File size: 15,532 Bytes
6a86ad5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
# Algorithms for computing the reduced row echelon form of a matrix.
#
# We need to choose carefully which algorithms to use depending on the domain,
# shape, and sparsity of the matrix as well as things like the bit count in the
# case of ZZ or QQ. This is important because the algorithms have different
# performance characteristics in the extremes of dense vs sparse.
#
# In all cases we use the sparse implementations but we need to choose between
# Gauss-Jordan elimination with division and fraction-free Gauss-Jordan
# elimination. For very sparse matrices over ZZ with low bit counts it is
# asymptotically faster to use Gauss-Jordan elimination with division. For
# dense matrices with high bit counts it is asymptotically faster to use
# fraction-free Gauss-Jordan.
#
# The most important thing is to get the extreme cases right because it can
# make a big difference. In between the extremes though we have to make a
# choice and here we use empirically determined thresholds based on timings
# with random sparse matrices.
#
# In the case of QQ we have to consider the denominators as well. If the
# denominators are small then it is faster to clear them and use fraction-free
# Gauss-Jordan over ZZ. If the denominators are large then it is faster to use
# Gauss-Jordan elimination with division over QQ.
#
# Timings for the various algorithms can be found at
#
#   https://github.com/sympy/sympy/issues/25410
#   https://github.com/sympy/sympy/pull/25443

from sympy.polys.domains import ZZ

from sympy.polys.matrices.sdm import SDM, sdm_irref, sdm_rref_den
from sympy.polys.matrices.ddm import DDM
from sympy.polys.matrices.dense import ddm_irref, ddm_irref_den


def _dm_rref(M, *, method='auto'):
    """
    Compute the reduced row echelon form of a ``DomainMatrix``.

    This function is the implementation of :meth:`DomainMatrix.rref`.

    Chooses the best algorithm depending on the domain, shape, and sparsity of
    the matrix as well as things like the bit count in the case of :ref:`ZZ` or
    :ref:`QQ`. The result is returned over the field associated with the domain
    of the Matrix.

    See Also
    ========

    sympy.polys.matrices.domainmatrix.DomainMatrix.rref
        The ``DomainMatrix`` method that calls this function.
    sympy.polys.matrices.rref._dm_rref_den
        Alternative function for computing RREF with denominator.
    """
    method, use_fmt = _dm_rref_choose_method(M, method, denominator=False)

    M, old_fmt = _dm_to_fmt(M, use_fmt)

    if method == 'GJ':
        # Use Gauss-Jordan with division over the associated field.
        Mf = _to_field(M)
        M_rref, pivots = _dm_rref_GJ(Mf)

    elif method == 'FF':
        # Use fraction-free GJ over the current domain.
        M_rref_f, den, pivots = _dm_rref_den_FF(M)
        M_rref = _to_field(M_rref_f) / den

    elif method == 'CD':
        # Clear denominators and use fraction-free GJ in the associated ring.
        _, Mr = M.clear_denoms_rowwise(convert=True)
        M_rref_f, den, pivots = _dm_rref_den_FF(Mr)
        M_rref = _to_field(M_rref_f) / den

    else:
        raise ValueError(f"Unknown method for rref: {method}")

    M_rref, _ = _dm_to_fmt(M_rref, old_fmt)

    # Invariants:
    #   - M_rref is in the same format (sparse or dense) as the input matrix.
    #   - M_rref is in the associated field domain and any denominator was
    #     divided in (so is implicitly 1 now).

    return M_rref, pivots


def _dm_rref_den(M, *, keep_domain=True, method='auto'):
    """
    Compute the reduced row echelon form of a ``DomainMatrix`` with denominator.

    This function is the implementation of :meth:`DomainMatrix.rref_den`.

    Chooses the best algorithm depending on the domain, shape, and sparsity of
    the matrix as well as things like the bit count in the case of :ref:`ZZ` or
    :ref:`QQ`. The result is returned over the same domain as the input matrix
    unless ``keep_domain=False`` in which case the result might be over an
    associated ring or field domain.

    See Also
    ========

    sympy.polys.matrices.domainmatrix.DomainMatrix.rref_den
        The ``DomainMatrix`` method that calls this function.
    sympy.polys.matrices.rref._dm_rref
        Alternative function for computing RREF without denominator.
    """
    method, use_fmt = _dm_rref_choose_method(M, method, denominator=True)

    M, old_fmt = _dm_to_fmt(M, use_fmt)

    if method == 'FF':
        # Use fraction-free GJ over the current domain.
        M_rref, den, pivots = _dm_rref_den_FF(M)

    elif method == 'GJ':
        # Use Gauss-Jordan with division over the associated field.
        M_rref_f, pivots = _dm_rref_GJ(_to_field(M))

        # Convert back to the ring?
        if keep_domain and M_rref_f.domain != M.domain:
            _, M_rref = M_rref_f.clear_denoms(convert=True)

            if pivots:
                den = M_rref[0, pivots[0]].element
            else:
                den = M_rref.domain.one
        else:
            # Possibly an associated field
            M_rref = M_rref_f
            den = M_rref.domain.one

    elif method == 'CD':
        # Clear denominators and use fraction-free GJ in the associated ring.
        _, Mr = M.clear_denoms_rowwise(convert=True)

        M_rref_r, den, pivots = _dm_rref_den_FF(Mr)

        if keep_domain and M_rref_r.domain != M.domain:
            # Convert back to the field
            M_rref = _to_field(M_rref_r) / den
            den = M.domain.one
        else:
            # Possibly an associated ring
            M_rref = M_rref_r

            if pivots:
                den = M_rref[0, pivots[0]].element
            else:
                den = M_rref.domain.one
    else:
        raise ValueError(f"Unknown method for rref: {method}")

    M_rref, _ = _dm_to_fmt(M_rref, old_fmt)

    # Invariants:
    #   - M_rref is in the same format (sparse or dense) as the input matrix.
    #   - If keep_domain=True then M_rref and den are in the same domain as the
    #     input matrix
    #   - If keep_domain=False then M_rref might be in an associated ring or
    #     field domain but den is always in the same domain as M_rref.

    return M_rref, den, pivots


def _dm_to_fmt(M, fmt):
    """Convert a matrix to the given format and return the old format."""
    old_fmt = M.rep.fmt
    if old_fmt == fmt:
        pass
    elif fmt == 'dense':
        M = M.to_dense()
    elif fmt == 'sparse':
        M = M.to_sparse()
    else:
        raise ValueError(f'Unknown format: {fmt}') # pragma: no cover
    return M, old_fmt


# These are the four basic implementations that we want to choose between:


def _dm_rref_GJ(M):
    """Compute RREF using Gauss-Jordan elimination with division."""
    if M.rep.fmt == 'sparse':
        return _dm_rref_GJ_sparse(M)
    else:
        return _dm_rref_GJ_dense(M)


def _dm_rref_den_FF(M):
    """Compute RREF using fraction-free Gauss-Jordan elimination."""
    if M.rep.fmt == 'sparse':
        return _dm_rref_den_FF_sparse(M)
    else:
        return _dm_rref_den_FF_dense(M)


def _dm_rref_GJ_sparse(M):
    """Compute RREF using sparse Gauss-Jordan elimination with division."""
    M_rref_d, pivots, _ = sdm_irref(M.rep)
    M_rref_sdm = SDM(M_rref_d, M.shape, M.domain)
    pivots = tuple(pivots)
    return M.from_rep(M_rref_sdm), pivots


def _dm_rref_GJ_dense(M):
    """Compute RREF using dense Gauss-Jordan elimination with division."""
    partial_pivot = M.domain.is_RR or M.domain.is_CC
    ddm = M.rep.to_ddm().copy()
    pivots = ddm_irref(ddm, _partial_pivot=partial_pivot)
    M_rref_ddm = DDM(ddm, M.shape, M.domain)
    pivots = tuple(pivots)
    return M.from_rep(M_rref_ddm.to_dfm_or_ddm()), pivots


def _dm_rref_den_FF_sparse(M):
    """Compute RREF using sparse fraction-free Gauss-Jordan elimination."""
    M_rref_d, den, pivots = sdm_rref_den(M.rep, M.domain)
    M_rref_sdm = SDM(M_rref_d, M.shape, M.domain)
    pivots = tuple(pivots)
    return M.from_rep(M_rref_sdm), den, pivots


def _dm_rref_den_FF_dense(M):
    """Compute RREF using sparse fraction-free Gauss-Jordan elimination."""
    ddm = M.rep.to_ddm().copy()
    den, pivots = ddm_irref_den(ddm, M.domain)
    M_rref_ddm = DDM(ddm, M.shape, M.domain)
    pivots = tuple(pivots)
    return M.from_rep(M_rref_ddm.to_dfm_or_ddm()), den, pivots


def _dm_rref_choose_method(M, method, *, denominator=False):
    """Choose the fastest method for computing RREF for M."""

    if method != 'auto':
        if method.endswith('_dense'):
            method = method[:-len('_dense')]
            use_fmt = 'dense'
        else:
            use_fmt = 'sparse'

    else:
        # The sparse implementations are always faster
        use_fmt = 'sparse'

        K = M.domain

        if K.is_ZZ:
            method = _dm_rref_choose_method_ZZ(M, denominator=denominator)
        elif K.is_QQ:
            method = _dm_rref_choose_method_QQ(M, denominator=denominator)
        elif K.is_RR or K.is_CC:
            # TODO: Add partial pivot support to the sparse implementations.
            method = 'GJ'
            use_fmt = 'dense'
        elif K.is_EX and M.rep.fmt == 'dense' and not denominator:
            # Do not switch to the sparse implementation for EX because the
            # domain does not have proper canonicalization and the sparse
            # implementation gives equivalent but non-identical results over EX
            # from performing arithmetic in a different order. Specifically
            # test_issue_23718 ends up getting a more complicated expression
            # when using the sparse implementation. Probably the best fix for
            # this is something else but for now we stick with the dense
            # implementation for EX if the matrix is already dense.
            method = 'GJ'
            use_fmt = 'dense'
        else:
            # This is definitely suboptimal. More work is needed to determine
            # the best method for computing RREF over different domains.
            if denominator:
                method = 'FF'
            else:
                method = 'GJ'

    return method, use_fmt


def _dm_rref_choose_method_QQ(M, *, denominator=False):
    """Choose the fastest method for computing RREF over QQ."""
    # The same sorts of considerations apply here as in the case of ZZ. Here
    # though a new more significant consideration is what sort of denominators
    # we have and what to do with them so we focus on that.

    # First compute the density. This is the average number of non-zero entries
    # per row but only counting rows that have at least one non-zero entry
    # since RREF can ignore fully zero rows.
    density, _, ncols = _dm_row_density(M)

    # For sparse matrices use Gauss-Jordan elimination over QQ regardless.
    if density < min(5, ncols/2):
        return 'GJ'

    # Compare the bit-length of the lcm of the denominators to the bit length
    # of the numerators.
    #
    # The threshold here is empirical: we prefer rref over QQ if clearing
    # denominators would result in a numerator matrix having 5x the bit size of
    # the current numerators.
    numers, denoms = _dm_QQ_numers_denoms(M)
    numer_bits = max([n.bit_length() for n in numers], default=1)

    denom_lcm = ZZ.one
    for d in denoms:
        denom_lcm = ZZ.lcm(denom_lcm, d)
        if denom_lcm.bit_length() > 5*numer_bits:
            return 'GJ'

    # If we get here then the matrix is dense and the lcm of the denominators
    # is not too large compared to the numerators. For particularly small
    # denominators it is fastest just to clear them and use fraction-free
    # Gauss-Jordan over ZZ. With very small denominators this is a little
    # faster than using rref_den over QQ but there is an intermediate regime
    # where rref_den over QQ is significantly faster. The small denominator
    # case is probably very common because small fractions like 1/2 or 1/3 are
    # often seen in user inputs.

    if denom_lcm.bit_length() < 50:
        return 'CD'
    else:
        return 'FF'


def _dm_rref_choose_method_ZZ(M, *, denominator=False):
    """Choose the fastest method for computing RREF over ZZ."""
    # In the extreme of very sparse matrices and low bit counts it is faster to
    # use Gauss-Jordan elimination over QQ rather than fraction-free
    # Gauss-Jordan over ZZ. In the opposite extreme of dense matrices and high
    # bit counts it is faster to use fraction-free Gauss-Jordan over ZZ. These
    # two extreme cases need to be handled differently because they lead to
    # different asymptotic complexities. In between these two extremes we need
    # a threshold for deciding which method to use. This threshold is
    # determined empirically by timing the two methods with random matrices.

    # The disadvantage of using empirical timings is that future optimisations
    # might change the relative speeds so this can easily become out of date.
    # The main thing is to get the asymptotic complexity right for the extreme
    # cases though so the precise value of the threshold is hopefully not too
    # important.

    # Empirically determined parameter.
    PARAM = 10000

    # First compute the density. This is the average number of non-zero entries
    # per row but only counting rows that have at least one non-zero entry
    # since RREF can ignore fully zero rows.
    density, nrows_nz, ncols = _dm_row_density(M)

    # For small matrices use QQ if more than half the entries are zero.
    if nrows_nz < 10:
        if density < ncols/2:
            return 'GJ'
        else:
            return 'FF'

    # These are just shortcuts for the formula below.
    if density < 5:
        return 'GJ'
    elif density > 5 + PARAM/nrows_nz:
        return 'FF'  # pragma: no cover

    # Maximum bitsize of any entry.
    elements = _dm_elements(M)
    bits = max([e.bit_length() for e in elements], default=1)

    # Wideness parameter. This is 1 for square or tall matrices but >1 for wide
    # matrices.
    wideness = max(1, 2/3*ncols/nrows_nz)

    max_density = (5 + PARAM/(nrows_nz*bits**2)) * wideness

    if density < max_density:
        return 'GJ'
    else:
        return 'FF'


def _dm_row_density(M):
    """Density measure for sparse matrices.

    Defines the "density", ``d`` as the average number of non-zero entries per
    row except ignoring rows that are fully zero. RREF can ignore fully zero
    rows so they are excluded. By definition ``d >= 1`` except that we define
    ``d = 0`` for the zero matrix.

    Returns ``(density, nrows_nz, ncols)`` where ``nrows_nz`` counts the number
    of nonzero rows and ``ncols`` is the number of columns.
    """
    # Uses the SDM dict-of-dicts representation.
    ncols = M.shape[1]
    rows_nz = M.rep.to_sdm().values()
    if not rows_nz:
        return 0, 0, ncols
    else:
        nrows_nz = len(rows_nz)
        density = sum(map(len, rows_nz)) / nrows_nz
        return density, nrows_nz, ncols


def _dm_elements(M):
    """Return nonzero elements of a DomainMatrix."""
    elements, _ = M.to_flat_nz()
    return elements


def _dm_QQ_numers_denoms(Mq):
    """Returns the numerators and denominators of a DomainMatrix over QQ."""
    elements = _dm_elements(Mq)
    numers = [e.numerator for e in elements]
    denoms = [e.denominator for e in elements]
    return numers, denoms


def _to_field(M):
    """Convert a DomainMatrix to a field if possible."""
    K = M.domain
    if K.has_assoc_Field:
        return M.to_field()
    else:
        return M