def add_canon(expr, args):
if expr.is_scalar():
return log_sum_exp(hstack(args)), []
rows = []
summands = [promote(s, expr.shape) if s.is_scalar() else s for s in args]
if len(expr.shape) == 1:
for i in range(expr.shape[0]):
row = []
row.append(
log_sum_exp(hstack([summand[i] for summand in summands])))
rows.append(row)
return reshape(bmat(rows), expr.shape), []
else:
for i in range(expr.shape[0]):
row = []
for j in range(expr.shape[1]):
row.append(
log_sum_exp(hstack([summand[i, j]
for summand in summands])))
rows.append(row)
return reshape(bmat(rows), expr.shape), []
def sigma_max_canon(expr, args):
A = args[0]
n, m = A.shape
shape = expr.shape
if not np.prod(shape) == 1:
raise RuntimeError('Invalid shape of expr in sigma_max canonicalization.')
t = Variable(shape)
tI_n = sp.eye(n) * t
tI_m = sp.eye(m) * t
X = bmat([[tI_n, A],
[A.T, tI_m]])
constraints = [PSD(X)]
return t, constraints
def normNuc_canon(expr, args):
A = args[0]
m, n = A.shape
# Create the equivalent problem:
# minimize (trace(U) + trace(V))/2
# subject to:
# [U A; A.T V] is positive semidefinite
constraints = []
U = Variable(shape=(m, m), symmetric=True)
V = Variable(shape=(n, n), symmetric=True)
X = bmat([[U, A], [A.T, V]])
constraints.append(X >> 0)
trace_value = 0.5 * (trace(U) + trace(V))
return trace_value, constraints
def log_det_canon(expr, args):
"""Reduces the atom to an affine expression and list of constraints.
Creates the equivalent problem::
maximize sum(log(D[i, i]))
subject to: D diagonal
diag(D) = diag(Z)
Z is upper triangular.
[D Z; Z.T A] is positive semidefinite
The problem computes the LDL factorization:
.. math::
A = (Z^TD^{-1})D(D^{-1}Z)
This follows from the inequality:
.. math::
\\det(A) >= \\det(D) + \\det([D, Z; Z^T, A])/\\det(D)
>= \\det(D)
because (Z^TD^{-1})D(D^{-1}Z) is a feasible D, Z that achieves
det(A) = det(D) and the objective maximizes det(D).
Parameters
----------
expr : log_det
args : list
The arguments for the expression
Returns
-------
tuple
(Variable for objective, list of constraints)
"""
A = args[0] # n by n matrix.
n, _ = A.shape
z = Variable(shape=(n * (n + 1) // 2, ))
Z = vec_to_upper_tri(z, strict=False)
d = diag_mat(Z) # a vector
D = diag_vec(d) # a matrix
X = bmat([[D, Z], [Z.T, A]])
constraints = [PSD(X)]
log_expr = log(d)
obj, constr = log_canon(log_expr, log_expr.args)
constraints += constr
return sum(obj), constraints
def mulexpression_canon(expr, args):
lhs = args[0]
rhs = args[1]
lhs_shape, rhs_shape, _ = mul_shapes_promote(lhs.shape, rhs.shape)
lhs = reshape(lhs, lhs_shape)
rhs = reshape(rhs, rhs_shape)
rows = []
# TODO(akshayka): Parallelize this for large matrices.
for i in range(lhs.shape[0]):
row = []
for j in range(rhs.shape[1]):
arr = hstack([lhs[i, k] + rhs[k, j] for k in range(lhs.shape[1])])
row.append(log_sum_exp(arr))
rows.append(row)
mat = bmat(rows)
if mat.shape != expr.shape:
mat = reshape(mat, expr.shape)
return mat, []
