Inverse iteration

Let us analyze the rate of convergence of the method.

The power method is known to converge linearly to the limit, more precisely:

$${\displaystyle \mathrm {Distance} (b^{\mathrm {ideal} },b_{\mathrm {Power~Method} }^{k})=O\left(\left|{\frac {\lambda _{\mathrm {subdominant} }}{\lambda _{\mathrm {dominant} }}}\right|^{k}\right),}$$

hence for the inverse iteration method similar result sounds as:

$${\displaystyle \mathrm {Distance} (b^{\mathrm {ideal} },b_{\mathrm {Inverse~iteration} }^{k})=O\left(\left|{\frac {\mu -\lambda _{\mathrm {closest~to~} \mu }}{\mu -\lambda _{\mathrm {second~closest~to~} \mu }}}\right|^{k}\right).}$$

This is a key formula for understanding the method's convergence. It shows that if ${\displaystyle \mu }$ is chosen close enough to some eigenvalue ${\displaystyle \lambda }$, for example ${\displaystyle \mu -\lambda =\epsilon }$ each iteration will improve the accuracy ${\displaystyle |\epsilon |/|\lambda +\epsilon -\lambda _{\mathrm {closest~to~} \lambda }|}$ times. (We use that for small enough ${\displaystyle \epsilon }$ "closest to ${\displaystyle \mu }$" and "closest to ${\displaystyle \lambda }$" is the same.) For small enough ${\displaystyle |\epsilon |}$ it is approximately the same as ${\displaystyle |\epsilon |/|\lambda -\lambda _{{\text{closest to }}\lambda }|}$. Hence if one is able to find ${\displaystyle \mu }$, such that the ${\displaystyle \epsilon }$ will be small enough, then very few iterations may be satisfactory.

The inverse iteration algorithm requires solving a linear system or calculation of the inverse matrix. For non-structured matrices (not sparse, not Toeplitz,...) this requires ${\displaystyle O(n^{3})}$ operations.

We can rewrite the formula in the following way:

$${\displaystyle (A-\mu I)b_{k+1}={\frac {b_{k}}{C_{k}}},}$$

emphasizing that to find the next approximation ${\displaystyle b_{k+1}}$ we may solve a system of linear equations. There are two options: one may choose an algorithm that solves a linear system, or one may calculate the inverse ${\displaystyle (A-\mu I)^{-1}}$ and then apply it to the vector. Both options have complexity O(n³), the exact number depends on the chosen method.

The choice depends also on the number of iterations. Naively, if at each iteration one solves a linear system, the complexity will be k O(n³), where k is number of iterations; similarly, calculating the inverse matrix and applying it at each iteration is of complexity k O(n³). Note, however, that if the eigenvalue estimate ${\displaystyle \mu }$ remains constant, then we may reduce the complexity to O(n³) + k O(n²) with either method. Calculating the inverse matrix once, and storing it to apply at each iteration is of complexity O(n³) + k O(n²). Storing an LU decomposition of ${\displaystyle (A-\mu I)}$ and using forward and back substitution to solve the system of equations at each iteration is also of complexity O(n³) + k O(n²).

Inverting the matrix will typically have a greater initial cost, but lower cost at each iteration. Conversely, solving systems of linear equations will typically have a lesser initial cost, but require more operations for each iteration.

If it is necessary to perform many iterations (or few iterations, but for many eigenvectors), then it might be wise to bring the matrix to the upper Hessenberg form first (for symmetric matrix this will be tridiagonal form). Which costs ${\textstyle {\frac {10}{3}}n^{3}+O(n^{2})}$ arithmetic operations using a technique based on Householder reduction), with a finite sequence of orthogonal similarity transforms, somewhat like a two-sided QR decomposition.[2][3] (For QR decomposition, the Householder rotations are multiplied only on the left, but for the Hessenberg case they are multiplied on both left and right.) For symmetric matrices this procedure costs ${\textstyle {\frac {4}{3}}n^{3}+O(n^{2})}$ arithmetic operations using a technique based on Householder reduction.[2][3]

Solution of the system of linear equations for the tridiagonal matrix costs ${\displaystyle O(n)}$ operations, so the complexity grows like ${\displaystyle O(n^{3})+kO(n)}$, where ${\displaystyle k}$ is the iteration number, which is better than for the direct inversion. However, for few iterations such transformation may not be practical.

Also transformation to the Hessenberg form involves square roots and the division operation, which are not universally supported by hardware.

On general purpose processors (e.g. produced by Intel) the execution time of addition, multiplication and division is approximately equal. But on embedded and/or low energy consuming hardware (digital signal processors, FPGA, ASIC) division may not be supported by hardware, and so should be avoided. Choosing ${\displaystyle C_{k}=2^{n_{k}}}$ allows fast division without explicit hardware support, as division by a power of 2 may be implemented as either a bit shift (for fixed-point arithmetic) or subtraction of ${\displaystyle k}$ from the exponent (for floating-point arithmetic).

When implementing the algorithm using fixed-point arithmetic, the choice of the constant ${\displaystyle C_{k}}$ is especially important. Small values will lead to fast growth of the norm of ${\displaystyle b_{k}}$ and to overflow; large values of ${\displaystyle C_{k}}$ will cause the vector ${\displaystyle b_{k}}$ to tend toward zero.

Typically, the method is used in combination with some other method which finds approximate eigenvalues: the standard example is the bisection eigenvalue algorithm, another example is the Rayleigh quotient iteration, which is actually the same inverse iteration with the choice of the approximate eigenvalue as the Rayleigh quotient corresponding to the vector obtained on the previous step of the iteration.

There are some situations where the method can be used by itself, however they are quite marginal.

The dominant eigenvalue can be easily estimated for any matrix. For any induced norm it is true that ${\displaystyle \left\|A\right\|\geq |\lambda |,}$ for any eigenvalue ${\displaystyle \lambda }$. So taking the norm of the matrix as an approximate eigenvalue one can see that the method will converge to the dominant eigenvector.

In some real-time applications one needs to find eigenvectors for matrices with a speed of millions of matrices per second. In such applications, typically the statistics of matrices is known in advance and one can take as an approximate eigenvalue the average eigenvalue for some large matrix sample. Better, one may calculate the mean ratio of the eigenvalues to the trace or the norm of the matrix and estimate the average eigenvalue as the trace or norm multiplied by the average value of that ratio. Clearly such a method can be used only with discretion and only when high precision is not critical. This approach of estimating an average eigenvalue can be combined with other methods to avoid excessively large error.