Section: Scientific Foundations
Identification and approximation
Identification typically consists in approximating experimental data by the prediction of a model belonging to some model class. It consists therefore of two steps, namely the choice of a suitable model class and the determination of a model in the class that fits best with the data. The ability to solve this approximation problem, often non-trivial and ill-posed, impinges on the effectiveness of a method.
Particular attention is payed within the team to the class of stable linear time-invariant systems, in particular resonant ones, and in isotropically diffusive systems, with techniques that dwell on functional and harmonic analysis. In fact one often restricts to a smaller class—e.g., rational models of suitable degree (resonant systems, see section 4.2 ) or other structural constraints—and this leads us to split the identification problem in two consecutive steps:
-
Seek a stable but infinite (numerically: high) dimensional model to fit the data. Mathematically speaking, this step consists in reconstructing a function analytic in the right half-plane or in the unit disk (the transfer function), from its values on an interval of the imaginary axis or of the unit circle (the band-width). We will embed this classical ill-posed issue (i.e., the inverse Cauchy problem for the Laplace equation) into a family of well-posed extremal problems, that may be viewed as a regularization scheme of Tikhonov-type. These problems are infinite-dimensional but convex (see section 3.1.1 ).
-
Approximate the above model by a lower order one reflecting further known properties of the physical system. This step aims at reducing the complexity while bringing physical significance to the design parameters. It typically consists of a rational or meromorphic approximation procedure with prescribed number of poles in certain classes of analytic functions. Rational approximation in the complex domain is a classical but difficult non-convex problem, for which few effective methods exist. In relation to system theory, two specific difficulties superimpose on the classical situation, namely one must control the region where the poles of the approximants lie in order to ensure the stability of the model, and one has to handle matrix-valued functions when the system has several inputs and outputs, in which case the number of poles must be replaced by the McMillan degree (see section 3.1.2 ).
When identifying elliptic (Laplace, conjugate-Beltrami) partial differential equations from boundary data, point 1. above can be recast as an inverse boundary-value problem with (overdetermined Dirichlet-Neumann) data on part of the boundary of a plane domain (recover a function, analytic in a domain, from incomplete boundary data). As such, it arises naturally in higher dimensions when analytic functions get replaced by gradients of harmonic functions (see section 4.1 ). Initial motivations of the team include:
-
free boundary problems in plasma control;
-
the recovery of sources, that arises for instance in magneto/electro-encephalography;
-
the detection of cracks and occlusions in non-destructive control.
We aim at generalizing this approach to the conjugate-Beltrami equation in dimension 2 (section 6.2.3 ) and to the Laplace equation in dimension 3 (section 6.2.1 ).
Step 2. above, i.e., meromorphic approximation with prescribed number of poles—is used to approach other inverse problems beyond harmonic identification. In fact, the way the singularities of the approximant (i.e., its poles) relate to the singularities of the approximated function is an all-pervasive theme in approximation theory: for appropriate classes of functions, the location of the poles of the approximant can be used as an estimator of the singularities of the approximated function (see section 6.2.2 ).
We provide further details on the two steps mentioned above in the sub-paragraphs to come.
Analytic approximation of incomplete boundary data
Participants : Laurent Baratchart, Slah Chaabi, Sylvain Chevillard, Yannick Fischer [Until November] , Juliette Leblond, Jean-Paul Marmorat, Jonathan Partington, Elodie Pozzi [Since October] , Fabien Seyfert.
Given a planar domain
A standard extremal problem on the disk is [61] :
(
) Let and ; find a function such that is of minimal norm in .
When seeking an analytic function in
(
) Let , a sub-arc of , , and ; find a function such that and is of minimal norm in under this constraint.
Here
To fix terminology we generically refer to (
Various modifications of
The above-mentioned problems can be stated on an annular geometry rather
than on a disk.
For
Continuing effort is currently payed by the team to carry over bounded extremal problems and their solution to more general settings.
Such generalizations are twofold: on the one hand
Apics considers 2-D diffusion equations with variable (but for now isotropic) conductivity,
on the other hand it investigates the ordinary Laplacian in
An isotropic diffusion equation in dimension 2 can be recast as a
so-called conjugate or real Beltrami equation [70] .
This way
analytic functions get replaced by “generalized” ones in problems (
At present, bounded extremal problems for the
Meromorphic and rational approximation
Participants : Laurent Baratchart, José Grimm, Martine Olivi, Edward Saff, Herbert Stahl [TFH Berlin] , Maxim Yattselev.
Let as before
A natural generalization of problem (
(
) Let , an integer, and ; find a function such that is of minimal norm in .
Problem (
Only for
The case
The former Miaou project (predecessor of Apics) has designed an
adapted steepest-descent algorithm
for the case
In order to establish convergence results of the algorithm to the global minimum, Apics has undergone a
long-haul study of the number and nature of critical points, in which
tools from differential topology and
operator theory team up with classical approximation theory.
