Cleanup for Optimal Control Ops (#1045)

* Blockwise optimal linear control ops

* Add jax rewrite to eliminate `BilinearSolveDiscreteLyapunov`

* set `solve_discrete_lyapunov` method default to bilinear

* Appease mypy

* restore method dispatching

* Use `pt.vectorize` on base `solve_discrete_lyapunov` case

* Apply JAX rewrite before canonicalization

* Improve tests

* Remove useless warning filters

* Fix local_blockwise_alloc rewrite

The rewrite was squeezing too many dimensions of the alloced value, when this didn't have dummy expand dims to the left.

* Fix float32 tests

* Test against complex inputs

* Appease ViPy (Vieira-py type checking)

* Remove condition from `TensorLike` import

* Infer dtype from `node.outputs.type.dtype`

* Remove unused mypy ignore

* Don't manually set dtype of output

Revert change to `_solve_discrete_lyapunov`

* Set dtype of Op outputs

---------

Co-authored-by: ricardoV94 <[email protected]>

Author: jessegrabowski

pytensor/tensor/rewriting/blockwise.py

@@ -127,8 +127,8 @@ def local_blockwise_alloc(fgraph, node):
                 value, *shape = inp.owner.inputs
 
                 # Check what to do with the value of the Alloc
-                squeezed_value = _squeeze_left(value, batch_ndim)
-                missing_ndim = len(shape) - value.type.ndim
+                missing_ndim = inp.type.ndim - value.type.ndim
+                squeezed_value = _squeeze_left(value, (batch_ndim - missing_ndim))
                 if (
                     (((1,) * missing_ndim + value.type.broadcastable)[batch_ndim:])
                     != inp.type.broadcastable[batch_ndim:]

pytensor/tensor/rewriting/linalg.py

@@ -4,9 +4,11 @@
 
 from pytensor import Variable
 from pytensor import tensor as pt
+from pytensor.compile import optdb
 from pytensor.graph import Apply, FunctionGraph
 from pytensor.graph.rewriting.basic import (
     copy_stack_trace,
+    in2out,
     node_rewriter,
 )
 from pytensor.scalar.basic import Mul
@@ -45,9 +47,11 @@
     Cholesky,
     Solve,
     SolveBase,
+    _bilinear_solve_discrete_lyapunov,
     block_diag,
     cholesky,
     solve,
+    solve_discrete_lyapunov,
     solve_triangular,
 )
 
@@ -966,3 +970,22 @@ def rewrite_cholesky_diag_to_sqrt_diag(fgraph, node):
             non_eye_input = pt.shape_padaxis(non_eye_input, -2)
 
     return [eye_input * (non_eye_input**0.5)]
+
+
+@node_rewriter([_bilinear_solve_discrete_lyapunov])
+def jax_bilinaer_lyapunov_to_direct(fgraph: FunctionGraph, node: Apply):
+    """
+    Replace BilinearSolveDiscreteLyapunov with a direct computation that is supported by JAX
+    """
+    A, B = (cast(TensorVariable, x) for x in node.inputs)
+    result = solve_discrete_lyapunov(A, B, method="direct")
+
+    return [result]
+
+
+optdb.register(
+    "jax_bilinaer_lyapunov_to_direct",
+    in2out(jax_bilinaer_lyapunov_to_direct),
+    "jax",
+    position=0.9,  # Run before canonicalization
+)

pytensor/tensor/slinalg.py

@@ -2,7 +2,7 @@
 import typing
 import warnings
 from functools import reduce
-from typing import TYPE_CHECKING, Literal, cast
+from typing import Literal, cast
 
 import numpy as np
 import scipy.linalg
@@ -11,7 +11,7 @@
 import pytensor.tensor as pt
 from pytensor.graph.basic import Apply
 from pytensor.graph.op import Op
-from pytensor.tensor import as_tensor_variable
+from pytensor.tensor import TensorLike, as_tensor_variable
 from pytensor.tensor import basic as ptb
 from pytensor.tensor import math as ptm
 from pytensor.tensor.blockwise import Blockwise
@@ -21,9 +21,6 @@
 from pytensor.tensor.variable import TensorVariable
 
 
-if TYPE_CHECKING:
-    from pytensor.tensor import TensorLike
-
 logger = logging.getLogger(__name__)
 
 
@@ -777,7 +774,16 @@ def perform(self, node, inputs, outputs):
 
 
 class SolveContinuousLyapunov(Op):
+    """
+    Solves a continuous Lyapunov equation, :math:`AX + XA^H = B`, for :math:`X.
+
+    Continuous time Lyapunov equations are special cases of Sylvester equations, :math:`AX + XB = C`, and can be solved
+    efficiently using the Bartels-Stewart algorithm. For more details, see the docstring for
+    scipy.linalg.solve_continuous_lyapunov
+    """
+
     __props__ = ()
+    gufunc_signature = "(m,m),(m,m)->(m,m)"
 
     def make_node(self, A, B):
         A = as_tensor_variable(A)
@@ -792,7 +798,8 @@ def perform(self, node, inputs, output_storage):
         (A, B) = inputs
         X = output_storage[0]
 
-        X[0] = scipy.linalg.solve_continuous_lyapunov(A, B)
+        out_dtype = node.outputs[0].type.dtype
+        X[0] = scipy.linalg.solve_continuous_lyapunov(A, B).astype(out_dtype)
 
     def infer_shape(self, fgraph, node, shapes):
         return [shapes[0]]
@@ -813,7 +820,41 @@ def grad(self, inputs, output_grads):
         return [A_bar, Q_bar]
 
 
+_solve_continuous_lyapunov = Blockwise(SolveContinuousLyapunov())
+
+
+def solve_continuous_lyapunov(A: TensorLike, Q: TensorLike) -> TensorVariable:
+    """
+    Solve the continuous Lyapunov equation :math:`A X + X A^H + Q = 0`.
+
+    Parameters
+    ----------
+    A: TensorLike
+        Square matrix of shape ``N x N``.
+    Q: TensorLike
+        Square matrix of shape ``N x N``.
+
+    Returns
+    -------
+    X: TensorVariable
+        Square matrix of shape ``N x N``
+
+    """
+
+    return cast(TensorVariable, _solve_continuous_lyapunov(A, Q))
+
+
 class BilinearSolveDiscreteLyapunov(Op):
+    """
+    Solves a discrete lyapunov equation, :math:`AXA^H - X = Q`, for :math:`X.
+
+    The solution is computed by first transforming the discrete-time problem into a continuous-time form. The continuous
+    time lyapunov is a special case of a Sylvester equation, and can be efficiently solved. For more details, see the
+    docstring for scipy.linalg.solve_discrete_lyapunov
+    """
+
+    gufunc_signature = "(m,m),(m,m)->(m,m)"
+
     def make_node(self, A, B):
         A = as_tensor_variable(A)
         B = as_tensor_variable(B)
@@ -827,7 +868,10 @@ def perform(self, node, inputs, output_storage):
         (A, B) = inputs
         X = output_storage[0]
 
-        X[0] = scipy.linalg.solve_discrete_lyapunov(A, B, method="bilinear")
+        out_dtype = node.outputs[0].type.dtype
+        X[0] = scipy.linalg.solve_discrete_lyapunov(A, B, method="bilinear").astype(
+            out_dtype
+        )
 
     def infer_shape(self, fgraph, node, shapes):
         return [shapes[0]]
@@ -849,83 +893,83 @@ def grad(self, inputs, output_grads):
         return [A_bar, Q_bar]
 
 
-_solve_continuous_lyapunov = SolveContinuousLyapunov()
-_solve_bilinear_direct_lyapunov = cast(typing.Callable, BilinearSolveDiscreteLyapunov())
+_bilinear_solve_discrete_lyapunov = Blockwise(BilinearSolveDiscreteLyapunov())
 
 
-def _direct_solve_discrete_lyapunov(A: "TensorLike", Q: "TensorLike") -> TensorVariable:
-    A_ = as_tensor_variable(A)
-    Q_ = as_tensor_variable(Q)
+def _direct_solve_discrete_lyapunov(
+    A: TensorVariable, Q: TensorVariable
+) -> TensorVariable:
+    r"""
+    Directly solve the discrete Lyapunov equation :math:`A X A^H - X = Q` using the kronecker method of Magnus and
+    Neudecker.
+
+    This involves constructing and inverting an intermediate matrix :math:`A \otimes A`, with shape :math:`N^2 x N^2`.
+    As a result, this method scales poorly with the size of :math:`N`, and should be avoided for large :math:`N`.
+    """
 
-    if "complex" in A_.type.dtype:
-        AA = kron(A_, A_.conj())
+    if A.type.dtype.startswith("complex"):
+        AxA = kron(A, A.conj())
     else:
-        AA = kron(A_, A_)
+        AxA = kron(A, A)
+
+    eye = pt.eye(AxA.shape[-1])
 
-    X = solve(pt.eye(AA.shape[0]) - AA, Q_.ravel())
-    return cast(TensorVariable, reshape(X, Q_.shape))
+    vec_Q = Q.ravel()
+    vec_X = solve(eye - AxA, vec_Q, b_ndim=1)
+
+    return cast(TensorVariable, reshape(vec_X, A.shape))
 
 
 def solve_discrete_lyapunov(
-    A: "TensorLike", Q: "TensorLike", method: Literal["direct", "bilinear"] = "direct"
+    A: TensorLike,
+    Q: TensorLike,
+    method: Literal["direct", "bilinear"] = "bilinear",
 ) -> TensorVariable:
     """Solve the discrete Lyapunov equation :math:`A X A^H - X = Q`.
 
     Parameters
     ----------
-    A
-        Square matrix of shape N x N; must have the same shape as Q
-    Q
-        Square matrix of shape N x N; must have the same shape as A
-    method
-        Solver method used, one of ``"direct"`` or ``"bilinear"``. ``"direct"``
-        solves the problem directly via matrix inversion.  This has a pure
-        PyTensor implementation and can thus be cross-compiled to supported
-        backends, and should be preferred when ``N`` is not large. The direct
-        method scales poorly with the size of ``N``, and the bilinear can be
+    A: TensorLike
+        Square matrix of shape N x N
+    Q: TensorLike
+        Square matrix of shape N x N
+    method: str, one of ``"direct"`` or ``"bilinear"``
+        Solver method used, . ``"direct"`` solves the problem directly via matrix inversion.  This has a pure
+        PyTensor implementation and can thus be cross-compiled to supported backends, and should be preferred when
+         ``N`` is not large. The direct method scales poorly with the size of ``N``, and the bilinear can be
         used in these cases.
 
     Returns
     -------
-        Square matrix of shape ``N x N``, representing the solution to the
-        Lyapunov equation
+    X: TensorVariable
+        Square matrix of shape ``N x N``. Solution to the Lyapunov equation
 
     """
     if method not in ["direct", "bilinear"]:
         raise ValueError(
             f'Parameter "method" must be one of "direct" or "bilinear", found {method}'
         )
 
-    if method == "direct":
-        return _direct_solve_discrete_lyapunov(A, Q)
-    if method == "bilinear":
-        return cast(TensorVariable, _solve_bilinear_direct_lyapunov(A, Q))
-
-
-def solve_continuous_lyapunov(A: "TensorLike", Q: "TensorLike") -> TensorVariable:
-    """Solve the continuous Lyapunov equation :math:`A X + X A^H + Q = 0`.
-
-    Parameters
-    ----------
-    A
-        Square matrix of shape ``N x N``; must have the same shape as `Q`.
-    Q
-        Square matrix of shape ``N x N``; must have the same shape as `A`.
+    A = as_tensor_variable(A)
+    Q = as_tensor_variable(Q)
 
-    Returns
-    -------
-        Square matrix of shape ``N x N``, representing the solution to the
-        Lyapunov equation
+    if method == "direct":
+        signature = BilinearSolveDiscreteLyapunov.gufunc_signature
+        X = pt.vectorize(_direct_solve_discrete_lyapunov, signature=signature)(A, Q)
+        return cast(TensorVariable, X)
 
-    """
+    elif method == "bilinear":
+        return cast(TensorVariable, _bilinear_solve_discrete_lyapunov(A, Q))
 
-    return cast(TensorVariable, _solve_continuous_lyapunov(A, Q))
+    else:
+        raise ValueError(f"Unknown method {method}")
 
 
-class SolveDiscreteARE(pt.Op):
+class SolveDiscreteARE(Op):
     __props__ = ("enforce_Q_symmetric",)
+    gufunc_signature = "(m,m),(m,n),(m,m),(n,n)->(m,m)"
 
-    def __init__(self, enforce_Q_symmetric=False):
+    def __init__(self, enforce_Q_symmetric: bool = False):
         self.enforce_Q_symmetric = enforce_Q_symmetric
 
     def make_node(self, A, B, Q, R):
@@ -946,9 +990,8 @@ def perform(self, node, inputs, output_storage):
         if self.enforce_Q_symmetric:
             Q = 0.5 * (Q + Q.T)
 
-        X[0] = scipy.linalg.solve_discrete_are(A, B, Q, R).astype(
-            node.outputs[0].type.dtype
-        )
+        out_dtype = node.outputs[0].type.dtype
+        X[0] = scipy.linalg.solve_discrete_are(A, B, Q, R).astype(out_dtype)
 
     def infer_shape(self, fgraph, node, shapes):
         return [shapes[0]]
@@ -960,14 +1003,16 @@ def grad(self, inputs, output_grads):
         (dX,) = output_grads
         X = self(A, B, Q, R)
 
-        K_inner = R + pt.linalg.matrix_dot(B.T, X, B)
-        K_inner_inv = pt.linalg.solve(K_inner, pt.eye(R.shape[0]))
-        K = matrix_dot(K_inner_inv, B.T, X, A)
+        K_inner = R + matrix_dot(B.T, X, B)
+
+        # K_inner is guaranteed to be symmetric, because X and R are symmetric
+        K_inner_inv_BT = solve(K_inner, B.T, assume_a="sym")
+        K = matrix_dot(K_inner_inv_BT, X, A)
 
         A_tilde = A - B.dot(K)
 
         dX_symm = 0.5 * (dX + dX.T)
-        S = solve_discrete_lyapunov(A_tilde, dX_symm).astype(dX.type.dtype)
+        S = solve_discrete_lyapunov(A_tilde, dX_symm)
 
         A_bar = 2 * matrix_dot(X, A_tilde, S)
         B_bar = -2 * matrix_dot(X, A_tilde, S, K.T)
@@ -977,30 +1022,45 @@ def grad(self, inputs, output_grads):
         return [A_bar, B_bar, Q_bar, R_bar]
 
 
-def solve_discrete_are(A, B, Q, R, enforce_Q_symmetric=False) -> TensorVariable:
+def solve_discrete_are(
+    A: TensorLike,
+    B: TensorLike,
+    Q: TensorLike,
+    R: TensorLike,
+    enforce_Q_symmetric: bool = False,
+) -> TensorVariable:
     """
     Solve the discrete Algebraic Riccati equation :math:`A^TXA - X - (A^TXB)(R + B^TXB)^{-1}(B^TXA) + Q = 0`.
 
+    Discrete-time Algebraic Riccati equations arise in the context of optimal control and filtering problems, as the
+    solution to Linear-Quadratic Regulators (LQR), Linear-Quadratic-Guassian (LQG) control problems, and as the
+    steady-state covariance of the Kalman Filter.
+
+    Such problems typically have many solutions, but we are generally only interested in the unique *stabilizing*
+    solution. This stable solution, if it exists, will be returned by this function.
+
     Parameters
     ----------
-    A: ArrayLike
+    A: TensorLike
         Square matrix of shape M x M
-    B: ArrayLike
+    B: TensorLike
         Square matrix of shape M x M
-    Q: ArrayLike
+    Q: TensorLike
         Symmetric square matrix of shape M x M
-    R: ArrayLike
+    R: TensorLike
         Square matrix of shape N x N
     enforce_Q_symmetric: bool
         If True, the provided Q matrix is transformed to 0.5 * (Q + Q.T) to ensure symmetry
 
     Returns
     -------
-    X: pt.matrix
+    X: TensorVariable
         Square matrix of shape M x M, representing the solution to the DARE
     """
 
-    return cast(TensorVariable, SolveDiscreteARE(enforce_Q_symmetric)(A, B, Q, R))
+    return cast(
+        TensorVariable, Blockwise(SolveDiscreteARE(enforce_Q_symmetric))(A, B, Q, R)
+    )
 
 
 def _largest_common_dtype(tensors: typing.Sequence[TensorVariable]) -> np.dtype: