def _covariance(self): # Derivation: https://sachinruk.github.io/blog/von-Mises-Fisher/ event_dim = tf.compat.dimension_value(self.event_shape[0]) if event_dim is None: raise ValueError( 'event shape must be statically known for _bessel_ive') # TODO(bjp): Enable this; numerically unstable. if event_dim > 2: raise ValueError( 'vMF covariance is numerically unstable for dim>2') concentration = self.concentration[..., tf.newaxis] safe_conc = tf.where(concentration > 0, concentration, tf.ones_like(concentration)) h = (_bessel_ive(event_dim / 2, safe_conc) / _bessel_ive(event_dim / 2 - 1, safe_conc)) intermediate = ( tf.matmul(self.mean_direction[..., :, tf.newaxis], self.mean_direction[..., tf.newaxis, :]) * (1 - event_dim * h / safe_conc - h**2)[..., tf.newaxis]) cov = tf.linalg.set_diag( intermediate, tf.linalg.diag_part(intermediate) + (h / safe_conc)) return tf.where( concentration[..., tf.newaxis] > tf.zeros_like(cov), cov, tf.linalg.eye(event_dim, batch_shape=self.batch_shape_tensor()) / event_dim)
def cholesky_concat(chol, cols, name=None): """Concatenates `chol @ chol.T` with additional rows and columns. This operation is conceptually identical to: ```python def cholesky_concat_slow(chol, cols): # cols shaped (n + m) x m = z x m mat = tf.matmul(chol, chol, adjoint_b=True) # batch of n x n # Concat columns. mat = tf.concat([mat, cols[..., :tf.shape(mat)[-2], :]], axis=-1) # n x z # Concat rows. mat = tf.concat([mat, tf.linalg.matrix_transpose(cols)], axis=-2) # z x z return tf.linalg.cholesky(mat) ``` but whereas `cholesky_concat_slow` would cost `O(z**3)` work, `cholesky_concat` only costs `O(z**2 + m**3)` work. The resulting (implicit) matrix must be symmetric and positive definite. Thus, the bottom right `m x m` must be self-adjoint, and we do not require a separate `rows` argument (which can be inferred from `conj(cols.T)`). Args: chol: Cholesky decomposition of `mat = chol @ chol.T`. cols: The new columns whose first `n` rows we would like concatenated to the right of `mat = chol @ chol.T`, and whose conjugate transpose we would like concatenated to the bottom of `concat(mat, cols[:n,:])`. A `Tensor` with final dims `(n+m, m)`. The first `n` rows are the top right rectangle (their conjugate transpose forms the bottom left), and the bottom `m x m` is self-adjoint. name: Optional name for this op. Returns: chol_concat: The Cholesky decomposition of: ``` [ [ mat cols[:n, :] ] [ conj(cols.T) ] ] ``` """ with tf.name_scope(name or 'cholesky_extend'): dtype = dtype_util.common_dtype([chol, cols], dtype_hint=tf.float32) chol = tf.convert_to_tensor(chol, name='chol', dtype=dtype) cols = tf.convert_to_tensor(cols, name='cols', dtype=dtype) n = prefer_static.shape(chol)[-1] mat_nm, mat_mm = cols[..., :n, :], cols[..., n:, :] solved_nm = linear_operator_util.matrix_triangular_solve_with_broadcast( chol, mat_nm) lower_right_mm = tf.linalg.cholesky( mat_mm - tf.matmul(solved_nm, solved_nm, adjoint_a=True)) lower_left_mn = tf.math.conj(tf.linalg.matrix_transpose(solved_nm)) out_batch = prefer_static.shape(solved_nm)[:-2] chol = tf.broadcast_to( chol, tf.concat([out_batch, prefer_static.shape(chol)[-2:]], axis=0)) top_right_zeros_nm = tf.zeros_like(solved_nm) return tf.concat([ tf.concat([chol, top_right_zeros_nm], axis=-1), tf.concat([lower_left_mn, lower_right_mm], axis=-1) ], axis=-2)
def _variance(self): # Because df is a scalar, we need to expand dimensions to match # scale_operator. We use ellipses notation (...) to select all dimensions # and add two dimensions to the end. df = self.df[..., tf.newaxis, tf.newaxis] x = tf.sqrt(df) * self._square_scale_operator() d = tf.expand_dims(tf.linalg.diag_part(x), -1) v = tf.square(x) + tf.matmul(d, d, adjoint_b=True) return v
def _covariance(self): total_count = tf.convert_to_tensor(self._total_count) concentration = tf.convert_to_tensor(self._concentration) scale = self._variance_scale_term(total_count, concentration) x = scale * self._mean(total_count, concentration) return tf.linalg.set_diag( -tf.matmul(x[..., tf.newaxis], x[..., tf.newaxis, :]), # outer prod self._variance(total_count, concentration))
def _covariance(self): concentration = tf.convert_to_tensor(self.concentration) total_concentration = tf.reduce_sum(concentration, axis=-1, keepdims=True) mean = concentration / total_concentration scale = tf.math.rsqrt(1. + total_concentration) x = scale * mean variance = x * (scale - x) return tf.linalg.set_diag( tf.matmul(-x[..., tf.newaxis], x[..., tf.newaxis, :]), variance)
def _forward(self, x): with tf.control_dependencies(self._assertions(x)): x_shape = tf.shape(x) identity_matrix = tf.eye(x_shape[-1], batch_shape=x_shape[:-2], dtype=dtype_util.base_dtype(x.dtype)) # Note `matrix_triangular_solve` implicitly zeros upper triangular of `x`. y = tf.linalg.triangular_solve(x, identity_matrix) y = tf.matmul(y, y, adjoint_a=True) return tf.linalg.cholesky(y)
def sparse_or_dense_matmul(sparse_or_dense_a, dense_b, validate_args=False, name=None, **kwargs): """Returns (batched) matmul of a SparseTensor (or Tensor) with a Tensor. Args: sparse_or_dense_a: `SparseTensor` or `Tensor` representing a (batch of) matrices. dense_b: `Tensor` representing a (batch of) matrices, with the same batch shape as `sparse_or_dense_a`. The shape must be compatible with the shape of `sparse_or_dense_a` and kwargs. validate_args: When `True`, additional assertions might be embedded in the graph. Default value: `False` (i.e., no graph assertions are added). name: Python `str` prefixed to ops created by this function. Default value: 'sparse_or_dense_matmul'. **kwargs: Keyword arguments to `tf.sparse_tensor_dense_matmul` or `tf.matmul`. Returns: product: A dense (batch of) matrix-shaped Tensor of the same batch shape and dtype as `sparse_or_dense_a` and `dense_b`. If `sparse_or_dense_a` or `dense_b` is adjointed through `kwargs` then the shape is adjusted accordingly. """ with tf.name_scope(name or 'sparse_or_dense_matmul'): dense_b = tf.convert_to_tensor(dense_b, dtype_hint=tf.float32, name='dense_b') if validate_args: assert_a_rank_at_least_2 = assert_util.assert_rank_at_least( sparse_or_dense_a, rank=2, message= 'Input `sparse_or_dense_a` must have at least 2 dimensions.') assert_b_rank_at_least_2 = assert_util.assert_rank_at_least( dense_b, rank=2, message='Input `dense_b` must have at least 2 dimensions.') with tf.control_dependencies( [assert_a_rank_at_least_2, assert_b_rank_at_least_2]): sparse_or_dense_a = tf.identity(sparse_or_dense_a) dense_b = tf.identity(dense_b) if isinstance(sparse_or_dense_a, (tf.SparseTensor, tf1.SparseTensorValue)): return _sparse_tensor_dense_matmul(sparse_or_dense_a, dense_b, **kwargs) else: return tf.matmul(sparse_or_dense_a, dense_b, **kwargs)
def _sparse_block_diag(sp_a): """Returns a block diagonal rank 2 SparseTensor from a batch of SparseTensors. Args: sp_a: A rank 3 `SparseTensor` representing a batch of matrices. Returns: sp_block_diag_a: matrix-shaped, `float` `SparseTensor` with the same dtype as `sparse_or_matrix`, of shape [B * M, B * N] where `sp_a` has shape [B, M, N]. Each [M, N] batch of `sp_a` is lined up along the diagonal. """ # Construct the matrix [[M, N], [1, 0], [0, 1]] which would map the index # (b, i, j) to (Mb + i, Nb + j). This effectively creates a block-diagonal # matrix of dense shape [B * M, B * N]. # Note that this transformation doesn't increase the number of non-zero # entries in the SparseTensor. sp_a_shape = tf.convert_to_tensor(_get_shape(sp_a, tf.int64)) ind_mat = tf.concat([[sp_a_shape[-2:]], tf.eye(2, dtype=tf.int64)], axis=0) indices = tf.matmul(sp_a.indices, ind_mat) dense_shape = sp_a_shape[0] * sp_a_shape[1:] return tf.SparseTensor(indices=indices, values=sp_a.values, dense_shape=dense_shape)
def _covariance(self): p = self._probs_parameter_no_checks() ret = -tf.matmul(p[..., None], p[..., None, :]) return tf.linalg.set_diag(ret, self._variance(p))
def _forward_log_det_jacobian(self, x): # Let Y be a symmetric, positive definite matrix and write: # Y = X X.T # where X is lower-triangular. # # Observe that, # dY[i,j]/dX[a,b] # = d/dX[a,b] { X[i,:] X[j,:] } # = sum_{d=1}^p { I[i=a] I[d=b] X[j,d] + I[j=a] I[d=b] X[i,d] } # # To compute the Jacobian dX/dY we must represent X,Y as vectors. Since Y is # symmetric and X is lower-triangular, we need vectors of dimension: # d = p (p + 1) / 2 # where X, Y are p x p matrices, p > 0. We use a row-major mapping, i.e., # k = { i (i + 1) / 2 + j i>=j # { undef i<j # and assume zero-based indexes. When k is undef, the element is dropped. # Example: # j k # 0 1 2 3 / # 0 [ 0 . . . ] # i 1 [ 1 2 . . ] # 2 [ 3 4 5 . ] # 3 [ 6 7 8 9 ] # Write vec[.] to indicate transforming a matrix to vector via k(i,j). (With # slight abuse: k(i,j)=undef means the element is dropped.) # # We now show d vec[Y] / d vec[X] is lower triangular. Assuming both are # defined, observe that k(i,j) < k(a,b) iff (1) i<a or (2) i=a and j<b. # In both cases dvec[Y]/dvec[X]@[k(i,j),k(a,b)] = 0 since: # (1) j<=i<a thus i,j!=a. # (2) i=a>j thus i,j!=a. # # Since the Jacobian is lower-triangular, we need only compute the product # of diagonal elements: # d vec[Y] / d vec[X] @[k(i,j), k(i,j)] # = X[j,j] + I[i=j] X[i,j] # = 2 X[j,j]. # Since there is a 2 X[j,j] term for every lower-triangular element of X we # conclude: # |Jac(d vec[Y]/d vec[X])| = 2^p prod_{j=0}^{p-1} X[j,j]^{p-j}. diag = tf.linalg.diag_part(x) # We now ensure diag is columnar. Eg, if `diag = [1, 2, 3]` then the output # is `[[1], [2], [3]]` and if `diag = [[1, 2, 3], [4, 5, 6]]` then the # output is unchanged. diag = self._make_columnar(diag) with tf.control_dependencies(self._assertions(x)): # Create a vector equal to: [p, p-1, ..., 2, 1]. if tf.compat.dimension_value(x.shape[-1]) is None: p_int = tf.shape(x)[-1] p_float = tf.cast(p_int, dtype=x.dtype) else: p_int = tf.compat.dimension_value(x.shape[-1]) p_float = dtype_util.as_numpy_dtype(x.dtype)(p_int) exponents = tf.linspace(p_float, 1., p_int) sum_weighted_log_diag = tf.squeeze(tf.matmul( tf.math.log(diag), exponents[..., tf.newaxis]), axis=-1) fldj = p_float * np.log(2.) + sum_weighted_log_diag # We finally need to undo adding an extra column in non-scalar cases # where there is a single matrix as input. if tensorshape_util.rank(x.shape) is not None: if tensorshape_util.rank(x.shape) == 2: fldj = tf.squeeze(fldj, axis=-1) return fldj shape = tf.shape(fldj) maybe_squeeze_shape = tf.concat([ shape[:-1], distribution_util.pick_vector(tf.equal( tf.rank(x), 2), np.array([], dtype=np.int32), shape[-1:]) ], 0) return tf.reshape(fldj, maybe_squeeze_shape)
