[qM\]](#qM){reference-type="ref" reference="qM"})) for a given \(M\). Hence, we can say that the most general symmetric detection efficiency threshold is upper bounded by \(1/K_{\text{D}}\), and it is an open problem whether this upper bound is tight or not.
# Methods {#Methodssec}
## Properties of the \(L_2\) and \(L_k\), \(k>2\) norm {#sec:L2proofs}
*Notations.*---We first introduce notation used throughout this subsection. Let \(A^n, n = 0, 1, 2, \dots\) be the set of \(n\) dimensional vectors over the set \(A\). Let \(v_i\) denote the \(i\)th element of \(v \in A^n\) (\(i = 1, 2, \dots, n\)). Let \(\_;\_: A^n \times A^m \rightarrow A^{n+m}\) denote the concatenation of vectors. Let \(()\) the singleton element of \(A^0\). Further let \((a) \in A^1\) if \(a \in A\). The parenthesis may be omitted so \((1); (2); (3) = 1; 2; 3 \in \mathbb R^3\), for example. Let \(\overline a^{n} = a; a;...; a \in A^n\) where \(a \in A\). We write \(\overline a\) instead of \(\overline a^{n}\) if \(n\) can be inferred from the context. We define \(\Mat{n}{m}\) as the set of real \(n\times m\) matrices. Matrices are represented as vectors of their row vectors, i.e. \(\Mat{n}{m} = (\mathbb R^m)^n\). Let \(\transp{M} \in \Mat{m}{n}\) be the transposition of \(M \in \Mat{n}{m}\) and let \(\Id{m} \in \Mat{m}{m}\) denote the \(m\times m\) identity matrix. Further, it is convenient to define by \(\W{n}{k} = \{\Id{k}_j \,|\, j = 1, 2, \dots, k\}^n \subset \Mat{n}{k}\) the set of matrices whose rows are all \(0\)s, but exactly one is \(1\). Let \(\Perm{n} \subset\Mat{n}{n}\) denote the set of permutation matrices. Let \(\SM{M} = \sum_{i=1}^n \left\|M_i\right\|_1\) denote the Manhattan norm of the matrix \(M \in \Mat{n}{m}\). *Definition of \(L_k\).*---We first give the definition of \(L_k\). Let \(k \in \mathbb N^{+}.\) \[\begin{aligned} &L_k: \Mat{n}{m} \rightarrow \mathbb R\\ &L_k(M) = \max_{W \in \W{n}{k}} \SM{\transp{W}M}. \end{aligned} \label{L2Def}\] Note that \(W\) is defined above in *Notations* and \(\transp{W}\) denotes the transposed matrix of \(W\). We prove below that Eq. ([\[L2Def\]](#L2Def){reference-type="ref" reference="L2Def"}) corresponds to Eq. ([\[L2Mv3\]](#L2Mv3){reference-type="ref" reference="L2Mv3"}) in the case of \(k=2\). The proof is as follows \[\begin{aligned}
\max_{a_x\in\{\pm 1\}}& \left(\Man{\sum_{x: a_x=+1} M_x} + \Man{\sum_{x: a_x=-1} M_x}\right)\\ &= \max_{W\in\W{n}{2}} \left(\Man{\transp{W}_1M}+ \Man{\transp{W}_2M}\right)\\ &= \max_{W\in\W{n}{2}} \SM{\transp{W}M}\\ &= L_2(M). \end{aligned}\] *Properties of \(L_k\).*---We prove several interesting properties of \(L_k\). Note that our focus in the main text is on \(k=2\). However, the general case \(k\ge 2\) is of interest for its own sake. Moreover, it is also motivated physically, corresponding to classical communication beyond bits . First we prove that \(L_k\) is a norm for any \(k\ge 2\). To this end, we prove its homogeneity, positive definiteness and subadditivity properties.
Homogeneity: \[\begin{aligned}
L_k(tM)&= \max_{W \in \W{n}{k}} \SM{\transp{W}(tM)}\\ &= \max_{W \in \W{n}{k}} |t|\SM{\transp{W}M}\\ &= |t|\max_{W \in \W{n}{k}} \SM{\transp{W}M}\\ &= |t|L_k(M), \end{aligned}\] where \(|t|\) denotes the absolute value of the scalar \(t\) and \(L_k\) is defined by ([\[L2Def\]](#L2Def){reference-type="ref" reference="L2Def"}). Positive definiteness: \[\begin{aligned}
L_k(M) &= 0\\ &\Rightarrow\max_{W \in \W{n}{k}} \SM{\transp{W}M} = 0\\ &\Rightarrow \forall W \in \W{n}{k}: \SM{\transp{W}M} = 0\\ &\Rightarrow \forall W \in \W{n}{k}: \transp{W}M = \overline{\overline 0}\\ &\Rightarrow \forall i: M_i = \overline 0\\ &\Rightarrow M = \overline{\overline 0}. \end{aligned}\] Triangle inequality: \[\begin{aligned}
&L_k(M+N) = \max_{W \in \W{n}{k}} \SM{\transp{W}(M+N)}\\ &\le \max_{W \in \W{n}{k}} (\SM{\transp{W}M} + \SM{\transp{W}N})\\ &\le \max_{W \in \W{n}{k}} \SM{\transp{W}M} + \max_{W \in \W{n}{k}} \SM{\transp{W}N}\\ &= L_k(M) + L_k(N). \end{aligned}\] 0◻Let us define \(L(M)\) as follows \[L(M) = \max_{v \in \{-1, +1\}^n} \Man{vM}. \label{LDef}\] The above definition is consistent with the one given in ([\[LMdef\]](#LMdef){reference-type="ref" reference="LMdef"}). \(L(M)\) is the local or classical bound of correlation Bell inequalities defined by the correlation matrix \(M\) in ([\[LDef\]](#LDef){reference-type="ref" reference="LDef"}). The \(L(M)\) quantity also appears in computer science literature under the name of \(K_{m,n}\)-quadratic programming . Let us note that recently an efficient computation of \(L(M)\) has been proposed in Ref. along with the code . First we prove the basic property that \(L_2(M)\ge L(M)\) for any \(M\). Next we prove that \(L_k(M)\le L_{k+1}(M)\) for \(k\ge 2\). Then we bound \(L_k(M)\) from above by the value of \(L(M)\) multiplied by \(k\). However, we do not know whether the bound can be saturated or not. The lemma stating our first claim is as follows
where the proof is given as the following chain of equations plus a single inequality invoked in the fourth line \[\begin{aligned}
L(M)&= \max_{v \in \{-1, +1\}^n} \Man{vM}\\ &= \max_{v \in \{-1, +1\}^n} \Man{\frac12(\overline 1 + v)M-\frac12(\overline 1-v)M}\\ &= \max_{W \in \W{n}{2}} \Man{\transp{W}_1M-\transp{W}_2M}\\ &\le \max_{W \in \W{n}{2}} (\Man{\transp{W}_1M} + \Man{\transp{W}_2M})\\ &= \max_{W \in \W{n}{2}} \SM{\transp{W}M}\\ &= L_2(M). \end{aligned}\] 0◻Our next lemma proves that \(L_k(M)\) is monotone increasing with \(k\).
The proof is given below as the following chain: \[\begin{aligned}
L_k(M)&= \max_{W \in \W{n}{k}} \SM{\transp{W}M}\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^k \Man{(\transp{W}M)_i}\\ &= \max_{W \in \W{n}{k}} (\sum_{i=1}^k \Man{(\transp{W}M)_i} + \Man{\overline 0M})\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^{k+1} \Man{((\transp{W}; \overline 0)M)_i}\\ &= \max_{W \in \W{n}{k}} \SM{(\transp{W}; \overline 0)M}\\ &\le \max_{W \in \mathcal W_{k+1,n}} \SM{\transp{W}M}\\ &= L_{k+1}(M). \end{aligned}\] 0◻Finally, we prove an upper bound on \(L_k(M)\). Our lemma reads as follows
*Proof.* \[\begin{aligned}
L_k(M)&=\max_{W \in \W{n}{k}} \SM{\transp{W}M}\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^k \Man{(\transp{W}M)_i}\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^k \Man{\transp{W}_iM}\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^k \frac12\Man{(2\transp{W}_i-\overline 1)M + \overline 1M}\\ &\le \max_{W \in \W{n}{k}} \sum_{i=1}^k \frac12(\Man{(2\transp{W}_i-\overline 1)M} + \Man{\overline 1M})\\ &\le \max_{W \in \W{n}{k}} \sum_{i=1}^k \frac12(L(M) + L(M))\\ &= \max_{W \in \W{n}{k}} \sum_{i=1}^k L(M)\\ &= \max_{W \in \W{n}{k}} kL(M)\\ &= kL(M). \end{aligned}\] To arrive at the sixth line, we invoked the definition ([\[LDef\]](#LDef){reference-type="ref" reference="LDef"}). 0◻It is an open question whether Lemma [\[LkkL\]](#LkkL){reference-type="ref" reference="LkkL"} is tight or not. However, we can find a family of matrices \(M^{(k)}\), \(k\ge 2\) such that the ratio \(L_k(M^{(k)})/L(M^{(k)})\) tends to infinity with increasing \(k\). More formally we have
The proof is based on an explicit construction of matrices \(M^k\), \(k=(2,\ldots,\infty)\) defined in Ref. . See also Refs. . *Proof.* Let \(M^k\in\Mat{k}{2^{k-1}}\), \(k = (1, 2, \dots,\infty)\) be a family of matrices such that \[M^k_{i,j} = (-1)^{\left\lfloor\frac{j}{2^{k-i-1}}\right\rfloor}.\] For example, \[M^4 = \left(\begin{array}{rrrrrrrr}1&1&1&1&1&1&1&1\\1&1&1&1&-1&-1&-1&-1\\1&1&-1&-1&1&1&-1&-1\\1&-1&1&-1&1&-1&1&-1\end{array}\right).\] Now by explicit calculations we find \[\frac{L(M^k)}{L_k(M^k)} = \frac{k\left(\begin{array}{c}k-1\\\left\lfloor\frac{k-1}2\right\rfloor\end{array}\right)}{k2^{k-1}} \sim \sqrt\frac2{\pi k}.\] 0◻Note that in the particular case of \(k=2\) the matrix \(M^{(k)}\) is the CHSH expression , in which case \(L(M^{(2)})=2\) and \(L_2(M^{(2)})=4\). Hence, for \(k=2\) the upper bound in Lemma [\[LkkL\]](#LkkL){reference-type="ref" reference="LkkL"} is tight. We conjecture that the bound is not tight for greater values of \(k\). Finally, we show how \(L_2\) and in general \(L_k\) behaves with the concatenation \((A; B)\) of two matrices \(A\) and \(B\), where we defined \[(A;B) = \left[\begin{array}{r}A\\B\end{array}\right].\]
*Proof.* \[\begin{aligned}
L_k(A; B)&= \max_{W \in \W{i+j}{k}} \SM{\transp{W}(A; B)}\\ &= \max_{S \in \W{i}{k}} \max_{T\in\W{j}{k}} \SM{\transp{(S; T)}(A; B)}\\ &= \max_{S \in \W{i}{k}} \max_{T\in\W{j}{k}} \SM{\transp{S}A + \transp{T}B}\\ &\le \max_{S \in \W{i}{k}} \max_{T\in\W{j}{k}} (\SM{\transp{S}A} + \SM{\transp{T}B})\\ &\le \max_{S \in \W{i}{k}} \SM{\transp{S}A} + \max_{T\in\W{j}{k}} \SM{\transp{T}B}\\ &= L_k(A) + L_k(B). \end{aligned}\] 0◻Note that \(L_k(A) \le L_k(A; B)\) does not hold in general. For example, let us have \[A = \left(\begin{array}{rr}1&1\\1&-1\end{array}\right)\] and \[B = \left(\begin{array}{rr}-1&0\end{array}\right).\] Then by explicit calculation we obtain \[4 \! = \! L_2(A) > L_2(A; B) \! = \! 3.\] Finally, it is shown that \(L_k\) relates to the cut norm \(C\), a matrix norm introduced by Frieze and Kannan in Ref. (see also for several applications in graph theory). This norm is defined as follows: \[C(M)=\max \sum_{x=1}^{n}\sum_{y=1}^{m} M_{x,y}a_xb_y, \label{CM}\] where the maximum is taken over all \(a_x, b_y\in\{0,1\}\). Note the similarity in the definition with the \(L(M)\) norm ([\[LMdef\]](#LMdef){reference-type="ref" reference="LMdef"}) which is equivalent to ([\[LDef\]](#LDef){reference-type="ref" reference="LDef"}). It has been shown that \(C(M)\) is related to \(L(M)\) as follows : \[C(M)\le L(M) \le 4C(M). \label{CMto4CM}\] Using the above relation ([\[CMto4CM\]](#CMto4CM){reference-type="ref" reference="CMto4CM"}) along with Lemma [\[LkkL\]](#LkkL){reference-type="ref" reference="LkkL"}, we find that \[C(M)\le L_k(M) \le 4kC(M), \label{CMto4kCM}\] and for the special case of \(L_2\) we have the following lower and upper bounds: \[C(M)\le L_2(M) \le 8C(M). \label{CMto8CM}\] *Generalization of the \(L_k\) norm.*---Below we generalize the norm \(L_k(M)\) to \(F_M\), which extension will prove to be a key property in the Branch-and-Bound implementation of the \(L_k\) algorithm. To do so, first we define the following function
In other words, \(F_M(P)\) is the maximum of \(\SM{\transp{W}M}\) where \(W \in\W{k}{n}\) and the prefix of \(W\) is \(P\). \(F_M\) can be considered as a generalization of \(L_k(M)\). The following lemma introduces a key property which is made use of in the Branch-and-Bound method.
*Proof.* \[\begin{aligned}
F_{A; B}(P)&= \max_{W \in \W{n-i}{k}} \SM{\transp{(P; W)}(A; B)}\\ &= \max_{W \in \W{n-i}{k}} \SM{(\transp{P}A + \transp{W}B)}\\ &\le \max_{W \in \W{n-i}{k}} (\SM{\transp{P}A} + \SM{\transp{W}B})\\ &= \SM{\transp{P}A} + \max_{W \in \W{n-i}{k}} \SM{\transp{W}B}\\ &= \SM{\transp{P}A} + L_k(B). \end{aligned}\] 0◻Let us now give the following definition further generalizing \(F_M(P)\):
The computation of \(f_M\) can be optimized such that for big enough \(c\) values \(f_M(P)(c)\) returns \(c\) without computing \(F_M(P)\). This is expressed by the following lemma.
The proof given below is split into three cases. Case 1: if \(P \in \W{n}{k}\), then \[\begin{aligned}
&f_M(P)(c)\\ &= \max\left(\max_{W \in \W{n-n}{k}} \SM{\transp{(P; W)}M}, c\right) \quad \text{by}\; (\ref{KDef})\, \text{and}\, (\ref{KPrimeDef})\\ &= \max(\SM{\transp{(P; ())}M}, c)\\ &= \max(\SM{\transp{P}M}, c). \end{aligned}\] Case 2: \[\begin{aligned}
\SM{\transp{P}A} &+ L_k(B) \le c\\ &\Rightarrow F_M(P) \le c \quad \text{by}\; (\ref{FfL2})\\ &\Rightarrow \max(F_M(P), c) = c\\ &\Rightarrow f_M(P)(c) = c \quad \text{by}\; (\ref{KPrimeDef}) \end{aligned}\] Case 3: \[\begin{aligned}
&f_M(P)(c)\\ &= \max\left(\max_{W \in \W{n-i}{k}} \SM{\transp{(P; W)}M}, c\right) \quad \text{by} \; (\ref{KDef}, \ref{KPrimeDef})\\ &= \max\left(\max_{S\in\W1k, W\in\W{n-i-1}{k}} \SM{\transp{(P; S; W)}M}, c\right)\\ &= \max\left(\max_j \max_{W \in \W{n-i-1}{k}} \SM{\transp{(P; \Id{k}_j; W)}M}, c\right)\\ &= \max(\max_j F_M(P;\Id{k}_j), c)\\ &= \max(F_M(P;\Id{k}_0), \max(F_M(P;\Id{k}_1), \max(..., c)))\\ &= (f_M(P;\Id{k}_0) \circ \dots \circ f_M(P;\Id{k}_k))(c). \end{aligned}\] 0◻Our last lemma in this subsection reads
and it can be proved as follows: \[\begin{aligned}
&f_M(())(0)\\ &= \max\left(\max_{W \in \W{n}{k}} \SM{\transp{((); W)}M}, 0\right) \quad \text{by}\, (\ref{KDef}, \ref{KPrimeDef})\\ &= \max_{W \in \W{n}{k}} \SM{\transp{W}M}\\ &= L_k(M) \end{aligned}\] 0◻
## Programming tips for the efficient implementation of the \(L_2\) and \(L_k\) codes {#sec:L2tips}
In this subsection, we give programming tips for the branch-and-bound implementation of the exact computation of \(L_k(M)\) for any \(k\ge 2\). For \(k=2\) and \(k=3\) our algorithms are even faster than the \(L_k\) solver for general \(k\) due to specialization which we detail below. First, we remind the reader of the notation defined in Sec. [3.1](#sec:L2proofs){reference-type="ref" reference="sec:L2proofs"}. The Haskell code can be downloaded from Github . Instructions installing and using the code (including parallel execution and using guessed results) can also be found there. *Branch-and-bound calculation of \(L_k\).*---The norm \(L_k(M)\) for \(k\ge 2\) can be calculated using the following definition and the following lemma.
The function \(f_M\) recursively calls itself with larger and larger \(P\) prefixes until the prefix size reaches \(n\). The middle case is a conditional exit from the recursion, which speeds up the computation crucially.
*Reducing cost by sharing sub-calculations.*---In the definition [\[KPrimeOptRecO\]](#KPrimeOptRecO){reference-type="ref" reference="KPrimeOptRecO"}, the most expensive calculations are \(L_k(B)\), \(\SM{\transp{P}M}\) and \(\SM{\transp{P}A}\). We show how to reduce the cost of these calculations. The cost of \(L_k(B)\) can be reduced by memoizing the previously computed \(L_k\) values in a table. If \(M = (v_1; v_2; v_3; \dots; v_n)\) then \(L_k(M)\) depends on \(L_k(v_i; v_{i+1}; \dots; v_n)\), where \(i = 2, 3, 4, \dots, n\). Note that \(L_k(v_i; v_{i+1}; \dots; v_n)\) itself depends on \(L_k(v_j; v_{j+1}; \dots; v_n)\), where \(j = i+1, i+2, \dots, n\). If we take into account all dependencies, the correct order of calculating \(L_k\) values is \(L_k(v_n)\), \(L_k(v_{n-1}, v_n)\), \(L_k(v_{n-2}, v_{n-1}, v_n)\), ..., \(L_k(v_2, v_3, \dots, v_n)\). There is an option to skip the \(\SM{\transp{P}A} + L_k(B) \le c\) test for large \(B\) matrices. This means that \(L_k(B)\) should not be calculated, and the trade-off is that we miss opportunities for exiting recursion. In our experience, skipping the test for \(B \in \Mat{k}{m}, k \ge (3n/4)\) results in about \(2\times\) speedup. The cost of calculating \(\SM{\transp{P}A}\) is \(O(km)\) if \(A\in\Mat{k}{m}\). Note that \[\transp{(P; Q)}(A; B) = \transp{P}A + \transp{Q}B. \label{share}\] \(\transp{P}A\) is already computed by the time when \(\transp{(P;\Id{k}_i)}(A;v)\) is needed, so the cost of \(\SM{\transp{P}A}\) can be reduced to \(O(m)\) by ([\[share\]](#share){reference-type="ref" reference="share"}). The cost of \(\transp{P}M\) can be reduced in the same way. This implies a considerable speedup; for example, for \(M \in \mathcal M_{70, 70}\) the calculation of \(L_2(M)\) can be made nearly 70 times faster by this optimization. The cost of \(\SM{\transp{P}A}\) can be further reduced by caching the previously calculated Manhattan norms of the rows of the matrix \(\transp{P}A\). *Reducing cost by symmetries.*---For all \(S\in\Perm{k}\) permutation matrices \[\SM{\transp{W}M} = \SM{\transp{(WS)}M}. \label{symm}\] The cost of \(L_2\) can be halved by ([\[symm\]](#symm){reference-type="ref" reference="symm"}) as follows. Let \[S = \left(\begin{array}{rr}0&1\\1&0\end{array}\right)\in\Perm{2}. \nonumber\] From \(\SM{\transp{(\Id{2}_2; W)}M} = \SM{\transp{(\Id{2}_1; WS)}M}\) it follows that \(f_M(\Id{2}_2, c) = f_M(\Id{2}_1, c)\). This means that we can skip the calculation of \(f_M(\Id{2}_2, c)\) for all \(c\), thus \(L_2(M) = f_M((\Id{2}_1), 0)\), i.e., we start the calculation with a non-empty prefix which saves work. Harnessing ([\[symm\]](#symm){reference-type="ref" reference="symm"}) in the general \(L_k\) case is a bit more complex. First we define the set of canonical prefixes. A prefix \(P = \Id{k}_{i_1}; \Id{k}_{i_2}; \dots \Id{k}_{i_j}\) is canonical if the first occurrences of the numbers in the indices \(i_1, i_2, \dots, i_j\) is the sequence \(1, 2, 3, \dots\). For example, the prefix \(\Id{k}_1; \Id{k}_2; \Id{k}_1; \Id{k}_3\) is canonical but \(\Id{k}_1; \Id{k}_3; \Id{k}_1; \Id{k}_2\) is non-canonical. For each prefix \(P\), there exists a permutation \(S\) such that \(PS\) is canonical, so that, \(f_M(P, c) = f_M(PS, c)\), which means that it is enough to examine only the canonical prefixes to compute \(L_k\). *Parallel and concurrent execution.*---For parallel execution one can use the following equation: \[f_M(P)(c) = \max_{i\in\{1,\dots,k\}} f_M(P;\Id{k}_i)(c) \label{fPar}\] We used Eq. ([\[fPar\]](#fPar){reference-type="ref" reference="fPar"}) for \(P\in\W{i}{k}, i<d\), where \(d\) is a "parallel depth" for fine-tuning the execution for different architectures. Higher depth is better for more cores. Parallel execution may miss opportunities of exiting recursion because there is no communication between threads about the best known \(L_k\) values at a certain point of time. Therefore we implemented concurrent execution where threads share the best known \(L_k\) values. *Reducing cost by guessed \(L_k\) values.*---Optionally, the computation can be sped up by providing a guessed \(L_k(M)\) value by the user. This value will be used instead of \(0\) in Eq. ([\[L2AlgO\]](#L2AlgO){reference-type="ref" reference="L2AlgO"}). The guessed value may be lower than \(L_k(M)\). Higher guessed values are better, unless the guessed value is higher than \(L_k(M)\), in which case \(f_M\) returns the guessed value. We compared the result of \(f_M\) with the witness \(W\) of the maximal \(\SM{\transp{W}M}\) value, to be able to detect whether the guessed value was too high or not.
## \(L\)-norm and \(L_2\)-norm are the same for a special family of matrices \(M'\) {#sec:LL2Mp}
We relate \(L(M')\) to \(L_2(M')\), where \(M'\) is given by the following matrix \[\begin{aligned}
M'=\begin{pmatrix} &M \\ -&M \end{pmatrix}, \label{Mv} \end{aligned}\] where \(M\) is a matrix of size \(n\times m\) with arbitrary real entries. Note that \(M'\) has size \(2n\times m\) and \(M'\) has rows such that \(M'_x=M_x\) and \(M'_{x+m}=-M_{x}\) for all \(x=1,\ldots,m\). Then the following lemma holds.
*Proof.* We fix a matrix \(M\) of dimension \(n\times m\) which specifies \(M'\) by the virtue of ([\[Mv\]](#Mv){reference-type="ref" reference="Mv"}). Let \(a_x\in\{-1,1\}\), \(x=1,\ldots,n\) and \(b_y\in\{-1,1\}\), \(y=1,\ldots,m\) be the optimal vectors giving \(L(M)\) in ([\[LDef\]](#LDef){reference-type="ref" reference="LDef"}). Note that these values are not unique in general, different optimal configurations may exist, however, we choose one such optimal vectors \(a_x\) and \(b_y\). We then choose \(a_{x+n}=-a_x\) for \(x=1,\ldots,n\), and \(b_y^+=b_y^-=b_y\) for \(y=1,\ldots,m\). With these values, we obtain the lower bound \(L_2(M')\ge 2L(M)\) on \(L_2(M')\) in ([\[L2M\]](#L2M){reference-type="ref" reference="L2M"}). Now we show the upper bound \(L_2(M')\le 2L(M)\), which implies \(L_2(M')=2L(M)\). As a contradiction of the lemma, assume that \(L_2(M')>2L(M)\). Then, not all \(a_x\) vectors corresponding to the \(L_2(M')\) value have the property \(a_{x+m}=-a_x\) for each \(x\). That is, there exists at least one \(x\), call it \(x'\), for which \(a_x'=a_{x'+n}\). Suppose that there is one such an \(x'\) (the proof for multiple \(x'\) indices for which \(a_x'=a_{x'+n}\) is very similar). Then in the formula ([\[L2Mv3\]](#L2Mv3){reference-type="ref" reference="L2Mv3"}) for \(L_2(M')\) the two rows \(x'\) and \(x'+n\) in question will appear within the same norm (either in the first or second norm, depending on whether \(a_x'=a_{x'+n}\) takes the value \(+1\) or \(-1\)). However, in both cases they cancel each other from the norm in question. As a result, two rows of \(M'\) in ([\[Mv\]](#Mv){reference-type="ref" reference="Mv"}) are eliminated, one from the matrix \(M\) and one from the matrix \(-M\). However, any matrix \(\pm M\) from which one row has been eliminated cannot have a local bound greater than \(L(M)\). The same applies to a matrix \(\pm M\) from which we have removed several rows. Therefore, \(L_2(M')>2L(M)\) cannot be true either. Thus we arrived at a contradiction. \(\qed\)
## Adapting the Gisin-Gisin model to the PM scenario {#sec:gisinmodel}
We now adapt the LHV model of Ref. which exploits the finite efficiency of the detectors to reproduce the quantum correlations of the singlet state exactly. We show that the LHV model in Ref. can be adapted to the PM communication scenario to produce the expectation value: \[\label{EGG} E(\vec a,\vec b)=P(b=+1|\vec a,\vec b)-P(b=-1|\vec a,\vec b)=\frac{\vec a\cdot\vec b + 1}{2},\] where \(\vec a\in S^2\) denotes the preparation Bloch vector and \(\vec b\in S^2\) denotes the measurement Bloch vector. First we show that the outcomes \(b=\pm 1\) giving the expectation value \[\label{EGG2} E(\vec a,\vec b)=P(b=+1|\vec a,\vec b)-P(b=-1|\vec a,\vec b)=\vec a\cdot\vec b\] can be obtained with probability \((1/2)\) and \(b=0\) outcome with probability \((1/2)\). Then by coarse-graining the above distribution by grouping \(b=0\) outcome with \(b=+1\), we obtain the expectation value ([\[EGG\]](#EGG){reference-type="ref" reference="EGG"}). The classical model, using one bit of classical communication from Alice to Bob, is as follows. *Protocol:* Alice and Bob share a classical variable, which is in the form of a unit vector \(\vec\lambda\), chosen uniformly at random from the unit sphere \(S^2\).
- *(Alice)* Alice sends a binary message \(c=\text{sgn}(\vec a\cdot\vec\lambda)\) to Bob. That is, \(c=+1\) if \(\vec a\cdot\vec\lambda\le 0\) and \(c=-1\) if \(\vec a\cdot\vec\lambda> 0\).
- *(Bob)* Bob outputs \(b=\text{sgn}(c\vec b\cdot\vec\lambda)\) with probability \(|\vec b\cdot\vec\lambda|\) (corresponding to the detection event \(b=\pm 1\)) and Bob outputs \(b=0\) with probability \(1-|\vec b\cdot\vec\lambda|\) (corresponding to the non-detection event).
Our claim is as follows. The above protocol yields the correlations \(E(\vec a,\vec b)=\vec a\cdot\vec b\), that is, it reproduces the correlations in Eq. ([\[EGG2\]](#EGG2){reference-type="ref" reference="EGG2"}) with probability \(1/2\) and returns \(b=0\) in the other cases. *Proof.*---We need to calculate the expectation value \(E(\vec a,\vec b)=P(b=+1|\vec a,\vec b)-P(b=-1|\vec a,\vec b)\) which according to the above protocol in the detection events \(b=\pm 1\) is given by ) \[E(\vec a,\vec b)=\int_{S^2}d\vec\lambda q(\vec\lambda|b=\pm 1)\text{sgn}(\vec a\cdot\vec\lambda)\text{sgn}(\vec b\cdot\vec\lambda), \label{Eab2}\] where \(q(\vec\lambda|b=\pm 1)\) is the conditional density probability distribution of choosing \(\vec\lambda\) given a detection event (either output \(b=+1\) or \(b=-1\)). This function can be calculated from \[q(\vec\lambda|b=\pm 1)=\frac{q(\vec\lambda \text{ and } b=\pm 1)}{p(b=\pm 1)}, \label{qlambdab}\] where the detection efficiency is \(\eta=p(b=\pm 1)\) and the probability of detection failure is \(1-\eta=p(b=0)\). The value of \(\eta\) is given by \[\eta=p(b=\pm 1)=\int_{S^2}\frac{d\vec\lambda}{4\pi}|\vec b\cdot\lambda|=\frac{1}{2},\] as stated and the protocol gives the density probability distribution \(q(\lambda \text{ and } b=\pm 1) = (\vec b\cdot\vec\lambda)/(4\pi)\). Inserting these values into ([\[qlambdab\]](#qlambdab){reference-type="ref" reference="qlambdab"}) gives \(q(\lambda|b=\pm 1)=(\vec b\cdot\vec\lambda)/(2\pi)\), which in turn is inserted into ([\[Eab2\]](#Eab2){reference-type="ref" reference="Eab2"}) to obtain the integral : \[E(\vec a,\vec b)=\frac{1}{2\pi}\int_{S^2}d\vec\lambda(\vec b\cdot\vec\lambda)\text{sgn}(\vec a\cdot\vec\lambda). \label{Eab3}\] The above integral can be calculated using spherical symmetries. In particular, one can choose w.l.o.g. the vectors \[\begin{aligned}
\vec a &= (0,0,1)\\ \vec b &=(\sin\alpha,0,\cos\alpha) \end{aligned}\] as in Ref. , and then obtain \[E(\vec a,\vec b)=\cos\alpha=\vec a\cdot\vec b \label{Eab4}\] with probability (1/2), which we wanted to prove. 0◻
## The modified Gilbert algorithm adapted to the PM scenario {#sec:gibi}
For a given \(\eta\in\left[1/2,1\right]\) and correlation matrix \(E(\eta)\) defined by ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}), the algorithm yields the following matrix \(M\) satisfying \[\label{hyper2} \sum_{x=1}^n\sum_{y=1}^m M_{xy}E_{xy}(\eta)>L_2(M).\] **Algorithm:** Input: The number of preparations \(n\) and the number of measurement settings \(m\) that define the setup. The unit vectors \(\{\vec a_x\}_{x=1}^n\) (i.e., the Bloch vectors of Alice's prepared states) and \(\{\vec{b_y}\}_{y=1}^m\) (i.e., the Bloch vectors of Bob's projective rank-1 measurements). The \((n\times m)\)-dimensional matrix \(E(\eta)\) given by the entries \(E_{xy}(\eta)\) in ([\[Exyeta\]](#Exyeta){reference-type="ref" reference="Exyeta"}). The values of \(\epsilon\) and \(i_{\text{max}}\) that define the stopping criteria. Output: The matrix \(M\) of size \(n\times m\).
1. Set \(i=0\) and set \(E^{(i)}\) the \(n\times m\) zero matrix.
2. Given a matrix \(E^{(i)}\) and the matrix \(E(\eta)\), run a heuristic oracle that maximizes the overlap \(\sum_{xy}(E_{xy}(\eta)-E_{xy}^{(i)})E_{xy}^{\text{det}}\) over all deterministic one-bit correlations \(E_{xy}^{\text{det}}\) in ([\[Exybit\]](#Exybit){reference-type="ref" reference="Exybit"}). The description of this heuristic (see-saw) oracle is given in Sec. [3.7](#sec:L2seesaw){reference-type="ref" reference="sec:L2seesaw"}. Denote the point \(E_{xy}^{\text{det}}\) returned by the oracle by \(E_{xy}^{\text{det,i}}\).
3. Find the convex combination \(E^{(i+1)}\) of \(E^{(i)}\) and \(E_{xy}^{\text{det,i}}\) that minimizes the distance \(\sqrt{\sum_{xy}\left(E_{xy}(\eta)-E_{xy}^{(i)}\right)^2}\). Let us denote this distance by \(\text{dist}(i)\).
4. Let \(i=i+1\) and go to Step 2 until \(\text{dist}(i)\le\epsilon\) or \(i=i_{\text{max}}\).
5. Return the matrix \(M\) with coordinates \(M_{xy}=E_{xy}(\eta)-E_{xy}^{(i)}\).
Note that \(\text{dist}(i)\) is a decreasing function of \(i\). Since maximizing the overlap of \(\sum_{xy}(E_{xy}(\eta)-E_{xy}^{(i)})E_{xy}^{\text{det}}\) over all deterministic one-bit correlations vectors is an NP-hard problem, in Step 2 we use a heuristic method to do it, which we describe in Sec. [3.7](#sec:L2seesaw){reference-type="ref" reference="sec:L2seesaw"}. On the other hand, the description of an exact branch-and-bound type algorithm can be found in Sec. [3.2](#sec:L2tips){reference-type="ref" reference="sec:L2tips"}. We use the exact method, which is generally more time-consuming than the see-saw method to check that the output matrix \(M\) satisfies the condition ([\[hyper2\]](#hyper2){reference-type="ref" reference="hyper2"}) with the chosen parameter \(\eta\). If this is true, then it implies the lower bound \(K_{\text{D}}\ge (1/\eta)\), as proved in Sec. [2.2](#sec:uplow){reference-type="ref" reference="sec:uplow"}. It should also be noted that the branch-and-bound-type algorithm is much faster than the brute force algorithm (the implemented algorithm using parallelism can be found in ). On a multi-core desktop computer, it can solve problems in range \(n=m=70\) in a day, while the brute force algorithm is limited to about \(n=m=40\) settings.
## Parameters and implementation of Gilbert algorithm {#sec:techgibi}
Here we specify the explicit parameters that are used to obtain the lower bound \(K_{\text{D}}\le 1.5682\). On the three-dimensional unit sphere, we choose the vectors \(\{\vec a_x\}_{x=1}^n\) and \(\{\vec b_y\}_{y=1}^m\) to be equal to each other, \(\vec v_i = \vec a_i=\vec b_i\) for \(i=1,\ldots,n\), where \(n=m=70\). The 70 unit vectors chosen define the optimal packing configuration in the Grassmannian space which can be downloaded from Neil Sloane's database . The advantage of this type of packing is that the points and their antipodal points are located as far apart as possible on the three-dimensional unit sphere. We implemented the modified Gilbert algorithm (of section [3.5](#sec:gibi){reference-type="ref" reference="sec:gibi"}) in Matlab with and without a memory buffer (see more details on the memory buffer in Ref. ). In the case of using memory buffer, the step 3 is modified in the algorithm so that instead of calculating the convex combination of the points \(E^{(i)}\) and \(E_{xy}^{\text{det,i}}\) (see section [3.5](#sec:gibi){reference-type="ref" reference="sec:gibi"}), we compute the convex combination of \(E^{(i)}\) and the points \(E_{xy}^{\text{det,i-j}}\), \(j=0,\ldots,m-1\), where \(m\) is the size of the memory buffer. In our explicit computations, we use a buffer size \(m=40\) and a stopping condition of \(k=2\times 10^5\) with \(\eta=0.665\). Details on the performance of this modification can be found in Ref. . In step 2 of the Gilbert algorithm, the oracle uses the see-saw heuristic described in Sec. [3.7](#sec:L2seesaw){reference-type="ref" reference="sec:L2seesaw"} to obtain a good (typically tight) lower bound to \(L_2(M)\). On the other hand, we used the branch-and-bound-type algorithm described in Sec. [3.2](#sec:L2tips){reference-type="ref" reference="sec:L2tips"} to calculate \(L_2(M)\) exactly for integer \(M\). The algorithm was implemented in Haskell. See the GitHub site for the downloadable version. The Matlab file `eta_70.m`, which can also be downloaded from GitHub (located in the subdirectory `L2_eta_70`) gives detailed results on the input parameters and the output matrix \(M\). In particular, it gives the integer matrix \(M\), the unit vectors \(\vec a_i=\vec b_i=\vec v_i\), the lower bound \(\sum_{xy}M_{xy}\vec a_x\cdot\vec b_y=\sum_{xy}M_{xy}\vec v_x\cdot\vec v_y\) to \(q(M)\) and the value \(L_2(M)\). The running time of the Gilbert algorithm (in Sec. [3.5](#sec:gibi){reference-type="ref" reference="sec:gibi"}) implemented in Matlab was about one week. Note, however, that most of the computation time was spent on the oracle (the see-saw part) described in Sec. [3.7](#sec:L2seesaw){reference-type="ref" reference="sec:L2seesaw"}. On the other hand, the Haskell code to compute the exact \(L_2(M)\) value of the \(70\times 70\) witness matrix \(M\) took about 8 hours to run on a HP Z8 workstation using 56 physical cores. The memory usage of the computation was negligible. The Matlab `eta_70.m` routine defines the \(70\times 70\) matrix \(M\), and gives the \(\vec v_i:=\vec a_i=\vec b_i\) the unit vectors from Sloane's database for all \(i=1,\ldots,70\). Note that \(M\) is integer (by multiplying the output \(M\) matrix in the Gilbert algorithm by \(1000\) and truncating the non-integer part). This calculation yields \(S(M)=\sum_{x,y}M_{x,y}=194369\) and \(Q(M)=\sum_{x,y}M_{x,y}\vec a_x\cdot\vec b_y\simeq5.3672235\times10^5\). On the other hand, the branch-and-bound-type Haskell code gives the exact value \(L_2(M)=412667\), which is matched by the see-saw search (in Sec. [3.7](#sec:L2seesaw){reference-type="ref" reference="sec:L2seesaw"}). From these numbers we then obtain \[K_{\text{D}}\ge\frac{Q(M)-S(M)}{L_2(M)-S(M)}>\frac{342353}{218298}=1.5682+\varepsilon\] and \(1/K_{\text{D}}=0.6377-\varepsilon'\) is the upper bound to the critical detection efficiency \(\eta_{\text{crit}}\), where \(\varepsilon\) and \(\varepsilon'\) are small positive numbers.
## Lower bound to \(L_2(M)\) using the see-saw iterative algorithm {#sec:L2seesaw}
Below we give an iterative algorithm based on see-saw heuristics to compute \(L_2(M)\). This algorithm forms the oracle part of step 2 of the Gilbert algorithm, which is described in Sec. [3.5](#sec:gibi){reference-type="ref" reference="sec:gibi"}. **Algorithm:** Input: Integer matrix \(M\) of size \(n\times m\). Output: Lower bound \(l_2(M)\) to \(L_2(M)\) defined by formula ([\[L2M\]](#L2M){reference-type="ref" reference="L2M"}).
1. Let \(l_2 = 0\).
2. Choose random assignments \(a_x = \pm 1\): That is, \(a_x\) are (random) elements of a vector \(a\) of size \(n\). Its elements are binary having value +1 or-1 only.
3. Set \(b^{+} = \text{sgn}(aM)\),where denotes the (modified) sign function: \(\text{sgn}(x)=+1\) if \(x\ge 0\) and \(-1\) otherwise. Let us transpose \(b^{+}\).
4. Set \(b^{-} = \text{sgn}(aM)\). Let us transpose \(b^{-}\).
5. Form the column vector \(s^+=Mb^+\) of size \(n\).
6. Form the column vector \(s^-=Mb^-\) of size \(n\).
7. Form the column vector \(s=\max(s^+,s^-)\) of size \(n\). That is, \(s_x=\max(s_x^+,s_x^-)\) for all \(x=1,\ldots,n\).
8. Form the \(\pm 1\)-valued column vector \(a\) as follows: Let \(a_x=+1\) if \(s_x^+\ge s_x^-\), otherwise let \(a_x=-1\) for all \(x=1,\ldots,n\).
9. Let \(l_2=\sum_{x=1}^{n}s_x\).
10. With the new vector \(a\), return to point 3. Repeat the algorithm until two values of \(l_2\) are equal in two consecutive iterations.
Note that at each iteration step, objective value \(l_2(M)\) is guaranteed not to decrease. Therefore, the output of the algorithm is a heuristic lower bound on the exact value of \(L_2(M)\).
# Discussion
We have tested the quantumness of two-dimensional systems in the prepare-and-measure (PM) scenario, with \(n\) preparations and \(m\) binary-outcome measurement settings, where \(n\) and \(m\) fall well into the range of 70. In the one-qubit PM scenario, a two-level system is transmitted from the sender to the receiver. In this setup, a real \(n\times m\) matrix \(M\) defines the coefficients of a linear witness. We denote by \(L_2(M)\) the exact value of the one-bit bound associated with matrix \(M\). We found efficient numerical algorithms for computing \(L_2(M)\). If this bound is exceeded, we can detect both the quantumness of the prepared qubits and the quantumness (i.e. incompatibility) of the measurements. We introduced new constants \(K_{\text{M}}\) and \(K_{\text D}\) which are related to the Grothendieck constant of order 3. Our large-scale tools are crucial for the efficient bounding of \(L_2(M)\) and hence for bounding of the constants \(K_{\text{M}}\) and \(K_{\text D}\). We further relate these new constants to the white noise resistance of the prepared qubits and the critical detection efficiency of the measurements performed. For large \(M\) matrices, we have given two algorithms for computing \(L_2(M)\): a simple iterative see-saw-type algorithm and a branch-and-bound-type algorithm. The former is a heuristic algorithm that usually gives a tight lower bound on \(L_2(M)\). However, sometimes it fails to find the exact value of \(L_2(M)\). This happens more and more often as the size of the matrix \(M\) gets larger and larger. In contrast, the latter branch-and-bound-type algorithm gives the exact value of \(L_2(M)\) and can be used to compute \(L_2(M)\) for matrix sizes as large as \(70\times 70\). As an application of the algorithms, we established the bounds \(1.5682\le K_{\text D}\le 2\) on the new constant and an upper bound of \(\eta_{\text{crit}}\le 0.6377\) on the critical detection efficiency of qubit measurements in the PM scenario. | {'timestamp': '2022-12-01T02:19:57', 'yymm': '2211', 'arxiv_id': '2211.17185', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17185'} |
null | null |
null | null |
null | null |
null | null |
# Formulation of the algorithm {#sec:Formulation_of_the_algorithm}
The Parareal method was first formulated in. Versions with averaging incorporated can be found in or. The next subsection [1.1](#subsec:Two-level_Parareal){reference-type="ref" reference="subsec:Two-level_Parareal"} will summarize the methods. The subsection [1.2](#subsec:Two-level_Parareal_averaging){reference-type="ref" reference="subsec:Two-level_Parareal_averaging"} will then present multi-level versions.
## Two-level Parareal and Two-level Parareal with averaging {#subsec:Two-level_Parareal}
The two-level Parareal method has two levels, denoted as level 1 and level 0. Level 1 is the coarse level where we do \(N\) timesteps on a coarse grid which has \(N+1\) grid points. Two neighboring grid points of the coarse grid form an interval. In total we have \(N\) such small intervals and on each small interval we introduce a fine grid. These are the fine grids of level 0. The Parareal method is a parallel-in-time method which has two basic solvers, a coarse propagator denoted as \(P^1\) and a fine propagator denoted as \(P^0\). The upper indices refer to the levels where the propagators are applied. The coarse and the fine propagators can be Runge-Kutta methods, but other choices are possible too. First, the coarse propagator is applied on level 1 to compute initial guess \(U^0_n\), where the index \(n\) counts the timesteps on level 1. The initial guess is computed in serial and shall be improved iteratively. The values provided through the initial guess are passed to the fine grids on level 0 as initial values. Then the fine propagator \(P^0\) is applied in parallel. The results of fine propagator are passed back to the coarse level 1. Applying the coarse propagator again, a Parareal iteration step can be computed \[U_{n+1}^{1} =P^1(U_{n}^{1}) + P^0 (U_{n}^{0})-P^1 (U_{n}^{0}).\] The upper index of the numerical solutions counts the iterations. Once the first iteration is computed, it can be used as a new initial guess and the next iteration can be done. This procedure is modified in the APinT method where averaging is incorporated, see and. Here the coarse propagator, denoted as \(\bar P^1\), provides a numerical solution to an averaged problem and not the original equations. Solving an averaged problem has the advantage that there are no fast oscillations in the solution, provided the averaging window \(\eta\) is chosen appropriately. Thus, the oscillatory stiffness is mitigated and taking large time steps is possible when the numerical method of the coarse propagator applied. This can be beneficial for the efficiency of the algorithm. We can formulate a Parareal iteration step with averaging incorporated as follows \[\label{eq:parareal_iteration_with_averaging} Y_{n+1}^{k+1} =\bar P^1 (Y_{n}^{k+1}) + P^0 (Y_{n}^{k})-\bar P^1(Y_{n}^{k}),\] where \(\bar P^1\) denotes the coarse propagator which provides a numerical solution to the averaged problem on the coarse level, level 1. In the averaged equation the right handside of [\[eq:modulation_equation\]](#eq:modulation_equation){reference-type="eqref" reference="eq:modulation_equation"} is replaced. In particular, when the coarse propagator is applied, an approximation to the following equation is computed \[\label{eq:averaged_problem} \frac{d \bar w}{dt} = \frac{1}{\eta} \int_{-\eta/2}^{\eta/2} \rho \left(\frac{s}{\eta} \right) \exp\left(\frac{L}{\epsilon} (s+t) \right) \mathcal{N} \left(\exp\left(-\frac{L}{\epsilon} (s+t) \right)\bar w(t)\right) ds,\] where \(\eta\) is the averaging window and \(\rho\) is the kernel function. We use \[\rho (s) = \frac{1}{\rho_0} e^{\frac{1}{(s-0.5)(s+0.5)} },\] where \(\rho_0\) normalizes the function. In the right handside of [\[eq:averaged_problem\]](#eq:averaged_problem){reference-type="eqref" reference="eq:averaged_problem"} we write \(\bar w(t)\) to emphasize the the equations depends on \(t\) only and not on \(s\). The oscillations in the equation with wavelength \(\le \eta\) are averaged. Averaging have been studied earlier, see for instance or in the context of heterogeneous multiscale methods or. Should we say more about how the averaging is done? Should more literature be mentioned?
## Multi-level Parareal and Multi-level Parareal with averaging {#subsec:Two-level_Parareal_averaging}
Here we state again the idea behind the Multi-level Parareal algorithm: The Multi-level Parareal algorithm uses a recursion in the levels. The Multi-level Parareal with \(L\) levels is a two-level Parareal algorithm where we compute a coarse solution witht the coarse propagator an the coarsest level and the fine propagator is a Multi-level Parareal algorithm with \(L-1\) levels applied on the finer levels. Then applying the Multi-level Parareal algorithm can be iterated until we reach the two-level case, where the two-level Parareal algorithm, described in the previous subsection, is applied. The algorithm with averaging included is illustrated in [\[fig:Multi_Level_Parareal\]](#fig:Multi_Level_Parareal){reference-type="ref" reference="fig:Multi_Level_Parareal"}. However we note here that it can be applied with and without averaging. When the method without averaging is applied, the coarse propagator solves the original problem and does not provide a solution to an averaged equation. To distinguish between the different levels, we introduce the subsequent notation: When we consider a Multi-level Parareal method with \(L\) levels, level \(L-1\) denotes the coarsest level and level \(0\) is finest level. If \(l_1 > l_2\) for \(l_1, l_2 \in \{ 0, \dots L-1 \}\), then level \(l_1\) is the coarser and level \(l_2\) the finer level. The propagator applied on level \(l\) is denoted by \(P_l\) or in the case when an averaged equation is solved by \(\bar P_l\). In the rest of the subsection we will write \(\bar P_l\). The case without averaging can be concluded straightforwardly. The only thing that must be done is replacing \(\bar P_l\) by \(P_l\). Let us now explain one iteration of the Multi-level Parareal method with \(L\) levels, assuming that every coarse propagator, \(\bar P_{L-1}, \dots,\bar P_1\), is a Parareal method with only one iteration too. On the coarsest level, level L-1, the coarse propagator, \(P_{L-1}\), provides the initial guess, which is the numerical solution to an averaged problem of the form [\[eq:averaged_problem\]](#eq:averaged_problem){reference-type="eqref" reference="eq:averaged_problem"} however with an indexed averaging window \(\eta_{L-1}\). This solution shall now be improved iteratively according to relation [\[eq:parareal_iteration_with_averaging\]](#eq:parareal_iteration_with_averaging){reference-type="eqref" reference="eq:parareal_iteration_with_averaging"}. This means we have to compute fine solutions. The fine solutions on level \(L-2\) are computed in parallel and the coarser solution from level \(L-1\) gives the initial values. In the case of \(L-2 =0\) the fine propagator is a basic numerical scheme, like a Runge-Kutta method, and not a Parareal method. In the other case, \(L-2 > 0\), the fine propagator is a Parareal method with \(L-1\) levels. Especially, we have another coarse propagator \(\bar P_{L-1}\). Although denoted as coarse \(\bar P_{L-2}\) uses finer time steps than the coarse propagator on level L-1, \(\bar P_{L-1}\). Moreover, \(\bar P_{L-2}\) computes a numerical solution to an averaged equation again of the form [\[eq:averaged_problem\]](#eq:averaged_problem){reference-type="eqref" reference="eq:averaged_problem"} but this time with averaging window \(\eta_{L-2}\). As level \(L-2\) is a finer level with finer grids than level \(L-1\), we want to resolve finer scales on level \(L-2\). This means that we resolve some of the oscillations on level \(L-2\) that were averaged on level \(L-1\). Accordingly, we have \(\eta_{L-2} < \eta_{L-1}\). Consequently, oscillations with wavelength \(>O(\eta_{L-2})\) are resolved. We continue the recursion in the levels until we reach level \(0\). Only on level 0 the unaveraged equation [\[eq:modulation_equation\]](#eq:modulation_equation){reference-type="eqref" reference="eq:modulation_equation"} is solved. Once a fine solution is computed, we can apply a Parareal iteration step according to equation [\[eq:parareal_iteration_with_averaging\]](#eq:parareal_iteration_with_averaging){reference-type="eqref" reference="eq:parareal_iteration_with_averaging"}, recursively until we reach level \(L-1\). Especially, level 0 provides the fine solutions for level 1, needed to apply equation [\[eq:parareal_iteration_with_averaging\]](#eq:parareal_iteration_with_averaging){reference-type="eqref" reference="eq:parareal_iteration_with_averaging"} to comoute a Parareal solution on level 1. Then, level 1 gives the fine solutions for level 2 and so on. [\[fig:MultiLevelParareal_InitialValues_FineSolution\]](#fig:MultiLevelParareal_InitialValues_FineSolution){reference-type="ref" reference="fig:MultiLevelParareal_InitialValues_FineSolution"} illustrates how the coarse solutions are passed as initial values to the propagators on the finer levels and the solutions on the finer levels are returned to the coarser levels so that the Parareal iteration step can be done. A basic principle of the Parareal method is that an initial guess is iteratively improved. These correction iterations, which are done on the levels \(L-1, L-2, \dots, 1\) in the multi-level case, lead to cycles. Illustrations of examples with three levels can be found in [\[fig:V_Cycle\]](#fig:V_Cycle){reference-type="ref" reference="fig:V_Cycle"} and [\[fig:W_Cycle\]](#fig:W_Cycle){reference-type="ref" reference="fig:W_Cycle"}. [\[fig:V_Cycle\]](#fig:V_Cycle){reference-type="ref" reference="fig:V_Cycle"} shows the case where one iteration is done on the levels 1 and 2, in Fig. [\[fig:W_Cycle\]](#fig:W_Cycle){reference-type="ref" reference="fig:W_Cycle"} we do two iterations on the levels 1 and 2. The algorithm with averaging is particularly promising for multi-scale problems. For each intrinsic scale of the problem we can introduce a level and resolve the properties of the system which are the specific for that scale, i.e. every scale is assigned a level and the scale specific properties are resolved on the level assigned. The key besides the correct choice of the time steps is to choose the averaging windows for the different levels such that the behavior on the finer scales is averaged but the properties of the model on the scale corresponding to the level considered are still present.
# Asymptotic convergence results {#sec:Asymptotic_convergence}
Here we state asymptotic convergence results for two-level Parareal methods which are known from the literature. The stated results hold when the fine propagator computes the exact solution. The authors in introduce subsequent conditions on the coarse propagator \(P^1\): \[\label{eq:(7)_Gander_Hairer} E(x)-P^1(x) = c_{p+1}(x) \Delta T^{p+1} + c_{p+2} (x) \Delta T^{p+2} + \dots,\] where \(E\) denotes the exact propagator, and \[\label{eq:(8)_Gander_Hairer} \|P^1(x)-P^1(y)\| \le (1+C_2\Delta T) \|x-y\|\] and give the following result
Another result with averaging incorporated is given in.
## Multi-level result without averaging {#subsec:Asymptotic_convergence_without_averaging}
The generalization of the result given in [\[thm:standard_two_level_estimate\]](#thm:standard_two_level_estimate){reference-type="ref" reference="thm:standard_two_level_estimate"} to the multi-level case has two important building blocks. First, a bound must be found when the fine propagator is not exact. This is necessary, as the fine propagator of the Multi-level Parareal algorithm is a Multi-level Parareal method (with one level less) and we want to investigate on which levels numerical errors emerge and how they develop as the fine solutions are passed to the coarser levels. Second, an inductive argument which takes the errors on all levels into account will be applied.
### A two-level result with non-exact fine propagator {#subsec:two-level_result_with_non-exact_fine_propagator}
The first step is to modify the result given in to the case when the fine propagator is not exact. This is necessary, because the fine propagator of a Multi-level Parareal algorithm with \(L\) levels is a Parareal algorithm with \(L-1\) levels and finally we want to estimate the error contribution emerging on the different levels. The time step on level 0 is denoted as \(\Delta t\) or \(\Delta T_0\) and on level 1 as \(\Delta T\) or \(\Delta T_1\).
To refer to the coarse time grid we used the notation \(T_0, T_1, \dots, T_N\). The bound in [\[eq:non-exact_fine_propagator\]](#eq:non-exact_fine_propagator){reference-type="ref" reference="eq:non-exact_fine_propagator"} is the two-level estimate from [\[thm:standard_two_level_estimate\]](#thm:standard_two_level_estimate){reference-type="ref" reference="thm:standard_two_level_estimate"} with an additional term depending on the accuracy of the fine propagator.
### Generalization to multiple levels {#subsec:multi-level_result}
The next step is to investigate the error estimate for the Multi-level Parareal method, and in particular how the errors introduced on the different levels influence the total error. We assume that we are given \(L\) levels, where level 0 corresponds to the finest level and level \(L-1\) is the level with coarsest grid. The proof for the Multi-Level method will be an inductive proof where we can apply similar arguments as in the two-level case. The difference is that the fine propagator is not a basic ODE solver, like a Runge-Kutta scheme, but it is the Multi-Level Parareal method on the finer grid. Only on the finest level, level 0, we apply a basic ODE solver. In particular, this changes the appearance of \(\delta\) in the proof. The size of the timestep is different for different levels. Therefore, \(\alpha, \beta, \gamma\) and \(\delta\) should have an index \(l\). More precisely, we write \(\alpha_l, \beta_l, \gamma_l\) and \(\delta_l\) instead of \(\alpha, \beta, \gamma\) and \(\delta\). In addition, the number of Parareal iteration does not have to be the same for each level. Consequently, \(k\) should also have an index \(l\) and we should write \(k_l\). Moreover, we introduce a coarsening factor \(N\) which relates the timesteps on the different levels. Let \(\Delta t = \Delta T_0\) be the timestep on the finest grids on level 0. Then, the next finest grids are on level 1 and the timestep is given by \(\Delta T_1 = N \Delta T_0\). For the next coarsest level, level 2, we have a timestep \(\Delta T_2 = N \Delta T_1= N^2 \Delta T_0\) and so on. See also [\[subsec:lemmata_multiple_levels\]](#subsec:lemmata_multiple_levels){reference-type="ref" reference="subsec:lemmata_multiple_levels"} and especially the relations [\[eq:relation_for_Delta_T\_l_1\]](#eq:relation_for_Delta_T_l_1){reference-type="ref" reference="eq:relation_for_Delta_T_l_1"} and [\[eq:relation_for_Delta_T\_l_2\]](#eq:relation_for_Delta_T_l_2){reference-type="ref" reference="eq:relation_for_Delta_T_l_2"}.
Especially, the derived error estimate is a sum. The different terms of the sum are composed of the error \(E_l\) emerging on level \(l\) and amplification factors \(A_l\) which amplify the errors made on the finer levels as the fine solutions are passed to the coarser levels.
# Asymptotic convergence results {#sec:Asymptotic_convergence}
Here we state asymptotic convergence results for two-level Parareal methods which are known from the literature. The stated results hold when the fine propagator computes the exact solution. The authors in introduce subsequent conditions on the coarse propagator \(P^1\): \[\label{eq:(7)_Gander_Hairer} E(x)-P^1(x) = c_{p+1}(x) \Delta T^{p+1} + c_{p+2} (x) \Delta T^{p+2} + \dots,\] where \(E\) denotes the exact propagator, and \[\label{eq:(8)_Gander_Hairer} \|P^1(x)-P^1(y)\| \le (1+C_2\Delta T) \|x-y\|\] and give the following result
Another result with averaging incorporated is given in.
## Multi-level result without averaging {#subsec:Asymptotic_convergence_without_averaging}
The generalization of the result given in [\[thm:standard_two_level_estimate\]](#thm:standard_two_level_estimate){reference-type="ref" reference="thm:standard_two_level_estimate"} to the multi-level case has two important building blocks. First, a bound must be found when the fine propagator is not exact. This is necessary, as the fine propagator of the Multi-level Parareal algorithm is a Multi-level Parareal method (with one level less) and we want to investigate on which levels numerical errors emerge and how they develop as the fine solutions are passed to the coarser levels. Second, an inductive argument which takes the errors on all levels into account will be applied.
### A two-level result with non-exact fine propagator {#subsec:two-level_result_with_non-exact_fine_propagator}
The first step is to modify the result given in to the case when the fine propagator is not exact. This is necessary, because the fine propagator of a Multi-level Parareal algorithm with \(L\) levels is a Parareal algorithm with \(L-1\) levels and finally we want to estimate the error contribution emerging on the different levels. The time step on level 0 is denoted as \(\Delta t\) or \(\Delta T_0\) and on level 1 as \(\Delta T\) or \(\Delta T_1\).
To refer to the coarse time grid we used the notation \(T_0, T_1, \dots, T_N\). The bound in [\[eq:non-exact_fine_propagator\]](#eq:non-exact_fine_propagator){reference-type="ref" reference="eq:non-exact_fine_propagator"} is the two-level estimate from [\[thm:standard_two_level_estimate\]](#thm:standard_two_level_estimate){reference-type="ref" reference="thm:standard_two_level_estimate"} with an additional term depending on the accuracy of the fine propagator.
### Generalization to multiple levels {#subsec:multi-level_result}
The next step is to investigate the error estimate for the Multi-level Parareal method, and in particular how the errors introduced on the different levels influence the total error. We assume that we are given \(L\) levels, where level 0 corresponds to the finest level and level \(L-1\) is the level with coarsest grid. The proof for the Multi-Level method will be an inductive proof where we can apply similar arguments as in the two-level case. The difference is that the fine propagator is not a basic ODE solver, like a Runge-Kutta scheme, but it is the Multi-Level Parareal method on the finer grid. Only on the finest level, level 0, we apply a basic ODE solver. In particular, this changes the appearance of \(\delta\) in the proof. The size of the timestep is different for different levels. Therefore, \(\alpha, \beta, \gamma\) and \(\delta\) should have an index \(l\). More precisely, we write \(\alpha_l, \beta_l, \gamma_l\) and \(\delta_l\) instead of \(\alpha, \beta, \gamma\) and \(\delta\). In addition, the number of Parareal iteration does not have to be the same for each level. Consequently, \(k\) should also have an index \(l\) and we should write \(k_l\). Moreover, we introduce a coarsening factor \(N\) which relates the timesteps on the different levels. Let \(\Delta t = \Delta T_0\) be the timestep on the finest grids on level 0. Then, the next finest grids are on level 1 and the timestep is given by \(\Delta T_1 = N \Delta T_0\). For the next coarsest level, level 2, we have a timestep \(\Delta T_2 = N \Delta T_1= N^2 \Delta T_0\) and so on. See also [\[subsec:lemmata_multiple_levels\]](#subsec:lemmata_multiple_levels){reference-type="ref" reference="subsec:lemmata_multiple_levels"} and especially the relations [\[eq:relation_for_Delta_T\_l_1\]](#eq:relation_for_Delta_T_l_1){reference-type="ref" reference="eq:relation_for_Delta_T_l_1"} and [\[eq:relation_for_Delta_T\_l_2\]](#eq:relation_for_Delta_T_l_2){reference-type="ref" reference="eq:relation_for_Delta_T_l_2"}.
Especially, the derived error estimate is a sum. The different terms of the sum are composed of the error \(E_l\) emerging on level \(l\) and amplification factors \(A_l\) which amplify the errors made on the finer levels as the fine solutions are passed to the coarser levels.
# Complexity {#sec:Efficency}
In this section, we distinguish between the total number of steps done on a level and the serial number of steps. Suppose on the coarsest level we do \(Y\) steps with a propagator, for example to compute the initial guess. Then the total number of steps on that level is \(Y\) and the serial number of steps on that level is \(Y\), too. Now, we go to the next finer level and the coarsening factor is \(N\). We then have \(NY\) as the total number of steps and \(N\) as the number of serial steps (done on \(Y\) grids in parallel). Refining again leads to \(N^2Y\) as the total number of steps and \(N\) serial steps (done on \(NY\) grids in parallel). We define the complexity of an algorithm as the number of serial steps. The complexity of the 3-level Parareal algorithm can be defined as \[C^3 = k_2 \left( N_2 + k_1 \left( N_1+N_0 \right) +N_1 \right),\] where \(N_i\) is the number of serial steps on level \(i\) and \(k_i\) is the number of iterations on level \(i\). For a V-cycle we have \(k_i=1\). We may hypothesize that increasing levels pays off when we have a broad range of scales or strong scale separation. We will give an example calculation below where the total number of steps on the finest level is given as \(X=10^6\). But first, we will derive an optimal coarsening factor, which minimizes the complexity. Suppose \(X\) is total number of fine steps on the finest level that must be done and \(N\) is the coarsening factor which relates the time steps of the different levels, i.e \(X\) would be the number of serial steps that must be done when a non-parallelizable basic ODE solver is applied. (We might assume that two neighboring levels are related by the different coarsening factors. To keep the computation simple we will not do that in this place and leave it for future work.) Then we can find a coarsening factor which minimizes the complexity and depends on \(X\) and the number of levels. For the complexity of a V-cycle, we have \[\begin{aligned}
\text{1 level: } f_1(N) &= X \text{ serial steps} \\ \text{2 levels: } f_2(N) &= 2N+X/N \text{ serial steps} \\ \text{3 levels: } f_3(N) &= 2N+2N+X/N^2 = 4N+X/N^2 \text{ serial steps} \\ \text{L levels: } f_L(N) &= 2(L-1)N+ X/N^{L-1} \text{ serial steps}. \\ \end{aligned}\] To find the coarsening factor \(N\) which minimizes the number of serial steps, depending on the number of levels, we compute the minimum. \[\begin{aligned}
f_L'(N) = 0\\ 0 = 2(L-1)-(L-1)X/N^L \\ N = (X/2)^{1/L}. \end{aligned}\] However, when we choose a coarsening factor \(N\) which is computed in the described way, we adapt the algorithm on the behavior of the model on the finest scale only. It may be necessary to account for the behavior of the model on other scales too, see the example in [5.4](#subsec:Swinging_spring){reference-type="ref" reference="subsec:Swinging_spring"}. Now, we come back to the example calculation. Let us assume \(X=10^6\) total steps must be done on the finest level. The ratio of the optimal complexities for a 2-and 3-level Paraeal scheme is roughly 6. This means that the 2-level scheme does roughly 6 times more serial steps compared to the 3-level scheme, provided optimal coarsening factors \(N\) are chosen for both schemes. This means we get a stable solution a lot faster with the 3-level scheme. The total number of evaluations of the right hand side for all the levels is not that much bigger than that of a serial time stepper. On the finest level we do the same number of evaluations of the right hand side that we would do with a serial time stepping scheme, provided we apply the Multi-level Parareal algorithm with averaging. When we do in total \(X\) evaluations of the right hand side on the finest level, level 0, we do \(X/N\) evaluations of the right hand side on level 1, \(X/N^2\) evaluations of the right hand side on level 2 and so on. That means that the number of evaluations of the right hand side decreases exponentially with the levels. Considering the Multi-level Parareal algorithm without averaging we can possibly not coarsen in the same way because the time step might become too large on the coarser levels and as a result the numerical solver might be unstable for the oscillatory problems.
# Numerical examples {#sec:Numerical_examples}
For the numerical examples the explicit midpoint rule (RK2) was used as a basic integrator. The results presented in [5.3](#subsec:System_with_three_scales){reference-type="ref" reference="subsec:System_with_three_scales"} and [5.4](#subsec:Swinging_spring){reference-type="ref" reference="subsec:Swinging_spring"} were computed solving the modulation equations of the systems.
## 1D single scale example {#subsec:1D_singe_scale_example}
Let us consider the problem \[\label{eq:non_oscillatory_problem} \frac{d x}{dt} =-x, \qquad x(0) = 1, \qquad t \in [0,2]\] which is to be solved with the Multi-level Parareal algorithm without averaging. We compute the error for one V-cycle and increase the number of levels. The time step on the coarsest level is for all the computations in the Table below constant \(\Delta T = 0.25\). We choose a coarsening factor of \(N=10\). This means: When we have a time step \(dt\) on level l, the time step on level \(l-1\) is \(dt/10\).
We see in [1](#tab:error_coarse_dt_const_){reference-type="ref" reference="tab:error_coarse_dt_const_"} that increasing the number of levels while keeping the time step on the coarsest level constant does barely change the error of the approximations of problem [\[eq:non_oscillatory_problem\]](#eq:non_oscillatory_problem){reference-type="ref" reference="eq:non_oscillatory_problem"}. This example leads to the following observation. The Multi-level Parareal algorithm with \(L\) levels (without averaging) is a two-level Parareal algorithm where the fine propagator is a Parareal algorithm with \(L-1\) levels. Therefore, when the dominant component of the error emerges on the coarsest level, we cannot not expect that the accuracy of the method increases when we only add additional levels and keep the timestep on the coarsest level constant.
## 1D example with fast oscillations {#sucsec:1D_example_with_fast_oscillations}
Let us now suppose the following situation. We have to solve a multi-scale problem with very fast oscillations on the finest scale. On the finest level we resolve the finest scale. Therefore the time step must be chosen small enough to resolve the very fast oscillations. This is independent of the total number of levels that the method has. In particular, the fastest scale dictates the time step on level 0. An example for such a problem is given by \[\label{eq:test_problem_eff} \frac{d w}{dt} =-\exp(irt)w^2\] for large values of \(r\). This equation can be interpreted as a one dimensional modulation equation. In the following we will consider \(r=100\), \(r=1000\) and \(r=10000\). The exact solution of the problem is given by \[w(t) = \frac{r w_0}{-iw_0 \exp(irt) + iw_0 +r}.\] The averaged problem obeys the following relation \[\frac{d \bar w}{dt} = (-1)\ \underbrace{\frac{1}{\eta_l} \int_{-\eta_l/2}^{\eta_l/2} \ \rho \left( \frac{s}{\eta_l} \right) \ \exp(irs) \ ds}_{\text{damping factor}} \ \exp(irt) \ w(t)^2.\] The problem [\[eq:test_problem_eff\]](#eq:test_problem_eff){reference-type="eqref" reference="eq:test_problem_eff"} will be solved with Multi-level Parareal methods including averaging. Numerical studies show that for \(r = 100\) the wavelength in the solution is of order \(0.1\). Therefore we choose averaging windows \(\eta_1 = 0.2 \text{ and } \eta_2 = 2\). In addition, for \(r = 1000\) the wavelength in the solution is of order \(0.01\). Thus the averaging windows \(\eta_1 = 0.02, \eta_2 = 0.2 \text{ and } \eta_3 = 2\) are a reasonable choice. For the case \(r = 10000\) the wavelength in the solution is of order \(0.001\). Therefore, the computations are done with the averaging windows \(\eta_1 = 0.002, \eta_2 = 0.02, \eta_3 = 0.2 \text{ and } \eta_4 = 2\). The length of the solution interval is always 1 and the coarsening factor is always 10. To solve problem [\[eq:test_problem_eff\]](#eq:test_problem_eff){reference-type="ref" reference="eq:test_problem_eff"} a V-cycle is applied. Depending on the value for \(r\) we choose different step size on the finest level
- r = 100: The step size on the finest level is \(\Delta t = 10^{-3}\)
- r = 1000: The step size on the finest level is \(\Delta t = 10^{-4}\)
- r = 10000: The step size on the finest level is \(\Delta t = 2.5 \cdot 10^{-5}\)
In [2](#tab:error_coarse_dt_const_3){reference-type="ref" reference="tab:error_coarse_dt_const_3"}, [3](#tab:error_coarse_dt_const_4){reference-type="ref" reference="tab:error_coarse_dt_const_4"} and [4](#tab:error_coarse_dt_const_5){reference-type="ref" reference="tab:error_coarse_dt_const_5"} we find the errors at time \(t=1\). The time step on the finest level was adapted to the fast oscillations in the problem which depend on the parameter \(r\). Fixing one value for \(r\), we see that the errors hardly change in the number of levels. We can draw the following conclusions. First, for the parameters chosen for the model runs the increased parallelism of the methods with 3, 4 or 5 levels should lead to a better efficiency. Second, having a third, fourth or fifth level does not change the error significantly. This might indicate that the error emerges on the finest level, where the fast oscillations are resolved and computed numerically. This means when we want to design more efficient algorithms with multi-level methods we have to exploit the increased parallelism when we have multiple levels. A gain in efficiency should come from a combination of increased parallelism and larger time steps on the coarser levels. For oscillatory problems larger times steps can lead to unstable behavior of the algorithm. This can be circumvented when the averaging is applied. Possibly more Parareal iterations can contribute to gain efficiency, too. The effect of more iterations on the efficiency shall be explored in future work.
## System with three scales {#subsec:System_with_three_scales}
We consider a simple model problem with a quadratic non-linearity which is inspired by an equation used to explain issues due to bi-linear terms arising in fluid modelling, see, \[\begin{pmatrix} \dot u_1 \\ \dot u_2 \\ \dot u_3 \end{pmatrix} + \begin{pmatrix} i \omega_1 &0& 0 \\0 & i \omega_2 & 0 \\0& 0 & i \omega_3 \end{pmatrix} \begin{pmatrix} u_1\\ u_2 \\ u_3 \end{pmatrix} + \begin{pmatrix} u_1u_1\\ u_2u_2 \\ u_3u_3 \end{pmatrix} = 0. \label{eq:three_timescales}\] We write \(\dot u_1\) instead of \(\frac{du_1}{dt}\). For strongly differing \(\omega_i\), \(i \in \{1,2,3\}\), the system has three well-separated timescales. Example [\[eq:three_timescales\]](#eq:three_timescales){reference-type="eqref" reference="eq:three_timescales"} is very convenient for numerical studies because the exact solution of the system and the averaged equations can be derived. The system is decoupled which would in principle allow us to treat the equations separately. However, we want to study the effect of scale separation in a problem and in general multi-scale systems it can be the case that one equation sees the effect of several scales, see for instance the example in [5.4](#subsec:Swinging_spring){reference-type="ref" reference="subsec:Swinging_spring"}. Therefore the system will be treated as a whole. Especially, we choose the same time steps and averaging windows for all the equations. In [\[sec:introduction\]](#sec:introduction){reference-type="ref" reference="sec:introduction"} we have already mentioned that when working with the modulation equation, the information about the phase is not in the initial data but in the exponential which is part of the right handside of the system. This is now illustrated in [\[fig:secondComp_InitGuess\]](#fig:secondComp_InitGuess){reference-type="ref" reference="fig:secondComp_InitGuess"} and [\[fig:secondComp_ThirdIt\]](#fig:secondComp_ThirdIt){reference-type="ref" reference="fig:secondComp_ThirdIt"}. In [\[fig:secondComp_InitGuess\]](#fig:secondComp_InitGuess){reference-type="ref" reference="fig:secondComp_InitGuess"} we see the strongly damped initial guess provided by the coarse propagator. This means that the fine propagator computes fine solutions on each small interval with almost the same initial data. Moreover, we see that the exact solution, denoted as the reference solution in the figure, can deviate strongly from the damped solution due to the oscillations. But as the coarse propagator only provides strongly damped solutions, it must be the fine propagator which gives the strong variation from the damped solution in the Parareal iteration. As [\[fig:secondComp_ThirdIt\]](#fig:secondComp_ThirdIt){reference-type="ref" reference="fig:secondComp_ThirdIt"} shows the variations from the damped solutions are reproduced correctly by the Parareal method. Therefore, as the fine propagator uses almost the same initial value on each small interval, the information about the phase must be in the right handside of the system and cannot be in the initial data. [\[fig:eq:three_timescales_error\]](#fig:eq:three_timescales_error){reference-type="ref" reference="fig:eq:three_timescales_error"} shows a study of the numerical errors made with a 3-level Parareal algorithm with averaging. The parameters in problem [\[eq:three_timescales\]](#eq:three_timescales){reference-type="eqref" reference="eq:three_timescales"} are \(\omega_1 =2, \omega_2 =20 \text{ and } \omega_3 =200\) and the solution is computed on the time interval \[0,6\]. The following averaging windows were chosen \(\eta_1 = 0.1 \text{ and } \eta_2 = 1\). In addition, a coarsening factor \(N=10\) related the different levels. We always have \(k_1=3\) iteration on level 1. In [\[fig:eq:three_timescales_error\]](#fig:eq:three_timescales_error){reference-type="ref" reference="fig:eq:three_timescales_error"}, the error is computed at time \(t=6\) for the second component. We can see that the accuracy of the solution increases with the number of iterations on level 2 for different coarse time steps on level 2. After several iterations the accuracy of the method does not increase anymore. The Multi-level Parareal method converges to the fine solution. In the case of the 3-level method with averaging the fine solution is the 2-level method with averaging.
## Swinging spring (Elastic Pendulum) {#subsec:Swinging_spring}
In this subsection we do a first test on a fluid-related, more complicated problem, the swinging spring, also denoted as the elastic pendulum. The relation of the system to geophysical flows was outlined for instance in. A perturbation analysis for the barotropic potential vorticity equation, a comlplex partial differential equation, leads to the three-wave equations. These equations also govern the dynamics of the swinging spring. Further investigations on the dynamics of the system can be found for example in. The swinging spring can be rewritten as a first order system and is given by \[\frac{d \bf w}{dt} = \begin{pmatrix} \dot x_1 \\ \dot x_2 \\\dot y_1 \\ \dot y_2\\ \dot z_1 \\ \dot z_2 \end{pmatrix} = \begin{pmatrix} 0 & 1& 0 & 0 & 0& 0 \\-\omega_R^2 & 0& 0 & 0 & 0& 0 \\ 0 & 0& 0 & 1 & 0& 0 \\ 0 & 0&-\omega_R^2 & 0&0 & 0 \\ 0 & 0& 0 & 0&0 & 1 \\ 0 & 0& 0 & 0 &-\omega_Z^2& 0 \\ \end{pmatrix} \begin{pmatrix} x_1 \\ x_2 \\ y_1 \\ y_2\\ z_1 \\ z_2 \end{pmatrix} + \begin{pmatrix} 0 \\ \lambda x_1 z_1 \\ 0 \\ \lambda y_1 z_1\\ 0 \\ \frac{1}{2} \lambda (x_1^2 + y_1^2) \end{pmatrix} .\] Multi-level Parareal algorithms with averaging are applied to solve the system. We compare a three-level method with a two-level method and investigate the errors for \(T_{max} =50\) for the first component of the system. For the two-level method we choose time steps \(\Delta T_1 = 5\) and \(\Delta T_0 = 0.05\). The time step \(\Delta T_0\) is small enough to resolve the fast oscillations in the problem. With the time step \(\Delta T_1\) the coarse dynamics is resolved. Additionally, the averaging window is \(\eta = 2\), which guarantees that the fast oscillations are averaged on the coarse level, level 1. When applying the three-level method, two iterations are done on the intermediate level, level 1. The number of iterations on the coarse level, level 2 is varied in the test. Moreover, the coarse and fine time steps are the same as in the two-level case. Especially, we have \(\Delta T_2 = 5\) and \(\Delta T_0 = 0.05\). But we introduce an intermediate level and increase the parallelism. The time step of the intermediate level, level 1, is \(\Delta T_1 = 0.5\). In particular, the coarsening factor is \(N=10\). For the coarsest level the averaging window is chosen as \(\eta_2 = 2\). We test different averaging windows on the intermediate level, namely \(\eta_1 = 0.2, 0.75 \text{ and } 2\). When we choose \(\eta_1 = 0.2\) we relate the averaging windows on the levels by the same coarsening factor that relates the time step. The choice \(\eta_1 = 2\) is inspired by the dynamics of the system, because this window ensures that we average the fast oscillations also on level 1. The value \(\eta_1 = 0.75\) is an intermediate choice to see how the error develops for changing averaging windows. The reference solution is computed with the RK2 method with a time steps of \(\Delta t = 0.001\). This is more accurate than the fine solver of the two-level method. The errors for the two-and three-level methods can be found in [5](#tab:error_coarse_dt_const_6){reference-type="ref" reference="tab:error_coarse_dt_const_6"}. We may observe that with an increasing number iterations on the coarse level the accuracy of both the two-and three-level methods increases until the methods are converged. How good the approximations of the three-level methods are strongly depends on the choice of the averaging window on the intermediate level, \(\eta_1\). For \(\eta_1 =2\) the convergence for the three-level method is almost the same as for the two-level method, whereas the accuracy becomes worse for smaller intermediate averaging windows. Consequently for the case \(\eta_1 =2\) we can benefit from more parallelism of the three-level method. [6](#tab:error_coarse_dt_const_7){reference-type="ref" reference="tab:error_coarse_dt_const_7"} shows the number of serial steps done for the two-and three-level methods depending on the iterations. Especially, we see that we do less serial steps when the three-level method is applied.
