Matrix Exponential Fréchet Sensitivity

The matrix exponential, exp(A), generalizes the ordinary exponential function to matrices. A critical question is how the matrix exponential responds to a small perturbation in a particular direction E. We define the Fréchet derivative of the matrix exponential at A in the direction E as the linear map L such that:

exp(A + \u03B5 E) \u2248 exp(A) + \u03B5 L

For non-commuting matrices (where AE ≠ EA), the simple calculus shortcut L = exp(A)E fails. Instead, the exact directional derivative involves an integral:

L = \u222B_0^1 exp((1 - u)A) E exp(uA) du

A remarkably elegant block-matrix identity allows computing this integral by evaluating a single larger matrix exponential. If we form the block matrix:

B = [[A, E], [], [0, A]]

Then its exponential yields the exact Fréchet derivative in the upper right block:

exp(B) = [[exp(A), L], [], [0, exp(A)]]

Worked Example

Let's manually compute the Fréchet derivative for a specific configuration. We freeze A as the Jordan block and perturb it in the Lower direction with time t = 1 and scale s = 1.


A = [0 1]   E = [0 0]   t = 1   s = 1
    [0 0]       [1 0]

The block matrix is:
B = [ tA  tsE ] = [ 0 1 0 0 ]
    [  0   tA ]   [ 0 0 1 0 ]
                  [ 0 0 0 1 ]
                  [ 0 0 0 0 ]

This is a nilpotent matrix (B^4 = 0). The exponential series simplifies:
exp(B) = I + B + B^2/2 + B^3/6

B^2 = [ 0 0 1 0 ]
      [ 0 0 0 1 ]
      [ 0 0 0 0 ]
      [ 0 0 0 0 ]

B^3 = [ 0 0 0 1 ]
      [ 0 0 0 0 ]
      [ 0 0 0 0 ]
      [ 0 0 0 0 ]

exp(B) = [ 1 1 1/2 1/6 ]
         [ 0 1  1  1/2 ]
         [ 0 0  1   1  ]
         [ 0 0  0   1  ]

The upper-right 2x2 block L is:


L = [ 1/2 1/6 ]
    [  1  1/2 ]

The entry at index 01 (row 0, col 1) is exactly 1/6 or roughly 0.16666666666666666.

Experiment

You can interact with the Fréchet derivative visually in the Matrix Exponential Fréchet Sensitivity module.

Set Matrix to "Jordan".
Set Direction to "Lower".
Set Time to 1 and Direction scale to 1.
Set Entry to "01".
Observe the "Fréchet derivative entry" which reads ~0.16666666666666666, confirming our derivation. Note the "Commuting shortcut entry" incorrectly predicts 0.0 because the matrices do not commute.

Limitations

This module implements the block-matrix algorithm using a direct Taylor series expansion. This approach is numerically stable for small matrices and small values of t, but it can suffer from cancellation errors (roundoff) for larger norms or ill-conditioned matrices. We use a threshold (1e-14) to normalize algebraically zero residuals. The series convergence is guaranteed but requires many terms for large scales.

References

The block-matrix identity for the Fréchet derivative of the matrix exponential is a standard result in matrix function theory. A comprehensive treatment of computing matrix functions and their derivatives can be found inHigham (2008), "Functions of Matrices: Theory and Computation".

Matrix Exponential Fréchet Sensitivity