Section: Scientific Foundations

Positive-dimensional and parametric systems

Participants : J.C. Faugère, D. Lazard, M. Safey El Din, D. Wang.

When a system is positive dimensional (with an infinite number of complex roots), it is no more possible to enumerate the solutions. Therefore, the solving process reduces to decomposing the set of the solutions into subsets which have a well-defined geometry. One may perform such a decomposition from an algebraic point of view or from a geometrical one, the latter meaning not taking the multiplicities into account (structure of primary components of the ideal is lost).

Although there exist algorithms for both approaches, the algebraic point of view is presently out of the possibilities of practical computations, and we restrict ourselves to geometrical decompositions.

When one studies the solutions in an algebraically closed field, the decompositions which are useful are the equidimensional decomposition (which consists in considering separately the isolated solutions, the curves, the surfaces, ...) and the prime decomposition (decomposes the variety into irreducible components). In practice, our team works on algorithms for decomposing the system into regular separable triangular sets, which corresponds to a decomposition into equidimensional but not necessarily irreducible components. These irreducible components may be obtained eventually by using polynomial factorization.

However, in many situations one is looking only for real solutions satisfying some inequalities (P i >0 or P i 0)(In the zero-dimensional case, inequations and inequalities are usually taken into account only at the end of the computation, to eliminate irrelevant solutions.). In this case, there are various kinds of decompositions besides the above ones: connected components, cellular or simplicial decompositions, ...

There are general algorithms for such tasks, which rely on Tarski's quantifier elimination. Unfortunately, these problems have a very high complexity, usually doubly exponential in the number of variables or the number of blocks of quantifiers, and these general algorithms are intractable. It follows that the output of a solver should be restricted to a partial description of the topology or of the geometry of the set of solutions, and our research consists in looking for more specific problems, which are interesting for the applications, and which may be solved with a reasonable complexity.

We focus on 2 main problems:

1. computing one point on each connected components of a semi-algebraic set;

2. solving systems of equalities and inequalities depending on parameters.

Critical point methods

The most widespread algorithm computing sampling points in a semi-algebraic set is the Cylindrical Algebraic Decomposition Algorithm due to Collins  [48] . With slight modifications, this algorithm also solves the problem of Quantifier Elimination. It is based on the recursive elimination of variables one after an other ensuring nice properties between the components of the studied semi-algebraic set and the components of semi-algebraic sets defined by polynomial families obtained by the elimination of variables. It is doubly exponential in the number of variables and its best implementations are limited to problems in 3 or 4 variables.

Since the end of the eighties, alternative strategies (see  [55] , [45] and references therein) with a single exponential complexity in the number of variables have been developed. They are based on the progressive construction of the following subroutines:

(a) solving zero-dimensional systems: this can be performed by computing a lexicographical Grobner basis;

(b) computing sampling points in a real hypersurface: after some infinitesimal deformations, this is reduced to problem (a) by computing the critical locus of a polynomial mapping reaching its extrema on each connected component of the real hypersurface;

(c) computing sampling points in a real algebraic variety defined by a polynomial system: this is reduced to problem (b) by considering the sum of squares of the polynomials;

(d) computing sampling points in a semi-algebraic set: this is reduced to problem (c) by applying an infinitesimal deformation.

On the one hand, the relevance of this approach is based on the fact that its complexity is asymptotically optimal. On the other hand, some important algorithmic developments have been necessary to obtain efficient implementations of subroutines (b) and (c).

During the last years, we focused on providing efficient algorithms solving the problems (b) and (c). The used method rely on finding a polynomial mapping reaching its extrema on each connected component of the studied variety such that its critical locus is zero-dimensional. For example, in the case of a smooth hypersurface whose real counterpart is compact choosing a projection on a line is sufficient. This method is called in the sequel the critical point method. We started by studying problem (b) [65] . Even if we showed that our solution may solve new classes of problems ( [66] ), we have chosen to skip the reduction to problem (b), which is now considered as a particular case of problem (c), in order to avoid an artificial growth of degree and the introduction of singularities and infinitesimals.

Putting the critical point method into practice in the general case requires to drop some hypotheses. First, the compactness assumption, which is in fact intimately related to an implicit properness assumption, has to be dropped. Second, algebraic characterizations of critical loci are based on assumptions of non-degeneracy on the rank of the Jacobian matrix associated to the studied polynomial system. These hypotheses are not satisfied as soon as this system defines a non-radical ideal and/or a non equidimensional variety, and/or a non-smooth variety. Our contributions consist in overcoming efficiently these obstacles and several strategies have been developed [2] , [7] .

The properness assumption can be dropped by considering the square of a distance function to a generic point instead of a projection function: indeed each connected component contains at least a point minimizing locally this function. Performing a radical and equidimensional decomposition of the ideal generated by the studied polynomial system allows to avoid some degeneracies of its associated Jacobian matrix. At last, the recursive study of overlapped singular loci allows to deal with the case of non-smooth varieties. These algorithmic issues allow to obtain a first algorithm [2] with reasonable practical performances.

Since projection functions are linear while the distance function is quadratic, computing their critical points is easier. Thus, we have also investigated their use. A first approach [7] consists in studying recursively the critical locus of projection functions on overlapped affine subspaces containing coordinate axes combined with the study of their set of non-properness. A more efficient one  [67] , avoiding the study of sets of non-properness is obtained by considering iteratively projections on generic affine subspaces restricted to the studied variety and fibers on arbitrary points of these subspaces intersected with the critical locus of the corresponding projection. The underlying algorithm is the most efficient we obtained.

In terms of complexity, we have proved in  [68] that when the studied polynomial system generates a radical ideal and defines a smooth algebraic variety, the output of our algorithms is smaller than what could be expected by applying the classical Bèzout bound and than the output of the previous algorithms. This has also given new upper bounds on the number of connected components of a smooth real algebraic variety which improve the classical Thom-Milnor bound. The technique we used, also allows to prove that the degree of the critical locus of a projection function is inferior or equal to the degree of the critical locus of a distance function. Finally, it shows how to drop the assumption of equidimensionality required in the aforementioned algorithms.

Parametric systems

Most of the applications we recently solved (celestial mechanics, cuspidal robots, statistics, etc.) require the study of semi-algebraic systems depending on parameters. Although we covered these subjects in an independent way, some general algorithms for the resolution of this type of systems can be proposed from these experiments.

The general philosophy consists in studying the generic solutions independently from algebraic subvarieties (which we call from now on discriminant varieties) of dimension lower than the semi-algebraic set considered. The study of the varieties thus excluded can be done separately to obtain a complete answer to the problem, or is simply neglected if one is interested only in the generic solutions, which is the case in some applications.

We recently proposed a new framework for studying basic constructible (resp. semi-algebraic) sets defined as systems of equations and inequations (resp. inequalities) depending on parameters. Let's consider the basic semi-algebraic set

𝒮={x n ,p 1 (x)=0,...,p s (x)=0,f 1 (x)>0,...f s (x)>0}

and the basic constructible set

𝒞={x n ,p 1 (x)=0,...,p s (x)=0,f 1 (x)0,...f s (x)0}

where p i ,f j are polynomials with rational coefficients.

  • [U,X]=[U 1 ,...U d ,X d+1 ,...X n ] is the set of indeterminates or variables, U=[U 1 ,...U d ] is the set of parameters and X=[X d+1 ,...X n ] the set of unknowns;

  • ={p 1 ,...p s } is the set of polynomials defining the equations;

  • ={f 1 ,...f l } is the set of polynomials defining the inequations in the complex case (resp. the inequalities in the real case);

  • For any uC d let φ u be the specialization Uu;

  • Π U : n d denotes the canonical projection on the parameter's space

    (u 1 ,,u d ,x d+1 ,...,x n )(u 1 ,,u d );

  • Given any ideal I we denote by 𝐕(I) n the associated (algebraic) variety. If a variety is defined as the zero set of polynomials with coefficients in we call it a -algebraic variety; we extend naturally this notation in order to talk about -irreducible components, -Zariski closure, etc.

  • for any set 𝒱 n , 𝒱 ¯ will denote its -Zariski closure in n .

In most applications, 𝐕(<φ u ()>)) as well as φ u (𝒞)=Π U -1 (u)𝒞 are finite and not empty for almost all parameter's u. Most algorithms that study 𝒞 or 𝒮 (number of real roots w.r.t. the parameters, parameterizations of the solutions, etc.) compute in any case a -Zariski closed set WC d such that for any u d W, there exists a neighborhood 𝒰 of u with the following properties :

  • (Π U -1 (𝒰)𝒞,Π U ) is an analytic covering of 𝒰; this implies that the elements of do not vanish (and so have constant sign in the real case) on the connected components of Π U -1 (𝒰)𝒞;

We recently [6] show that the parameters' set such that there doesn't exist any neighborhood 𝒰 with the above analytic covering property is a -Zariski closed set which can exactly be computed. We name it the minimal discriminant variety of 𝒞 with respect to Π U and propose also a definition in the case of non generically zero-dimensional systems.

Being able to compute the minimal discriminant variety allows to simplify the problem depending on n variables to a similar problem depending on d variables (the parameters) : it is sufficient to describe its complementary in the parameters' space (or in the closure of the projection of the variety in the general case) to get the full information about the generic solutions (here generic means for parameters' values outside the discriminant variety).

Then being able to describe the connected components of the complementary of the discriminant variety in d becomes a main challenge which is strongly linked to the work done on positive dimensional systems. Moreover, rewriting the systems involved and solving zero-dimensional systems are major components of the algorithms we plan to build up.

We currently propose several computational strategies. An a priori decomposition into equidimensional components as zeros of radical ideals simplifies the computation and the use of the discriminant varieties. This preliminary computation is however sometimes expensive, so we are developing adaptive solutions where such decompositions are called by need. The main progress is that the resulting methods are fast on easy problems (generic) and slower on the problems with strong geometrical contents.

The existing implementations of algorithms able to "solve" (to get some information about the roots) parametric systems do all compute (directly or indirectly) discriminant varieties but none computes optimal objects (strict discriminant variety). This is the case, for example of the Cylindrical Algebraic Decomposition adapted to [48] , of algorithms based on "Comprehensive Gröbner bases" [72] , [73] , [71] or of methods that compute parameterizations of the solutions (see [69] for example). The consequence is that the output (case distinctions w.r.t. parameters' values) are huge compared with the results we can provide.