This Mathematica script generates the following picture indicating a typical behavior of steepest descent method: the zigzag in two perpendicular directions. In fact, in 2D this is the only case because one has only two choices, determined by starting direction, as I showed at the beginning. It is natural to ask whether one can use only two step with single change of direction to reach the minimization point. This introduces **conjugate direction method**.

Assume and the iteration starts with direction . The method of conjugate direction tries to reach a point along with condition that in the rest of iteration descent along direction would never happen again, so does for any other following steps. To make sure this happen, we can require that the iterations after always are performed on the subspace perpendicular to . Remembering that exact solution should always be in the subspace in which the coming iterations are to be performed, we come up this scheme:

If we view the above as two equations with three unknowns (), we know that it’s not going to work. However, since is SPD matrix, by Cholesky decomposition we have in which is an upper-triagonal matrix. Then if we consider the orthogonality between and , instead of and , we have

Finally we have

Of course this can be obtained by using the requirement as well, but derivation I used here gives better geometric explanation. In order to utilize the fact that we do not know but we know , we impose the orthogonality requirement in the space transformed by . By now one can see that to consider as the effect of applied on can often serve the similar purpose of the error itself, that’s why sometimes we call a similar name: **residual**.

The orthogonalilty condition is also called -orthogonality. And by is -orthogonal to the following one can easily see that is -orthogonal to the following . Since now we have the formular for , with a set of -orthogonal we will be done, and this set can be found by **Gram-Schimdt conjugation**, a -orthogonal version of Gram-Schimdt orthogonazation process. With those tools in hand, it is not hard to show that this conjugate direction method, whose name comes from that the directions are -orthogonal, or, -conjugate to each other, can indeed reach within steps. In fact, the goal in each interation in conjugate direction method achieved is to eliminate one dimension of the space in which the solution lies and confine the rest of search to a lower dimension hyperplane.