The main discovery is that
the nature of the critical points
(e.g., local minima, saddles...)
depends on the decrease of the interpolation
error to
A common feature to all these problems is that critical point equations express non-Hermitian orthogonality relations for the denominator of the approximant. This is used in an essential manner to assess the behavior of the poles of the approximants to functions with branched singularities, which is of particular interest for inverse source problems (cf. section 6.2.2 ).
In higher dimensions, the analog of problem (
Certain constrained rational approximation problems, of special interest
in identification
and design of passive systems, arise when putting additional
requirements on the approximant, for instance that it should be smaller than 1
in modulus.
Such questions have become over years an increasingly significant
part of the team's
activity (see section
4.2 ).
When translated over to the circle, a prototypical formulation
consists in approximating the modulus of a given function by the modulus of a
rational function of degree
Behavior of poles of meromorphic approximants and inverse problems for the Laplacian
Participants : Laurent Baratchart, Herbert Stahl [TFH Berlin] , Maxim Yattselev.
We refer here to the behavior of the poles of best
meromorphic approximants, in the
Generally speaking, the behavior of poles is particularly important in meromorphic approximation to obtain error rates as the degree goes large and also to tackle constructive issues like uniqueness. However, the original motivation of Apics is to consider this issue in connection with the approximation of the solution to a Dirichlet-Neumann problem, so as to extract information on the singularities. The general theme is thus how do the singularities of the approximant reflect those of the approximated function? The approach to inverse problem for the 2-D Laplacian that we outline here is attractive when the singularities are zero- or one-dimensional (see section 4.1 ). It can be used as a computationally cheap preliminary step to obtain the initial guess of a more precise but heavier numerical optimization.
As regards crack detection or source recovery, the approach in question boils down to analyzing the behavior of best meromorphic approximants of a function with branch points. We were able to prove ([6] , [9] ) that the poles of the approximants accumulate in a neighborhood of the geodesic hyperbolic arc that links the endpoints of the crack, or the sources [46] . Moreover, the asymptotic density of the poles turns out to be the equilibrium distribution on the geodesic arc of the Green potential and this distribution puts heavy charge at the end points, that are thus well localized if one is able to compute sufficiently many zeros (this is where the method could fail). The case of more general cracks, as well as situations when three or more sources, require handling a finite but arbitrary number of branch points. These are outstanding open questions for applications to inverse problems (see section 6.2 ), as for the problem of a general singularity, that may be two dimensional.
Results of this type open new perspectives in non-destructive control, in that they link issues of current interest in approximation theory (the behavior of zeroes of non-Hermitian orthogonal polynomials) to some classical inverse problems for which a dual approach is thereby proposed: to approximate the boundary conditions by true solutions of the equations, rather than approximating (by discretization) the equation itself.
We wish to point out that rational or meromorphic approximation to the Cauchy transform of measure
can be viewed as discretization of the logarithmic potential of that measure.
If approximation takes place in the
Matrix-valued rational approximation
Participants : Laurent Baratchart, Martine Olivi, José Grimm, Jean-Paul Marmorat, Bernard Hanzon, Ralf Peeters [Univ. Maastricht] .
Matrix-valued approximation is necessary for handling systems with several inputs and outputs, and it generates substantial additional difficulties with respect to scalar approximation, theoretically as well as algorithmically. In the matrix case, the McMillan degree (i.e., the degree of a minimal realization in the System-Theoretic sense) generalizes the degree.
The problem we want to consider reads:
Let
The approximation algorithm
designed in the scalar case generalizes to
the matrix-valued situation [60] . The
first difficulty here consists in the parametrization
of transfer matrices of given
McMillan degree
The set of inner matrices of given degree has the structure of a smooth manifold that allows one to use differential tools as in the scalar case. In practice, one has to produce an atlas of charts (parametrization valid in a neighborhood of a point), and we must handle changes of charts in the course of the algorithm. Such parametrization can be obtained from interpolation theory and Schur type algorithms, the parameters being interpolation vectors or matrices ( [35] , [15] , [16] ). Some of these parametrizations have a particular interest for computation of realizations ([15] , [16] ), involved in the estimation of physical quantities for the synthesis of resonant filters. Two rational approximation software codes (see sections 5.2 and 5.5 ) have been developed in the team.
Problems relative to multiple local minima naturally arise in the matrix-valued case as well, but deriving criteria that guarantee uniqueness is even more difficult than in the scalar case. The already investigated case of rational functions of the sought degree (the consistency problem) was solved using rather heavy machinery [10] . The case of matrix-valued Markov functions, the first example beyond rational functions, has undergone progress only recently [40] .
Let us stress that the algorithms mentioned above are first to handle rational approximation in the matrix case in a way that converges to local minima, while meeting stability constraints on the approximant.