We define the Matrix Exponential, denoted by , of a Matrix by the power series , which converges absolutely in any submultiplicative matrix norm since (see Calculating the Matrix Exponential). For a generic Matrix pair , the matrix exponential satisfies:
- , and , where is the entrywise complex conjugate, is the Transpose and is the conjugate transpose.
- holds .
- , where is the Trace
- , so is always an invertible matrix
- (Lie product formula)
Calculating the exponential explicitly
- Diagonalisable: if with , property 3 gives . For normal the matrix can be chosen unitary.
- Jordan form: any is similar to a direct sum of Jordan blocks with nilpotent, for a block of size . Since and commute, property 5 gives the finite sum
- closed form: write with , so that is traceless. By Cayley-Hamilton , so splitting the series into even and odd powers,
Here at . Any branch of the square root works since the expression is even in . For the Pauli matrices this specialises to for a unit vector , since .
- Polynomial interpolation: by Cayley-Hamilton , where the polynomial agrees with on each eigenvalue, together with its first derivatives at an eigenvalue whose largest Jordan block has size . For distinct eigenvalues this is Sylvester’s formula .
- Numerically, eigendecomposition is unstable for nearly defective matrices. Standard libraries (
scipy.linalg.expm) use scaling and squaring, , with a Padé approximant for .
Calculating the logarithm explicitly
A logarithm of is any with ; details in Calculating the matrix logarithm.
- By property 6, only invertible matrices have logarithms, and every invertible complex matrix has one (take a logarithm of each Jordan block, as below). Logarithms are never unique, since .
- If has no eigenvalue on , there is a unique principal logarithm whose eigenvalues have imaginary part in .
- Diagonalisable: , for any choice of branches.
- Jordan block: , using the Mercator series for , which terminates as is nilpotent.
- Near the identity, converges for and inverts the exponential there.
- Real matrices: a real invertible has a real logarithm if and only if each Jordan block of a negative eigenvalue occurs an even number of times. In particular is necessary but not sufficient.