X-Git-Url: http://gitweb.michael.orlitzky.com/?a=blobdiff_plain;f=mjo%2Fcone%2Fcone.py;h=f3543a147ad8da3c5000015f3c53e837781180e5;hb=fbaecc56ec029d6f813d76e26bd8891a41416bf0;hp=3a1e190cb2ebe41f57810f04726f8123294c55cd;hpb=db71f895bcdf4550bc4212d0e2fae41c8de22d6a;p=sage.d.git diff --git a/mjo/cone/cone.py b/mjo/cone/cone.py index 3a1e190..f3543a1 100644 --- a/mjo/cone/cone.py +++ b/mjo/cone/cone.py @@ -8,77 +8,235 @@ addsitedir(abspath('../../')) from sage.all import * -def project_span(K, K2 = None): +def _basically_the_same(K1, K2): r""" - Return a "copy" of ``K`` embeded in a lower-dimensional space. + Test whether or not ``K1`` and ``K2`` are "basically the same." - By default, we will project ``K`` into the subspace spanned by its - rays. However, if ``K2`` is not ``None``, we will project into the - space spanned by the rays of ``K2`` instead. + This is a hack to get around the fact that it's difficult to tell + when two cones are linearly isomorphic. We have a proposition that + equates two cones, but represented over `\mathbb{Q}`, they are + merely linearly isomorphic (not equal). So rather than test for + equality, we test a list of properties that should be preserved + under an invertible linear transformation. - EXAMPLES:: + OUTPUT: + + ``True`` if ``K1`` and ``K2`` are basically the same, and ``False`` + otherwise. + + EXAMPLES: + + Any proper cone with three generators in `\mathbb{R}^{3}` is + basically the same as the nonnegative orthant:: + + sage: K1 = Cone([(1,0,0), (0,1,0), (0,0,1)]) + sage: K2 = Cone([(1,2,3), (3, 18, 4), (66, 51, 0)]) + sage: _basically_the_same(K1, K2) + True + + Negating a cone gives you another cone that is basically the same:: + + sage: K = Cone([(0,2,-5), (-6, 2, 4), (0, 51, 0)]) + sage: _basically_the_same(K, -K) + True + + TESTS: + + Any cone is basically the same as itself:: + + sage: K = random_cone(max_ambient_dim = 8) + sage: _basically_the_same(K, K) + True + + After applying an invertible matrix to the rows of a cone, the + result should be basically the same as the cone we started with:: + + sage: K1 = random_cone(max_ambient_dim = 8) + sage: A = random_matrix(QQ, K1.lattice_dim(), algorithm='unimodular') + sage: K2 = Cone( [ A*r for r in K1.rays() ], lattice=K1.lattice()) + sage: _basically_the_same(K1, K2) + True + + """ + if K1.lattice_dim() != K2.lattice_dim(): + return False - sage: K = Cone([(1,0,0), (0,1,0)]) - sage: project_span(K) - 2-d cone in 2-d lattice N - sage: project_span(K).rays() - N(1, 0), - N(0, 1) - in 2-d lattice N + if K1.nrays() != K2.nrays(): + return False - sage: K = Cone([(1,0,0), (0,1,0)]) - sage: K2 = Cone([(0,1)]) - sage: project_span(K, K2).rays() + if K1.dim() != K2.dim(): + return False + + if K1.lineality() != K2.lineality(): + return False + + if K1.is_solid() != K2.is_solid(): + return False + + if K1.is_strictly_convex() != K2.is_strictly_convex(): + return False + + if len(LL(K1)) != len(LL(K2)): + return False + + C_of_K1 = discrete_complementarity_set(K1) + C_of_K2 = discrete_complementarity_set(K2) + if len(C_of_K1) != len(C_of_K2): + return False + + if len(K1.facets()) != len(K2.facets()): + return False + + return True + + + +def _restrict_to_space(K, W): + r""" + Restrict this cone a subspace of its ambient space. + + INPUT: + + - ``W`` -- The subspace into which this cone will be restricted. + + OUTPUT: + + A new cone in a sublattice corresponding to ``W``. + + EXAMPLES: + + When this cone is solid, restricting it into its own span should do + nothing:: + + sage: K = Cone([(1,)]) + sage: _restrict_to_space(K, K.span()) == K + True + + A single ray restricted into its own span gives the same output + regardless of the ambient space:: + + sage: K2 = Cone([(1,0)]) + sage: K2_S = _restrict_to_space(K2, K2.span()).rays() + sage: K2_S + N(1) + in 1-d lattice N + sage: K3 = Cone([(1,0,0)]) + sage: K3_S = _restrict_to_space(K3, K3.span()).rays() + sage: K3_S N(1) in 1-d lattice N + sage: K2_S == K3_S + True - """ - # Allow us to use a second cone to generate the subspace into - # which we're "projecting." - if K2 is None: - K2 = K + TESTS: + + The projected cone should always be solid:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: _restrict_to_space(K, K.span()).is_solid() + True + + And the resulting cone should live in a space having the same + dimension as the space we restricted it to:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: K_P = _restrict_to_space(K, K.dual().span()) + sage: K_P.lattice_dim() == K.dual().dim() + True + + This function should not affect the dimension of a cone:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: K.dim() == _restrict_to_space(K,K.span()).dim() + True + + Nor should it affect the lineality of a cone:: - # Use these to generate the new cone. - cs1 = K.rays().matrix().columns() + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: K.lineality() == _restrict_to_space(K, K.span()).lineality() + True - # And use these to figure out which indices to drop. - cs2 = K2.rays().matrix().columns() + No matter which space we restrict to, the lineality should not + increase:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: S = K.span(); P = K.dual().span() + sage: K.lineality() >= _restrict_to_space(K,S).lineality() + True + sage: K.lineality() >= _restrict_to_space(K,P).lineality() + True - perp_idxs = [] + If we do this according to our paper, then the result is proper:: - for idx in range(0, len(cs2)): - if cs2[idx].is_zero(): - perp_idxs.append(idx) + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim = 8) + sage: K_S = _restrict_to_space(K, K.span()) + sage: K_SP = _restrict_to_space(K_S.dual(), K_S.dual().span()).dual() + sage: K_SP.is_proper() + True + sage: K_SP = _restrict_to_space(K_S, K_S.dual().span()) + sage: K_SP.is_proper() + True - solid_cols = [ cs1[idx] for idx in range(0,len(cs1)) - if not idx in perp_idxs - and not idx >= len(cs2) ] + Test the proposition in our paper concerning the duals and + restrictions. Generate a random cone, then create a subcone of + it. The operation of dual-taking should then commute with + _restrict_to_space:: + + sage: set_random_seed() + sage: J = random_cone(max_ambient_dim = 8) + sage: K = Cone(random_sublist(J.rays(), 0.5), lattice=J.lattice()) + sage: K_W_star = _restrict_to_space(K, J.span()).dual() + sage: K_star_W = _restrict_to_space(K.dual(), J.span()) + sage: _basically_the_same(K_W_star, K_star_W) + True + + """ + # First we want to intersect ``K`` with ``W``. The easiest way to + # do this is via cone intersection, so we turn the subspace ``W`` + # into a cone. + W_cone = Cone(W.basis() + [-b for b in W.basis()], lattice=K.lattice()) + K = K.intersection(W_cone) + + # We've already intersected K with the span of K2, so every + # generator of K should belong to W now. + K_W_rays = [ W.coordinate_vector(r) for r in K.rays() ] + + L = ToricLattice(W.dimension()) + return Cone(K_W_rays, lattice=L) - m = matrix(solid_cols) - L = ToricLattice(len(m.rows())) - J = Cone(m.transpose(), lattice=L) - return J def discrete_complementarity_set(K): r""" - Compute the discrete complementarity set of this cone. + Compute a discrete complementarity set of this cone. - The complementarity set of this cone is the set of all orthogonal - pairs `(x,s)` such that `x` is in this cone, and `s` is in its - dual. The discrete complementarity set restricts `x` and `s` to be - generators of their respective cones. + A discrete complementarity set of `K` is the set of all orthogonal + pairs `(x,s)` such that `x \in G_{1}` and `s \in G_{2}` for some + generating sets `G_{1}` of `K` and `G_{2}` of