1: %%%%% Artikel mit Gregorio ueber multi-hom. Bezoutzahl
2: %%%%% volle Version
3: %%%%% letzte Version: 29.4.2004
4:
5: \documentclass[10pt]{article}
6: \usepackage{amsmath, amssymb, amsthm, amscd}
7: \usepackage{amsrefs}
8: \usepackage{graphics}
9:
10:
11:
12: \textheight20cm
13: \textwidth13.3cm
14:
15:
16: \oddsidemargin1.3cm
17: \evensidemargin0.6cm
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29: \newcommand{\binomial}[2] {{\left(\begin{array}{c}#1\\#2\end{array}\right)}}
30:
31:
32: \newtheorem{theorem} {Theorem}
33: \newtheorem{restatedtheorem}{Theorem}
34: \newtheorem{subtheorem} {Theorem}
35: \newtheorem*{theorem*} {Theorem}
36: \newtheorem*{main} {Main Theorem}
37: \newtheorem{lemma} {Lemma}
38: \newtheorem{lowerbound} {Lower bound}
39: \newtheorem{proposition} {Proposition}
40: \newtheorem{corollary} {Corollary}
41: %\numberwithin{corollary}{theorem}
42:
43: \newtheorem{conjecture} {Conjecture}
44: \newtheorem{hypothesis} {Working Hypothesis}
45: \newtheorem{assumption} {Assumption}
46: \newtheorem*{claim} {Claim}
47: \newtheorem{algorithm} {Algorithm}
48:
49:
50: \theoremstyle{definition}
51: \newtheorem{definition} {Definition}
52: \newtheorem{example} {Example}
53: \newtheorem{remark} {Remark}
54: \newtheorem{induction} {Induction Hypothesis}
55: \newtheorem{question} {Question}
56: \newtheorem{problem} {Problem}
57: \newtheorem{openproblem} {Open Problem}
58:
59:
60:
61:
62:
63:
64: %%%%%%%%%%%%% Definitions KLaus
65:
66: \newcommand{\R}{\mathbb{R}}
67: \newcommand{\Z}{\mathbb{Z}}
68: \newcommand{\N}{\mathbb{N}}
69: \newcommand{\Q}{\mathbb{Q}}
70: \newcommand{\C}{\mathbb{C}}
71:
72:
73: %%%%%%%%%%%%% Redefinitions Greg. Use math mode
74: %%%%%%%%%%%%% for better results.
75:
76: \newcommand{\CompClass}[1]{\ensuremath{\mathbf {#1}}}
77: \newcommand{\NPO}{\CompClass{NPO}}
78: \newcommand{\PTAS}{\CompClass{PTAS}}
79: \newcommand{\APX}{\CompClass{APX}}
80:
81: \newcommand{\NP}{\CompClass{NP}}
82: \newcommand{\BPP}{\CompClass{BPP}}
83: \newcommand{\Po}{\CompClass{P}}
84: \newcommand{\slashp}{\CompClass{\# P}}
85: %%%%%%%%%%%%%%%
86:
87:
88: \newcommand{\bez}{\mathrm{B\acute ez}}
89: \newcommand{\vol}{\mathrm{Vol}}
90: \newcommand{\conv}{\mathrm{Conv}}
91:
92:
93:
94: \newcommand{\boldalpha}{\boldsymbol{\alpha}}
95: \newcommand{\partition}[1]{\boldsymbol #1}
96:
97: \newcommand{\bydef}{\ensuremath{\stackrel{\mbox{\scriptsize def}}{=}}}
98:
99:
100:
101:
102: \begin{document}
103: \author{Gregorio Malajovich\footnote{
104: {\bf Postal address:}
105: Departamento de Matem\'atica Aplicada,
106: Universidade Federal do Rio de Janeiro.
107: Caixa Postal 68530, CEP 21945-970, Rio de Janeiro,
108: RJ, Brasil. {\bf email:}{\tt gregorio@ufrj.br}.
109: {\bf url:} {\tt www.labma.ufrj.br//\~{ }gregorio}.
110: {\bf Acknowledgements:} G.M. is partially supported
111: by CNPq (Brasil). This
112: work was done while visiting
113: Syddansk Universitet at Odense, thanks to the
114: generous support of the
115: Villum Kann Rasmussen Fond.} \\
116: \and Klaus Meer
117: \footnote{{\bf Postal address:}
118: Department of Mathematics and Computer Science.
119: University of Southern Denmark,
120: Campusvej 55,
121: DK-5230, Odense M.
122: Denmark.
123: {\bf email:}{\tt meer@imada.sdu.dk}.
124: {\bf url:} {\tt www.imada.sdu.dk//\~{ }meer}.
125: {\bf Acknowledgements:}
126: Partially supported
127: by the EU Network of Excellence PASCAL
128: Pattern Analysis, Statistical Modelling and Computational Learning
129: and by the Danish Natural Science Research Council SNF.}}
130:
131:
132: \title{Computing Multi-Homogeneous B\'ezout Numbers is Hard}
133:
134: \date{29 april 2004}
135: \maketitle
136:
137: \begin{abstract}
138: The multi-homogeneous B\'ezout number is a bound for
139: the number of solutions of a system of multi-homogeneous
140: polynomial equations, in a suitable product of projective
141: spaces.
142:
143: Given an arbitrary, not necessarily multi-homogeneous
144: system, one can ask for the optimal multi-homogenization
145: that would minimize the B\'ezout number.
146:
147: In this paper, it is proved that the problem of computing,
148: or even estimating the optimal multi-homogeneous B\'ezout
149: number is actually $\NP$-hard.
150:
151: In terms of approximation theory for combinatorial
152: optimization, the problem of computing the best
153: multi-homogeneous structure does not belong to $\APX$,
154: unless $\Po = \NP$.
155:
156: Moreover, polynomial time algorithms for estimating
157: the minimal multi-homo\-ge\-neous B\'ezout number up to a
158: fixed factor cannot exist even in a randomized setting,
159: unless $\BPP \supseteq \NP$.
160: \end{abstract}
161:
162: \section{Introduction}
163:
164: The multi-homogeneous B\'ezout number is a bound for
165: the number of solutions of a system of multi-homogeneous
166: polynomial equations.
167:
168: Estimating the number of isolated solutions of a polynomial
169: system is useful for the design and analysis of
170: homotopy algorithms~\cite{LI}. Applications include
171: problems in engineering like the design of certain mechanisms
172: ~\cites{MORGAN, WMS} or others, such as computational
173: geometry.
174:
175: An application of
176: multi-homogeneous B\'ezout bounds
177: outside the realm
178: of algebraic equation solving is discussed in~\cite{DMS},
179: where the number of roots
180: is used to bound geometrical quantities such as
181: volume and curvature.
182:
183: There is an important connection between root-counting and
184: $\NP$-completeness theory. Indeed, it is easy to
185: reduce an $\NP$-complete or $\NP$-hard
186: problem such as SAT, the Traveling
187: Salesman problem, Integer Programming (and thus all
188: other
189: $\NP$ problems as well) to the question whether certain
190: polynomial
191: systems have a common zero.
192:
193: \bigskip
194: \par
195:
196:
197:
198: The best-known example giving an estimate for the number of roots
199: of a polynomial equation is the
200: Fundamental Theorem of Algebra. It was generalized
201: to multivariate polynomial systems at the
202: end of the 18th century by Etienne B\'ezout. The B\'ezout number
203: bounds the number of (isolated) complex solutions of a polynomial
204: $f : \mathbb C^n \mapsto \mathbb C^n$ from above by the product of the degrees
205: of the involved polynomials. However, in many cases this estimate is
206: far from optimal. A well known example is given by the eigenvalue
207: problem: Given a $n \times n$
208: matrix $M$, find the eigenpairs $(\lambda, u) \in
209: \mathbb C \times \mathbb C^n$ such that
210: $M u - \lambda u = 0$. If we
211: equate $u_n$ to $1$,
212: the classical B\'ezout number becomes $2^{n-1},$ though of course only
213: $n$ solutions exist.
214:
215: The multi-homogeneous B\'ezout number provides a sharper bound on
216: the number of isolated solutions of a system of equations,
217: in a suitable product of projective spaces.
218: The multi-homogeneous B\'ezout bound depends on
219: the choice of a {\em multi-homogeneous structure},
220: that is of a partition of the variables $(\lambda, u)$
221: into several groups.
222:
223:
224: In the eigenvalue example,
225: the eigenvector $u$ is defined up to a multiplicative
226: constant, so it makes sense to define it as an element
227: of $\mathbb P^{n-1}$. With respect to the eigenvector
228: $\lambda$, we need to introduce a homogenizing variable.
229: We therefore rewrite the equation as:
230: $\lambda_0 M u - \lambda_1 u = 0$, and $\lambda =
231: \lambda_1 / \lambda_0$. Now the pair $(\lambda_0 :
232: \lambda_1)$ is an element of $\mathbb P^1$. The
233: multi-homogeneous B\'ezout number for this system
234: is precisely $n$.
235:
236:
237: %%%%%%%%%%%%%
238: %
239: %The multi-homogeneous B\'ezout number is a bound on
240: %the number of isolated solutions of a system of
241: %multi-homogeneous equations, in a suitable product
242: %of projective spaces.
243: %
244: %\medskip
245: %
246: %
247: %A well-known example (\cite[Eq. (2.2)]{LI-BAI}) is
248: %given by the eigenvalue problem: given a $n \times n$
249: %matrix $M$, find the eigenpairs $(\lambda, u) \in
250: %\mathbb C \times \mathbb C^n$ such that
251: %$M u - \lambda u = 0$.
252: %
253: %The eigenvector $u$ is defined up to a multiplicative
254: %constant, so it makes sense to define it as an element
255: %of $\mathbb P^{n-1}$. With respect to the eigenvalue
256: %$\lambda$, we need to introduce a homogenizing variable.
257: %We therefore rewrite the equation as:
258: %$\lambda_0 M u - \lambda_1 u = 0$, and $\lambda =
259: %\lambda_1 / \lambda_0$. Now the pair $(\lambda_0 :
260: %\lambda_1)$ is an element of $\mathbb P^1$. The
261: %multi-homogeneous B\'ezout number for this system
262: %is precisely $n$.
263:
264: \medskip
265:
266: Better bounds on the root number
267: are known, such as
268: Kushnirenko's~\cite{KUSHNIRENKO} or Bernstein's
269: ~\cite{BERNSTEIN}.
270: However, interest in computing the multi-homogeneous B\'ezout
271: number stems from the fact that hardness results are
272: known for those sharper bounds
273: (see section~\ref{sec:volume}
274: for details).
275:
276: Another reason of interest is that in many cases, a
277: natural multi-homogeneous structure is known or may
278: be found with some additional human work.
279:
280: \medskip
281:
282: In this paper, we consider the following problem.
283: {\em
284: Let $n \in \mathbb N$ and a finite $A \subset \mathbb N^n$
285: be given as input. Find the minimal multi-homogeneous
286: B\'ezout number, among all choices of a
287: multi-homogeneous structure for a polynomial system
288: with support $A$:}
289: \begin{equation}
290: \label{eq:*}
291: \left\{
292: \begin{array}{lcl}
293: f_1(z) &=& \sum_{\mathbf \boldalpha \in A} f_{1 \mathbf \boldalpha}
294: z_1^{\alpha_1} z_2^{\alpha_2} \cdots z_n^{\alpha_n} \\
295: & \vdots & \\
296: f_n(z) &=& \sum_{\mathbf \boldalpha \in A} f_{n \mathbf \boldalpha}
297: z_1^{\alpha_1} z_2^{\alpha_2} \cdots z_n^{\alpha_n} \ .\\
298: \end{array}
299: \right.
300: \end{equation}
301:
302: where the $f_{i \mathbf \alpha}$ are non-zero complex coefficients.
303:
304: Geometrically, this minimal B\'ezout number is an
305: upper bound for the number of isolated roots of the
306: system~(\ref{eq:*}) in $\mathbb C^n$.
307: \medskip
308:
309:
310: The main result in this paper (restated formally in section~\ref{sec:bezout}
311: below) is:
312:
313: \begin{theorem}\label{th:A}
314: There cannot possibly exist a polynomial
315: time algorithm to approximate the minimal multi-homogeneous
316: B\'ezout number for (\ref{eq:*}) up to any fixed factor,
317: unless $\Po = \NP$.
318: \end{theorem}
319:
320:
321: This means that computing or even approximating the
322: minimal B\'ezout number up to a fixed factor is
323: \NP-hard. In terms of the hierarchy of approximation
324: classes (see \cite{AUSIELLO} and section~\ref{sec:approx}),
325: the minimal multi-homogeneous B\'ezout number does not
326: belong to the class $\APX$ unless $\Po = \NP$.
327: \medskip
328:
329:
330: Motivated by what is known on volume approximation
331: (see section~\ref{sec:volume}), one could
332: ask whether allowing for randomized algorithms would
333: be of any improvement.
334:
335: \begin{theorem}\label{th:B}
336: There cannot possibly exist a randomized
337: polynomial time algorithm to approximate the
338: minimal multi-homogeneous
339: B\'ezout number for (\ref{eq:*}) up to any fixed factor,
340: with probability of failure $\epsilon < 1/4$,
341: unless $\BPP \supseteq \NP$.
342: \end{theorem}
343:
344:
345: While the conjecture $\BPP \not \supseteq \NP$ is less
346: widely known outside the computer science community
347: than the conjecture $\Po \neq \NP$, its
348: failure would imply the existence of probabilistic
349: polynomial time algorithms for solving problems such
350: as the factorization of large integers or the
351: discrete logarithm. Most widespread cryptographic
352: schemes are based on the assumption that those two
353: problems are hard.
354:
355:
356: \section{Background and Statement of Main Results}
357: \label{sec:background}
358: \subsection{B\'ezout numbers}
359: \label{sec:bezout}
360:
361: In the definition of (\ref{eq:*}), we assumed
362: for simplicity that each equation had the same
363: {\em support} $A$. In general, a system $f(z)$ of
364: $n$ polynomial equations with support $(A_1, \dots, A_n)$
365: is a system of the form:
366:
367: \begin{equation}
368: \label{eq:**}
369: \left\{
370: \begin{array}{lcl}
371: f_1(z) &=& \sum_{\boldalpha \in A_1} f_{1 \mathbf \boldalpha}
372: z_1^{\alpha_1} z_2^{\alpha_2} \cdots z_n^{\alpha_n} \\
373: & \vdots & \\
374: f_n(z) &=& \sum_{\boldalpha \in A_n} f_{n \mathbf \boldalpha}
375: z_1^{\alpha_1} z_2^{\alpha_2} \cdots z_n^{\alpha_n} \ ,\\
376: \end{array}
377: \right.
378: \end{equation}
379: where the coefficients $f_{i \mathbf \boldalpha}$ are non-zero
380: complex numbers.
381:
382: A {\em multi-homogeneous structure} is given by a
383: partition of $\{1, \dots , n\}$ into (say) $k$ sets
384: $I_1, \dots, I_k$. Then for each set $I_j$, we consider
385: the group of variables $Z_j = \{ z_i : i \in I_j \}$.
386:
387: The degree of $f_i$ in the group of variables $Z_j$ is
388: \[
389: d_{ij} \bydef \max_{\boldalpha \in A_i} \ \sum_{l \in I_j}
390: \ \alpha_l
391: \]
392:
393: When for some $j$, for all $i$, the maximum $d_{ij}$ is
394: attained for all $\boldalpha \in A_i$, we say that
395: (\ref{eq:**}) is homogeneous in the variables $Z_j$.
396: The dimension of the projective space associated to
397: $Z_j$ is:
398: \[
399: a_j \bydef \left\{
400: \begin{array}{ll}
401: \# I_j - 1& \text{if (\ref{eq:**}) is homogeneous in $Z_j$, and}
402: \\
403: \# I_j & \text{otherwise.}
404: \end{array}
405: \right.
406: \]
407:
408: We assume that $n = \sum_{j=1}^k a_j$. Otherwise,
409: we would have an undetermined ($n < \sum_{j=1}^k a_j$)
410: or overdetermined ($n > \sum_{j=1}^k a_j$)
411: polynomial system, and multi-homogeneous B\'ezout
412: numbers would have no meaning.
413:
414:
415: The multi-homogeneous B\'ezout number
416: $\bez(A_1, \dots, A_n; I_1, \dots, I_k)$
417: is the coefficient of
418: $\prod_{j=1}^k \zeta_j^{a_j}$ in the formal
419: expression $\prod_{i=1}^n \sum_{j=1}^k d_{ij} \zeta_j$
420: (see~\cites{MORGAN-SOMMESE,LI,SHAFAREVICH}). It bounds
421: the maximal number of
422: isolated roots of (\ref{eq:**}) in
423: $\mathbb P^{a_1} \times \cdots \times \mathbb P^{a_k}$.
424: Therefore it also bounds the number of {\em finite} roots
425: of (\ref{eq:**}), i.e. the roots in $\mathbb C^n$.
426:
427: \medskip
428:
429:
430: In the particular case where $A = A_1 = \cdots = A_n$
431: there is a simpler expression for the multi-homogeneous
432: B\'ezout number $\bez(A;I_1, \dots, I_k) \bydef
433: \bez(A_1, \dots, A_n;$ $I_1, \dots, I_k)$, namely:
434: \begin{equation}\label{eq:bezfromI}
435: \bez(A;I_1, \dots, I_k) =
436: \binomial {n}{a_1\ a_2\ \cdots\ a_k}
437: \
438: \prod_{j=1}^k d_j^{a_j} \ ,
439: \end{equation}
440:
441: where $d_j = d_{ij}$ (equal for each $i$) and the
442: multinomial coefficient
443: \[
444: \binomial {n}{a_1\ a_2\ \cdots\ a_k}
445: \bydef
446: \frac{n!}{a_1!\ a_2!\ \cdots\ a_k!}
447: \]
448: is the coefficient of $\prod_{j=1}^k \zeta_j^{a_k}$
449: in $(\zeta_1 + \cdots + \zeta_k)^{n}$ (recall that
450: $n = \sum_{j=1}^k a_j$).
451:
452: \medskip
453: \par
454:
455: Heuristics for computing a suitable multi-homogeneous
456: structure $(I_1, \dots, I_k)$ given $A_1, \dots, A_n$
457: are discussed in \cites{LI-BAI, LI-LIN-BAI}. Surprisingly
458: enough, there seems to be no theoretical results available
459: on the complexity of computing the minimal B\'ezout number.
460: It was conjectured in \cite[p.78]{LI-BAI} that
461: computing the minimal multi-homogeneous B\'ezout
462: number is $\NP$-hard.
463:
464: Even, no polynomial time algorithm for computing the multi-
465: homogeneous B\'ezout number {\em given a multi-homogeneous
466: structure} seems to be known (see~\cite[p.240]{LI-LIN-BAI}).
467:
468: This is why in this paper, we restrict ourselves to
469: the case $A=A_1 = \cdots = A_n$. This is a particular
470: subset of the general case, and any hardness result for
471: this particular subset implies the same hardness
472: result in the general case.
473:
474:
475: More formally, we adopt the Turing model of computation and
476: we consider the function:
477: \[
478: \bez: n,k,A, I_1, \dots, I_k \mapsto \bez (A; I_1, \dots,
479: I_k)
480: \]
481: where all integer numbers are in binary representation,
482: and $A$ is a list of $n$-tuples
483: $(\alpha_1, \dots, \alpha_n)$, and each $I_j$ is a list
484: of its elements. In particular, the input size is
485: bounded below by $n \# A_i$ and by $\max_{\boldalpha, i}
486: \lceil \log_2 \alpha_i \rceil$.
487: Therefore, $\bez(A; I_1, \cdots, I_k)$ can be computed
488: in polynomial time by a straight-forward application of
489: formula (\ref{eq:bezfromI}). As a matter of fact, it
490: can be computed in time polynomial in the size of $A$.
491:
492: \begin{problem}[Discrete optimization problem]
493: \label{prob:opt}
494: Given $n$ and $A$, compute
495: \[
496: \min_{\partition I} \bez(A; {\partition I}) \ ,
497: \]
498: where $\partition I =(I_1, \dots, I_k)$ ranges over all the partitions
499: of $\{1, \dots , n\}$.
500: \end{problem}
501:
502:
503: \begin{problem}[Approximation problem]
504: \label{prob:approx}
505: Let $C>1$ be fixed. Given $n$ and $A$, compute
506: some $B$ such that
507: \[
508: BC^{-1} < \min \bez(A; \partition I) < BC
509: \]
510: Again, $\partition I = (I_1, \dots, I_k)$ ranges over all the partitions
511: of $\{1, \dots , n\}$.
512: \end{problem}
513:
514: In the problems above, we are not asking for the actual
515: partition.
516:
517: \begin{restatedtheorem}[restated]
518: Problem~\ref{prob:approx} is $\NP$-hard.
519: \end{restatedtheorem}
520:
521: This is actually stronger than the conjecture by
522: Li and Bai~\cite{LI-BAI}, that corresponds to
523: the following immediate
524: corollary:
525:
526: \begin{corollary}
527: Problem~\ref{prob:opt} is $\NP$-hard.
528: \end{corollary}
529:
530: \subsection{Other bounds for the number of roots}
531: \label{sec:volume}
532:
533: Kushnirenko's Theorem~\cite{KUSHNIRENKO} bounds
534: the number of isolated solutions of (\ref{eq:*})
535: in $(\mathbb C^*)^n$ by $n!\ \vol\ \conv A$,
536: where $\conv A$ is the smallest convex polytope
537: containing all the points of $A$.
538:
539: This bound is sharper than the B\'ezout bound, but
540: the known hardness results are far more
541: dramatic: In~\cite{KHACHIYAN}, Khachiyan proved that
542: computing the volume of a polytope given by a set
543: of vertices is $\slashp$-hard.
544:
545: There is a large literature on algorithms for approximating
546: the volume of a convex body given by a separation
547: oracle. The problem of approximating the volume of a polytope
548: in vertex representation can be reduced to the latter by standard
549: linear programming techniques.
550:
551: It is known that no deterministic algorithm can approximate the
552: volume in polynomial time~(\cite{LOVASZ}). However, randomized
553: polynomial time
554: algorithms are known for the same problem~\cites{KLS, WW}.
555:
556: The same situation seems to be the case regarding the estimation
557: of the {\em mixed volume}~\cite{DGH}, which gives the actual
558: number of solutions in $(\mathbb C^ *)^n$ for
559: generic polynomials of the form (\ref{eq:**})~\cite{BERNSTEIN}.
560:
561:
562: \subsection{Probabilistic algorithms}
563: \label{sec:probabilistic}
564:
565: A {\em probabilistic machine} is a machine that has
566: access to {\em random} bits of information, each
567: random bit costing one unit of time. Each random bit
568: is an independent, uniformly distributed random variable
569: in $\{0,1\}$. In that sense,
570: a probabilistic machine is a machine that flips a
571: fair coin, as many times as necessary, spending
572: one unit of time at each flip.
573:
574: We can therefore speak of the probability that the
575: machine returns a correct result.
576:
577: The class $\BPP$
578: is the class of decision problems $(X, X_{\mathrm{yes}})$
579: such that there is a probabilistic machine and a constant
580: $\epsilon < 1/2$ that will:
581: \begin{itemize}
582: \item[(i)] Decide in polynomial time if $x \in X$.
583: \item[(ii)] Output YES or NO, in polynomial time.
584: \item[(iii)] For every $x$, the output is the correct
585: answer to the question: does $x \in X_{\mathrm{yes}}$?
586: with probability $\ge 1-\epsilon$.
587: \end{itemize}
588:
589: Notice that we can improve the probability that the
590: result is correct by running the same machine several
591: times. Therefore, in the definition above, we may as
592: well take $\epsilon = 1/4$.
593:
594: More generally, a probabilistic machine solves a
595: certain problem (e.g. Problem~\ref{prob:approx}) in
596: polynomial time with probability $\ge 1 - \epsilon$ if
597: and only if it always terminates in polynomial time,
598: and the answer is correct with probability $1-\epsilon$.
599:
600: \begin{restatedtheorem}[restated]
601: There is no $\epsilon < 1/2$ and no
602: probabilistic machine solving
603: Problem~\ref{prob:approx} with probability $1-\epsilon$,
604: unless $\BPP \supseteq \NP$.
605: \end{restatedtheorem}
606: \subsection{Approximation classes}
607: \label{sec:approx}
608:
609: A theory of complexity classes appropriate for
610: the study of combinatorial optimization problems
611: is described in ~\cite{AUSIELLO}. Problem~\ref{prob:opt}
612: fits naturally in the class of combinatorial
613: optimization problems. In this context,
614: Problem~\ref{prob:opt} is characterized by:
615:
616: \begin{itemize}
617: \item[(i)] A set of {\em instances}, given by the
618: set of pairs $(n,A)$,
619: $n \in \mathbb N, \ A \subset \mathbb N^n$
620: finite and non-empty.
621: \item[(ii)] For every instance $(n,A)$, a set of
622: feasible solutions, namely the set of partitions
623: ${\partition{I}} = (I_1, \dots, I_k)$ of $\{1, \dots, n \}$.
624: \item[(iii)] An objective function (to minimize),
625: $\bez(A; {\partition I})$.
626: \end{itemize}
627:
628: The class $\NPO$ of combinatorial optimization problems
629: is analogous to the class $\NP$ of decision problems.
630: Problem~\ref{prob:opt} belongs to that class:
631:
632: \begin{itemize}
633: \item [(1)] The size of each feasible solution is
634: polynomially bounded on the size of each instance.
635: \item [(2)] Given an instance $(n,A)$ and a string $w$,
636: it can be decided in time polynomial in $(n,A)$ whether
637: $w$ encodes a feasible solution $\partition I = (I_1,
638: \dots, I_k)$.
639: \item [(3)] The objective function can be computed in
640: polynomial time.
641: \end{itemize}
642:
643: The class $\APX$ of approximable problems in $\NPO$
644: is defined as the subset of $\NPO$ for which there
645: is some $C>1$ and a polynomial time algorithm such that,
646: given an instance of the problem (say $n, A$)) produces
647: a feasible solution $\partition I$ such that
648: the objective function applied to that solution
649: approximates the minimum up to
650: a factor of $C$.
651:
652: Theorem~\ref{th:B} admits as a corollary:
653:
654: \begin{corollary}
655: Problem~\ref{prob:opt} does not belong to $\APX$,
656: unless $\Po = \NP$.
657: \end{corollary}
658:
659: Our result actually holds even if we do not require the
660: algorithms to compute a feasible solution.
661:
662: \section{Proof of the Main Theorems}
663:
664: \subsection{From graph theory to systems of equations.}
665:
666: \begin{definition}
667: A {\em $k$-coloring} of a graph $G=(V,E)$ is
668: a partition of the set of vertices $V$ into $k$
669: disjoint subsets (``colors'') $I_j$, so that adjacent
670: vertices do not belong to the same ``color'' $I_j$.
671: \end{definition}
672:
673: \begin{problem}[Graph 3-Coloring] \label{graph3}
674: Given a graph $G=(V,E)$, decide if there exists a
675: $3$-coloring of $G$.
676: \end{problem}
677:
678: It is known since~\cite{KARP} that the Graph 3-Coloring
679: Problem is $\NP$-hard (see also~\cite{GAREY-JOHNSON}).
680: We will actually need to consider an equivalent
681: formulation of the
682: Graph 3-coloring problem.
683:
684: Recall that the cartesian product of two graphs
685: $G_1 = (V_1, E_1)$ and $G_2 = (V_2, E_2)$ is the
686: graph $G_1 \times G_2 = (V_1 \times V_2, E)$ with
687: $( (v_1, v_2), (v_1',v_2') ) \in E$ if and only if
688: $v_1 = v_1'$ and $(v_2, v_2') \in E_2$ or
689: $v_2 = v_2'$ and $(v_1, v_1') \in E_1$.
690:
691: Also, let $K_3$ denote the complete graph with 3 vertices.
692:
693: \begin{lemma} \label{lem:equiv}
694: The graph $G$ admits a 3-coloring if and only if
695: the graph $G \times K_3$ admits a 3-coloring
696: $\partition I = (I_1,I_2,I_3)$ with $\# I_1 = \# I_2 = \# I_3 = |G|$.
697: \end{lemma}
698:
699: \begin{proof}
700: $G$ admits a 3-coloring if and only if $G \times K_3$
701: admits a 3-coloring.
702: Moreover, any coloring $\partition I$
703: of $G \times K_3$ satisfies
704: $\# I_1 = \# I_2 = \# I_3$.
705: \end{proof}
706:
707:
708:
709: To each graph $H = (V,E)$ we will associate
710: two spaces of polynomial systems. Each of those spaces
711: is characterized by a support set $A = A(H)$ (resp.
712: $A(H)^l$) to be constructed and corresponds
713: to the space of polynomials of the form (\ref{eq:*})
714: with complex coefficients. Of particular interest will
715: be graphs of the form $H = G \times K_3$.
716:
717: We start by identifying the set $V$ of vertices of $H$
718: to the set $\{1, \dots, m \}$.
719: Let $K_s$ denote the complete graph of size $s$, i.e.
720: the graph with $s$ vertices all of them pairwise
721: connected by edges.
722:
723: To each copy of $K_s$, $s=0, \dots, 3$ that can
724: be embedded as a subgraph of $H$ (say the subgraph
725: generated by $\{v_1, \cdots, v_s\}$) we associate
726: the monomial
727: \[
728: z_{v_1} z_{v_2} \cdots z_{v_s}
729: \]
730: \begin{figure}
731: \centerline{\resizebox{5cm}{!}{\includegraphics{hard-graph.eps}}}
732: \caption{\label{fig:graph}
733: In this example, $A(H) = \{ (0,0,0,0), (1,0,0,0), (0,1,0,0),$
734: $(0,0,1,0),$ $(0,0,0,1),$ $(1,1,0,0),$ $(1,0,1,0),$ $(0,1,1,0),$
735: $(0,0,1,1),$ $(1,1,1,0) \}$. A possible polynomial with that support
736: would be $1+v_1+v_2+v_3+v_4+v_1v_2+v_1v_3 +v_2v_3 + v_3v_4 + v_1v_2v_3$.}
737: \end{figure}
738: (the empty graph $K_0$ corresponds to the constant
739: monomial).
740: Then we consider the linear space generated by all those
741: monomials (Figure~\ref{fig:graph}). Therefore, the support
742: $A(H)$ is the set
743: of all
744: $e_{v_1} + \cdots + e_{v_s} \subset \mathbb N^m$ such
745: that $0 \le s \le 3$ and
746: $\{v_1, \dots, v_s\}$ induces a copy of $K_s$
747: as a subgraph of $H$. Here, $e_i$ denotes the $i$-th
748: vector of the canonical basis of $\mathbb R^n$.
749:
750: Given a set $A$, we denote by $A^l$ the $l$-fold i
751: cartesian product of $A$.
752:
753: The two spaces of polynomial systems associated to
754: a graph $H$ will be the polynomial systems with
755: support $A(H)$ and $A(H)^l$.
756:
757: Remark that none of the two classes of systems
758: above is homogeneous in any possible group of
759: variables (because we introduced a constant
760: monomial). Therefore, in the calculation of
761: the B\'ezout number for a partition $\partition I$,
762: we can set $a_j = \# I_j$.
763:
764: \begin{lemma}\label{Aiseasy}
765: Let $l$ be fixed. Then, there is
766: a polynomial time algorithm to compute $A(H)$ and
767: $A(H)^l$, given $H$.
768: \end{lemma}
769:
770: \subsection{A gap between B\'ezout numbers}
771:
772: In case the graph $H$
773: admits a $3$-coloring
774: $\partition I = (I_1, I_2, I_3)$, any corresponding
775: polynomial system is always trilinear (linear in each
776: set of variables). If moreover $H$ is of the
777: form $H=G \times K_3$ with $|G|=n$, the cardinality
778: of the
779: $I_j$ is always $n$, and formula ~(\ref{eq:bezfromI})
780: becomes:
781:
782: \[
783: \bez (A(G \times K_3); {\partition I})
784: = \binomial{3n}{n \ n \ n}
785: \]
786:
787: The crucial step in the proof of Theorem~\ref{th:A}
788: is to show that
789:
790: \[
791: \bez (A(G \times K_3); {\partition I}) \ge \frac{4}{3}
792: \binomial{3n}{n \ n \ n}
793: \]
794: unless $k=3$ and ${\partition I}$ is a 3-coloring
795: of $G \times K_3$.
796:
797:
798: In order to do that, we introduce the following
799: cleaner abstraction for the B\'ezout number:
800: if $k \in \mathbb N$ and ${\mathbf a}
801: = (a_1, \dots, a_k) \in \mathbb N^k$
802: are such that $\sum_{j=1}^k a_j = 3n$, we set
803: \[
804: B(\mathbf a) \bydef
805: \binomial{3n}{a_1 \ a_2 \ \cdots \ a_k}
806: \
807: \prod_{j=1}^k
808: \left \lceil \frac{a_j}{n} \right \rceil ^{a_j}
809: \]
810:
811: \begin{lemma} \label{lem:bezB}
812: If $H = G \times K_3$ and
813: ${\partition I}=(I_1, \dots, I_k)$ is a partition
814: of the set $\{1, \dots , 3n\}$ of vertices of $H$, then
815: \[
816: \bez(A(H); {\partition I}) \ge B( \mathbf a)
817: \]
818: with $a_j = \# I_j$.
819: \end{lemma}
820:
821: \begin{proof}
822: Consider the $n$ disjoint copies of $K_3$ in
823: $H=G \times K_3$ induced by the nodes of $G$.
824: By the pigeonhole principle,
825: there is at least one of those copies with
826: at least $\lceil a_j / n \rceil$ elements of
827: $I_j$. Hence, the degree $d_j$ in the $j$-th group
828: of variables is at least $\lceil a_j / n \rceil$.
829: \end{proof}
830:
831:
832: The main step towards establishing the ``gap'' is the
833: following Proposition:
834:
835: \begin{proposition}\label{prop1}
836: Let $n, k \in \mathbb N^n$ and let $a_1 \ge a_2 \ge
837: \cdots \ge a_k \ge 1$ be such that $\sum_{j=1}^k a_j
838: = 3n$. Then, either $k=3$ and $a_1 = a_2 = a_3 = n$,
839: or:
840: \[
841: B(\mathbf a) \ge \frac{4}{3} B(n,n,n) \ .
842: \]
843: Moreover, this bound is sharp.
844: \end{proposition}
845:
846: The proof of Proposition~\ref{prop1} is postponed
847: to section~\ref{sec:prop1}.
848: \medskip
849: \par
850:
851: Putting it all together,
852:
853: \begin{lemma} \label{lemma:key}
854: Let $G$ be a graph and $n = |G|$. If $G$ admits a
855: 3-coloring, then
856: \[
857: \min_{\partition I} \bez (A(G \times K_3); \partition I)
858: = \binomial{3n}{n \ n \ n}
859: \]
860: Otherwise,
861: \[
862: \min_{\partition I} \bez (A(G \times K_3); \partition I) \ge \frac{4}{3}
863: \binomial{3n}{n \ n \ n}
864: \]
865: \end{lemma}
866:
867: \begin{proof}
868: According to Lemma~\ref{lem:equiv}, $G$ admits a
869: 3-coloring if and only if $G \times K_3$
870: admits a 3-coloring.
871:
872: If ${\partition I} = (I_1, I_2, I_3)$ is a 3-coloring of
873: $G \times K_3$, then
874: \[
875: \bez(A(G \times K_3); \partition I)
876: =
877: \binomial{3n}{n \ n \ n}
878: \]
879:
880:
881: If $\partition I = (I_1, \dots, I_k)$ is not a 3-coloring of
882: $G \times K_3$, then we distinguish two cases.
883:
884: We set $a_j = \# I_j$.
885:
886: {\bf Case 1:} ${\mathbf a}=(n,n,n)$ and hence $k=3$.
887: Then the degree in at least one group of variables is
888: $\ge 2$, and
889: \[
890: \bez(A(G \times K_3); \partition I)
891: \ge 2^n
892: \binomial{3n}{n \ n \ n}
893: \]
894:
895: {\bf Case 2:} ${\mathbf a} \neq (n,n,n)$. Then
896:
897: \[
898: \bez(A(G \times K_3); \partition I)
899: \ge
900: B(a_1, \dots, a_k)
901: \ge
902: \frac{4}{3}
903: \binomial{3n}{n \ n \ n}
904: \ ,
905: \]
906: where the first inequality follows from
907: Lemma~\ref{lem:bezB} and the
908: second from Proposition~\ref{prop1}.
909:
910: In both cases,
911: \[
912: \min_{\partition I} \bez (A(G \times K_3), \partition I) \ge \frac{4}{3}
913: \binomial{3n}{n \ n \ n}.
914: \]
915: \end{proof}
916:
917: Lemma~\ref{lemma:key} would be sufficient to prove
918: a weaker version of Theorem~\ref{th:A}, where the
919: factor $C$ in problem~\ref{prob:approx} is less
920: than $4/3$.
921:
922: \subsection{Improving the gap}
923:
924: In order to obtain a proof valid for any $C$
925: the idea is to increase the gap by considering several copies
926: of a polynomial system, but each copy in a new set of
927: variables. This idea works out because of the special multiplicative
928: structure of the multi-homogeneous B\'ezout number.
929: We will need:
930:
931: \begin{proposition} \label{prop2}
932: Let $m,l \in \mathbb N$. Let $A \subset \mathbb N^m$
933: be finite and assume that $0 \in A$. Then,
934: \[
935: \min_{\partition J} \bez (A^l; \partition J)
936: =
937: \binomial{lm}{m \ m \ \cdots \ m}
938: \ \left( \min_{\partition I} \bez (A; \partition I) \right)^l
939: \]
940: \end{proposition}
941:
942: \begin{proof}
943: 1. Let $\partition I=(I_1, \cdots, I_k)$ be the partition of
944: $\{1, \dots, m\}$ where the minimal B\'ezout
945: number for $A$ is attained.
946:
947: This induces a partition $\partition J = (J_{js})_{1\le j\le k, 1 \le s \le l}$ of $\{1, \dots, m\}
948: \times \{ 1, \dots, l\}$, given by
949: $J_{js} = I_j \times \{s\}$. Identifying each pair
950: $(i,s)$ with $i+ms$, the $J_{js}$ are also a partition
951: of $\{1, \dots, lm \}$.
952:
953: By construction of $A^l$, the degree $d_{js}$
954: in the variables corresponding to $J_{js}$ is
955: equal to the degree $d_j$ of the variables $I_j$
956: in $A$.
957:
958: The systems corresponding to $A$ and $A^l$ cannot
959: be homogeneous for any partition, since $0 \in A$ and
960: $0 \in A^l$. Then we have $a_j = \#I_j = a_{js}$ for any
961: $s$. Therefore,
962:
963: \begin{eqnarray*}
964: \min_{\partition K} \bez (A^l, \partition K)
965: &\le&
966: \bez (A^l, \partition J)
967: \\
968: &=&
969: \binomial{lm}
970: {\underbrace{a_1 \ \cdots \ a_1}_{l \text{\ times}}
971: \ \cdots \
972: \underbrace{a_k \ \cdots \ a_k}_{l \text{\ times}}}
973: \prod_{s=1}^l \prod_{j=1}^k d_j^{a_j} \\
974: &=&
975: \binomial{lm}{m\ m\ \cdots \ m}
976: \
977: \left(
978: \binomial{m}{a_1\ a_2\ \cdots \ a_k}
979: \prod_{j=1}^k d_j^{a_j} \right)^l \\
980: &=&
981: \binomial{lm}{m \ m \ \cdots \ m}
982: \ \left( \min_{\partition I} \bez (A; \partition I) \right)^l
983: \end{eqnarray*}
984:
985: 2. Now, suppose that the minimal B\'ezout number
986: for $A^l$ is attained for a partition $\partition J=(J_1,
987: \cdots, J_r)$.
988: We claim that each $J_t$ fits into exactly one of
989: the $l$ sets $\{1, \dots , m\} \times \{s\}$.
990:
991: Suppose this is not the case. Assume without loss
992: of generality that $J_1$ splits into $K \subset
993: \{1, \dots , m\} \times \{1\}$ and $L \subset
994: \{1, \dots , m\} \times \{2, \dots, l\}$,
995: both $K$ and $L$ non-empty.
996:
997: If $d_K$ denotes the degree in the $K$-variables
998: and $d_L$ the degree in the $L$ variables, then
999: $d_1 = d_K + d_L$. Also, $a_1 = a_K + a_L$ where
1000: $a_K$ is the size of $K$ and $a_L$ is the size of $L$.
1001: The multi-homogeneous B\'ezout
1002: number corresponding to the partition ${\partition {J'}} =
1003: (K,L,J_2, \cdots, J_r)$ is:
1004: \[
1005: \bez(A^l; \partition {J' }) =
1006: \binomial{3lm}{a_K \ a_L \ a_2 \ \cdots \ a_r}
1007: d_K^{a_K} d_L^{a_L} \prod_{j=2}^r d_j^{a_j}
1008: \]
1009:
1010: Therefore,
1011: \[
1012: \frac
1013: {\bez(A^l; \partition{J' })}
1014: {\bez(A^l, \partition{J})}
1015: =
1016: \frac {\binomial{a_1}{a_K} d_K^{a_K} d_L^{a_L}}
1017: {(d_K + d_L)^{a_1}}
1018: < 1
1019: \]
1020: and the B\'ezout number was not minimal, thus
1021: establishing the claim.
1022:
1023:
1024: 3. Denote by $\partition J = \cup_{s=1}^l \partition {J^{(s)}}$
1025: the partition minimizing
1026: the B\'ezout number corresponding to $A^l$. In the
1027: notation above, we
1028: assume that ${\partition {J^{(s)}}}$ is a partition
1029: of $\{1, \dots, m\} \times \{s\}$.
1030:
1031: In that case,
1032: \begin{eqnarray*}
1033: \bez(A^l; \partition J)
1034: &=&
1035: \binomial{lm}{m\ m\ \cdots \ m}
1036: \prod_{s=1}^l
1037: \left(
1038: \binomial{m}{a_{1}^{(s)}\ \cdots \ a_{k}^{(s)}}
1039: \prod_{j=1}^k (d_{j}^{(s)})^{a_{j}^{(s)}}
1040: \right) \\
1041: &=&
1042: \binomial{lm}{m\ m\ \cdots \ m}
1043: \prod_{s=1}^l
1044: \bez(A, \partition{J^{(s)}})
1045: \\
1046: &\ge&
1047: \binomial{lm}{m\ m\ \cdots \ m}
1048: \left( \min_{\partition I} \bez (A ; \partition{I}) \right)^l
1049: \end{eqnarray*}
1050:
1051: \end{proof}
1052:
1053:
1054: Combining Lemma~\ref{lemma:key} and Proposition~\ref{prop2},
1055: we established that:
1056:
1057: \begin{lemma} \label{lem:final}
1058: Let $G$ be a graph and $n = |G|$. Let $l \in \mathbb N$.
1059: If $G$ admits a 3-coloring, then
1060: \[
1061: \min_{\partition J} \bez (A(G \times K_3)^l, \partition J) =
1062: \binomial{3nl}{3n \ 3n \ \cdots \ 3n}
1063: \binomial{3n}{n \ n \ n}^l
1064: \]
1065: Otherwise,
1066: \[
1067: \min_{\partition J} \bez (A(G \times K_3)^l, \partition J) \ge
1068: \left( \frac{4}{3} \right)^l
1069: \binomial{3nl}{3n \ 3n \ \cdots \ 3n}
1070: \binomial{3n}{n \ n \ n}^l
1071: \]
1072: \end{lemma}
1073:
1074:
1075: \begin{proof}[Proof of Theorem~\ref{th:A}]
1076: Assume that {\tt ApproxB\'ez} is a deterministic,
1077: polynomial time algorithm for solving
1078: problem~\ref{prob:approx}, i.e., for
1079: estimating the B\'ezout number up to
1080: a factor of $C$.
1081:
1082: Then the following algorithm decides
1083: Graph 3-coloring (Problem~\ref{graph3})
1084: in polynomial time:
1085:
1086: \begin{algorithm}[Decides Graph 3-coloring problem]
1087: \label{alg1}
1088: \end{algorithm}
1089: \begin{trivlist}
1090: \item [] {\tt Input: a graph} $G$ {\tt of size} $n$.
1091: \item [] {\tt Output: } YES if $G$ admits a 3-coloring, NO otherwise.
1092: \item [] {\tt Constants: }
1093: $l =
1094: \left \lceil \frac{\log C }{2 \log 4/3} \right \rceil$.
1095: \item [\text{1.}] {\tt Compute }
1096: \[
1097: \rho \leftarrow
1098: \frac{ {\text{\sc ApproxB\'ez}}
1099: \left(A(G \times K_3)^l\right) }
1100: {
1101: \binomial{3nl}{3n \ 3n \ \cdots \ 3n}
1102: \binomial{3n}{n \ n \ n}^l
1103: }
1104: \]
1105: \item [\text{2.}] {\tt \bf If }$\rho^2 < C$
1106: {\tt {\bf then} Output YES,
1107: {\bf else} Output NO.}
1108: \end{trivlist}
1109:
1110: By our choice of the constant $l$, $\sqrt{C} \le (4/3)^l$.
1111: Therefore, Lemma~\ref{lem:final} asserts that the output
1112: of algorithm~\ref{alg1} is correct.
1113:
1114: The bit-size of the numbers that occur when computing
1115: the denominator of line 2 are bounded above by
1116: $O(3nl \log (3nl))$. The size of the graph $G \times K_3$
1117: is $O(n)$, and Lemma~\ref{Aiseasy} says that $A^l$ can
1118: be computed in polynomial time.
1119:
1120: It follows that Algorithm~\ref{alg1} runs in polynomial
1121: time. Since Graph 3-coloring is $\NP$-complete, we
1122: deduce that $\Po = \NP$.
1123: \end{proof}
1124:
1125: \begin{proof}[Proof of Theorem~\ref{th:B}]
1126: Assume now that {\tt ApproxB\'ez} is a probabilistic
1127: polynomial time algorithm for solving
1128: problem~\ref{prob:approx}, which returns a correct
1129: result with probability $1-\epsilon$, $\epsilon < 1/4$.
1130:
1131: Then Algorithm~\ref{alg1} will return the correct answer
1132: for the Graph 3-coloring Problem, with probability at
1133: least $1-\epsilon$. This implies that
1134: Problem~\ref{graph3} is actually in $\BPP$.
1135: \end{proof}
1136:
1137: \section{Proof of Proposition \ref{prop1}}
1138: \label{sec:prop1}
1139:
1140: We will need the following trivial Lemma in the proof of
1141: Proposition~\ref{prop1}:
1142:
1143: \begin{lemma}\label{lem:triv}
1144: Let $x, n \in \mathbb N$. Then,
1145: \[
1146: \left( \left\lceil \frac{x}{n} \right\rceil \frac{n}{x}
1147: \right)^x
1148: \ge
1149: 1 + \left(\rule{0em}{2ex} (n-x) \mod n \right)
1150: \ .\]
1151: In particular, the left-hand side is $\ge 2$ whenever
1152: $n \not \: \mid x$, and is always $\ge 1$.
1153: \end{lemma}
1154:
1155: \begin{proof} Since $n \left\lceil \frac{x}{n} \right\rceil
1156: = x + (n-x) \mod n$,
1157: we have:
1158: \[
1159: \left( \left\lceil \frac{x}{n} \right\rceil \frac{n}{x}
1160: \right)^x
1161: =
1162: \left(
1163: 1 + \frac{(n-x) \mod n }{x}
1164: \right)^x
1165: \ge
1166: 1 + \left(\rule{0em}{2ex} (n-x) \mod n \right)
1167: \]
1168: \end{proof}
1169:
1170: Also, we will make use of the Stirling
1171: Formula~\cite[(6.1.38)]{HANDBOOK}:
1172: \begin{equation} \label{Stirling}
1173: x! = \sqrt{2 \pi} \ x^{x+\frac{1}{2}} \
1174: e^{-x+ \frac{\theta(x)}{12x}} \ ,
1175: \end{equation}
1176: where $0 < \theta(x) < 1$.
1177:
1178: \begin{proof}[Proof of Proposition~\ref{prop1}]
1179: The ratio between $B(\mathbf a)$ and
1180: $B(n,n,n)$ is:
1181: \[
1182: \frac{B(\mathbf a)}{B(n,n,n)}
1183: =
1184: \prod_{j=1}^k
1185: \left\lceil \frac{a_j}{n} \right\rceil^{a_j}
1186: \
1187: \frac{ n! \ n! \ n! } {a_1! \ a_2! \ \cdots \ a_k!}
1188: \]
1189:
1190:
1191: From Stirling formula~(\ref{Stirling}) it follows
1192: immediately that:
1193:
1194: \begin{equation}\label{ratio}
1195: \frac{B(\mathbf a)}{B(n,n,n)}
1196: =
1197: \sqrt{2\pi}^{3-k}
1198: \
1199: \prod_{j=1}^k
1200: \left\lceil \frac{a_j}{n} \right\rceil^{a_j}
1201: \
1202: \frac{ n^{3n + \frac{3}{2}} }
1203: { \prod_{j=1}^k a_j^{a_j + \frac{1}{2}} }
1204: \
1205: e^{ \frac{\theta(n)}{4n}-\sum \frac{\theta(a_j)}{12 a_j}}
1206: \end{equation}
1207:
1208: Now we distinguish the cases $k=1$, $k=2$, and $k \ge 3$.
1209: The first two cases are easy:
1210: \medskip
1211:
1212:
1213:
1214:
1215: {\bf Case 1:} If $k=1$, then $a_1 = 3n$ and (\ref{ratio})
1216: becomes:
1217: \[
1218: \frac{B(\mathbf a)}{B(n,n,n)}
1219: =
1220: 2 \pi \ \frac{n}{\sqrt{3}} \
1221: e^{ \frac{\theta(n)}{4n}-\frac{\theta(3n)}{36n}}
1222: \]
1223: which is bounded below by $\frac{2 \pi}{\sqrt{3}} e^{-1/36}
1224: \simeq 3.528218766$.
1225: \medskip
1226:
1227:
1228: {\bf Case 2:} If $k=2$, Lemma~\ref{lem:triv} implies
1229: that
1230: \[
1231: \frac{B(\mathbf a)}{B(n,n,n)}
1232: \ge
1233: \sqrt{2 \pi} \ \frac{n^{\frac{3}{2}}}{\sqrt{a_1 \ a_2}}
1234: \ e^{-1/6}
1235: \]
1236:
1237: Since $\sqrt{a_1 \ a_2} \le \frac{a_1 + a_2}{2}
1238: = \frac{3n}{2}$, we obtain:
1239:
1240: \[
1241: \frac{B(\mathbf a)}{B(n,n,n)}
1242: \ge
1243: \frac{2}{3} \sqrt{2 \pi} e^{-1/6} \simeq
1244: 1.414543350
1245: \]
1246: \medskip
1247:
1248:
1249:
1250: {\bf Case 3:} Let $k \ge 3$. If $a_3 = n$, then
1251: $k=3$ and $a_1 = a_2 = a_3 = n$, so there is nothing
1252: to prove. Therefore, we assume from now on that $a_3 < n$.
1253:
1254: We separate the right-hand side of (\ref{ratio})
1255: into two products, the first for $j=1,2,3$ and the
1256: second for $j \ge 4$. Equation (\ref{ratio}) becomes
1257: now:
1258:
1259: \begin{equation}\label{ratio2}
1260: \begin{split}
1261: \frac{B(\mathbf a)}{B(n,n,n)}
1262: =
1263: \left(
1264: \prod_{j=1}^3
1265: \left(
1266: \left\lceil \frac{a_j}{n} \right\rceil
1267: \frac{n}{a_j}
1268: \right)^{a_j}
1269: \
1270: \frac{ n^{\frac{3}{2}} }
1271: { \sqrt{a_1 a_2 a_3 }}
1272: \
1273: e^{ \frac{\theta(n)}{4n}-\sum_{j=1}^{3}
1274: \frac{\theta(a_j)}{12 a_j}}
1275: \right)
1276: \\
1277: \left(
1278: \sqrt{2\pi}^{3-k}
1279: \
1280: \prod_{j=4}^k
1281: \frac{n^{a_j}}{a_j^{a_j + \frac{1}{2}}}
1282: \
1283: e^{ -\sum_{j=4}^k \frac{\theta(a_j)}{12 a_j}}
1284: \right)
1285: \end{split}
1286: \end{equation}
1287: using the fact that $a_j < n$ for $j \ge 4$. In case
1288: $k=3$, the second factor in equation (\ref{ratio2}) above is
1289: equal to one.
1290:
1291: Since $a_3 < n$, $n \not \: \mid a_3$ and
1292: Lemma~\ref{lem:triv} implies that for $a_3 < n$
1293: \[
1294: \prod_{j=1}^3
1295: \left(
1296: \left\lceil \frac{a_j}{n} \right\rceil
1297: \frac{n}{a_j}
1298: \right)^{a_j}
1299: \ge 2
1300: \]
1301:
1302: Moreover, $\sqrt[3]{a_1 a_2 a_3} \le (a_1+a_2+a_3)/3 \le n$,
1303: so the first factor of the right-hand side of
1304: (\ref{ratio2}) can be bounded below by
1305: \[
1306: \prod_{j=1}^k
1307: \left(
1308: \left\lceil \frac{a_j}{n} \right\rceil
1309: \frac{n}{a_j}
1310: \right)^{a_j}
1311: \
1312: \frac{ n^{\frac{3}{2}} }
1313: { \sqrt{a_1 a_2 a_3 }}
1314: \
1315: e^{ \frac{\theta(n)}{4n}-\sum_{j=1}^{3}
1316: \frac{\theta(a_j)}{12 a_j}}
1317: \ge
1318: 2 e^{-1/4}
1319: \simeq
1320: 1.557601566
1321: \]
1322:
1323: If $k=3$ we are done. Otherwise, we notice that
1324: since the $a_j$ are non-increasing,
1325: $a_j \le \frac{3n}{4}$ for all $j \ge 4$. In order
1326: to bound the second factor of (\ref{ratio2}),
1327: we will need the following technical Lemma:
1328:
1329: \begin{lemma} \label{case-a-small}
1330: Let $n,x \in \mathbb N$ and let
1331: $x \le \frac{3n}{4}$. Then,
1332: unless $(n,x) \in \{(2,1),(3,2),(4,3),\\
1333: (6,4),(7,5),(8,6)\}$, we have:
1334: \[
1335: \frac{ n^{x} } {\sqrt{2 \pi} x^{x + \frac{1}{2}}}
1336: e^{-\frac{1}{12x}}
1337: >
1338: 1
1339: \]
1340: \end{lemma}
1341:
1342: (Proof is postponed).
1343:
1344: Therefore, unless some of the pairs $(n, a_j)$,
1345: $j \ge 4$ belong to the exceptional subset
1346: $\{(2,1),(3,2),(4,3), (6,4),(7,5),(8,6) \}$, we
1347: have:
1348: \[
1349: \frac{B(\mathbf a)}{B(n,n,n)}
1350: \ge
1351: 2 e^{-\frac{1}{4}} \simeq 1.557601566 \ .
1352: \]
1353:
1354: Finally, we consider the values of $n$ and $\mathbf a$
1355: where some $(n,a_j)$, $j \ge 4$,
1356: is in the exceptional subset. All the possible values
1357: of $n$ and $\mathbf a$ are listed in table~\ref{tab1}.
1358: \begin{table}
1359: \centerline{
1360: \begin{tabular}{||c|c|c|l|c|r|r|c||}
1361: \hline
1362: \hline
1363: $n$ & $a_j$& $3n$ & $\mathbf a$ & j
1364: & $B(\mathbf a)$ & $B(n,n,n)$&
1365: $\frac{B(\mathbf a)}{B(n,n,n)}$\\
1366: \hline
1367: \hline
1368: 2 & 1 & 6 & 1 1 1 {\bf 1} {\bf 1} {\bf 1}
1369: & 4,5,6
1370: & 720 & 90
1371: & 8\\
1372: & & & 2 1 1 {\bf 1} {\bf 1}
1373: & 4,5
1374: & 360 &
1375: &4\\
1376: & & & 2 2 1 {\bf 1} & 4 & 180 &
1377: &2\\
1378: & & & 3 1 1 {\bf 1} & 4 & 120 &
1379: &$\frac{4}{3}$\\
1380: \hline
1381: 3 & 2 & 9 & 2 2 2 {\bf 2} 1 & 4 & 22680 & 1680
1382: &$\frac{27}{2}$\\
1383: & & & 3 2 2 {\bf 2} & 4 & 7560 &
1384: &$\frac{9}{2}$\\
1385: \hline
1386: 4 & 3 &12 & 3 3 3 {\bf 3} & 4 & 369600 & 34650
1387: &$\frac{32}{3}$\\
1388: \hline
1389: 6 & 4 & 18& 4 4 4 {\bf 4} 1 1 & 4 & 19297278000 & 17153136
1390: &1125\\
1391: & & & 4 4 4 {\bf 4} 2 & 4 & 9648639000 &
1392: &$\frac{1125}{2}$\\
1393: & & & 5 4 4 {\bf 4} 1 & 4 & 3859455600 &
1394: &225\\
1395: & & & 5 5 4 {\bf 4} & 4 & 771891120 &
1396: &45\\
1397: & & & 6 4 4 {\bf 4} & 4 & 643242600 &
1398: &$\frac{75}{2}$\\
1399: \hline
1400: 7 & 5 & 21& 5 5 5 {\bf 5} 1 & 4 & 246387645504 & 399072960
1401: &$\frac{3087}{5}$\\
1402: & & & 6 5 5 {\bf 5} & 4 & 41064607584 &
1403: &$\frac{1029}{10}$\\
1404: \hline
1405: 8 & 6 & 24& 6 6 6 {\bf 6} & 4 & 2308743493056 & 9465511770
1406: &$\frac{10976}{45}$\\
1407: \hline
1408: \hline
1409: \end{tabular}
1410: }
1411: \caption{\label{tab1}
1412: Ratios for all the exceptional pairs $(n, \mathbf a)$.}
1413: \end{table}
1414: The ratio is always $\ge 4/3$, and the value of $4/3$
1415: is attained for $n=2$ and $\mathbf a = (3,1,1,1)$.
1416: \end{proof}
1417:
1418:
1419:
1420: \begin{proof}[Proof of Lemma~\ref{case-a-small}]
1421: \begin{figure}
1422: \includegraphics{hard-gandh.eps}
1423: \caption{\label{fig-gandh}Plots of $g_n(x)$ and $h(x)$.}
1424: \end{figure}
1425:
1426: Let
1427: \begin{eqnarray*}
1428: g_n(x) &=& \log \left(
1429: \frac{ n^x }{\sqrt{2 \pi} x^{x + \frac{1}{2}}}
1430: e^{-\frac{1}{12x}} \right) \\
1431: &=&
1432: x \log n
1433: - x \log x
1434: - \frac{1}{2} \log x
1435: - \frac{1}{12x}
1436: - \frac{1}{2} \log 2\pi
1437: \end{eqnarray*}
1438:
1439: (see figure~\ref{fig-gandh}).
1440: We first consider values of $x \ge 7$.
1441: By hypothesis, $n/x \ge 4/3$ so
1442: $\log n - \log x \ge \log(4/3)$, and therefore $g_n(x)
1443: \ge h(x)$, where:
1444: \[
1445: h(x) = x \log(4/3)
1446: - \frac{1}{2} \log x
1447: - \frac{1}{12x}
1448: - \frac{1}{2} \log 2\pi
1449: \]
1450:
1451: (see Figure~\ref{fig-gandh} also).
1452: Notice that $h(x)$ is independent of $n$.
1453: The derivative of $h$ is
1454: \[
1455: h '(x) =
1456: \log(4/3)
1457: - \frac{1}{2x}
1458: + \frac{1}{12x^2}
1459: =
1460: \frac{12 \log(4/3) x^2 - 6x + 1}{12x^2}
1461: \]
1462:
1463: The numerator vanishes at
1464: \[
1465: x = \frac{1 \pm 1\sqrt{1-4/3 \log(4/3)}}{4 \log(4/3)}
1466: \]
1467: Numerically, $x \simeq 0.1867281114$ or
1468: $x \simeq 1.551301638$.
1469: Therefore, the function $h(x)$ is increasing for $x \ge 2$.
1470: Again, numerically $h(7) \simeq 0.1099761345$ and therefore,
1471: if $x \ge 7$ we always have:
1472: \[
1473: e^{g_n(x)} \ge 1.1162 > 1
1474: \]
1475:
1476:
1477: Now we consider $x \le 6$.
1478: Having $g_n > 0$,
1479: is equivalent to:
1480: \[
1481: n > n_0(x) =
1482: x
1483: e^{\frac{1}{2x} \log x
1484: +\frac{1}{12x^2}
1485: +\frac{1}{2x} \log 2\pi}
1486: \]
1487:
1488: At this point, we proved that
1489: $g_n(x)$ is positive, except possibly for
1490: pairs $(n,x)$ with $1 \le x \le 6$ and
1491: $\frac{4}{3}x \le n \le n_0(x)$. The values
1492: of $n_0$ are tabulated in Table~\ref{tabn0}.
1493: From Table~\ref{tabn0} it is clear that the
1494: only exceptions are those listed
1495: in the hypothesis.
1496:
1497: \begin{table}[ht]
1498: \centerline{
1499: \begin{tabular}{||l|l|l|c||}
1500: \hline
1501: \hline
1502: $x$ & $\frac{4}{3}x$& $n_0(x)$ &Possible $n$'s \\
1503: \hline
1504: 1 & 1.333333333 & 2.724464424 & 2 \\
1505: 2 & 2.666666666 & 3.844857634 & 3 \\
1506: 3 & 4 & 4.939610298 & 4 \\
1507: 4 & 5.333333333 & 6.016610872 & 5 \\
1508: 5 & 6.666666666 & 7.081620345 & 6 \\
1509: 6 & 8 & 8.137996302 & 8 \\
1510: \hline
1511: \hline
1512: \end{tabular}}
1513: \caption{\label{tabn0}Possible values of $n$ for $x$ small}
1514: \end{table}
1515:
1516: \end{proof}
1517:
1518:
1519: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1520:
1521:
1522:
1523: \begin{bibsection}
1524:
1525:
1526: \renewcommand{\PrintEditorsA}[1]{%
1527: \PrintNames{editor}{}{ (ed\Plural{s}.)}{#1}%
1528: }
1529: \renewcommand{\ReviewList}[1]{%
1530: % \PrintSeries{\review}{}{,}{ }{,}{ }{,}{ }{}{#1}{}%
1531: }
1532: \renewcommand{\ISBNList}[1]{%
1533: % \PrintSeries{\isbn}{}{,}{ }{,}{ }{,}{ }{ISBN }{#1}{}%
1534: }
1535: \begin{biblist}
1536:
1537:
1538: \bib{HANDBOOK}{collection}{
1539: title={Handbook of mathematical functions with formulas, graphs, and
1540: mathematical tables},
1541: editor={Abramowitz, Milton},
1542: editor={Stegun, Irene A.},
1543: note={Reprint of the 1972 edition},
1544: publisher={Dover Publications Inc.},
1545: place={New York},
1546: date={1992},
1547: pages={xiv+1046},
1548: isbn={0-486-61272-4},
1549: review={MR 94b:00012},
1550: }
1551:
1552:
1553: \bib{AUSIELLO}{book}{
1554: author={Ausiello, G.},
1555: author={Crescenzi, P.},
1556: author={Gambosi, G.},
1557: author={Kann, V.},
1558: author={Marchetti-Spaccamela, A.},
1559: author={Protasi, M.},
1560: title={Complexity and approximation},
1561: note={Combinatorial optimization problems and their approximability
1562: properties;
1563: With 1 CD-ROM (Windows and UNIX)},
1564: publisher={Springer-Verlag},
1565: place={Berlin},
1566: date={1999},
1567: pages={xx+524},
1568: isbn={3-540-65431-3},
1569: review={MR 2001f:68002},
1570: }
1571:
1572: \bib{BERNSTEIN}{article}{
1573: author={Bernstein, D. N.},
1574: title={The number of roots of a system of equations},
1575: language={Russian},
1576: journal={Funkcional. Anal. i Prilo\v zen.},
1577: volume={9},
1578: date={1975},
1579: number={3},
1580: pages={1\ndash 4},
1581: review={MR 55 \#8034},
1582: }
1583:
1584: \bib{DMS}{article}{
1585: author={Dedieu, Jean-Pierre},
1586: author={Malajovich, Gregorio},
1587: author={Shub, Mike},
1588: title={On the curvature of the central path of linear programming theory},
1589: year={2003},
1590: eprint={arXiv:math.OC/0312083},
1591: }
1592:
1593: \bib{DGH}{article}{
1594: author={Dyer, Martin},
1595: author={Gritzmann, Peter},
1596: author={Hufnagel, Alexander},
1597: title={On the complexity of computing mixed volumes},
1598: journal={SIAM J. Comput.},
1599: volume={27},
1600: date={1998},
1601: number={2},
1602: pages={356\ndash 400 (electronic)},
1603: issn={1095-7111},
1604: review={MR 99f:68092},
1605: }
1606:
1607: \bib{GAREY-JOHNSON}{book}{
1608: author={Garey, Michael R.},
1609: author={Johnson, David S.},
1610: title={Computers and intractability},
1611: note={A guide to the theory of NP-completeness;
1612: A Series of Books in the Mathematical Sciences},
1613: publisher={W. H. Freeman and Co.},
1614: place={San Francisco, Calif.},
1615: date={1979},
1616: pages={x+338},
1617: isbn={0-7167-1045-5},
1618: review={MR 80g:68056},
1619: }
1620:
1621: \bib{KLS}{article}{
1622: author={Kannan, Ravi},
1623: author={Lov{\'a}sz, L{\'a}szl{\'o}},
1624: author={Simonovits, Mikl{\'o}s},
1625: title={Random walks and an $O\sp *(n\sp 5)$ volume algorithm for convex
1626: bodies},
1627: journal={Random Structures Algorithms},
1628: volume={11},
1629: date={1997},
1630: number={1},
1631: pages={1\ndash 50},
1632: issn={1042-9832},
1633: review={MR 99h:68078},
1634: }
1635: \bib{KARP}{article}{
1636: author={Karp, Richard M.},
1637: title={Reducibility among combinatorial problems},
1638: booktitle={Complexity of computer computations (Proc. Sympos., IBM Thomas
1639: J. Watson Res. Center, Yorktown Heights, N.Y., 1972)},
1640: pages={85\ndash 103},
1641: publisher={Plenum},
1642: place={New York},
1643: date={1972},
1644: review={MR 51 \#14644},
1645: }
1646:
1647:
1648:
1649: \bib{KHACHIYAN}{article}{
1650: author={Khachiyan, L. G.},
1651: title={The problem of calculating the volume of a polyhedron is
1652: enumeratively hard},
1653: language={Russian},
1654: journal={Uspekhi Mat. Nauk},
1655: volume={44},
1656: date={1989},
1657: number={3(267)},
1658: pages={179\ndash 180},
1659: issn={0042-1316},
1660: review={MR 91e:68073},
1661: }
1662:
1663:
1664: \bib{KUSHNIRENKO}{article}{
1665: author={Kushnirenko, A.G.},
1666: title={Newton Polytopes and the B\'ezout Theorem},
1667: journal={Funct. Anal. Appl.},
1668: volume={10},
1669: year={1976},
1670: pages={233\ndash 235},
1671: }
1672:
1673: \bib{LI-BAI}{article}{
1674: author={Li, Tiejun},
1675: author={Bai, Fengshan},
1676: title={Minimizing multi-homogeneous B\'ezout numbers by a local search
1677: method},
1678: journal={Math. Comp.},
1679: volume={70},
1680: date={2001},
1681: number={234},
1682: pages={767\ndash 787 (electronic)},
1683: issn={0025-5718},
1684: review={MR 2002b:65085},
1685: }
1686:
1687: \bib{LI}{article}{
1688: author={Li, T. Y.},
1689: title={Numerical solution of multivariate polynomial systems
1690: by
1691: homotopy continuation methods},
1692: booktitle={Acta numerica, 1997},
1693: series={Acta Numer.},
1694: volume={6},
1695: pages={399\ndash 436},
1696: publisher={Cambridge Univ. Press},
1697: place={Cambridge},
1698: date={1997},
1699: review={MR 2000i:65084},
1700: }
1701:
1702:
1703: \bib{LI-LIN-BAI}{article}{
1704: author={Li, Ting},
1705: author={Lin, Zhenjiang},
1706: author={Bai, Fengshan},
1707: title={Heuristic methods for computing the minimal multi-homogeneous
1708: B\'ezout number},
1709: journal={Appl. Math. Comput.},
1710: volume={146},
1711: date={2003},
1712: number={1},
1713: pages={237\ndash 256},
1714: issn={0096-3003},
1715: review={2 007 782},
1716: }
1717:
1718: \bib{LOVASZ}{book}{
1719: author={Lov{\'a}sz, L{\'a}szl{\'o}},
1720: title={An algorithmic theory of numbers, graphs and convexity},
1721: series={CBMS-NSF Regional Conference Series in Applied Mathematics},
1722: volume={50},
1723: publisher={Society for Industrial and Applied Mathematics (SIAM)},
1724: place={Philadelphia, PA},
1725: date={1986},
1726: pages={iv+91},
1727: isbn={0-89871-203-3},
1728: review={MR 87m:68066},
1729: }
1730:
1731:
1732:
1733: \bib{MORGAN}{book}{
1734: author={Morgan, Alexander},
1735: title={Solving polynomial systems using continuation for engineering
1736: and scientific problems},
1737: publisher={Prentice Hall Inc.},
1738: place={Englewood Cliffs, NJ},
1739: date={1987},
1740: pages={xiv+546},
1741: isbn={0-13-822313-0},
1742: review={MR 91c:00014},
1743: }
1744:
1745:
1746: \bib{MORGAN-SOMMESE}{article}{
1747: author={Morgan, Alexander},
1748: author={Sommese, Andrew},
1749: title={A homotopy for solving general polynomial systems that respects
1750: $m$-homogeneous structures},
1751: journal={Appl. Math. Comput.},
1752: volume={24},
1753: date={1987},
1754: number={2},
1755: pages={101\ndash 113},
1756: issn={0096-3003},
1757: review={MR 88j:65110},
1758: }
1759:
1760: \bib{SHAFAREVICH}{book}{
1761: author={Shafarevich, I. R.},
1762: title={Basic algebraic geometry},
1763: edition={Springer Study Edition},
1764: note={Translated from the Russian by K. A. Hirsch;
1765: Revised printing of Grundlehren der mathematischen
1766: Wissenschaften, Vol. 213, 1974},
1767: publisher={Springer-Verlag},
1768: place={Berlin},
1769: date={1977},
1770: pages={xv+439},
1771: review={MR 56 \#5538},
1772: }
1773:
1774:
1775: \bib{WMS}{article}{
1776: author={Wampler, Charles},
1777: author={Morgan, Alexander},
1778: author={Sommese, Andrew},
1779: title={Numerical continuation methods for solving polynomial systems
1780: arising in kinematics},
1781: journal={Journal Mechanical Design},
1782: volume={112},
1783: date={1990},
1784: pages={59\ndash 68},
1785: }
1786:
1787:
1788: \bib{WW}{article}{
1789: author={Werschulz, A. G.},
1790: author={Wo{\'z}niakowski, H.},
1791: title={What is the complexity of volume calculation?},
1792: note={Algorithms and complexity for continuous problems/Algorithms,
1793: computational complexity, and models of computation for
1794: nonlinear and multivariate problems (Dagstuhl/South Hadley, MA,
1795: 2000)},
1796: journal={J. Complexity},
1797: volume={18},
1798: date={2002},
1799: number={2},
1800: pages={660\ndash 678},
1801: issn={0885-064X},
1802: review={MR 2003k:68048},
1803: }
1804:
1805: \end{biblist}
1806: \end{bibsection}
1807: \end{document}
1808: