TensorFlow 1 version

View source on GitHub

LinearOperator acting like a circulant matrix.

tf.linalg.LinearOperatorCirculant(
    spectrum, input_output_dtype=tf.dtypes.complex64, is_non_singular=None,
    is_self_adjoint=None, is_positive_definite=None, is_square=True,
    name='LinearOperatorCirculant'
)

This operator acts like a circulant matrix A with shape [B1,...,Bb, N, N] for some b >= 0. The first b indices index a batch member. For every batch index (i1,...,ib), A[i1,...,ib, : :] is an N x N matrix. This matrix A is not materialized, but for purposes of broadcasting this shape will be relevant.

Description in terms of circulant matrices

Circulant means the entries of A are generated by a single vector, the convolution kernel h: A_{mn} := h_{m-n mod N}. With h = [w, x, y, z],

A = |w z y x|
    |x w z y|
    |y x w z|
    |z y x w|

This means that the result of matrix multiplication v = Au has Lth column given circular convolution between h with the Lth column of u.

Description in terms of the frequency spectrum

There is an equivalent description in terms of the [batch] spectrum H and Fourier transforms. Here we consider A.shape = [N, N] and ignore batch dimensions. Define the discrete Fourier transform (DFT) and its inverse by

DFT[ h[n] ] = H[k] := sum_{n = 0}^{N - 1} h_n e^{-i 2pi k n / N}
IDFT[ H[k] ] = h[n] = N^{-1} sum_{k = 0}^{N - 1} H_k e^{i 2pi k n / N}

From these definitions, we see that

H[0] = sum_{n = 0}^{N - 1} h_n
H[1] = "the first positive frequency"
H[N - 1] = "the first negative frequency"

Loosely speaking, with * element-wise multiplication, matrix multiplication is equal to the action of a Fourier multiplier: A u = IDFT[ H * DFT[u] ]. Precisely speaking, given [N, R] matrix u, let DFT[u] be the [N, R] matrix with rth column equal to the DFT of the rth column of u. Define the IDFT similarly. Matrix multiplication may be expressed columnwise:

Operator properties deduced from the spectrum.

Letting U be the kth Euclidean basis vector, and U = IDFT[u]. The above formulas show thatA U = H_k * U. We conclude that the elements of H are the eigenvalues of this operator. Therefore

• This operator is positive definite if and only if Real{H} > 0.

A general property of Fourier transforms is the correspondence between Hermitian functions and real valued transforms.

Suppose H.shape = [B1,...,Bb, N]. We say that H is a Hermitian spectrum if, with % meaning modulus division,

• This operator corresponds to a real matrix if and only if H is Hermitian.
• This operator is self-adjoint if and only if H is real.

See e.g. "Discrete-Time Signal Processing", Oppenheim and Schafer.

Example of a self-adjoint positive definite operator

# spectrum is real ==> operator is self-adjoint
# spectrum is positive ==> operator is positive definite
spectrum = [6., 4, 2]

operator = LinearOperatorCirculant(spectrum)

# IFFT[spectrum]
operator.convolution_kernel()
==> [4 + 0j, 1 + 0.58j, 1 - 0.58j]

operator.to_dense()
==> [[4 + 0.0j, 1 - 0.6j, 1 + 0.6j],
     [1 + 0.6j, 4 + 0.0j, 1 - 0.6j],
     [1 - 0.6j, 1 + 0.6j, 4 + 0.0j]]

Example of defining in terms of a real convolution kernel

# convolution_kernel is real ==> spectrum is Hermitian.
convolution_kernel = [1., 2., 1.]]
spectrum = tf.signal.fft(tf.cast(convolution_kernel, tf.complex64))

# spectrum is Hermitian ==> operator is real.
# spectrum is shape [3] ==> operator is shape [3, 3]
# We force the input/output type to be real, which allows this to operate
# like a real matrix.
operator = LinearOperatorCirculant(spectrum, input_output_dtype=tf.float32)

operator.to_dense()
==> [[ 1, 1, 2],
     [ 2, 1, 1],
     [ 1, 2, 1]]

Example of Hermitian spectrum

# spectrum is shape [3] ==> operator is shape [3, 3]
# spectrum is Hermitian ==> operator is real.
spectrum = [1, 1j, -1j]

operator = LinearOperatorCirculant(spectrum)

operator.to_dense()
==> [[ 0.33 + 0j,  0.91 + 0j, -0.24 + 0j],
     [-0.24 + 0j,  0.33 + 0j,  0.91 + 0j],
     [ 0.91 + 0j, -0.24 + 0j,  0.33 + 0j]

Example of forcing real dtype when spectrum is Hermitian

# spectrum is shape [4] ==> operator is shape [4, 4]
# spectrum is real ==> operator is self-adjoint
# spectrum is Hermitian ==> operator is real
# spectrum has positive real part ==> operator is positive-definite.
spectrum = [6., 4, 2, 4]

# Force the input dtype to be float32.
# Cast the output to float32.  This is fine because the operator will be
# real due to Hermitian spectrum.
operator = LinearOperatorCirculant(spectrum, input_output_dtype=tf.float32)

operator.shape
==> [4, 4]

operator.to_dense()
==> [[4, 1, 0, 1],
     [1, 4, 1, 0],
     [0, 1, 4, 1],
     [1, 0, 1, 4]]

# convolution_kernel = tf.signal.ifft(spectrum)
operator.convolution_kernel()
==> [4, 1, 0, 1]

Performance

Suppose operator is a LinearOperatorCirculant of shape [N, N], and x.shape = [N, R]. Then

• operator.matmul(x) is O(R*N*Log[N])
• operator.solve(x) is O(R*N*Log[N])
• operator.determinant() involves a size N reduce_prod.

If instead operator and x have shape [B1,...,Bb, N, N] and [B1,...,Bb, N, R], every operation increases in complexity by B1*...*Bb.

Matrix property hints

This LinearOperator is initialized with boolean flags of the form is_X, for X = non_singular, self_adjoint, positive_definite, square. These have the following meaning:

• If is_X == True, callers should expect the operator to have the property X. This is a promise that should be fulfilled, but is not a runtime assert. For example, finite floating point precision may result in these promises being violated.
• If is_X == False, callers should expect the operator to not have X.
• If is_X == None (the default), callers should have no expectation either way.

References:

Toeplitz and Circulant Matrices - A Review: Gray, 2006 (pdf)

A = |w z y x|
|x w z y|
|y x w z|
|z y x w|

A = |W Z Y X|
|X W Z Y|
|Y X W Z|
|Z Y X W|

Methods

add_to_tensor

View source

add_to_tensor(
    x, name='add_to_tensor'
)

Add matrix represented by this operator to x. Equivalent to A + x.

adjoint

View source

adjoint(
    name='adjoint'
)

Returns the adjoint of the current LinearOperator.

Given A representing this LinearOperator, return A*. Note that calling self.adjoint() and self.H are equivalent.

assert_hermitian_spectrum

View source

assert_hermitian_spectrum(
    name='assert_hermitian_spectrum'
)

Returns an Op that asserts this operator has Hermitian spectrum.

This operator corresponds to a real-valued matrix if and only if its spectrum is Hermitian.

assert_non_singular

View source

assert_non_singular(
    name='assert_non_singular'
)

Returns an Op that asserts this operator is non singular.

This operator is considered non-singular if

ConditionNumber < max{100, range_dimension, domain_dimension} * eps,
eps := np.finfo(self.dtype.as_numpy_dtype).eps

assert_positive_definite

View source

assert_positive_definite(
    name='assert_positive_definite'
)

Returns an Op that asserts this operator is positive definite.

Here, positive definite means that the quadratic form x^H A x has positive real part for all nonzero x. Note that we do not require the operator to be self-adjoint to be positive definite.

assert_self_adjoint

View source

assert_self_adjoint(
    name='assert_self_adjoint'
)

Returns an Op that asserts this operator is self-adjoint.

Here we check that this operator is exactly equal to its hermitian transpose.

batch_shape_tensor

View source

batch_shape_tensor(
    name='batch_shape_tensor'
)

Shape of batch dimensions of this operator, determined at runtime.

If this operator acts like the batch matrix A with A.shape = [B1,...,Bb, M, N], then this returns a Tensor holding [B1,...,Bb].

block_shape_tensor

View source

block_shape_tensor()

Shape of the block dimensions of self.spectrum.

cholesky

View source

cholesky(
    name='cholesky'
)

Returns a Cholesky factor as a LinearOperator.

Given A representing this LinearOperator, if A is positive definite self-adjoint, return L, where A = L L^T, i.e. the cholesky decomposition.

cond

View source

cond(
    name='cond'
)

Returns the condition number of this linear operator.

convolution_kernel

View source

convolution_kernel(
    name='convolution_kernel'
)

Convolution kernel corresponding to self.spectrum.

The D dimensional DFT of this kernel is the frequency domain spectrum of this operator.

determinant

View source

determinant(
    name='det'
)

Determinant for every batch member.

diag_part

View source

diag_part(
    name='diag_part'
)

Efficiently get the [batch] diagonal part of this operator.

If this operator has shape [B1,...,Bb, M, N], this returns a Tensor diagonal, of shape [B1,...,Bb, min(M, N)], where diagonal[b1,...,bb, i] = self.to_dense()[b1,...,bb, i, i].

my_operator = LinearOperatorDiag([1., 2.])