its dual. Polyhedral + convex cones are input in terms of their generators, so "the" (this + particular) discrete complementarity set corresponds to ``G1 + == K.rays()`` and ``G2 == K.dual().rays()``. OUTPUT: A list of pairs `(x,s)` such that, - * `x` is in this cone. - * `x` is a generator of this cone. - * `s` is in this cone's dual. - * `s` is a generator of this cone's dual. + * Both `x` and `s` are vectors (not rays). + * `x` is one of ``K.rays()``. + * `s` is one of ``K.dual().rays()``. * `x` and `s` are orthogonal. + REFERENCES: + + .. [Orlitzky/Gowda] M. Orlitzky and M. S. Gowda. The Lyapunov Rank of an + Improper Cone. Work in-progress. + EXAMPLES: The discrete complementarity set of the nonnegative orthant consists @@ -109,24 +267,43 @@ def discrete_complementarity_set(K): sage: discrete_complementarity_set(K) [] + Likewise when this cone is trivial (its dual is the entire space):: + + sage: L = ToricLattice(0) + sage: K = Cone([], ToricLattice(0)) + sage: discrete_complementarity_set(K) + [] + TESTS: The complementarity set of the dual can be obtained by switching the components of the complementarity set of the original cone:: - sage: K1 = random_cone(max_dim=10, max_rays=10) + sage: set_random_seed() + sage: K1 = random_cone(max_ambient_dim=6) sage: K2 = K1.dual() sage: expected = [(x,s) for (s,x) in discrete_complementarity_set(K2)] sage: actual = discrete_complementarity_set(K1) - sage: actual == expected + sage: sorted(actual) == sorted(expected) True + The pairs in the discrete complementarity set are in fact + complementary:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=6) + sage: dcs = discrete_complementarity_set(K) + sage: sum([x.inner_product(s).abs() for (x,s) in dcs]) + 0 + """ V = K.lattice().vector_space() - # Convert the rays to vectors so that we can compute inner - # products. + # Convert rays to vectors so that we can compute inner products. xs = [V(x) for x in K.rays()] + + # We also convert the generators of the dual cone so that we + # return pairs of vectors and not (vector, ray) pairs. ss = [V(s) for s in K.dual().rays()] return [(x,s) for x in xs for s in ss if x.inner_product(s) == 0] @@ -192,33 +369,39 @@ def LL(K): [0 0 1] ] + If our cone is the entire space, then every transformation on it is + Lyapunov-like:: + + sage: K = Cone([(1,0), (-1,0), (0,1), (0,-1)]) + sage: M = MatrixSpace(QQ,2) + sage: M.basis() == LL(K) + True + TESTS: The inner product `\left< L\left(x\right), s \right>` is zero for every pair `\left( x,s \right)` in the discrete complementarity set of the cone:: - sage: K = random_cone(max_dim=8, max_rays=10) + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) sage: C_of_K = discrete_complementarity_set(K) sage: l = [ (L*x).inner_product(s) for (x,s) in C_of_K for L in LL(K) ] sage: sum(map(abs, l)) 0 - Try the formula in my paper:: - - sage: K = random_cone(max_dim=15, max_rays=25) - sage: actual = lyapunov_rank(K) - sage: K_S = project_span(K) - sage: J_T1 = project_span(K, K_S.dual()) - sage: J_T2 = project_span(K_S.dual()).dual() - sage: J_T2 = Cone(J_T2.rays(), lattice=J_T1.lattice()) - sage: J_T1 == J_T2 - True - sage: J_T = J_T1 - sage: l = K.linear_subspace().dimension() - sage: codim = K.lattice_dim() - K.dim() - sage: expected = lyapunov_rank(J_T) + K.dim()*(l + codim) + codim**2 - sage: actual == expected + The Lyapunov-like transformations on a cone and its dual are related + by transposition, but we're not guaranteed to compute transposed + elements of `LL\left( K \right)` as our basis for `LL\left( K^{*} + \right)` + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) + sage: LL2 = [ L.transpose() for L in LL(K.dual()) ] + sage: V = VectorSpace( K.lattice().base_field(), K.lattice_dim()^2) + sage: LL1_vecs = [ V(m.list()) for m in LL(K) ] + sage: LL2_vecs = [ V(m.list()) for m in LL2 ] + sage: V.span(LL1_vecs) == V.span(LL2_vecs) True """ @@ -226,7 +409,7 @@ def LL(K): C_of_K = discrete_complementarity_set(K) - tensor_products = [s.tensor_product(x) for (x,s) in C_of_K] + tensor_products = [ s.tensor_product(x) for (x,s) in C_of_K ] # Sage doesn't think matrices are vectors, so we have to convert # our matrices to vectors explicitly before we can figure out how @@ -257,7 +440,7 @@ def LL(K): def lyapunov_rank(K): r""" - Compute the Lyapunov (or bilinearity) rank of this cone. + Compute the Lyapunov rank (or bilinearity rank) of this cone. The Lyapunov rank of a cone can be thought of in (mainly) two ways: @@ -276,16 +459,7 @@ def lyapunov_rank(K): An integer representing the Lyapunov rank of the cone. If the dimension of the ambient vector space is `n`, then the Lyapunov rank will be between `1` and `n` inclusive; however a rank of `n-1` is - not possible (see the first reference). - - .. note:: - - In the references, the cones are always assumed to be proper. We - do not impose this restriction. - - .. seealso:: - - :meth:`is_proper` + not possible (see [Orlitzky/Gowda]_). ALGORITHM: @@ -302,6 +476,9 @@ def lyapunov_rank(K): cone and Lyapunov-like transformations, Mathematical Programming, 147 (2014) 155-170. + .. [Orlitzky/Gowda] M. Orlitzky and M. S. Gowda. The Lyapunov Rank of an + Improper Cone. Work in-progress. + .. [Rudolf et al.] G. Rudolf, N. Noyan, D. Papp, and F. Alizadeh, Bilinear optimality constraints for the cone of positive polynomials, Mathematical Programming, Series B, 129 (2011) 5-31. @@ -321,6 +498,15 @@ def lyapunov_rank(K): sage: lyapunov_rank(octant) 3 + The full space `\mathbb{R}^{n}` has Lyapunov rank `n^{2}` + [Orlitzky/Gowda]_:: + + sage: R5 = VectorSpace(QQ, 5) + sage: gs = R5.basis() + [ -r for r in R5.basis() ] + sage: K = Cone(gs) + sage: lyapunov_rank(K) + 25 + The `L^{3}_{1}` cone is known to have a Lyapunov rank of one [Rudolf et al.]_:: @@ -334,7 +520,30 @@ def lyapunov_rank(K): sage: lyapunov_rank(L3infty) 1 - The Lyapunov rank should be additive on a product of cones + A single ray in `n` dimensions should have Lyapunov rank `n^{2} - n + + 1` [Orlitzky/Gowda]_:: + + sage: K = Cone([(1,0,0,0,0)]) + sage: lyapunov_rank(K) + 21 + sage: K.lattice_dim()**2 - K.lattice_dim() + 1 + 21 + + A subspace (of dimension `m`) in `n` dimensions should have a + Lyapunov rank of `n^{2} - m\left(n - m)` [Orlitzky/Gowda]_:: + + sage: e1 = (1,0,0,0,0) + sage: neg_e1 = (-1,0,0,0,0) + sage: e2 = (0,1,0,0,0) + sage: neg_e2 = (0,-1,0,0,0) + sage: z = (0,0,0,0,0) + sage: K = Cone([e1, neg_e1, e2, neg_e2, z, z, z]) + sage: lyapunov_rank(K) + 19 + sage: K.lattice_dim()**2 - K.dim()*K.codim() + 19 + + The Lyapunov rank should be additive on a product of proper cones [Rudolf et al.]_:: sage: L31 = Cone([(1,0,1), (0,-1,1), (-1,0,1), (0,1,1)]) @@ -360,19 +569,34 @@ def lyapunov_rank(K): TESTS: - The Lyapunov rank should be additive on a product of cones + The Lyapunov rank should be additive on a product of proper cones [Rudolf et al.]_:: - sage: K1 = random_cone(max_dim=10, max_rays=10) - sage: K2 = random_cone(max_dim=10, max_rays=10) + sage: set_random_seed() + sage: K1 = random_cone(max_ambient_dim=8, + ....: strictly_convex=True, + ....: solid=True) + sage: K2 = random_cone(max_ambient_dim=8, + ....: strictly_convex=True, + ....: solid=True) sage: K = K1.cartesian_product(K2) sage: lyapunov_rank(K) == lyapunov_rank(K1) + lyapunov_rank(K2) True + The Lyapunov rank is invariant under a linear isomorphism + [Orlitzky/Gowda]_:: + + sage: K1 = random_cone(max_ambient_dim = 8) + sage: A = random_matrix(QQ, K1.lattice_dim(), algorithm='unimodular') + sage: K2 = Cone( [ A*r for r in K1.rays() ], lattice=K1.lattice()) + sage: lyapunov_rank(K1) == lyapunov_rank(K2) + True + The dual cone `K^{*}` of ``K`` should have the same Lyapunov rank as ``K`` itself [Rudolf et al.]_:: - sage: K = random_cone(max_dim=10, max_rays=10) + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) sage: lyapunov_rank(K) == lyapunov_rank(K.dual()) True @@ -382,7 +606,10 @@ def lyapunov_rank(K): trivial cone in a trivial space as well. However, in zero dimensions, the Lyapunov rank of the trivial cone will be zero:: - sage: K = random_cone(max_dim=10, strictly_convex=True, solid=True) + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8, + ....: strictly_convex=True, + ....: solid=True) sage: b = lyapunov_rank(K) sage: n = K.lattice_dim() sage: (n == 0 or 1 <= b) and b <= n @@ -390,5 +617,120 @@ def lyapunov_rank(K): sage: b == n-1 False + In fact [Orlitzky/Gowda]_, no closed convex polyhedral cone can have + Lyapunov rank `n-1` in `n` dimensions:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) + sage: b = lyapunov_rank(K) + sage: n = K.lattice_dim() + sage: b == n-1 + False + + The calculation of the Lyapunov rank of an improper cone can be + reduced to that of a proper cone [Orlitzky/Gowda]_:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) + sage: actual = lyapunov_rank(K) + sage: K_S = _restrict_to_space(K, K.span()) + sage: K_SP = _restrict_to_space(K_S.dual(), K_S.dual().span()).dual() + sage: l = K.lineality() + sage: c = K.codim() + sage: expected = lyapunov_rank(K_SP) + K.dim()*(l + c) + c**2 + sage: actual == expected + True + + The Lyapunov rank of any cone is just the dimension of ``LL(K)``:: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8) + sage: lyapunov_rank(K) == len(LL(K)) + True + + We can make an imperfect cone perfect by adding a slack variable + (a Theorem in [Orlitzky/Gowda]_):: + + sage: set_random_seed() + sage: K = random_cone(max_ambient_dim=8, + ....: strictly_convex=True, + ....: solid=True) + sage: L = ToricLattice(K.lattice_dim() + 1) + sage: K = Cone([ r.list() + [0] for r in K.rays() ], lattice=L) + sage: lyapunov_rank(K) >= K.lattice_dim() + True + + """ + beta = 0 + + m = K.dim() + n = K.lattice_dim() + l = K.lineality() + + if m < n: + # K is not solid, restrict to its span. + K = _restrict_to_space(K, K.span()) + + # Non-solid reduction lemma. + beta += (n - m)*n + + if l > 0: + # K is not pointed, restrict to the span of its dual. Uses a + # proposition from our paper, i.e. this is equivalent to K = + # _rho(K.dual()).dual(). + K = _restrict_to_space(K, K.dual().span()) + + # Non-pointed reduction lemma. + beta += l * m + + beta += len(LL(K)) + return beta + + + +def is_lyapunov_like(L,K): + r""" + Determine whether or not ``L`` is Lyapunov-like on ``K``. + + We say that ``L`` is Lyapunov-like on ``K`` if `\left\langle + L\left\lparenx\right\rparen,s\right\rangle = 0` for all pairs + `\left\langle x,s \right\rangle` in the complementarity set of + ``K``. It is known [Orlitzky]_ that this property need only be + checked for generators of ``K`` and its dual. + + INPUT: + + - ``L`` -- A linear transformation or matrix. + + - ``K`` -- A polyhedral closed convex cone. + + OUTPUT: + + ``True`` if it can be proven that ``L`` is Lyapunov-like on ``K``, + and ``False`` otherwise. + + .. WARNING:: + + If this function returns ``True``, then ``L`` is Lyapunov-like + on ``K``. However, if ``False`` is returned, that could mean one + of two things. The first is that ``L`` is definitely not + Lyapunov-like on ``K``. The second is more of an "I don't know" + answer, returned (for example) if we cannot prove that an inner + product is zero. + + REFERENCES: + + .. [Orlitzky] M. Orlitzky. The Lyapunov rank of an + improper cone (preprint). + + EXAMPLES: + + todo. + + TESTS: + + todo. + """ - return len(LL(K)) + return all([(L*x).inner_product(s) == 0 + for (x,s) in discrete_complementarity_set(K)])