Department of Mathematics

Dresden University of Technology

The Contribution of Heinz Unger (1914 -- 2007)

It seems that Heinz Unger was the first (after Jacobi, who much earlier did something like a linearization of the eigenvalue equation) who used Newton techniques for solving eigenvalue problems. In his groundbreaking short paper

Heinz Unger: **Nichtlineare Behandlung von Eigenwertaufgaben**
(Nonlinear Treatment of Eigenvalue Problems)

Z. Angew. Math. Mech. 30 (1950), pp. 281--282

he considered the eigenvalue equation

M( λ ) x = ( Ax - λ I ) x = 0 (EE)

with a real (n x n)-matrix A and added to it the normalization condition

w^{T}x = 1
(NC)

where w = e_{m} is an appropriately chosen coordinate vector, i.e.,
he required the m-th component of x being equal to 1:
x_{m} = 1. Without restriction of generality Unger
sets m = 1, i.e., x_{1} = 1, so that (EE) reduces to the nonlinear
(n x n)-system

g( λ,y) =
g( λ,x_{2}, ... ,x_{n}) =
M( λ )( 1 ,x_{2}, ... ,x_{n})^{T} =
m_{1}( λ ) 1 +
m_{2}( λ ) x_{2} + ... +
m_{n}( λ ) x_{n} = 0
(REE)

( λ^{(k)},y^{(k)} ) --->
( λ^{(k+1)} = λ^{(k)}+ μ^{(k)},
y^{(k+1)}= y^{(k)}+ t^{(k)} )

g( λ^{(k)},y^{(k)} ) +
g'( λ^{(k)},y^{(k)} )
(μ^{(k)},t^{(k)}^{T} )^{T} = 0

This is equivalent to applying Newton's method to the extended (n+1)x(n+1)-system

(A - λ I) x = 0

w^{T}x - 1 = 0

It turns out that the Newton step yields the new eigenvector approximation

x^{(k+1)} =
(A - λ^{(k)} I) ^{-1} x^{(k)} μ^{(k)}

where μ^{(k)} plays the role of a scaling factor, i.e., the direction of the Newton successor to x^{(k)} is the same as the direction obtained by performing one step of inverse iteration with shift λ^{(k)} to x^{(k)} , cf. formula (31) on page 165 of the German textbook

Rudolf Zurmühl: **Praktische Mathematik für Ingenieure und Physiker**

Springer, Berlin 1953

which only three years later presented Unger's idea. By the way, I learned some parts of numerical linear algebra from this widely used book when I studied Mathematics at TH Dresden in the early sixties of the last century.

Unger pointed out in his 1950 paper that his approach can also be used for general linear eigenproblems

M( λ ) x = ( Ax - λ B ) x = 0

and for general nonlinear eigenproblemsM( λ ) x = 0

with a nonlinear matrix valued mapping M( λ ), e.g., for quadratic eigenproblems
M( λ ) x =
( A λ^{2} + B λ + C ) x = 0.

Unger's approach is also described by Lothar Collatz in subsection 19.4: *Das Newtonsche Verfahren bei Eigenwertaufgaben (Newton's Method for Eigenvalue Problems)* of his monograph

Lothar Collatz: **Funktionalanalysis und numerische Mathematik**

Springer, Berlin 1964

There instead of Ax = λ x the general linear eigenvalue problem Ax = λ Bx is considered. Moreover, since the eigenvalue equation is bilinear in x and λ, hence the second F-derivative is constant, applying the third order Euler-Tschebyscheff method which uses second derivatives is proposed there.

Unfortunately, the English speaking numerical linear algebra community did not recognize Unger's basic paper. For instance, also Peters and Wilkinson did not mention it in their famous survey

Peters, G. and Wilkinson, J. H.: **Inverse Iteration, Ill-Conditioned Equations and Newton's Method**

SIAM Rev. 21 (1979), pp. 339 -- 360

For colleagues who are interested in this early source I include here

- a copy of Unger's original paper
- an English translation using \mathfrak letters for vectors and matrices as in the original paper
- an English translation using \mathbf letters for vectors and matrices for better readability

**A short CV of Heinz Unger**

- Born June 10, 1914 in Nordhausen
- 1941 Diplom (Master) at TH Darmstadt
- April 24, 1944 Dr.rer.nat. (PhD) at TH Darmstadt under the supervision of Alwin Walther

Title of his Dissertation (Thesis):*Numerische Behandlung von Anfangswertproblemen bei gewöhnlichen linearen Differentialgleichungen 2. Ordnung*

English translation:*Numerical treatment of initial value problems for ordinary linear differential equations of 2nd order* - 1948 Habilitation at TH Darmstadt
- 1955 Full Professor at TH Hanover
- 1958 Full Professor on the first Chair for Applied Mathematics at the University of Bonn
- Founding of the GMD (Gesellschaft für Mathematik und Datenverarbeitung), one of the two first CEO's of GMD
- 1979 Professor Emeritus at the University of Bonn
- Died November 6, 2007
(Obituary notice from
*Bonner Generalanzeiger*)