betaFeedback

WorksMathematics

Rational barycentric coordinates and low degree adjoints

Abstract

We study rational generalized barycentric coordinates (rational GBCs) on polytopes. We recall their essential theory and clarify some statements from the literature. We study polytopes that admit GBCs with low degree denominators (and hence, low degree (affine) adjoint polynomials). We give a full classification of polytopes that admit polynomial barycentric coordinates (i.e. constant adjoints).

Introduction

Given a convex polytope P⊆RdP \subseteq\mathbb{R}^{d}, every point x∈Px \in P can be expressed as a convex combination of the polytopes vertices. If PP is a simplex, then this representation is unique and the convex coefficients are known as the barycentric coordinates of xx. In contrast, if PP has more than d+1d+1 vertices, there is no longer a unique such representation, but there are many. This motivates the following definition: a system of generalized barycentric coordinates on PP (GBCs for short) consists of a family of maps α={αv:P→R}\boldsymbol{\alpha} = \{\alpha_v : P \to\mathbb{R}\}, one per vertex vv of PP, so that for all x∈Px \in P holds

(positivity)αv(x)≥0,\text{(positivity)} \qquad\alpha_v(x) \geq0,
(normalization)∑vαv(x)=1,\text{(normalization)} \qquad\sum_v \alpha_v(x) = 1,
(precision)∑vαv(x)v=x.\text{(precision)} \qquad\sum_v \alpha_v(x)v = x.

All sums range over the vertices vv of PP. In other words, GBCs make a consistent choice of convex representation across all points of the polytope.

GBCs provide a solution to the ubiquitous problem of interpolating data given only at the vertices of a polytope, as it appears, for example, in finite element analysis, computer graphics or geometric modeling. An overview of different well-studied GBCs and some applications can be found in [5]. For many of these applications it is advantageous to use GBCs that are easy to represent and evaluate. This constraint naturally favors algebraic over transcendental expressions and motivates the study of polynomial and rational GBCs.

In this paper we study such polynomial and rational GBCs, with a particular focus on understanding their degrees.

Polynomial GBCs

Polynomial GBCs have αv∈R[x]\alpha_v \in\mathbb{R}[x] a multivariate polynomial in x∈Rdx \in\mathbb{R}^d. The classical barycentric coordinates on simplices are linear, and hence examples thereof. Other examples are obtained by taking products of polytopes and GBCs (cf. Theorem 2.4).

As pointed out by Wachspress already in 1975, most polytopes do not admit polynomial GBCs:

Theorem 1.1 ([9 Chapter 2, §2.1]). A quadrangle has polynomial GBCs if and only if it is a parallelogram.

The polynomial GBCs on a parallelogram are obtain as affine transformations of the product GBCs on squares. We likewise obtain polynomial GBCs on all affine transformations of products of simplices. In § 4.1 we prove that this list is complete:

Theorem 1.2. A polytope admits polynomial GBCs if and only if it is affinely equivalent to a product of simplices.

This resolves [4 Conjecture 4.8].

Rational GBCs

In his work on the finite element method Wachspress gave a construction for rational GBCs that exist for all convex polygons and simple 3-polytopes ([9] (i.e., αv=pv/q\alpha_v = p_v/q where pv,q∈R[x]p_v, q \in\mathbb{R}[x] are multivariate polynomials in x∈Rdx \in\mathbb{R}^d). Warren later extended the construction to polytopes of general dimension and combinatorics [10]. These coordinates came to be known as the Wachspress coordinates and are of great practical and theoretical interest (cf. Theorem 2.3).

We emphasize that the typical definition of rational GBCs (and a main reason for the nice theory) is that all coordinates pv/qp_v/q are over a common denominator qq. We will assume throughout that qq does not vanish on the interior of the polytope.

In § 2 and § 3 we recall and complement Warren’s theory of denominator degrees in rational GBCs. We provide corrections and completions to important points in Warren’s arguments.

In § 4 we transition to studying the denominator degrees of rational GBCs. We in particular focus on low degree denominators. In particular, the denominator degree is zero if and only if the GBCs are polynomial. In § 4.1 we apply this theory to provide the classification of polytopes with polynomial GBCs (Theorem 1.2). This amounts to the classification of polytopes with degree-zero adjoints.

On Warren's theory of rational GBCs

This section focuses primarily on the theory of rational GBCs developed by Warren, and in particular, on reproving degree bounds and the uniqueness of Wachspress coordinates. For this we essentially follow [11]. At times this section goes beyond the classical work and provides deeper insight into the structure of degrees and vanishing orders of general rational GBCs.

Remark 2.1. While [11] contains the essential ideas, it also contains a number of inaccuracies and gaps that have confused readers over the years. Some of them are not trivial to patch. A few stem from being imprecise about whether the setting is affine or projective, which leads to neglecting the consequences of parallelisms in the face structure of polytopes. In fact, there are true subtleties to the degree theory once we work in affine space (e.g. the denominator degree is not bounded from below by #facets − dim − 1). Other gaps appear in central proofs, such as local-global arguments for proving degree bounds. We include comments throughout the section (in footnotes and remarks) to point out whenever we repair or elaborate on [11].

Throughout, let P⊂RdP \subset\mathbb{R}^{d} be a dd-dimensional polytope (i.e., with non-empty interior) and αv={pv/q}\alpha_v = \{p_v/q\} rational GBCs on PP.

Affine precision

On rational GBCs normalization and precision take on the following form:

q=∑vpvandxq=∑vpvv.(1)q = \sum_v p_v \quad\text{and} \quad xq = \sum_v p_v v. \tag*{(1)}

The left sum states that qq is determined by the numerator polynomials. A standard argument shows that the pair (1) is equivalent to the single property called affine precision: for any affine linear form LL holds

L(x)q(x)=∑vL(v)pv(x).(2)L(x)q(x) = \sum_v L(v)p_v(x). \tag*{(2)}

Some authors define GBCs (rational or not) in terms of this single property.

Positivity

The third defining property, positivity, is used essentially in the following way: if F⊂PF \subset P is a facet and v∈Pv \in P is a vertex that does not lie in FF, then we necessarily have αv(x)=0\alpha_v(x) = 0 for all x∈Fx \in F, or equivalently, pv(x)=0p_v(x) = 0 for all x∈Fx \in F. In still other words, the facet linear form ℓF(x)\ell_F(x) (characterized by F=P∩{ℓF(x)=0}F = P \cap\{\ell_F(x) = 0\}) divides the numerator polynomial pvp_v. We conclude that the numerators of general rational GBCs factors as follows:

pv(x)=p^v(x)∏F∌vℓF(x),(3)p_v(x) = \widehat{p}_v(x) \prod_{F \not\ni v} \ell_F(x), \tag*{(3)}

where the product is over all facets F⊂PF \subset P that do not contain the vertex vv, and p^v\widehat{p}_v is some polynomial whose form we do not yet know.

Interestingly, (3) is the only consequence of positivity that we will make use of. Hence, at least for our purpose, one could replace positivity by (3).

Degrees

There are essentially three ways to measure complexity of GBCs by their degrees: numerator degrees, denominator degrees and rational degrees.

Numerator degrees have been the primary focus of Warren’s work. For example, consider GBCs on a simple polytope PP. Then from (3) follows that the degree of pvp_v must be at least m−dm-d, where mm is the number of facets of PP. There are three major insights from Warren’s theory:

  1. the lower bound of m−dm-d on the numerator degree applies to all polytopes, not just simple polytopes (we reprove this in Theorem 2.9);

  2. this lower bound is actually attained by some GBCs (see Theorem 2.3);

  3. the rational GBCs that attain this lower bound are essentially unique (we prove this in Theorem 3.10).

These unique degree minimizers are known as the Wachspress coordinates αW\alpha^{\mathrm{W}} of PP and have many interesting properties and use cases. We emphasize again that this extraordinarily nice theory is, to a decent extent, a consequence of the fact that all coordinates of GBCs share the same denominator (see also Theorem 2.10 and Theorem 2.11).

Denominator degrees, in contrast, are much less well-understood. The right sum in (1) gives

deg⁡(q)≤max⁡vdeg⁡(pv)−1(4)\deg(q) \le\max_{v}\deg(p_v)-1 \tag*{(4)}

The denominator of Wachspress coordinates is called the adjoint or adjoint polynomial adj⁡P\operatorname{adj}_{P} of PP. There are several ways to define this polynomial even without the Wachspress coordinates. This often happens in a projective setting, where one can show that the (homogeneous) degree is exactly deg⁡(adj⁡P)=m−d−1\deg(\operatorname{adj}_{P})=m-d-1 [10]. One of the oversights of Warren in [12] was that this does not carry over to the affine setting. For a general polytope, the degree of adj⁡P\operatorname{adj}_{P} can be much smaller than m−d−1m-d-1. And so, while for a generic polytope the Wachspress coordinates satisfy (4) with equality, this is not the case for general polytopes, not even if they are simple or simplicial. This “deficiency in degree”

drop⁡(P):=m−d−1−deg⁡(adj⁡P)\operatorname{drop}(P):=m-d-1-\deg(\operatorname{adj}_{P})

was studied in detail in [2], where it was found to subtly relate to parallelisms in the face structure of PP.

For general GBCs, no nice theory (as compared with the numerator degrees) has emerged so far. In contrast, we found polytopes PP with GBCs {pv/q}\{p_v/q\} so that (1) deg⁡(q)=deg⁡(adj⁡P)\deg(q)=\deg(\operatorname{adj}_{P}), where {pv/q}\{p_v/q\} are not the Wachspress coordinates. (2) deg⁡(q)<deg⁡(adj⁡P)\deg(q)<\deg(\operatorname{adj}_{P}).

In § 4 we introduce deg⁡(P)\deg(P) as the minimally possible denominator degree of rational GBCs on PP.

Lastly, one can study the rational degree deg⁡(αv):=deg⁡(pv)−deg⁡(q)\deg(\alpha_v):=\deg(p_v)-\deg(q). By (4) deg⁡(αv)≥1\deg(\alpha_v)\ge1. We did not pursue this route. We note that in the case of Wachspress coordinates, deg⁡(αvW)=m−d−deg⁡(adj⁡P)\deg(\alpha_v^W)=m-d-\deg(\operatorname{adj}_{P}) is precisely drop⁡(P)+1\operatorname{drop}(P)+1 and its study is therefore subsumed by [2].

