1: \documentclass{article}
2: \textheight=230mm
3: \voffset=-15mm
4: \title{A sufficient condition for global invertibility of Lipschitz
5: mapping} \author {Sergey P. Tarasov \thanks{Supported in part by RFBR
6: grant 02-01-00716} \\ Computing center of RAS \\ Vavilova 40 \\ 117967
7: Moscow GSP-1, Russia.\\ {\tt e-mail: sergey@ccas.ru} } \date{ }
8: \newtheorem{theorem}{Theorem}
9: \newtheorem{definition}{Definition}
10: \newtheorem{lemma}{Lemma}
11: \newtheorem{rK}{Remark}
12: \newtheorem{statement}{Statement}
13: \newtheorem{proposition}{Proposition}
14: \newtheorem {pF}{Proof}
15: \newtheorem {corollary}{Corollary}
16: \renewcommand \thepF {}
17: \newenvironment {proof}{\begin{pF}\rm}{\end{pF}}
18: \newenvironment {remark}{\begin{rK}\rm}{\end{rK}}
19:
20: \let\ra\rightarrow
21: \let\eps\varepsilon
22: \def\RR{{\bf R}}
23: \def\B{{\cal B}}
24: \def\T{{\cal T}}
25: \def\F{{\cal F}}
26: \def\J{{\cal J}}
27: \def\K{{\cal K}}
28: \def\M{{\cal M}}
29: \def\N{{\cal N}}
30: \def\Q{{\cal Q}}
31: \def\P{{\cal P}}
32: \def\X{{\cal X}}
33: \def\BB{{\bf B}}
34:
35: \begin{document}
36: \maketitle
37:
38: \begin{abstract}
39: We show that S.Vavasis' sufficient
40: condition for
41: global invertibility of a polynomial mapping can be easily generalized
42: to the case of a general Lipschitz mapping.
43:
44: Keywords: Invertibility conditions, generalized Jacobian,
45: nonsmooth analysis.
46: \end{abstract}
47:
48: In applications, in particular in finite-element analysis it is useful
49: to have some sufficient conditions for testing invertibility or
50: injectivity of a mapping $f : \RR^n \ra \RR^n$.
51: Usually $f(\cdot)$ is defined on some simple region $I$, say cube or
52: simplex, in the reference domain. The image $f(I)$ is some region with
53: more complicated geometry (a grid cell) in the physical domain.
54:
55: In a paper \cite{Vavasis}
56: S.Vavasis proposed a sufficient
57: condition for global invertibility of a polynomial mapping in $\RR^n$.
58: Here we show that this condition can be more naturally formulated in
59: the framework of {\em nonsmooth analysis} (see, e.g. \cite{Clarke}),
60: and this enables to spread the results of \cite{Vavasis} and
61: translate them almost word-for-word into a seemingly
62: much more general setting (at least in the linguistic sense). Thus
63: this note may be regarded as a feedback to \cite{Vavasis}.
64:
65: We need several definitions. The standard
66: reference on nonsmooth analysis is \cite{Clarke}.
67: Our discussion is restricted to a
68: finite-dimensional case though an essential part of \cite{Clarke} is
69: devoted to the infinite-dimensional generalizations.
70:
71: Recall that nonsmooth analysis works with Lipschitz functions that
72: are almost everywhere differentiable.
73: Roughly speaking, the {\em generalized (or Clarke's)} gradient
74: $\partial F$ of a
75: Lipschitz function $f(\cdot): R^n \rightarrow R$ at the point $x \in R^n$
76: is defined as a convex hull of (almost) all converging sequences of the
77: gradients: $\partial f(x) \stackrel{def}{=} conv(\lim \nabla(f(x_i))$,
78: for $x_i \rightarrow x$ and $f(\cdot)$ is differentiable at points $x_i$
79: and the sequence $\nabla f(x_i)$ converges. It is essential that
80: at the points of smoothness of $f(\cdot)$ the generalized gradient
81: coincides with gradient, and for a convex function--- with its subgradient.
82:
83: Similarly
84: \footnote{Actually, Clarke's definitions of the generalized gradient
85: and the generalized Jacobian are somewhat less restrictive. We may
86: assume that the points of the sequences ${x_i}$ or $JF(x_i)$ involved
87: do not belong not only to nonsmooth point of the map but additionally
88: {\em they do not belong to an arbitrary set of Lebesque measure zero}.
89: Such definitions are technically more convenient.}, the {\em
90: generalized Jacobian} $\partial F$ of a Lipschitz mapping $F(\cdot):
91: R^n \rightarrow R^m$ at the point $x$ is a convex hull of all $m \times n$
92: matrices obtained as limits of sequences $JF(x_i)$
93: (its Jacobian matrix at the point of smoothness $x_i$),
94: where $x_i \rightarrow x$ and $F(\cdot)$ is differentiable at
95: $x_i$.
96:
97: The generalized Jacobian {\em has maximum rank at point $x_0$}
98: if each matrix from $\partial F(x_0)$ has maximum rank.
99:
100: Nonsmooth inverse function theorem \cite[Th.7.1.1]{Clarke}
101: states that Lipschitz mapping $F: R^n \rightarrow R^n$, whose
102: generalized Jacobian has maximum rank at point $x_0$ is locally
103: Lipschitz invertible.
104:
105: In \cite{Vavasis} S.Vavasis proposed a sufficient
106: condition for the nondegeneracy of a matrix family $\M$
107: consisting of
108: square $n \times n$ matrices. To present this
109: result in a slightly more general form we need some (more or
110: less standard) definitions.
111:
112: {\em A cone} $\K \subseteq \RR^n$ is a set with a property that for all
113: $x \in \K$ and for any $\lambda \geq 0, \; \lambda x \in \K$, i.e. $\K$
114: contains all intersecting rays through the origin.
115:
116: A cone $\K$ is {\em convex} if sum of any points of $\K$ belongs to
117: $\K$.
118:
119: A convex cone is {\em polyhedral} it can be represented in the form:
120: $\K=\{x \,|\, x= \sum_{i=1}^m{\lambda_i a_i}, \; \lambda_i \geq 0, \;
121: a_i \in \RR^n \}$.
122:
123: A convex cone $\K$ is {\em acute} if it contains no nontrivial
124: subspaces or equivalently, if no finite set of elements of $\K$ sums to
125: zero.
126:
127: If $A \subseteq \RR^N$ then $cone\; hull(A)$ is a union of all rays
128: through the origin intersecting $A$.
129:
130: Let formulate {\em Vavasis' condition}.
131: Denote by $R_i
132: \subseteq \RR^n$
133: the set of all $i$-th columns of the matrices
134: belonging to $\M$ and assume that all $R_i, \; i=1,\dots, n$ are
135: separated from the origin, i.e. for some $\delta>0 \; R_i \cap \{x \,
136: | \, \|x\| \leq \delta\} = \emptyset$ (here $\| \cdot \|$ denotes the
137: Euclidean norm). Set $K_i = cone \; hull(R_i),\; i=1,\dots,n$ and let
138: any cone out of the $2^{n-1}$ cones $K_1 \pm K_2 \pm \dots \pm
139: K_n$ be {\em acute}.
140:
141: By definition, $V$-{\em family} is any matrix family satisfying these
142: conditions. It follows from the above that all matrices in $V$-family
143: are nondegenerate.
144:
145: If {\em additionally}, any cone involved
146: $K_1 \pm K_2 \pm \dots \pm K_n$
147: is contained in a cone $K_a=\{x
148: \in \RR^n \,| \, ax \geq \varepsilon \| x \|\}$ for some {\em
149: certificate vector} $a\in \RR^n$ and $\eps > 0$
150: then such family is defined as {\em strict
151: $V$-family}. In particular, if all $K_i$ (or $R_i$), $i=1,\dots, n$ are
152: polyhedral then by Farkas lemma all $V$-families are strict. Moreover,
153: checking that some matrix family is $V$-family is reduced to solving
154: $2^{n-1}$ linear programs, and thus the overall test could be performed
155: in linear time with respect to the input in any fixed dimension.
156:
157: \begin{theorem} Let $F(\cdot)=(f^1(\cdot),\dots,f^n(\cdot)): U
158: \subseteq \RR^n \ra \RR^n$ be any Lipschitz mapping defined on the
159: convex reference domain $U$. If the set of the generalized Jacobians
160: $\J=\{J \in \partial F(u),\, u \in U\}$ forms a strict $V$-family then
161: $F(\cdot)$\ is globally invertible on $U$. \end{theorem}
162:
163: {\bf Proof.} Actually the demonstration is a direct
164: translation into the new setting of the original proof from
165: \cite{Vavasis}.
166:
167: At first, local invertibility of $F(\cdot)$ at any point $u \in U$ of
168: the reference domain follows from the nonsmooth inverse function
169: theorem as the generalized Jacobian $\partial F(u)$ has
170: maximum rank at any point by construction.
171:
172: Secondly, to show global invertibility it is enough to check
173: injectivity of the mapping. Take any different points $u, v \in U$
174: from the reference domain. By convexity of $U$ and by the nonsmooth
175: analog of the Lagrange formula \cite[Th.2.6.5]{Clarke} for almost
176: all pairs $u, v \in U$ the following equality holds:
177: $F(v)-F(u)=\int^1_0{JF(u + t(v-u))(v-u)dt}$.
178: Assume w.l.o.g. that the first
179: coordinate of the vector $v-u$ is nonnegative (otherwise, exchange $v$
180: and $u$). Set $K_i= cone\; hull(\partial f^i(x), \, x \in U)$. Now
181: assume for a moment that $v-u \geq 0$. Take the corresponding
182: certificate vector $a$ for the cone $K_1+\dots+K_n$. By
183: construction, for all $s \in U, \; (a, \partial f_i(s)) > \eps \delta,
184: \; i=1,\dots,n$. Hence, $(a,F(v)-F(u)) > \eps \delta
185: \|u-v\| > 0$ $ $ and the injectivity follows. In general case, we
186: take any certificate vector for the cone $K_1 \pm K_2 \pm \dots
187: \pm K_n$ {\em with the same sign pattern as the sign pattern of the
188: coordinates of the vector $v-u$}.
189: \medskip
190:
191: Informally, Vavasis sufficiency condition assumes that the columns of
192: the corresponding Jacobian matrices are independent and thus may be
193: extremely restrictive
194: but, on the other hand, it has attractive enough {\em decomposition
195: feature}. Namely, let the reference domain $I$ be subdivided into
196: several simple regions, say, $I$ is a cube that is partitioned into
197: parallelepipedal patches $I=\cup_{i=1}^M I^{i}$ by several axis-
198: parallel hyperplanes. Now let $F(\cdot)=(f^1(\cdot),\dots,f^n(\cdot)):
199: I \ra \RR^n$ be any Lipschitz mapping. Let $\Delta_i= \cup_{x \in
200: I_i}JF(x)$ (the union of all generalized Jacobians in the {\em closure}
201: of the patch $I_i$). Obviously these sets can be computed {\em
202: separately} for any {\em closed} patch $I_i \; i=1, \dots,
203: M$. As above we obtain the following sufficient condition: if each of
204: the sets $\Delta_i, \, i=1, \dots, M$ forms a strict $V$-family and
205: there exists a certificate vector {\em common to all} $\Delta_i$ then
206: the mapping $F(\cdot)$ is globally invertible. Equivalently, as
207: Vavasis condition is insensible to taking convex hulls, $F(\cdot)$ is
208: globally invertible on $I$ if the set $\Delta=conv(\cup_{i=1}^M
209: \Delta_i)$ forms a strict $V$-family.
210:
211: For a simple example, assume
212: that $F(\cdot)$ is a continuous piecewise polynomial (product)
213: Bernstein-Bezier (BB) mapping as proposed in
214: \cite{Vavasis}\footnote{The reader should take into account that the
215: example below is almost explicit in \cite{Vavasis} and could be easily
216: recovered from the arguments therein {\em but with some additional
217: formal arguments.}} , i.e. in any patch $I_q, \, q=1,\dots,
218: M$ the mapping is given by
219: $$F(\xi_1,\dots,\xi_n)=\sum_{i_1=0}^p\dots\sum_{i_n=0}^p f^q_{i_1,\dots
220: i_n} {p\choose i_1}\xi_1^{i_1}(1-\xi_1)^{p-i_1} \dots {p\choose
221: i_1}\xi_n^{i_n}(1-\xi_n)^{p-i_n},$$
222: where $(\xi_1,\dots,\xi_n)$ are
223: local coordinates in $I_q=\{0\leq \xi_1 \leq 1, \dots, 0\leq \xi_n \leq
224: 1\}$ and the vectors $f^q_{i_1,\dots i_n} \in \RR^n, \,
225: i_1,\dots,i_n=1,\dots,n$ are called the {\em control points}. It is
226: well known that the image $F(I_q)$ of BB mapping is contained in the
227: convex hull of the control points. It is also well known that the
228: partial derivatives of $F(\cdot)$ can be put into BB form (and the
229: resulting control points for the derivatives are effectively computable
230: linear combinations of the control points for $F(\cdot)$). Thus for any
231: $i=1,\dots,n$ the set of all $i$-th columns of the Jacobian matrices
232: $JF(I_q)=\{JF(x), \; x \in I_q\}$ is contained in some {\em explicitly
233: computable polytope} $P^q_i$. Hence, all possible $i$-th columns of the
234: Jacobian matrices $JF(x),\; x \in I$ are contained in
235: $P_i=conv\{P^1_i,\dots, P^M_i\}, \; i=1,\dots,n$. (Here we use the
236: continuity of $F(\cdot)$ as at any point belonging to the intersection
237: of some patches $x \in I_{i_1} \cap \dots \cap I_{i_k}$ the generalized
238: Jacobian $JF(x)$ is by construction contained in
239: $conv\{JF(I_{i_1},\dots,JF(I_{i_k})\}$.) Thus if the matrix family
240: $\P$, whose $i$-th column belongs to $P_i, \; i=1,\dots,n$, form a
241: strict $V$-family then $F(\cdot)$ is globally invertible on the whole
242: domain $I$.
243:
244: Here we show that invertibility test in \cite{Vavasis} is valid for a
245: much more ample class of mappings and thus should be
246: {\em simultaneously robust and restrictive} enough. Therefore it
247: would be nice if some new arguments would be applied to the
248: invertibility problem even in the simplest case of bilinear polynomials.
249:
250:
251: \begin{thebibliography}{9}
252:
253: \bibitem{Vavasis} S.Vavasis.
254: A Bernstein-Bezier sufficient condition for invertibility
255: of polynomial mapping function, http://www.cs.cornell.edu/~vavasis/
256:
257: \bibitem{Clarke} F.H.Clarke. Optimization and nonsmooth analysis, NY,
258: John Wiley \& Sons, 1983.
259:
260: \bibitem{Voronoi} M.Vyalyi, E.Gordeev, and S.Tarasov. The
261: stability of the Voronoi diagram. Comp.Maths Math.Phys., 1996, {\bf
262: 36}, 3, 405-414.
263:
264: \end{thebibliography}
265: \end{document}
266:
267: