1: % Linear precision for parametric patches
2: %
3: % Luis Garcia-Puente
4: % Frank Sottile
5: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
6: \documentclass[12pt]{amsart}
7: \usepackage{amssymb,amsmath}
8: \usepackage{graphicx}
9: \usepackage{colordvi}
10:
11: %%%%%%%%%%%%%%%%%%%%% Layouts %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
12: %
13: % US size paper
14: %
15: \headheight=8pt % \topmargin=40pt
16: %\textheight=611pt \textwidth=461pt
17: \textheight=621pt \textwidth=461pt
18: \oddsidemargin=6pt \evensidemargin=6pt
19: %%%%%%%%%%%%%%%%%%%%% Layouts %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
20: %
21: % A4 paper (not so big margins)
22: %
23: %\headheight=8pt \topmargin=-20pt
24: %\textheight=720pt \textwidth=460pt
25: %\oddsidemargin=-1pt \evensidemargin=-1pt
26: %
27: %%%%%%%%%%%%%%%%%%%%% Environments %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
28: \numberwithin{equation}{section}
29:
30: \newtheorem{theorem}[equation]{Theorem}
31: \newtheorem{lemma}[equation]{Lemma}
32: \newtheorem{proposition}[equation]{Proposition}
33: \newtheorem{corollary}[equation]{Corollary}
34: \newtheorem{conjecture}[equation]{Conjecture}
35:
36: \theoremstyle{definition}
37: \newtheorem{definition}[equation]{Definition}
38:
39: %\theoremstyle{example}
40: \newtheorem{example}[equation]{Example}
41:
42: %\theoremstyle{remark}
43: \newtheorem{remark}[equation]{Remark}
44: \newcounter{FNC}[page]
45: \def\fauxfootnote#1{{\addtocounter{FNC}{2}$^\fnsymbol{FNC}$%
46: \let\thefootnote\relax\footnotetext{$^\fnsymbol{FNC}$#1}}}
47: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
48: \hyphenation{pa-ram-e-trized}
49:
50: \newcommand{\rat}{\ {\relbar\rightarrow}\ }
51: \newcommand{\longrat}{{\;\relbar\relbar\rightarrow\;}}
52:
53: \newcommand{\zwei}[2]{\left[\!\begin{array}{r}\!#1\!\\\!#2\!\end{array}\!\right]}
54: \newcommand{\kzwei}[2]{\left[\substack{#1\\#2}\right]}
55: \newcommand{\smi}{\hspace{0.6em}}
56:
57: \newcommand{\ba}{{\bf a}}
58: \newcommand{\bb}{{\bf b}}
59: \newcommand{\bc}{{\bf c}}
60: \newcommand{\bd}{{\bf d}}
61: \newcommand{\be}{{\bf e}}
62: \newcommand{\bp}{{\bf p}}
63: \newcommand{\bk}{{\bf k}}
64: \newcommand{\bx}{{\bf x}}
65: \newcommand{\bv}{{\bf v}}
66:
67: \newcommand{\calA}{{\mathcal A}}
68: \newcommand{\calB}{{\mathcal B}}
69:
70: \newcommand{\C}{{\mathbb C}}
71: \renewcommand{\P}{{\mathbb P}}
72: \newcommand{\R}{{\mathbb R}}
73: \newcommand{\Z}{{\mathbb Z}}
74:
75: \newcommand{\QED}{\hfill\raisebox{-5pt}{\includegraphics[height=14pt]{figures/QED.eps}}\vspace{6pt}}
76: \renewcommand{\qed}{\includegraphics[height=14pt]{figures/QED.eps}}
77:
78: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
79: \begin{document}
80:
81: \title{Linear precision for parametric patches}
82:
83: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
84: \author{Luis David Garcia-Puente}
85: \address{Department of Mathematics and Statistics\\
86: Sam Houston State University\\
87: Huntsville\\
88: TX \ 77341\\
89: USA}
90: \email{lgarcia@shsu.edu}
91: \urladdr{http://www.shsu.edu/\~{}ldg005}
92: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
93: \author{Frank Sottile}
94: \address{Department of Mathematics\\
95: Texas A\&M University\\
96: College Station\\
97: TX \ 77843\\
98: USA}
99: \email{sottile@math.tamu.edu}
100: \urladdr{http://www.math.tamu.edu/\~{}sottile}
101: \thanks{Work of Sottile supported by NSF CAREER grant DMS-0538734,
102: by the Institute for Mathematics and its Applications with funds provided by the
103: National Science Foundation, and by Peter Gritzmann of the Technische
104: Universit\"at M\"unchen.}
105: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
106: \subjclass[2000]{65D17, 14M25}
107: %68U07 Computer aided design
108: %65D17 Computer aided design (modeling of curves and surfaces)
109: \keywords{tensor product B\'ezier surfaces; triangular B\'ezier surface patches;
110: Barycentric coordinates; Iterative proportional fitting}
111: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
112: \begin{abstract}
113: % We study linear precision for multi-sided parametric patches of any dimension,
114: % showing that every proper parametric patch has a unique reparametrization which
115: % has linear precision and giving a geometric criterion for when this
116: % reparametrization is by rational functions.
117: % For toric patches, this geometric criterion is equivalent to a certain toric
118: % differential defining a birational map.
119: % While the reparametrization of a general toric patch having linear precision is not
120: % necessarily a rational function, we show that it is computed by iterative proportional
121: % fitting, a numerical algorithm from statistics.
122: %
123: We give a precise mathematical formulation for the notions of a parametric patch and linear
124: precision, and establish their elementary properties.
125: We relate linear precision to the geometry of a particular linear projection, giving
126: necessary (and quite restrictive) conditions for a patch to possess linear precision.
127: A main focus is on linear precision for Krasauskas' toric patches, which we show is
128: equivalent to a certain rational map on $\C\P^d$ being a birational isomorphism.
129: Lastly, we establish the connection between linear precision for toric surface patches
130: and maximum likelihood degree for discrete exponential families in algebraic statistics,
131: and show how iterative proportional fitting may be used to compute toric patches.
132: %
133: \end{abstract}
134:
135: %
136: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
137: \maketitle
138: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
139:
140: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
141: \section*{Introduction}
142: B\'ezier curves and surfaces are the fundamental units in geometric modeling.
143: There are two basic shapes for surfaces---triangular B\'ezier patches
144: and rectangular tensor product patches.
145: Multi-sided patches are needed for some applications, and there are
146: several control point schemes for $C^\infty$ multi-sided patches.
147: These include the $S$-patches of Loop and DeRose~\cite{LdR89},
148: Warren's hexagon~\cite{Wa92}, Kar{\v{c}}iauskas's $M$-patches~\cite{Ka03},
149: and the multi-sided toric B\'ezier patches of
150: Krasauskas~\cite{KR02}.
151: (Relationships between these and other patches are discussed
152: in~\cite{KK00}.)
153: Parametric patches are general control point schemes for $C^\infty$
154: patches whose shape is a polygon or polytope.
155: They include the patch schemes just mentioned, as well as
156: barycentric coordinates for polygons and polytopes~\cite{Fl,Wa75,Wa96}.
157:
158: The success and widespread adoption of B\'ezier and tensor-product patches is
159: due in part to their possessing many useful mathematical properties.
160: Some, such as affine invariance and the convex hull property, are
161: built into their definitions and also hold for the more general
162: parametric patches.
163: Other properties, such as de Castlejau's algorithm
164: for computing B\'ezier patches, come from the specific form of their
165: Bernstein polynomial blending functions.
166: Linear precision is the ability of a parametric patch to replicate linear
167: functions.
168: When the blending functions of a parametric patch correspond to the
169: vertices of a polytope, these blending functions give barycentric coordinates precisely
170: when the patch has linear precision.
171: In this way, blending functions for a parametric patch having linear precision
172: are barycentric coordinates for general control point schemes.
173:
174: For us, a
175: %%%%%%%%%%%%%
176: (parametric) \Blue{{\it patch}} is a collection of non-negative blending
177: functions indexed by a finite set $\calA$ of points in $\R^d$,
178: where the common domain of the blending functions is the convex hull $\Delta$
179: of $\calA$.
180: A collection of control points in $\R^\ell$ indexed by $\calA$ is used to
181: define a map from $\Delta$ to $\R^\ell$.
182: The blending functions determine the internal structure of this map
183: and the basic shape, $\Delta$, of its image, while the control points
184: determine how the image lies in $\R^\ell$.
185: Choosing the control points to be the points of $\calA$ gives the
186: tautological map, and the patch has linear precision when this
187: tautological map is the identity on $\Delta$.
188: We show that every patch has a unique reparametrization (composing the
189: blending functions with a homeomorphism of $\Delta$) having linear
190: precision (Theorem~\ref{T:Lin_Prec}).
191: This generalizes Theorem~8.5 of~\cite{So03}, which
192: was for toric patches.
193:
194: The blending functions of a patch may be arbitrary
195: non-negative $C^\infty$ functions.
196: A patch is rational if it has a reparametrization having polynomial
197: blending functions.
198: For a rational patch, a map given by control points corresponds to a
199: linear projection of a projective algebraic variety associated to its
200: blending functions.
201: Its unique reparametrization having linear precision has rational blending
202: functions if and only if a certain canonical map defined on
203: this variety is a birational isomorphism (Theorem~\ref{Th:LP=birational}),
204: which implies that this variety has a maximally degenerate position with
205: respect to a canonical linear subspace given by the set $\calA$ (Theorem~\ref{Th:Necessary}).
206:
207:
208: We apply this analysis to Krasauskas's toric patches, which are rational.
209: The shape of a toric patch is a pair $(\calA,w)$ where $\calA$ is a set
210: of integer points in $\Z^d$ and $w$ is a collection of positive real numbers indexed by
211: $\calA$.
212: We show that iterative proportional fitting~\cite{DR72}, a simple numerical algorithm from
213: statistics, computes the blending functions that have linear precision.
214: This algorithm was suggested to us by Bernd Sturmfels.
215: It may form a basis for algorithms to manipulate these patches.
216:
217: Krasauskas~\cite[Problem 3]{Kr06} asked whether any toric patches (besides the classical B\'ezier
218: simploids~\cite{dRGHM}) admit a rational reparametrization having linear precision.
219: A toric patch of shape $(\calA,w)$ corresponds to a homogeneous polynomial $f=f_{\calA,w}$
220: whose dehomogenization is the sum of monomials with exponents $\calA$ and coefficients $w$.
221: The toric patch admits a rational reparametrization having linear precision if and only if the toric
222: differential
223: \[
224: D_{\mbox{\scriptsize\rm toric}}f\ :=\
225: \Big[x_0\tfrac{\partial}{\partial x_0}f\;\colon\;
226: x_1\tfrac{\partial}{\partial x_1}f\;\colon\;\dotsb
227: \;\colon\;x_d\tfrac{\partial}{\partial x_d}f\Big]\
228: \colon\ \C\P^d\ \longrat\ \C\P^d
229: \]
230: defines a birational isomorphism (Corollary~\ref{Cor:Reformulation}).
231: This analysis of linear precision is used in~\cite{BRS}
232: to classify which toric surface patches can have linear precision.
233:
234: In Krasauskas's question about linear precision, he allowed the points $\calA$ indexing
235: the blending functions of toric patches to move within their (fixed) convex hull,
236: keeping the same blending functions (See Example~\ref{Ex:Pentagon}).
237: Our analysis of linear precision for general parametric
238: patches will help to address that version of his question.
239: \bigskip
240:
241:
242: In Section 1, we define parametric patches and show
243: that every patch has a unique reparametrization that has linear
244: precision.
245: In Section 2 we show that if a
246: rational patch has linear precision, then an algebraic variety we
247: obtain from the blending functions has an exceptional
248: position with respect to a certain linear space.
249: We define toric patches in Section 3, and in Section 4 we
250: explain how iterative proportional fitting computes the blending functions for toric
251: patches that have linear precision.
252:
253:
254: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
255: %
256: %
257: %
258: \section{Linear precision for parametric patches}\label{S:PP}
259:
260: We interpret the standard definition of a mapping via control
261: points and blending functions (see for example \cite[\S 2]{KK00})
262: in a general form convenient for our discussion.
263: All functions here are smooth ($C^\infty$) where defined and
264: real-valued unless otherwise stated.
265: Let $\R_>$ be the set of strictly positive real numbers and $\R_\geq$ the
266: set of non-negative real numbers.
267:
268: Let $\calA$ be a finite set of points in $\R^d$, which we shall use as
269: geometrically meaningful indices.
270: A control point scheme for parametric patches, or
271: \Blue{({\it parametric}) {\it patch}} for short, is a collection
272: $\beta=\{\beta_\ba\mid\ba\in\calA\}$ of non-negative functions, called
273: \Blue{{\it blending functions}}.
274: The common domain of the blending functions is the convex hull $\Delta$ of
275: $\calA$, which we call the \Blue{{\it domain polytope}}.
276: We will always assume that $\Delta$ is full dimensional in that it has dimension $d$.
