1: \documentclass{elsart}
2: \usepackage{epsfig}
3: \usepackage{amssymb}
4:
5: \newcommand{\sink}{\mathop{\rm sink}}
6: \newcommand{\R}{{\mathbf{R}}}
7: \renewcommand{\P}{\mathrm{P}}
8: \newcommand{\V}{{\mathcal V}}
9: \newcommand{\E}{{\mathcal E}}
10: \newcommand{\G}{{\mathcal G}}
11: \newcommand{\keyw}[1]{\textnormal{\textrm{\texttt{#1}}}}
12: \newcommand{\vi}{{\sf V}}
13: \newcommand{\bfaceq}{{{\mathcal B}/\!\sim}}
14: \newcommand{\figeps}[3]{
15: \begin{figure}
16: \begin{center}
17: \epsfig{figure=#1.eps, width=#2cm}
18: \end{center}
19: \caption{#3}
20: \label{fig:#1}
21: \end{figure}
22: }
23:
24: \begin{document}
25: \begin{frontmatter}
26:
27: \title{Violator Spaces: Structure and Algorithms
28: \thanksref{label1}}
29: \thanks[label1]{The first and the third author acknowledge support
30: from the Swiss Science Foundation (SNF), Project No.\ 200021-100316/1.
31: The fourth author acknowledges support from the Czech Science Foundation
32: (GACR), Grant No.\ 201/05/H014.}
33: \author[label2]{B. G\"artner},
34: \ead{gaertner@inf.ethz.ch}
35: \author[label3]{J. Matou\v{s}ek},
36: \ead{matousek@kam.mff.cuni.cz}
37: \author[label2]{L. R\"ust},
38: \ead{ruestle@inf.ethz.ch}
39: \author[label3]{P. \v{S}kovro\v{n}}
40: \ead{xofon@kam.mff.cuni.cz}
41: \address[label2]{
42: Institute of Theoretical Computer Science,
43: ETH Z\"urich,
44: 8092 Z\"urich, Switzerland
45: }
46: \address[label3]{
47: Department of Applied Mathematics and Institute of Theoretical Computer Science,
48: Charles University,
49: Malostransk\'{e} n\'{a}m. 25,
50: 118~00~~Praha~1, Czech Republic
51: }
52:
53: \begin{abstract}
54: Sharir and Welzl introduced an abstract framework for optimization
55: problems, called {\em LP-type problems\/} or also {\em generalized
56: linear programming problems}, which proved useful in algorithm design.
57: We define a new, and as we believe, simpler and more natural framework:
58: {\em violator spaces}, which constitute a proper generalization of LP-type
59: problems. We show that Clarkson's randomized algorithms for low-dimensional
60: linear programming work in the context of violator spaces. For example,
61: in this way we obtain the fastest known algorithm for the \emph{$\P$-matrix
62: generalized linear complementarity problem} with a constant number of blocks.
63: We also give two new characterizations of LP-type problems: they are
64: equivalent to {\em acyclic\/} violator spaces, as well as to {\em concrete\/}
65: LP-type problems (informally, the constraints in a concrete LP-type
66: problem are subsets of a linearly ordered ground set, and the value
67: of a set of constraints is the minimum of its intersection).
68: \end{abstract}
69:
70: \begin{keyword}
71: LP-type problem \sep generalized linear programming \sep violator
72: space \sep Clarkson's algorithms \sep unique sink orientation \sep
73: generalized linear complementarity problem
74: \end{keyword}
75:
76: \end{frontmatter}
77:
78: \section{Introduction}
79: The framework of LP-type problems,
80: invented by Sharir and Welzl in 1992 \cite{sw-cblpr-92},
81: has become a well-established tool
82: in the field of geometric optimization. Its origins are in linear
83: programming: Sharir and Welzl developed a randomized variant of the
84: dual simplex algorithm for linear programming and showed that this
85: algorithm actually works for a more general class of problems they
86: called LP-type problems.
87:
88: For the theory of linear programming, this
89: algorithm constituted an important progress, since it was later shown
90: to be \emph{subexponential} in the RAM model \cite{msw-sblp-92}.
91: Together with a similar result independently obtained by Kalai
92: \cite{k-srsa-92}, this was the first linear programming algorithm
93: provably requiring a number of arithmetic operations subexponential
94: in the dimension and number of constraints (independent of the
95: precision of the input numbers).
96:
97: For many other geometric optimization problems in fixed dimension,
98: the algorithm by Sharir and Welzl was the first to achieve expected
99: linear runtime, simply because these problems could be formulated
100: as LP-type problems. The class of LP-type problems for example
101: includes the problem of computing the minimum-volume ball or
102: ellipsoid enclosing a given point set in $\R^d$, and the problem
103: of finding the distance of two convex polytopes in $\R^d$. Many
104: other problems have been identified as LP-type problems over the
105: years \cite{msw-sblp-92,a-bbhdn-94,a-httgl-94,bsv-dsapg-01,halman}.
106:
107: Once it is shown that a particular optimization problem is an LP-type
108: problem, and certain algorithmic primitives are implemented for it,
109: several efficient algorithms are immediately at our disposal: the
110: Sharir--Welzl algorithm, two other randomized optimization algorithms due
111: to Clarkson \cite{c-lvali-95} (see \cite{gw-lpraf-96,cm-ltdao-96} for
112: a discussion of how it fits the LP-type framework), a deterministic
113: version of it \cite{cm-ltdao-96}, an algorithm for computing the
114: minimum solution that violates at most $k$ of the given $n$
115: constraints \cite{m-gofvc-95}, and probably more are to come in
116: the future.
117:
118: The framework of LP-type problems is not only a prototype for concrete
119: optimization problems, it also serves as a mathematical tool by
120: itself, in algorithmic \cite{GWSampl01,ChanTukey} and non-algorithmic
121: contexts \cite{a-spiht-96}.
122:
123: An (abstract) LP-type problem is given by a finite set $H$ of
124: {\em constraints\/} and a {\em value\/} $w(G)$ for every subset
125: $G\subseteq H$. The values can be real numbers or, for technical convenience,
126: elements of any other linearly ordered set. Intuitively, $w(G)$ is
127: the minimum value of a solution that satisfies all constraints in $G$.
128: The assignment $G\mapsto w(G)$ has to obey the axioms in the
129: following definition.
130:
131: \begin{defn}\label{def:lptype}
132: An {\em abstract LP-type problem} is a quadruple $(H,w,W,\leq)$, where
133: $H$ is a finite set, $W$ is a set linearly ordered by $\leq$, and
134: $w\colon 2^H\to W$ is a mapping satisfying the following two conditions:
135:
136: \begin{tabular}{ll}
137: Monotonicity: & for all $F\subseteq G\subseteq H$ we have
138: $w(F)\leq w(G)$, and\\
139: Locality: & for all $F\subseteq G\subseteq H$ and all $h\in H$ with
140: $w(F)=w(G)$ and\\
141: &$w(G)<w(G\cup\{h\})$, we have $w(F)<w(F\cup\{h\})$.\\
142: \end{tabular}
143: \end{defn}
144:
145: As our running example, we will use the smallest enclosing ball problem,
146: where $H$ is a finite point set in $\R^d$ and $w(G)$ is the radius of
147: the smallest ball that encloses all points of $G$.
148: In this case monotonicity is obvious, while verifying locality requires
149: the nontrivial but well known geometric result that the smallest
150: enclosing ball is unique for every set.
151:
152: It seems that the order $\leq$ of subsets is crucial; after all,
153: LP-type problems model \emph{optimization problems}, and indeed,
154: the subexponential algorithm for linear programming and other
155: LP-type problems \cite{msw-sblp-92} heavily relies on such an
156: order.
157:
158: A somewhat deeper look reveals that often, we only care whether two
159: subsets have the \emph{same} value, but not how they compare under
160: the order $\leq$. The following definition is taken from
161: \cite{sw-cblpr-92}:
162:
163: \begin{defn}\label{def:lp_type_basis}
164: Consider an abstract LP-type problem $(H,w,W,\leq)$. We say that
165: $B\subseteq H$ is a {\em basis} if for all proper subsets $F\subset B$ we
166: have $w(F)\neq w(B)$. For $G\subseteq H$, a \emph{basis of $G$} is a minimal
167: subset $B$ of $G$ with $w(B)=w(G)$.
168: \end{defn}
169:
170: We observe that a minimal subset $B\subseteq G$ with $w(B)=w(G)$ is
171: indeed a basis.
172:
173: Solving an abstract LP-type problem $(H,w,W,\leq)$ means to find a
174: basis of $H$. In the smallest enclosing ball problem, a basis of $H$
175: is a minimal set $B$ of points such that the smallest enclosing ball
176: of $B$ has the same radius (and is in fact the same) as the smallest
177: enclosing ball of $H$, $w(B)=w(H)$.
178:
179: In defining bases, and in saying what it means to solve an LP-type
180: problem, we therefore do not need the order $\leq$. The
181: main contribution of this paper is that many of the things one can
182: prove about LP-type problems do not require a concept of order.
183:
184: We formalize this by defining the new framework of
185: \emph{violator spaces}.
186: Intuitively, a violator space is an LP-type problem without order.
187: This generalization of LP-type problems is proper, and we can exactly
188: characterize the violator spaces that ``are'' LP-type problems.
189: In doing so, we also establish yet
190: another equivalent characterization of
191: LP-type problems that is closer to the applications than the abstract
192: formulation of Definition \ref{def:lptype}. In a concrete LP-type
193: problem, the constraints are not just elements of a set, but
194: they are associated with subsets of some linearly ordered ground
195: set $X$, with the minimal elements in the intersections of such
196: subsets corresponding to ``solutions''.
197: The framework of concrete LP-type problems is similar to the model
198: presented in \cite{a-httgl-94} as a mathematical programming problem, with
199: a few technical differences.
200:
201: These are our main findings on the structural side. Probably the most
202: surprising insight on the algorithmic side is that
203: Clarkson's algorithms \cite{c-lvali-95} work for violator spaces
204: of fixed dimension, leading to an expected linear-time algorithm
205: for ``solving'' the violator space.
206: Clarkson's algorithms were originally developed for linear
207: programs with small dimension. They can be
208: generalized for LP-type problems \cite{gw-lpraf-96,cm-ltdao-96}. The
209: fact that the scheme also works for violator spaces may come as a
210: surprise since the structure of
211: violator spaces is not acyclic in general (in contrast to LP-type
212: problems). The LP-type algorithm from \cite{msw-sblp-92} is also
213: applicable to violator spaces, but its analysis breaks down.
214:
215: We give an application of Clarkson's algorithms in the more general
216: setting by linking our new violator space
217: framework to well-known abstract and concrete frameworks in
218: combinatorial optimization. For this, we show that any \emph{unique
219: sink orientation} (USO) of the cube
220: \cite{SW,grid_uso,M02,MatUSO,MCube,Develin,MS,SchSz,SS,lptouso}
221: or the more general grid \cite{grid_uso} gives rise to a violator
222: space, but not to an LP-type problem in general.
223: Grid USO capture some important problems
224: like linear programming over products of simplices, \emph{generalized
225: linear complementarity problems} over $\mathrm{P}$-matrices \cite{grid_uso}
226: or games like parity, mean-payoff, and simple stochastic games
227: \cite{sweden1,sweden2,ssg_pglcp}.
228:
229: We show that we can find the sink in a unique sink orientation by
230: solving the violator space, for example with Clarkson's
231: algorithms. A concrete new result is obtained by applying this to
232: $\P$-matrix generalized linear complementarity problems.
233: These problems are not known to be polynomial-time solvable, but
234: NP-hardness would imply NP=co-NP \cite{Meg,grid_uso}.
235: Since any $\P$-matrix generalized linear
236: complementarity problem gives rise to a unique sink orientation
237: \cite{grid_uso}, we may use violator spaces and Clarkson's algorithms
238: to solve the problem in expected linear time in the (polynomially
239: solvable) case of a \emph{fixed} number of \emph{blocks}. This is
240: optimal and beats all previous algorithms.
241:
242: The rest of the paper is organized as follows. In Section
243: \ref{sec:basics}, we formally define the frameworks of
244: concrete LP-type problems and violator spaces, along with their
245: essential terminology. Then we state our main structural result.
246:
247: In Section \ref{sec:mainproof}, we prove this result by deriving the
248: equivalence of abstract and concrete LP-type problems, and of
249: \emph{acyclic} violator spaces.
250:
251: Section \ref{sec:clarkson} shows that Clarkson's algorithms work for
252: (possibly cyclic) violator spaces. Section \ref{grid_uso}, finally,
253: shows how unique sink orientations induce violator spaces. A
254: unique sink orientation can be cyclic, and a cyclic orientation gives
255: rise to a cyclic violator space. Unique sink orientations are
256: therefore nontrivial examples of possibly cyclic violator spaces.
257:
258: \section{Structural Results}\label{sec:basics}
259:
260: \subsection{Concrete LP-type problems.}
261: Although intuitively one thinks about $w(G)$ as the value of an
262: optimal solution of an optimization problem, the solution itself is
263: not explicitly represented in Definition \ref{def:lptype}. In
264: specific geometric examples, the constraints can usually be
265: interpreted as a subset of some ground set $X$ of points, and the
266: optimal solution for $G$ is the point with the smallest value in the
267: intersection of all constraints in $G$. For example, in linear
268: programming, the constraints are halfspaces, the value is given by the
269: objective function, and the optimum is the point with minimum value in
270: the admissible region, i.e., the intersection of the halfspaces. In
271: order to have a unique optimum for every set of constraints (which is
272: needed for $w$ to define an LP-type problem), one assumes that the
273: points are linearly ordered by the value; for linear programming, we
274: can always take the lexicographically smallest optimal solution, for
275: instance.
276:
277: Such an interpretation is possible for the smallest enclosing ball problem
278: too, although it looks a bit artificial. Namely, the ``points'' of $X$
279: are all balls in $\R^d$, where the ordering can be an arbitrary linear
280: extension of the partial ordering of balls by radius. The ``constraint''
281: for a point $h\in H$ is the set of all balls containing $h$.
282:
283: The following definition captures this approach to LP-type problems.
284:
285: \begin{defn}
286: \label{def:concrete_LPtype}
287: A {\em concrete LP-type problem} is a triple $(X,\preceq,{\mathcal H})$,
288: where $X$ is a set linearly ordered by $\preceq$,
289: ${\mathcal H}$ is a finite multiset whose elements are subsets of
290: $X$, and for any
291: ${\mathcal G}\subseteq{\mathcal H}$, if the intersection
292: $\bigcap{\mathcal G}:=\bigcap_{G\in\mathcal{G}}G$ is nonempty, then it
293: has a minimum element with respect to $\preceq$ (for ${\mathcal G}=\emptyset$
294: we define $\bigcap{\mathcal G}:=X$).
295: \end{defn}
296:
297: The definition allows $\mathcal H$ to be a multiset, i.e., a constraint
298: set $A\subseteq X$ may be included several times.
299: For example, in an instance of linear programming, some constraints
300: can be the same, which we can reflect by this.
301: In Subsection \ref{sec:examples} we provide an example of an abstract
302: LP-type problem, for which the multiplicity comes in handy to represent it as
303: a concrete LP-type problem.
304:
305: A similar model has been presented in \cite{a-httgl-94} (mathematical
306: programming problem). The slight difference is that it allows several
307: points to have the same value but the constraints form a set rather than
308: a multiset.
309:
310: Bases are defined analogously to Definition \ref{def:lp_type_basis}.
311:
312: \begin{defn}\label{def:concrete_basis}
313: Consider a concrete LP-type problem $(X,\preceq,{\mathcal H})$. We say that
314: $\mathcal{B}\subseteq\mathcal{H}$ is a {\em basis} if for all proper submultisets
315: $\mathcal{F}\subset\mathcal{B}$ we have $\min(\bigcap\mathcal{F})\prec
316: \min(\bigcap\mathcal{B})$. For $\mathcal{G}\subseteq\mathcal{H}$, a
317: \emph{basis of $\mathcal{G}$} is a minimal $\mathcal{B}\subseteq\mathcal{G}$
318: with $\min(\bigcap\mathcal{B})=\min(\bigcap\mathcal{G})$.
319: \end{defn}
320:
321: As before, a minimal $\mathcal{B}\subseteq\mathcal{G}$ with
322: $\min(\bigcap\mathcal{B})=\min(\bigcap\mathcal{G})$ is indeed a basis.
323:
324: Given any concrete LP-type problem ${\mathcal P}=(X,\preceq,{\mathcal H})$,
325: we obtain an abstract LP-type problem $P=({\mathcal H},w,X,\preceq)$
326: according to Definition \ref{def:lptype} by putting
327: $w({\mathcal G})=\min(\bigcap {\mathcal G})$ (or $w({\mathcal G})=+\infty$,
328: if $\bigcap{\mathcal G}$ is empty), as is easy to check (proof omitted).
329: It is clear that $\mathcal{B}\subseteq\mathcal{G}$ is a basis of
330: $\mathcal{G}$ in $P$ if and only if $\mathcal{B}$ is a basis of
331: $\mathcal{G}$ in $\mathcal{P}$. We say that $P$ is
332: \emph{basis-equivalent} to $\mathcal{P}$.
333:
334: Somewhat surprising is the converse, which we prove below in
335: Theorem~\ref{thm:main}:
336: Any abstract LP-type problem $(H,w,W,\leq)$ has a
337: ``concrete representation'', that is, a concrete LP-type problem
338: that is basis-equivalent to $(H,w,W,\leq)$.
339:
340: Strictly speaking, if the multiset ${\mathcal H}$ in the
341: concrete LP-type problem has elements with
342: multiplicity bigger than 1, then ${\mathcal H}$ cannot be used
343: as the set of constraints for the abstract LP-type problem (since it
344: is not a set).
345: However, we can bijectively map $\mathcal H$ to a set, i.e., we take
346: any set $H$ with $|H|=|{\mathcal H}|$ and a mapping
347: $f\colon H\to{\mathcal H}$ such that for any $\bar h\in{\mathcal H}$,
348: the number of elements $h\in H$ that map to $\bar h$ is equal to the
349: multiplicity of $\bar h$. For $G\subseteq H$ we then define
350: $w(G)=\min(\bigcap_{g\in G} f(g))$ which gives us a fair abstract LP-type
351: problem $P=(H,w,X,\preceq)$ basis-equivalent to $\mathcal{P}$.
352: In this case, by basis-equivalence we mean the existence of a
353: suitable mapping $f$ together with the condition that $B\subseteq G$
354: is a basis of $G$ in $P$ if and only if the multiset
355: $\{f(b)\colon b\in B\}$ is a basis of $\{f(g)\colon g\in G\}$ in
356: $\mathcal{P}$.
357:
358: \subsection{Violator spaces.} Let $(H,w,W,\leq)$ be an abstract
359: LP-type problem. It is natural to define that a constraint $h\in H$
360: {\em violates\/} a set $G\subseteq H$ of constraints if
361: $w(G\cup\{h\})>w(G)$. For example, in the smallest enclosing ball problem,
362: a point $h$ violates a set $G$ if it lies outside of the smallest ball
363: enclosing $G$ (which is unique).
364:
365: \begin{defn}
366: The \emph{violator mapping} of $(H,w,W,\leq)$ is defined by
367: $\vi(G)=\{h\in H\colon w(G\cup\{h\})>w(G)\}$. Thus, $\vi(G)$ is the set
368: of all constraints violating $G$.
369: \end{defn}
370:
371: It turns out that the knowledge of
372: $\vi(G)$ for all $G\subseteq H$ is enough to describe the ``structure''
373: of an LP-type problem. That is, while we cannot reconstruct $W$, $\leq$,
374: and $w$ from this knowledge, it is natural to consider two LP-type problems
375: with the same mapping $\vi\colon 2^H\to 2^H$ the same (isomorphic). Indeed,
376: the algorithmic primitives needed for implementing the Sharir--Welzl
377: algorithm and the other algorithms for LP-type problems mentioned above
378: can be phrased in terms of testing violation (does $h\in \vi(G)$ hold for
379: a certain set $G\subseteq H$?), and they never deal explicitly with the
380: values of $w$.
381:
382: We now introduce the notion of {\em violator space}:
383:
384: \begin{defn}\label{def:vs}
385: A {\em violator space} is a pair $(H,\vi)$, where $H$ is a finite set
386: and $\vi$ is a mapping $2^H\to2^H$ such that
387:
388: \begin{tabular}{ll}
389: Consistency: & $G\cap \vi(G)=\emptyset$ holds for all $G\subseteq H$, and\\
390: Locality: & for all $F\subseteq G\subseteq H$, where
391: $G\cap \vi(F)=\emptyset$, we have\\
392: & $\vi(G)=\vi(F)$.\\
393: \end{tabular}
394: \end{defn}
395:
396: A basis of a violator space is defined in analogy to a basis of an
397: LP-type problem.
398:
399: \begin{defn}\label{def:vbasis}
400: Consider a violator space $(H,\vi)$. We say that $B\subseteq H$
401: is a {\em basis} if for all proper subsets $F\subset B$ we have
402: $B\cap \vi(F)\neq\emptyset$. For $G\subseteq H$, a \emph{basis of $G$}
403: is a minimal subset $B$ of $G$ with $\vi(B)=\vi(G)$.
404: \end{defn}
405:
406: Observe that a minimal subset $B\subseteq G$ with $\vi(B)=\vi(G)$ is
407: indeed a basis: Assume for contradiction that there is a set
408: $F\subset B$ such that $B\cap \vi(F)=\emptyset$. Locality then
409: yields $\vi(B)=\vi(F)=\vi(G)$, which contradicts minimality of $B$.
410:
411: We will check in Subsection~\ref{sec:abstract->acyclicVS}
412: that the violator mapping of an abstract LP-type
413: problem satisfies the two axioms above. Consistency is immediate:
414: since $w(G)=w(G\cup\{h\})$ for $h\in G$, no element in $G$ violates
415: $G$. The locality condition has the following intuitive interpretation:
416: adding only non-violators to a set does not change the value.
417:
418: We actually show more: given an abstract LP-type problem
419: $(H,w,W,{\leq})$, the pair $(H,\vi)$, with $\vi$ being the violator mapping,
420: is an \emph{acyclic} violator space. (Acyclicity of a violator space
421: will be defined later in Definition \ref{def:vsetord}.) It turns out
422: in Subsection \ref{sec:acyclicVS->concrete}
423: that acyclicity already characterizes the
424: violator spaces obtained from LP-type problems, and thus any acyclic
425: violator space can be represented as an LP-type problem (abstract or
426: concrete). These equivalences are stated in our main theorem.
427:
428: \begin{thm}\label{thm:main}
429: The axioms of abstract LP-type problems, of concrete LP-type problems,
430: and of acyclic violator spaces are equivalent. More precisely, every
431: problem in one of the three classes has a basis-equivalent problem
432: in each of the other two classes.
433: \end{thm}
434:
435: The construction is illustrated on simple instances of problems of
436: linear programming and the smallest enclosing ball in
437: Subsection \ref{sec:examples}. Several more results
438: concerning violator spaces have been achieved in the MSc. thesis of
439: the fourth author \cite{Skovron}.
440:
441: \section{Equivalence of LP-type Problems and Acyclic Violator Spaces}
442: \label{sec:mainproof}
443: In this section we prove Theorem \ref{thm:main}.
444:
445: \subsection{Preliminaries on Violator Spaces}\label{sec:preliminaries}
446:
447: To show that every acyclic violator space $(H,\vi)$ originates from
448: some concrete LP-type problem, we need an appropriate linearly
449: ordered set $X$ of ``points'', and then we will identify the elements of
450: $H$ with certain subsets of $X$.
451:
452: What set $X$ will we take? Recall that for smallest enclosing balls,
453: $X$ is the set of all balls, and the subset for $h\in H$ is the subset
454: of balls containing $h$. It is not hard to see that we may restrict
455: $X$ to smallest enclosing balls of \emph{bases}; in fact, we may
456: choose $X$ as the set of bases, in which case the subset for $h$
457: becomes the set of bases not violated by $h$.
458:
459: This also works for general acyclic violator spaces, with bases suitably
460: ordered. The only blemish is that we may get several minimal bases for
461: $G\subseteq H$; for smallest enclosing balls, this corresponds to the
462: situation in which several bases define the same smallest enclosing
463: ball. To address this, we will declare such bases as equivalent
464: and choose $X$ as the set of all equivalence classes instead.
465:
466: In the following, we fix a violator space $(H,\vi)$. The set of all
467: bases in $(H,\vi)$ will be denoted by $\mathcal B$.
468:
469: \begin{defn}\label{def:vequiv}
470: $B,C\in \mathcal B$ are \emph{equivalent}, $B\sim C$, if $\vi(B)=\vi(C)$.
471: \end{defn}
472:
473: Clearly, the relation $\sim$ defined on $\mathcal{B}$ is an equivalence
474: relation. The equivalence class containing a basis $B$ will be
475: denoted by $[B]$.
476:
477: Now we are going to define an ordering of the bases, and we derive from this
478: an ordering of the equivalence classes as well as the notion of acyclicity
479: in violator spaces.
480:
481: \begin{defn}\label{def:vsetord}
482: For $F,G\subseteq H$ in a violator space $(H,\vi)$, we say that
483: $F\leq_0 G$ ($F$ is {\em locally smaller} than $G$) if
484: $F\cap \vi(G)=\emptyset$.
485:
486: For equivalence classes $[B],[C]\in\bfaceq$, we say that
487: $[B]\leq_0 [C]$ if there exist $B'\in[B]$ and $C'\in[C]$ such
488: that $B'\leq_0 C'$.
489:
490: We define the relation $\leq_1$ on the equivalence classes as
491: the transitive closure of $\leq_0$. The relation $\leq_1$
492: is clearly reflexive and transitive. If it is antisymmetric, we say that the
493: violator space is {\em acyclic}, and we define the relation $\leq$
494: as an arbitrary linear extension of $\leq_1$.
495: \end{defn}
496:
497: The intuition of the \emph{locally-smaller} notion comes from LP-type
498: problems: if no element of $F$ violates $G$, then $G\cup F$ has the
499: same value as $G$ (this is formally proved in Lemma
500: \ref{lem:conseqloc} below), and monotonicity yields that value-wise,
501: $F$ is smaller than or equal to $G$.
502:
503: Note that in the definition of $[B]\leq_0[C]$ we do not require
504: $B'\leq_0 C'$ to hold for {\em every} $B'$ and $C'$. In fact
505: $B'\not\leq_0 C'$ may happen for some bases $B'$ and $C'$,
506: but $C'\leq_0 B'$ can not hold (which can easily be shown).
507: %
508: %Assume for contradiction
509: %that $[B]\leq_0 [C]$ and $[C]\leq_0 [B]$. Then $B'\cap V(C')=\emptyset$
510: %and $C'\cap V(B')=\emptyset$ for some $B'$ and $C'$ out of the
511: %respective equivalence classes (because all bases in the same
512: %equivalence class have the same violators). This is the same as
513: %$(B'\cup C')\cap V(C')=\emptyset$ and $(B'\cup C')\cap V(B')=\emptyset$,
514: %which with locality yields the contradiction $V(C')=V(B'\cup C')=V(B')$.
515:
516: To show that acyclicity does not always hold, we conclude this
517: section with an example of a cyclic violator space.
518:
519: We begin with an intuitive geometric description; see Figure \ref{fig:cyclic}.
520: We consider a triangle without the center point. We say that a point is
521: ``locally smaller'' if it is farther clockwise with respect to the center.
522: The constraints in our violator space are the three halfplanes $f,g,h$.
523:
524: \figeps{cyclic}{5}{A cyclic violator space.}
525:
526: The locally smallest point within each halfplane is marked, and a halfplane
527: violates a set of halfplanes if it does not contain the locally smallest point
528: in their intersection.
529:
530: Now we specify the corresponding violator space formally.
531: We have $H=\{f,g,h\}$, and $\vi$ is given by the following table:
532: \begin{center}
533: \tabcolsep0.3cm
534: \begin{tabular}{|c||c|c|c|c|c|c|c|c|}\hline
535: $G$ & $\emptyset$ & $f$ & $g$ & $h$ & $f,g$ & $f,h$ & $g,h$ & $f,g,h$ \\ \hline
536: $\vi(G)$ & $f,g,h$ & $h$ & $f$ & $g$ & $h$ & $g$ & $f$ & $\emptyset$ \\ \hline
537: \end{tabular}
538: \end{center}
539: This $(H,\vi)$ is really a violator space, since we can easily check both
540: consistency and locality. The bases are $\emptyset$, one-element sets, and
541: $H$. We have $\{f\}\leq_0\{h\}\leq_0\{g\}\leq_0\{f\}$, but none of
542: the one-element bases are equivalent; i.e., $\leq_1$ is not antisymmetric.
543:
544: \subsection{Abstract LP-type Problems yield Acyclic Violator Spaces}
545: \label{sec:abstract->acyclicVS}
546: In this subsection, we show that the violator mapping of
547: an abstract LP-type problem is an acyclic violator space.
548: To this end, we need the following two lemmas.
549:
550: \begin{lem}\label{lem:conseqloc}
551: Consider an abstract LP-type problem $(H,w,W,\leq)$ with violator mapping
552: $\vi$. Let $A,B\subseteq H$, where $B$ is not violated by any $h\in A$
553: ($A\cap \vi(B)=\emptyset$). Then $w(A\cup B)=w(B)$.
554: \end{lem}
555:
556: \pf From monotonicity, we immediately obtain the inequality ``$\geq$''.
557: The inequality ``$\leq$'' can be shown by induction on $|A|$.
558: If $|A|=1$, i.e., $A=\{h\}$, then $w(B\cup\{h\})>w(B)$ would imply that
559: $B$ is violated by $h\in A$, a contradiction.
560:
561: Let $|A|>1$ and $A=A_0\mathbin{\dot\cup}\{h\}$ (disjoint union).
562: From the induction hypothesis we have $w(B\cup A_0)=w(B)$. Now, if
563: $w(B\cup A_0)<w(B\cup A_0\cup \{h\})$, then by locality (for $B\cup A_0$,
564: $B$ and $h$) we get $w(B)<w(B\cup\{h\})$. This means that $h\in\vi(B)$,
565: and since $h\in A$ we have $h\in A\cap\vi(B)$, a contradiction.
566: So $w(B)=w(B\cup A_0) \geq w(B\cup A_0\cup\{h\})=w(B\cup A)$. We have
567: proved $w(A\cup B)\leq w(B)$.
568: \qed
569:
570: \begin{lem}\label{lem:separbasis}
571: Consider an abstract LP-type problem $(H,w,W,\leq)$ with violator mapping
572: $\vi$. Then for any $A,B\subseteq H$ with $\vi(A)=\vi(B)$ we have $w(A)=w(B)$.
573: Conversely, $w(A)=w(B)=w(A\cup B)$ implies $\vi(A)=\vi(B)$. In particular,
574: if $A\subseteq B$ and $w(A)=w(B)$, then $\vi(A)=\vi(B)$.
575: \end{lem}
576:
577: Note that the condition $w(A)=w(B)$ generally does not suffice for
578: $\vi(A)=\vi(B)$. For example, having any $H$, we can define $w$ by $w(G)=|G|$
579: for all $G\subseteq H$ (it can be checked that it is an abstract LP-type
580: problem). Then any $G$'s of the same size
581: have the same $w$, however, $\vi(G)=H\setminus G$, and so no distinct
582: $G$'s share the value of $\vi$. Roughly speaking, the equality
583: $w(A)=w(B)$ may hold just ``by accident''. This is one way in which
584: we can see that $w$ by itself does not reflect the combinatorial structure
585: of the problem in a natural way.
586:
587: \begin{pf*}{PROOF of Lemma \ref{lem:separbasis}.} Let $w(A)\neq w(B)$.
588: Without loss of generality we assume $w(A)>w(B)$ (note that here we use
589: the linearity of the ordering $\leq$). If $A\cap \vi(B)=\emptyset$, from
590: Lemma \ref{lem:conseqloc} we would get $w(A\cup B)=w(B)$, which contradicts
591: $w(A\cup B)\geq w(A)>w(B)$. So there necessarily exists $h\in A\cap \vi(B)$,
592: but since $h\in A$, we have $h\not\in \vi(A)$. So $\vi(A)\neq \vi(B)$.
593:
594: Conversely, suppose $w(A)=w(B)=w(A\cup B)$. We want to show $\vi(A)=\vi(B)$,
595: i.e., that $w(A)<w(A\cup\{h\})$ holds iff $w(B)<w(B\cup\{h\})$
596: holds. By symmetry, it suffices to show only one of the implications.
597: We assume $w(A)<w(A\cup\{h\})$. Then $w(A\cup B)=w(A)<w(A\cup\{h\}) \leq
598: w(A\cup B\cup\{h\})$. Since $B\subseteq A\cup B$ and $w(B)=w(A\cup B)$, we
599: may use locality, which gives $w(B)<w(B\cup\{h\})$. So the desired
600: equivalence holds.
601: \qed
602: \end{pf*}
603:
604: \begin{prop}\label{prop:abstract->acyclicVS}
605: Consider an abstract LP-type problem $(H,w,W,\leq)$, and let $\vi$ be its
606: violator mapping. Then $(H,\vi)$ is an acyclic violator space.
607: Moreover, $(H,\vi)$ is basis-equivalent to $(H,w,W,\leq)$.
608: \end{prop}
609:
610: \pf Clearly $G\cap\vi(G)=\emptyset$, since $w(G\cup\{g\})=w(G)$
611: for any $g\in G$, so consistency holds. If $G\cap \vi(F)=\emptyset$
612: for $F \subseteq G$, then by Lemma \ref{lem:conseqloc} we get
613: $w(F\cup G)=w(F)$. Since $F\subseteq G$, we have $F \cup G=G$ and so
614: $w(G)=w(F)$. Lemma \ref{lem:separbasis} then yields $\vi(G)=\vi(F)$,
615: so locality holds.
616:
617: We proceed to prove acyclicity of $(H,\vi)$. Fix $[B]$ and $[C]$,
618: $[B]\neq [C]$, with $[B]\leq_0[C]$, that is $B'\cap \vi(C')=\emptyset$ for
619: some $B'\in [B]$ and $C'\in [C]$. Lemma \ref{lem:conseqloc} implies
620: $w(C')=w(B'\cup C')$. For contradiction, assume $w(B')\geq w(C')$; then
621: $w(B')\geq w(B'\cup C')$ which with monotonicity yields $w(B')=w(B'\cup C')=w(C')$.
622: Lemma \ref{lem:separbasis} gives $\vi(B')=\vi(C')$, a contradiction to
623: $[B]\neq [C]$.
624: Thus $[B]\leq_0[C]$ for $[B]\neq [C]$ implies $w(B')<w(C')$ for some
625: bases $B'$ and $C'$ out of the respective equivalence classes. By
626: Lemma \ref{lem:separbasis}, $w(B')$ is the same for all $B'\in [B]$
627: (because all bases in $[B]$ have the same violators). Therefore, by chaining
628: several $\leq_0$'s we also get $w(B')<w(C')$ for $[B]\leq_1[C]$. This
629: proves that $\leq_1$ is necessarily antisymmetric (since $\leq$ is an
630: ordering of $W$).
631:
632: Finally, observe that by Lemma \ref{lem:separbasis}, $B\subseteq G$ is an
633: inclusion-minimal subset of $G$ with $w(B)=w(G)$ if and only if $B$ is
634: an inclusion-minimal subset of $G$ with $\vi(B)=\vi(G)$. So, $B\subseteq G$
635: is a basis of $G$ in $(H,w,W,\leq)$ if and only if $B$ is a basis of $G$
636: in $(H,\vi)$. Thus $(H,\vi)$ is basis-equivalent to $(H,w,W,\leq)$.
637: \qed
638:
639: At first glance, one might think that for $F\subseteq G$ we should
640: have $\vi(F)\supseteq \vi(G)$. Unfortunately, this is not the case, as
641: the linear programming example in Figure \ref{fig:counterexample} shows (the
642: $y$-coordinate is to be minimized).
643:
644: \figeps{counterexample}{6}{A linear programming example ($F=\{h_1,h_2\}
645: \subseteq G=\{h_1,h_2,h_3\}$ with $\vi(G)\not\subseteq\vi(F)$).}
646:
647: We put $F=\{h_1,h_2\}$ and $G=\{h_1,h_2,h_3\}\supseteq F$. The point $1$
648: is minimum in the intersection of $F$, and $2$ is minimum in the
649: intersection of $G$. We have $1\in h^*$, $2\not\in h^*$, and so
650: $h^*\not\in\vi(F)$ and $h^*\in\vi(G)$.
651:
652: \subsection{Acyclic Violator Spaces yield Concrete LP-type Problems}
653: \label{sec:acyclicVS->concrete}
654: The following proposition is the last ingredient for Theorem \ref{thm:main}.
655: \begin{prop}\label{prop:acyclicVS->concrete}
656: Every acyclic violator space $(H,\vi)$ can be represented as a concrete
657: LP-type problem that is basis-equivalent to $(H,\vi)$.
658: \end{prop}
659:
660: \pf We are given an acyclic violator space $(H, \vi)$ and we define the
661: mapping $S\colon H\to 2^\bfaceq$ that will act as a ``concretization''
662: of the constraints in $H$:
663: \[
664: S(h)=\{[B]\colon B\in{\mathcal B},\, h\not\in \vi(B)\}.
665: \]
666: Further, let ${\mathcal H}$ be the image of the mapping $S$ taken as
667: a multiset, i.e.,
668: \[
669: {\mathcal H}=\{S(h)\colon h\in H\}.
670: \]
671: Thus, $S$ is a bijection between $H$ and $\mathcal H$.
672: By saying that a mapping $S$ is a bijection between a set and
673: a multiset we mean that for any $\bar h\in{\mathcal H}$, the number of
674: $h\in H$ that map to $\bar h$ is equal to the multiplicity of $\bar h$.
675: Note that we cannot use some common properties of set bijections;
676: for instance we have to avoid using the inverse mapping $S^{-1}$.
677:
678: Additionally, let
679: $\sigma$ be the induced bijection of $2^H$ and $2^{\mathcal H}$
680: defined by $\sigma(G)=\{S(h)\colon h\in G\}$, for $G\subseteq H$.
681:
682: Now, consider the triple $(\bfaceq,\leq,{\mathcal H})$, where $\leq$ is an
683: arbitrary linear extension of $\leq_1$ (such an extension exists since
684: $(H,\vi)$ is acyclic and $\leq_1$ therefore antisymmetric). This is a
685: concrete LP-type problem: The only thing to check is the existence of
686: a minimal element of every nonempty intersection $\bigcap{\mathcal G}$
687: ($\mathcal G \subseteq \mathcal H$), which is guaranteed by the linearity
688: of $\leq$ (remember from Definition \ref{def:concrete_LPtype} that
689: $\bigcap{\mathcal G}:=\bigcap_{G\in\mathcal{G}}G$).
690:
691: It remains to prove basis-equivalence, which we do with the following two
692: lemmas.
693:
694: \begin{lem}\label{lem:equ1}
695: If $B$ is an inclusion-minimal subset of $G$ with $\vi(B)=\vi(G)$ in $(H,\vi)$
696: (that is, $B$ is a basis of $G$), then $\min(\bigcap\sigma(B))=
697: \min(\bigcap\sigma(G))$ in $(\bfaceq,\leq,{\mathcal H})$.
698: \end{lem}
699:
700: \pf It is clear that $[B]\in \bigcap\sigma(G)$. Therefore, showing
701: that there is no other basis in $\bigcap\sigma(G)$ that is locally
702: smaller than $[B]$ proves the lemma, because then
703: $\min(\bigcap\sigma(G))=[B]=\min(\bigcap\sigma(B))$ (the second
704: equality holds since $B$ is a basis of $B$; just replace $G$ by $B$ in
705: the following proof). Assume for contradiction that a $C$ with
706: $[C]\neq [B]$, $[C]\in\bigcap\sigma(G)$ and $C\leq_0[B]$ exists.
707: By $[C]\in\bigcap\sigma(G)$ we have $G\cap\vi(C)=\emptyset$, which is
708: equivalent to
709: \[
710: (G\cup C)\cap\vi(C)=\emptyset,
711: \]
712: and by $C\leq_0[B]$ we have $C\cap\vi(B)=\emptyset$ which is equivalent
713: to (because $B$ is a basis of $G$)
714: \[
715: (G\cup C)\cap\vi(B)=\emptyset.
716: \]
717: Applying locality in $(H,\vi)$ to these two equations tells us that
718: $\vi(C)=\vi(B)$, a contradiction to $[C]\neq [B]$.
719: \qed
720:
721: \begin{lem}\label{lem:equ2}
722: If $\sigma(B)$ is an inclusion-minimal submultiset of $\sigma(G)$ with
723: $\min(\bigcap\sigma(B))=\min(\bigcap\sigma(G))$ in
724: $(\bfaceq,\leq,{\mathcal H})$ (that is, $\sigma(B)$ is a basis of
725: $\sigma(G)$), then $\vi(B)=\vi(G)$ in $(H,\vi)$.
726: \end{lem}
727:
728: \pf Let $A$ be a basis of $B$, so $\vi(A)=\vi(B)$. Note that
729: $[A]\in\bigcap\sigma(B)$. Let $[C]=\min(\bigcap\sigma(B))$,
730: thus $B\cap\vi(C)=\emptyset$ and therefore also $A\cap\vi(C)=\emptyset$.
731: This means that $[A]\leq_0 [C]$ from which we conclude that $[A]=[C]$.
732: From $\min(\bigcap\sigma(G))=[C]$ we get $G\cap\vi(C)=\emptyset$ which
733: is equivalent to
734: \[
735: G\cap\vi(B)=\emptyset.
736: \]
737: As $\sigma(B)\subseteq\sigma(G)$ if and only if $B\subseteq G$, we can apply
738: locality and derive $\vi(B)=\vi(G)$ as needed.
739: \qed
740:
741: Lemmas \ref{lem:equ1} and \ref{lem:equ2} prove that $(H,\vi)$ and
742: $(\bfaceq,\leq,{\mathcal H})$ are basis-equivalent, in the sense
743: that $B$ is a basis of $G$ in $(H,\vi)$ if and only if $\sigma(B)$
744: is a basis of $\sigma(G)$ in $(\bfaceq,\leq,{\mathcal H})$:
745: Starting with a basis $B$ of $G$ in $(H,\vi)$, Lemma \ref{lem:equ1}
746: yields $\min(\bigcap\sigma(B))=\min(\bigcap\sigma(G))$. This
747: $\sigma(B)$ is inclusion-minimal w.r.t. $\sigma(G)$, since otherwise
748: Lemma \ref{lem:equ2} would yield a contradiction to the
749: inclusion-minimality of $B$ w.r.t. $G$ (where we again use that
750: $\sigma(B)\subseteq\sigma(G)$ if and only if $B\subseteq G$). The
751: reasoning for inclusion-minimality in the other direction is analogous.
752: This concludes the proof of Proposition \ref{prop:acyclicVS->concrete}.
753: \qed
754:
755: Propositions \ref{prop:abstract->acyclicVS} and \ref{prop:acyclicVS->concrete},
756: together with the fact that every concrete LP-type problem can
757: be transformed into an abstract one (as described below
758: Definition \ref{def:concrete_basis}), yield Theorem \ref{thm:main}.
759:
760: \subsection{Examples}\label{sec:examples}
761: Here we present some particular abstract LP-type problems and we
762: demonstrate the construction (via acyclic violator spaces) of their
763: concrete representations.
764:
765: Let $a$, $b$, $c$ and $d$ be the vertices of a unit square (in the
766: counterclockwise order); let $H=\{a,b,c,d\}$. For $G\subseteq H$ let
767: $w(G)$ be the radius of the smallest circle enclosing all the points
768: of $G$ (for $G=\emptyset$ put $w(G)=-\infty$). The corresponding
769: acyclic violator space is described by the following table:
770: \begin{center}
771: \tabcolsep0.4cm
772: \begin{tabular}{|c||c|c|c|c|c|c|c|c|}\hline
773: $G$ & $\emptyset$ & $a$ & $b$ & $c$ & $d$ & $ab$ & $ac$ & $ad$ \\ \hline
774: $\vi(G)$ & $abcd$ & $bcd$ & $acd$ & $abd$ & $abc$ & $cd$ & $\emptyset$ & $bc$ \\ \hline \hline
775: $G$ & $bc$ & $bd$ & $cd$ & $abc$ & $abd$ & $acd$ & $bcd$ & $abcd$ \\ \hline
776: $\vi(G)$ & $ad$ & $\emptyset$ & $ab$ & $\emptyset$ & $\emptyset$ & $\emptyset$ & $\emptyset$ & $\emptyset$ \\ \hline
777: \end{tabular}
778: \end{center}
779: The bases are $\emptyset$, $a$, $b$, $c$, $d$, $ab$, $ac$, $ad$, $bc$,
780: $bd$, $cd$; the only equivalent pair is $ac\sim bd$. There is no
781: inconvenience concerning differences between $\leq_0$ on sets and
782: equivalence classes and $\leq_1$; the ordering $\leq_1$ is given by the
783: Hasse diagram in Figure \ref{fig:hasse}.
784:
785: \figeps{hasse}{4}{Hasse diagram from smallest enclosing circle of the
786: vertices of a square.}
787:
788: As the linear extension $\leq$ of $\leq_1$ we may choose
789: $\emptyset<a<b<c<d<ab<bc<cd<ad<[ac]$.
790: Finally, the concrete representation $S$ is as follows:
791: \begin{center}
792: \tabcolsep0.4cm
793: \begin{tabular}{|c||c|c|c|c|}\hline
794: $h$ & $a$ & $b$ & $c$ & $d$ \\ \hline
795: $S(h)$ & $a,ab,ad,[ac]$ & $b,ab,bc,[ac]$ & $c,bc,cd,[ac]$
796: & $d,cd,ad,[ac]$ \\ \hline
797: \end{tabular}
798: \end{center}
799: In the geometric view that we have mentioned earlier, $S(a)$ corresponds to
800: the set of all ``canonical'' (i.e., basic) balls that contain the point
801: $a$ (inside or on the boundary). The same holds for the other points.
802:
803: As the other example, consider the following LP problem in the positive
804: orthant (rotated by 45 degrees for convenience). Beside the restriction
805: to the positive orthant, the constraints are the four halfplanes depicted in
806: Figure \ref{fig:lp_example}. The optimization direction is given by the arrow.
807:
808: \figeps{lp_example}{6}{Illustration example -- linear programming.}
809:
810: Here the violator space bases are $\emptyset$, $a$, $b$, $c$, $d$, $ac$, $ad$, $bc$,
811: $bd$; the equivalence classes are $O=\emptyset$, $A=\{a\}$, $B=\{b\}$,
812: $C=\{c\}$, $D=\{d\}$ and $Q=\{ac\}\sim\{ad\}\sim\{bc\}\sim\{bd\}$.
813: Note that the equivalence classes correspond to the points in the plane.
814: We have $O\leq_1 B\leq_1 A\leq_1 Q$ and $O\leq_1 C\leq_1 D\leq_1 Q$; we
815: choose $\leq$ to be $O<B<A<C<D<Q$. The concrete representation is
816: \begin{center}
817: \tabcolsep0.4cm
818: \begin{tabular}{|c||c|c|c|c|}\hline
819: $h$ & $a$ & $b$ & $c$ & $d$ \\ \hline
820: $S(h)$ & $A,Q$ & $A,B,Q$ & $C,D,Q$ & $D,Q$ \\ \hline
821: \end{tabular}.
822: \end{center}
823: Here we may interpret $S(a)$ as the set of all ``canonical'' points
824: lying in the halfplane $a$.
825:
826: To see why we allow $\mathcal H$ in the definition of a concrete
827: LP-type problem to be a multiset, consider the abstract LP-type
828: problem with $H=\{a,b\}$ and $w(G)=0$ for every $G\subseteq H$.
829: The only basis is $\emptyset$ and it is not violated by any
830: $h\in H$. Thus we have $S(a)=S(b)=\{[\emptyset]\}$. If we do not allow
831: $\mathcal H$ to be a multiset, we have ${\mathcal H}=\{S(a),S(b)\} =
832: \bigl\{ \{[\emptyset]\} \bigr\}$ with only one constraint; it seems
833: improper to define this to be basis-equivalent to $H$.
834: We could alter the construction of $\mathcal H$ and get
835: $S(a)=\{0\}$, $S(b)=\{0,1\}$,
836: which does represent the original abstract LP-type problem with
837: $\mathcal H$ being a set; however, we believe that our definition
838: catches the structure in a more straightforward way, although it may
839: seem unusual at first glance.
840:
841: \section{Clarkson's Algorithms}\label{sec:clarkson}
842: We show that Clarkson's randomized reduction scheme, originally
843: developed for linear programs with many constraints and few variables,
844: actually works for general (possibly cyclic) violator spaces. The two
845: algorithms of Clarkson involved in the reduction have been analyzed
846: for LP and LP-type problems before
847: \cite{c-lvali-95,gw-lpraf-96,cm-ltdao-96}; the analysis we give below
848: is almost identical on the abstract level. Our new contribution is
849: that the combinatorial properties underlying Clarkson's algorithms
850: also hold for violator spaces.
851:
852: We start off by deriving these combinatorial properties; the
853: analysis of Clarkson's reduction scheme is included for completeness.
854:
855: \subsection{Violator spaces revisited}
856: We recall that an abstract LP-type problem is of the form $(H,w,W,\leq)$.
857: In this subsection we will view a violator space as
858: an ``LP-type problem without the order $\leq$'', i.e., we will
859: only care whether two subsets $F$ and $G$, $F\subseteq G\subseteq H$, have
860: the same value (and therefore the same violators, see Lemma
861: \ref{lem:separbasis}), but not how they compare under the order $\leq$.
862: It turns out that the order is irrelevant for Clarkson's algorithms.
863:
864: Even without an order, we can talk about monotonicity in violator
865: spaces:
866: \begin{lem}
867: Any violator space $(H,\vi)$ satisfies
868:
869: \begin{tabular}{ll} Monotonicity: & $\vi(F)=\vi(G)$ implies
870: $\vi(E)=\vi(F)=\vi(G)$,\\
871: & for all sets $F\subseteq E\subseteq G\subseteq H$.
872: \end{tabular}
873: \label{lem:monotonicity}
874: \end{lem}
875:
876: \pf Assume $\vi(E)\neq \vi(F),\vi(G)$. Then locality yields
877: $\emptyset \neq E\cap \vi(F) = E\cap \vi(G)$ which contradicts
878: consistency. \qed
879:
880: Recall Definition \ref{def:vbasis}: A basis is a set $B$ satisfying
881: $B\cap \vi(F)\neq\emptyset$ for all proper subsets $F$ of $B$. A basis
882: of $G$ is an inclusion-minimal subset of $G$ with the same violators.
883: This can be used to prove the following observation, well-known to
884: hold for LP-type problems \cite{gw-lpraf-96}.
885:
886: \begin{obs}\label{obs:extreme}
887: Let $(H,\vi)$ be a violator space. For $R\subseteq H$ and all $h\in H$, we have
888: \begin{enumerate}
889: \item[(i)] $\vi(R)\neq \vi(R\cup\{h\})$ if and only if $h\in \vi(R)$, and
890: \item[(ii)] $\vi(R)\neq \vi(R\setminus\{h\})$ if and only if $h$ is contained
891: in every basis of $R$.
892: \end{enumerate}
893: An element $h$ such that (ii) holds is called \emph{extreme} in $R$.
894: \end{obs}
895:
896: \pf (i) If $h\notin\vi(R)$, we get $\vi(R)=\vi(R\cup\{h\})$ by
897: Lemma \ref{lem:conseqloc}. If $h\in\vi(R)$, then $\vi(R)\neq \vi(R\cup\{h\})$
898: is a consequence of consistency applied to $G=R\cup\{h\}$. (ii) if
899: $\vi(R)=\vi(R\setminus\{h\})$, there is a basis $B$ of $R\setminus\{h\}$,
900: and this basis is also a basis of $R$ not containing $h$. Conversely,
901: if there is some basis $B$ of $R$ not containing $h$, then
902: $\vi(R)=\vi(R\setminus\{h\})$ follows from monotonicity.
903: \qed
904:
905: We are particularly interested in violator spaces with small bases.
906:
907: \begin{defn}
908: Let $(H,\vi)$ be a violator space. The size of a largest basis is called
909: the \emph{combinatorial dimension} $\delta=\delta(H,\vi)$ of $(H,\vi)$.
910: \end{defn}
911:
912: Observation \ref{obs:extreme} implies that in a violator space of
913: combinatorial dimension $\delta$, every set has at most $\delta$
914: extreme elements. This in turn yields a bound for the \emph{expected}
915: number of violators of a random subset of constraints, using the
916: \emph{sampling lemma} \cite{GWSampl01}.
917:
918: \begin{lem}{\rm\cite{GWSampl01}}
919: Consider a triple $(H,w,W)$, where $w$ is a function mapping subsets of
920: the set $H$ to the set $W$ (not necessarily ordered).
921: For $R\subseteq H$, we define
922: \begin{eqnarray*}
923: V(R) &:=& \{h\in H\setminus R: w(R)\neq w(R\cup\{h\}), \\
924: X(R) &:=& \{h\in R: w(R)\neq w(R\setminus\{h\}).
925: \end{eqnarray*}
926: For $0\leq r\leq |H|$, let $v_r$ be the expected value of $|V(R)|$,
927: for $R$ chosen uniformly at random among all subsets of $H$ with
928: $r$ elements. $x_r$ is defined similarly as the expected value
929: of $|X(R)|$. Then for $0\leq r<n$, the following equality holds.
930: \[\frac{v_r}{n-r}=\frac{x_{r+1}}{r+1}.\]
931: \end{lem}
932:
933: To apply this in our situation, we fix a set $W\subseteq H$, and we
934: define $w(R)=\vi(W\cup R)$. Since then $|X(R)|\leq\delta$ for all
935: $R$, the following Corollary is obtained.
936:
937: \begin{cor}\label{cor:sampling}
938: Let $(H,\vi)$ be a violator space of combinatorial dimension $\delta$
939: and $W\subseteq H$ some fixed set. Let $v_r$ be the expected number of
940: violators of the set $W\cup R$, where $R\subseteq H$ is a random subset
941: of size $r< n=|H|$. Then
942: \[
943: v_r \leq \delta \frac{n-r}{r+1}.
944: \]
945: \end{cor}
946:
947: \subsection{The Trivial Algorithm}
948: Given a violator space $(H,\vi)$ of combinatorial dimension $\delta$,
949: the goal is to find a basis of $H$. For this, we assume availability
950: of the following primitive.
951:
952: \begin{primitive}\label{prim:viol}
953: Given $G\subseteq H$ and $h\in H\setminus G$, decide whether $h\in \vi(G)$.
954: \end{primitive}
955:
956: Given this primitive, the problem can be solved in a brute-force manner
957: by going through all sets of size $\leq \delta$, testing each of them for
958: being a basis of $H$. More generally, $B\subseteq G$ is a basis of $G$ if
959: and only if
960: \[
961: \begin{array}{rcll}
962: h&\in& \vi(B\setminus\{h\}), \quad & \forall h\in B, \\
963: h&\notin& \vi(B), & \forall h\in G\setminus B.
964: \end{array}
965: \]
966: Consequently, the number of times the primitive needs to be invoked
967: in order to find a basis of $H$ is at most
968: \[
969: n \sum_{i=0}^{\delta}{n\choose i} = O(n^{\delta+1}).
970: \]
971: The next two subsections show that this can be substantially improved.
972:
973: \subsection{Clarkson's First Algorithm}
974: Fix a violator space $(H,\vi)$ of combinatorial dimension $\delta$,
975: implicitly specified through Primitive \ref{prim:viol}. Clarkson's
976: first algorithm calls Clarkson's second algorithm (\keyw{Basis2})
977: as a subroutine. Given $G\subseteq H$, both algorithms compute a basis
978: $B$ of $G$.
979:
980: \begin{tabbing}
981: \quad \=\quad \=\quad \=\quad \=\quad \kill
982: \keyw{Basis1}$(G)$: \\
983: \> (* computes a basis $B$ of $G$ *) \\
984: \> \keyw{IF} $|G|\leq 9\delta^2$ \keyw{THEN} \\
985: \>\> \keyw{RETURN} \keyw{Basis2}$(G)$ \\
986: \> \keyw{ELSE} \\
987: \>\> $r:=\lfloor\delta\sqrt{|G|}\rfloor$ \\
988: \>\> $W:=\emptyset$ \\
989: \>\> \keyw{REPEAT} \\
990: \>\>\> choose $R$ to be a random $r$-element subset
991: of $G$, $R\in{G\choose r}$ \\
992: \>\>\> $C:=\keyw{Basis2}(W\cup R)$ \\
993: \>\>\> $\vi:=\{h\in G\setminus C\colon h\in \vi(C)\}$\\
994: \>\>\> \keyw{IF} $|\vi|\leq 2\sqrt{|G|}$ \keyw{THEN} \\
995: \>\>\>\> $W:=W\cup \vi$ \\
996: \>\>\> \keyw{END} \\
997: \>\> \keyw{UNTIL} $\vi=\emptyset$ \\
998: \>\> \keyw{RETURN} $C$ \\
999: \> \keyw{END} \\
1000: \end{tabbing}
1001:
1002: Assuming \keyw{Basis2} is correct, this algorithm is correct as well:
1003: if $B$ is a basis of $W\cup R\subseteq G$ that in addition has no violators
1004: in $G$, $B$ is a basis of $G$. Moreover, the algorithm augments the working
1005: set $W$ at most $\delta$ times, which is guaranteed by the following
1006: observation.
1007:
1008: \begin{obs}\label{obs:basis}
1009: If $C\subseteq G$ and $G\cap \vi(C)\neq \emptyset$, then $G\cap \vi(C)$
1010: contains at least one element from every basis of $G$.
1011: \end{obs}
1012:
1013: \pf Let $B$ be a basis of $G$. Assuming
1014: \[
1015: \emptyset = B\cap G\cap \vi(C) = B \cap \vi(C),
1016: \]
1017: consistency yields $C\cap \vi(C)=\emptyset$, implying
1018: $(B\cup C)\cap \vi(C)=\emptyset$. From locality and monotonicity (Lemma
1019: \ref{lem:monotonicity}), we get
1020: \[
1021: \vi(C)=\vi(B\cup C)=\vi(G),
1022: \]
1023: meaning that $G\cap \vi(G)= G\cap \vi(C)=\emptyset$, a contradiction.
1024: \qed
1025:
1026: It is also clear that \keyw{Basis2} is called only
1027: with sets of size at most $3\delta\sqrt{|G|}$. Finally, the expected
1028: number of iterations through the \keyw{REPEAT} loop is bounded by
1029: $2\delta$: by Corollary \ref{cor:sampling} (applied to $(G,\vi|_G)$) and the
1030: Markov inequality, the expected number of calls to \keyw{Basis2}
1031: before we next augment $W$ is bounded by $2$.
1032:
1033: \begin{lem}
1034: Algorithm \keyw{Basis1} computes a basis of $G$ with an expected number
1035: of at most $2\delta|G|$ calls to Primitive \ref{prim:viol}, and an expected
1036: number of at most $2\delta$ calls to \keyw{Basis2}, with sets of size
1037: at most $3\delta\sqrt{|G|}$.
1038: \end{lem}
1039:
1040: \subsection{Clarkson's Second Algorithm}
1041: This algorithm calls the trivial algorithm as a subroutine.
1042: Instead of adding violated constraints to a working set,
1043: it gives them larger probability of being selected in
1044: further iterations. Technically, this is done by maintaining
1045: $G$ as a multiset, where $\mu(h)$ denotes the multiplicity
1046: of $h$ (we set $\mu(F):=\sum_{h\in F}\mu(h)$). Sampling from
1047: $G$ is done as before, imagining that $G$ contains $\mu(h)$
1048: copies of the element $h$.
1049:
1050: \begin{tabbing}
1051: \quad \=\quad \=\quad \=\quad \=\quad \kill
1052: \keyw{Basis2}$(G)$: \\
1053: \> (* computes a basis $B$ of $G$ *) \\
1054: \> \keyw{IF} $|G|\leq 6\delta^2$ \keyw{THEN} \\
1055: \>\> \keyw{RETURN} \keyw{Trivial}$(G)$ \\
1056: \> \keyw{ELSE} \\
1057: \>\> $r:=6\delta^2$ \\
1058: \>\> \keyw{REPEAT} \\
1059: \>\>\> choose random $R\in{G\choose r}$ \\
1060: \>\>\> $C:=\keyw{Trivial}(R)$ \\
1061: \>\>\> $\vi:=\{h\in G\setminus C\colon h\in \vi(C)\}$ \\
1062: \>\>\> \keyw{IF} $\mu(\vi)\leq \mu(G)/3\delta$ \keyw{THEN} \\
1063: \>\>\>\> $\mu(h) := 2\mu(h), \quad h\in \vi$ \\
1064: \>\>\> \keyw{END} \\
1065: \>\> \keyw{UNTIL} $\vi=\emptyset$ \\
1066: \>\> \keyw{RETURN} $C$ \\
1067: \> \keyw{END} \\
1068: \end{tabbing}
1069:
1070: Invoking Corollary \ref{cor:sampling} again (which also applies to
1071: multisets as we use them), we see that the expected number of
1072: calls to \keyw{Trivial} before we next reweight elements
1073: (a \emph{successful} iteration), is bounded by $2$. It remains
1074: to bound the number of successful iterations.
1075:
1076: \begin{lem}
1077: Let $k$ be a positive integer. After $k\delta$ successful iterations,
1078: we have \[
1079: 2^k \leq \mu(B)\leq |G|e^{k/3},
1080: \]
1081: for every basis $B$ of $G$. In particular, $k< 3\ln |G|$.
1082: \end{lem}
1083:
1084: \pf Every successful iteration multiplies the total weight
1085: of elements in $G$ by at most $(1+1/3\delta)$, which gives the upper
1086: bound (not only for $\mu(B)$ but actually for $\mu(G)$). For the
1087: lower bound, we use Observation \ref{obs:basis} again to argue
1088: that each successful iteration doubles the weight of some element in $B$,
1089: meaning that after $k\delta$ iterations, one element has been doubled at
1090: least $k$ times. Because the lower bound exceeds the upper bound for
1091: $k\geq 3\ln |G|$, the bound on $k$ follows.
1092: \qed
1093:
1094: Summarizing, we get the following lemma.
1095:
1096: \begin{lem}
1097: Algorithm \keyw{Basis2} computes a basis of $G$ with an expected number
1098: of at most $6\delta |G|\ln|G|$ calls to Primitive \ref{prim:viol}, and
1099: expected number of at most $6\delta\ln |G|$ calls to \keyw{Trivial},
1100: with sets of size $6\delta^2$.
1101: \end{lem}
1102:
1103: \subsection{Combining the Algorithms}
1104: \begin{thm}\label{thm:runtime}
1105: Using a combination of the above two algorithms, a basis of $H$ in a violator
1106: space $(H,\vi)$ can be found calling Primitive \ref{prim:viol} expected
1107: \[
1108: O\left(\delta n + \delta^{O(\delta)}\right)
1109: \]
1110: many times.
1111: \end{thm}
1112:
1113: \pf Using the above bound for the trivial algorithm, \keyw{Basis2}
1114: can be implemented to require an expected number of at most
1115: \[O\left(\delta\log |G|(|G| + \delta^{O(\delta)})\right)\]
1116: calls to the primitive. Applying this as a subroutine in
1117: \keyw{Basis1}$(H)$ with $|H|=n$, $|G|$ is bounded by
1118: $3\delta\sqrt{n}$, and we get an overall expected complexity of
1119: \[
1120: O\left(\delta n + \delta^2 (\log n(\delta \sqrt{n} +
1121: \delta^{O(\delta)}))\right)
1122: \]
1123: in terms of the number of calls to Primitive \ref{prim:viol}. The
1124: terms $\delta^2\log n~ \delta\sqrt{n}$ and $\delta^2\log n~ \delta^{O(\delta)}$
1125: are asymptotically
1126: dominated by either $\delta n$ or $\delta^{O(\delta)}$, and we
1127: get the simplified bound of
1128: $O\left(\delta n + \delta^{O(\delta)}\right)$.
1129: \qed
1130:
1131: \section{Grid USO as Models for Violator Spaces}\label{grid_uso}
1132: We show in this section that the problem of finding the sink
1133: in a $\delta$-dimensional \emph{grid unique sink orientation}
1134: \cite{grid_uso} can be reduced to the problem of finding the (unique)
1135: basis of a violator space of combinatorial dimension $\delta$.
1136:
1137: Unique sink orientations of grids arise from various problems,
1138: including linear programming over products of simplices and generalized
1139: linear complementarity problems (GLCP) over $\P$-matrices \cite{grid_uso}.
1140: The GLCP has been introduced by Cottle and Dantzig \cite{cd-glcp-70}
1141: as a generalization of the well known LCP \cite{lcp}.
1142: There are also applications in game theory; for instance
1143: \cite{sweden1,sweden2,ssg_pglcp} show how parity, mean-payoff,
1144: and simple stochastic games are related to grid USO.
1145:
1146:
1147: \subsection{Grid USO}
1148: Fix a partition
1149: \[
1150: \Pi=(\Pi_1,\ldots,\Pi_{\delta})
1151: \]
1152: of the set $H:=\{1,\ldots,n\}$ into $\delta$ nonempty subsets,
1153: where we refer to $\Pi_i$ as the \emph{block $i$}. A subset
1154: $J\subseteq H$ is called a \emph{vertex} if $|J\cap \Pi_i|=1$
1155: for all $i$. The vertices naturally correspond to the Cartesian
1156: product of the $\Pi_i$. Let $\V$ be the set of all vertices.
1157:
1158: In the following definition, we introduce the \emph{grid spanned}
1159: by subsets $\Pi'_i$ whose union is $G\subseteq H$. The vertex set
1160: of this grid contains all vertices $J\subseteq G$ ($J\in \V$),
1161: with two vertices
1162: being adjacent whenever they differ in exactly two elements.
1163:
1164: \begin{defn}
1165: The $\delta$-dimensional \emph{grid} spanned by $G \subseteq H$
1166: is the undirected graph $\G(G)=(\V(G),\E(G))$, with
1167: \[
1168: \V(G):=\{J\in \V\colon J\subseteq G\},\quad
1169: \E(G):=\{\{J,J'\}\subseteq\V(G)\colon |J\oplus J'|=2\}.
1170: \]
1171: Here, $\oplus$ is the symmetric difference of sets.
1172: \end{defn}
1173:
1174: $\V(G)$ is in one-to-one correspondence with the Cartesian product
1175: \[
1176: \prod_{i=1}^{\delta}G_i, \quad G_i:=G\cap \Pi_i,
1177: \]
1178: and the edges in $\E(G)$ connect vertices in $\V(G)$ whose
1179: corresponding tuples differ in exactly one coordinate.
1180: See Figure \ref{fig:grid} left for an example of a grid.
1181:
1182: \figeps{grid}{13}{A 3-dimensional grid $\G(H)$ with $H=\{1,\ldots,7\}$
1183: where $\Pi=(\{1,2,3\},\{4,5\},\{6,7\})$ and a USO of it.}
1184:
1185: Note that $\G(G)$ is the empty graph whenever
1186: $G_i=G\cap \Pi_i=\emptyset$ for some $i$. We say that such a
1187: $G$ is \emph{not $\Pi$-valid}, and it is \emph{$\Pi$-valid}
1188: otherwise.
1189:
1190: A \emph{subgrid} of $\G(G)$ is any graph of the form $\G(G')$,
1191: for $G'\subseteq G$.
1192:
1193: \begin{defn}
1194: An orientation $\psi$ of the graph $\G:=\G(H)$ is called a
1195: \emph{unique sink orientation} (USO) if all nonempty subgrids of
1196: $\G$ have unique sinks w.r.t.\ $\psi$.
1197: \end{defn}
1198:
1199: We are interested in finding the sink in a USO of $\G$ as fast as possible,
1200: since the sink corresponds to the solution of the underlying problem
1201: (the $\P$-matrix GLCP, for example). Our measure of complexity
1202: will be the expected number of \emph{edge evaluations}, see \cite{grid_uso}.
1203: An edge evaluation returns the orientation of the considered edge and
1204: can typically be implemented to run in polynomial time (depending on
1205: the underlying problem). In the remainder of this paper, we derive
1206: the following theorem.
1207:
1208: \begin{thm}\label{thm:uso_algo}
1209: The sink of a unique sink grid orientation can be found by evaluating
1210: expected $O\left(\delta n + \delta^{O(\delta)}\right)$ edges.
1211: \end{thm}
1212:
1213: Note that a USO $\psi$ can be cyclic (see the thick edges in
1214: Figure \ref{fig:grid} right).
1215: If $\psi$ induces the directed edge $(J,J')$, we also write
1216: $J\stackrel{\psi}{\rightarrow}J'$. Any USO can be specified by associating
1217: each vertex $J$ with its outgoing edges. Given $J$ and $j\in H\setminus J$,
1218: we define $J\rhd j$ to be the unique vertex $J'\subseteq J\cup\{j\}$
1219: that is different from $J$, and we call $J'$ the \emph{neighbor} of
1220: $J$ \emph{in direction} $j$. Note that $J$ is a neighbor of $J'$ in
1221: some direction different from $j$.
1222:
1223: \begin{defn}
1224: Given an orientation $\psi$ of $\G$, the
1225: function $s_{\psi}:\V\rightarrow 2^{H}$, defined by
1226: \begin{equation}
1227: s_{\psi}(J) := \{j\in H\setminus J\colon J\stackrel{\psi}{\rightarrow} J\rhd j\},
1228: \end{equation}
1229: is called the \emph{outmap} of $\psi$.
1230: \end{defn}
1231: By this definition, any sink w.r.t.\ $\psi$ has empty outmap value.
1232:
1233: \subsection{Reduction to Violator Spaces}
1234: Let us fix a unique sink orientation $\psi$ of $\G$.
1235: Given a $\Pi$-valid subset $G\subseteq H$, we define $\sink(G)\in
1236: \V(G)$ to be the unique sink vertex in $\G(G)$. For a subset $G$ that
1237: is not $\Pi$-valid, let
1238: \[
1239: \bar{G}:=\bigcup_{i\colon\, G_i=\emptyset}\Pi_i.
1240: \]
1241: Thus $\bar{G}$ is the set of elements occurring in blocks of $\Pi$
1242: disjoint from $G$.
1243:
1244: \begin{defn}\label{def:gridtoviol}
1245: For $G\subseteq H$, define
1246: \[
1247: \vi(G)=
1248: \left\{
1249: \begin{array}{ll}
1250: s_{\psi}(\sink(G)), &\quad\textit{if $G$ is $\Pi$-valid}\\
1251: \bar{G}, &\quad\textit{if $G$ is not $\Pi$-valid.}
1252: \end{array}
1253: \right.
1254: \]
1255: \end{defn}
1256:
1257: \begin{thm}
1258: The pair $(H,\vi)$ from Definition \ref{def:gridtoviol}
1259: is a violator space of combinatorial dimension $\delta$.
1260: Moreover, for all $\Pi$-valid $G\subseteq H$, the unique sink of the subgrid
1261: $\mathcal{G}(G)$ corresponds to the unique basis of $G$ in $(H,\vi)$.
1262: \end{thm}
1263:
1264: \pf For every $G \subseteq H$, consistency holds by
1265: definition of $\sink(G), s_{\psi}(J)$ and $\bar{G}$.
1266: In order to prove locality for $F\subseteq G\subseteq H$,
1267: we look at three different cases.
1268:
1269: \paragraph*{$\mathbf{G}$ is not $\mathbf{\Pi}$-valid.} Then, $F\subseteq G$
1270: is not $\Pi$-valid either. The condition
1271: $\emptyset=G\cap \vi(F)=G\cap\bar{F}$ means that
1272: $F$ is disjoint from the same blocks as $G$. This
1273: implies $\bar{G}=\bar{F}$, hence $\vi(G)=\vi(F)$.
1274:
1275: \paragraph*{$\mathbf{G}$ and $\mathbf{F}$ are both
1276: $\mathbf{\Pi}$-valid.} Then $\G(F)$ is a nonempty subgrid of $\G(G)$, and
1277: $G \cap \vi(F)=\emptyset$ means that the sink of $\G(F)$ has no outgoing
1278: edges into $\G(G)$. Thus the unique sink of $\G(F)$ is also a sink of
1279: $\G(G)$ and therefore the unique one. This means that
1280: $\sink(G)=\sink(F)$, from which $\vi(G)=\vi(F)$ follows.
1281:
1282: \paragraph*{$\mathbf{G}$ is $\mathbf{\Pi}$-valid,
1283: $\mathbf{F}$ is not $\mathbf{\Pi}$-valid.} Then the condition
1284: $G \cap \vi(F)=\emptyset$ can never be satisfied since
1285: $\vi(F)=\bar{F}$ contains at least one full block $\Pi_i$,
1286: and $G_i=G\cap\Pi_i\neq \emptyset$.
1287:
1288: Next we prove that a largest basis in $(H,\vi)$ has at most $\delta$
1289: elements. For this, let $G\subseteq H$ be a set of size larger than
1290: $\delta$. If $G$ is $\Pi$-valid, we have
1291: \[\vi(G):=s_{\psi}(\sink(G))=s_{\psi}(\sink(\sink(G)))=:\vi(\sink(G)),\]
1292: since $J=\sink(J)$ for any vertex $J$. This means that $G$ has a subset
1293: of size $\delta$ with the same violators, so $G$ is not a basis.
1294:
1295: If $G$ is not $\Pi$-valid, we consider some subset $B$ that contains
1296: exactly one element from every block intersected by $G$. By definition,
1297: we have $\bar{G}=\bar{B}$ and $\vi(G)=\vi(B)$. Since $B$ has less than
1298: $\delta$ elements, $G$ cannot be a basis in this case, either.
1299:
1300: It remains to prove that for $G$ being $\Pi$-valid, the vertex
1301: $\sink(G)$ is the unique basis of $G$ in $(H,\vi)$. We have
1302: already shown that $\vi(G)=\vi(\sink(G))$ must hold in this
1303: case. Moreover, $\vi(\sink(G))$ contains no full block $\Pi_i$.
1304: On the other hand, any proper subset $F$ of $\sink(G)$ is not
1305: $\Pi$-valid, so its violator set \emph{does} contain at least
1306: one full block. It follows that $V(F)\neq V(\sink(G))$, so
1307: $\sink(G)$ is a basis of $G$. The argument is complete when
1308: we can prove that no other vertex $J\subseteq G$ is a basis of
1309: $G$. Indeed, such a vertex $J$ is not a sink in $\G(G)$, meaning
1310: that $G\cap \vi(J)\neq \emptyset$. This implies $\vi(J)\neq \vi(G)$.
1311: \qed
1312:
1313: Note that the global sink of the grid USO corresponds to the unique
1314: $\delta$-element (and $\Pi$-valid) set $B$ with $\vi(B)=\emptyset$.
1315: This is exactly the set output by the call \texttt{Basis1}$(H)$ of
1316: Clarkson's algorithms, when we apply it to the violator space
1317: constructed in Definition \ref{def:gridtoviol}.
1318:
1319: Primitive \ref{prim:viol} corresponds to one edge evaluation in the USO
1320: setting. With Theorem \ref{thm:runtime}, we therefore have proved
1321: Theorem \ref{thm:uso_algo}. For small $\delta$, the running time given
1322: in the theorem is faster than the one from the \emph{Product Algorithm}
1323: \cite{grid_uso} which needs expected $O(\delta!n+H_n^{\delta})$ edge
1324: evaluations, where $H_n$ is the $n$-th harmonic number.
1325:
1326: \section{Conclusions}
1327: We introduced violator spaces as a new framework for optimization
1328: problems and showed that acyclic violator spaces are
1329: equivalent to abstract and concrete LP-type problems.
1330: It turned out that the explicit ordering inherent to
1331: LP-type problems is not necessary in order to capture
1332: the structure of the underlying optimization
1333: problem. Violator spaces are more general than LP-type problems,
1334: yet Clarkson's algorithms still work on them.
1335:
1336: The Sharir-Welzl algorithm is also applicable for violator spaces
1337: in a straightforward way. However, the most obvious translation
1338: of this algorithm to the setting of violator spaces
1339: is not even guaranteed to finish, since for a general violator space
1340: it may run in a cycle and the subexponential analysis thus breaks down.
1341:
1342: We have seen that unique sink orientations are models for possibly cyclic
1343: violator spaces, and with Clarkson's algorithms we therefore
1344: have a fast scheme to solve fixed dimensional USO problems like
1345: the generalized linear complementarity problem with a $\P$-matrix.
1346: The GLCP with a $\P$-matrix has in general a cyclic structure and
1347: therefore gives rise to a cyclic USO. A violator space obtained from
1348: a cyclic USO is again cyclic. It is interesting that there are no cycles
1349: in a 2-dimensional grid USO \cite{grid_uso}. Whether the same is true
1350: for violator spaces of combinatorial dimension 2 is an open question.
1351:
1352: \section*{Acknowledgment}
1353: We thank an anonymous referee for useful comments.
1354: The second author would like to thank Nina Amenta for discussions
1355: concerning LP-type problems, possibly already forgotten by her as
1356: they took place many years ago, but nevertheless helpful for reaching
1357: the results in this paper.
1358:
1359: \begin{thebibliography}{10}
1360: \expandafter\ifx\csname url\endcsname\relax
1361: \def\url#1{\texttt{#1}}\fi
1362: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
1363:
1364: \bibitem{sw-cblpr-92}
1365: M.~Sharir, E.~Welzl, A combinatorial bound for linear programming and related
1366: problems, in: Proc. 9th Symposium on Theoretical Aspects of Computer Science
1367: (STACS), Vol. 577 of Lecture Notes in Computer Science, Springer-Verlag,
1368: 1992, pp. 569--579.
1369:
1370: \bibitem{msw-sblp-92}
1371: J.~Matou\v{s}ek, M.~Sharir, E.~Welzl, A subexponential bound for linear
1372: programming, Algorithmica 16 (1996) 498--516.
1373:
1374: \bibitem{k-srsa-92}
1375: G.~Kalai, A subexponential randomized simplex algorithm, in: Proc. 24th Annual
1376: ACM Symposium on Theory of Computing (STOC), 1992, pp. 475--482.
1377:
1378: \bibitem{a-bbhdn-94}
1379: N.~Amenta, Bounded boxes, {H}ausdorff distance, and a new proof of an
1380: interesting {H}elly-type theorem, in: Proc. 10th Annual Symposium on
1381: Computational Geometry (SCG), ACM Press, 1994, pp. 340--347.
1382:
1383: \bibitem{a-httgl-94}
1384: N.~Amenta, {H}elly theorems and generalized linear programming, Discrete and
1385: Computational Geometry 12 (1994) 241--261.
1386:
1387: \bibitem{bsv-dsapg-01}
1388: H.~Bj{\"o}rklund, S.~Sandberg, S.~Vorobyov, A discrete subexponential algorithm
1389: for parity games, in: Proc. 20th Annual Symposium on Theoretical Aspects of
1390: Computer Science (STACS), Springer-Verlag, 2003, pp. 663--674.
1391:
1392: \bibitem{halman}
1393: N.~Halman, Discrete and lexicographic {H}elly theorems and their relations to
1394: {LP}-type problems, Ph.D. thesis, Tel-Aviv University (2004).
1395:
1396: \bibitem{c-lvali-95}
1397: K.~L. Clarkson, {L}as {V}egas algorithms for linear and integer programming,
1398: Journal of the ACM 42 (1995) 488--499.
1399:
1400: \bibitem{gw-lpraf-96}
1401: B.~G{\"a}rtner, E.~Welzl, Linear programming - randomization and abstract
1402: frameworks, in: Proc. 13th Annual Symposium on Theoretical Aspects of
1403: Computer Science (STACS), Springer-Verlag, London, UK, 1996, pp. 669--687.
1404:
1405: \bibitem{cm-ltdao-96}
1406: B.~Chazelle, J.~Matou\v{s}ek, On linear-time deterministic algorithms for
1407: optimization problems in fixed dimension, Journal of Algorithms 21 (1996)
1408: 579--597.
1409:
1410: \bibitem{m-gofvc-95}
1411: J.~Matou\v{s}ek, On geometric optimization with few violated constraints,
1412: Discrete and Computational Geometry 14 (1995) 365--384.
1413:
1414: \bibitem{GWSampl01}
1415: B.~G{\"a}rtner, E.~Welzl, A simple sampling lemma - analysis and applications
1416: in geometric optimization, Discrete and Computational Geometry 25~(4) (2001)
1417: 569--590.
1418:
1419: \bibitem{ChanTukey}
1420: T.~Chan, An optimal randomized algorithm for maximum {T}ukey depth, in: Proc.
1421: 15th ACM-SIAM Symposium on Discrete Algorithms (SODA), 2004, pp. 423--429.
1422:
1423: \bibitem{a-spiht-96}
1424: N.~Amenta, A short proof of an interesting {H}elly-type theorem, Discrete and
1425: Computational Geometry 15 (1996) 423--427.
1426:
1427: \bibitem{SW}
1428: T.~Szab\'o, E.~Welzl, Unique sink orientations of cubes, in: Proc. 42nd IEEE
1429: Symposium on Foundations of Computer Science (FOCS), 2000, pp. 547--555.
1430:
1431: \bibitem{grid_uso}
1432: B.~G{\"a}rtner, W.~D. Morris{, Jr.}, L.~R{\"u}st, Unique sink orientations of
1433: grids, in: Proc. 11th Conference on Integer Programming and Combinatorial
1434: Optimization (IPCO), Vol. 3509 of Lecture Notes in Computer Science,
1435: Springer-Verlag, 2005, pp. 210--224.
1436:
1437: \bibitem{M02}
1438: W.~D. Morris{, Jr.}, Randomized principal pivot algorithms for {P}-matrix
1439: linear complementarity problems, Mathematical Programming, Series A 92 (2002)
1440: 285--296.
1441:
1442: \bibitem{MatUSO}
1443: J.~Matou\v{s}ek, The number of unique sink orientations of the hypercube,
1444: Combinatorica, to appear (2006).
1445:
1446: \bibitem{MCube}
1447: W.~D. Morris{, Jr.}, Distinguishing cube orientations arising from linear
1448: programs, Manuscript (2002).
1449:
1450: \bibitem{Develin}
1451: M.~Develin, {LP}-orientations of cubes and crosspolytopes, Advances in Geometry
1452: 4 (2004) 459--468.
1453:
1454: \bibitem{MS}
1455: J.~Matou\v{s}ek, T.~Szab\'o, {R}andom {E}dge can be exponential on abstract
1456: cubes, in: Proc. 45th Annual IEEE Symposium on Foundations of Computer
1457: Science (FOCS), 2004, pp. 92--100.
1458:
1459: \bibitem{SchSz}
1460: I.~Schurr, T.~Szab{\'o}, Finding the sink takes some time, Discrete and
1461: Computational Geometry 31 (2004) 627--642.
1462:
1463: \bibitem{SS}
1464: I.~Schurr, T.~Szab\'o, Jumping doesn't help in abstract cubes, in: Proc. 11th
1465: Conference on Integer Programming and Combinatorial Optimization (IPCO), Vol.
1466: 3509 of Lecture Notes in Computer Science, Springer-Verlag, 2005, pp.
1467: 225--235.
1468:
1469: \bibitem{lptouso}
1470: B.~G{\"a}rtner, I.~Schurr, Linear programming and unique sink orientations, in:
1471: Proc. 17th Annual Symposium on Discrete Algorithms (SODA), 2006, pp.
1472: 749--757.
1473:
1474: \bibitem{sweden1}
1475: H.~Bj{\"o}rklund, S.~Vorobyov, Combinatorial structure and randomized
1476: subexponential algorithms for infinite games, Theoretical Computer Science
1477: (in press).
1478:
1479: \bibitem{sweden2}
1480: H.~Bj{\"o}rklund, S.~Sandberg, S.~Vorobyov, A combinatorial strongly
1481: subexponential strategy improvement algorithm for mean payoff games, in:
1482: Proc. 29th International Symposium on Mathematical Foundations of Computer
1483: Science (MFCS), Vol. 3153 of Lecture Notes in Computer Science,
1484: Springer-Verlag, 2004, pp. 673--685.
1485:
1486: \bibitem{ssg_pglcp}
1487: B.~G{\"a}rtner, L.~R{\"u}st, Simple stochastic games and {P}-matrix generalized
1488: linear complementarity problems, in: Proc. 15th International Symposium on
1489: Fundamentals of Computation Theory (FCT), Vol. 3623 of Lecture Notes in
1490: Computer Science, Springer-Verlag, 2005, pp. 209--220.
1491:
1492: \bibitem{Meg}
1493: N.~Megiddo, A note on the complexity of {P}-matrix {LCP} and computing an
1494: equilibrium, Tech. rep., IBM Almaden Research Center, San Jose (1988).
1495:
1496: \bibitem{Skovron}
1497: P.~{\v S}kovro{\v n}, Generalized linear programming, Master's thesis, Charles
1498: University, Prague, Faculty of Mathematics and Physics (2002).
1499: \newline\urlprefix\url{http://kam.mff.cuni.cz/\~{}xofon/diplomka/}
1500:
1501: \bibitem{cd-glcp-70}
1502: R.~W. Cottle, G.~B. Dantzig, A generalization of the linear complementarity
1503: problem, Journal on Combinatorial Theory 8 (1970) 79--90.
1504:
1505: \bibitem{lcp}
1506: R.~W. Cottle, J.~Pang, R.~E. Stone, The Linear Complementarity Problem,
1507: Academic Press, 1992.
1508:
1509: \end{thebibliography}
1510:
1511: \end{document}
1512:
1513: