From: Michael Orlitzky <michael@orlitzky.com>
Date: Sat, 9 Mar 2013 01:51:40 +0000 (-0500)
Subject: Add some step length functions, untested.
X-Git-Url: http://gitweb.michael.orlitzky.com/?p=octave.git;a=commitdiff_plain;h=69d4adf22828f26858e89de560078e1b5ed1714a;ds=sidebyside

Add some step length functions, untested.
---

diff --git a/optimization/step_length_cgm.m b/optimization/step_length_cgm.m
new file mode 100644
index 0000000..60f7cd4
--- /dev/null
+++ b/optimization/step_length_cgm.m
@@ -0,0 +1,40 @@
+function alpha = step_length_cgm(r, A, p)
+  ##
+  ## Compute the step length for the conjugate gradient method (CGM).
+  ## The CGM attempts to solve,
+  ##
+  ##   Ax = b
+  ##
+  ## or equivalently,
+  ##
+  ##   min[phi(x) = (1/2)<Ax,x> - <b,x>]
+  ##
+  ## where ``A`` is positive-definite. In the process, we need to
+  ## compute a number of search directions ``p`` and optimal step
+  ## lengths ``alpha``; i.e.,
+  ##
+  ##   x_{k+1} = x_{k} + alpha_{k}*p_{k}
+  ##
+  ## This function computes alpha_{k} in the formula above.
+  ##
+  ## INPUT:
+  ##
+  ##   - ``r`` -- The residual, Ax - b, at the current step.
+  ##
+  ##   - ``A`` -- The matrix ``A`` in the formulation above.
+  ##
+  ##   - ``p`` -- The current search direction.
+  ##
+  ## OUTPUT:
+  ##
+  ##   - ``alpha`` -- The minimizer of ``f(x) = x + alpha*p`` along ``p`.
+  ##
+  ## NOTES:
+  ##
+  ## All vectors are assumed to be *column* vectors.
+  ##
+
+  ## A simple calculation should convince you that the gradient of
+  ## phi(x) above is Ax - b == r.
+  alpha = step_length_positive_definite(r, A, p);
+end
diff --git a/optimization/step_length_positive_definite.m b/optimization/step_length_positive_definite.m
new file mode 100644
index 0000000..b1e5c48
--- /dev/null
+++ b/optimization/step_length_positive_definite.m
@@ -0,0 +1,36 @@
+function alpha = step_length_positive_definite(g, Q, p)
+  ##
+  ## Find the minimizer alpha of,
+  ##
+  ##   phi(alpha) = f(x + alpha*p)
+  ##
+  ## where ``p`` is a descent direction,
+  ##
+  ##   f(x) = (1/2)<Qx,x> - <b,x>
+  ##
+  ## and ``Q`` is positive-definite.
+  ##
+  ## The closed-form solution to this problem is given in Nocedal and
+  ## Wright, (3.55).
+  ##
+  ## INPUT:
+  ##
+  ##   - ``g`` -- The gradient of f.
+  ##
+  ##   - ``Q`` -- The positive-definite matrix in the definition of
+  ##     ``f`` above.
+  ##
+  ##   - ``p`` -- The direction in which ``f`` decreases. The line
+  ##     along which we minimize f(x + alpha*p).
+  ##
+  ## OUTPUT:
+  ##
+  ##   - ``alpha`` -- The value which causes ``f`` to decrease the
+  ##     most.
+  ##
+  ## NOTES:
+  ##
+  ## All vectors are assumed to be *column* vectors.
+  ##
+  alpha = -(g' * p)/(p' * Q * p)
+end