Chapter 1

Introduction

• Overview
• A Brief History
• Classification
• A Glimpse of Some Known Results
• Conclusion

1
2 Introduction

Overview

• Often a physical process is described by a mathematical

model of which parameters represent important physi-
cal quantities.
¦ Direct analysis — Analyze or predict the behavior
of the underlying physical process from the param-
eters.
¦ Inverse analysis — Validate, determine, or estimate
the parameters adaptively from the behavior of the
physical process.
Overview 3

Inverse Eigenvalue Problem (IEP)

• The mathematical model involves matrices whose spec-

tral properties determine the dynamics of the physical
system.
• Reconstruct a matrix from prescribed spectral data.
¦ Spectral data may involve a mixture information of
eigenvalues or eigenvectors.
¦ Sometimes complete information is difficult to ob-
tain. Only partial information is available.
¦ For feasibility, often necessary to restrict the con-
struction to special classes of matrices.
4 Introduction

Fundamental Questions

• Solvability:
¦ Determine a necessary or a sufficient condition under
which an IEP has a solution.
• Computability:
¦ Develop a scheme through which, knowing a priori
that the given spectral data are feasible, a matrix
can be constructed numerically.
• Sensitivity:
¦ Quantify how a solution to an IEP is subject to
changes of the spectral data.
• Applicability:
¦ Differentiate whether the given data are exact or
approximate, complete or incomplete, and whether
only an estimation of the parameters of the system
is sufficient.
¦ Decide between physical realizability and physical
uncertainty which constraint of the problem should
be enforced.
Brief History 5

Brief History

• Studies of IEP’s have been quite extensive

¦ Engineering application.
¦ Algebraic theorization.
• Mathematical techniques employed in the study are
quite sophisticated:
¦ Algebraic curves.
¦ Degree theory.
¦ Differential geometry.
¦ Matrix theory.
¦ Differential equations.
¦ Functional analysis.
¦ ..
• Results are quite few and scattered even within the
same field of discipline.
6 Introduction

Literature Review

• Inverse Sturm-Liouville problem:

¦ Ambartsumyan’29
¦ Krein’33
¦ Borg’46, Levinson’49
¦ Gel’fand&Levitan’51
¦ Kac’66 (Can one hear the shape of a drum?)
¦ Hochstadt’73, Barcilon’74, McLaughlin’76, Hald’78
¦ Zayed’82, Issacson et al’83, McLaughlin’86, An-
dersson’88
¦ Lowe et al’95, Rundell’97
• Matrix theory:
¦ Downing&Householder’56, Mirsky’58
¦ Hochstadt’67
¦ de Oliveira’70, Hald’72, Golub’73, Friedland’77, de
Boor&Golub’78
¦ Biegler-König’81, Shapiro’83, Barcilon’86, Sun’86,
Boley&Golub’87
¦ Landau’94, Chu’98
Brief History 7

• Applied Mechanics:
¦ Barcilon’74
¦ Gottlieb’83, Gladwell’86
¦ Ram’91, Gladwell’96, Nylen&Uhlig’97
• Computation:
¦ Morel’76, Boley&Golub’77
¦ Nocedal et al’83, Friedland et al’88, Laurie’88
¦ Chu’90, Zhou&Dai’91, Trench’97, Xu’98
8 Introduction

Applications

• System identification and control theory.

¦ State/output feedback pole assignment problems.
• Applied mechanics and structure design.
¦ Construct a model of a (damped) mass-spring sys-
tem with prescribed natural frequencies/modes.
• Applied physics.
¦ Compute the electronic structure of an atom from
measured energy levels.
¦ Neutron transport theory.
• Numerical analysis.
¦ Preconditioning.
¦ Computing B-stable RK methods with real poles.
¦ Gaussian quadratures.
• Mathematical analysis.
¦ Inverse Sturm-Liouville problems.
Brief History 9

An Example

m3
m2
m1 m4
x3
x2
x1
F x4 F

• Vibration of equally spaced particles (with spacing h

and mass mi) on a string subject to a constant hori-
zontal tension F .
• Equation of motion for 4 particles:
d2x1 x1 x2 − x 1
m1 2 = −F + F
dt h h
d2x2 x2 − x1 x3 − x2
m2 2 = −F +F
dt h h
d2x3 x3 − x2 x4 − x3
m3 2 = −F +F
dt h h
d2x4 x4 − x3 x4
m4 2 = −F −F
dt h h
10 Introduction