## 1D Rotating Shallow Water Equations (RSWE) {#subsec:RSWE}
Here solutions to the RSWE given by the equations \[\begin{aligned}
\frac{dv_1}{dt} + \frac{1}{\epsilon} \left(-v_2+ F^{-1/2} \frac{dh}{dx}\right) + v_1 \frac{dv_1}{dx} &=\mu \partial_x^4 v_1 \\ \frac{dv_2}{dt} + \frac{1}{\epsilon} v_1 + v_1 \frac{dv_2}{dx} &= \mu \partial_x^4 v_2 \\ \frac{dh}{dt} + \frac{ F^{-1/2}}{\epsilon} \frac{dv_1}{dx} + \frac{\partial}{\partial x} (hv_1) &= \mu \partial_x^4 h \end{aligned}\] shall be computed. This is the same example investigated in where convergence of the two-level method with averaging was shown. With \(h(x,t)\) we denote the surface height and \(v1(x,t), v2(x,t)\) are the horizontal velocity. We choose the parameter \(\epsilon = 1/10\) which is the Rossby number and the Froude number \(F^{1/2} \epsilon=\). Moreover, a hyperviscosity term with diffusion coefficient \(\mu = 10^{-4}\) is used. Additionally, we impose periodic boundary conditions and use the following initial data \[\begin{split} & v1=0, \qquad v2=0 \\ & h = c_1 \left( e^{(-4(x-\pi/4)^2)} \sin(3(x-\pi/2)) + e^{(-2(x-\pi)^2)} \sin(8(x-\pi)) \right) + c_0, \end{split}\] where the constants c1 and c2 are chosen such such that \[\int_0^{2\pi} h(x,0) dx = 0, \qquad \max_x | h(x,0) | = 1.\] To solve the RSWE numerically a pseudo-spectral method with 128 spatial Fourier modes is be applied. For the time stepping we use a three-level Parareal scheme with averaging in two different parameter regimes. The Rossby number \(\epsilon\) is chosen 0.1 in both cases. We vary the Burger number which is 1 in the first case and 100 in the second case. The tests were done with coarsening factors \(N=10,20,40\). The time step on the finest level is 1/2000. The time steps on level 1 and 2 are determined by the fine time step and the coarsening factor. The averaging windows on the levels 1 and 2 are chosen to be equal to the time steps on the levels 1 and 2. We see in [\[fig:3level_error_quasigeostophic\]](#fig:3level_error_quasigeostophic){reference-type="ref" reference="fig:3level_error_quasigeostophic"} and [\[fig:3level_error_incompressible\]](#fig:3level_error_incompressible){reference-type="ref" reference="fig:3level_error_incompressible"} the numerical method converges with with increasing number iterations. For the case \(F=100\) the method converges faster than for the case \(F=1\). Suppose we were given a two-level method to solve the case with \(F=1\). The two-level method uses a coarsening factor \(N=40\) and two correction iterations are done. The solution interval in time is \[0,48\]. Then we want to answer the question if it is possible to design a three-level method which is more efficient than the two-level method. The two-level method gives an error of 1.7963565539455182e-05, where 7280 serial steps are done. We compare it to the three-level method with a coarsening factor of \(N=20\). With \(k\) we denote the number of steps an the coarse level and we do 3 iterations on the intermediate levels. Then the three-level method converges to a two-level method with three iterations which guarantees that the three-level method to a two-level method which is more exact the two-level method with two iterations. When \(k<28\) we do less serial steps with the three-level method than with two-level method with two iterations. With the three-level method we reach the accuracy of the two-level method with two iterations after 13 iterations already. Therefore we can fins a three-level method which is more efficient than two-level method with two iterations.
# Discussion and conclusion {#sec:Discussion_conclusion}
In the present study, a Multi-level Parareal method with and without averaging for oscillatory problems was proposed and investigated. For that we adopted the terminology of cycles used in multigrid-methods. The averaging plays a central role when oscillatory problems shall be solved because it mitigates the oscillatory stiffness. Therefore big time steps can be used on the coarse levels, which was later demonstrated in the numerical examples. For future studies a change of the integration kernels might be of interest to adapt the method to dissipative problems. One important result is the derivation of asymptotic error bounds building on results in the literature and underpinning theoretically the convergence of the method. In particular, the basic time stepping methods applied on the levels cause an error, denoted as \(E_l\) or \(\delta_0\). The errors are amplified as a solution from one level is passed to coarser grids. In the error bound the amplification is given by the amplifications factors \(A_j\). Increasing the number of iterations on a level \(l\) changes \(E_l\). Additionally, computational complexity of the Multi-level Parareal method was discussed. Finally, several numerical examples were studied. The investigation of the numerical examples included a discussion when multiple levels can be more efficient than the Parareal method with only two levels. A gain of efficiency can be difficult to accomplish if the algorithms without averaging are used, see the example in [5.1](#subsec:1D_singe_scale_example){reference-type="ref" reference="subsec:1D_singe_scale_example"}. However, combined with the averaging the multi-level approach seems promising, as it allows us to take big time steps on the coarse levels and we do not have to satisfy severe stability constraints, see the other examples. Example [5.4](#subsec:Swinging_spring){reference-type="ref" reference="subsec:Swinging_spring"} also shows that a good choice of the averaging windows is crucial for the accuracy. The examinations on the example in [5.4](#subsec:Swinging_spring){reference-type="ref" reference="subsec:Swinging_spring"} also inspire an interesting variant of the Multi-level Parareal method. When introducing an intermediate level, we might use a constant function as initial guess on the grids of level 1. In particular, the constant functions on level 1 would admit the initial values provided by level 2. Possibly, this would not increase the error significantly because the level 1 solutions only have to capture the coarse dynamics which is already resolved by the solution from level 2. Additionally, we would save the serial steps needed to compute the initial guess on level 1. The investigations concerning computational complexity and efficiency are by far not exhaustive and shall be continued in the future on more complex examples including many scales or a continuous range of scales. | {'timestamp': '2022-12-01T02:21:16', 'yymm': '2211', 'arxiv_id': '2211.17239', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17239'} |
# Introduction {#sec:introduction}
From Darwinian evolution, a process of local optimization, emerged highly complex organisms. An organism is, in a sense, a program encoded by DNA determining physical actions based on observed data. It is therefore natural to ask whether it would also be possible to evolve through natural selection computer code that solves a given problem. This is the task at the heart of Genetic Programming. This paper is aimed at describing a hill-climbing algorithm for evolving a programs, taking an approach that is quite different from established Genetic Programming methods. As a starting point, let us consider an extremely simple hill-climbing algorithm for evolving programs for a programming language \(\mathscr L\). For that algorithm, we will need
(1) an initial program \(P_0 \in \mathscr L\),[^1]
(2) a scoring function \(S: \mathscr L \to [0, 1]\) assessing how well a program is performing at solving a given task with \(1\) meaning perfect,[^2] and
(3) for each program \(P \in \mathscr L\), a probability distribution \(M(P)\) on \(\mathscr L\) of of \(P\).
The (non-deterministic) algorithm produces a sequence of programs \(P_0, P_1, \ldots \in \mathscr L\) where \(P_0\) is the initial program and for \(k \geq 1\) \[P_k \defeq \begin{cases} P_k', & \text{if \(S(P_k') > S(P_{k-1})\);} \\ P_{k-1}, & \text{otherwise.} \end{cases}\] where \(P_k'\) is randomly sampled from \(M(P_{k-1})\). The algorithm terminates with solution \(P_k\) when \(S(P_k) = 1\), which may or may not happen eventually. A good choice of the distribution \(M(P)\) should minimize the expected score loss \(L(P) \defeq S(P)-\mathbb E[S(M(P))].\) In particular, likely mutations should ideally only affect the program behavior slightly. However, small changes in code often have major effect on the behavior of the program, so typically \(L > 0\). The objective of this paper is to describe ways to improve and refine this primitive idea to make it feasible:
(1) In [2](#sec:local){reference-type="ref" reference="sec:local"}, we shall see how assign a score to each expression in a program as opposed to the entire program at once. In this way, we get a more precise understanding of which parts of a program are buggy. Later, these are leveraged to choose mutations so that the expressions that are the most buggy are the most likely to be mutated.
(2) In [3](#sec:suppose){reference-type="ref" reference="sec:suppose"}, we introduce \(\mathbf{suppose}\)-expressions. These are expressions that are waiting to be promoted into \(\mathbf{if}\)-conditionals. This two-stage process gives an efficient and precise way to evolve conditions for \(\mathbf{if}\)-conditionals.
(3) In [4](#sec:principles){reference-type="ref" reference="sec:principles"}, we discuss greedy criteria for when a mutation should be regarded successful. Specifically, we will mutate one expression at a time, and the mutation is regarded successful if the score of the new expression is strictly greater than the score of the old expression.
(4) In [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}, we describe an evolutionary process which consists of *cycles*. Each cycle is divided into four stages: stretching, mutation, rewinding, and compression. This design allows us to probe new program structures while avoiding endless program expansion.
(5) In [6](#sec:poc){reference-type="ref" reference="sec:poc"}, we present a basic proof-of-concept implementation of these ideas and show that it is able to generate correct code for several small problems. On one well-known problem, the even-parity problem, we also compare our implementation to some found in the literature and find that our methods outperforms the best of the considered implementations by more than an order of magnitude.
(6) In [7](#sec:future){reference-type="ref" reference="sec:future"}, we discuss future theoretical and practical directions, including integration with Deep Learning techniques.
## Functional Programming as a setting for Genetic Programming
This paper is focused on functional programs, although some of its techniques are applicable to other paradigms as well. The lack of side effects lends itself to Genetic Programming as it means that changing a part of the code has a more predictable effect on the program behavior: we need not worry about how a state change could affect the behavior of expressions elsewhere in the program.
# Local Scoring {#sec:local}
When evolving programs, we need to be able to quantify how well a program is performing. As we discussed in the introduction, one could do that by assigning the program a score based on how well it adheres to its test cases. However, with such an approach to scoring, if a test case fails, we gain little insight into what parts of the code could be the root cause of the failure. If we were able to trace back such failure to specific parts of the code, we could focus on mutating those parts. In this section, we explain a simple method for assigning each expression a score. We assume that every type is an Algebraic Data Type so that values can be viewed as trees; each node corresponds to a value constructor and its children correspond to its arguments. Then, the idea is to augment each node in the value tree with a *trace*, which is a list of expressions that could affect the value at that node. To state this method precisely, we formally describe a small programming language implementing this idea.
## A programming language with traced values {#subsec:language}
As the setting for this paper, we consider a simplistic, functional programming language whose values are . We first describe its syntax: \[\begin{aligned}
f \Coloneqq\ & g_1 \mid g_2 \mid \cdots & \text{function} \\ c \Coloneqq\ & \mathbf{True} \mid \mathbf{False} \mid d_1 \mid d_2 \mid \cdots & \text{value constructor} \\ x \Coloneqq\ & y_1 \mid y_2 \mid \cdots & \text{variable} \\ t \Coloneqq\ & 1 \mid 2 \mid \cdots & \text{expression tag} \\ E \Coloneqq\ & x \mid c(e, \ldots, e) \mid f(e, \ldots, e) & \\ \mid\ & \mathbf{case}\mathop{} e \mathop{}\mathbf{of}\mathop{} c(x, \ldots, x) \to e; \cdots; c(x, \ldots, x) \to e \mathop{}\mathbf{end} \\ \mid\ & \mathbf{if}\mathop{} e \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{else}\mathop{} e \mathop{}\mathbf{end} \\ \mid\ & \mathbf{suppose}\mathop{} e \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end} \\ e \Coloneqq\ & E^{(t)} & \text{tagged expression} \\ D \Coloneqq\ & f(x, \ldots, x) = e & \text{function definition} \\ P \Coloneqq\ & D \mid D; P & \text{program} \end{aligned}\] We moreover stipulate that
(1) in a program \(P\), each function \(g_i\) may be declared at most once,
(2) in a program \(P\), each participating expression has a unique tag, that is, among all the bodies of functions and all their subexpressions, each tag \(i\) appears at most once,
(3) cases in a \(\mathbf{case}\)-expression should be disjoint, that is, the constructors appearing on the left-hand side of the arms should be pairwise distinct,
(4) a variable can be bound at most once in each case of a \(\mathbf{case}\)-expression, that is, the left-hand side of the arm takes the form \(c(x_1, \ldots, x_n)\) where \(x_1, \ldots, x_n\) are pairwise distinct variables, and
(5) variables bound by a case of a \(\mathbf{case}\)-expression cannot overshadow variables that are already bound.
We now describe the typing rules. We use a simple nominal type system with types \[\tau \Coloneqq \mathbf{Bool} \mid \tau_1 \mid \tau_2 \mid \cdots.\] A *typing context* \(\Gamma\) is a collection of judgments of the forms
(1) a variable typing: \(x: \tau\),
(2) a function signature: \(f: (\tau, \ldots, \tau) \to \tau\), or
(3) a type definition: \(\tau \Coloneqq c(\tau, \ldots, \tau) \mid \cdots \mid c(\tau, \ldots, \tau)\).
The typing rules are \[\inferrule[TrueType] {} {\Gamma \vdash \mathbf{True}^{(t)}: \mathbf{Bool}} \qquad \inferrule[FalseType] {} {\Gamma \vdash \mathbf{False}^{(t)}: \mathbf{Bool}} \quad \inferrule[VarType] {x: a \in \Gamma} {\Gamma \vdash x^{(t)}: a}\] \[\inferrule[FunctionType] { \Gamma \vdash e_1: a_1, \ldots, e_n: a_n \\ f: (a_1, \ldots, a_n) \to b \in \Gamma } {\Gamma \vdash f(e_1, \ldots, e_n)^{(t)}: b}\] \[\inferrule[ConstructorType] { \Gamma \vdash e_1: a_1, \ldots, e_n: a_n \\ a \Coloneqq \cdots \mid c(a_1, \ldots, a_n) \mid \cdots \in \Gamma } {\Gamma \vdash c(e_1, \ldots, e_n)^{(t)}: a}\] \[\inferrule[CaseType] { a \Coloneqq c_1(a_{1, 1}, \ldots, a_{1, n_1}) \mid \cdots \mid c_m(a_{m, 1}, \ldots, a_{m, n_m}) \in \Gamma \\ \Gamma \vdash d: a \\ \Gamma, \forall j. x_{1, j}: a_{1, j} \vdash e_1: b \\ \cdots \\ \Gamma, \forall j. x_{m, j}: a_{m, j} \vdash e_m: b } { \Gamma \vdash \mathbf{case}\mathop{} d \mathop{}\mathbf{of}\mathop{} c_1(x_{1, 1}, \ldots, x_{1, n_1}) \to e_1; \cdots; c_m(x_{m, 1}, \ldots, x_{m, n_m}) \to e_n \mathop{}\mathbf{end}^{(t)} : b }\] \[\inferrule[IfType] {\Gamma \vdash e_1: a, e_2: a, c: \mathbf{Bool}} { \Gamma \vdash \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} e_1 \mathop{}\mathbf{else}\mathop{} e_2 \mathop{}\mathbf{end}^{(t)} : a } \qquad \inferrule[SupposeType] {\Gamma \vdash e: a, c: \mathbf{Bool}} { \Gamma \vdash \mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end}^{(t)} : a }\] We shall often write just \(P\) when we mean a well-typed program \((P, \Gamma)\), letting the typing context \(\Gamma\) containing all the type definitions and function signatures be implicit. At last, we explain the semantics of our programming language. The values of the programming language are \[v \Coloneqq c^{[\pm t, \ldots, \pm t]}(v, \ldots, v)\] where \(\pm \Coloneqq + \mid-\). Conceptually, each value constructor is tagged with a *trace*, a list of expressions that could affect the value. Its trace need not contain the traces of the arguments \((v, \ldots, v)\) whose traces in turn consist of expressions that could affect the arguments but not the parent value constructor. A sign \(\pm\) should be thought of as measuring the of an expression: \(+\) means that if the value is bad then the expression should be punished, and \(-\) means that if the value is bad then the expression should be rewarded. The relevance of \(-\) will be revealed in [3](#sec:suppose){reference-type="ref" reference="sec:suppose"}. An *evaluation context* \(\sigma\) is a collection of bindings of the form \[B \Coloneqq x = v \mid f(x, \ldots, x) = e\] such that each variable \(x\) and each function \(f\) appears at most once in the left-hand sides. For a program \(P\), one extracts an evaluation context \(\sigma(P)\) by collecting all its function definitions. We use the notation \(\sigma \vdash e \downarrow v\) to denote that an expression \(e\) evaluates to a value \(v\) in the evaluation context \(\sigma\). Furthermore, we make use of the notation \[\(c^{[t_{k + 1}, \ldots, t_n]}(v_1, \ldots, v_k)\)^{[t_0, \ldots, t_k,]} \defeq c^{[t_0, t_1, \ldots, t_n]}(v_1, \ldots, v_k)\] where \(t_i\) are signed tags. We shall sometimes elide the \(+\) in \(+ t\) for notational convenience. The rules for evaluation are as follows. \[\inferrule[Var] {} {\sigma, x = v \vdash x^{(t)} \downarrow v^{[t,]}} \qquad \inferrule[Cons] {\sigma \vdash e_1 \downarrow v_1, \ldots, e_n \downarrow v_n} {\sigma \vdash c(e_1, \ldots, e_n)^{(t)} \downarrow c^{[t]}(v_1, \ldots, v_n)}\] \[\inferrule[Call] { \sigma \vdash e_1 \downarrow v_1, \ldots, e_n \downarrow v_n \\ x_1 = v_1, \ldots, x_n = v_n \vdash e \downarrow v } {\sigma, f(x_1, \ldots, x_n) = e \vdash f(e_1, \ldots, e_n)^{(t)} \downarrow v^{[t,]}}\] \[\inferrule[Case] { \sigma \vdash d \downarrow c^{[t_1, \ldots, t_k]}(v_1, \ldots, v_n) \\ \sigma, x_1 = v_1, \ldots, x_n = v_n \vdash e \downarrow v } { \sigma \vdash \mathbf{case}\mathop{} d \mathop{}\mathbf{of}\mathop{} \cdots; c(x_1, \ldots, x_n) \to e; \cdots \mathop{}\mathbf{end}^{(t_0)} \downarrow v^{[t_0, t_1, \ldots, t_k,]} }\] \[\inferrule[IfTrue] { \sigma \vdash c \downarrow \mathbf{True}^{[t_1, \ldots, t_n]}, e_1 \downarrow v } { \sigma \vdash \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} e_1 \mathop{}\mathbf{else}\mathop{} e_2 \mathop{}\mathbf{end}^{(t_0)} \downarrow v^{[t_0, t_1, \ldots, t_n,]} }\] \[\inferrule[IfFalse] { \sigma \vdash c \downarrow \mathbf{False}^{[t_1, \ldots, t_n]}, e_2 \downarrow v } { \sigma \vdash \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} e_1 \mathop{}\mathbf{else}\mathop{} e_2 \mathop{}\mathbf{end}^{(t_0)} \downarrow v^{[t_0, t_1, \ldots, t_n,]} }\] We postpone stating the semantics of \(\mathbf{suppose}\)-expressions to [3](#sec:suppose){reference-type="ref" reference="sec:suppose"}. Note that whenever an expression is evaluated, the value it produces is tagged with a unique identifier for that expression. The rules are defined in such a way that each value ends up carrying traces of expressions that could affect the value. To illustrate, let us elaborate a bit on the [Case]{.smallcaps} rule. The rule says that what a \(\mathbf{case}\)-expression evaluates to may be affected by
(1) the \(\mathbf{case}\)-expression \(t_0\) itself since, trivially, replacing the entire \(\mathbf{case}\)-expression could change its value,
(2) those expressions \(t_1, \ldots, t_n\) that could affect the entirety, even the value constructor, of the value attained by \(d\) that we are casing on, and
(3) those expressions that could affect the value attained by the body \(e\) of the chosen arm.
The rules [IfTrue]{.smallcaps} and [IfFalse]{.smallcaps} are quite similar.[^3]
## Extracting local scores {#subsec:local_score_algo}
We shall need to consider values with other kinds of tags. Define \[v^u \Coloneqq c^{u}(v^u, \ldots, v^u).\] Of particular use are untagged values \(v^{()}\), traced values \(v = v^{[\pm t, \ldots, \pm t]}\) (recall the definitions from [2.1](#subsec:language){reference-type="ref" reference="subsec:language"}), and score-tagged values \(v^{[0, 1]}\). Fix a base typing context \(\Gamma_{\f{types}}\) containing just the type definitions we shall use. To assign scores to expressions in a program \((P, \Gamma)\) with \(\Gamma_{\f{types}} \subseteq \Gamma\), we will need a *test criterion* consisting of the data
(1) a name and signature \(f: (a_1, \ldots, a_n) \to b\) for the function we are testing, where \(a_1, \ldots, a_n, b\) are types defined in \(\Gamma_{\f{types}}\),
(2) untagged input values \((v_{1, k}: a_1, \ldots, v_{n, k}: a_n)\) for \(k = 1, \ldots, N\), and
(3) for each \(k = 1, \ldots, N\), an *assessment function* \[s_{(v_{1, k}, \ldots, v_{n, k})}: \{w \in v^{()} \mid w: b\} \to \{w \in v^{[0, 1]} \mid w: b\}\] which takes a \(b\)-typed value and sets the tag at each node to a score (but does not change the underlying value) such that all the tags in \(s_{(v_{1, k}, \ldots, v_{n, k})}(w)\) are \(1\) if and only if \(w\) is the desired result of \(f(v_{1, k}, \ldots, v_{n, k})\).[^4]
We call the function specified in (1) the *primary function*. For example, suppose \(P\) consisted of just one function definition \(\text{swap}(x, y) = \cdots\) and \(\Gamma\) consisted of the type definition \(\mathbf{Bool2} \Coloneqq \mathbf{Pair}(\mathbf{Bool}, \mathbf{Bool})\) and the function signature \(\text{swap}: \mathbf{Bool2} \to \mathbf{Bool2}\). If we were to assess how well \(\text{swap}\) swaps the booleans in the pair like \((x, y) \mapsto (y, x)\), the test criterion could consist of
- the name and signature \(\text{swap}: \mathbf{Bool2} \to \mathbf{Bool2},\)
- the inputs \(\mathbf{Pair}^{()}(\mathbf{True}^{()}, \mathbf{True}^{()})\), \(\mathbf{Pair}^{()}(\mathbf{False}^{()}, \mathbf{True}^{()})\),\
\(\mathbf{Pair}^{()}(\mathbf{True}^{()}, \mathbf{False}^{()})\), \(\mathbf{Pair}^{()}(\mathbf{False}^{()}, \mathbf{False}^{()})\), and
- the assessment functions \[s_{\mathbf{Pair}(a_1^{()}, a_2^{()})}: \mathbf{Pair}^{()}(b_1^{()}, b_2^{()}) \mapsto \mathbf{Pair}^{(\delta(b_1, a_2) + \delta(b_2, a_1))/2} (b_1^{\delta(b_1, a_2)}, b_2^{\delta(b_2, a_1)})\] where \(\delta(x, y)\) is \(1\) if \(x = y\) and \(0\) otherwise.
The assessment functions are defined in such a way that the resulting tag to the \(\mathbf{Pair}\) constructor measures how close the value is to the desired value \(\mathbf{Pair}(a_2^{()}, a_1^{()})\), and the resulting tag to the boolean in each coordinate measures how close the respective coordinate is to the desired value. Notice how the assessment functions give perfect scores when the behavior is as desired: \[s_{\mathbf{Pair}(a_1^{()}, a_2^{()})}(\mathbf{Pair}^{()}(a_2^{()}, a_1^{()})) = \mathbf{Pair}^{1}(a_2^{1}, a_1^{1}).\] Let us return to the general setup \((1)\)-\((3)\) and explain how to compute a score for each expression in the program \(P\). We will assume that \(f\) always terminates.[^5] The process for obtaining expression-level scores is as follows.
(a) For each \(k = 1, \ldots, N\), evaluate the function call \(f(\f{tr}(v_{1, k}), \ldots, \f{tr}(v_{n, k}))\) and call its (traced) value \(w_k\). Here, \(\f{tr}(v)\) is the traced value obtained from the untagged value \(v\) by setting each tag to the empty trace \([]\).
(b) For each \(k = 1, \ldots, N\), compute \(w_k' \defeq s_{(v_{1, k}, \ldots, v_{n, k})}(\f{untag}(w_k))\) where \(\f{untag}(w_k)\) is the untagged value obtained from the traced value \(w_k\) by replacing each tag with \(()\).
(c) For each node in \(w_k\), extract the positively charged tags \(+ t_1, \ldots, + t_p\) from its trace, the negatively charged tags \(-u_1, \ldots,-u_q\) from its trace, and the tag \(s \in [0, 1]\) of the corresponding node in \(w_k'\); then collect all the tuples \((t_i, s)\) for \(i = 1, \ldots, p\) and \((u_i, 1-s)\) for \(i = 1, \ldots, q\) into a list \(\ell\).
(d) An expression in \(P\) with the tag \(t\) is assigned the score, \[S_P(t) \defeq \frac {\sum_{(t, s) \in \ell} s} {\sum_{(t, s) \in \ell} 1}\] averaging over all the scores in \(\ell\) associated to the tag \(t\).
Note how a positively charged tag \(+t\) in some trace in \(w_k\) contribute to the average defining \(S_P(t)\) with the score produced by \(s_{(v_{1, k}, \ldots, v_{n, k})}\) corresponding to the trace whereas a negatively charged tag \(-t\) contribute with the inversion of that score. The net effect is that an expression with tag \(t\) that appears in a trace positively charged (resp. negatively charged) is incentivized to maximize (resp. minimize) the score corresponding to that trace if we are to maximize the score \(S_P(t)\). For the rest of the paper, we shall fix a test criterion (with respect to the fixed \(\Gamma_{\f{types}}\)). Moreover, from now on, we will only consider programs \(P\) that contain the type definitions \(\Gamma_{\f{types}}\) and a function definition for the primary function (1). In particular, for each such program \(P\), we get a scoring function \(S_P\). If \(e\) is an expression in \(P\) with expression tag \(t\), we write \(S_P(e) \defeq S_P(t)\). Also, when the context resolves ambiguity, we shall omit the subscript \(P\), writing \(S\) instead of \(S_P\).
# `suppose`-expressions {#sec:suppose}
Let us informally consider the problem of evolving code for the minimum function \(\f{min}: \Z \times \Z \to \Z.\) We may assess a candidate implementation based on some example arguments \(A \subseteq \Z \times \Z\) ranging over small numbers. Start with the initial program \[\tag{\(P_0\)} \f{min}(n, m) = 0.\] This is not faring particularly well: perhaps, there are a few \((n, m) \in A\) with \(0 = n \leq m\) or \(0 = m \leq n\) but they make up an increasingly small fraction as \(|A|\) grows. If we try to mutate \(P_0\) by replacing the body of \(\f{min}(n, m)\) by various randomly chosen expressions and testing it on \(A\), it is not hard to find the improved program \[\tag{\(P_1\)} \f{min}(n, m) = n,\] which performs a lot better: it is correct whenever \(n \leq m\) which is roughly \(\frac12\) of the time. However, at this point, we stagnate. The next program we want is \[\tag{\(P_2\)} \f{min}(n, m) = \mathbf{if}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{else}\mathop{} m \mathop{}\mathbf{end},\] but this is quite a leap from \(P_1\) and would therefore take a long time to discover through picking out random expressions. More complex examples are even less tractable. How do we overcome this problem of evolving \(\mathbf{if}\)-conditionals? \(\mathbf{suppose}\)-expressions allow us to separate the problem of evolving the condition from the problem of evolving the \(\mathbf{else}\)-branch. In the above informal example, small intermediate steps between \(P_1\) and \(P_2\) could be \[\begin{aligned}
&\f{min}(n, m) = n \\ \becomes\ & \f{min}(n, m) = \mathbf{suppose}\mathop{} \mathbf{True} \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{end} \\ \becomes\ & \f{min}(n, m) = \mathbf{suppose}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{end} \\ \becomes\ & \f{min}(n, m) = \mathbf{if}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{else}\mathop{} n \mathop{}\mathbf{end}, \\ \becomes\ & \f{min}(n, m) = \mathbf{if}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{else}\mathop{} m \mathop{}\mathbf{end}, \end{aligned}\] Conceptually, \(\mathbf{suppose}\)-expressions encourage the condition, initially constant \(\mathbf{True}\), to evolve to be \(\mathbf{True}\) whenever the body is correct and \(\mathbf{False}\) otherwise. When the condition for the \(\mathbf{suppose}\)-expression is good enough, we will promote the \(\mathbf{suppose}\)-expression into an \(\mathbf{if}\)-conditional: \[\mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end} \quad \becomes \quad \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{else}\mathop{} e \mathop{}\mathbf{end}.\] This evolutionary strategy is established in [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}. With this discussion in mind, let us now state the semantics of \(\mathbf{suppose}\)-expressions. We use the notation established in [2](#sec:local){reference-type="ref" reference="sec:local"}. Define \({-}{-}t \defeq +t\). The rules are then as follows. \[\inferrule[SupposeTrue] { \sigma \vdash c \downarrow \mathbf{True}^{[t_1, \ldots, t_n]}, e \downarrow v } { \sigma \vdash \mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end}^{(t_0)} \downarrow v^{[t_0, t_1, \ldots, t_n,]} }\] \[\inferrule[SupposeFalse] { \sigma \vdash c \downarrow \mathbf{False}^{[t_1, \ldots, t_n]}, e \downarrow v } { \sigma \vdash \mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end}^{(t_0)} \downarrow v^{[t_0,-t_1, \ldots,-t_n,]} }\] Thus, if we disregard traces, \(\mathbf{suppose}\)-expressions just evaluate their bodies. However, if we follow the Local Scoring model set out in [2](#sec:local){reference-type="ref" reference="sec:local"}, we see that if \(c\) evaluates to \(\mathbf{True}\), the better \(e\) performs, the higher score is assigned to \(c\), whereas if \(c\) evaluates to \(\mathbf{False}\), the better \(e\) performs, the lower score is assigned to \(c\). We will need the following lemma in [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}.
# Some mutation principles {#sec:principles}
We discuss here some simple principles for how to mutate programs by leveraging Local Scoring. These will be used in combination with the methods presented in [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}.
## Mutually exclusive mutation {#subsec:mutex}
Changing a single expression has a high chance of breaking the program, but changing multiple expressions has an even higher chance of doing so. It therefore makes sense to change just one expression in each mutation step. A simple scheme to mutate a program \(P\) is to first pick a random expression. The probability for picking a given expression \(e\) should be proportional to \(1-S(e)\), so that we are more likely to mutate what is broken and we never mutate what works perfectly.[^6] This expression is then replaced by a new, randomly generated expression of the same type (with expression tags disjoint from the expression tags used in \(P\)), giving a new program \(P'\), the mutant of \(P\).
## Greedy Principle of Mutation {#subsec:greedy}
Suppose we use a hill-climbing algorithm which has as state a program \(P\) that it repeatedly tries to mutate until it finds a good mutant, a successor, which then replaces \(P\), repeating this process until the program solves some task perfectly. How do we determine if a given mutant \(P'\) should replace \(P\) as its successor? If we go purely by some kind of global score based on the behavior of the program, as in [1](#sec:introduction){reference-type="ref" reference="sec:introduction"}, we fail to evolve the conditions for \(\mathbf{suppose}\)-expressions changing which has no exterior effect on the program behavior. We therefore propose another criterion. Recall that we mutate a program \(P\) to a program \(P'\) by replacing some expression \(e\) with a new expression \(e'\). The *Greedy Principle of Mutation* stipulates that \(P'\) should be the successor to \(P\) if and only if \(S_{P'}(e') > S_P(e)\). The principle is since the mutation can (and usually does) change the scores of expressions that are not \(e\) and yet we only focus on how the score of \(e\) changes. Notice how the condition is \(S_{P'}(e') > S_P(e)\) as opposed to \(S_{P'}(e') \geq S_P(e)\). This is a deliberate choice. It means that we prefer doing no mutation to mutating: we only accept mutations that strictly improve the expression they change. We shall refer to this property as *mutation conservativity*.
# Cyclic Evolution {#sec:cyclic}
We have described \(\mathbf{suppose}\)-expressions but we have yet to explain how \(\mathbf{suppose}\mathop{}\) should be introduced during evolution and when these should be promoted into \(\mathbf{if}\)-conditionals. When mutating a program \(P\), we would like to sometimes replace an expression \[e \quad \becomes \quad \mathbf{suppose}\mathop{} \mathbf{True} \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end}\] after which mutations can begin to improve \(\mathbf{True}\) to a more meaningful condition. If we were to just follow the simplistic scheme set out in [4](#sec:principles){reference-type="ref" reference="sec:principles"}, such replacement might happen but the resulting program will never succeed the existing program because the score of the replaced expression is the same as the score of the original expression \(e\). One could circumvent this issue by making an exception to mutation conservativity, making this specific type of mutation always succeed. This, however, is not a very good solution to the problem at hand: the program would then grow unbounded as we add more and more redundant \(\mathbf{suppose}\)-expressions. What we propose instead is to divide the evolution into phases of growth, phases of mutation, and phases of reduction; similar to how a person who is trying to gain muscle might divide their eating into phases of bulking and phases of cutting. Specifically, instead of proceeding one mutation at a time, the basic unit of the algorithm will be a *cycle* consisting of four phases:
(1) *Stretching* in which a program is replaced by an equivalent, more redundant program by introducing redundant \(\mathbf{suppose}\)-expressions, redundant \(\mathbf{if}\)-conditionals, redundant \(\mathbf{case}\)-expressions, and redundant new functions.
(2) *Mutation* in which a number small changes are made to the program, replacing the program whenever a mutation satisfying the Greedy Principle of Mutation is found.
(3) *Rewinding* in which we try to undo the stretches done in (a).
(4) *Compression* in which we try to rewrite the program to an equivalent, smaller form.
The algorithm we propose evolves programs by repeatedly applying cycles until a program performing perfectly (in accord with the test criterion) is found. We describe basic ways to implement each phase. We emphasize that each can be extended and refined in many ways, some of which will be discussed in [7](#sec:future){reference-type="ref" reference="sec:future"}.
## Stretching {#subsec:stretch}
The stretching phase makes the program more redundant, setting the scene for the mutation phase to turn that redundancy into useful structure. The stretching phase consists of \(N_\f{stretch}\) *stretches* for some fixed constant \(N_\f{stretch} > 0\). Each stretch picks out a random expression in the program with the probability of choosing an expression \(e\) being proportional to \(1-S(e)\). The stretched program \(P'\) is constructed from \(P\) by picking randomly among the following possibilities:
(1) replacing \(e \becomes \mathbf{suppose}\mathop{} \mathbf{True} \mathop{}\mathbf{then}\mathop{} e \mathop{}\mathbf{end},\)
(2) if \(e = \mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} e' \mathop{}\mathbf{end}\) and \(S(c) > S(e')\), replacing \(e \becomes \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} e' \mathop{}\mathbf{else}\mathop{} e' \mathop{}\mathbf{end}\),
(3) picking a variable \(x: \tau\) bound in the context of \(e\) and replacing \[\begin{aligned}
e \becomes\ &\mathbf{case}\mathop{} x \mathop{}\mathbf{of}\mathop{} \\& c_1(x_1, \ldots, x_{n_1}) \to e[x/c_1(x_1, \ldots, x_{n_1})]; \\& \cdots; \\& c_k(x_1, \ldots, x_{n_k}) \to e[x/c_k(x_1, \ldots, x_{n_k})] \\ &\mathop{}\mathbf{end}, \end{aligned}\] where (a) \(x_1, \ldots\) are variables that do not participate in \(e\), (b) \(c_1, \ldots, c_k\) are the value constructors for \(\tau\): \[\tau \Coloneqq c_1(\cdots \text{(\(n_1\) arguments)}) \mid \cdots \mid c_k(\cdots \text{(\(n_k\) arguments)}),\] and (c) \(e[x/e']\) denotes \(e\) with all occurrences of \(x\) replaced by \(e'\), or
(4) for \(x_1: a_1, \ldots, x_n: a_n\) the variables bound in the context of \(e: b\), adding a new function \[\begin{aligned}
&f: (b, a_1, \ldots, a_n) \to b \\ &f(y, x_1, \ldots, x_n) = y \end{aligned}\] and replacing \(e \becomes f(e, x_1, \ldots, x_n).\)
In each case, the replacing expression should be tagged with some tags disjoint from those used in \(P\), but we elide this for notational simplicity. Let us elaborate on the condition \(S(c) > S(e')\) in stretch (2). The justification for this condition comes from [\[lem:suppose_true\]](#lem:suppose_true){reference-type="ref" reference="lem:suppose_true"} which precludes stretch (2) from happening when \(c = \mathbf{True}\). In this way, when a stretch (1) introduces a \(\mathbf{suppose}\)-expression, the condition must first improve (which can happen in the mutation phase) before a stretch (2) can promote it into an \(\mathbf{if}\)-conditional. As such, we don't end up immediately creating a lot of redundant \(\mathbf{if}\)-conditionals.
## Mutation
The mutation phase consists of \(N_\f{mutate}\) mutations, successful or not, where \(N_\f{mutate} > 0\) is some fixed constant. A mutation changes the program in the way described [4.1](#subsec:mutex){reference-type="ref" reference="subsec:mutex"}. A mutation is considered successful if it satisfies the Greedy Principle of Mutation (see [4.2](#subsec:greedy){reference-type="ref" reference="subsec:greedy"}) in which case the mutant program replaces the program and the search continues.
## Rewinding
The purpose of stretching was to create new program structures that the mutation phase could make use of. However, not all stretches create useful program structures. If we return to our informal example of [3](#sec:suppose){reference-type="ref" reference="sec:suppose"} with evolving the \(\min\) function, we could have the stretch \[\begin{aligned}
&\min(n, m) = \mathbf{if}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{else}\mathop{} n \mathop{}\mathbf{end} \\ \becomes\ & \min(n, m) = \mathbf{if}\mathop{} n \leq m \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{else}\mathop{} (\mathbf{suppose}\mathop{} \mathbf{True} \mathop{}\mathbf{then}\mathop{} n \mathop{}\mathbf{end}) \mathop{}\mathbf{end} \end{aligned}\] But that wouldn't be of much use because in the \(\mathbf{else}\)-branch \(n\) is never the right expression. So, unless the second \(n\) is mutated, the condition for the \(\mathbf{suppose}\)-expression will just converge to \(\mathbf{False}\). We must be able to deal effectively with such to stop the program that is being evolved from accumulating redundancy. After the mutation phase has completed, we propose to clean up the stretches by them. Specifically, for each stretching operation \(s\), we define a corresponding *rewinding operation* \(\overline s\), which is a transformation of programs that undoes (is left-inverse to) \(s\). Then, if stretches \(s_1, \ldots, s_{N_\f{stretch}}\) were made in the stretching phase in that order, the rewinding phase applies rewinding operations \(\overline s_{N_\f{stretch}}, \ldots, \overline s_1\) in that order (notice the reversal of order). For a stretch \(s\) out of the \(N_\f{stretch}\) stretches done in the stretching phase, we will now describe its corresponding rewinding operation \(\overline s\). \(\overline s\) depends on which type (1)-(4) of stretch \(s\) is:
- If \(s\) is a stretch of type (1), \(\overline s\) is the following program transformation. Let \(e_1, \ldots, e_n\) be the \(\mathbf{suppose}\)-expressions that originated from \(s\). Let \(c_1, \ldots, c_n\) be their respective condition expressions and \(e_1', \ldots, e_n'\) their respective bodies. For \(i = 0, \ldots, n\), replace \(e_i \becomes e_i'\) if \(S(c_i) \leq S(e_i)\).
- If \(s\) is a stretch of type (2), \(\overline s\) is the following program transformation. Let \(e_1, \ldots, e_n\) be the \(\mathbf{if}\)-conditionals that originated from \(s\). Let \(c_1, \ldots, c_n\) be their respective condition expressions and \(e_1', \ldots, e_n'\) their respective \(\mathbf{True}\)-branches. For \(i = 0, \ldots, n\), replace \(e_i \becomes \mathbf{suppose}\mathop{} c_i \mathop{}\mathbf{then}\mathop{} e_i' \mathop{}\mathbf{end}\) if the latter expression attains a score in the program with this replacement made at least the score \(S(e_i)\).
- If \(s\) is a stretch of type (3), \(\overline s\) is the following program transformation. Let \(x\) be the variable chosen in \(s\). Let \(e_1, \ldots, e_n\) be the expressions \(\mathbf{case}\mathop{} x \mathop{}\mathbf{of}\mathop{} \cdots\) that originated from \(s\). Let \(c_1(x_1, \ldots, x_{n_1}), \ldots, c_k(x_1, \ldots, x_{n_k})\) be the cases that appear in each of those \(\mathbf{case}\)-expressions. For \(i = 1, \ldots, n\), do the following.
(a) Look for a \(1 \leq j \leq k\) such that (i) \(n_j > 0\), (ii) the expression \(e_{i, j}' \defeq e_{i, j}[c_j(x_1, \ldots, x_{n_j})/x],\) where \(e_{i, j}\) is the body of the \(\mathbf{case}\)-expression \(e_i\) to the case \(c_j\), does not reference the variables \(x_1, \ldots, x_{n_j}\), and (iii) after replacing \(e_i\) with \(e_{i, j}'\), the new \(e_{i, j}'\) attains a score at least the score \(S(e_i)\). If such a \(j\) exists, replace \(e_i \becomes e_{i, j}'\) in the program.
(b) If not, look instead for a \(1 \leq j \leq k\) such that (i) \(n_j = 0\) and (ii) after replacing \(e_i\) with \(e_{i, j}\), the new \(e_{i, j}\) attains a score at least the score \(S(e_i)\). If such a \(j\) exists, replace \(e_i \becomes e_{i, j}\) in the program.
(c) If such a \(j\) could not be found either, do nothing.
- If \(s\) is a stretch of type (4), \(\overline s\) is the following program transformation. Let \(e_1, \ldots, e_n\) be the call-expressions that originated from \(s\). Let \(f\) be the name of the function they are calling. Suppose its definition in the program is \(f(x_1, \ldots, x_k) = e.\) If \(e\) contains a subexpression \(f(\cdots)\), do nothing. Otherwise, do the following. For each \(i = 1, \ldots, n\), replace \(e_i = f(a_1, \ldots, a_k) \becomes e[x_1/a_1]\cdots[x_k/a_k]\). If after this transformation the expression \(f(\cdots)\) appears nowhere in the program, furthermore remove the function definition \(f(x_1, \ldots, x_k) = e\) from the program.
As before, we elide tags in notation, implicitly tagging expressions with unique tags. And again, \(e[a/b]\) denotes the expression \(e\) with all occurrences of \(a\) replaced by \(b\). Note that in each of these rewinding operations, we talk of possibly multiple expressions \(e_1, \ldots, e_n\). This is because an expression created by one stretch may be duplicated by another (stretches of type (2) or (3)). It is therefore necessary to track what expressions originate in what stretches. This can be done by tagging each expression by some extra bookkeeping data of which we shall not deal with the details. The condition \(S(c_i) \leq S(e_i)\) in \((\overline 1)\) is again justified by [\[lem:suppose_true\]](#lem:suppose_true){reference-type="ref" reference="lem:suppose_true"}. In particular, the condition is only fulfilled if the condition \(c_i\) is at least as bad as the initial condition \(\mathbf{True}\). In effect, we throw away a \(\mathbf{suppose}\mathop{}\) if its condition didn't manage to improve at all during the mutation phase.
## Compression {#subsec:compression}
The final phase is compression, replacing the program by an equivalent, shorter form. There are indeed many ways to approach this problem of removing excess; we present one which we found works reasonably well in practice. As input for the algorithm, we need to run the program on the examples in the test criterion that we fixed in [2.2](#subsec:local_score_algo){reference-type="ref" reference="subsec:local_score_algo"} (see step (a)) and collect for each expression \(e\) an *input--output table*: the set with an element \((\sigma, \f{untag}(v))\) for each evaluation \(\sigma \vdash e \downarrow v\) done. We say that an expression \(e\) is *linear* if either \(e = x^{(t)}\) for \(x\) a variable or \(e = c(e_1, \ldots, e_n)^{(t)}\) for \(c\) a value constructor and \(e_1, \ldots, e_n\) linear expressions. Recall that we defined a program to consist of a list of function declarations \(f(x_1, \ldots, x_n) = e\). The compression algorithm first applies the following procedure to each of the bodies \(e\).
(1) If \(e = c(e_1, \ldots, e_n)^{(t)}\) for \(c\) a value constructor, apply the procedure recursively to \(e_1, \ldots, e_n\).
(2) If \(e = \mathbf{suppose}\mathop{} c \mathop{}\mathbf{then}\mathop{} b \mathop{}\mathbf{end}^{(t)}\), apply the procedure recursively to \(c\) and \(b\); and if after doing so \(e\) became \(e' = \mathbf{suppose}\mathop{} c' \mathop{}\mathbf{then}\mathop{} b' \mathop{}\mathbf{end}^{(t)}\) where \(c'\) admits no variable \(x^{(u)}\) as a subexpression, further replace \(e' \becomes b'\).
(3) If neither of the above cases apply, try to find a linear expression \(\ell\) with the behavior specified by the input--output table of \(e\). (The algorithm for finding \(\ell\) is postponed to [5.4.1](#subsubsec:lexprs){reference-type="ref" reference="subsubsec:lexprs"}.) If such \(\ell\) exists, pick one and replace \(e \becomes \ell\) (in which we tag \(\ell\) with expression tags disjoint from those used in \(P\)), and we're done. If not, continue to the next steps.
(4) If \(e = \mathbf{if}\mathop{} c \mathop{}\mathbf{then}\mathop{} a \mathop{}\mathbf{else}\mathop{} b \mathop{}\mathbf{end}^{(t)}\), apply the procedure recursively to \(c\), \(a\), and \(b\). Suppose that after doing so \(e\) became \(e' = \mathbf{if}\mathop{} c' \mathop{}\mathbf{then}\mathop{} a' \mathop{}\mathbf{else}\mathop{} b' \mathop{}\mathbf{end}^{(t)}\). If \(c' = \mathbf{True}^{(u)}\) for some \(u\), further replace \(e' \becomes a'\). Or, if \(c' = \mathbf{False}^{(u)}\) for some \(u\), further replace \(e' \becomes b'\).
(5) If \(e = f(e_1, \ldots, e_n)^{(t)}\) for \(f\) a function, apply the procedure recursively to \(e_1, \ldots, e_n\).
After this, the algorithm removes from the program function definitions \(g(\cdots) = \cdots\) for functions \(g\) that are not reachable from the primary function \(f\) in the call graph. The resulting program is the result of the compression phase. Note that, in each step the \(w_k\) used in the definition of \(S\) in [2.2](#subsec:local_score_algo){reference-type="ref" reference="subsec:local_score_algo"} remain unchanged if we disregard traces. This is the sense in which the original program is equivalent to the compressed program, that is, they are equivalent on known input.
### An algorithm for finding linear expressions {#subsubsec:lexprs}
We are presented with the problem of listing the linear expressions \(\ell\) satisfying a given specification. Let us be more precise. We are given a type \(\tau\), a typing context \(\Gamma\) containing a definition of \(\tau\), and an input--output table \(T = \{(\sigma_1, v_1), \ldots, (\sigma_n, v_n)\}\) such that for each \(i = 1, \ldots, n\), (i) \(\sigma_i\) is an evaluation context and \(v_i\) an untagged value of type \(\tau\) and (ii) all variables found in \(\sigma_i\) are found in \(\Gamma\) and vice versa. The task is then to compute the (untagged) linear expressions \(\ell\) that evaluate[^7]to \(v_i\) in context \(\sigma_i\) for every \(i = 1, \ldots, n\). If \(n = 0\) (that is, the input--output table is empty), the problem is equivalent to finding all linear expressions of type \(\tau\). This possibly infinite list can easily be computed by a breadth-first search applied to the recursive definition of linear expressions given near the beginning of [5.4](#subsec:compression){reference-type="ref" reference="subsec:compression"}. Now, suppose \(n > 0\). For convenience, we shall rename the variables in \(\Gamma\) so that they are called \(x_1, \ldots, x_n\). Recall the notation for tagged values established in [2.2](#subsec:local_score_algo){reference-type="ref" reference="subsec:local_score_algo"}. An *\(\ell\)-value* is a tagged value \(\in v^{\mathscr P(\{\circ, 1, \ldots, n\})}\) where \(\mathscr P\) denotes the power set operator. Moreover, given an \(\ell\)-value \(w = c^A(w_1, \ldots, w_k)\), define \[\ell(w) \defeq \{x_i^{()} \mid i \in \N \cap A\} \cup \begin{cases} \{c^{()}(w_1', \ldots, w_k') \mid w_i' \in \ell(w_i)\}, & \text{if \(\circ \in A\);} \\ \emptyset, & \text{otherwise.} \end{cases}\]
To an input--output pair \((\sigma, v)\) where \(v = c^{()}(v_1, \ldots, v_k),\) we associate an \(\ell\)-value \[w(\sigma, v) \defeq c^{\{\circ\} \cup \{i \mid (x_i = v) \in \sigma\}} (w(\sigma, v_1), \ldots, w(\sigma, v_k)).\]
Given \(\ell\)-values \(w_1 = c_1^A(w_{1, 1}, \ldots, w_{1, p})\) and \(w_2 = c_2^B(w_{2, 1}, \ldots, w_{2, q})\), define \[w_1 * w_2 \defeq \begin{cases} c_1^{A \cap B}(w_{1, 1} * w_{2, 1}, \ldots, w_{1, p} * w_{2, p}), & \text{if \(c_1 = c_2\);} \\ c_1^{A \cap B \setminus \{\circ\}}(w_{1, 1}, \ldots, w_{1, p}), & \text{otherwise.} \end{cases}\]
We can now describe the algorithm for finding linear expressions. Combining the lemmata above, we find that the set \[\ell(w(\sigma_1, v_1) * \cdots * w(\sigma_n, v_n))\] contains exactly the linear expressions coforming to the input--output table \(T\). Since the definitions of \(\ell\), \(w\), and \(*\) deal with finite data in finite numbers of steps, we can compute this set in its entirety by unfolding definitions, thus giving an algorithm.
# A proof-of-concept implementation {#sec:poc}
We provide a proof-of-concept implementation which can be found at <https://gitlab.com/maxvi/ssgp>. Excluding comments and blank lines, it consists of a little less than 2700 lines of Haskell code. Great care was taken to document the API and every piece of code, amounting to over 1300 lines of comments. Various QuickCheck tests are also included. We emphasize that the implementation is a *minimum viable product*; for instance, it is completely unoptimized (see [7.2](#subsec:optimize){reference-type="ref" reference="subsec:optimize"}). The implementation deviates from a naïve implementation of the theoretical ideas described above in a number of ways, many of which we shall cover in this section. It produces code written in what is, sparring \(\mathbf{suppose}\)-expressions, a subset of Haskell.
## Results {#subsec:results}
We have chosen a few examples, tabulated below, which highlight some strengths and weaknesses of our implementation. Each example was run five times. Each run was set to have a maximum of 10 cycles (see [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}), after which we would declare the evolution to not converge. The constants used were \(N_\f{stretch} = 3\) and \(N_\f{mutate} = 300\). The last column records the average number of iterations (stretches, mutations, rewindings, compressions) that occured before evolution converged to correct code. At each iteration, a new program is proposed and this program is then run on a pre-fixed set of examples through which local scores are assigned. Every type is an Algebraic Data Type. Integers are encoded by their binary representation. Lists are defined recursively in a left-biased way as \([\Z] \Coloneqq \mathbf{Cons}(\Z, [\Z]) \mid \mathbf{Nil}.\) There is no standard library; each example can only use the functions specified in the third column, nothing else. If one were to add more useful functions there, one would get better convergence rates. Let us elaborate on the behavior of some of the tabulated functions. picks out the smallest number from a list, defaulting to \(0\) if the list is empty. inserts a number into a list of numbers before the first entry that is greater than or equal to the number. lists the ASCII codes for the string . The rest are self-explanatory.
## Comparison with other works {#subsec:compare}
Koza established the *even-parity problem* as a benchmark for Genetic Programming. The goal is to evolve code for the function \([\mathbf{Bool}] \to \mathbf{Bool}\) which returns \(\mathbf{True}\) if the list has even number of \(\mathbf{True}\)s and \(\mathbf{False}\) otherwise. Briggs and O'Neill use this problem to compare a number of approaches found in literature. We reproduce their benchmarks below, adding a benchmark of our own proof-of-concept implementation. The last column records estimates for the smallest number of program assessments needed during evolution for an algorithm to find a correct solution with 99% probability. A *program assessment* scores a program by running it on a number of examples. We ran our proof-of-concept implementation 100 times on the even-parity problem. Every run eventually converged on correct code, with the average number of programs tested before convergence being 335.3. The best run needed to test just 79 programs and the worst 677 programs. Generalizing from our data, we find that if we stop evolution after 674 programs, we arrive at a correct solution with \(99\%\) probability. The only built-in function made available from the onset was the binary operation, \(\f{not}: \mathbf{Bool} \to \mathbf{Bool}.\)
## Local scores and deep recursion
For scores \(S(e)\), we use instead a weighted average in which each summand is weighted by the reciprocal of the call-stack depth. In effect, deep recursive calls have less influence on scores.
## Continuous scoring {#subsec:continuous}
The Hello World example uses a scoring function, meaning that it takes into account how close the string that \(\f{helloworld}()\) produces is to the desired string , instead of just rewarding \(1\) when correct and \(0\) when incorrect (in which case evolving the correct code would be virtually impossible). Local scoring means that if \(\f{helloworld}()\) returned the string , we trace the problematic character, G, back to the expression that produced it and then change that expression little by little, inching in on the correct character, W.
## Timeouts and restrictions to recursion {#subsec:rec}
A fundamental issue is that the problem of bounding the running time of an arbitrary program in a Turing-complete language is undecidable as was famously shown by Turing in 1936. To overcome this issue, we imposed restrictions on what is allowed during execution:
(1) We impose a limit on how many function calls can be made in one program execution. If exceeded, the local score of every expression that was evaluated during the execution gets punished.
(2) We impose restrictions on what forms of recursions are allowed. Specifically, we shall only allow a call \(f(v_1, \ldots, v_n)\) if there is an \(i\) such that for all calls \(f(w_1, \ldots, w_n)\) on the call-stack, \(|v_i| < |w_i|\). Here, \(|v|\) denotes the of a value \(v\), that is, its number of nodes. If this condition is violated, the local score of every expression on the (the expression-level analogue of the call-stack) gets punished.
The restriction (2) alone ensures termination, as the following lemma shows.
## Controlled Recursion {#subsec:controlled}
When a recursive function is broken, it is usually catastrophically broken. This is because errors propagate (and multiply) upwards through recursion. This affects negatively the quality of the local scores which we want to reflect to a solution. For example, if we were to define natural numbers inductively using successors, \(\N \Coloneqq 0 \mid S(\N),\) then the recursive definition \[\begin{aligned}
0 \oplus m &\defeq m \\ S(n) \oplus m &\defeq n \oplus m \end{aligned}\] is only one \(S\) away from being the correct definition of addition (\(S(n) \oplus m = S(n \oplus m)\)), yet in terms of actual behavior it is quite far away: \(n \oplus m = m\). To deal with this issue, we employ a technique we shall call *Controlled Recursion*, in which we answer recursive calls using training data. For example, if we were trying to evolve \(\oplus\) to be addition, we would, when producing the values \(w_1, \ldots, w_N\) in [2.2](#subsec:local_score_algo){reference-type="ref" reference="subsec:local_score_algo"}, evaluate the recursive call \(n \oplus m\) in the body of the second equation above as \(n + m\) instead of using the actual code for \(\oplus\). That means that the output only differs by \(1\) from the desired output, and the score is accordingly better (that is, if we use a continuous scoring function as in [6.4](#subsec:continuous){reference-type="ref" reference="subsec:continuous"}) than if we were to answer recursive calls using the actual code. To discourage degenerate self-referential definitions like \(S(n) \oplus m \defeq S(n) \oplus m,\) we only answer recursive calls like this that are than the calls on the call-stack, a la the restriction (2) we imposed in [6.5](#subsec:rec){reference-type="ref" reference="subsec:rec"}.
## Stretches
Since the examples did not appear to need it, we disabled the stretch (4) described in [5.1](#subsec:stretch){reference-type="ref" reference="subsec:stretch"}. Moreover, for stretch (1), we only introduce \(\mathbf{suppose}\)-expressions at the top of bodies of equations.[^8]
## Limitations
A significant issue is the performance suboptimality of our proof-of-concept implementation. With optimized code execution, we could perhaps try many times more mutations in the same amount of time, thereby searching much deeper for new mutations. We discuss this further in [7.2](#subsec:optimize){reference-type="ref" reference="subsec:optimize"}. Another striking issue is that of overfitting. For example, overfitting obstructs discovery of correct code for the function \(\f{reverse}: [\Z] \to [\Z]\) reversing the items of a list. Instead of attacking the problem algorithmically, it generates more and more cases: one for the empty list \([]\), one for \([x_1]\), one for \([x_1, x_2]\), and so on. While it obviously does get closer to a correct solution in terms of behavior, it does not approach a correct solution in terms of code. The issue is that lists are defined recursively in a left-biased way, making operations on the beginning of the list far easier to discover than operations on the end. A human is able to immediately tell that such code is bad style; perhaps if one taught the same intuition to a neural network, that could be used to stir away from overfitting. See [7.3](#subsec:ml){reference-type="ref" reference="subsec:ml"} for more discussion. Overfitting was not an issue in any of the converging examples from [6.1](#subsec:results){reference-type="ref" reference="subsec:results"}.
# Extensions and future work {#sec:future}
## Crossover and parallelism
In contrast to a lot of the literature (see [6.2](#subsec:compare){reference-type="ref" reference="subsec:compare"}), we use what is a kind of hill-climbing algorithm: at every point in the algorithm, we have exactly one program. It is typical to instead evolve a population of programs, mutating them and using a procedure called *crossover* to combine two programs producing one or more offspring programs. We remark that this could be added on top of our hill-climbing method, making for highly parallelized Genetic Programming: one could do \(C\) cycles of hill-climbing (see [5](#sec:cyclic){reference-type="ref" reference="sec:cyclic"}) for each program in the population \(\{P_{(1)}, \ldots, P_{(n)}\}\) in parallel, yielding a new population \(\{P_{(1)}', \ldots, P_{(n)}'\}\) which is then extracted for crossover combining programs at random and natural selection killing of all but \(n\) of the resulting programs. This procedure is repeated until a solution is found. With this, many hill-climbing processes run in parallel but still once in a while, avoiding too much duplicated effort.
## Optimizations {#subsec:optimize}
By far the most significant component performance-wise is that which executes programs. As we were more focused on the theoretical background, no serious attempt at optimizing this was made in our proof-of-concept implementation written in Haskell. Let us discuss some techniques and ideas that could be used in a performant implementation.
(1) Use a more performant language such as Rust or C.
(2) Currently, we represent code as an Abstract Syntax Tree which is quite inefficient for a number of reasons. For instance, execution requires dereferencing a lot of pointers resulting in many cache misses. A more efficient, flat representation of code, such as a stack machine, could be used instead.
(3) Expressions that attain high scores could be compiled, making for extremely performant execution of those parts of the program that are unlikely to change. Compilation is typically a heavy process and it is therefore unlikely that it is worth it to recompile the entire program on every mutation, but perhaps a good balance could be struck.
(4) Since we are running the same (or similar) code again and again, it would make sense to make heavy use of memoization, trading CPU cycles for increased memory usage.
(5) A significant amount of time is spent on testing out bad mutations, typically scoring extremely poorly. One could use heuristics for quickly discarding large swaths of bad mutations with only few false positives. Local Scoring is particularly apt for this. For example, if one was to test a mutation \(e \becomes e'\) of an expression \(e\), one could first compute the score of \(e'\) with respect to a small, randomly chosen subset of \(w_1, \ldots, w_N\) (see [2.2](#subsec:local_score_algo){reference-type="ref" reference="subsec:local_score_algo"}) chosen so that for each \(w_i\) in this subset the expression \(e'\) appears somewhere in one of its traces, and then preemptively discard the mutation if this is significantly smaller than \(S(e)\).[^9]
(6) While our proof-of-concept implementation does ensure that the exact same program isn't tried multiple times, one could employ more advanced ways of ensuring that equivalent mutations aren't attempted multiple times. One would need to specify algorithms that are able to tell if a mutation on a program \(P\) corresponds to a mutation on a program \(P'\) equivalent to \(P\) by a number of stretches (see [5.1](#subsec:stretch){reference-type="ref" reference="subsec:stretch"}). Perhaps a hash function could be devised for this purpose.
## Integration with Deep Learning methods {#subsec:ml}
We are choosing new code to try out essentially at random instead of doing so intelligently and drawing from an intuitive understanding of what changes are needed in what situations. Recently, there has been major breakthroughs on the subject of using Deep Learning methods to generate code (e.g., and ). An interesting direction would be to hybridize the approaches, combinining the *overview* of Local Scoring with the *intuition* of Deep Learning. Let us speculate a bit on what that could look like.
(1) One could incorporate into the scores an assessment of the of code, as determined by a Deep Learning model, rewarding code that looks like human-written code. This could help guide the search in directions of known patterns, combating issues like overfitting and getting stuck in local minima.
(2) What kind of mutation or stretch to do could be chosen using a neural network that is trained (on a fixed training set of code) to predict what is most likely to cause improvements. Deep Learning-based code generation could be used to propose candidates for the replacing expressions that we use when mutating instead of generating these at random.
(3) Additionally, since we use mutation conservativity (see [4.2](#subsec:greedy){reference-type="ref" reference="subsec:greedy"}), there will typically be long periods of attempting mutations before a successful mutation is found. One could collect data from these failed mutations. An idea is to then use Deep Learning to analyze these attempted mutations and their resulting local scores in order to get a sense of which mutations do well, which could then be leveraged to pick out qualified guesses for new mutations making use of problem-specific knowledge. Perhaps, one could even train a Deep Learning model during evolution on this data to predict what mutations are likely successful.
[^1]: For example, this could be a randomly generated program or a program that we are trying to debug.
[^2]: \(S\) will typically work by running the program on a pre-defined list of inputs and then comparing the actual outputs to the desired outputs, assigning a score based on how many of these cases are correct.
[^3]: The reason that we include \(\mathbf{if}\)-conditionals in spite of having \(\mathbf{case}\)-expressions is that they simplify our discussion of \(\mathbf{suppose}\)-expressions in [3](#sec:suppose){reference-type="ref" reference="sec:suppose"}.
[^4]: Moreover, scores above \(0.5\) ought to represent and scores below \(0.5\) . (This will later become important for \(\mathbf{suppose}\)-expressions.)
[^5]: This restriction can be dealt with by imposing limits to the number of execution steps allowed or by restricting what kinds of recursion are allowed. The minutiae are dealt with in our proof-of-concept implementation; see [6.5](#subsec:rec){reference-type="ref" reference="subsec:rec"}.
[^6]: In particular, under this scheme, dead code is never mutated.
[^7]: according to semantics like those given in [2.1](#subsec:language){reference-type="ref" reference="subsec:language"} but ignoring traces
[^8]: It would probably be better to allow them to be introduced wherever but have a mechanism pushing them up the expression tree () before they are promoted into \(\mathbf{if}\)-conditionals so as to not have to reinvent the same conditions for various subexpressions.
[^9]: Or perhaps, instead of \(S(e)\), it would be better to use the corresponding approximate score computed for \(e\). This would necessitate storing not just the local scores but also all summands of the averages, so that averages of subsets can be computed at a later point. | {'timestamp': '2022-12-01T02:21:07', 'yymm': '2211', 'arxiv_id': '2211.17234', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17234'} |
null | null |
# Conclusion {#sec:conclusion}
In this paper, we introduce GENNAPE, or Generalized Neural Architecture Performance Estimators, to address issues present in neural predictor design. Namely, that neural predictors typically operate within a fixed search space, and are not generalizable to unseen architecture families. GENNAPE receives Computation Graphs representing arbitrary network architectures as input, before using a Contrastive Learning encoder to generate embeddings. The embeddings pass through an MLP Ensemble, and we compute the prediction using a weighted summation according to Fuzzy C-Means clustering memberships. We extensively test GENNAPE against a number of known neural predictors on NB-101 and show that it yields high MAE and SRCC performance. Experimental results demonstrate the generalizability of our scheme in zero-shot and fine-tuning contexts in terms of SRCC and NDCG\@10. When applied to search, GENNAPE can improve upon existing, high-performance NB-101, NB-201 and OFA-ResNet architectures. Finally, we introduce three new challenge families: HiAML, Inception and Two-Path, as open-source benchmarks.
# Architecture Families {#sec:fam}
We describe the neural network families used in this paper. We start with publicly available benchmarks before introducing three new families of networks. Table [\[table:datasets\]](#table:datasets){reference-type="ref" reference="table:datasets"} provides a overall summary of each family.
## Public NAS Benchmark Families {#sec:pubData}
**NAS-Bench-101** (NB-101) is one of the first and largest benchmarks for NAS. It consists of 423k unique architectures, individually evaluated on CIFAR-10. The architectures are cell-based, where each cell is a Directed Acyclic Graph (DAG) containing operations, stacked repeatedly to form a network. We sample 50k random architectures from this family to form our CG training family. **NAS-Bench-201** (NB-201) and **NAS-Bench-301** (NB-301) are two additional benchmarks. Like NB-101, architectures consist of a fixed topology of cells, except they follow the DARTS search space. Additionally, NB-201 only contains 15.6k evaluated architectures, while NB-301 is a surrogate predictor for the DARTS search space. Therefore, we treat both as test families. **ProxylessNAS** (PN) and **Once-for-All-MobileNetV3** (OFA-MBv3) are based on the MobileNet architecture families, with ProxylessNAS and OFA implementing versions 2 and 3, respectively. **Once-for-All-ResNet** (OFA-RN) is based on the classical ResNet topology. All three evaluate on ImageNet. Architectures consist of searchable, macro features where the number of blocks is variable. We refer the reader to for further details regarding the PN, OFA-MBv3 and OFA-RN families.
## Introduced NAS Benchmark Families {#sec:newData}
We introduce the key attributes of the new network families and provide additional information in the supplementary. For all families, we define an *operator* as a bundle of primitive operations, e.g., Conv3x3-BN-ReLU. A *block* is a set of operators with different connection topologies, while a *stage* contains a repetition of the same block type and the *backbone* determines how we connect blocks to form distinct networks. For these network families, accuracy values are rounded to four decimal places. That is, we would represent 91.23% as a terminating 0.9123. Figure [\[fig:all_hists\]](#fig:all_hists){reference-type="ref" reference="fig:all_hists"} provides accuracy and FLOPs histograms. **HiAML** is a custom family of CIFAR-10 networks inspired by NAS-Bench-101. Networks are built using 14 pre-defined operation block structures found by GA-NAS . Specifically, a pair of identical blocks form a stage, and there are 4 stages per network. Since August 2021, HiAML networks have been in use as the backbone feature extractor in Huawei's mobile Facial Landmark Detection application, i.e., the family is not limited to Image Classification. HiAML networks are efficient enough to run entirely on the less powerful Ascend Tiny Core, which is a part of Da Vinci Architecture 2.0 in Huawei's Neural Processing Unit (NPU). Figure [\[fig:hiaml\]](#fig:hiaml){reference-type="ref" reference="fig:hiaml"} provides a backbone illustration of HiAML. As shown in Table [\[table:datasets\]](#table:datasets){reference-type="ref" reference="table:datasets"}, we observe that HiAML has the narrowest performance profile among all families considered, as 4.6k architectures fit into an accuracy range of only 2.33%. As a comparison, OFA-MBv3 has the second smallest accuracy range of 4.25%. This characteristic makes accurate prediction on HiAML networks very difficult as there are many ties as Figure [\[fig:all_hists\]](#fig:all_hists){reference-type="ref" reference="fig:all_hists"}(a) shows. **Inception** takes inspiration from the Inception-v4 classification networks in literature. Our custom Inception networks consist of a single-path backbone with 3 stacked stages, and inside each stage, there is a repetition of 2 to 4 blocks. Like the original Inception, each block may contain up to 4 branching paths. The main advantage of Inception is that channels are divided among the paths, leading to lower computation costs. The backbone follows a single-path topology with 3 stages. **Two-Path** networks are designed to be more lightweight than other families. As the name suggests, the backbone consists of two branching paths. On each path, there are 2 to 4 blocks, each a sequence of up to 3 operators. One may view the Two-path family as the complement of the Inception family. An Inception network has a single-path backbone with branching paths inside blocks. But a Two-Path network has two branching paths of blocks, while each block is a single path of operators.
# Introduction {#sec:intro}
Understanding and predicting the performance of neural networks is crucial to automated neural architecture design and neural architecture search (NAS). In fact, the amount of time and computational resources required by a NAS scheme is critically dependent on how evaluation is performed. Early methods rely on training individual architectures from scratch via intensive GPU usage. Weight-sharing methodologies like DARTS propose supernet models that superimpose the features of all architectures in a search space, substantially reducing the cost of search. DARTS further reduces this cost by searching for an architecture on a small image classification dataset like CIFAR-10 before transferring the found architecture to a larger benchmark dataset, like ImageNet . In contrast, Once-for-All (OFA) employs strategies to effectively pre-train a re-usable supernet once on ImageNet prior to performing search. Newer methods have seen the introduction of Zero-Cost Proxies which estimate performance from forward pass gradients. Advancements in neural predictors allow for faster inference time, which significantly lowers the architecture evaluation cost incurred in NAS. However, one major trade-off associated with using neural predictors is a lack of generalizability. The input to a neural predictor is usually a vector encoding of the architecture, whose scope is confined to the designated search space the predictor is meant to operate in. For example, to facilitate architecture search, Once-for-All, SemiNAS and BANANAS design specialized predictors, each performing in an individually defined search space. Such a restriction seriously limits the scope that a search algorithm can explore, regardless of how effective it is. Furthermore, as a separate new predictor must be trained for each new search space created, there is an implied additional heavy cost associated with the laborious and resource-intensive task of labelling the performance of a suitable amount of neural architecture samples in this search space, since these architecture samples must be trained from scratch for evaluation. Ideally, a generalized neural predictor would not be subject to these constraints. Not only would it be able to receive input from multiple search spaces, but it would learn an architecture representation that is robust enough to accurately estimate and rank the performance of networks from unseen search spaces. In this paper we propose a **Gen**eralized **N**eural **A**rchitecture **P**erformance **E**stimator (GENNAPE), to perform a feasibility check on pretraining a robust and generalizable neural predictor based only on open neural network benchmarks and transferring it to a wide range of unseen architecture families for accurate accuracy ranking. GENNAPE accepts input architectures from any search space through a generic Computation Graph representation, and performs a two-stage process of graph encoding and cluster-based prediction to achieve transferability to a new neural network design space under no or minimum fine-tuning. In designing GENNAPE, we make the following contributions: First, we propose a generic embedding scheme based on *Computation Graph* (CG) representations of any neural networks and a self-supervised Contrastive Learning (CL) loss as the starting point of our predictor. Specifically, we represent an input neural network as a graph of atomic neural operations, e.g., convolutions, pooling, activation functions, rather than macro units like MBConv blocks that are specific to architecture families. This allows us to feed any neural network into our predictor. We then pass the CG through an encoder to obtain a fixed-length embedding. The graph encoder is trained by a carefully designed CL loss based on Laplacian Eigenvalues and dropout data augmentation, to distinguish graphical features between architecture families.
[\[tab:zcp\]]{#tab:zcp label="tab:zcp"}
Second, we propose to learn multiple prediction heads, each being an MLP and focusing on learning the characteristics of a local cluster of architectures in the training family. We use Fuzzy C-Means (FCM) clustering to partition the training family (in our case, NAS-Bench-101 due to its abundance of labelled samples) into soft clusters, where a given input architecture can find soft memberships among the clusters. Based on the FCM partitions, we construct a Multi-Layer Perceptron-Ensemble (MLP-E) that makes prediction for an architecture by soft aggregating the MLP heads according to the architecture's membership. We verify the performance of GENNAPE on a wide range of public NAS families. Specifically, we pretrain our predictor on NAS-Bench-101, and treat the search spaces of NAS-Bench-201, NAS-Bench-301, ProxylessNAS , OFA-MobileNetV3, OFA-ResNet as unseen target families. We show that our predictor is able to achieve over 0.85 SRCC and over 0.9 NDCG on all of these target families with minimal fine-tuning. When integrated into a downstream NAS search algorithm, GENNAPE further improves high-accuracy architectures from NAS-Bench-101, NAS-Bench-201 and OFA-ResNet on ImageNet120. To stress test the generalizability of GENNAPE, we further introduce three new challenging architecture benchmark sets: HiAML, Inception and Two-Path. These families contain certain properties, such as narrow accuracy range and ties, which make it difficult for a predictor to accurately learn architecture rankings, and thus serve as challenging benchmarks for future NAS research. We show that GENNAPE can obtain over 0.78 NDCG on all three families. Finally, we open-source[^1] these new benchmarks to facilitate further research on generalizable neural predictors.
# Methodology {#sec:method}
Figure [\[fig:gennape\]](#fig:gennape){reference-type="ref" reference="fig:gennape"} visualizes our overall scheme. A predictor that generalizes to arbitrary architecture families should accept a generalizable neural network representation as input. We consider the Computation Graph (CG) of a network structure based on the graph structures deep learning libraries like TensorFlow and PyTorch use to facilitate backpropagation. Specifically, nodes refer to atomic operators in a network, e.g., Convolution, Linear layers, Pooling layers, non-linearities and tensor operations like mean, add and concat. Edges are directed and featureless. This approach allows us to encode architectures from different families using a common format that represents all primitive network operations and the connections between them. For example, both NB-101 and NB-201 include 'Conv3x3' and 'Conv1x1' as candidate operations. These candidates are actually sequences of three primitive operations: Convolution, ReLU and Batch Normalization (BN), which are arranged differently for each search space. Wheras NB-101 consider a 'Conv-BN-ReLU' ordering, NB-201 use 'ReLU-Conv-BN'. Moreover, NB-101 encodes operations as nodes, while NB-201 follow DARTS and encode operations on edges. Our CG format reconciles these differences by representing each primitive as its own node with edges defined by the forward-pass of the network. Figure [\[fig:simpleCG\]](#fig:simpleCG){reference-type="ref" reference="fig:simpleCG"} illustrates a simple CG, where each node is a primitive operation. Node features consist of a one-hot category for operation type as well as the height-width-channel (HWC) size of the input and output tensors. If an operation contains trainable weights, e.g., convolution and linear layers, we include a feature for the weight tensor dimensions, and a boolean for whether bias is enabled. As graphs can contain an arbitrary number of nodes and edges, it is necessary to process them into a common, fixed-length format prior to prediction. Therefore, in this section, we introduce our Contrastive Loss-based (CL) graph embeddings scheme, the Fuzzy C-Means (FCM) ensemble that forms the GENNAPE predictor and describe a FLOPs-based accuracy transform.
## Embeddings from Contrastive Learning {#sec:cl}
We use Contrastive Learning (CL) to learn vector representations of CGs by adapting the two key components of SimCLR . The first is data augmentation, where perturbations produce two different copies of every image in a batch. The second is the CL loss function, which aims to identify positive and negative pairs within a batch. Unlike images, random perturbations of CGs are unlikely to belong to the same architecture family. Therefore, to generate embedding variation, we adopt dropout augmentation from . Consider a batch of \(N\) CGs. We use dropout and apply a graph encoder twice to make \(2\) copies of each CG. We then use an MLP head to project these embeddings of length \(e\) into a lower dimensional space \(p\) to optimize the CL loss. Let \(I=\{h_1,h_2,\dots,h_{2N}\}\subset\mathbb{R}^e\) be the vector representations, and let \(z_i=proj(h_i)\in \{||z||=1: z \in \mathbb{R}^p\}\) be their projections. We define the cosine similarity \(sim(z_i,z_j)=z_i\cdot z_j/\tau\), with temperature \(\tau>0\) and \(\cdot\) as the dot product. Then, our CL loss function is \[\centering \label{eq:cl_loss} \mathcal{L}_{CL} =-\sum_{i\in I} \sum_{\ell \not=i} \alpha_\ell^{(i)} \log \frac{\exp(sim(z_i,z_\ell))}{\sum_{r\not=i} \exp(sim(z_i,z_r))},\] where \(\alpha_\ell^{(i)} \ge 0\) and \(\sum_{\ell \not= i} \alpha_\ell^{(i)}=1\). For each computational graph \(i\), we calculate the convex combination \(\alpha_*^{(i)}\) using the spectral distance of the underlying undirected graph as a similarity measure. We use topological properties to group architectures in the same family or separate architectures from different families, because the arrangement of operations differs for each family. Therefore, given a computation graph, let \(G\) be its underlying undirected graph. The normalized Laplacian of \(G\) encodes important connectivity features. Furthermore, it allows us to define a (pseudo) distance between computation graphs by evaluating the Euclidean distance between the corresponding \(q = 21\) smallest eigenvalues . We use \(\sigma(g_1,g_2)\) to denote the distance between computation graphs \(g_1,g_2\) and define \(\alpha_*^{(i)}\) to be the softmax of \(\sigma(i,*)\). We train a CL encoder on NB-101 and use it to infer embeddings for all test families. A 2-dimensional representation of the embeddings is shown in Figure [\[fig:tsne\]](#fig:tsne){reference-type="ref" reference="fig:tsne"}. We note two advantages of the CL encoder. First, it effectively separates each test family into distinct clusters with minimal overlap. By contrast, the CL encoder divides the training family, NB-101, into many smaller clusters. This is an important finding. While NB-101 contains many architectures, it only contains 3 candidate operations, meaning that most of the diversity amongst these architectures comes from topological differences, i.e., Laplacian eigenvalues, which play an important role in training our CL encoder. As such, each NB-101 cluster may reside closer to a given test family than others. Next, we exploit these observations to build a predictor ensemble.
## Soft-Clustering Ensemble {#sec:fcm}
We use MLP Ensembles (MLP-E) in GENNAPE to make predictions. Each head in the MLP-E is is responsible for a region of the embedding space, although these regions can partially overlap with each other. That is because we use Fuzzy C-Means (FCM) clustering to divide up the embedding space, and FCM represents data using continuous *memberships* with respect to all clusters. The output of the MLP-E is defined as, \[\centering \label{eq:mlp_ensemble} y'_i = \sum_{j=1}^{C}U_{i,j}f^{j}(h_i),\] where \(h_i\) and \(y'_i\) are the graph embedding and target prediction a CG, of which there are \(N\) in total, \(C\) is the number of heads, \(f^j\) is a head and \(U_{i, j}\) is the membership for CG \(i\) across cluster \(j\). FCM membership values \(U \in (0, 1)^{N \times C}\) are positive numbers that sum to 1 for a given sample; \(\forall i \sum_{j=1}^{C} U_{i, j} = 1\). Like other distance-based clustering algorithms, FCM consists of centroids that are computed alongside membership values by minimizing an optimization. Let \(d_{i,j}\) be the Euclidean distance between CG embedding \(h_i\) and FCM centroid \(v_j\). FCM converges by alternating between updating membership values and centroid locations, respectively, according to the following equations, \[\centering \label{eq:fcmMembership} U_{i, k} = (\sum_{j=1}^{C}(\dfrac{d_{i,k}^2}{d_{i,j}^2})^{\dfrac{1}{m-1}})^{-1},\] \[\centering \label{eq:fcmCentroids} v_k = \dfrac{\sum_{i=1}^N U_{i,k}^m h_i}{\sum_{i=1}^N U_{i,k}^m},\] where \(m > 1\) is known as the *fuzzification coefficient*, and is typically a value of 2 or greater , which controls the degree of overlap between the clusters. The FCM algorithm alternates between Equations [\[eq:fcmMembership\]](#eq:fcmMembership){reference-type="ref" reference="eq:fcmMembership"} and [\[eq:fcmCentroids\]](#eq:fcmCentroids){reference-type="ref" reference="eq:fcmCentroids"} until either a pre-specified number of iterations has been reached, or the following stopping criteria is satisfied: \[\centering \label{eq:fcmStop} \max_{i, j} \lvert U_{i, j}(t+1)-U_{i,j}(t) \rvert \leq \epsilon,\] where \(t\) is a time parameterization of \(U\). Essentially, clustering convergence is reached once the maximum change in membership across all \(N\) data samples and \(C\) clusters is less than \(\epsilon = 1e^{-9}\). When applying FCM to GENNAPE, we compute cluster centroids on the training family. For test families, we infer membership from existing centroids using Equation [\[eq:fcmMembership\]](#eq:fcmMembership){reference-type="ref" reference="eq:fcmMembership"}. A simple MLP is not guaranteed to generalize to features of families it has not seen. However, we can aid generalizability by only having the MLP learn a specific region in the feature space and weighing its predictions based on the distance between the unseen features and the learnt region. Moreover, we can divide up the feature space with an MLP-E. Each head learns to specialize on a subset of the embedding features for the training family, as well an association with nearby, unseen features, depending on memberships.
## FLOPs-based Features and Transforms {#sec:FLOPs}
Every network has an associated FLOPs value, or floating point operations required for the forward pass. GENNAPE uses FLOPs to transform the regression targets While the downstream goal of any regressor is to estimate the accuracy of a neural network, GENNAPE estimates the following, \[\centering \label{eq:accTrans} y_i = \mathcal{Z}(\dfrac{A_i}{\texttt{Log}_{10}(F_i+1)+1}),\] where \(y_i\) is the label, \(A_i\) is the accuracy of a network, \(F_i\) is the FLOPs, measured in gigaFLOPs and \(\mathcal{Z}\) is the Z-Score. The inverse of this transformation can be applied to a prediction to obtain the corresponding accuracy prediction. We use the standardization statistics from the NB-101 training set for all test families, and we prune out NB-101 architectures whose accuracy is below 80% when implementing Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"}. This leaves us with 49k architectures.
