betaFeedback

WorksMathematics

The maximum likelihood degree of no-three-way interaction models

Abstract

We prove that for all I,J,K≥2I,J,K\ge2 the maximum likelihood degree of the no-three-way interaction model of I×J×KI\times J\times K contingency tables is (−1)(I−1)(J−1)∫β(MA)K dχ(A), (-1)^{(I-1)(J-1)}\int\beta(M_A)^K\,d\chi(A), an integral with respect to the Euler characteristic over the I×JI\times J matrices AA with nonzero entries modulo row and column scaling, where β(MA)\beta(M_A) is the beta invariant of the matroid of the columns of [\Id_I\mid A]. By a theorem of Clarke, Hoşten, Kushnerchuk, and Oldekop, β(MA)\beta(M_A) is the maximum likelihood degree of the independence model scaled by AA. Closed forms were previously known only when one variable is binary, and the formula recovers them. For fixed II and JJ the maximum likelihood degree is an exponential polynomial in KK; the coefficient of its largest base (I+J−2I−1)\binom{I+J-2}{I-1} is the maximum likelihood degree of the moduli space X(I,I+J)X(I,I+J) of points in linearly general position, which counts solutions to the generalized scattering equations of Cachazo, Early, Guevara, and Mizera. The same argument for a generically scaled model expresses the degree of MI,J,K\mathcal M_{I,J,K} as an Euler-characteristic integral of products of beta invariants. For 3×33\times3 matrices the integral reduces to counting strata of line arrangements, and the maximum likelihood degree of the 3×3×K3\times3\times K model is 26⋅6K−60⋅5K+36⋅4K+6⋅3K−9⋅2K+126\cdot6^K-60\cdot5^K+36\cdot4^K+6\cdot3^K-9\cdot2^K+1; in particular the 3×3×33\times3\times3 model has maximum likelihood degree 511511.

Introduction

The no-three-way interaction model for an I×J×KI \times J \times K contingency table consists of the distributions proportional to aijbikcjka_{ij}b_{ik}c_{jk}, and their limits: each pair of the three variables interacts, and there is no three-factor interaction [3, 4]. Its maximum likelihood estimate has no closed form and is computed iteratively, for instance by iterative proportional fitting [11, 10]. The algebraic complexity of the estimate is measured by the maximum likelihood degree: the number of complex critical points of the likelihood function on the model for generic data [13, 2]. Only when one variable is binary is it known in closed form. Améndola et al. conjectured that the maximum likelihood degree of the 2×2×K2 \times2 \times K model is 2K−12^K - 1 [2 Conj. 36], and Brysiewicz and Maraj proved this and gave a closed formula for every 2×J×K2 \times J \times K model by identifying it with a Lawrence lifting and expressing its maximum likelihood degree as an evaluation of a Tutte polynomial [5 Thm. 23, Cor. 32, Cor. 38]. They noted that the degree of the toric variety M3,3,3\mathcal{M}_{3,3,3} is 2457 and left the models with no binary variable to future work [5 Rem. 39]. In the binary case, the maximum likelihood estimate is also given by explicit determinants: for 2×2×K2 \times2 \times K tables it is determined by a root of the determinant of a 2K×2K2^K \times2^K matrix, which gives an elementary proof that the maximum likelihood degree is 2K−12^K - 1, and for 2×J×K2 \times J \times K tables the likelihood equations form a multiparameter eigenvalue problem [14].

We give a formula for every I×J×KI \times J \times K model (Theorem 2.1):

mld⁡(MI,J,K)=(−1)(I−1)(J−1)∫β(MA)K dχ(A),\operatorname{mld}(\mathcal{M}_{I,J,K}) = (-1)^{(I-1)(J-1)} \int\beta(\mathcal{M}_{A})^{K}\,\mathrm{d}\chi(A),

where AA runs over the I×JI \times J matrices with nonzero entries modulo row and column scaling, MA\mathcal{M}_{A} is the matroid of the columns of [IdI∣A][\mathrm{Id}_{I} \mid A], β\beta is Crapo’s beta invariant [9], and the integral is with respect to the Euler characteristic [15]. By a theorem of Clarke, Hoşten, Kushnerchuk, and Oldekop [8], β(MA)\beta(\mathcal{M}_{A}) is the maximum likelihood degree of the independence model of an I×JI \times J table scaled by AA, so the maximum likelihood degree of the three-way model is a signed Euler-characteristic sum of the KK-th powers of those of the scaled two-way models. For fixed II and JJ it is an exponential polynomial in KK, whose bases are the beta invariants that occur and whose coefficients are Euler characteristics of matroid strata.

The proof applies Huh’s theorem that the maximum likelihood degree of a smooth very affine variety is its signed Euler characteristic [13] to the model with its coordinate hyperplanes and the hyperplane ∑p=0\sum p = 0 removed, together with the observation that for a fixed interaction AA the model decomposes into KK independent slices whose contributions multiply. Theorem 2.1 is stated for the model XTX_{T} scaled by any tensor TT, with β(MA)K\beta(\mathcal{M}_{A})^{K} replaced by ∏kβ(MTk∘A)\prod_{k}\beta(\mathcal{M}_{T_{k}\circ A}); for generic TT this gives the degree of the toric variety, and in particular an expression for the degree 2457 of M3,3,3\mathcal{M}_{3,3,3} (Corollary 4.1). The formula recovers the result of Brysiewicz and Maraj (Corollary 4.3): for I=2I = 2 the strata are indexed by set partitions of [J][J], and the integral becomes a Stirling-number sum equivalent to their Tutte-polynomial evaluation. Since the model is symmetric in its three variables, the formula also yields an identity between integrals over matrices of different shapes (Corollary 4.2).

Since β(MA)\beta(\mathcal{M}_{A}) depends only on the matroid, the integral is a finite sum of Euler characteristics of matroid strata. For 3×33 \times3 matrices we compute them exactly, by fibering over the second column of AA: the fibers are strata of arrangements of lines through pairs of at most five points in the plane, whose combinatorics depends only on which three of these points are collinear. This gives the maximum likelihood degree of every 3×3×K3 \times3 \times K model (Proposition 5.1); in particular the 3×3×33 \times3 \times3 model has maximum likelihood degree 511. For 3×43 \times4 matrices we evaluate the integral by counting points over finite fields; the counts are polynomial in qq for all primes tested, and their values at q=1q = 1 give a formula for the 3×4×K3 \times4 \times K models (§ 5). Up to sign, the coefficient of the largest base (I+J−2I−1)\binom{I+J-2}{I-1} is the Euler characteristic of the moduli space X(I,I+J)X(I,I+J) of points in linearly general position, whose likelihood equations are the scattering equations of Cachazo, Early, Guevara, and Mizera [6, 1]; hence the leading coefficient of the maximum likelihood degree in KK is the number of solutions of these equations (Proposition 6.1).

