lecture_18
lecture_18
Lecture 18
Lecturer: Pablo A. Parrilo Scribe: ???
Quantifier elimination (QE) is a very powerful procedure for problems involving firstorder formulas
over real fields. The cylindrical algebraic decomposition (CAD) is a technique for the “efficient” im
plementation of QE, that effectively reduces an seemingly infinite problem into a finite (but potentially
large) instance. For much more information about QE and CAD (including a reprint of Tarski’s original
1930 work), we recommend the book [CJ98].
1 Quantifier elimination
A quantifierfree formula is an expression consisting of polynomial equations (f (x) = 0) and inequalities
(f (x) ≤ 0) combined using the Boolean operators ∧ (and), ∨ (or), and ⇒ (implies). We often also allow
strict inequalities f (x) > 0 and inequations f (x) �= 0, since these are just shorthands for particular
boolean combinations of equations and inequalities.
In general, a formula (in prenex form) is an expression in the variables x = (x1 , ..., xn ) of the following
type:
(Q1 x1 )...(Qs xs ) F (f1 (x), ..., fr (x)) (1)
where Qi is one of the quantifiers ∀ (for all) and ∃ (there exists). Furthermore, F (f1 (x), ..., fr (x)) is
assumed to be a quantifierfree formula. If there is a quantifier corresponding to the variable xi , we say
that xi is quantified, or free otherwise.
Example 1. The following are valid formulas
The first formula has two free variables (since the variables a and b are unquantified), while for the other
two all variables are quantified.
We will interpret the symbols in a formula as taking only real values. Notice that a formula without
free variables (usualled called a closed formula or a sentence) is either true or false. For instance, the
last two expressions in Example 1 are sentences, with the first one being false and the second being true.
Notice also that the truth value may depend on the order of the quantifiers.
Tarski showed that for every formula including quantifiers there is always an equivalent quantifier
free formula. Obtaining the latter from the former is called quantifier elimination.
Theorem 2 (TarskiSeidenberg). For every firstorder formula over the real field there exists an equiv
alent quantifierfree formula. Furthermore, there is an explicit algorithm to compute this quantifierfree
formula.
The TarskiSeidenberg theorem is an extremely powerful result, since it provides a complete charac
terization and algorithmic technique for an extremely large collection of problems involving polynomials.
Unfortunately, there are very serious computational barriers to the efficient practical implementation of
these ideas, since the resulting algorithms have extremely poor scaling properties, with respect to the
number of variables (towers of exponentials). Newer methods, such as the (partial) cylindrical algebraic
decomposition (CAD) technique due to Collins and described below, or the critical point method, are by
comparison much better. Nevertheless, by necessity they still behave exponentially (or worse) in terms
of the number of variables.
181
2 TarskiSeidenberg
Example 3. Consider the quantified firstorder formula:
(a ≥ 0) ∧ (a3 − 8a − 16 ≤ 0),
which defines the interval [0, a� ], where a� ≈ 3.538. Thus, the original expression (2) is true only for
a ∈ [0, a� ].
2.2 Applications
Static output feedback An early application of TarskiSeidenberg in control theory was the “so
lution” of the static output feedback stabilization problem in [ABJ75]. Given matrices A ∈ Rn×n ,
B ∈ Rn×m , we want to find a matrix K ∈ Rm×n such that the matrix A + BK is Hurwitz, i.e., all its
eigenvalues are in the lefthand plane. Since the existence of such a matrix can be easily expressed as a
formula in first order logic1 , the decidability and existence of an effective (but not efficient) algorithm
immediately follows.
Simultaneous stabilization A very interesting result by Blondel [Blo94, BG93] shows that the si
multaneous stabilization of three linear timeinvariant systems is not decidable (and thus, cannot be
semialgebraic). Notice however that, for any given bound on the degree of the controller, the problem
is decidable.
notation, since for a fully real formulation, we should split x and λ into real and imaginary parts. There are many other
equivalent expressions, using for instance a Lyapunov equation, or the Routh array.
182
3.0.1 Description
Given a set P of multivariate polynomials in n variables, a CAD is a special partition of Rn into com
ponents, called cells, over which all the polynomials have constant signs. The algorithm for computing
a CAD also provides a point in each cell, called sample point, which can be used to determine the sign
of the polynomials in the cell.
A cell is called cylindrical if it has the form S × Rk , for some k ≤ n. A decomposition of Rn is a
CAD if all polynomials have constant sign on each cell, and all cells are cylindrical.
The CAD associated to the formula (1) depends only on its quantifier free part F (f1 (x), ..., fr (x)).
Since all possible truth values of the formula are in correspondence with the values at the sample points,
we can use the CAD to evaluate its truth value, and to perform quantifier elimination.
The basic CAD construction consists of two steps: projection and lifting (plus an additional third
one, if formula construction is desired).
In the first projection phase, we computes successive sets of polynomials in n − 1, n − 2, ..., 1 variables.
The main idea is, given an input set of polynomials, to compute at each step a new set of polynomials
obtained by eliminating one variable at a time. In general, the elimination order does matter and a good
choice leads to lower computational complexity.
The second phase (lifting) constructs a decomposition of R, at the lowest level of projection, af
ter all but one variable have been eliminated. This decomposition of R is successively extended to a
decomposition of Rn .
The basic operations necessary in the construction of CADs are (sub)resultants and (sub)discriminants.
Complete ToDo
An implementation of (an improved version of) the CAD method for quantifier elimination is the
software package QEPCAD [Bro03].
References
[ABJ75] B. D. O. Anderson, N. K. Bose, and E. I. Jury. Output feedback stabilization and related
problems—solution via decision methods. IEEE Transactions on Automatic Control, 20:53–
66, 1975.
[BG93] V. Blondel and M. Gevers. Simultaneous stabilizability of three linear systems is rationally
undecidable. Mathematics of Control, Signals, and Systems, 6(2):135–145, 1993.
[Blo94] V. Blondel. Simultaneous stabilization of linear systems, volume 191 of Lecture Notes in Control
and Information Sciences. SpringerVerlag London Ltd., London, 1994.
[BPR03] S. Basu, R. Pollack, and M.F. Roy. Algorithms in real algebraic geometry, volume 10 of
Algorithms and Computation in Mathematics. SpringerVerlag, Berlin, 2003.
[Bro03] C.W. Brown. QEPCAD Quantifier Elimination by Partial Cylindrical Algebraic Decomposi
tion, 2003. Available from http://www.cs.usna.edu/~qepcad/B/QEPCAD.html.
[CJ98] B. F. Caviness and J. R. Johnson, editors. Quantifier elimination and cylindrical algebraic de
composition, Texts and Monographs in Symbolic Computation, Vienna, 1998. SpringerVerlag.
[Col75] G. E. Collins. Quantifier elimination for real closed fields by cylindrical algebraic decomposition.
In Automata theory and formal languages (Second GI Conf., Kaiserslautern, 1975), pages 134–
183. Lecture Notes in Comput. Sci., Vol. 33. Springer, Berlin, 1975.
183
[Ren91] J. Renegar. Recent progress on the complexity of the decision problem for the reals. In
Discrete and computational geometry (New Brunswick, NJ, 1989/1990), volume 6 of DIMACS
Ser. Discrete Math. Theoret. Comput. Sci., pages 287–308. Amer. Math. Soc., Providence, RI,
1991.
184