# Related Work {#sec:related}
Neural predictors rely on benchmark datasets, namely NAS-Bench-101 , 201 and 301 , which allow for performance *querying*. The downside to these methods is that every architecture must be trained from scratch, possibly multiple times. Arguably, supernet-based methods like OFA fall under a unique category of benchmark, where sampled architectures can be immediately evaluated. However, this is still less resource intensive than training the network from scratch. Therefore, while benchmarks provide a rich repository of architecture statistics, an upfront resource cost is incurred, motivating the need for generalized predictors that can perform well on unseen families. One method of low-resource neural prediction are Zero-Cost Proxies (ZCP); outline many variants. ZCP schemes involve gradient calculations as well as parameter salience. As Table [\[tab:zcp\]](#tab:zcp){reference-type="ref" reference="tab:zcp"} shows, while ZCP methods may achieve good results on some architecture families, their performance is inconsistent across families. In addition, ZCP methods require network instantiation to perform forward passes to compute corresponding measures, while neural predictors do not. Contrastive Learning (CL) is based on distinguishing pairs of similar and dissimilar objects. This approach has been recently used successfully in image classification . Specifically, CL classifiers improve transferability on unseen data. Furthermore, fine-tuning on a very small fraction of the labeled data from a new dataset is enough to produce state of the art classifiers. further extend this approach to the case when image labels are known during training. We focus on applying CL methods for finding vector representations of graphs. Some deep learning schemes use Fuzzy C-Means clustering (FCM) to design interpretable models. substitute the last MLP for an FCM clustering layer and classify images according to the highest membership value. By contrast, GENNAPE incorporates FCM into an MLP ensemble, using the membership values to perform a weighted sum of MLP heads.
# Experimental Results {#sec:exp}
In this Section, we describe our experimental procedure and results in terms of regression and correlation metrics. Finally, we apply GENNAPE to NAS and perform search.
## NAS-Bench-101 Comparison {#sec:setup}
We use NB-101 as the training family and treat all others as unseen test families. We split the NB-101 architectures into a training set with 80% of the data, and two separate validations sets that contain 10% each. We apply Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"} using the mean and standard deviation from the training partition to standardize the validation sets. The model trains for 40 epochs. We evaluate it on the first validation set at every epoch without applying the inverse of Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"} to track loss statistics on transformed labels. Once training is complete, we further evaluate the model on the second validation set, this time applying the inverse of Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"} on predictions to calculate performance compared to the ground truth accuracy. For each scheme, we train a model 5 times using different random seeds. The overall best model is the one that achieves the highest rank correlation on the second NB-101 validation set. We provide hyperparameters and other training details in the supplementary materials. We consider several types of regressor variants in our experiments. 'CL+FCM+T' uses CL graph embeddings as well as the FCM MLP-E to predict performance labels transformed using Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"}. 'CL+MLP+T' removes the FCM clustering ensemble and uses a single MLP while 'CL+FCM' keeps the keeps the clustering ensemble while removing Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"}. Finally, 'CL+MLP' drops both Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"} and the FCM ensemble to predict accuracy directly using the CL embeddings. As baselines, we consider several GNN regressors such as \(k\)-GNN , Graph Convolutional Networks (GCN) and Graph Isomorphism Networks (GIN) and train them end-to-end on CGs. Finally, we also consider several related neural predictors, such as Neural Predictor for NAS (NPN) , BANANAS and TNASP . These predictors operate on the NB-101 family, however since they do not represent architectures using CGs, they are not transferable outside of NB-101. We consider two predictor performance metrics. The first is Mean Absolute Error (MAE), which measures predictor accuracy with respect to ground-truth labels. We report MAE as a percentage where lower is better. The second is Spearman's Rank Correlation (SRCC) which judges a predictor's ability to rank a population of architectures. SRCC ranges from-1 to 1 and in our case higher is better as that represents positive agreement with the ground-truth. Table [\[tab:nb101_compare\]](#tab:nb101_compare){reference-type="ref" reference="tab:nb101_compare"} enumerates our results for NB-101. We note the effectiveness of Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"}, which allows us to achieve MAE scores below 1% on average. Moreover, the inclusion of CL and FCM further improve this result to a minimum of 0.59%. In terms of ranking correlation, CL+FCM+T achieves the second highest SRCC at 0.930. While NPN obtains a slightly higher SRCC, that is the limit of its capabilities, since NPN on NB-101 cannot perform predictions for architectures from other families. Finally, in terms of other GNN CG baselines, the \(k\)-GNN excels. While it does not obtain the best MAE or SRCC, it is distinctly better than both the GCN and GIN. Therefore, we continue to use it as a generalizability baseline when comparing performance on other families.
## Transferability Prediction Error {#sec:family_tests}
When evaluating on test families, we consider two cases. The first is zero-shot transfer performance, where we measure a given metric on all architecture CGs in a family without fine-tuning the predictor. In the second case, we sample 50 architectures at random, fine-tune for 100 epochs and then evaluate on the remaining, held-out architectures. We perform fine-tuning 5 separate times on different seeds, ensuring that we select the same 50 architectures per seed. Evaluating MAE performance on test families is applicable in the fine-tuning setting as that provides the predictor some information on the actual distribution of target labels. When fine-tuning using Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"}, we use the mean and standard deviation information from the NB-101 training partition to standardize labels for the 50 fine-tuning samples, and to recover accuracy on the remaining samples at inference. Table [\[tab:mae\]](#tab:mae){reference-type="ref" reference="tab:mae"} provides test family MAE results for the \(k\)-GNN baseline and several CL-based regressors. Once again, we observe how regressors that use Equation [\[eq:accTrans\]](#eq:accTrans){reference-type="ref" reference="eq:accTrans"} achieve the lowest MAE on 3 of 9 families, including all ImageNet families and Two-Path. The \(k\)-GNN achieves good performance on NB-201, likely because it has the second largest range after NB-101. On HiAML, Inception and Two-Path, the results generally very close, showing how difficult it can be to predict accuracy for them.
## Unseen Architecture Ranking Performance {#sec:srcc_ndcg}
GENNAPE consists of a weighted average of predictions from several models. The first two are the aformentioned 'CL+FCM+T' and 'CL+MLP+T' regressors. We also include two pairwise classifiers, that, instead of estimating accuracy, predict which CG in a pair has higher accuracy. We label these as 'CL+Pairwise+FCM' and 'CL+Pairwise', differentiated by whether they have FCM and an MLP ensemble. We also include the \(k\)-GNN and the FLOPs predictor from Table [\[tab:zcp\]](#tab:zcp){reference-type="ref" reference="tab:zcp"}. For zero-shot performance, GENNAPE weighs the output of all predictors equally. When fine-tuning, we calculate weights by taking the softmax of the Kendall's Tau (KT) result each individual predictor obtains on the fine-tuning samples. The supplementary materials contains details and an ablation study of these components. Table [\[tab:srcc\]](#tab:srcc){reference-type="ref" reference="tab:srcc"} provides SRCC results across all test families. We observe how GENNAPE achieves SRCC values above 0.85 on all public benchmarks. In fact, the only public family GENNAPE does not achive zero-shot SRCC above 0.5 on is NB-301, however, at 0.3214 this is still much higher than the \(k\)-GNN and any of the ZCP from Table [\[tab:zcp\]](#tab:zcp){reference-type="ref" reference="tab:zcp"}. Moreover, fine-tuning greatly improves the performance of GENNAPE on NB-301. Meanwhile, the \(k\)-GNN does achieve higher zero-shot SRCC on OFA-RN, but the improvement is less than 0.1 SRCC. On all other public benchmarks it fails to achieve over 0.5 SRCC in the zero-shot setting. GENNAPE achieves the best zero-shot performance on HiAML, Inception and Two-Path. With fine-tuning, it manages to achieve above 0.5 SRCC on Inception and close to 0.5 SRCC on Two-Path. By contrast the \(k\)-GNN fails to achieve positive correlation on HiAML and Inception in the zero-shot setting, and while it achieves better HiAML SRCC with fine-tuning, once again the performance gap is very small, below 0.015 SRCC and less than the GENNAPE zero-shot SRCC of 0.4331. Finally, GENNAPE does lose some ranking performance when fine-tuning on HiAML, but that is a feature of the family: a narrow accuracy range with many ties that is hard to predict on. We provide additional results using Kendall's Tau in the supplementary materials.
SRCC measures correlation over all \(N\) samples and assigns equal weight to each. However, the concerns of NAS relate to finding high-performance architectures. Therefore, for further evaluation of GENNAPE, we consider another metric that assigns disproportionate importance. Following AceNAS , we adopt a metric from the field of Information Retrieval (IR), where queried objects are assigned a relevance score, and it is more important to properly order objects with higher relevance. This goal is analogous to the concerns of NAS. Accuracy is relevance. Specifically, we adopt Normalized Discounted Cumulative Gain (NDCG), which prioritizes the ordering of the top-\(k\) architectures by a ground truth ranking. NDCG is reported in the range \([0, 1]\), where higher is better. We consider the case of \(k=10\) in our experiments, which corresponds to ordering the very best architectures. Details of the NDCG calculation are in the supplementary materials. We list our findings in Table [\[tab:ndcg\]](#tab:ndcg){reference-type="ref" reference="tab:ndcg"}. GENNAPE clearly outperforms the \(k\)-GNN baseline in both contexts. The sole exception is OFA-R50, likely due to NB-101 architectures being partially based on ResNets in terms of operation choice. Still, in the zero-shot transfer context GENNAPE achieves over 0.65 NDCG on *all* families, and over 0.9 NDCG on all public families with fine-tuning. Moreover, GENNAPE now overtakes the \(k\)-GNN on HiAML in both the zero-shot and fine-tuning contexts. Results are also more favorable on Inception and Two-Path, where GENNAPE achieves over 0.8 NDCG in both the zero-shot and fine-tuning settings. This aligns with our earlier intuition: Although the small accuracy range makes global ranking difficult, there are only a few architectures at the tail of the distribution and GENNAPE is better at determining which ones they are. We report additional NDCG results in the supplementary materials. In sum, we show that GENNAPE is a robust neural predictor that can easily generalize to unseen architecture families. The NDCG results we show are of particular relevance to the problem of NAS, where neural predictors fill the role of performance evaluation. This is because, ultimately, the downstream performance of any search algorithm relies upon having a performance estimator capable of identifying high-performance architectures. Next, we apply GENNAPE to the problem of NAS, to demonstrate its applied capability and the flexibility of our CG architecture representation.
## Applying GENNAPE to NAS {#sec:search}
We conduct NAS experiments where we directly modify the CGs of existing optimal classification networks. We aim to reduce FLOPs while improving or maintaining accuracy using a simple location search algorithm and provide details in the supplementary materials. After search, we train and evaluate the original optimal architecture and the one and found by search on classification datasets. Specifically, we apply this search routine to NB-101 and NB-201 on CIFAR-10. To reduce carbon footprints, we also search on OFA-RN but use a subset of ImageNet, which only contains the first 120 classes; ImageNet120. We use the fine-tuning CL+FCM+T predictors in these experiments. Search takes less than 3 hours. Inference on our predictors is fast. The primary bottleneck stems from the search algorithm determining whether newly found CGs are feasible as classification neural networks. Table [\[tab:search\]](#tab:search){reference-type="ref" reference="tab:search"} summarizes the results, demonstrating the power of GENNAPE to further improve upon existing, high-performance architectures, in terms of accuracy and FLOPs, on three public families. Notably, the OFA-ResNet architecture found by search achieves 22% FLOPs reduction while improving accuracy by 0.4%. Moreover, these results demonstrate the utility of our CG framework, as the atomic operation mutations result in architectures that are not exact fits to the original search spaces. Using CGs, we can modify sections like the stem or head that are typically fixed, and introduce new operation sequences. For example, the NB-201 architecture we compare to and outperform was already the best on CIFAR-10. Therefore, a typical neural predictor for NB-201 would not be able to adequately process and predict performance for architectures found by our search algorithm.
## Introduced Neural Network Benchmarks {#sec:customFamilies}
Figure [\[fig:families\]](#fig:families){reference-type="ref" reference="fig:families"} provides a detailed depiction of the three new benchmark families we propose: HiAML, Inception and Two-Path. We re-iterate the key attributes of these new benchmarks. Specifically, we define an *operator* as a bundle of primitive operations, e.g., Conv3x3-BN-ReLU. A *block* is a set of operators with different connection topologies, a *stage* contains a repetition of the same block type and the *backbone* determines how we connect blocks to form distinct networks.
### The HiAML family
(Fig. [\[fig:families\]](#fig:families){reference-type="ref" reference="fig:families"}(a)) first constructs a network backbone with 4 stages and exactly 2 block repetitions per stage. Different from other families, the block structures are not randomly sampled. Instead, we first run GA-NAS on a search space similar to NAS-Bench-101, i.e., we allow up to 4 operators per block with any connection patterns in-between. After running NAS we find 14 blocks, and to construct the HiAML family we uniformly sample from these 14 blocks for each stage. Reduction is performed in the stem Conv3x3 and the first block of stages 2 and 4. We label 4629 networks for this family on CIFAR-10.
### The Inception family
(Fig. [\[fig:families\]](#fig:families){reference-type="ref" reference="fig:families"}(b)) mimics the Inception-v4 classification networks. Our Inception networks all contain 3 stages. Inside each stage, we have 2-4 stacked blocks. Like HiAML, the blocks inside a stage have the same internal structure. Blocks in different stages will likely can have different internal structures. When sampling a block, one first chooses the number of branching paths. We allow up to 4 branching paths and inside each path, there are up to 4 operators. An operator could be a Conv3x3-BN-ReLU combination. We adopt a wide range of candidate operators, including regular Convolution, Depth-wise Separable Convolution, MBConv, etc. Figure [\[fig:families\]](#fig:families){reference-type="ref" reference="fig:families"} (b) shows an example of a 2-path block topology. Additional paths will follow a similar layout. Note that we also divide the candidate operators into two groups: Conv operators (Conv op) and Supplementary operators (Supp op). The supplementary operators set includes all 1x1 convolution and pooling operations, whereas the Conv operators set only includes convolution with kernel size \(>\)`<!-- -->`{=html}1. A path must have at least one 1 Conv op, and at most one Supp op. Input channels will be further reduced based on the number of paths, e.g., if there are 2 paths then the channel size of a path will be half of the input channels, achieved via projection by the first operator of the path. In our Inception networks, we perform reductions on the input image in each stage by halving the height and width of the input and doubling the channels. This is achieved by the first block of each stage. We label 580 networks for this family on CIFAR-10.
### The Two-Path family
(Fig. [\[fig:families\]](#fig:families){reference-type="ref" reference="fig:families"}(c)) constructs a simpler structure where an entire network contains only two branching paths. Inside a path, there are 2-4 blocks. Each block has a single-path topology with 1-3 operators. The definition of an operator is the same as the Inception family. In fact, one may view the Two-path family as the complement of the Inception family. An Inception network has a single-path backbone with branching paths inside blocks. But a Two-Path network has two branching paths of blocks, while each block is a single path of operators. Reduction on the input image is performed in the second stem Conv3x3 and also by two randomly chosen blocks of each path. Since we do not allow block repetitions, Two-Path networks are usually lightweight and shallower than the other families. We label 6890 networks for this family on CIFAR-10.
## Model and Hyperparameter Details
### Baseline GNNs {#sec:gnnDetails}
Our GNNs (e.g., \(k\)-GNN, as well as GCN and GIN ) consist of an initial embedding layer that transforms CG node features into continuous vectors of size 32. Graph encoders consist of 6 layers. We derive the graph embedding by averaging all node embeddings. Regressors consist of 4 layers with a hidden size of 32. Baseline GNNs train predictors for 40 epochs with an initial learning rate of \(1e^{-4}\) and a batch size of 32.
### Contrastive Learning Graph Encoder {#sec:clDetails}
The CL encoder consists of a GNN and Transformer whose output is concatenated to produce CG embeddings. The GNN consists of 4 layers, while the Transformer consists of a single layer with two attention heads trained with a dropout rate of 0.1. Both modules produce output vectors of length 64, so the overall embedding size is \(m = 128\). We train the CL encoder using a batch size of 128, consider the \(q = 21\) smallest eigenvalues and use a temperature \(\tau = 0.05\). For regularization, we add a second projection head that predicts the FLOPs of a CG.
### Details of the Fuzzy C-Means Ensemble {#sec:fcmDetails}
We use a simple grid search to determine the number of clusters \(C \in [10, 20]\) and fuzzification coefficient \(m \in [2.0, 4.0]\). As \(m\) is continuous we consider increments of \(0.5\) and select the model that achieves the highest SRCC the second NB-101 validation set. For regressors, these were \([C, m] = [16, 4.0]\) and for classifiers it was \([C, m] = [11, 3.5]\). Prior to clustering, we standardize FLOPs using the mean and standard deviation from the NB-101 training set and concatenate them with the CL graph embeddings. We also use Principle Component Analysis (PCA) to reduce embedding size down to 32. FCM runs for 10k iterations or until the stopping criteria \(\epsilon = 1e^{-9}\) is satisifed.
### Details of the CL+MLP and CL+FCM {#sec:mlpDetails}
MLP regressors contain 4 hidden layers with a hidden size of 256 each, while pairwise classifiers contain 128. FCM memberships in the classifier sum to produce a latent vector of size 16 for each CG in a pair. We concatenate these and use an MLP layer to produce a prediction. We use ReLU as an activation function and train all predictors for 40 epochs with an initial learning rate of \(1e^{-4}\) and a batch size of 32. We perform fine-tuning for 100 epochs using 50 samples with a batch size of 1. Finally, we obtain the SRCC of pairwise classifiers using mergesort with the model as a comparater to generate a list of rankings.
## Calculation of NDCG {#sec:ndcg}
We provide a brief explanation on the calculation of Normalized Discounted Cumulative Gain (NDCG) for the top-\(k\) best architectures in a population. Given lists of predictions and ground-truth accuracies, sort into descending order according to the prediction. Each prediction shall be given an index \(1, 2, 3..., N\) while the labels will be assigned a relevance score \(rel_1, rel_2, rel_3..., rel_N\). Following AceNAS , we assign relevancy by re-scaling accuracy labels into the range \([0, 20]\). We first compute the Discounted Cumulative Gain (DCG) as \[\centering \label{eq:dcg} DCG = \sum_{i=1}^{k}\dfrac{2^{rel_i}-1}{\texttt{log}_2(i+1)},\] where \(k \leq N\) allows one to control how many items we consider. We then compute NDCG as \[\centering \label{eq:ndcg} NDCG = \dfrac{DCG}{IDCG},\] where \(IDCG\) is \(DCG\) under the ideal ordering, e.g., descending sort according to the ground truth where \(rel_i \geq rel_j\) if \(i < j\).
## Search Algorithm Details {#sec:search_details}
We pair our fine-tuned CL+FCM+T predictors with a simple local search algorithm which runs for 6 iterations. In each iteration, we collect the current top-\(k\) best CGs based on predicted performance and FLOPs. For each top-\(k\) CG, we perform mutation by randomly replacing a sequence of primitive operations with a sub-graph of new operations, ensuring that previous HWC channel constraints are met. Finally, we rank the CGs based on the predicted performance and collect the new top-\(k\) best CGs for the next iteration. When searching on NB-101 and NB-201, we select the optimal architectures on CIFAR-10. For OFA-ResNet, we use the best architecture found by.
## Ablation Studies {#sec:ablate}
We ablate the performance of several GENNAPE components, specifically the CL+MLP, CL+MLP+T and CL+FCM+T regressors as well as the CL+Pairwise and CL+FCM+Pairwise classifiers. We consider the zero-shot transfer and fine-tuning contexts, for all target families across three different metrics.
### Spearman's Rank Correlation
Table [\[tab:srcc_ablate\]](#tab:srcc_ablate){reference-type="ref" reference="tab:srcc_ablate"} ablates SRCC results across all families. Overall, our findings demonstrate the generalizablity of GENNAPE as it still attains the best correlation performance on many families. Sometimes CL+FCM+T may outperform the full GENNAPE on a family (e.g., OFA-MBv3, HiAML, Inception) in the zero-shot transfer context, yet reinforcing the efficacy of our FCM and MLP-E scheme.
### Kendall's Tau
We consider Kendall's Tau (KT), which counts the number of concordant and disconcordant pairs in a set and can handle ties. Table [\[tab:kendall_ablate\]](#tab:kendall_ablate){reference-type="ref" reference="tab:kendall_ablate"} summarizes our findings. GENNAPE clearly outperforms other configurations on this metric in many scenarios. We note that in comparison to SRCC results in Table [\[tab:srcc_ablate\]](#tab:srcc_ablate){reference-type="ref" reference="tab:srcc_ablate"}, the best performing predictors for any family and context do not change much. This is because KT is a correlation metric like SRCC that considers all samples equally, yet because it calculates over \(N^2\) pairs rather than \(N\) samples, performance numbers are generally closer to 0.
### Normalized Discounted Cumulative Gain
NDCG calculations involve selecting the top-\(k\) samples, e.g., NDCG@\(k\). Before, we considered NDCG@\(10\), but now broaden the scope of our ablation results to include NDCG@\(50\). Table [\[tab:ndcg_ablate\]](#tab:ndcg_ablate){reference-type="ref" reference="tab:ndcg_ablate"} presents our findings. For both values of \(k\), GENNAPE most frequently achieves the highest performance, specifically for NDCG@\(50\). However, as we decrease \(k\) to \(10\), several individual CL-based predictors attain the best performance, specifically the ensemble-driven CL+FCM+T.
## Hardware and Software Details {#sec:hardware}
We perform our experiments on rack servers that use Intel Xeon Gold 6140 CPUs with 756GB of RAM running Ubuntu 20.04.4 LTS. Each server holds a set of 8 NVIDIA Tesla V100 GPUs with 32GB each. To create Compute Graphs from models we use `TensorFlow==1.15.0`, reproducing `PyTorch` models using `Keras` components when necessary. To train predictor models we use `PyTorch==1.8.1` as well as `Pytorch-Geometric`. Additionally, we use `PyTorch-Fast-Transformers` to implement the transformer module in the CL graph encoder.
[^1]: https://github.com/Ascend-Research/GENNAPE | {'timestamp': '2022-12-01T02:21:03', 'yymm': '2211', 'arxiv_id': '2211.17226', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17226'} |
null | null |
null | null |
null | null |
null | null |
# Introduction
Coarse-grained reconfigurable arrays (CGRAs) have been studied heavily in recent years as a promising configurable accelerator architecture . The end of Moore's law necessitates the creation of specialized hardware accelerators to enable running increasingly complex image processing and machine learning applications. While a variety of hardware accelerator architectures exist, CGRAs have emerged as an interesting midpoint between the flexibility of an FPGA and the performance of an application-specific accelerator. A CGRA can achieve high energy efficiency and performance due to word-level arithmetic operations and interconnect, while maintaining enough flexibility to run a variety of applications that evolve over time. CGRAs, as well as other spatial accelerator architectures, often have hundreds of compute cores and memory cores. These compute cores (called processing elements or PEs) and memory (MEM) cores are laid out spatially in a grid of tiles and are connected through a configurable interconnect. An example is shown in Fig. [\[fig:cgra-diagram\]](#fig:cgra-diagram){reference-type="ref" reference="fig:cgra-diagram"}. The reconfigurable interconnect contains switch boxes (SBs), which connect the PE/MEM outputs to the tracks in the interconnect, and connection boxes (CBs), which connect the interconnect tracks to the inputs of the cores. While having a large number of compute cores enables very high performance, the reconfigurable interconnect connecting these cores can constitute over 50% of the CGRA area and 25% of the CGRA energy. Design space exploration of the interconnect is necessary to achieve high performance with lower energy and area costs. There are many interconnect design choices that directly impact the power, performance, and area of the resulting accelerator, including the number and bitwidth of tracks in the interconnect, how the processing elements and memories are arranged in the array, how those elements are connected, and how the interconnect is configured. An agile approach for specifying and generating the interconnect is needed for efficient design space exploration. In this paper, we present Canal, a Python-embedded domain-specific language (eDSL) and compiler for specifying and generating reconfigurable interconnects for CGRAs using a graph-based intermediate representation. The major contributions of our paper are:
1. We describe a graph-based intermediate representation (IR) for CGRA interconnects that is capable of representing and generating a variety of topologies.
2. We propose an embedded domain-specific language (eDSL) called Canal that can compile an interconnect architecture specification into the graph-based IR.
3. We propose an interconnect generator system that can take the IR and automatically produce hardware, place and route collateral, and a bitstream generator.
4. We explore various design space choices using Canal and demonstrate its effectiveness in generating an efficient CGRA design.
# Related Work
Previous attempts have been made to create an interconnect generator for CGRAs, but new demands in CGRA design necessitate a more flexible and powerful system. VPR is one of the most established FPGA architecture research tools . It allows users to adjust various design aspects of the FPGA and observe the effects on final application performance, such as timing and area usage. However, VPR does not offer an RTL generator and users have to design their FPGA independently and hand-write the VPR architecture file accordingly. CGRA-ME is a CGRA architecture research tool similar to Canal in that it also offers integrated RTL generation and place and route tools. One of the major differences is the architecture specification. CGRA-ME opts for a more rigid XML-based input whereas Canal takes in a Python eDSL program, which is more flexible and readable. FastCGRA is a similar CGRA architecture exploration tool that uses an eDSL to construct the hardware. However, with FastCGRA, users explicitly construct multiplexers and switches, from which RTL is generated. Canal abstracts away the notion of hardware primitives and lets the compiler backend choose how to generate the hardware. In summary, Canal supports more tools, allows for more flexibility, and enables easier design space exploration than previous attempts at an interconnect generator.
# System Design
In this section, we introduce the design of the Canal system, including the graph-based IR for representing interconnects, the Canal eDSL, the static interconnect generation used to translate the IR into hardware, and finally how that IR interfaces with the application place and route (PnR) algorithms. Fig. [\[fig:system-diagram\]](#fig:system-diagram){reference-type="ref" reference="fig:system-diagram"} summarizes how the Canal interconnect generator interfaces with the PE and memory core designs, application PnR, RTL generation, and bitstream generation.
## Graph-Based Intermediate Representation
The primitives in Canal's intermediate representation (IR) are nodes, which represent anything that can be connected in the underlying hardware, and edges, which are wires connecting the nodes together. An example of the IR for a switch box is shown in Fig. [\[fig:graph-representation\]](#fig:graph-representation){reference-type="ref" reference="fig:graph-representation"}. All edges are unidirectional so the IR represents a directed graph. Nodes in the graph can have multiple incoming edges which, when translated into hardware, transform into multiplexers. Each node also has attributes that provide additional information for type checking and hardware generation. This intermediate representation is flexible enough to represent a wide variety of interconnect topologies and handle an arbitrarily complex set of CGRA cores. We will discuss a few design space exploration experiments that exploit this flexibility in Section [9](#sec:results){reference-type="ref" reference="sec:results"}.
## The Canal Language
The Canal language is a Python-embedded domain-specific language (eDSL) that constructs the interconnect intermediate representation described in the previous section. The Canal language translates the Python description of an interconnect into this IR, so at the lowest level, a designer could instantiate nodes in the Canal language and wire them together. As Canal is embedded in Python, we have also built a layer on top of the basic primitives of Canal, which simplifies the IR construction. For instance, for creating a uniform interconnect (all switch boxes have the same topology) with no diagonal connections, we provide a simple helper function that produces different interconnect topologies by varying function parameters such as height and width of the array, switch box topology, number of tracks, bit width of tracks, and density of pipeline registers. An example of creating a low level Canal node and using a higher level Canal helper function is shown in Fig. [\[fig:canal_example\]](#fig:canal_example){reference-type="ref" reference="fig:canal_example"}. The Canal eDSL allows designers to easily conduct design space exploration by varying the parameters in the helper functions, or by generating an entirely new interconnect. Canal can easily be integrated with other DSLs. For example, one could use a DSL for specifying individual CGRA tiles and then integrate them together using Canal to generate the interconnect.
## Generating Interconnect Hardware {#sec:compile-hardware}
Because Canal's IR only describes the connectivity among the different nodes, it is up to the hardware compiler backend to decide how to lower the IR. We implemented two different hardware compiler backends that lower the IR into
. This NoC has data channels along with a ready-valid interface and routing is configured statically before the application runs. We use magma as our hardware circuit implementation, but this could be extended to any hardware generator framework. To generate a static mesh interconnect, we adopt the following principles to generate hardware:
1. Nodes with hardware attributes (e.g. a processing element core) generate the specified hardware.
2. Directed edges are translated into wires.
3. Nodes with multiple incoming edges generate multiplexers.
We also use attributes associated with each node to lower the node to different hardware components. For instance, a register node will be lowered into a physical register. A port node will be lowered to a CB (with an internal multiplexer) where the output of the CB connects to the port of the core. These translations are mechanical and can be accomplished through a compiler pass. Reusing the same IR to generate a statically configured NoC has several challenges. First, the application graph may have fanouts, that is, one output port of a node is connected to multiple input ports. While this is simple to handle in a static interconnect, we now need an area-efficient way to handle control signals. Since valid signals flow in the same direction as the data, generating hardware for valid channels follows the same strategy as that used for the data channels. However, since ready signals flow in a direction opposite to that of the data channels, we need a way to merge ready signals at the fan-in point. A naive solution is to implement a lookup table (LUT) that encodes the fan-in information, as shown in Fig. [\[fig:ready-valid-join\]](#fig:ready-valid-join){reference-type="ref" reference="fig:ready-valid-join"}. During configuration, we statically encode the ready signal joining logic into the LUT. However, building a LUT for each multiplexer is expensive and it also bloats the configuration space. To optimize this logic, we leverage the fact that the data multiplexers in the switch box are AOI multiplexers, which internally use a decoder to convert the mux selection bits into a one-hot vector, which represents the active routing information. The ready signal joining logic needs this routing information to properly join multiple ready signals. We can reuse these one-hot decoder signals to compute the joining logic without introducing expensive LUT based solutions, as shown at the bottom of Fig. [\[fig:ready-valid-join\]](#fig:ready-valid-join){reference-type="ref" reference="fig:ready-valid-join"}. In this figure, the one-hot selection signals are \(n\_sel\_oh\), \(w\_sel\_oh\), and \(s\_sel\_oh\). We know that if bit 2 of \(n\_sel\_oh\) is high, than the \(data\_in\) is routed to the north. We OR the inversion of \(n\_sel\_oh[2]\) with the \(n\_ready\) signal, giving us a signal that is high when \(n\_ready\) is high or when that route is not used. We can repeat this logic for the remaining two directions and AND them all together to form the final \(ready\) output. Another challenge when generating a statically configured NoC is that a ready-valid NoC needs FIFOs present in the interconnect to buffer data when a downstream tile is not yet ready. While we can easily generate a fixed-size FIFO for a register IR node, the area cost of those FIFOs can be quite high, as shown in Fig [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}. Therefore, we also need a way to reduce FIFO size. To reduce the area overhead introduced by FIFOs while maintaining backward compatibility with a static interconnect, we realize that we can combine two registers from adjacent tiles into a single size-two FIFO. We call this a split FIFO. The first register's FIFO control signals are passed from the first tile into the second tile and its register, as shown in Figure [\[fig:split-fifo\]](#fig:split-fifo){reference-type="ref" reference="fig:split-fifo"}. We can also chain more registers together into a deeper FIFO using the same logic. Since each register's FIFO logic is slightly different depending on its location inside the FIFO pipeline, we need to configure them differently during place and route. The drawback, however, is that these control signals cannot be registered at the tile boundary; the longer the FIFO is chained, the longer the combinational delay on the path. However, if most of the target applications do not require a deep FIFO between different nodes, using this scheme can significantly reduce the silicon area. The area impact of this optimization is evaluated in Fig. [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}. After the graph is translated into RTL, Canal verifies structural correctness by comparing the connectivity of the hardware with that of the IR by parsing the generated RTL. In addition, Canal also has a built in configuration sweep test suite that exhaustively tests every possible connection in IR on the CGRA. This ensures correctness of the design. The methodology described here also applies to generating dynamic NoCs. Instead of lowering a node into a configurable multiplexer to select among incoming data tracks, we can generate a router whose routing table is computed based on the same connectivity information.
## Place and Route using Canal Interconnect
This section describes how the Canal system integrates with a PnR algorithm backend (see Fig. [\[fig:system-diagram\]](#fig:system-diagram){reference-type="ref" reference="fig:system-diagram"}) to enable running applications on a given interconnect. During the translation from a Canal program into the directed graph representation, information regarding important hardware characteristics, like core or wire delays, can be embedded into the graph as shown in Fig. [\[fig:pnr\]](#fig:pnr){reference-type="ref" reference="fig:pnr"}. The Canal system then executes PnR in three stages: packing, placement, and routing. The remainder of this section describes the PnR backend we use in our results. During the packing stage, both the interconnect graph and the application graph (represented as a dataflow graph) are loaded into the PnR tool. Constants and registers in the application are analyzed to identify any packing opportunities. For example, a pipeline register that feeds directly into a PE can be packed within that PE, eliminating the need to place that register on the configurable interconnect. After packing, the placement tool places the tiles in the application onto the interconnect in two stages: global placement and detailed placement. Global placement uses an analytical algorithm that leverages the standard conjugate gradient method on the summation of the cost of each net (Equation [\[eqn:globalplacement-function\]](#eqn:globalplacement-function){reference-type="ref" reference="eqn:globalplacement-function"}) . The cost of a net is the combination of its half-perimeter wire length (HPWL) and a legalization term for memory tiles. In global placement, we use L2 distance to approximate the HPWL to speed up the algorithm. CGRAs typically have fewer rows or columns of memory tiles (compared to PE tiles), so the legalization term is needed to ensure that memory tiles are only placed in those rows/columns. \[\label{eqn:globalplacement-function} \text{Cost}_{net} = \text{HPWL}_{net, estimate} + \text{MEM}_{potential}\] After global placement, we perform detailed placement based on simulated annealing . The cost function for simulated annealing (see Equation [\[eqn:detailedplacement-function\]](#eqn:detailedplacement-function){reference-type="ref" reference="eqn:detailedplacement-function"}) is the total wirelength of the application, calculated by summing the HPWL cost for each net, and an additional term for penalizing pass-through tiles. Pass-through tiles are those that are only used for routing, which need to be powered on despite not computing anything for the application. \(\gamma\) and \(\alpha\) are hyperparameters. Some CGRAs have tile-level power gating that enables the ability to turn off tiles that are used for neither the application's computation nor as pass-through tiles. A higher value of \(\gamma\) penalizes pass-through tiles more, which encourages the placement algorithm to use already-used tiles for routing, rather than powering on otherwise unused tiles. A higher value for \(\alpha\) will penalize longer potential routes, thereby encouraging shorter critical paths after routing. We find that sweeping \(\alpha\) from 1 to 20 and choosing the best result post-routing results in short application critical paths. \[\label{eqn:detailedplacement-function} \text{Cost}_{net} = (\text{HPWL}_{net}-\gamma \times (\text{Area}_{net} \cap \text{Area}_{existing}))^\alpha\] After global and detailed placement, we route using an iteration-based routing algorithm . During each iteration, we compute the slack on a net and determine how critical it is given global timing information. Then we route using the A\* algorithm on the weighted graph. The weights for each edge are based on historical usage, net slack, and current congestion. This allows us to balance both routing congestion and timing criticality. Similarly to detailed placement, we also adjust the wire cost functions to discourage the use of unused tiles in favor of tracks within already-used tiles. We finish routing when a legal routing result is produced.
# Evaluation {#sec:results}
We evaluate the Canal system by first exploring the optimizations of interconnect FIFOs described in Section [8.3](#sec:compile-hardware){reference-type="ref" reference="sec:compile-hardware"} and then by using the Canal system to conduct design space exploration of a CGRA interconnect.
## Interconnect FIFO Optimizations {#sec:results-fifo}
We evaluate the effect of introducing FIFOs in the interconnect on switch box area. As described in Section [8.3](#sec:compile-hardware){reference-type="ref" reference="sec:compile-hardware"}, we need to include FIFOs in the configurable routing when running applications with ready-valid signaling. As a baseline, we compare against a fully static interconnect with five 16-bit routing tracks containing PEs with two outputs and four inputs, synthesized in Global Foundries 12 nm technology. As shown in Fig. [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}, adding these depth two FIFOs to the baseline design introduces a 54% area overhead. Splitting the FIFO between multiple switch boxes results in only a 32% area overhead over the baseline. This optimization allows for much more efficient implementation of an interconnect that supports ready-valid signaling.
## Interconnect Design Space Exploration
We use Canal to explore three important design space axes of a configurable interconnect: switch box topology, number of routing tracks, and number of switch box and connection box port connections. We find that Canal's automation greatly simplifies the procedure to explore each option in the following subsections.
### Exploring Switch Box Topologies and Number of Routing Tracks
The switch box topology defines how each track on each side of the switch box connects to the tracks on the remaining sides of the switch box. The choice of topology affects how easily nets can be routed on the interconnect. High routability generally corresponds to shorter routes and shorter critical paths in applications. This allows the CGRA to be run at higher frequencies, which decreases application run time. For these experiments we investigate two different switch box topologies illustrated in Fig. [\[fig:sb-topos\]](#fig:sb-topos){reference-type="ref" reference="fig:sb-topos"}: Wilton and Disjoint . These switch box topologies have the same area, as they both connect each input to each of the other sides once. We found that the Wilton topology performs much better than the Disjoint topology, which failed to route in all of our test cases. The Disjoint topology is worse for routability because every incoming connection on track \(i\) has a connection only to track \(i\) on the three other sides of the SB. This imposes a restriction that if you want to route a wire from any point on the array to any other point on the array starting from a certain track number, you must only use that track number. In comparison, the Wilton topology does not have this restriction resulting in many more choices for the routing algorithm, and therefore much higher routability . We also vary the number of routing tracks in the interconnect. This directly affects the size of both the connection box and switch box and the amount of routing congestion. For these experiments we measure the area of the connection box and switch box as well as the run time of applications running on the CGRA. In this experiment we use an interconnect with five 16-bit tracks and PE tiles that have 4 inputs and 2 outputs. As shown in Fig. [\[fig:sb-topo-tracks-area\]](#fig:sb-topo-tracks-area){reference-type="ref" reference="fig:sb-topo-tracks-area"}, the area of both the switch box and connection box scale with the number of tracks. From Fig. [\[fig:sb-topo-tracks-runtime\]](#fig:sb-topo-tracks-runtime){reference-type="ref" reference="fig:sb-topo-tracks-runtime"}, we can see that the run time of the applications generally decreases as the number of tracks increases, although the benefits are less than 25%.
### Exploring Switch Box and Connection Box Port Connections
Finally, we explore how varying the number of switch box and connection box port connections affects the area of the interconnect and the run time of applications executing on the CGRA. In Canal, we have the ability to specify how many of the incoming tracks from each side of the tile are connected to the inputs/outputs of the PE/MEM cores. Decreasing these connections should reduce the area of the interconnect, but may decrease the number of options that the routing algorithm has. For these experiments, we vary the number of connections from the incoming routing tracks through the connection box to the inputs of the PE/MEM core, and vary the number of connections from the outputs of the core to the outgoing ports of the switch box. At maximum, we can have 4 SB sides, with connections from the core output to the four sides of the switch box. We then decrease this by removing the connections facing east for a total of three sides with connections, and finally we also remove the connections facing south for a total of two sides with connections. This is shown in Fig. [\[fig:sb-conns\]](#fig:sb-conns){reference-type="ref" reference="fig:sb-conns"}. We do the same for the connection box. As shown in Fig. [\[fig:port-conn-area\]](#fig:port-conn-area){reference-type="ref" reference="fig:port-conn-area"}, as the number of connections from the core to the switch box decreases, we see a decrease in switch box area. From Fig. [\[fig:port-conn-sb-runtime\]](#fig:port-conn-sb-runtime){reference-type="ref" reference="fig:port-conn-sb-runtime"}, we can see that this generally has a small negative effect on the run time of the applications. In this case, a designer could choose to trade some performance for a decrease in switch box area. As shown in Fig. [\[fig:port-conn-area\]](#fig:port-conn-area){reference-type="ref" reference="fig:port-conn-area"}, as the number of connections from the connection box to the tile inputs decreases, we see a larger decrease in connection box area. From Fig. [\[fig:port-conn-cb-runtime\]](#fig:port-conn-cb-runtime){reference-type="ref" reference="fig:port-conn-cb-runtime"}, we can see that this has a larger negative effect on the run time of the applications. Again, a designer could choose to trade some performance for a decrease in connection box area.
# Conclusion
We have developed Canal, a domain-specific language and interconnect generator for coarse-grained reconfigurable arrays. The Canal language allows a designer to easily specify a complex configurable interconnect, while maintaining control over the low-level connections. The hardware generator, placer and router, and bitstream generator help to facilitate design space exploration of CGRA interconnects. We demonstrate the flexibility of Canal by creating a hybrid ready-valid interconnect and demonstrate the design space exploration capabilities of Canal by evaluating different switch box topologies, number of interconnect routing tracks, and number of SB and CB port connections. The power and flexibility that Canal provides will enable more designers to create and explore diverse and interesting CGRA architectures.
# Introduction
Coarse-grained reconfigurable arrays (CGRAs) have been studied heavily in recent years as a promising configurable accelerator architecture . The end of Moore's law necessitates the creation of specialized hardware accelerators to enable running increasingly complex image processing and machine learning applications. While a variety of hardware accelerator architectures exist, CGRAs have emerged as an interesting midpoint between the flexibility of an FPGA and the performance of an application-specific accelerator. A CGRA can achieve high energy efficiency and performance due to word-level arithmetic operations and interconnect, while maintaining enough flexibility to run a variety of applications that evolve over time. CGRAs, as well as other spatial accelerator architectures, often have hundreds of compute cores and memory cores. These compute cores (called processing elements or PEs) and memory (MEM) cores are laid out spatially in a grid of tiles and are connected through a configurable interconnect. An example is shown in Fig. [\[fig:cgra-diagram\]](#fig:cgra-diagram){reference-type="ref" reference="fig:cgra-diagram"}. The reconfigurable interconnect contains switch boxes (SBs), which connect the PE/MEM outputs to the tracks in the interconnect, and connection boxes (CBs), which connect the interconnect tracks to the inputs of the cores. While having a large number of compute cores enables very high performance, the reconfigurable interconnect connecting these cores can constitute over 50% of the CGRA area and 25% of the CGRA energy. Design space exploration of the interconnect is necessary to achieve high performance with lower energy and area costs. There are many interconnect design choices that directly impact the power, performance, and area of the resulting accelerator, including the number and bitwidth of tracks in the interconnect, how the processing elements and memories are arranged in the array, how those elements are connected, and how the interconnect is configured. An agile approach for specifying and generating the interconnect is needed for efficient design space exploration. In this paper, we present Canal, a Python-embedded domain-specific language (eDSL) and compiler for specifying and generating reconfigurable interconnects for CGRAs using a graph-based intermediate representation. The major contributions of our paper are:
1. We describe a graph-based intermediate representation (IR) for CGRA interconnects that is capable of representing and generating a variety of topologies.
2. We propose an embedded domain-specific language (eDSL) called Canal that can compile an interconnect architecture specification into the graph-based IR.
3. We propose an interconnect generator system that can take the IR and automatically produce hardware, place and route collateral, and a bitstream generator.
4. We explore various design space choices using Canal and demonstrate its effectiveness in generating an efficient CGRA design.
# Related Work
Previous attempts have been made to create an interconnect generator for CGRAs, but new demands in CGRA design necessitate a more flexible and powerful system. VPR is one of the most established FPGA architecture research tools . It allows users to adjust various design aspects of the FPGA and observe the effects on final application performance, such as timing and area usage. However, VPR does not offer an RTL generator and users have to design their FPGA independently and hand-write the VPR architecture file accordingly. CGRA-ME is a CGRA architecture research tool similar to Canal in that it also offers integrated RTL generation and place and route tools. One of the major differences is the architecture specification. CGRA-ME opts for a more rigid XML-based input whereas Canal takes in a Python eDSL program, which is more flexible and readable. FastCGRA is a similar CGRA architecture exploration tool that uses an eDSL to construct the hardware. However, with FastCGRA, users explicitly construct multiplexers and switches, from which RTL is generated. Canal abstracts away the notion of hardware primitives and lets the compiler backend choose how to generate the hardware. In summary, Canal supports more tools, allows for more flexibility, and enables easier design space exploration than previous attempts at an interconnect generator.
# System Design
In this section, we introduce the design of the Canal system, including the graph-based IR for representing interconnects, the Canal eDSL, the static interconnect generation used to translate the IR into hardware, and finally how that IR interfaces with the application place and route (PnR) algorithms. Fig. [\[fig:system-diagram\]](#fig:system-diagram){reference-type="ref" reference="fig:system-diagram"} summarizes how the Canal interconnect generator interfaces with the PE and memory core designs, application PnR, RTL generation, and bitstream generation.
## Graph-Based Intermediate Representation
The primitives in Canal's intermediate representation (IR) are nodes, which represent anything that can be connected in the underlying hardware, and edges, which are wires connecting the nodes together. An example of the IR for a switch box is shown in Fig. [\[fig:graph-representation\]](#fig:graph-representation){reference-type="ref" reference="fig:graph-representation"}. All edges are unidirectional so the IR represents a directed graph. Nodes in the graph can have multiple incoming edges which, when translated into hardware, transform into multiplexers. Each node also has attributes that provide additional information for type checking and hardware generation. This intermediate representation is flexible enough to represent a wide variety of interconnect topologies and handle an arbitrarily complex set of CGRA cores. We will discuss a few design space exploration experiments that exploit this flexibility in Section [9](#sec:results){reference-type="ref" reference="sec:results"}.
## The Canal Language
The Canal language is a Python-embedded domain-specific language (eDSL) that constructs the interconnect intermediate representation described in the previous section. The Canal language translates the Python description of an interconnect into this IR, so at the lowest level, a designer could instantiate nodes in the Canal language and wire them together. As Canal is embedded in Python, we have also built a layer on top of the basic primitives of Canal, which simplifies the IR construction. For instance, for creating a uniform interconnect (all switch boxes have the same topology) with no diagonal connections, we provide a simple helper function that produces different interconnect topologies by varying function parameters such as height and width of the array, switch box topology, number of tracks, bit width of tracks, and density of pipeline registers. An example of creating a low level Canal node and using a higher level Canal helper function is shown in Fig. [\[fig:canal_example\]](#fig:canal_example){reference-type="ref" reference="fig:canal_example"}. The Canal eDSL allows designers to easily conduct design space exploration by varying the parameters in the helper functions, or by generating an entirely new interconnect. Canal can easily be integrated with other DSLs. For example, one could use a DSL for specifying individual CGRA tiles and then integrate them together using Canal to generate the interconnect.
## Generating Interconnect Hardware {#sec:compile-hardware}
Because Canal's IR only describes the connectivity among the different nodes, it is up to the hardware compiler backend to decide how to lower the IR. We implemented two different hardware compiler backends that lower the IR into
. This NoC has data channels along with a ready-valid interface and routing is configured statically before the application runs. We use magma as our hardware circuit implementation, but this could be extended to any hardware generator framework. To generate a static mesh interconnect, we adopt the following principles to generate hardware:
1. Nodes with hardware attributes (e.g. a processing element core) generate the specified hardware.
2. Directed edges are translated into wires.
3. Nodes with multiple incoming edges generate multiplexers.
We also use attributes associated with each node to lower the node to different hardware components. For instance, a register node will be lowered into a physical register. A port node will be lowered to a CB (with an internal multiplexer) where the output of the CB connects to the port of the core. These translations are mechanical and can be accomplished through a compiler pass. Reusing the same IR to generate a statically configured NoC has several challenges. First, the application graph may have fanouts, that is, one output port of a node is connected to multiple input ports. While this is simple to handle in a static interconnect, we now need an area-efficient way to handle control signals. Since valid signals flow in the same direction as the data, generating hardware for valid channels follows the same strategy as that used for the data channels. However, since ready signals flow in a direction opposite to that of the data channels, we need a way to merge ready signals at the fan-in point. A naive solution is to implement a lookup table (LUT) that encodes the fan-in information, as shown in Fig. [\[fig:ready-valid-join\]](#fig:ready-valid-join){reference-type="ref" reference="fig:ready-valid-join"}. During configuration, we statically encode the ready signal joining logic into the LUT. However, building a LUT for each multiplexer is expensive and it also bloats the configuration space. To optimize this logic, we leverage the fact that the data multiplexers in the switch box are AOI multiplexers, which internally use a decoder to convert the mux selection bits into a one-hot vector, which represents the active routing information. The ready signal joining logic needs this routing information to properly join multiple ready signals. We can reuse these one-hot decoder signals to compute the joining logic without introducing expensive LUT based solutions, as shown at the bottom of Fig. [\[fig:ready-valid-join\]](#fig:ready-valid-join){reference-type="ref" reference="fig:ready-valid-join"}. In this figure, the one-hot selection signals are \(n\_sel\_oh\), \(w\_sel\_oh\), and \(s\_sel\_oh\). We know that if bit 2 of \(n\_sel\_oh\) is high, than the \(data\_in\) is routed to the north. We OR the inversion of \(n\_sel\_oh[2]\) with the \(n\_ready\) signal, giving us a signal that is high when \(n\_ready\) is high or when that route is not used. We can repeat this logic for the remaining two directions and AND them all together to form the final \(ready\) output. Another challenge when generating a statically configured NoC is that a ready-valid NoC needs FIFOs present in the interconnect to buffer data when a downstream tile is not yet ready. While we can easily generate a fixed-size FIFO for a register IR node, the area cost of those FIFOs can be quite high, as shown in Fig [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}. Therefore, we also need a way to reduce FIFO size. To reduce the area overhead introduced by FIFOs while maintaining backward compatibility with a static interconnect, we realize that we can combine two registers from adjacent tiles into a single size-two FIFO. We call this a split FIFO. The first register's FIFO control signals are passed from the first tile into the second tile and its register, as shown in Figure [\[fig:split-fifo\]](#fig:split-fifo){reference-type="ref" reference="fig:split-fifo"}. We can also chain more registers together into a deeper FIFO using the same logic. Since each register's FIFO logic is slightly different depending on its location inside the FIFO pipeline, we need to configure them differently during place and route. The drawback, however, is that these control signals cannot be registered at the tile boundary; the longer the FIFO is chained, the longer the combinational delay on the path. However, if most of the target applications do not require a deep FIFO between different nodes, using this scheme can significantly reduce the silicon area. The area impact of this optimization is evaluated in Fig. [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}. After the graph is translated into RTL, Canal verifies structural correctness by comparing the connectivity of the hardware with that of the IR by parsing the generated RTL. In addition, Canal also has a built in configuration sweep test suite that exhaustively tests every possible connection in IR on the CGRA. This ensures correctness of the design. The methodology described here also applies to generating dynamic NoCs. Instead of lowering a node into a configurable multiplexer to select among incoming data tracks, we can generate a router whose routing table is computed based on the same connectivity information.
## Place and Route using Canal Interconnect
This section describes how the Canal system integrates with a PnR algorithm backend (see Fig. [\[fig:system-diagram\]](#fig:system-diagram){reference-type="ref" reference="fig:system-diagram"}) to enable running applications on a given interconnect. During the translation from a Canal program into the directed graph representation, information regarding important hardware characteristics, like core or wire delays, can be embedded into the graph as shown in Fig. [\[fig:pnr\]](#fig:pnr){reference-type="ref" reference="fig:pnr"}. The Canal system then executes PnR in three stages: packing, placement, and routing. The remainder of this section describes the PnR backend we use in our results. During the packing stage, both the interconnect graph and the application graph (represented as a dataflow graph) are loaded into the PnR tool. Constants and registers in the application are analyzed to identify any packing opportunities. For example, a pipeline register that feeds directly into a PE can be packed within that PE, eliminating the need to place that register on the configurable interconnect. After packing, the placement tool places the tiles in the application onto the interconnect in two stages: global placement and detailed placement. Global placement uses an analytical algorithm that leverages the standard conjugate gradient method on the summation of the cost of each net (Equation [\[eqn:globalplacement-function\]](#eqn:globalplacement-function){reference-type="ref" reference="eqn:globalplacement-function"}) . The cost of a net is the combination of its half-perimeter wire length (HPWL) and a legalization term for memory tiles. In global placement, we use L2 distance to approximate the HPWL to speed up the algorithm. CGRAs typically have fewer rows or columns of memory tiles (compared to PE tiles), so the legalization term is needed to ensure that memory tiles are only placed in those rows/columns. \[\label{eqn:globalplacement-function} \text{Cost}_{net} = \text{HPWL}_{net, estimate} + \text{MEM}_{potential}\] After global placement, we perform detailed placement based on simulated annealing . The cost function for simulated annealing (see Equation [\[eqn:detailedplacement-function\]](#eqn:detailedplacement-function){reference-type="ref" reference="eqn:detailedplacement-function"}) is the total wirelength of the application, calculated by summing the HPWL cost for each net, and an additional term for penalizing pass-through tiles. Pass-through tiles are those that are only used for routing, which need to be powered on despite not computing anything for the application. \(\gamma\) and \(\alpha\) are hyperparameters. Some CGRAs have tile-level power gating that enables the ability to turn off tiles that are used for neither the application's computation nor as pass-through tiles. A higher value of \(\gamma\) penalizes pass-through tiles more, which encourages the placement algorithm to use already-used tiles for routing, rather than powering on otherwise unused tiles. A higher value for \(\alpha\) will penalize longer potential routes, thereby encouraging shorter critical paths after routing. We find that sweeping \(\alpha\) from 1 to 20 and choosing the best result post-routing results in short application critical paths. \[\label{eqn:detailedplacement-function} \text{Cost}_{net} = (\text{HPWL}_{net}-\gamma \times (\text{Area}_{net} \cap \text{Area}_{existing}))^\alpha\] After global and detailed placement, we route using an iteration-based routing algorithm . During each iteration, we compute the slack on a net and determine how critical it is given global timing information. Then we route using the A\* algorithm on the weighted graph. The weights for each edge are based on historical usage, net slack, and current congestion. This allows us to balance both routing congestion and timing criticality. Similarly to detailed placement, we also adjust the wire cost functions to discourage the use of unused tiles in favor of tracks within already-used tiles. We finish routing when a legal routing result is produced.
# Evaluation {#sec:results}
We evaluate the Canal system by first exploring the optimizations of interconnect FIFOs described in Section [8.3](#sec:compile-hardware){reference-type="ref" reference="sec:compile-hardware"} and then by using the Canal system to conduct design space exploration of a CGRA interconnect.
## Interconnect FIFO Optimizations {#sec:results-fifo}
We evaluate the effect of introducing FIFOs in the interconnect on switch box area. As described in Section [8.3](#sec:compile-hardware){reference-type="ref" reference="sec:compile-hardware"}, we need to include FIFOs in the configurable routing when running applications with ready-valid signaling. As a baseline, we compare against a fully static interconnect with five 16-bit routing tracks containing PEs with two outputs and four inputs, synthesized in Global Foundries 12 nm technology. As shown in Fig. [\[fig:sb-fifo-area\]](#fig:sb-fifo-area){reference-type="ref" reference="fig:sb-fifo-area"}, adding these depth two FIFOs to the baseline design introduces a 54% area overhead. Splitting the FIFO between multiple switch boxes results in only a 32% area overhead over the baseline. This optimization allows for much more efficient implementation of an interconnect that supports ready-valid signaling.
## Interconnect Design Space Exploration
We use Canal to explore three important design space axes of a configurable interconnect: switch box topology, number of routing tracks, and number of switch box and connection box port connections. We find that Canal's automation greatly simplifies the procedure to explore each option in the following subsections.
### Exploring Switch Box Topologies and Number of Routing Tracks
The switch box topology defines how each track on each side of the switch box connects to the tracks on the remaining sides of the switch box. The choice of topology affects how easily nets can be routed on the interconnect. High routability generally corresponds to shorter routes and shorter critical paths in applications. This allows the CGRA to be run at higher frequencies, which decreases application run time. For these experiments we investigate two different switch box topologies illustrated in Fig. [\[fig:sb-topos\]](#fig:sb-topos){reference-type="ref" reference="fig:sb-topos"}: Wilton and Disjoint . These switch box topologies have the same area, as they both connect each input to each of the other sides once. We found that the Wilton topology performs much better than the Disjoint topology, which failed to route in all of our test cases. The Disjoint topology is worse for routability because every incoming connection on track \(i\) has a connection only to track \(i\) on the three other sides of the SB. This imposes a restriction that if you want to route a wire from any point on the array to any other point on the array starting from a certain track number, you must only use that track number. In comparison, the Wilton topology does not have this restriction resulting in many more choices for the routing algorithm, and therefore much higher routability . We also vary the number of routing tracks in the interconnect. This directly affects the size of both the connection box and switch box and the amount of routing congestion. For these experiments we measure the area of the connection box and switch box as well as the run time of applications running on the CGRA. In this experiment we use an interconnect with five 16-bit tracks and PE tiles that have 4 inputs and 2 outputs. As shown in Fig. [\[fig:sb-topo-tracks-area\]](#fig:sb-topo-tracks-area){reference-type="ref" reference="fig:sb-topo-tracks-area"}, the area of both the switch box and connection box scale with the number of tracks. From Fig. [\[fig:sb-topo-tracks-runtime\]](#fig:sb-topo-tracks-runtime){reference-type="ref" reference="fig:sb-topo-tracks-runtime"}, we can see that the run time of the applications generally decreases as the number of tracks increases, although the benefits are less than 25%.
### Exploring Switch Box and Connection Box Port Connections
Finally, we explore how varying the number of switch box and connection box port connections affects the area of the interconnect and the run time of applications executing on the CGRA. In Canal, we have the ability to specify how many of the incoming tracks from each side of the tile are connected to the inputs/outputs of the PE/MEM cores. Decreasing these connections should reduce the area of the interconnect, but may decrease the number of options that the routing algorithm has. For these experiments, we vary the number of connections from the incoming routing tracks through the connection box to the inputs of the PE/MEM core, and vary the number of connections from the outputs of the core to the outgoing ports of the switch box. At maximum, we can have 4 SB sides, with connections from the core output to the four sides of the switch box. We then decrease this by removing the connections facing east for a total of three sides with connections, and finally we also remove the connections facing south for a total of two sides with connections. This is shown in Fig. [\[fig:sb-conns\]](#fig:sb-conns){reference-type="ref" reference="fig:sb-conns"}. We do the same for the connection box. As shown in Fig. [\[fig:port-conn-area\]](#fig:port-conn-area){reference-type="ref" reference="fig:port-conn-area"}, as the number of connections from the core to the switch box decreases, we see a decrease in switch box area. From Fig. [\[fig:port-conn-sb-runtime\]](#fig:port-conn-sb-runtime){reference-type="ref" reference="fig:port-conn-sb-runtime"}, we can see that this generally has a small negative effect on the run time of the applications. In this case, a designer could choose to trade some performance for a decrease in switch box area. As shown in Fig. [\[fig:port-conn-area\]](#fig:port-conn-area){reference-type="ref" reference="fig:port-conn-area"}, as the number of connections from the connection box to the tile inputs decreases, we see a larger decrease in connection box area. From Fig. [\[fig:port-conn-cb-runtime\]](#fig:port-conn-cb-runtime){reference-type="ref" reference="fig:port-conn-cb-runtime"}, we can see that this has a larger negative effect on the run time of the applications. Again, a designer could choose to trade some performance for a decrease in connection box area.
# Conclusion
We have developed Canal, a domain-specific language and interconnect generator for coarse-grained reconfigurable arrays. The Canal language allows a designer to easily specify a complex configurable interconnect, while maintaining control over the low-level connections. The hardware generator, placer and router, and bitstream generator help to facilitate design space exploration of CGRA interconnects. We demonstrate the flexibility of Canal by creating a hybrid ready-valid interconnect and demonstrate the design space exploration capabilities of Canal by evaluating different switch box topologies, number of interconnect routing tracks, and number of SB and CB port connections. The power and flexibility that Canal provides will enable more designers to create and explore diverse and interesting CGRA architectures. | {'timestamp': '2022-12-01T02:20:26', 'yymm': '2211', 'arxiv_id': '2211.17207', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17207'} |
null | null |
# Introduction {#section:intro}
By Alexander's theorem, every link in \(S^3\) can be realized as the closure of some braid. Some properties of braids can be ported into important topological properties of their closures. For example, if a braid \(\beta\) is *homogeneous* (i.e. for every \(i\), every \(\sigma_i\) appearing in \(\beta\) has the same sign), then the closure is a fibered link in \(S^3\). In a different vein, the closures of *quasipositive braids* are exactly the links arising as the transverse intersections of smooth algebraic curves \(f^{-1}(0) \subset \mathbb{C}^2\) (here, \(f\) is a non-constant polynomial) with the unit sphere \(S^3 \subset \mathbb{C}^2\). Another property of braids with important topological consequences is *twist positivity*:
A full twist on three strands is seen in ; here, \(\Delta^2 = (\sigma_2 \sigma_1)^3\). Twist positivity is a well studied property: perhaps most notably, Morton and Franks-Williams proved that if \(K\) is twist positive on \(n\) strands, then the braid index of \(K\) is \(n\). In this paper, we show that twist positivity unifies classical results in knot theory with modern concordance invariants. Our first theorem is about the Alexander polynomial of twist positive knots.
The strength and utility of is leveraged by way of the connection between the Alexander polynomial and knot Floer homology: not only does knot Floer homology categorify the Alexander polynomial, but it also provides a suite of powerful knot invariants. We recall: an *L-space knot* in \(S^3\) is a non-trivial knot which admits a positive Dehn surgery to an *L-space*, the closed, connected, oriented 3-manifolds with simple Heegaard Floer homology. For L-space knots, both the Alexander polynomial and knot Floer homology take on a restricted form. This connection leads to our first application of .
Lorenz knots exhibit rich phenomena from the dynamical, geometric, and braid theoretic perspectives. They have been intensely studied since their introduction in 1963; see, for example,, or for a survey. In , we observe L-space Lorenz knots include the well-studied *algebraic knots*, *1-bridge braids*, and some *twisted torus knots*, amongst others. Therefore, generalizes results of both Juhász-Miller-Zemke (who prove for torus knots) and Hom-Lidman-Park, who use input from bordered Floer homology to prove it for algebraic knots. Our results suggests that twist positivity is the important topological property guiding this behavior. Birman-Williams noticed that Lorenz knots are always twist positive on some number of strands, and conjectured that this quantity was actually the braid index. Indeed, this conjecture follows with the subsequent announcement of the Franks-Williams result. To the best of our knowledge, the bridge index of Lorenz knots has not been investigated. proves that often, the bridge index and the braid index for a Lorenz knot agree, hence the Lorenz presentation cannot be minimized further to decrease the bridge index. and suggest that twist positivity is not only an important 3-dimensional property, but that it also provides some insights into 3.5-and 4-dimensional phenomena. Indeed, we next study concordance. Baker observed a novel relationship between concordance properties of fibered knots and Fox's *slice-ribbon conjecture*, which asks whether every slice knot admits a ribbon presentation in \(S^3\). In particular, Baker conjectured that *if two fibered, strongly quasipositive knots are concordant, then they are isotopic*. Baker observes that either his conjecture is true, or the Slice-Ribbon conjecture is false[^1]. Therefore, studying the concordance properties of fibered, strongly quasipositive knots is of central interest. L-space knots form a prominent class of fibered, strongly quasipositive knots. As noted by Abe-Tagami, if Baker's conjecture is true, then every concordance class should have at most one L-space knot.
is proved by using the braid index, \(\tau(K)\), and the Seifert genus in concert. Baader-Dehorney-Liechti proved that only finitely many positive braid knots can be concordant to each other (c.f. ). However, to the best of our knowledge, this is the first explicit infinite family of positive braid knots that we know are distinct in concordance, where as \(g\) tends to infinity, the number of *hyperbolic* knots of genus \(g\) gets arbitrarily large. While L-space knots are well-studied, a classification remains elusive; nevertheless, collecting their properties remains a central goal. L-space knots are "simple" from many perspectives: they have simplest possible knot Floer homology, "simple" knot exteriors (their exteriors are realized as mapping tori), and, heuristically, they seem to have "simple" monodromy. The bridge index and the braid index provide different measures of complexity of a knot, as the former is a lower bound on the latter. The difference between these quantities is itself a measure of complexity. We conjecture that this defect is trivial for L-space knots.
This conjecture is true for the most well known families of L-space knots:
- The only L-space two-bridge knots are \(T(2,n)\) torus knots..
- The only L-space Montesinos knots are the \(P(-2,3,q)\), \(q \geq 1\) pretzel knots. They are all twist positive on either two or three strands, hence applies.
- As proved in this article, the braid and bridge index agree for *1-bridge braids*.
- Baker-Kegel constructed an infinite family of hyperbolic L-space knots where the first knot is provably not braid positive, but the braid positivity status of the remaining knots remains unknown. We prove for the Baker-Kegel family in .
- Combining work of Hedden and Hom, it is known that \(K_{p,q}\) is an L-space knot if and only if \(K\) is and \(q \geq p(2g(K)-1)\). In , we show that if is true for an L-space knot \(K\), it is true for all of its L-space cables
The Alexander polynomials of L-space knots have been of interest ever since Ozsváth-Szabó proved that the coefficients of these polynomials take values in \(\{-1,0,1\}\). Hedden-Watson later proved that when \(K\) is an L-space knot, \(\Delta_K(t) = 1-t+\ldots+ t^{2g-1} + t^{2g}\), where \(g\) is the Seifert genus of \(K\). The results we summarized above provide evidence for the following:
Hyperbolicity is a key assumption in : the conjecture is false even for the (2,3)-cable of \(T(2,3)\), which is an L-space knot. In a different direction, we also predict:
is true for all positive braid knots in the knot tables, but quickly becomes false for even slightly broader classes of knots. For example, the fibered, strongly quasipositive knot \(10_{145}\) is not braid positive, and it has braid index is four while the bridge index is three. It is worth noting that there are positive braid knots which do not admit a positive braid representative on a minimal number of strands, indicating a subtlety to .
## Conventions
- We let \(B_n\) denote the braid group on \(n\) strands.
- Unless stated otherwise, we assume that all braid closures are knots.
- When drawn vertically (resp. horizontally), our braids are oriented from north to south (resp. from west to east).
- We write \(\beta_1 \approx \beta_2\) when braids are related by conjugation. We write \(\beta_1 = \beta_2\) if they are related by braid relations. We underline the subword of a braid that is being modified.
- Throughout, we denote the braid (resp. bridge) index of a knot \(K\) by \(i(K)\) (resp. \(br(K)\)).
- We always work within the smooth category. In particular, we only study knots up to smooth concordance.
## Organization
We prove , , and in , , and , respectively. provides evidence for .
# Twist Positivity and the Alexander Polynomial {#AlexPolyResult}
This section establishes a relationship between twist positivity and the Alexander polynomial. **.** *If \(K\) is twist positive, then \(\Delta_K(t) = 1-t + t^n +t^{n+1}R(t)\), where \(R(t) \in \mathbb Z[t]\).* The proof of requires the *reduced Burau matrix*, \(B(t)\), which is associated to a braid \(\beta \in B_n\). We quickly recall the relevant details here, and recommend for a full account. Let \(\beta\) be a braid on \(n\) strands. We write \(\displaystyle \beta =\prod_{r=1}^m \sigma_{i_r}^{\epsilon_{i_r}}\), i.e. as the product of Artin generators To \(\beta\), we can associate its reduced Burau matrix, \(B(t)\), which is the product of \((n-1)\times (n-1)\) matrices \(\sigma_i(t)\), where \(\sigma_i(t)\) differs from the \((n-1)\times (n-1)\) identity matrix \(I_{n-1}\) only in row \(i\), as in . (Note that this matrix is truncated appropriately when \(i=1,n-1\).) In particular, \(\displaystyle B(t)=\prod_{r=1}^m \sigma_{i_r}^{\epsilon_{i_r}}(t)\). Suppose \(\widehat{\beta}=K\), a knot. The Alexander polynomial for \(K\), denoted \(\Delta_K(t)\), is derived from \(B(t)\) as follows: \[\frac{\det(I_{n-1}-B(t))}{1-t^n}=\frac{\Delta_K(t)}{1-t}\] In particular, \[\Delta_K(t)=(1-t)\cdot(\det(I_{n-1}-B(t)))\cdot(1+t^n+t^{2n}+\cdots).\] We collect some important properties of \(B(t)\):
- In general, the entries in a reduced Burau matrix lie in \(\mathbb Z[t^{\pm 1}]\). However, when \(\beta\) is a positive braid, the entries of \(B(t)\) are genuine polynomial entries.
- Recall: the full twist on \(n\) strands is denoted by \(\Delta^2\), where \(\Delta^2 = (\sigma_1 \ldots \sigma_{n-1})^n \approx (\sigma_{n-1} \ldots \sigma_1)^n\). When \(\beta = \Delta^2\), we have \(B(t) = t^n I_{n-1}\). In particular, if \(\beta\) is a twist positive braid, then \(\beta = \Delta^2 \gamma\), so the reduced Burau matrix associated to \(\beta\) is \(B(t) = (t^n I_{n-1})(C(t)) = t^nC(t)\), where \(C(t)\) is the reduced Burau matrix associated with \(\gamma\).
- Suppose \(\beta\) is a positive braid on \(n\) strands. We define the polynomial \(q_{B(t)}(x)\) as follows: \[\begin{aligned}
q_{B(t)}(x):=\det(I_{n-1}-xB(t)) = 1+a_1(t)x+a_2(t)x^2 + \ldots + x^{n-1}a_{n-1}(t) \label{eqn:Burau} \end{aligned}\] We claim that the values of \(a_1(t)\) and \(a_{n-1}(t)\) can be quickly determined from \(B_(t)\): if we let \(p_{B(t)}(x):=\det(xI_{n-1}-B(t))\) denote the characteristic polynomial of \(B(t)\), then \(q_{B(t)}(x)=x^{n-1}p_{B(t)}(x^{-1})\). That is, \(q_{B(t)}(x)\) is obtained by \(p_{B(t)}(x)\) by exchanging the coefficients of \(x^i\) and \(x^{(n-1)-i}\) for each \(i \leq \lfloor (n-1)/2 \rfloor\). In particular, in \(q_{B(t)}(t)\), \(a_{n-1}(t) = (-1)^{n-1}\det(B(t))\) and \(a_1(t)=-\text{tr}(B(t))\).
We observe that substituting \(x=t^n\) into yields: \[\begin{aligned}
\det(I_{n-1}-t^nB(t)) &= 1+a_1(t)t^n+a_2(t)t^{2n} + \ldots + t^{n^2-n}a_{n-1}(t) \nonumber \\ &= 1+a_1(t)t^n+t^{n+1} R(t) \label{eqn:PolyFullTwist} \end{aligned}\] where \(R(t) \in \mathbb Z[t]\) collects all the higher order terms. With these preliminaries established, we can prove the result.
With this claim in hand, we can finish proving . In particular, since \(a_1(0)=0\), then \(a_1(t) = e_1t + e_2t^2 + \ldots + e_kt^k\). Plugging this into , we have \[\begin{aligned}
\Delta_K(t) &= 1-t+(1+a_1(t))t^n + t^{n+1}S(t) \\ &= 1-t+(1 + (e_1t + e_2t + \ldots + e_kt^k))t^n + t^{n+1}S(t) \\ &= 1-t+t^n + t^{n+1}V(t) \end{aligned}\] for some \(V(t) \in \mathbb Z[t]\). This is exactly what we wanted to show. ◻
# Proof of {#section:Lorenz}
First, we establish that some well-studied families of knots are twist positive:
## Twist positive L-space knots
We begin by proving . **.** *If \(K\) is a twist positive L-space knot, then its braid index and bridge index agree. Moreover, \(K\) realizes the minimal braid and bridge indices of its concordance class.*
Let \(\alpha_0, \ldots, \alpha_{2m}\) be the non-zero degrees appearing in \(\Delta_K(t)\), written in decreasing order. Juhász-Miller-Zemke showed that for L-space knots, \(Ord_v(K) = \max \{\alpha_{i-1}-\alpha_i | 1 \leq i \leq 2m \}\) . Since \(n-1\) appears as a gap in the degrees of the exponents of \(\Delta_K(t)\), we have \(n-1 \leq Ord_v(K)\). The bridge index, \(br(K)\), is a lower bound on the braid index, \(i(K)\), so: \[\begin{aligned}
n-1 &\leq Ord_v(K) \leq br(K)-1 \leq i(K)-1 = n-1 \label{eqn:Inequalities} \end{aligned}\] But this means we must have equalities throughout, hence \(br(K) = n = i(K)\), as desired. Next, we prove that twist positive L-space knots realize the minimal braid index of their concordance classes. The proof of this is nearly identical to the proof of. Suppose the knot \(J\) is concordant to a twist positive L-space knot \(K\), and let \(N(K)\) denote the concordance invariant defined by Dai-Hom-Truong-Stoffregen. We have the following string of inequalities (justification for the respective (in)equalities are in the paragraph following): \[\begin{aligned}
i(J) \geq br(J) &\geq Ord_v(J)+1 \\ & \geq N(J) + 1 \\ &= N(K) + 1 \\ &= Ord_v(K) + 1 \\ &= br(K) = i(K) \end{aligned}\] The first equality of (3.2) is standard; the second is a restatement of . Dai-Hom-Truong-Stoffregen showed that \(N(K)\) is a lower bound on \(Ord_v(K)\), yielding the inequality in (3.3). Since \(N(K)\) is a concordance invariant, we have the equality in (3.4). showed that for L-space knots, the invariants \(N(\star)\) and \(Ord_v(\star)\) agree; this yields the equality in (3.5). The equalities of (3.5) follow from those in ([\[eqn:Inequalities\]](#eqn:Inequalities){reference-type="ref" reference="eqn:Inequalities"}). ◻
We briefly justify why we may assume that \(k\) is at least 4: if \(k=1, 2\), then \(T(3,k; 2m)\) is isotopic to a \(T(2,r)\) torus knot. Additionally, for all \(m \geq 0\) and \(k \equiv 0 \mod 3\), \(T(3,k;2m)\) is a link. Thus, we may assume \(k \geq 4\). We note that when \(k \geq 4\), \(T(3,k; 2m)\) is twist positive on three strands.
In the next lemma, we see that it suffices to study \(T(3,k;2m)\) where \(k \equiv 1 \mod 3\).
In fact, our proof of will show that the \(\lfloor \frac{q}{3}\rfloor-1\) twisted torus knots are pairwise distinct from \(T(3,q)\) and from each other.
Our long-term goal is to show that, despite there being many twisted torus knots with the same \(\tau\) invariants, these knots represent distinct concordance classes. This is proved by computing the signatures of 3-stranded twisted torus knots.
## Computing the signature via the Goeritz matrix
To compute the signature, we use the Gordon-Litherland method, which now we recall: let \(\mathcal{K}\) be a knot in \(S^3\), and fix a regular projection \(K\) of \(\mathcal{K}\). First, we checkerboard color \(K\), where the checkerboard has black and white regions (without loss of generality, we assume that the unbounded region is white). Each white region of the diagram is labelled \(X_i\), and we set \(X_0\) to be the unbounded white region. To each double point \(D\) in \(K\), we assign a value \(\eta(D) = \pm 1\), as in . Each double point \(D\) is also assigned a *type*, i.e. it is either *Type I* or *Type II*; see . This information is organized into a symmetric, integral matrix \(G'(K)\), with entries \(g_{ij}\), where: \[\begin{aligned}
g_{ij} = \begin{cases}-\displaystyle \sum \eta(D) & \text{summed over double points \(D\) incident to \(X_i\) and \(X_j\), where \(i \neq j\)} \\ \\ \displaystyle-\sum_{k, k \neq i} g_{ik} & \text{if \(i = j\)} \end{cases} \end{aligned}\] The *Goeritz matrix* \(G(K)\) is constructed from \(G'(K)\) by deleting the \(0^{th}\) row and column of \(G'(K)\). Thus, \(G\) is a symmetric, integral, \(n \times n\) matrix, where \(n\) is the number of bounded white regions in \(K\). Gordon-Litherland proved that \[\begin{aligned}
\label{eqn:GordonLitherland} \sigma(\mathcal{K}) = \text{sign}(G(K))-\mu(K) \end{aligned}\] where \(\mu(K) = \sum \eta(D)\), where \(D\) is a Type II double point in \(K\). We are now ready to compute the signature of \(T(3,q;2m)\) where \(q \geq 4\) and \(m \geq 0\).
The matrices \(G'(K)\) and \(G(K)\) for the twisted torus knot \(T(3,7;6)\) from are below. \[\hspace{-0.5cm} G'(K) = \begin{blockarray}{ccccccccc} & \scriptscriptstyle{\textcolor{gray}{0}} & \scriptscriptstyle{\textcolor{gray}{1}} & \scriptscriptstyle{\textcolor{gray}{2}} & \scriptscriptstyle{\textcolor{gray}{3}} & \scriptscriptstyle{\textcolor{gray}{4}} & \scriptscriptstyle{\textcolor{gray}{5}} & \scriptscriptstyle{\textcolor{gray}{6}} & \scriptscriptstyle{\textcolor{gray}{7}} \\ \begin{block}{c(cccccccc)} \scriptscriptstyle{\textcolor{gray}{0}}&11 &-1 &-1 &-1 &-1 &-1 &-1 &-5 \\ \scriptscriptstyle{\textcolor{gray}{1}}&-1 &-1 & +1 & 0 & 0 & 0 & 0 & +1 \\ \scriptscriptstyle{\textcolor{gray}{2}}&-1 & +1 &-1 & +1 & 0 & 0 & 0 & 0 \\ \scriptscriptstyle{\textcolor{gray}{3}}&-1 & 0 & +1 &-1 & +1 & 0 & 0 & 0 \\ \scriptscriptstyle{\textcolor{gray}{4}}&-1 & 0 & 0 & +1 &-1 & +1 & 0 & 0\\ \scriptscriptstyle{\textcolor{gray}{5}}&-1 & 0 & 0 & 0 & +1 &-1 & +1 & 0 \\ \scriptscriptstyle{\textcolor{gray}{6}}&-1 & 0 & 0 & 0 & 0 & +1 &-1 & +1 \\ \scriptscriptstyle{\textcolor{gray}{7}}&-5 & +1 & 0 & 0 & 0 & 0 & +1 & +3 \\ \end{block} \end{blockarray} \qquad G(K) = \begin{pmatrix} \begin{array}{ccccccc} -1 & +1 & 0 & 0 & 0 & 0 & +1 \\ +1 &-1 & +1 & 0 & 0 & 0 & 0 \\ 0 & +1 &-1 & +1 & 0 & 0 & 0 \\ 0 & 0 & +1 &-1 & +1 & 0 & 0\\ 0 & 0 & 0 & +1 &-1 & +1 & 0 \\ 0 & 0 & 0 & 0 & +1 &-1 & +1 \\ +1 & 0 & 0 & 0 & 0 & +1 & +3 \\ \end{array} \end{pmatrix}\] Our goal is to compute the signature of \(G(K)\). We will need a powerful theorem, typically refereed to as "Sylvester's Law of Inertia": *two symmetric, \(n \times n\) matrices have the same number of positive, negative, and zero eigenvalues if and only if they are congruent*. (Recall that two matrices \(A\) and \(B\) are congruent if there exists a non-singular matrix \(S\) such that \(B = SAS^T\).) We will prove that \(G(K)\) is congruent to a diagonal matrix \(D\). Since, for diagonal matrices, the eigenvalues are exactly the diagonal entries, we can compute the signature of \(G(K)\) from \(D\).
To prove this result, we will first establish some preliminaries.
We note: the first parameter in the subscript identifies the dimension of the matrix, while the second index records the value \(\epsilon\). (right) displays \(N_{1,\epsilon}\) and \(P_{1, \epsilon}\). We note that the entries of \(P_{\ell, \epsilon}\) and \(N_{\ell, \epsilon}\) agree in all but two entries: \(n_{1, 3\ell+1} = n_{3\ell+1,1}=-1\) while \(p_{1, 3\ell+1} = p_{3\ell+1,1}=+1\) (indeed, \(N\) and \(P\) stand for "negative" and "positive", respectively).
*Proof of :* We begin by showing \(N_{1,\epsilon}\) can be diagonalized into the claimed form. The sequence of simultaneous row and column operations is written out explicitly in . In Step 1 and 2, we simultaneously clear the non-zero, non-diagonal entries in the first row/column. In Step 3, we simultaneously swap the second and third rows and columns. In Steps 4 and 5, we clear the non-zero, non-diagonal entries of the second row/column. The analogous sequence of operations used to diagonalize \(P_(1,\epsilon)\) are in . \(\Box\) This proves . Next, we establish the base cases required to perform strong induction.