Statement

Fix integers I,J,K≥2I,J,K \ge2. The no-three-way interaction model MI,J,K\mathcal{M}_{I,J,K} is the toric variety X⊆PIJK−1X \subseteq\mathbb{P}^{IJK-1}, the closure of the image of

(a,b,c)⟼(aijbikcjk)1≤i≤I, 1≤j≤J, 1≤k≤K,(a,b,c)∈(C∗)IJ×(C∗)IK×(C∗)JK.(a,b,c) \longmapsto(a_{ij}b_{ik}c_{jk})_{1\le i\le I,\,1\le j\le J,\,1\le k\le K}, \qquad(a,b,c) \in(\mathbb{C}^{*})^{IJ} \times(\mathbb{C}^{*})^{IK} \times(\mathbb{C}^{*})^{JK}.

More generally, for T=(tijk)∈(C∗)IJKT = (t_{ijk}) \in(\mathbb{C}^{*})^{IJK} let XTX_{T} be the closure of the image of (a,b,c)⟼(tijkaijbikcjk)(a,b,c) \longmapsto(t_{ijk}a_{ij}b_{ik}c_{jk}), the model scaled by TT, so that X=X1X = X_{1}, and write Tk=(tijk)i,jT_{k} = (t_{ijk})_{i,j} for its slices. Let H⊆PIJK−1\mathcal{H} \subseteq\mathbb{P}^{IJK-1} be the union of the coordinate hyperplanes and the hyperplane ∑ijkpijk=0\sum_{ijk}p_{ijk} = 0. The maximum likelihood degree mldeg⁡(XT)\operatorname{mldeg}(X_{T}) is the number of complex critical points of the log-likelihood function ∑ijkuijklog⁡pijk−u+++log⁡∑ijkpijk\sum_{ijk}u_{ijk}\log p_{ijk} - u_{+++}\log\sum_{ijk}p_{ijk}, where u+++=∑ijkuijku_{+++} = \sum_{ijk}u_{ijk}, on the smooth locus of XT∖HX_{T} \setminus\mathcal{H}.

For an I×JI \times J matrix A=(aij)A=(a_{ij}) with nonzero entries, let MAM_A be the matroid on I+JI+J elements represented by the columns of [IdI∣A][\mathrm{Id}_I \mid A], and let β(M)\beta(M) be the beta invariant of Crapo [9],

β(M)=(−1)rk⁡M∑S⊆E(−1)∣S∣rk⁡S.\beta(M)=(-1)^{\operatorname{rk} M}\sum_{S\subseteq E}(-1)^{\lvert S\rvert}\operatorname{rk} S.

The torus G=((C∗)I×(C∗)J)/C∗G=((\mathbb{C}^{*})^I\times(\mathbb{C}^{*})^J)/\mathbb{C}^{*} acts freely on (C∗)I×J(\mathbb{C}^{*})^{I\times J} by row and column scaling, and MAM_A is constant on orbits; the quotient is identified with the torus TI,J=(C∗)(I−1)(J−1)T_{I,J}=(\mathbb{C}^{*})^{(I-1)(J-1)} of matrices with a1j=ai1=1a_{1j}=a_{i1}=1. The function A↦β(MA)A\mapsto\beta(M_A) is constructible, since the matroid strata are defined by the vanishing and nonvanishing of minors, and β(MA)≥1\beta(M_A)\geq1, since MAM_A is connected on I+J≥4I+J\geq4 elements: each column of AA has full support, so e1,…,eIe_1,\ldots,e_I and any column of AA form a circuit, and β\beta of a connected matroid on at least two elements is positive [9].

For a constructible function φ\varphi on a complex algebraic variety YY that takes finitely many values,

∫Yφ dχ=∑tt χ(φ−1(t)),\int_Y \varphi\,d\chi=\sum_t t\,\chi\left(\varphi^{-1}(t)\right),

where χ\chi is the topological Euler characteristic [15].

Theorem 2.1. For all I,J,K≥2I,J,K\geq2 and T∈(C∗)IJKT\in(\mathbb{C}^{*})^{IJK},

mldeg⁡(XT)=(−1)(I−1)(J−1)∫TI,J∏k=1Kβ(MTk∘A) dχ(A),\operatorname{mldeg}(X_T)=(-1)^{(I-1)(J-1)}\int_{T_{I,J}}\prod_{k=1}^{K}\beta(M_{T_k\circ A})\,d\chi(A),

where ∘\circ is the entrywise product. In particular,

mldeg⁡(MI,J,K)=(−1)(I−1)(J−1)∫TI,Jβ(MA)K dχ(A).\operatorname{mldeg}(\mathcal{M}_{I,J,K})=(-1)^{(I-1)(J-1)}\int_{T_{I,J}}\beta(M_A)^K\,d\chi(A).

When T=1T=1, β(MA)\beta(M_A) is the maximum likelihood degree of the independence model of an I×JI\times J table scaled by AA [8 Thm. 1.3] (this also follows from Lemma 3.5 below and Huh’s theorem), so the integrand is its KK-th power. Grouping matrices by the value of β\beta, Theorem 2.1 reads

mldeg⁡(MI,J,K)=(−1)(I−1)(J−1)∑s≥1WI,J(s)sK,WI,J(s)=χ({A∈TI,J:β(MA)=s}),\operatorname{mldeg}(\mathcal{M}_{I,J,K})=(-1)^{(I-1)(J-1)}\sum_{s\geq1}W_{I,J}(s)s^K,\qquad W_{I,J}(s)=\chi\left(\{A\in T_{I,J}:\beta(M_A)=s\}\right),

an exponential polynomial in KK whose bases are the beta invariants that occur.

Proof of Theorem 2.1

We use two properties of the Euler characteristic of complex algebraic varieties: it is additive over finite partitions into locally closed subvarieties, and for a morphism π:Y→B\pi:Y\to B,

χ(Y)=∫Bχ(π−1(x)) dχ(x).(1)\chi(Y)=\int_B\chi\left(\pi^{-1}(x)\right)\,d\chi(x). \tag*{(1)}

where x↦χ(π−1(x))x\mapsto\chi(\pi^{-1}(x)) is constructible [15]. Viro’s calculus uses the Euler characteristic with compact supports, which for complex algebraic varieties coincides with χ\chi (see [7] and the references there). In particular χ\chi is multiplicative on products, and χ(C∗)=0\chi(\mathbb{C}^{*})=0.

The proof proceeds in four steps. We first identify XT∖HX_T\setminus\mathcal{H} with a smooth hypersurface V={f=1}V=\{f=1\} in a torus T′T' (Lemma 3.1) and apply Huh’s theorem. Projecting VV onto A∈TI,JA\in T_{I,J} reduces χ(V)\chi(V) to the Euler characteristic of a level set {g1+⋯+gK=1}\{g_1+\cdots+g_K=1\} of a sum of KK homogeneous functions, which factors into single-slice contributions (Lemmas 3.2 and 3.3); each slice then evaluates to a signed beta invariant via hyperplane arrangement complements (Lemmas 3.4 and 3.5).

Lemma 3.1. Let d=(I−1)(J−1)+(I−1)K+JKd=(I-1)(J-1)+(I-1)K+JK and let T′=(C∗)dT'=(\mathbb{C}^*)^d have coordinates aija_{ij} (i,j≥2i,j\ge2), bikb_{ik} (i≥2i\ge2), and cjkc_{jk}, and put a1j=ai1=b1k=1a_{1j}=a_{i1}=b_{1k}=1. Then XT∖HX_T\setminus H is isomorphic to the hypersurface

V={x∈T′:f(x)=1},f=∑k=1K∑i,jtijkaijbikcjk=∑k=1KbkT(Tk∘A)c⋅k.V=\left\{x\in T':f(x)=1\right\},\qquad f=\sum_{k=1}^{K}\sum_{i,j}t_{ijk}a_{ij}b_{ik}c_{jk}=\sum_{k=1}^{K}b_k^{\mathsf T}(T_k\circ A)c_{\cdot k}.

which is smooth, of dimension d−1d-1.

Proof. The map ψ:T′→(C∗)IJK\psi:T'\to(\mathbb{C}^*)^{IJK}, x↦(tijkaijbikcjk)x\mapsto(t_{ijk}a_{ij}b_{ik}c_{jk}), is injective with an inverse given by monomials on its image: with qijk=pijk/tijkq_{ijk}=p_{ijk}/t_{ijk}, cjk=q1jkc_{jk}=q_{1jk}, bik=qi1k/q11kb_{ik}=q_{i1k}/q_{11k}, and aij=qij1q111/(qi11q1j1)a_{ij}=q_{ij1}q_{111}/(q_{i11}q_{1j1}). Every (a,b,c)(a,b,c) in the full torus can be rescaled by aij↦aijλiμja_{ij}\mapsto a_{ij}\lambda_i\mu_j, bik↦bikλi−1νkb_{ik}\mapsto b_{ik}\lambda_i^{-1}\nu_k, cjk↦cjkμj−1νk−1c_{jk}\mapsto c_{jk}\mu_j^{-1}\nu_k^{-1}, which preserves every product aijbikcjka_{ij}b_{ik}c_{jk}, so as to satisfy a1j=ai1=b1k=1a_{1j}=a_{i1}=b_{1k}=1. Hence ψ(T′)\psi(T') is the image of the full torus, a translate of a closed subtorus of (C∗)IJK(\mathbb{C}^*)^{IJK} that is stable under scaling (by scaling cc); it is therefore X^T∩(C∗)IJK\widehat{X}_T\cap(\mathbb{C}^*)^{IJK}, where X^T⊆CIJK\widehat{X}_T\subseteq\mathbb{C}^{IJK} is the affine cone over XTX_T, and ψ\psi is an isomorphism onto it. A point of XT∖HX_T\setminus H has a unique representative pp in this cone with ∑p=1\sum p=1, and ∑ψ(x)=f(x)\sum\psi(x)=f(x). The function ff is homogeneous of degree one under c↦λcc\mapsto\lambda c, so ∑jkcjk∂f/∂cjk=f=1\sum_{jk}c_{jk}\partial f/\partial c_{jk}=f=1 on VV, and df≠0df\ne0 there. □

By Lemma 3.1, XT∖H≅VX_T\setminus H\cong V is smooth, so it coincides with the set (XT)reg∖H(X_T)_{\mathrm{reg}}\setminus H on which critical points are counted, and through the coordinates pp it is a closed subvariety of (C∗)IJK(\mathbb{C}^*)^{IJK}, namely the torus coset X^T∩(C∗)IJK\widehat{X}_T\cap(\mathbb{C}^*)^{IJK} intersected with {∑p=1}\{\sum p=1\}. On VV the term u+++log⁡∑pu_{+++}\log\sum p vanishes, and since the log-likelihood is invariant under scaling pp, its critical points on XT∖HX_T\setminus H correspond to the critical points on VV of the master function ∏pijkuijk\prod p_{ijk}^{u_{ijk}}. By Huh’s theorem [13 Thm. 1] (see also [8 Thm. 2.3] for this form), for generic uu

mldeg⁡(XT)=(−1)d−1χ(V).(2)\operatorname{mldeg}(X_T)=(-1)^{d-1}\chi(V). \tag*{(2)}

Lemma 3.2. For m≥1m\ge1 let Hm={v∈(C∗)m:v1+⋯+vm=1}H_m=\{v\in(\mathbb{C}^*)^m:v_1+\cdots+v_m=1\}. Then χ(Hm)=(−1)m−1\chi(H_m)=(-1)^{m-1}.

Proof. For m=1m=1, H1H_1 is a point. For m≥2m\ge2 project to vmv_m. Over vm∈C∗∖{1}v_m\in\mathbb{C}^*\setminus\{1\} the fiber is {v′∈(C∗)m−1:∑v′=1−vm}≅Hm−1\{v'\in(\mathbb{C}^*)^{m-1}:\sum v'=1-v_m\}\cong H_{m-1}, and (v′,vm)↦(v′/(1−vm),vm)(v',v_m)\mapsto(v'/(1-v_m),v_m) trivializes this family over C∖{0,1}\mathbb{C}\setminus\{0,1\}. Over vm=1v_m=1 the fiber {v′∈(C∗)m−1:∑v′=0}\{v'\in(\mathbb{C}^*)^{m-1}:\sum v'=0\} is empty for m=2m=2 and isomorphic to C∗×Hm−2\mathbb{C}^*\times H_{m-2} for m≥3m\ge3, by v′↦(vm−1′,−(v1′,…,vm−2′)/vm−1′)v'\mapsto(v'_{m-1},-(v'_1,\ldots,v'_{m-2})/v'_{m-1}), so its Euler characteristic is 0. Hence χ(Hm)=χ(C∖{0,1})χ(Hm−1)=−χ(Hm−1)\chi(H_m)=\chi(\mathbb{C}\setminus\{0,1\})\chi(H_{m-1})=-\chi(H_{m-1}). □

Lemma 3.3. For k=1,…,Kk=1,\ldots,K let YkY_k be a variety with an algebraic C∗\mathbb{C}^*-action and gk:Yk→Cg_k:Y_k\to\mathbb{C} a regular function with gk(λ⋅y)=λgk(y)g_k(\lambda\cdot y)=\lambda g_k(y). Put zk=χ(gk−1(0))z_k=\chi(g_k^{-1}(0)) and ek=χ(gk−1(1))e_k=\chi(g_k^{-1}(1)). Then zk=χ(Yk)z_k=\chi(Y_k) and

χ({y∈∏kYk:∑kgk(yk)=1})=∏kzk−∏k(zk−ek).\chi\left(\left\{y\in\prod_kY_k:\sum_k g_k(y_k)=1\right\}\right)=\prod_k z_k-\prod_k(z_k-e_k).

Proof. The map y↦(gk(y),gk(y)−1⋅y)y\mapsto(g_k(y),g_k(y)^{-1}\cdot y) is an isomorphism Yk∖gk−1(0)→C∗×gk−1(1)Y_k\setminus g_k^{-1}(0)\to\mathbb{C}^*\times g_k^{-1}(1), so χ(Yk)=zk\chi(Y_k)=z_k. Partition the set on the left by P={k:gk(yk)≠0}P=\{k:g_k(y_k)\ne0\}, which is nonempty. The part with a given PP is isomorphic to ∏k∉Pgk−1(0)×∏k∈Pgk−1(1)×H∣P∣\prod_{k\notin P}g_k^{-1}(0)\times\prod_{k\in P}g_k^{-1}(1)\times H_{|P|}, by y↦((yk)k∉P,(gk(yk)−1⋅yk)k∈P,(gk(yk))k∈P)y\mapsto((y_k)_{k\notin P},(g_k(y_k)^{-1}\cdot y_k)_{k\in P},(g_k(y_k))_{k\in P}), with inverse (y′,e,v)↦y(y',e,v)\mapsto y, where yk=yk′y_k=y'_k for k∉Pandk\notin P andy_k=v_k\cdot e_k$ for k∈Pk\in P. By Lemma 3.2 the left side is

∑∅≠P⊆[K](−1)∣P∣−1∏k∉Pzk∏k∈Pek=∏kzk−∏k(zk−ek).\sum_{\varnothing\ne P\subseteq[K]}(-1)^{|P|-1}\prod_{k\notin P}z_k\prod_{k\in P}e_k=\prod_k z_k-\prod_k(z_k-e_k).

:::

Lemma 3.4. Let v1,…,vn∈Crv_1,\ldots,v_n \in\mathbb{C}^r span Cr\mathbb{C}^r, let MM be their matroid, and let UM⊆Pr−1U_M \subseteq\mathbb{P}^{r-1} be the complement of the hyperplanes {x:ve⋅x=0}\{x:v_e\cdot x=0\}. Then χ(UM)=(−1)r−1β(M)\chi(U_M)=(-1)^{r-1}\beta(M).

Proof. We use the following properties of β\beta [9]: β(M)=β(M∖e)+β(M/e)\beta(M)=\beta(M\setminus e)+\beta(M/e) if ee is neither a loop nor a coloop; β(M)=0\beta(M)=0 if MM has a loop, or a coloop and at least two elements; and β(M)=1\beta(M)=1 if MM is loopless of rank one. If MM has a loop, UMU_M is empty. If r=1r=1 and MM is loopless, UM=P0U_M=\mathbb{P}^0 is a point. For r≥2r\ge2 induct on nn, for all ranks simultaneously. If every element is a coloop, UM≅(C∗)r−1U_M\cong(\mathbb{C}^*)^{r-1} and β(M)=0\beta(M)=0. Otherwise choose ee neither a loop nor a coloop; then M∖eM\setminus e has rank rr and M/eM/e has rank r−1≥1r-1\ge1. Then UM∖eU_{M\setminus e} is the disjoint union of UMU_M and the complement in the hyperplane He≅Pr−2H_e\cong\mathbb{P}^{r-2} of the traces of the remaining hyperplanes, which is UM/eU_{M/e}, since the trace of {vf⋅x=0}\{v_f\cdot x=0\} on HeH_e is defined by the image of vfv_f in Cr/⟨ve⟩\mathbb{C}^r/\langle v_e\rangle. Hence χ(UM)=χ(UM∖e)−χ(UM/e)=(−1)r−1(β(M∖e)+β(M/e))\chi(U_M)=\chi(U_{M\setminus e})-\chi(U_{M/e})=(-1)^{r-1}\bigl(\beta(M\setminus e)+\beta(M/e)\bigr).

Lemma 3.5. For an I×JI\times J matrix AA with nonzero entries let Y={y∈(C∗)I:y1=1}×(C∗)JY=\{y\in(\mathbb{C}^*)^I:y_1=1\}\times(\mathbb{C}^*)^J and g(y,γ)=yTAγg(y,\gamma)=y^{\mathsf{T}}A\gamma. Then χ(g−1(1))=(−1)I+Jβ(MA)\chi(g^{-1}(1))=(-1)^{I+J}\beta(M_A).

Proof. Project g−1(1)g^{-1}(1) to yy and put w=ATyw=A^{\mathsf{T}}y. If Z={j:wj=0}Z=\{j:w_j=0\}, the fiber is {γ∈(C∗)J:∑j∉Zwjγj=1}\{\gamma\in(\mathbb{C}^*)^J:\sum_{j\notin Z}w_j\gamma_j=1\}, which is isomorphic to (C∗)Z×HJ−∣Z∣(\mathbb{C}^*)^Z\times H_{J-|Z|} by γj↦wjγj\gamma_j\mapsto w_j\gamma_j for j∉Zj\notin Z and the identity on γj\gamma_j for j∈Zj\in Z when Z≠[J]Z\ne[J], and is empty when Z=[J]Z=[J]. Its Euler characteristic is (−1)J−1(-1)^{J-1} if Z=∅Z=\varnothing and 00 otherwise, and it is constant on the locally closed sets of yy with a given ZZ. By (1), χ(g−1(1))=(−1)J−1χ(UA)\chi(g^{-1}(1))=(-1)^{J-1}\chi(U_A), where

UA={y∈(C∗)I:y1=1, (ATy)j≠0 for all j}U_A=\{y\in(\mathbb{C}^*)^I:y_1=1,\ (A^{\mathsf{T}}y)_j\ne0\text{ for all }j\}

is the complement in PI−1\mathbb{P}^{I-1} of the hyperplanes whose normal vectors are the columns of [IdI∣A][\mathop{\mathrm{Id}}\nolimits_I\mid A]. By Lemma 3.4, χ(UA)=(−1)I−1β(MA)\chi(U_A)=(-1)^{I-1}\beta(M_A).

Proof of Theorem 2.1. Project VV to A∈TI,JA\in T_{I,J}. The fiber over AA is {(b,c):∑kgk(b⋅k,c⋅k)=1}\{(b,c):\sum_k g_k(b_{\cdot k},c_{\cdot k})=1\}, where gkg_k is the function of Lemma 3.5 for the matrix Tk∘AT_k\circ A on Yk=YY_k=Y, with C∗\mathbb{C}^* acting on the γ\gamma factor. Since YY is a torus of positive dimension, zk=χ(Y)=0z_k=\chi(Y)=0, and Lemmas 3.3 and 3.5 give the fiber Euler characteristic

−∏k(−ek)=−(−1)(I+J−1)K∏kβ(MTk∘A).-\prod_k(-e_k)=-(-1)^{(I+J-1)K}\prod_k\beta(M_{T_k\circ A}).

By (1) and (2),

mldeg⁡(XT)=(−1)d(−1)(I+J−1)K∫TI,J∏kβ(MTk∘A) dχ(A),\operatorname{mldeg}(X_T)=(-1)^d(-1)^{(I+J-1)K}\int_{T_{I,J}}\prod_k\beta(M_{T_k\circ A})\,d\chi(A),

and d+(I+J−1)K=(I−1)(J−1)+2(I+J−1)K≡(I−1)(J−1)(mod2)d+(I+J-1)K=(I-1)(J-1)+2(I+J-1)K\equiv(I-1)(J-1)\pmod2.

Consequences

Corollary 4.1 (Degree). For generic T∈(C∗)IJKT\in(\mathbb{C}^*)^{IJK},

deg⁡(MI,J,K)=(−1)(I−1)(J−1)∫TI,J∏k=1Kβ(MTk∘A) dχ(A).\deg(\mathcal{M}_{I,J,K})=(-1)^{(I-1)(J-1)}\int_{T_{I,J}}\prod_{k=1}^{K}\beta(M_{T_k\circ A})\,d\chi(A).

Proof. For a generic scaling the maximum likelihood degree of a toric variety equals its degree [2 Cor. 8]; apply Theorem 2.1.

Brysiewicz and Maraj noted that they have no combinatorial interpretation of the degree 2457 of M3,3,3\mathcal{M}_{3,3,3} [5, Rem. 39; Corollary 4.1 expresses it as an Euler-characteristic integral of products of beta invariants over the common refinement of the matroid stratifications of the matrices Tk∘AT_k \circ A. For example, for I=J=2I=J=2 write AA in terms of ψ=a22\psi=a_{22} and let ρk=t11kt22k/(t12kt21k)\rho_k=t_{11k}t_{22k}/(t_{12k}t_{21k}) be the odds ratio of TkT_k. Then β(MTk∘A)=2\beta(M_{T_k\circ A})=2 unless ψρk=1\psi\rho_k=1, where it is 1. For generic TT the ρk\rho_k are distinct, the stratum C∗∖{ρ1−1,…,ρK−1}\mathbb{C}^* \setminus\{\rho_1^{-1},\ldots,\rho_K^{-1}\} has Euler characteristic −K-K, and the integral is −K 2K+K 2K−1=−K 2K−1-K\,2^K+K\,2^{K-1}=-K\,2^{K-1}, so deg⁡(M2,2,K)=K 2K−1\deg(M_{2,2,K})=K\,2^{K-1} [2 Prop. 34].

Corollary 4.2 (Symmetry). For all I,J,K≥2I,J,K\geq2,

(−1)(I−1)(J−1)∫(C∗)(I−1)(J−1)β(MA)K dχ(A)=(−1)(I−1)(K−1)∫(C∗)(I−1)(K−1)β(MB)J dχ(B),(-1)^{(I-1)(J-1)}\int_{(\mathbb{C}^*)^{(I-1)(J-1)}}\beta(M_A)^K\,\mathrm{d}\chi(A)=(-1)^{(I-1)(K-1)}\int_{(\mathbb{C}^*)^{(I-1)(K-1)}}\beta(M_B)^J\,\mathrm{d}\chi(B),

where AA ranges over I×JI\times J and BB over I×KI\times K matrices with first row and column equal to 1.

Proof. Both sides equal mldeg⁡(MI,J,K)\operatorname{mldeg}(M_{I,J,K}), by Theorem 2.1 applied to MI,J,KM_{I,J,K} and to the isomorphic model MI,K,JM_{I,K,J}. □

The symmetry in II and JJ is visible directly: the matrix [−AT∣Id⁡J][-A^{\mathsf{T}}\mid\operatorname{Id}_J] represents the dual of MAM_A with the same labels, and after scaling the first block by −1-1 and exchanging the two blocks it becomes [Id⁡J∣AT][\operatorname{Id}_J\mid A^{\mathsf{T}}], which represents MATM_{A^{\mathsf{T}}}; transposition maps the I×JI\times J chart onto the J×IJ\times I chart, and β(M∗)=β(M)\beta(M^*)=\beta(M) for matroids on at least two elements [9]. The symmetry between JJ and KK is not apparent from the integral.

Corollary 4.3 (Brysiewicz and Maraj [5 Cor. 32 and Cor. 38], in an equivalent form).

mldeg⁡(M2,J,K)=(−1)J−1∑m=1J(−1)m−1(m−1)! S(J,m) mK,\operatorname{mldeg}(M_{2,J,K})=(-1)^{J-1}\sum_{m=1}^{J}(-1)^{m-1}(m-1)!\,S(J,m)\,m^K,

where S(J,m)S(J,m) is a Stirling number of the second kind.

Proof. For I=2I=2, AA is determined by xj=a2jx_j=a_{2j} with x1=1x_1=1, and MAM_A consists of the points [1:0][1:0], [0:1][0:1], and [1:xj][1:x_j] of P1\mathbb{P}^1. If the xjx_j take mm distinct values, MAM_A is a rank-2 matroid whose simplification is the uniform matroid on m+2m+2 elements, so β(MA)=m\beta(M_A)=m (deleting a parallel element in a loopless matroid leaves β\beta unchanged, since contraction produces a loop). The matrices inducing a given partition of [J][J] into mm blocks are parametrized by the values of the m−1m-1 blocks not containing 1: distinct points of C∖{0,1}\mathbb{C}\setminus\{0,1\}, whose configuration space has Euler characteristic (−1)(−2)⋯(−(m−1))(-1)(-2)\cdots(-(m-1)). There are S(J,m)S(J,m) partitions, and the sign is (−1)(2−1)(J−1)(-1)^{(2-1)(J-1)}. □

Three-level variables

For I=3I=3, the matroid MAM_A is that of J+3J+3 points in P2\mathbb{P}^2: the coordinate points and the columns of AA. When the columns are distinct and no three of the J+3J+3 points are collinear, MAM_A is uniform of rank three and β(MA)=(J+12)\beta(M_A)=\binom{J+1}{2}; this is the largest value that occurs, since β(MA)\beta(M_A) is the maximum likelihood degree of the scaled independence model [8 Thm. 1.3], which is at most the degree (J+12)\binom{J+1}{2} of P2×PJ−1\mathbb{P}^2\times\mathbb{P}^{J-1} [2 Cor. 8].

We use two facts about β\beta. If MM is a simple matroid of rank three on nn elements represented over C\mathbb{C}, then

β(M)=(n−22)−∑L(∣L∣−12),(3)\beta(M)=\binom{n-2}{2}-\sum_L\binom{|L|-1}{2}, \tag*{(3)}

the sum over the lines LL of MM with at least three elements. By Lemma 3.4, β(M)\beta(M) is the Euler characteristic of the complement in P2\mathbb{P}^2 of the nn lines dual to the elements, which is 3−2n+∑L(∣L∣−1)3-2n+\sum_L(|L|-1), the sum now over all rank-two flats, the points where ∣L∣|L| of the lines meet; and ∑L(∣L∣−1)=(n2)−∑∣L∣≥3(∣L∣−12)\sum_L(|L|-1)=\binom{n}{2}-\sum_{|L|\geq3}\binom{|L|-1}{2}. Second, if ee and ff are parallel elements of a matroid MM without loops, then β(M)=β(M∖e)\beta(M) = \beta(M \setminus e), since ee is not a loop, is not a coloop because ff spans it, and M/eM/e has the loop ff.

Proposition 5.1. The weights W3,3(1),…,W3,3(6)W_{3,3}(1),\ldots,W_{3,3}(6) are 1,−9,6,36,−60,261,-9,6,36,-60,26, and W3,3(s)=0W_{3,3}(s)=0 for s>6s>6. Consequently, for all K≥2K\ge2,

mldeg⁡(M3,3,K)=26⋅6K−60⋅5K+36⋅4K+6⋅3K−9⋅2K+1;\operatorname{mldeg}(\mathcal{M}_{3,3,K})=26\cdot6^K-60\cdot5^K+36\cdot4^K+6\cdot3^K-9\cdot2^K+1;

in particular mldeg⁡(M3,3,3)=511\operatorname{mldeg}(\mathcal{M}_{3,3,3})=511.

Proof. Write A=[u∣v∣w]A=[u\mid v\mid w] with u=(1,1,1)Tu=(1,1,1)^{\mathsf{T}} and v=(1,v2,v3)Tv=(1,v_2,v_3)^{\mathsf{T}}, and regard e1,e2,e3,u,v,we_1,e_2,e_3,u,v,w as points of P2\mathbb{P}^2; MAM_A is their matroid. We apply (1) to each set {A:β(MA)=s}\{A:\beta(M_A)=s\} and the projection A↦vA\mapsto v; the fibers of the projection are copies of the torus (C∗)2(\mathbb{C}^*)^2 of points ww. For each vv let φv(s)\varphi_v(s) be the Euler characteristic of the set of ww in the torus with β(MA)=s\beta(M_A)=s. We show below that φv\varphi_v depends only on which of three types vv belongs to; then W3,3(s)W_{3,3}(s) is the sum over the types of the Euler characteristic of the set of such vv times φv(s)\varphi_v(s). We record φv\varphi_v via the exponential polynomial Φv(K)=∑sφv(s)sK\Phi_v(K)=\sum_s\varphi_v(s)s^K.

Let Pv={e1,e2,e3,u,v}P_v=\{e_1,e_2,e_3,u,v\}. Since uu and vv have no zero coordinate, three points of PvP_v are collinear only if they are ei,u,ve_i,u,v for some ii, that is, if vv lies on one of the lines v2=v3v_2=v_3, v3=1v_3=1, v2=1v_2=1 through uu and e1,e2,e3e_1,e_2,e_3, respectively, which meet only at uu. This gives three types of vv: the three sets SiS_i of v≠uv\ne u on the line through eie_i and uu, each isomorphic to C∗∖{1}\mathbb{C}^*\setminus\{1\}, with χ(Si)=−1\chi(S_i)=-1; the point v=uv=u; and the complement S0S_0 of the three lines, with χ(S0)=χ((C∗)2)−3χ(S1)−1=2\chi(S_0)=\chi((\mathbb{C}^*)^2)-3\chi(S_1)-1=2.

For fixed vv, let Av\mathcal{A}_v be the arrangement of lines spanned by pairs of points of PvP_v. The matroid MAM_A is determined by the set of lines of Av\mathcal{A}_v that contain ww, since ww equals a point aa of PvP_v exactly when it lies on two lines through aa. Two lines of Av\mathcal{A}_v share at most one point of PvP_v; if three lines passed through a point outside PvP_v, the sets of points of PvP_v on them would be pairwise disjoint, which needs six points. Hence the points lying on at least two lines of Av\mathcal{A}_v, which we call intersection points, are the points of PvP_v and, outside PvP_v, the points where two lines containing disjoint sets of points of PvP_v meet, each of which lies on exactly two lines. By additivity, the open part of a line ℓ\ell, the set of points on ℓ\ell and on no other line, has Euler characteristic 22 minus the number of intersection points on ℓ\ell, and the complement of Av\mathcal{A}_v has Euler characteristic 33 minus the sum of the Euler characteristics of the open parts minus the total number of intersection points. The point ww lies in the torus exactly when it avoids the three coordinate lines eieje_i e_j; in particular, since each intersection point outside PvP_v lies on exactly two lines, it lies in the torus unless one of its two lines is a coordinate line. All the counts below are combinatorial, so φv\varphi_v is constant on each type of vv.

If v∈S0v\in S_0, no three points of PvP_v are collinear, Av\mathcal{A}_v has 10 lines, and there are 15 intersection points outside PvP_v, so that each line contains five intersection points, two of them in PvP_v; nine of them lie on the coordinate lines, each of which meets the three lines spanned by the other three points. By (3) with the parallel-element rule, the strata of w∈(C∗)2w\in(\mathbb{C}^*)^2 are:

  • the complement of Av\mathcal{A}_v (χ=3+30−20=13\chi=3+30-20=13, β=6\beta=6);

  • the open parts of the seven noncoordinate lines (χ=−3\chi=-3 each, β=5\beta=5);

  • the six intersection points in the torus (χ=6\chi=6, β=4\beta=4);

  • the two points w∈{u,v}w\in\{u,v\} (χ=2\chi=2, β=3\beta=3, for five points with no three collinear).

Thus Φv(K)=13⋅6K−21⋅5K+6⋅4K+2⋅3K\Phi_v(K)=13\cdot6^K-21\cdot5^K+6\cdot4^K+2\cdot3^K.

If v∈Siv\in S_i, permuting the coordinates of P2\mathbb{P}^2 we may take i=1i=1. Then Av\mathcal{A}_v consists of the line mm through e1,u,ve_1,u,v, the three coordinate lines, and the four lines e2u,e2v,e3u,e3ve_2u,e_2v,e_3u,e_3v. The pairs of lines with disjoint sets of points of PvP_v give seven intersection points outside PvP_v: m∩e2e3m\cap e_2e_3, the four points where e1e2e_1e_2 meets e3u,e3ve_3u,e_3v and e1e3e_1e_3 meets e2u,e2ve_2u,e_2v, and the two points e2u∩e3ve_2u\cap e_3v and e2v∩e3ue_{2v} \cap e_{3u}, which lie in the torus. The line e2e3e_{2}e_{3} contains three intersection points and every other line four. By (3) with the parallel-element rule, the strata of w∈(C∗)2w \in(\mathbb{C}^{*})^{2} are:

  • the complement of Av\mathcal{A}_{v} (χ=3+15−12=6\chi= 3 + 15 - 12 = 6, β=5\beta= 5);

  • the open parts of e2u,e2v,e3u,e3ve_{2u}, e_{2v}, e_{3u}, e_{3v} (χ=−2\chi= -2 each, β=4\beta= 4);

  • the open part of mm (χ=−2\chi= -2, β=3\beta= 3) and the two intersection points in the torus (χ=2\chi= 2, β=3\beta= 3);

  • the two points w∈{u,v}w \in\{u,v\} (χ=2\chi= 2, β=2\beta= 2).

Thus Φv(K)=6⋅5K−8⋅4K+2⋅2K\Phi_v(K) = 6 \cdot5^K - 8 \cdot4^K + 2 \cdot2^K, the terms in 3K3^K canceling.

If v=uv = u, then β(MA)=β(M[u∣w])\beta(M_A) = \beta(M_{[u|w]}) by the parallel-element rule, so Φu(K)\Phi_u(K) is ∑sW3,2(s)sK=2⋅3K−3⋅2K+1\sum_s W_{3,2}(s)s^K = 2 \cdot3^K - 3 \cdot2^K + 1, by Corollary 4.3 and the symmetry W3,2=W2,3W_{3,2} = W_{2,3} noted after Corollary 4.2.

Adding 2 times the first of these, −3-3 times the second, and the third gives the stated weights, and Theorem 2.1 with (−1)(I−1)(J−1)=1(-1)^{(I-1)(J-1)} = 1 gives the formula. □

For J=4J = 4 we evaluate the weights by counting points over finite fields. The numbers of points over Fq\mathbb{F}_q of the strata {A∈T3,4:β(MA)=s}\{A \in T_{3,4} : \beta(M_A) = s\}, with β(MA)\beta(M_A) computed from the ranks of submatrices over Fq\mathbb{F}_q, agree for each ss with a polynomial in qq of degree at most 6 for all primes 7≤q≤377 \le q \le37, including at least two primes beyond those needed to determine it. If the numbers of points are given by these polynomials for all sufficiently large finite fields, the value of the polynomial at 1 is the Euler characteristic of the stratum, by a theorem of Katz [12, Thm. 6.1.2.3; see Thm. 2.1.8]. The entries for J=4J = 4 in Table 1, and the formula and values derived from them below, rest on this polynomiality, which we have verified for the primes listed.

Table 1. The weights W3,J(s)W_{3,J}(s), the Euler characteristics of the sets of AA with β(MA)=s\beta(M_A) = s. For J=2J = 2 they are those of Corollary 4.3, using W3,2=W2,3W_{3,2} = W_{2,3}; for J=3J = 3 they are given by Proposition 5.1; for J=4J = 4 they are the values at q=1q = 1 of the point counts.

ss12345678910
W3,2(s)W_{3,2}(s)1−3-32
W3,3(s)W_{3,3}(s)1−9-9636−60-6026
W3,4(s)W_{3,4}(s)1−21-2114204−288-288252−2088-20884878−4224-42241272

Table 1.

With these weights, Theorem 2.1 gives

mldeg⁡(M3,4,K)=1272⋅10K−4224⋅9K+4878⋅8K−2088⋅7K+252⋅6K−288⋅5K+204⋅4K+14⋅3K−21⋅2K+1.\operatorname{mldeg}(\mathcal{M}_{3,4,K}) = 1272 \cdot10^K - 4224 \cdot9^K + 4878 \cdot8^K - 2088 \cdot7^K + 252 \cdot6^K \\ - 288 \cdot5^K + 204 \cdot4^K + 14 \cdot3^K - 21 \cdot2^K + 1.

Table 2 lists the first values of this formula and of Proposition 5.1.

Table 2. Maximum likelihood degrees of the 3×3×K3 \times3 \times K and 3×4×K3 \times4 \times K models.

KK23456
3×3×K3 \times3 \times K31511575552711426811
3×4×K3 \times4 \times K1155755172951379749168382295

Table 2.

These values are consistent with the symmetry of the model. For K=2K = 2 they agree with the 2×3×32 \times3 \times3 and 2×3×42 \times3 \times4 values of Corollary 4.3, and the 3×4×33 \times4 \times3 and 3×3×43 \times3 \times4 models, computed from different weights, both give 5755. The weights W4,3W_{4,3}, computed in the same way, coincide with W3,4W_{3,4}, as the remark following Corollary 4.2 requires. For K=1K = 1 every formula gives 1, the maximum likelihood degree of the saturated model.

The leading term

The leading coefficient is an Euler characteristic that appears in the theory of generalized scattering amplitudes. Let X(I,n)X(I,n) be the moduli space of nn points in linearly general position in PI−1\mathbb{P}^{I-1}, the quotient by (C∗)n(\mathbb{C}^{*})^n of the open subset of the Grassmannian Gr⁡(I,n)\operatorname{Gr}(I,n) where all Plücker coordinates are nonzero [6]. It is a smooth very affine variety of dimension (I−1)(n−I−1)(I-1)(n-I-1).

Proposition 6.1. For every A∈TI,JA \in T_{I,J}, β(MA)≤(I+J−2I−1)\beta(M_A) \le\binom{I+J-2}{I-1}, with equality if and only if every minor of AA is nonzero, and the set of such AA is isomorphic to X(I,I+J)X(I,I+J). Consequently, as K→∞K \to\infty,

mldeg⁡(MI,J,K)=mldeg⁡(X(I,I+J))(I+J−2I−1)K+O(((I+J−2I−1)−1)K).\operatorname{mldeg}(M_{I,J,K}) = \operatorname{mldeg}(X(I,I+J)) \binom{I+J-2}{I-1}^{K} + O\left(\left(\binom{I+J-2}{I-1}-1\right)^{K}\right).

Proof. By [8 Thm. 1.3], β(MA)\beta(M_A) is the maximum likelihood degree of the independence model of an I×JI \times J table scaled by AA, whose degree is (I+J−2I−1)\binom{I+J-2}{I-1}. By [2 Thm. 2] the maximum likelihood degree is smaller than the degree exactly when the scaling is a zero of the principal AA-determinant, which for PI−1×PJ−1\mathbb{P}^{I-1} \times\mathbb{P}^{J-1} is the product of all minors of the scaling matrix [8 Prop. 2.1]. A point of Gr⁡(I,I+J)\operatorname{Gr}(I,I+J) with all Plücker coordinates nonzero is the row space of a unique matrix [Id∣A][I_d \mid A], whose maximal minors are, up to sign, the minors of AA; scaling the columns of [Id∣A][I_d \mid A] and then restoring the identity block by scaling rows acts on AA by row and column scaling. Hence the set of A∈TI,JA \in T_{I,J} with all minors nonzero is X(I,I+J)X(I,I+J), and its Euler characteristic is the coefficient of the largest base in Theorem 2.1. Since dim⁡X(I,I+J)=(I−1)(J−1)\dim X(I,I+J) = (I-1)(J-1), Huh’s theorem gives (−1)(I−1)(J−1)χ(X(I,I+J))=mldeg⁡(X(I,I+J))(-1)^{(I-1)(J-1)}\chi(X(I,I+J)) = \operatorname{mldeg}(X(I,I+J)). The other bases in Theorem 2.1 are integers smaller than (I+J−2I−1)\binom{I+J-2}{I-1}, which gives the error term. □

The leading coefficient mldeg⁡(X(I,I+J))\operatorname{mldeg}(X(I,I+J)) is positive in every case in which it is known. For I=2I = 2, X(2,J+2)X(2,J+2) is the moduli space M0,J+2M_{0,J+2} of J+2J+2 points on the line, whose maximum likelihood degree is (J−1)!(J-1)!, the top term of Corollary 4.3. For I=3I = 3, the maximum likelihood degrees of X(3,n)X(3,n) for n=5,…,9n = 5,\ldots,9 are 2, 26, 1272, 188112, and 74570400 [6, 1]; they are the numbers of solutions of the scattering equations on X(3,n)X(3,n). The first three are the last entries of the rows of Table 1, and the next two are the leading coefficients of mldeg⁡(M3,5,K)\operatorname{mldeg}(M_{3,5,K}) and mldeg⁡(M3,6,K)\operatorname{mldeg}(M_{3,6,K}). In particular, a closed formula for the maximum likelihood degree of the 3×J×K3 \times J \times K model would give one for the Euler characteristic of X(3,J+3)X(3,J+3). The duality X(I,n)≅X(n−I,n)X(I,n) \cong X(n-I,n) [6] identifies X(I,I+J)X(I,I+J) with X(J,I+J)X(J,I+J), in accordance with the symmetry of the model in II and JJ; the first case not covered by I≤3I \leq3 is the self-dual X(4,8)X(4,8), which governs the 4×4×K4 \times4 \times K models, and whose maximum likelihood degree was computed numerically to be 5211816, with a certified lower bound of 5211598 [1 Thm. 6.1].

References

References

  1. [1]D. Agostini, T. Brysiewicz, C. Fevola, L. Kühne, B. Sturmfels, and S. Telen, Likelihood degenerations, with an appendix by T. Lam, Adv. Math. 414 (2023), Paper No. 108863, doi:10.1016/j.aim.2023.108863.DOI
  2. [2]C. Améndola, N. Bliss, I. Burke, C. R. Gibbons, M. Helmer, S. Hoşten, E. D. Nash, J. I. Rodriguez, and D. Smolkin, The maximum likelihood degree of toric varieties, J. Symbolic Comput. 92 (2019), 222–242, doi:10.1016/j.jsc.2018.04.016; numbering as in arXiv:1703.02251.DOI
  3. [3]M. W. Birch, Maximum likelihood in three-way contingency tables, J. Roy. Statist. Soc. Ser. B 25 (1963), no. 1, 220–233, doi:10.1111/j.2517-6161.1963.tb00504.x.DOI
  4. [4]Y. M. M. Bishop, S. E. Fienberg, and P. W. Holland, Discrete Multivariate Analysis: Theory and Practice, MIT Press, Cambridge, MA, 1975; reprinted by Springer, New York, 2007, doi:10.1007/978-0-387-72806-3.
  5. [5]T. Briesewicz and A. Maraj, Lawrence lifts, matroids, and maximum likelihood degrees, Algebr. Stat. 16 (2025), no. 2, 217–242, doi:10.2140/astat.2025.16.217; numbering as in arXiv:2310.13064.
  6. [6]F. Cachazo, N. Early, A. Guevara, and S. Mizera, Scattering equations: from tropical Grassmannians to tropical Grassmannians, J. High Energy Phys. (2019), no. 6, 039, doi:10.1007/JHEP06(2019)039.DOI
  7. [7]S. E. Cappell, L. Maxim, and J. L. Shaneson, Euler characteristics of algebraic varieties, Comm. Pure Appl. Math. 61 (2008), no. 3, 409–421, doi:10.1002/cpa.20201; arXiv:math/0606654.DOI
  8. [8]O. Clarke, S. Hoşten, N. Khusrenchuk, and J. Oldekop, Matroid stratification of ML degrees of independence models, Algebr. Stat. 15 (2024), no. 2, 199–223, doi:10.2140/astat.2024.15.199; numbering as in arXiv:2312.10010.DOI
  9. [9]H. H. Crapo, A higher invariant for matroids, J. Combinatorial Theory 2 (1967), 406–417, doi:10.1016/S0021-9800(67)80051-6.DOI
  10. [10]I. Csiszár, I-divergence geometry of probability distributions and minimization problems, Ann. Probab. 3 (1975), no. 1, 146–158, doi:10.1214/aop/1176996454.DOI
  11. [11]W. E. Deming and F. Stephan, On a least squares adjustment of a sampled frequency table when the expected marginal totals are known, Ann. Math. Statist. 11 (1940), no. 4, 427–444, doi:10.1214/aoms/1177731829.DOI
  12. [13]J. Huh, The maximum likelihood degree of a very affine variety, Compos. Math. 149 (2013), no. 8, 1245–1266, doi:10.1112/S0010437X13007057.DOI
  13. [14]S. Sivarajan, An explicit determinantal formula for limits of iterative proportional fitting with a binary variable, hexagon:2611.00705 (2026).
  14. [15]O. Ya. Viro, Some integral calculus based on Euler characteristic, in: Topology and Geometry, Rohlin Seminar, Lecture Notes in Math. 1346, Springer, Berlin, 1988, 127–138, doi:10.1007/BFb0082775.

Paper details

vlnc:W_msqynzpj
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