Avoid divide-by-zero in preconditioned_conjugate_gradient_method().

[octave.git] / optimization / preconditioned_conjugate_gradient_method.m

diff --git a/optimization/preconditioned_conjugate_gradient_method.m b/optimization/preconditioned_conjugate_gradient_method.m
index eb2089f5b4e50d367c8237be9c166d5065f70982..cd886084a9766e1d785df59185c358ae69d75cc7 100644 (file)
--- a/optimization/preconditioned_conjugate_gradient_method.m
+++ b/optimization/preconditioned_conjugate_gradient_method.m
@@ -1,8 +1,8 @@
-function [x, k] = preconditioned_conjugate_gradient_method(Q,
-                                                          M,
-                                                          b,
-                                                          x0,
-                                                          tolerance,
+function [x, k] = preconditioned_conjugate_gradient_method(Q, ...
+                                                          M, ...
+                                                          b, ...
+                                                          x0, ...
+                                                          tolerance, ...
                                                           max_iterations)
   %
   % Solve,
@@ -58,8 +58,15 @@ function [x, k] = preconditioned_conjugate_gradient_method(Q,
   % REFERENCES:
   %
   %   1. Guler, Osman. Foundations of Optimization. New York, Springer,
-  %   2010.
+  %      2010.
   %
+  %   2. Shewchuk, Jonathan Richard. An Introduction to the Conjugate
+  %      Gradient Method Without the Agonizing Pain, Edition 1.25.
+  %      August 4, 1994.
+  %
+
+  % We use this in the inner loop.
+  sqrt_n = floor(sqrt(length(x0)));
 
   % Set k=0 first, that way the references to xk,rk,zk,dk which
   % immediately follow correspond (semantically) to x0,r0,z0,d0.
@@ -70,14 +77,7 @@ function [x, k] = preconditioned_conjugate_gradient_method(Q,
   zk = M \ rk;
   dk = -zk;
 
-  for k = [ 0 : max_iterations ]
-
-    if (norm(rk) < tolerance)
-      % Check our stopping condition. This should catch the k=0 case.
-      x = xk;
-      return;
-    end
-
+  while (k <= max_iterations && norm(rk, 'inf') > tolerance)
     % Used twice, avoid recomputation.
     rkzk = rk' * zk;
 
@@ -85,15 +85,36 @@ function [x, k] = preconditioned_conjugate_gradient_method(Q,
     % do them both, so we precompute the more expensive operation.
     Qdk = Q * dk;
 
-    % After substituting the two previously-created variables, the
-    % following algorithm occurs verbatim in the reference.
-    alpha_k = rkzk/(dk' * Qdk);
+    % We're going to divide by this quantity...
+    dkQdk = dk' * Qdk;
+
+    % So if it's too close to zero, we replace it with something
+    % comparable but non-zero.
+    if (abs(dkQdk) < eps)
+      dkQdk = sign(dkQdk)*eps;
+    end
+
+    alpha_k = rkzk/dkQdk;
     x_next = xk + (alpha_k * dk);
-    r_next = rk + (alpha_k * Qdk);
+
+    % The recursive definition of r_next is prone to accumulate
+    % roundoff error. When sqrt(n) divides k, we recompute the
+    % residual to minimize this error. This modification was suggested
+    % by the second reference.
+    if (mod(k, sqrt_n) == 0)
+      r_next = Q*x_next - b;
+    else
+      r_next = rk + (alpha_k * Qdk);
+    end
+
     z_next = M \ r_next;
     beta_next = (r_next' * z_next)/rkzk;
     d_next = -z_next + beta_next*dk;
 
+    % We potentially just performed one more iteration than necessary
+    % in order to simplify the loop. Note that due to the structure of
+    % our loop, we will have k > max_iterations when we fail to
+    % converge.
     k = k + 1;
     xk = x_next;
     rk = r_next;
@@ -101,7 +122,6 @@ function [x, k] = preconditioned_conjugate_gradient_method(Q,
     dk = d_next;
   end
 
-  % The algorithm didn't converge, but we still want to return the
-  % terminal value of xk.
+  % If we make it here, one of the two stopping conditions was met.
   x = xk;
 end