*Proof of :* We first prove the claim for \(P_{2,\epsilon}\). We perform a sequence of elementary matrix operations in . In Step 1, we simultaneously clear the non-zero, non-diagonal entries of the first row and column. In Step 2, we swap the second and third rows and columns. In Step 3, we clear the non-zero, non-diagonal entries of the second row and column. In Step 4, we clear the non-zero, non-diagonal entries of the third row and column. The lower right block of the final matrix is exactly \(N_1\). The proof for \(N_{2,\epsilon}\) is nearly identical, so we suppress the steps for brevity. \(\Box\) and establish the base cases for our induction. To prove the lemma, we will use strong induction: we assume that for all \(2 \leq i \leq k\), we have that \(P_{i,\epsilon}\) and \(N_{i, \epsilon}\) are congruent to block diagonal matrices of the form in . Therefore, to prove the claim, it suffices to establish the inductive step.
*Proof of :* We prove the claim for \(P_{k+1, \epsilon}\); the steps are in . We entries that stay unchanged throughout the process are labelled with a \"\(\star\)\". The suppressed block with the \"\(\star\)\" entries is a \(((3k-1) \times (3k-1))\) matrix. In Step 1, we simultaneously clear the non-zero, non-diagonal entries of the first row/column. In Step 2, we swap the second and third rows and columns. In Step 3, we clear the non-zero, non-diagonal entries of the second row/column. In Step 4, we clear the non-zero, non-diagonal entries of the third row/column. Finally, we redraw our guidelines to see that the resulting matrix contains \(B\) and \(N_{k,\epsilon}\) as factors. The analogous manipulation on \(N_{k+1,\epsilon}\) yields the result. \(\Box\) We can now finish the proof of : *for all \(q \geq 1\), \(P_{q, \epsilon}\) is congruent to a diagonal matrix with \(q\)-many \(B\) factors, and a single \(1 \times 1\) factor with entry \(\epsilon \pm 2\).* We begin by diagonalizing the first three rows and columns of \(P_{q,\epsilon}\) to produce a block diagonal matrix with factors \(B\) and \(N_{q-1, \epsilon}\). Then, we factor \(N_{q-1, \epsilon}\) into a block diagonal matrix with factors \(B\) and \(P_{q-2, \epsilon}\). Exhausting this procedure eventually yields a factorization with \(q\)-many \(B\) blocks. The remaining block is a \(1 \times 1\) matrix. When \(q\) is even, this entry will be \(\epsilon+2\), and when \(q\) is odd, the entry will be \(\epsilon-2\). ◻
## Signatures and the Seifert genus
The following lemmas study how the signatures of twisted torus knots of a fixed genus are related.
To prove , we need one last lemma which analyzes the case where a twisted torus knot and a torus knot have the same Seifert genus and signature.
We can now prove our last main result. *Let \(\mathcal{S} = \mathcal{T} \cup \mathcal{L}\), where \(\mathcal{T}\) is the set of all positive torus knots, and \(\mathcal{L}\) is the set of L-space knots of braid index three. Every concordance contains at most one knot from \(\mathcal{S}\).* As we will see, the knot invariants \(i(K)\), \(\tau(K)\), and \(\sigma(K)\), when combined, are sufficient to show that every knot in \(\mathcal{S}\) lies in a distinct concordance class.
# Discussion {#Discussion}
In this section, we justify some of the claims we made in the introduction. Namely, we provide evidence for : *for any L-space knot, the braid index and bridge index agree*.
To conclude, we justify why holds for certain cables of L-space knots.
We hope these observations provide some motivation for investigating further.
# Introduction {#section:intro}
By Alexander's theorem, every link in \(S^3\) can be realized as the closure of some braid. Some properties of braids can be ported into important topological properties of their closures. For example, if a braid \(\beta\) is *homogeneous* (i.e. for every \(i\), every \(\sigma_i\) appearing in \(\beta\) has the same sign), then the closure is a fibered link in \(S^3\). In a different vein, the closures of *quasipositive braids* are exactly the links arising as the transverse intersections of smooth algebraic curves \(f^{-1}(0) \subset \mathbb{C}^2\) (here, \(f\) is a non-constant polynomial) with the unit sphere \(S^3 \subset \mathbb{C}^2\). Another property of braids with important topological consequences is *twist positivity*:
A full twist on three strands is seen in ; here, \(\Delta^2 = (\sigma_2 \sigma_1)^3\). Twist positivity is a well studied property: perhaps most notably, Morton and Franks-Williams proved that if \(K\) is twist positive on \(n\) strands, then the braid index of \(K\) is \(n\). In this paper, we show that twist positivity unifies classical results in knot theory with modern concordance invariants. Our first theorem is about the Alexander polynomial of twist positive knots.
The strength and utility of is leveraged by way of the connection between the Alexander polynomial and knot Floer homology: not only does knot Floer homology categorify the Alexander polynomial, but it also provides a suite of powerful knot invariants. We recall: an *L-space knot* in \(S^3\) is a non-trivial knot which admits a positive Dehn surgery to an *L-space*, the closed, connected, oriented 3-manifolds with simple Heegaard Floer homology. For L-space knots, both the Alexander polynomial and knot Floer homology take on a restricted form. This connection leads to our first application of .
Lorenz knots exhibit rich phenomena from the dynamical, geometric, and braid theoretic perspectives. They have been intensely studied since their introduction in 1963; see, for example,, or for a survey. In , we observe L-space Lorenz knots include the well-studied *algebraic knots*, *1-bridge braids*, and some *twisted torus knots*, amongst others. Therefore, generalizes results of both Juhász-Miller-Zemke (who prove for torus knots) and Hom-Lidman-Park, who use input from bordered Floer homology to prove it for algebraic knots. Our results suggests that twist positivity is the important topological property guiding this behavior. Birman-Williams noticed that Lorenz knots are always twist positive on some number of strands, and conjectured that this quantity was actually the braid index. Indeed, this conjecture follows with the subsequent announcement of the Franks-Williams result. To the best of our knowledge, the bridge index of Lorenz knots has not been investigated. proves that often, the bridge index and the braid index for a Lorenz knot agree, hence the Lorenz presentation cannot be minimized further to decrease the bridge index. and suggest that twist positivity is not only an important 3-dimensional property, but that it also provides some insights into 3.5-and 4-dimensional phenomena. Indeed, we next study concordance. Baker observed a novel relationship between concordance properties of fibered knots and Fox's *slice-ribbon conjecture*, which asks whether every slice knot admits a ribbon presentation in \(S^3\). In particular, Baker conjectured that *if two fibered, strongly quasipositive knots are concordant, then they are isotopic*. Baker observes that either his conjecture is true, or the Slice-Ribbon conjecture is false[^2]. Therefore, studying the concordance properties of fibered, strongly quasipositive knots is of central interest. L-space knots form a prominent class of fibered, strongly quasipositive knots. As noted by Abe-Tagami, if Baker's conjecture is true, then every concordance class should have at most one L-space knot.
is proved by using the braid index, \(\tau(K)\), and the Seifert genus in concert. Baader-Dehorney-Liechti proved that only finitely many positive braid knots can be concordant to each other (c.f. ). However, to the best of our knowledge, this is the first explicit infinite family of positive braid knots that we know are distinct in concordance, where as \(g\) tends to infinity, the number of *hyperbolic* knots of genus \(g\) gets arbitrarily large. While L-space knots are well-studied, a classification remains elusive; nevertheless, collecting their properties remains a central goal. L-space knots are "simple" from many perspectives: they have simplest possible knot Floer homology, "simple" knot exteriors (their exteriors are realized as mapping tori), and, heuristically, they seem to have "simple" monodromy. The bridge index and the braid index provide different measures of complexity of a knot, as the former is a lower bound on the latter. The difference between these quantities is itself a measure of complexity. We conjecture that this defect is trivial for L-space knots.
This conjecture is true for the most well known families of L-space knots:
- The only L-space two-bridge knots are \(T(2,n)\) torus knots..
- The only L-space Montesinos knots are the \(P(-2,3,q)\), \(q \geq 1\) pretzel knots. They are all twist positive on either two or three strands, hence applies.
- As proved in this article, the braid and bridge index agree for *1-bridge braids*.
- Baker-Kegel constructed an infinite family of hyperbolic L-space knots where the first knot is provably not braid positive, but the braid positivity status of the remaining knots remains unknown. We prove for the Baker-Kegel family in .
- Combining work of Hedden and Hom, it is known that \(K_{p,q}\) is an L-space knot if and only if \(K\) is and \(q \geq p(2g(K)-1)\). In , we show that if is true for an L-space knot \(K\), it is true for all of its L-space cables
The Alexander polynomials of L-space knots have been of interest ever since Ozsváth-Szabó proved that the coefficients of these polynomials take values in \(\{-1,0,1\}\). Hedden-Watson later proved that when \(K\) is an L-space knot, \(\Delta_K(t) = 1-t+\ldots+ t^{2g-1} + t^{2g}\), where \(g\) is the Seifert genus of \(K\). The results we summarized above provide evidence for the following:
Hyperbolicity is a key assumption in : the conjecture is false even for the (2,3)-cable of \(T(2,3)\), which is an L-space knot. In a different direction, we also predict:
is true for all positive braid knots in the knot tables, but quickly becomes false for even slightly broader classes of knots. For example, the fibered, strongly quasipositive knot \(10_{145}\) is not braid positive, and it has braid index is four while the bridge index is three. It is worth noting that there are positive braid knots which do not admit a positive braid representative on a minimal number of strands, indicating a subtlety to .
## Conventions
- We let \(B_n\) denote the braid group on \(n\) strands.
- Unless stated otherwise, we assume that all braid closures are knots.
- When drawn vertically (resp. horizontally), our braids are oriented from north to south (resp. from west to east).
- We write \(\beta_1 \approx \beta_2\) when braids are related by conjugation. We write \(\beta_1 = \beta_2\) if they are related by braid relations. We underline the subword of a braid that is being modified.
- Throughout, we denote the braid (resp. bridge) index of a knot \(K\) by \(i(K)\) (resp. \(br(K)\)).
- We always work within the smooth category. In particular, we only study knots up to smooth concordance.
## Organization
We prove , , and in , , and , respectively. provides evidence for .
# Twist Positivity and the Alexander Polynomial {#AlexPolyResult}
This section establishes a relationship between twist positivity and the Alexander polynomial. **.** *If \(K\) is twist positive, then \(\Delta_K(t) = 1-t + t^n +t^{n+1}R(t)\), where \(R(t) \in \mathbb Z[t]\).* The proof of requires the *reduced Burau matrix*, \(B(t)\), which is associated to a braid \(\beta \in B_n\). We quickly recall the relevant details here, and recommend for a full account. Let \(\beta\) be a braid on \(n\) strands. We write \(\displaystyle \beta =\prod_{r=1}^m \sigma_{i_r}^{\epsilon_{i_r}}\), i.e. as the product of Artin generators To \(\beta\), we can associate its reduced Burau matrix, \(B(t)\), which is the product of \((n-1)\times (n-1)\) matrices \(\sigma_i(t)\), where \(\sigma_i(t)\) differs from the \((n-1)\times (n-1)\) identity matrix \(I_{n-1}\) only in row \(i\), as in . (Note that this matrix is truncated appropriately when \(i=1,n-1\).) In particular, \(\displaystyle B(t)=\prod_{r=1}^m \sigma_{i_r}^{\epsilon_{i_r}}(t)\). Suppose \(\widehat{\beta}=K\), a knot. The Alexander polynomial for \(K\), denoted \(\Delta_K(t)\), is derived from \(B(t)\) as follows: \[\frac{\det(I_{n-1}-B(t))}{1-t^n}=\frac{\Delta_K(t)}{1-t}\] In particular, \[\Delta_K(t)=(1-t)\cdot(\det(I_{n-1}-B(t)))\cdot(1+t^n+t^{2n}+\cdots).\] We collect some important properties of \(B(t)\):
- In general, the entries in a reduced Burau matrix lie in \(\mathbb Z[t^{\pm 1}]\). However, when \(\beta\) is a positive braid, the entries of \(B(t)\) are genuine polynomial entries.
- Recall: the full twist on \(n\) strands is denoted by \(\Delta^2\), where \(\Delta^2 = (\sigma_1 \ldots \sigma_{n-1})^n \approx (\sigma_{n-1} \ldots \sigma_1)^n\). When \(\beta = \Delta^2\), we have \(B(t) = t^n I_{n-1}\). In particular, if \(\beta\) is a twist positive braid, then \(\beta = \Delta^2 \gamma\), so the reduced Burau matrix associated to \(\beta\) is \(B(t) = (t^n I_{n-1})(C(t)) = t^nC(t)\), where \(C(t)\) is the reduced Burau matrix associated with \(\gamma\).
- Suppose \(\beta\) is a positive braid on \(n\) strands. We define the polynomial \(q_{B(t)}(x)\) as follows: \[\begin{aligned}
q_{B(t)}(x):=\det(I_{n-1}-xB(t)) = 1+a_1(t)x+a_2(t)x^2 + \ldots + x^{n-1}a_{n-1}(t) \label{eqn:Burau} \end{aligned}\] We claim that the values of \(a_1(t)\) and \(a_{n-1}(t)\) can be quickly determined from \(B_(t)\): if we let \(p_{B(t)}(x):=\det(xI_{n-1}-B(t))\) denote the characteristic polynomial of \(B(t)\), then \(q_{B(t)}(x)=x^{n-1}p_{B(t)}(x^{-1})\). That is, \(q_{B(t)}(x)\) is obtained by \(p_{B(t)}(x)\) by exchanging the coefficients of \(x^i\) and \(x^{(n-1)-i}\) for each \(i \leq \lfloor (n-1)/2 \rfloor\). In particular, in \(q_{B(t)}(t)\), \(a_{n-1}(t) = (-1)^{n-1}\det(B(t))\) and \(a_1(t)=-\text{tr}(B(t))\).
We observe that substituting \(x=t^n\) into yields: \[\begin{aligned}
\det(I_{n-1}-t^nB(t)) &= 1+a_1(t)t^n+a_2(t)t^{2n} + \ldots + t^{n^2-n}a_{n-1}(t) \nonumber \\ &= 1+a_1(t)t^n+t^{n+1} R(t) \label{eqn:PolyFullTwist} \end{aligned}\] where \(R(t) \in \mathbb Z[t]\) collects all the higher order terms. With these preliminaries established, we can prove the result.
With this claim in hand, we can finish proving . In particular, since \(a_1(0)=0\), then \(a_1(t) = e_1t + e_2t^2 + \ldots + e_kt^k\). Plugging this into , we have \[\begin{aligned}
\Delta_K(t) &= 1-t+(1+a_1(t))t^n + t^{n+1}S(t) \\ &= 1-t+(1 + (e_1t + e_2t + \ldots + e_kt^k))t^n + t^{n+1}S(t) \\ &= 1-t+t^n + t^{n+1}V(t) \end{aligned}\] for some \(V(t) \in \mathbb Z[t]\). This is exactly what we wanted to show. ◻
# Proof of {#section:Lorenz}
First, we establish that some well-studied families of knots are twist positive:
## Twist positive L-space knots
We begin by proving . **.** *If \(K\) is a twist positive L-space knot, then its braid index and bridge index agree. Moreover, \(K\) realizes the minimal braid and bridge indices of its concordance class.*
Let \(\alpha_0, \ldots, \alpha_{2m}\) be the non-zero degrees appearing in \(\Delta_K(t)\), written in decreasing order. Juhász-Miller-Zemke showed that for L-space knots, \(Ord_v(K) = \max \{\alpha_{i-1}-\alpha_i | 1 \leq i \leq 2m \}\) . Since \(n-1\) appears as a gap in the degrees of the exponents of \(\Delta_K(t)\), we have \(n-1 \leq Ord_v(K)\). The bridge index, \(br(K)\), is a lower bound on the braid index, \(i(K)\), so: \[\begin{aligned}
n-1 &\leq Ord_v(K) \leq br(K)-1 \leq i(K)-1 = n-1 \label{eqn:Inequalities} \end{aligned}\] But this means we must have equalities throughout, hence \(br(K) = n = i(K)\), as desired. Next, we prove that twist positive L-space knots realize the minimal braid index of their concordance classes. The proof of this is nearly identical to the proof of. Suppose the knot \(J\) is concordant to a twist positive L-space knot \(K\), and let \(N(K)\) denote the concordance invariant defined by Dai-Hom-Truong-Stoffregen. We have the following string of inequalities (justification for the respective (in)equalities are in the paragraph following): \[\begin{aligned}
i(J) \geq br(J) &\geq Ord_v(J)+1 \\ & \geq N(J) + 1 \\ &= N(K) + 1 \\ &= Ord_v(K) + 1 \\ &= br(K) = i(K) \end{aligned}\] The first equality of (3.2) is standard; the second is a restatement of . Dai-Hom-Truong-Stoffregen showed that \(N(K)\) is a lower bound on \(Ord_v(K)\), yielding the inequality in (3.3). Since \(N(K)\) is a concordance invariant, we have the equality in (3.4). showed that for L-space knots, the invariants \(N(\star)\) and \(Ord_v(\star)\) agree; this yields the equality in (3.5). The equalities of (3.5) follow from those in ([\[eqn:Inequalities\]](#eqn:Inequalities){reference-type="ref" reference="eqn:Inequalities"}). ◻
We briefly justify why we may assume that \(k\) is at least 4: if \(k=1, 2\), then \(T(3,k; 2m)\) is isotopic to a \(T(2,r)\) torus knot. Additionally, for all \(m \geq 0\) and \(k \equiv 0 \mod 3\), \(T(3,k;2m)\) is a link. Thus, we may assume \(k \geq 4\). We note that when \(k \geq 4\), \(T(3,k; 2m)\) is twist positive on three strands.
In the next lemma, we see that it suffices to study \(T(3,k;2m)\) where \(k \equiv 1 \mod 3\).
In fact, our proof of will show that the \(\lfloor \frac{q}{3}\rfloor-1\) twisted torus knots are pairwise distinct from \(T(3,q)\) and from each other.
Our long-term goal is to show that, despite there being many twisted torus knots with the same \(\tau\) invariants, these knots represent distinct concordance classes. This is proved by computing the signatures of 3-stranded twisted torus knots.
## Computing the signature via the Goeritz matrix
To compute the signature, we use the Gordon-Litherland method, which now we recall: let \(\mathcal{K}\) be a knot in \(S^3\), and fix a regular projection \(K\) of \(\mathcal{K}\). First, we checkerboard color \(K\), where the checkerboard has black and white regions (without loss of generality, we assume that the unbounded region is white). Each white region of the diagram is labelled \(X_i\), and we set \(X_0\) to be the unbounded white region. To each double point \(D\) in \(K\), we assign a value \(\eta(D) = \pm 1\), as in . Each double point \(D\) is also assigned a *type*, i.e. it is either *Type I* or *Type II*; see . This information is organized into a symmetric, integral matrix \(G'(K)\), with entries \(g_{ij}\), where: \[\begin{aligned}
g_{ij} = \begin{cases}-\displaystyle \sum \eta(D) & \text{summed over double points \(D\) incident to \(X_i\) and \(X_j\), where \(i \neq j\)} \\ \\ \displaystyle-\sum_{k, k \neq i} g_{ik} & \text{if \(i = j\)} \end{cases} \end{aligned}\] The *Goeritz matrix* \(G(K)\) is constructed from \(G'(K)\) by deleting the \(0^{th}\) row and column of \(G'(K)\). Thus, \(G\) is a symmetric, integral, \(n \times n\) matrix, where \(n\) is the number of bounded white regions in \(K\). Gordon-Litherland proved that \[\begin{aligned}
\label{eqn:GordonLitherland} \sigma(\mathcal{K}) = \text{sign}(G(K))-\mu(K) \end{aligned}\] where \(\mu(K) = \sum \eta(D)\), where \(D\) is a Type II double point in \(K\). We are now ready to compute the signature of \(T(3,q;2m)\) where \(q \geq 4\) and \(m \geq 0\).
The matrices \(G'(K)\) and \(G(K)\) for the twisted torus knot \(T(3,7;6)\) from are below. \[\hspace{-0.5cm} G'(K) = \begin{blockarray}{ccccccccc} & \scriptscriptstyle{\textcolor{gray}{0}} & \scriptscriptstyle{\textcolor{gray}{1}} & \scriptscriptstyle{\textcolor{gray}{2}} & \scriptscriptstyle{\textcolor{gray}{3}} & \scriptscriptstyle{\textcolor{gray}{4}} & \scriptscriptstyle{\textcolor{gray}{5}} & \scriptscriptstyle{\textcolor{gray}{6}} & \scriptscriptstyle{\textcolor{gray}{7}} \\ \begin{block}{c(cccccccc)} \scriptscriptstyle{\textcolor{gray}{0}}&11 &-1 &-1 &-1 &-1 &-1 &-1 &-5 \\ \scriptscriptstyle{\textcolor{gray}{1}}&-1 &-1 & +1 & 0 & 0 & 0 & 0 & +1 \\ \scriptscriptstyle{\textcolor{gray}{2}}&-1 & +1 &-1 & +1 & 0 & 0 & 0 & 0 \\ \scriptscriptstyle{\textcolor{gray}{3}}&-1 & 0 & +1 &-1 & +1 & 0 & 0 & 0 \\ \scriptscriptstyle{\textcolor{gray}{4}}&-1 & 0 & 0 & +1 &-1 & +1 & 0 & 0\\ \scriptscriptstyle{\textcolor{gray}{5}}&-1 & 0 & 0 & 0 & +1 &-1 & +1 & 0 \\ \scriptscriptstyle{\textcolor{gray}{6}}&-1 & 0 & 0 & 0 & 0 & +1 &-1 & +1 \\ \scriptscriptstyle{\textcolor{gray}{7}}&-5 & +1 & 0 & 0 & 0 & 0 & +1 & +3 \\ \end{block} \end{blockarray} \qquad G(K) = \begin{pmatrix} \begin{array}{ccccccc} -1 & +1 & 0 & 0 & 0 & 0 & +1 \\ +1 &-1 & +1 & 0 & 0 & 0 & 0 \\ 0 & +1 &-1 & +1 & 0 & 0 & 0 \\ 0 & 0 & +1 &-1 & +1 & 0 & 0\\ 0 & 0 & 0 & +1 &-1 & +1 & 0 \\ 0 & 0 & 0 & 0 & +1 &-1 & +1 \\ +1 & 0 & 0 & 0 & 0 & +1 & +3 \\ \end{array} \end{pmatrix}\] Our goal is to compute the signature of \(G(K)\). We will need a powerful theorem, typically refereed to as "Sylvester's Law of Inertia": *two symmetric, \(n \times n\) matrices have the same number of positive, negative, and zero eigenvalues if and only if they are congruent*. (Recall that two matrices \(A\) and \(B\) are congruent if there exists a non-singular matrix \(S\) such that \(B = SAS^T\).) We will prove that \(G(K)\) is congruent to a diagonal matrix \(D\). Since, for diagonal matrices, the eigenvalues are exactly the diagonal entries, we can compute the signature of \(G(K)\) from \(D\).
To prove this result, we will first establish some preliminaries.
We note: the first parameter in the subscript identifies the dimension of the matrix, while the second index records the value \(\epsilon\). (right) displays \(N_{1,\epsilon}\) and \(P_{1, \epsilon}\). We note that the entries of \(P_{\ell, \epsilon}\) and \(N_{\ell, \epsilon}\) agree in all but two entries: \(n_{1, 3\ell+1} = n_{3\ell+1,1}=-1\) while \(p_{1, 3\ell+1} = p_{3\ell+1,1}=+1\) (indeed, \(N\) and \(P\) stand for "negative" and "positive", respectively).
*Proof of :* We begin by showing \(N_{1,\epsilon}\) can be diagonalized into the claimed form. The sequence of simultaneous row and column operations is written out explicitly in . In Step 1 and 2, we simultaneously clear the non-zero, non-diagonal entries in the first row/column. In Step 3, we simultaneously swap the second and third rows and columns. In Steps 4 and 5, we clear the non-zero, non-diagonal entries of the second row/column. The analogous sequence of operations used to diagonalize \(P_(1,\epsilon)\) are in . \(\Box\) This proves . Next, we establish the base cases required to perform strong induction.
*Proof of :* We first prove the claim for \(P_{2,\epsilon}\). We perform a sequence of elementary matrix operations in . In Step 1, we simultaneously clear the non-zero, non-diagonal entries of the first row and column. In Step 2, we swap the second and third rows and columns. In Step 3, we clear the non-zero, non-diagonal entries of the second row and column. In Step 4, we clear the non-zero, non-diagonal entries of the third row and column. The lower right block of the final matrix is exactly \(N_1\). The proof for \(N_{2,\epsilon}\) is nearly identical, so we suppress the steps for brevity. \(\Box\) and establish the base cases for our induction. To prove the lemma, we will use strong induction: we assume that for all \(2 \leq i \leq k\), we have that \(P_{i,\epsilon}\) and \(N_{i, \epsilon}\) are congruent to block diagonal matrices of the form in . Therefore, to prove the claim, it suffices to establish the inductive step.
*Proof of :* We prove the claim for \(P_{k+1, \epsilon}\); the steps are in . We entries that stay unchanged throughout the process are labelled with a \"\(\star\)\". The suppressed block with the \"\(\star\)\" entries is a \(((3k-1) \times (3k-1))\) matrix. In Step 1, we simultaneously clear the non-zero, non-diagonal entries of the first row/column. In Step 2, we swap the second and third rows and columns. In Step 3, we clear the non-zero, non-diagonal entries of the second row/column. In Step 4, we clear the non-zero, non-diagonal entries of the third row/column. Finally, we redraw our guidelines to see that the resulting matrix contains \(B\) and \(N_{k,\epsilon}\) as factors. The analogous manipulation on \(N_{k+1,\epsilon}\) yields the result. \(\Box\) We can now finish the proof of : *for all \(q \geq 1\), \(P_{q, \epsilon}\) is congruent to a diagonal matrix with \(q\)-many \(B\) factors, and a single \(1 \times 1\) factor with entry \(\epsilon \pm 2\).* We begin by diagonalizing the first three rows and columns of \(P_{q,\epsilon}\) to produce a block diagonal matrix with factors \(B\) and \(N_{q-1, \epsilon}\). Then, we factor \(N_{q-1, \epsilon}\) into a block diagonal matrix with factors \(B\) and \(P_{q-2, \epsilon}\). Exhausting this procedure eventually yields a factorization with \(q\)-many \(B\) blocks. The remaining block is a \(1 \times 1\) matrix. When \(q\) is even, this entry will be \(\epsilon+2\), and when \(q\) is odd, the entry will be \(\epsilon-2\). ◻
## Signatures and the Seifert genus
The following lemmas study how the signatures of twisted torus knots of a fixed genus are related.
To prove , we need one last lemma which analyzes the case where a twisted torus knot and a torus knot have the same Seifert genus and signature.
We can now prove our last main result. *Let \(\mathcal{S} = \mathcal{T} \cup \mathcal{L}\), where \(\mathcal{T}\) is the set of all positive torus knots, and \(\mathcal{L}\) is the set of L-space knots of braid index three. Every concordance contains at most one knot from \(\mathcal{S}\).* As we will see, the knot invariants \(i(K)\), \(\tau(K)\), and \(\sigma(K)\), when combined, are sufficient to show that every knot in \(\mathcal{S}\) lies in a distinct concordance class.
# Discussion {#Discussion}
In this section, we justify some of the claims we made in the introduction. Namely, we provide evidence for : *for any L-space knot, the braid index and bridge index agree*.
To conclude, we justify why holds for certain cables of L-space knots.
We hope these observations provide some motivation for investigating further.
[^1]: Baker phrases his conjecture in terms of *tight fibered knots*, but by, this is equivalent to studying fibered, strongly quasipositive knots.
[^2]: Baker phrases his conjecture in terms of *tight fibered knots*, but by, this is equivalent to studying fibered, strongly quasipositive knots. | {'timestamp': '2022-12-01T02:18:10', 'yymm': '2211', 'arxiv_id': '2211.17109', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17109'} |
null | null |
null | null |
# Introduction
Most of our non-perturbative understanding of quantum gravity comes from dualities and black holes. What makes these two phenomena special is that they allows us to perform a perturbative calculation for strongly coupled objects of the theory. In this note we will focus specifically on black holes. In the case of black holes, Hawking's field theory calculation gives the density of states at deep UV where the states have strong gravitational screening (i.e. they are black holes). In the absence of a complete theory of quantum gravity, it is challenging to sharply define what \"quantum gravity\" means. However, we can expect a quantum theory gravity in Minkowski spacetime to provide a consistent description of particles[^1] and their high-energy scattering amplitudes, especially those of graviton. Any such description would be naturally limited since at some energy scale \(\Lambda_{BH}\), the massive would-be-particles are replaced by black holes. String theory incorporates this transition in a very elegant way. But it is natural to ask, what do we expect from a \"quantum gravity\" beyond the energy scale \(\Lambda_{BH}\) when black holes become relevant? The good news about black hole is that many of their properties follow from IR calculations and are universal. We can use those properties as a litmus test for any theory of quantum gravity. This note focuses on how string theory passes black hole related litmus tests. Before going further, let us clarify the difference between a particle and a black hole. For a particle, the effective field theory breaks before its would-be horizon. For example, take the electron. Due to the screening effect (or the running of the charge), the classical Maxwell equations are only reliable for length scales greater than its Compton wavelength, which is much greater than electron's Schwarzschild radius. Therefore, we call the electron a *particle* rather than a black hole. Similarly, for strings, the effective field theory can only be trusted for length scales greater than the string length \(l_s\), which is much greater than the Schwarzschild radius for masses \(m\ll (l_sg_s^2)^{-1}\), where \(g_s\) is the string coupling. In this sense, we can think of black holes as would-be particles that can be described by effective field theory all the way up to their horizons, and perhaps beyond that. This suggests that in a theory of quantum gravity, we need to adopt a frame-work that includes both black holes and particles, where at some mass scale the two transition to each other. In string theory specifically, the light particles are described by excitations of a string. Therefore, we expect a transition between string states and black holes. We can test this continuity in multiple ways. One is to look for thermal properties of black holes in strings, and another is to look for scattering properties of strings in black holes. The first approach is commonly referred to as string/black hole transition, however, the second approach is equally important given that scattering amplitudes is a big chunk of the data that constitutes what we call string theory. In this note we try to review some properties of strings and black holes both from thermodynamic as well as scattering theory. One of the main goals of this note is to present some of the existing literature in a coherent way that connects it to the string/black hole transition. The organization of this paper is as follows. In section [2](#st){reference-type="ref" reference="st"} we review the common lore and some recent developments about string/black hole transition from the matching of their thermal properties. In section [3](#ss){reference-type="ref" reference="ss"} we study the string/black hole transition in light of scattering amplitudes. We continue that study in more detail in section [4](#ss2){reference-type="ref" reference="ss2"} and discuss some of the nuanced differences between fixed-angle scattering and fixed impact parameter scattering that often leads to contradictory statements in the existing literature. We argue that the impact parameter is a more suitable approach to understand the string/black hole transition and opens new doors for further studies. In section [5](#CS){reference-type="ref" reference="CS"}, we discuss some of the challenges for interpreting the string/black hole transition as a correspondence between back hole states and string states.
# String/black hole transition: entropy {#st}
Let us start with reviewing the string/black hole transition. The transition refers to the matching of multiple energy scales which suggests a phase transition between a weakly interacting strings and black holes .
- The energy scale at which the Schwarzschild radius of the string is the string length. At this energy scale, the Einstein field equations are reliable at the Schwarzschild radius. Therefore, the string excitations, beyond this energy scale, if they exist, must be described by black holes. The Schwarzschild radius of a string excitation with mass \(M\) is given by \[\begin{aligned}
r_S\sim (m l_s^{d-2}g_s^2)^{\frac{1}{d-3}}. \end{aligned}\] The crossover energy scale is where \(r_S\sim l_s\) which corresponds to the mass \[\begin{aligned}
M_c\sim M_s/g_s^2, \end{aligned}\] where \(M_s=l_s^{-1}\) is the string mass.
- Energy at which the number of free string microstates of a given energy matches those of a black hole of the same energy. It was shown in the early years of string theory that the leading term in the entropy scales linearly with energy (see and references therein). Here, we give a simple combinatorial argument for unfamiliar readers. The mass of a string excitation at high energies is proportional to \(\sqrt{N}\) where \(N=\sum_i i\cdot n_i\) is the level and \(n_i\) is overall excitation number of all the \(i\)-th modes. Let us first find a lower bound for the number of microstates. Consider a string at level \(N\), where the only exited modes are those with \(i<\sqrt{N}\). Given that any single excitation of such modes can at most contribute \(\sqrt{N}\) to the level, we need the sum of the excitation numbers \(n_i\) to be at least \(\sqrt{N}\). Consider an arbitrary set of \(n_i\geq0\) with \(1<i<\sqrt{N}\) that add up to \(\simeq \sqrt{N}\), and choose \(n_1\) to account for the rest of the level. \[\begin{aligned}
n_1\simeq N-\sum_{i>1}i\cdot n_i. \end{aligned}\] These states are only a subset of total string excitations, but they are very easy to count. Their number is equal to the number of different ways that one can partition \(\simeq \sqrt{N}\) into \(\simeq \sqrt{N}\) numbers which is given by \(\sim {2\sqrt{N}\choose \sqrt{N}}>2^{2\sqrt{N}}/(2\sqrt{N})\). Therefore, the number of microstates are expected to increase exponentially with \(M\sim M_s\sqrt{N}\). Upon doing a more careful calculation, one can see that the states that we did not consider (\(i>\sqrt{N}\)), do not alter the exponential behavior. This is because, although considering the modes \(i>\sqrt{N}\) will increase the number of possibilities. Each such mode will have a very large contribution to the level and thus bring down the total number of remaining excitations. The leading contribution to the number of string states with mass \(M\) goes like \(\text{N}(M)\propto (M/M_s)^{-\gamma}\exp(\delta\cdot(M/M_s))\) for some positive constants \(\gamma\) and \(\delta\). So far we calculated the number of single string microstate. A more accurate comparison would be to compute the number of microstate with energy \(M\) in a free string gas. The number of microstates with \(k\) strings each of energy \(\sim E/k\) is \(\simeq \text{N}(E/k)^k\sim (\frac{E}{M_s})^{-k\gamma}\exp(-\delta\cdot(E/M_s))\). Due to the polynomial prefactor, the number of single string states (\(k=1\)) dominates the microstates and the leading term in the entropy of the string gas is, \[\begin{aligned}
S_{string}(M)\simeq \beta_H M, \end{aligned}\] where \(\beta_H\sim 1/M_s\) is a constant. It is easy to see that the free string partition function \(\int dM S(M)e^{-\beta M}\) diverges at \(\beta=\beta_H\) which is called the Hagedorn temperature. This divergence of the partition function can be used for a more precise evaluation of \(\beta_H\) in different string theories. Now let us calculate the entropy of a d-dimensional black hole in string units. The black hole entropy is proportional to its area in Planck units which goes like \[\begin{aligned}
r_S^{D-2}\sim M^\frac{d-2}{d-3}, \end{aligned}\] in Planck units. After going to the string units via \(l_P^{d-2}=l_S^{d-2}g_s^2\), we find \[\begin{aligned}
S_\text{black hole}\sim (g_s^2l_s^{d-2}M^{d-2})^{\frac{1}{d-3}}. \end{aligned}\] By equating the two entropies we find that the crossover energy scale is \[\begin{aligned}
M_c\sim M_s/g_s^2, \end{aligned}\] regardless of the spacetime dimension.
The matching of all these energy scales paints a consistent picture. It suggests that in the corners of moduli space where string theories are reliable, there is transition between strings and black hole. The transition happens at \(M\sim M_s/g_s^2\). The particles with masses below this mass are interacting weakly enough to be described perturbatively by string theory. More importantly, their Schwarzschild radius is smaller than the string length and therefore it cannot be probed using scattering. On the other hand, \"*particles*\" with masses higher than \(M_s/g_s^2\) are no longer weakly interacting. They have strong gravitational screening to the point that they collapse into black holes. Moreover, the black hole geometry near the horizon is reliable for them given their large enough radii. In the above discussion we discussed the continuous transition in the entropy as we go from free string gas to black holes. However, there are other macroscopic quantities that we can compare between the two and one of them is the size of the system. The size of a free string with mass \(M\) is \(l\sim l_s\sqrt{M/M_s}\). It is easy to see that this size does not match that of the black hole at the transition point. However, this is not an issue because, as we increase the energy of a string, we expect the free string approximation to break down. The self-gravitational interaction of the string should gradually set in to the point where after the transition, the correct description of the thermal system is in terms of black holes. As we approach the Hagedorn temperature \(T_H\), the string gas will start backreacting on the background and curve it towards becoming a black hole. We can study this curved background by looking at string partition function. One way to understand a classical background of a black hole is that if we Wick rotate the background into Euclidean signature, it would determine the free energy of the corresponding thermal system. One can calculate the free energy of black hole as a thermal system using this trick. Going back to strings, to find the background sourced by a self-gravitating string, we can calculate its thermal partition function and see if corresponds to any Euclidean \(d\) dimensional background. It was shown in that the thermal partition of strings can be calculated by evaluating its vacuum amplitude on \(\mathbb{R}^{d-1}\times S^1\) where the radius of \(S^1\) at spatial infinity goes to \(1/(2\pi T)\) where \(T\) is the temperature of the black hole. We can estimate this partition function by looking at its saddle point. We can think of this saddle point as the configuration that solves the equations of motion when we compactify Euclidean \(d\)-dimensional string theory on a circle with anti-periodic boundary condition for fermions. String theory on a small circle is well-approximated by a \((d-1)\) dimensional field theory. Thus, in order to find the saddle point, we can extremize the \((d-1)\) dimensional effective action in the non-compact dimensions. Let us review the light fields that would admit a field theory description. The \((d-1)\) dimensional field theory would contain the light fields resulted from the usual dimensional reduction of higher dimensional fields. However, near the Hagedorn temperature, there will be a new light field. This is the string winding mode. At temperatures above the Hagedorn temperature, this mode becomes tachyonic which signals a breakdown in the stringy treatment and a phase transition. The resulting background was found by Horowitz and Polchinski for temperatures near the Hagedorn temperature in dimensions \(d<7\). To connect this background to the black hole background, one can think of the \(d-1\) non-compact dimensions as the analog of the \(d-1\) non-compact dimensions of Euclidean black holes. However, here we have an additional light \((d-1)\) dimensional field which is the winding mode. Unlike the black hole which is localized at a point in the Euclidean signature, this solution is spread out and it is reasonable to expect that the Lorentzian continuation of this solution would have a size of the same order. It turns out in rotating the Euclidean background to a Lorentzian background, the field values exit the perturbative regime of string theory and is therefore, not trustable. But as far as thermodynamic properties are concerned, we can use the Euclidean solution to do calculations. In, authors did precise calculations for the size and entropy of the Horowitz-Polchinski solution and demonstrated that it provides a continuous bridge between free strings and black holes. However, they found a surprising feature for the solution. In type II theories, it seems that there is no smooth transition between this background and the black hole background, if they are viewed as the target space of the string worldsheet CFT. However, this does not invalidate the string/black hole transition as it could be that along the transition, the purterbative description of the worldsheet breaks down.
# String/black hole transition: scattering amplitudes {#ss}
Another way that we can study the string/black hole transition is through scattering amplitudes. One of the main features of black holes is their universal thermal properties. Since quantum gravity in Minkowski space is usually formulated in terms of an S-matrix, it is natural to expect that the thermal properties of black holes have some meaning in terms of the scattering amplitudes. We review an argument that postulates that at small impact parameters, high energy scattering events go through black hole formation/evaporation. In that case, since black hole evaporate thermally, we expect the amplitude for a particular out-going state to be \(\sim \exp(-cS_{BH})\). But this is only at small enough impact parameters and large energies that allow black hole formations. Fortunately, we can calculate high energy scattering amplitudes below that energy threshold with string theory. In that case, scattering processes go though string states rather than black hole states. If string black hole transition is correct, we expect a smooth transition between the amplitude. In fact, this transition has been demonstrated which is what we will review in this section.
## Scattering through black holes
Consider a 2-particle in-going state with a center of mass energy \(E\) and impact parameter \(b\). Suppose \(E\) is much larger than the mass of the smallest black hole, and \(b\) is much smaller than the Schwarzschild radius of a black hole with mass \(E\). In that case, when the particles distance from each other becomes much smaller than the Schwarzschild radius, we can still trust the effective field theory description which tells us that the geometry collapses into a black hole. Now the outcome of this scattering event is determined by the evaporation of the black hole. Due to the thermodynamic nature of black holes, we expect that the amplitude of a typical 2 particle out-going state to be \(\sim\exp(-S_{BH})\). However, even though the Hawking radiation at finite time interval is thermal, assuming the black hole evaporation respects unitarity, the overall out-going state must be pure and cannot be thermal. In other words, it is only when we look at part of the Hawking radiation and trace over the rest of the quantum system that we expect the radiation to be thermal. Therefore, the overall amplitude is not exactly given by a black body radiation ensemble. However, we can break up the Hawking radiation into smaller intervals (early radiation and late radiation) and approximate the density matrix of each with an almost thermal density matrix with entropy \(\propto S_{BH}\). In that case, a \(2\rightarrow 2\) corresponds to the event where each part of the radiation is a single particle. This allows us to estimate the amplitude of a \(2\rightarrow 2\) event as \[\begin{aligned}
\mathcal{M}(2\rightarrow 2)\sim \exp(-c(E/M_P)^\frac{d-2}{d-3}), \end{aligned}\] where \(c\sim \mathcal{O}(\log E)\) accounts for any polynomial correction. The above argument is explained more precisely in Appendix [\[A1\]](#A1){reference-type="ref" reference="A1"}.
## Gross Mende saddle points
In the previous subsection we saw that the thermodynamic properties of black holes gives a universal behavior for the scattering amplitudes at impact parameters \(b\ll r_S(E)\) where \(r_S\) is the Schwarzschild radius of a black hole with mass \(E\). This analysis is only valid when the center of mass energy is larger than the mass of the smallest black hole. From our discussion of string/black hole transition, we expect the mass of the smallest black hole to be \(\sim M_s/g_s^2\). Such a black hole would have a Schwarzschild radius of \(l_s\) which is the smallest length scale that the perturbative string can detect. Therefore, the black hole analysis, in conjunction with the string/black hole transition suggests that the scattering amplitude of events with impact parameter \(b\sim l_s\) and energies \(E\sim M_s/g_s^2\) is given by \[\begin{aligned}
\label{BHSA} \mathcal{M}\sim \exp(c(\frac{M_s/g_s^2}{M_P})^\frac{d-2}{d-3})\sim \exp(cg_s^{-2}). \end{aligned}\] For smaller energies, we can trust string perturbation and if the transition is correct, we should be able to reproduce the above scattering amplitude from string theory. In the following, we review the work of Gross and Mende which leads to the following equation for string amplitudes. \[\begin{aligned}
\mathcal{M}\sim \exp(c(E/M_s)), \end{aligned}\] where \(c\sim\mathcal{O}(\ln(E))\). As one can see, the above equation reproduces the black hole result of [\[BHSA\]](#BHSA){reference-type="eqref" reference="BHSA"} at energies \(E\sim M_s/g_s^2\). Gross and Mende solved for the saddle points of the string wordlsheet amplitudes. Once these saddle points are found, we can think of them as the classical string processes whose amplitudes receive quantum corrections from other worldsheet configurations. Gross-Mende saddle points are K-sheet covers of the sphere with \(N\) branch points that are connected by branch cuts. The branch-points correspond to the external particles. They showed that such worldsheet saddle points correspond to the solutions to 2d electrostatic on the sphere and the total energy corresponds to the worldsheet action. The multi-sheet covers correspond to fractional charges for which the electric potential will not be single valued. The genus of the corresponding worldsheet can be written in terms of \(K\) and \(N\) as follows. \[\begin{aligned}
G=\frac{1}{2}(K-1)(N-2), \end{aligned}\] In the following we focus on the \(2\rightarrow 2\) scattering for which \(N=4\). Given that higher \(K\) corresponds to smaller fractional electric charges, we expect the worldsheet action to be inversely proportional to \(K\). Therefore, using the result for the tree-level string amplitude, we find that the amplitude of the Gross-Mende saddle point behaves like \[\begin{aligned}
\exp(-S_{K})\sim \exp(-\frac{1}{4K}(s\ln(s)+t\ln(t)+u\ln(u))), \end{aligned}\] in string units. The full string amplitude contains the sum over all of these classical saddles and goes as \[\begin{aligned}
\mathcal{M}\sim \sum_{K} g_s^{2G-2}e^{\frac{1}{4K}(s\ln(s)+t\ln(t)+u\ln(u))+\mathcal{O}(\ln(E))}. \end{aligned}\] Mende and Ooguri showed that this series is Borel resummable. However, without reserving to the full summation, we can easily estimate the order of magnitude of this series by looking at the term with the largest contribution. The largest contribution comes from \(G\sim E=\sqrt{s}\) and the corresponding amplitude is \[\begin{aligned}
\label{GMSR} \mathcal{M}\sim e^{-c(E/M_s)}. \end{aligned}\] The Gross-Mende saddle points correspond to a scattering processes where the intermediate state is \(K\) coincident strings. The impact parameter of the scattering event can be estimated by the size of the intermediate strings which goes as \(E/K\). For the scattering event with the dominant contribution we have \(K\sim E/M_s\) which means the impact parameter is of the order of string length. \[\begin{aligned}
b_{\text{GM}}\sim l_s. \end{aligned}\] Since this is the Schwarzschild radius of the smallest black hole, we expect the Gross-Mende result to match with the black hole result at the transition point. As one can see this is indeed the case. \[\begin{aligned}
\mathcal{M}_{string}(2\rightarrow 2)&\sim \exp(-c(E/M_s)),\nonumber\\ \mathcal{M}_{BH}(2\rightarrow 2)&\sim \exp(-c(E/M_P)^\frac{d-2}{d-3}), \end{aligned}\] and at the transition point \(E\sim M_s/g_s^2\) we have \[\begin{aligned}
\ln(\mathcal{M}_{string})\sim g_s^{-2}\sim \ln ( \mathcal{M}_{BH}). \end{aligned}\] What is perhaps more surprising is that both results go like \(\exp(-S_{ent})\) of the corresponding thermal system. We will come back to this point in the next subsection. In, the above matching between the stringy calculation and the black hole result was generalized to \(2\rightarrow N\) scattering.
## Relationship between scattering amplitudes and entropy {#RAC}
As we saw in the previous subsection, high energy scattering amplitude at small enough impact parameters always seem to go as \(\exp(-cS)\) where \(S\) is the thermodynamic entropy of the system at energy \(E\). For energies below \(M_s/g_s^2\), \(S\) is the entropy of a string gas, while for energies \(E>M_s/g_s^2\), it is the entropy of a black hole with mass \(E\). In the black hole case, there is a clear interpretation for the relationship between the scattering amplitude and the entropy. We can think of black holes as highly complex quantum systems where the information of the incoming state gets scrambled in the black hole Hilbert space and creates a thermal ensemble. Let us call this feature of scattering amplitudes, *thermalization* [^2]. The fact that the scattering amplitudes go as \(\exp(-cS)\) in the stringy regime suggest that some form of thermalization happens in the stringy regime as well. In the following, we will try to understand this stringy thermalization from high energy scattering of strings. If thermalization as described above happens, we expect that the string scattering at low impact parameters (\(b\sim l_s\)) to mostly go through a localized stringy system with \(S_{ent}\) degrees of freedoms. However, \(S_{ent}\) is the entropy of the string gas and it is unclear why a string gas must form. Moreover, the microstates of a free string gas are dominated by the 1-particle string states, however, Gross-Mende analysis suggest that string amplitudes are dominated by saddle points with a large number of intermidiate strings. So how do we resolve this inconsistency and explain the appearance of the entropy in the scattering amplitude? The key feature is that even though the states with higher number of strings have polynomially less microstates, they can have an entropy of the same order. This is beacuse entropy is the logarithm of the number of microstates. So it could be that the entropy appearing in the scattering amplitude is not that of the full string gas, but a subsystem with a large number of strings with a similar entropy. For example, the entropy of a localized \(k\)-string gas with energy \(E\) is \[\begin{aligned}
S_k(E)\simeq \ln[(\frac{E}{M_s})^{-d\cdot k}\exp(\beta_H E)], \end{aligned}\] where the equality is up to subleading corrections. for \(k\lesssim E\), we have \[\begin{aligned}
S_k(E)\simeq S_{string}(E). \end{aligned}\] However, for larger values of \(k\) the entropy quickly drops. Therefore, the string/black hole transition suggests the following.
This notion of thermalization carries over to black hole's thermodynamic properties after the string/black hole transition. The worldsheet saddles are the stringy version of high energy scatterings with \(E>M_s/g_s^2\) and \(b<r_S\) that go through the classical saddle of black hole formation. Just as the scattering amplitudes in the black hole case is inversely proportional to the black hole microstates. The high energy string amplitude is inversely proportional to the microstates with \(k\lesssim E\) strings. Now let us check the above hypothesis for string thermalization using the Gross-Mende saddle points. In the Gross-Mende saddle points, the impact parameter goes as \(E/K\). Therefore, the processes with impact parameter less than \(l_s\) correspond to processes with less than \(E\) strings that have an entropy comparable to the free-string gas. Therfore, the stringy amplitude is expected to exhibit thermal properties even pre-transition. To see this more precisely, the contribution of a Gross-Mende saddle is \[\begin{aligned}
\sim g_s^{2G}\exp(-cE^2/k), \end{aligned}\] where \(2G=(k-1)(N-2)\) is the genus and \(N\) is the number of external particles. Among \(k\lesssim E\) configurations which have impact parameters \(b\sim l_s\), the amplitude is dominated by \(K\sim E\) \[\begin{aligned}
\mathcal{M}(2\rightarrow N)\sim \exp(-cE/M_s) \end{aligned}\] as we would expect from the entropy argument. Therefore, we see that for all the low impact parameter processes, the dominant contribution goes through a classical saddle with an entropy comparable to the thermodynamic entropy. This is to say that the intermediate states are starting to explore more and more of the Hilbert space as the complexity of the time-evolution is increasing. The increase in the complexity was shown to manifest itself in the form of erratic behavior of the string S-matrix for highly excited strings. One can think of our argument above as an evidence for Hawking-like radiation for \(k\) string states where the average energy of the strings is \(E/k \gtrsim>M_s\). The reason we expect this thermal radiation, is that the entropy of this system is proportional to that of the string gas. So another test for the string/black hole transition would be to see if a single energetic string emits a black-body radiation. This was calculated and confirmed in which makes another strong case for the string/black hole transition.
# High energy scattering {#ss2}
## fixed angle vs fixed impact parameter
So far, whenever we have talked about high energy scattering, it has been about scattering events with small impact parameters. In this subsection we discuss fixed angle scattering and how black holes manifest themselves in the fixed angle scattering. Let us consider a generic \(2\rightarrow 2\) scattering scattering amplitude. In addition to the scattering energy \(\sqrt{s}\), we can choose one of impact parameter \(b\) or scattering angle \(\theta\simeq 2\sin^{-1}(\sqrt{-t/s})\) to describe the out-going state, but not both. This is due to the fact that impact parameter is a measure of angular momentum, and it does not commute with the momenta of the out-going particles. In the following, we often use the Mandelstam variable \(t\) instead of the scattering angle \(\theta\simeq 2\sin^{-1}(\sqrt{-t/s})\). The transformation between amplitudes in \((s,t)\) space and \((s,b)\) space is given by the following Fourier transform. \[\begin{aligned}
\label{tb} \frac{1}{s}A(s,t)=(\epsilon_a\cdot\epsilon_d)(\epsilon_b\cdot\epsilon_c)\cdot4\int d^{D-2}b e^{i\bf{q\cdot b}}a(s,b), \end{aligned}\] where \(\bf{q}\) is the momentum exchange \(t=-\bf{q}^2\). In the above formula, \(\epsilon_i\) are the polarization vectors of the external particles and \(a(s,b)\) is the polarization-independent part of the amplitude. As one can see, a fixed angle scattering involves a complex averaging over the fixed impact parameter amplitudes. For large impact parameters, we can think of impact parameter as a measure of angular momentum \(l\simeq Eb/2\). However, for small impact parameters \(b\ll E^{-1}\), the quantization of the angular momentum becomes important and this approximation breaks down. The partial wave expansion of the fixed angle amplitude is the analog of equation [\[tb\]](#tb){reference-type="eqref" reference="tb"} for \(l\). The \(d\)-dimensional partial wave expansion is given by \[\begin{aligned}
\label{gfa} \mathcal{A}(s,t)=2^{2d-3}\pi^\frac{d-3}{2}\Gamma(\frac{d-3}{2})s^\frac{4-d}{2}\sum_{l=0}^\infty (l+\frac{d-3}{2})C_l^\frac{d-3}{2}(\cos(\theta))f_l(s), \end{aligned}\] where \(C_l^\frac{d-3}{2}\) are Gegenbauer polynomials. For energies larger than the black hole threshold \(E\gg M_s/g_s^2\), and impact parameters smaller than the Schwarschild radius \(b\simeq 2l/E\ll r_S(E)\), we expect the cross-section to be exponentially small due to black hole formation and evaporation. \[\begin{aligned}
b\simeq 2l/E\ll r_S(E):~~\ln(a(s,b))\sim \ln(f_l(E))\sim-S_{BH}(E). \end{aligned}\] However, if the amplitude is polynomial for impact parameters greater than the Schwarzschild radius, the overall sum [\[tb\]](#tb){reference-type="eqref" reference="tb"} or [\[gfa\]](#gfa){reference-type="eqref" reference="gfa"} comes out polynomial. This is in fact what happens. For large enough impact parameters one can use the Born approximation to find a polynomial amplitude. Therefore, the overall fixed angle amplitude is expected to be polynomial. Another way to see this polynomial behavior is that due to the exponential suppression of the amplitude \(f_l(E)\) for \(b<r_S(E)\), the partial waves can no longer interfere to create a delta function. Now the overall amplitude looks like that of a scattering off of a black disc with radius \(\sim r_S(E)\). The shadow effect from the edge of the disc gives a polynomial cross-section for any given angle. This effect was calculated more precisely in. Note that there are multiple different regimes between the black hole regime and the Born regime based on the value of \(\ln(b)/\ln(E)\). We will discuss these different regimes in more detail in the next subsection. The above discussion shows that the fixed angle scattering amplitude is much less sensitive to the black hole physics than the fixed impact parameter amplitudes. In the next subsection we will focus on finding the perturbative scattering amplitudes in the E--b plane. Let us take a moment to comment on the Gross-Mende saddles. In principle, since the Gross-Mende solution is a saddle, we can use it regardless of what parameters we use for the scattering. However, as we will see in the next subsection, when the scattering angle is too small, there is another contribution from higher impact parameters that has a more dominant contribution. Therefore, the Gross-Mende result should be only used to estimate the fixed angle scattering at large enough angles. There is a very intuitive explanation for this. As we discussed earlier, the Gross-Mende saddle with dominant contribution has impact parameter \(\lesssim l_s\) which suggests that it is the stringy version of the black hole formation/evaporation. Scattering events that go through black hole formation/evaporation are hard scattering events that emit particles at all angles. However, scatterings at larger impact parameters are expected to be softer and have a more focused amplitudes at small scattering angles. Therefore, just like the black hole example, we should expect the small angle scattering to always be dominated by soft processes that do not go through any saddle (Gross-Mende or black hole). In the previous section, we tested the string/black hole transition by studying the continuity of the amplitude across the \(E\sim M_s/g_s^2,~b\sim l_s\) point. However, when we look at scattering amplitudes in terms of the \(E-b\) plane, we see that the black hole region has another boundary at \(E>M_s/g_s^2,~~b\sim r_S(E)\). One could hope that the scattering amplitudes at larger impact parameters are purterbutively calculable and one can check the string/black hole via continuity of the scattering amplitudes across this boundary as well. In a series of papers, Veneziano et al. developed a method to reformulate string perturbation theory in terms of \(E\) and \(b\) and calculate the string amplitude in the \(b\gg r_S(E)\) regime. We will review these results in the next subsection and discuss their connection with the string/black hole transition.
## ACV approach
First let us summarize the Amati-Ciafaloni-Veneziano (ACV) approach. The idea is to first calculate the string theory amplitudes in the Regge limit \(t\)-fixed and \(s\rightarrow\infty\), and express those results in terms of the impact parameter. Then we try to push the approximation toward the fixed angle scattering regime as much as possible. ACV show that similar to the Eikonal regime in field theory where the ladder diagrams become dominant, in string theory the diagrams factorize as well. In particular, the genus \(h\) amplitude goes like \[\begin{aligned}
A^{(h)}(s,t)\rightarrow \epsilon_a\cdot\epsilon_d\epsilon_b\cdot\epsilon_c4 (\frac{i}{2s})^h &\int\frac{d^{D-2}\bf{q}_1...d^{D-2}\bf{q}_{h+1}}{(h+1)!(2\pi)^{(D-2)h}}\delta^{D-2}(\bf{q}-\sum_i\bf{q}_i)\nonumber\\ &\times \Pi_{j=1}^{j=h+1}a_{tree}(s,t_j)\langle 0|\int\Pi_j\frac{d\sigma_j^ud\sigma_j^d}{(2\pi)^2}:e^{i\bf{q}_j\cdot(\hat X^u(\sigma_j^u)-\hat X^d(\sigma_j^d)}:|0\rangle. \end{aligned}\] Note that the operator \(:e^{i\bf{q}_j\cdot(\hat X^u(\sigma_j^u)-\hat X^d(\sigma_j^d)}:\) captures the effect of stringy excitations that are absent in the field theory calculations. If we sum over \(h\) in the above formula and plug in the output in [\[tb\]](#tb){reference-type="eqref" reference="tb"} we find a genus expansion for \(a(s,b)\), \[\begin{aligned}
a^{(h)}(s,b)=\frac{(2i)^h}{(h+1)!}\langle 0|\hat\delta^{h+1}|0\rangle, \end{aligned}\] where, \[\begin{aligned}
\hat\delta&=\int\frac{d\bf{q}}{(2\pi)^{D-2})}\frac{a_{tree}(s,t)}{s}\int\langle 0|\int\frac{d\sigma_ud\sigma_d}{(2\pi)^2}:e^{i\bf{q}\cdot(\hat X^u(\sigma_u)-\hat X^d(\sigma_d)}:|0\rangle.\nonumber\\ &=\int\frac{d\sigma_ud\sigma_d}{(2\pi)^2}:a_{tree}(s,b+\hat X^u(\sigma_u)-\hat X^d(\sigma_d)):. \end{aligned}\] Therefore, by summing over genus \(h\) we find \[\begin{aligned}
a(s,b)=\sum_{h\geq 0} a^{(h)}(s,b)=\langle 0|\frac{1}{2i}(e^{2i\hat\delta}-1)|0\rangle. \end{aligned}\] We need to compute the vacuum expectation value of \(e^{2i\hat\delta}\). For that we can start with by approximating that expectation value with \(\hat X=0\). This approximation ignores string excitations and therefore must reproduce the result of graviton ladder diagrams in field theory calculation, i.e., the normal gravitational deflection. We can read \(\delta(b,s)\) from the tree-level amplitude (Veneziano amplitude) as follows. \[\begin{aligned}
\delta(b,s)\equiv \hat \delta|_{\hat X=0}=\frac{\pi g^2s}{(4\pi)^{D/2}}\int^{\infty+i\epsilon}_{-\infty-i\epsilon}\frac{dy}{y^{D/2-1}}\exp(-\frac{b^2}{4y})\times F(s/2,s/2;1;\exp(-y)). \end{aligned}\] The above integral has very different behaviors depending on \(b/\sqrt{Y}\) where \(Y=\ln(s)\)[^3]. For \(b\gg \sqrt{Y}\) we find, \[\begin{aligned}
\text{Re} \delta&\simeq \frac{g^2s}{8\pi}\frac{1}{\Omega_{D-4}}b^{D-4}\nonumber\\ \text{Im} \delta&\simeq \frac{\pi g^2 s}{8(4\pi Y)^{D/2-2}}\exp(\frac{b^2}{4Y}). \end{aligned}\] From this point on we refer to the above expressions as elastic expressions since it ignores string excitations. For \(b\ll \sqrt{Y}\) we find, \[\begin{aligned}
\label{inel} \text{Re} \delta&\simeq \frac{g^2s}{8\pi}\frac{1}{(4\pi Y)^{D/2-2}}(\frac{1}{D-4}-\frac{b^2}{4Y(D-2)}+...)\nonumber\\ \text{Im} \delta&\simeq \frac{\pi g^2 s}{8(4\pi Y)^{D/2-2}}\exp(\frac{b^2}{4Y}). \end{aligned}\] We can enhance our approximation by looking at the quantum fluctuations of \(\hat\delta\). \[\begin{aligned}
\exp(i\delta)\exp(i\Delta_\bot: \bar{\hat {X}}_{\bot u^2}:i\Delta_\parallel: \bar{\hat {X}}_{\parallel u^2}:+(u\leftrightarrow d)), \end{aligned}\] where \(\Delta_{ij}=\partial^2 \delta/\partial b_i\partial b_j\). \[\begin{aligned}
\Delta_\parallel=-(D-3)\Delta_\bot=\frac{(D-3)g^2s}{4\Omega_{D-2}b^{D-2}}. \end{aligned}\] Using our results for elastic scattering we find for \(b\gg\sqrt{Y}\) \[\begin{aligned}
\text{Im}\delta=\text{Im}\delta_{el}+\frac{(D-3)\pi g^2s}{4\Omega_{D-2}b^{D-2}}. \end{aligned}\] The transition between the dominance of \(\text{Im}\delta_{el}\) and \(\frac{(D-3)\pi g^2s}{4\Omega_{D-2}b^{D-2}}\) happens at \[\begin{aligned}
b_{DE}=(\frac{2\pi^2(D-3)}{\Omega_{D-2}}sg^2)^\frac{1}{D-2}l_s^{\frac{D}{D-2}}. \end{aligned}\] Before calculating the fixed angle scattering amplitudes, let us summarize the above results in the \((s,b)\) plane so far. For \(b\gg b_{DE}\), \(a(s,b)\) is polynomial while for \(\sqrt{Y}\ll b<b_{DE}\), we have \(a(s,b)\sim \exp(-Bg^2s/b^{D-2})\). Finally, for \(b\ll\sqrt{\ln(s)}\), from [\[inel\]](#inel){reference-type="eqref" reference="inel"}, we have \(a(s,b)\sim \exp(-A g^2s/\ln(s)^{D/2-2})\). In order to calculate the fixed angle amplitude [\[tb\]](#tb){reference-type="eqref" reference="tb"} we can use saddle point approximation. So for every angle \(\theta\), first we find the impact parameter \(\bf{b_s}\) where \[\begin{aligned}
q_i=-\frac{\text{Re} \delta}{\partial b_i}|_{\bf {b=b_s}} ~~~,~~~ \theta=\frac{|\bf{q}|}{\sqrt{s}}=-\frac{2}{\sqrt{s}}\frac{\text{Re}\delta}{\partial b}|_{\bf{b=b_s}}. \end{aligned}\] Then the fixed angle amplitude is given by \[\begin{aligned}
A(s,t)\sim \exp(i\bf{q.b_s})e^{i\delta(b_s,E)}, \end{aligned}\] up to polynomial prefactors. It turns out such a saddle point only exists for \(\theta\sim l_s^{D-2}Eg_s^2/b^{D-3}\). In the \(\sqrt{Y}\ll b<b_{DE}\) regime we have \[\begin{aligned}
\theta=\frac{D-2}{D-3}(\frac{r_S(E)}{b})^{D-3}, \end{aligned}\] where \(r_S(E)=(Eg_s^2l_s^{D-2})^{1/(D-3)}\) is the corresponding Schwarzschild radius. The amplitude would then read \[\begin{aligned}
A(s,\theta)\sim \exp(-Cg_s^\frac{-2}{D-3}(E/M_s)^\frac{D-4}{D-3}\theta^{-\frac{D-2}{D-3}}). \end{aligned}\] Although there is also a saddle point in \(b<\sqrt{Y}\), \[\begin{aligned}
\theta\sim\frac{bEg_s^2}{Y^{D/2-1}} \end{aligned}\] its contribution to the fixed angle amplitude is not dominant since the amplitude goes like \(\sim \exp(-A g^2s/\ln(s)^{D/2-2})\) which is very similar to the Gross-Mende saddle point. However, Gross-Mende saddle point exists for all angles as opposed to ACV saddle points. It is important to note that the dominant Gross-Mende saddle point is also at \(b<\sqrt{Y}\). According to the ACV analysis, Gross-Mende result is a saddle point that contributes to the fixed angle scattering but is not the dominant contribution. It most naturally compares to the ACV saddles found for \(b<\sqrt{Y}\) where the amplitude exponentially decays. In this picture the amplitude for small impact parameters goes like \(\exp (-S_{ent})\) and changes continiously accross the \(E=M_s/g_s^2\) line. Note that depending on the dimension, there could also be a minimum on the scattering angle of the saddle point in the region of exponential decay. When \[\begin{aligned}
\theta\lesssim \theta_{DE}=E^{\frac{3D-12}{D-2}}g_s^{\frac{4D-14}{D-2}}, \end{aligned}\] the saddle point lies in the elastic regime where amplitude is polynomial. In 4d, the LHS is \(\theta_{DE}=g_s\) while in higher dimensions it grows with energy. Thus, in 4d, the ACV analysis suggests exponentially decaying amplitudes for fixed angle \(\theta>g_s\) scattering. \[\begin{aligned}
A\sim \exp(-CE^\frac{D-4}{D-3}). \end{aligned}\] However, in higher dimesnions, for sufficiently high energies the amplitude is always polynomial. The reason behind this mismatch in four dimensions is that ACV by its use of the saddle point approximation focuses on the phase of the amplitude rather than the absolute value of it for different values of \(b\). In other words, ACV calculation assumes that because of the fastly oscilating phase of amplitudes as a function of \(b\) or \(l\) in [\[tb\]](#tb){reference-type="eqref" reference="tb"} or [\[gfa\]](#gfa){reference-type="eqref" reference="gfa"}, the amplitudes cancel each other out at values of \(b\) or \(l\) far away from the saddle point.
# Transition or correspondence? {#CS}
In the previous sections, we studied two aspects of strings and black holes that continuously match across the transition point. Namely, we looked at the transition between the scattering amplitudes and entropies. However, there are two other tests of continuity which challenge the correspondence between strings and black holes in its naive form. Those are the continuity of size and greybody factors. In fact these two properties are very closely connected for any gravitational bound state. The challenges we discuss in this section do not question the string/black hole transition, whether thermodynamically or as dominant scattering resonances. However, they question a particular interpretation of that transition which identifies black hole microstates with string states at the transition point. We refer to this interpretation as string/black hole correspondence to differentiate it from the more conservative observation of string/black hole transition.
## Size
The size of a free string of energy \(E\) which is given by \(\sqrt{\langle{\overline{(X-X_{\text{avg}})^2}}\rangle}\) can be approximated as \(l\propto l_s\sqrt{E\cdot l_s}\). Therefore, at the transition scale \(E\sim M_s/g_s^2\), the size of a free string is \(l\sim l_s/g_s\). This length is much greater than the size of the corresponding black hole which is \(l_s\). There are two ways to resolve this mismatch.
1. As we increase the mass of the string and approach the transition point, the self-gravitating effects become important and shrink the size of the string to match it with that of the black hole.
2. The states that continuously connect to black hole microstates at \(M\sim M_s/g_s^2\) are non-perturbative objects rather than strings. However, because they have smaller number of microstates than strings for \(M<M_s/g_s^2\), they do not significantly contribute to entropy or amplitudes below the transition point. In this picture, the string black hole transition is a transition between string microstates to other strongly coupled objects that become more thermodynamically likely at the transition point.
## Greybody factor
Highly massive strings and black holes both emit a near black-body radiation (for thermal radiation of heavy strings, see ). However, there is a fundamental difference between the two which lies in the deviation from the black-body spectrum. Black hole's Hawking radiation comes with a grebody factor which suppresses the radiation of higher spin particles. The origin of the grey-body factor is the gravitational potential created around the black hole which partially blocks higher spin particles from escaping to infinity. Since a free string does not have such a structure around it, it lacks the greybody factors. For a detailed account of the mismatch between the emission rates of strings and black holes see. In fact, if the back reaction of the string creates a structure with a size that matches that of the black hole, the matching of the grey-body factors will automatically follow. This is because the greybody factors of a Schwarzschild black hole, are universal consequences of its spherically symmetric gravitational profile. Therefore, the question of continuity of the size and the grey-body factors are in fact the same question. In order to see how string theory can account for the greybody factor, it is useful to look at setups where the black holes microstates are known to correspond to BPS branes with some compact dimensions. In those setups, the lower-spin particles correspond to polarizations of higher-dimensional fields along the compact dimension. In other words, the existence of the compact dimension is responsible makes a fundamental distinction between particles of different spins (see for more detail). However, such a structure is absent for a free string with no compact dimensions. In fact, as we saw in the previous section, in scatterings with energies below \(M_s/g_s^2\), a full thermalization does not happen. In other words, the majority of the scattering amplitude comes from intermediate states with large number of strings, which make up a negligible part of the string gas Hilbert space. At least from the perspective of scattering amplitudes, it seems that the string gas Hilbert space does not continuously connect to black hole Hilbert space.
## Horowitz-Polchinski solution and its limitations
The Horowitz-Polchinski solution, which is often said to describe a self-gravitating string, is a Euclidean EFT background that bridges black holes and free strings in terms of size. However, it is important to note a few points about it.
- The Horowitz-Polchinski solution only exists in \(d\leq6\) spacetime dimensions. Recently, it was argued that using a dimension expansion in \(d-7\), one can extend the results to higher dimensions . However, the nature of the solution still remains different in dimensions smaller and larger than \(7\).
- The transition between the Horowitz-Polchinski background and the black hole background as worldsheet theories, seems to be obstructed for type II theories.
- The Horowitz-Polchinski solution can be thought of as a string propagating in a thermal background, rather than a single self-gravitating string. In other words, it does not necessarily tell us that the background itself is a single string. For example, the insertion of a D-brane in a background is captured by the worldsheet theory of the fundamental string via boundary conditions. However, the fact that the background is described by a worldsheet theory does not imply that the background is a single string. This is why we have been cautious to use the the phrase string/black hole transition rather than string/black hole correspondence, since the transition of parameters do not imply that we can view the black hole microstates as extension of string states. In fact, in the known examples where the microscopic origin of the BPS black holes are known, the microstates correspond to non-perturbative objects rather than strings. However, nonetheless, those non-perturbative objects can be understood as worldsheet theories in weak coupling limits.
# Conclusions
We reviewed the string/black hole transition and how it can be understood in terms of the properties of high energy scattering amplitudes in string theory. These two topics are closely related, mainly due to the fact that all of the known formulations of quantum gravity are in terms of some boundary observables. In flat spacetime, these boundary observables take the form of an S-matrix with a rich structure. One example of an S-matrix that leads to a consistent theory of quantum gravity is string perturbation theory. Therefore, one should expect the universal properties of black holes to be manifested in string scattering amplitudes. For that, it is helpful to see black holes as resonances that dominate scattering amplitudes when the impact parameter is much less than the Schwarzschild radius. In this regime, the scattering is given by black hole formation and evaporation. Since the Hawking radiation is thermal at short time scales, the amplitude of a given event is exponentially suppressed as \(\ln(\mathcal{M})\sim-S_{BH}\) where \(S_{BH}\) is the black hole entropy. Even though string perturbation breaks down in the regime where black hole processes are dominant, we can hope to access its boundary from string perturbation and see how the black hole behavior starts to form at the boundary. The string/black hole transition suggests that the mass of the smallest black hole is \(M_s/g_s^2\). Therefore, the black hole regime is \(E\gg M_s/g_s^2\) and \(b\ll r_s(E)\). This region has two boundaries, one at \(E=M_s/g_s^2\) where string states supposedly start to form black holes, and the other is \(b\sim r_s\) for \(E\gg M_s/g_s^2\) where the black hole already has a sharp meaning in the effective field theory. As we cross the second boundary, we should expect a sharp change in the scattering amplitudes as they lose their angular dependence and become exponentially suppressed to match the Hawking radiation. This is indeed the case and the stringy results at \(b\gg r_S\) does not match the black hole result at \(b\ll r_S\). This makes it challenging to study large black holes from string theory amplitudes. However, for the \(E\sim M_s/g_s^2\) boundary, we expect to see a continuous transition motivated by the first order phase transition between strings and black holes. This is indeed the case. In the stringy regime (\(E\lesssim M_s/g_s^2\)), the scattering events are mediated by resonances that are described by classical Gross-Mende saddle points. These scattering events are mediated by \(\sim E/M_s\) localized strings and impact parameter \(\sim l_s\). This is very similar to the black hole regime (\(E\gtrsim M_s/g_s^2\)) where the scattering events are mediated by black holes that can be seen as the saddle point of the classical theory. Interestingly, the stringy result is very similar to the black hole result in that it is also given by \(\mathcal{M}\sim\exp(-cS)\) where this time \(S\) is the entropy of a string gas. A priori, the relationship between scattering amplitude and entropy is not clear in the stringy regime, because as opposed to black holes, the intermediate state does not thermalize along the scattering event. The explanation of this matching is a little subtle in the stringy regime. As we showed in [3.3](#RAC){reference-type="ref" reference="RAC"}, the entropy of the string gas with fixed \(k\lesssim E/M_s\) number of strings, is of the same order of a string gas with energy \(E\) which matches with the black hole entropy at the transition point. Based on this observation, we propose a thermal interpretation for the string/black hole transition. As we increase the energy, most of the scattering goes through saddles that correspond to a subsystem of entropy \(\sim S\). Moreover, given that the dependence of the amplitude to the individual outgoing momenta is small, these are hard scattering processes that seem insensitive to the information of a individual out-going particles. In other words, the information is highly scrambled and is hidden in the entanglement between the particles. As we increase the energy, the Hilbert space that is accessible to the in-going states eventually is described by black holes. This fits nicely with the string/black hole transition and allows one to study the information paradox just below the transition using string theory S-matrix. Moreover, the string S-matrix opens the door to study the complexity of the black hole via the emergence of chaos in the high energy string scatterings. We also reviewed two different interpretations of the string/black hole transition in section [5](#CS){reference-type="ref" reference="CS"}. The first interpretation views the string/black hole transition as a phase transition between string states and black hole states in terms of dominance in thermodynamics and scattering analysis. The second interpretation views the transition as an identification between the string states and the black hole microstates at the correspondence point. The first interpretation is more conservative and is supported by the known examples in string theory. This is due to the fact that the known microscopic descriptions of black hole microstates in string theory are not in terms of excitations of weakly strings, but non-perturbative objects such as D-branes. | {'timestamp': '2022-12-01T02:19:25', 'yymm': '2211', 'arxiv_id': '2211.17162', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17162'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# [\[sec:Intro\]]{#sec:Intro label="sec:Intro"} Introduction
There are about 250 stable isotopes and over 3000 unstable ones are known. One way that unstable atomic nuclei decay into stable ones is \(\beta\)-decay. As a result, the \(\beta\)-decay energy \(Q_\beta\) that goes with it is a fundamental property of unstable atomic nuclei. \(Q_\beta\) values can be determined via several methods, such as \(\beta\) endpoint measurements, counting in coincidence with annihilation radiation, electron capture \(EC/\beta^+\) ratio method, \(\gamma\) absorption with X-ray coincidence. This process is complex to explore provided that its energy spectrum has a continuous structure. The decay energy seems to simply relate to the proton number and mass of the atomic nuclei. Experimental \(Q_\beta\) energies of nuclei can be verified in the \(Z \in [4,\ 82]\), \(N \in [4,\ 126]\) regions, whereas unknown energies can be calculated using \(\alpha\)-\(\beta\) energy cycles by means of reliable \(\alpha\) spectroscopic data. Moreover, some \(\beta\)-decay energies can be obtained with the help of energy cycles from \(\alpha\)-decay energy systematic .
## Explainable Machine Learning {#subsec:Explainable-ML}
Machine Learning (ML) has recently emerged as an important tool for understanding physical phenomena, focusing on understanding how a model makes a particular decision, rather than just accepting the output as a black-box. In the supervised approach, an algorithm is used as a functional approximation \(\mathcal{F}(x)\) for an unknown function \(F(x)\). Thus, it is possible to obtain the desired output for the collected physical data \(x\in \mathbb{R}^n\). Classical statistical algorithms such as linear regression models or Principal Components Regression (PCR) allow for characterizing the relevant predicting features of the model, thus providing an understanding of the physical phenomena underlying the data. However, they impose specific restrictions on the studied manifold, and thus, they are not well suited to most of the complex phenomena observed in nature. On the other hand, modern approaches such as deep learning (DL) allow for a comprehensive representation of physical observables, as they work as universal function approximators [@chen2018universal; @scarselli1998universal]. The trade-off comes when it is necessary to understand what the algorithms are considering. To address this issue, explainable ML allows identifying of underlying characteristics such as feature importance [@casalicchio2018visualizing; @wojtas2020feature; @rengasamy2021mechanistic], model the dynamics of complex systems, conservation law discovery and symbolic expression extraction [@cranmer2020discovering; @udrescu2020ai; @bakarji2022discovering]. Besides, simple ML models have previously been used in the study of the \(\beta\)-decay [@niu2019predictions; @bayram; @cruz2019clustering; @costiris2007global; @gao2021machine]; however, there is still the need to obtain physical intuition out of these techniques. Nuclear physics is not an exception, as determination of one and two proton separation energies, developing nuclear mass systematics , determination of ground state energies of the nuclei, identification of impact parameters in heavy-ion collisions, estimation of fusion reaction cross-section , estimating nuclear RMS charge radii, decoding \(\beta\)-decay systematic and determination of fusion barrier heights have already been studied via this tooling.
## ML for \(\beta\)-decay {#subsec:ML_beta_decay}
To be precise, the \(\beta\)-decay energies were estimated by considering the neutron and proton numbers of atomic nuclei as the only input parameters and the data values have been obtained based on the Hartree--Fock--BCS method with the Skyrme force MSk7. By contrast, here, the recent AME2020 experimental data is considered and the input parameters are augmented by considering a full range of physical properties of atoms (see Table [\[tab:table_desc\]](#tab:table_desc){reference-type="ref" reference="tab:table_desc"} below). By using the experimental \(Q_\beta\) values available in the literature, the systematics of atomic nuclei related to this energy was obtained here by performing ML, considerably decreasing the root mean squared error (RMSE) of previous approaches by a factor of 2.5. In addition, by calculating the importance scores, the fundamental parameters of atomic nuclei that have an impact on the \(Q_\beta\) energies are highlighted.
## The atomic mass table {#subsec:Intro_A}
The dataset consists of the classical nuclear variables, numerically described in Table [\[tab:table_desc\]](#tab:table_desc){reference-type="ref" reference="tab:table_desc"}, whence entries with any non-computable data entrances were removed, leaving a total of 2813 unstable isotopes. The relevant dataset variables are the number of neutrons in each isotope \(N\), number of protons in each isotope \(Z\), mass number \(A\), difference between measured mass and the mass number \(M_e\) and binding energy \(B_E\). Here, notice that a majority of the features are highly skewed. Further description of the raw data is shown in Appendix [\[sec:datadist\]](#sec:datadist){reference-type="ref" reference="sec:datadist"}.
It is possible to identify high correlations between some of the variables provided that some of the features are computable from the others (such as the mass number from the atomic and nuclear numbers). This needs to be addressed since it is known that algorithms that are susceptible to multicollinear features generate unreliable predictions. Therefore, an exploratory dimensionality reduction analysis is implemented using Principal Component Analysis (PCA) with the normalized variables and evaluating the performance of a naive predictor model. Here, both the explained variance and a principal component regression (PCR) used a 5-fold cross-validation scheme . As expected, more than two principal components are necessary to give a complete sense of \(Q_\beta\) by evaluating the RMSE of a linear regression model. In addition, Fig. [\[fig:r2pca\]](#fig:r2pca){reference-type="ref" reference="fig:r2pca"} depicts the asymptotic behavior that is reached using more than 3 principal components. Nevertheless, neither the second nor the third principal components are negligible. The PCA loadings, obtained as shown in Fig. [\[fig:loadings_pca\]](#fig:loadings_pca){reference-type="ref" reference="fig:loadings_pca"}, were used to understand the importance of each of the parameters. Interestingly enough, the second orthogonal component loads more heavily the uncertainties, rather than the components themselves, with a major role in the binding energy \(B_\mathrm{E}\). This suggests that there exists a significant contribution of the features related to the experimental uncertainties reported in the original dataset. However, the correlation analysis shown in Fig. [\[fig:correlations\]](#fig:correlations){reference-type="ref" reference="fig:correlations"} accounts for part of this effect. Nonetheless, it is expected that Universal Approximators can easily overcome this issue.