• In matrix form:
d2x
= −DAx
dt2
¦ x = [x1, x2, x3, x4]T
 
2 −1 0 0
 −1 2 −1 0 
¦A=  0 −1 2 −1 


0 0 −1 2
¦ D = diag(d1, d2, d3, d4) with di = F
mi h .
• Eigenvalues of DA are the squares of the so called nat-
ural frequencies of the system.
• Want to place weights mi so that the system has a
prescribed set of natural frequencies.
¦ A is symmetric and tridiagonal.
¦ D is diagonal.
¦ This is a multiplicative inverse eigenvalue problem.
• Open Question: Can such a string have arbitrarily pre-
scribed natrual frequencies by adjusting the diagonal
matrix D?
Classification 11

Classification

• Based on constraint.
¦ Spectral constraint.
¦ Structure constraint.
• Based on physical suitability.
¦ Physical realizability.
¦ Physical uncertainty.
• Based on discipline.
¦ Essentially mathematical problem.
¦ Essentially engineering problem.
• Based on expectation.
¦ Determination problem.
¦ Estimation problem.
12 Introduction

Via Algebraic Characteristics

MVIEP

(single variate)

LSIEP
PIEP

MIEP
AIEP
SIEP

PDIEP

MVIEP = Multi-Variate IEP

LSIEP = Least Squares IEP
PIEP = Parameterized IEP
SIEP = Structured IEP
PDIEP = Partially Described IEP
AIEP = Additive IEP
MIEP = Multiplicative IEP
Classification 13

PIEP

• Generic form:
¦ Given
. A family of matrices A(c) ∈ M with c ∈ Fm,
. A set of scalars Ω ⊂ F,
¦ Find
. Values of parameter c such that
σ(A(c)) ⊂ Ω
• Remarks:
¦ Not necessarily m = n.
¦ Commonly used Ω:
. Ω = {λ∗k }nk=1.
. Ω = left-half complex plan.
. Ω = anything but must have a specific number of
zeros.
14 Introduction

Some Special PIEP’s

Pn
• A(c) = A0 + i=1 ci Ai

¦ Ai ∈ R(n), F = R.
¦ Ai ∈ S(n), F = R.
• (AIEP) A(c) = A(X) = A0 + X, X ∈ N .
¦ A0 ∈ C(n), F = C, N = DC (n).
• (MIEP) A(c) = A(X) = XA0, X ∈ N .
¦ Preconditioning?
Pq
• A(c) = A(K1, . . . , Kq ) = A0 + i=1 Bi Ki Ci .

¦ Pole assignment problem.

Classification 15

SIEP

• Generic form:
¦ Given
. A set N of specially structured matrices,
. A set of scalars {λ∗k }nk=1 ∈ F,
¦ Find
. X ∈ N such that
σ(X) = {λ∗k }nk=1.
• Some special cases:
¦N = {Toeplitz matrices in S(n)}.
¦N = {Persymmetric Jacobi matrices in S(n)}.
¦N = {Nonnegative matrices in S(n)}.
¦N = {Row-stochastic matrices in R(n)}.
16 Introduction

A Few More Special SIEP’s

• Given scalars λ∗i ≤ µi ≤ λ∗i+1, i = 1, . . . , n − 1, find a

Jacobi matrix J such that
σ(J) = {λ∗k }nk=1
σ(J(1 : n − 1, 1 : n − 1)) = {µ1, . . . , µn−1}.
• Given scalars {λ1, . . . , λ2n} and {µ1, . . . , µ2n−2} ∈ C,
find tridiagonal symmetric matrices C and K for the
λ-matrix Q(λ) = λ2I + λC + K so that
σ(Q) = {λ1, . . . , λ2n},
σ(Q(1 : n − 1, 1 : n − 1)) = {µ1, . . . , µ2n−2}.
• Given distinct scalars {λ1, . . . , λ2n} ⊂ R and a Jacobi
matrix Jn ∈ R(n), find a Jacobi matrix a Jacobi matrix
J2n ∈ R(2n) so that
σ(J2n) = {λ1, . . . , λ2n},
J2n(1 : n, 1 : n) = Jn.
• Given a family of matrices A(c) ∈ Rm×n, with c ∈ Rn,
m ≥ n, find a parameter c such that the singular values
of A(c) are precisely the same as a prescribed set of
nonnegative real values {σ1, . . . , σn}.
Classification 17

LSIEP

• Maintain the structure, approximate the eigenvalues:

¦ Given
. A set of scalars {λ∗1 , . . . , λ∗m} ⊂ F (m ≤ n),
. A set N of specially structured matrices,
¦ Find
. A matrix X ∈ N
. An index subset σ = {σ1 < . . . < σm} such that
1X
m
F (X, σ) := (λσi (X) − λ∗i )2,
2 i=1
is minimized.
18 Introduction

• Maintain the spectrum, approximate the structure:

¦ Given
. A set M of spectrally constrained matrices,
. A set N of specially structured matrices,
. A projection P from M onto N ,
¦ Find
. X ∈ M that minimizes
1
F (X) := kX − P (X)k2.
2
Classification 19

Via Physical Characteristics

• By mechanical types:
¦ Continuous vs. discrete.
¦ Damped vs. undamped.
• By data type:
¦ Spectral, modal, or nodal.
¦ Complete vs incomplete.
20 Introduction

A Glimpse of Some Major Issues

• Studies on IEP’s have been intensive, ranging from ac-

quiring a pragmatic solution to a real-world application
dealing the metaphysical theory of an abstract fromu-
lation.
• Results are scattered even within the same field of dis-
cipline.
• Only a handful of the problems have been completely
understood.
• Many interesting yet challenging questions remain to
be answered.
A Glimpse of Some Major Issues 21

Complex Solvability

• Solving an IEP over complex field amounts to solving a

polynomial system with complex coefficients. Generally
speaking, the system is generically solvable.
• Given A0 ∈ C(n) and arbitrary {λ∗k }nk=1 ⊂ C,
¦ There exists D ∈ DC (n) such that
σ(A0 + D) = {λ∗k }nk=1
and there are at most n! solutions.
¦ If det(A0(1 : j, 1 : j) 6= 0, j = 1, . . . , n, then there
exists D ∈ DC (n) such that
σ(DA0) = {λ∗k }nk=1
and there are at most n! solutions.
22 Introduction

Real Solvability

• Solving an IEP over real field is a much harder problem.

Sufficient conditions are generally quite restrictive.
• Assume all matrices involved are real,
¦ If the prescribed real eigenvalues are sufficiently dif-
ferent, then there exist c1, . . . , cn ∈ R such that
X
n
σ(A0 + ciAi) = {λ∗k }nk=1.
i=1

¦ The inverse eigenvalue problem associated with

X
n
A0 + ciAi
i=1

is unsolvable almost everywhere if and only if any

of the prescribed eigenvalues has multiplicity great
than 1.
• Symmetric Toeplitz matrices can have arbitrary spec-
tra.
A Glimpse of Some Major Issues 23

Numerical Methods

• Direct methods
¦ Lanczos method.
¦ Orthogonal reduction methods.
• Iterative methods
¦ Newton-type iteration.
• Continuous methods:
¦ Homotopy approach.
¦ Projected gradient method.
¦ ASVD approach.
24 Introduction

Sensitivity Analysis

• Assume all matrices are symmetric and the PIEP for

Xn
A(c) = A0 + ciAi
i=1
is solvable.
• Assume A(c) = Q(c)diag{λ∗k }nk=1Q(c)T and define
£ T
¤
J(c) = qi(c) Aj qi(c) , i, j = 1, . . . , n,
£ T T
¤T
b = q1(c) A0q1(c), . . . , qn A0qn(c) .
• If
X
n
δ = kλ − λ̃k∞ + kAi − Ãik2
i=0
is sufficiently small, then
¦ The PIEP associated with Ãi, i = 0, . . . , n and
{λ̃1, . . . , λ̃n} is solvable.
¦ There is a solution c̃ near to c,
à Pn !
kc − c̃k∞ kλ − λ̃k∞ + kA0 − Ã0 k2 i=1 kAi − Ãi k2
≤ κ∞ (J(c)) + + O(δ 2 ).
kck∞ kλ − bk∞ kJ(c)k∞
A Glimpse of Some Major Issues 25

Summary

• An IEP concerns the reconstruction of a matrix satis-

fying two constraints.
¦ Spectral constraint – the prescribed spectral data.
¦ Structural constraint – the desirable structure.
• Different constraints define a variety of IEP’s.
• Studies on IEP’s have been intensive, ranging from en-
gineering application to algebraic theorization.
¦ Many unanswered yet interesting questions.
• A common phenomenon in all applications is that the
physical parameters of a certain system are to be re-
constructed from knowledge of its dynamical behavior,
in particular, of its natural frequencies/modes.
¦ Sometimes the constraints can be precisely deter-
mined.
¦ Sometimes the constraints are only approximate and
often incomplete.