1: \documentclass[12pt]{article}
2: \usepackage{amsmath,amssymb,amsthm}
3:
4: \newcommand{\R}{\mathbb{R}}
5: \newcommand{\cl}[1]{\mathrm{cl}(#1)}
6: \newcommand{\inte}[1]{\mathrm{int}(#1)}
7: \newcommand{\bd}[1]{\mathrm{bd}(#1)}
8: \newcommand{\vac}{V}
9: \newcommand{\wac}{W}
10: \newcommand{\sa}[1]{\mathit{SA}_{#1}}
11: \newcommand{\ball}{\blacksquare}
12: \newcommand{\sphere}{\square}
13: \newcommand{\spheredot}{\boxdot}
14: \renewcommand{\leq}{\leqslant}
15: \renewcommand{\geq}{\geqslant}
16: \renewcommand{\emptyset}{\varnothing}
17: \newcommand{\Aff}{\mathbb A}
18:
19: \newtheorem*{blanktheorem}{Theorem}
20: \newtheorem{theorem}{Theorem}
21: \newtheorem{lemma}{Lemma}
22: \newtheorem*{conjecture}{Conjecture}
23: \newtheorem*{fact}{Fact}
24: \title{Two- versus three-dimensional connectivity testing
25: of first-order queries \\ to semi-algebraic sets}
26: \author{Floris Geerts\thanks{Contact author. Address:
27: Helsinki Institute for Information Technology, Basic Research
28: Unit, PO Box 26 (Teollisuuskatu 23), FIN-00014, Finland. Phone:
29: +358 9 19144037. Fax: +358 9 191 44441. Email:
30: floris.geerts@cs.helsinki.fi},
31: Lieven Smits,
32: and Jan Van den Bussche\thanks{Limburgs Universitair Centrum,
33: Diepenbeek, Belgium.}}
34: \date{}
35:
36: \begin{document}
37: \maketitle
38:
39: \begin{abstract}
40: This paper addresses the question whether one can
41: determine the connectivity of a semi-algebraic set in three dimensions by
42: testing the connectivity of a finite number of two-dimensional ``samples'' of
43: the set, where these samples are defined by first-order queries.
44: The question is answered negatively for two classes of first-order queries:
45: cartesian-product-free, and positive one-pass.
46: \end{abstract}
47:
48: \section{Introduction}
49:
50: Semi-algebraic sets provide a useful model for spatial datasets \cite{cdbbook}.
51: First-order logic over the reals (FO) then provides
52: a basic query language for expressing queries about such spatial data.
53: The power of FO, however, is too limited. In particular, testing whether a
54: set in $\R^n$ is topologically connected is not expressible in FO
55: for $n \geq 2$ (for $n=1$ it is easily expressed).
56:
57: The obvious reaction to this limitation of FO is to enrich it with an explicit
58: operator for testing connectivity, as proposed by Giannella and Van Gucht
59: \cite{giann} and by Benedikt et al.~\cite{bgls}. This operator can be
60: applied not just to the dataset itself, but also to any set derived from
61: the original set by an FO query.
62:
63: The question now arises whether the connectivity of a set in $\R^n$ can
64: be tested by testing the connectivity of a finite number of
65: sets in $\R^{n-1}$, constructed from the original set by FO queries.
66: For $n=2$, the answer is clearly negative, because connectivity in $\R^1$ is
67: expressible in FO, and therefore a positive answer would imply that
68: also connectivity in $\R^2$ would be expressible in FO, which we know is not
69: true. It is intuitive to conjecture
70: that the answer is negative for all $n \geq 2$.
71:
72: While this conjecture in its generality remains open (and seems very hard to
73: prove), we have proven it for two fragments of FO\@. In the first fragment,
74: cartesian product is disallowed. In the second fragment, negation is
75: disallowed, and the query must be ``one pass'' in a sense that can be made
76: precise. Our treatment of
77: the second fragment is for $n=3$ only.
78:
79: \section{Preliminaries}
80:
81: \paragraph{Semi-algebraic sets}
82: A \emph{semi-algebraic
83: set in $\R^n$} is a finite union of sets definable by conditions of
84: the form $ f_1(\vec{x})=\cdots=f_k(\vec{x})=0$,
85: $g_1(\vec{x})>0$, \dots, $g_\ell(\vec{x})>0$, with
86: $\vec{x}=(x_1,\ldots,x_n)\in\R^n$, and where $f_1(\vec{x})$,
87: \dots, $f_k(\vec{x})$, $g_1(\vec{x})$, \dots,
88: $g_\ell(\vec{x})$ are multivariate polynomials in the variables
89: $x_1,\ldots, x_n$ with real coefficients.
90:
91: Semi-algebraic sets form a very robust class; for example, any set definable
92: by a formula with quantifiers
93: in first-order logic
94: over the reals is semi-algebraic (i.e., definable also without quantifiers;
95: this is the Tarski--Seidenberg principle \cite{bcr}).
96:
97: \paragraph{Relational algebra} To express first-order queries about
98: a set $S$ in $\R^n$, we use not the formalism of first-order logic,
99: but the equivalent formalism of relational algebra expressions
100: (RAEs). These are inductively defined as follows. The symbol $S$
101: is a RAE, of \emph{arity} $n$. Any constant semi-algebraic set in
102: $\R^k$, for any $k$, is a RAE of arity $k$. If $e_1$ and $e_2$
103: are REAs of arities $k_1$ and $k_2$ respectively, then the cartesian
104: product $(e_1\times e_2)$ is a RAE of arity $k_1+k_2$, and provided
105: that $k_1=k_2=k$, the union $(e_1\cup e_2)$, the intersection $(e_1 \cap e_2)$
106: and the difference
107: $(e_1-e_2)$ are RAEs of arity $k$. Finally, if $e$ is a RAE of
108: arity $k$, and $i_1,\ldots,i_p\in\{1,\ldots,k\}$, then the projection
109: $\pi_{i_1,\ldots,i_p}(e)$ is a RAE of arity $p$.
110:
111: When applied to a given set $A$ in $\R^n$, a RAE
112: $e$ of arity $k$ evaluates in the natural way to a set $e(A)$ in $\R^k$.
113: When $A$ is semi-algebraic, $e(A)$ is too,
114: by the Tarski--Seidenberg principle.
115:
116: \paragraph{Notation}
117: We will use the following notations.
118: \begin{itemize}
119: \item
120: The topological closure
121: of a set $A\subseteq\R^n$ is denoted by $\cl{A}$, its interior is
122: denoted by $\inte{A}$ and its boundary $\cl{A}-\inte{A}$ is
123: denoted by $\bd{A}$.
124: \item
125: The $n$-dimensional closed unit ball
126: centered around the origin is denoted by $\ball$;
127: the $n$-dimensional unit sphere centered around the origin by
128: $\sphere$; and
129: the union $\sphere \cup \{\vec{0}\}$ by $\spheredot$.
130: \item
131: The set of affine transformations from $\R^n$ to $\R^n$ (compositions
132: of a scaling and a translation) is denoted by $\Aff$.