# Modeling \(Q_\beta\): A regression approach
To start the modeling of the phenomena in a systematic and replicable way, training/testing/validation sets were generated with the dataset using a 70/20/10 split. A variety of tabular models were fitted in an automatic manner, reporting the scores, and using only the training and validation splits. In this approach, in each one of the steps, the best-performing models are selected and their hyper-parameters are tuned via Bayesian optimization techniques as designed by Ref. . Based on a collection of regression metrics (such as explained variance, RSME, and their volatility among different splits of the training set), the models obtaining high scores while keeping explainability were selected and further explored via hyperparameter tuning and feature importance measurement techniques. To have a baseline, the previous work was replicated using a similar approach. Here, a vanilla fully-connected Multi-Layer Perceptron (MLP) was implemented, but with the AME2020 dataset, which was trained only using the \(N\) and \(Z\) of each one of the entries. The architecture consists of a single hidden layer with 100 neurons, hyperbolic-tangent non-linearities, and classic Stochastic Gradient Descent (SGD).
## ML and AI models {#subsec:ML_AI_models}
### Ordinary Linear Regression {#subsec:ordinary_regression}
Besides its simplicity, this approach is well-suited to scientific tasks because of the transparency, a general regression allows. Thus, it serves both as a benchmark and brings insights into the process. The Ordinary Least Squares (OLS) approach uses a projection model to fit the input vector \(\mathbf{X}\) containing the selected features to obtain an estimator \(\mathcal{Q}_\beta\) via \[\label{equ:q_beta_fit} \mathcal Q_\beta = \mathbf{X}^T \cdot \mathbf{k} + \varepsilon.\] The optimal parameters for the vector of coefficients \(\mathbf{k}\) and the scalar \(\varepsilon\), which minimizes the difference between the observation \(Q_\beta\) and the outputs \(\mathcal Q_\beta\) are obtained via the Ordinary Least Squares procedure. Moreover, this technique allows for obtaining a statistical measurement of the fitted coefficients and their uncertainties, thus, getting insight into the relevant features and characteristics of the model.
### Modern MLP
Multilayer Perceptrons (MLPs) are a type of Artificial Neural Network (ANN) in which neurons are connected via a directed acyclic graph. They have one or more inputs, more than one hidden layer, and one output layer. The hidden neurons simply propagate information forward to the next layer, each layer is a linear transformation of the previous one via non-linearities, which are required to allow a general approximation of a black box function such as the studied \(\mathcal Q_\beta = \mathcal Q_\beta (A, Z, N...)\). Such functions are referred to as the layer activation function and the output layer corresponds to a prediction of \(Q_\beta\). The Neural Network (NN) architecture schema is shown in Fig. [\[fig:nn_architecture\]](#fig:nn_architecture){reference-type="ref" reference="fig:nn_architecture"}. The Gaussian Error Linear Unit (GeLU) activation function was included as a non-linearity, given that it has been shown that helps reduce several problems associated with small datasets . In addition, an inter-layer dropout is included, which discards random connections between layers with a probability of 0.1. This has been shown also to help in the optimization procedure . Several experiments modifying the hyper-parameters via optimization methods were performed to obtain a suitable configuration of both the number of neurons per layer and the batch size . A proper learning rate was also obtained using the method of cyclical learning rates as proposed in Ref. . The training process used a decaying learning rate and was implemented using the PyTorch Lightning library. This used mini-batch gradient descent to optimize the Mean Squared Error Loss (MSELoss) with the Adam optimizer . Before starting the actual training process, three warm-up rounds were used.
### Gradient boosting machines
Besides being increasingly known and used in different areas of physics, ANN are far from being the only available option for an explainable and accessible model, a clear example is an extreme gradient boosting (with [XGBoost]{.smallcaps}) algorithm , which uses the gradient boosted trees (BDT) method for fitting several tree regressors (or weak linear learners) using a regularized loss function. These are compared and later assembled with the approach of ML boosting. They perform especially well on tabular data and in most cases out-performing deep-learning approaches, requiring only fractions of the computational cost . Here, the XGBoost model with the optimization of hyper-parameters with Optuna was implemented and several hyper-parameters such as the depth, learning rate, and the number of learners were inspected. Moreover, callbacks to avoid over-fitting were added to the loss minimization process.
### Attention Networks
The *attention mechanism* is implemented here to properly learning weights and relations within the input via learning queries and keys. Besides, this approach was found to be successful for unstructured data in several tasks such as protein unfolding and natural language processing, it had not been especially useful for tabular data until the the appearance of the TabNet model . In this approach, the TabNet model was used with Weighted Adam optimizer and early stopping callbacks to avoid overfitting.
### Ensemble model
Even though the fitting generated with the methods above might fit accurately to the datasets, the models trained can generally either underfit, overfit or just be poorly configured. The aim of the ensemble method is to facilitate the best of several Base Models to train a strong Prediction Model. This is achieved via a Voting Ensemble Regression (VER), a linear combination of the predictions is done by weighting the individual models according to their performance on the validation split. Specifically, the weighting of the BDTs, the ANN, and the TabNet model is performed by evaluating both the RMSE and the standard deviation over a 5-fold in the validation set.
## Data Augmentation Techniques[\[sec:usage_unc\]]{#sec:usage_unc label="sec:usage_unc"}
From the initial evaluation of the feature importance, experimental uncertainties arose as noteworthy in the regression of \(Q_\beta\). Here, they come from the combination of both epistemic and systematic errors. Also, a characteristic often makes it difficult for the models to improve their performance is the limited experimental data they are training on. Therefore, a method for obtaining more diverse and accurate data is useful. A Monte Carlo sampling on each one of the parameters with registered uncertainty is proposed. Assuming that each one of the variables has an error normally distributed and centered on the reported data. This allows obtaining a significant data set for further improving the ML models with increased robustness. This is only done in the training entries so that no isotopes are overlapped between the split, ensuring no data leakage. Notice that this approach can be followed by the same procedure as an augmentation at testing time, as it has been shown to considerably improve robustness . Following this procedure, the original training data was overpopulated, from \(1912\) to a total of \(573600\) samples, followed by the removal of the columns related to uncertainties, since this information has already been taken into account. The models that were trained on the original data (which will continue to be referred to as AME2020) were also trained and tuned on the augmented dataset.
## Periodic Feature Injection[\[sec:usage_priodic\]]{#sec:usage_priodic label="sec:usage_priodic"}
In the spirit of the augmentation technique presented in Sect. [\[sec:usage_unc\]](#sec:usage_unc){reference-type="ref" reference="sec:usage_unc"}, physics-inspired features were added to both datasets (i.e to the AME2020 and to the augmented dataset from Sect [\[sec:usage_unc\]](#sec:usage_unc){reference-type="ref" reference="sec:usage_unc"}). The periodic group was added to the data as a one-hot encoded feature. Moreover, the *magic numbers* such that they are arranged into complete shells in the nucleus were also taken for each one of the isotopes, both for \(N\) and \(Z\). For this, the dataset includes the difference to the closes magic number (which we denote as \(m_N\) and \(m_Z\), respectively). In addition, the ratio \(m_Z/(m_N+1)\) and \(Z/N\) were also included. Given that this method heavily relies on the current atomic model, we denote the generated data as *Injected*.
# Results
For each one of the models described above, the coefficient of determination (\(R^2\)) of the fit, the Mean Squared Error (MSE), the Root The Mean Squared Error (RMSE), and the Mean Absolute Error (MAE) were calculated. We use this collection of metrics to get an estimation of how well distributed and big the deviations between \(\mathcal{Q}_\beta\) and \(Q_\beta\) is for the unseen testing dataset. Below, are the results obtained using the approaches of different ML models with several datasets.
### AME2020 {#ame2020 .unnumbered}
After fitting several regressors on the original split, modern architecture clearly outperforms the previous state-of-the-art RMSE. Notably, results obtained by the OLS are not considerably bigger compared to the baseline ANN. Moreover, experiments suggest that the number of parameters that have to be trained on the TabNet requires considerably more than one entry per isotope, thus the metrics are not taken into account.
\
### Augmented Dataset {#augmented-dataset .unnumbered}
Since the amount of data in this used in this approach is considerably big, the OLS was not implemented.
\
\
:::
This approach demonstrates the importance of generating features based on theoretical models, as higher-level atomic features clearly allow models to fit the function better. This is noticeable in the distribution of deviations. As expected, this approach allows a considerable reduction in errors, especially for the critical isotopes in not injected approaches, and with the highest error at a low number of isotopes, as illustrated in Fig. [\[fig:ensemble_inj_error_NZ\]](#fig:ensemble_inj_error_NZ){reference-type="ref" reference="fig:ensemble_inj_error_NZ"}. Notably, the two isotopes with an error above the percentile 90 correspond to Helium (isotope with \(A=6\)) and Nitrogen (isotope with \(A=24\)).
# Explaining the Black-Box
Up to this point, ML techniques have been shown to provide a highly accurate regression of the \(\beta\)-decay energies. However, it is possible further to inspect the impact of each one of the features on the output of the model. This can be done via the evaluation of the Shapley Additive explanations (SHAP), a model-agnostic explainability method inherited from game theory, where the feature importance of the feature is calculated via a global analysis of the marginals per each input feature . For this end, the KernelSHAP technique implemented in Ref. is applied. The results are presented in Fig. [\[fig:shaps\]](#fig:shaps){reference-type="ref" reference="fig:shaps"} for the case of BDT, which was fully trained and hyper-optimized on the AME2020 training data. In accordance with our explorations with the PCA approach, the uncertainties play a significant role in explaining the model output and that in general, the mass excess is considered charged. The feature importance after training on the augmented the dataset was also studied. It was found that the importance scores for the physical features (\(N\), \(Z\), Mass Excess, \(B_E\), Atomic Mass, and \(A\)) kept their assigned importance as expected from an orthogonal permutational scoring. Moreover, the feature weights inside the trees for the BDT and the feature permutation importance for the ANN, which matches the overall order was also investigated. The models thus fit the coefficients with a major importance of \(Z\) and \(N\), which could in fact explain the good results obtained in Ref. with such a simple architecture. Furthermore, the relation found by the black boxes, in general, is inversely proportional for \(Z\) and direct for \(N\). More importantly, the fact that the groups were detrimental to the performance of the regressors was validated by the permutational scoring. By considering the periodic injected features, the high importance of the ratios \(N/Z\), and the magic numbers give insights into why the critical isotopes near the magic numbers were predicted with lower errors (cf [\[fig:f_scores\]](#fig:f_scores){reference-type="ref" reference="fig:f_scores"}). Remarkably, the feature importance averaged over 5 runs for the models show to have a relative low \(F\) score, experiments show that injecting periodic variables allows the BDT model to achieve the same metrics as the augmented dataset, and to increase metrics on the augmented dataset with virtually no overfitting, as illustrated in Fig. [\[fig:losses_xgb\]](#fig:losses_xgb){reference-type="ref" reference="fig:losses_xgb"}. Further inspection of the injected features also shows interesting behavior arising from physical phenomena. For instance, both for \(mZ\) (absolute difference between \(Z\) and the closest magic number) and \(mN\) (difference between \(N\) and the closest magic number) have alternating feature importance with parity as shown in Fig [\[fig:altern_shap\]](#fig:altern_shap){reference-type="ref" reference="fig:altern_shap"}. The oddness in the number of nucleons in relation to the magic numbers, in general, results in lower \(B_E\), which is in excellent agreement with the Nuclear Shell Model as well as theoretical approximations of nuclear structure such as the Weizsäcker formula . This alternation in the SHAP importance value is also larger near the magic numbers, which might explain the improvement in deviations near magic numbers with the feature injection, and is clear evidence that the model is already capable of getting insights into the nuclear structures.
# Conclusions and Outlooks
In this work, an ML model was developed to predict the \(\beta\)-decay energies of isotopes using the experimental dataset AME2020 . Using modern techniques and data augmentation strategies, it was shown that our models can considerably decrease the error in estimating \(Q_\beta\) compared with previous ML approaches while maintaining interpretability. The importance of considering experimental uncertainties was illustrated as the models incorporating sampling techniques become considerably more robust to noise. Furthermore, it was demonstrated that by incorporating physical features such as magic numbers, models can be highly improved. Using explainability tools it is shown that the models have learned fundamental features of the atomic structure. We look forward to generating clearer physical intuition from the explainability extracted from several ML techniques. Clearly, the intersection between ML and physics is only starting to showcase its potential. | {'timestamp': '2022-12-01T02:18:42', 'yymm': '2211', 'arxiv_id': '2211.17136', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17136'} |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction and Motivation {#sec:introduction}
The ever more widespread use of social media and user-contributed content also causes an increase of toxic or offensive language and other forms of unwanted contributions which may need to get detected and removed. In this paper, we present work aimed at supporting moderators of a large daily Austrian (German language) newspaper which allows registered users to discuss the articles published on its web-site. Users post some 20K to 50K comments per day. An analysis of commenting behaviour has shown that only a third of the users participating in the online discussion are women and that one important reason why women avoid participating in article forum discussions is the presence of sexist comments. The aim therefore was to use automatic classification of sexist/misogynist comments to support moderators in detecting such comments in order to provide a more welcoming and safer climate of discussion especially for female users. For this, a corpus of 6600 comments was collected and annotated and subsequently used to train a classifier used to flag comments or entire discussion forums with a high number of suspected misogynist comments.
# Related Work {#sec:relatedwork}
Together with work on toxic and offensive language classification in recent years, there has also been increasing work on the classification of sexist or misogynist language, sometimes as part of a more general toxic language classification task. give an overview over earlier work and describe a dataset of Tweets containing abusive sexist terms. present work on creating a dataset of tweets, subsequently used as part of IberEval-2018 and EvalIta-2018 challenges for sexism classification. give an overview over misogyny detection in social media, specifically Twitter. describe work on a corpus of 16K tweets for detecting toxic and hate speech including sexist slurs or defending sexism (3383 tweets with sexist content). present work on the datasets described in and. include some more detailed description of sexist language and is based on another dataset of English language tweets. describe work on categorizing accounts of sexism from the Everyday Sexism Project website through fine-grained multilabel classification. Other datasets are described or used in (12K tweets in French) and (tweets, work-related quotes, press quotes and other sources), and (Youtube comments in Indian English, Hindi and Bengla), (Spanish language tweets) and (Danish language dataset sampled from several social media sites). The EXIST task at IberLEF 2021 addresses the identification and categorization of sexism in English and Spanish language tweets and postings from [Gab.com](Gab.com). Most recently, SemEval 2023 Task 10[^1] provides an English language corpus of 20000 texts sampled from Gab and Reddit, annotated with 3 hierarchical labels.
# Corpus Creation {#sec:corpuscreation}
The aim of manual corpus annotation was to reflect the judgement of moderators in their everyday work. For this reason, the manual annotations were carried out by 8 annotators of which 7 were experienced moderators. The 8th annotator is a natural language processing and corpus linguistics expert. One of the annotators is among the authors of this paper. There were 3 male and 5 female annotators. Since the phenomenon of \"sexism\"/\"misogyny\" is complex, guidelines were created to describe the most important kinds of sexism relevant for the annotation task. For this we used the categorization from as inspiration. The guidelines also attempt to clarify some of the difficulties likely to be encountered: how to decide if there is not enough context, what if the sexist remark is aimed at a man or men in general, how to treat \"reported sexism\". However, the guidelines follow the aim of providing help for annotating in a way that reflects the daily work of moderators and the newspaper's editorial concept. They are not meant as an accurate abstract definition of sexism and misogyny. The (German language) guidelines are available online[^2]. We will refer to the task as \"misogyny classification\" in the rest of the paper. Postings were annotated by assigning one of 5 possible labels 0.. 4, corresponding to 0=absence of misogyny and 4 levels of \"severity\" of the expressed misogyny as perceived by the individual annotators, with 1=mild, 2=present, 3=strong, 4=extreme. This was done on the one hand to reflect the personal aspect in the assessment of misogyny, and on the other hand to identify the instances with the biggest disagreements among annotators. Postings to be annotated were collected from several different sources: (1) a collection of postings which had been reported with a (free text) reporting reason that included a keyword related to sexism/misogyny, (2) postings which were reported with a different reason, (3) postings randomly sampled from all available postings, (4) a subset of postings (2) preclassified with an early version of the binary classifier trained on the first 2800 annotated postings and (5) postings from 24 article forums which have been identified to contain an above-average number of postings considered sexist, preclassified with the same early binary classifier. Preclassified postings were selected from the highest probability label \"1\" postings (to correct false positives) as well as those label \"0\" and \"1\" postings with close to 0.5 probability (to add what may be hard to classify instances). Postings were given to annotators in batches of 100, by creating a spreadsheet from the posting texts and preparing a selection field for selecting one of the 5 possible labels. The first batch of 100 postings was given to all 8 annotators and then analysed to find postings with the biggest disagreement: for this we calculated a heuristic disagreement score based on all pairwise distances between the labels, where the distance between labels 0 and 1 were defined to be 4, and distances between labels \(l_i, l_j >0\) defined to be \(|l_i-l_j|\). Examples with high scores were then discussed among annotators to clarify the annotation guidelines or clear up misunderstandings. After this, each round of 100 postings was given to a random selection of 3 annotators available at the time. This was done in order to compromise between annotating as many postings as possible given the available time and resources and still get enough annotators for each posting to identify disagreements. In total, 66 rounds with 6600 postings were annotated (20300 annotations). The overall distribution of assigned labels is shown in Table [1](#table:dist5){reference-type="ref" reference="table:dist5"}.
[\[table:dist5\]]{#table:dist5 label="table:dist5"}
# Annotator Agreement and Corpus Analysis {#sec:corpusanalysis}
Krippendorff Alpha over all annotations was 0.36 (nominal scale) and 0.64 (ordinal scale). After binarization of the 5 possible annotations into 0 (for no sexism) and 1 (for labels \(1\dots 4\)), Krippendorff Alpha was 0.60. Overall agreement was 0.65 / 0.83 (binary) if macro averaged over all agreements of pairs of users, and 0.66 / 0.82 (binary) if micro averaged over all pairs of annotations. F1.0 macro over all 22300 pairs of annotations was 0.390 / 0.803 (binary). Overall Cohen's Kappa was 0.39 / 0.63 (binary) if macro averaged over all kappas for pairs of users. This indicates that there was considerable difference of opinion or difficulty in assigning the fine-grained labels. As shown in Table [1](#table:dist5){reference-type="ref" reference="table:dist5"} the overall rate of assigned positive labels (classes \(1\dots 4\)) was 0.335. Looking at all pairs of annotations in the dataset the relative frequencies of annotation pairs (confusion matrix) is shown in Table [2](#table:conf1){reference-type="ref" reference="table:conf1"}. This illustrates the large rate of disagreement among annotators, especially on estimating the fine-grained degree of mysogyny (labels \(1\dots 4\)).
[\[table:conf1\]]{#table:conf1 label="table:conf1"}
We deliberately did not decide on a single \"best\" way to decide on a final single label as the judgement on misogyny may depend on personal opinion and we believe it would be wrong to assume there is a single correct value for each instance. Other instances may have received labels by mistake (e.g. misinterpretation of the text or of the annotation guidelines). For creating a training set we used different strategies to resolve disagreements (see Section [5](#sec:model){reference-type="ref" reference="sec:model"}). We aim to make the corpus available for academic research, the necessary steps for doing this legally and in accordance with Austrian and EU data protection law are still ongoing. This is not only the first German language corpus related to misogyny detection but also the first corpus in any language which covers such a wide range of often very subtle ways a comment may express mysogyny, and which contains many comments where human annotators will have different opinions on which label is most adequate.
# Sexism Classification Model {#sec:model}
The main purpose of the deployed classification model is to alert moderators both of individual sexist comments and article forums with a high rate of potentially sexist comments. For this reason, our main interest is a binary classifier (original label 0 vs original labels 1 to 4). However we also studied the performance of a model for predicting the original label, both as seen as a multiclass classification task and as an ordinal regression task. Finally, we investigated if combining both the binary and multiclass tasks into a multi-task model would impair or improve the performance of the individual tasks. All models are based on a transformer architecture with one or two classification heads on top of the pooling layer. We used the pretrained German BERT models `gbert-base`[^3] and `gbert-large`[^4]. For all models, accuracy and F1.0-macro metrics where estimated using 5-fold cross-validation with class-stratified folds. 10% of the training set were used as dev-set and this split was class-stratified as well. Hyper-parameters where selected in a step-wise process where manually chosen values for a few hyper-parameters were evaluated using grid search. For many parameters, however, there was no clear best selection as the range of F1.0-macro estimation results introduced by different random seeds was larger or comparable to the changes in F1.0-macro estimation for different parameter values. In such cases we chose an intermediate value among those with similar estimates. The final set of hyper-parameters used for all models described below was: BERT maximum sequence length 192, batch size 8, gradient accumulation: 1 batch, learning rate 7.5e-06, language model dropout rate 0.1 and no layer-wise learning rate adjustment. The AdamW optimizer with weight decay 0.01 and linear warm-up during 200 training steps was used. For all classification heads we used an addition layer with 768 hidden units, ReLU and no dropout before the actual output layer. We evaluated the following single-head models: Bin (binary classification), Multi (multiclass classification), Coral (multiclass classification using an implementation of the CORAL ordinal regression model ); and the following dual-head models: BinMulti (binary and multiclass heads combined), BinCoral (binary and CORAL heads combined). For each of these 5 models we evaluated a variant based on the `bert-base` model (/B) and one based on the `bert-large` model (/L). Table [3](#table:mostfreq){reference-type="ref" reference="table:mostfreq"} shows the estimation results (accuracy and F1.0-macro) on a training set where the original and binarized targets where selected as the most frequently assigned labels, falling back to the highest most frequently assigned label or the maximum assigned label. For dual-head multitask models there are two lines, showing the binary model as head 1 (:1) and the multiclass model as head 2 (:2).
The dual-head model combining binary and CORAL ordinal regression heads shows the best accuracy estimates both for the binary and the multiclass tasks, and also the best F1.0-macro for the binary tasks while the performance of the base single-head multiclass model shows best F1.0-macro for the malticlass task. Interestingly, the models based on `bert-large` did not always perform better than those based on `bert-base`.
We also trained the same set of models on a training corpus where both the binary and multi-class target was always selected as the highest label assigned by any annotator (\"when in doubt, treat it as misogynist\"). The results for this experiment are shown in Table [4](#table:choosemax){reference-type="ref" reference="table:choosemax"}. On this data, the single-head binary model performed best both with respect to accuracy and F1.0-macro, while the single-head multiclass classification model performed best on the multiclass task with respect to F1.0-macro. In order to get an impression for how a binary model could be used to indicate forums where moderator intervention may be necessary because of a high number of misogynist postings, moderators selected 6 forums of which 3 have been found to have a high observed rate of misogynist postings and 3 without. The model BinCoral/L:1 from Table [3](#table:mostfreq){reference-type="ref" reference="table:mostfreq"} assigned a rate of 0.23, 0.17, and 0.27 of postings in each of the first 3 forums to the misogyny class but only a rate of 0.01, 0.02 and 0.07 of the second 3 forums, thus giving a good indication of which of the forums moderators should prioritize. Looking at the actual postings labeled to be misogynist shows that in addition to many postings which were correctly identified, there was also a high number of false positives. However, many of these false positives are texts which are related to topics often raised in misogynist comments. In future work, the corpus will be expanded by providing manual annotations for the false positives detected in those evaluation runs. The code for all experiments is based on the FARM library[^5] and is available online[^6]. Additional future work will analyse the performance of the different classification models on comments based on source, annotator disagreement or the kind of misogyny and improve the approach both by further improving and extending the corpus and the models used for training. This work was conducted as part of project FemDwell[^7] supported through FemPower IKT 2018[^8].
[^1]: <https://codalab.lisn.upsaclay.fr/competitions/7124>
[^2]: <https://ofai.github.io/femdwell/annotation_guidelines_v1.pdf>
[^3]: <https://huggingface.co/deepset/gbert-base/tree/main>
[^4]: <https://huggingface.co/deepset/gbert-large/tree/main>
[^5]: https://github.com/deepset-ai/FARM
[^6]: URL withheld for anonymous review
[^7]: <https://ofai.github.io/femdwell/>
[^8]: <https://austrianstartups.com/event/call-fempower-ikt-2018> | {'timestamp': '2022-12-01T02:19:26', 'yymm': '2211', 'arxiv_id': '2211.17163', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17163'} |
# Introduction {#intro .unnumbered}
It is a long been asked Diophantine question: which cube-free integers
can be expressed as a sum of cubes of two rational numbers? If
is a cube-free integer and
denotes the Mordell-Weil rank of the elliptic curve
over
, then it is a well-known fact that
can be written as a sum of two rational cubes if and only if
For a prime
a conjecture classically attributed to Sylvester (see for example ) predicts that
One can reformulate the conjecture above as follows:
be a prime and either
or
We have
1. If
::: {.small}
\(n \not\equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(\text{rk } E_{-432n^2}(\Q) = \begin{cases} 0, & \text{ if } \ell \equiv 2 \pmod{3},\\ 1, & \text{ if } \ell \equiv 1 \pmod{3}. \end{cases}\)
:::
2. If
::: {.small}
\(n \equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(\text{rk } E_{-432n^2}(\Q) = \begin{cases} 1, & \text{ if } \ell \equiv 2 \pmod{3},\\ \{0, 2 \}, & \text{ if } \ell \equiv 1 \pmod{3}. \end{cases}\)
:::
have been done by many mathematicians including Sylvester, Satgé, Selmer and many others. For example, the case when
is a prime such that
then the conjecture above was proved by Sylvester himself. If
, then proved that
and
are cube sums under the hypothesis that \(3\) is not a cube modulo
In, the authors showed that there are infinitely many primes congruent to
which are sum of two rational cubes. Many other authors have contributed towards the Sylvester's conjecture and the cube sum problem, for specific congruence classes of primes modulo \(9\) ,. However, the problem is still open in full generality. In, we proved Sylvester's Conjecture under the hypothesis that the Tate-Shafarevich group
satisfies
is even. For a general cube-free integer
recently, announced a proof of the fact that a positive proportion of integers are cube sums and a positive proportion are not. Apart from this, very little is known about this problem in general. If
has two distinct prime factors, we know some partial results. Satgé proved that
(resp.
) is a cube sum when
(resp.
); and also study this situation. In, we have also examined the integers of type
and
for their solubility as a cube sum, where
is a prime. In this article, we generalize the situation to consider the cube sum problem for cube-free integers
which are co-prime to \(3\) and have exactly two distinct prime factors, that is
or
or
where
and
are two distinct primes, none of which is equal to \(3\). We start by stating a conjecture which is slightly more general in nature. By the work of Birch-Stephens, we know that for a cube-free
the global root number of
is given by
In particular, when
is a cube-free integer co-prime to \(3\), this formula for the global root number simplifies to
(resp. to
) if
(resp. if
); here
is the number of distinct primes
which divide
and are congruent to \(2\) modulo \(3\). Looking at the sign of the functional equation of
we predict:
be a cube-free integer co-prime to \(3\). Assume that the set of prime divisors of
is given by
where
and
Let
We have
1. If
::: {.small}
\(n \not\equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(r_n \le 2k_1+k_2-1\)
:::
and
::: {.small}
\(r_n \equiv k_2-1 \pmod 2\).
:::
2. If
::: {.small}
\(n \equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(r_n \le 2k_1+k_2\)
:::
and
::: {.small}
\(r_n \equiv k_2 \pmod 2\)
:::
.
and
defined over
It is a well-known fact that there exists a rational \(3\)-isogeny
given by
This implies that
Hence, for determining
, it is enough to obtain
Let
and
where
denotes the primitive third root of unity. Observe that
and so, it follows that
and
are isomorphic over
Composing
with this isomorphism, we obtain a \(3\)-isogeny
defined over
given by
Further, let
be the set of all finite primes of
and
Note that
It follows from Corollary [\[corforrk\]](#corforrk){reference-type="ref" reference="corforrk"} that
(resp.
) if
(resp. if
). From these upper bounds on the Mordell-Weil rank of
over
along with the global root number, the parity conjecture implies Generalised Sylvester's Conjecture.\
Sylvester's Conjecture is a special case of the Generalised Sylvester's Conjecture for
In this paper, we study the case
that is, when
has exactly two distinct prime divisors both different from \(3\). In this special case, the Generalised Sylvester's Conjecture can be restated as follows:
and
be two distinct primes both different from \(3\). Assume that
is of the form either
or
or
We have
1. If
::: {.small}
\(n \not\equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(\text{rk } E_{-432n^2}(\Q) = \begin{cases} \{0, 2\}, & \text{ if } \ell_1\ell_2 \equiv 2 \pmod{3},\\ 1, & \text{ if } \ell_1 \equiv \ell_2 \equiv 2 \pmod{3},\\ \{1,3\}, & \text{ if } \ell_1 \equiv \ell_2 \equiv 1 \pmod{3}.\end{cases}\)
:::
2. If
::: {.small}
\(n \equiv \pm 1 \pmod{9}\),
:::
then
::: {.small}
\(\text{rk } E_{-432n^2}(\Q) = \begin{cases} \{1,3\}, & \text{ if } \ell_1\ell_2 \equiv 2 \pmod{3},\\ \{0, 2 \}, & \text{ if } \ell_1 \equiv \ell_2 \equiv 2 \pmod{3},\\ \{0, 2, 4 \}, & \text{ if } \ell_1 \equiv \ell_2 \equiv 1 \pmod{3}.\end{cases}\)
:::
-dimension of the
-Selmer group of
over
and prove that it has the same parity as
(resp.
) provided
(resp.
). Now a simple application of Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"} proves Conjecture [\[conjforrank2\]](#conjforrank2){reference-type="ref" reference="conjforrank2"} (assuming
is even). We remark that the Tate-Shafarevich conjecture predicts that
is finite, which in turn implies that
is even (see for a more detailed discussion). To compute
explicitly in all possible scenarios, we broadly divide them in three cases, namely
The computations for each of these cases are quite involved and challenging. From these computations we observe that
depends on equivalence classes of
modulo
, as well as on certain cubic residues (if
). See §[2](#sec2){reference-type="ref" reference="sec2"} for precise statements. Finally, Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"} provides us with bounds on
and we show that our bounds on
are sharp. See §[3](#rankcomp){reference-type="ref" reference="rankcomp"} for precise statements. One of the consequences of our explicit computation of the Selmer groups is that it leads us to obtain the following unconditional result:
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
We have
i.e.
is not a rational cube sum in the following cases:
1. ::: {.small}
\(n=\ell_1\ell_2\)
:::
or
::: {.small}
\(n=\ell_1^2\ell_2^2\)
:::
with
::: {.small}
\(\ell_1 \equiv 2 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 5 \pmod 9\),
:::
2. ::: {.small}
\(n=\ell_1^2\ell_2\)
:::
with
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 2,5 \pmod 9\)
:::
,
3. ::: {.small}
\(n=\ell_1\ell_2\)
:::
or
::: {.small}
\(n=\ell_1^2\ell_2\)
:::
with
::: {.small}
\(n \equiv 2,5 \pmod 9\)
:::
where
::: {.small}
\(\ell_1 \equiv 1 \pmod 3, \ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\)
:::
,
4. ::: {.small}
\(n=\ell_1^2\ell_2^2\)
:::
or
::: {.small}
\(n=\ell_1\ell_2^2\)
:::
with
::: {.small}
\(n \equiv 4,7 \pmod 9\)
:::
where
::: {.small}
\(\ell_1 \equiv 1 \pmod 3, \ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\)
:::
.
is even, we also provide a criterion for natural numbers
for which
(cf. Theorem [\[thmtosylvesterrk1\]](#thmtosylvesterrk1){reference-type="ref" reference="thmtosylvesterrk1"}) in terms of congruences modulo \(9\) and cubic residues. The article is structured as follows: In §[1](#prelims){reference-type="ref" reference="prelims"}, we discuss the basic set-up, recall some definitions and results from which are used subsequently. Further, we explore the relationship between dimension of
-Selmer group over
and rank of elliptic curve over
This is done in Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}, which is one of the key components of the paper. In §[2](#sec2){reference-type="ref" reference="sec2"}, we restrict ourselves to cube-free natural numbers co-prime to \(3\) with exactly two distinct prime divisors and give an explicit description of the
-Selmer group of
through computing its generators. Finally in §[3](#rankcomp){reference-type="ref" reference="rankcomp"}, using Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}, we relate
with
thereby comment about the solubility of
as a sum of two rational cubes.\
# Basic set up {#prelims}
First, we fix the notation that we will be using throughout this paper. Let
and
denote its ring of integers. Further, let
be the set of all finite primes of
and
denote the completion at
of
. If we set
, then
is the unique prime of
dividing \(3\). We denote the ideal
and its uniformizer
both by
itself. There is a norm map
given by
We denote the kernel of this norm map by
. The map
induces a norm map
We recall the following from:
The group
is defined similarly by replacing
with
, the ring of integers of
, in the above definition.
we define
we denote the ring of
-integers of
and define, in a similar fashion as in the Definition [\[defnofAq\]](#defnofAq){reference-type="ref" reference="defnofAq"}, the group
.\
Recall that
is the unique of
above \(3\). For the local field
and its uniformizer
, we set
Recall the following result from:
and
where
is defined
The map
given by
is then an isomorphism. The same map identifies
with
◻
and a prime
, the local Kummer map
, corresponding to the isogeny
in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}, is defined as follows: For
. Then one has
1. If
::: {.small}
\({\mathfrak{q}} \nmid 3\),
:::
then
::: {.small}
\(\text{Im } \delta_{\phi,K_{\mathfrak{q}}}= \big(A_{\mathfrak{q}}^*/A_{\mathfrak{q}}^{*3}\big)_{N=1}\)
:::
and
::: {.small}
\(|\text{Im } \delta_{\phi,K_{\mathfrak{q}}}|=3\).
:::
2. If
::: {.small}
\({\mathfrak{q}}={\mathfrak{p}} \mid 3\),
:::
then
::: {.small}
\(V_3/A_{\mathfrak{p}}^{*3} \subset \text{Im } \delta_{\phi,K_{\mathfrak{p}}} \subset \big(A_{\mathfrak{p}}^*/A_{\mathfrak{p}}^{*3}\big)_{N=1}\)
:::
and
::: {.small}
\(|\text{Im } \delta_{\phi,K_{\mathfrak{p}}}|=9\).
:::
On the other hand, for
, hence
. 0◻
:
be a cube-free integer co-prime to \(3\). We have
1. For
::: {.small}
\(n \not\equiv \pm 1 \pmod 9,\)
:::
::: {.small}
\(\text{Im } \delta_{\phi,K_{\mathfrak{p}}} = \big\langle(\overline{n}^2,\overline{n}), (\overline{(1+{\mathfrak{p}}^3)^2},\overline{(1+{\mathfrak{p}}^3)}) \big\rangle.\)
:::
Here,
::: {.small}
\(\text{Im } \delta_{\phi,K_{\mathfrak{p}}} \cong \frac{U^2_{K_{\mathfrak{p}}}}{U^4_{K_{\mathfrak{p}}}}\).
:::
As a consequence, we see that
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \notin \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\).
:::
2. For
::: {.small}
\(n \equiv \pm 1 \pmod 9,\)
:::
::: {.small}
\(\text{Im }\delta_{\phi,K_{\mathfrak{p}}}=\big\langle(\overline{\zeta^2},\overline{\zeta}), (\overline{(1+{\mathfrak{p}}^3)^2},\overline{(1+{\mathfrak{p}}^3)}) \big\rangle.\)
:::
In this situation,\
::: {.small}
\(\text{Im } \delta_{\phi,K_{\mathfrak{p}}} \cap f\big(\frac{U^2_{K_{\mathfrak{p}}}}{U^4_{K_{\mathfrak{p}}}}\big) \cong \frac{U^3_{K_{\mathfrak{p}}}}{U^4_{K_{\mathfrak{p}}}}\),
:::
where
::: {.small}
\(f\)
:::
is the isomorphism from Lemma [\[V3U1U4\]](#V3U1U4){reference-type="ref" reference="V3U1U4"}.
and so
but
Note that
and
Since
and
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}, we get
It is then easy to see that
\(2\) Now if
, then
Hence, there exists
such that
Thus, the point
Observe that
Now
as
for some
Consequently,
As
and
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}, we see that
Note that
so
◻
, we define the
Selmer group of
as a subgroup of
, which we denote by
as follows (cf. for details):
Using the following result, one obtains a trivial bound on
:
be an elliptic curve and
be the \(3\)-isogeny over
in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}. Then
. As a consequence we obtain that
with
:
be a cube-free natural number co-prime to \(3\) and
be the set of all distinct primes dividing
Further, assume that
for
and
for
Then
for
remain inert in
whereas the primes
for
split in
say
(One can make an explicit choice of
(resp.
) as follows: Write
and define
and
) Therefore,
and hence
So, it follows from Lemma [\[selmerbound\]](#selmerbound){reference-type="ref" reference="selmerbound"} that
Now recall from that
Thus,
Note that
, so it follows that
\
In particular, if
then we claim that
and hence
in this situation. Observe that the group
where
denotes the set of
-units in
has
dimension
by Dirichlet's \(S\)-units theorem. In particular,
On the other hand, note from Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"} that
whenever
Consequently,
, so
Thus, one obtains
◻
be an elliptic curve and
the \(3\)-isogeny over
defined in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}. If
then
\
Moreover, if
is even, then
from [\[eq:expforrk\]](#eq:expforrk){reference-type="eqref" reference="eq:expforrk"}. So if
then
Recall that
since
On the other hand, consider the rational \(3\)-isogeny
defined in the Introduction and its dual isogeny
. Note that
Set
. Then by we have the following exact sequence
So,
since
has no non-trivial \(3\)-torsion points. Now, we have
by. Therefore, we obtain
. Then the two expressions for
composedly imply
We know that
is even by. Now if we assume
is even, then
◻
that are co-prime to \(3\) and have exactly two distinct prime factors, that is
is of the form
or
or
where
and
are two distinct primes. In the language of the Generalised Sylvester's Conjecture, we are studying the case
We consider the cases
and
separately in three different subsections. In each of the three subsections, we obtain the
-dimensions of
for integers
where
and
are two distinct primes both different from \(3\). However, we furnish the proofs only for
to avoid redundancy of literature, as the proofs for other types of
are similar in nature.
## The Case \(k_1=0\), i.e., \(\ell_1 \equiv \ell_2 \equiv 2 \pmod 3\)
In this subsection, we assume that
We have the following results regarding
be distinct primes and
is either
or
. Let
be the \(3\)-isogeny over
given in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}. Then
Observe that
By Lemma [\[selmerbound\]](#selmerbound){reference-type="ref" reference="selmerbound"}, we have
and
The group of
-units in
is
Hence,
Note that
and
for all
Hence,
Thus,
We now examine the remaining generators of
for their containment in
First, we discuss the generator
. For
we know that
and
. Then it follows that
for
Hence,
belongs to both
and
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Also
by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(2) in this situation. For primes
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. So
, when
On the other hand, if
then we know that
and
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Hence,
and so
when
\
We now investigate the remaining generator
of
Note that for primes
one has
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Thus, we only need to check whether
for
Note that the cubic residue
as both
and
are primes congruent to
Hence,
and so
Similarly using
we get
Thus, it reduces to check whether
Let us first take
In this case,
i.e.
Since
by Lemmas [\[V3U1U4\]](#V3U1U4){reference-type="ref" reference="V3U1U4"} and [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}, it follows that
Hence,
and
when
Next, let
and notice that
i.e.
Hence by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(1),
and it follows that
So,
whenever
Finally for
and
observe that
This implies by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(2) that
and hence
There are four subgroups of
of order \(3\) corresponding to the elements
and
To conclude that
we need to show that neither
nor
are contained in
Recall that
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Using
we see that
Since
we have
and so it is not in
A similar argument works for
This shows that
when
and
This completes the proof for
◻
be distinct primes and
. For the \(3\)-isogeny
over
given in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}, we have
co-prime to \(3\), all of whose prime divisors
are congruent to \(2\) modulo \(3\). In this situation we have,
To see this, observe that if
then
whereas
for each
If
then
and for every
Finally, if
and if some of the
writing
one checks that
if and only if
and
An easy combinatorial argument shows that there are
such elements.
is congruent to
while its other prime divisor is congruent to
For simplicity of notation, we assume that
and
. Note that
splits as
in
, while
remains inert in
We denote by
the cubic residue symbol. Note that
if and only if
is a cube in
be an integer of the form either
or
or
or
, where
and
are primes such that
For the \(3\)-isogeny
defined in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}, we have
and mention that the proof in other cases for
given above can be obtained via a similar computation. We have
and by Lemma [\[selmerbound\]](#selmerbound){reference-type="ref" reference="selmerbound"},
and
The group of
-units in
is
Hence,
Note that
and
by the Definition [\[eq:kummap\]](#eq:kummap){reference-type="eqref" reference="eq:kummap"} of the local Kummer maps. Hence,
If
then we know that
and
by Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Hence,
On the other hand, if
then
So by a similar argument
Thus in either cases, we get that
and so
We now investigate the remaining generators
and
case by case. Recall that one can write
and
.
1. Let us first consider the case
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3= 1\)
:::
and investigate the generator
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})\)
:::
of
::: {.small}
\(\Big(\frac{A_{S_{n}}^*}{A_{S_{n}}^{*3}}\Big)_{N=1}\).
:::
Note that for primes
::: {.small}
\({\mathfrak{q}} \notin S_{n} \cup \{{\mathfrak{p}}\}\), \((\overline{\ell_2^2},\overline{\ell_2}) \in \text{Im } \delta_{\phi,K_{\mathfrak{q}}}\)
:::
due to Lemma [\[imgofkummap\]](#imgofkummap){reference-type="ref" reference="imgofkummap"}. Thus, we only need to check whether or not
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2}) \in \text{Im } \delta_{\phi,K_{\mathfrak{q}}}\)
:::
for
::: {.small}
\({\mathfrak{q}} \in S_{n} \cup \{{\mathfrak{p}}\}\).
:::
Notice that the cubic residue
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3=1\)
:::
implies
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})=(\overline{1},\overline{1}) \in \big({A_{\pi_{\ell_1}}^*}/{A_{\pi_{\ell_1}}^{*3}}\big)_{N=1}\).
:::
So, it is in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\).
:::
Similarly,
::: {.small}
\(\big(\frac{\ell_2}{\pi'_{\ell_1}}\big)_3=1\)
:::
gives
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2}) \in \text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Now using
::: {.small}
\(\big(\frac{\ell_1}{{\ell_2}}\big)_3=1\),
:::
one gets
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})=(\overline{n}^2,\overline{n}) \in \text{Im } \delta_{\phi,K_{\ell_2}}\).
:::
Finally,
::: {.small}
\(\ell_2 \equiv 2 \pmod 3\)
:::
implies
::: {.small}
\(\ell_2^2 \in U^2_{K_{\mathfrak{p}}}\)
:::
. So by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(1),
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2}) \in \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\)
:::
and hence
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2}) \in {\rm Sel}^\phi(E_{16n^2}/K).\)
:::
Next, we draw our attention to the generator
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})\)
:::
of
::: {.small}
\(\Big(\frac{A_{S_{n}}^*}{A_{S_{n}}^{*3}}\Big)_{N=1}\)
:::
and note that the cubic residue
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi'_{\ell_1}}\big)_3=1\)
:::
implies
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})=(\overline{1},\overline{1}) \in \big({A_{\pi'_{\ell_1}}^*}/{A_{\pi'_{\ell_1}}^{*3}}\big)_{N=1}\).
:::
This gives us
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \in \text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Similarly,
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\ell_2}\big)_3=1\)
:::
gives
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \in \text{Im } \delta_{\phi,K_{{\ell_2}}}\).
:::
Additionally, one has
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})=(\overline{n}^2,\overline{n}) \in \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\)
:::
due to the assumption
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3= 1\).
:::
Finally, writing
::: {.small}
\(\pi_{\ell_1}=\frac{1}{2}(L+3\sqrt{-3}M)\)
:::
, we have
::: {.small}
\(\pi_{\ell_1}^2 \in U^2_{K_{\mathfrak{p}}}\)
:::
which implies
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \in \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\)
:::
by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(1). Hence,
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\)
:::
and we get
::: {.small}
\(\dim_{{\mathbb{F}}_3}{\rm Sel}^\phi(E_{16n^2}/K)=3\)
:::
when
::: {.small}
\(n=\ell_1\ell_2 \not\equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3= 1\).
:::
\
2. We now consider the case
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\).
:::
There are \(13\) subgroups of order \(3\) of
::: {.small}
\(\Big(\frac{A_{S_{n_{ }}}^*}{A_{S_{n}}^{*3^{ }}}\Big)_{N=1} \Big/ \big\langle(\overline{n}^2,\overline{n})\big\rangle\)
:::
corresponding to the elements:
::: {.small}
\(\zeta\), \(\ell_2\), \(\pi_{\ell_1},\)
:::
::: {.small}
\(\zeta\ell_2\), \(\zeta\ell_2^2\), \(\zeta\pi_{\ell_1}\), \(\zeta\pi_{\ell_1}^2\), \(\ell_2\pi_{\ell_1}\), \(\ell_2^2\pi_{\ell_1}\), \(\zeta\ell_2\pi_{\ell_1},\)
:::
::: {.small}
\(\zeta^2\ell_2\pi_{\ell_1}\), \(\zeta\ell_2^2\pi_{\ell_1}\)
:::
and
::: {.small}
\(\zeta\ell_2\pi_{\ell_1}^2\).
:::
Thus, to arrive at the conclusion that
::: {.small}
\(\dim_{{\mathbb{F}}_3}{\rm Sel}^\phi(E_{16n^2}/K)=1\),
:::
one requires to show that none of these \(13\) subgroups is contained in
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\).
:::
The fact that
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \notin {\rm Sel}^\phi(E_{16n^2}/K)\)
:::
is already discussed above. One can easily see that
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \neq (\overline{1},\overline{1}) \in \big(A_{\ell_2}^*/A_{\ell_2}^{*3}\big)_{N=1}\),
:::
since
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3=\big(\frac{\pi_{\ell_1}}{{\ell_2}}\big)_3 \neq 1\)
:::
by assumption. Thus,
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{\ell_2}}\).
:::
Similar arguments using
::: {.small}
\(\big(\frac{\ell_2}{\pi'_{\ell_1}}\big)_3 \neq 1\)
:::
and
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi'_{\ell_1}}\big)_3=1\)
:::
prove that
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})\), \((\overline{\ell_2^2\pi_{\ell_1}^2},\overline{\ell_2\pi_{\ell_1}})\)
:::
and
::: {.small}
\((\overline{\ell_2\pi_{\ell_1}^2},\overline{\ell_2^2\pi_{\ell_1}})\)
:::
are all not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
For all the remaining elements, one can see that they are not in
::: {.small}
\({U^2_{K_{\mathfrak{p}}}}/{U^4_{K_{\mathfrak{p}}}}.\)
:::
Hence, they do not belong to
::: {.small}
\(\text{Im } \delta_{\phi,K_{{\mathfrak{p}}}}\)
:::
by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(1). Therefore, we deduce that
::: {.small}
\(\dim_{{\mathbb{F}}_3}{\rm Sel}^\phi(E_{16n^2}/K)=1\).
:::
0◻
be an integer of the form either
or
or
or
, where
and
are primes such that
For the \(3\)-isogeny
defined in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}, we have
and by Lemma [\[selmerbound\]](#selmerbound){reference-type="ref" reference="selmerbound"},
and
As in Prop. [\[sylvesterpq2\]](#sylvesterpq2){reference-type="ref" reference="sylvesterpq2"}, we have
and
for all
Hence,
and
1. Let us first study the case
::: {.small}
\(\ell_1 \equiv 1 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 8 \pmod 9\).
:::
Note that
::: {.small}
\(\zeta_9\)
:::
belongs to each of
::: {.small}
\({\mathcal{O}}_{K_{\ell_2}}^*\), \({\mathcal{O}}_{K_{\pi_{\ell_1}}}^*\)
:::
and
::: {.small}
\({\mathcal{O}}_{K_{\pi'_{\ell_1}}}^*\).
:::
Hence,
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\)
:::
belongs to each of
::: {.small}
\(\text{Im } \delta_{\phi,K_{\ell_2}}\), \(\text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\)
:::
and
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Also by Lemma [\[imgofdeltaatp\]](#imgofdeltaatp){reference-type="ref" reference="imgofdeltaatp"}(2), we know that
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \in \text{Im } \delta_{\phi,K_{{\mathfrak{p}}}}\).
:::
Thus,
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \in {\rm Sel}^\phi(E_{16n^2}/K)\)
:::
in this situation. Now if
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3=1\)
:::
, then giving a similar argument as in the proof of (1) in Prop. [\[sylvesterpq2\]](#sylvesterpq2){reference-type="ref" reference="sylvesterpq2"}, we obtain that both
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})\)
:::
and
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})\)
:::
belong to
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Hence,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=4\)
:::
when
::: {.small}
\(\ell_1 \equiv 1 \pmod 9\), \(\ell_2 \equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3=1.\)
:::
Next, let
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\).
:::
Then there are four subgroups of
::: {.small}
\(\Big(\frac{A_{S_{n}}^*}{A_{S_{n}}^{*3}}\Big)_{N=1} \Big/ \big\langle(\overline{n}^2,\overline{n}), (\overline{\zeta^2},\overline{\zeta})\big\rangle\)
:::
of order \(3\) corresponding to the elements
::: {.small}
\(\ell_2,\) \(\pi_{\ell_1},\) \(\ell_2\pi_{\ell_1}\)
:::
and
::: {.small}
\(\ell_2\pi_{\ell_1}^2\)
:::
. To conclude that
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 2\),
:::
we need to show that
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\)
:::
does not contain any of these subgroups. As
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\)
:::
by assumption, we get that
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\).
:::
Similarly,
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3=\big(\frac{\pi_{\ell_1}}{\ell_2}\big)_3\)
:::
implies
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{{\ell_2}}}\).
:::
Next, using
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi'_{\ell_1}}\big)_3 = 1\)
:::
we identify
::: {.small}
\((\overline{\ell_2^2\pi_{\ell_1}^2},\overline{\ell_2\pi_{\ell_1}})\)
:::
with
::: {.small}
\((\overline{{\ell_2}^2},\overline{{\ell_2}})\)
:::
in
::: {.small}
\(\big({A_{\pi'_{\ell_1}}^*}/{A_{\pi'_{\ell_1}}^{*3}}\big)_{N=1}.\)
:::
Then
::: {.small}
\((\overline{{\ell_2}^2},\overline{{\ell_2}}) \neq (\overline{1},\overline{1})\)
:::
in
::: {.small}
\(\big({A_{\pi'_{\ell_1}}^*}/{A_{\pi'_{\ell_1}}^{*3}}\big)_{N=1}\)
:::
gives
::: {.small}
\((\overline{{\ell_2}^2\pi_{\ell_1}^2},\overline{{\ell_2}\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
A similar argument shows that
::: {.small}
\((\overline{\ell_2^2\pi_{\ell_1}},\overline{\ell_2\pi_{\ell_1}^2}) \notin \text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}.\)
:::
Thus one can deduce that
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) =2\),
:::
when
::: {.small}
\(\ell_1 \equiv 1 \pmod 9\),
:::
::: {.small}
\(\ell_2 \equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\).
:::
\
2. Let
::: {.small}
\(\ell_1 \equiv 4 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 2 \pmod 9\).
:::
Proving that
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) =2\)
:::
involves checking of the \(13\) subgroups mentioned in the proof of (2) in Prop. [\[sylvesterpq2\]](#sylvesterpq2){reference-type="ref" reference="sylvesterpq2"} for their containment in
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\).
:::
We observe the following:
1. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 = 1,\)
:::
then
::: {.small}
\((\overline{\ell_2^2\pi_{\ell_1}},\overline{\ell_2\pi_{\ell_1}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
The elements
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}^2},\overline{\zeta\ell_2\pi_{\ell_1}}),\)
:::
::: {.small}
\((\overline{\zeta\ell_2^2\pi_{\ell_1}^2},\overline{\zeta^2\ell_2\pi_{\ell_1}})\), \((\overline{\zeta^2\ell_2\pi_{\ell_1}^2},\overline{\zeta\ell_2^2\pi_{\ell_1}})\)
:::
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}},\overline{\zeta\ell_2\pi_{\ell_1}^2})\)
:::
and
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\)
:::
are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\),
:::
whereas all the other elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{{\mathfrak{p}}}}\).
:::
So,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
2. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 =\zeta,\)
:::
then
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}},\overline{\zeta\ell_2\pi_{\ell_1}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Note that
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{{\ell_2}}}\)
:::
and
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}^2},\overline{\zeta\ell_2\pi_{\ell_1}})\), \((\overline{\zeta^2\ell_2^2},\overline{\zeta\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\),
:::
while all the other elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
3. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 =\zeta^2,\)
:::
then
::: {.small}
\((\overline{\zeta^2\ell_2\pi_{\ell_1}^2},\overline{\zeta\ell_2^2\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Notice that
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{{\ell_2}}}\)
:::
and
::: {.small}
\((\overline{\zeta\ell_2^2\pi_{\ell_1}^2},\overline{\zeta^2\ell_2\pi_{\ell_1}})\), \((\overline{\zeta\ell_2^2},\overline{\zeta^2\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\),
:::
whereas all the remaining elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
\
3. Let
::: {.small}
\(\ell_1 \equiv 7 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 5 \pmod 9\).
:::
Checking the \(13\) subgroups mentioned in the proof of (2) in Prop. [\[sylvesterpq2\]](#sylvesterpq2){reference-type="ref" reference="sylvesterpq2"} for their containment in
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\)
:::
yields:
1. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 = 1,\)
:::
then
::: {.small}
\((\overline{\ell_2^2\pi_{\ell_1}},\overline{\ell_2\pi_{\ell_1}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Note that
::: {.small}
\((\overline{\zeta^2\ell_2^2},\overline{\zeta\ell_2})\),\((\overline{\zeta\ell_2^2},\overline{\zeta^2\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\)
:::
and
::: {.small}
\((\overline{\ell_2^2},\overline{\ell_2})\), \((\overline{\pi_{\ell_1}^2},\overline{\pi_{\ell_1}})\), \((\overline{\ell_2^2\pi_{\ell_1}^2},\overline{\ell_2\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\),
:::
whereas all the remaining elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
2. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 =\zeta,\)
:::
then
::: {.small}
\((\overline{\zeta^2\ell_2\pi_{\ell_1}^2},\overline{\zeta\ell_2^2\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Observe that
::: {.small}
\((\overline{\pi_{\ell_1}^2},\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{{\ell_2}}}\),
:::
::: {.small}
\((\overline{\zeta\ell_2^2\pi_{\ell_1}^2},\overline{\zeta^2\ell_2\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\)
:::
and
::: {.small}
\((\overline{\zeta\ell_2^2},\overline{\zeta^2\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\),
:::
whereas all the other elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
3. If
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 =\zeta^2,\)
:::
then
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}},\overline{\zeta\ell_2\pi_{\ell_1}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Notice that
::: {.small}
\((\overline{\pi_{\ell_1}^2},\overline{\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{{\ell_2}}}\),
:::
::: {.small}
\((\overline{\zeta^2\ell_2^2\pi_{\ell_1}^2},\overline{\zeta\ell_2\pi_{\ell_1}}) \notin \text{Im } \delta_{\phi,K_{\mathfrak{p}}}\)
:::
and
::: {.small}
\((\overline{\zeta^2\ell_2^2},\overline{\zeta\ell_2}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_1}}}\),
:::
while all the remaining elements are not in
::: {.small}
\(\text{Im } \delta_{\phi,K_{\pi'_{\ell_1}}}\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K)=2\).
:::
This shows that
in all the cases, except when
in which case it is \(4\). This completes the proof for
. ◻
and
are primes congruent to
Let
split as
for
in
and denote by
the cubic residue symbol.
be an integer of the form
or
or
where
are distinct primes and
be the \(3\)-isogeny defined in [\[eq:defofphi\]](#eq:defofphi){reference-type="eqref" reference="eq:defofphi"}. Then
1. If
::: {.small}
\(n \equiv 1 \pmod 9\),
:::
then
::: {.small}
\[\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) =\begin{cases} 5, & \text{if } \ell_1 \equiv \ell_2 \equiv 1 \pmod 9,\text{ } \big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1 \text{ and } \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1, \\ 3, & \text{otherwise.} \end{cases}\]
:::
2. If
::: {.small}
\(n \equiv 4 \text{ or } 7 \pmod 9\),
:::
then
::: {.small}
\[\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) =\begin{cases} 4, & \text{if } \big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1 \text{ and } \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1, \\ 2, & \text{otherwise.} \end{cases}\]
:::
Here
and the group
Hence, we have
which implies by Lemma [\[selmerbound\]](#selmerbound){reference-type="ref" reference="selmerbound"} that
As before, we observe that the element
in every case. In the cases where we claim
is \(2\) or \(3\), in order to determine
precisely, one has to investigate the \(40\) subgroups of order \(3\) of
corresponding to the following elements:
\
A similar but much tedious computation as in Props. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}-[\[sylvesterpq3\]](#sylvesterpq3){reference-type="ref" reference="sylvesterpq3"} shows that
1. All of
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\), \((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})\), \((\overline{\pi'^{2}_{\ell_1}},\overline{\pi'_{\ell_1}})\), \((\overline{\pi_{\ell_2}}^2,\overline{\pi_{\ell_2}})\)
:::
belong to
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\),
:::
when
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 9\)
:::
and both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
are \(1\). Hence,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 5\)
:::
here.\
2. Let
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 9\)
:::
such that at least one of
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is not \(1\). We see that apart from
::: {.small}
\((\overline{n}^2, \overline{n})\)
:::
, out of the \(40\) subgroups stated above, only the subgroups generated by elements mentioned below are contained in the
::: {.small}
\(\phi\)-
:::
Selmer group:
1. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 \neq 1\),
:::
then
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\)
:::
,
::: {.small}
\((\overline{\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\pi'_{\ell_1}\pi_{\ell_2}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
2. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta \text{ or } \zeta^2\),
:::
then
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\)
:::
,
::: {.small}
\((\overline{{\ell_1}\pi'_{\ell_1}},\overline{{\ell_1}\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
3. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta^2\),
:::
then
::: {.small}
\((\overline{\zeta^2},\overline{\zeta})\)
:::
,
::: {.small}
\((\overline{{\ell_1}^2\pi_{\ell_2}},\overline{{\ell_1}\pi_{\ell_2}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 3\)
:::
in these cases.\
3. Next, let
::: {.small}
\(\ell_1 \equiv 4 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 7 \pmod 9\).
:::
Then the elements given below, along with
::: {.small}
\((\overline{n}^2, \overline{n})\)
:::
, generate
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\):
:::
1. If both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\),
:::
then
::: {.small}
\((\overline{\ell_1 \pi'_{\ell_1}},\overline{\ell_1\pi_{\ell_1}})\)
:::
,
::: {.small}
\((\overline{\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\pi'_{\ell_1}\pi_{\ell_2}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
2. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta\),
:::
then
::: {.small}
\((\overline{\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\pi'_{\ell_1}\pi_{\ell_2}})\)
:::
,
::: {.small}
\((\overline{\zeta^2{\ell_1}\pi'_{\ell_1}},\overline{\zeta{\ell_1}\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
3. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta^2\),
:::
then
::: {.small}
\((\overline{\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\pi'_{\ell_1}\pi_{\ell_2}})\)
:::
,
::: {.small}
\((\overline{\zeta^2{\ell_1}\pi_{\ell_1}},\overline{\zeta{\ell_1}\pi'_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
4. If both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta\),
:::
then
::: {.small}
\((\overline{{\ell_1}\pi'_{\ell_1}},\overline{{\ell_1}\pi_{\ell_1}})\),
:::
::: {.small}
\((\overline{\zeta^2\pi'_{\ell_1}\pi_{\ell_2}},\overline{\zeta\pi'^{2}_{\ell_1}\pi_{\ell_2}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
5. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta^2\),
:::
then
::: {.small}
\((\overline{{\ell_1}^2\pi_{\ell_2}},\overline{{\ell_1}\pi_{\ell_2}^2}),\)
:::
::: {.small}
\((\overline{\zeta^2{\ell_1}\pi'_{\ell_1}},\overline{\zeta{\ell_1}\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
6. If both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta^2\),
:::
then
::: {.small}
\((\overline{{\ell_1}\pi'_{\ell_1}},\overline{{\ell_1}\pi_{\ell_1}})\),
:::
::: {.small}
\((\overline{\zeta^2\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\zeta\pi'_{\ell_1}\pi_{\ell_2}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Hence,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 3\)
:::
in all of these cases.\
4. When
::: {.small}
\(n=\ell_1\ell_2 \not\equiv 1 \pmod 9,\)
:::
then at least one of
::: {.small}
\(\ell_1\)
:::
and
::: {.small}
\(\ell_2\)
:::
is not
::: {.small}
\(1 \pmod 9.\)
:::
We know that
::: {.small}
\(\zeta_9 \notin K_{\pi_{\ell_i}}^*\)
:::
whenever
::: {.small}
\(\ell_i \not\equiv \pm 1 \pmod 9.\)
:::
Hence, we get that
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \notin \text{Im } \delta_{\phi,K_{\pi_{\ell_i}}}\)
:::
and so
::: {.small}
\((\overline{\zeta^2},\overline{\zeta}) \notin {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
We also obtain that all of
::: {.small}
\((\overline{\pi_{\ell_1}}^2,\overline{\pi_{\ell_1}})\), \((\overline{\pi'^{2}_{\ell_1}},\overline{\pi'_{\ell_1}})\), \((\overline{\pi_{\ell_2}}^2,\overline{\pi_{\ell_2}})\)
:::
belong to
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\),
:::
if
::: {.small}
\(n \not\equiv 1 \pmod 9\)
:::
and each of
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is \(1\). Hence,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 4\)
:::
in this case.\
5. Finally, when
::: {.small}
\(n \not\equiv 1 \pmod 9\)
:::
and at least one of
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is not \(1\), we obtain that
::: {.small}
\({\rm Sel}^\phi(E_{16n^2}/K)\)
:::
is generated by
::: {.small}
\((\overline{n}^2, \overline{n})\)
:::
and the elements given below:
1. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 \neq 1\),
:::
then
::: {.small}
\((\overline{\pi'^{2}_{\ell_1}\pi_{\ell_2}^2},\overline{\pi'_{\ell_1}\pi_{\ell_2}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
2. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta \text{ or } \zeta^2\),
:::
then
::: {.small}
\((\overline{{\ell_1}\pi'_{\ell_1}},\overline{{\ell_1}\pi_{\ell_1}}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
3. If
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = \zeta^2\),
:::
then
::: {.small}
\((\overline{{\ell_1}^2\pi_{\ell_2}},\overline{{\ell_1}\pi_{\ell_2}^2}) \in {\rm Sel}^\phi(E_{16n^2}/K)\).
:::
Thus,
::: {.small}
\(\dim_{{\mathbb{F}}_3} {\rm Sel}^\phi(E_{16n^2}/K) = 2\)
:::
in all of these cases. 0◻
# Computation of the rank of the elliptic curve \(E_{16n^2}/\Q\) {#rankcomp}
We are now ready to obtain results on the Mordell-Weil rank of the curves
over
We use Props. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}-[\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"} and Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"} to obtain the results stated in Conjecture [\[conjforrank2\]](#conjforrank2){reference-type="ref" reference="conjforrank2"}. Recall that
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
We have
i.e.
is not a rational cube sum in the following cases:
1. ::: {.small}
\(n=\ell_1\ell_2\)
:::
or
::: {.small}
\(n=\ell_1^2\ell_2^2\)
:::
with
::: {.small}
\(\ell_1 \equiv 2 \pmod 9\)
:::
and
::: {.small}
\(\ell_2 \equiv 5 \pmod 9\),
:::
2. ::: {.small}
\(n=\ell_1^2\ell_2\)
:::
with
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 2,5 \pmod 9\)
:::
,
3. ::: {.small}
\(n=\ell_1\ell_2\)
:::
or
::: {.small}
\(n=\ell_1^2\ell_2\)
:::
with
::: {.small}
\(n \equiv 2,5 \pmod 9\)
:::
where
::: {.small}
\(\ell_1 \equiv 1 \pmod 3,\) \(\ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\)
:::
,
4. ::: {.small}
\(n=\ell_1^2\ell_2^2\)
:::
or
::: {.small}
\(n=\ell_1\ell_2^2\)
:::
with
::: {.small}
\(n \equiv 4,7 \pmod 9\)
:::
where
::: {.small}
\(\ell_1 \equiv 1 \pmod 3,\) \(\ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\)
:::
.
if
satisfies any of the conditions (1) to (4) above. Then Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"} implies
◻
is even. The following results are then immediate consequences of Props. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}-[\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"} and Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}:
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
Then we have
in the following cases.
1. ::: {.small}
\(n \not\equiv \pm 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 2 \pmod 3\),
:::
2. ::: {.small}
\(n \equiv \pm 1 \pmod 9\), \(\ell_1 \equiv 1 \pmod 9\), \(\ell_2 \equiv 8 \pmod 9\)
:::
but
::: {.small}
\(\ell_2\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\),
:::
3. ::: {.small}
\(n \equiv \pm 1 \pmod 9\), \(\ell_2 \equiv 2 \pmod 3\), \(\ell_1 \equiv 4 \text{ or } 7 \pmod 9\),
:::
4. ::: {.small}
\(n \not\equiv \pm 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 1 \pmod 3\)
:::
and at least one among
::: {.small}
\(\pi_{\ell_1}, \pi_{\ell_1}'\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_2}\).
:::
Further, if we assume that
is even in the cases above, then
i.e.
is a sum of two rational cubes.
if
satisfies any one of the conditions (1) to (4) stated above, from Props. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}-[\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"}. The statement follows from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
whereas these theorems allow us to find
precisely. We establish this claim in the examples below: Let
and
or
. Then by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(1), we get that
Whereas using, one can only obtain
. Let
and
Then for
, by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(2), we get that
Whereas, only predicts that
. Let
and
Then for
, by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(2), we get that
Whereas, only predicts that
. Let
and
Then it is easy to see that
is not a cube in
So for
or
, by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(3), we get that
On the other hand, only provides the bounds
. Let
and
Then one can see that
is not a cube in
Hence for
or
, by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(3), we get that
On the other hand, only provides the bounds
. Let
and
Then it can be easily seen that
is not a cube in
Hence for
or
, by Theorem [\[thmtosylvester\]](#thmtosylvester){reference-type="ref" reference="thmtosylvester"}(4), we get that
On the other hand, only provides the bounds
. Let
and
Then for
we get
by Theorem [\[thmtosylvesterrk1\]](#thmtosylvesterrk1){reference-type="ref" reference="thmtosylvesterrk1"}(1). On the other hand, we obtain
using, when
Hence, we deduce that
Let
and
We obtain
using. One can see that
is not a cube in
, so by Theorem [\[thmtosylvesterrk1\]](#thmtosylvesterrk1){reference-type="ref" reference="thmtosylvesterrk1"}(2) we have
and hence we conclude that
. Let
and
Then for
we get
by Theorem [\[thmtosylvesterrk1\]](#thmtosylvesterrk1){reference-type="ref" reference="thmtosylvesterrk1"}(3). On the other hand, we get
using, when
Thus,
Let
and
Let
One can see that
is not a cube in
, which implies at least one of
is not a cube in
. Thus by Theorem [\[thmtosylvesterrk1\]](#thmtosylvesterrk1){reference-type="ref" reference="thmtosylvesterrk1"}(4), we have
. We obtain
using and hence, we conclude that
.
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
In the following cases, we have
1. ::: {.small}
\(n \equiv \pm 1 \pmod 9\), \(\ell_1 \equiv 1 \pmod 9\), \(\ell_2 \equiv 8 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\),
:::
2. ::: {.small}
\(n \not\equiv \pm 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 1 \pmod 3\)
:::
and both
::: {.small}
\(\pi_{\ell_1}, \pi_{\ell_1}'\)
:::
are cubes in
::: {.small}
\({\mathbb{F}}_{\ell_2}\).
:::
Further, if we assume that
is even in the above cases, then we have
i.e.
is a rational cube sum.
satisfies any one of the conditions (1) or (2) written above, then we know that
from Props. [\[sylvesterpq3\]](#sylvesterpq3){reference-type="ref" reference="sylvesterpq3"}-[\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"}. The statement follows from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
and
then we see that
is a cube in
. For
we have
So, by Theorem [\[thmtosylvesterrk3\]](#thmtosylvesterrk3){reference-type="ref" reference="thmtosylvesterrk3"}(1), it then follows that
. Further, if we assume that
is even, then we have
. Using, we obtain
and
for
Let
and
then
and
. Notice that
is a cube in
. Next observe that there exists an
such that
. Writing
, we see that
. As a consequence, we see that both
are cubes in
. From Theorem [\[thmtosylvesterrk3\]](#thmtosylvesterrk3){reference-type="ref" reference="thmtosylvesterrk3"}(2), it then follows that for
or
, we have
. Further, if we assume that
is even, then we have
. Using, we obtain
and
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
In the following cases we have
1. ::: {.small}
\(n \equiv \pm 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 8 \pmod 9\),
:::
2. ::: {.small}
\(n \not\equiv \pm 1 \pmod 9\), \(\ell_1 \equiv 1 \pmod 3\), \(\ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
is a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\),
:::
3. ::: {.small}
\(n \equiv 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 1 \pmod 3\)
:::
and
::: {.small}
\(\ell_1 \equiv 4 \text{ or } 7 \pmod 9\),
:::
4. ::: {.small}
\(n \equiv 1 \pmod 9\), \(\ell_1 \equiv \ell_2 \equiv 1 \pmod 9\)
:::
and at least one of
::: {.small}
\(\pi_{\ell_1}, \pi_{\ell_1}'\)
:::
is not a cube in
::: {.small}
\({\mathbb{F}}_{\ell_2}\).
:::
Moreover, under the assumption that
is even in the cases above, we have
is either \(0\) or \(2\).
in all the above-mentioned cases. The statement follows from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
and
we have
and
by Theorem [\[thmtosylvesterrk2\]](#thmtosylvesterrk2){reference-type="ref" reference="thmtosylvesterrk2"}(1). Further, assuming
is even, we get
. gives
and
for
. Take
. Then
and
is a cube in
For
we have
by Theorem [\[thmtosylvesterrk2\]](#thmtosylvesterrk2){reference-type="ref" reference="thmtosylvesterrk2"}(2). Further, assuming
is even, we get
. Using,
and
for
. Take
and
Then we have
by Theorem [\[thmtosylvesterrk2\]](#thmtosylvesterrk2){reference-type="ref" reference="thmtosylvesterrk2"}(3). Further, the assumption that
is even implies
. Using,
and
. Consider
. Since
is not a cube modulo
, it follows that at least one of
is not a cube in
Thus by [\[thmtosylvesterrk2\]](#thmtosylvesterrk2){reference-type="ref" reference="thmtosylvesterrk2"}(4), for
and
, we have
Further, if we assume that
is even, then we have
. From, we see that
be a cube-free integer co-prime to \(3\) having exactly two distinct prime factors
and
If
and both
is are cubes in
then
Further, assume that
is even. Then we have
in the above mentioned case. The statement follows from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
. Notice that
is a cube in
. Next observe that there exists an
such that
. Writing
, we see that
. As a consequence, both
are cubes in
. From Theorem [\[thmtosylvesterrk4\]](#thmtosylvesterrk4){reference-type="ref" reference="thmtosylvesterrk4"}, we have
for
. Further, if we assume that
is even, then we have
. From, we see that
Sylvester's conjecture is unable to pinpoint which integers
of the form
or
are rational cube sum. This question was studied by. The Birch and Swinnerton-Dyer conjecture predicts that
where \(*\) is an non-zero number (depending on
) and
if and only if
is a rational cube sum (note that if
is not a rational cube sum, then one expects that
and
which is expected to be a square integer). provided efficient methods to numerically test for which of these integers
of the form
or
with
In a similar fashion, one can ask: Let
be a cube-free integer co-prime to \(3\) with exactly two distinct prime factors
and
Can you provide an efficient method to numerically test which of the following integers
are rational cube sum?
- ::: {.small}
\(n \equiv 1 \pmod 9\)
:::
with
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 3\),
:::
- ::: {.small}
\(n \equiv \pm 1 \pmod 9\)
:::
with
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 8 \pmod 9\),
:::
- ::: {.small}
\(n \not\equiv \pm 1 \pmod 9\)
:::
with
::: {.small}
\(\ell_1 \equiv-\ell_2 \equiv 1 \pmod 3\)
:::
and
::: {.small}
\(\ell_2\)
:::
a cube in
::: {.small}
\({\mathbb{F}}_{\ell_1}\).
:::
where
and
are two distinct primes none of which is \(3\) and satisfy one of the following
1. ::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 2 \pmod 3\)
:::
and
::: {.small}
\(n \equiv 4,7 \pmod 9\),
:::
2. ::: {.small}
\(\ell_1 \equiv 1 \pmod 3\), \(\ell_2 \equiv 2 \pmod 3\)
:::
such that
::: {.small}
\(n \equiv 8 \pmod 9\)
:::
and either
::: {.small}
\(\ell_1 \equiv 4,7 \pmod 9\)
:::
or the cubic residue
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\),
:::
where
::: {.small}
\(\pi_{\ell_1} \mid \ell_1\)
:::
is a prime of
::: {.small}
\(K\)
:::
,
3. ::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 3\)
:::
such that
::: {.small}
\(n \equiv 4,7 \pmod 9\)
:::
and at least one of the cubic residues
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is not \(1\), where
::: {.small}
\(\pi_{\ell_i} \mid \ell_i\)
:::
is a prime of
::: {.small}
\(K\)
:::
for
::: {.small}
\(i=1,2\)
:::
.
Assume that
is even. Then
i.e.
can be written as a sum of two rational cubes.
if
satisfy any of the conditions above. The result is then immediate from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
where
and
are distinct primes none of which is \(3\). If
is even, then
if
by Props. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}, [\[sylvesterpq2\]](#sylvesterpq2){reference-type="ref" reference="sylvesterpq2"} and [\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"}, whereas
if
by Prop. [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"}, [\[sylvesterpq3\]](#sylvesterpq3){reference-type="ref" reference="sylvesterpq3"} and [\[sylvesterpq4\]](#sylvesterpq4){reference-type="ref" reference="sylvesterpq4"}. The result is then immediate from Lemma [\[rankbound\]](#rankbound){reference-type="ref" reference="rankbound"}. ◻
and
be the rational \(3\)-isogenies defined above. Note that
and
Set
gives the exact sequence:
So,
; the last equality is due to Lemma [\[ono\]](#ono){reference-type="ref" reference="ono"}. By Theorem [\[sylvesterpq\]](#sylvesterpq){reference-type="ref" reference="sylvesterpq"},
Since
is even, we obtain
Hence,
if
and
Now, if
then
implies either
or
It is easy to see that
implies
Finally, for
if
then
On the other hand, if
then either
; equivalently, either
◻
when either
and
or
where
and
are primes. We have the following
1. If
::: {.small}
\(\ell_1\ell_2 \not\equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\),
:::
then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) =0\)
:::
and
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]=\Sh(E_{-432\ell_1^2\ell_2^2}/K)[3]=0\).
:::
2. If
::: {.small}
\(\ell_1 \ell_2 \equiv 8 \pmod 9\)
:::
such that either
::: {.small}
\(\ell_1 \not\equiv 1 \pmod 9\)
:::
or
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 \neq 1\),
:::
then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong \Q_3/{\mathbb{Z}}_3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q)=1\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q)=1\)
:::
and
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]=0\).
:::
3. If
::: {.small}
\(\ell_1 \ell_2 \not\equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 = 1\),
:::
then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong \Q_3/{\mathbb{Z}}_3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\).
:::
4. If
::: {.small}
\(\ell_1 \equiv 1 \pmod 9\),
:::
::: {.small}
\(\ell_2 \equiv 8 \pmod 9\)
:::
and
::: {.small}
\(\big(\frac{\ell_2}{\pi_{\ell_1}}\big)_3 = 1\),
:::
then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong\)
:::
one of
::: {.small}
\(\Q_3/{\mathbb{Z}}_3\)
:::
and
::: {.small}
\(\big(\Q_3/{\mathbb{Z}}_3\big)^3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{1,3\}\)
:::
and
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3] \in \{0,2\}\).
:::
where
and
are primes. We have the following
1. If
::: {.small}
\(\ell_1\ell_2 \not\equiv 1 \pmod 9\)
:::
and at least one of
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is not \(1\), then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong \Q_3/{\mathbb{Z}}_3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q)=1\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q)=1\)
:::
and
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]=0\).
:::
2. Let either
::: {.small}
\(\ell_1 \ell_2 \equiv 1 \pmod 9\)
:::
but
::: {.small}
\(\ell_1 \not\equiv 1 \pmod 9\)
:::
or
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 9\)
:::
and at least one of
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
and
::: {.small}
\(\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3\)
:::
is not \(1\), then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong \Q_3/{\mathbb{Z}}_3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\).
:::
3. If
::: {.small}
\(\ell_1\ell_2 \not\equiv 1 \pmod 9\)
:::
and both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 =\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\),
:::
then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong\)
:::
one of
::: {.small}
\(\Q_3/{\mathbb{Z}}_3\)
:::
and
::: {.small}
\(\big(\Q_3/{\mathbb{Z}}_3\big)^3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2\}\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{1,3\}\)
:::
.
4. If
::: {.small}
\(\ell_1 \equiv \ell_2 \equiv 1 \pmod 9\)
:::
such that both
::: {.small}
\(\big(\frac{\pi_{\ell_1}}{\pi_{\ell_2}}\big)_3 =\big(\frac{\pi'_{\ell_1}}{\pi_{\ell_2}}\big)_3 = 1\),
:::
then either
::: {.small}
\(\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3^\infty] \cong\)
:::
one of
::: {.small}
\(\Q_3/{\mathbb{Z}}_3\)
:::
and
::: {.small}
\(\big(\Q_3/{\mathbb{Z}}_3\big)^3\)
:::
or
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q)\in \{0,2,4\}\)
:::
. In particular, if
::: {.small}
\(\dim_{{\mathbb{F}}_3}\Sh(E_{-432\ell_1^2\ell_2^2}/\Q)[3]\)
:::
is even, then
::: {.small}
\(\text{rk } E_{-432\ell_1^2\ell_2^2}(\Q) \in \{0,2,4\}\)
:::
.
::: | {'timestamp': '2022-12-01T02:18:20', 'yymm': '2211', 'arxiv_id': '2211.17118', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17118'} |
null | null |
{'timestamp': '2022-12-01T02:20:57', 'yymm': '2211', 'arxiv_id': '2211.17224', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17224'} |
|
# Introduction
On a global basis, more than 60\(\%\) of the energy consumed is ultimately lost via heat to the environment. It was reported that buildings contribute somewhere between 20 to 40\(\%\) of the globally generated energy and energy related environmental emissions, with a sizeable amount of this energy going towards heating and cooling requirements of buildings. The overwhelming influence of climate change as well as the increasing energy demand in urban areas calls for more buildings to maximise their potential energy harvesting, distribution, storage and efficient usage. Thermoelectric (TE) units can make use of the unavoidable loses to generate free electric power, because of their ability to partially recover heat by converting it into electricity.\
In thermoelectric materials, a temperature gradient leads to a movement of charge carriers, which results into a difference of potential at the two ends of an open circuit. In this way, thermal energy can be partially converted into electrical energy with a finite efficiency ultimately bounded by the second low of the thermodynamics. A quick estimate of the energy conversion efficiency of a thermoelectric material is given by the dimensionless figure of merit, \(ZT\): \[\label{zt} ZT = \frac{S^2 \sigma}{\kappa}T,\] where \(S\) is the Seebeck coefficient, \(\sigma\) is the electrical conductivity, \(T\) is the absolute temperature, and the total thermal conductivity \(\kappa= \kappa_e + \kappa_l\) is made up of contributions from charge carriers \(\kappa_e\) and lattice vibrations \(\kappa_l\). To maximise the figure of merit, the Seebeck coefficient and electrical conductivity must be maximised, while at the same time, the thermal conductivity must be minimised. In the limit of \(ZT\to\infty\), the efficiency of a Carnot engine is retrieved. For practical applications, a figure of merit \(ZT\gg1\) is typically required.\
Cementitious materials are complex heterogeneous composites formed by the admixture of calcareous (mostly calcium carbonate) material such as limestone with silica-, alumina-and iron-based material, which are calcinated until they fuse together. Cement-based materials like concrete are not only used as structural materials for buildings, but also for other applications like nuclear waste disposal and massive structures like water reservoirs and dams.\
In recent years, attention has turned to composite cement-based materials. One such composite, carbon fiber reinforced cement (CFRC), has ignited great interest in the scientific community as a promising thermoelectric material. While standard cement formulations show only a mild thermoelectric effect, the inclusion of additives such as carbon and steel fibers, inorganic compounds like graphite and/or metallic oxides such as ZnO, Bi\(_2\)O\(_3\) or Fe\(_2\)O\(_3\), or standard thermoelectric materials like Bi\(_2\)Te\(_3\), can enhance the thermometric performance considerably. The challenges and opportunities presented by these composite materials have been discussed extensively in the literature, and in a recent review. While rapid progress has been made in improving the Seebeck coefficient within CFRCs, the overall figure of merit has not been improved due, understandably, to reportedly small values of electrical conductivity.\
The most important hydration product of cementitious materials is Calcium-Silicate-Hydrate gel, denoted C-S-H within the cement chemistry community, where C=CaO, S=SiO\(_2\) and H=H\(_2\)O. C-S-H gel has an average calcium to silicon (Ca/Si) ratio of 1.7, with local Ca/Si ratio fluctuations of 0.67 to 2.0. The description of the microscopic structure of cement is particularly complex due to the disordered and inhomogeneous nature of the material, which comprises several coexisting phases like alite, belite, etc. After the work of Taylor, the consensus is that C-S-H based materials have a layered atomic structure akin to those of tobermorite, T14Å \[Ca\(_5\)Si\(_6\)O\(_{16}\)(OH)\(_2\).7H\(_2\)O\] and jennite \[Ca\(_9\)Si\(_6\)O\(_{18}\)(OH)\(_6\).8H\(_2\)O\], with a calcium to silicon ratio of 0.83 and 1.5, respectively. Tobermorite minerals are generally characterized by their interlayer spacing and are named accordingly as tobermorite, T9Å, tobermorite, T11Å and tobermorite, T14Å. Extensive structural characterization via XRD experiments of C-S-H based materials has been carried out by the group of Merlino and Bonaccorsi. So far the first principles density functional theory calculations carried out on the C-S-H based materials such as jennite structure and tobermorite minerals has been centered around the understanding of their structural properties, elastic constants and anisotropic effect, average mechanical properties, vibrational properties and infrared spectra and/or NMR shift investigations. Recently, some properties of C-S-H within the context of the civil nuclear industry have been assessed, like its ability to trap and hold radioactive fission products like Sr and its daughters, and hydrogen gas production as a consequence of irradiation. Since the C-S-H gel is responsible for the cohesive strength and durability of cementitious materials, our drive is to optimize the thermoelectric transport coefficient of some C-S-H based tobermorite materials with the aim of improving their energy efficiency.\
Motivated by the experimental results on cement-based materials, we present here first-principles calculations of the structural and electronic transport properties of the C-S-H analogs tobermorite, T9Å \[Ca\(_5\)Si\(_6\)O\(_{16}\)(OH)\(_2\)\], tobermorite, T11Å \[Ca\(_4\)Si\(_6\)O\(_{15}\)(OH)\(_2\).5H\(_2\)O\], and tobermorite, T14Å \[Ca\(_5\)Si\(_6\)O\(_{16}\)(OH)\(_2\).7H\(_2\)O\] as shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. We are mainly interested in the thermoelectric properties in the low temperature regime since most of the experiments have been carried out at temperatures below 100 \(^{\circ}\)C. In particular, we focus on the effects of temperature and doping on electrical and thermal conductivity, Seebeck coefficient, power factor, and figure of merit. In this work, doping is modelled in terms of an intrinsic carrier concentration determined by the electronic chemical potential via the rigid band model. To the best of our knowledge, until now no calculations have been performed to understand the thermoelectric behavior of calcium silicate hydrate (C-S-H)-based materials. Our initial first-principles study can guide further investigations of the electronic transport properties of this class of materials.
# Computational Methods
## Details of electronic structure calculations
All our calculations were carried out within the framework of density functional theory (DFT) using a plane wave basis set as implemented in the open-source package Quantum ESPRESSO. All structural optimizations and electronic property calculations were carried out using the Perdew-Burke-Ernzerhof (PBE) exchange-correlation functional belonging to the generalized gradient approximation (GGA) family. Ultra-soft pseudopotentials were used and along with a plane wave basis with an energy cutoff of 70 Ry and density cutoff of 560 Ry for T9Å \[Ca\(_5\)Si\(_6\)O\(_{16}\)(OH)\(_2\)\] while energy cutoff of 80 Ry and density cutoff 640 Ry was used for T11Å \[Ca\(_4\)Si\(_6\)O\(_{15}\)(OH)\(_2\).5H\(_2\)O\] and T14Å \[Ca\(_5\)Si\(_6\)O\(_{16}\)(OH)\(_2\).7H\(_2\)O\], respectively. The Brillouin zone was sampled using the Monkhorst-Pack scheme, with a regular mesh of 3\(\times\)`<!-- -->`{=html}3\(\times\)`<!-- -->`{=html}3 for T9Å and T11Å, and 3\(\times\)`<!-- -->`{=html}3\(\times\)`<!-- -->`{=html}1 for T14Å. Electronic occupations were smeared using the Marzari-Vanderbilt scheme with broadening of \(0.03\) Ry and the SCF convergence threshold for the electronic wavefunctions was set to \(10^{-10}\) Ry for both the variable-cell and ionic relaxation. Both jennite and tobermorite models have layered structures, with layers bound by non-bonded interactions. To include the effect of these interactions and obtain realistic interlayer distances during lattice constant optimization, the semiempirical dispersion correction of Grimme *et al.* for the PBE functional (DFT-D3) was used. Atomic positions and cell parameters were fully optimized at 0 K using the Broyden-Fretcher-Goldfarb-Shanno method. We optimized all stress and force components to less than \(0.5\) kbar and \(10^{-6}\) Ry/Bohr, respectively.\
Finally we performed a single point energy calculation in order to ensure the robustness of our results with much denser \(k\)-point grid of 5\(\times\)`<!-- -->`{=html}5\(\times\)`<!-- -->`{=html}5 for all the tobermorite models. Since the electronic transport coefficients are strongly dependent on the band structure energies in the Brillouin zone, we performed non-self consistent calculations for all the tobermorite models using 9\(\times\)`<!-- -->`{=html}9\(\times\)`<!-- -->`{=html}9 Monkhorst-Pack \(k\)-points grid (365 \(k\)-points in the irreducible Brillouin zone), while maintaining the energy criterion above. These non-SCF calculations were used to obtain the band structures and transport coefficients of the tobermorite models.
## Transport Calculations {#subsec1}
Electronic transport coefficients were computed from the solution of the linearized Boltzmann transport equation in conjunction with the rigid-band approximation (RBA) and constant relaxation time approximations (CRTA) as implemented in the BoltzTraP code. In the rigid band approximation, it is assumed that only the chemical potential, not the band shape of the host compound, is changed according to the nominal doping and/or temperature. Hence, only the band energies are extrapolated for each calculation. The electronic band energies of all the tobermorite models were calculated on a 9\(\times\)`<!-- -->`{=html}9\(\times\)`<!-- -->`{=html}9 Monkhorst-Pack \(k\)-point grid and Fourier interpolated over a denser grid containing \(15\) times as many \(k\)-points for better numerical evaluation of the integrals, to obtain the thermoelectric transport coefficients. In the semi-classical Boltzmann approach, the energy-dependent transport distribution function-the kernel of all transport coefficients is expressed as: \[\begin{aligned}
\label{eqn1} \Sigma_{\alpha\beta}(\epsilon)&=&\frac{q^2}{V} \sum_{ik} \tau_{ik}v_\alpha(i,k)v_\beta(i,k) \delta(\epsilon-\epsilon(i,k)), \end{aligned}\] where \(\alpha\) and \(\beta\) are Cartesian components and the subscripts \(i\) and \(k\) are the band and wave vector indices, respectively. \(V\) is the unit cell volume, \(\epsilon(i,k)\) represents the \(k\)-dependent band energies, \(q\) is the electron charge, \(\tau_{ik}\) is the relaxation time and \(v_\alpha(i,k)\) is the \(\alpha\) component of the electron group velocity written as: \[\begin{aligned}
v_\alpha(k) &=& \frac{1}{\hbar}\frac{\partial \epsilon(i,k)}{\partial k_\alpha}. \end{aligned}\] The advantage of the RBA in conjunction with the CTRA, is that the transport distribution function \(\Sigma_{\alpha\beta}\left(\epsilon\right)\) does not depend on either temperature \(T\) or electron chemical potential (doping) \(\mu\). Therefore, by integrating over a fixed transport distribution function, \(\Sigma_{\alpha\beta}\left(\epsilon\right)\), the doping and/or temperature-dependent generalised transport coefficient of the \(p\)-th order, which is solely due to the Fermi-Dirac distribution function \(f_{0}\) is obtained as: \[\begin{aligned}
L^p(\mu, T)&=& \int\Sigma_{\alpha\beta}(\mu, T)(\epsilon(i,k)-\mu)^p\left(-\frac{\delta f_0(\epsilon,\mu,T)}{\delta\epsilon}\right)\delta\epsilon, \end{aligned}\] From where the electrical conductivity \(\sigma\), Seebeck coefficient \(S\), and electronic thermal conductivity \(\kappa_e\) are extracted as a function of temperature, \(T\), and electronic chemical potential, \(\mu\). These transport coefficients are written as: \[\begin{aligned}
\sigma(\mu,T)&=&\frac{1}{V}\int\Sigma_{\alpha\beta}(\mu,T)\left(- \frac{\partial f_0(\epsilon,\mu,T)}{\partial \epsilon}\right)d\epsilon, \end{aligned}\] \[\begin{aligned}
\label{eqn5} S(\mu,T)&=& \dfrac{1}{VqT\sigma(\mu,T)}\int\Sigma_{\alpha\beta}(\mu,T)(\epsilon-\mu)\left(-\frac{\partial f_0(\epsilon, \mu,T)}{\partial \epsilon}\right)d\epsilon, \end{aligned}\] and \[\begin{aligned}
\kappa_e(\mu,T)&=&\frac{1}{VT}\int\Sigma_{\alpha\beta}(\mu,T)(\epsilon-\mu)^2\left(-\frac{\partial f_0(\epsilon,\mu,T)}{\partial \epsilon}\right)d\epsilon-TS^2\sigma \end{aligned}\] At zero temperature, \(\mu\) is equal to the Fermi energy, \(E_F\). At fixed doping, the chemical potential steadily deviates from \(E_F\) with increasing temperature as a result of the energy dependence of the density of states. The electrical conductivity is related to the electronic thermal conductivity \(\kappa_e\) through the Wiedemann-Franz law, \(\kappa_e=L\sigma T\), where \(L\) is the Lorenz number. The electrical conductivity is given as a function of \(\tau(i,k)\) (Eq. [\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}) and, therefore, the relaxation time \(\tau\) must be included as a parameter. To obtain \(\tau\), one needs to calculate the scattering of electrons by ionized impurities, piezoelectric scattering, and/or acoustic phonon scattering by deformation. From Eq. [\[eqn1\]](#eqn1){reference-type="ref" reference="eqn1"}, it follows that \(\tau \propto (\epsilon-\epsilon(i,k))^{r-1/2}\), which sometimes has been used to understand the electron transport processes but has not been established in many materials. For a simple parabolic band model, \(\tau\) is expressed as a power-law in a reduced carrier energy \(\eta=E_F/k_{B}T\), \(\tau \propto \epsilon^{r-1/2}\), where \(r\) is the scattering parameter. Within the rigid-band approximation, \(\tau\) is typically assumed to be energy-independent and constant for simplicity. In this case, \(\tau\) cancels out in the expression for the Seebeck coefficient, which is one of the key quantities for the thermoelectric figure of merit, \(ZT\). The calculated transport coefficients form the essential part of the figure of merit \(ZT\), which can be rearranged as: \[\begin{aligned}
\label{eqn3} ZT &=& \frac{S^2 \sigma~T}{\kappa_e} \left(\frac{\kappa_e}{\kappa_e+\kappa_l} \right) \end{aligned}\] Two contributions to \(ZT\) can be identified in Eq. ([\[eqn3\]](#eqn3){reference-type="ref" reference="eqn3"}): (i) the electronic figure of merit (\(Z_eT = S^2 \sigma T/\kappa_e\)), which is independent of the relaxation time and obtained when the heat transferred to the material through lattice vibrations (\(\kappa_l\)) is completely neglected; and (ii) a scaling factor determined by the ratio of \(\kappa_e/(\kappa_e+\kappa_l)\). The electronic figure of merit \(Z_eT\) approaches \(ZT\) when \(\kappa_l\) becomes much smaller than \(\kappa_e\). Note that the final thermoelectric transport coefficients were evaluated from their averaged trace components.
# Results and discussion
## Structural details of Tobermorite models
A representation of the different tobermorite crystal structures (T9Å, T11Å and T14Å) is shown in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. These layered tobermorite minerals share similar structural arrangements, but are distinguished by the interlayer spacing distance. The notation 9Å, 11Å and 14Å indicates the characteristic interlayer spacings of \(0.93\) nm, \(1.13\) nm and \(1.4\) nm. This spacing is affected by the degree of hydration, a property that changes through heating. Each layer consists of continuous seven-fold edge-sharing Ca-O polyhedra stretched along the b-direction. All the oxygen atoms of the polyhedra are shared with silicon atoms to form corner-sharing \[SiO\(_4\)\]\(^{4-}\) tetrahedra chains. These chains are either dreierketten or wollastonite-like type with repeating units of three-fold tetrahedra. Two of the tetrahedra chains share edges with the Ca-polyhedron in a paired form while the third tetrahedra chain is linked to the calcium polyhedra through their apical oxygen atoms in a bridging form. Interlayer calcium polyhedra are present in all three tobermorite models. In T11Å and T14Å the interlayer spacing is filled with water molecules and additional calcium atoms. T11Å and T14Å are both different from T9Å because of the formation of double (condensed) dreierketten bridging silicate chains. Single tetrahedra silicate chains were also reported by Hamid for tobermorite 11 Å with differing calcium to silicon (Ca/Si) ratios of 0.67, 0.83 and 1.0. Experimental lattice parameters and atomic positions for T9Å (riverside), T11Å (tobermorite) and T14Å (plombierite) were obtained from the American Mineralogist Crystal Structure Database. These materials crystallize in triclinic (space group \(C\Bar{1}\)), monoclinic (space group of \(B11m\)) and monoclinic (space group \(B11b\)) Bravais lattices containing \(62\), \(88\), and \(104\) atoms per unit cell, respectively. The parent compound of tobermorite T11 Å and T14 Å shows partial occupancy for water/oxygen (WO) sites in its experimentally determined unit cell structure (i.e., the Crystallographic Information File just has WO, as it from XRD), which hinders a direct description via simulation. Refining atomic positions of the parent compound of T11Å and T14Å, leads to a breakdown in the nominal monoclinic symmetry to \(P1\) trinclinic Bravais lattice. The angles and lattice vectors were fixed to that of the parent compound, which were used as inputs for structural optimization. The calculated structural parameters are explicitly listed in Table [1](#tab1){reference-type="ref" reference="tab1"}. All parameters were determined using the DFT-D3 method to correct for dispersion interactions, and are compared with the available experimental data and previous DFT calculations in Table [1](#tab1){reference-type="ref" reference="tab1"}.
:::
The good agreement in the calculated lattice parameters relative to the available experimental values and the results cited in previous DFT calculations (deviation \(\pm\)`<!-- -->`{=html}1% overall) gives confidence in the relaxed tobermorite structures. In T11Å and T14Å, there is a small deviation in the angles \(\alpha\) and \(\beta\) from the experimental measured values because of the symmetry breakdown mentioned above.
## Electronic structure calculations
Accurate electronic structure calculations were carried out to obtain the transport properties. These calculations were performed using the relaxed cell parameters listed in Table [1](#tab1){reference-type="ref" reference="tab1"}. The calculated band structures of the tobermorite models along some high symmetry paths for triclinic lattices are presented in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The tobermorite models are insulators and generally expected to have wide energy band gaps. The band structures of all the tobermorite models showed a direct energy band gap at the \(\Gamma\)-point. The resulting gaps (\(E_{gap}\)) listed in the last row of Table [1](#tab1){reference-type="ref" reference="tab1"}, i.e. 4.6 eV, 4.2 eV and 4.0 eV for T9Å, T11Å and T14Å, respectively, were found to be in very good agreement with previous DFT-LDA results. Our band structure calculations showed that as the water content increases in T11Å and T14Å structures, their band gaps decrease as compared to T9Å, which contains no water molecules. Most of the bands near the valence band maximum (VBM) are flatter (heavy bands) than the bands near the conduction band minimum (CBM), which indicate larger effective mass for the holes than for the electrons. The resulting heavy bands near the VBM should lead to a large Seebeck coefficient in \(p\)-type thermoelectric materials. It is interesting to note that the conduction bands (Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}) of the tobermorite models exhibit obvious parabolic character, which suggests that their transport properties can also be modeled using the simple parabolic band approach. To gain a better understanding of the nature of the states in the VBM and CBM, we have plotted the element-resolved partial density of states (pDOS) of the unit cells of T9Å, T11Å and T14Å, respectively, as shown in Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}. The sharp peaks in the electronic density of states below the Fermi level for all three compounds originate almost entirely from the non-water O states, with small contributions from the water O-H bonds, Si, Ca and H\(_2\)O states, respectively. From Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, it can be seen that non-water oxygen O-\(p\) states contribute significantly to the pDOS between-8.5 eV to-2.0 eV, hybridizing with the Si-\(p\) states and forming strong covalent bonds that lead to the formation of SiO\(_4\) tetrahedra. The Ca-\(p\) states exhibit a relatively small contribution to the density of states (from-7 eV to-3 eV), which indicates a negligible contribution of the alkaline-earth metal to the covalent bond but stresses the ionic bonding within the interlayer calcium polyhedra. This ionic bond character is greater in T9Å than in T11Å and T14Å models due to the absence of H\(_2\)O molecules. The hybridized pDOS is expected to be beneficial to the electronic transport properties. In the tobermorite models, the water molecules are confined to the interlayer region and interact in the sense of hybridising with all other elements within the crystal. The peak in the lower valence band composed mainly of the Ca-\(s\) states was observed to make significant contributions from-24 eV to-20 eV and their interaction with the non-water O contributes to the ionic bonding within the interlayers of Ca-O polyhedra. Turning to the conduction bands, the Ca and Si states dominate the conduction band minimum. These bands were observed to be generally more dispersive than those found in the valence bands. Dispersive bands are considered to contribute to high electron mobility because of the small effective mass of the carriers.
## Electronic transport properties
### Seebeck Coefficient {#seebeck-coefficient .unnumbered}
Within the constant relaxation time approximation (CRTA), the Seebeck coefficient \(S\) is independent of \(\tau\). Without any adjustable parameters, \(S\) results exclusively from a first principles electronic structure calculation. For a parabolic band model of a wide band-gap or non-degenerate semiconductor, the Seebeck coefficient is given by: \[\begin{aligned}
\label{eqn8} S &=&\pm \frac{k_{B}}{q}\left[\eta-\left(r + \frac{5}{2}\right) \right], \end{aligned}\] where \(\eta\) is the reduced carrier energy, \(r\) is the scattering parameter, \(q\) is the electron charge and \(k_B\) is the Boltzmann constant. Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} shows the temperature dependence of the averaged Seebeck coefficient of all tobermorite models for different values of carrier concentrations ranging from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\). These calculations were performed for both the \(p\)-type and \(n\)-type doping. The averaged Seebeck coefficient is a conductivity-weighted average of its directional components: \[\begin{aligned}
S_{avg} = \dfrac{S_a\sigma_a + S_b\sigma_b + S_c\sigma_c}{\sigma_a+\sigma_b+\sigma_c}, \end{aligned}\] which stem from the definition of \(S\) in terms of the Onsager coefficients: \(S=L_1/L_0T\). Here, \(a\), \(b\), and \(c\) are the Cartesian axes. This equation is the direct expression of the directional average of the transport coefficient in Eq. ([\[eqn5\]](#eqn5){reference-type="ref" reference="eqn5"}), provided that the relaxation time is independent of the direction. It is important to note that the negative sign of the Seebeck coefficient indicates electron carriers while the positive sign is for hole carriers. For the sake of simplicity, our discussion will be focused on the values rather than the signs. The calculated values of Seebeck coefficient are almost invariant in both \(p\)-type and \(n\)-type doping and grow monotonically with temperature as the carrier concentration increases from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\) for T9Å, T11Å and T14Å, respectively. In all the tobermorite models, the calculated Seebeck coefficients show a linear decrease in their absolute values with increasing carrier concentrations as required by Eq. ([\[eqn8\]](#eqn8){reference-type="ref" reference="eqn8"}). The optimal doping for thermoelectric performance for these models expected at \(10^{17}\) cm\(^{-3}\) for both \(p\)-type and \(n\)-type materials. Furthermore, from Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} the calculated absolute Seebeck coefficients for \(p\)-type doping in the tobermorite models were observed to be larger than those of the \(n\)-type doping. This is the direct consequence of the heavy valence bands present in the band structure of the tobermorite models in which the effective mass of the hole carriers keeps increasing. T14Å (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(c)) shows higher \(S\) values at lower carrier concentration for \(p\)-type doping than those of T9Å (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(a)) and T11Å (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(b)), because of its smaller energy band gap compared to the other two models. These values of \(S\) for tobermorite models are considered the first results, since there are no available experimental or theoretical data to compare them with.
### Relaxation Time {#Sec:RelaxTime .unnumbered}
The BoltzTraP code returns the electric conductivity divided by the constant relaxation time \(\sigma/\tau\). To obtain a quantitative estimate of the electrical conductivity \(\sigma\), we need to calculate the relaxation time \(\tau\) independently. Considering the high computational cost and the size of the unit cells of the tobermorite models, only the energy-dependent relaxation time for the T9Å model---*i.e.*, the model with the fewest atoms in the unit cell---was calculated using the simple parabolic band (SPB) approach, as a function of temperature, and for carrier concentrations ranging from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\) for both electrons and holes. In the SPB approximation, the relaxation time is assumed to follow a power-law expression: \(\tau=\tau_{0}\left(T\right) \left(\epsilon-\epsilon_0\right)^{r-1/2}\), where \(\epsilon_0\) is a reference energy, *e.g.* the valence band maximum (VBM) or the conduction band minimum (CBM). The value \(r=0\) is appropriate for scattering from acoustic phonons at a temperature larger than the Debye temperature. Note that the prefactor \(\tau_{0}\) is temperature dependent (see below). The SPB approach has been successfully applied to study the thermoelectric properties of materials as reported in the literature. In the case of (implicit) \(n\)-type doping, the Fermi energy \(E_F\) was first determined from the carrier concentration given by \[\begin{aligned}
n\left(T\right) &=& \frac{1}{2\pi^2}\left(\frac{2m^*_dk_BT}{\hbar^2}\right)^{\frac{3}{2}}F_{\frac{1}{2}}(\eta), \end{aligned}\] and the Fermi integral \(F_j(\eta)\) is defined as \[\begin{aligned}
\label{eqn4} F_j(\eta) &=& \int_0^\infty \frac{\epsilon^j}{1+\exp{\left(\frac{\epsilon}{k_BT}-\eta\right)}}d\epsilon\;, \end{aligned}\] where \(\eta=(E_F-\epsilon_0)/k_BT\). For electrons, \(\epsilon_0\) is the energy of the CBM. The density of states effective mass is defined as \(m^*_d = N_v^{\frac{2}{3}}(m^*_a m^*_b m^*_c)^{\frac{1}{3}}\), where \(N_v\) is the degeneracy of the band valley and \(m^*_a\), \(m^*_b\), \(m^*_c\) are the principal effective masses with respect to the Cartesian directions at the band valley. The principal effective masses were evaluated from the Hessian matrix of band energies at the band edges using: \[\begin{aligned}
\frac{1}{m^*_{ij}}=\frac{1}{\hbar^2}\frac{\partial^2\epsilon}{\partial k_i\partial k_j} \end{aligned}\] where \(i,j\) are Cartesian components. The elements of the Hessian matrix were calculated using the finite difference method with a \(5\)-point central difference stencil. The calculated effective masses at the conduction band minimum are \(m^*_{k_a}=0.149 m_0\), \(m^*_{k_b}=133 m_0\) and \(m^*_{k_c}=0.155 m_0\), where \(m_0=9.11\cdot10^{-31}\) kg is the free electron mass. Similar calculations were done in the case of (implicit) \(p\)-type doping. In this case the calculated effective masses at the valence band maximum are \(m^*_{k_a}=0.309m_0\), \(m^*_{k_b}=1.161m_0\) and \(m^*_{k_c}=0.989m_0\). According to the SPB approach, the energy dependent relaxation time for phonon scattering is expressed as: \[\begin{aligned}
\label{tau_acoustic} \tau &=& \frac{\sqrt{2}\pi \hbar^4 v_s^2\rho}{3E_d^2\left(m^*k_BT\right)^{\frac{3}{2}}}\frac{F_0(\eta)}{F_{\frac{1}{2}}(\eta)}, \end{aligned}\] where we consider acoustic phonons because their scattering is dominant at temperatures larger than the Debye temperature and moderate doping. Eq. ([\[tau_acoustic\]](#tau_acoustic){reference-type="ref" reference="tau_acoustic"}) is consistent with the SPB approximation if we take \(1/\tau_0\propto k_BT\), i.e., the scattering rate proportional to the number of phonons. Here, \(\rho\) is the mass density and the longitudinal sound velocity \(v_s\) is calculated from the elastic constants values of \((C_{11}/\rho)^{\frac{1}{2}}\), \((C_{22}/\rho)^{\frac{1}{2}}\) and \((C_{33}/\rho)^{\frac{1}{2}}\) in the direction of the wave vector. The elastic constants (\(C_{11}\)=158 GPa, \(C_{22}\)=154 GPa and \(C_{33}\)=74 GPa) were obtained from a DFT calculation at the gamma point. \(E_d\) is the acoustic deformation energy defined by \(E^i_d = \Delta\epsilon_i/(\Delta V/V_0)\) and calculated from the energy change (\(\Delta\epsilon_i\)) of the \(i\)-\(th\) band with volume dilation (\(\Delta V/V_0\)) along the Cartesian directions. We calculated the band structure by varying the volume from \(0.98V_0\) to \(1.01V_0\) with \(\pm\)`<!-- -->`{=html}5.0% increase of the cell volume. To calculate the deformation energy, we adopted the method proposed by Xi \(et\) \(al.\) by taking the energy change of the conduction band minimum (CBM) and the valence band maximum (VBM) for electrons and holes, respectively. The evaluated deformation energy for T9Å is 10.428 eV for the VBM and 5.381 eV for the CBM. In Figs. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}(i) to [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}(v), we plot the relaxation time \(\tau\) as a function of temperature and carrier concentration obtained from the SPB model for T9Å, along the three Cartesian axes. A glance at Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} shows that at low temperatures \(\tau\) is large and decreases with increasing temperature. As we move from \(10^{17}\) cm\(^{-3}\) to higher carrier concentrations, the relaxation time \(\tau\) decreases linearly with temperature for both \(p\)-type and \(n\)-type doping. The \(\tau\) values along the \(x\) and \(y\) directions are comparable because the longitudinal sound velocity is nearly the same. Over the entire carrier concentration range, the \(n\)-type relaxation time \(\tau\) exhibits higher values compared to those of the \(p\)-type. This is because the electrons have a much smaller effective mass than the holes, which results in a lower density of states and hence a longer relaxation time \(\tau\).
### Electrical Conductivity {#electrical-conductivity .unnumbered}
The trace of the electrical conductivity \({\sigma}_{avg}\)=\(\left(\sigma_a+\sigma_b+\sigma_c\right)/3\) was obtained from BoltzTraP for T9Å using the calculated relaxation time computed in Sec. [\[Sec:RelaxTime\]](#Sec:RelaxTime){reference-type="ref" reference="Sec:RelaxTime"}. The resulting electrical conductivity \(\sigma\) for \(p\)-type and \(n\)-type doping as a function of temperature and fixed carrier concentrations (\(10^{17}\) to \(10^{21}\) cm\(^{-3}\)) is presented in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}. Both \(p\)-type and \(n\)-type T9Å exhibit an increase in electrical conductivity with increasing carrier concentration, which is consistent with the dependence of electrical conductivity on the carrier concentration. At \(300\) K, we found \(p\)-type electrical conductivity values of \(22.49\), \(224.43\), \(2190.98\), \(18534.20\) and \(93440.60\) \(\Omega^{-1}\)m\(^{-1}\) for carrier concentrations increasing from \(10^{17}\) cm\(^{-3}\) to \(10^{21}\) cm\(^{-3}\). For \(n\)-type doping, electrical conductivity values of \(367.62\), \(3666.00\), \(30631.10\), \(289992.00\) and \(1.044\times10^6\) \(\Omega^{-1}\)m\(^{-1}\) were obtained for the same concentration range at \(300\) K. A further glance at Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} shows that at low temperature \(\sigma\) is higher and then decreases as temperature increases, a behaviour attributed to the temperature dependence of the electron-phonon scattering described by a T\(^{-3/2}\) power-law. However, the electrical conductivity increases overall with increasing carrier concentration from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\) in both the \(p\)-type and \(n\)-type, respectively. This is typical of a lightly doped semiconductor where the electrical conductivity decreases with temperature when most of the carriers are ionised, as a result of the thermal scattering of carriers by the vibrating lattice. The overall \(n\)-type electrical conductivity \(\sigma\) values are significantly higher than those of the \(p\)-type doping at the same carrier concentration and temperature. This is due to the smaller effective mass at the conduction band maximum of the \(n\)-type doping.\
Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} shows the temperature-dependent electrical conductivity \(\sigma/\tau\) versus carrier concentration obtained from BoltzTraP, in both \(p\)-type and \(n\)-type doping for T11Å and T14Å models. The general observation from Figs. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}(i) and [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}(ii), is that \(\sigma/\tau\) decreases at a negligible rate with temperature over the entire carrier concentrations range, reaching a maximum value of \(<0.1\times10^{6}\) \((10^{14}(\Omega ms)^{-1})\) for \(p\)-type and \(>0.2\times10^{6}\) \((10^{14}(\Omega ms)^{-1})\) for \(n\)-type at carrier concentrations of \(10^{21}\) cm\(^{-3}\), for T11Å and T14Å models, respectively. In the whole carrier concentration range, the \(n\)-type values are higher than those of \(p\)-type doping in both T11Å and T14Å. This difference can be attributed to the heavy valence band maximum in their band structures (see Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(b) and [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(c)), which causes a reduction in the mobility of the hole carriers. However, as shown in Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} the electrical conductivity of T14Å is slightly higher than that of T11Å for the same temperature and carrier concentrations in both the \(p\)-type and \(n\)-type doping.
### Electronic Thermal Conductivity {#electronic-thermal-conductivity .unnumbered}
The electronic component of the total thermal conductivity is given by the Wiedemann-Franz relation: \[\kappa_e=L\sigma T,\label{eqn7}\] where \(L\) is the Lorenz number and takes the value \(L=2(k_B/q)^2\) for the non-degenerate case. We evaluated the averaged electronic thermal conductivity \(\kappa_{e}\) as a function of temperature and carrier concentration for T9Å for both \(p\)-type and \(n\)-type doping using the values of \(\tau\) computed from Sec. [\[Sec:RelaxTime\]](#Sec:RelaxTime){reference-type="ref" reference="Sec:RelaxTime"}. The results are presented in Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}(ii). For \(p\)-type doping \(\kappa_e\) decreases slowly with temperature as the carrier concentration increases, except at carrier concentration 10\(^{20}\) cm\(^{-3}\), where the magnitude of \(\kappa_e\) initially increases up to 325 \(K\) and then slightly decreases. In contrast, for the \(n\)-type \(\kappa_e\) slightly increases (1%) with temperature at carrier concentrations of \(10^{18}\), \(10^{19}\) and \(10^{21}\) cm\(^{-3}\), and decreases with temperature at carrier concentrations of \(10^{17}\) and \(10^{20}\) cm\(^{-3}\). The average electronic thermal conductivity for the tobermorite 11Å and 14Å models was also evaluated and the plots are presented in Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}. It exhibits a monotonic increase with carrier concentration in the entire temperature range when the former increases from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\), for both \(p\)-type and \(n\)-type doping and for both T11Å and T14Å. This linear increase in \(\kappa_e/\tau\) can easily be explained through the Wiedemann-Franz law, Eq. ([\[eqn7\]](#eqn7){reference-type="ref" reference="eqn7"}), i.e. multiplying the decreasing values of \(\sigma/\tau\) by the operating temperature, raises the values of \(\kappa_e/\tau\). The electronic thermal conductivity exhibits a monotonic increase with temperature, the \(n\)-type values being an order of magnitude higher than those of the \(p\)-type doping over the entire carrier concentration range, for both T11Å and T14Å.
### Power Factor {#power-factor .unnumbered}
The power factor, \(S^2\sigma/\tau\) was evaluated as a function of temperature and carrier concentration for the T11Å and T14Å models using the values of \(S\) and \(\sigma/\tau\) obtained from BoltzTraP. Results are presented in Fig. [\[fig9\]](#fig9){reference-type="ref" reference="fig9"}. Both \(p\)-type and \(n\)-type power factors for T11Å and T14Å increase monotonically with temperature over the considered carrier concentration range. This is due to \(S^2\) increasing faster than \(\sigma/\tau\) decreases. However, at carrier concentrations of \(10^{17}\) and \(10^{21}\) cm\(^{-3}\) in T11Å, the power factor is increasing with temperature at a very negligible rate as compared to that of \(n\)-type doping, due to the almost zero \(S\) and very low \(\sigma/\tau\) values. A similar behaviour was observed in T14Å at \(10^{17}\) cm\(^{-3}\). In \(p\)-type T11Å (T14Å) at \(10^{20}\) cm\(^{-3}\) concentration, the power factor rises to a maximum of \(\approx 0.36\times 10^{14}\) mW/mK\(^2\)s (\(\approx 0.45\times 10^{14}\) mW/mK\(^2\)s) at about 400 K. In \(n\)-type doping, the optimal power factors are slightly lower, hence suggesting that \(p\)-type doping holds the promise for a slightly better thermoelectric performance in tobermorite models. The temperature-dependent power factor \(S^2\sigma\) versus carrier concentrations plot for T9Å is also shown in Fig. [\[fig9\]](#fig9){reference-type="ref" reference="fig9"}(iii). In both \(n\)-type and \(p\)-type power factor is seen to decrease as temperature increases except in higher carrier concentration values of \(10^{20}\) and \(10^{21}\) cm\(^{-3}\). However, the power factor generally improved when the values of scattering time was added.
### Figure of merit ZT {#figure-of-merit-zt .unnumbered}
Using the calculated Seebeck coefficient and the values of \(\sigma\) and \(\kappa_e\), we calculated \(Z_eT=S^2\sigma T/\kappa_e\), which is the figure of merit that ignores the lattice thermal conductivity \(\kappa_l\), for \(p\)-type and \(n\)-type T9Å. This is reported in Fig. [\[fig10\]](#fig10){reference-type="ref" reference="fig10"}(i). The calculated \(Z_eT\) values of \(p\)-type T9Å are larger than those of \(n\)-type doping, because of the higher Seebeck coefficient (see Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(a)). The calculated \(Z_eT\) decreases for increasing carrier concentration, with the optimal values at \(10^{17}\) cm\(^{-3}\) in both \(p\)-type and \(n\)-type T9Å. The maximum \(Z_eT\) value at \(10^{17}\) cm\(^{-3}\) for \(p\)-type doping is \(0.983\), which is constant from 275-400 K, while for \(n\)-type doping the value is 0.956 from 375-400 K. At 300 K, \(Z_eT\) takes the value 0.951 for \(n\)-type doping at \(10^{17}\) cm\(^{-3}\). In Fig. [\[fig10\]](#fig10){reference-type="ref" reference="fig10"} we report the carrier concentration and temperature dependence of \(Z_eT\) for \(p\)-type and \(n\)-type T11Å and T14Å. These were calculated using the Seebeck coefficient, electrical conductivity, \(\sigma/\tau\) and electronic thermal conductivity, \(\kappa_e/\tau\) obtained from BoltzTraP. Notice that the relaxation time \(\tau\) is not needed to compute \(Z_eT\), as it cancels out in the ratio \(\sigma/\kappa_e=(\sigma/\tau)/(\kappa_e/\tau)\). From Fig. [\[fig10\]](#fig10){reference-type="ref" reference="fig10"}(ii) (T11Å) and [\[fig10\]](#fig10){reference-type="ref" reference="fig10"}(iii) (T14Å), it is clear that \(p\)-type exhibits larger \(Z_eT\) compared to \(n\)-type over the whole range of carrier concentrations. The optimal values of \(Z_eT\) slightly decrease with carrier concentration, but overall they remain close to 1, being optimal at \(400\) K for \(p\)-type T11Å. For \(p\)-type T14Å the optimal \(Z_eT\) values of 1.14, 1.17 and 1.20 correspond to a temperature of \(225\) K and carrier concentrations from \(10^{17}\) to \(10^{19}\) cm\(^{-3}\). For \(n\)-type doping, the highest values of \(Z_eT\) are again slightly below 1 and decreasing mildly with concentration for both T11Å and T14Å. These results suggest that these tobermorite cement-based material would perform better as a \(p\)-type thermoelectric material in the low temperature regime.\
To illustrate the potential thermoelectric performance of cement-based composites, their overall figure of merit, \(ZT\), we use in Eq. ([\[eqn3\]](#eqn3){reference-type="ref" reference="eqn3"}) the following representative values for the total thermal conductivity, \(\kappa = \kappa_e + \kappa_l = 1.15\) W/mK and doping, \(n = 10^{19}\) cm\(^{-3}\). Carrier densities as high as \(1.626\cdot10^{18}\) cm\(^{-3}\) have been reported in the literature. Values of the total thermal conductivity in the interval \(0.53\)-\(1.15\) W/mK have been reported for plain (*i.e.*, with no additives) hydrated cement paste.. We also fix the scattering rate, \(\tau=10^{-14}\) s for tobermorite cement models T11Å and T14Å. In Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}, based on our assumptions, our calculated results (ZT) for a typical plain hydrated cement paste show that the contribution of the lattice thermal conductivity, \(\kappa_l\) is very significant and that the figure of merit is of the order \(<0.1\) for lower carrier density \(10^{17}\) to \(10^{19}\) cm\(^{-3}\) and a scaling factor \(>0.1\) for higher carrier density. This means that the actual values of \(ZT\) may be one tenth or slightly more of what we reported in the figures for the electronic figure of merit, \(Z_eT\). From Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"} (i), it can also be seen that the inclusion of the scattering rate obtained via the simple parabolic band method, play a significant role as well in the determination of the figure of merit. Overall, our calculations indicate that \(p\)-type cement materials are the promising thermoelectric material.
# Conclusions
We have, for the first time, employed first principles methods and the semi-classical Boltzmann theory to systematically investigate the electronic and transport properties of tobermorite minerals as an analog for cement for use in cement-based thermoelectrics. The calculated structural parameters obtained were seen to be comparable with the available experimental and DFT data. The electronic structure of these materials showed that they are direct band gap insulators, whose band gap (already mentioned in Table [1](#tab1){reference-type="ref" reference="tab1"}) decreases as the concentration of chemically bound water molecules increases. The calculated electronic relaxation time using a simple parabolic band approach showed a decreasing trend with increasing temperature, over the entire range of carrier concentrations. The electronic relaxation time \(\tau\) along the \(a\) and \(b\) crystallographic directions changes quite slowly as carrier concentration increases from \(10^{17}\) to \(10^{21}\) cm\(^{-3}\), because of the small difference in their longitudinal sound velocities. To the best of our knowledge there are no experimental data reported on transport properties of tobermorites, hence the results presented here are to be taken as predictions. However these results can also be affected by uncertainties especially the figure of merit \(ZT\), which is based on the total thermal conductivity. As we tried to clarify these uncertainties based on the assumptions we made in Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"}, it can be seen that the contribution from lattice vibrations of these cement materials is essential and can not be neglected. We further saw from Fig. [\[fig11\]](#fig11){reference-type="ref" reference="fig11"} (i) that the contribution from the scattering time \(\tau\) is appropriate from carrier concentration of \(10^{19}\) to \(10^{21}\) for both \(p\)-type and \(n\)-type materials but not much effective at carrier concentration of \(10^{21}\) cm\(^{-3}\). We found that the electronic transport properties of T9Å, T11Å and T14Å favor the \(p\)-type thermoelectric behaviour. This is because of the presence of heavy, and thus flat, valence bands, which lead to a higher effective mass and hence a larger Seebeck coefficient. In contrast, the effective mass for electrons in the conduction band is an order of magnitude smaller, hence leading to smaller Seebeck coefficients in \(n\)-doped tobermorites. The small value of the electronic thermal conductivity suggests that the total thermal conductivity is mainly dominated by the intrinsic lattice thermal conductivity.\
Our study provides a blueprint for further computational studies of emerging cement composite materials of interest for thermoelectric applications, in particular those combined with carbon fibre. We have demonstrated from first principles how the Seebeck coefficient of a candidate material can be predicted *in silico*, allowing for theoretical modifications to the material to be proposed and evaluated rapidly, in order to maximise the thermoelectric figure of merit. | {'timestamp': '2022-12-01T02:18:27', 'yymm': '2211', 'arxiv_id': '2211.17128', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17128'} |
null | null |
null | null |
# Introduction
Many task-oriented dialogue (TOD) datasets and models have been proposed along with the advancement of dialogue system research. However, the diversity in data formats and ontologies between datasets brings inconvenience to model adaptation and uniform evaluation, which potentially hinders the study of model generalization and knowledge transfer across datasets. To address this issue, we define a unified data format that serves as the adapter between TOD datasets and models: datasets are first transformed to the unified format and then loaded by models. Following this design, we present ConvLab-3, a flexible dialogue system toolkit supporting many datasets and models. As illustrated in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}, a unified dataset in ConvLab-3 consists of dialogues, ontology, and database (or API interface). We define the unified format over user goal, dialogue acts, state, API result, etc. to support common tasks, while keeping the dataset-specific annotation (e.g., emotion) intact. Once a dataset is transformed into the unified format, it can be immediately used by models on the tasks it supports. Similarly, a model supporting the unified format can access all transformed datasets. This feature reduces the cost of adapting \(M\) models to \(N\) datasets from \(M\times N\) to \(M+N\), saving the time of (1) researchers to conduct (especially transfer learning) experiments, (2) developers to build an agent with custom datasets, and (3) community contributors to consistently add models, datasets, and uniform data processing functions. Compared with other dialogue toolkits such as PyDial, Rasa, ParlAI, and Nemo, ConvLab-3 and its predecessors provide recent powerful models for all components of a dialogue system, which allows people to assemble conversational agents with different recipes and to evaluate the agent with various user simulators. ConvLab-3 inherits the framework of ConvLab-2 but provides support to the unified data format for existing datasets, commonly used and state-of-the-art models, and task evaluation scripts. Thanks to the unified data format, we further add many TOD datasets and powerful transformer-based models, including two transferable user simulators. In addition, we highly advanced the RL toolkit by 1) simplifying the process of building the dialogue system and user environment with different modules, 2) improving the RL training process, and 3) adding evaluation and analysis tools. Since the formats of different datasets are unified, it is convenient to perform dataset-agnostic evaluation and investigate the effects of different transfer learning settings such as domain adaptation, domain adaptive pre-training, and continual learning. In this paper, we explore several transfer strategies and conduct comprehensive experiments on multiple tasks and settings. Experimental results show that the benefit of transfer learning is extremely large when labeled data in the target domain is insufficient but gradually decreases as the number of labeled data increases. Moreover, we conduct several RL experiments with multiple data scenarios, different user simulators, and the usage of uncertainty. The results show that pre-training on different data is not necessarily beneficial for initial performance but leads to better final performance, training with different user simulators leads to different dialogue system behavior, and dialogue systems that leverage uncertainty estimates in the state prediction request information from users more frequently in order to resolve uncertainty.
``` {.python language="Python" numbers="none" caption="Example usage of unified datasets."}
from convlab.util import * dataset_name = "multiwoz21" # load dataset: a dict maps data_split to dialogues dataset = load_dataset(dataset_name) # load dataset in a predefined order with a custom # split ratio for reproducible few-shot experiments dataset = load_dataset(dataset_name, \ dial_ids_order=0, split2ratio={"train": 0.01}) # load ontology and database similarly ontology = load_ontology(dataset_name) database = load_database(dataset_name) # query the database with domain and state state = {"hotel": {"area": "east", \ "price range": "moderate"}} res = database.query("hotel", state, topk=3) # Example functions based on the unified format # load the user turns in the test set for NLU task nlu_data = load_nlu_data(dataset, "test", "user") # dataset-agnostic delexicalization dataset, delex_vocab = create_delex_data(dataset)
```
[\[listing:util\]]{#listing:util label="listing:util"}
# Unified Data Format {#unified-data-format-section}
Our goal is to unify the format of annotations included in many datasets and commonly used by dialogue models (e.g., dialogue acts, state), while keeping the original format of annotations that only appear in specific datasets (e.g., emotion). As we integrate more datasets in the future, we will transform more kinds of common annotations into the unified format. In our unified data format, a dataset consists of (1) an **ontology** that defines the annotation schema, (2) **dialogues** with transformed annotations, and (3) a **database or API interface** that links to external knowledge sources. Further, the metadata of the dataset such as the format transformation process and data statistics are described in a dataset card. As shown in Listing [\[listing:util\]](#listing:util){reference-type="ref" reference="listing:util"}, we provide utility functions to process the unified datasets, such as delexicalization, splitting data for few-shot learning, and loading data for specific tasks. Based on the unified format, evaluations of common tasks are standardized. We list datasets and models supporting the unified data format in Table [\[tab:datasets\]](#tab:datasets){reference-type="ref" reference="tab:datasets"} and [\[tab:tasks\]](#tab:tasks){reference-type="ref" reference="tab:tasks"} respectively.
## Ontology
Following, an ontology consists of: (1) Domains and their slots in a hierarchical format. Each slot has a Boolean flag indicating whether it is a categorical slot (whose value set is small and fixed). (2) All possible intents in dialogue acts. (3) Possible dialogue acts appeared in the dialogues. Each act is comprised of intent, domain, slot, and originator (i.e., system or user). (4) Template dialogue state. We also provide a natural language description, if any, for each domain, slot, and intent to facilitate few-shot learning and domain transfer.
## Dialogues
For dialogue in the unified format, dialogue-level information includes the dataset name, data split, unique dialogue ID, involved domains, user goal, etc. Following MultiWOZ, a user goal has informable slot-value pairs, requestable slots, and a natural language instruction summarizing the goal. Turn-level information includes speaker, utterance, dialogue acts, state, API result, etc. Each dialogue act is a tuple consisting of intent, domain, slot, and value. According to the value, we divide dialogue acts into three groups: (1) **categorical** for categorical slots whose value set is predefined in the ontology (e.g., inform the weekday of a flight). (2) **non-categorical** for non-categorical slots whose values can not be enumerated (e.g., inform the address of a hotel). (3) **binary** for intents without actual values (e.g., request the address of a hotel). The state is initialized by the template state as defined by the ontology and updated during the conversation, containing slot-value pairs of involved domains. An API result is a list of entities retrieved from the database or other knowledge sources. We list common annotations included in the unified data format and the tasks they support in Table [\[tab:tasks\]](#tab:tasks){reference-type="ref" reference="tab:tasks"}. Other dataset-specific annotations are retained in their original formats.
## Database/API Interface
To unify the interaction with different types of databases, we define a base class `BaseDatabase` that has an abstract `query` function to be customized. The `query` function takes the current domain, dialogue state, and other custom arguments as input and returns a list of top-k candidate entities which is consistent with the format of the API result in dialogues. By inheriting `BaseDatabase` and overriding the `query` function, we can access different databases/APIs in a similar fashion.
# New Models
Besides models already included in ConvLab-2, we integrate new transformer-based models supporting the unified data format as shown in Table [\[tab:tasks\]](#tab:tasks){reference-type="ref" reference="tab:tasks"}: SUMBT, SetSUMBT and TripPy for dialogue state tracking (DST), DDPT and LAVA for policy learning, SC-GPT for natural language generation (NLG), SOLOIST with T5 as backbone model for end-to-end modeling (End2End), as well as TUS that outputs user dialogue acts and GenTUS that outputs both user dialogue acts and response for user simulation (US).
Inspired by that converts all NLP tasks into a text-to-text format, we provide serialization and deserialization methods for annotations in the unified data format, as shown in Table [1](#tab:serialization){reference-type="ref" reference="tab:serialization"}. In this way, we can apply a text-to-text model to solve any task in Table [\[tab:tasks\]](#tab:tasks){reference-type="ref" reference="tab:tasks"}. As an example, we provide a unified training script for T5, producing a family of models including T5-NLU, T5-DST, T5-NLG, etc. Other pre-trained language models such as GPT-2 could similarly be used as backbone models. We regard these models as base models to solve a task since they are not involved with dataset/task-specific model design and are simple to use. Developers that are unfamiliar with dialogue models could conveniently use the base models for their custom datasets.
# Reinforcement Learning Toolkit
Compared to ConvLab-2, we improve the reinforcement learning toolkit for the dialogue policy module in ConvLab-3. We simplify the process of building the dialogue system and its RL environment, extend the modularity of the vectorizer codebase, incorporate multiple US, and offer a wide range of evaluation metrics.
## Pipeline Setup
A dialogue system agent can be comprised of a wide range of sub-modules as mentioned in Table [\[tab:tasks\]](#tab:tasks){reference-type="ref" reference="tab:tasks"}. ConvLab-3 fully supports the straightforward building of any combination and allows easy substitution of components. This also includes the definition of the agent environment given by the choice of user policy and its components. In ConvLab-3 various components can be easily combined using an easy-to-use configuration file.
## Vectorizer Class {#subsection:vectoriser_class}
The dialogue policy module trained using RL obtains the output of the DST as input and typically outputs a semantic action. Just as ConvLab-2, ConvLab-3 uses a vectorizer class as a communication module that translates the semantic actions and DST output into vectorized representations using the given ontology. We extend and modify that module and make the vectorizer easily interchangeable in order to use different vectorization strategies. Firstly, we add the possibility for masking certain actions as in to aid policy learning or to prevent the policy from taking invalid actions. Secondly, the module leverages the functionalities described in Section [2](#unified-data-format-section){reference-type="ref" reference="unified-data-format-section"} to allow easy adaptation to new datasets. As the vectorizer module takes care of producing a suitable input representation, it is possible to use any off-the-shelf neural network architecture for the policy module and only modify the vectorizer to fit the required input. As an example, defines a belief tracking module that outputs a semantic state accompanied by relevant uncertainty scores, which are used as input to the policy. In order to implement the change, the policy module can remain untouched while only a new vectorizer (inheriting from the base class) needs to be implemented with one required class method. This simplifies the integration of new policy modules and vectorization strategies.
## Evaluation Tools
The core metrics for a dialogue policy trained using RL are *success* and *return*. Nevertheless, as these policies are built for interacting with humans, it is necessary to provide an in-depth analysis of their behavior. For instance, is the policy providing too many actions in a turn, which a simulator can handle well but lead to an information overload for real humans? Does the policy have a tendency to only use some specific intents but neglect others? How are the actions taken by the system distributed in general? In order to easily answer these questions, we provide plotting tools to conveniently compare algorithms with each other, both in terms of performances and action distributions. Moreover, ConvLab-3 provides an evaluation function to inspect generated dialogues, thereby enabling full understanding of the dialogue policy behavior. We will showcase the additional analysis tools in Section [6.2](#rl-analysis){reference-type="ref" reference="rl-analysis"}. Lastly, we use a stricter form of task success that we call *strict success*, which extends the previous success definition. While the previous version checked whether the necessary information was provided and an entity was booked if required, it did not check whether the entity fulfills the booking constraints. As a consequence, it was possible to book an entity for a wrong day, with a wrong number of people but still be successful. In summary, these changes allow easy building, extension, and evaluation of dialogue policies, making them more accessible to researchers.
# Experimental Setup
To show how the new features of our platform may promote TOD research, we conduct comprehensive experiments, which can be categorized into reinforcement learning for the dialogue policy module and supervised learning for other modules.
## Supervised Learning
Based on the unified data format, conducting experiments on multiple TOD datasets is convenient. We believe this feature will encourage researchers to build general dialogue models that perform well on different datasets and investigate knowledge transfer between them. To provide insight into how to transfer knowledge from other datasets, we explore three typical transfer learning settings on various tasks, namely (1) pre-train on other datasets and then fine-tune on the target dataset, (2) joint training on the target dataset and other datasets, and (3) retrieve samples that are similar to the samples of the target dataset from other datasets as additional training data or as in-context learning examples.
#### Datasets
We use MultiWOZ 2.1, the Schema-Guided Dialogue (SGD) dataset, and Taskmaster-1/2/3 for our experiments since they are large and have rich annotations. We regard MultiWOZ 2.1 as the target dataset.
#### Metrics
We evaluate our newly integrated models covering various tasks with classic metrics: turn accuracy (ACC) and dialogue act F1 score for NLU, joint goal accuracy (JGA) and slot F1 score for DST, BLEU and slot error rate (SER) for NLG, BLEU and Combined score (Comb.) for End2End, turn accuracy and slot-value F1 score for user simulators that output user dialogue acts only (US-DA), and slot-value F1 score and SER for user simulators that output both user dialogue acts and response (US-NL). Note that the calculation of Combined score for End2End is not the exactly same as those in previous works because we use our own dataset-agnostic delexicalization instead of the MultiWOZ-specific one. Since SGD and Taskmaster do not have goal annotations, we calculate the turn-level slot-value F1 score between original responses and generated responses instead of Combined score on these datasets. For NLG, we ignore utterances with empty dialogue acts.
### Pre-training then Fine-tuning
In this experiment, we explore knowledge transfer through pre-training on other TOD datasets before fine-tuning. We pre-trained models on SGD and Taskmaster datasets and then fine-tuned the models on MultiWOZ in full-data or low-resource (1%, 10% of the entire training set) settings. As shown in Listing [\[listing:util\]](#listing:util){reference-type="ref" reference="listing:util"}, it is easy to control the data to use in few-shot learning experiments through the data loading function. For low-resource fine-tuning, we set the data ratios of both training and validation set to 1%/10% and run the training 3 times with different subsets of samples. If not mentioned otherwise, for existing models we follow the settings of their original papers. We use `T5-Small` for T5-NLU, T5-DST, and T5-NLG. We initialize SC-GPT and SOLOIST with `GPT2-Medium` and `T5-Base` respectively instead of their checkpoints pre-trained on many TOD datasets. For pre-training, we merge SGD and Taskmaster datasets. Training details and hyper-parameters can be found in our GitHub repository.
### Joint Training
In this experiment, we investigate the effect of training a model on multiple datasets jointly instead of separately. For joint training, we merge MultiWOZ, SGD, and Taskmaster datasets into one and train a single model, which requires the model to handle datasets with different ontologies. Intuitively, the advantage of joint training is that knowledge transfer is bi-directional and persists for the whole training period, while the disadvantage is that there may be inconsistent labels for similar inputs on different datasets, potentially confusing the models. To avoid confusion, for T5-NLU, T5-DST, and T5-NLG, we prepend the dataset name to the original input to distinguish data from different datasets. For SetSUMBT, we only predict the state of the target dataset. Since SGD may have several services for one domain, we normalize the service name to the domain name (e.g., `Restaurant_1` to `Restaurant`) when evaluating NLU and DST. However, similar slots of different services (e.g., `city` and `location`) will still confuse the model. While further normalization may help, we are aiming to compare independent training and joint training instead of achieving SOTA performance. For Taskmaster-1/2/3, we evaluate each sample with the corresponding ontology and then calculate the metrics on all test samples of three datasets. In addition, on SGD and Taskmaster, we build pseudo user goals for TUS and GenTUS by accumulating constraints and requirements in user dialogue acts during conversations.
### Retrieval Augmentation
We further explore transferring knowledge from other datasets through retrieval-based data augmentation. Here we only consider the single-turn NLU task where the input is an utterance, since utterance-level similarity is easier to model than dialogue-level similarity. For each utterance in the target dataset, we retrieve the top-k (k \(\in \{1,3\}\)) most similar utterances from other datasets measured by the MiniLMv2 model using Sentence Transformers We then use retrieved samples in two ways:
1. Augment training data. Models are trained on both original training data and retrieved data.
2. Additionally input the retrieved samples as in-context examples, including retrieved utterances and their dialogue acts, as shown in Table [2](#tab:in-context){reference-type="ref" reference="tab:in-context"}. Different from, the retrieved samples are from other datasets instead of the target dataset and we will train models on augmented samples.
Since different datasets have different ontologies (i.e., definitions of intent, domain, slot), we prepend the corresponding dataset name to an input utterance as in the joint training experiment. We use the T5-NLU model and try two model sizes `T5-Small` and `T5-Large`. We fine-tune the models on MultiWOZ using the same settings as in the pre-training-then-fine-tuning experiment.
[\[tab:in-context\]]{#tab:in-context label="tab:in-context"}
## Reinforcement Learning
In order to showcase our RL toolkit, we conduct several experiments. The first set of experiments leverages the convenient unified format for pre-training policies before starting the RL training. The second set of experiments regards training with multiple user simulators as well as incorporating uncertainty estimates in state representation.
### Pre-training then RL Training
In this experiment, we simulate different data scenarios before starting the RL training. We leverage the implemented DDPT policy model for training as its flexibility allows transfer from one ontology to another. We test four different scenarios:
1. **Scratch**: the policy starts the RL training with random initialization. This simulates the zero data scenario.
2. **SGD**: the policy is pre-trained on the full SGD data before the RL training starts. This simulates having related data.
3. **1%MWOZ**: the policy is pre-trained on 1% of MultiWOZ 2.1 data. This simulates having scarce in-domain data.
4. **SGD-\>1%MWOZ**: the policy is pre-trained on the full SGD data, then fine-tuned on 1% of MultiWOZ 2.1 data before the RL training starts. This simulates having related data as well as scarce in-domain data.
The unified format conveniently allows us to pre-train on different datasets by only specifying the dataset name. The experiments are conducted on the semantic level, leveraging the rule-based dialogue state tracker and the rule-based user simulator of ConvLab-3.
### Training with Different User Simulators
To ensure a policy learns a general skill that can solve the task with diverse users instead of overfitting to a certain user simulator, it is important to test the policy with various user simulators. ConvLab-3 provides the SOTA user simulators TUS and GenTUS and allows easy usage through the simplified pipeline setup. In this set of experiments, we compare TUS, GenTUS and the rule-based simulator ABUS by training a dialogue policy in interaction with each of them and performing cross-model evaluation afterwards. The dialogue policy used is a simple multilayer perceptron and optimized using PPO. The simulators and policy are pre-trained on MultiWOZ 2.1 and experiments are run on the semantic level using the rule-based dialogue state tracker.
### Training with Uncertainty Features
The simplified vectorizer module in ConvLab-3 makes it possible to easily include dialogue related features that might benefit dialogue policy learning. One such example is given due to the problem of resolving ambiguities in conversations. Humans naturally identify these ambiguities and resolve the uncertainty resulting from them. For a dialogue system to be robust to ambiguity, it is crucial to identify and resolve these uncertainties. ConvLab-3 provides the SetSUMBT dialogue belief tracker which achieves SOTA performance in terms of the accuracy of its uncertainty estimates. Using the vectorizer class to incorporate these features we can train a policy using the uncertainty features obtained from SetSUMBT. To illustrate the effectiveness of uncertainty features during RL, we train a PPO policy using these features. The template-based NLG in ConvLab-3 also allows for the inclusion of noise in generated responses which allows for uncertainties to arise during the conversation, simulating a more realistic conversation.
# Result Analysis
## Supervised Learning
### Pre-training then Fine-tuning
[\[tab:transfer\]]{#tab:transfer label="tab:transfer"}
We compare direct fine-tuning and pre-training-then-fine-tuning in Table [\[tab:transfer\]](#tab:transfer){reference-type="ref" reference="tab:transfer"}. Generally, pre-training is beneficial, especially in the low-resource setting. However, as the fine-tuning data size increases, the benefit gradually decreases in most cases. When fine-tuning on the full MultiWOZ, whether or not pre-training is conducted does not matter for NLU, DST, and NLG models, while for the End2End model SOLOIST, pre-training can still improve performance. The reason might be that end-to-end modeling is a more complex task, such that the full MultiWOZ data is still not enough to reach saturated performance. Compared with other models, the user simulators TUS and GenTUS behave differently. Pre-training largely boosts the performance of TUS because it uses domain-agnostic features for semantic actions, thus it can focus on the cross-domain user behavior on the semantic level. Pre-training-then-fine-tuning on 1% MultiWOZ data is even better than directly fine-tuning on the full MultiWOZ data. For GenTUS, the observation is quite different. With pre-training, the F1 score is improved in the 1% data fine-tuning setting but is largely restricted when more fine-tuning data is available, while the SER is consistently improved. Note that SER is calculated using self-generated dialogue acts instead of the golden ones, measuring the faithfulness of utterance generation. This indicates that pre-training biases GenTUS towards a user behavior that differs from users in MultiWOZ but makes utterance generation more accurate.
[\[tab:multitask\]]{#tab:multitask label="tab:multitask"}
### Joint Training
We compare independent training and joint training in Table [\[tab:multitask\]](#tab:multitask){reference-type="ref" reference="tab:multitask"}. MultiWOZ, SGD, and Taskmaster have 8K, 16K, and 43K dialogues for training respectively. Joint training on these datasets does not lead to substantial performance drops in most cases, indicating that models have sufficient capacity to encode knowledge of different datasets simultaneously. However, joint training does not always improve performance either. It consistently improves the End2End model SOLOIST but makes no difference to T5-NLU. For other models, the gains vary with the dataset. Associating with the previous pre-training-then-fine-tuning experiment, we think the difference may be attributed to the varying task complexity on different datasets. When the original data of a certain dataset are sufficient for a model to solve the task, including other datasets via joint training may not bring further benefit.
### Retrieval Augmentation
[\[tab:retrieval\]]{#tab:retrieval label="tab:retrieval"}
We compare retrieval augmentation methods with direct fine-tuning (Baseline) and pre-training-then-fine-tuning as shown in Table [3](#tab:retrieval){reference-type="ref" reference="tab:retrieval"}. We can see that the performances of different models mainly differ in the 1% data setting, and retrieving top-1 or top-3 samples does not have a large effect. In the 1% data setting, using retrieved samples to augment inputs for training is even worse, while augmenting the training set with retrieved samples is beneficial for `T5-Small` but disadvantageous for `T5-Large`. Using other TOD datasets for pre-training is better than or on par with using them for two retrieval augmentation methods and direct fine-tuning. The reason may be that different datasets' ontologies for similar utterances consistently confuse the models when training on retrieved samples or augmented input, while pre-trained models will no longer be distracted by other datasets' ontologies during fine-tuning. From the example in Table [2](#tab:in-context){reference-type="ref" reference="tab:in-context"}, although retrieved samples may extract important values, they assign different intent, domain, and slot names to these values, which are not transferable to the target dataset, calling for a better way to use other TOD datasets.
## Reinforcement Learning {#rl-analysis}
### Pre-training Then RL Training
Let us first take a look at the metrics *strict success rate* and *number of turns*, plotted in Figure [\[pretrain-rl\]](#pretrain-rl){reference-type="ref" reference="pretrain-rl"}(a) and (b). We can observe that pre-training on SGD does not yield an advantage for the starting performance, compared to training from scratch. Nevertheless, looking at the final performance, we can see that the model pre-trained on SGD is different from the scratch model and reaches the same final performance as when MultiWOZ data is used. We hypothesize that pre-training on SGD effectively puts the initial policy parameters in an appropriate parameter space for dialogue learning. We can also observe that it does not seem to make a difference whether or not the policy is pre-trained on SGD before being fine-tuned on MultiWOZ, which suggests that the SGD information has been overwritten. Figure [\[pretrain-rl\]](#pretrain-rl){reference-type="ref" reference="pretrain-rl"}(c) shows the average number of actions taken in a turn, which is an important evaluation measure, as exposing a user to too much information can lead to information overload. Interestingly, the model only pre-trained on SGD initially takes much fewer actions in a turn, which exactly reflects what can be observed in the SGD data. In addition to the plots depicted in Figure [\[pretrain-rl\]](#pretrain-rl){reference-type="ref" reference="pretrain-rl"}, the new toolkit allows us to have a look at the average intent distribution within a turn. Figure [\[pretrain-rl-distributions\]](#pretrain-rl-distributions){reference-type="ref" reference="pretrain-rl-distributions"} (a), (b) and (c) depict the average probability of taking a request, offer, or inform intent in a turn, respectively. We can see that SGD pre-training initially leads to many offer intents, while mostly ignoring request intents compared to pre-training on MultiWOZ. The number of inform intents is in a similar range for all models apart from the scratch model. Nevertheless, the probabilities eventually converge to the same level, which generally holds for all intents.
[\[tab:cross-model\]]{#tab:cross-model label="tab:cross-model"}
### Training with Different User Simulators
The result of strict success rates of PPO policies trained with ABUS, TUS, or GenTUS is shown in Table [\[tab:cross-model\]](#tab:cross-model){reference-type="ref" reference="tab:cross-model"}. The policy trained with ABUS only outperforms policies trained with the other USs when evaluated with ABUS. On the other hand, the policy trained with GenTUS outperforms the other policies when evaluated not only with GenTUS, but also with TUS. It also gets comparable performance with policies trained with ABUS when evaluated with ABUS, indicating that GenTUS trains the most robust policy. According to the result, a policy that performs well when evaluated with the US it is trained on does not necessarily perform well on other USs. suggest evaluating a policy with different types of USs to provide a more holistic view and show the average performance correlates well with human evaluation. In ConvLab-3, we provide not only the rule-based US, i.e., ABUS, but also data-driven USs, e.g., TUS and GenTUS, to build a richer environment for training and evaluation. Moreover, with the two transferable USs TUS and GenTUS, researchers can easily evaluate their policy on custom datasets.
### Training with Uncertainty Features
Figure [\[rl-uncertainty\]](#rl-uncertainty){reference-type="ref" reference="rl-uncertainty"} (a) reveals that the policy trained using uncertainty features performs at least as well as the policy trained without these features. further showed that the policy trained with uncertainty features performs significantly better in conversation with humans than the policy trained without. This is an indication that the policy using uncertainty features can handle ambiguities in conversation better than the policy without uncertainty modeling. To investigate how this policy resolves uncertainty we analyze the action distributions of the policy using the new RL toolkit evaluation tools, which provide new insights into the behavior of dialogue policy modules. Figure [\[rl-uncertainty\]](#rl-uncertainty){reference-type="ref" reference="rl-uncertainty"} (b) and (c) show that the policy trained using uncertainty features utilizes significantly more request actions than the policy without these features. This indicates that the policy aims to resolve uncertainty by requesting information from the user. For instance, if the policy recognizes uncertainty regarding the price range a user has requested, it can resolve this through the use of a request. See for example dialogues with humans where this can be observed.
# Conclusion
In this paper, we present the dialogue system toolkit ConvLab-3, which puts a large number of datasets under one umbrella through our proposed unified data format. The usage of the unified format facilitates comparability and significantly reduces the implementation cost required for conducting experiments on multiple datasets. In addition, we provide recent powerful models for all components of a dialogue system and improve the RL toolkit which enables researchers to easily build, train and evaluate dialogue systems. We showcase the advantages of the unified format and RL toolkit in a large number of experiments, ranging from pre-training, joint training, and retrieval augmentation to RL training with different user simulators and state uncertainty estimation. We hope that ConvLab-3 can aid and accelerate the development of models that can be trained with different datasets and learning paradigms, empowering the community to develop the next generation of task-oriented dialogue systems.
#### Code implementation
Qi Zhu proposed and implemented the unified data format, transformed all datasets, adapted BERTNLU and MILU, and implemented T5 series models. Christian Geishauser was the main contributor to the RL toolkit and adapted DDPT, PPO, PG, and MLE models. Hsien-chin Lin adapted TUS and GenTUS, allowing training policy with different user simulators. Carel van Niekerk adapted SUMBT and SetSUMBT, contributed to the RL toolkit, and incorporated uncertainty estimates with policy learning. Baolin Peng adapted SOLOIST and implemented End2End evaluation metrics. Zheng Zhang adapted SC-GPT and implemented NLG evaluation metrics. Michael Heck adapted TripPy. Nurul Lubis adapted LAVA. Dazhen Wan contributed to user simulators and SC-GPT. Xiaochen Zhu contributed to the RL toolkit and SC-GPT. Experiments of each model were performed by the corresponding author.
#### Paper writing
Qi Zhu, Christian Geishauser, Hsien-chin Lin, and Carel van Niekerk finished most part of the paper. Baolin Peng and Zheng Zhang added the necessary details. Michael Heck, Nurul Lubis, and Jianfeng Gao proofread and polished the paper. | {'timestamp': '2022-12-01T02:19:07', 'yymm': '2211', 'arxiv_id': '2211.17148', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17148'} |
null | null |
null | null |
null | null |
# Introduction
The last decade has seen two major breakthroughs in the field of experimental astronomy. First, the detection in 2015 of gravitational waves from a binary black hole merger heralded the arrival of gravitational wave astronomy as an experimental science. Second, the picture from the shadow of the M87 supermassive black hole from the Event Horizon Telescope provided the first direct observation of a black hole shadow. By providing the first direct time and spatial signals from black holes, these measurements have taken black holes from a purely theoretical realm into concrete physical objects in our matter-filled, inhomogeneous Universe. Hence, it is now necessary to study black holes not only as isolated objects but also within their environment. This is by no means a new consideration and there is extensive literature on the study of black hole embedded into rich and varied environments (see for example the review articles , motivated by the detection of gravitational waves). A particular emphasis was put on the impact of dark matter surrounding black holes and their imprint of gravitational waveforms, (however see also, e.g., for studies of the interaction of black holes with dark energy). Studies of black holes surrounded by matter have focused on two directions: i) the environmental impact on the inspirals of compact objects and ii) the ringdown emission . The consensus on the latter was that while the resonance spectrum of black holes surrounded with extra structures could be widely different from that of isolated black holes, the modifications were somehow irrelevant for practical considerations and could not be seen. Interestingly, this conclusion is being revisited with the recent identification and characterisation of the so-called QNMs spectral instability . In the present study, we consider the particular case of what is referred to in the literature as a *dirty black hole* (DBH), that is, a black hole surrounded by a thin-shell of matter. DBHs were introduced as practical toy models to investigate the impact of a local environment on black hole effects. Surprisingly and despite the many prospects offered by gravitational wave astronomy, the quasinormal mode (QNM) spectrum of DBHs has received very limited attention. Generic properties of the QNM spectrum were discussed in and the concrete case of a DBH was considered by Leung et al.İn their study, Leung et al. adopted a perturbative approach which allowed for an analytical investigation but restricted their conclusion to DBH spacetimes where the shell had a limited impact. Similar DBH configurations (that is with the shell having a limited impact on the scattering) were investigated recently and the impact of the shell on both the absorption and scattering cross section were computed. These recent studies as well as the need for concrete resonance calculations beyond the perturbative regime are the main motivations for this study. There are two main goals of this paper: First, to investigate wave scattering by DBHs beyond the perturbative cases found in the literature. We therefore consider physical DBH configurations where the shell has a substantial effect on the scattering of waves leaving an imprint on measurable quantities. Second, to compute the spectrum of resonances of DBHs from the Regge pole (RP) paradigm. RPs and the associated complex angular momentum (CAM) technique are the counterparts to QNMs. While QNMs have a real angular momentum and a complex frequency, the RPs have a real frequency and a complex angular momentum. This paradigm allows for a different interpretation of resonances in a system. The RP and CAM approaches have shed new light in many different domains of physics involving resonant scattering theory, notably, in quantum mechanics, electromagnetism, optics, nuclear physics, seismology and high-energy physics (see, for example, [@deAlfaro:1965zz; @Newton:1982qc; @Watson18; @Sommerfeld49; @Nussenzveig:2006; @Grandy2000; @Uberall1992; @AkiRichards2002; @Gribov69; @Collins77; @BaronePredazzi2002; @DonnachieETAL2005] and references therein), and have been successfully extended to black hole physics [@Andersson_1994; @Andersson_1994b; @Decanini:2002ha; @Decanini:2011xi; @Folacci:2018sef; @Folacci:2019cmc; @Folacci:2019vtt]. As an illustration of the power of the CAM approach, we note that it provides a unifying framework describing the glory and orbiting effects of black hole scattering. The paper is organised as follows: in Sec. [2](#sec:DBH_spacetime){reference-type="ref" reference="sec:DBH_spacetime"}, we review the derivation of the DBH spacetime and show that there exist static configurations fully characterised by the mass of the shell and its equation of state; in Sec. [3](#sec:geodesics){reference-type="ref" reference="sec:geodesics"} we give a qualitative description of the geodesic motion in DBH spacetimes which paves the way for the future interpretation of our results; in Sec. [4](#sec:waves_in_DBH){reference-type="ref" reference="sec:waves_in_DBH"} we review the description of scalar wave propagation on a DBH spacetime. Sec. [5](#sec:resonances){reference-type="ref" reference="sec:resonances"} contains the main results of the paper, that is, the calculation of the Regge pole spectrum of DBH's. We reveal in particular that in contrast with the isolated black hole case, the DBH resonance spectrum contains several branches which correspond to the extra structure and complexity of the spacetime. The spectra are computed numerically by adapting the continued fraction method originally developed by Leaver. We supplement our analysis with a WKB calculation to verify our numerical results. In Sec. [6](#sec:scattering){reference-type="ref" reference="sec:scattering"}, we compute numerically the scattering cross section of planar waves impinging on a DBH and show that characteristic oscillations appear in some configurations. The cross sections are then described using the RP spectrum identified in Sec. [5](#sec:resonances){reference-type="ref" reference="sec:resonances"} via the CAM representation. Finally we conclude with a discussion of our results and future prospects in Sec. [7](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}. We use natural units \((c=G=\hbar=1)\) throughout the paper.
# Dirty Black Hole spacetime {#sec:DBH_spacetime}
Here we briefly review the Israel formalism applied to our case where there are two distinct Schwarzschild geometries: \[ds^2 = f_\pm(r) dt_\pm^2-\frac{dr^2}{f_\pm(r)}-r^2 d\Omega^2\] where \(d\Omega^2 = d\theta^2+ \sin^2\theta d\phi^2\) is the line element on a unit sphere, \(f_\pm(r) = 1-2M_\pm/r\) is the Schwarzschild potential for each side of the shell, and we have already applied the knowledge that the shell is at a given radius \(r=R_s(\tau)\), which implies that the angular and radial coordinates are the same on each side. At this point, we are not assuming the wall is static, hence there is a different local time coordinate on each side of the shell, which has, in general, a time dependent trajectory given by \(\left (t_\pm(\tau),R_s(\tau)\right)\) with \(\tau\) the proper time on the shell: \(f_\pm {\dot t}_\pm^2-\dot{R}_s^2/f_\pm = 1\). The Israel junction conditions read \[\Delta K_{ab}-\Delta K h_{ab} = 8\pi S_{ab} = Eu_au_b-P \left ( h_{ab}-u_au_b \right)\] where \(h_{ab} = g_{ab} + n_a n_b\) is the induced metric on the shell (with \(n_\pm^a = ( \dot{R}_s, \dot{t}_\pm )\) an outward pointing unit normal on each side of the shell), \(\Delta K_{ab} = K^+_{ab}-K^-_{ab}\) is the jump in extrinsic curvature across the shell, and \(S_{ab} = \int_-^+ T_{ab}\) is the energy momentum of the wall, here assumed to take a perfect fluid form. Inputting the form of the geometry into the Israel equations results in two independent "cosmological" equations \[\begin{aligned} \dot{R}_s^2 + 1 &= (4\pi E)^2R_s^2 + \frac{(M_++M_-)}{R_s} + \frac{(M_+-M_-)^2}{(8\pi E)^2R_s^4}, \\ \dot{E} &+ 2 \frac{\dot{R}_s}{R_s} (E+P) =0. \end{aligned}\] A static shell requires both \(\dot{R}_s\) and \(\ddot{R}_s\) to be zero, which places constraints between the values of the mass and shell energy-momentum. As is conventional, we assume an equation of state for the shell \[P=wE\;, \qquad \Rightarrow \qquad E = \frac{\rho_0}{4\pi R_s^{2(1+w)}}\] which then gives \[\begin{aligned} M_\mp &= \frac{4w(1+2w)R_s-(1+4w)^2M_\pm}{f_\pm(R_s)(1+4w)^2},\\ \rho_0^2 &= \frac{4R_s^{4w}(2wR_s-(1+4w)M_\pm)^2}{f_\pm(R_s)(1+4w)^2}. \end{aligned}\] For a given equation of state, the minimum value of \(R_s\) is when the shell has vanishing energy (and \(M_+=M_-\) trivially) \[R_{s,min} = \frac{(1+4w)}{2w} M_\pm. \label{minshell}\] As \(R_s\) increases, the shell gradually contributes more mass to the spacetime, leading to a larger disparity between \(M_+\) and \(M_-\). (If \(R_s<R_{s,min}\), \(M_+<M_-\), and the energy of the shell is negative.) An important feature of a black hole spacetime of relevance to QNM's and RP's is the *light ring*, or the unstable null circular geodesic orbit at \(r =3M\). We now see three different possibilities for the shell location depending on where it is situated with respect to the light-rings of the interior and exterior Schwarzschild masses (\(3M_\pm\)). (i) \(R_s < 3M_-\): If \(w>1/2\), i.e. the equation of state is stiffer than radiation, then [\[minshell\]](#minshell){reference-type="eqref" reference="minshell"} shows that it is possible for the shell to lie 'inside' the light-ring of the black hole. Since the local Schwarzschild mass is \(M_+\) outside the shell, this means that the \(3M_-\) light-ring no longer exists, and the spacetime will have only one light-ring at \(r=3M_+\). Conversely, if \(w\leq1/2\), then [\[minshell\]](#minshell){reference-type="eqref" reference="minshell"} shows that the shell can never lie inside the inner light-ring. (ii) \(3M_-\leq R_S \leq 3M_+\): This configuration, where the shell lies between the light-rings, is possible for equations of state with \(w\in(\frac{\sqrt{3}-1}{4},1]\), and the spacetime will have both light-rings present. (iii) \(R_s>3M_+\): If the shell lies outside the light-ring of the exterior geometry, then again there is only one light-ring in the spacetime, although it is now *inside* the shell at \(3M_-\). Demanding positivity of \(R_s-3M_+\) for \(R_s\geq R_{s,min}\) shows that this requires the equation of state parameter \(w<1/2\). Having derived all the possible configurations of the shell, we conclude this section by altering our notation slightly to align with that of Macedo et al. in which a global coordinate system \((t,r,\theta,\phi)\) is used: \[\label{lineele} ds^2 = A(r)dt^2-B(r)^{-1}dr^2-r^2d\Omega^2.\] Note, we did not begin with this global system as it is not possible to define a global time coordinate unless the shell is static, and we wanted to derive all possibly consistent static configurations commensurate with physical equations of state. Replacing \(M_-= M_{\text{\tiny BH}}\) and \(M_+ = M_\infty\), these metric functions are: \[\begin{aligned} A(r) &= \begin{cases} \alpha(1-2M_{\text{\tiny BH}}/r), & r < R_s \\ 1-2M_{\infty}/r, & r > R_s, \end{cases} \\ B(r) &= \begin{cases} 1-2M_{\text{\tiny BH}}/r, & r < R_s \\ 1-2M_{\infty}/r, & r > R_s, \end{cases} \end{aligned} \label{AandB}\] where \(\alpha\) is inserted to ensure the induced metric on the shell is well defined from each side \[\alpha = \frac{R_s-2M_\infty}{R_s-2M_{\text{\tiny BH}}} \;,\] and \(B\) contains a jump at the shell to represent the discontinuity in the extrinsic curvature. A cartoon of the spacetime geometry is given in Fig. [\[fig:Schematic_diagram_DBH\]](#fig:Schematic_diagram_DBH){reference-type="ref" reference="fig:Schematic_diagram_DBH"}.
# Geodesic motion {#sec:geodesics}
DBH spacetimes can exhibit critical effects of geometrical optics. We will see later that this rich geodesic structure can be used to understand features in the scattering of waves presented in this paper. The important notion emerging from the geodesic motion relevant to the study of resonances is the *light-ring* or *photon sphere*. It corresponds to a local maximum and an unstable equilibirum point of the effective geodesic potential and offers an intuitive link to QNMs . It is therefore natural to ask about the structure of the light-rings in DBH spacetimes. We note that the link between QNMs and light-rings is however not exact and a simple geodesic analysis is not always sufficient to characterise the resonances of a system, as we will see later. Geodesic motion on static DBH spacetimes has been studied in, e.g.,, hence we simply note the key steps of the calculation. The equation of motion of a null geodesic can, by a judicious choice of affine parameter, be brought into the form \[\label{eq:geo_pot} \dot{r}^2 + B(r) \left ( \frac{L^2}{r^2}-\frac{1}{A(r)}\right) = \dot{r}^2 + U_{\text{\tiny{eff}}}(r) = 0\] where without loss of generality the geodesic is taken to lie in the equatorial plane, \(A\) and \(B\) were given in [\[AandB\]](#AandB){reference-type="eqref" reference="AandB"}, and \(L = r^2\dot{\phi}\) is a conserved (not necessarily integer) quantity along the geodesic related to the angular momentum (the affine parameter is chosen so that \(A\dot{t}=1\)). The discontinuity of \(B\) across the shell means this potential is also discontinuous, however the interpretation of the level of the potential as a "kinetic" energy remains, and we can use the graph of \(U_{\text{\tiny{eff}}}(r)\) to not only interpret geodesic motion, but also to infer more general scattering phenomena. The main feature of the potential, resulting from the discontinuity, is the possible existence of two local maxima. A further interesting feature of these geometries is the possibility of geodesics that remain trapped between the light-rings. While these obviously will not be visible from far away, we might expect the existence of these resonant trajectories to correspond to some imprint in the scattering cross sections. Indeed, even for the case of a single light-ring, when that is located inside the shell, we see the sharp discontinuity in the potential at the shell location gives rise to a sharp "dip" in the potential (see Fig. [\[fig:geodesic_potential\]](#fig:geodesic_potential){reference-type="ref" reference="fig:geodesic_potential"}). While this dip results in a change of direction for the light ray, it does not 'trap' the (classical) geodesic, however a quantum mechanical system would display an effect. The scattering of scalar waves of the dirty black hole might therefore be expected to detect this dip. As we will see in Sec. [5](#sec:resonances){reference-type="ref" reference="sec:resonances"}, the properties of the local maxima of the geodesic potential can be seen in the resonance spectrum leading to two separated branches. We also identify a third branch related precisely to the dip referred to above-indicative of quasibound states trapped between the light-ring and the shell. Before turning our attention to the spectrum of resonances of DBH, we first anticipate the effects expected to be present in the scattering based on the geodesic structure presented above. The link between geodesic motion and scattering effect is best seen through the deflection angle which we now discuss.
## Deflection angle and classical scattering
The deflection angle is an important geometrical quantity that allows us to analyze in the classical limit the differential scattering cross section. In the case of DBH, the classical scattering cross section for null geodesics can be defined as (see also ) \[\label{classical_scatt_cross_sec} \frac{d\sigma}{d\Omega} =\frac{b}{\sin(\theta)}\frac{1}{\Big{|} \frac{d\Theta_\text{geo}(b)}{db}\Big{|}}\] and the geodesic deflection angle is given by \[\label{Deflection_angle} \Theta_\text{geo}(b) = 2\int_{0}^{u_0} du\left[B(u)\frac{1-A(u)b^2u^2}{A(u)b^2}\right]^{-1/2}-\pi,\] where \(u=1/r\) and \(u_0 = 1/r_0\) with \(r_0\) being a turning point. In Eqs. [\[classical_scatt_cross_sec\]](#classical_scatt_cross_sec){reference-type="eqref" reference="classical_scatt_cross_sec"} and [\[Deflection_angle\]](#Deflection_angle){reference-type="eqref" reference="Deflection_angle"}, the geodesic deflection angle is related to the scattering angle \(\theta\) by \[\label{rela_def_angle_scatt_angle} \Theta(b) +2n\pi =\pm \theta\] with \(n \in \mathbb{N}\) such that \(\theta\) remains in "its interval of definition", i.e. \([0,\pi]\) and \(b\) is the impact parameter of the scattered null geodesic. There are four main scattering effects, each associated with a specific property of the deflection angle that we briefly review for the reader before qualitatively describing the scattering of waves in DBH spacetimes. They can be cast into two classes: the divergent class or the interference class.
### The divergent class: glories and rainbows
*-The glory:* If the deflection function passes smoothly through \(0\) or \(\pi\), i.e. if geodesics are scattered in the forward or backward direction, then the semi-classical cross section contributions from these geodesics will diverge. Expanding the deflection function around the glory point to linear order, one can cure this singularity and obtain a semi-classical expression for the glory in terms of Bessel functions. The typical behaviour associated with the glory is therefore an increase of the scattering amplitude in the forward or backward direction. The glory effect is well known in black hole physics and also appears in these DBH spacetimes. Fig. [\[fig:critical_geo\]](#fig:critical_geo){reference-type="ref" reference="fig:critical_geo"} a) shows one backscattered ray contributing to the glory effect in the case of a DBH studied in the following. *-Rainbow scattering:* From [\[classical_scatt_cross_sec\]](#classical_scatt_cross_sec){reference-type="eqref" reference="classical_scatt_cross_sec"}, we can see that \(d\sigma/d\Omega\) can diverge for \(\theta\neq 0\) or \(\pi\) if \(d\Theta_{geo}/db = 0\). An impact parameter \(b_r\), for which the deflection function is stationary defines an extremal angle, called the rainbow angle \(\theta_r\). For \(b \sim b_r\), no rays can be deflected beyond \(\theta_r\). This will result in interferences on one side of the rainbow angle (known as the illuminated side) and to exponential decay on the other (known as the dark side). Similarly to the glory, one can expand the deflection function to second order in the vicinity of the rainbow point to obtain a semi-classical description built on the Airy function. The typical behaviour of rainbow scattering is an enhanced amplitude near the rainbow angle with oscillations on one side of the rainbow and exponential decay on the other side. This effect was shown to appear in astrophysical settings as well as in gravitational analogues. Fig. [\[fig:critical_geo\]](#fig:critical_geo){reference-type="ref" reference="fig:critical_geo"} b) shows rainbow rays in the case of a DBH studied in the following.
### The interference class: orbiting and grazing
*-Orbiting:* Orbiting belongs to the interference class of critical effects and is associated with a critical impact parameter \(b_c\) for which the deflection angle diverges \(\lim_{b\rightarrow b_c}\Theta_{geo} =\infty\). This implies that geodesics with an impact parameter close to \(b_c\) can be deflected with arbitrary angles. In other words, they orbit around the scatterer. This divergence also implies that there will be infinitely many geodesics deflected at any angles, causing interference at arbitrary angles. Orbiting is well known in black hole physics and is associated with the existence of the light-ring previously mentioned. Fig. [\[fig:critical_geo\]](#fig:critical_geo){reference-type="ref" reference="fig:critical_geo"} c) shows the orbiting of geodesics in a DBH spacetime. Since orbiting allows geodesics to be deflected to arbitrary angles, it is naturally linked to the glory effect. Indeed it was shown, using the CAM approach, that the two effects can be incorporated in a unified semi-classical formula built around the properties of surface waves propagating on the light-ring . *-Grazing:* Grazing is another critical effect belonging to the interference class. Grazing (also known as edge effects) appear when the scatterer exhibits a discontinuity. The discontinuity of the scatterer will lead to a singular point in the deflection function which defines an extremal angle beyond which rays cannot be deflected. This is similar to the rainbow effect with the main difference being the fact that the deflection function is not stationary at the grazing angle. Therefore, grazing will lead to interferences but does not necessarily imply an increase of the scattering amplitude as was the case in rainbow scattering. Fig. [\[fig:critical_geo\]](#fig:critical_geo){reference-type="ref" reference="fig:critical_geo"} d) shows grazing rays in a DBH spacetime studied in the following.
## Critical effects in DBH spacetimes
Equipped with the terminology and intuition from the semi-classical description of critical effects, we now turn our attention to scattering in DBH spacetimes. As we noted in the previous section, there are three main geometric cases classified by the location of the shell relative to the light-ring radii of the inner and outer masses.\
**Case 1: \(R_S<3M_{\rm{BH}}\).** In this case, the shell is located inside the inner light-ring and therefore cannot be probed by geodesics escaping to infinity. From the point of view of geodesics, this case is therefore similar to the one of an isolated black hole.\
**Case 2: \(3M_{\rm{BH}}<R_S<3M_{\infty}\).** If the shell lies between the light-rings, then there is the possibility for geodesics to probe each light-ring separately. Each light-ring is associated with a divergence of the deflection angle, leading to orbiting. In the case where both light-rings are accessible to geodesics, we expect two such divergences to be present, associated with orbiting around the inner and outer light-rings. Hence we further anticipate that there exists a value of \(b_{\ell_-}<b<b_{\ell_+}\) such that the deflection angle is extremal. This implies the presence of a fold caustic and leads to rainbow scattering . The deflection angle revealing these critical effects is represented in Fig. [\[fig:deflection_angle_rs_4\]](#fig:deflection_angle_rs_4){reference-type="ref" reference="fig:deflection_angle_rs_4"} for the configuration where the shell is located between the two light-rings. In this configuration, and for the choice of parameters \(R_S=4M_{\rm{BH}}\) and \(M_{\infty}=1.5M_{\rm{BH}}\), the rainbow angle \(\theta_r\approx 35.8^{\circ}\) and is associated with the impact parameter \(b_r\). The presence of rainbow scattering in DBH spacetimes was noted in but its characteristic amplification was not clearly observed in their simulation due to the specific choice of DBH configurations.\
\
**Case 3: \(R_S > 3M_\infty\)**. If the shell lies outside \(3M_\infty\), a single light-ring is present in the geometry, which is now located *inside* the shell. In this case, the orbiting effect is qualitatively the same as for an isolated black hole and is associated with the critical impact parameter \(b = 3\sqrt{3}M_\text{BH}/\sqrt{\alpha}\). In this configuration, a second local maximum of the potential is present. Contrary to the previous case, this maximum is not associated with the second light-ring but rather to the presence of the shell (see Fig. [\[fig:geodesic_potential\]](#fig:geodesic_potential){reference-type="ref" reference="fig:geodesic_potential"}). Since this second local maximum is not an (unstable) equilibrium point, it will not be associated with a divergence of the deflection angle but rather with a singular point (i.e. the deflection function is not differentiable). It is important to note that the singular point occurring at \(b=b_{R_S} = \sqrt{R_S^3/(R_S-2M_\infty)}\) is a local maximum of the deflection angle, but does not satisfy the condition \(d\Theta_{geo}/db = 0\), which implies that it is *not* associated to rainbow scattering as stated in . Instead, this singular point will lead to grazing as discussed previously. Similar to the previous case, there is an impact parameter associated to rainbow scattering between the orbiting impact parameter and the grazing one. In this configuration, and for the choice of parameters \(R_S=5M_{\rm{BH}}\) and \(M_{\infty}=1.5M_{\rm{BH}}\), the rainbow angle \(\theta_r\approx 157.8^{\circ}\). Fig. [\[fig:deflection_angle_rs_5\]](#fig:deflection_angle_rs_5){reference-type="ref" reference="fig:deflection_angle_rs_5"} depicts the deflection function of such a configuration and Fig. [\[fig:ray_optic_Rs_5\]](#fig:ray_optic_Rs_5){reference-type="ref" reference="fig:ray_optic_Rs_5"} depicts a congruence of geodesics in this DBH set-up. The red and dark green rays in Fig. [\[fig:ray_optic_Rs_5\]](#fig:ray_optic_Rs_5){reference-type="ref" reference="fig:ray_optic_Rs_5"} represent the rainbow and grazing ray respectively and are both associated with an extremal angle. Note the qualitative difference between the rainbow and grazing ray which resides in the concentration of geodesics around each critical trajectory.
# Waves on a dirty black hole spacetime {#sec:waves_in_DBH}
We consider a scalar field, \(\Phi(x)\), propagating on the DBH spacetime, governed by the Klein-Gordon equation \[\Box \Phi \equiv \frac{1}{\sqrt{-g}} \partial_{\mu} \left( \sqrt{-g} g^{\mu \nu} \partial_{\nu} \Phi \right) = 0\] where \(g^{\mu \nu}\) is the inverse metric and \(g\) is the metric determinant. Performing a standard separation of variables, \[\Phi = \frac{1}{r} \sum_{\omega \ell m} \phi_{\omega \ell}(r) Y_{\ell m}(\theta, \phi) e^{-i \omega t}, \label{eq:sepvariables}\] leads to a radial equation of the form \[\label{H_Radial_equation} \left[\frac{d^{2}}{dr_{\ast}^{2}}+\omega^{2}-V_{\ell}(r)\right]\phi_{\omega\ell}= 0,\] where \(V_{\ell}(r)\) is the effective potential which is given by \[\label{Potential} \begin{aligned} V_{\ell}(r) = &A(r) \Bigl[ \frac{\ell(\ell+1)}{r^2}\\ &+\frac{2}{r^3} \left ( M_{\text{\tiny BH}} \Theta(R_s-r) + M_\infty \Theta(r-R_s) \right) \Bigr] \end{aligned}\] with \(\Theta\) being the Heaviside step function, and \(r_{\ast}\) denotes the *tortoise coordinate* defined by \[\label{tortoise_coord} \frac{dr_*}{dr} = \frac{1}{\sqrt{A(r)B(r)}},\] *viz.* \[\label{tortoise_coordinate} r_{\ast}(r) = \left\{ \begin{aligned} &\frac{1}{\sqrt{\alpha}}\left[r +M_{\text{\tiny BH}}\ln\left(\frac{r}{M_{\text{\tiny BH}}}-1\right)\right]+ \kappa & \scriptstyle{M_{\text{\tiny BH}} < r \leq R_s,}\\ & r + 2M_\infty\ln\left(\frac{r}{2M_\infty}-1\right) & \scriptstyle{r > R_s,} \end{aligned} \right.\] where the constant \(\kappa\) is fixed so that \(r_{\ast}(r)\) is continuous. In the following, we introduce the modes \(\phi_{\omega\ell}^{\text{in}}\) which are solutions of [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"} and are defined by their behaviour at the horizon \(r =2M_{\text{\tiny BH}}\) (i.e., for \(r_* \to-\infty\)) and at spatial infinity \(r \to +\infty\) (i.e., for \(r_* \to +\infty\)): \[\begin{aligned}
\label{bc_in} & & \phi_{\omega \ell}^{\mathrm {in}}(r_{*}) \sim \left\{ \begin{aligned} &\displaystyle{e^{-i\omega r_\ast}} & \, (r_\ast \to-\infty),\\ &\displaystyle{A^{(-)}_\ell (\omega) e^{-i\omega r_\ast} + A^{(+)}_\ell (\omega) e^{+i\omega r_\ast}} & (r_\ast \to +\infty). \end{aligned} \right. \nonumber\\ & & \end{aligned}\] Here, the coefficients \(A^{(-)}_\ell (\omega)\) and \(A^{(+)}_\ell (\omega)\) appearing in [\[bc_in\]](#bc_in){reference-type="eqref" reference="bc_in"} are complex amplitudes and allow us to define the scattering \(S\)-matrix elements, \[\label{Matrix_S} S_{\ell}(\omega) = e^{i(\ell+1)\pi} \, \frac{A_{\ell}^{(+)}(\omega)}{A_{\ell}^{(-)}(\omega)}.\] It should be noted that, due to the choice of coordinates (keeping \(r\) as the areal radius on both sides of the shell) the transverse metric is discontinuous at the shell meaning that we have to also place a boundary condition on the eigenfunctions \(\phi_{\omega\ell}\) at the shell: \[\begin{aligned} \bigl[ &\sqrt{B(R_s)} \left (R_s \phi_{\omega \ell}^\prime(R_s)-\phi_{\omega \ell}(R_s) \right )\bigr]_{+} \\ &\qquad = \left[ \sqrt{B(R_s)} \left (R_s \phi_{\omega \ell}^\prime(R_s)-\phi_{\omega \ell}(R_s) \right )\right]_{-} \end{aligned} \label{jump_condition}\]
# Resonances of the dirty black holes {#sec:resonances}
## Quasinormal modes and Regge poles
Resonant modes are characteristic solutions of the wave equation, [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"}, satisfying purely ingoing/outgoing boundary conditions at the horizon/infinity. Their spectrum is the set of zeros of the scattering matrix \(S_{\ell}(\omega)\) [\[Matrix_S\]](#Matrix_S){reference-type="eqref" reference="Matrix_S"}, i.e., a simple pole of \(A_\ell^{(-)}(\omega)\). It can be seen as a set of frequencies \(\omega_{\ell n}\) in the complex-\(\omega\) plane at which the scattering matrix \(S_{\ell}(\omega)\) has a simple pole for \(\ell \in \mathbb{N}\) and \(\omega_{\ell n} \in \mathbb{C}\) (the so-called *quasinormal mode spectrum*), or a set of angular momenta \(\lambda_n(\omega)\equiv \ell_n+1/2\) in the complex-\(\lambda\) plane at which the scattering matrix has a simple pole for \(\omega \in \mathbb{R}\) and \(\lambda_{n}(\omega) \in \mathbb{C}\) (the so-called *Regge pole spectrum*). Here \(n = 1,2,3,\ldots\) labels the different elements of the spectrum and is referred to as the overtone number. The quasinormal mode spectrum of DBHs has been investigated for configurations where \(M_\infty-M_{\text{\tiny BH}}\ll 1\) using perturbative techniques. We shall now consider, for the first time, the Regge poles of generic DBHs.
## Numerical method
To compute the QNM/Regge pole spectrum of dirty black holes, we follow the method of Ould El Hadj et al. who calculated the Regge pole spectrum of scalar and gravitational waves (in the axial sector) for a gravitating compact body. Their method is an extension of the original continued fraction method developed originally by Leaver . The method involves writing the solution to the wave equation [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"} as a power series around a point \(b\) located outside the shell, \[\label{ansatz_CFM} \phi_{\omega,l}(r) = e^{i\omega r_*(r)} \sum_{n=0}^{+\infty} a_n \left(1-\frac{b}{r}\right)^n\,,\] where the coefficients \(a_n\) obey a four-term recurrence relation: \[\label{Recurrence_4_terms} \alpha_n a_{n+1} + \beta_n a_{n} +\gamma_{n} a_{n-1} +\delta_{n} a_{n-2} = 0, \quad \forall n\geq 2,\] where
The initialisation coefficients, \(a_0\) and \(a_1\), are found directly from [\[ansatz_CFM\]](#ansatz_CFM){reference-type="eqref" reference="ansatz_CFM"}, \[\begin{aligned}
\label{Initial_Conds} && a_0 = e^{-i\omega r_*(b)}\phi_{\omega_\ell}(b), \\ && a_1 = b e^{-i\omega r_*(b)}\left(\frac{d}{dr}\phi_{\omega_\ell}(r) \Big{|}_{r=b}-\frac{i\omega b}{b-2M_\infty}\phi_{\omega_\ell}(b)\right). \end{aligned}\] In practise, the coefficients \(a_0\) and \(a_1\) are found numerically by integrating [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"} from the horizon up to \(r = b> R_s\). In order to apply Leaver's method, we first perform a Gaussian elimination step in order to reduce the 4-term recurrence relation to a 3-term recurrence relation: \[\hat{\alpha}_n a_{n+1} + \hat{\beta}_n a_n + \hat{\gamma}_n a_{n-1} = 0,\] where we have defined the new coefficients, for \(n\geq 2\):
The series expansion [\[ansatz_CFM\]](#ansatz_CFM){reference-type="eqref" reference="ansatz_CFM"} is convergent outside the shell provided that \(a_n\) is a minimal solution to the recurrence relation and \(b/2<R_s<b\). The existence of a minimal solution implies that the following continued fraction holds: \[\label{eq:CF} \frac{a_1}{a_0} = \frac{-\hat{\gamma}_1}{\hat{\beta}_1-}\, \frac{\hat{\gamma}_2 \hat{\alpha}_1}{\hat{\beta}_2-} \, \frac{\hat{\gamma}_3 \hat{\alpha}_2}{\hat{\beta}_3-}...\] The above relation (or any of its inversions) is the equation, written in the standard form of continued fractions, we are solving in order to find the RP/QNM spectrum. In practise, we fix \(\omega\) (equivalently \(\ell\)), and define a function \(f(\ell\in \mathbb{C})\) (equivalently \(f(\omega \in \mathbb{C})\)), which gives the difference between the left-hand side and right-hand side of the condition [\[eq:CF\]](#eq:CF){reference-type="eqref" reference="eq:CF"}. We then find the zeros of the function \(f\) starting from an initial guess. An alternative method to the continued fraction expression is to use the Hill determinant , where one looks for the zeros of the following determinant \[\label{Determinant_Hill_4_termes} D = \begin{vmatrix} \beta_0 & \alpha_0 & 0 & 0 & 0 & \ldots & \ldots & \ldots \\ \gamma_1 & \beta_1 & \alpha_1 & 0 & 0 & \ldots & \ldots & \ldots \\ \delta_2 & \gamma_2 & \beta_2 & \alpha_2 & 0 & \ldots & \ldots & \ldots \\ \vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \ldots & \ldots \\ \vdots & \vdots & \delta_{n-1} & \gamma_{n-1} & \beta_{n-1} & \alpha_{n-1} & \ddots & \ldots \\ \vdots & \vdots & \vdots & \delta_n & \gamma_{n} & \beta_{n} & \alpha_{n} & \ddots \\ \vdots & \vdots & \vdots & \vdots & \ddots & \ddots & \ddots & \ddots \end{vmatrix} = 0.\] By assuming \(D_n\) \[\label{derecurrence_4_termes} D_n=\beta_n D_{n-1}-\gamma_{n}\alpha_{n-1}D_{n-2} + \delta_n \alpha_{n-1} \alpha_{n-2} D_{n-3},\] to be the determinant of the \(n \times n\) submatrix of \(D\) with the initial conditions \[\label{Determinant_initial_conds} \begin{split} D_0 &=\beta_0, \\ D_1 &=\beta_1\beta_0-\gamma_1\alpha_0, \\ D_2 &=\beta_0(\beta_1 \beta_2-\alpha_1 \gamma_2)-\alpha_0(\alpha_1 \delta_2-\gamma_1 \beta_2). \end{split}\] the Regge poles (QNM frequencies) are found by solving numerically the roots \(\lambda_n\) (\(\omega_n\)) of \(D_n\). We note that in order to confirm our results, we have used both methods and they give the same results, up to the numerical precision.
## Results: The Regge pole spectrum
In order to check the robustness of our numerical code, we first compute the QNM spectrum for a configuration of DBH considered by Leung et al. i.e., for \(M_\infty = 1.02 M_{\text{\tiny BH}}\), \(R_s = 2,52 M_{\text{\tiny BH}}\) and assuming \(M_{\text{\tiny BH}} =1/2\).
## The WKB approximation
In order to gain insight into the physical origin of the broad resonances identified in Figs. [\[fig:rp1\]](#fig:rp1){reference-type="ref" reference="fig:rp1"} and [\[fig:rp2\]](#fig:rp2){reference-type="ref" reference="fig:rp2"}, we now turn to the WKB approximation.
### Propagation of WKB modes
We begin by constructing the WKB solution to [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"} by assuming that solutions are rapidly oscillating functions with slowly varying amplitudes: \[\phi_\ell (r_*) = A(r_*)e^{i \int p(r'_*) dr_*'}.\] Inserting this ansatz into the wave equation, and assuming that \(|p'| \ll |p^2|\) and \(|A'| \ll |pA|\), we get that \[p(r_*)^2 + V(r_*) = 0, \quad A = a |p|^{-1/2},\] where \(a\) is a constant coefficient, \(V = V_\ell-\omega^2\), and we suppress \(\omega\) and \(\ell\) indices for clarity. The first of the above equations is quadratic in \(p\), thus it will admit two solutions that can be interpreted as waves travelling radially inward and outward. Taking into account contributions from both modes, we can write the following WKB solution to the wave equation: \[\phi_\ell = |p|^{-1/2}\left( a^{in} e^{-i\int{p dr_*}} + a^{out} e^{i\int{p dr_*}}\right).\] We see from the above expression that, in regions where the WKB approximation holds, the only change to the amplitudes of the modes comes from the variation of the function \(p\). It is also clear that the WKB approximation breaks down at those points which satisfy \(V(r_*) = 0\), since it implies that \(p = 0\), leading to an infinite amplitude of the WKB modes. The points satisfying \(V(r_*) = 0\), are called turning points and one can write a WKB solution on both sides of the turning points, with different coefficients \((a^{in},a^{out})\) on each side. One then relates the WKB modal coefficients on each side of the turning points by applying connection formulae (see ). Between turning points, the WKB modes do not mix and their propagation between two points \(r_{*i}\) and \(r_{*j}\) is captured by the following propagation matrices: \[\binom{a^{out}_i}{a^{in}_i} = P_{ij}\binom{a^{out}_j}{a^{in}_j},\] where \[P_{ij} = \begin{pmatrix} e^{-iS_{ij}} & 0 \\ 0 & e^{iS_{ij}} \end{pmatrix} \quad \text{when}\quad p^2(r_{*i}<r_*<r_{*j}) > 0,\] or \[P_{ij} = \begin{pmatrix} 0 & e^{-S_{ij}} \\ e^{S_{ij}} & 0 \end{pmatrix} \quad \text{when}\quad p^2(r_{*i}<r_*<r_{*j}) < 0.\] In the above formula, \(S_{ij}\) is the WKB action between the points \(r_{*i}\) and \(r_{*j}\) and is given by \[S_{ij} = \int_{r_{*i}}^{r_{*j}} |p(r_*)|dr_*.\] Combining the propagation matrix for modes below the potential barrier and the connection matrices across isolated turning points, we define the tunneling matrix, \(T_{12}\), which connects WKB modes on the outside of two turning points \(r_{*_1}\) and \(r_{*_2}\): \[T_{12} = \begin{pmatrix} 1/\mathcal{T} &-\mathcal{R}/\mathcal{T}, \\ \mathcal{R}^*/\mathcal{T} & 1/\mathcal{T}^*. \end{pmatrix},\] where \(\mathcal{R}\) and \(\mathcal{T}\) are the local reflection and transmission coefficients across the potential barrier and are given by: \[\begin{aligned}
\mathcal{R} &=&-i\frac{1-e^{-2S_{12}}/4}{1+e^{-2S_{12}}/4}\\ \mathcal{T} &=& \frac{e^{-2S_{12}}}{1+e^{-2S_{12}}/4} \end{aligned}\]
### WKB modes across the shell
In order to construct the full WKB solution across the entire radial range, we also need to connect WKB modes on both sides of the shell. Since we are only interested in the qualitative behaviour of the WKB modes in order to interpret the broad resonance, we assume here that the jump discontinuity in the potential is much smaller than the frequencies of the WKB modes, \(\omega\), and that this frequency does not vary significantly near the shell. Within this framework, we write the WKB solution for \(r\sim R_s\) as: \[\phi_{\omega,l}(r_*) = \begin{cases} a^{out}_{L}e^{i\omega r_*} + a^{in}_L e^{-i\omega r_*} \ \text{for} \ r_* \leq R_s^*, \\ a^{out}_{R}e^{i\omega r_*} + a^{in}_R e^{-i\omega r_*} \ \text{for} \ r_* \geq R_s^*. \end{cases}\] From the junction condition at the shell given in [\[jump_condition\]](#jump_condition){reference-type="eqref" reference="jump_condition"}, we can define the connection matrix, \(\mathcal{C}\), which connects the WKB modal coefficients on the left and right sides of the shell: \[\mathcal{C} = \begin{pmatrix} 1 &-e^{-2i\omega R_s} \frac{\Delta}{2i\omega} \\ e^{2i\omega R_s} \frac{\Delta}{2i\omega} & 1 \end{pmatrix},\] where \(\Delta = \frac{\sqrt{AB}_+-\sqrt{AB}_{-}}{R_s}\). Note that \(\Delta\) is not the discontinuity of the potential but rather the coefficient entering in the discontinuity of the field's derivative. From the connection matrix, we can also define a reflection coefficient across the shell: \[\begin{aligned}
\mathcal{R}_s = \frac{\Delta}{2i\omega}. \end{aligned}\] Note here that the reflection across the discontinuity is of order \(\mathcal{O}(\omega^{-1})\). This is due to the fact that the metric itself is discontinuous at the shell and is in contrast with the case of a compact object which usually has a continuous metric and for which the reflection coefficient at the surface of the object is at least of order \(\mathcal{O}(\omega^{-2})\) .
### WKB estimate of the broad resonances
We now have all the necessary quantities to relate the modal coefficient at \(r_{*}\rightarrow-\infty\) to the one at \(r_* \rightarrow + \infty\). This is done by combining the propagating, tunneling and connection matrix as follows (see Fig. [\[fig:WKB\]](#fig:WKB){reference-type="ref" reference="fig:WKB"} for illustration): \[\label{eq:connection_WKB} \binom{a^{out}_{-\infty}}{a^{in}_{-\infty}} = P_{-\infty 1}\times T_{12} \times P_{2R_s} \times \mathcal{C} \times P_{R_s\infty} \binom{a^{out}_{\infty}}{a^{in}_\infty}.\] Solving the above system for outgoing boundary conditions, that is \(a^{out}_{-\infty} = a^{in}_{\infty}=0\), leads to the following condition: \[e^{-2iS_{2R_s}-2i\omega R_s} = \mathcal{R}\mathcal{R}_s.\] We now anticipate that broad resonances will be slowly damped and therefore the real part of their RPs will be much greater than their imaginary part. We write \(\lambda = \lambda_R + i\Gamma\), with \(\Gamma \ll \lambda_r\). The real and imaginary parts of the above condition give \[\sin\left( S_{2R_S}(\lambda_R) + \omega R_s-\frac{\pi}{4} \right) = 0 \ \ \text{and} \ \ \Gamma = \frac{\ln(|\mathcal{R}\mathcal{R}_s|)}{2\partial_\lambda S_{2Rs} \big|_{\lambda_R}}.\] Hence the location of the shell sets the real part of the broad resonances part of the RP spectrum while its matter content sets their imaginary part. Contrary to the other branches of the RP spectrum, the above systems admit a finite number of solutions. In particular we must have that the turning points be located between the outer light-ring and the shell, which limits the range of \(\lambda_R\). Solving the above condition numerically for the shell parameters of Fig. [\[fig:rp1\]](#fig:rp1){reference-type="ref" reference="fig:rp1"}, we find that there are only two solutions, \(\lambda_{\mathrm{WKB}}^{(1)} = 35.6946 + 1.79675 i\) and \(\lambda_{\mathrm{WKB}}^{(2)} = 38.2285 + 2.8487 i\), which differ from the values found numerically by \(2.2\%\) and \(3.7\%\) respectively.
# Wave Scattering by a dirty black hole {#sec:scattering}
In this section, we compute the differential scattering cross sections \(d\sigma/d\Omega\) for plane monochromatic scalar waves impinging upon a DBH using the partial wave expansion, and we compare with the results constructed by CAM representations of these cross sections by means of the Sommerfeld-Watson transform and Cauchy theorem .
## The differential scattering cross section: Partial waves expansion
The differential scattering cross section for a scalar field is given by \[\label{Scalar_Scattering_diff} \frac{d\sigma}{d\Omega} = |f(\omega,\theta)|^2\] where \[\label{Scalar_Scattering_amp} f(\omega,\theta) = \frac{1}{2 i \omega} \sum_{\ell = 0}^{\infty} (2\ell+1) [S_{\ell}(\omega)-1]P_{\ell}(\cos\theta)\] denotes the scattering amplitude. In [\[Scalar_Scattering_amp\]](#Scalar_Scattering_amp){reference-type="eqref" reference="Scalar_Scattering_amp"}, the functions \(P_{\ell}(\cos\theta)\) are the Legendre polynomials and the \(S\)-matrix elements \(S_{\ell}(\omega)\) were given by [\[Matrix_S\]](#Matrix_S){reference-type="eqref" reference="Matrix_S"}.
## CAM representation of the scattering amplitude
Following the steps in, we construct the CAM representation of \(f(\theta)\) using a Sommerfeld-Watson transformation \[\label{SWT_gen} \sum_{\ell=0}^{+\infty} (-1)^\ell F(\ell)= \frac{i}{2} \int_{\cal C} d\lambda \, \frac{F(\lambda-1/2)}{\cos (\pi \lambda)},\] where \(F(\cdot)\) is any function without singularities on the real axis \(\lambda\). By means of [\[SWT_gen\]](#SWT_gen){reference-type="eqref" reference="SWT_gen"}, we replace the discrete sum over the ordinary angular momentum \(\ell\) in [\[Scalar_Scattering_amp\]](#Scalar_Scattering_amp){reference-type="eqref" reference="Scalar_Scattering_amp"} with a contour integral in the complex \(\lambda\) plane (i.e., in the complex \(\ell\)-plane with \(\lambda = \ell +1/2\)). By noting that \(P_\ell (\cos \theta) =(-1)^\ell P_\ell (-\cos \theta)\), we obtain \[\begin{aligned}
\label{SW_Scalar_Scattering_amp} & & f(\omega,\theta) = \frac{1}{2 \omega} \int_{\cal C} d\lambda \, \frac{\lambda} {\cos (\pi \lambda)} \nonumber \\ && \qquad\qquad \times \left[ S_{\lambda-1/2} (\omega)-1 \right]P_{\lambda-1/2} (-\cos \theta). \end{aligned}\] It should be noted that, in [\[SWT_gen\]](#SWT_gen){reference-type="eqref" reference="SWT_gen"} and [\[SW_Scalar_Scattering_amp\]](#SW_Scalar_Scattering_amp){reference-type="eqref" reference="SW_Scalar_Scattering_amp"}, the integration contour encircles counterclockwise the positive real axis of the complex \(\lambda\)-plane, and \(P_{\lambda-1/2} (z)\) denotes the analytic extension of the Legendre polynomials \(P_\ell (z)\) which is defined in terms of hypergeometric functions by \[\label{Def_ext_LegendreP} P_{\lambda-1/2} (z) = F[1/2-\lambda,1/2+\lambda;1;(1-z)/2].\] Here, \(S_{\lambda-1/2} (\omega)\) is given by \[see [\[Matrix_S\]](#Matrix_S){reference-type="eqref" reference="Matrix_S"}\] \[\label{Matrix_S_CAM} S_{\lambda-1/2}(\omega) = e^{i(\lambda + 1/2)\pi} \, \frac{A_{\lambda-1/2}^{(+)}(\omega)}{A_{\lambda-1/2}^{(-)}(\omega)}\] and denotes "the" analytic extension of \(S_\ell (\omega)\) where the complex amplitudes \(A^{(\pm)}_{\lambda-1/2} (\omega)\) are defined from the analytic extension of the modes \(\phi_{\omega \ell}\), i.e., from the function \(\phi_{\omega ,\lambda-1/2}\). It is also important to recall that the Regge poles \(\lambda_n(\omega)\) of \(S_{\lambda-1/2}(\omega)\) lie in the first and third quadrants, symmetrically distributed with respect to the origin \(O\), and are defined as the zeros of the coefficient \(A^{(-)}_{\lambda-1/2} (\omega)\) \[see [\[Matrix_S\_CAM\]](#Matrix_S_CAM){reference-type="eqref" reference="Matrix_S_CAM"}\] \[\label{PR_def_Am} A^{(-)}_{\lambda_n(\omega)-1/2} (\omega)=0,\] with \(n=1,2,3,\ldots\)., and the associated residues at the poles \(\lambda=\lambda_n(\omega)\) are defined by \[see [\[Matrix_S\_CAM\]](#Matrix_S_CAM){reference-type="eqref" reference="Matrix_S_CAM"}\] \[\label{residues_RP} r_n(\omega)=e^{i\pi [\lambda_n(\omega)+1/2]} \left[ \frac{A_{\lambda-1/2}^{(+)} (\omega)}{\frac{d}{d \lambda}A_{\lambda-1/2}^{(-)}(\omega)} \right]_{\lambda=\lambda_n(\omega)}.\] In order to collect the Regge pole contributions, we deform the contour \({\cal C}\) in [\[SW_Scalar_Scattering_amp\]](#SW_Scalar_Scattering_amp){reference-type="eqref" reference="SW_Scalar_Scattering_amp"} while using the Cauchy theorem to obtain \[\label{CAM_Scalar_Scattering_amp_tot} f (\omega, \theta) = f^\text{\tiny{B}} (\omega, \theta) + f^\text{\tiny{RP}} (\omega, \theta)\] where
The second term in [\[CAM_Scalar_Scattering_amp_tot\]](#CAM_Scalar_Scattering_amp_tot){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_tot"} \[\begin{aligned}
\label{CAM_Scalar_Scattering_amp_decomp_RP} & & f^\text{\tiny{RP}} (\omega, \theta) =-\frac{i \pi}{\omega} \sum_{n=1}^{+\infty} \frac{ \lambda_n(\omega) r_n(\omega)}{\cos[\pi \lambda_n(\omega)]} \nonumber \\ && \qquad\qquad \qquad\qquad \times P_{\lambda_n(\omega)-1/2} (-\cos \theta), \end{aligned}\] is a sum over the Regge poles lying in the first quadrant of the CAM plane. Of course, the CAM representation of the scattering amplitude \(f (\omega, \theta)\) for the scalar field given by [\[CAM_Scalar_Scattering_amp_tot\]](#CAM_Scalar_Scattering_amp_tot){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_tot"} and [\[CAM_Scalar_Scattering_amp_decomp\]](#CAM_Scalar_Scattering_amp_decomp){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_decomp"} is equivalent to the initial partial wave expansion [\[Scalar_Scattering_amp\]](#Scalar_Scattering_amp){reference-type="eqref" reference="Scalar_Scattering_amp"}. From this CAM representation, we extract the contribution \(f^\text{\tiny{RP}}(\omega, \theta)\) given by [\[CAM_Scalar_Scattering_amp_decomp_RP\]](#CAM_Scalar_Scattering_amp_decomp_RP){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_decomp_RP"} which is only an approximation of \(f(\omega, \theta)\), and which provides us with a corresponding approximation of the differential scattering cross section via ([\[Scalar_Scattering_diff\]](#Scalar_Scattering_diff){reference-type="ref" reference="Scalar_Scattering_diff"}).
## Computational methods
In order to construct the scattering amplitude [\[Scalar_Scattering_amp\]](#Scalar_Scattering_amp){reference-type="eqref" reference="Scalar_Scattering_amp"}, and the Regge pole contribution [\[CAM_Scalar_Scattering_amp_decomp_RP\]](#CAM_Scalar_Scattering_amp_decomp_RP){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_decomp_RP"}, it is necessary first to obtain the function \(\phi_{\omega\ell}^{\text{\tiny in}}(r)\), the coefficients \(A_\ell^{(\pm)}(\omega)\), and the \(S\)-matrix elements \(S_\ell(\omega)\) by solving [\[H_Radial_equation\]](#H_Radial_equation){reference-type="eqref" reference="H_Radial_equation"} with conditions [\[bc_in\]](#bc_in){reference-type="eqref" reference="bc_in"}, and second to compute the Regge poles \(\lambda_n(\omega)\) of [\[PR_def_Am\]](#PR_def_Am){reference-type="eqref" reference="PR_def_Am"} and the associated residues [\[residues_RP\]](#residues_RP){reference-type="eqref" reference="residues_RP"}. To do this, we use the numerical methods of (see Secs. III B and IVA of these papers). It is important to note that, the scattering amplitude [\[Scalar_Scattering_amp\]](#Scalar_Scattering_amp){reference-type="eqref" reference="Scalar_Scattering_amp"} suffers a lack of convergence due to the long range nature of the field propagating on the Schwarzschild spacetime (outside the thin shell) and to accelerate the convergence of this sum, we have used the method described in the Appendix of. All numerical calculations were performed using *Mathematica*.
## Numerical Results and comments: Scattering cross sections
We present in Figs [\[fig:CAM_Sec_2Mw_6\_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_6_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_6_PRs_Rs_4"}, [\[fig:CAM_Sec_2Mw_16_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_16_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_PRs_Rs_4"} and [\[fig:CAM_Sec_2Mw_16_Rs_5\_PRs\]](#fig:CAM_Sec_2Mw_16_Rs_5_PRs){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_Rs_5_PRs"} various scattering cross sections constructed from the CAM approach, and compare with results obtained from the partial wave expansion method. As in Sec. [5](#sec:resonances){reference-type="ref" reference="sec:resonances"}, we focus on two configurations of the DBH: (i) a DBH with \(3M_{\text{\tiny BH}} < R_s < 3M_\infty\), and (ii) DBH with \(R_s >3M_\infty\). Fig. [\[fig:CAM_Sec_2Mw_6\_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_6_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_6_PRs_Rs_4"} shows the scattering cross section constructed from the CAM approach at \(2M_\infty\omega =6\) for a shell configuration with \(M_\infty=1.5 M_{\text{\tiny BH}}\) and \(R_s=4M_{\text{\tiny BH}}\), i.e. for a DBH configuration with \(3 M_{\text{\tiny BH}} < R_s< 3M_\infty\). We can see that it gets progressively closer to the scattering cross section constructed from partial wave expansion by including increasingly more Regge poles in the sum [\[CAM_Scalar_Scattering_amp_decomp_RP\]](#CAM_Scalar_Scattering_amp_decomp_RP){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_decomp_RP"}. Indeed, the sum over only the first two Regge poles in equation [\[CAM_Scalar_Scattering_amp_decomp_RP\]](#CAM_Scalar_Scattering_amp_decomp_RP){reference-type="eqref" reference="CAM_Scalar_Scattering_amp_decomp_RP"} does not reproduce the cross section, however, summing over \(54\) Regge poles, but without including the background integral, the agreement is perfect with that constructed from the partial wave expansion for \(\theta \gtrsim 20^\circ\). Fig. [\[fig:CAM_Sec_2Mw_16_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_16_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_PRs_Rs_4"} illustrates, for the same DBH configuration, the scattering cross section at high frequency \(2M_\infty\omega =16\). In this case, with just \(18\) Regge poles the glory and the orbiting oscillations are very well reproduced. With \(59\) Regge poles and without adding the background integral the result obtained from the CAM approach is again indistinguishable from partial wave expansion for \(\theta \gtrsim 30^\circ\) on the plot. Fig. [\[fig:CAM_Sec_2Mw_16_Rs_5\_PRs\]](#fig:CAM_Sec_2Mw_16_Rs_5_PRs){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_Rs_5_PRs"} shows the scattering cross section for a DBH configuration with \(R_s> 3M_\infty\) at \(2M_\infty\omega = 16\). Here, we have the shell configuration with \(M_\infty=1.5 M_{\text{\tiny BH}}\) and \(R_s=5M_{\text{\tiny BH}}\). In this case, to construct the result from CAM approach, we sum over three different branches and with \(17\) Regge poles the glory as well as the orbiting oscillations are captured. By including \(48\) Regge poles, but no background integral, the agreement is excellent with the partial wave expansion result for \(\theta \gtrsim 30^\circ\). Through Figs. [\[fig:CAM_Sec_2Mw_6\_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_6_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_6_PRs_Rs_4"}, [\[fig:CAM_Sec_2Mw_16_PRs_Rs_4\]](#fig:CAM_Sec_2Mw_16_PRs_Rs_4){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_PRs_Rs_4"} and [\[fig:CAM_Sec_2Mw_16_Rs_5\_PRs\]](#fig:CAM_Sec_2Mw_16_Rs_5_PRs){reference-type="ref" reference="fig:CAM_Sec_2Mw_16_Rs_5_PRs"}, we have shown that the scattering cross section can be described and understood in terms of the Regge Poles, i.e., in terms of various contributions from different types of resonances. In particular, we have identified different branches associated with the properties of the light-rings and the shell position and/or its matter content. To conclude our description of the differential scattering cross section, we return to the discussion of critical effects associated with geodesic motion described in Sec. [3](#sec:geodesics){reference-type="ref" reference="sec:geodesics"}. We can indeed understand qualitatively the observed scattering cross sections from the various critical effect. As in the case of the isolated black hole, the orbiting effect and the glory (associated with a single light-ring) will result in oscillations in the scattering cross section as well as an increase for \(\theta \sim 180^\circ\). In DBH spacetimes, we have shown that we may expect modulations due to the rainbow effect and to grazing (creeping modes) or secondary orbiting. The rainbow effect will lead to a local amplification of the scattering cross section around the rainbow angle. Grazing, i.e., the edge rays (creeping modes) or second orbiting will lead to further oscillations and global modulations of the scattering cross section. This is illustrated in Fig. [\[fig:comparison_critical\]](#fig:comparison_critical){reference-type="ref" reference="fig:comparison_critical"}. Note the qualitative difference that for configurations where the shell is outside the outer light-ring, the scattering cross is mainly modulated around the rainbow angle while for the case where the shell lies between the two light-rings the scattering cross sections exhibits noticeable modulation at all angles. Isolating and identifying quantitatively the different contribution from each critical effect would require asymptotic formulae from the RPs and their residues which are beyond the scope of this work.
# Conclusion and discussion {#sec:conclusion}
In this paper we have investigated the scattering of planar waves incident on a dirty black hole spacetime, that is, a black hole surrounded by a thin shell of matter. We have considered cases where the shell contributes substantially to the scattering, thereby complementing the analyses presented in. We have focused our attention on the spectrum of resonances of DBHs and calculated the Regge pole spectrum of different DBH configurations. This was done by extending the continued fraction method originally introduced by Leaver, and following the approach of Ould El Hadj et al.Ẇe have identified two key properties of the RP spectrum of DBHs: i) the existence of two branches and ii) the possible existence of a third branch with a nearly constant imaginary part. The first two branches are associated with the inner light-ring and with either the outer light-ring or the shell depending on the DBH configuration. It appears that the two branches emerge from the original isolated black hole spectrum. The separation from the original branch to the two distinct ones appears first for large overtone numbers while the two branches may overlap for low overtone. The separation is also more prominent for high frequencies and for DBH configuration where the shell contributes significantly. The existence of the two branches results in modulation of the scattering cross section. Such modulations were not seen in previous studies due to the specific DBH configuration considered by the authors. In addition to these two branches, we have identified a third branch corresponding to broad resonances trapped between the inner light-ring and the thin shell. This interpretation was supported by a WKB analysis of the RP spectrum and is reminiscent of the resonances obtained for compact objects . Building on the identification of the RP spectrum, we have constructed the complex angular momentum representation of the scattering. We have shown that the RP spectrum and their associated residues accurately describes the scattering cross section, including the modulations and deviations from the case of an isolated black hole. The accurate reconstruction of the scattering cross section also confirms that the resonances have been correctly identified and further validates our results. Our study fits into the general goal of modelling and understanding black holes within their environments. We have considered here a toy model of a black hole surrounded by a static and spherically symmetric thin shell of matter. In order to account for even more realistic scenarios, one would need to account for asymmetric distributions of matter and take into account the dynamical evolution of the environment. The thin shell model provides a practical toy model for the latter as the dynamics of the shell can be found exactly. It would be an interesting extension of our work to investigate the resonances of dynamical dirty black holes. Indeed the literature of resonances of time dependent black hole is scarce with only numerical results which still lack a complete theoretical description . We note the interesting recent developments in this direction . Another interesting avenue to explore would be the impact of the environment and its dynamics on the resonance spectrum is to mimic gravitational effects using condensed matter platforms. This field of research, known as analogue gravity , has allowed us to observe and better understand several effects predicted from field theory in curved spacetimes. Recently, analogue simulators have investigated experimentally the ringdown phase of analogue black holes , and more experiments are being developed within the Quantum Simulators for Fundamental Physics research programme (https://www.qsimfp.org/). In analogue experiments, environmental and dynamical effects will inherently be present and accurate modelling of their impact will be necessary. These simulators will provide valuable insights to build the necessary tools to describe dynamical dirty black holes. | {'timestamp': '2022-12-01T02:19:01', 'yymm': '2211', 'arxiv_id': '2211.17147', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17147'} |
{'timestamp': '2022-12-01T02:21:03', 'yymm': '2211', 'arxiv_id': '2211.17225', 'language': 'es', 'url': 'https://arxiv.org/abs/2211.17225'} |
|
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction
Reinforcement learning (RL) has seen remarkable successes in recent years, many of which fall into the multi-agent setting, such as playing multi-agent games, smart grid, queueing networks, etc. In this work, we focus on a form of *networked* Multi-Agent RL (MARL) where the agents interact according to a given network graph. Compared to single-agent RL, MARL faces many additional challenges. First of all, the curse of dimensionality (which is already a major challenge in single-agent RL) becomes a more severe issue in MARL because the complexity of the problem scales exponentially with the number of agents. This is because the size of the state and action space scales exponentially with the number of agents and, as a result, the dimension of the value/\(Q\)-functions and the policies all scale exponentially with the number of agents, which is hard to compute and store in moderately large networks. Moreover, since the agents are coupled by the global state, the training process requires extensive communication among the agents in the entire network. To overcome the aforementioned challenges, the existing literature considers performing MARL with only local information. For example, it has been proposed that each agent's policy depends only on the states of itself and, potentially, its neighboring agents. An example of this type is the independent learners approach, where each agent learns a policy for itself while treating the states and actions of other agents as part of the environment. Another example is the recent work of, which focuses on learning localized policies where each agent is allowed to choose its action based on its own and neighbors' states. Beyond the localization in policy, it has also been proposed to approximate each agent's value or \(Q\) functions in a way that they only depend on the local and nearby agents' states and actions, as opposed to the full state. These approaches greatly relieve the computation and communication burden both empirically and theoretically . Despite this progress, there are still major limitations. As discussed above, in order to speed up learning, many previous works, e.g., , restrict consideration to localized policies where each agent makes decisions only based on the local state of the agent and, potentially, neighbors, as opposed to the global state. This leads to a fundamental performance gap between the class of localized policies considered and the optimal centralized policy. Even when the best localized policy can be found, there is a performance degradation compared to the best centralized policy, as the centralized policy class is a strict superset of the localized policy class. An important open question that remains is the following:
This question has received increasing attention in linear control settings, but is still open in MARL settings. In this work, we provide the first bounds on the gap between localized and centralized policies under a MARL model in networked systems. Another limitation of prior work studying MARL is that, while existing results have provided convergence bounds for local policies, the convergence of their methods is typically only to a suboptimal policy in the localized policy class. For example, in, the policies are shown to converge to a stationary point of the objective function defined on localized policies, as opposed to the global optimum. This is unsatisfying as converging to a stationary point does not even guarantee converging to the best localized policy. Therefore, another important and open question that remains is the following:
This question has received increasing attention in single-agent settings, where people have studied the convergence to the global optimum for policy-gradient methods under various policy classes. However, it is open in the context of learning localized policies in MARL and in this work, we provide an affirmative answer to this question under an MARL model in networked systems.
## Related Literature
Reinforcement Learning (RL) studies a dynamical environment where the agent decides its current action based on the current state and the current state/action affects the distribution of its next state. This environment is usually modeled as a *Markov Decision Process* (MDP) where the key assumption is the *Markov Property*, which means the current state and action are statistically sufficient for deciding the next state (see, e.g., ). Specifically, an MDP (with discounted accumulative reward) can be characterized as a tuple \(\langle \mathcal{S}, \mathcal{A}, P, \gamma, r\rangle\), where \(\mathcal{S}\) and \(\mathcal{A}\) denote the state/action space. The transition function \(P: \mathcal{S} \times \mathcal{A}\to \Delta_\mathcal{S}\) characterizes the transition probability \(\mathbf{P}(s'\mid s, a)\) from the current state/action pair to next state \(s'\). \(\gamma \in (0, 1)\) denotes the discount factor, and the reward function \(r: \mathcal{S} \times \mathcal{A} \to [0, \bar{r}]\) is nonnegative for the agent with the state/action pair \((s, a)\). In single-agent RL, the goal of the agent is to learn a policy \(\zeta: \mathcal{S} \to \Delta(\mathcal{A})\) to maximize the expected discounted accumulative reward \(J(\zeta) := \mathbb{E}_{s(0) \sim \rho, a(t) \sim \zeta(\cdot\mid s(t))}\left[\sum_{t=0}^T \gamma^t r(s(t), a(t))\right].\) Two classic learning algorithms have been proposed to learn the optimal policy when the transition probabilities are known: value iteration and policy iteration. *Value Iteration* (VI) iteratively updates the estimated optimal value function by the Bellman equation and derives the final policy from the learned optimal value function. In contrast, *Policy Iteration* (PI) works by evaluating the current policy and iteratively updating the policy using this policy's value function. Our algorithm, \(\text{LPI}\), can be viewed as a generalization of PI to a networked setting where the transition probabilities are unknown. Lower bound results show tabular RL algorithms inevitably suffer from *the curse of dimensionality* when the transition probabilities are unknown (see, e.g., ), which means the sample complexity (i.e., the number of samples to find a near-optimal policy) grows with respect to the size of state space \(\mathcal{S}\) and action space \(\mathcal{A}\). Multi-Agent Reinforcement Learning (MARL) generalizes the single-agent RL setting to a situation where the global MDP evolves based on the joint action/policy of \(n\) agents. The generalization to include more agents is necessary for solving practical problems that involve large-scale networks and/or games. Specifically, each agent \(i\) has its local action space \(\mathcal{A}_i\) and its local reward function \(r_i\), and the global action space \(\mathcal{A} = \mathcal{A}_1 \times \cdots \times \mathcal{A}_n\). At every time step \(t\), agent \(i\) receives a partial observation \(o_i(s(t))\) to decide its local action \(a_i(t)\). More details about this setting can be found in . Depending on each agent's learning objective, MARL can be divided to two categories: cooperative or competitive MARL (see for a survey). In *cooperative MARL*, all agents work together to optimize a shared global objective, which is also the setting studied in this work. In *competitive MARL*, each agent optimizes its local accumulative reward, and the goal is to find an approximate Nash Equilibrium. A major challenge for cooperative MARL is that the size of the global action space \(\mathcal{A}\) can grow exponentially with respect to the number of agents \(n\), which makes centralized training intractable. Fully distributed training is not practical either, because its reward and transition probability depend on other agents' policies. To address this challenge and derive finite-sample complexity bounds that are not exponential in \(n\), a common assumption made by previous works is that the global state is observable to all agents (i.e., \(o_i(s) = s\)), which thus eliminates the need for communicating local states in the network. This works in special settings such as cooperative navigation and Predator-Prey where a joint state space is easy to observe or all agents share the same state space. Nevertheless, such an assumption is not practical in more general MARL settings like the one studied in our paper, where each agent has its own state space and the global state space \(\mathcal{S} = \mathcal{S}_1 \times \cdots \times \mathcal{S}_n\) is exponentially large with respect to \(n\). In this setting, huge communication costs are usually unavoidable and thus it is more challenging and less studied. Our work is most related to a class of cooperative MARL problems where the agents are located in a network graph \(\mathcal{G}\), which makes similar structural assumptions on the MDP as this paper, i.e., each agent has its own local state/action space and the local transition probabilities depend only on the agent's neighbors. Each agent \(i\) observes the local states of the agents whose graph distance to \(i\) is less than or equal to \(\kappa\) to decide its local action (i.e., \(o_i(s) = s_{\mathcal{N}_i^\kappa}\), where \(\mathcal{N}_i^\kappa := \{j \mid d_{\mathcal{G}}(i, j) \leq \kappa\}\). \(d_{\mathcal{G}}(i, j)\) denotes the length of the shortest path between \(i\) and \(j\) on \(\mathcal{G}\)). As a remark, all local policies become centralized when the dependence parameter \(\kappa\) exceeds the diameter of graph \(\mathcal{G}\). Previous works propose decentralized policy iteration algorithms to learn local policies for each agent[^1] and prove finite-time convergence bounds. However, there is still an open gap between the localized policy and centralized policy where each agent makes decisions based on the global state. Specifically, before our work, there is no result on how large the gap between the optimal \(\kappa\)-hop localized policy and the optimal centralized policy is, nor how fast it decays with respect to \(\kappa\). Besides, even if the objective is defined with respect to the \(\kappa\)-hop localized policy class, the algorithms proposed by are only guaranteed to converge to stationary points of their local objective functions rather than global optima of the objective. Another challenge widely encountered in RL is that the policy might rapidly become deterministic during the training process, which further leads to a slow convergence. To speed up the convergence and also encourage exploration for the training algorithm to escape suboptimal points, entropy regularization has often been added to the value function. This approach has yielded both good empirical performance and strong theoretical guarantees. In this paper, we focus on the entropy regularized problem in multi-agent reinforcement learning.
# Model & Preliminaries {#sec:model_pre}
In this section, we introduce the model we study and provide preliminaries that are used throughout our paper. In particular, we introduce networked multi-agent Markov decision processes in .
[^1]: The local policy defined in these papers focuses on a local agent that chooses local actions based only on the agent's state. | {'timestamp': '2022-12-01T02:18:16', 'yymm': '2211', 'arxiv_id': '2211.17116', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17116'} |
null | null |
# Introduction {#S:one .unnumbered}
As many engineering applications have become increasingly reliant on user data, data privacy has become a concern that data aggregators and curators must take into consideration. In numerous applications, such as healthcare , energy systems , transportation systems and Internet of Things (IoT) , the data gathered to support system operation often contains sensitive individual information. Differential privacy has emerged as a standard privacy framework that can be used in such applications to protect sensitive data while allowing privatized data to remain useful. Differential privacy is a statistical notion of privacy that provides privacy guarantees by adding carefully calibrated noise to sensitive data or functions of sensitive data. Its key features include: (i) it is robust to side information, in that any additional knowledge about data-producing entities does not weaken their privacy by much , and (ii) it is immune to post-processing, in that any transformation of private data stays private . Well-known mechanisms for the enforcement of differential privacy include the Laplace mechanism , the Gaussian mechanism , and the exponential mechanism . Other mechanisms have been developed for specific applications, in some cases by building upon or modifying these well-known mechanisms. A representative sample includes the Dirichlet mechanism on the unit simplex , mechanisms for sensitive words of symbolic data , the matrix-variate Gaussian mechanism for matrix-valued queries , and the XOR mechanism for binary-valued data . As the use of differential privacy has grown, these and other mechanisms have been needed to respond to privacy needs in settings with constraints on allowable data. The Gaussian mechanism has been used for some classes of numerical data, although it may also require modifications for some types of sensitive data because the Gaussian mechanism adds unbounded noise. For example, the Gaussian mechanism may generate negative values for data such as ages, salaries, and weights, and such negative values are not meaningful in these contexts. One attempt to solve this problem is through projecting out-of-domain results back to the closest value in the given domain. Although this procedure does not weaken differential privacy because it is only post-processing on private data, it has been observed to lead to low accuracy in applications and thus is undesirable. Nonetheless, the Gaussian mechanism has been used in numerous applications, including deep learning ,, convex optimization , filtering and estimation problems and cloud control , all of which can use data that is inherently bounded in some way. On the other hand, compared to the Laplace mechanism, which is another popular privacy mechanism designed for numerical data, the Gaussian mechanism is able to use lower-variance noise to provide the same privacy level for high-dimensional data. This is because the variance of privacy noise is an increasing function of the sensitivity of a given query, and the Gaussian mechanism allows the use of the \(L_2\) sensitivity which, in applications like federated learning and deep learning , is much lower than the \(L_1\) sensitivity used in Laplace mechanism. Therefore, in this paper we develop a new type of Gaussian mechanism that accounts for the boundedness of semantically valid data in a given application. In the univariate case we consider data confined to a closed interval, and in the multivariate case we consider data confined to a product of closed intervals. For both cases we show that the bounded Gaussian mechanism provides \(\epsilon\)-differential privacy, rather than \((\epsilon, \delta)\)-differential privacy as is provided by the ordinary Gaussian mechanism. We also present algorithms for finding the minimum variance necessary to enforce \(\epsilon\)-differential privacy using the bounded Gaussian mechanism. The mechanisms we develop do not rely on projections and thus avoid the harms to accuracy that can accompany projection-based approaches. Besides, point out that the original Gaussian mechanism's variance bound is far from tight when the privacy parameter \(\epsilon\) approaches either \(0\) or \(\infty\). This is because the original Gaussian mechanism calibrates noise with a worst-case bound which is only tight for a small range of \(\epsilon\). In this work, we present bounded mechanisms that address this limitation by leveraging the boundedness of the domains we consider. This boundedness excludes the tail of each Gaussian distribution from our analysis and allows for the development of tight bounds on the variance of noise required for all \(\epsilon\). Related work in developed a bounded Laplace mechanism, and in this work we bring the ability to bound private data to the Gaussian mechanism. Developments in include a generalized Gaussian mechanism with bounded support, and we show in Section [4](#S:algo){reference-type="ref" reference="S:algo"} that the mechanism developed in this paper requires significantly lower variance of noise to attain differential privacy and thus provides improved accuracy. **Notation** Let \(\mathbb{N}\) denote the set of all positive integers and let \(I\) denote the identity matrix. We use non-bold letters to denote scalars, e.g., \(a\in\mathbb{R}\), and we use bold letters denote vectors, e.g., \(\mathbf{b}\in\mathbb{R}^n\). All intervals of the form \([a, b]\) are assumed to have \(a < b\). For \(\mathbf{a},\mathbf{b} \in \mathbb{R}^m\) we define \([\mathbf{a}, \mathbf{b}] = [a_1, b_1] \times [a_2, b_2] \times \cdots \times [a_m, b_m]\). We also use the function \[\text{erf}(z) = \frac{2}{\sqrt{\pi}}\int_0^z e^{-t^2}dt.\]
# Preliminaries and Problem Statement
This section gives background on differential privacy and problem statements that are the focus of the remainder of the paper.
## Differential Privacy Background
Differential privacy is enforced by a *mechanism*, which is a randomized map. We let \(S\) denote the space of sensitive data of interest. For nearby pieces of sensitive data, a mechanism must produce outputs that are approximately distinguishable. The definition of "nearby" is given by an adjacency relation.
For numerical queries \(Q:S^n\rightarrow \mathbb{R}^m\), \(m \in \mathbb{N}\), the sensitivity of a query \(Q\), denoted \(\Delta Q\), is used to calibrate the variance of privacy noise used to implement privacy. Throughout this paper we will use the \(\ell_2\)-sensitivity.
The \(\ell_2\) sensitivity captures the largest magnitude by which the outputs of \(Q\) can change across two adjacent input databases. We next introduce the definition of differential privacy itself.
The privacy parameter \(\epsilon\) sets the strength of privacy protections, and a smaller \(\epsilon\) implies stronger privacy. The parameter \(\delta\) can be interpreted as a relaxation parameter, in the sense that it is the probability that \(\epsilon\)-differential privacy fails to hold (though it may hold with a different value of \(\epsilon\)). If \(\delta=0\), then a mechanism is said to be \(\epsilon\)-differentially private. In the literature, \(\epsilon\) has ranged from \(0.01\) to \(10\) (). A differential privacy mechanism guarantees that the randomized outputs of two adjacent databases will be made approximately indistinguishable to any recipient of their privatized forms, including any eavesdroppers. One widely used differential privacy mechanism is the Gaussian mechanism. The standard Gaussian mechanism is given by the following theorem:
As discussed in the introduction, the Gaussian mechanism has been favored over the Laplace mechanism in several applications, including deep learning, empirical risk minimization, and various problems in control theory and optimization. Simultaneously, these applications may have bounds on the data they use, such as bounds on training data for a learning algorithm or bounds on states in a control system, though the Gaussian mechanism has infinite support and hence produces unbounded private outputs.
## Problem Statement
In this work we seek a Gaussian mechanism that respects given bounds on data, and we formalize the development of this mechanism in the next two problem statements.
We solve Problem [\[prob:univ_bd_gaus\]](#prob:univ_bd_gaus){reference-type="ref" reference="prob:univ_bd_gaus"} in Section [2](#sec:bounded_gaussian_mechanism){reference-type="ref" reference="sec:bounded_gaussian_mechanism"}, and we solve Problem [\[prob:multi_bd_gaus\]](#prob:multi_bd_gaus){reference-type="ref" reference="prob:multi_bd_gaus"} in Section [3](#sec:multivariate_mechanism){reference-type="ref" reference="sec:multivariate_mechanism"}.
# Univariate Bounded Gaussian mechanism {#sec:bounded_gaussian_mechanism}
In this section, we develop the bounded Gaussian mechanism for a bounded domain \(D=[a,b]\subset\mathbb{R}\). We now give a formal statement of the univariate bounded Gaussian mechanism, and this will be the focus of the rest of this section:
Since \(Q(d)=s\), the density function \(p_B(x)\) is not the same for each database, and data-dependent noise is known to be problematic in general for differential privacy . Therefore using the parameters of the standard Gaussian mechanism is no longer guaranteed to satisfy differential privacy for the bounded Gaussian mechanism. We next provide the parameters required for the bounded Gaussian mechanism to provide differential privacy by first introducing some preliminary results.
## Preliminary Results
This leads to the following lemma.
## Main Result on the Univariate Bounded Gaussian Mechanism
We now proceed to the main result of this section, which bounds the variance required for the bounded Gaussian mechanism to provide \(\epsilon\)-differential privacy.
We observe that in Equation [\[eq:diff_pvt_one_d\]](#eq:diff_pvt_one_d){reference-type="eqref" reference="eq:diff_pvt_one_d"} the privacy noise variance \(\sigma\) appears on both sides of the equation. Given a value of \(\epsilon > 0\), it is desirable to use the smallest \(\sigma\) as this corresponds to using the smallest variance of privacy noise for a given privacy level. To do so, in the next section, we form a zero finding problem to find the minimum \(\sigma\) that satisfies Equation [\[eq:diff_pvt_one_d\]](#eq:diff_pvt_one_d){reference-type="eqref" reference="eq:diff_pvt_one_d"}.
## Calculating \(\sigma\) {#sec:calculate_sigma}
Let \(\sigma^*\) be the smallest admissible value of \(\sigma\). The calculation of \(\sigma^*\) can be formulated as a zero finding problem, where we require \[f(\sigma^*)=(\sigma^*)^2-\frac{\left[(b-a)+\frac{\Delta Q}{2}\right]\Delta Q}{\epsilon-\ln(\Delta C(\sigma^*))}=0.\label{eq:function_f}\] We now present some technical lemmas concerning the function \(f\) with respect to a point \[\sigma_0=\sqrt{\frac{\left[(b-a)+\frac{\Delta Q}{2}\right]\Delta Q}{\epsilon}}.\label{eq:sigma_0}\]
Next, Lemma [\[lem:lemma_f\_sigma_0\]](#lem:lemma_f_sigma_0){reference-type="ref" reference="lem:lemma_f_sigma_0"} shows the value of \(f(\sigma_0)\) and Lemma [\[lem:lemma_f\_decreasing\]](#lem:lemma_f_decreasing){reference-type="ref" reference="lem:lemma_f_decreasing"} shows \(f\) is a monotonically increasing function on \([\sigma_0,\infty)\).
By combining Lemma [\[lem:lemma_f\_sigma_0\]](#lem:lemma_f_sigma_0){reference-type="ref" reference="lem:lemma_f_sigma_0"} and Lemma [\[lem:lemma_f\_decreasing\]](#lem:lemma_f_decreasing){reference-type="ref" reference="lem:lemma_f_decreasing"}, we can use Algorithm 1 in , given below as Algorithm [\[alg:zero_finding_algorithm\]](#alg:zero_finding_algorithm){reference-type="ref" reference="alg:zero_finding_algorithm"}, to find the exact fixed point of \(f\), which in our case is \(\sigma^*\).
The output of Algorithm [\[alg:zero_finding_algorithm\]](#alg:zero_finding_algorithm){reference-type="ref" reference="alg:zero_finding_algorithm"} can be used in Definition [\[def:bounded_guassian\]](#def:bounded_guassian){reference-type="ref" reference="def:bounded_guassian"} to form the univariate bounded Gaussian mechanism, and Theorem [\[thm:main_result_1\]](#thm:main_result_1){reference-type="ref" reference="thm:main_result_1"} shows that doing so provides \(\epsilon\)-differential privacy. We next present an analogous mechanism for multi-variate query responses.
# Multivariate Bounded Gaussian mechanism {#sec:multivariate_mechanism}
We begin by formally defining the multivariate bounded Gaussian mechanism, and it will be the focus of the remainder of this section.
Before introducing the this mechanism's differential privacy guarantee we first require some preliminary results.
## Preliminary Results
Now we define \(\Delta C_m(\sigma_m,\mathbf{c}^*)\), where \[\begin{aligned}
\frac{C_m(\mathbf{a},\sigma_m)}{C_m(\mathbf{a}+\mathbf{c},\sigma_m)}\leq \frac{C_m(\mathbf{a},\sigma_m)}{C_m(\mathbf{a}+\mathbf{c}^*,\sigma_m)}=:\Delta C_m(\sigma_m,\mathbf{c}^*), \end{aligned}\] where \(\mathbf{c}^*\in\mathbb{R}^m\) is the optimal value of \(\mathbf{c}=(c_1,\dots,c_m)^T\) in the following optimization problem: \[\begin{aligned}
\max_{\mathbf{c}\in\mathbb{R}^m}\,\,\,&\Delta C_m(\sigma_m,\mathbf{c})\label{eq:objective_function}\\ \text{subject to } \,\, &0\leq c_i\leq b_i-a_i \text{ for each }i\nonumber\\ &||\mathbf{c}||_2 \leq \Delta Q.\nonumber \end{aligned}\]
## Main Results
We now proceed to the main result of this section, which bounds the variance required for the multivariate bounded Gaussian mechanism to provide \(\epsilon\)-differential privacy.
As with the univariate bounded Gaussian mechanism, we see that \(\sigma_m\) appears on both sides of [\[eq:sigma_m\_multi\]](#eq:sigma_m_multi){reference-type="eqref" reference="eq:sigma_m_multi"}. Hence, we will use numerical methods to compute it.
## Calculating \(\sigma_m\)
We will follow similar steps to those in Section [2.3](#sec:calculate_sigma){reference-type="ref" reference="sec:calculate_sigma"}. Let \(\sigma_m^*\) be the minimal admissible value of \(\sigma_m\) that satisfies [\[eq:sigma_m\_multi\]](#eq:sigma_m_multi){reference-type="eqref" reference="eq:sigma_m_multi"}. Then for \(D\subset \mathbb{R}^m\) we have \[f_m(\sigma_m^*)=(\sigma_m^*)^2-\frac{\left[||\mathbf{b}-\mathbf{a}||_2+\frac{\Delta Q}{2}\right]\Delta Q}{\epsilon-\ln(\Delta C_m(\sigma_m^*,\mathbf{c}^*))}=0.\label{eq:f_n}\] We now present some lemmas concerning the function \(f_n\) with respect to the point \[\sigma_{m,0}=\sqrt{\frac{\left[||\mathbf{b}-\mathbf{a}||_2+\frac{\Delta Q}{2}\right]\Delta Q}{\epsilon}}.\label{eq:sigma_0_nd}\]
By combining Lemma [\[lem:lemma_f\_sigma_0\_nd\]](#lem:lemma_f_sigma_0_nd){reference-type="ref" reference="lem:lemma_f_sigma_0_nd"} and Lemma [\[lem:lemma_f\_decreasing_nd\]](#lem:lemma_f_decreasing_nd){reference-type="ref" reference="lem:lemma_f_decreasing_nd"} we can use Algorithm 1 in to find the exact fixed point of \(f_m\), namely \(\sigma_m^*\), which is shown in Algorithm [\[alg:zero_finding_algorithm_2\]](#alg:zero_finding_algorithm_2){reference-type="ref" reference="alg:zero_finding_algorithm_2"}. The output of Algorithm [\[alg:zero_finding_algorithm_2\]](#alg:zero_finding_algorithm_2){reference-type="ref" reference="alg:zero_finding_algorithm_2"} can be combined with the mechanism in Definition [\[def:bounded_guassian_nd\]](#def:bounded_guassian_nd){reference-type="ref" reference="def:bounded_guassian_nd"}, and Theorem [\[thm:main_result_2\]](#thm:main_result_2){reference-type="ref" reference="thm:main_result_2"} shows that this combination gives \(\epsilon\)-differential privacy for multivariate queries.
# Numerical Results {#S:algo}
This section presents simulation results. We consider queries of the properties of a graph. Specifically, the bounded data we consider is (i) the average degree in an undirected graph, and (ii) the second-smallest eigenvalue of the Laplacian of an undirected graph. This eigenvalue is also called the Fiedler value of the graph. It is known that an undirected graph is connected if and only if its Fiedler value is positive , and the Fiedler value also sets the rate of convergence of various dynamical processes over graphs . We let \(G=(V,E)\) denote a connected, undirected graph on \(10\) nodes. In this experiment we have a two-dimensional query \(Q\) to compute (i) the algebraic connectivity, equal to the second smallest eigenvalue \(\lambda_{2}\) of the Laplacian matrix of \(G\), which satisfies \(\lambda_{2}\in[0,10]\) here, and (ii) the degree of a fixed but arbitrary node \(i\), denoted \(N^i\), which satisfies \(N^i \in [1,9]\) here. Therefore, for a graph \(G\) we have \(Q(G)=[\lambda_{2},N_{\text{out}}]^T\in D\), where \(D= [\mathbf{a,b}]\subset\mathbf{R}^2\), with \(\mathbf{a}=[0,1]^T\) and \(\mathbf{b}=[10,9]^T\). For a fixed \(k \in \mathbb{N}\), we say that two graphs are adjacent if they have the same node set but differ in \(k\) edges. Let \(\Delta Q_{\lambda_2}\) denote the sensitivity of \(\lambda_2\), let \(\Delta Q_{N^i}\) denote the sensitivity of node \(i\)'s degree, and let \(\Delta Q=||(\Delta Q_{\lambda_2},\Delta Q_{N^i})^T||_2\) denote the sensitivity of the query \(Q\). From we have \(\Delta Q_{\lambda_2}=2k\) and \(\Delta Q_{N^i}=k\). Thus, \(\Delta Q = \sqrt{5}k\). In our simulations we set \(k=2\). Table [1](#table:1){reference-type="ref" reference="table:1"} gives some example variances computed for using the multivariate bounded Gaussian mechanism. In Figure [\[fig:Normal_distribution\]](#fig:Normal_distribution){reference-type="ref" reference="fig:Normal_distribution"}, there is a general decrease in the variance of the bounded Gaussian mechanism as \(\epsilon\) grows. This decrease agrees with intuition because a larger \(\epsilon\) implies weaker privacy protection and thus the private output distribution has a higher peak on the true query answer. We now compare the proposed mechanism with generalized Gaussian mechanism . Let \(\sigma_{GG}^2\) be the variance of generalized Gaussian mechanism and \(\sigma_{BG}^2\) be the bounded Gaussian mechanism. Then we define the Percent Reduction in variance as \[\text{Percent Reduction} = \frac{\sigma_{GG}^2-\sigma_{BG}^2}{\sigma_{GG}^2}\times 100\%.\] Both Figure [\[fig:Normal_distribution\]](#fig:Normal_distribution){reference-type="ref" reference="fig:Normal_distribution"} and Table [1](#table:1){reference-type="ref" reference="table:1"} show that the bounded Gaussian mechanism always generates smaller variance, i.e., \(\sigma_{BG}^2<\sigma_{GG}^2\) and the \(\text{Percent Reduction}>0\) for all \(\epsilon\). In other words, compared to the generalized Gaussian mechanism , the bounded Gaussian mechanism generates private outputs with less noise but with the same level of privacy protection.
# Conclusion {#S:item}
This paper presented two differential privacy mechanisms, namely the univariate and multivariate bounded Gaussian mechanisms, for bounded domain queries of a database of sensitive data. Compared to the existing generalized Gaussian mechanism, the bounded Gaussian mechanisms we present generate private outputs with less noise and better accuracy for the same privacy level. Future work will apply this mechanism to real world applications, such as privately forecasting epidemic propagation, federated learning and optimization, and explore privacy and performance trade-off. Besides, future work will design a denoise post-processing procedure to generate more accurate private outputs. | {'timestamp': '2022-12-01T02:21:04', 'yymm': '2211', 'arxiv_id': '2211.17230', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17230'} |
# Introduction[\[sec:introduction\]]{#sec:introduction label="sec:introduction"}
Time-dependent brightness variations of celestial objects may be caused by different phenomena: intrinsic physical changes, such as pulsations, eruptions, and cataclysmic outbursts, or extrinsic reasons that depend on the direction of observation, such as eclipsing binaries, stars rotating with spots or with ellipsoidal shapes, and microlensing events, as shown in fig. 1 of. The detection of variability requires multi-epoch observations and, depending on the signal sampling, a certain set of classes can be identified. *Gaia* As the time span of *Gaia* For brevity, we refer to Table [\[tab:training\]](#tab:training){reference-type="ref" reference="tab:training"} for selected publications related to these classes, with representatives identified in various surveys. Machine learning is a practical tool to automate classification tasks that involve multiple known classes and a possibly high number of attributes to identify such classes and distinguish them from others. Herein, we present how supervised classification was applied to *Gaia* In particular, we describe the details concerning the construction of the training set and of the classifiers, the verification of results, and the generation of an overall classification score. Selection procedures, parameter distributions, and assessments of candidates are presented for each class. Some of the classification results are further processed by Specific Object Studies (SOS) dedicated to single classes, typically describing in detail a subset of the most reliable candidates. Such single-class processing modules are available in DR3 for active galactic nuclei, Cepheids, compact companions, eclipsing binaries, long-period variables, main-sequence oscillators, planetary transits, and RR Lyrae stars. Other SOS modules, such as microlensing events, short timescale variables, and solar-like rotation modulation stars, were executed independently of classification results, as they relied on their own candidate selection. A summary of the variability results from all modules is presented in. This article is organised as follows. The classification input data are outlined in Sect. [\[sec:data\]](#sec:data){reference-type="ref" reference="sec:data"}, the preparation, application, and verification of supervised learning procedures are described in Sect. [\[sec:method\]](#sec:method){reference-type="ref" reference="sec:method"}, results for each class are presented in Sect. [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, and conclusions are drawn in Sect. [\[sec:conclusions\]](#sec:conclusions){reference-type="ref" reference="sec:conclusions"}. Special training selections applied to a subset of classes are detailed in Appendix [\[app:training\]](#app:training){reference-type="ref" reference="app:training"}, selected classification attributes are listed in Appendix [\[app:attributes\]](#app:attributes){reference-type="ref" reference="app:attributes"}, additional class labels from the literature (among the false positive classes listed in Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}) are defined in Appendix [\[app:FP_labels\]](#app:FP_labels){reference-type="ref" reference="app:FP_labels"}, some examples of queries to facilitate the exploitation of classification results in the *Gaia* and common diagrams for all classes, including a summary of trained and classified sources, an assessment of results with respect to the literature, and sample light curves are presented in Appendix [\[app:plots\]](#app:plots){reference-type="ref" reference="app:plots"}. All table names in the *Gaia*
# Data[\[sec:data\]]{#sec:data label="sec:data"}
As part of the *Gaia* Before any computation, sources and associated epoch FoV transits were processed by the chain of operators described in sect. 10.2.3 of the *Gaia* DR3 All time series and derived statistical numbers hereafter refer to these cleaned time series. The median number of FoV measurements in the three photometric bands is between 40 and 44 per source, within a time span of typically 900--1000 days in the \(G\) While the processing of *Gaia* Among instrumental effects, scan-angle dependent signals were induced mainly by asymmetric extended sources (such as barred spiral galaxies and tidally distorted stars) and multiple close pairs (\(\lesssim 1\arcsec\)) of point-like sources. Although such signals helped the identification of galaxies from photometric variations, in general data artefacts might interfere with the correct identification of classes with genuine variability, especially those associated with low signal-to-noise ratios. The classification of variables employed also astrometrically derived parameters such as parallax and proper motion. However, *Gaia* DR3 A subset of classified sources were analysed in more detail by subsequent SOS modules, typically focusing on specific classes, as mentioned in Sect. [\[sec:introduction\]](#sec:introduction){reference-type="ref" reference="sec:introduction"}. The results of all variability modules were subject to additional source filtering before their ingestion into the public *Gaia* Statistical parameters of all the photometric time series published in *Gaia* DR3
# Method[\[sec:method\]]{#sec:method label="sec:method"}
For *Gaia* DR3 Known variables in the literature are cross-matched with *Gaia* An extensive cross-match of *Gaia* The robustness of the cross-match method, which included astrometric and photometric information in the identification of matches, and the verification of the genuineness of literature classifications ensured the reliability of training sources (critical to supervised classification) and of the validation of the results.