# Efficiently get the diagonal
my_operator.diag_part()
==> [1., 2.]

# Equivalent, but inefficient method
tf.linalg.diag_part(my_operator.to_dense())
==> [1., 2.]

domain_dimension_tensor

View source

domain_dimension_tensor(
    name='domain_dimension_tensor'
)

Dimension (in the sense of vector spaces) of the domain of this operator.

Determined at runtime.

If this operator acts like the batch matrix A with A.shape = [B1,...,Bb, M, N], then this returns N.

eigvals

View source

eigvals(
    name='eigvals'
)

Returns the eigenvalues of this linear operator.

If the operator is marked as self-adjoint (via is_self_adjoint) this computation can be more efficient.

inverse

View source

inverse(
    name='inverse'
)

Returns the Inverse of this LinearOperator.

Given A representing this LinearOperator, return a LinearOperator representing A^-1.

log_abs_determinant

View source

log_abs_determinant(
    name='log_abs_det'
)

Log absolute value of determinant for every batch member.

matmul

View source

matmul(
    x, adjoint=False, adjoint_arg=False, name='matmul'
)

Transform [batch] matrix x with left multiplication: x --> Ax.

# Make an operator acting like batch matrix A.  Assume A.shape = [..., M, N]
operator = LinearOperator(...)
operator.shape = [..., M, N]

X = ... # shape [..., N, R], batch matrix, R > 0.

Y = operator.matmul(X)
Y.shape
==> [..., M, R]

Y[..., :, r] = sum_j A[..., :, j] X[j, r]

matvec

View source

matvec(
    x, adjoint=False, name='matvec'
)

Transform [batch] vector x with left multiplication: x --> Ax.

# Make an operator acting like batch matrix A.  Assume A.shape = [..., M, N]
operator = LinearOperator(...)

X = ... # shape [..., N], batch vector

Y = operator.matvec(X)
Y.shape
==> [..., M]

Y[..., :] = sum_j A[..., :, j] X[..., j]

range_dimension_tensor

View source

range_dimension_tensor(
    name='range_dimension_tensor'
)

Dimension (in the sense of vector spaces) of the range of this operator.

Determined at runtime.

If this operator acts like the batch matrix A with A.shape = [B1,...,Bb, M, N], then this returns M.

shape_tensor

View source

shape_tensor(
    name='shape_tensor'
)

Shape of this LinearOperator, determined at runtime.

If this operator acts like the batch matrix A with A.shape = [B1,...,Bb, M, N], then this returns a Tensor holding [B1,...,Bb, M, N], equivalent to tf.shape(A).

solve

View source

solve(
    rhs, adjoint=False, adjoint_arg=False, name='solve'
)

Solve (exact or approx) R (batch) systems of equations: A X = rhs.

The returned Tensor will be close to an exact solution if A is well conditioned. Otherwise closeness will vary. See class docstring for details.

Examples:

# Make an operator acting like batch matrix A.  Assume A.shape = [..., M, N]
operator = LinearOperator(...)
operator.shape = [..., M, N]

# Solve R > 0 linear systems for every member of the batch.
RHS = ... # shape [..., M, R]

X = operator.solve(RHS)
# X[..., :, r] is the solution to the r'th linear system
# sum_j A[..., :, j] X[..., j, r] = RHS[..., :, r]

operator.matmul(X)
==> RHS

solvevec

View source

solvevec(
    rhs, adjoint=False, name='solve'
)

Solve single equation with best effort: A X = rhs.

The returned Tensor will be close to an exact solution if A is well conditioned. Otherwise closeness will vary. See class docstring for details.

Examples:

# Make an operator acting like batch matrix A.  Assume A.shape = [..., M, N]
operator = LinearOperator(...)
operator.shape = [..., M, N]

# Solve one linear system for every member of the batch.
RHS = ... # shape [..., M]

X = operator.solvevec(RHS)
# X is the solution to the linear system
# sum_j A[..., :, j] X[..., j] = RHS[..., :]

operator.matvec(X)
==> RHS

tensor_rank_tensor

View source

tensor_rank_tensor(
    name='tensor_rank_tensor'
)

Rank (in the sense of tensors) of matrix corresponding to this operator.

If this operator acts like the batch matrix A with A.shape = [B1,...,Bb, M, N], then this returns b + 2.

to_dense

View source

to_dense(
    name='to_dense'
)

Return a dense (batch) matrix representing this operator.

trace

View source

trace(
    name='trace'
)

Trace of the linear operator, equal to sum of self.diag_part().

If the operator is square, this is also the sum of the eigenvalues.

__matmul__

View source

__matmul__(
    other
)