I have a question about the conjugate gradient (CG) algorithm in particular, and possibly about interative solvers in general. My main reference here will be section 5.1 of Numerical Optimization by Nocedal and Wright (hereafter NW).

Let \(H\) be a symmetric \(D \times D\) positive definite matrix. All lower-case letters will denote \(D\)-vectors. The conjugate gradient algorithm solves the linear system

\[x = H^{-1} a,\]

using only a black-box routine to compute the product \(Hv\) for any vector \(v\). To provide some intuition for how this is possible, note that by defining

\[\phi(x) := \frac{1}{2} x^T H x - x^T a,\]

the optimization problem \(\mathrm{argmin}_x \phi(x)\) has the unique solution \(H^{-1} a\) (cf NW Equation 5.2). Furthermore, the gradient of \(phi\) is given by

\[\nabla \phi(x) = Hx - a,\]

evaluation of which requires only the matrix-vector product. Consequently, one can perfor gradient descent to solve \(\mathrm{argmin}_x \phi(x)\) using only matrix-vector products. (In fact, this is nearly what CG does— CG performs gradient descent on \(\phi(x)\) with the constraint that the search directions are all mutually conjugate with respect to \(H\).)

Now, suppose that you are not actually interested in the full vector \(x\), but only in the value of \(x\) in some subspace expressed by an inner product, \(s^T x\), where \(s\) is a unit vector. (This is typically the case in sensitivity anaylysis, which is the source of my interest in this question.) Is there a way to monitor the progress of CG for the quantity \(s^T x\)? No matter how you look at it, it seems to me that CG is not set up to answer such a question, especially when \(s\) is aligned with a small eigenvalue of \(H\).

Typically, one monitors the convergence of CG using the residual \(Hx - a\). Of course, the value of \(s^T (Hx - a)\) tells you little about \(s^T x\) in general. The theoretical guanrantees for CG (as given, for example, in NW Theorem 5.5) are given in terms of the norm \(\left\| v\right\|_H^2 := v^T H v\). If \(Hs\) is very small, then it is possible for \(\left\| v\right\|_H^2\) to be small but for \(s^T v\) to be very large. Similarly, it is possible for the optimization objective \(\phi(\hat x)\) to be near \(0\) while \(s^T (\hat x - H^{-1} a)\) is still very large.

It might be pointed out that calculating \(s^T H^{-1} a\) is equal to \(a^T H^{-1} s\), so that the roles of \(a\) and \(s\) are symmetric here.

What I believe this comes down to is the obvious-enough fact that an iterative method based on products \(Hv\) will not do well for directions \(s\) in which \(Hs\) is small. Of course, in theory and practice we expect our linear solvers to work best when the condition number is not too large. However, I would be very interested to know whether there is a practical variant of CG that provides a guide for terminating the algorithm if the quantity of interest is \(s^T H^{-1}a\) and not \(Hx - a\).