## Training set[\[ssec:training\]]{#ssec:training label="ssec:training"}
Potential training sources from literature were vetted for each class to ensure the correct class membership. This was repeated for every catalogue that was deemed trustable for training the class under investigation. The reliance of supervised classification on known objects makes it vulnerable to biases from the literature, for instance, related to their data acquisition and classification methods. Thus, in addition to class verification, the cross-matched objects were probed in several dimensions to identify intrinsic biases, such as limited sky coverage or apparent magnitude range with respect to the ones of *Gaia*
### Published classes[\[sssec:classes\]]{#sssec:classes label="sssec:classes"}
Since it was difficult to know a priori the full list of classes that could be identified in *Gaia* DR3 The published variability classes, corresponding acronyms, including the types trained within class groups or a (non-comprehensive) list of subtypes for some classes, are presented as follows:
1. \(\alpha^2\) Canum Venaticorum (ACV) or (magnetic) chemically peculiar (MCP, CP) or rapidly oscillating Am/Ap-type (ROAM, ROAP) or SX Arietis (SXARI) star (collectively denominated as ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI);
2. \(\alpha\) Cygni-type star (ACYG);
3. active galactic nucleus (AGN); from the perspective of variability, the general term AGN is favoured with respect to quasar (or QSO), as brightness variability is caused by the activity of galactic nuclei, from processes in the accretion disc around a supermassive black hole (such as in Seyfert galaxies and quasars), which can lead to the formation of relativistic plasma jets (identified as blazars when directed towards us);
4. \(\beta\) Cephei variable (BCEP);
5. B-type emission line (BE) star or \(\gamma\) Cassiopeiae (GCAS) or S Doradus (SDOR) or Wolf-Rayet (WR) star (denoted as BE\|GCAS\|SDOR\|WR);
6. Cepheid (CEP), including anomalous Cepheid (ACEP), BL Herculis variable (BLHER, also known as CWB), W Virginis variable (CW), \(\delta\) Cephei star (DCEP), RV Tauri-type star (RV), and generic type II Cepheid (T2CEP);
7. cataclysmic variable (CV), excluding supernova and symbiotic star (mentioned separately);
8. \(\delta\) Scuti (DSCT) or \(\gamma\) Doradus (GDOR) or SX Phoenicis (SXPHE) star, including hybrid \(\delta\) Scuti + \(\gamma\) Doradus (DSCT+GDOR) stars (labelled as DSCT\|GDOR\|SXPHE);
9. eclipsing binary (ECL), including Algol (\(\beta\) Persei) type (EA), \(\beta\) Lyrae type (EB), and W Ursae Majoris type (EW);
10. ellipsoidal variable (ELL);
11. star with exoplanet transits (EP);
12. long-period variable (LPV), including long secondary period variable (LSP), Mira (\(o\) Ceti) type (M), Mira or semi-regular variable (M\|SR), OGLE small amplitude red giant (OSARG), small amplitude red giant (SARG), and semi-regular variable (SR) of subtypes SRA, SRB, SRC, SRD, and SRS;
13. microlensing event (MICROLENSING);
14. R Coronae Borealis variable (RCB);
15. RR Lyrae star (RR), including fundamental-mode (RRAB), first overtone (RRC), double-mode (RRD) RR Lyrae star (and anomalous double-mode, ARRD);
16. RS Canum Venaticorum variable (RS);
17. short-timescale object (S);
18. subdwarf B star (SDB) of type V1093 Herculis (V1093HER) or V361 Hydrae (V361HYA);
19. supernova (SN);
20. solar-like star (SOLAR_LIKE), including BY Draconis type (BY), rotating spotted star (ROT), BY and/or ROT star (BY\|ROT), and flaring star (FLARES);
21. slowly pulsating B-type variable (SPB);
22. symbiotic variable star (SYST), including Z Andromedae type (ZAND);
23. variable white dwarf (WD), including a generic class (ZZ), objects of spectral type DA, such as ZZ Ceti (known as ZZA or DAV), DB, such as V777 Herculis (known as ZZB, DBV, V777HER, GD358), or DO, such as GW Virginis pre-white dwarf (comprising ZZO, DOV, GWVIR, PG1159); extremely low mass and hot ZZ Ceti variables were labelled as ELM_ZZA and HOT_ZZA, respectively;
24. young stellar object (YSO), including dipper stars (DIP), eruptive YSOs such as FU Orionis type variables (FUOR), pulsating pre-main-sequence stars (PULS_PMS), Herbig Ae/Be types (HAEBE), including UX Orionis stars (UXOR), and T Tauri star (TTS), among which classical (CTTS), weak-lined (WTTS), and late G to early K type pre-main-sequence (GTTS) stars. [\[itm:yso\]]{#itm:yso label="itm:yso"}
A brief definition of the 24 variability class labels is presented also in the `vari_classifier_class_definition` table of the *Gaia* These labels identify the best classification of a variable source in the field `best_class_name` of table `vari_classifier_result`. In addition, galaxies (labelled as GALAXY) are classified not because they are intrinsically variable, but because they appear to be photometrically variable in the *Gaia* To prevent misinterpretation, galaxies are published exclusively in the extragalactic `galaxy_candidates` table, together with results on galaxies from other *Gaia*
### Source selection[\[sssec:selection\]]{#sssec:selection label="sssec:selection"}
For each class in, two sets of sources were selected: one for training and one for testing purposes, generally with no sources in common, except for classes that were insufficiently represented. For such classes, higher priority was given to the quality of results (by training with all known objects) rather than to their assessment, so the independence of training and test sets of poorly represented classes was not pursued. For both training and test sets, up to six subsets of sources were created with different sample sizes (up to about 500 for the first one and then approximately 1000, 2000,\..., 5000, depending on the available number of sources per class), in order to have, for a given class, a set of the chosen size ready for use, while preserving the representation of the full magnitude range and sky coverage (see Sect. [\[sssec:trainingclasses\]](#sssec:trainingclasses){reference-type="ref" reference="sssec:trainingclasses"}). General conditions and selection procedures were applied to all sources of known variables from the cross-match with *Gaia*
1. At least 5 FoV transit observations in the \(G\)
2. Due to the importance of colour information, at least one observation in \(G_{\rm BP}\)
3. The distribution of angular distances of the cross-match was verified for each class and catalogue: potential spurious matches at significantly higher angular distances than most were removed, unless their correlation with parallax or median \(G\)
4. The minimum standard deviation vs median magnitude in the \(G\)
5. For catalogues whose distribution of sources in the sky had over/under-represented regions due to, for example, limitations of ground-based observations or special targeted regions (and not because of the natural distribution of such objects), sources were sampled to prevent or minimise biases resulting from these limitations. The same procedure was also applied to classes that were extremely rich with respect to the few thousand sources per class needed for training (not to overwhelm the identification of rare classes trained with only a few tens or hundreds of sources). In particular, the sky was subdivided into 3072 Hierarchical Equal Area iso-Latitude Pixelization pixels (corresponding to a resolution of 4, with mean angular spacing of 3.6645 deg)[^1] and the \(k\) nearest neighbours (with \(k\) depending on catalogue) to each HEALPix centre were identified within a radius of 5 degrees. This radius allowed for sources to be drawn from nearby HEALPix as well, in order to avoid gaps or distributions reflecting single HEALPix boundaries. Sources were then randomly selected from such \(k\) neighbours according to the desired downsampling and possible duplicated sources were removed.
6. The steep increase in the number of objects towards faint magnitudes can make the bright end unnoticeable by classifiers, so sources were sampled in median \(G\) In particular, the magnitude distribution was binned in 100 intervals and unique sources were randomly sampled up to a maximum number per bin. The latter depended linearly on the magnitude by a custom (positive) coefficient and offset per class. These coefficient and offset were adjusted for each of the subsample sizes introduced earlier (for training and testing). This procedure made it possible to retain the brightest and the faintest objects, as well as an indication of the most represented magnitudes, with a supplementary option to keep full catalogues of special relevance before adding sources from other catalogues.
Additional per-class selections were applied in specific cases to ensure genuine class representation (see Appendix [\[app:training\]](#app:training){reference-type="ref" reference="app:training"}). The distribution of training sources in various diagrams is presented in Appendix [\[app:plots\]](#app:plots){reference-type="ref" reference="app:plots"}, for each published class, as listed in Sect. [\[sssec:classes\]](#sssec:classes){reference-type="ref" reference="sssec:classes"}.
### Training classes[\[sssec:trainingclasses\]]{#sssec:trainingclasses label="sssec:trainingclasses"}
Commonly confused or physically related (sub)types were combined as listed in the second column of Table [\[tab:training\]](#tab:training){reference-type="ref" reference="tab:training"}, where the values in parentheses indicate an approximate number of representatives or an upper limit when fewer sources were available (see Sect. [\[sssec:selection\]](#sssec:selection){reference-type="ref" reference="sssec:selection"}). In general, larger sample sizes were assigned to more frequent types, although not with a realistic occurrence relative to other types (as known from the literature), otherwise (unweighted) decision-tree based classifiers might optimise models that neglected rare types. As some types overlapped in part with other ones (such as CP and MCP), some of the sets had sources in common, thus duplicated sources were removed. For one per cent of the training set, some of the literature classifications were in conflict with other classes (represented in different class groups). In such cases, duplicated sources that were part of generic class groups (such as S and SOLAR_LIKE) were removed in favour of specific classes and the remaining few hundred sources in conflict (none of which belonged to rare classes) were removed from the training set. Special objects that were wished not to be missed, such as class prototypes, were added to the training set for several types. A similar procedure (except for the addition of special objects) was followed for the test set. Classes that were trained, classified, but not published in *Gaia* DR3 Such omitted classes are not listed in Table [\[tab:training\]](#tab:training){reference-type="ref" reference="tab:training"} to prevent false expectations.
### Attributes[\[sssec:attributes\]]{#sssec:attributes label="sssec:attributes"}
Classification attributes were used to characterize the light variations and the general properties of sources. About 60 attributes were generated, including time series statistics, photometric colours, astrometric parameters, periodicity indicators, combination of photometric and astrometric quantities, comparisons of statistics and correlations between the \(G_{\rm BP}\) The choices of method for attribute selection and of training-set downsampling were due to the necessity of computational efficiency, given the limited time available. The final list of attributes and their definitions are presented in Appendix [\[app:attributes\]](#app:attributes){reference-type="ref" reference="app:attributes"}. The adaptation of parameters (`qso_variability` and `non_qso_variability`) to the *Gaia* The predictive power of classification attributes was limited by the unaccounted reddening of colours, extinguished magnitudes, and no priors on parallax (as a function of celestial location), in addition to the effects of residual outliers, artificial signals, and other photometric imperfections.
## Classifiers[\[ssec:classifier\]]{#ssec:classifier label="ssec:classifier"}
The numbers of classified variable sources and related classes are significantly higher in *Gaia* DR3 The richness of variability types was a major challenge and it was addressed with a multitude of classifiers, which provided different perspectives on the peculiarities of each type of variable object. Classifiers were trained with two machine learning algorithms, as implemented in the H2O platform: Distributed Random Forest and eXtreme Gradient Boosting. Both methods had their pros and cons, DRF results seemed more robust with realistic posterior probability distributions, while XGBoost was more effective in the identification of rare and subtle classes (such as planetary transits). According to the attribute selection (Sect. [\[sssec:attributes\]](#sssec:attributes){reference-type="ref" reference="sssec:attributes"}), optimal results were achieved with 300 trees and with the square root of the number of attributes as the number of randomly sampled attributes to test at each split of any DRF tree. The same number of trees was used for XGBoost too. The following types of classifiers were trained:
1. multi-class DRF and XGBoost classifiers with all classes at the same level (two versions, with minor updates to some class groups),
2. binary DRF (unweighted and weighted) and XGBoost classifiers of one class versus all other classes,
3. meta-classifiers that aimed to combine the best per-class results of the DRF and XGBoost methods, applied to both multi-class and binary classifiers,
4. a two-stage classifier for main-sequence pulsating OBAF-type stars,
which lead to over 100 classifiers in total, with up to 12 classifiers per class. For some of the least represented classes, there were classifiers that could not recover any training sources, so they were excluded. Also, results from all the available classifiers were not necessarily used for each class. In the case of binary classifiers, DRF was more sensitive to the class imbalance than XGBoost, especially for classes that were two orders of magnitude less numerous than the rest of the training set. However, higher weights could be associated with the less represented classes, so DRF was executed also by weighting the targeted class such that the latter became as relevant as all other classes grouped into one, that is, equivalent to a binary classifier with two perfectly balanced classes. In the two-stage classifier for main-sequence pulsating OBAF-type stars, the first stage separated ACV, BCEP, BE, CP, DSCT, GCAS, GDOR, MCP, PULS_PMS, SPB, SXARI, and SXPHE stars (including variability types that may be confused with the pulsating OBAF-type stars, when using only *Gaia* As a result, for each source, only the combined solution of all these different classifiers is recorded in the `best_class_name` and `best_class_score` fields of the `vari_classifier_result` table (as explained in Sect. [\[ssec:verification\]](#ssec:verification){reference-type="ref" reference="ssec:verification"}), and the `vari_classifier_definition` table lists only a single 'combined' classifier instead of the over 100 classifiers that were used to compile the classifications of all sources. As a side note, the training set for the preceding module of general variability detection included the training set described in Sects. [\[sssec:selection\]](#sssec:selection){reference-type="ref" reference="sssec:selection"} and [\[sssec:trainingclasses\]](#sssec:trainingclasses){reference-type="ref" reference="sssec:trainingclasses"} (except for the constant objects) as a single 'variable' class, for the targeted variable sources for *Gaia* DR3
## Verification and filtering of results[\[ssec:verification\]]{#ssec:verification label="ssec:verification"}
After the automated execution of the classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), trained with the selected sources (Sect. [\[sssec:selection\]](#sssec:selection){reference-type="ref" reference="sssec:selection"}) and attributes (Sect. [\[sssec:attributes\]](#sssec:attributes){reference-type="ref" reference="sssec:attributes"}), the results of all classifiers for a given class were verified, alleviated from contaminants, and assessed. The classes whose value was deemed sufficient to be published (separately or merged with other classes) are listed in Sect. [\[sssec:classes\]](#sssec:classes){reference-type="ref" reference="sssec:classes"}, while unconvincing results of other classes were excluded (Sect. [\[sssec:trainingclasses\]](#sssec:trainingclasses){reference-type="ref" reference="sssec:trainingclasses"}). Depending on class, different selective conditions were applied to the classifier results, such as minimum thresholds on the posterior probability to belong to a given class, adjustments of the minimum variability level, colour and/or magnitude cuts, conditions on astrometric and time series parameters, and limitations on environment crowding, among other restrictions, often guided by the distributions of known objects (among the classified ones) and of the candidates. The list of such filters and thresholds are presented for each class in Sect. [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, to help the interpretation of the results where they are described. Exceptionally, following the feedback from SOS modules, a small fraction of sources (of particular interest) might override the general selection conditions (such as the special objects mentioned in Sect. [\[sssec:trainingclasses\]](#sssec:trainingclasses){reference-type="ref" reference="sssec:trainingclasses"} or bona fide sources with peculiar behaviour), provided they were classified correctly. Additional sources were removed following a further post-processing verification, typically involving suspect features for a very small fraction of candidates.
## Classification score[\[ssec:class_score\]]{#ssec:class_score label="ssec:class_score"}
After the verification filters, the remaining candidates were assigned a classification score, which was derived from the posterior probabilities of the classifiers used to identify the sources of a given class. Classifier posterior probabilities were not calibrated, so their values were not directly comparable, because of the use of different methods and classifier structures. In order to treat the posterior probabilities of all classifiers on the same footing for a given class, they were converted into normalised ranks (for each classifier), which could then be compared across different classifiers. For each classifier, this normalised rank was computed by sorting posterior probabilities of a certain class in ascending order. The source rank depended on its location in this ordered list and identical probabilities corresponded to identical ranks. Denoting the number of different probabilities (that is, the number of sources with posterior probabilities of a given class minus that of sources with duplicated probability values for that class) by \(N\), the possible score values were represented by the ranks of the probabilities, from the first to the \(N\)-th, normalised by \(N\): {1, 2,\..., \(N\)} \(/\) \(N\) (some of which were repeated in presence of duplicated probabilities). Thus, for each classifier, the score associated with a source of a given class ranged from a minimum value greater than zero to a maximum of one. The median of such normalised ranks from different classifiers, for a given source and class (`best_class_name`), was stored as `best_class_score` in the `vari_classifier_result` table. Exceptionally, for BCEP and WD, the normalised ranks from probabilities of the general variability detection classifier (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) were included in the score evaluation, when higher variability implied higher reliability. Occasional gaps in the distribution of the classification score of some classes were created by the removal of sources associated with multiple classes (Sect. [\[ssec:overlaps\]](#ssec:overlaps){reference-type="ref" reference="ssec:overlaps"}) and also of some questionable candidates, following the post-processing verification. The average and extreme values of the classification scores are presented in Table [1](#tab:results){reference-type="ref" reference="tab:results"} for each class.
## Multi-class source reduction[\[ssec:overlaps\]]{#ssec:overlaps label="ssec:overlaps"}
After the verification filters (Sect. [\[ssec:verification\]](#ssec:verification){reference-type="ref" reference="ssec:verification"}) and assignment of classification scores (Sect. [\[ssec:class_score\]](#ssec:class_score){reference-type="ref" reference="ssec:class_score"}), the combination of all per-class candidates revealed about 110 thousand sources (less than 1 % of all classified sources) associated with multiple (usually two) classes. The vast majority of them were not due to genuine concurrent variability types (as it may happen, for example, in an eclipsing binary system with intrinsically variable components, perhaps with transiting exoplanets too). Multiple variability was not pursued in *Gaia* DR3 The following set of rules was applied to multi-class sources in order to reduce them to a single class per source.
1. The least numerous classes would be unfairly reduced if their candidates were left to compete with those of the most numerous classes, so rare classes of ACYG, BCEP, BE\|GCAS\|SDOR\|WR, EP, MICROLENSING, RCB, SDB, SN, SPB, SYST, and WD were safeguarded against any other alternative class.
2. [\[item:score-xm\]]{#item:score-xm label="item:score-xm"} For 1290 sources that remained with multiple classes after the application of all the other rules, the class corresponding to the highest classification score was kept.
3. Candidates of some classes were considered dispensable because they were byproducts of classification (such as galaxies), or related to SOS modules that did not rely on classification and were expected to provide higher quality candidates than a general classification (with expert procedures dedicated to a single class). Such classes included S, GALAXY, SOLAR_LIKE, and RS, ordered from the most to the least dispensable class. Dispensable classes were dropped from multi-class sources, provided at least one non-dispensable class per source remained. For sources with all dispensable multi-classes, only the least dispensable one was kept (following the above mentioned list ordered by decreasing dispensability).
4. For the special objects (Sect. [\[sssec:trainingclasses\]](#sssec:trainingclasses){reference-type="ref" reference="sssec:trainingclasses"}) with multiple classes, only the known class was kept, when available among the classifications, otherwise the standard treatment of other multi-class sources was followed.
5. For multi-class sources that were present also in at least one SOS module, different scenarios were possible:
1. the classified classes included at least one class for which a corresponding SOS module did not exist or whose source was absent from the corresponding SOS module:
1. all classes with the source absent from the corresponding SOS module, provided the latter existed, were removed (e.g. for a source classified as BE+AGN and present in the upper main-sequence oscillator SOS module but not in the AGN SOS one, the AGN class was removed);
2. the classes with the source present in the corresponding SOS modules were kept and all other classes were removed (e.g. for a source classified as BE+AGN+RR and present only in the AGN SOS module, the BE and RR classes were removed);
2. all classes had the source present in the SOS modules: classes unmatched by the corresponding SOS classes were removed, as long as at least one of the other classes matched an SOS class, otherwise item ([\[item:sos-similarity\]](#item:sos-similarity){reference-type="ref" reference="item:sos-similarity"}) was applied;
3. [\[item:sos-similarity\]]{#item:sos-similarity label="item:sos-similarity"} for all classes that had the source present in SOS modules but that did not match any of the SOS classes, the class matching criterion was extended to include the similarity of such classes with the following SOS modules, for partial validation:
1. BE\|GCAS\|SDOR\|WR and GALAXY sources in the AGN SOS module,
2. CEP sources in the RR Lyrae and the long-period variable SOS modules,
3. ECL sources in the RR Lyrae SOS module,
4. ELL and EP sources in the eclipsing binary SOS module,
5. RR sources in the Cepheid or eclipsing binary SOS modules,
6. S in the eclipsing binary or the RR Lyrae SOS modules,
7. YSO in the rotational modulation SOS module;
classes not similar to those of SOS modules were removed, as long as at least one of the other classes was similar to an SOS class, otherwise item ([\[item:score-xm\]](#item:score-xm){reference-type="ref" reference="item:score-xm"}) was applied.
The existence of SOS modules matching some of the multiple classes, but without the multi-class sources present in any of them, added no information to the selection rules (bona fide candidates can be excluded from the corresponding SOS modules for various reasons, such as insufficient sampling for reliable model results).
## Classification vs SOS modules[\[ssec:class-sos\]]{#ssec:class-sos label="ssec:class-sos"}
Although SOS modules provided purer class samples than those from classification, the latter did not necessarily represent a superset of the corresponding SOS modules. As shown in Table [1](#tab:results){reference-type="ref" reference="tab:results"}, in addition to the 'extra' sources in classification and the ones in 'common' with respect to the SOS modules, some sources were classified as 'other' classes, different from the expected SOS modules, and others were 'missed' in the published classification results. The reasons for the 'other' and 'missed' classifications are listed in the following.
1. Some SOS modules did not depend on classification but relied on special variability detection or extraction of candidates (as for microlensing events, short timescales, and solar-like rotation modulation).
2. For the SOS modules that depended on classification:
1. they received input candidates before the multi-class source treatment described in Sect. [\[ssec:overlaps\]](#ssec:overlaps){reference-type="ref" reference="ssec:overlaps"}, so sources could be assigned to 'other' classes as a consequence of enforcing a single class per source;
2. some of these SOS modules included classification candidates from multiple classes, due to similar features (such as long-period variables and symbiotic stars);
3. given the advanced class-specific SOS verification, permissive classifier probability thresholds or their combination with other parameters allowed for a larger initial set of candidates (before verification cuts) than the one considered for classification, for a given class.
3. The results of several SOS modules were not mutually exclusive, while only one class per source was required by classification.
# Results[\[sec:results\]]{#sec:results label="sec:results"}
From the about 400 million sources identified as potentially variable by the preceding general variability detection module (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), 12 428 245 objects were selected for the *Gaia* DR3 The classified variable sources are identified by `source_id`, they are assigned class labels (`best_class_name`), and they are associated with classification scores (`best_class_score`) in the `vari_classifier_result` table. The galaxies whose classification was based on photometric time series are labelled as GALAXY in the field `vari_best_class_name`, with the classification score stored in the field `vari_best_class_score` of the `galaxy_candidates` table (see Appendix [\[app:queries\]](#app:queries){reference-type="ref" reference="app:queries"}). Figure [\[fig:histo\]](#fig:histo){reference-type="ref" reference="fig:histo"} depicts the number of classified sources per class group, from the most to the least numerous class (galaxies and R Coronae Borealis stars, respectively). Besides common and rare classes, certain class groups such as ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI and S were published for exploratory use to the benefit of the community. Table [1](#tab:results){reference-type="ref" reference="tab:results"} summarises the classification results for each class: the source counts, the distribution of classification scores (Sect. [\[ssec:class_score\]](#ssec:class_score){reference-type="ref" reference="ssec:class_score"}), completeness (the ratio of identified to all known sources of a given class), contamination (the fraction of contaminants among the classifications of a given class, i.e. 1\(-\) purity), the \(F_1\) score (the harmonic mean of completeness and purity), the maximum \(F_1\) value and the corresponding minimum classification score (for an optimal balance between completeness and contamination), and a comparison with SOS modules (explained in Sect. [\[ssec:class-sos\]](#ssec:class-sos){reference-type="ref" reference="ssec:class-sos"}). Sample ADQL queries in Appendix [\[app:queries\]](#app:queries){reference-type="ref" reference="app:queries"} include indications of how to retrieve all the candidates of a given class (from classification and SOS modules) and how to reproduce the comparison of results from classification vs SOS modules in Table [1](#tab:results){reference-type="ref" reference="tab:results"}. Completeness and contamination rates are computed for each class globally, with no restriction on sky location, magnitude, amplitude, signal-to-noise, period range, or other parameters, thus they might be more conservative than the detailed estimates presented in the papers describing the SOS module results. Typically, such rates depend on surveys from the literature in optical bands (or with detectable counterparts in *Gaia* More details on the class composition of the true and unknown positives, in addition to the false positives and negatives, are presented in Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}. For an overview on the identified classes of the least sampled sources, Fig. [\[fig:leastsampled_histo\]](#fig:leastsampled_histo){reference-type="ref" reference="fig:leastsampled_histo"} shows the occurrence of classified objects as a function of median \(G\) Similarly, the distributions in the sky of variable classifications with up to 10, 15, and 20 \(G\)
The top-five most common classes among the classified sources with the lowest number of observations are listed as follows (with the number of sources satisfying the sampling conditions indicated in parentheses):
1. GALAXY (43 428), LPV (30 203), SN (2515), CV (818), and AGN (579), for sources with up to 10 \(G\)
2. GALAXY (273 603), LPV (104 133), AGN (20 846), RR (4295), and SN (2833), for sources with up to 15 \(G\)
3. GALAXY (643 935), LPV (170 681), ECL (104 043), AGN (96 760), and RR (22 005), for sources with up to 20 \(G\)
Galaxies are the most common sources with few observations because they are the faintest classified objects and thus susceptible to reduced detectability at the faint end of the *Gaia* The following subsections present more details for each class, among which the number of selected classifiers (as an upper limit as not all sources were necessarily classified as a given class by all classifiers), the verification cuts, special considerations on completeness or contamination (when applicable), and the references to relevant articles. A selection of diagrams illustrating the general properties of classified sources for each class is presented in Appendix [\[app:plots\]](#app:plots){reference-type="ref" reference="app:plots"}, followed by a visualisation of completeness vs contamination rates as a function of minimum classification score, \(F_1\) score vs minimum classification score, and samples of light curves in the \(G\) An overview of all variability results (including classifications), with figures and tables combining metrics of several classes, is presented in.
## ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI stars[\[ssec:acv\]]{#ssec:acv label="ssec:acv"}
This class group accounts for 10 779 variable sources from a set of different types, which are particularly challenging and share similar features (according to the *Gaia* DR3 This group of variability types includes also some SPB stars, which are not mentioned in this class group denomination because of the attempt to split them into their own class (Sect. [\[ssec:spb\]](#ssec:spb){reference-type="ref" reference="ssec:spb"}). This class was considered by the upper main-sequence oscillator SOS module, but no source satisfied its requirements. These candidates were selected from three multi-class and binary meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with some minimum probability level. The following additional conditions were required (employing field names in the `vari_summary` table).
1. The values of `std_dev_mag_g_fov` were set above the third quartile of the standard deviations in \(G\)
2. To further reduce contaminants, a higher level of variability probability than the one used for the general variability detection (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was required.
3. As predictions still tended to be less variable than those exhibited in the literature, a minimum threshold was set for the single-band Stetson index: `stetson_mag_g_fov` \(>\) 1.5.
4. The reddened colour `median_mag_bp - median_mag_rp` was set to be bluer than 1 mag, in order to remove sparse outliers, whose presence among the training sources (see Fig. [\[fig:app:ACV_trn\]](#fig:app:ACV_trn){reference-type="ref" reference="fig:app:ACV_trn"}d) was subsequently deemed questionable. In fact, the bulk of results was bluer than 0.5 mag and redder objects were associated with low scores (Fig. [\[fig:app:ACV\]](#fig:app:ACV){reference-type="ref" reference="fig:app:ACV"}b).
5. While training objects reduced steeply beyond a median \(G\)
According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional candidates of this class group may be found among the false positives of the SPB (221) and ACYG (7) classes, where the number of known sources indicated in parentheses is a lower limit.
## \(\alpha\) Cygni stars (ACYG)[\[ssec:acyg\]]{#ssec:acyg label="ssec:acyg"}
The classification of the \(\alpha\) Cygni-type variables comprised 329 candidates, selected from two binary meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with some minimum probability threshold, with no further condition. It is one of the rare types for which the training set sources were included in the assessment of completeness and contamination rates (in Table [1](#tab:results){reference-type="ref" reference="tab:results"}) because of the lack of known stars of this type. Among the contaminating classes, GCAS stars are the most common, as expected from the partial overlap in the bright and blue part of the colour--magnitude diagram (see Figs. [\[fig:app:ACYG\]](#fig:app:ACYG){reference-type="ref" reference="fig:app:ACYG"}b and [\[fig:app:BE\]](#fig:app:BE){reference-type="ref" reference="fig:app:BE"}b). From the diagrams that include `median_mag_g_fov` from the `vari_summary` table (as in Figs. [\[fig:app:ACYG\]](#fig:app:ACYG){reference-type="ref" reference="fig:app:ACYG"}b,c,g), two clumps of candidates separated at `median_mag_g_fov` \(\approx\) 9.5 mag are clearly visible; the fainter one is also associated with low scores and, as apparent from the sky map in Fig. [\[fig:app:ACYG\]](#fig:app:ACYG){reference-type="ref" reference="fig:app:ACYG"}a, 82 % of them (73/89) are projected in the direction of the Magellanic Clouds, with an average `best_class_score` of 0.3 (in the `vari_classifier_result` table). The bright and faint clumps were not so obvious from the training sources, as the relative occurrence in the Magellanic Clouds was strongly underrepresented and consequently related to lower scores than Galactic sources. The upper main-sequence oscillator SOS module did not consider ACYG classifications as input, but it implicitly did so as its candidates were selected before the implementation of the rules that reduced multi-class sources to a single source (Sect. [\[ssec:overlaps\]](#ssec:overlaps){reference-type="ref" reference="ssec:overlaps"}) and eventually five ACYG candidates became upper main-sequence oscillator candidates too.
## Active galactic nuclei (AGN)[\[ssec:agn\]]{#ssec:agn label="ssec:agn"}
The 1 035 207 AGN candidates were selected from 11 multi-class and binary meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) following some minimum probability level and were filtered according to the following conditions (employing field names in the `vari_summary` and `gaia_source` tables), guided mostly by the *Gaia*
1. A higher level of variability probability than the one used for the general variability detection (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was required in order to focus on truly variable AGN, considering the possible contribution of artefacts such as the one described by, when the host galaxy is detectable; the consequent offset with respect to the general threshold is visible in Fig. [\[fig:app:AGN\]](#fig:app:AGN){reference-type="ref" reference="fig:app:AGN"}e.
2. The renormalised unit weight error `ruwe` was set to be lower than 1.2, as the astrometric measurements of the vast majority of AGN fit well the single-source model of the astrometric solution (97 % of *Gaia*
3. The `abbe_mag_g_fov` parameter was set to be lower than 0.9, as long timescale variations with respect to with *Gaia*
4. The number of sources within 100 arcsec from each AGN candidate (computed by excluding the contribution of the AGN source at the centre) was set to be less than 126, in order to avoid crowded stellar fields in the foreground, as the steep increase in the number of stars per solid angle can lead to regions with excessive false positive rates, sometimes even where very few AGN are detectable in the optical wavelengths following an enhanced interstellar extinction. The number density threshold corresponds to a suspicious increase of the fraction of unknown-to-known AGN towards higher crowding. The most evident sky regions affected by this conditions are the Magellanic Clouds and the Galactic plane, as shown in Fig. [\[fig:app:AGN\]](#fig:app:AGN){reference-type="ref" reference="fig:app:AGN"}a. Occasional AGN candidates appear at low Galactic latitudes, especially towards the Galactic anti-centre, most of which are associated with low classification scores.
5. For sources with existing parallax, the latter was assumed to be insignificant, as for any extragalactic object, after correcting it for a global systematic offset of \(-\)`<!-- -->`{=html}0.017 mas, consequently its ratio with an assumed Gaussian uncertainty was expected to follow a standard normal distribution (with zero average and unit variance). We required parallax measurements not to deviate beyond five sigma: \| `parallax`+0.017 mas \| \(/\) `parallax_error` \(<\) 5. Only 282 of the published AGN candidates had no parallax (nor proper motion) available, while the remaining 1 034 925 sources had (at least) five-parameter astrometric solutions.
6. The same rationale was applied to AGN classifications with existing proper motion, leading to the following condition, which accounts also for the correlation between the measured proper motion along the right ascension and declination directions:\
`pmra`\(^2\)/`pmra_error`\(^2+\) `pmdec`\(^2\) \(/\) `pmdec_error`\(^2\)+\(-\) 2 (`pmra`/`pmra_error`) \(\times\) (`pmdec` \(/\) `pmdec_error`) \(\times\[\times\) `pmra_pmdec_corr` \(<\) 5\(^2\) (1 \(-\) `pmra_pmdec_corr`\(^2\)).
7. Three conditions were set in the *Gaia*
1. `median_mag_g_fov` \(-\) `median_mag_rp` \(>\]-\) 3.7 (`median_mag_bp` \(-\) `median_mag_g_fov`) \(-\) 0.7,
2. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\)\(-\) 0.75 (`median_mag_bp` \(-\) `median_mag_g_fov`) \(+\) 1.55,
3. `median_mag_g_fov` \(-\) `median_mag_rp` \(>\) (`median_mag_bp` \(-\) `median_mag_g_fov`) \(-\) 0.6,
whose impact is clearly visible in Fig. [\[fig:app:AGN\]](#fig:app:AGN){reference-type="ref" reference="fig:app:AGN"}c.
The set of conditions listed above proved efficient in the selection of the bulk of AGN candidates while keeping low contamination, however, peculiar objects may not conform to these general rules. In order to recuperate known AGN with different behaviour from average and those of particular interest for which *Gaia* Similarly, about 50 thousand AGN candidates were recovered following the feedback from the AGN SOS module. According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional candidates may be found among the false positives of the GALAXY (1884) and CV (37) classes, where the number of known sources indicated in parentheses is a lower limit. In comparison with the AGN SOS module requirements described in, the conditions applied to classification results were generally more permissive, in particular:
1. at least 5 (instead of 20) FoV transits in the \(G\)
2. no cut in the `ruwe` vs `abbe_mag_g_fov` plane, but a slightly more restrictive limit on the `ruwe` parameter (less than 1.2 instead of 1.3);
3. no constraint on the amount of scan-angle dependent signal;
4. no requirements on the following parameters, which are defined and published in the `vari_agn` table: `fractional_variability_g`, `structure_function_index`, `qso_variability` and `non_qso_variability`.
Parallax was used as classification attribute (among others listed in Appendix [\[app:attributes\]](#app:attributes){reference-type="ref" reference="app:attributes"}) and consequently it introduced a bias in the distribution of the parallax significance of AGN candidates with respect to `best_class_score` (in the `vari_classifier_result` table). Figure [\[fig:AGN_plx_pmra_pmdec\]](#fig:AGN_plx_pmra_pmdec){reference-type="ref" reference="fig:AGN_plx_pmra_pmdec"}a illustrates a comparison between the top-150 000 (high-score) AGN classifications and the bottom-150 000 (low-score) candidates: there is a distinct deficit of sources with positive parallax among the most reliable AGN candidates, clearly yielding to stellar objects in the Galaxy (often associated with positive parallax), while the low-score sample is biased in the opposite direction, in addition to being the main contributors to thicker than Gaussian tails at both ends. Eventually, the interplay among high and low score AGN leads to an overall distribution that approximates a standard normal distribution. As expected, no training AGN source and only 13 low-score AGN candidates fulfil the condition of `parallax_over_error` \(>\) 5 required for observational Hertzsprung--Russell diagrams (Figs. [\[fig:app:AGN_trn\]](#fig:app:AGN_trn){reference-type="ref" reference="fig:app:AGN_trn"}d and [\[fig:app:AGN\]](#fig:app:AGN){reference-type="ref" reference="fig:app:AGN"}d). A different distribution of high vs low score candidates is expected for every classification attribute that can discriminate AGN from other classes, such as the \(G\) Proper motion was not among the classification attributes and Figs. [\[fig:AGN_plx_pmra_pmdec\]](#fig:AGN_plx_pmra_pmdec){reference-type="ref" reference="fig:AGN_plx_pmra_pmdec"}b,c do not show evident biases in the distribution of proper motion components, except for the expected thicker tails and more extreme values from low-score candidates with respect to those from sources in the high-score sample. The asymmetric proper motion distribution of low-score AGN, in particular in the declination direction (Fig. [\[fig:AGN_plx_pmra_pmdec\]](#fig:AGN_plx_pmra_pmdec){reference-type="ref" reference="fig:AGN_plx_pmra_pmdec"}c), is characteristic of the motion of stars in the Galaxy and thus it suggests stellar contamination. Table [3](#tab:agn){reference-type="ref" reference="tab:agn"} summarises the averages and standard deviations of the distributions illustrated in Fig. [\[fig:AGN_plx_pmra_pmdec\]](#fig:AGN_plx_pmra_pmdec){reference-type="ref" reference="fig:AGN_plx_pmra_pmdec"}, in order to help assess biases and similarity with respect to a standard normal distribution.
While the completeness rate of AGN candidates in Table [1](#tab:results){reference-type="ref" reference="tab:results"} is consistent with the assessment of, the contamination rate is deemed underestimated, given the large set of 200 081 'unknown' positives stated in Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}. Comparing the 1 034 925 AGN candidates with at least five-parameter astrometric solutions with 17 catalogues from the literature as described in, only about 45 thousand sources were unknown. While many of them may still be genuine AGN, literature catalogues may also have false positives, thus an approximate estimate for the contamination is \(\sim\) 5 %. A comparison of the sources from all of the classes identified by variability that overlap with those from QSO modules published by other *Gaia*
## \(\beta\) Cephei stars (BCEP)[\[ssec:bcep\]]{#ssec:bcep label="ssec:bcep"}
The classified \(\beta\) Cephei type candidates include 1475 sources, 174 of which are shared with the upper main-sequence oscillator SOS module, although multi-periodicity may not be easily detectable with the current average number of observations per source. More information on the upper main-sequence oscillators is provided in. Training sources for the \(\beta\) Cephei class included a tail of objects at the blue end with lower intrinsic luminosity than possible for this class (see Fig. [\[fig:app:BCEP\]](#fig:app:BCEP){reference-type="ref" reference="fig:app:BCEP"}d), explaining the contamination by GCAS mentioned in Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}. Figures [\[fig:app:BCEP\]](#fig:app:BCEP){reference-type="ref" reference="fig:app:BCEP"}b,e,g, show that candidates with fainter apparent magnitudes tend to have lower scores, mirroring the decreasing occurrence of training sources towards faint magnitudes (only three training objects had \(G\) The BCEP candidates were selected from 11 binary, multi-class, multi-stage, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) according to some minimum probability threshold and two conditions were set in the *Gaia*
1. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\)model(\(G-G_{\rm RP}\)
2. `median_mag_g_fov` \(-\) `median_mag_rp` \(>\)model(\(G-G_{\rm RP}\)
employing field names in the `vari_summary` table. Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional BCEP candidates may be found among the false positives of the following classes: ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (3), BE\|GCAS\|SDOR\|WR (1), CV (1), and RS (1), where the number of known sources indicated in parentheses is a lower limit.
## BE\|GCAS\|SDOR\|WR stars[\[ssec:be\]]{#ssec:be label="ssec:be"}
This class group includes a set of 8560 eruptive variables of the following types: B-type emission line star, \(\gamma\) Cassiopeiae, S Doradus, and Wolf-Rayet star. Stars of types BE\|GCAS, SDOR, and WR can significantly overlap in the blue and bright end of the Hertzsprung--Russell diagram and share similarities in the irregular light changes. Therefore, although candidates of these types were selected independently (including the verification filters mentioned below), they were subsequently merged, preserving their original classification scores. The bimodal distribution in apparent magnitude, clearly visible in several panels of Figs. [\[fig:app:BE_trn\]](#fig:app:BE_trn){reference-type="ref" reference="fig:app:BE_trn"}b,e,g and [\[fig:app:BE\]](#fig:app:BE){reference-type="ref" reference="fig:app:BE"}b,e,g, is not related to different class components, but to sources (mostly BE\|GCAS) located in the Galaxy (bright clump) vs the Magellanic Clouds (faint clump). The latter is also associated with the clump with scattered colours in Figs. [\[fig:app:BE_trn\]](#fig:app:BE_trn){reference-type="ref" reference="fig:app:BE_trn"}c and [\[fig:app:BE\]](#fig:app:BE){reference-type="ref" reference="fig:app:BE"}c. The BE\|GCAS candidates were obtained from two binary and multi-class classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with some minimum probability level. They were further constrained by the following criteria (employing field names in the `vari_summary` and `gaia_source` tables).
1. A higher level of variability probability than the one used for the general variability detection (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was required, in order to prioritise high-amplitude candidates.
2. The renormalised unit weight error `ruwe` was restricted to values lower than 1.4, which was consistent with the vast majority of known objects of this class.
3. The values of the Abbe parameter were constrained for the three *Gaia*
1. `abbe_mag_g_fov` \(<\) 0.8,
2. `abbe_mag_bp` \(<\) 1,
3. `abbe_mag_rp` \(<\) 1.
Exceptionally, probabilities from the general variability detection classifier were included in the computation of the classification score, to further increase the weight of high-amplitude candidates. Given the similarities between Be pulsators and SPB stars and related confusion from classifiers, the upper main-sequence oscillator SOS module considered, among other input sources, classifications of BE\|GCAS stars (223 of which were published as upper main-sequence oscillators). The SDOR classifications were derived from 10 binary, multi-class, and meta-classifiers with minimum probability thresholds, some of which were combined with the maximum brightness limits of `median_mag_g_fov` \(<\) 13 or 17 mag, to exclude a suspicious component characterized by very faint and low-probability candidates, some of which however extended above the chosen minimum probability. The condition of `ruwe` \(<\) 2 was applied to all candidates, as no known SDOR star existed above such a limit. Eight further questionable candidates were removed following visual inspections. The WR candidates were selected from four binary and meta-classifiers according to minimum probability thresholds. The number of sources within 100 arcsec from each WR candidate (computed by excluding the contribution of the WR source at the centre) was set to be greater than 157, as these relatively young stars were expected to be within the Galactic disc and no WR star was known to exist in lower star density fields. According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional BE\|GCAS\|SDOR\|WR candidates may be found among the false positives of the following classes: BCEP (55), SPB (42), ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (34), and ACYG (20), where the number of known sources indicated in parentheses is a lower limit.
## Cepheids (CEP)[\[ssec:cep\]]{#ssec:cep label="ssec:cep"}
The classification of Cepheid variables included 16 141 stars of types \(\delta\) Cephei, anomalous Cepheid, and type-II Cepheid. For this class, the selection of Cepheids from the relevant SOS module was followed, with the addition of 1154 candidates from three binary and multi-class classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with strict probability conditions, including about 400 sources that were rejected from the SOS module because reliable periods (and thus models) could not be achieved with the *Gaia* The pulsating nature of these objects is clearly visible in Figs. [\[fig:app:CEP_trn\]](#fig:app:CEP_trn){reference-type="ref" reference="fig:app:CEP_trn"}g and [\[fig:app:CEP\]](#fig:app:CEP){reference-type="ref" reference="fig:app:CEP"}g (until noise prevails at the faint end), indicating greater amplitudes in the \(G_{\rm BP}\) Low-amplitude candidates (`std_dev_mag_g_fov` \(<\) 0.06 mag) and those at the faint end (`median_mag_g_fov` \(<\) 20.3 mag) are associated with low classification scores (Fig. [\[fig:app:CEP\]](#fig:app:CEP){reference-type="ref" reference="fig:app:CEP"}e). The main classes responsible of the few percent of contamination are represented by RR Lyrae, eclipsing binaries, and long-period variables (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}). Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional CEP candidates may be found among the false positives of the following classes: RS (494), RR (274), LPV (199), ECL (114), DSCT\|GDOR\|SXPHE (33), YSO (13), BE\|GCAS\|SDOR\|WR (11), CV (7), RCB (4), S (4), AGN (3), ELL (2), SYST (2), ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (1), and BCEP (1), where the number of known sources indicated in parentheses is a lower limit.
## Cataclysmic variables (CV)[\[ssec:cv\]]{#ssec:cv label="ssec:cv"}
The 7306 candidates of cataclysmic variables (excluding supernovae and symbiotic stars, described in Sects. [\[ssec:sn\]](#ssec:sn){reference-type="ref" reference="ssec:sn"} and [\[ssec:syst\]](#ssec:syst){reference-type="ref" reference="ssec:syst"}, respectively) were selected from five binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), after the application of minimum probability thresholds. The verification of CV classifications lead to the following conditions (employing field names in the `vari_summary` table).
1. A higher level of variability probability than the one used for the general variability detection was required to omit doubtful low signal-to-noise candidates.
2. A suspicious clump of extremely blue candidates was filtered out by `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) \(-\) 4 mag.
3. A thin tail of blue outliers was excluded by the condition `median_mag_bp` \(-\) `median_mag_rp` \(>\) \(-\) 0.5 mag.
4. The number of sources within 100 arcsec from each CV candidate (computed by excluding the contribution of the CV source at the centre) was set to be less than 408, to prevent the increase of candidates towards the end of the tail of the distribution of CV from the literature.
5. The following two conditions served to exclude two clumps of candidates in the `outlier_median_g_fov` vs `skewness_mag_g_fov` space that were only marginally populated by literature instances:
1. `outlier_median_g_fov` \(>\) 20,
2. `skewness_mag_g_fov` \(>\) \(-\) 4.
The two features with skewness \< \(-\) 3 that appear in the plot of `skewness_mag_g_fov` vs `abbe_mag_g_fov` (Fig. [\[fig:app:CV\]](#fig:app:CV){reference-type="ref" reference="fig:app:CV"}f) are related to faint sources, with median \(G\) According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional CV candidates of this class may be found among the false positives of the SN (22) and WD (12) classes, where the number of known sources indicated in parentheses is a lower limit.
## DSCT\|GDOR\|SXPHE stars[\[ssec:dsct\]]{#ssec:dsct label="ssec:dsct"}
This class group includes 748 058 candidates of types \(\delta\) Scuti, \(\gamma\) Doradus, and SX Phoenicis, in particular as DSCT and SXPHE types could not be distinguished without metallicity or other indicators of type I vs type II star. These classifications were also the main contributors to the upper main-sequence oscillator SOS module, which were further analysed in. This sample is dominated by low amplitude candidates (`std_dev_mag_g_fov` \(<\) 0.01 mag for 537 769 sources), which are expected to be significantly contaminated, as suggested from their distribution of `std_dev_mag_bp` / `std_dev_mag_rp` \(<\) 1 between `median_mag_g_fov` of 13 and 17 mag, as shown in Fig. [\[fig:app:DSCT\]](#fig:app:DSCT){reference-type="ref" reference="fig:app:DSCT"}g. About seven thousand candidates with \(G_{\rm BP}-G_{\rm RP}\) The DSCT\|GDOR\|SXPHE candidates were extracted from 10 binary, multi-class, multi-stage, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. The following additional criteria were applied (employing field names in the `vari_summary` table).
1. Removal of outliers in the *Gaia*
1. `median_mag_bp` \(-\) `median_mag_rp` \(>\) \(-\)`<!-- -->`{=html}0.5 mag,
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) \(-\)`<!-- -->`{=html}1 mag,
3. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\) 1 mag.
2. Constraints on the colour--colour scatter around the mean relation (with a model of \(G-G_{\rm RP}\)
1. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\) model(\(G-G_{\rm RP}\)
2. `median_mag_g_fov` \(-\) `median_mag_rp` \(>\) model(\(G-G_{\rm RP}\)
The conditions described above were overridden by candidates that existed also in the upper main-sequence oscillator SOS module. According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional DSCT\|GDOR\|SXPHE candidates may be found among the false positives of the following classes: RS (792), AGN (275), ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (49), BCEP (11), WD (4), and SDB (1), where the number of known sources indicated in parentheses is a lower limit.
## Eclipsing binaries (ECL)[\[ssec:ecl\]]{#ssec:ecl label="ssec:ecl"}
The classification of eclipsing binaries included 2 184 356 systems of types \(\beta\) Persei (Algol), \(\beta\) Lyrae, and W Ursae Majoris. They were obtained from four meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) and followed the selection of the corresponding SOS module described by, including the following conditions (employing field names in the `vari_summary` table).
1. A minimum apparent brightness in the \(G\)
2. A minimum value of the time series skewness, computed from \(G\)
3. A minimum number of clean FoV transits in the \(G\)
4. Further constraints on period and model properties listed in the `vari_eclipsing_binary` table.
The ratio between \(G_{\rm BP}\) Low-amplitude and low-skewness candidates tend to be associated with low classification scores (Fig. [\[fig:app:ECL\]](#fig:app:ECL){reference-type="ref" reference="fig:app:ECL"}e,f). According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional ECL candidates may be found among the false positives of the following classes: RS (10 908), S (5713), RR (4318), DSCT\|GDOR\|SXPHE (3402), LPV (2507), ELL (2135), SOLAR_LIKE (286), YSO (219), BE\|GCAS\|SDOR\|WR (152), CEP (104), ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (61), CV (28), ACYG (6), EP (3), SDB (3), WD (3), MICROLENSING (2), where the number of known sources indicated in parentheses is a lower limit.
## Ellipsoidal variables (ELL)[\[ssec:ell\]]{#ssec:ell label="ssec:ell"}
The classification of 65 300 ellipsoidal variables targeted input sources for the compact companion SOS module. The training objects were not sufficiently representative of the whole sky (Fig. [\[fig:app:ELL_trn\]](#fig:app:ELL_trn){reference-type="ref" reference="fig:app:ELL_trn"}a) and thus low scores were assigned to candidates beyond the Galactic bulge region (Fig. [\[fig:app:ELL\]](#fig:app:ELL){reference-type="ref" reference="fig:app:ELL"}a). Most of them are likely contaminants, considering that low score classifications contribute negligibly to completeness but they double the contamination rate (see Fig. [\[fig:app:ELL_cc\]](#fig:app:ELL_cc){reference-type="ref" reference="fig:app:ELL_cc"}a). As expected, most of the contaminants are represented by W Ursae Majoris eclipsing binaries. The ELL candidates were selected from 10 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. They fulfil the following conditions (employing field names in the `vari_summary` and `gaia_source` tables).
1. The renormalised unit weight error `ruwe` was set to values lower than 1.2, to restrict an excess of candidates with high `ruwe` values (and thus unreliable astrometric solutions).
2. In order to better fit the distribution of known ELL stars in standard deviation vs median \(G\)
1. `std_dev_mag_g_fov` \(>\) 1.5 times the third quartile of the standard deviations in \(G\)
2. `std_dev_over_rms_err_mag_g_fov` \(>\) 4.
3. Additional constraints removed candidates from colour and magnitude ranges that were poorly represented in the literature:
1. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) 0.3 mag,
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) (`median_mag_bp` \(-\) `median_mag_rp`)\(^2\) \(+\) (`median_mag_bp` \(-\) `median_mag_rp`) \(-\) 0.65 mag,
3. `median_mag_g_fov` \(<\) 19 mag.
4. The number of sources within 100 arcsec from each ELL candidate (computed by excluding the contribution of the ELL source at the centre) was set to be greater than 314, to avoid a suspicious concentration of candidates around the Galactic anti-centre.
These criteria were overridden by candidates that existed also in the compact companion SOS module (including the intention of the last item above). According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional ELL candidates may be found among the false positives of the ECL (1056), LPV (731), SPB (9), BCEP (3), and ACYG (2) classes, where the number of known sources indicated in parentheses is a lower limit.
## Exoplanets (EP)[\[ssec:ep\]]{#ssec:ep label="ssec:ep"}
The 214 stars classified with exoplanet transits have percent-level variations in the \(G\) The weakness of the signal and the typically small number of observations in transit implied the necessity of a dedicated period search algorithm such as that used in the planetary transit SOS module, rather than the generic computationally efficient generalised Lomb-Scargle method used for all the classes. Consequently, the EP classifications were limited to the selection of the corresponding SOS module and their scores were set to one. A binary XGBoost classifier (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was used to extract a sample of 18 383 potential candidates, without the application of a minimum probability threshold. All of these sources were processed by the planetary transit SOS module, which returned 214 objects, including known stars with planetary transits as well as new candidates. According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, at least 93 additional EP candidates may be found among the false positives of the SOLAR_LIKE class.
## Long-period variables (LPV)[\[ssec:lpv\]]{#ssec:lpv label="ssec:lpv"}
The classification of long-period variables included 2 325 775 stars, among which long secondary period variables, \(o\) Ceti (Mira) stars, (OGLE) small amplitude red giants, and semi-regular types. They were obtained from four meta-classifiers (of binary and multi-class classifiers, see Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds and selected according to the criteria defined in the LPV SOS module, some of which were more permissive for classification candidates, as indicated in the following (employing field names in the `vari_summary` and `gaia_source` tables).
1. A minimum number of clean FoV transits in the \(G\)
2. The *Gaia*
3. A minimum fraction of clean FoV transits in the \(G_{\rm RP}\)
4. A minimum amplitude assessed from the 5th to the 95th percentile of the magnitude distribution of \(G\)
5. A minimum signal-to-noise level or apparent brightness in the \(G\)
6. The number of groups of observations in the \(G\)
For further details, including an analysis of the difference between the SOS module and the classification LPV candidates, see. Faint and/or blue candidates in Figs. [\[fig:app:LPV\]](#fig:app:LPV){reference-type="ref" reference="fig:app:LPV"}b,d tend to be associated with low scores and might include spurious candidates, as their distribution in the sky suggests scanning law features (Fig. [\[fig:app:LPV\]](#fig:app:LPV){reference-type="ref" reference="fig:app:LPV"}a) and the step around \(G\) According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional LPV candidates may be found among the false positives of the following classes: RS (5761), ELL (443), YSO (234), SYST (215), CEP (73), AGN (50), RCB (42), BE\|GCAS\|SDOR\|WR (20), and MICROLENSING (3), where the number of known sources indicated in parentheses is a lower limit.
## Microlensing events[\[ssec:microlensing\]]{#ssec:microlensing label="ssec:microlensing"}
The classification of 254 candidate microlensing events was performed independently from the microlensing SOS module and the latter did not consider these classification results as input of potential candidates, as it already extracted all possible candidates with methods specific to this class. Nevertheless, the classification candidates had the advantage of fewer assumptions and requirements than the SOS counterparts, in particular regarding the time series modelling. Among the microlensing candidates, 187 are in common with the SOS module and 67 are unique to classification. Microlensing sources were selected from 10 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. The following additional criteria were applied to such candidates (employing field names in the `vari_summary` and `gaia_source` tables).
1. In order to focus on high amplitude events, the `std_dev_mag_g_fov` parameter was set to be greater than 0.06 mag, as only a handful of the known sources, (among the classified ones) were below this threshold.
2. To further support the rationale of the previous item, at the cost of completeness, the most significant candidates were selected by `outlier_median_g_fov` \(>\) 100.
3. The ratio of the mean magnitude, weighted by squared uncertainties, to the unweighted mean magnitude was required to be less than one, as expected from the bias introduced by weighting towards brighter measurements (this condition was fulfilled by almost all known events).
4. The renormalised unit weight error `ruwe` was set to be lower than 1.4, considering the distribution of `ruwe` for known microlensing events peaked at close to one.
5. The colour `median_mag_g_fov` \(-\) `median_mag_rp` was set bluer than 1.7 mag, following the distribution of known events.
Some of the criteria listed above were overridden by visual inspections of a selection of candidates. As the probability of source-lens alignment along the line-of-sight to the observer increases in high-density regions, microlensing candidates are found, as expected, prevalently towards the Galactic bulge and a minority of cases in the Galactic disc (see Fig. [\[fig:app:MICROLENSING\]](#fig:app:MICROLENSING){reference-type="ref" reference="fig:app:MICROLENSING"}a). They populate the region of negative `skewness_mag_g_fov` and low `abbe_mag_g_fov` values that are typical of bright and time-dependent outburst-like events (Fig. [\[fig:app:MICROLENSING\]](#fig:app:MICROLENSING){reference-type="ref" reference="fig:app:MICROLENSING"}f). Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, at least three additional microlensing candidates may be found among the false positives of the LPV class.
## R Coronae Borealis stars (RCB)[\[ssec:rcb\]]{#ssec:rcb label="ssec:rcb"}
The 153 stars classified as R Coronae Borealis variables were obtained from two meta-classifiers (of binary and multi-class classifiers, see Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), after filtering out low probability candidates and confirmation by visual inspection. These high-amplitude variables are characterized by sudden fading by up to several magnitudes (Fig. [\[fig:app:RCB\]](#fig:app:RCB){reference-type="ref" reference="fig:app:RCB"}e), followed by an irregular recovery (sample light curves are shown in Figs. [\[fig:app:RCB_cc\]](#fig:app:RCB_cc){reference-type="ref" reference="fig:app:RCB_cc"}c,d,r,f). They are simultaneously eruptive and pulsating, although the amplitude of the latter is an order of magnitude lower. As a consequence of the long timescale variations, these objects are also characterized by low `abbe_mag_g_fov` values (Fig. [\[fig:app:RCB\]](#fig:app:RCB){reference-type="ref" reference="fig:app:RCB"}f) and long-period variables are the main source of contamination (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}). Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional RCB candidates may be found among the false positives of the LPV (5) and RS (4) classes, where the number of known sources indicated in parentheses is a lower limit.
## RR Lyrae stars (RR)[\[ssec:rr\]]{#ssec:rr label="ssec:rr"}
The classification of RR Lyrae stars includes 297 778 variables of fundamental-mode, first-overtone, double mode (and anomalous double mode) types. The selection of candidates followed the one of the relevant SOS module. Additional candidates were obtained from four binary and multi-class classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) after the application of strict minimum probability thresholds. Among the 26 202 extra RR candidates in classification with respect to the RR Lyrae SOS module, 9239 are known RR Lyrae stars in the literature compilation of (or 8276 if counting only those flagged by their `selection` field), for which the correct period could not be recovered with the *Gaia* DR3 Other candidates were found to be eclipsing binaries (as already noted in Sect. [\[ssec:ecl\]](#ssec:ecl){reference-type="ref" reference="ssec:ecl"}), which constitute the main source of contamination for this class (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}). About 800 AGN and galaxy contaminants, as well as dubious RR candidates, are listed in tables 5 and 6 of. Faint sources, affected by extinction in the Galactic disc and bulge, and low-amplitude candidates tend to be associated with low classification scores (Figs. [\[fig:app:RR\]](#fig:app:RR){reference-type="ref" reference="fig:app:RR"}a,b,e). Other low-score candidates are found between the main sequence and the white dwarf sequence (Fig. [\[fig:app:RR\]](#fig:app:RR){reference-type="ref" reference="fig:app:RR"}d) and are located mainly in the Galactic bulge and disc (some in the Magellanic Clouds). They are characterized by lower values of the renormalised unit weight error (`ruwe`) than the other candidates: among all RR classifications with `parallax_over_error` \(>\) 5, about 80 % have `ruwe` \(<\) 1.22, while only half of the ones between the main and white-dwarf sequences fulfil this condition. The pulsating nature of the RR candidates is confirmed in Fig. [\[fig:app:RR\]](#fig:app:RR){reference-type="ref" reference="fig:app:RR"}g, until the contribution of noise overcomes the expected \(G_{\rm BP}\) According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional RR candidates may be found among the false positives of the following classes: ECL (5833), DSCT\|GDOR\|SXPHE (525), AGN (264), GALAXY (244), S (168), CEP (158), CV (41), and BCEP (3), where the number of known sources indicated in parentheses is a lower limit.
## RS Canum Venaticorum stars (RS)[\[ssec:rs\]]{#ssec:rs label="ssec:rs"}
The classification of RS Canum Venaticorum type binary systems includes 742 263 candidates. Although they are presented separately from SOLAR_LIKE stars, any confusion between these two classes does not constitute real contamination, as at least one of the components of an RS binary system is a solar-like star. Thus, as expected, 20 050 RS sources are included also in the `vari_rotation_modulation` table of the solar-like SOS module and the top RS false positive classes listed in Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"} are ROT and BY, which should not be considered as contaminants. The RS candidates were selected from 10 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. The following additional criteria were applied (employing field names in the `vari_summary` and `gaia_source` tables).
1. The renormalised unit weight error `ruwe` was set to be lower than 1.1, to exclude many candidates in a range sparsely populated by literature RS instances.
2. To better fit the bulk of known RS stars in standard deviation vs median \(G\)
3. To exclude a clump of negatively skewed candidates (by artefact) and for symmetry reasons, the following condition was applied: \| `skewness_mag_g_fov` \| \(<\) 3 (as visible in Fig. [\[fig:app:RS\]](#fig:app:RS){reference-type="ref" reference="fig:app:RS"}f).
4. Two conditions in the *Gaia*
1. `median_mag_bp` \(-\) `median_mag_g_fov` \(<\)model(\(G_{\rm BP}-G\)
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\)model(\(G_{\rm BP}-G\)
According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional RS candidates may be found among the false positives of the following classes: ECL (10 791), DSCT\|GDOR\|SXPHE (1598), YSO (1542), ELL (210), RR (165), SOLAR_LIKE (148), and CEP (36), where the number of known sources indicated in parentheses is a lower limit.
## Short-timescale objects (S)[\[ssec:s\]]{#ssec:s label="ssec:s"}
This class was trained with stars exhibiting rapid light variations that however were not well studied in the literature. It resulted in 512 005 short timescale candidates, originally targeting possible input for the short timescale SOS module. Eventually, the similar goal was pursued independently (with different data types, as the SOS module employed per-CCD photometry), leading to two rather complementary sets. This class is meant to be exploratory and considered as a sample where not well-defined short timescale sources (including those from other classes published in *Gaia* DR3 The assessment of completeness and contamination in Table [1](#tab:results){reference-type="ref" reference="tab:results"} is not truly meaningful for this class, as completeness is relative to objects that might be better resolved in *Gaia* The S candidates were obtained from seven binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. The lowest probability candidates excluded by such thresholds corresponded also to the least sampled sources, causing the appearance of regions (intersected by the Ecliptic) almost devoid of candidates, as a consequence of the *Gaia* Additional selection criteria are described in the following (employing field names in the `vari_summary` table).
1. To remove a suspicious peak of low-amplitude candidates without counterpart in the literature, a minimum standard deviation threshold was set: `std_dev_mag_g_fov` \(>\) 0.05 mag, as noticeable in Fig. [\[fig:app:S\]](#fig:app:S){reference-type="ref" reference="fig:app:S"}e.
2. The tails of colour distributions of S candidates that were overly represented in proportion to the literature were reduced:
1. `median_mag_bp` \(-\) `median_mag_rp` \(<\) 2.6 mag,
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(<\) 1.3 mag,
3. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\) 1.6 mag (Fig. [\[fig:app:S\]](#fig:app:S){reference-type="ref" reference="fig:app:S"}c).
3. To exclude a population of candidates with magnitude time series associated with extremely negative skewness and a few positively skewed outliers, only `skewness_mag_g_fov` values between \(-\)`<!-- -->`{=html}1.4 and 4 were accepted (Fig. [\[fig:app:S\]](#fig:app:S){reference-type="ref" reference="fig:app:S"}f).
4. The number sources within a radius of 3 arcsec of each S candidate (excluding the contribution of the candidate at the centre) was set to be less than 5, to exclude candidates in the most crowded environments.
5. The minimum ratio of two (unpublished) spectral shape components in the \(G_{\rm BP}\)
6. A constraint on the amount of scan-angle dependent signal, quantified by the Spearman correlation \(r_{{\rm ipd},G}\) between the \(G\)
Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional S sources from the literature are found among the following classes: ECL (24), RR (12), DSCT\|GDOR\|SXPHE (9), CV (3), RS (2), WD (2), and AGN (1), where the number of known sources indicated in parentheses is a lower limit.
## Subdwarf B-type stars (SDB)[\[ssec:sdb\]]{#ssec:sdb label="ssec:sdb"}
The classification of subdwarf B variables returned 893 candidates meant to represent stars of types V1093 Herculis and V361 Hydrae. While they occupy the expected location in the observational Hertzsprung--Russell diagram (Fig. [\[fig:app:SDB\]](#fig:app:SDB){reference-type="ref" reference="fig:app:SDB"}d), their variability might not be due to pulsation only and the presence of the latter cannot be assured when concurrent high-amplitude phenomena exist, such as:
1. the reflection effect due to irradiation of a cool companion by a hot subdwarf primary and consequent re-radiation from the illuminated side of the cool companion;
2. the tidal distortion of the cool companion, which generates ellipsoidal-like flux variations (and possible mass transfer);
3. the possible presence of spots on a rotating SDB star.
The first two items are related to binary systems with a hot subdwarf and a cool companion. They cause larger variations in the red band than in the blue band. The difference between the \(G_{\rm BP}\) A multi-periodic analysis of the SDB candidates should be performed to account for the effects of binarity and/or spots, in addition to the stellar oscillations. These candidates were selected from 10 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. Additional conditions are described as follows (employing field names in the `vari_summary` and `gaia_source` tables).
1. To highlight candidates with clear signals, the minimum level of variability was raised by requiring `std_dev_mag_g_fov` to be greater than the third quartile of the standard deviations in \(G\)
2. Sparse outliers with respect to the general colour--colour relation followed by all other candidates were removed by the condition `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) 0.33 (`median_mag_bp` \(-\) `median_mag_rp`) \(-\) 0.03 mag.
3. The renormalised unit weight error `ruwe` was restricted to values lower than 1.15, to remove a tail of SDB candidates at high `ruwe` values that were marginally represented in the literature.
4. The possible association of SDB candidates with crowded regions (mostly around the Galactic bulge) was limited by setting the number of sources within 100 arcsec from each SDB candidate (computed by excluding the contribution of the SDB source at the centre) to be less than 471.
Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, at least one additional SDB candidate may be found among the false positives of the WD class.
## Supernovae (SN)[\[ssec:sn\]]{#ssec:sn label="ssec:sn"}
The 3029 classified supernovae represent some of the most extreme types of cataclysmic variables. As anticipated in Sect. [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, SN candidates are the least sampled sources, given their transient detectability, with an average number of clean observations in the \(G\) The \(G-G_{\rm RP}\) Classification scores of the candidates tend to be high for sources with high signal-to-noise (Fig. [\[fig:app:SN\]](#fig:app:SN){reference-type="ref" reference="fig:app:SN"}e). The linear features observed in the `skewness_mag_g_fov` vs `abbe_mag_g_fov` diagram (Fig. [\[fig:app:SN\]](#fig:app:SN){reference-type="ref" reference="fig:app:SN"}f) are due to sources with the least amount of observations (typically five). The SN candidates were obtained from 10 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. Additional verification filters are described as follows (employing field names in the `vari_summary` and `gaia_source` tables).
1. Given the slow SN luminosity decay with respect to *Gaia*
2. Obvious contamination-dominated candidates at low Galactic latitudes (\(b\)) were removed by requiring \| `b` \| \(>\) 7 degrees (Fig. [\[fig:app:SN\]](#fig:app:SN){reference-type="ref" reference="fig:app:SN"}a).
3. In order to favour candidates with a clear signal, the single-band Stetson index `stetson_mag_g_fov` was set to be greater than 8.
4. To remove a small number of outliers in the \(G_{\rm BP}-G_{\rm RP}\)
5. The minimum number of clean measurements in the \(G\)
6. The number of sources within 100 arcsec from each SN candidate (computed by excluding the contribution of the SN at the centre) was set to be less than 314, to remove a high-density tail of candidates prone to contamination.
According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, at least 106 additional SN candidates may be found among the false positives of the GALAXY class.
## Solar-like stars (SOLAR_LIKE)[\[ssec:solar_like\]]{#ssec:solar_like label="ssec:solar_like"}
The classification of 1 934 844 stars with solar-like variability, such as flaring and rotating spotted stars, was obtained independently of the rotational modulation SOS module. Further insights on stellar chromospheric activity using *Gaia* The scanning law features in the sky map of candidates (Fig. [\[fig:app:SOLAR\]](#fig:app:SOLAR){reference-type="ref" reference="fig:app:SOLAR"}a) reflect the dependence of the solar-like identifications on the number of observations, which was learnt from the distribution of training sources that included *Gaia* The contamination by constant stars (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}) is not unexpected, given the low-amplitude signal (for example, see Figs. [\[fig:app:SOLAR_LIKE_cc\]](#fig:app:SOLAR_LIKE_cc){reference-type="ref" reference="fig:app:SOLAR_LIKE_cc"}c--f), and the true nature of such objects depends on the photometric precision of those sources in *Gaia* DR3 The solar-like candidates were obtained from 11 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds and two additional conditions in the *Gaia*
1. `median_mag_g_fov` \(-\) `median_mag_rp` \(<\)model(\(G-G_{\rm RP}\)
2. `median_mag_g_fov` \(-\) `median_mag_rp` \(>\)model(\(G-G_{\rm RP}\)
employing field names in the `vari_summary` table. The impact of these conditions are also apparent in other colour--colour diagrams, such as the one in Fig. [\[fig:app:SOLAR\]](#fig:app:SOLAR){reference-type="ref" reference="fig:app:SOLAR"}c. According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional SOLAR_LIKE candidates may be found among the false positives of the following classes: RS (41913), ECL (5804), YSO (1922), DSCT\|GDOR\|SXPHE (1052), RR (314), WD (3), and SDB (1), where the number of known sources indicated in parentheses is a lower limit.
## Slowly pulsating B-type stars (SPB)[\[ssec:spb\]]{#ssec:spb label="ssec:spb"}
The classified slowly pulsating B stars include 1228 sources, among which 434 contributed to the upper main-sequence oscillator SOS module. Detailed analyses on this type of objects are presented in. The classification SPB candidates were obtained from eight binary, multi-class, multi-stage, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), after the application of minimum probability thresholds and of an additional condition on the \(G_{\rm BP}-G_{\rm RP}\) Most SPB contaminants originated from classes in the ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI group (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}), as expected from the presence of known SPB stars contaminating this group of types. Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional SPB candidates may be found among the false positives of the following classes: DSCT\|GDOR\|SXPHE (17), ACV\|CP\|MCP\|ROAM\|ROAP\|SXARI (13), and RS (1), where the number of known sources indicated in parentheses is a lower limit.
## Symbiotic stars (SYST)[\[ssec:syst\]]{#ssec:syst label="ssec:syst"}
The classification of 649 symbiotic variable stars was obtained from a single binary meta-classifier (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}), after the application of a minimum probability threshold that selected the high probability component (to which almost all known symbiotic stars belonged) of a bimodal distribution. The SYST candidates are prevalently distributed in high stellar density regions, such as the Galactic bulge, disc, and Magellanic Clouds (Fig. [\[fig:app:SYST\]](#fig:app:SYST){reference-type="ref" reference="fig:app:SYST"}a). The red clump of stars in the colour--magnitude diagram in Fig. [\[fig:app:SYST\]](#fig:app:SYST){reference-type="ref" reference="fig:app:SYST"}b are mostly in the Galactic disc and bulge. Given the long timescale variations typical of this class (see Figs. [\[fig:app:SYST_cc\]](#fig:app:SYST_cc){reference-type="ref" reference="fig:app:SYST_cc"}c--f for a sample of light curves), the SYST candidates tend to have small values of the `abbe_mag_g_fov` parameter (Fig. [\[fig:app:SYST\]](#fig:app:SYST){reference-type="ref" reference="fig:app:SYST"}f) and long-period variables constitute their main source of contamination (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}). Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional SYST candidates may be found among the false positives of the LPV (6) and RS (1) classes, where the number of known sources indicated in parentheses is a lower limit.
## Variable white dwarfs (WD)[\[ssec:wd\]]{#ssec:wd label="ssec:wd"}
The classification of 910 white dwarf variables intended to target pulsating stars of types ZZ Ceti, V777 Herculis, and GW Virginis. However, training sources, in particular the ones based simply on variability and location in the observational Hertzsprung--Russell diagram, most likely included photometric variations due to binarity and spots (see Sect. [\[ssec:sdb\]](#ssec:sdb){reference-type="ref" reference="ssec:sdb"}), and not necessarily pulsation. Because of this reason, there are fewer WD candidates than the corresponding training sources (which account for less than a quarter of the selected candidates). The reflection and tidal effects in binary sytems with a WD and a cooler companion cause larger variations in \(G_{\rm RP}\) The selection of the most variable WD candidates implicitly favoured effects of binarity or spots with respect to those from stellar oscillations. Given the low intrinsic brightness of WDs, the ones detectable by *Gaia* The computation of the classification score included probabilities from one multi-class meta-classifier and from the general variability detection classifier, to increase the relevance of high-amplitude candidates. The main contaminating classes are represented by those of cataclysmic variables and post-common envelope binaries (Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}), both of which consist of systems that include a WD. The variable WD candidates were obtained from 11 binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. Additional verification filters are described in the following (employing field names in the `vari_summary` and `gaia_source` tables).
1. Since WD candidates tended to be on average less variable than instances from the literature, a higher level of variability probability than the one used for the general variability detection (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was required (Fig. [\[fig:app:WD\]](#fig:app:WD){reference-type="ref" reference="fig:app:WD"}e).
2. A significant clump of candidates at \(G_{\rm BP}-G_{\rm RP}\)
3. The renormalised unit weight error `ruwe` was set to be lower than 1.15, to remove contaminants towards the Magellanic Clouds (only one known WD lost among the classified ones).
4. The corrected \(G_{\rm BP}\)
Although not listed in the top-6 false positive types of Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional variable WD candidates may be found among the false positives of the DSCT\|GDOR\|SXPHE (4) and ECL (2) classes, where the number of known sources indicated in parentheses is a lower limit.
## Young stellar objects (YSO)[\[ssec:yso\]]{#ssec:yso label="ssec:yso"}
The classification of young stellar objects included 79 375 sources of several types (listed in item [\[itm:yso\]](#itm:yso){reference-type="ref" reference="itm:yso"} of Sect. [\[sssec:classes\]](#sssec:classes){reference-type="ref" reference="sssec:classes"}). These candidates were validated in detail by. The low completeness and high contamination (mainly by RS and SOLAR_LIKE stars; see Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}) are expected for YSOs identified in the optical wavelengths. While complementary observations in the infrared bands would help reduce the confusion between YSO and solar-like stars, the contamination rate and false positives listed in Tables [1](#tab:results){reference-type="ref" reference="tab:results"} and [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, respectively, are significantly overestimated, considering subsequent studies on the RS and BY classifications of, as reported in. The YSO candidates were selected from three binary and multi-class classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. Additional verification filters are described as follows (employing field names in the `vari_summary` and `gaia_source` tables), often with a common criterion that trimmed one or both ends of a distribution, when a high fraction of unknown YSO candidates was removed at the cost of a low fraction of known YSOs (among the classified sources).
1. The minimum variability probability of the general variability detection (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) was set to be greater than the 5th percentile of known YSOs (Fig. [\[fig:app:YSO\]](#fig:app:YSO){reference-type="ref" reference="fig:app:YSO"}e).
2. The \(G\)
3. The `std_dev_over_rms_err_mag_g_fov` parameter was set to be greater than the 5th percentile of known YSOs.
4. The single-band `stetson_mag_g_fov` index was set to be greater than the 5th percentile of known YSOs.
5. The proper motion components in the right ascension and declination directions (`pmra` and `pmdec`, respectively), were restricted between the 1st and the 99th percentiles of known YSOs.
6. The `parallax_over_error` ratio was set to be greater than 3, for a minimal significance of parallax and also for a selection of sources that were not too distant for YSOs to be observable.
7. Two conditions in \(G_{\rm BP}-G\)
1. `median_mag_bp` \(-\) `median_mag_g_fov` \(<\)model(\(G_{\rm BP}-G\)
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\)model(\(G_{\rm BP}-G\)
According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional YSO candidates may be found among the false positives of the following classes: AGN (328), CV (20), BE\|GCAS\|SDOR\|WR (9), and RCB (5), where the number of known sources indicated in parentheses is a lower limit.
## Galaxies (GALAXY, in `galaxy_candidates`)[\[ssec:galaxy\]]{#ssec:galaxy label="ssec:galaxy"}
The classification of 2 451 364 galaxies was made possible by their apparent variability in the *Gaia* As mentioned in Sect. [\[sec:data\]](#sec:data){reference-type="ref" reference="sec:data"}, galaxies can be affected by spurious signals peculiar to *Gaia* The main role of galaxies for the classification of variable objects was to reduce the impact of artificial variations on the identification of candidates of genuine variability types, as already noticed in *Gaia* Unlike stars, galaxies occupy the red end in \(G-G_{\rm RP}\) As expected, only a few training sources or galaxy candidates fulfil the condition of `parallax_over_error` \(>\) 5 required for observational Hertzsprung--Russell diagrams (Figs. [\[fig:app:GALAXY_trn\]](#fig:app:GALAXY_trn){reference-type="ref" reference="fig:app:GALAXY_trn"}d and [\[fig:app:GALAXY_ALL\]](#fig:app:GALAXY_ALL){reference-type="ref" reference="fig:app:GALAXY_ALL"}d). Galaxy candidates were obtained from three binary, multi-class, and meta-classifiers (Sect. [\[ssec:classifier\]](#ssec:classifier){reference-type="ref" reference="ssec:classifier"}) with minimum probability thresholds. They were further filtered by the following conditions (employing field names in the `vari_summary` and `gaia_source` tables).
1. Removal of extremely blue and red outliers:
1. `median_mag_bp` \(-\) `median_mag_rp` in the range from 0 to 3 mag,
2. `median_mag_bp` \(-\) `median_mag_g_fov` \(>\) \(-\) 4 mag,
3. `median_mag_g_fov` \(-\) `median_mag_rp` in the range from 0.5 to 5.5 mag.
2. The distribution of the \(G_{\rm BP}\)
3. With respect to point sources, extended objects are often associated with higher positional uncertainty (further amplified by the causes of the spurious photometric variations) and the condition `astrometric_excess_noise` \(>\) 7 mas matched the distribution of known galaxies.
4. The number of sources within 100 arcsec from each GALAXY candidate (computed by excluding the contribution of the galaxy at the centre) was set to be less than 314, affecting mostly candidates around the Galactic plane and behind the Magellanic Clouds (Fig. [\[fig:app:GALAXY_ALL\]](#fig:app:GALAXY_ALL){reference-type="ref" reference="fig:app:GALAXY_ALL"}a).
5. It was further required that `num_selected_rp` \(>\) 5.
6. Two (unpublished) spectral shape components in the \(G_{\rm BP}\)
According to Table [2](#tab:results_details){reference-type="ref" reference="tab:results_details"}, additional GALAXY candidates may be found among the false positives of the following classes: LPV (1364), S (244), AGN (48), and SN (2), where the number of known sources indicated in parentheses is a lower limit. In addition to those of galaxy contaminants in other classes, as mentioned in the beginning of Sect. [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, only the light curves of the sources included in the *Gaia* A comparison of the sources from all of the classes identified by variability that overlap with those from galaxy modules published by other *Gaia* An example of a query to extract our GALAXY candidates from the `galaxy_candidates` table is presented in Appendix [\[app:queries\]](#app:queries){reference-type="ref" reference="app:queries"}.
# Conclusions[\[sec:conclusions\]]{#sec:conclusions label="sec:conclusions"}
The *Gaia* DR3 This combination of number of sources and classes made it one of the largest and most uniformly constructed variable source catalogues in the literature. The cross-match of *Gaia* A multi-classifier approach made it possible to obtain suitable models for a large variety of variability classes, among which several types of pulsating stars, eclipsing binaries, ellipsoidal variables, spotted stars, eruptive and cataclysmic phenomena, stochastic variations of galactic nuclei, microlensing events, and planetary transits. Almost half of the genuine variable sources (4.7 million) and several classes are available uniquely as classification results (in the `vari_classifier_result` table), while the other variable sources are included (also) among the SOS module results. Galaxies were detected by an artificial signal of *Gaia* In *Gaia*
[^1]: <https://lambda.gsfc.nasa.gov/toolbox/pixelcoords.html> | {'timestamp': '2022-12-01T02:21:15', 'yymm': '2211', 'arxiv_id': '2211.17238', 'language': 'en', 'url': 'https://arxiv.org/abs/2211.17238'} |
Subsets and Splits