Purdue University

EAS 657

Geophysical Inverse Theory

Robert L. Nowack

Lecture 6

 

Spectral Decompositions and Singular Value Decompositions

 

            We will first investigate finite dimensional mappings represented by square matrices

 

A:RN  RN

 

An eigenvector for the operator A is a vector whose direction remains unchanged upon the operation of operator A.  Thus

 

 

where  is a scalar called the eigenvalue associated with u.  Also, a zero eigenvalue can be associated with the equation

 

Au = 0

 

The eigenvalue, , is in general a complex scalar.  u is an eigenvector if  with .  Thus, (A-I) must be singular (not invertible).  This will occur when Det(A-I) = 0.  Det(A-I) = 0 is called the characteristic equation and can be factored as

 

 

            Ex)  A:R2  R2           A =

 

 

Back solving this for the eigenvectors for A gives

 

 

            We now define the eigenvectors of the adjoint operator A*, w, and the corresponding eigenvalues .  Since transposing doesn’t change the determinant, the eigenvalues of A* are the complex conjugates of the eigenvalues of A.  Thus

 

 

Now since

 

 ,

 

if ui is an eigenvector of A and wj is an eigenvector of A*, then

 

 

Thus, if , then wj must be perpendicular to ui.  This is called the biorthogonality relation for eigenvectors of A and A* for distinct .

 

Theorem:  Let  be distinct eigenvalues of A.  Then the corresponding eigenvectors of A are independent.  Thus (u1, u2, … uN) can be used as a basis for U.  This is a major utility of eigenvectors.

 

            For each eigenvector then,

 

 

or putting all the eigenvectors into one equation

 

 

If the ui are all linearly independent (which is not guaranteed for repeated eigenvalues), then

 

 

and

 

 

This is called a spectral representation of the linear operator A with independent eigenvectors.  Of course, if the eigenvectors are not independent, then one cannot perform the inverse of M.

 

            Ex) 

 

but

 

 

Thus, for this example, u1 is not distinct from u2 and a spectral decomposition cannot be performed on this matrix.

 

            Matrices of the form

 

 

are called Jordan Blocks and have all eigenvectors for  the same.  For these type of matrices, spectral decompositions using eigenvectors cannot be performed.

 

            This type of problem, with eigenvectors used as a basis for square operators, can be extended for more general decomposition of operators using the singular value decomposition.  However, eigenvector expansions are very useful for self adjoint operators where

 

A* = A.

 

In this case, .  Thus, the eigenvalues for any self adjoint (Hermitian matrix) operator are real.  Also, the eigenvectors of A and A* are the same.  Thus, ui = wi.  In this case, for  then, ui is perpendicular uj.  Thus, the eigenvectors are directly orthogonal and not biorthogonal.

 

            We have only showed this for distinct eigenvalues, but in fact, for an arbitrary case of repeated eigenvalues, we can always construct an orthonormal set of eigenvectors for a Hermitian matrix.  (Hermitian matrices will never reduce to Jordan Blocks.)  We can, therefore, always perform spectral decompositions for self adjoint operators (for Hermitian matrices)!

 

            Any matrix with orthonormal columns is called unitary and has the property that

 

M-1 = M*

 

M*M = MM* = I

 

Thus, since M is unitary for a symmetric Hermitian matrices, then

 

 

Directly in terms of the eigenvectors for a square Hermitian Matrix A,

 

 

where  denotes an outer product.  This is called the spectral theorem.  Thus,

 

 

This scales each ui by  and the projection of x on each ui.

 

            For all nondefective square matrices (no Jordan Blocks), we can define functions of matrices as

 

 

where  is the ith power of the matrix A and  is a function of the matrix.  This corresponds to the scalar expansion for  as

 

 

Since

 

 

then

 

 

In general, for a nondefective square matrix,

 

 

This is called Sylvester’s theorem.

 

            Ex)  Most linear physical systems can be written as a system of first order differential equations

 

 

The formed solution to this is eAt where A is a matrix.  But what is eAt?  By analogy to the 1-D case

 

 has a solution

 

Using Sylvester’s theorem,

 

 

Solving systems of differential equations is another major use of eigenvectors and eigenvalues.

 

            Ex)  The stress “tensor” is a matrix quantity at a point in space where one index relates to the coordinate face on which the traction acts and the second index is the traction vector component.

 

 

 

 

The traction vector acting on a plane with normal  can be written

 

 sum on j

 

            Now we want to find the vectors  such that

 

 

or

 

 eigenvector problem

 

On these faces, only normal tractions act.   are the principle stress directions and  are the principle stresses.  The  define three orthonormal directions since  is symmetric, then

 

 

where M is  are the principle directions, and  is a diagonal matrix of principle stresses.

 

 

 

Generalized Matrix Analysis

 

            Only Hermitian matrices are guaranteed not to have problems in standard eigenvector/eigenvalue analysis and non-square matrices are excluded.  Fortunately, many differential operators are self adjoint which allow this sort of symmetric matrix analysis.

 

To avoid problems in other cases, Lanczos (1959) devised a trick of imbedding a general non-square matrix A within a larger Hermitian matrix

 

 

S is now Hermitian and, therefore, has a complete set of eigenvectors and real eigenvalues.

 

            If A is an MxN matrix

 

 

 

 

Then S is an (M+N) x (M+N) matrix.  Thus,

 

 

 

 

            Solutions to  define the eigenvectors and eigenvalues for S.  Since we are interested in A, we will split the eigenvectors w into parts

 

 

where  is length M and  is length N.  Then

 

 

This can also be written

 

 

This is a coupled eigenvalue problem through A and its adjoint A*.  Note that if

 

 

is an eigenvector for , then

 

 

is an eigenvector for .

 

            The eigenvalues of S are found from

 

 

If there are 2P non-zero eigenvalues, then

 

 

            Since the eigenvector set for S is orthogonal (or made orthogonal for repeated eigenvalues)

 

 

then

 

 

and

 

 

 

Thus, ui is perpendicular to uj and vi is perpendicular to vj where

 

 

 

We will normalize these to 1.

 

 

            The eigenvalues, , separate the coupled eigenvalue problem into

 

 

 

The , i = 1, N-P, span the nullspace N(A) of A, thus

 

 

and correspondingly , i = 1, M-P, span the nullspace N(A*), thus

 

 

Since these equations are now decoupled, we can now organize the eigenvectors of S.  for nonzero

 

 

 

 

 

where P is the rank of A and also the dimension of Nperp(A) or R(A).  For the zero eigenvalues,

 

 

 

Thus, by playing this trick of constructing S, we have been able to span the domain and range of A by eigenvector-like orthonormal vectors.

 

            Now lets look again at the coupled eigenvector problem

 

 

Operate on the first equation by A*, then

 

 

and the second equation by A, then

 

 

where A*A is an (NxN) matrix and AA* is an (MxM) matrix.  These are ordinary eigenvalue problems for (A*A) and (AA*).  Note that since

 

(x, A*Ay) = (A*Ax, y)

 

(x, AA*y) = (AA*x, y)

 

both A*A and AA* are self adjoint (Hermitian matrices).  Thus {u}, i = 1,M, and {v}, i = 1,N, form complete eigenvector systems for the operators {AA*} and {A*A}.  The number of non-zero positive eigenvalues  is P where .

 

Define

 

 

These are the “singular values” and are always positive real numbers.  We can also separate the eigenvectors into matrices.

 

 

 

 

 

and

 

 

where  are the non-zero eigenvalues for A*A and AA*.  Now let

 

V = [Vp, V0]

 

U = [Up, U0]

 

Since the columns of V and U form complete orthonormal sets of eigenvectors for A*A and AA*, then V* = V-1 and U* = U-1 are unitary matrices.  Thus,

 

UU* = U*U = IMxM

 

VV* = V*V = INxN

 

For the reduced set Up (a MxP matrix) and Vp (a NxP matrix) by orthogonality of the colums of Up, then

 

Up* Up = Ipxp             Vp* Vp = Ipxp

 

where Up* is PxM, Up is MxP, Vp* is PxN and Vp is NxP.  But in general, UpUp*  IMxM (unless P = M, and VpVp*  INxN (unless P = N).

 

 

            Now we want to assemble this into one representation for A.

 

 

 

where A is MxN, Vp is NxP, Up is MxP, V0 is Nx(N-P), A* is NxM,  is PxP, and U0 is
Mx(M-P).  We can then write

 

 

and

 

 

If we multiple this through, then the only non-zero term is

 

 

This is called the singular value decomposition of the operator or matrix where

 

 

 

 

The columns of Up and Vp are orthonormal vectors that span R(A) and Nperp(A) respectively.  This is a general representation of any matrix regardless of whether it is square or non-square, or whether it is complete or defective.  The subspaces spanned by U0 and V0 can be thought of as “blindspots” not illuminated by A.

 

            Ex)  Let

v1 + v2 = 1

 

v3 = 2

 

-v3 = 1

 

This can be written in matrix form as

 

 

where A:R3  R3  

 

 

 

 

Now,

 

 

 

In U space:  , i = 1, 2, 3.  The eigenvalues are found from

 

 

 

Thus,

 

 

The u1 and u2 eigenvectors are found by solving AA*ui = 2ui for u1 and u2 and normalizing   Thus,

 

 

The third eigenvalue is zero, so

 

 

Thus,

 

 

 

Now we can use a shortcut to solve for the Vp space by

 

 

Find the V0 space vector by solving A*Av3 = 0

 

 

then,

 

 

This is the singular value decomposition of the matrix A.