277: We assume that the blending functions do not vanish simultaneously
278: at any point of $\Delta$.
279: That is, the blending functions have no base points in their domain.
280:
281: A set $\{\bb_\ba\mid\ba\in\calA\}\subset\R^\ell$ of \Blue{{\it control points}}
282: gives a map $F \colon\Delta\to\R^\ell$ defined by
283: %
284: \begin{equation}\label{Eq:P_Patch}
285: F(x)\ :=\
286: \frac{\sum_{\ba\in\calA} \beta_\ba(x)\,\bb_\ba }
287: {\sum_{\ba\in\calA} \beta_\ba(x)}\ .
288: \end{equation}
289: %
290: The denominator in~\eqref{Eq:P_Patch} is positive on $\Delta$ and so
291: the map $F$ is well-defined.
292:
293: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
294: \begin{remark}\label{R:Comments}
295: Positive weights, or shape parameters,
296: $\{w_\ba\in\R_>\mid\ba\in\calA\}$ scaling the blending functions are often included
297: in the definition of the map $F$.
298: We have instead absorbed them into our notion of blending functions.
299: See Remarks~\ref{R:Restrictive} and \ref{Rem:weights}.\hfill\QED
300: \end{remark}
301: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
302:
303: The blending functions $\{\beta_\ba\mid\ba\in\calA\}$ are
304: \Blue{{\it normalized}} if they form a partition of unity,
305: \[
306: \sum_{\ba\in\calA} \beta_\ba(x)\ =\ 1\,.
307: \]
308: As the denominator in~\eqref{Eq:P_Patch} is (strictly) positive on $\Delta$, we may divide
309: each blending function by this denominator to obtain normalized blending functions.
310: \[
311: \mbox{Redefine}\quad
312: \beta_\ba\ :=\ \frac{\beta_\ba}{\sum_{\ba\in\calA} \beta_\ba}\,.
313: \qquad\mbox{Then}\quad
314: \beta_\ba(x)\ \geq\ 0
315: \quad\mbox{and}\quad
316: \sum_{\ba\in\calA} \beta_\ba(x)\ =\ 1\,,
317: \]
318: for $x\in\Delta$.
319: For normalized blending functions, the formula~\eqref{Eq:P_Patch} becomes
320: %
321: \begin{equation}\label{Eq:Conv_Combin}
322: F(x)\ =\ \sum_{\ba\in\calA} \beta_\ba(x) \bb_\ba\,.
323: \end{equation}
324: %
325: We deduce two fundamental properties of parametric patches.
326: \begin{itemize}
327: \item {\bf Convex hull property.}
328: For $x\in\Delta$, $F(x)$ is a convex combination of the control
329: points.
330: Thus $F(\Delta)$ lies in the convex hull of the control points.
331: \item {\bf Affine invariance.} If $\Lambda$ is an affine function on
332: $\R^\ell$, then $\Lambda(F(x))$ is the map with the same blending
333: functions, but with control points
334: $\{\Lambda(\bb_\ba)\mid\ba\in\calA\}$.
335: \end{itemize}
336:
337: A patch is \Blue{{\it non-degenerate}} if its blending functions are
338: linearly independent.
339: This implies in particular that $F(\Delta)$ cannot collapse into a point
340: unless all the control points are equal.
341:
342: We make our key definition.
343:
344: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
345: \begin{definition}\label{DE:LP}
346: A patch
347: $\{\beta_\ba\mid\ba\in\calA\}$ has \Blue{{\it linear precision}} if
348: for every affine function $\Lambda(x)$ defined on $\R^d$,
349: \[
350: \Lambda(x)\ =\
351: \sum_{\ba\in\calA} \beta_\ba(x)\,\Lambda(\ba)\,.
352: \]
353: That is, if the patch can replicate affine functions.
354: This notion depends strongly on the positions of the points in
355: $\calA$. \hfill\QED
356: \end{definition}
357: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
358:
359:
360: The \Blue{{\it tautological map $\tau$}} of a given patch is the
361: map~\eqref{Eq:Conv_Combin} when the control points are taken
362: to be the corresponding points of $\calA$,
363: %
364: \begin{equation}\label{Eq:Tautological}
365: \tau(x)\ :=\
366: \sum_{\ba\in\calA} \beta_\ba(x)\,\ba\,.
367: \end{equation}
368: %
369: By the convex hull property, $\tau(\Delta)\subset \Delta$.
370: By affine invariance, the patch has linear precision if and
371: only if $\tau$ is the identity function, $\tau(x)=x$.
372: We record this fact.
373:
374: \begin{proposition}\label{P:LP_criterion}
375: A patch has linear precision if and only if its tautological
376: map is the identity map on the domain polytope $\Delta$.
377: \end{proposition}
378:
379: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
380: \begin{remark}\label{R:Restrictive}
381: This is a more restrictive notion of linear precision
382: than is typically considered.
383: It is more common to define a patch to be a collection
384: $\{\beta_i\mid i=1,\dotsc,n\}$ of
385: blending functions, and this has linear precision if we have
386: \[
387: x\ =\ \frac{\sum_{i=1}^n w_i\,\beta_i(x)\, \ba_i}
388: {\sum_{i=1}^n w_i\,\beta_i(x)}\ ,
389: \]
390: for some non-negative \Blue{{\it weights}} $\{w_i\mid i=1,\dotsc,n\}$ and
391: some points $\calA=\{\ba_i\mid i=1,\dotsc,n\}$ whose convex hull is the
392: domain polytope $\Delta$ of the blending functions.
393:
394: Our more restrictive definition of patches and linear precision, where we
395: incorporate the weights and the points $\calA$ into our definition of blending functions,
396: allows us to derive precise criteria with which to study linear precision.
397: Our intention is to employ these criteria to study patches in the
398: generality in which other authors had worked. \hfill\QED
399: \end{remark}
400: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
401:
402: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
403: \begin{example}\label{Ex:BC}
404: Fix an integer $n>0$ and let $\calA:=\{\frac{i}{n}\mid \,0\leq i\leq n\}$
405: be the set of $n{+}1$ equally distributed points in the unit interval.
406: The $i$th \Blue{{\it Bernstein polynomial\/}} is $\beta_i:=\binom{n}{i}x^i(1{-}x)^{n-i}$,
407: which we associate to the point $\frac{i}{n}\in\calA$.
408: These form a partition of unity,
409: \[
410: \sum_{i=0}^n \beta_i(x)\ =\
411: \sum_{i=0}^n \binom{n}{i}x^i(1{-}x)^{n-i}\ =\
412: \bigl(x\, +\, (1{-}x)\bigr)^n\ =\ 1\,,
413: \]
414: and are therefore normalized blending functions.
415: Given control points $\bb_{i}\in\R^3$, formula~\eqref{Eq:P_Patch}
416: becomes
417: \[
418: F(x)\ =\
419: \sum_{i=0}^n \binom{n}{i}x^i(1-x)^{n-i} \bb_i \,,
420: \]
421: which is the classical formula for a B\'ezier curve of degree $n$ in $\R^3$.
422:
423: This patch has linear precision.
424: First, note that $\frac{i}{n}\cdot\binom{n}{i}= \binom{n-1}{i-1}$.
425: Then $\tau(x)$ is
426: %
427: \begin{eqnarray*}
428: \sum_{i=0}^n\binom{n}{i}x^i(1-x)^{n-i}\cdot \frac{i}{n}
429: &=&
430: \sum_{i=1}^n\binom{n-1}{i-1}x^i(1-x)^{n-i}\\
431: &=& x\cdot
432: \sum_{j=0}^{n-1}\binom{n-1}{j}x^j(1-x)^{n-1-j}
433: \quad=\quad x\,.\makebox[.1cm][l]{\hspace{36pt}$\QED$}
434: \end{eqnarray*}
435: %
436: \end{example}
437: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
438:
439: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
440: \begin{example}\label{Ex:Wach}
441: Let $\calA$ be the vertices of the hexagon, given
442: %
443: % DO not remove. This is in case we need to make it a figure again
444: %
445: %in Figure~\ref{F:Wach}
446: below by its defining inequalities.
447: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
448: %\begin{figure}[htb]
449: %
450: \[
451: \begin{picture}(250,172)(-51,-11)
452: \put( 0, 0){\includegraphics[height=150pt]{figures/hexagon.eps}}
453: \put(120.5,58){$1$} \put(60,120){$1$}
454: \put(83,23.5){$-1$} \put(13.8,84){$-1$}
455:
456: \put(100,153){$1-y>0$}
457: \put(-35,122){$1+x-y>0$} \put(148, 95){$1-x>0$}
458: \put(-51, 47){$1+x>0$} \put(113, 22){$1+y-x>0$}
459: \put( 2,-10){$1+y>0$}
460:
461: \put(111.5,148){\vector(-1,-2){9.5}}
462: \put(31.5,118.5){\vector(1,-1){14.5}}\put(146, 98){\vector(-1, 0){20}}
463: \put( 1 ,50){\vector(1, 0){20}} \put(118.5, 31.5){\vector(-1, 1){14.5}}
464: \put(38.5, 2){\vector(1, 2){9.5}}
465: \end{picture}
466: % \qquad\qquad \raisebox{60.3pt}{$\displaystyle
467: % \left\{
468: % \begin{array}{rcl}
469: % 1+x&>&0\\
470: % 1+x-y&>&0\rule{0pt}{13.1pt}\\
471: % 1-y&>&0 \rule{0pt}{13.1pt}\\
472: % 1-x&>&0 \rule{0pt}{13.1pt}\\
473: % 1+y-x&>&0\rule{0pt}{13.1pt}\\
474: % 1+y&>&0 \rule{0pt}{13.1pt}
475: % \end{array} \right.$ }
476: \]
477: % \caption{A Hexagon.\label{F:Wach}}
478: %\end{figure}
479: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
480: Let the blending function associated to a vertex be the product of the linear forms
481: defining edges that do not contain that vertex.
482: %
483: \begin{equation}\label{Eq:WHex}
484: % \begin{eqnarray*}
485: \begin{array}{rcl}
486: \beta_{\kzwei{-1}{-1}}&:=&(1+x-y)(1-y)(1-x)(1+y-x)\\\rule{0pt}{14pt}
487: \beta_{\kzwei{-1}{\smi 0}} &:=&(1-y)(1-x)(1+y-x)(1+y) \\\rule{0pt}{14pt}
488: \beta_{\kzwei{0}{1}} &:=&(1-x)(1+y-x)(1+y)(1+x) \\\rule{0pt}{14pt}
489: \beta_{\kzwei{1}{1}} &:=&(1+y-x)(1+y)(1+x)(1+x-y) \\\rule{0pt}{14pt}
490: \beta_{\kzwei{1}{0}} &:=&(1+y)(1+x)(1+x-y)(1-y) \\\rule{0pt}{14pt}
491: \beta_{\kzwei{\smi 0}{-1}} &:=&(1+x)(1+x-y)(1-y)(1-x)
492: \end{array}
493: \end{equation}
494: %
495: The normalized blending functions have denominator
496: $\varphi(x,y):=2(3+xy-x^2-y^2)$, which is strictly positive on
497: $\Delta$.
498: These normalized blending functions have linear precision.
499: %%
500: \[
501: % \beta_{\kzwei{-1}{-1}} \zwei{-1}{-1}
502: % +\beta_{\kzwei{-1}{\smi 0}}\zwei{-1}{0}
503: % +\beta_{\kzwei{0}{1}} \zwei{0}{1}
504: % +\beta_{\kzwei{1}{1}} \zwei{1}{1}
505: % +\beta_{\kzwei{1}{0}} \zwei{1}{0}
506: % +\beta_{\kzwei{\smi 0}{-1}}\zwei{0}{-1}\ =\
507: \sum_{\ba\in\calA} \beta_{\ba}(x,y)\,\ba\ =\
508: \varphi(x,y) \zwei{x}{y} \ .
509: % \bigl( \beta_{(-1,-1)} + \beta_{(-1,0)}
510: % + \beta_{(0,1)} + \beta_{(1,1)}
511: % + \beta_{(1,0)} + \beta_{(0,-1)} \bigr)\ .
512: \]
513: In fact, these normalized blending functions are
514: Wachspress's barycentric coordinates~\cite{Wa75} for this hexagon.\hfill\QED
515: \end{example}
516: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
517:
518: When $\calA$ is the set of vertices of $\Delta$
519: (so that $\calA$ is in convex position), then normalized blending
520: functions of a patch with linear precision are
521: \Blue{{\it barycentric coordinates}} for $\Delta$.
522: By this we mean in the sense of~\cite{Fl,Wa75,Wa96}: nonnegative functions indexed by the
523: extreme points of $\Delta$ that have linear precision in the sense of
524: Definition~\ref{DE:LP}.\smallskip
525:
526: A \Blue{{\it reparametrization}} of a patch
527: $\{\beta_\ba\mid\ba\in\calA\}$ by a homeomorphism
528: $\psi\colon \Delta\to\Delta$
529: is a new patch with blending functions
530: $\beta_\ba\circ \psi$ for $\ba\in\calA$.
531: A patch is \Blue{{\it proper}} if its tautological
532: map~\eqref{Eq:Tautological} is a homeomorphism.
533: This condition is necessary for the patch to have linear precision.
534: If a patch is proper, then reparametrizing it by
535: the inverse $\tau^{-1}$ of $\tau$ gives a new patch whose
536: tautological map is $\tau\circ\tau^{-1}$, the identity
537: function, and $\tau^{-1}$ is the unique homeomorphism of $\Delta$ with
538: this property.
539: We record this straightforward, yet fundamental result about linear precision.
540:
541: %%%%%%%%%%%%%%%
542: \begin{theorem}\label{T:Lin_Prec}
543: A proper patch has a unique reparametrization that has linear precision.
544: \end{theorem}
545: %%%%%%%%%%%%%
546:
547: Theorem~\ref{T:Lin_Prec} suggests that it will be fruitful to discuss
548: patches up to reparametrization.
549: In Section~\ref{S:Geometry}, we show that a natural geometric object
550: $X_\beta$ associated to a patch is invariant under
551: reparametrizations and thus represents the patch up to reparametrization.
552:
553: A patch is \Blue{{\it rational}} if it has a reparametrization whose
554: normalized blending functions are rational functions (quotients of
555: polynomials).
556: A patch has \Blue{{\it rational linear precision}} if
557: it is proper and its reparametrization having linear precision has blending
558: functions that are rational functions.
559: Such a patch is necessarily rational.
560: We seek criteria that determine when a rational patch has
561: rational linear precision.
562:
563: Theorem~\ref{T:Lin_Prec} and our preceding discussion concerns
564: reparametrizations of a patch having linear precision.
565: We may also alter a patch by multiplying its blending functions by
566: positive real numbers, typically called weights, or
567: by moving the points of $\calA$ that are not vertices of the domain
568: polytope $\Delta$.
569: This second change does not alter any map~\eqref{Eq:P_Patch} given by
570: control points, but it will change the tautological map
571: $\tau$~\eqref{Eq:Tautological}.
572: Our analysis in the next section may be helpful in addressing
573: whether it is possible to tune a given patch (using weights or
574: moving the points in $\calA$) to obtain one that has
575: rational linear precision.
576: Example~\ref{Ex:Pentagon} (due to Krasauskas and Kar{\v{c}}iauskas)
577: shows how one toric patch may be tuned to achieve linear precision.
578:
579:
580: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
581: %
582: %
583: %
584: \section{The geometry of linear precision}\label{S:Geometry}
585:
586: We introduce an algebraic-geometric formulation of patches to clarify the
587: discussion in Section~\ref{S:PP} and to provide tools with which to
588: understand and apply Theorem~\ref{T:Lin_Prec}.
589: This leads us to discuss the relevance for geometric modeling of the
590: subtle difference between rational varieties and unirational varieties.
591: Lastly, we use this geometric formulation to give a geometric
592: characterization of when a patch has rational linear precision.
593: We recommend the text~\cite{CLO} for additional background on algebraic
594: geometry.
595: We first review linear projections, which are the geometric counterpart
596: of control points.
597:
598:
599: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
600: \subsection{Linear projections}
601:
602: We consider $\R^\ell$ as a subset of the
603: $\ell$-dimensional real projective space $\R\P^\ell$ via the embedding
604: \[
605: \R^\ell\ \ni\ z\ \longmapsto [1,z]\ \in\ \R\P^\ell\,.
606: \]
607: A point $[z_0,z_1,\dotsc,z_\ell]\in\R\P^\ell$ lies in this copy of
608: $\R^\ell$ if and only if $z_0\neq 0$.
609: In that case, the corresponding point of $\R^\ell$ is
610: %
611: \begin{equation}\label{Eq:CoordsRd}
612: \Bigl( \frac{z_1}{z_0},\, \frac{z_2}{z_0}, \,\dotsc,\,
613: \frac{z_\ell}{z_0}\Bigr)\,.
614: \end{equation}
615:
616: Let $\calA\subset\R^d$ be a finite set of points and
617: write $\R\P^\calA$ for the real projective
618: space whose coordinates are indexed by $\calA$.
619: Let $b=\{\bb_\ba\in\R^\ell\mid \ba\in\calA\}$ be a collection of control
620: points, which we regard as points in $\R\P^\ell$ so that
621: $b=\{[1,\bb_\ba]\in\R\P^\ell\mid \ba\in\calA\}$.
622:
623: Given a point $y=[y_\ba\mid\ba\in\calA]\in \R\P^\calA$, if the sum
624: %
625: \begin{equation}\label{Eq:sum}
626: \sum_{\ba\in\calA} y_\ba \cdot(1, \bb_\ba) \ \in \ \R^{\ell+1}
627: \end{equation}
628: %
629: is non-zero then it represents a point
630: $\pi_b(y)$ in $\R\P^\ell$.
631: This map $\pi_{b}$ is a \Blue{{\it linear projection}}
632: %
633: \begin{equation}\label{Eq:Lin_Proj}
634: \pi_b\ \colon\ \R\P^\calA\ \longrat\ \R\P^\ell\,.
635: \end{equation}
636: %
637: We use a broken arrow $\rat$ to indicate that $\pi_b$ is defined only
638: on the complement of the set
639: $E_b\subset\R\P^\calA$ where the sum~\eqref{Eq:sum} vanishes.
640: This linear subspace is the \Blue{{\it center of projection}}.
641:
642:
643: When the control points affinely span $\R^\ell$,
644: the center of projection $E_b$ has codimension $\ell\!+\!1$ in $\R\P^\calA$.
645: The inverse image $\pi_b^{-1}(x)$ of a point $x\in\R\P^\ell$
646: is the set $H\setminus E_b$, where $H\subset\R\P^\calA$ is a
647: linear subspace of codimension $\ell$ containing $E_b$.
648: If $L$ is any $\ell$-dimensional linear subspace of $\R\P^\calA$ that
649: does not meet $E_b$, then $\pi_b$ maps $L$ isomorphically to
650: $\R\P^\ell$.
651: Identifying $L$ with $\R\P^\ell$ gives an explicit
652: description of the map $\pi_b$.
653: If $y\in\R\P^\calA\setminus E_b$, then
654: \[
655: \pi_b(y)\ =\ L\ \cap\ \left(\overline{ y,\,E_b }\right)\,,
656: \]
657: the intersection with $L$ of the linear span of $y$ and the center
658: $E_b$.
659: If $y\in L$, then its inverse image under $\pi_b$ is
660: $(\overline{ y,\,E_b })\setminus E_b$.
661:
662: For example, Figure~\ref{F:cubic_Projection}
663: shows the effect of a linear projection $\pi$
664: on a cubic space curve $C$.
665: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
666: \begin{figure}[htb]
667: \[
668: \begin{picture}(210,150)(0,-2)
669: \put(0,0){\includegraphics{figures/Cubic_Projection.eps}}
670:
671: \put(10,130){$\pi$}\put(20,133){\vector(1,0){20}}
672:
673: \put(0,62){$E$}\put(12,64){\vector(2,-1){20}}
674: \put(18,79){$B$}\put(30,82){\vector(2,-1){20}}
675:
676: \put(135,114){$\overline{ y,\, E}$}
677: \put(145,109){\vector(-1,-4){5}}
678:
679: \put(207,79){$y$}\put(204,81){\vector(-1,0){20}}
680: \put(207,119){$L$}\put(204,121){\vector(-1,0){20}}
681:
682: \put(155,8){\Blue{$C$}}
683:
684: \put(100,12.5){\vector(-3, 4){21}}
685: \put(116.5,12.5){\vector(0,1){62}}
686: \put(92,-1){$\pi^{-1}(y)$}
687:
688: \end{picture}
689: \]
690: \caption{A linear projection $\pi$ with center $E$.\label{F:cubic_Projection}}
691:
692: \end{figure}
693: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
694: The center of projection is a line, $E$, which meets the curve in a
695: point, $B$.
696:
697:
698:
699: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
700: \subsection{Geometric formulation of a patch}
701: Consider \eqref{Eq:P_Patch} from the point of view of
702: projective geometry.
703: Given a patch
704: $\beta=\{\beta_\ba(x)\mid\ba\in\calA\}$ with $\calA\subset\R^d$, then
705: %
706: \begin{equation}\label{Eq:blending_param}
707: x\ \longmapsto\ [\beta_\ba(x)\mid \ba\in\calA]
708: \end{equation}
709: %
710: defines a map $\beta\colon\Delta\to\R\P^\calA$ whose $\ba$th coordinate
711: is the blending function $\beta_\ba(x)$.
712:
713: The map $\beta$ is unchanged if we
714: multiply all blending functions by the same positive function $\varphi(x)$
715: defined on $\Delta$.
716: Thus we
717: obtain the same map if we use instead the normalized blending functions.
718: The image $X_\beta$ of $\Delta$ under $\beta$ is a closed subset of
719: $\R\P^\calA$ (but not Zariski closed!), as $\Delta$ is compact and the map $\beta$ is
720: continuous.
721: It is non-degenerate (does not lie in a hyperplane) exactly when the
722: patch is non-degenerate.
723:
724: Suppose now that we have control points
725: $b=\{\bb_\ba\in\R^\ell\mid \ba\in\calA\}$ and
726: consider the composition of the map $\beta$ and the linear projection $\pi_b$
727: \[
728: \Delta\ \xrightarrow{\ \beta\ }\ X_\beta\ \subset\
729: \R\P^\calA\ \stackrel{\pi_b}{\longrat}\ \R\P^\ell\,.
730: \]
731: By our assumption on the positivity of the blending functions,
732: formula~\eqref{Eq:CoordsRd} implies that the image lies in the standard copy of $\R^\ell$, and
733: there it is given by the formula~\eqref{Eq:P_Patch}.
734: Thus this image, $F(\Delta)$, is the image of $X_\beta$
735: under the projection $\pi_b$ given by the control points.
736:
737:
738: A reparametrization of the patch $\beta$ by a homeomorphism
739: $\psi\colon\Delta\to\Delta$ gives a different map
740: with the same image in $\R\P^\calA$
741: %
742: \[
743: \widetilde{\beta}\ \colon\
744: \Delta\ \xrightarrow{\ \psi\ }\
745: \Delta\ \xrightarrow{\ \beta\ }\ X_\beta\,.
746: \]
747: %
748: Thus $X_\beta$ is an invariant of the patch modulo reparametrization.
749:
750: If we take $\calA$ to be our set of control points where the element $\ba\in\calA$ is the
751: control point associated to the blending function $\beta_\ba$, then
752: the resulting linear projection $\pi_\calA$ is the
753: \Blue{{\it tautological projection}},
754: %
755: \begin{equation}\label{Eq:Taut_proj}
756: \pi_\calA\ \colon\ \R\P^\calA\ \longrat\ \R\P^d\,.
757: \end{equation}
758: %
759: Here, the coordinate points $(e_\ba\mid \ba\in\calA)$ are indexed by elements of $\calA$
760: and $\pi_\calA(e_\ba)=\ba\in\R\P^d$.
761: The tautological map $\tau$~\eqref{Eq:Tautological} is the composition
762: \[
763: \Delta\ \xrightarrow{\ \beta\ }\ X_\beta\ \subset\
764: \R\P^\calA\ \stackrel{\pi_\calA}{\longrat}\ \R\P^d\,.
765: \]
766: The patch is proper if this map is a homeomorphism onto its image
767: $\Delta\subset\R\P^d$, and the patch has linear precision if the
768: composition is the identity map, by Proposition~\ref{P:LP_criterion}.
769:
770: With these definitions, we have the following identification of the reparametrization
771: having linear precision, which is a more precise version of Theorem~\ref{T:Lin_Prec}.
772:
773: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
774: \begin{theorem}\label{Th:LP}
775: Suppose that $X_{\beta}$ is a proper patch and $\pi_{\calA}\colon X_{\beta}\to\Delta$
776: is the tautological projection restricted to $X_{\beta}$.
777: Then the blending functions for $X_{\beta}$ that have linear precision are given by
778: the coordinates of the inverse of $\pi_{\calA}$.
779: \end{theorem}
780: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
781:
782:
783:
784:
785: \begin{remark}\label{R:toric}
786: This geometric perspective, where a patch is first an embedding
787: of $\Delta$ into $\R\P^\calA$ followed by a linear projection, may have been introduced into
788: geometric modeling in~\cite{dR91}.
789: It is fundamental for Krasauskas's toric
790: B\'ezier patches~\cite{KR02}, and was reworked in the
791: tutorial~\cite{So03}.
792: There, an analog of Theorem~\ref{T:Lin_Prec} was formulated.
793: We take this opportunity to correct an error in notation.
794: The tautological projection is an algebraic
795: version of the moment map of the toric variety underlying
796: a toric patch, and not the actual moment map from symplectic geometry.\hfill\QED
797: \end{remark}
798:
799:
800: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
801: \subsection{Rational varieties}
802:
803: We study the algebraic relaxation of our previous notions, replacing the real numbers by
804: the complex numbers so that we may use notions from algebraic geometry.
805: A rational patch has a parame\-trization by rational functions.
806: Clearing their denominators gives a new collection
807: $\{\beta_\ba\mid\ba\in\calA\}$ of blending functions which are
808: polynomials.
809: These polynomials define complex-valued functions on $\C^d$, and so the
810: blending functions give a map
811: \[
812: \beta_\C\ \colon\ \C^d\setminus B\ \longrightarrow\ \C\P^\calA\,,
813: \]
814: which is defined on the complement of the set
815: $B$ where all the blending functions vanish.
816: This is called the \Blue{{\it base locus}} of the map $\beta_\C$.
817: This map $\beta_\C$ extends the map $\beta\colon\Delta\to X_\beta$.
818: We write $\beta$ for $\beta_\C$, $\P^\calA$ for $\C\P^\calA$, and in general
819: use the same notation for maps defined on complex algebraic varieties as for
820: their restrictions to subsets of their real points.
821:
822: As with linear projections, we write $\beta\colon\C^d\rat\P^\calA$
823: to indicate that $\beta$ is only defined on $\C^d\setminus B$.
824: Such a map between algebraic varieties that is given by polynomials and
825: defined on the complement
826: of an algebraic subset is called a \Blue{{\it rational map}}.
827:
828: Let $Y_\beta$ be the Zariski closure of the image of $\beta$.
829: One important consequence of the patch being rational is that
830: $X_\beta$ is a full-dimensional subset of the real points of the algebraic variety
831: $Y_\beta$.
832: In particular, $X_\beta$ is defined locally in $\R\P^\calA$ by the
833: vanishing of some polynomials, and this remains true for any image
834: $F(\Delta)$ of a map $F$~\eqref{Eq:P_Patch} given by our original blending
835: functions and any choice of control points.
836: This is important in modeling, for these implicit equations are used
837: to compute intersections of patches.
838: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
839:
840: Suppose that $X_\beta$ has rational linear precision and that $\beta=\{\beta_\ba\mid\ba\in\calA\}$
841: are rational blending functions with linear precision.
842: By Proposition~\ref{P:LP_criterion}, the composition
843: \[
844: \Delta\ \xrightarrow{\ \beta\ }\
845: X_\beta \ \stackrel{\pi_\calA}{\longrat}\ \Delta
846: \]
847: is the identity map.
848: Since $\Delta$ has full dimension in $\R^d$, it is Zariski dense in $\C^d$, and so
849: the composition of rational functions
850: %
851: \begin{equation}\label{Eq:comp}
852: \C^d\ \stackrel{\beta}{\longrat}\ Y_\beta\
853: \stackrel{\pi_\calA}{\longrat}\ \C^d
854: \end{equation}
855: %
856: is also the identity map (where it is defined).
857:
858:
859: We introduce some terminology to describe this situation.
860: A rational map $\varphi\colon Y\rat Z$ between complex algebraic
861: varieties is a \Blue{{\it birational isomorphism}} if there is another map
862: $\psi\colon Z\rat Y$ such that $\varphi\circ\psi$ and $\psi\circ\varphi$
863: are the identity maps where they are defined.
864: In particular, $Y$ has a Zariski open subset $U$ and $Z$ a Zariski open
865: subset $V$ such that $\varphi|_U$ is one-to-one between $U$ and
866: $V$.
867: The projection $\pi\colon C\rat L$ of Figure~\ref{F:cubic_Projection}
868: is not a birational isomorphism.
869: Indeed, if $x\in C$ is a point in $\pi^{-1}(p)$, then
870: $\overline{p,E}=\overline{x,E}$ and this plane meets
871: the complement $C\setminus B$ of the base locus in $x$ and one other point.
872:
873:
874:
875: Thus if the composition~\eqref{Eq:comp} is the identity map, then the map
876: $\pi_\calA$ is a birational isomorphism from $Y_\beta\rat\C\P^d$.
877: We deduce the algebraic-geometric version of Proposition~\ref{P:LP_criterion}.
878:
879: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
880: \begin{theorem}\label{Th:LP=birational}
881: A patch $\beta=\{\beta_\ba\mid \ba\in\calA\}$ has rational linear precision if
882: and only if the complexified tautological projection
883: \[
884: \pi_\calA\ \colon\ Y_\beta\ \ \longrat\ \ \C\P^d
885: \]
886: is a birational isomorphism.
887: \end{theorem}
888: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
889:
890: Using well-known properties of birational projections, we use this
891: characterization of linear precision to deduce
892: necessary conditions for a patch to have linear precision.
893: A \Blue{{\it rational variety}} of dimension $d$ is one that is birational to
894: $\C^d$.
895:
896: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
897: \begin{theorem}\label{Th:Necessary}
898: If a patch $\beta=\{\beta_\ba\mid \ba\in\calA\}$ has rational linear precision,
899: then
900: \begin{enumerate}
901: \item $Y_\beta$ is a rational variety,
902: \item almost all codimension $d$ planes $L$ containing the center $E_\calA$ of the tautological
903: projection meet
904: $Y_\beta$ in at most one point outside of $E_\calA$.
905: \end{enumerate}
906: \end{theorem}
907: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
908:
909: By condition (2), $Y_\beta$ has an exceptionally singular position with respect
910: to $E_\calA$.
911: Typically, $Y_\beta$ does not meet a given codimension $d{+}1$ plane and its intersection
912: with a given codimension $d$ plane consists of $\deg(Y_\beta)$ points, counted with
913: multiplicity.
914: By Condition (2), not only does $Y_\beta$ meet $E_\calA$, but if $L$ has codimension $d$
915: and contains $E_\calA$, then most of the $\deg(Y_\beta)$ points in $Y_\beta\cap L$
916: lie in $E_\calA$.
917: We will see this in Examples~\ref{Ex:WachH} and~\ref{Ex:Pentagon}.
918: \smallskip
919:
920:
921: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
922: \noindent{\it Proof.}
923: Statement (1) is immediate from Theorem~\ref{Th:LP=birational}.
924:
925: Since the birational isomorphism
926: $\pi_\calA\colon Y_\beta\,\;\rat\;\,\C\P^d$ is the restriction of a linear
927: projection with center $E_\calA$,
928: then for almost all points $y\in Y_\beta\setminus E_\calA$, we have
929: \[
930: \bigl(\overline{y,\,E_\calA}\bigr) \cap Y_\beta\ =\
931: \{y\}\cup \bigl( E_\calA\cap Y_\beta\bigr)\,.
932: \]
933: Since all codimension $d$ planes $L$ that contain $E_\calA$ have the form
934: $(\overline{y,\,E_\calA})$ for some $y\not\in E_\calA$ and almost all meet $Y_\beta$,
935: statement (2) follows.
936: \hfill\QED
937:
938:
939: Consider the geometric situation of Figure~\ref{F:cubic_Projection}.
940: Suppose now that $E$ is the line tangent to $C$ at the point $B$.
941: Then $E\cap C$ is the point $B$, but counted with multiplicity 2.
942: A linear projection with center $E$ restricts to a birational isomorphism of $C$
943: with $\P^1$.
944:
945:
946: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
947: \subsection{Unirational varieties}
948:
949:
950: The necessary condition of Theorem~\ref{Th:Necessary}(1)
951: gives an important but subtle geometric restriction on patches that have
952: rational linear precision.
953:
954: The map $\beta$ provides a parametrization of an open subset of $Y_\beta$
955: by an open subset of $\C^d$.
956: Such parametrized algebraic varieties are called
957: \Blue{{\it unirational}}.
958: Unirational curves and surfaces are also rational, but these two notions
959: differ for varieties of dimension three and higher.
960: Clemens and Griffiths~\cite{CG72} showed that a smooth
961: hypersurface of degree 3 in $\P^4$ is not rational (these were classically
962: known to be unirational, e.g. by Max Noether).
963:
964: Thus parametric patches of dimension at least 3 will in general be unirational and not
965: rational.
966: This does not occur for B\'ezier simploids or toric
967: patches.
968:
969:
970:
971: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
972: \begin{example}\label{Ex:WachH}
973: Consider again the Wachspress coordinates~\eqref{Eq:WHex}
974: of Example~\ref{Ex:Wach}, which have linear precision.
975: Let $Y_\beta\subset\P^5$ be the image of the blending functions~\eqref{Eq:WHex}.
976: We study the base locus of the tautological projection $\pi_\calA$ of this patch.
977: We first find equations for $Y_\beta$ as an algebraic subvariety of $\P^5$.
978:
979: If we divide
980: $\beta_{\kzwei{-1}{-1}}\beta_{\kzwei{0}{1}}
981: +2\beta_{\kzwei{0}{1}}\beta_{\kzwei{\smi 0}{-1}}
982: +\beta_{\kzwei{1}{1}}\beta_{\kzwei{\smi 0}{-1}}$
983: by the product of the linear forms defining the hexagon
984: we get
985: \[
986: (1-x)(1+y-x) + 2(1+x)(1-x) + (1+x)(1+x-y)
987: \ =\ 4-2xy\,,
988: \]
989: which is symmetric in $x$ and $y$,
990: and so
991: %
992: \begin{equation}\label{Eq:beta_relation}
993: \beta_{\kzwei{-1}{-1}}\beta_{\kzwei{0}{1}}
994: +2\beta_{\kzwei{0}{1}}\beta_{\kzwei{\smi 0}{-1}}
995: +\beta_{\kzwei{1}{1}}\beta_{\kzwei{\smi 0}{-1}}
996: \ =\
997: \beta_{\kzwei{-1}{-1}}\beta_{\kzwei{1}{0}}
998: +2\beta_{\kzwei{1}{0}}\beta_{\kzwei{-1}{\smi 0}}
999: +\beta_{\kzwei{1}{1}}\beta_{\kzwei{-1}{\smi 0}}\,.
1000: \end{equation}
1001: %
1002: If we let $(y_\ba\mid\ba\in\calA)$ be natural coordinates for $\P^\calA$,
1003: then~\eqref{Eq:beta_relation} gives the quadratic polynomial
1004: which vanishes on $Y_\beta$,
1005: %
1006: \[
1007: y_{\kzwei{-1}{-1}}y_{\kzwei{0}{1}}
1008: +2y_{\kzwei{0}{1}}y_{\kzwei{\smi 0}{-1}}
1009: +y_{\kzwei{1}{1}}y_{\kzwei{\smi 0}{-1}}
1010: \ -\
1011: \bigl(y_{\kzwei{-1}{-1}}y_{\kzwei{1}{0}}
1012: +2y_{\kzwei{1}{0}}y_{\kzwei{-1}{\smi 0}}
1013: +y_{\kzwei{1}{1}}y_{\kzwei{-1}{\smi 0}}\bigr)\,.
1014: \]
1015: Cyclically permuting the vertices of the hexagon gives two other
1016: quadratics that vanish on $Y_\beta$, but these three sum to 0.
1017: There is another quadratic polynomial vanishing on $Y_\beta$,
1018: \[
1019: y_{\kzwei{-1}{-1}}y_{\kzwei{0}{1}} +
1020: y_{\kzwei{0}{1}}y_{\kzwei{1}{0}}+
1021: y_{\kzwei{1}{0}}y_{\kzwei{-1}{-1}}\ -\
1022: \bigl( y_{\kzwei{1}{1}}y_{\kzwei{\smi 0}{-1}} +
1023: y_{\kzwei{\smi 0}{-1}}y_{\kzwei{-1}{\smi 0}}+
1024: y_{\kzwei{-1}{\smi 0}}y_{\kzwei{1}{1}}\bigr)\ .
1025: \]
1026:
1027: There is an additional cubic relation among the blending functions,
1028: which gives a cubic polynomial that vanishes on $Y_\beta$
1029: %
1030: \begin{equation}\label{Eq:cubic}
1031: y_{\kzwei{-1}{-1}}y_{\kzwei{0}{1}}y_{\kzwei{1}{0}}\ -\
1032: y_{\kzwei{1}{1}}y_{\kzwei{\smi 0}{-1}}y_{\kzwei{-1}{\smi 0}}\,.
1033: \end{equation}
1034: %
1035: These relations, three independent quadratics and one cubic, define $Y_\beta$ as an
1036: algebraic subset of $\P^\calA=\P^5$.
1037: The quadratic equations define $Y_\beta$, together with the
1038: 2-dimensional linear space cut out by the three linear equations.
1039: %
1040: \begin{equation}\label{Eq:lin_forms}
1041: \begin{array}{rcl}
1042: 0&=&
1043: y_{\kzwei{-1}{-1}} + y_{\kzwei{0}{1}} + y_{\kzwei{1}{0}} +
1044: y_{\kzwei{1}{1}} + y_{\kzwei{\smi 0}{-1}} +y_{\kzwei{-1}{\smi 0}}\\
1045: 0&=&\rule{0pt}{13pt}
1046: y_{\kzwei{1}{1}} -y_{\kzwei{-1}{-1}}+y_{\kzwei{1}{0}}-y_{\kzwei{-1}{\smi 0}}\\
1047: 0&=&\rule{0pt}{13pt}
1048: y_{\kzwei{1}{1}} -y_{\kzwei{-1}{-1}}+y_{\kzwei{0}{1}}-y_{\kzwei{\smi 0}{-1}}
1049: \end{array}
1050: \end{equation}
1051:
1052: Thus $Y_\beta$ has degree $7=2^3-1$.
1053: A general codimension $2$ plane $L$ containing the center $E_\calA$ of the tautological projection
1054: $\pi_\calA$
1055: will meet $Y_\beta$ in at most 1 point outside of $E_\calA$, by Theorem~\ref{Th:Necessary}(2).
1056: Since $L\cap Y_\beta$ has degree at least 7,
1057: the other 6 points must lie in the base locus
1058: $E_\calA\cap Y_\beta$, which is in fact
1059: a reducible cubic plane curve.
1060:
1061: The three linear forms~\eqref{Eq:lin_forms} also define the tautological
1062: projection, $\pi_\calA$, so this linear subspace is the center of projection
1063: $E_\calA$, and the base locus $B$ is defined in $E_\calA$ by the cubic~\eqref{Eq:cubic}
1064: defining $Y_\beta$.
1065: Let us parametrize the center $E_\calA$ as follows.
1066: For $[a,b,c]\in\P^2$, set
1067: \[
1068: \begin{array}{rclcrcl}
1069: y_{\kzwei{-1}{-1}} &:=& a+b && y_{\kzwei{1}{1}} &:=& -a+b\\
1070: y_{\kzwei{0}{1}} &:=& a-c && y_{\kzwei{\smi 0}{-1}}&:=& -a-c\\
1071: y_{\kzwei{1}{0}} &:=& a-b+c&& y_{\kzwei{-1}{\smi 0}}&:=& -a-b+c
1072: \end{array}
1073: \]
1074: Then we have
1075: \[
1076: y_{\kzwei{-1}{-1}}y_{\kzwei{0}{1}}y_{\kzwei{1}{0}}\ -\
1077: y_{\kzwei{1}{1}}y_{\kzwei{\smi 0}{-1}}y_{\kzwei{-1}{\smi 0}}\ =\
1078: 2a(a^2-b^2+bc-c^2)\, ,
1079: \]
1080: the product of a linear and a quadratic form.
1081: \hfill\QED
1082: \end{example}
1083: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1084:
1085:
1086:
1087: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1088: %
1089: %
1090: %
1091: \section{Toric Patches}
1092:
1093: Toric patches, which were introduced by Krasauskas~\cite{KR02}, are a
1094: class of patches based on certain
1095: special algebraic varieties, called toric varieties.
1096: For a basic reference on toric varieties as parametrized varieties, see the
1097: book~\cite{GBCP}, particularly Chapters 4 and 13, and the articles~\cite{Cox03,So03},
1098: which are introductions aimed at people in geometric modeling.
1099: We first define toric patches, give some examples, reinterpret
1100: Theorem~\ref{Th:LP=birational} for toric patches, and then state the main open problem
1101: about linear precision for toric patches.
1102:
1103: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1104: \subsection{Toric Patches}
1105: We regard elements $\ba=(a_1,\dotsc,a_d)\in\Z^d$ as exponent vectors for
1106: monomials in the variables $x=(x_1,\dotsc,x_d)$
1107: \[
1108: x^\ba\ :=\ x_1^{a_1} x_2^{a_2} \dotsb x_d^{a_d}\,.
1109: \]
1110: Let $\calA\subset\Z^d$ be a finite set of integer vectors whose convex hull $\Delta$ is a
1111: polytope of dimension $d$ in $\R^d$.
1112: This implies that $\calA$ affinely spans $\R^d$.
1113: Suppose that $w=\{w_\ba\in\R_>\mid \ba\in\calA\}$ is a set of positive
1114: weights indexed by $\calA$.
1115: These data $(\calA,w)$ define a map
1116: $\varphi_{\calA,w}\colon(\C^\times)^d\to\P^\calA$,
1117: %
1118: \begin{equation}\label{Eq:monom_parametrize}
1119: \varphi_{\calA,w}\ \colon\ x \ \longmapsto
1120: \ [\,w_\ba x^\ba\mid \ba\in\calA\,]\,.
1121: \end{equation}
1122: %
1123: Let $Y_{\calA,w}$ be the closure of the image of
1124: $(\C^\times)^d$ under the map $\varphi_{\calA,w}$.
1125: When the weights $w_\ba$ are equal, this is the toric variety
1126: parametrized by the monomials in $\calA$.
1127: In general, it is the translate of that toric variety by the element
1128: $w$ of the positive real torus $\R_>^\calA$, which acts on $\P^\calA$ by
1129: scaling the coordinates.
1130:
1131: Let \Blue{$X_{\calA,w}}\subset\R\P^\calA$ be the closure $\varphi_{\calA,w}(\R_>^d)$.
1132: This is the non-negative part of the translated toric variety
1133: $Y_{\calA,w}$.
1134: That is, if $\Blue{\R\P^\calA_\geq}\subset \P^\calA$ is the set of points having
1135: non-negative (real) homogeneous coordinates, then
1136: $X_{\calA,w}=Y_{\calA,w}\cap\R\P^\calA_\geq$.
1137: It is homeomorphic to the convex hull $\Delta$ of the
1138: vectors in $\calA$, as a manifold with corners~\cite[\S~4]{Fu93}.
1139:
1140: The set $\calA$ is \Blue{{\it primitive}} if it affinely spans $\Z^d$.
1141: We may assume that $\calA$ is primitive, for there is always some primitive set
1142: $\calA'\subset\Z^d$ with $Y_{\calA,w}=Y_{\calA',w}$.
1143: First note that translating every vector in $\calA$ by a fixed
1144: vector $\ba'$ multiplies each coordinate of the map $\varphi_{\calA,w}$ by the
1145: monomial $x^{\ba'}$, which does not change $\varphi_{\calA,w}$ as a map to projective space.
1146: Thus we may assume that $\calA$ contains the zero vector.
1147: Since $\calA$ affinely spans $\R^d$, it generates a subgroup $\Z\calA$ of $\Z^d$ which is
1148: isomorphic to $\Z^d$.
1149: Let $\bb_1,\dotsc,\bb_d\in\Z\calA$ be the basis corresponding to the basis of $\Z^d$
1150: under an isomorphism $\Z\calA\xrightarrow{\,\sim\,}\Z^d$.
1151: If we let $\calA'\subset\Z^d$ be the image of $\calA$ under this isomorphism
1152: and $\calB:=\{\bb_1,\dotsc,\bb_d\}$,
1153: then the map $\varphi_{\calA,w}$ factors
1154: \[
1155: (\C^\times)^d\ \xrightarrow{\ \chi_\calB\ }
1156: (\C^\times)^d\ \xrightarrow{\ \varphi_{\calA',w}\ }\P^\calA\,,
1157: \]
1158: where $\chi_\calB$ is the surjective map that sends $x$ to
1159: $(x^{\bb_1},\dotsc,x^{\bb_d})$.
1160: Note that $\varphi_{\calA', w}$ is injective but that $\chi_{\calB}$ has
1161: fibers of size $|\Z^{d}/\Z\calA|$.
1162: Thus $\varphi_{\calA,w}$ is not injective unless $\calA$ is primitive.
1163:
1164:
1165: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1166: \begin{definition}\label{DE:TP}
1167: A \Blue{{\it toric patch}} of \Blue{{\it shape $(\calA,w)$}} is any patch
1168: $\{\beta_\ba\mid\ba\in\calA\}$
1169: such that the image, $X_\beta$, of
1170: $\Delta$ under the map~\eqref{Eq:blending_param} given by the blending
1171: functions is equal to the non-negative part $X_{\calA,w}$ of the
1172: translated toric variety $Y_{\calA,w}$.\hfill\QED
1173: \end{definition}
1174: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1175:
1176: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1177: \begin{remark}\label{Rem:weights}
1178: Our geometric perspective that the fundamental object is the image $X_\beta$ requires us
1179: to absorb the weights $w$ into our definition and distinguishes patches with different
1180: choices of weights.\hfill\QED
1181: \end{remark}
1182: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1183:
1184:
1185: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1186: \begin{example}\label{Ex:TBC}
1187: Let $\calA:=\{0,1,\dotsc,n\}\subset\Z$ and let $w\in\R_>^{n+1}$ be any
1188: set of weights.
1189: Then the map $\varphi_{\calA,w}$ is
1190: \[
1191: x\ \longmapsto\ [w_0,\, w_1x,\, w_2x^2,\,\dotsc,\, w_nx^n]\,.
1192: \]
1193: The map $x\mapsto \frac{x}{n-x}$ sends $\Delta=[0,n)$ to $[0,\infty)$.
1194: Precomposing $\varphi_{\calA,w}$ with this map gives a
1195: toric patch of shape $(\calA,w)$.
1196: %
1197: \newcommand{\xnx}{\bigl({\textstyle\frac{x}{n-x}}\bigr)}
1198: %
1199: \begin{eqnarray*}
1200: [0,n]\ \ni\ x&\mapsto&\Bigl[w_0,\,w_1\xnx,\,w_2\xnx^2,\,\dotsc,\,w_n\xnx^n\Bigr]\\
1201: &=&[w_0(n-x)^n,\, w_1x(n-x)^{n-1},\,w_2x^2(n-x)^{n-2},\,
1202: \dotsc,w_nx^n]\,.
1203: \end{eqnarray*}
1204: %
1205: Observe that if we choose weights $w_i=\binom{n}{i}$, substitute
1206: $ny$ for $x$, and remove the common factors of $n^n$, then we obtain the
1207: blending functions for the B\'ezier curve of Example~\ref{Ex:BC}.
1208: Note that replacing $x$ by $ny$ also replaces
1209: $i\in\calA$ by $\frac{i}{n}$.
1210: \hfill\QED
1211: \end{example}
1212:
1213: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1214: \begin{example}\label{Ex:TBP}
1215: Krasauskas~\cite{KR02} generalized the classical B\'ezier
1216: parametrization to any polytope with integer vertices.
1217: A polytope $\Delta$ is defined by its facet inequalities
1218: \[
1219: \Delta\ =\ \{x\in\R^d\mid h_i(x)\geq 0\,, i=1,\dotsc,N\}\,.
1220: \]
1221: Here, $\Delta$ has $N$ facets and for each $i=1,\dotsc,N$,
1222: $h_i(x)=\langle \bv_i,x\rangle + c_i$ is the linear function defining
1223: the $i$th facet, where $\bv_i\in\Z^d$ is the (inward oriented) primitive vector normal to
1224: the facet and $c_i\in\Z$.
1225: (Compare these to the functions in Example~\ref{Ex:Wach}.)
1226:
1227: Let $\calA\subset\Delta\cap\Z^d$ be any subset of the integer points of
1228: $\Delta$ that includes its vertices.
1229: Let $w=\{w_\ba\mid \ba\in\calA\}\subset \R_>$ be a collection of
1230: positive weights.
1231: For every $\ba\in\calA$, Krasauskas defined \Blue{{\it toric B\'ezier functions}}
1232: %
1233: \begin{equation}\label{Eq:toric-Bezier}
1234: \beta_\ba(x)\ :=\
1235: w_\ba h_1(x)^{h_1(\ba)}h_2(x)^{h_2(\ba)}\dotsb h_N(x)^{h_N(\ba)}\,.
1236: \end{equation}
1237: %
1238: For $x\in\Delta$, these parametrize a patch $X_\beta$.
1239:
1240: Krasauskas defined a toric patch to be any reparametrization of such a
1241: patch.
1242: Our definition of a toric patch (Definition~\ref{DE:TP}) agrees with Krasauskas's,
1243: when $\calA=\Delta\cap\Z^d$ consists of all the integer points in a polytope:
1244: Observe that the map $\beta$ in~\eqref{Eq:toric-Bezier}
1245: is the composition of a map $H:\Delta\to\R^{N}$ given by
1246: $x\mapsto(h_{1}(x),\dots,h_{N}(x))$
1247: with a rational map $\varphi: \C^{N} \rat\ \R\P^{\calA}$ given by
1248: %
1249: \[
1250: \varphi\ \colon\ (u_{1},\dots,u_{N})\ \longmapsto\
1251: [w_{\ba}u_{1}^{h_{1}(\ba)}\dotsm u_{N}^{h_{N}(\ba)} \mid \ba \in \calA]\,.
1252: \]
1253: %
1254: This map $\varphi$ factors through the map $\varphi_{\calA,w}$.
1255: Indeed, define a map
1256: %
1257: \begin{eqnarray*}
1258: f_\Delta\ \colon\ (\C^\times)^N&\longrightarrow& (\C^\times)^n\\
1259: u=(u_1,\dotsc,u_N)&\longmapsto&t=(t_1,\dotsc,t_n)\quad\mbox{where}\quad
1260: t_j:=\prod_{i=1}^N u_i^{\langle \bv_i,\be_j\rangle}
1261: \end{eqnarray*}
1262: %
1263: Then
1264: \[
1265: \prod_{i=1}^N u_i^{h_i(\ba)}\ =\ \prod_{i=1}^N u_i^{c_i}\cdot
1266: \prod_{i=1}^N u_i^{\langle \bv_i,\ba\rangle}
1267: \ =\ u^c\cdot t^{\ba}\,.
1268: \]
1269: And so $\varphi(u)=u^c\cdot\varphi_{\calA,w}(f_\Delta(u))$.
1270: This shows that $X_\beta\subset X_{\calA,w}$, and Krasauskas~\cite[\S 4]{Kr06} shows
1271: that this inclusion is an equality.
1272: \hfill\QED
1273: \end{example}
1274:
1275:
1276: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1277: \begin{example}\label{Ex:CBP}
1278: A \Blue{{\it standard $d$-simplex of degree $n$}} is the simplex in
1279: $\R^d$,
1280: \[
1281: n\Delta_d\ :=\
1282: \{x\in \R^d\mid x_1+\dotsb+x_d\leq n\mbox{\ and\ }x_i\geq 0\}\,.
1283: \]
1284: For a point $\ba=(a_1,\dotsc,a_d)$ in $n\Delta_d\cap\Z^d$ set $|\ba|:=a_1+\dotsb+a_d$
1285: and let $w_\ba:=\binom{n}{\ba}=\frac{n!}{a_1!\dotsb a_d! (n-|\ba|)!}$ be the multinomial
1286: coefficient.
1287: This gives a system of weights for $\calA= n\Delta\cap\Z^d$.
1288: Then Krasauskas's toric B\'ezier
1289: patch for $n\Delta_d$ has blending functions
1290: \[
1291: \beta_\ba\ :=\ {\textstyle \binom{n}{\ba}}
1292: (n-{\textstyle \sum_i x_i})^{n-\sum_i a_i}\,
1293: \prod_{i=1}^d x_i^{a_i}\,.
1294: \]
1295: If we substitute $ny_i$ for $x_i$ and then remove
1296: the common factors of $n^n$, we recover the classical Bernstein
1297: polynomials~\cite[\S 4.1]{Fa97}. \hfill\QED
1298: \end{example}
1299: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1300:
1301: \begin{example}
1302: The \Blue{{\it B\'ezier simploids}}~\cite{dRGHM} are toric patches based on
1303: products of B\'ezier simplices.
1304: Suppose that $d_1,\dotsc,d_m$ and $n_1,\dotsc,n_m$ are positive integers.
1305: Set $d:=d_1+\dotsb+d_m$.
1306: Write $\R^d$ as a direct sum $\R^{d_1}\oplus\dotsb\oplus\R^{d_m}$ and suppose that for each
1307: $i=1,\dotsc,m$, the scaled simplex $n_i\Delta_{d_i}$ lies in the $i$th summand $\R^{d_i}$.
1308: Consider the Minkowski sum
1309: \[
1310: \Delta\ :=\ n_1\Delta_{d_1} + n_2\Delta_{d_2} + \dotsb + n_m \Delta_{d_m}\ \subset\ \R^d\,,
1311: \]
1312: which is just the product of the simplices $n_i\Delta_{d_i}$.
1313:
1314: If $\bx_i:=(x_{i,1},\dotsc,x_{i,d_i})$ are coordinates on $\R^{d_i}$, then
1315: \[
1316: \Delta\ =\ \{ \bx\ =\ (\bx_1,\dotsc,\bx_m) \mid
1317: x_{i,j}\geq 0\ \quad\mbox{and}\quad\
1318: x_{i,1}+\dotsb+x_{i,d_i}\leq n_i\quad i=1,\dotsc,m\}\,.
1319: \]
1320: Let $\calA:=\Delta\cap\Z^d$ be the set of integer points of $\Delta$.
1321: Given an integer point $\ba=(\ba_1,\dotsc,\ba_m)\in\calA$,
1322: set $w_{\ba}:=\prod_{i=1}^m \binom{n_i}{\ba_i}$,
1323: the product of multinomial coefficients.
1324: This gives a system of weights, and let $Y_{\calA,w}$ be the resulting toric patch.
1325: Krasauskas's toric patch of shape $\Delta$ has blending
1326: functions
1327: \[
1328: \beta_{\ba}(\bx)\ :=\ \prod_{i=1}^m \beta_{\ba_i}(\bx)\,,
1329: \]
1330: where $\beta_{\ba_i}(\bx_i)$ for $\bx_i\in\R^{d_i}$ are the Bernstein polynomials.
1331: The resulting patch is called a \Blue{{\it B\'ezier simploid}} in~\cite{dRGHM}.
1332: \hfill\QED
1333: \end{example}
1334: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1335:
1336: We state our main result about linear precision for toric patches,
1337: which is a useful reformulation of linear precision for toric patches.
1338:
1339: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1340: \begin{theorem}\label{Th:toric_diff}
1341: Let $\calA\subset\Z^d$ be a primitive collection of exponent vectors,
1342: $w\in\R_>^{\calA}$ be a system of weights, and define the Laurent polynomial
1343: \[
1344: f\ =\ \Blue{f_{\calA,w}}\ :=\ \sum_{a\in\calA} w_{a} x^a\,.
1345: \]
1346: A toric patch of shape $(\calA,w)$ has rational linear precision if and only if the
1347: rational function $\psi_{\calA,w}\colon\C^d\rat\C^d$
1348: defined by
1349: %
1350: \begin{equation}\label{Eq:toric_differential}
1351: \frac{1}{f}\Bigl( x_1 \frac{\partial}{\partial x_1}f,\
1352: x_2 \frac{\partial}{\partial x_2}f,\ \dotsc,\
1353: x_d \frac{\partial}{\partial x_d}f \Bigr)
1354: \end{equation}
1355: %
1356: is a birational isomorphism.
1357: \end{theorem}
1358: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1359:
1360: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1361: \begin{remark}
1362: The map~\eqref{Eq:toric_differential} has an interesting reformulation in terms of toric
1363: derivatives.
1364: The $i$th \Blue{{\it toric derivative}} of a Laurent polynomial $f$ is
1365: $x_i\frac{\partial}{\partial x_i}f$.
1366: The toric differential $D_{\mbox{\scriptsize\rm torus}}f$ is the vector whose components are the
1367: toric derivatives of $f$.
1368: Thus the map $\psi_{\calA,w}$ is the logarithmic toric differential
1369: $D_{\mbox{\scriptsize\rm torus}} \log f$.
1370:
1371:
1372: In the proof below, we show that the map $\psi_{\calA,w}$ comes from a map $\C^d\;\rat\;\C\P^d$,
1373: %
1374: \begin{equation}\label{Eq:MAP}
1375: x\ \longmapsto\
1376: \Bigr[f\;:\; x_1 \frac{\partial}{\partial x_1}f\;:\;
1377: x_2 \frac{\partial}{\partial x_2}f\;:\ \dotsb\ :\;
1378: x_d \frac{\partial}{\partial x_d}f \bigr]
1379: \end{equation}
1380: %
1381: Let $F_{\calA,w}$ be the homogenization of $f$ with respect to a new variable $x_0$.
1382: Then
1383: \[
1384: \deg(F_{\calA,w})F_{\calA,w}\ =\ \sum_{i=0}^d x_i\frac{\partial}{\partial x_i}F_{\calA,w}\,.
1385: \]
1386: Thus, after homogenizing and a linear change of coordinates, the map~\eqref{Eq:MAP}
1387: is defined by the formula
1388: \[
1389: \Blue{D_{\mbox{\scriptsize\rm torus}}F_{\calA,w}}\ =\
1390: \Bigr[x_0 \frac{\partial}{\partial x_0}F_{\calA,w}\;:\;
1391: x_1 \frac{\partial}{\partial x_1}F_{\calA,w}\;:\ \dotsb\ :\
1392: x_d \frac{\partial}{\partial x_d}F_{\calA,w} \bigr]
1393: \makebox[.1cm][l]{\hspace{2cm}\QED}
1394: \]
1395: \end{remark}
1396: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1397:
1398: This last form of the map~\eqref{Eq:MAP} gives an appealing reformulation
1399: of Theorem~\ref{Th:toric_diff} which we feel is the most useful for
1400: further analysis of linear precision for toric patches.
1401:
1402: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1403: \begin{corollary}\label{Cor:Reformulation}
1404: The toric patch of shape $(\calA,w)$ has linear precision if and only if the map
1405: $D_{\mbox{\scriptsize\rm torus}}F_{\calA,w}\colon\C\P^{d}\rat\C\P^{d}$
1406: is a birational isomorphism.
1407: \end{corollary}
1408: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1409:
1410:
1411:
1412:
1413: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1414: \noindent{\it Proof of Theorem~$\ref{Th:toric_diff}$.}
1415: By Theorem~\ref{Th:LP=birational}, a toric patch of shape $(\calA,w)$
1416: has linear precision if and only if the complexified tautological projection
1417: \[
1418: \pi_\calA\ \colon\ Y_{\calA,w}\ \ \longrat\ \ \C\P^d
1419: \]
1420: is a birational isomorphism.
1421: Precomposing this with the defining parametrization
1422: of $Y_{\calA,w}$, $\varphi_{\calA_w}\colon(\C^\times)^d\hookrightarrow Y_{\calA,w}$
1423: (which is a rational map $\C^d\rat Y_{\calA,w}$), gives a map
1424: %
1425: \begin{equation}\label{Eq:composite}
1426: \begin{array}{rcccl}
1427: \C^d\ &\ \stackrel{\varphi_{\calA,w}}{\longrat}\ &Y_{\calA,w}&
1428: \ \stackrel{\pi_{\calA}}{\longrat}\ &\C\P^d \\
1429: x\ &\ \longmapsto & [w_\ba x^\ba\mid \ba\in\calA]& \longmapsto & \rule{0pt}{15pt}
1430: {\displaystyle \sum_{\ba\in\calA} w_\ba x^\ba (1,\ba)}\,.
1431: \end{array}
1432: \end{equation}
1433: %
1434:
1435: The initial (0th) coordinate of this map is
1436: \[
1437: \sum_{\ba\in\calA} w_\ba x^\ba
1438: \]
1439: which is the polynomial $f_{\calA,w}$.
1440: For $i>0$ the $i$th coordinate of this composition is
1441: \[
1442: \sum_{\ba\in\calA} a_i\cdot w_\ba x^\ba\ =\ x_i\frac{\partial}{\partial x_i}f_{\calA,w}\,.
1443: \]
1444: If we divide the composite map~\eqref{Eq:composite} by its 0th coordinate, $f_{\calA,w}$,
1445: we obtain the map $\psi_{\calA,w}$.
1446: This map is birational if and only if $\pi_\calA$ is birational, as $\varphi_{\calA,w}$ is
1447: birational.
1448: \hfill\QED
1449:
1450: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1451: \begin{example}
1452: Let $\calA:=\{0,1,\dotsc,n\}\subset\Z$ as in Example \ref{Ex:TBC}.
1453: Let $w\in\R_>^{n+1}$ be any set of weights.
1454: Then the polynomial $f = f_{\calA,w}$ is
1455: \[
1456: f(x) \ = \ w_0 + w_1x + \dotsm + w_nx^n\,.
1457: \]
1458:
1459: Suppose that this toric patch has rational linear precision.
1460: By Theorem~\ref{Th:toric_diff}, the logarithmic toric differential
1461: $xf'(x)/f(x)$ is a birational isomorphism.
1462: Thus the fraction $xf'(x)/f(x)$ reduces to a quotient of linear
1463: polynomials and so $f(x)$ is necessarily a pure power, say $a(x+\alpha)^n$
1464: with $\alpha\neq 0$.
1465:
1466: We conclude that the weights are $w_i=\alpha^{n-i}\binom{n}{i}$, which are not quite the
1467: weights in the B\'ezier curve~\ref{Ex:BC}.
1468: If we rescale $x$ (which does not change $Y_{\calA,w}$), setting $y=x/\alpha$, then the
1469: weights become $|\alpha|^n\binom{n}{i}$.
1470: Removing the common factor of $|\alpha|^n$, reveals that this
1471: toric patch is the classical B\'ezier curve of Example~\ref{Ex:BC}.
1472: \hfill\QED
1473: \end{example}
1474:
1475: We close with two problems.\medskip
1476:
1477: \noindent{\bf Problem 1.}
1478: Classify the toric patches of dimension $d$ that have linear precision.\medskip
1479:
1480: The analysis of this section was recently used to classify toric surface patches ($d=2$)
1481: having linear precision~\cite{BRS}.
1482: It remains an open problem to understand how to tune a toric patch (moving the points
1483: $\calA$) to achieve linear precision.\medskip
1484:
1485: \noindent{\bf Problem 2.}
1486: Classify the toric patches of dimension $d$ that may be tuned to have linear
1487: precision.\medskip
1488:
1489:
1490: This is a strictly larger class of patches.
1491: For example, the classification of~\cite{BRS} shows if a toric surface patch has linear
1492: precision, then it is triangular or quadrilateral,
1493: but it is possible to tune a pentagonal patch to achieve linear precision.
1494:
1495: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1496: \begin{example}\label{Ex:Pentagon}
1497: We discuss Krasauskas's version~\cite{Kr06} of Kar{\v{c}}iauskas's pentagonal
1498: $M$-patch~\cite{Ka03}, which has linear precision.
1499: Let $\calA$ be the integer points in the
1500: pentagon~\includegraphics{figures/pent.eps} shown below,
1501: and let $\{w_\ba\mid \ba\in\calA\}$ be the indicated weights.
1502: \[
1503: \begin{picture}(125,120)(-10,-18)
1504: \put(17,11){\includegraphics{figures/pentagon.eps}}
1505: \put(-10,89){$(0,2)$} \put(40,89){$(1,2)$}
1506: \put(-12,44){$(0,1)$} \put(40,55){$(1,1)$}\put(92,44){$(2,1)$}
1507: \put(-10,-1){$(0,0)$} \put(40,-1){$(1,0)$}\put(87, -1){$(2,0)$}
1508: \put(10,-18){Lattice points $\calA$}
1509: \end{picture}
1510: \qquad\qquad\quad
1511: \begin{picture}(110,120)(-10,-18)
1512: \put(4,11){\includegraphics{figures/pent_weights.eps}}
1513: \put(2,78){$2$} \put(37,78){$2$}
1514: \put(2,43){$5$} \put(37,43){$7$} \put(72,43){$2$}
1515: \put(2, 8){$3$} \put(37, 8){$5$} \put(72, 8){$2$}
1516: \put(-10,-18){weights $\{w_\ba\mid\ba\in\calA\}$}
1517: \end{picture}
1518: \]
1519: These define a toric patch $X_{\calA,w}$ and translated toric variety $Y_{\calA,w}$.
1520: This subvariety of $\P^7$ has degree 7 and is defined by 14 quadratic equations, all of
1521: the form
1522: \[
1523: y_\ba y_\bb-y_\bc y_\bd\qquad
1524: \mbox{where}\quad \ba,\bb,\bc,\bd\in\calA
1525: \quad\mbox{with}\quad \ba+\bb=\bc+\bd\,.
1526: \]
1527:
1528: This has a monomial parametrization
1529: \[
1530: \varphi_{\calA,w}\ \colon\ (s,t)\ \longmapsto\
1531: [s^{a_1}t^{a_2}\,\mid\, \ba=(a_1,a_2)\in\calA]\,.
1532: \]
1533: Its composition with the tautological projection $\pi_\calA$ gives a
1534: map $g\colon\P^2\rat\P^2$ of degree 4, with
1535: 3 base points at $(-1,-1)$, $(-1,-\frac{1}{2})$, and $(-\frac{1}{2},-1)$
1536:
1537: We tune this patch (or rather the projection $\pi_\calA$) by moving the
1538: non-extreme points of $\calA$ within the pentagon.
1539: Let $\calB$ consist of the vertices of the pentagon, together with the
1540: three non-vertices
1541: \[
1542: \begin{picture}(125,120)(-10,-20)
1543: \put(17,11){\includegraphics{figures/pent_new.eps}}
1544: \put(-10,89){$(0,2)$} \put(40,89){$(1,2)$}
1545: \put(-14,51){$(0,\frac{6}{5})$}
1546: \put(30,37){$(\frac{8}{7},\frac{8}{7})$}\put(92,45){$(2,1)$}
1547: \put(-10,-2){$(0,0)$} \put(47,-2){$(\frac{6}{5},0)$}\put(85,-2){$(2,0)$}
1548: \put(10,-20){Lattice points $\calB$}
1549: \end{picture}
1550: \]
1551: The composition of $\varphi_{\calA,w}$ with this new tautological projection $\pi_\calB$
1552: gives the rational map
1553: \[
1554: (s,t)\ \longmapsto\
1555: \Bigl( \frac{2s(2s+t+3)}{(s+1)(2s+2t+3)}\,,\
1556: \frac{2t(s+2t+3)}{(t+1)(2s+2t+3)} \Bigr)\,.
1557: \]
1558: This is given by quadrics and has 3 base points $(-1,-1)$, $(0,-\frac{3}{2})$,
1559: and $(-\frac{3}{2},0)$,
1560: and so it has degree $1=2\cdot 2-3$, and is therefore a birational isomorphism.
1561: By Theorem~\ref{Th:LP=birational}, the tuned toric patch
1562: (with the lattice points $\calA$ replaced by $\calB$) has linear precision.
1563:
1564: Thus there is a rational parametrization
1565: $\beta\colon\includegraphics[height=10pt]{figures/pent.eps}\to X_{\calA,w}$
1566: whose composition with the tautological projection $\pi_\calB$ is the identity map.
1567: Indeed, Krasauskas~\cite{Kr06} gives the following modification of
1568: the toric B\'ezier functions and shows that the map
1569: $\pi_\calB\circ\beta$ is the identity.
1570: %
1571: \begin{equation}\label{Eq:NTBF}
1572: \begin{array}{rcl}
1573: \beta_{\kzwei{0}{0}} &:=& 3(2-s)^2(3-s-t)^3(2-t)^2 \\\rule{0pt}{14pt}
1574: \beta_{\kzwei{1}{0}} &:=& 5(2-s)(3-s-t)^2(2-t)^2s(3-s-\tfrac{t}{2}) \\\rule{0pt}{14pt}
1575: \beta_{\kzwei{0}{1}} &:=& 5(3-\tfrac{s}{2}-t)t(2-s)^2(3-s-t)^2(2-t) \\\rule{0pt}{14pt}
1576: \beta_{\kzwei{1}{1}} &:=& 7(3-\tfrac{s}{2}-t)t(2-t)(3-s-t)(2-s)s(3-s-\tfrac{t}{2})
1577: \\\rule{0pt}{14pt}
1578: \beta_{\kzwei{2}{0}} &:=& 2(3-s-t)(2-t)^2s^2(3-s-\tfrac{t}{2})^2 \\\rule{0pt}{14pt}
1579: \beta_{\kzwei{0}{2}} &:=& 2(3-\tfrac{s}{2}-t)^2t^2(2-s)^2s(3-s-t) \\\rule{0pt}{14pt}
1580: \beta_{\kzwei{2}{1}} &:=& 2(3-\tfrac{s}{2}-t)t(2-t)s^2(3-s-\tfrac{t}{2})^2 \\\rule{0pt}{14pt}
1581: \beta_{\kzwei{1}{2}} &:=& 2(3-\tfrac{s}{2}-t)^2t^2(2-s)s(3-s-\tfrac{t}{2})
1582: \end{array}
1583: \end{equation}
1584: %
1585: This is a modification of the toric B\'ezier functions for the pentagon.
1586: If, in the definition~\eqref{Eq:toric-Bezier}, we replace the form $s$
1587: by $s(3-s-\tfrac{t}{2})$ and the form $t$ by
1588: $(3-\tfrac{s}{2}-t)t$, then we get these modified blending
1589: functions~\eqref{Eq:NTBF}. \smallskip
1590:
1591:
1592: Let us consider the geometry of linear precision for this tuned patch in the spirit of
1593: Section~\ref{S:Geometry}.
1594: The projection $\pi_\calB$ is defined by the three linear forms.
1595: %
1596: \begin{eqnarray*}
1597: X&:=& \tfrac{6}{5}y_{\kzwei{1}{0}}\;+\;2y_{\kzwei{2}{0}}\;+\;
1598: \tfrac{8}{7}y_{\kzwei{1}{1}}\;+\;2y_{\kzwei{2}{1}}\;+\;y_{\kzwei{1}{2}}\\
1599: Y&:=& \tfrac{6}{5}y_{\kzwei{0}{1}}\;+\;2y_{\kzwei{0}{2}}\;+\;
1600: \tfrac{8}{7}y_{\kzwei{1}{1}}\;+\;2y_{\kzwei{1}{2}}\;+\;y_{\kzwei{2}{1}}\\
1601: Z&:=&y_{\kzwei{0}{0}}\;+\;y_{\kzwei{1}{0}}\;+\;y_{\kzwei{0}{1}}\;+\;y_{\kzwei{1}{1}}\;+\;
1602: y_{\kzwei{2}{0}}\;+\;y_{\kzwei{0}{2}}\;+\;y_{\kzwei{2}{1}}\;+\;y_{\kzwei{1}{2}}
1603: \end{eqnarray*}
1604: %
1605: Its center $E_\calB$ is defined by the vanishing of these three forms.
1606: A general linear subspace $L$ of codimension 2 containing $E_\calB$ is defined by
1607: equations of the form
1608: \[
1609: X\ =\ x Z\qquad\mbox{and}\qquad Y\ =\ yZ\,.
1610: \]
1611: This subspace meets $Y_{\calA,w}$ in four points.
1612: One point lies outside of $E_\calB$ while three points lie in the center $E_\calB$.
1613: We give the three points.
1614: \[
1615: \begin{picture}(110,120)(0,-18)
1616: \put(4,11){\includegraphics{figures/pent_weights.eps}}
1617: \put(2,78){$3$} \put(37,78){$0$}
1618: \put(-7,43){$-5$} \put(37,43){$0$} \put(72,43){$0$}
1619: \put(2, 8){$2$} \put(37, 8){$0$} \put(72, 8){$0$}
1620: \put(4,-18){multiplicity 1}
1621: \end{picture}
1622: \qquad
1623: \begin{picture}(110,120)(0,-18)
1624: \put(4,11){\includegraphics{figures/pent_weights.2.eps}}
1625: \put(2,78){$0$} \put(37,78){$0$}
1626: \put(2,43){$0$} \put(37,43){$0$} \put(72,43){$0$}
1627: \put(2, 8){$2$} \put(28, 8){$-5$} \put(72, 8){$3$}
1628: \put(4,-18){multiplicity 1}
1629: \end{picture}
1630: \qquad
1631: \begin{picture}(110,120)(0,-18)
1632: \put(4,11){\includegraphics{figures/pent_weights.3.eps}}
1633: \put(2,78){$2$} \put(28,78){$-2$}
1634: \put(-7,43){$-5$} \put(37,43){$7$} \put(63,43){$-2$}
1635: \put(2, 8){$3$} \put(28, 8){$-5$} \put(72, 8){$2$}
1636: \put(4,-18){multiplicity 4}
1637: \end{picture}
1638: \]
1639: We write the coordinate $y_{\ba}$ of a point $y\in\R\P^{\calA}$ at the
1640: position of the point $\ba$.\hfill\QED
1641: \end{example}
1642: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1643:
1644:
1645: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1646: %
1647: %
1648: %
1649: \section{Iterative proportional fitting for toric patches}
1650:
1651: The toric patch $X_{\calA,w}$ from geometric modeling appears naturally in algebraic
1652: statistics in the form of a toric model, which leads to a dictionary between
1653: the subjects.
1654: We show that a toric patch has rational linear precision if and only if it has maximum
1655: likelihood degree 1 as a statistical model.
1656: As a consequence, we present a new family
1657: of toric patches with rational linear precision.
1658: Finally, we propose the iterative proportional fitting algorithm from
1659: statistics as a tool to compute the unique reparametrization of a toric patch having
1660: linear precision.
1661:
1662:
1663: In algebraic statistics, the image of $\R^{d}_{>}$ under the map
1664: $\varphi_{\calA,w}$~\eqref{Eq:monom_parametrize}
1665: is known as a \Blue{{\it toric model}}~\cite[\S 1.2]{PS05}.
1666: They are more commonly called \Blue{{\it log-linear models}}, as the logarithms of the
1667: coordinates of $\varphi_{\calA,w}$ are linear functions in the logarithms of the
1668: coordinates of $\R^{d}_{>}$, or \Blue{{\it discrete exponential families}}, as
1669: the coordinates of $\varphi_{\calA,w}$ are exponentials of the coordinates of $\R^d$, which
1670: are themselves logarithms of the coordinates of $\R^{d}_{>}$.
1671:
1672: We identify the non-negative orthant $\R\P^\calA_{\geq}$ with the
1673: probability simplex
1674: \[
1675: \Blue{\Delta^\calA}\ :=\
1676: \{y\in\R^\calA_{\geq0}\mid |y|:=\sum_{a\in\calA} y_a=1\}\,,
1677: \]
1678: and so we may also regard $X_{\calA,w}$ as a subvariety of $\Delta^\calA$.
1679:
1680: The tautological map $\pi_{\calA}$~\eqref{Eq:Taut_proj} appears in statistics.
1681: Given (normalized) data $q \in \Delta^{\calA}$, the problem of
1682: \Blue{{\it maximum likelihood estimation}}
1683: is to find the toric model parameters (a point in $\R^{d}_{>}$) that best explain the
1684: data $q$.
1685: By Lemma 4 in \cite{DR72}, the maximum likelihood estimate for
1686: the log-linear model $X_{\calA,w}$ is the unique point $p \in X_{\calA,w}$ such that
1687: \[
1688: \pi_{\calA}(p)\ =\ \pi_{\calA}(q)
1689: \quad \mbox{so that}\quad p\ =\ \pi^{-1}_{\calA}(\pi_{\calA}(q)).
1690: \]
1691: Thus inverting the tautological projection is necessary for maximum likelihood
1692: estimation.
1693:
1694: Catanese, {\it et.~al.}~\cite{CHKS} defined the \Blue{{\it maximum likelihood degree}}
1695: of the model $X_{\calA,w}$ to be the degree of $\pi_{\calA}^{-1}$, as an
1696: algebraic function.
1697: Equivalently, this is the number of complex solutions to the critical equations of the
1698: likelihood function, which is the degree of the tautological map $\pi_{\calA}$ from
1699: the Zariski closure $Y_{\calA,w}$ of $X_{\calA,w}$ to $\C\P^d$.
1700: By definition, a toric patch has rational linear precision if $\pi_{\calA}^{-1}$ is
1701: a rational function (an algebraic function of degree 1).
1702:
1703: \begin{proposition}
1704: A toric patch $X_{\calA, w}$ has rational linear precision if and only if $X_{\calA, w}$
1705: has maximum likelihood degree $1$.
1706: \end{proposition}
1707:
1708: An important family of toric models called \Blue{{\it decomposable graphical models}}
1709: are known to have maximum likelihood degree 1~\cite[p.~91]{Lauritzen}.
1710: Therefore, these models, which are not in general B\'ezier simploids, have linear
1711: precision.
1712: They typically have large $d>2$ dimension.
1713:
1714: Darroch and Ratcliff~\cite{DR72} introduced the numerical algorithm of iterative
1715: proportional fitting, also known as generalized iterative scaling,
1716: for computing the inverse of the tautological projection $\pi_{\calA}^{-1}$.
1717: Their algorithm requires that the data $\calA$ be in a normal, homogeneous form.
1718:
1719: Observe first that the toric patch $X_{\calA,w}$ does not change if we translate
1720: all elements of $\calA$ by a fixed vector $\bb$, ($\ba\mapsto \ba+\bb$),
1721: so we may assume that $\calA$ lies in the positive orthant $\R^d_>$.
1722: Scaling the exponent vectors in $\calA$ by a fixed positive scalar $t\in\R_>$ also does
1723: not change $X_{\calA,w}$ as $x\mapsto x^t$ is a homeomorphism of $\R_>$ that
1724: extends to a homeomorphism of $\R^d_>$.
1725: Thus we may assume that $\calA$ lies in the standard simplex $\Delta_d$ in $\R^d$:
1726: \[
1727: \Delta_d\ =\ \{ x\in\R^d\mid x\geq 0\quad\mbox{and}\quad |x|\leq 1\}\,.
1728: \]
1729: Lastly, we lift this to the probability simplex $\Delta^+_d\in\R^{d+1}_{\geq}$,
1730: \[
1731: \Blue{\Delta^+_d}\ :=\
1732: \{ y\in \R^{1+d}\mid y_i\geq 0\quad\mbox{and}\quad |y|=1\}\,,
1733: \]
1734: by
1735: \[
1736: \calA\ni \ba \ \longmapsto\ \ba^+:=(1{-}|\ba|,\,\ba) \in \Delta^+_d\,.
1737: \]
1738: Since for $t\in\R_>$ and $x\in\R^d_>$,
1739: \[
1740: (t,tx)^{\ba^+}\ =\ t^{1-|\ba|}(tx)^\ba\ =\
1741: t x^\ba\,,
1742: \]
1743: we see that replacing $\calA$ by its lifted version $\calA^+$ also does not change
1744: $X_{\calA,w}$.
1745: We remark that this is just a way to homogenize the data for our problem.
1746:
1747: Since these are affine transformations, the property of
1748: affine invariance for patches shows that the affine map that is the composition
1749: of these transformations intertwines the original tautological projection with the new one.
1750:
1751: We describe the algorithm of \Blue{{\it iterative proportional fitting}}, which is Theorem~1
1752: in~\cite{DR72}.
1753:
1754: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1755: \begin{proposition}\label{P:IPF}
1756: Suppose that $\calA\subset\Delta_d$ and $y\in\mbox{conv}(\calA)$.
1757: Then the sequence
1758: \[
1759: \{ \bp^{(n)}\;\mid\; n=0,1,2\dotsc\} %\ \subset\ \Delta^\calA\,,
1760: \]
1761: whose $\ba$-coordinates are defined by $p_\ba^{(0)}:=w_\ba$ and, for
1762: $n\geq 0$,
1763: \[
1764: p_\ba^{(n+1)}\ :=\ p_\ba^{(n)}\cdot \frac{y^\ba}{(\pi_{\calA}(\bp^{(n)}))^\ba}\,,
1765: \]
1766: converges to the unique point $\bp\in X_{\calA,w}$ such that
1767: $\pi_{\calA}(\bp)=\pi_{\calA}(y)$.
1768: \end{proposition}
1769: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1770:
1771: We remark that if $\calA$ is not homogenized then to compute $\pi_{\calA}^{-1}(y)$ for
1772: $y\in \mbox{conv}(\calA)$, we first put
1773: $\calA$ into homogeneous form $\calA^+$ using an affine map $\psi$, and then
1774: use iterative proportional fitting to compute
1775: $\pi_{\calA^{+}}^{-1}(\psi(y))=\pi_{\calA}^{-1}(y)$.
1776: We also call this modification of the algorithm of Proposition~\ref{P:IPF}
1777: iterative proportional fitting.
1778: Thus iterative proportional fitting computes the inverse image of the tautological
1779: projection, which, by Theorem~\ref{Th:LP} gives the unique parametrization of
1780: $X_{\calA,w}$ having linear precision.
1781:
1782: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1783: \begin{corollary}
1784: Iterative proportional fitting computes the unique parametrization of a toric patch having
1785: linear precision.
1786: \end{corollary}
1787: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1788:
1789: Little is known about the covergence of iterative proportional fitting.
1790: We think that it is an interesting question to investigate this convergence, particularly
1791: how it may relate to maximum likelihood degree.
1792:
1793: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1794: \bibliographystyle{amsplain}
1795: \bibliography{bibl}
1796:
1797: \end{document}
1798:
1799:
1800: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1801: \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
1802: \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
1803: % \MRhref is called by the amsart/book/proc definition of \MR.
1804: \providecommand{\MRhref}[2]{%
1805: \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
1806: }
1807: \providecommand{\href}[2]{#2}
1808: \begin{thebibliography}{10}
1809:
1810: \bibitem{CHKS}
1811: Fabrizio Catanese, Serkan Ho{\c{s}}ten, Amit Khetan, and Bernd Sturmfels,
1812: \emph{The maximum likelihood degree}, Amer. J. Math. \textbf{128} (2006),
1813: no.~3, 671--697.
1814:
1815: \bibitem{CG72}
1816: C.~Herbert Clemens and Phillip~A. Griffiths, \emph{The intermediate {J}acobian
1817: of the cubic threefold}, Ann. of Math. (2) \textbf{95} (1972), 281--356.
1818:
1819: \bibitem{CLO}
1820: D.~Cox, J.~Little, and D.~O'Shea, \emph{Ideals, varieties, algorithms: An
1821: introduction to computational algebraic geometry and commutative algebra},
1822: UTM, Springer-Verlag, New York, 1992.
1823:
1824: \bibitem{Cox03}
1825: David Cox, \emph{What is a toric variety?}, Topics in algebraic geometry and
1826: geometric modeling, Contemp. Math., vol. 334, Amer. Math. Soc., Providence,
1827: RI, 2003, pp.~203--223.
1828:
1829: \bibitem{DR72}
1830: J.~N. Darroch and D.~Ratcliff, \emph{Generalized iterative scaling for
1831: log-linear models}, Ann. Math. Statist. \textbf{43} (1972), 1470--1480.
1832:
1833: \bibitem{dRGHM}
1834: T.~DeRose, R.~Goldman, H.~Hagen, and S.~Mann, \emph{Functional composition
1835: algorithms via blossoming}, ACM Trans. on Graphics \textbf{12} (1993),
1836: 113--135.
1837:
1838: \bibitem{dR91}
1839: Tony~D. DeRose, \emph{Rational {B}\'ezier curves and surfaces on projective
1840: domains}, NURBS for curve and surface design (Tempe, AZ, 1990), SIAM,
1841: Philadelphia, PA, 1991, pp.~35--45.
1842:
1843: \bibitem{Fa97}
1844: Gerald Farin, \emph{Curves and surfaces for computer-aided geometric design},
1845: Computer Science and Scientific Computing, Academic Press Inc., San Diego,
1846: CA, 1997.
1847:
1848: \bibitem{Fu93}
1849: William Fulton, \emph{Introduction to toric varieties}, Annals of Mathematics
1850: Studies, vol. 131, Princeton University Press, Princeton, NJ, 1993, The
1851: William H. Roever Lectures in Geometry.
1852:
1853: \bibitem{BRS}
1854: Hans-Christian Graf~van Bothmer, Kristian Ranestad, and Frank Sottile, \emph{Linear
1855: precision for toric surface patches}, 2007, in preparation.
1856:
1857: \bibitem{Ka03}
1858: K{\c{e}}stutis Kar{\v{c}}iauskas, \emph{Rational {$M$}-patches and
1859: tensor-border patches}, Topics in algebraic geometry and geometric modeling,
1860: Contemp. Math., vol. 334, Amer. Math. Soc., Providence, RI, 2003,
1861: pp.~101--128.
1862:
1863: \bibitem{KK00}
1864: K{\c{e}}stutis Kar{\v{c}}iauskas and Rimvydas Krasauskas, \emph{Comparison of
1865: different multisided patches using algebraic geometry}, Curve and Surface
1866: Design: Saint-Malo 1999 (P.-J. Laurent, P.~Sablonniere, and L.L. Schumaker,
1867: eds.), Vanderbilt University Press, Nashville, 2000, pp.~163--172.
1868:
1869: \bibitem{KR02}
1870: Rimvydas Krasauskas, \emph{Toric surface patches}, Adv. Comput. Math.
1871: \textbf{17} (2002), no.~1-2, 89--133, Advances in geometrical algorithms and
1872: representations.
1873:
1874: \bibitem{Kr06}
1875: \bysame, \emph{B\'ezier patches on almost toric surfaces}, Algebraic geometry
1876: and geometric modeling, Math. Vis., Springer, Berlin, 2006, pp.~135--150.
1877:
1878: \bibitem{Lauritzen}
1879: Steffen~L. Lauritzen, \emph{Graphical models}, Oxford Statistical Science
1880: Series, vol.~17, The Clarendon Press Oxford University Press, New York, 1996,
1881: Oxford Science Publications.
1882:
1883: \bibitem{LdR89}
1884: Charles~T. Loop and Tony~D. DeRose, \emph{A multisided generalization of
1885: {B}{\'e}zier surfaces}, ACM Trans. Graph. \textbf{8} (1989), no.~3, 204--234.
1886:
1887: \bibitem{PS05}
1888: Lior Pachter and Bernd Sturmfels (eds.), \emph{Algebraic statistics for
1889: computational biology}, Cambridge University Press, New York, 2005.
1890:
1891: \bibitem{So03}
1892: Frank Sottile, \emph{Toric ideals, real toric varieties, and the moment map},
1893: Topics in algebraic geometry and geometric modeling, Contemp. Math., vol.
1894: 334, Amer. Math. Soc., Providence, RI, 2003, pp.~225--240.
1895:
1896: \bibitem{GBCP}
1897: Bernd Sturmfels, \emph{Gr\"obner bases and convex polytopes}, American
1898: Mathematical Society, Providence, RI, 1996.
1899:
1900: \bibitem{Wa75}
1901: Eugene~L. Wachspress, \emph{A rational finite element basis}, Academic Press,
1902: Inc. [A subsidiary of Harcourt Brace Jovanovich, Publishers], New York, 1975,
1903: Mathematics in Science and Engineering, Vol. 114.
1904:
1905: \bibitem{Wa92}
1906: Joe Warren, \emph{Creating multisided rational {B}{\'e}zier surfaces using base
1907: points}, ACM Trans. Graph. \textbf{11} (1992), no.~2, 127--139.
1908:
1909: \bibitem{Wa96}
1910: \bysame, \emph{Barycentric coordinates for convex polytopes}, Adv. Comput.
1911: Math. \textbf{6} (1996), no.~2, 97--108 (1997).
1912:
1913: \end{thebibliography}
1914:
1915: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1916:
1917: \end{document}
1918:
1919: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1920: \bibliographystyle{amsplain}
1921: \bibliography{bibl}
1922:
1923: \end{document}
1924:
1925: