]>
gitweb.michael.orlitzky.com - sage.d.git/blob - mjo/cone/symmetric_psd.py
2 The positive semidefinite cone `$S^{n}_{+}$` is the cone consisting of
3 all symmetric positive-semidefinite matrices (as a subset of
4 `$\mathbb{R}^{n \times n}$`
9 def is_symmetric_psd(A
):
11 Determine whether or not the matrix ``A`` is symmetric
12 positive-semidefinite.
16 - ``A`` - The matrix in question
20 Either ``True`` if ``A`` is symmetric positive-semidefinite, or
25 sage: from mjo.cone.symmetric_psd import is_symmetric_psd
29 Every completely positive matrix is symmetric
30 positive-semidefinite::
32 sage: v = vector(map(abs, random_vector(ZZ, 10)))
33 sage: A = v.column() * v.row()
34 sage: is_symmetric_psd(A)
37 The following matrix is symmetric but not positive semidefinite::
39 sage: A = matrix(ZZ, [[1, 2], [2, 1]])
40 sage: is_symmetric_psd(A)
43 This matrix isn't even symmetric::
45 sage: A = matrix(ZZ, [[1, 2], [3, 4]])
46 sage: is_symmetric_psd(A)
49 The trivial matrix in a trivial space is trivially symmetric and
50 positive-semidefinite::
52 sage: A = matrix(QQ, 0,0)
53 sage: is_symmetric_psd(A)
58 if A
.base_ring() == SR
:
59 msg
= 'The matrix ``A`` cannot be symbolic.'
60 raise ValueError.new(msg
)
62 # First make sure that ``A`` is symmetric.
63 if not A
.is_symmetric():
66 # If ``A`` is symmetric, we only need to check that it is positive
67 # semidefinite. For that we can consult its minimum eigenvalue,
68 # which should be zero or greater. Since ``A`` is symmetric, its
69 # eigenvalues are guaranteed to be real.
71 # A is trivial... so trivially positive-semudefinite.
74 return min(A
.eigenvalues()) >= 0
77 def unit_eigenvectors(A
):
79 Return the unit eigenvectors of a symmetric positive-definite matrix.
83 - ``A`` -- The matrix whose unit eigenvectors we want to compute.
87 A list of (eigenvalue, eigenvector) pairs where each eigenvector is
88 associated with its paired eigenvalue of ``A`` and has norm `1`. If
89 the base ring of ``A`` is not algebraically closed, then returned
90 eigenvectors may (necessarily) be over its algebraic closure and not
91 the base ring of ``A`` itself.
95 sage: from mjo.cone.symmetric_psd import unit_eigenvectors
99 sage: A = matrix(QQ, [[0, 2, 3], [2, 0, 0], [3, 0, 0]])
100 sage: unit_evs = list(unit_eigenvectors(A))
101 sage: bool(unit_evs[0][1].norm() == 1)
103 sage: bool(unit_evs[1][1].norm() == 1)
105 sage: bool(unit_evs[2][1].norm() == 1)
109 return ( (val
,vec
.normalized())
110 for (val
,vecs
,multiplicity
) in A
.eigenvectors_right()
117 Factor a symmetric positive-semidefinite matrix ``A`` into
122 - ``A`` - The matrix to factor. The base ring of ``A`` must be either
123 exact or the symbolic ring (to compute eigenvalues), and it
124 must be a field so that we can take its algebraic closure
125 (necessary to e.g. take square roots).
129 A matrix ``X`` such that `A = XX^{T}`. The base field of ``X`` will
130 be the algebraic closure of the base field of ``A``.
134 Since ``A`` is symmetric and positive-semidefinite, we can
135 diagonalize it by some matrix `$Q$` whose columns are orthogonal
136 eigenvectors of ``A``. Then,
140 From this representation we can take the square root of `$D$`
141 (since all eigenvalues of ``A`` are nonnegative). If we then let
142 `$X = Q*sqrt(D)*Q^{T}$`, we have,
144 `$XX^{T} = Q*sqrt(D)*Q^{T}Q*sqrt(D)*Q^{T} = Q*D*Q^{T} = A$`
148 In principle, this is the algorithm used, although we ignore the
149 eigenvectors corresponding to the eigenvalue zero. Thus if `$rank(A)
150 = k$`, the matrix `$Q$` will have dimention `$n \times k$`, and
151 `$D$` will have dimension `$k \times k$`. In the end everything
156 sage: from mjo.cone.symmetric_psd import factor_psd
160 Create a symmetric positive-semidefinite matrix over the symbolic
163 sage: A = matrix(SR, [[0, 2, 3], [2, 0, 0], [3, 0, 0]])
164 sage: X = factor_psd(A)
165 sage: A2 = (X*X.transpose()).simplify_full()
169 Attempt to factor the same matrix over ``RR`` which won't work
170 because ``RR`` isn't exact::
172 sage: A = matrix(RR, [[0, 2, 3], [2, 0, 0], [3, 0, 0]])
174 Traceback (most recent call last):
176 ValueError: The base ring of ``A`` must be either exact or symbolic.
178 Attempt to factor the same matrix over ``ZZ`` which won't work
179 because ``ZZ`` isn't a field::
181 sage: A = matrix(ZZ, [[0, 2, 3], [2, 0, 0], [3, 0, 0]])
183 Traceback (most recent call last):
185 ValueError: The base ring of ``A`` must be a field.
189 if not A
.base_ring().is_exact() and not A
.base_ring() is SR
:
190 msg
= 'The base ring of ``A`` must be either exact or symbolic.'
191 raise ValueError(msg
)
193 if not A
.base_ring().is_field():
194 raise ValueError('The base ring of ``A`` must be a field.')
196 if not A
.base_ring() is SR
:
197 # Change the base field of ``A`` so that we are sure we can take
198 # roots. The symbolic ring has no algebraic_closure method.
199 A
= A
.change_ring(A
.base_ring().algebraic_closure())
202 # Get the eigenvectors, and filter out the ones that correspond to
203 # the eigenvalue zero.
204 all_evs
= unit_eigenvectors(A
)
205 evs
= [ (val
,vec
) for (val
,vec
) in all_evs
if not val
== 0 ]
207 d
= ( val
.sqrt() for (val
,vec
) in evs
)
208 root_D
= diagonal_matrix(d
).change_ring(A
.base_ring())
210 Q
= matrix(A
.base_ring(), ( vec
for (val
,vec
) in evs
)).transpose()
212 return Q
*root_D
*Q
.transpose()
215 def random_symmetric_psd(V
, accept_zero
=True, rank
=None):
217 Generate a random symmetric positive-semidefinite matrix over the
218 vector space ``V``. That is, the returned matrix will be a linear
219 transformation on ``V``, with the same base ring as ``V``.
221 We take a very loose interpretation of "random," here. Otherwise we
222 would never (for example) choose a matrix on the boundary of the
223 cone (with a zero eigenvalue).
227 - ``V`` - The vector space on which the returned matrix will act.
229 - ``accept_zero`` - Do you want to accept the zero matrix (which
230 is symmetric PSD? Defaults to ``True``.
232 - ``rank`` - Require the returned matrix to have the given rank
237 A random symmetric positive semidefinite matrix, i.e. a linear
238 transformation from ``V`` to itself.
242 The matrix is constructed from some number of spectral projectors,
243 which in turn are created at "random" from the underlying vector
246 If no particular ``rank`` is desired, we choose the number of
247 projectors at random. Otherwise, we keep adding new projectors until
248 the desired rank is achieved.
250 Finally, before returning, we check if the matrix is zero. If
251 ``accept_zero`` is ``False``, we restart the process from the
256 sage: from mjo.cone.symmetric_psd import (is_symmetric_psd,
257 ....: random_symmetric_psd)
261 Well, it doesn't crash at least::
263 sage: V = VectorSpace(QQ, 2)
264 sage: A = random_symmetric_psd(V)
265 sage: A.matrix_space()
266 Full MatrixSpace of 2 by 2 dense matrices over Rational Field
267 sage: is_symmetric_psd(A)
270 A matrix with the desired rank is returned::
272 sage: V = VectorSpace(QQ, 5)
273 sage: A = random_symmetric_psd(V,False,1)
276 sage: A = random_symmetric_psd(V,False,2)
279 sage: A = random_symmetric_psd(V,False,3)
282 sage: A = random_symmetric_psd(V,False,4)
285 sage: A = random_symmetric_psd(V,False,5)
289 If the user asks for a rank that's too high, we fail::
291 sage: V = VectorSpace(QQ, 2)
292 sage: A = random_symmetric_psd(V,False,3)
293 Traceback (most recent call last):
295 ValueError: The ``rank`` must be between 0 and the dimension of ``V``.
299 # We construct the matrix from its spectral projectors. Since
300 # there can be at most ``n`` of them, where ``n`` is the dimension
301 # of our vector space, we want to choose a random integer between
302 # ``0`` and ``n`` and then construct that many random elements of
308 # Choose one randomly
309 rank_A
= ZZ
.random_element(n
+1)
310 elif (rank
< 0) or (rank
> n
):
311 # The rank of ``A`` can be at most ``n``.
312 msg
= 'The ``rank`` must be between 0 and the dimension of ``V``.'
313 raise ValueError(msg
)
315 # Use the one the user gave us.
318 if n
== 0 and not accept_zero
:
319 # We're gonna loop forever trying to satisfy this...
320 raise ValueError('You must have accept_zero=True when V is trivial')
322 # Loop until we find a suitable "A" that will then be returned.
324 # Begin with the zero matrix, and add projectors to it if we
326 A
= matrix
.zero(V
.base_ring(), n
, n
)
328 # Careful, begin at idx=1 so that we only generate a projector
329 # when rank_A is greater than zero.
330 while A
.rank() < rank_A
:
331 v
= V
.random_element()
332 A
+= v
.column()*v
.row()
334 if accept_zero
or not A
.is_zero():
335 # We either don't care what ``A`` is, or it's non-zero, so