133: \end{itemize}
134:
135: \section{Cartesian-product-free queries}
136:
137: A RAE is called \emph{cartesian-product-free} if it does not use cartesian
138: product. An example of such a RAE is
139: $$ \pi_{1,2}((S \cap \Gamma_1) \cup (\Gamma_2-S)) - \pi_{1,3}(S
140: \cup \Gamma_3)$$ where $\Gamma_1$, $\Gamma_2$ and $\Gamma_3$ can
141: be arbitrary semi-algebraic sets in $\R^3$ and $S$ is ternary (i.e., stands
142: for a set in $\R^3$).
143:
144: In this section, we prove that
145: the connectivity of a semi-algebraic set in $\R^3$ cannot be determined by
146: sampling it using a finite number of binary
147: cartesian-product-free RAEs.
148: \begin{theorem} \label{cartfree}
149: Let $S$ range over sets in $\R^3$.
150: For any finite collection $e_1$, \dots, $e_\ell$ of binary
151: cartesian-product-free RAEs over $S$, there exist two
152: semi-algebraic sets $A$ and $B$ in $\R^3$ such that
153: \begin{enumerate}
154: \item
155: $A$ is connected;
156: \item
157: $B$ is disconnected;
158: \item
159: $e_i(A) = e_i(B)$ for $i=1,\dots,\ell$.
160: \end{enumerate}
161: \end{theorem}
162:
163: Toward the proof, we start with the following observation.
164: \begin{lemma}\label{lem:findset}
165: Let $\Lambda_0$, $\Lambda_1$, \dots, $\Lambda_k$ be nonempty
166: semi-algebraic sets in $\R^3$, where $\Lambda_0$ is open. Then
167: there exists a partition $\{I,J\}$ of
168: $\{1,\ldots,k\}$ and an
169: open semi-algebraic set $\vac\subseteq\Lambda_0$ such that
170: \begin{itemize}
171: \item
172: $\vac\subseteq \Lambda_i$ for $i \in I$, and
173: \item
174: $\vac\cap\Lambda_j=\emptyset$ for $j \in J$.
175: \end{itemize}
176: \end{lemma}
177: \begin{proof}
178: By induction on $k$.
179: If $k=0$, set
180: $I=\{0\}$, $J=\emptyset$, and $\vac=\Lambda_0$.
181:
182: If $k>0$,
183: consider the set
184: $\{\Lambda_0,\Lambda_1,\ldots,\Lambda_{k-1}\}$. Then by the
185: induction hypothesis, there is a partition $\{I',J'\}$ of
186: $\{1,\ldots,k-1\}$ and an open set $\vac'\subseteq\Lambda_0$
187: satisfying the condition as stated in the lemma for $k-1$.
188: Since
189: $\vac'=(\vac'\setminus\Lambda_k)\cup(\vac'\cap\Lambda_k)$, since
190: $\dim{\vac'}=3$, and since $\dim(A\cup B)=\max\{\dim{A},\dim{B}\}$
191: for semi-algebraic sets $A$ and $B$, at least one of the following two cases
192: occurs:
193: \begin{enumerate}
194: \item $\dim(\vac'\setminus\Lambda_k)=3$, in which case we choose
195: $\vac$ an open subset of $\vac'\setminus\Lambda_k$, and set
196: $I=I'$ and $J=J'\cup\{k\}$.
197: \item
198: $\dim(\vac'\cap\Lambda_k)=3$, in which case we choose
199: $\vac$ an open subset of $\vac'\cap\Lambda_k$, and set
200: $I=I'\cup\{k\}$ and $J=J'$. \qedhere
201: \end{enumerate}
202: \end{proof}
203:
204: The following lemma is the crucial element in our proof of the theorem.
205: \begin{lemma}\label{lem:projsimple}
206: For a given open semi-algebraic set $U\subseteq\R^3$, and any
207: ternary cartesian-product-free RAE $e$, there exists an open set
208: $V\subseteq U$ such that $e$ is equivalent to an expression of one of the four
209: possible forms $$ \Gamma, \quad S, \quad S \cup \Gamma, \quad \Gamma - S $$
210: on all sets $S \subseteq V$, where $\Gamma$ denotes a constant set in $\R^3$.
211: Moreover, in the last form, $V$ is included in the interior of\/ $\Gamma$.
212: \end{lemma}
213: \begin{proof}
214: Since both the input $S$ to $e$ and the output of $e$ are ternary, and $e$ is
215: cartesian-product-free, $e$ must be projection-free as well.
216: By rewriting $(e_1\cap e_2)$ as $(e_1-(e_2-e_1))$ we can ignore the
217: intersection operator.
218: We now proceed by induction on the structure of $e$.
219: The base cases where $e$ is $S$ or $e$ is constant are already in the right
220: form.
221:
222: For the cases $e = (e_1\cup e_2)$ and $e=(e_1
223: - e_2)$, by induction we can find an open set $V_1 \subseteq U$
224: such that $e_1$ has one of the four possible forms within
225: $V_1$, and we can further find an open set $V_2 \subseteq V_1$ such that
226: $e_2$ has one of the four possible forms within $V_2$.
227: This means that we have to consider $2\times 4 \times 4$ possibilities
228: (actually less, as there are symmetries), shown
229: in Tables \ref{tabelu} and \ref{tabelv}.
230:
231: \begin{table}
232: \caption{Proof of Lemma~\protect\ref{lem:projsimple}, possibilities for $e_1
233: \cup e_2$.} \label{tabelu}
234: $$ \begin{array}{c|cc}
235: \cup & S & \Gamma_2 \\
236: \hline
237: S & S & S \cup \Gamma_2 \\
238: \Gamma_1 & S \cup \Gamma_1 & \Gamma_1 \cup \Gamma_2 \\
239: S \cup \Gamma_1 & S \cup \Gamma_1 & S \cup (\Gamma_1 \cup \Gamma_2) \\
240: \Gamma_1-S & \Gamma_1 & \{\Gamma_1\cup\Gamma_2, (\Gamma_1\cup\Gamma_2)-S\} \\
241: \hline
242: \hline
243: \cup & S \cup \Gamma_2 & \Gamma_2 - S \\
244: \hline
245: S & S \cup \Gamma_2 & \Gamma_2 \\
246: \Gamma_1 &
247: S \cup (\Gamma_1 \cup \Gamma_2) &
248: \{\Gamma_1\cup\Gamma_2, (\Gamma_1\cup\Gamma_2)-S\} \\
249: S \cup \Gamma_1 &
250: S \cup (\Gamma_1 \cup \Gamma_2) & S \cup (\Gamma_1 \cup \Gamma_2) \\
251: \Gamma_1-S &
252: S \cup (\Gamma_1 \cup \Gamma_2) & (\Gamma_1\cup\Gamma_2)-S
253: \end{array} $$
254: \end{table}
255:
256: \begin{table}
257: \caption{Proof of Lemma~\protect\ref{lem:projsimple}, possibilities for $e_1
258: - e_2$.} \label{tabelv}
259: $$ \begin{array}{c|cc}
260: - & S & \Gamma_2 \\
261: \hline
262: S & \emptyset & \{\emptyset,S\} \\
263: \Gamma_1 & \Gamma_1-S & \Gamma_1-\Gamma_2 \\
264: S \cup \Gamma_1 & \{\Gamma_1,\Gamma_1-S\} &
265: \{\Gamma_1-\Gamma_2,(\Gamma_1-\Gamma_2)\cup S\} \\
266: \Gamma_1 - S & \Gamma_1-S &
267: \{\Gamma_1-\Gamma_2, (\Gamma_1-\Gamma_2) - S\} \\
268: \hline
269: \hline
270: - & S \cup \Gamma_2 & \Gamma_2 - S \\
271: \hline
272: S & \emptyset & S \\
273: \Gamma_1 &
274: \{\Gamma_1-\Gamma_2, (\Gamma_1-\Gamma_2) - S\} &
275: \{\Gamma_1-\Gamma_2, (\Gamma_1-\Gamma_2) \cup S\} \\
276: S \cup \Gamma_1 &
277: \{\Gamma_1-\Gamma_2,(\Gamma_1-\Gamma_2)- S\} &
278: S \cup (\Gamma_1-\Gamma_2) \\
279: \Gamma_1 - S &
280: \{\Gamma_1-\Gamma_2, (\Gamma_1-\Gamma_2) - S\} &
281: \{\Gamma_1-\Gamma_2, (\Gamma_1-\Gamma_2) - S\}
282: \end{array} $$
283: \end{table}
284:
285: Take, for example,
286: $e=(\Gamma_1 - S) \cup \Gamma_2$.
287: By applying
288: Lemma~\ref{lem:findset} to $\Lambda_0 = V_2$ and $\Lambda_1 = \Gamma_1$,
289: we get a $V \subseteq V_2$ such that either $V \subseteq \Gamma_1$
290: or $V \cap \Gamma_1 = \emptyset$. In the latter case, $e$ is equivalent to
291: $\Gamma_1\cup\Gamma_2$ within $V$. In the former case, $e$ is equivalent to
292: $(\Gamma_1\cup\Gamma_2)-S$ within $V$,
293: and we can always shrink $V$ a bit so that it is
294: included in the interior
295: of $\Gamma_1\cup\Gamma_2$, in accordance with the statement of the lemma.
296: In both cases $e$ is in a desired form. We summarize this in the
297: corresponding entry in Table~\ref{tabelu}. All other entries in the tables
298: are proven similarly, or are trivial.
299: \end{proof}
300:
301: We are now ready for the
302:
303: \begin{proof}[Proof of Theorem~\ref{cartfree}]
304: A binary cartesian-product-free RAE $e$ over ternary $S$ can be viewed as an
305: expression built up, using the operators $\cup$ and $-$,
306: from binary constant sets and binary projections of
307: ternary cartesian-product-free RAEs. If $\pi_{i,j}(c)$ is such a projection
308: occurring in $e$, we call $c$ a \emph{component} of $e$.
309:
310: By a series of applications of Lemma~\ref{lem:projsimple}, we can get all
311: components of all the given binary expressions $e_1$, \dots, $e_\ell$ in one
312: of the four normal forms mentioned in the lemma. The first application starts
313: with $U=\R^3$, and every next application takes as $U$ the $V$ produced by
314: the previous application. Within the $V$ produced by the final application,
315: all components are in normal form.
316:
317: Choose $\tau\in \Aff$ such that $\tau(\ball)\subset V$, and consider the
318: sets $A = \tau(\sphere)$ (which is connected) and $B = \tau(\spheredot)$ (which is
319: disconnected). Now any binary projection $\pi_{i,j}$ of a component $c$ in
320: normal form yields the same result whether applied to $A$ or to $B$. Indeed,
321: if $c$ is of the form $\Gamma$, $S$, or $S \cup \Gamma$ this is clear; if $c$
322: is of the form $\Gamma - S$ then we recall that Lemma~\ref{lem:projsimple}
323: guarantees that $V$ is fully included in the interior of $\Gamma$, so
324: $\pi_{i,j}(\Gamma-S) = \pi_{i,j}(\Gamma)$.
325:
326: We can thus conclude that $e_i(A)=e_i(B)$ for $i=1,\dots,\ell$ as desired.
327: \end{proof}
328:
329: For simplicity of exposition, in this section, we have stated and proved
330: Theorem~\ref{cartfree} in three dimensions only.
331: However, the argument readily generalizes to prove for any $n>2$ that
332: the connectivity of a semi-algebraic set in $\R^n$ cannot be determined by
333: sampling it using a finite number of $n-1$-ary
334: cartesian-product-free RAEs.
335:
336: \begin{blanktheorem}
337: Let $n>2$, and let $S$ range over sets in $\R^n$.
338: For any finite collection $e_1$, \dots, $e_\ell$ of $n-1$-ary
339: cartesian-product-free RAEs over $S$, there exist two
340: semi-algebraic sets $A$ and $B$ in $\R^n$ such that
341: \begin{enumerate}
342: \item
343: $A$ is connected;
344: \item
345: $B$ is disconnected;
346: \item
347: $e_i(A) = e_i(B)$ for $i=1,\dots,\ell$.
348: \end{enumerate}
349: \end{blanktheorem}
350:
351: \section{Positive one-pass queries}
352:
353: A RAE is called \emph{positive one-pass} if it does not use the difference
354: operator, and mentions $S$ only once. An example is
355: $$ \pi_{3,5} \big ( \Lambda_1\cup (\Lambda_2 \cap (S \times \R^2)) \big) $$
356: where $S$ is ternary, and $\Lambda_1$ and $\Lambda_2$
357: are arbitrary semi-algebraic sets in $\R^5$.
358: As a matter of fact, this example is very representative, in view of the
359: following:
360: \begin{lemma} \label{normalform}
361: Every binary positive one-pass RAE can be written in the
362: form $$
363: \pi_{i_1,i_2} \big ( \Lambda_1\cup (\Lambda_2 \cap (S \times \R^k)) \big). $$
364: \end{lemma}
365: More generally, it can be verified by induction that every $p$-ary positive
366: one-pass RAE can be written in the form of the above lemma, with
367: $\pi_{i_1,i_2}$ replaced by $\pi_{i_1,\dots,i_p}$.
368:
369: In this section, we prove that
370: the connectivity of a semi-algebraic set in $\R^3$ cannot be determined by
371: sampling it using a finite number of binary
372: positive one-pass RAEs.
373: \begin{theorem} \label{posone}
374: Let $S$ range over sets in $\R^3$.
375: For any finite collection $e_1$, \dots, $e_\ell$ of binary
376: positive one-pass RAEs over $S$, there exist two
377: semi-algebraic sets $A$ and $B$ in $\R^3$ such that
378: \begin{enumerate}
379: \item
380: $A$ is connected;
381: \item
382: $B$ is disconnected;
383: \item
384: $e_i(A) = e_i(B)$ for $i=1,\dots,\ell$.
