betaFeedback

WorksMathematics

Matrix-point interpolation for positive real odd functions in one and several variables

Abstract

We study the single matrix point interpolation problem for positive real matrix valued functions on the right half-plane, both in the single variable setting and in the noncommutative multivariable setting. In the single variable case, a careful analysis is done for when the solution may or must have a pole at infinity. In the multivariable setting, as the pole structure at infinity can be more complicated, we focus on the case when there is no pole at infinity. The results also yield numerical algorithms for finding a solution.

Introduction

We write C(z)\mathbb{C}(z) and R(z)\mathbb{R}(z) for the rational functions in the complex variable zz with complex and real coefficients, respectively. We write Pm×m\mathcal{P}^{m \times m} for the class of positive real m×mm \times m rational matrix functions ff without poles in the open right half-plane C+:={z∈C:Re⁡z>0}\mathbb{C}_{+} := \{z \in\mathbb{C} : \operatorname{Re} z > 0\}, i.e., such that

Re⁡f(z)≥0whenever Re⁡z>0.\operatorname{Re} f(z) \ge0 \quad\text{whenever } \operatorname{Re} z > 0.

Next, Om×m\mathcal{O}^{m \times m} denotes the class of odd m×mm \times m rational matrix functions ff, i.e.,

f(z)=−f(−z)∗.f(z) = -f(-z)^*.

In addition, we define the class of positive real odd matrix functions POm×m:=Pm×m∩Om×m\mathcal{PO}^{m \times m} := \mathcal{P}^{m \times m} \cap\mathcal{O}^{m \times m}. If the superscript m×mm \times m is missing, then m=1m = 1. We also have real variants of these classes of functions, such as PR:=P∩R(z)\mathcal{PR} := \mathcal{P} \cap\mathbb{R}(z), PRO:=PO∩R(z)\mathcal{PRO} := \mathcal{PO} \cap\mathbb{R}(z), etc. Scalar valued positive real odd functions were first introduced in the electrical engineering literature (see Foster [12], Cauer [7] and Brune [6]; see also [21, 5, 2]).

A function f∈C(z)f \in\mathbb{C}(z), written as f(z)=p(z)q(z)f(z) = \frac{p(z)}{q(z)} for co-prime polynomials pp and qq is evaluated in a matrix A∈Cn×nA \in\mathbb{C}^{n \times n} via f(A)=p(A)q(A)−1f(A) = p(A)q(A)^{-1} whenever q(A)−1q(A)^{-1} exists, that is, whenever the intersection of the eigenvalues of AA and the roots of qq is empty. By standard functional calculus, the definitions of the classes

2020 Mathematics Subject Classification. 47A57, 32E30, 15B48, 93B28.

Key words and phrases. Positive real functions, Odd functions, Matrix-point interpolation, Noncommutative variables, Completely positive maps, Arveson extension.

This work is based on the research supported in part by the National Research Foundation of South Africa (Grant Number 127364). In addition, HW is partially supported by NSF grant DMS-2348720. Pm×m\mathcal{P}^{m\times m} and Om×m\mathcal{O}^{m\times m} extend to matrix-point evaluation. Hence f∈Pm×mf\in\mathcal{P}^{m\times m} if and only if for each n∈Nn\in\mathbb{N} and each A∈Cn×nA\in\mathbb{C}^{n\times n} with Re⁡A:=12(A+A∗)>0\operatorname{Re}A:=\frac{1}{2}(A+A^*)>0, we have Re⁡f(A)≥0\operatorname{Re}f(A)\geq0, whereas f∈Om×mf\in\mathcal{O}^{m\times m} holds if and only if for each n∈Nn\in\mathbb{N} and A∈Cn×nA\in\mathbb{C}^{n\times n} we have −f(A)=f(−A∗)∗-f(A)=f(-A^*)^*.

The main problem in this paper concerns the single matrix point interpolation problem, which in the case of one variable is the following.

Problem 1.1. For A∈Cn×nA\in\mathbb{C}^{n\times n} and B∈Cnm×nmB\in\mathbb{C}^{nm\times nm}, when does there exist a function ff in Pm×m\mathcal{P}^{m\times m}, POm×m\mathcal{PO}^{m\times m} or PROm×m\mathcal{PRO}^{m\times m} such that f(A)=Bf(A)=B? If so, provide a way to find a/all solutions.

Basic functional calculus shows that BB should be in the bicommutant of AA, and in that case, when AA and BB are also Lyapunov regular, one can reduce to the case where AA is nonderogatory and Re⁡A\operatorname{Re}A positive definite. Indeed, in the Jordan canonical form of AA one can simply remove any blocks with repeated eigenvalues, leaving only the largest one, and reflect eigenvalues of AA with negative real part over the imaginary axis, making appropriate adjustments in BB.

This problem was first considered in the real scalar case (thus, PRO\mathcal{PRO}) by N. Cohen and I. Lewkowicz [9, 10], for the case that AA and BB are Lyapunov regular, that is, in both cases the eigenvalues λj\lambda_j satisfy λj+λi‾≠0\lambda_j+\overline{\lambda_i}\neq0 for all ii and jj. More recently, S. ter Horst and A. van der Merwe [15] studied the problem in PO\mathcal{PO} and PRO\mathcal{PRO} and identified the so-called Hill-Pick matrix HA,B\mathbb{H}_{A,B} of the pair (A,B)(A,B) and showed that solvability of the interpolation problem requires HA,B\mathbb{H}_{A,B} to be positive semidefinite. With the additional condition that rank⁡HA,B\operatorname{rank}\mathbb{H}_{A,B} equals the dimension of the real bicommutant of AA, they showed in [14] that a solution exists and provided a constructive way to find the solution. This solution f(z)f(z) has the additional property lim⁡z→∞f(z)=0\lim_{z\to\infty}f(z)=0, which is referred to as the solution having no pole at infinity.

In what follows we build on the results by Ter Horst and Van der Merwe, and generalize them in several directions. Namely, in the matrix case, we prove existence of a solution without the condition on rank HA,B\mathbb{H}_{A,B}, in which a pole at ∞\infty can occur, and we clarify in detail when that happens, as well as when there exists a unique solution. We then extend some of the results to the multivariable noncommutative function setting. Here, with a domain described via a linear matrix inequality Q(z)>0Q(z)>0, we first establish necessary and sufficient condition for the existence of an interpolant with a realization of the form

f(z)=D+ℓ∗(Is⊗Q(z)−M)−1ℓ,D=−D∗, M=−M∗.(1)f(z)=D+\ell^*(I_s\otimes Q(z)-M)^{-1}\ell,\qquad D=-D^*,\ M=-M^*. \tag*{(1)}

These conditions involve complete positivity of a map defined on an operator system and the existence of an Arveson extension, where the extension is required to satisfy an additional constraint; see Theorem 4.2 for details. Next, we consider the special case where Q(z)=diag⁡(zj)j=1dQ(z)=\operatorname{diag}(z_j)_{j=1}^d, that is, the case when the domain is the standard noncommutative poly-half-plane. In this case, we can give a functional theoretic characterization of rational matrix functions in noncommutative variables of the form (1.1); see Theorem 5.1. We end the paper with providing a computational algorithm (using SemiDefinite Programming) to find a multivariable interpolant of the form (1.1) and illustrate the results in some examples.

The paper is organized as follows. In Section 2 we will discuss in detail the one variable case, where already many results exist. Putting the matrix AA in Jordan canonical form allows for detailed computational descriptions. In Section 3, we provide a view of the one variable case that includes completely positive maps, which can then be used to develop partial results in the multivariable noncommutative case, which we will treat in Sections 4 (generalized poly-half planes) and 5 (standard poly-half plane). The approach here is a more coordinate free one, in part as there is no convenient canonical form for a tuple of noncommuting matrices. Along the way, we illustrate the results with examples and provide constructions to find a solution.

The one variable case

In this section, we discuss the one variable interpolation results. The results are treated in different parts of the literature, so this section serves as a convenient recap. As most of the results are not new, we do not provide all details in the proofs but rather provide the main ideas. In the next section we will approach the one variable case from a different and coordinate free perspective, as a lead into the multivariable case. The spectrum of a square matrix MM is denoted by σ(M)\sigma(M).

Background on state space realizations

More details and proofs on the results we recall in this subsection can be found in [16, 4, 11, 18]. We let rational functions over F∈{C,R}\mathbb{F} \in\{\mathbb{C}, \mathbb{R}\} be denoted by F(z)\mathbb{F}(z), which is the field of fractions of the polynomial ring F[z]\mathbb{F}[z] of one variable.

Definition 2.1 (Minimal realization). Let

G(z)∈F(z)p×mG(z) \in\mathbb{F}(z)^{p \times m}

be a matrix-valued rational function over a field F\mathbb{F}. A state-space realization of GG is a quadruple

(A,B,C,D),(A, B, C, D),

with

A∈Fn×n,B∈Fn×m,C∈Fp×n,D∈Fp×m,A \in\mathbb{F}^{n \times n}, \qquad B \in\mathbb{F}^{n \times m}, \qquad C \in\mathbb{F}^{p \times n}, \qquad D \in\mathbb{F}^{p \times m},

such that

G(z)=D+C(zI−A)−1B.G(z) = D + C(zI-A)^{-1}B.

The realization (A,B,C,D)(A, B, C, D) is called minimal if its state dimension nn is the smallest among all realizations of G(z)G(z); that is, if there does not exist another realization

(A~,B~,C~,D~)(\widetilde{A}, \widetilde{B}, \widetilde{C}, \widetilde{D})

with state dimension n~<n\tilde{n}<n satisfying

G(s)=D~+C~(sI−A~)−1B~.G(s)=\tilde{D}+\tilde{C}(sI-\tilde{A})^{-1}\tilde{B}.

The minimal state dimension is called the McMillan degree of GG.

Theorem 2.2. A realization (A,B,C,D)(A,B,C,D) of the rational matrix G(s)G(s) is minimal if and only if it is both controllable and observable. Equivalently,

rank⁡(B  AB  ⋯  An−1B)=n,\operatorname{rank}\left(B\ \ AB\ \ \cdots\ \ A^{n-1}B\right)=n,

and

rank⁡(CCA⋮CAn−1)=n.\operatorname{rank}\begin{pmatrix} C \\ CA \\ \vdots\\ CA^{n-1} \end{pmatrix}=n.

Rational matrix function given by a realization as in Definition 2.1 have the limitation that they cannot have a pole at ∞\infty. Since the functions we consider in Problem 1.1 can have poles at ∞\infty, we will also consider descriptor systems which we define next; the following can be found in [11 Sections 2-2, 2-3].

Definition 2.3 (Minimal descriptor realization). Let

G(z)∈F(z)p×mG(z)\in\mathbb{F}(z)^{p\times m}

be a matrix-valued rational function. A descriptor realization (or singular realization) of GG is a quintuple

(E,A,B,C,D),(E,A,B,C,D),

where

E,A∈Fn×n,B∈Fn×m,C∈Fp×n,D∈Fp×m,E,A\in\mathbb{F}^{n\times n},\qquad B\in\mathbb{F}^{n\times m},\qquad C\in\mathbb{F}^{p\times n},\qquad D\in\mathbb{F}^{p\times m},

such that the matrix pencil zE−AzE-A is regular, i.e.,

det⁡(zE−A)≠0,\det(zE-A)\ne0,

and

G(z)=D+C(zE−A)−1B.G(z)=D+C(zE-A)^{-1}B.

The descriptor realization (E,A,B,C,D)(E,A,B,C,D) is called minimal if its state dimension nn is the smallest among all regular descriptor realizations of G(z)G(z). The minimal state dimension is called the McMillan degree of GG.

Definition 2.4. Consider the descriptor realization (E,A,B,C,D)(E,A,B,C,D) of a rational matrix function G(z)∈F(z)p×mG(z)\in\mathbb{F}(z)^{p\times m}.

(1) The triple (E,A,B)(E,A,B) is said to be finite controllable if

rank⁡(λE−A  B)=n\operatorname{rank}\left(\lambda E-A\ \ B\right)=n

for every finite generalized eigenvalue λ\lambda of the pencil sE−AsE-A.

(2) The triple (E,A,B)(E,A,B) is said to be impulse controllable if

rank⁡(E00AEB)=n+rank⁡E.\operatorname{rank}\begin{pmatrix} E & 0 & 0 \\ A & E & B \end{pmatrix}=n+\operatorname{rank}E.

(3) The triple (E,A,C)(E,A,C) is said to be finite observable if

rank⁡(λE−AC)=n\operatorname{rank}\begin{pmatrix} \lambda E-A \\ C \end{pmatrix}=n

for every finite generalized eigenvalue λ\lambda of the pencil sE−AsE-A.

(4) The triple (E,A,C)(E,A,C) is said to be impulse observable if

rank⁡(EA0E0C)=n+rank⁡E.\operatorname{rank}\begin{pmatrix} E & A \\ 0 & E \\ 0 & C \end{pmatrix}=n+\operatorname{rank}E.

References

  1. [1]D. Alpay and D. S. Kalyuzhnyi-Verbovetzkii. Matrix-J-unitary non-commutative rational formal power series. In The state space method generalizations and applications, volume 161 of Oper. Theory Adv. Appl., pages 49–113. Birkhäuser, Basel, 2006.DOI
  2. [2]B. D. O. Anderson and S. Vongpanitlerd. Networks Analysis and Synthesis: A Modern Systems Theory Approach. Prentice-Hall, New Jersey, 1973.
  3. [3]T. Ando. Generalized Schur complements. Linear Algebra Appl., 27:173–186, 1979.DOI
  4. [4]Joseph A. Ball, Israel Gohberg, and Leiba Rodman. Interpolation of rational matrix functions, volume 45 of Operator Theory: Advances and Applications. Birkhäuser Verlag, Basel, 1990.DOI
  5. [5]V. Belevitch. Classical Network Theory. Holden-Day, San Francisco, 1968.
  6. [6]Otto Brune. Synthesis of a finite two-terminal network whose driving-point impedance is a prescribed function of frequency. Journal of Mathematics and Physics, 10:191–236, 1931.DOI
  7. [7]Wilhelm Cauer. Die Verwirklichung der Wechselstromwiderstände Vorgeschriebener Frequenzabhängigkeit. Archiv für Elektrotechnik, 17:355–388, 1926.
  8. [8]Man Duen Choi. Completely positive linear maps on complex matrices. Linear Algebra Appl., 10:285–290, 1975.DOI
  9. [9]N. Cohen and I. Lewkowicz. Convex invertible cones and positive real analytic functions. Linear Algebra Appl., 425:797–813, 2007.DOI
  10. [10]N. Cohen and I. Lewkowicz. The Lyapunov order for real matrices. Linear Algebra Appl., 430:1849–1866, 2009.DOI
  11. [11]Limin Dai. Singular Control Systems, volume 118 of Lecture Notes in Control and Information Sciences. Springer-Verlag, Berlin, 1989.
  12. [12]R. M. Foster. A reactance theorem. Bell System Technical Journal, 3:259–267, 1924.DOI
  13. [13]S. ter Horst and A. Naudé. Realization formulas for involutions of matrix-valued positive real odd functions. IFAC-PapersOnLine, 54(9):172–177, 2021.
  14. [14]S. ter Horst and A. van der Merwe. Convex invertible cones and Nevanlinna–Pick interpolation: The suboptimal case. IFAC-PapersOnLine, 55(30):186–193, 2022.
  15. [15]S. ter Horst and A. van der Merwe. A Hill-Pick matrix criterion for the Lyapunov order. J. Math. Anal. Appl., 515(1):Paper No. 126401, 33, 2022.
  16. [16]Thomas Kailath. Linear Systems. Prentice-Hall, Englewood Cliffs, NJ, 1980.
  17. [17]Dmitry S. Kaliuzhnyi-Verbovetskyi and Victor Vinnikov. Foundations of free non-commutative function theory, volume 199 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2014.
  18. [18]Peter Kunkel and Volker Mehrmann. Differential-Algebraic Equations: Analysis and Numerical Solution. EMS Textbooks in Mathematics. European Mathematical Society, Zürich, 2006.
  19. [19]P. Lancaster and M. Tismenetsky. The Theory of Matrices. Computer Science and Applied Mathematics. Academic Press, Orlando, FL, 2 edition, 1985.
  20. [20]Peter Lancaster and Leiba Rodman. Canonical forms for symmetric/skew-symmetric real matrix pairs under strict equivalence and congruence. Linear Algebra and its Applications, 406:1–76, 2005.DOI
  21. [21]R. W. Newcomb. Linear Multiport Synthesis. McGraw-Hill, 1966.
  22. [22]Victor Vinnikov and Hugo J. Woerdeman. Strictly stable Hurwitz polynomials and their determinantal representations. Indag. Math. (N.S.), 37(2):610–638, 2026.arxiv.org/abs/2411.17526
  23. [23]Jan C. Willems. Dissipative dynamical systems. II. Linear systems with quadratic supply rates. Arch. Rational Mech. Anal., 45:352–393, 1972.DOI
  24. [24]D. C. Youla and M. Saito. Interpolation with positive-real functions. Journal of the Franklin Institute, 284:77–108, 1967.DOI

Paper details

vlnc:W_yeyb44q6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

GPT-5.6 Sol (bare)

View statement
Looking for related papers…
Contents