Basic examples

The classical barycentric coordinates are defined on simplices. They are linear functions and are the unique GBCs with this property. We give a definition based on cone volumes:

Example 2.2 (barycentric coordinates). Let Δ\Delta be a dd-dimensional simplex. Let Fv⊂ΔF_v\subset\Delta be the facet opposite to vertex vv. For a point x∈Δx\in\Delta let Cv:=Fv∨xC_v:=F_v\vee x be the cone with base FvF_v and apex xx. Define αv:=vol⁡(Cv)/vol⁡(Δ)\alpha_v:=\operatorname{vol}(C_v)/\operatorname{vol}(\Delta). Since the cones CvC_v partition Δ\Delta, we have ∑vαv=1\sum_v\alpha_v=1. We also have αv(w)=1\alpha_v(w)=1 if v=wv=w and 00 otherwise. Lastly, it is easy to see that αv(x)\alpha_v(x) is linear in xx. It follows that these are the classical barycentric coordinates on Δ\Delta.

Interestingly, the Wachspress coordinates can be defined in terms of cone volumes as well, but they require the polar dual.

Example 2.3 (Wachspress coordinates). Given a polytope P⊂RdP\subset\mathbb{R}^{d} with x∈int⁡(P)x\in\operatorname{int}(P). Consider its polar dual at xx:

(P−x)∘:={y∈Rd∣⟨y,z−x⟩≤1 for all z∈P}.(P-x)^{\circ}:=\{y\in\mathbb{R}^{d}\mid\langle y,z-x\rangle\le1\text{ for all }z\in P\}.

Let Fv∘F_v^{\circ} be the facet of (P−x)∘(P-x)^{\circ} that is dual to the vertex v∈Pv\in P. Let CvC_v be the cone with base Fv∘F_v^{\circ} and apex at the origin. Define αv:=vol⁡(Cv)/vol⁡(P−x)∘\alpha_v:=\operatorname{vol}(C_v)/\operatorname{vol}(P-x)^{\circ}. These are the Wachspress coordinates of PP at xx. The fact that these are indeed GBCs is a consequence of Minkowski’s balancing condition. This geometric cone volume

construction goes back to Ju, Schaefer, Warren and Desbrun [6]. Evidently, they are rational. We do here not reverify the degree but point to the literature.

An important elementary operations to construct new GBCs from old ones is the product construction:

Example 2.4 (product coordinates). Given polytope Pi⊂RdiP_i \subset\mathbb{R}^{d_i} with i∈{1,2}i \in\{1,2\} with GBCs αi\alpha^i, one can define the product coordinates on P1×P2P_1 \times P_2. The vertices of P1×P2P_1 \times P_2 are of the form (v1,v2)(v_1,v_2), where vi∈Piv_i \in P_i. Then

(α1×α2)v1,v2:=αv11⋅αv22.(\alpha^1 \times\alpha^2)_{v_1,v_2} := \alpha^1_{v_1} \cdot\alpha^2_{v_2}.

One can verify that these do indeed give GBCs.

Taking the product of barycentric coordinates on simplices gives polynomial GBCs on products of simplices. As we will show in § 4.1, these are the only polytopes that admit polynomial GBCs. The simplest non-simplicial example is the square:

Example 2.5 (square).

P=[0,1]2=conv⁡{v00,v10,v11,v01}.P = [0,1]^2 = \operatorname{conv}\{v_{00},v_{10},v_{11},v_{01}\}.

Let ℓij\ell_{ij} be the linear form of the facet spanned by viv_i and vjv_j. Then

ℓ00,10=x2,ℓ10,11=1−x1,ℓ11,01=1−x2,ℓ01,00=x1.\ell_{00,10}=x_2,\quad\ell_{10,11}=1-x_1,\quad\ell_{11,01}=1-x_2,\quad\ell_{01,00}=x_1.

and

pv00=ℓ10,11ℓ11,01,pv10=ℓ11,01ℓ01,00,p_{v_{00}}=\ell_{10,11}\ell_{11,01},\quad p_{v_{10}}=\ell_{11,01}\ell_{01,00},
pv11=ℓ01,00ℓ00,10,pv01=ℓ00,10ℓ10,11.p_{v_{11}}=\ell_{01,00}\ell_{00,10},\quad p_{v_{01}}=\ell_{00,10}\ell_{10,11}.

and q=p00+p10+p11+p01=2q=p_{00}+p_{10}+p_{11}+p_{01}=2.

Vanishing orders and the face arrangement

A central tool for proving degree lower bounds for the numerators and denominators is to look at the multiplicities of their zeroes. Let oE(p)o_E(p) denote the vanishing order of the polynomial pp along a non-empty affine subspace EE. Clearly, deg⁡(p)≥oE(p)\deg(p) \ge o_E(p). For a face f⊆Pf \subseteq P we abbreviate of(p):=oaff⁡(f)(p)o_f(p):=o_{\operatorname{aff}(f)}(p).

Lemma 2.6. Given a δ\delta-face f⊆Pf \subseteq P and vertex v∈Pv \in P:

(i) if v∈fv \in f then of(pv)=of(p^v)o_f(p_v)=o_f(\hat{p}_v).

(ii) of(pv)≥of(q)o_f(p_v) \ge o_f(q) with equality if and only if v∈fv \in f.

Proof. Part (i): by the factorization (3) we have

ov(pv)=ov(p^v)+ov(∏F∌vℓF).o_v(p_v)=o_v(\hat{p}_v)+o_v\left(\prod_{F\not\ni v}\ell_F\right).

But the latter term vanishes since none of the linear forms ℓF\ell_F, F∌vF\not\ni v is zero on vv.

Part (ii): the limits of pv(x)/q(x)p_v(x)/q(x) as xx approaches the face ff from within PP exist and yield GBCs on ff. The limit towards a point in the relative interior of ff is zero if and only if v∉fv \notin f. □

The face arrangement F(P)\mathcal{F}(P) of a polytope PP is the arrangement of the affine subspaces generated by face spans (including of PP and the empty face) closed under intersection. The elements of the arrangement are called flats. The ones generated by faces are called facial flats. The others are called residual flats. Alternatively, a flat is facial if and only if it is spanned by the vertices v∈Pv \in P that it contains.

Face arrangements are only implicit in Warren’s treatment of the subject. They are more visible in his work on the adjoint polynomial [10]. The modern treatment of rational GBCs seems to favor this language, especially in the form of the residual arrangement [8].

Lemma 2.7. If E∈F(P)E \in\mathcal{F}(P) is a flat then

oE(q)≥min⁡v∉EoE(pv)−{1if E is a facial flat,0if E is a residual flat.o_{E}(q) \ge\min_{v \notin E} o_{E}(p_{v}) - \begin{cases} 1 & \text{if } E \text{ is a facial flat},\\ 0 & \text{if } E \text{ is a residual flat}. \end{cases}

Proof. Let E′⊂EE' \subset E be the affine span of the vertices v∈Pv \in P that lie in EE. Let LL be a generic affine linear form that vanishes on E′E'. Then

oE(L)+oE(q)=oE(Lq)=(∗)oE(∑vL(v)pv)=oE(∑v∉EL(v)pv)≥(∗∗)min⁡v∉EoE(pv).\begin{aligned} o_{E}(L) + o_{E}(q) &= o_{E}(Lq) \overset{(*)}{=} o_{E}\left(\sum_{v} L(v)p_{v}\right) = o_{E}\left(\sum_{v \notin E} L(v)p_{v}\right) \overset{(**)}{\ge} \min_{v \notin E} o_{E}(p_{v}). \end{aligned}

where in (∗)(*) we used affine precision, and (∗∗)(**) is the well-known inequality for the vanishing order of sums of polynomials. Since LL is linear, oE(L)≤1o_{E}(L) \le1. If E′E' is residual, then it is not spanned by the contained vertices, and so E′⊊EE' \subsetneq E. Since LL was chosen generic, it does not vanish on EE, and so oE(L)=0o_{E}(L) = 0. □

Facial flats

Given a flat E∈F(P)E \in\mathcal{F}(P), its excess ex⁡(E)\operatorname{ex}(E) counts how many more facet hyperplanes of PP contain EE than one would generically except from only knowing its dimension. More precisely, the number of facet hyperplanes that contain EE is codim⁡(E)+ex⁡(E)\operatorname{codim}(E) + \operatorname{ex}(E), where codim⁡(E):=d−dim⁡(E)\operatorname{codim}(E) := d - \dim(E).

Lemma 2.8. Given rational GBCs {pv/q}\{p_v/q\} on PP and a face f⊆Pf \subseteq P, it holds

of(q)≥ex⁡(f).o_f(q) \ge\operatorname{ex}(f).

Proof. We use induction on codim⁡(f)\operatorname{codim}(f). If codim⁡(f)=0\operatorname{codim}(f) = 0, then f=Pf = P and ex⁡(f)=0\operatorname{ex}(f) = 0, and the statement is immediate. Assume now codim⁡(f)≥1\operatorname{codim}(f) \ge1.

By Theorem 2.7 we have of(q)≥min⁡v∉fof(pv)−1o_f(q) \ge\min_{v \notin f} o_f(p_v) - 1 and so it suffices to show that of(pv)≥ex⁡(f)+1o_f(p_v) \ge\operatorname{ex}(f) + 1 for all v∉fv \notin f. Fix such a v∉fv \notin f. Let fv:=f∨vf_v := f \vee v be the minimal face of PP that contains both ff and vv. To bound of(pv)o_f(p_v) we consider two independent contributions to its value: first, the vanishing order of pvp_v on the larger face fv⊃ff_v \supset f; second, additional contributions from the facet linear forms ℓF\ell_F where FF contains ff but not vv (cf. the factorization (3)):

of(pv)≥ofv(pv)+#{facets of P that contain f but not v}.(5)o_f(p_v) \ge o_{f_v}(p_v) + \#\{\text{facets of } P \text{ that contain } f \text{ but not } v\}. \tag*{(5)}

The left term we bound using Theorem 2.6 (ii) and the induction hypothesis. The latter applies due to codim⁡(fv)<codim⁡(f)\operatorname{codim}(f_v) < \operatorname{codim}(f):

ofv(pv)=(ii)ofv(q)≥ex⁡(fv).o_{f_v}(p_v) \overset{(ii)}{=} o_{f_v}(q) \ge\operatorname{ex}(f_v).

The second term we bound as follows:

#{facets that contain f but not v}=#{facets that contain f but not fv}=(codim⁡(f)+ex⁡(f))−(codim⁡(fv)+ex⁡(fv))=(ex⁡(f)−ex⁡(fv))+(codim⁡(f)−codim⁡(fv))≥(ex⁡(f)−ex⁡(fv))+1,\begin{aligned} \#\{\text{facets that contain } f \text{ but not } v\} &= \#\{\text{facets that contain } f \text{ but not } f_v\}\\ &= \bigl(\operatorname{codim}(f) + \operatorname{ex}(f)\bigr) - \bigl(\operatorname{codim}(f_v) + \operatorname{ex}(f_v)\bigr)\\ &= \bigl(\operatorname{ex}(f) - \operatorname{ex}(f_v)\bigr) + \bigl(\operatorname{codim}(f) - \operatorname{codim}(f_v)\bigr)\\ &\ge\bigl(\operatorname{ex}(f) - \operatorname{ex}(f_v)\bigr) + 1, \end{aligned}

where in the first equality we use that fvf_v was chosen minimal; in the second equality we use that the facets that contain fvf_v are a subset of the facets that contain ff, that ff is contained in codim⁡(f)+ex⁡(f)\operatorname{codim}(f)+\operatorname{ex}(f) facets, and fvf_v is contained in codim⁡(fv)+ex⁡(fv)\operatorname{codim}(f_v)+\operatorname{ex}(f_v) facets.

Combining both estimations in (5) finishes the proof. □

We obtain our first major insight for numerator degrees:

Theorem 2.9. Let P⊂RdP \subset\mathbb{R}^d be a polytope with mm facets and {pv/q}\{p_v/q\} rational GBCs on PP. Then deg⁡(pv)≥m−d\deg(p_v) \ge m-d for all vertices vv.

Proof. Suppose that vv lies on d+kd+k facets, that is, ex⁡(v)=k\operatorname{ex}(v)=k. By Theorem 2.6 (i) and (ii), as well as Theorem 2.8, we have

deg⁡(p^v)≥ord⁡v(p^v)=(i)ord⁡v(pv)=(ii)ord⁡v(q)≥ex⁡(v)=k.\deg(\widehat{p}_v) \ge\operatorname{ord}_v(\widehat{p}_v) \overset{(i)}{=} \operatorname{ord}_v(p_v) \overset{(ii)}{=} \operatorname{ord}_v(q) \ge\operatorname{ex}(v)=k.

And hence

deg⁡(pv)=(2.3)deg⁡(p^v)+deg⁡(∏F∌vℓF)≥(∗)k+(m−(d+k))=m−d.\deg(p_v) \overset{(2.3)}{=} \deg(\widehat{p}_v)+\deg\left(\prod_{F\not\ni v}\ell_F\right) \overset{(*)}{\ge} k+(m-(d+k))=m-d.

□

The lower bound of Theorem 2.9 should be understood as a consequence of writing our rational GBCs pv/qp_v/q over a common denominator polynomial qq. Even though the Wachspress coordinates satisfy gcd⁡{pv,adj⁡P}=1\gcd\{p_v,\operatorname{adj}_P\}=1, individual pvp_v may still share factors with the adjoint, and canceling these factors would result in lower numerator degrees.

Example 2.10 (Pyramid). Given a polytope PP with GBCs α={pv/q}\alpha=\{p_v/q\}. Let P⋆P_\star be the pyramid over PP with apex ⋆\star at the origin. The following gives GBCs β\beta of P⋆P_\star:

βv(x,t)=tαv(xt),for vertices v∈P,\beta_v(x,t)=t\alpha_v\left(\frac{x}{t}\right), \qquad\text{for vertices }v\in P,
β⋆(x,t)=1−t.\beta_\star(x,t)=1-t.

One can verify that these are indeed rational GBCs of P⋆P_\star. Note that deg⁡(β⋆)=1\deg(\beta_\star)=1, independent of the combinatorics of P⋆P_\star. Hence, if m−d>1m-d>1, then by Theorem 2.9 factors in numerator and denominator must have canceled.

In fact, it is possible that pvp_v and adj⁡P\operatorname{adj}_P share factors at all vertices vv.

Example 2.11 (Joins). Let PP and QQ be polytopes with rational GBCs αP\alpha^P and αQ\alpha^Q. Each point of the join P∗QP\ast Q can be written as (1−t)x+ty(1-t)x+ty for some x∈Px\in P, y∈Qy\in Q and t∈[0,1]t\in[0,1]. The following defines GBCs β\beta on P∗QP\ast Q:

βv(x,y,t)=(1−t)αvP(x1−t),for all v∈P,\beta_v(x,y,t)=(1-t)\alpha^P_v\left(\frac{x}{1-t}\right), \qquad\text{for all }v\in P,
βw(x,y,t)=tαwQ(yt),for all w∈Q.\beta_w(x,y,t)=t\alpha^Q_w\left(\frac{y}{t}\right), \qquad\text{for all }w\in Q.

Again, one can verify that these are indeed rational GBCs of P⋆QP\star Q. Note that, as written here, the βv\beta_v only depend on xx, whereas the βw\beta_w depend only on yy. Hence, unless αP\alpha^P and βP\beta^P had constant denominator, writing β\beta over a common denominator requires adding shared factors at all vertices.

The following fact, that GBCs cannot vanish on facets, is of use later.

Lemma 2.12. Let F⊂PF\subset P be a facet, and assume either of the following:

(i) gcd⁡{q,pv}=1\gcd\{q,p_v\}=1 for some vertex v∈Pv\in P, or

(ii) gcd⁡{pv∣v∈P}=1\gcd\{p_v \mid v \in P\}=1,

Then oF(q)=oF(pw)=0o_F(q)=o_F(p_w)=0 for all vertices w∈Fw \in F.

Proof. Part (i): If oF(q)>0o_F(q)>0, then oF(pv)>0o_F(p_v)>0 by Theorem 2.6 (ii). But this means ℓF∣q\ell_F \mid q, as well as ℓF∣pv\ell_F \mid p_v, contradicting the coprimality assumption. Consequently, oF(q)=0o_F(q)=0. Via Theorem 2.6 (ii) also follows oF(pw)=0o_F(p_w)=0 for all w∈Fw \in F. Part (ii): As in the proof of (i), if oF(q)>0o_F(q)>0, then ℓF∣pv\ell_F \mid p_v for all vv. Therefore, ℓF∣gcd⁡{pv}\ell_F \mid\gcd\{p_v\}.

Results stating non-vanishing on faces of PP will be important later, are however hard to come by in this generality. Theorem 2.12 is among the strongest claims we can make under current assumptions. Even assuming gcd⁡{q,pv∣v∈V}=1\gcd\{q,p_v \mid v \in V\}=1, it is still possible that qq vanishes on a face of codimension two, or that pvp_v vanishes on a facet F∌vF \not\ni v to an order higher than one.

Residual flats

Lemma 2.13. If PP is a polytope with rational GBCs {pv/q}\{p_v/q\} and a residual flat EE, then oE(q)≥1o_E(q) \ge1. In other words, qq vanishes on the residual flats of PP.

In the literature this is often phrased as “qq vanishes on the residual arrangement of PP”.

Proof. Theorem 2.7 states oE(q)≥min⁡v∉EoE(pv)o_E(q) \ge\min_{v\notin E} o_E(p_v), thus it suffices to show oE(pv)≥1o_E(p_v) \ge1 for all v∉Ev \notin E. Fix such a v∉Ev \notin E. Since EE is cut out by facet hyperplanes, there is some facet F⊂PF \subset P whose span contains EE but not vv. Hence ℓF∣pv\ell_F \mid p_v by (3), and oE(pv)≥oE(ℓF)=1o_E(p_v) \ge o_E(\ell_F)=1.

The following conjecture is a natural common generalization of Theorem 2.13 and Theorem 2.8.

Conjecture 2.14. If E∈F(P)E \in\mathcal{F}(P) is a flat, then

oE(q)≥ex⁡(E)+{0if E is a facial flat1if E is a residual flato_E(q) \ge\operatorname{ex}(E) + \begin{cases} 0 & \text{if } E \text{ is a facial flat} \\ 1 & \text{if } E \text{ is a residual flat} \end{cases}

In fact, in the case of Wachspress coordinates, where the denominator is the adjoint polynomial adj⁡P\operatorname{adj}_P, these vanishing orders are exact [3]:

oE(adj⁡P)=ex⁡(E)+{0if E is a facial flat1if E is a residual flato_E(\operatorname{adj}_P) = \operatorname{ex}(E) + \begin{cases} 0 & \text{if } E \text{ is a facial flat} \\ 1 & \text{if } E \text{ is a residual flat} \end{cases}

One could expect to prove Theorem 2.14 by combining ideas from Theorem 2.8 and Theorem 2.13. The proof of the latter relies on the fact that we can estimate the vanishing orders of the pvp_v. But so far we know how to do this only if v∈Ev \in E. In fact, already an analogue of Theorem 2.13 for numerator polynomials fails for sufficiently complicated examples (see Theorem A.4). Analogous can be proven in special cases, for example, see Theorem 3.3.

Minimal numerator degree

In this section we focus specifically on rational GBCs {pv/q}\{p_v/q\} of minimal possible numerator degree, that is, deg⁡(pv)=m−d\deg(p_v)=m-d according to Theorem 2.9. The goal is to show that {pv/q}\{p_v/q\} must be the Wachspress coordinates. Again, we broadly follow Warren’s work [11], though we make corrections at several points in the proof.

It turns out that quite a number of our statements have a “per-vertex” version, that is, need to assume deg⁡(pv)=m−d\deg(p_v)=m-d only for a single vertex vv rather than all vertices at the same time. We make an effort to provide per-vertex versions whenever this is possible and reasonable given our setup. The following is a first example:

Corollary 3.1. Suppose deg⁡(pv)=m−d\deg(p_v)=m-d for some vertex v∈Pv\in P. Then

(i) gcd⁡{pv∣v∈P}=1\gcd\{p_v\mid v\in P\}=1.

(ii) oF(q)=0o_F(q)=0 for each facet F⊂PF\subset P.

Proof. To prove (i) suppose that r∣pwr\mid p_w for all vertices ww. Then r∣∑wpw=qr\mid\sum_w p_w=q and rr can be canceled from all pw/qp_w/q, yielding GBCs of lower numerator degrees. This contradicts Theorem 2.9 at vv.

To prove (ii) observe that if oF(q)>0o_F(q)>0, Theorem 2.6 (ii) implies oF(pw)>0o_F(p_w)>0 for all w∈Fw\in F. From (3) we also get oF(pw)>0o_F(p_w)>0 for all w∉Fw\notin F. In other words, ℓF∣pw\ell_F\mid p_w for all ww, contradicting (i). □

Numerators on residual flats

Recall p^v\hat{p}_v as define in (3).

Lemma 3.2. If deg⁡(pv)=m−d\deg(p_v)=m-d for some vertex v∈Pv\in P, then

deg⁡(p^v)=ov(p^v)=ov(pv)=ov(q)=ex⁡(v).(6)\deg(\hat{p}_v)=o_v(\hat{p}_v)=o_v(p_v)=o_v(q)=\operatorname{ex}(v). \tag*{(6)}

In particular, p^v\hat{p}_v is homogeneous at vv of degree ex⁡(v)\operatorname{ex}(v) and is (up to a factor) the homogeneous part of pvp_v of lowest degree.

This lemma is later strengthened to general faces (see Theorem 3.9), though that requires more preparation.

Proof of Theorem 3.2. From minimal degree and the factorization (3) we get

m−d=deg⁡(pv)=deg⁡(p^v)+#{facets that do not contain v}=deg⁡(p^v)+m−#{facets that do contain v}=deg⁡(p^v)+m−(d+ex⁡(v)),\begin{aligned} m-d &= \deg(p_v) \\ &= \deg(\hat{p}_v)+\#\{\text{facets that do not contain }v\} \\ &= \deg(\hat{p}_v)+m-\#\{\text{facets that do contain }v\} \\ &= \deg(\hat{p}_v)+m-(d+\operatorname{ex}(v)), \end{aligned}

which rearranges to (∗)(*) deg⁡(p^v)=ex⁡(v)\deg(\hat{p}_v)=\operatorname{ex}(v). Using also Theorem 2.6 (i) and Theorem 2.6 (ii), as well as Theorem 2.8 we get the following cyclic chain of inequalities:

ex⁡(v)=(∗)deg⁡(p^v)≥ov(p^v)=(i)ov(pv)=(ii)ov(q)≥2.8ex⁡(v).\operatorname{ex}(v)\overset{(*)}{=}\deg(\hat{p}_v)\ge o_v(\hat{p}_v)\overset{(i)}{=}o_v(p_v)\overset{(ii)}{=}o_v(q)\overset{2.8}{\ge}\operatorname{ex}(v).

Thus all terms must be equal, giving the claimed identities in (6). □

The following is an anlogue of Theorem 2.13 for numerator polynomials:

Lemma 3.3. Let E∈F(P)E\in\mathcal{F}(P) be a residual flat. If deg⁡(pv)=m−d\deg(p_v)=m-d for some vertex v∈Pv\in P, then

(i) oE(pv)≥1o_E(p_v) \ge1.

(ii) if v∈Ev \in E, then oE(p^v)≥1o_E(\hat{p}_v) \ge1.

Proof. Case v∉Ev \notin E: The flat EE is the intersection of some facet hyperplanes. Thus, v∉Fv \notin F for at least one such facets FF. We get ℓF∣pv\ell_F \mid p_v, and oE(pv)≥oE(ℓF)=1o_E(p_v) \ge o_E(\ell_F) = 1.

Case v∈Ev \in E: Normalization gives us

q=pv+∑w≠vpw(7)q = p_v + \sum_{w \ne v} p_w \tag*{(7)}

This identity must hold in each homogeneous degree (defined at vv), and so we may restrict to degree ex⁡(v)\operatorname{ex}(v). Write pex⁡p^{\operatorname{ex}} for the homogeneous part of degree ex⁡(v)\operatorname{ex}(v) of a polynomial pp. By Theorem 3.2 we have pvex⁡=p^vp_v^{\operatorname{ex}} = \hat{p}_v. Using Theorem 2.6 (ii), for w≠vw \ne v we have ov(pw)>ov(q)=ex⁡(v)o_v(p_w) > o_v(q) = \operatorname{ex}(v), and hence pwex⁡=0p_w^{\operatorname{ex}} = 0. In degree ex⁡(v)\operatorname{ex}(v) the identity (7) therefore reduces to qex⁡=p^vq^{\operatorname{ex}} = \hat{p}_v. Now restrict both sides to EE. Recall that restriction to homogeneous parts and restriction to affine subspaces (through vv) commutes:

p^v∣E=qex⁡∣E=(q∣E)ex⁡≡0\hat{p}_v|_E = q^{\operatorname{ex}}|_E = (q|_E)^{\operatorname{ex}} \equiv0

where q∣E=0q|_E = 0 by Theorem 2.13. We conclude oE(pv)=oE(p^v)≥1o_E(p_v) = o_E(\hat{p}_v) \ge1. □

Remark 3.4. Theorem 3.2 combined with Theorem 3.3 (ii) states that p^v\hat{p}_v is a homogeneous polynomial of minimal degree that vanishes on the residual flats of the tangent cone at vv. The reader familiar with the literature on adjoints might then recognize that p^v\hat{p}_v must be the adjoint of the tangent cone [3, 8]. This insight could shortcut our proof of uniqueness for minimal degree rational GBCs. Note however that the known proofs of uniqueness for the adjoint requires advanced tools, such as canonical forms. We emphasize that our goal is to reconstruct Warren’s elementary proof of uniqueness of GBCs without significant external input.

Restriction to facets

Later, we will need to perform induction on the dimension of the polytope by stepping from PP to its faces. We therefore need to understand how rational GBCs (in particular, their degrees) behave under restriction to faces.

If α={pv/q}\alpha= \{p_v/q\} are rational GBCs and F⊂PF \subset P is a facet, then the restriction α∣F\alpha|_F yields rational GBCs on FF:

αv∣F=pv∣Fq∣Ffor all v∈F.\alpha_{v|F} = \frac{p_v|_F}{q|_F} \qquad\text{for all } v \in F.

We emphasize that this representation is valid only because we have q∣F≠0q|_F \ne0 due to Theorem 3.1 (ii). This representation might however not be reduced, i.e., numerator and denominator may still share factors.

Lemma 3.5. Given a facet F⊂PF \subset P and a vertex v∈Fv \in F with deg⁡(pv)=m−d\deg(p_v) = m-d, the following polynomial divides both pv∣Fp_v|_F and q∣Fq|_F:

rF:=∏G⊂PG is a facet neither equal nor adjacent to FℓG∣F.r_F := \prod_{\substack{G \subset P \\ G\text{ is a facet neither equal nor adjacent to }F}} \ell_G|_F.

Observe that the shared factor rFr_F depends neither on the precise GBCs nor the vertex v∈Pv \in P. This will be important in Theorem 3.10.

Proof of Theorem 3.5. We show that each restricted linear factor ℓG∣F\ell_G|_F divides q∣Fq|_F and all pv∣Fp_v|_F. If GG is parallel to FF, then ℓG∣F\ell_G|_F is constant, and the claim is trivial. If GG is not parallel, then since it is not adjacent to FF, it defines a (d−2)(d-2)-dimensional residual flat EG:=aff⁡(G)∩aff⁡(F)E_G := \operatorname{aff}(G) \cap\operatorname{aff}(F). By Theorem 2.13 and Theorem 3.3 qq and pvp_v vanish on EGE_G. Since EGE_G is the zero locus of ℓG∣F\ell_{G|F} on aff⁡(F)\operatorname{aff}(F), the claim follows. □\square

We can then cancel the common factor rFr_F provided by Theorem 3.5 and define the polynomials pvF:=pv∣F/rFp_v^F := p_{v|F}/r_F and qF:=q∣F/rFq^F := q_{|F}/r_F. We call the representation

αv∣F=pvFqFfor all v∈F\alpha_{v|F} = \frac{p_v^F}{q^F} \quad\text{for all } v \in F

the reduced restriction of {pv/q}\{p_v/q\} on FF. From this definition alone it is not clear (and in general not true) that we removed all shared factors. It is however sufficient for preparing the facet induction argument.

Lemma 3.6. Given a facet F⊂PF \subset P and a vertex v∈Fv \in F with deg⁡(pv)=m−d\deg(p_v)=m-d, then deg⁡(pvF)=mF−(d−1)\deg(p_v^F)=m_F-(d-1), where mFm_F denotes the number of facets of FF.

Proof. By definition we have deg⁡(pvF)=deg⁡(pv∣F)−deg⁡(rF)\deg(p_v^F)=\deg(p_{v|F})-\deg(r_F). We are going to compute the summands on the right individually.

From (3) we get the factorization

pv∣F=p^v∣F⋅∏G∌vℓG∣F.p_{v|F}=\widehat{p}_{v|F}\cdot\prod_{G\not\ni v}\ell_{G|F}.

From Theorem 2.6 (i)(i) and (ii)(ii), as well as Theorem 3.1 we get

oF(p^v)=(ii)oF(pv)=(i)oF(q)=3.10.o_F(\widehat{p}_v)\mathrel{\overset{(ii)}{=}}o_F(p_v)\mathrel{\overset{(i)}{=}}o_F(q)\mathrel{\overset{3.1}{=}}0.

Hence, p^v∣F≠0\widehat{p}_{v|F}\ne0. By Theorem 3.2, p^v\widehat{p}_v is homogeneous at v∈Fv\in F. The restriction of a homogeneous polynomial to a containing affine subspace is either zero or of the same degree. Hence we get

deg⁡(p^v∣F)=deg⁡(p^v).\deg(\widehat{p}_{v|F})=\deg(\widehat{p}_v).

Assume for now that there is no facet GG parallel to FF.2 Then deg⁡(ℓG∣F)=1=deg⁡(ℓG)\deg(\ell_{G|F})=1=\deg(\ell_G) for all facets G≠FG\ne F. We get

deg⁡(pv∣F)=deg⁡(p^v∣F)+∑F∌vdeg⁡(ℓG∣F)\deg(p_{v|F})=\deg(\widehat{p}_{v|F})+\sum_{F\not\ni v}\deg(\ell_{G|F})
=deg⁡(p^v)+∑F∌vdeg⁡(ℓG)=deg⁡(pv)=m−d.=\deg(\widehat{p}_v)+\sum_{F\not\ni v}\deg(\ell_G)=\deg(p_v)=m-d.

Likewise deg⁡(rF)=m−mF−1\deg(r_F)=m-m_F-1. The claim follows by subtraction.

If, however, G≠FG\ne F is a facet that is parallel to FF, then deg⁡(ℓG∣F)=0\deg(\ell_{G|F})=0, changing the computed value of both deg⁡(pv∣F)\deg(p_{v|F}) and deg⁡(rF)\deg(r_F). But both values change by exactly 1, and so there is no net change to the value of deg⁡(pvF)\deg(p_v^F). □\square

Corollary 3.7. Let F⊂PF\subset P be a facet. If {pv/q}\{p_v/q\} are GBCs on PP of minimal numerator degree deg⁡(pv)=m−d\deg(p_v)=m-d, then the reduced restrictions {pvF/qF}\{p_v^F/q^F\} form GBCs on FF of minimal numerator degree deg⁡(pvF)=mF−(d−1)\deg(p_v^F)=m_F-(d-1).

Remark 3.8. It is this point at which we stop proving per-vertex variants. For completeness we note that Theorem 3.9 and Theorem 3.10 below do admit such variants, but their proofs go too far beyond the intended focus of this article.

The following result provides an equality analogue of Theorem 2.8, and a strengthening of Theorem 3.1 (ii)(ii) to general faces.

Lemma 3.9. Suppose deg⁡(pv)=m−d\deg(p_v)=m-d for all vv. For a face f⊆Pf\subseteq P with v∈fv\in f holds

of(q)=of(pv)=ex⁡(f).o_f(q)=o_f(p_v)=\operatorname{ex}(f).

This result is stronger than needed for what follows. Subsequently we only need this statement for when ff is a facet or a ridge of PP. In both cases ex⁡(f)=0\operatorname{ex}(f)=0, that is, Theorem 3.9 gives us that pvp_v and qq do not vanish on these faces.

Proof of Theorem 3.9. The proof proceeds by induction on codim⁡(f)\operatorname{codim}(f). The statement is clear for f=Pf=P. If ff is a facet then this is precisely Theorem 3.1 (ii).

Assume now codim⁡(f)≥2\operatorname{codim}(f)\geq2. We have of(pv)=of(q)≥ex⁡(f)o_f(p_v)=o_f(q)\geq\operatorname{ex}(f) by Theorem 2.6 (ii) and Theorem 2.8. Thus, it suffices to show of(q)≤ex⁡(f)o_f(q)\leq\operatorname{ex}(f). Fix a facet F⊃fF\supset f. Then

of(q)≤of(q∣F)=of(qF)+of(rF).o_f(q)\leq o_f(q|_F)=o_f(q^F)+o_f(r_F).

We rewrite the summands individually.

First, we rewrite of(qF)o_f(q^F). Theorem 3.7 gives us that {pvF/qF}\{p_v^F/q^F\} form GBCs on FF of minima numerator degree deg⁡(pvF)=mF−(d−1)\deg(p_v^F)=m_F-(d-1). Also codim⁡F(f)=codim⁡P(f)−1\operatorname{codim}_F(f)=\operatorname{codim}_P(f)-1 (here and in the following we use subscripts to indicate the polytopes with respect to which a quantity is expressed). Hence, the induction hypothesis applies and gives us of(qF)=ex⁡F(f)o_f(q^F)=\operatorname{ex}_F(f).

To rewrite of(rF)o_f(r_F) we use that rFr_F is the product of restricted linear forms ℓG∣F\ell_G|_F, where G⊂PG\subset P is a facet that is neither identical nor adjacent to FF. Such a linear form contributes to of(rF)o_f(r_F) exactly if f⊂Gf\subset G. Write Fac⁡P(f)\operatorname{Fac}_P(f) (resp. Fac⁡F(f)\operatorname{Fac}_F(f)) for the number of facets of PP (resp. of FF) that contain ff. Recall that the facets G⊂PG\subset P adjacent to FF correspond to facets of FF. Then

of(rF)=Fac⁡P(f)−Fac⁡F(f)−1=(codim⁡P(f)+ex⁡P(f))−(codim⁡F(f)+ex⁡F(f))−1=(ex⁡P(f)−ex⁡F(f))+(codim⁡P(f)−codim⁡F(f)−1)=ex⁡P(f)−ex⁡F(f).\begin{aligned} o_f(r_F)&=\operatorname{Fac}_P(f)-\operatorname{Fac}_F(f)-1\\ &=(\operatorname{codim}_P(f)+\operatorname{ex}_P(f))-(\operatorname{codim}_F(f)+\operatorname{ex}_F(f))-1\\ &=(\operatorname{ex}_P(f)-\operatorname{ex}_F(f))+(\operatorname{codim}_P(f)-\operatorname{codim}_F(f)-1)\\ &=\operatorname{ex}_P(f)-\operatorname{ex}_F(f). \end{aligned}

This finishes the proof. □

We arrive at the main result of the section (cf. [11 Theorem 6]).

Theorem 3.10. If α={pv/q}\alpha=\{p_v/q\} and α~={p~v/q~}\tilde{\alpha}=\{\tilde{p}_v/\tilde{q}\} are rational GBCs with minimal numerator degrees deg⁡(pv)=deg⁡(p~v)=m−d\deg(p_v)=\deg(\tilde{p}_v)=m-d for all vertices v∈Pv\in P, then α=α~\alpha=\tilde{\alpha}. In particular, the Wachspress coordinates are the unique rational GBCs of minimal numerator degree.

Proof. We proceed by induction on the dimension dd of PP. The case d≤1d\leq1 is clear because then PP is a simplex on which only barycentric coordinates exist.

We now assume d≥2d\geq2. Fix a facet F⊂PF\subset P. By Theorem 3.7 the reduced restrictions {pvF/qF}\{p_v^F/q^F\} and {p~vF/q~F}\{\tilde{p}_v^F/\tilde{q}^F\} are GBCs of minimal numerator degree on FF. Hence, the induction hypothesis applies, that is, {pvF/qF}\{p_v^F/q^F\} and {p~vF/q~F}\{\tilde{p}_v^F/\tilde{q}^F\} are equivalent and there is a factor λF\lambda_F so that pvF=λFp~vFp_v^F=\lambda_F\tilde{p}_v^F and qF=λFq~Fq^F=\lambda_F\tilde{q}^F. For all we know, λF\lambda_F might depend on the facet F⊂PF\subset P. We show that it does not.3

From Theorem 3.5 we get

q∣F=rFqF=rFλFq~F=λFrFq~F=λFq~∣F.(8)q|_F=r_Fq^F=r_F\lambda_F\tilde{q}^F=\lambda_Fr_F\tilde{q}^F=\lambda_F\tilde{q}|_F. \tag*{(8)}

Now choose a ridge R⊂FR \subset F. Restricting both sides of (8) to RR yields

q∣R=q∣F∣R=λFq~∣F∣R=λFq~∣R⟹λF=q∣Rq~∣Rq|_{R}=q|_{F}|_{R}=\lambda_{F}\widetilde{q}|_{F}|_{R}=\lambda_{F}\widetilde{q}|_{R}\quad\Longrightarrow\quad\lambda_{F}=\frac{q|_{R}}{\widetilde{q}|_{R}}

which is valid since q~∣R≢0\widetilde{q}|_{R}\not\equiv0 by Theorem 3.9. We see that λF\lambda_{F} only depends on the ridge. In particular, adjacent facets share the same value for the factor. Since the facet-adjacency graph is connected, it follows that λ:=λF\lambda:=\lambda_{F} is independent of FF.

At this point we can conclude

rFpv∣F=pvF=λp~vF=λrFp~v∣F⟹rF(pv−λp~v)∣F=0.r_{F}p_{v}|_{F}=p_{v}^{F}=\lambda\widetilde{p}_{v}^{F}=\lambda r_{F}\widetilde{p}_{v}|_{F}\quad\Longrightarrow\quad r_{F}(p_{v}-\lambda\widetilde{p}_{v})|_{F}=0.

for all facets F⊂PF\subset P. Since rFr_{F} does not vanish on FF, we conclude that each facet linear form ℓF\ell_{F} divides pv−λp~vp_{v}-\lambda\widetilde{p}_{v}. There are mm of these. Since this exceeds the degree of pvp_{v} and p~v\widetilde{p}_{v}, we conclude pv=λp~vp_{v}=\lambda\widetilde{p}_{v}. The same argument proves q=λq~q=\lambda\widetilde{q}.

Induction on Theorem 3.7 then also yields

Corollary 3.11. The restriction of the Wachspress coordinates to some face f⊂Pf\subset P yields the Wachspress coordinates on ff.

Intermediate degree GBCs

After proving [11 Theorem 6], Warren notes

[ . . . ] the last part of the proof allows for the possibility of rational coordinates functions of degree nn [ mm in our notation ] (or greater) that are barycentric. Constructing such a wider class of such coordinates functions is an interesting problem for future research.

suggesting that numerator degrees strictly between m−dm-d and mm are excluded by the same proof. This is not correct. The following example demonstrates that such intermediate degree GBCs (i.e., numerator degree m−d+km-d+k with 0<k<d0<k<d) do indeed exist and are abundant:

Example 3.12 (Intermediate degree). Consider the cube [0,1]3=[0,1]2×[0,1][0,1]^{3}=[0,1]^{2}\times[0,1]. The parallelogram has GBCs of degree 44 (cf. ??) and the line has GBCs of degree 11. The product GBCs on the cube therefore have numerator degree 5<6=m−d−15<6=m-d-1. To generalize this we can consider the parallelogram together with a simplex of any dimension. Then we can construct GBCs of degree 55 for any dimension.

We do agree that the construction of a wider class of rational GBCs is interesting. For example, initial investigations (that are beyond the scope of this article) suggest that every polytope that is not a simplex can be equipped with rational GBCs of numerator degree deg⁡(pv)=m−d+k\deg(p_{v})=m-d+k for all k≥0k\geq0. It appears interesting and challenging to study the space of rational GBCs on PP for a fixed kk. For example, the question for k=1k=1 seems related to Kohn and Ranestad’s [8 Theorem 7], which in turn suggests that there is an nn-dimensional space of such rational GBCs (here, nn is the number of vertices of PP).

Low denominator degrees

From this section on we transition to the study of denominators of rational GBCs. The focus is on polytopes that admit rational GBCs of especially low denominator degree. For a polytope P⊂RdP\subset\mathbb{R}^{d} let deg⁡(P)\deg(P) be the minimally possible denominator degree for rational GBCs on PP. We call it the degree of PP. Note that deg⁡(P)=0\deg(P)=0 if and only if PP admits polynomial GBCs, which we address in § 4.1.

Recall that the denominator of the Wachspress coordinates is called the adjoint polynomial adjP⁡\operatorname{adjP} of PP and we have deg⁡(adjP⁡)≤m−d−1\deg(\operatorname{adjP}) \le m-d-1 (cf. § 2.3), where mm is the number of facets of PP. Hence

deg⁡(P)≤deg⁡(adjP⁡)≤m−d−1.\deg(P) \le\deg(\operatorname{adjP}) \le m-d-1.

A core message of [2] is that, in general, deg⁡(adjP⁡)\deg(\operatorname{adjP}) can be much smaller than m−d−1m-d-1. The main factor seems to be parallelities in the face structure of PP, though the precise geometric properties are not well understood. In particular, and in contrast to numerators, the degree of the Wachspress denominator does not depend only on the combinatorics of PP. It is known that deg⁡(adjP⁡)≥m−2d\deg(\operatorname{adjP}) \ge m-2d [2 Theorem 3.5 (iii)].

Conjecture 4.1. For all polytopes P⊂RdP \subset\mathbb{R}^{d} holds deg⁡(P)≥m−2d\deg(P) \ge m-2d.

In general, understanding the degree of polytopes seems to be hard. We are aware of the following negative results:

  1. There are polytopes where det⁡(P)<det⁡(adjP⁡)\det(P) < \det(\operatorname{adjP}). That is, the Wachspress coordinates are not necessarily the GBCs of minimal denominator degree (cf. Theorem A.2).

  2. The GBCs of minimal denominator degree det⁡(P)\det(P) might not be unique (cf. Theorem A.3).

We collect some results about the polytope degree.

Corollary 4.2. If f⊆Pf \subseteq P is a face, then deg⁡(f)≤deg⁡(P)\deg(f) \le\deg(P).

Proof. It suffices to prove this for when ff is a facet, other faces follow by induction. If now {pv/q}\{p_v/q\} are rational GBCs on PP with deg⁡(P)=deg⁡(q)\deg(P)=\deg(q), they are in particular maximally canceled. By Theorem 2.12 qq does not vanish on a facet. We can therefore restrict {pv/q}\{p_v/q\} to ff and obtain rational GBCs of a denominator degree that cannot be larger than deg⁡(q)\deg(q).

Low degree implies combinatorial restrictions:

Lemma 4.3. Let PvP_v be the vertex figure of PP at vertex v∈Pv \in P. Every vertex w∈Pvw \in P_v lies in all but at most deg⁡(P)+1\deg(P)+1 facets of PvP_v.

Proof. Let ww be a vertex of PvP_v. It corresponds to an edge e=vwˉe=v\bar{w} of PP, where wˉ\bar{w} denotes the other endpoint of ee. Let NN be the number of facets of PvP_v that do not contain ww. We show that N≤deg⁡(P)+1N \le\deg(P)+1.

Choose rational GBCs {pz/q}\{p_z/q\} on PP such that deg⁡(q)=deg⁡(P)\deg(q)=\deg(P), and set r:=oe(q)r:=o_e(q). The point of introducing rr is that qq may vanish identically on ee, so we cannot simply restrict the rational functions pz/qp_z/q to ee.

Parametrize aff⁡(e)\operatorname{aff}(e) by xs:=(1−s)v+swˉx_s:=(1-s)v+s\bar{w}, and choose a generic vector u∈Rdu \in\mathbb{R}^{d} transverse to aff⁡(e)\operatorname{aff}(e). Since oe(q)=ro_e(q)=r, we may write

q(xs+tu)=trqˉ(s)+O(tr+1),pwˉ(xs+tu)=trpˉ(s)+O(tr+1),q(x_s+tu)=t^r\bar{q}(s)+O(t^{r+1}), \qquad p_{\bar{w}}(x_s+tu)=t^r\bar{p}(s)+O(t^{r+1}),

where qˉ∈R[s]\bar{q} \in\mathbb{R}[s] is nonzero and deg⁡(qˉ)≤deg⁡(q)\deg(\bar{q}) \le\deg(q). The second expansion is possible because Theorem 2.6 (ii) gives oe(pwˉ)=oe(q)=ro_e(p_{\bar{w}})=o_e(q)=r.

We first relate pˉ\bar{p} and qˉ\bar{q}. Choose an affine linear form LL with L(v)=0L(v)=0 and L(wˉ)=1L(\bar{w})=1. Affine precision says that L(x)q(x)=∑zL(z)pz(x)L(x)q(x)=\sum_z L(z)p_z(x). Substituting x=xs+tux=x_s+tu, we have L(xs)=sL(x_s)=s, and hence L(xs+tu)=s+O(t)L(x_s+tu)=s+O(t). Thus the coefficient of trt^r on the left-hand side is sqˉ(s)s\bar{q}(s).

On the right-hand side, the term corresponding to vv vanishes because L(v)=0L(v)=0, while the term corresponding to wˉ\bar{w} contributes pˉ(s)\bar{p}(s) because L(wˉ)=1L(\bar{w})=1. If z∉ez \notin e, then Theorem 2.6 (ii) gives oe(pz)>ro_e(p_z)>r, so these terms do not contribute to the coefficient of trt^r. Comparing coefficients therefore yields

pˉ(s)=sqˉ(s).\bar{p}(s)=s\bar{q}(s).

We now use the factorization (3):

pwˉ=p^wˉ∏F∌wˉℓF.p_{\bar{w}}=\hat{p}_{\bar{w}}\prod_{F\not\ni\bar{w}}\ell_F.

For every facet FF occurring in this product, wˉ∉F\bar{w}\notin F, and hence e⊈Fe\not\subseteq F. Thus ℓF\ell_F does not vanish identically on ee, so oe(ℓF)=0o_e(\ell_F)=0. Since oe(pwˉ)=ro_e(p_{\bar{w}})=r, it follows that oe(p^wˉ)=ro_e(\hat{p}_{\bar{w}})=r.

We may therefore write p^wˉ(xs+tu)=trh(s)+O(tr+1)\hat{p}_{\bar{w}}(x_s+tu)=t^rh(s)+O(t^{r+1}) for some h∈R[s]h\in\mathbb{R}[s]. Since each ℓF\ell_F is affine linear, we also have ℓF(xs+tu)=ℓF(xs)+O(t)\ell_F(x_s+tu)=\ell_F(x_s)+O(t). Taking the coefficient of trt^r in the factorization above gives

pˉ(s)=h(s)∏F∌wˉℓF(xs).\bar{p}(s)=h(s)\prod_{F\not\ni\bar{w}}\ell_F(x_s).

From this point on, divisibility is understood in the univariate polynomial ring R[s]\mathbb{R}[s].

Facets of PvP_v correspond to facets FF of PP containing vv, and such a facet of PvP_v contains ww precisely when the corresponding facet FF contains ee. Hence the NN facets of PvP_v not containing ww correspond exactly to the facets FF satisfying v∈Fv\in F and wˉ∉F\bar{w}\notin F.

For each such facet, ℓF(v)=0\ell_F(v)=0 and ℓF(wˉ)≠0\ell_F(\bar{w})\ne0. Since xs=(1−s)v+swˉx_s=(1-s)v+s\bar{w}, we have ℓF(xs)=sℓF(wˉ)\ell_F(x_s)=s\ell_F(\bar{w}). Thus each of these NN factors contributes a factor ss to the product above, and hence sN∣pˉ(s)s^N\mid\bar{p}(s). Together with pˉ(s)=sqˉ(s)\bar{p}(s)=s\bar{q}(s), this gives sN−1∣qˉ(s)s^{N-1}\mid\bar{q}(s) whenever N≥1N\ge1. Since qˉ≠0\bar{q}\ne0,

N−1≤deg⁡(qˉ)≤deg⁡(q)=deg⁡(P).N-1\le\deg(\bar{q})\le\deg(q)=\deg(P).

Thus N≤deg⁡(P)+1N\le\deg(P)+1, as claimed. □\square

Polynomial GBCs

In this section we prove the following main result

Theorem 1.2. A polytope P⊂RdP\subset\mathbb{R}^d admits polynomial GBCs if and only if it is affinely equivalent to a product of simplices.

As a corollary we obtain

Corollary 4.4. If deg⁡(P)=0\deg(P)=0 then deg⁡(adj⁡P)=0\deg(\operatorname{adj}_P)=0.

As alluded to in § 4.1, polynomial GBCs on cartesian products of simplices are constructed as product coordinates. Polynomiality is moreover preserved under affine transformation. Thus, it suffices to focus on the “only if” part of Theorem 1.2.

Throughout this section, let P⊂RdP\subset\mathbb{R}^d be a polytope with polynomial GBCs. By the following theorem it suffices to show that PP is simple and affinely equivalent to a 0/1-polytope:

Theorem 4.5 ([7 Theorem 1]). A 0/1-polytope is simple if and only if it is a product of 0/1-simplices.

That PP is simple is a direct consequence of Theorem 4.3: for denominator degree zero, vertex figures must be simplices:

Corollary 4.6. Polytopes with polynomial GBCs are simple.

To show that PP is affinely equivalent to a 0/10/1-polytope requires some preparation. For this, recall that the slack matrix SPS_P of PP is the matrix whose rows are indexed by the vertices, columns are indexed by the facets, and entries are given by

(SP)v,F=ℓF(v).(S_P)_{v,F}=\ell_F(v).

The slack matrix depends on our choice of facet linear forms ℓF\ell_F, which are determined by PP only up to a positive factor. Hence, we can rescale the columns of SPS_P freely and still obtain a slack matrix of PP. Given the following theorem, it therefore suffices to show that the columns can be rescaled to 0/10/1-vectors:

Theorem 4.7 ([1 Theorem 5]). PP is affinely equivalent to the convex hull of the rows of its slack matrix SPS_P.

With this we have all ingredients for a proof.

Proof of Theorem 1.2. Suppose that PP has polynomial GBCs {pv∈R[x]}\{p_v\in\mathbb{R}[x]\}. Fix a facet F⊂PF\subset P. We show that across all vertices vv, the linear form ℓF(v)\ell_F(v) only ever attains two values, depending on whether v∈Fv\in F or v∉Fv\notin F. Since ℓF(v)=0\ell_F(v)=0 for all v∈Fv\in F, the corresponding column of the slack matrix SPS_P can therefore be rescaled to a 0/10/1-vector. The proof then concludes as explained above, using Theorem 4.7, Theorem 4.6 and Theorem 4.5.

It remains to show that ℓF(v)\ell_F(v) takes on the same value for all v∉Fv\notin F. Since the part of the edge graph of PP induced on the vertices outside of FF is still connected, it suffices to show that adjacent vertices attain the same value under ℓF\ell_F.

Suppose then that v,w∉Fv,w\notin F are adjacent via an edge e⊂Pe\subset P and ℓF(v)≠ℓF(w)\ell_F(v)\ne\ell_F(w). Then the restriction ℓF∣e\ell_F|_e is not constant. Moreover, there exists a facet F′⊂PF'\subset P that contains ww but not vv. Then ℓF′(w)=0≠ℓF′(v)\ell_{F'}(w)=0\ne\ell_{F'}(v), and the restriction ℓF′∣e\ell_{F'}|_e is also not constant. Restricting the representation (3) to ee yields

pv∣e=p~v∣e⋅∏G∌vℓG∣ep_v|_e=\widetilde{p}_v|_e\cdot\prod_{G\not\ni v}\ell_G|_e

and we see that both ℓF∣e\ell_F|_e and ℓF′∣e\ell_{F'}|_e are independent non-constant factors of pv∣ep_v|_e. The latter is therefore at least quadratic. However, the pv∣ep_v|_e form GBCs on the edge, and are therefore necessarily linear. This is a contradiction, and thus ℓF(v)=ℓF(w)\ell_F(v)=\ell_F(w).

Statement on AI use

AI was used only for finding several of the examples in Section A. The article is entirely human written and AI had no part in the presented research results outside the appendix.

Funding

Martin Winter is funded by the SPP 2458 “Combinatorial Synergies” (project ID 539851419), funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation).

Appendix A. Examples

We collect here examples to support the main text.

Example A.1 (perturbations). Given rational GBCs {αv}\{\alpha_v\}, one can construct new rational GBCs βv:=αv+rv\beta_v := \alpha_v + r_v using a suitable family of rational maps {rv:P→R}\{r_v : P \to\mathbb{R}\}. Preserving normalization and precision is equivalent to

∑vrv=0and∑vrvv=0.(9)\sum_v r_v = 0 \quad\text{and} \quad\sum_v r_v v = 0. \tag*{(9)}

Positivity is harder to guarantee. We need at least ℓF∣rv\ell_F \mid r_v whenever v∉Fv \notin F since we need to preserve the zero loci derived from positivity in § 2.2. This is automatic if we instead perturb like

βv:=αv+rv∏FℓF.\beta_v := \alpha_v + r_v \prod_F \ell_F.

Consider the square from Theorem 2.5. We can choose r00=r11=1r_{00} = r_{11} = 1 and r01=r10=−1r_{01} = r_{10} = -1. This yields GBCs of numerator degree 4:

βv1=ℓ2ℓ3+ℓ1ℓ2ℓ3ℓ4,βv2=ℓ3ℓ4−ℓ1ℓ2ℓ3ℓ4,βv3=ℓ4ℓ1+ℓ1ℓ2ℓ3ℓ4,βv4=ℓ1ℓ2−ℓ1ℓ2ℓ3ℓ4.\begin{aligned} \beta_{v_1} &= \ell_2\ell_3 + \ell_1\ell_2\ell_3\ell_4,\\ \beta_{v_2} &= \ell_3\ell_4 - \ell_1\ell_2\ell_3\ell_4,\\ \beta_{v_3} &= \ell_4\ell_1 + \ell_1\ell_2\ell_3\ell_4,\\ \beta_{v_4} &= \ell_1\ell_2 - \ell_1\ell_2\ell_3\ell_4. \end{aligned}

Example A.2 (Adjoint degree is not minimal). We construct a polytope PP with deg⁡(P)<deg⁡(adj⁡P)\deg(P) < \deg(\operatorname{adj}P), in dimension d=3d = 3 with fully explicit GBCs. The idea is to start with a polytope with a small adjoint, such as a zonotope, and to move facets parallel to themselves. This keeps the normals and parallelisms, but the adjoint jumps back to its generic degree m−d−1m - d - 1. Other GBCs can still profit from cancellations that the adjoint does not see.

Let Z⊂R3Z \subset\mathbb{R}^3 be the truncated octahedron

Z={∣x∣≤6, ∣y∣≤6, ∣z∣≤6, ∣±x±y±z∣≤9}.Z = \{ |x| \le6,\ |y| \le6,\ |z| \le6,\ |\pm x \pm y \pm z| \le9 \}.

It is a zonotope with m=14m = 14 facets, so deg⁡(adj⁡Z)=m−2d=8\deg(\operatorname{adj}Z) = m - 2d = 8. Let PP be obtained from ZZ by moving three facets parallel to themselves: replace y≤6y \le6 by y≤8y \le8, and x+y+z≤9x + y + z \le9 and x+y−z≤9x + y - z \le9 by x+y+z≤10x + y + z \le10 and x+y−z≤10x + y - z \le10 (see Figure 1). Then PP is simple with 24 vertices and the same normals, and deg⁡(adj⁡P)=10=m−d−1\deg(\operatorname{adj}P) = 10 = m - d - 1. There are rational GBCs {pv/q}\{p_v/q\} on PP with deg⁡(q)=9\deg(q) = 9 and pv=p^v∏F∌vℓFp_v = \hat{p}_v \prod_{F \not\ni v} \ell_F, where each p^v\hat{p}_v is a quadratic polynomial that is positive on PP. Hence again deg⁡(P)≤9<10=deg⁡(adj⁡P)\deg(P) \le9 < 10 = \deg(\operatorname{adj}P). The exact GBCs are complicatd and are provided on Zenodo.

The polytopes of Theorem A.2 in dimension three: the truncated octahedron $Z$ and the deformed polytope $P$.

Figure 1. The polytopes of Theorem A.2 in dimension three. Left: the truncated octahedron ZZ. Right: the polytope PP obtained by moving three facets (red edges) parallel to themselves. Both have 14 facets and the same normals, but deg⁡(adj⁡Z)=8\deg(\operatorname{adj}_Z)=8 and deg⁡(adj⁡P)=10\deg(\operatorname{adj}_P)=10, while PP admits rational GBCs with denominator of degree 9.

Example A.3 (Minimal denominator degree is not unique). Consider the trapezoid P=conv⁡{v1,v2,v3,v4}⊂R2P = \operatorname{conv}\{v_1, v_2, v_3, v_4\} \subset\mathbb{R}^2 with vertices v1=(0,0)v_1 = (0,0), v2=(2,0)v_2 = (2,0), v3=(1,1)v_3 = (1,1), v4=(0,1)v_4 = (0,1). It has m=4m = 4 facets, namely {y=0}\{y = 0\}, {x+y=2}\{x + y = 2\}, {y=1}\{y = 1\}, {x=0}\{x = 0\}, and dimension n=2n = 2, so m−n=2m - n = 2. We construct rational GBCs on PP whose numerators have the intermediate degree m−n+1=3m - n + 1 = 3 and the denominator has degree one, but is not equal to the adjoint.

Let

pv1=(2−x−y)(1−y)(2+x),pv2=x(1−y)(2+x+y),pv3=2xy(x+y),pv4=2y(2−x−y)(1+x).\begin{aligned} p_{v_1} &= (2 - x - y)(1 - y)(2 + x), \qquad p_{v_2} = x(1 - y)(2 + x + y),\\ p_{v_3} &= 2xy(x + y), \qquad p_{v_4} = 2y(2 - x - y)(1 + x). \end{aligned}

and q=2(2+x−y)q = 2(2+x-y). We claim that {pvi/q}\{p_{v_i}/q\} are rational GBCs on PP. Indeed, one checks by expanding that

∑i=14pvi=q,2pv2+pv3=xq,pv3+pv4=yq.\sum_{i=1}^{4}p_{v_i}=q,\qquad2p_{v_2}+p_{v_3}=xq,\qquad p_{v_3}+p_{v_4}=yq.

The first identity is normalization. The other two are linear precision, since ∑ipvivi=(2pv2+pv3,pv3+pv4)\sum_i p_{v_i}v_i=(2p_{v_2}+p_{v_3},p_{v_3}+p_{v_4}). On PP we have 0≤x0\le x, 0≤y≤10\le y\le1 and x+y≤2x+y\le2. Hence every factor of every pvip_{v_i} is non-negative, and q=2(2+x−y)>0q=2(2+x-y)>0. So pvi/qp_{v_i}/q is well defined and non-negative on PP.

Degrees. We have deg⁡(pvi)=3=m−n+1\deg(p_{v_i})=3=m-n+1 for all ii and deg⁡(q)=1\deg(q)=1. The representation is maximally canceled: qq is an irreducible linear form that does not divide pv1=(2−x−y)(1−y)(2+x)p_{v_1}=(2-x-y)(1-y)(2+x), so qq and the pvip_{v_i} have no common non-constant factor. Hence the numerator degree m−n+1m-n+1 is not caused by a cancelable common factor of numerators and denominator.

Residual flats and the adjoint. The Wachspress denominator of PP is adj⁡P=2−y\operatorname{adj}_P=2-y, which also has degree 1. However, qq is not a multiple of adj⁡P\operatorname{adj}_P. The pair of facets {x=0}\{x=0\} and {x+y=2}\{x+y=2\} is not adjacent, so E={(0,2)}E=\{(0,2)\} is a residual flat, and q(0,2)=0q(0,2)=0 as required by Theorem 2.13; see Figure 2. But qq does not vanish at the point at infinity of the two parallel facets {y=0}\{y=0\} and {y=1}\{y=1\}: in homogeneous coordinates [x:y:z][x:y:z] this point is [1:0:0][1:0:0], and the homogenization 2(2z+x−y)2(2z+x-y) of qq takes the value 2≠02\ne0 there. So qq vanishes on the affine, but not on the projective residual arrangement of PP. The adjoint, in contrast, vanishes on the projective residual arrangement.

The trapezoid $P$ with its facet lines, affine residual flat, and zero set of $q$

Figure 2. The trapezoid PP of Theorem A.3 with its facet lines (dotted), the affine residual flat (0,2)(0,2) (red point) and the zero set of qq (orange). The parallel facets {y=0}\{y=0\} and {y=1}\{y=1\} meet only at infinity.

Example A.4. Theorem 2.13 states that the denominator qq vanishes on residual flats, and Theorem 3.3 gives the analogue for a numerator pvp_v with v∈Ev\in E under the assumption deg⁡(pv)=m−d\deg(p_v)=m-d. The following example shows that this assumption cannot be dropped. There are rational GBCs, a residual flat EE, and a vertex v∈Ev\in E with pv∣E≢0p_v|_E\not\equiv0.

The polytope. Let Q⊂R3Q \subset\mathbb{R}^{3} be the square pyramid with apex (0,0,1)(0,0,1),

Q={(x,y,z)∈R3:0≤z≤1, ∣x∣≤1−z, ∣y∣≤1−z},Q=\{(x,y,z)\in\mathbb{R}^{3}:0\le z\le1,\ |x|\le1-z,\ |y|\le1-z\},

and let P=Q×[0,1]2⊂R5P=Q\times[0,1]^{2}\subset\mathbb{R}^{5} with coordinates (x,y,z,s,t)(x,y,z,s,t). It has m=5+4=9m=5+4=9 facets, dimension d=5d=5, and 20 vertices, namely (σ,τ,0,i,j)(\sigma,\tau,0,i,j) and (0,0,1,i,j)(0,0,1,i,j) with σ,τ∈{±1}\sigma,\tau\in\{\pm1\} and i,j∈{0,1}i,j\in\{0,1\}. Thus m−d=4m-d=4.

The residual flat. The facets F=P∩{x=1−z}F=P\cap\{x=1-z\} and F‾=P∩{x=−(1−z)}\overline{F}=P\cap\{x=-(1-z)\} of PP intersect in

F∩F‾={(0,0,1)}×[0,1]2,aff⁡(F∩F‾)={x=0, y=0, z=1},F\cap\overline{F}=\{(0,0,1)\}\times[0,1]^{2},\qquad\operatorname{aff}(F\cap\overline{F})=\{x=0,\ y=0,\ z=1\},

whereas

E:=aff⁡(F)∩aff⁡(F‾)={x=0, z=1}⊋aff⁡(F∩F‾).E:=\operatorname{aff}(F)\cap\operatorname{aff}(\overline{F})=\{x=0,\ z=1\}\supsetneq\operatorname{aff}(F\cap\overline{F}).

So EE is an intersection of facet-spanning flats that is not spanned by the vertices it contains, i.e., a residual flat. The vertices in EE are the four apex vertices (0,0,1,i,j)(0,0,1,i,j).

The GBCs. On QQ we use the GBCs

αapex=z,α(σ,τ)=(1−z+σx)(1−z+τy)4(1−z),σ,τ∈{±1},\alpha_{\mathrm{apex}}=z,\qquad\alpha_{(\sigma,\tau)}=\frac{(1-z+\sigma x)(1-z+\tau y)}{4(1-z)},\qquad\sigma,\tau\in\{\pm1\},

where (σ,τ)(\sigma,\tau) denotes the base vertex (σ,τ,0)(\sigma,\tau,0). (One checks directly that they are non-negative on QQ, sum to 1, and reproduce x,y,zx,y,z.) On [0,1]2[0,1]^{2} we use the bilinear coordinates

λ00=(1−s)(1−t),λ10=s(1−t),λ11=st,λ01=(1−s)t,\lambda_{00}=(1-s)(1-t),\qquad\lambda_{10}=s(1-t),\qquad\lambda_{11}=st,\qquad\lambda_{01}=(1-s)t,

and set B:=λ00λ11=λ10λ01=s(1−s)t(1−t)B:=\lambda_{00}\lambda_{11}=\lambda_{10}\lambda_{01}=s(1-s)t(1-t). The product coordinates α×λ\alpha\times\lambda of Theorem 2.4 are GBCs on PP. We perturb them at the apex vertices:

γ(σ,τ,0,i,j)=α(σ,τ)λij,\gamma_{(\sigma,\tau,0,i,j)}=\alpha_{(\sigma,\tau)}\lambda_{ij},
γ(0,0,1,i,j)=zλij+(−1)i+jzy2B1−z.\gamma_{(0,0,1,i,j)}=z\lambda_{ij}+(-1)^{i+j}\frac{zy^{2}B}{1-z}.

These are rational GBCs on PP (verified below) with common denominator q=4(1−z)q = 4(1-z) and numerators

p(σ,τ,0,i,j)=(1−z+σx)(1−z+τy)λij,p(0,0,1,i,j)=4z(1−z)λij+(−1)i+j4zy2B.\begin{aligned} p_{(\sigma,\tau,0,i,j)} &= (1-z+\sigma x)(1-z+\tau y)\lambda_{ij},\\ p_{(0,0,1,i,j)} &= 4z(1-z)\lambda_{ij}+(-1)^{i+j}4zy^{2}B. \end{aligned}

The base numerators have degree 4=m−d4=m-d, the apex numerators have degree 7≠m−d7\ne m-d.

Restriction to $E. Consider the vertex v=(0,0,1,0,0)∈Ev=(0,0,1,0,0)\in E. On EE we have x=0x=0 and z=1z=1, so the first summand of pvp_v vanishes and

pv∣E=4y2s(1−s)t(1−t)≠0.\left.p_v\right|_{E}=4y^{2}s(1-s)t(1-t)\ne0.

For instance, at the point (0,1,1,12,12)∈E(0,1,1,\frac{1}{2},\frac{1}{2})\in E it equals 14\frac{1}{4}. This point lies in EE but not in PP; on E∩P=F∩F‾E\cap P=F\cap\overline{F} we have y=0y=0, so pvp_v does vanish there. Since deg⁡(pv)=7≠m−d\deg(p_v)=7\ne m-d, this does not contradict Theorem 3.3. △

Normalization. Since ∑i,jλij=1\sum_{i,j}\lambda_{ij}=1 and ∑i,j(−1)i+jB=B+B−B−B=0\sum_{i,j}(-1)^{i+j}B=B+B-B-B=0, the apex coordinates sum to zz. The base coordinates sum to ∑σ,τα(σ,τ)=1−z\sum_{\sigma,\tau}\alpha(\sigma,\tau)=1-z. Hence ∑vγv=1\sum_v\gamma_v=1.

Linear precision. The perturbation (−1)i+jzy2B1−z(-1)^{i+j}\frac{zy^{2}B}{1-z} does not change ∑vγvv\sum_v\gamma_v v: it sums to 00 over (i,j)(i,j), also when weighted with ii (the terms for (1,0)(1,0) and (1,1)(1,1) cancel) and with jj (those for (0,1)(0,1) and (1,1)(1,1) cancel), and the apex has x=y=0x=y=0. So ∑vγvv\sum_v\gamma_vv equals the corresponding sum for the product coordinates, which is (x,y,z,s,t)(x,y,z,s,t).

Non-negativity. The coordinates γ(σ,τ,0,i,j)\gamma_{(\sigma,\tau,0,i,j)} are products of non-negative factors. On the apex, γ(0,0,1,0,0)\gamma_{(0,0,1,0,0)} and γ(0,0,1,1,1)\gamma_{(0,0,1,1,1)} are sums of non-negative terms. For the remaining two,

γ(0,0,1,1,0)=zλ10(1−y2λ011−z)≥0,\gamma_{(0,0,1,1,0)}=z\lambda_{10}\left(1-\frac{y^{2}\lambda_{01}}{1-z}\right)\ge0,

because y21−z≤1−z≤1\frac{y^{2}}{1-z}\le1-z\le1 on QQ and λ01≤1\lambda_{01}\le1; the vertex (0,0,1,0,1)(0,0,1,0,1) is analogous.

Continuity at z=1z=1. On PP we have ∣zy2B1−z∣≤(1−z)B≤1−z16→0\left|\frac{zy^{2}B}{1-z}\right|\le(1-z)B\le\frac{1-z}{16}\to0 for z→1z\to1. Hence γ\gamma extends continuously to the fiber {(0,0,1)}×[0,1]2\{(0,0,1)\}\times[0,1]^2 over the apex, where it equals the product coordinates.

References

  1. [1]M. Aprile, M. Conforti, Y. Faenza, S. Fiorini, T. Huynh, and M. Macchia. Recognizing cartesian products of matrices and polytopes. Graphs and Combinatorial Optimization: from Theory to Applications: CTW2020 Proceedings, pages 361–373, 2021.arxiv.org/abs/2002.02264
  2. [2]T. Baumbach, A. Freyer, J. Weigert, and M. Winter. The canonical form, scissors congruence and adjoint degrees of polytopes, 2025.arxiv.org/abs/2508.04275
  3. [3]C. Brüser and J. Weigert. Geometry of adjoint hypersurfaces for polytopes. arXiv preprint arXiv:2511.13537, 2025.
  4. [4]P. Clarke and D. A. Cox. Moment maps, strict linear precision, and maximum likelihood degree one. Advances in Mathematics, 370:107233, 2020.arxiv.org/abs/1810.03672
  5. [5]M. S. Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161–214, 2015.DOI
  6. [6]T. Ju, S. Schaefer, J. D. Warren, and M. Desbrun. A geometric construction of coordinates for convex polyhedra using polar duals. In Symposium on Geometry Processing, pages 181–186, 2005.DOI
  7. [7]V. Kaibel and M. Wolff. Simple 0/1-polytopes. European Journal of Combinatorics, 21(1):139–144, 2000.DOI
  8. [8]K. Kohn and K. Ranestad. Projective geometry of wachspress coordinates. Foundations of Computational Mathematics, 20(5):1135–1173, 2020.arxiv.org/abs/1904.02123
  9. [9]E. L. Wachspress. A rational finite element basis. 1975.DOI
  10. [10]J. Warren. Barycentric coordinates for convex polytopes. Advances in Computational Mathematics, 6(1):97–108, 1996.DOI
  11. [11]J. Warren. On the uniqueness of barycentric coordinates. Contemporary Mathematics, 334:93–100, 2003.DOI
  12. [12]M. Winter. Can a polytopal graph be “centrally symmetric” in more than one way? https://mathoverflow.net/q/439787. version: 2023-02-01.

Paper details

vlnc:W_7nz4by3m
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents