Week3: Rewrote Solvers to include a max_iter parameter

week3/solvers.py

@@ -1,113 +1,99 @@
 #!/usr/bin/env python3
 
 import numpy as np
+from itertools import count as count
 
 def diff(a, b):
     return np.amax(np.abs(a-b))
 
-def jacobi(A, b, eps):
+def jacobi(A, b, eps, max_iter = None):
+    """ Use the Jacobi Method to solve a Linear System. """
+
     A = np.array(A, dtype=np.float64)
     b = np.array(b, dtype=np.float64)
 
-
+    # Determine Diagonal and Upper and Lower matrices
     D = np.diag(A)
     L = -np.tril(A, -1)
     U = -np.triu(A, 1)
 
     D_inv = np.diagflat(np.reciprocal(D))
 
-    # initially x_f = x_(i-1)
+    x_0 = D_inv @ b
-    # this changes when in the loop
+    for i in count():
-    x_i = np.dot(D_inv, b)
+        x_1 = D_inv @ ( L + U) @ x_0
-    x_f = np.zeros(len(A))
-    k = 1
 
-    while diff(x_i, x_f) >= eps:
+        # Are we close enough?
-        k += 1
+        if diff(x_0, x_1) < eps:
+            return x_1, i
 
-        # Save the previous solution vector as x_f
+        # Running out of iterations
-        x_f = x_i
+        if max_iter is not None and max_iter >= i:
+            raise RuntimeError("Did not converge in {} steps".format(max_iter))
 
-        # Create new solution vector
+        # Set values for next loop
-        x_i = np.dot(np.dot(D_inv, ( L + U )), x_f ) + np.dot(D_inv, b)
+        x_0 = x_1
 
-    return x_i, k
+def steepest_descent(A, b, eps, max_iter = None):
-
+    """ Use Steepest Descent to solve a Linear System. """
-def steepest_descent(A, b, eps):
     A = np.array(A, dtype=np.float64)
     b = np.array(b, dtype=np.float64)
    
-    # initially x_f = x_(i-1)
-    # this changes when in the loop
-    x_f = np.zeros(len(A), dtype=np.float64)
-    k = 1
 
-    v_f = b
+    x_0 = np.zeros(len(A), dtype=np.float64)
-    t = np.dot(v_f, v_f) / np.dot(v_f, np.dot(A, v_f))
-    x_i = x_f + t*v_f
 
-    while diff(x_i, x_f) >= eps:
+    for i in count():
-        k += 1
+        Ax = A @ x_0
+        v = b - Ax
+        t = np.dot(v,v) / np.dot(v, A @ v )
 
-        # Pre calculate v_f and t
+        x_1 = x_0 + t*v
-        v_f = b - np.dot(A, x_i)
 
-        t = np.dot(v_f, v_f) / np.dot(v_f, np.dot(A, v_f))
+        # Are we close enough?
+        if diff(x_0, x_1) < eps:
+            return x_1, i
 
-        # Save the previous solution vector as x_f
+        # Running out of iterations
-        x_f = x_i
+        if max_iter is not None and max_iter >= i:
+            raise RuntimeError("Did not converge in {} steps".format(max_iter))
 
-        # Create new solution vector
+        # Set values for next loop
-        x_i = x_f + t * v_f
+        x_0 = x_1
 
-    return x_i, k
+def conjugate_gradient(A, b, eps, max_iter = None):
-
+    """ Use the Conjugate Gradient Method to solve a Linear System. """
-def conjugate_gradient(A, b, eps):
     A = np.array(A, dtype=np.float64)
     b = np.array(b, dtype=np.float64)
 
-    # initially x_f = x_(i-1)
-    # this changes when in the loop
-    x_f = np.zeros(len(A), dtype=np.float64)
-    r_f = b - np.dot(A, x_f)
-    v_f = r_f
 
-    k = 1
+    # Setup vectors
+    x_0 = np.zeros(len(A), dtype=np.float64)
+    r_0 = b - A @ x_0
+    v = r_0.copy()
 
-    # Calculate first iteration
+    for i in count():
-    t = np.dot(r_f, r_f) / np.dot(v_f, np.dot(A, v_f))
+        Ax = A @ x_0
-    x_i = x_f + t*v_f
+        Av = A @ v
         
-    r_i = r_f - t * np.dot(A, v_f)
+        r_0_square = np.dot(r_0, r_0)
-    s = np.dot(r_i, r_i) / np.dot(r_f, r_f)
 
-    v_i = r_i + s*v_f
+        t = r_0_square / np.dot(v, Av )
+        x_1 = x_0 + t*v
 
-    # Set r and v vectors for next loop
+        r_1 = r_0 - t * Av
-    r_f = r_i
-    v_f = v_i
 
-    while diff(x_i, x_f) >= eps:
+        s = np.dot(r_1, r_1) / r_0_square
-        k += 1
+        v = r_1 + s*v
-        
-        t = np.dot(r_f, r_f) / np.dot(v_f, np.dot(A, v_f))
-        # Save the previous solution vector as x_f
-        x_f = x_i
-
-        # Create new solution vector
-        x_i = x_f + t*v_f
-
-        # Calculate r and v vectors
-
-        r_i = r_f - t * np.dot(A, v_f)
-        s = np.dot(r_i, r_i) / np.dot(r_f, r_f)
-        v_i = r_i + s*v_f
-        
-        # Save r and v vectors for next loop
-        r_f = r_i
-        v_f = v_i
 
 
-    return x_i, k
+        # Are we close enough?
+        if diff(x_0, x_1) < eps:
+            return x_1, i
 
+        # Running out of iterations
+        if max_iter is not None and max_iter >= i:
+            raise RuntimeError("Did not converge in {} steps".format(max_iter))
 
+        # Set values for next loop
+        x_0 = x_1
+        r_0 = r_1