Matrix functions
1. We can easily define a natural power of a square matrix X as a product:
For a non-singular matrix, a negative power can be defined as a corresponding positive power of the matrix inverse:
X–n = (X–1)n
or, alternatively, as the matrix inverse of the positive power:
X–n = (Xn)–1
Exercise: Demonstrate that the two definitions are equivalent.
The zeroth power of any matrix is the identity matrix by definition:
X0 = I
Exercise: Demonstrate the following properties of matrix powers:.
XnXm = Xn+m
(Xn)m = Xnm
3. If y is an eigenvector of matrix X: Xy = y, then X2y = XXy = Xy = Xy = 2y. Analogously, for any power: Xny = ny. Therefore, the n-th power of matrix X has the same eigenvectors as the original matrix X, with the corresponding eigenvalues being n. This is also true for any negative power, including the matrix inverse: X–1y = y XX–1y = Xy Xy = (1/)y.
A diagonalizable matrix X can be expressed as follows (similarity transformation):
X = TT–1
where = diag(1,2,...N) is the diagonal matrix of eigenvalues and T is the eigenvector matrix (column by column). Then, for the X2 matrix we obtain:
where = diag(12,22,...N2). In the case of normal matrices (X†X = XX†), the eigenvectors are orthogonal: T–1 = T†. Analogously, any power Xn can be expressed as follows:
Xn = TnT–1
n = diag(1n,2n,...Nn).
4. If a given real function can be expressed as a Taylor series:
where
we can construct a similar power series substituting the scalar argument x by a square matrix X:
This is a definition of function f of a matrix X.
For instance, the exponent of matrix X will be given by the following series:
5. The above Taylor expansion can be used as such to iteratively compute a matrix function, but for diagonalizable matrices it is usually more efficient to employ the following algorithm:
a. Diagonalize matrix X, obtain the matrix of eigenvectors T and the diagonal matrix of eigenvalues . If X is not a normal matrix, we will also need to invert T. This gives a similarity transformation X = TT–1, where = diag(1,2,...N).
b. The exponent of matrix X is given by the following expression:
The latter formula gives us a non-iterative method of calculating exp(X) through diagonalization. In general, any matrix function can be calculated this way:
f(X) = T diag(f(1), f(2),... f(N))T–1.
Obviously, f(X) has the same eigenvectors as the original matrix X, with the corresponding eigenvalues being f().
6. The square root X1/2 of matrix X is defined through the expression (X1/2)2 = X. The calculation of X1/2 can be also carried out by way of the diagonalization technique:
Exercise: Demonstrate that the above formula indeed yields X1/2.
6. The inverse square root X–1/2 of matrix X is defined through the expression (X–1/2)2 = X–1. The calculation of X–1/2 can be done analogously:
Of course, this is only possible if all eigenvalues are non-zero. Otherwise, X is singular and X–1/2 does not exist. Note that the functions or can not be represented by Taylor series. Nevertheless, the above algorithms work.
The inverse square root plays a crucial role in the Löwdin orthogonalization method.
7. Note that if matrix X is Hermitian, all its powers Xn and matrix functions f(X) will be also Hermitian. This refers, in particular, to matrix inverse X–1, matrix square root X1/2, and the inverse square root X–1/2.
ABSTRAK PEMBUATAN ALSIC METAL MATRIX COMPOSOITES DENGAN METODE LANXIDE
ACCOMMODATIONS MATRIX FOR DAILY ACTIVITIES READ WRITE HDWR
ACCOUNTING MATRIX STRANDS CONCEPTUAL PROCESSING REPORTING INTERPRETATION
Tags: define a, natural, functions, easily, define, matrix