def _forward(self, x): with tf.control_dependencies(self._assertions(x)): # For safety, explicitly zero-out the upper triangular part. x = tf.linalg.band_part(x, -1, 0) return tf.matmul(x, x, adjoint_b=True)
def lu_reconstruct(lower_upper, perm, validate_args=False, name=None): """The inverse LU decomposition, `X == lu_reconstruct(*tf.linalg.lu(X))`. Args: lower_upper: `lu` as returned by `tf.linalg.lu`, i.e., if `matmul(P, matmul(L, U)) = X` then `lower_upper = L + U - eye`. perm: `p` as returned by `tf.linag.lu`, i.e., if `matmul(P, matmul(L, U)) = X` then `perm = argmax(P)`. validate_args: Python `bool` indicating whether arguments should be checked for correctness. Default value: `False` (i.e., don't validate arguments). name: Python `str` name given to ops managed by this object. Default value: `None` (i.e., 'lu_reconstruct'). Returns: x: The original input to `tf.linalg.lu`, i.e., `x` as in, `lu_reconstruct(*tf.linalg.lu(x))`. #### Examples ```python import numpy as np from tensorflow_probability.python.internal.backend import numpy as tf import tensorflow_probability as tfp; tfp = tfp.experimental.substrates.numpy x = [[[3., 4], [1, 2]], [[7., 8], [3, 4]]] x_reconstructed = tfp.math.lu_reconstruct(*tf.linalg.lu(x)) tf.assert_near(x, x_reconstructed) # ==> True ``` """ with tf.name_scope(name or 'lu_reconstruct'): lower_upper = tf.convert_to_tensor(lower_upper, dtype_hint=tf.float32, name='lower_upper') perm = tf.convert_to_tensor(perm, dtype_hint=tf.int32, name='perm') assertions = lu_reconstruct_assertions(lower_upper, perm, validate_args) if assertions: with tf.control_dependencies(assertions): lower_upper = tf.identity(lower_upper) perm = tf.identity(perm) shape = tf.shape(lower_upper) lower = tf.linalg.set_diag( tf.linalg.band_part(lower_upper, num_lower=-1, num_upper=0), tf.ones(shape[:-1], dtype=lower_upper.dtype)) upper = tf.linalg.band_part(lower_upper, num_lower=0, num_upper=-1) x = tf.matmul(lower, upper) if (tensorshape_util.rank(lower_upper.shape) is None or tensorshape_util.rank(lower_upper.shape) != 2): # We either don't know the batch rank or there are >0 batch dims. batch_size = tf.reduce_prod(shape[:-2]) d = shape[-1] x = tf.reshape(x, [batch_size, d, d]) perm = tf.reshape(perm, [batch_size, d]) perm = tf.map_fn(tf.math.invert_permutation, perm) batch_indices = tf.broadcast_to( tf.range(batch_size)[:, tf.newaxis], [batch_size, d]) x = tf.gather_nd(x, tf.stack([batch_indices, perm], axis=-1)) x = tf.reshape(x, shape) else: x = tf.gather(x, tf.math.invert_permutation(perm)) x.set_shape(lower_upper.shape) return x
def pinv(a, rcond=None, validate_args=False, name=None): """Compute the Moore-Penrose pseudo-inverse of a matrix. Calculate the [generalized inverse of a matrix]( https://en.wikipedia.org/wiki/Moore%E2%80%93Penrose_inverse) using its singular-value decomposition (SVD) and including all large singular values. The pseudo-inverse of a matrix `A`, is defined as: 'the matrix that 'solves' [the least-squares problem] `A @ x = b`,' i.e., if `x_hat` is a solution, then `A_pinv` is the matrix such that `x_hat = A_pinv @ b`. It can be shown that if `U @ Sigma @ V.T = A` is the singular value decomposition of `A`, then `A_pinv = V @ inv(Sigma) U^T`. [(Strang, 1980)][1] This function is analogous to [`numpy.linalg.pinv`]( https://docs.scipy.org/doc/numpy/reference/generated/numpy.linalg.pinv.html). It differs only in default value of `rcond`. In `numpy.linalg.pinv`, the default `rcond` is `1e-15`. Here the default is `10. * max(num_rows, num_cols) * np.finfo(dtype).eps`. Args: a: (Batch of) `float`-like matrix-shaped `Tensor`(s) which are to be pseudo-inverted. rcond: `Tensor` of small singular value cutoffs. Singular values smaller (in modulus) than `rcond` * largest_singular_value (again, in modulus) are set to zero. Must broadcast against `tf.shape(a)[:-2]`. Default value: `10. * max(num_rows, num_cols) * np.finfo(a.dtype).eps`. validate_args: When `True`, additional assertions might be embedded in the graph. Default value: `False` (i.e., no graph assertions are added). name: Python `str` prefixed to ops created by this function. Default value: 'pinv'. Returns: a_pinv: The pseudo-inverse of input `a`. Has same shape as `a` except rightmost two dimensions are transposed. Raises: TypeError: if input `a` does not have `float`-like `dtype`. ValueError: if input `a` has fewer than 2 dimensions. #### Examples ```python from tensorflow_probability.python.internal.backend import numpy as tf import tensorflow_probability as tfp; tfp = tfp.experimental.substrates.numpy a = tf.constant([[1., 0.4, 0.5], [0.4, 0.2, 0.25], [0.5, 0.25, 0.35]]) tf.matmul(tfp.math.pinv(a), a) # ==> array([[1., 0., 0.], [0., 1., 0.], [0., 0., 1.]], dtype=float32) a = tf.constant([[1., 0.4, 0.5, 1.], [0.4, 0.2, 0.25, 2.], [0.5, 0.25, 0.35, 3.]]) tf.matmul(tfp.math.pinv(a), a) # ==> array([[ 0.76, 0.37, 0.21, -0.02], [ 0.37, 0.43, -0.33, 0.02], [ 0.21, -0.33, 0.81, 0.01], [-0.02, 0.02, 0.01, 1. ]], dtype=float32) ``` #### References [1]: G. Strang. 'Linear Algebra and Its Applications, 2nd Ed.' Academic Press, Inc., 1980, pp. 139-142. """ with tf.name_scope(name or 'pinv'): a = tf.convert_to_tensor(a, name='a') assertions = _maybe_validate_matrix(a, validate_args) if assertions: with tf.control_dependencies(assertions): a = tf.identity(a) dtype = dtype_util.as_numpy_dtype(a.dtype) if rcond is None: def get_dim_size(dim): if tf.compat.dimension_value(a.shape[dim]) is not None: return tf.compat.dimension_value(a.shape[dim]) return tf.shape(a)[dim] num_rows = get_dim_size(-2) num_cols = get_dim_size(-1) if isinstance(num_rows, int) and isinstance(num_cols, int): max_rows_cols = float(max(num_rows, num_cols)) else: max_rows_cols = tf.cast(tf.maximum(num_rows, num_cols), dtype) rcond = 10. * max_rows_cols * np.finfo(dtype).eps rcond = tf.convert_to_tensor(rcond, dtype=dtype, name='rcond') # Calculate pseudo inverse via SVD. # Note: if a is symmetric then u == v. (We might observe additional # performance by explicitly setting `v = u` in such cases.) [ singular_values, # Sigma left_singular_vectors, # U right_singular_vectors, # V ] = tf.linalg.svd(a, full_matrices=False, compute_uv=True) # Saturate small singular values to inf. This has the effect of make # `1. / s = 0.` while not resulting in `NaN` gradients. cutoff = rcond * tf.reduce_max(singular_values, axis=-1) singular_values = tf.where(singular_values > cutoff[..., tf.newaxis], singular_values, np.array(np.inf, dtype)) # Although `a == tf.matmul(u, s * v, transpose_b=True)` we swap # `u` and `v` here so that `tf.matmul(pinv(A), A) = tf.eye()`, i.e., # a matrix inverse has 'transposed' semantics. a_pinv = tf.matmul(right_singular_vectors / singular_values[..., tf.newaxis, :], left_singular_vectors, adjoint_b=True) if tensorshape_util.rank(a.shape) is not None: a_pinv.set_shape(a.shape[:-2].concatenate( [a.shape[-1], a.shape[-2]])) return a_pinv
def _sample_n(self, n, seed): batch_shape = self.batch_shape_tensor() event_shape = self.event_shape_tensor() batch_ndims = tf.shape(batch_shape)[0] ndims = batch_ndims + 3 # sample_ndims=1, event_ndims=2 shape = tf.concat([[n], batch_shape, event_shape], 0) stream = SeedStream(seed, salt="Wishart") # Complexity: O(nbk**2) x = tf.random.normal(shape=shape, mean=0., stddev=1., dtype=self.dtype, seed=stream()) # Complexity: O(nbk) # This parametrization is equivalent to Chi2, i.e., # ChiSquared(k) == Gamma(alpha=k/2, beta=1/2) expanded_df = self.df * tf.ones( self.scale_operator.batch_shape_tensor(), dtype=dtype_util.base_dtype(self.df.dtype)) g = tf.random.gamma(shape=[n], alpha=self._multi_gamma_sequence( 0.5 * expanded_df, self.dimension), beta=0.5, dtype=self.dtype, seed=stream()) # Complexity: O(nbk**2) x = tf.linalg.band_part(x, -1, 0) # Tri-lower. # Complexity: O(nbk) x = tf.linalg.set_diag(x, tf.sqrt(g)) # Make batch-op ready. # Complexity: O(nbk**2) perm = tf.concat([tf.range(1, ndims), [0]], 0) x = tf.transpose(a=x, perm=perm) shape = tf.concat( [batch_shape, [event_shape[0]], [event_shape[1] * n]], 0) x = tf.reshape(x, shape) # Complexity: O(nbM) where M is the complexity of the operator solving a # vector system. For LinearOperatorLowerTriangular, each matmul is O(k^3) so # this step has complexity O(nbk^3). x = self.scale_operator.matmul(x) # Undo make batch-op ready. # Complexity: O(nbk**2) shape = tf.concat([batch_shape, event_shape, [n]], 0) x = tf.reshape(x, shape) perm = tf.concat([[ndims - 1], tf.range(0, ndims - 1)], 0) x = tf.transpose(a=x, perm=perm) if not self.input_output_cholesky: # Complexity: O(nbk**3) x = tf.matmul(x, x, adjoint_b=True) return x
def _sample_n(self, num_samples, seed=None, name=None): """Returns a Tensor of samples from an LKJ distribution. Args: num_samples: Python `int`. The number of samples to draw. seed: Python integer seed for RNG name: Python `str` name prefixed to Ops created by this function. Returns: samples: A Tensor of correlation matrices with shape `[n, B, D, D]`, where `B` is the shape of the `concentration` parameter, and `D` is the `dimension`. Raises: ValueError: If `dimension` is negative. """ if self.dimension < 0: raise ValueError( 'Cannot sample negative-dimension correlation matrices.') # Notation below: B is the batch shape, i.e., tf.shape(concentration) seed = SeedStream(seed, 'sample_lkj') with tf.name_scope('sample_lkj' or name): concentration = tf.convert_to_tensor(self.concentration) if not dtype_util.is_floating(concentration.dtype): raise TypeError( 'The concentration argument should have floating type, not ' '{}'.format(dtype_util.name(concentration.dtype))) concentration = _replicate(num_samples, concentration) concentration_shape = tf.shape(concentration) if self.dimension <= 1: # For any dimension <= 1, there is only one possible correlation matrix. shape = tf.concat( [concentration_shape, [self.dimension, self.dimension]], axis=0) return tf.ones(shape=shape, dtype=concentration.dtype) beta_conc = concentration + (self.dimension - 2.) / 2. beta_dist = beta.Beta(concentration1=beta_conc, concentration0=beta_conc) # Note that the sampler below deviates from [1], by doing the sampling in # cholesky space. This does not change the fundamental logic of the # sampler, but does speed up the sampling. # This is the correlation coefficient between the first two dimensions. # This is also `r` in reference [1]. corr12 = 2. * beta_dist.sample(seed=seed()) - 1. # Below we construct the Cholesky of the initial 2x2 correlation matrix, # which is of the form: # [[1, 0], [r, sqrt(1 - r**2)]], where r is the correlation between the # first two dimensions. # This is the top-left corner of the cholesky of the final sample. first_row = tf.concat([ tf.ones_like(corr12)[..., tf.newaxis], tf.zeros_like(corr12)[..., tf.newaxis] ], axis=-1) second_row = tf.concat([ corr12[..., tf.newaxis], tf.sqrt(1 - corr12**2)[..., tf.newaxis] ], axis=-1) chol_result = tf.concat([ first_row[..., tf.newaxis, :], second_row[..., tf.newaxis, :] ], axis=-2) for n in range(2, self.dimension): # Loop invariant: on entry, result has shape B + [n, n] beta_conc = beta_conc - 0.5 # norm is y in reference [1]. norm = beta.Beta(concentration1=n / 2., concentration0=beta_conc).sample(seed=seed()) # distance shape: B + [1] for broadcast distance = tf.sqrt(norm)[..., tf.newaxis] # direction is u in reference [1]. # direction shape: B + [n] direction = _uniform_unit_norm(n, concentration_shape, concentration.dtype, seed) # raw_correlation is w in reference [1]. raw_correlation = distance * direction # shape: B + [n] # This is the next row in the cholesky of the result, # which differs from the construction in reference [1]. # In the reference, the new row `z` = chol_result @ raw_correlation^T # = C @ raw_correlation^T (where as short hand we use C = chol_result). # We prove that the below equation is the right row to add to the # cholesky, by showing equality with reference [1]. # Let S be the sample constructed so far, and let `z` be as in # reference [1]. Then at this iteration, the new sample S' will be # [[S z^T] # [z 1]] # In our case we have the cholesky decomposition factor C, so # we want our new row x (same size as z) to satisfy: # [[S z^T] [[C 0] [[C^T x^T] [[CC^T Cx^T] # [z 1]] = [x k]] [0 k]] = [xC^t xx^T + k**2]] # Since C @ raw_correlation^T = z = C @ x^T, and C is invertible, # we have that x = raw_correlation. Also 1 = xx^T + k**2, so k # = sqrt(1 - xx^T) = sqrt(1 - |raw_correlation|**2) = sqrt(1 - # distance**2). new_row = tf.concat( [raw_correlation, tf.sqrt(1. - norm[..., tf.newaxis])], axis=-1) # Finally add this new row, by growing the cholesky of the result. chol_result = tf.concat([ chol_result, tf.zeros_like(chol_result[..., 0][..., tf.newaxis]) ], axis=-1) chol_result = tf.concat( [chol_result, new_row[..., tf.newaxis, :]], axis=-2) if self.input_output_cholesky: return chol_result result = tf.matmul(chol_result, chol_result, transpose_b=True) # The diagonal for a correlation matrix should always be ones. Due to # numerical instability the matmul might not achieve that, so manually set # these to ones. result = tf.linalg.set_diag( result, tf.ones(shape=tf.shape(result)[:-1], dtype=result.dtype)) # This sampling algorithm can produce near-PSD matrices on which standard # algorithms such as `tf.cholesky` or `tf.linalg.self_adjoint_eigvals` # fail. Specifically, as documented in b/116828694, around 2% of trials # of 900,000 5x5 matrices (distributed according to 9 different # concentration parameter values) contained at least one matrix on which # the Cholesky decomposition failed. return result