385: \end{enumerate}
386: \end{theorem}
387:
388: The following lemma essentially proves the theorem.
389:
390: \begin{lemma}\label{lem:projexpr}
391: For a given open semi-algebraic set $U\subseteq\R^3$, any
392: semi-algebraic sets $\Lambda_1$ and $\Lambda_2$ in $\R^{3+k}$, and
393: any $i_1,i_2\in\{1,2,3,\allowbreak\ldots,\allowbreak k+3\}$, we can
394: always find an open set $V\subseteq U$ such that for any
395: $\tau\in \Aff$ with $\tau(\ball)\subset V$,
396: $$
397: \pi_{i_1,i_2}\big(\Lambda_1\cup(\Lambda_2\cap
398: (\tau(\sphere)\times\R^k))\big)
399: =\pi_{i_1,i_2}\big(\Lambda_1\cup(\Lambda_2\cap
400: (\tau(\ball)\times\R^k))\big).
401: $$
402: \end{lemma}
403:
404: Assuming this lemma, we can give the
405:
406: \begin{proof}[Proof of Theorem~\ref{posone}]
407: By a series of applications of Lemma~\ref{lem:projexpr}, we obtain a $V$ such
408: that for any $\tau\in \Aff$ for which $\tau(\ball)\subset V$, we have
409: $e_i(\tau(\sphere))=e_i(\tau(\ball))$ for $i=1,\dots,\ell$.
410: Since every $e_i$ is positive
411: (does not use the difference operator), every $e_i$ is monotone with respect
412: to the subset order. Hence, $e_i(\tau(\sphere)) \subseteq
413: e_i(\tau(\spheredot)) \subseteq
414: e_i(\tau(\ball))$ and thus $e_i(\tau(\sphere)) = e_i(\tau(\spheredot))$.
415: Taking $A = \tau(\sphere)$ and
416: $B = \tau(\spheredot)$ thus proves the theorem.
417: \end{proof}
418:
419: To prove Lemma~\ref{lem:projexpr} we will use the
420: regular cell decomposition of semi-algebraic sets, whose definition we recall
421: next. A function $f:C\rightarrow\R$, where $C\subseteq\R^n$, is called
422: \emph{regular} if it is continuous and for each
423: $i\in\{1,\ldots,n\}$ either strictly increasing, strictly
424: decreasing, or constant in the $i^{\mathrm{th}}$ coordinate.
425: (Which of these three cases holds may depend on $i$.)
426: Also, we call $f$ semi-algebraic if its graph is semi-algebraic.
427:
428: We define a \emph{regular cell} by induction on the number of dimensions.
429: Regular cells in $\R$ are singletons $\{a\}$, or open intervals
430: $(a,b)$, $(-\infty,a)$, or $(a,+\infty)$.
431: Now assume that $C\subseteq\R^n$ is a regular cell, and
432: $f,g:C\rightarrow\R$ are regular semi-algebraic functions on $C$,
433: with $f(\vec{x})<g(\vec{x})$ for all $\vec{x}\in C$. Then the sets
434: $\{(\vec{x},f(\vec{x}))\mid \vec{x}\in C\}$ and $\{(\vec{x},r)\mid
435: \vec{x}\in C, f(\vec{x})< r < g(\vec{x})\}$ are regular cells in
436: $\R^{n+1}$. In the latter case, $f$ can be $-\infty$, and
437: $g$ can be $\infty$.
438:
439: A \emph{regular cell decomposition of $\R^n$} is a special kind of
440: partition of
441: $\R^n$ into a finite number of regular cells. This is also defined
442: by induction on $n$. A regular decomposition of $\R$ is just any finite
443: partition of $\R$ in regular cells. For $n>1$, a regular cell
444: decomposition of $\R^n$ is a finite partition
445: $\{\mathcal{S}_1,\allowbreak\dots,\allowbreak\mathcal{S}_k\}$ of
446: $\R^n$ in regular cells such that
447: $\{\pi(\mathcal{S}_1),\ldots,\pi(\mathcal{S}_k)\}$ is a regular
448: cell decomposition of $\R^{n-1}$. Here, $\pi : (x_1,\dots,x_n)
449: \mapsto (x_1,\dots,x_{n-1})$ is the natural projection of $\R^n$
450: onto $\R^{n-1}$.
451:
452: Let $A$ be a semi-algebraic set in $\R^n$. A regular cell
453: decomposition of $\R^n$ is said to be \emph{compatible} with $A$
454: if $A$ is a union of regular cells from this decomposition.
455:
456: \begin{fact}[\cite{dries}]\label{regdecomp}
457: For every semi-algebraic set $A$ in $\R^n$ there exists a regular
458: cell decomposition of $\R^n$ compatible with $A$.
459: \end{fact}
460:
461: Toward the proof of Lemma~\ref{lem:projexpr},
462: we start with the following observation.
463:
464: \begin{lemma}\label{lem:extreme}
465: Let $A\subseteq\R^3$ be a compact semi-algebraic set and let
466: $f:A\rightarrow\R$ be a regular function. Then
467: $\min_{A} f=\min_{\bd{A}} f$ and
468: $\max_A f=\max_{\bd{A}} f$.
469: If moreover $\bd{A}$ is
470: connected, then
471: $f(\bd A)$ equals the interval $[\min_A f, \allowbreak \max_A f]$.
472: \end{lemma}
473:
474: \begin{proof}
475: Since $A$ is closed, $\bd A \subseteq A$ and thus
476: $\min_A f \geq \min_{\bd A} f$. To show the reverse
477: inequality, we need to find for any point in $A
478: - \bd A$ another point in $\bd A$ with the same or lower $f$-value.
479: Take a point $(x,y,z) \in A - \bd A$, and shoot a straight ray out of that
480: point in any direction. Since $A$ is bounded, the ray will intersect $\bd
481: A$. Let us focus on the two rays orthogonal to the $xy$ plane.
482: If $f$ is strictly decreasing in $z$, shoot the ray in
483: increasing $z$ direction to obtain an intersection point with $\bd A$ with
484: lower $f$-value as desired.
485: If $f$ is strictly increasing, follow the converse
486: direction, and if $f$ is constant, any direction will do to find a point in
487: $\bd A$ with the same $f$-value.
488: The equality $\max_A f = \max_{\bd A} f$ is proven in the same way.
489:
490: Now assume $\bd{A}$ is connected. Choose $\vec{x}_{\max}\in\bd{A}$
491: with maximal $f$-value, and choose
492: $\vec{x}_{\min}\in\bd{A}$ with minimal $f$-value.
493: Since
494: for semi-algebraic sets, connectivity coincides with path
495: connectivity~\cite{bcr}, there is a continuous path
496: $\gamma:[0,1]\rightarrow\bd{A}$ such that
497: $\gamma(0)=\vec{x}_{\min}$ and $\gamma(1)=\vec{x}_{\max}$. Since
498: $f$ is continuous, so is the composition
499: $f\circ\gamma$. Since $[0,1]$ is closed and connected,
500: $f\circ\gamma([0,1])$ must be a closed and connected set in $\R$, and must
501: therefore equal the interval
502: $[\min_A f,\max_A f]$.
503: \end{proof}
504:
505: We are now ready to embark on the
506:
507: \begin{proof}[Proof of Lemma~\ref{lem:projexpr}]
508: First note that
509: $\pi_{i_1,i_2}\big(\Lambda_1\cup
510: (\Lambda_2\cap(S\times\R^k))\big)$
511: is equivalent to $\pi_{i_1,i_2}(\Lambda_1) \cup
512: \pi_{i_1,i_2}(\Lambda_2\cap (S\times\R^k))$. So we may focus on
513: expressions of the form
514: \begin{equation}
515: \pi_{i_1,i_2}(\Lambda\cap(S\times\R^k)).\label{expr}
516: \end{equation}
517: We only need to prove the inclusion
518: \begin{equation}\label{eq:toprove}
519: \pi_{i_1,i_2}(\Lambda\cap(\tau(\ball)\times\R^k))\subseteq
520: \pi_{i_1,i_2}(\Lambda\cap(\tau(\sphere)\times\R^k)),
521: \end{equation} the
522: other direction being trivial.
523: The proof consists of several cases depending on the indices $i_1,i_2$.
524:
525: \subsection*{Case 1: $i_1,i_2\in\{1,2,3\}$}
526: Expression~(\ref{expr}) is equivalent to
527: $\pi_{i_1,i_2} (E)$, where $E$ is
528: $\pi_{1,2,3}(\Lambda\cap(S\times\R^k))$.
529: Applying Lemma~\ref{lem:findset} to
530: $\Lambda_0=\R^3$ and
531: $\Lambda_1=\pi_{1,2,3}(\Lambda)$,
532: we get
533: an open set $\vac$ such that we are in one of the following two
534: cases.
535: \begin{enumerate}
536: \item $\vac\cap\Lambda_1=\emptyset$.
537:
538: Within $\vac$, expression $E$,
539: and hence also~(\ref{expr}),
540: reduces to the empty set, so the
541: inclusion~(\ref{eq:toprove}) to be proven trivially holds within $\vac$.
542: \item $\vac\subset\Lambda_1$.
543:
544: Within $\vac$, expression $E$ now reduces to $S$, so
545: expression~(\ref{expr}) reduces to $\pi_{i_1,i_2}(S)$. In
546: particular this holds for both $S=\tau(\ball)$ and $S=\tau(\sphere)$,
547: where $\tau\in \Aff$ such that $\tau(\ball)\subset\vac$. Since
548: $\pi_{i_1,i_2}\tau(\ball)=\pi_{i_1,i_2}\tau(\sphere)$, the
549: inclusion~(\ref{eq:toprove}) holds within $V$.
550: \end{enumerate}
551:
552: \subsection*{Case 2: $i_1\in\{1,2,3\}$, $i_2\notin\{1,2,3\}$}
553: Expression~(\ref{expr}) is now equivalent to
554: $\pi_{i_1,4} (E)$, where $E$ now is
555: $\pi_{1,2,3,i_2}(\Lambda\cap(S\times\R^k))$.
556: Applying Lemma~\ref{lem:findset} to
557: $\Lambda_0=\R^3$ and $\Lambda_1=\pi_{1,2,3}(\Lambda)$, we get
558: an open set $\vac^{(0)}$ such that we are in one of the
559: following cases.
560: \begin{enumerate}
561: \item $\vac^{(0)}\cap\Lambda_1=\emptyset$.
562:
563: Within $\vac^{(0)}$, expression $E$,
564: and hence also~(\ref{expr}), reduces to the empty set,
565: so the inclusion~(\ref{eq:toprove}) holds within
566: $\vac^{(0)}$.
567: \item $\vac^{(0)}\subset\Lambda_1$.
568:
569: Within $\vac^{(0)}$, expression $E$ now reduces to
570: $A\cap(S\times\R)$, with $A=\pi_{1,2,3,i_2}(\Lambda)$.
571: Consider a regular cell
572: decomposition of $\R^4$ compatible with $A$, and write the projection
573: of this decomposition onto $\R^3$ as $\{C_1,\dots,C_\ell\}$.
574: Applying Lemma~\ref{lem:findset} to
575: $\Lambda_0^{(1)}=\vac^{(0)}$, and $\Lambda_i^{(1)}=C_i\cap \vac^{(0)}$
576: for $i=1,\dots,\ell$, we
577: get an open set $\vac^{(1)}\subset\vac^{(0)}$ contained in a unique cell
578: $C_j$.
579: Due to our regular cell decomposition, in particular the parts based on $C_j$,
580: within $\vac^{(1)}$ the
581: expression $E = A\cap(S\times\R)$
582: can now be written as a union of sets of the
583: form
584: \begin{align*}
585: & E_1 = \{(x,y,z,v)\mid (x,y,z)\in S \land v=f(x,y,z)\} \\
586: \intertext{or}
587: & E_2 = \{(x,y,z,v)\mid (x,y,z)\in S \land f(x,y,z)<v<g(x,y,z)\},
588: \end{align*}
589: where $f$ and $g$ are regular functions.
590:
591: Assume that $i_1=3$, so that the inclusion~(\ref{eq:toprove}) to be proven
592: becomes $\pi_{3,4}(E(\tau(\ball))) \subseteq \pi_{3,4}(E(\tau(\sphere)))$.
593: The cases $i_1=1$ and $i_1=2$ are analogous.
594: Since the projection of a union is the union of the projections,
595: we can restrict attention to the cases $E=E_1$ and $E=E_2$.
596: \begin{enumerate}
597: \item $E = E_1$. \label{22a}
598:
599: Let $\tau\in \Aff$ such that $\tau(\ball)\subset\vac^{(1)}$.
600: Take an arbitrary element $(z_0,f(x_0,y_0,z_0)) \in \pi_{3,4}(E(\tau(\ball)))$.
601: Since $\{(x,y,z)\in \tau(\ball) \mid z=z_0\}$ is compact with connected
602: boundary, we can apply Lemma~\ref{lem:extreme} to obtain $(x_1,y_1,z_0) \in
603: \tau(\sphere)$ with $f(x_1,y_1,z_0)=f(x_0,y_0,z_0)$. Hence,
604: $(z_0,f(x_0,y_0,z_0)) \in \pi_{3,4}(E(\tau(\sphere)))$ as desired.
605: \item $E = E_2$. \label{22b}
606:
607: By continuity of $f$ and $g$, and because $f<g$,
608: there exists an open set $\vac^{(2)}\subset\vac^{(1)}$ within which
609: $f < C < g$ for some constant $C$.
610: Within $\vac^{(2)}$, we can then break up $E_2$
611: into three sets
612: \begin{align*}
613: B_1&=\{(x,y,z,v)\mid (x,y,z)\in S\land f(x,y,z)<v<C\} \\
614: B_2&=\{(x,y,z,v)\mid (x,y,z)\in S\land C<v<g(x,y,z)\} \\
615: B_3&=\{(x,y,z,v)\mid (x,y,z)\in S\land v=C\}
616: \end{align*}
617:
618: The set $B_3$ is an instance of case~(\ref{22a}).
619: We now show that
620: the set $B_1$ (and, analogously, $B_2$)
621: can be reduced to that case as well.
622: Indeed, within $\vac^{(2)}$,
623: $$
624: B_1 = \bigcup_{t\in(0,1)} \{(x,y,z,v)\mid (x,y,z)\in S\land
625: v=tf(x,y,z)+(1-t)C\}.
626: $$
627: We now observe that for any $t\in(0,1)$, the function
628: $tf + (1-t)C$ is regular, so case~(\ref{22a}) applies to
629: each $t$ individually.
630: \end{enumerate}
631: \end{enumerate}
632:
633: \subsection*{Case 3: $i_1,i_2\notin\{1,2,3\}$}
634:
635: Expression~(\ref{expr}) is now equivalent to
636: $\pi_{4,5} (E)$, where $E$ now is
637: $\pi_{1,2,3,i_1,i_2}(\Lambda\cap(S\times\R^k))$.
638: Applying, as always, Lemma~\ref{lem:findset} to
639: $\Lambda_0=\R^3$ and $\Lambda_1=\pi_{1,2,3}(\Lambda)$, we get
640: an open set $\vac^{(0)}$ such that either
641: $\vac^{(0)}\cap\Lambda_1=\emptyset$
642: or
643: $\vac^{(0)}\subset\Lambda_1$.
644:
645: If $\vac^{(0)}\cap\Lambda_1=\emptyset$,
646: within $\vac^{(0)}$, expression $E$,
647: and hence also~(\ref{expr}), reduces to the empty set,
648: so the inclusion~(\ref{eq:toprove}) holds within
649: $\vac^{(0)}$.
650:
651: So we can assume that
652: $\vac^{(0)}\subset\Lambda_1$.
653: Within $\vac^{(0)}$, expression $E$ now reduces to
654: $A\cap(S\times\R^2)$, with $A=\pi_{1,2,3,i_1,i_2}(\Lambda)$.
655: Consider a regular cell
656: decomposition of $\R^5$ compatible with $A$, and write the projection
657: of this decomposition onto $\R^3$ as $\{C_1,\dots,C_\ell\}$.
658: Applying Lemma~\ref{lem:findset} to
659: $\Lambda_0^{(1)}=\vac^{(0)}$, and $\Lambda_i^{(1)}=C_i\cap \vac^{(0)}$
660: for $i=1,\dots,\ell$, we
661: get an open set $\vac^{(1)}\subset\vac^{(0)}$ contained in a unique cell
662: $C_j$.
663: Due to our regular cell decomposition, in particular the parts based on $C_j$,
664: within $\vac^{(1)}$ the
665: expression $E = A\cap(S\times\R^2)$
666: can now be written as a union of sets of the
667: form
668: \begin{tabbing}
669: $E_1 = \{(x,y,z,u,v)\mid (x,y,z)\in S\land
670: u=f(x,y,z) \land v=g(x,y,z,u)\}$,
671: \\
672: $E_2 =
673: \{(x,y,z,u,v)\mid (x,y,z)\in S$\=${}\land u=f(x,y,z)$ \\
674: \>${}\land g_1(x,y,z,u)< v <g_2(x,y,z,u)\}$,
675: \\
676: $E_3 =
677: \{(x,y,z,u,v)\mid (x,y,z)\in S$\=${}\land f_1(x,y,z)<u<f_2(x,y,z)$ \\
678: \>${}\land v=g(x,y,z,u)\}$, or \\
679: $E_4 =
680: \{(x,y,z,u,v)\mid (x,y,z)\in S$\=${}\land f_1(x,y,z)<u<f_2(x,y,z)$ \\
681: \>${}\land g_1(x,y,z,u)<v<g_2(x,y,z,u)\}$,
682: \end{tabbing}
683: where $f$, $f_1$, $f_2$, $g$, $g_1$, and $g_2$ are regular functions.
684:
685: We need to prove $\pi_{4,5}(E(\tau(\ball))) \subseteq
686: \pi_{4,5}(E(\tau(\sphere)))$.
687: Since the projection of an union is the union of the projections,
688: we can restrict attention to the cases $E=E_1$, $E=E_2$, $E=E_3$, and $E=E_4$.
689: \begin{enumerate}
690: \item $E=E_1$. \label{31}
691: \begin{enumerate}
692: \item
693: If $f$ is constant, with value $u_0$, $\pi_{4,5}(E_1)$ reduces to
694: $$ \{(u_0,g(x,y,z,u_0)) \mid (x,y,z) \in S\} $$ which can be handled by the
695: same reasoning as in Case~2,~(\ref{22a}).
696: \item \label{31b}
697: Now assume $f$ is not constant in $x$; the cases $y$ and $z$ are
698: analogous. Look at the projection
699: $\pi_{2,3,4,5}(E_1)$:
700: \begin{equation*}
701: \{(y,z,u,v)\mid \exists x((x,y,z)\in S \land u=f(x,y,z)\land
702: v=g(x,y,z,u))\}.
703: \end{equation*}
704: This set can be written as
705: \begin{equation}
706: E'_1 = \{(y,z,u,v)\mid (y,z,u)\in h(S)\land v=k(y,z,u)\},
707: \end{equation}
708: where
709: $h:(x,y,z)\mapsto (y,z,f(x,y,z))$, and
710: $$k:(y,z,u)\mapsto g(h_x^{-1}(y,z,u),y,z,u), $$ where
711: $h_x^{-1}$ is the function defined by
712: $h(h_x^{-1}(y,z,u),y,z)=(y,z,u)$. This
713: inverse function exists; in fact, because $f$ is regular and non-constant in
714: $x$, $h$ is a homeomorphism within $\vac^{(1)}$.
715:
716: Within $\wac^{(0)}=h(\vac^{(1)})$, we
717: can find an open set $\wac^{(1)}\subset\wac^{(0)}$ such that $k$
718: is regular. Since $h$ is a homeomorphism, we also can
719: find an open set $\vac^{(2)}\subset\vac^{(1)}$ such that
720: $h(\vac^{(2)})\subset\wac^{(1)}$.
721: Since $\pi_{4,5}(E_1)$ reduces to $\pi_{3,4}(E'_1)$,
722: we can, within $\vac^{(2)}$,
723: now again reason analogously as in Case~2,~(\ref{22a}).
724: \end{enumerate}
725:
726: \item $E=E_2$. \label{32}
727:
728: By continuity of $f$, $g_1$ and $g_2$, and because $g_1 < g_2$,
729: we can find
730: an open set $\vac^{(2)}\subseteq \vac^{(1)}$ within which
731: $$g_1(x,y,z,f(x,y,z))< C <g_2(x,y,z,f(x,y,z))$$ for some constant $C$.
732: Now reason repeatedly as in case~(\ref{31}) on the following sets:
733: \begin{tabbing}
734: $B_1=\{(x,y,z,u,v)\mid (x,y,z)\in S\land u=f(x,y,z) \land v=g_1(x,y,z,u)\}$
735: \\
736: $B_2=\{(x,y,z,u,v)\mid (x,y,z)\in S\land u=f(x,y,z) \land v=g_2(x,y,z,u)\}$
737: \\
738: $B_3=\{(x,y,z,u,v)\mid (x,y,z)\in S\land u=f(x,y,z) \land v=C)$
739: \end{tabbing}
740: We thus obtain $\vac^{(3)} \subseteq \vac^{(2)}$ within which
741: $\pi_{4,5}(B_i(\tau(\ball))) \subseteq \pi_{4,5}(B_i(\tau(\sphere)))$ for
742: $i=1,2,3$.
743: We then break up $E_2$ in $B_3 \cup B_4 \cup B_5$, where
744: \begin{tabbing}
745: $B_4=\{(x,y,z,u,v)\mid (x,y,z)\in S\land u=f(x,y,z) \land
746: g_1(x,y,z,u)< v < C \}$
747: \\
748: $B_5=\{(x,y,z,u,v)\mid (x,y,z)\in S\land
749: u=f(x,y,z) \land C< v < g_2(x,y,z,u) \}$
750: \end{tabbing}
751:
752: It remains to treat $B_4$ and $B_5$, and we do this as follows,
753: similarly to what we did in Case~2,~(\ref{22b}).
754: Reasoning as in case~(\ref{31}), we have written $\pi_{4,5}(B_1)$ as
755: $\pi_{3,4}(\{(y,z,u,v)\mid (y,z,u)\in
756: h(S)\land u=k(y,z,u)\})$. We then can write $\pi_{4,5}(B_4)$ as
757: $$
758: \bigcup_{t\in(0,1)}\pi_{3,4}(\{(y,z,u,v)\mid
759: (y,z,u)\in h(S)\land v=tk(y,z,u)+(1-t)C\}).
760: $$
761: We now observe
762: that for any $t\in(0,1)$, the
763: function $t k + (1-t)C$ is regular, so we can reason analogously as in
764: Case~2,~(\ref{22a})
765: for each $t\in(0,1)$ individually. We treat $B_5$ in the same way, now using
766: the $h$ and $k$ from $B_2$.
767:
768: \item $E=E_3$. \label{33}
769:
770: We begin again by determining an open set
771: $\vac^{(2)}\subset\vac^{(1)}$ within which $f_1 < C < f_2$
772: for some constant $C$,
773: and break up $E_3$ in the following sets:
774: \begin{tabbing}
775: $B_1=\{(x,y,z,u,v)\mid (x,y,z)\in S\land f_1(x,y,z)<u<C\land v=g(x,y,z,u)\}$
776: \\
777: $B_2=\{(x,y,z,u,v)\mid (x,y,z)\in S\land C<u<f_2(x,y,z)\land
778: v=g(x,y,z,u)\}$
779: \\
780: $B_3=\{(x,y,z,u,v)\mid (x,y,z)\in S\land u=C
781: \land v=g(x,y,z,u)\}$
782: \end{tabbing}
783: On $B_3$ we can reason as in case~(\ref{31}) and obtain
784: an open set
785: $\vac^{(3)}\subset\vac^{(2)}$ within which $\pi_{4,5}(B_3(\tau(\ball)))
786: \subseteq \pi_{4,5}(B_3(\tau(\sphere)))$.
787:
788: We show how to treat $B_1$; the treatment of $B_2$ is analogous.
789: Within a certain open set $\vac$ to be determined, we are going to break up
790: $B_1$ in a special way in two overlapping parts of the following form:
791: \begin{tabbing}
792: $\displaystyle
793: B_{1,1}=\bigcup_{t\in(0,\delta)} \{(x,y,z,u,v)\mid
794: \begin{aligned}[t] (x,y,z)\in S
795: & \land u=f(x,y,z)+t \\
796: & \land v=g(x,y,z,u)\} \end{aligned}$
797: \\
798: $\displaystyle
799: B_{1,2}=\bigcup_{t\in(c_{\vac},C)} \{(x,y,z,u,v)\mid (x,y,z)\in
800: S\land u=t\land v=g(x,y,z,u)\}$
801: \end{tabbing}
802: for certain $\delta$ and $c_{\vac}$, which we are now going to
803: define.
804:
805: If $f$ is constant, then $\delta:=0$, and $c_V$ is the
806: constant value of $f$.
807:
808: So, suppose that $f$ is not constant in $x$; the cases $y$ and $z$ are
809: analogous. Then
810: $h_t:(x,y,z)\mapsto (y,z,f(x,y,z)+t)$ is a homeomorphism for
811: every $t$. Let
812: $$k: (y,z,u,t) \mapsto g((h_t)_x^{-1}(y,z,u),y,z,u),$$
813: where $(h_t)_x^{-1}$ is the function defined by
814: $h_t((h_t)_x^{-1}(y,z,u),y,z)=u$.
815:
816: We now want to find a $\delta$ such that $k_t : (y,z,u) \mapsto k(y,z,u,t)$
817: is regular for every
818: $t\in(0,\delta)$. Thereto, consider the (semi-algebraic) set
819: $$
820: D=\{(y,z,u,t)\mid (y,z,u)\in h_0(\vac^{(3)}) \land 0<t<1 \land
821: \frac{\partial k}{\partial y}(y,z,u,t)=0 \}
822: $$
823: Using a cell decomposition of $\R^4$ compatible with $D$, we can
824: find an open set $\wac^{(0)}\subseteq h_0(\vac^{(3)})$ and
825: a $\delta^{(0)}>0$ such that on $\wac^{(0)}\times(0,\delta^{(0)})$
826: either $\frac{\partial k}{\partial y}=0$,
827: i.e., $k$ is constant in $y$,
828: or $\frac{\partial k}{\partial y}\neq 0$, i.e., $k$ is strictly
829: monotone in $y$.
830: Proceeding similarly, we can find $\wac^{(2)} \subseteq \wac^{(1)} \subseteq
831: \wac^{(0)}$ and $0<\delta^{(2)}<\delta^{(1)}<\delta^{(0)}$ such that
832: $k$ is either constant or strictly monotone in $z$
833: on $\wac^{(1)}\times(0,\delta^{(1)})$, and
834: $k$ is either constant or strictly monotone in $u$
835: on $\wac^{(2)}\times(0,\delta^{(2)})$. Hence, within $\wac^{(2)}$,
836: $k_t$ is regular for every $t \in (0,\delta^{(2)})$.
837:
838: Next, choose an open set $\vac^{(4)}\subset\vac^{(3)}$ and
839: $0<\delta^{(3)}<\delta^{(2)}$ such that
840: $h_t(\vac^{(4)})\subset\wac^{(2)}$ for every $t\in(0,\delta^{(3)})$.
841: We then restrict $\vac^{(4)}$ even further to an open set
842: $\vac^{(5)}$, and simultaneously choose $\delta^{(4)}$
843: such that the following conditions are satisfied:
844: \begin{gather*}
845: C-\sup_{\vac^{(5)}}f>\delta^{(4)}>0
846: \\
847: \sup_{\vac^{(5)}}f - \inf_{\vac^{(5)}}f <
848: \min\{\delta^{(3)},\delta^{(4)}\}
849: \end{gather*}
850: It is now clear that, within $\vac:=\vac^{(5)}$,
851: we have $B_1 = B_{1,1} \cup B_{1,2}$ where
852: we put $c_{\vac}:=\sup_{\vac^{(5)}}f$ and
853: $\delta:=\min\{\delta^{(3)},\delta^{(4)}\}$.
854:
855: It remains to deal with $B_{1,1}$ and $B_{1,2}$, but this poses no longer any
856: problems:
857: \begin{description}
858: \item[$B_{1,1}$:]
859: By construction, $k(y,z,u,t)$ is regular for every $t\in(0,\delta)$.
860: This implies that we can work with the sets
861: $$
862: \{(y,z,u,v)\mid (y,z,u)\in h_t(S) \land v=k(y,z,u,t)\}
863: $$
864: as in case~(\ref{31}).
865: \item[$B_{1,2}$:]
866: Here, for every $t$, we are back in Case~2,~(\ref{22a}).
867: \end{description}
868: \item $E = E_4$.
869: We begin again by determining an open set
870: $\vac^{(2)}\subset\vac^{(1)}$ within which $f_1 < C < f_2$
871: for some constant $C$,
872: and break up $E_4$ in the following sets:
873: \begin{tabbing}
874: $B_1=\{(x,y,z,u,v)\mid (x,y,z)\in S$\=${} \land f_1(x,y,z)<u<C$ \\
875: \>${} \land g_1(x,y,z,u)<v<g_2(x,y,z,u)\}$
876: \\
877: $B_2=\{(x,y,z,u,v)\mid (x,y,z)\in S$\=${} \land C<u<f_2(x,y,z)$ \\
878: \>${} \land g_1(x,y,z,u)<v<g_2(x,y,z,u)\}$
879: \\
880: $B_3=\{(x,y,z,u,v)\mid (x,y,z)\in S$\=${} \land u=C$ \\
881: \>${} \land g_1(x,y,z,u)<v<g_2(x,y,z,u)\}$
882: \end{tabbing}
883: On $B_3$ we can reason as in case~(\ref{32}) and obtain
884: an open set
885: $\vac^{(3)}\subset\vac^{(2)}$ within which $\pi_{4,5}(B_3(\tau(\ball)))
886: \subseteq \pi_{4,5}(B_3(\tau(\sphere)))$.
887:
888: We show how to treat $B_1$; the treatment of $B_2$ is analogous.
889: By the same procedure as in case~(\ref{33}), but now working with two
890: functions $k_1$ and $k_2$ (one for $g_1$ and one for $g_2$), we break up
891: $B_1$ within a certain open set $\vac$:
892: \begin{align*}
893: & B_{1,1}=\bigcup_{t\in(0,\delta)} \{(x,y,z,u,v)\mid
894: \begin{aligned}[t] & (x,y,z)\in S \land u=f(x,y,z)+t \\
895: & \quad \land g_1(x,y,z,u)<v<g_2(x,y,z,u)\} \end{aligned}
896: \\
897: & B_{1,2}=\bigcup_{t\in(c_{\vac},C)} \{(x,y,z,u,v)\mid
898: \begin{aligned}[t] & (x,y,z)\in S \land u=t \\
899: & \quad \land g_1(x,y,z,u)<v<g_2(x,y,z,u)\} \end{aligned}
900: \end{align*}
901: We finally deal with $B_{1,1}$ and $B_{1,2}$ as follows:
902: \begin{description}
903: \item[$B_{1,1}$:]
904: By construction, $(k_1)_t$ and $(k_2)_t$ are regular for every
905: $t\in(0,\delta)$. Writing $\pi_{4,5}(B_{1,1})$ as
906: $$
907: \bigcup_{t \in (0,\delta)} \pi_{3,4}(\{(y,z,u,v) \mid
908: \begin{aligned}[t] & (y,z,u) \in h_t(S) \\
909: & \quad \land k_1(y,z,u,t) < v < k_2(y,z,u,t)\}) \end{aligned}
910: $$
911: we can therefore reason analogously as in Case~2,~(\ref{22b}) for every $t$
912: individually.
913: \item[$B_{1,2}$:]
914: Here, for every $t$ individually, we are straight back in Case~2,~(\ref{22b}).
915: \end{description}
916: \end{enumerate}
917: The proof of Lemma~\ref{lem:projexpr} is complete.
918: \end{proof}
919:
920: \section{Concluding remarks}
921:
922: We have treated the positive-one pass queries for three-dimensional datasets
923: only. Our proof uses only fairly
924: elementary mathematics. By using more heavy machinery, one can probably prove
925: our Theorem~\ref{posone} in general for $n$-dimensional datasets and $n-1$-ary
926: queries. Conceivably this generalisation can also be performed starting from
927: our own proof, but that will be exceedingly laborious.
928:
929: Extending our proof technique to larger classes of RAEs is not obvious to us.
930: For instance, when relaxing the one-pass restriction, it is not clear how to
931: find a good $\tau\in \Aff$ such that $\tau(\ball) \times \tau(\ball)$ is nicely
932: located. When negation is allowed, the normal form of Lemma~\ref{normalform}
933: becomes much more complex, with consequences for the case analysis.
934:
935: Ultimately, one can even go further than the problem posed in the
936: Introduction, and throw in
937: connectivity testing of \emph{parameterized} queries, which can then even be
938: nested \cite{bgls,giann}.
939:
940: \begin{thebibliography}{BGLS}
941:
942: \bibitem[BGLS]{bgls}
943: M.~Benedikt, M.~Grohe, L.~Libkin, and L.~Segoufin.
944: \newblock Reachability and connectivity queries in constraint databases.
945: \newblock {\em Journal of Computer and System Sciences}, 66(1):169-206, 2003.
946:
947: \bibitem[BCR]{bcr}
948: J.~Bochnak, M.~Coste, and M.-F. Roy.
949: \newblock {\em Real Algebraic Geometry}.
950: \newblock Springer, 1998.
951:
952: \bibitem[GVG]{giann}
953: C.~Giannella and D.~Van Gucht.
954: \newblock Adding a path connectedness operator to {FO}+poly (linear).
955: \newblock {\em Acta Informatica}, 38(9):621--648, 2002.
956:
957: \bibitem[CDB]{cdbbook}
958: G.M. Kuper, J.~Paredaens, and L.~Libkin, editors.
959: \newblock {\em Constraint Databases}.
960: \newblock Springer, 2000.
961:
962: \bibitem[VDD]{dries}
963: L.~{van den Dries}.
964: \newblock {\em Tame Topology and O-minimal Structures}.
965: \newblock Cambridge University Press, 1998.
966:
967: \end{thebibliography}
968: \end{document}
969: