Pass associative scan arg through linalg funcs (#34)

* Pass associative scan arg through linalg funcs

* Fix style

* Add docstring for assoc scan arg

Author: denismelanson

thermox/linalg.py

@@ -12,6 +12,7 @@ def solve(
     dt: float = 1.0,
     burnin: int = 0,
     key: Array = None,
+    associative_scan: bool = True,
 ) -> Array:
     """
     Obtain the solution of the linear system
@@ -29,6 +30,7 @@ def solve(
         burnin: Time-step index corresponding to the end of the burn-in period.
             Samples before this step are not collected.
         key: JAX random key
+        associative_scan: If True, uses jax.lax.associative_scan.
 
     Returns:
         Approximate solution, x, of the linear system.
@@ -37,7 +39,9 @@ def solve(
         key = random.PRNGKey(0)
     ts = jnp.arange(burnin, burnin + num_samples) * dt
     x0 = jnp.zeros_like(b)
-    samples = sample_identity_diffusion(key, ts, x0, A, jnp.linalg.solve(A, b))
+    samples = sample_identity_diffusion(
+        key, ts, x0, A, jnp.linalg.solve(A, b), associative_scan
+    )
     return jnp.mean(samples, axis=0)
 
 
@@ -47,6 +51,7 @@ def inv(
     dt: float = 1.0,
     burnin: int = 0,
     key: Array = None,
+    associative_scan: bool = True,
 ) -> Array:
     """
     Obtain the inverse of a matrix A by
@@ -60,6 +65,7 @@ def inv(
         burnin: Time-step index corresponding to the end of the burn-in period.
             Samples before this step are not collected.
         key: JAX random key
+        associative_scan: If True, uses jax.lax.associative_scan.
 
     Returns:
         Approximate inverse of A.
@@ -69,7 +75,7 @@ def inv(
     ts = jnp.arange(burnin, burnin + num_samples) * dt
     b = jnp.zeros(A.shape[0])
     x0 = jnp.zeros_like(b)
-    samples = sample(key, ts, x0, A, b, 2 * jnp.eye(A.shape[0]))
+    samples = sample(key, ts, x0, A, b, 2 * jnp.eye(A.shape[0]), associative_scan)
     return jnp.cov(samples.T)
 
 
@@ -80,6 +86,7 @@ def expnegm(
     burnin: int = 0,
     key: Array = None,
     alpha: float = 0.0,
+    associative_scan: bool = True,
 ) -> Array:
     """
     Obtain the negative exponential of a matrix A by
@@ -95,6 +102,7 @@ def expnegm(
         key: JAX random key
         alpha: Regularization parameter to ensure diffusion matrix
             is symmetric positive definite.
+        associative_scan: If True, uses jax.lax.associative_scan.
 
     Returns:
         Approximate negative matrix exponential, exp(-A).
@@ -108,7 +116,7 @@ def expnegm(
     ts = jnp.arange(burnin, burnin + num_samples) * dt
     b = jnp.zeros(A.shape[0])
     x0 = jnp.zeros_like(b)
-    samples = sample(key, ts, x0, A_shifted, b, B)
+    samples = sample(key, ts, x0, A_shifted, b, B, associative_scan)
     return autocovariance(samples) * jnp.exp(alpha)
 
 
@@ -119,6 +127,7 @@ def expm(
     burnin: int = 0,
     key: Array = None,
     alpha: float = 1.0,
+    associative_scan: bool = True,
 ) -> Array:
     """
     Obtain the exponential of a matrix A by
@@ -134,11 +143,12 @@ def expm(
         key: JAX random key
         alpha: Regularization parameter to ensure diffusion matrix
             is symmetric positive definite.
+        associative_scan: If True, uses jax.lax.associative_scan.
 
     Returns:
         Approximate matrix exponential, exp(A).
     """
-    return expnegm(-A, num_samples, dt, burnin, key, alpha)
+    return expnegm(-A, num_samples, dt, burnin, key, alpha, associative_scan)
 
 
 def autocovariance(samples: Array) -> Array: