LORENTZIAN POLYNOMIALS

PETTER BRAND¨ EN´ AND JUNE HUH

ABSTRACT. We study the class of Lorentzian polynomials. The class contains homogeneous stable polynomials as well as volume polynomials of convex bodies and projective varieties. We prove that the Hessian of a nonzero Lorentzian polynomial has exactly one positive eigenvalue at any point on the positive orthant. This property can be seen as an analog of the Hodge–Riemann rela- tions for Lorentzian polynomials. Lorentzian polynomials are intimately connected to matroid theory and negative dependence properties. We show that matroids, and more generally M-convex sets, are characterized by the Lorentzian property, and develop a theory around Lorentzian polynomials. In particular, we pro- vide a large class of linear operators that preserve the Lorentzian property and prove that Lorentzian measures enjoy several negative dependence properties. We also prove that the class of tropicalized Lorentzian polynomials coincides with the class of M-convex functions in the sense of discrete con- vex analysis. The tropical connection is used to produce Lorentzian polynomials from M-convex functions. We give two applications of the general theory. First, we prove that the homogenized multi- variate Tutte polynomial of a matroid is Lorentzian whenever the parameter q satisfies 0 ă q ď 1. Consequences are proofs of the strongest Mason’s conjecture from 1972 and negative dependence properties of the random cluster model in statistical physics. Second, we prove that the multivariate characteristic polynomial of an M-matrix is Lorentzian. This refines a result of Holtz who proved that the coefficients of the characteristic polynomial of an M-matrix form an ultra log-concave se- quence.

CONTENTS

1. Introduction 2 2. Basic theory 6

arXiv:1902.03719v4 [math.CO] 1 Jul 2021 2.1. The space of Lorentzian polynomials6 2.2. Hodge–Riemann relations for Lorentzian polynomials 15 2.3. Independence and negative dependence 17 2.4. Characterizations of Lorentzian polynomials 22 3. Advanced theory 26 3.1. Linear operators preserving Lorentzian polynomials 26 3.2. Matroids, M-convex sets, and Lorentzian polynomials 31 3.3. Valuated matroids, M-convex functions, and Lorentzian polynomials 33 2 PETTER BRAND¨ EN´ AND JUNE HUH

4. Examples and applications 42 4.1. Convex bodies and Lorentzian polynomials 42 4.2. Projective varieties and Lorentzian polynomials 45 4.3. Potts model partition functions and Lorentzian polynomials 47 4.4. M-matrices and Lorentzian polynomials 50 4.5. Lorentzian probability measures 54 References 56

1. INTRODUCTION

d Let Hn be the space of degree d homogeneous polynomials in n variables with real coef- ficients. Inspired by Hodge’s index theorem for projective varieties, we introduce a class of ˚2 2 polynomials with remarkable properties. Let Ln Ď Hn be the open subset of quadratic forms with positive coefficients that have the Lorentzian signature p`, ´,..., ´q. For d larger than 2, we ˚d d define an open subset Ln Ď Hn by setting ˚d d ˚d´1 Ln “ f P Hn | Bif P Ln for all i ,

! ) ˚d where Bi is the partial derivative with respect to the i-th variable. Thus f belongs to Ln if and ˚2 ˚d only if all polynomials of the form Bi1 Bi2 ¨ ¨ ¨ Bid´2 f belongs to Ln. The polynomials in Ln are called strictly Lorentzian, and the limits of strictly Lorentzian polynomials are called Lorentzian. We show that the class of Lorentzian polynomials contains the class of homogeneous stable polynomials (Section 2.1) as well as volume polynomials of convex bodies and projective vari- eties (Sections 4.1 and 4.2). 2 2 Lorentzian polynomials link discrete and continuous notions of convexity. Let Ln Ď Hn be the closed subset of quadratic forms with nonnegative coefficients that have at most one positive ˚2 2 n d eigenvalue, which is the closure of Ln in Hn. We write supppfq Ď N for the support of f P Hn, the set of monomials appearing in f with nonzero coefficients. For d larger than 2, we define d d Ln Ď Hn by setting d d d´1 Ln “ f P Mn | Bif P Ln for all i , d d ! ) where Mn Ď Hn is the set of polynomials with nonnegative coefficients whose supports are M- convex in the sense of discrete convex analysis [Mur03]: For any index i and any α, β P supppfq

whose i-th coordinates satisfy αi ą βi, there is an index j satisfying

αj ă βj and α ´ ei ` ej P supppfq and β ´ ej ` ei P supppfq,

n d d´1 where ei is the i-th standard unit vector in N . Since f P Mn implies Bif P Mn , we have

d d 2 Ln “ f P Mn | Bi1 Bi2 ¨ ¨ ¨ Bid´2 f P Ln for all i1, i2, . . . , id´2 . ! ) LORENTZIAN POLYNOMIALS 3

d d Our central result states that Ln is the set of Lorentzian polynomials in Hn (Theorem 2.25). d ˚d d To show that Ln is contained in the closure of Ln, we construct a Nuij-type homotopy for Ln in d Section 2.1. The construction is used in Section 2.2 to prove that all polynomials in Ln satisfy a formal version of the Hodge–Riemann relations: The Hessian of any nonzero polynomial d in Ln has exactly one positive eigenvalue at any point on the positive orthant. To show that d ˚d Ln contains the closure of Ln, we develop the theory of c-Rayleigh polynomials in Section 2.3. Since homogeneous stable polynomials are Lorentzian, the latter inclusion generalizes a result of Choe et al. that the support of any homogenous multi-affine stable polynomial is the set of bases of a matroid [COSW04]. In Section 2.4, we use the above results to show that the classes of strongly log-concave [Gur09], completely log-concave [AOVI], and Lorentzian polynomials are identical for homogeneous polynomials (Theorem 2.30). This enables us to affirmatively answer two questions of Gurvits on strongly log-concave polynomials (Corollaries 2.31 and 2.32). Lorentzian polynomials are intimately connected to matroid theory and discrete convex anal- ysis. We show that matroids, and more generally M-convex sets, are characterized by the d d Lorentzian property. Let PHn be the projectivization of the vector space Hn, and let LJ be the set d d ˚d d of polynomials in Ln with nonempty support J. We denote the images of Ln, Ln, and LJ in PHn d ˚d by PLn, PLn, and PLJ respectively, and write d PLn “ PLJ, J ž where the union is over all nonempty M-convex subsets of the d-th discrete simplex in Nn. The d d d space PLn is homeomorphic to the intersection of Ln with the unit sphere in Hn for the Euclidean d norm on the coefficients. We prove that PLn is a compact contractible set with contractible ˚d 1 interior PLn (Theorem 2.28). In addition, we show that PLJ is nonempty and contractible for d every nonempty M-convex set J (Theorem 3.10 and Proposition 3.25). Similarly, writing Hn d for the space of multi-affine degree d homogeneous polynomials in n variables and Ln for the corresponding set of multi-affine Lorentzian polynomials, we have d PLn “ PLB, B ž d where the union is over all rank d matroids on the n-element set rns. The space PLn is compact and contractible, and PLB is nonempty and contractible for every matroid B (Remark 3.6). The latter fact contrasts the case of stable polynomials. For example, there is no stable polynomial whose support is the set of bases of the Fano plane [Bra07¨ ]. In Section 3.1, we describe a large class of linear operators preserving the class of Lorentzian polynomials, thus providing a toolbox for working with Lorentzian polynomials. We give a Lorentzian analog of a theorem of Borcea and Brand¨ en´ for stable polynomials [BB09], who char- acterized linear operators preserving stable polynomials (Theorem 3.2). It follows from our re- sult that any homogeneous linear operator that preserves stable polynomials and polynomials with nonnegative coefficients also preserves Lorentzian polynomials (Theorem 3.4).

1 d We conjecture that PLn is homeomorphic to the closed Euclidean ball of the same dimension (Conjecture 2.29). 4 PETTER BRAND¨ EN´ AND JUNE HUH

In Section 3.3, we strengthen the connection between Lorentzian polynomials and discrete convex analysis. For a function ν : Nn Ñ R Y t8u, we write dompνq Ď Nn for the effective domain of ν, the subset of Nn where ν is finite. For a positive real parameter q, we consider the generating function qνpαq f ν pwq “ wα, w “ pw , . . . , w q. q α! 1 n α ν Pdomÿ p q ν The main result here is Theorem 3.14, which states that fq is a Lorentzian polynomial for all 0 ă q ď 1 if and only if the function ν is M-convex in the sense of discrete convex analysis

[Mur03]: For any index i and any α, β P dompνq whose i-th coordinates satisfy αi ą βi, there is an index j satisfying

αj ă βj and νpαq ` νpβq ě νpα ´ ei ` ejq ` νpβ ´ ej ` eiq.

n 1 α In particular, J Ď N is M-convex if and only if its exponential generating function αPJ α! w is a Lorentzian polynomial (Theorem 3.10). Another special case of Theorem 3.14 is the statement ř that a homogeneous polynomial with nonnegative coefficients is Lorentzian if the natural loga- rithms of its normalized coefficients form an M-concave function (Corollary 3.16). Working over the field of formal Puiseux series K, we show that the tropicalization of any Lorentzian polyno- mial over K is an M-convex function, and that all M-convex functions are limits of tropicaliza- tions of Lorentzian polynomials over K (Corollary 3.28). This generalizes a result of Brand¨ en´ [Bra10¨ ], who showed that the tropicalization of any homogeneous stable polynomial over K is M-convex.2 In particular, for any matroid M with the set of bases B, the Dressian of all valuated matroids on M can be identified with the tropicalization of the space of Lorentzian polynomials over K with support B. In Sections 4.1 and 4.2, we show that the volume polynomials of convex bodies and projective d varieties are Lorentzian. It follows that, for any convex bodies K1,..., Kn in R , the set of all α P Nn satisfying the conditions

α1 ` ¨ ¨ ¨ ` αn “ d and V pK1,..., K1,..., Kn,..., Knq ‰ 0

α1 αn

is M-convex, where the symbol V stands for thelooooomooooon mixed volumelooooomooooon of convex bodies in Rd. Similarly, for any d-dimensional projective variety Y and any nef divisors H1,..., Hn on Y , the set of all α P Nn satisfying the conditions

α1 ` ¨ ¨ ¨ ` αn “ d and pH1 ¨ ... ¨ H1 ¨ ... ¨ Hn ¨ ... ¨ Hnq ‰ 0

α1 αn is M-convex, where the symbol ¨ stands forlooooomooooon the intersectionloooooomoooooon product of Cartier divisors on Y . The problem of finding a Lorentzian polynomial that is not a volume polynomial remains open. For a precise formulation, see Question 4.9.

2In [Bra10¨ ], the field of formal Puiseux series with real exponents was used. The tropicalization used in [Bra10¨ ] differs from ours by a sign. LORENTZIAN POLYNOMIALS 5

In Section 4.3, we use the basic theory developed in Section2 to show that the homogenized multivariate Tutte polynomial of any matroid is Lorentzian. We use the Lorentzian property to prove a conjecture of Mason from 1972 on the enumeration of independent sets [Mas72]: For any matroid M on rns and any positive integer k, I pMq2 I pMq I pMq k ě k`1 k´1 , n 2 n n k k`1 k´1 where IkpMq is the number of k-element` ˘ independent` ˘ sets` of˘ M. More generally, the Lorentzian property reveals several inequalities satisfied by the coefficients of the classical Tutte polynomial

rkMprnsq´rkMpAq |A|´rkMpAq TMpx, yq “ px ´ 1q py ´ 1q , A n ÿĎr s n where rkM : t0, 1u Ñ N is the rank function of M. For example, if we write q n wrkMprnsqT 1 ` , 1 ` w “ ckpMqwk, M w q k“0 ´ ¯ ÿ k 3 then the sequence cq pMq is ultra log-concave for every 0 ď q ď 1. In Section 4.4, we show that the multivariate characteristic polynomial of any M-matrix is Lorentzian.4 This strengthens a theorem of Holtz [Hol05], who proved that the coefficients of the characteristic polynomial of any M-matrix form an ultra log-concave sequence. In Section 4.5, we define a class of discrete probability measures, called Lorentzian measures, properly containing the class of strongly Rayleigh measures studied in [BBL09]. We show that Lorentzian measures enjoy several negative dependence properties and prove that the class of Lorentzian measures is closed under the symmetric exclusion process. As an example, we show n that the uniform measure µM on t0, 1u concentrated on the independent sets of a matroid M on rns is Lorentzian (Proposition 4.25). A conjecture of Kahn [Kah00] and Grimmett–Winkler [GW04] states that, for any graphic matroid M and distinct elements i and j, PrpF contains i and jq ď PrpF contains iq PrpF contains jq,

3Nima Anari, Kuikui Liu, Shayan Oveis Gharan and Cynthia Vinzant have independently developed methods that partially overlap with our work in a series of papers [AOVI, ALOVII, ALOVIII]. They study the class of completely log-concave polynomials. For homogenous polynomials this class agrees with the class of Lorentzian polynomials, see Theorem 2.30 in this paper. The main overlap is an independent proof of Mason’s conjecture in [ALOVIII]. The man- uscript [BH], which is not intended for publication, contains a short self-contained proof of Mason’s conjecture which was published on arXiv simultaneously as [ALOVIII]. In addition, the authors of [AOVI] prove that the basis generating polynomial of any matroid is completely log-concave, using results of Adiprasito, Huh, and Katz [AHK18]. An equiva- lent statement on the Hessian of the basis generating polynomial can be found in [HW17, Remark 15]. A self-contained proof of the complete log-concavity of the basis generating polynomial, based on an implication similar to p3q ñ p1q of Theorem 2.30 in this paper, appears in [ALOVII, Section 5.1]. The authors of [ALOVII] apply these results to design an FPRAS to count the number of bases of any matroid given by an independent set oracle, and to prove the conjecture of Mihail and Vazirani that the bases exchange graph of any matroid has expansion at least 1. 4An nˆn matrix is an M-matrix if all the off-diagonal entries are nonpositive and all the principal minors are positive. The class of M-matrices shares many properties of hermitian positive definite matrices and appears in mathematical economics and computational biology [BP94]. 6 PETTER BRAND¨ EN´ AND JUNE HUH where F is an independent set of M chosen uniformly at random. The Lorentzian property of the measure µM shows that, for any matroid M and distinct elements i and j,

PrpF contains i and jq ď 2PrpF contains iq PrpF contains jq,

where F is an independent set of M chosen uniformly at random. Acknowledgments. Petter Brand¨ en´ is a Wallenberg Academy Fellow supported by the Knut and Alice Wallenberg Foundation and Vetenskapsradet.˚ June Huh was supported by NSF Grant DMS-1638352 and the Ellentuck Fund. Special thanks go to anonymous referees and Claus Scheiderer’s reading group, whose valuable comments significantly improved the quality of the paper.

2. BASICTHEORY

2.1. The space of Lorentzian polynomials. Let n and d be nonnegative integers, and set rns “ d t1, . . . , nu. We write Hn for the set of degree d homogeneous polynomials in Rrw1, . . . , wns. d d d We define a topology on Hn using the Euclidean norm for the coefficients, and write Pn Ď Hn for the open subset of polynomials all of whose coefficients are positive. The Hessian of f P Rrw1, . . . , wns is the symmetric matrix n Hf pwq “ BiBjf , i,j“1 ´ ¯ B n where Bi stands for the partial derivative . For α P , we write Bwi N n n α “ αiei and |α|1 “ αi, i“1 i“1 ÿ ÿ n where αi is a nonnegative integer and ei is the standard unit vector in N , and set

α α1 αn α α1 αn w “ w1 ¨ ¨ ¨ wn and B “B1 ¨ ¨ ¨ Bn .

d n We define the d-th discrete simplex ∆n Ď N by

d n ∆n “ α P N | |α|1 “ d , ! ) and define the Boolean cube t0, 1un Ď Nn by

n n t0, 1u “ ei P N | S Ď rns . # iPS + ÿ The intersection of the d-th discrete simplex and the Boolean cube will be denoted n “ t0, 1un X ∆d . d n „  n n The cardinality of d is the binomial coefficient d . We often identify a subset S of rns with the e n wS w zero-one vector i“PS‰ i in N . For example, we write` ˘ for the square-free monomial iPS i. ř ś LORENTZIAN POLYNOMIALS 7

˚0 0 ˚1 1 Definition 2.1 (Lorentzian polynomials). We set Ln “ Pn, Ln “ Pn, and

˚2 2 Ln “ f P Pn | Hf is nonsingular and has exactly one positive eigenvalue . ! ) ˚d For d larger than 2, we define Ln recursively by setting

˚d d ˚d´1 Ln “ f P Pn | Bif P Ln for all i P rns . ! ) ˚d The polynomials in Ln are called strictly Lorentzian, and the limits of strictly Lorentzian polyno- mials are called Lorentzian.

˚d d ˚2 Clearly, Ln is an open subset of Hn, and the space Ln may be identified with the set of n ˆ n symmetric matrices with positive entries that have the Lorentzian signature p`, ´,..., ´q. Un- winding the recursive definition, we have

˚d d α ˚2 d´2 Ln “ f P Pn | B f P Ln for every α P ∆n . ! ) ˚d Proposition 2.2 below on stable polynomials shows that Ln is nonempty for every n and d. An important subclass of Lorentzian polynomials is homogeneous stable polynomials, which play a guiding role in many of our proofs. Recall that a polynomial f in Rrw1, . . . , wns is stable if f is non-vanishing on Hn or identically zero, where H is the open upper half plane in C. d Let Sn be the set of degree d homogeneous stable polynomials in n variables with nonnegative d d coefficients. Hurwitz’s theorem shows that Sn is a closed subset of Hn [Wag11, Section 2]. When f is homogeneous and has nonnegative coefficients, the stability of f is equivalent to any one of the following statements on univariate polynomials in the variable x [BBL09, Theorem 4.5]:

n n – For any u P Rą0, fpxu ´ vq has only real zeros for all v P R . n n – For some u P Rą0, fpxu ´ vq has only real zeros for all v P R . n n – For any u P Rě0 with fpuq ą 0, fpxu ´ vq has only real zeros for all v P R . n n – For some u P Rě0 with fpuq ą 0, fpxu ´ vq has only real zeros for all v P R .

We refer to [Wag11] and [Pem12] for background on the class of stable polynomials. We will use d d the fact that any polynomial f P Sn is the limit of polynomials in the interior of Sn, that is, of strictly stable polynomials [Nui68].

d Proposition 2.2. Any polynomial in Sn is Lorentzian.

d ˚d Proof. We show that the interior of Sn is a subset of Ln by induction on d. When d “ 2, the statement follows from Lemma 2.5 below. The general case follows from the fact that Bi is an d d´1 open map sending Sn to Sn [Wag11, Lemma 2.4]. 

All the nonzero coefficients of a homogeneous stable polynomial have the same sign [COSW04, Theorem 6.1]. Thus, any homogeneous stable polynomial is a constant multiple of a Lorentzian 8 PETTER BRAND¨ EN´ AND JUNE HUH polynomial. For example, determinantal polynomials of the form

fpw1, . . . , wnq “ detpw1A1 ` ¨ ¨ ¨ ` wnAnq, where A1,...,An are positive semidefinite matrices, are stable [BB08, Proposition 2.4], and hence Lorentzian.

Example 2.3. Consider the homogeneous bivariate polynomial with positive coefficients d k d´k f “ akw1 w2 . k“0 ÿ Computing the partial derivatives of f reveals that f is strictly Lorentzian if and only if 2 a ak 1 ak 1 k ą ´ ` for all 0 ă k ă d. d 2 d d k k´1 k`1

On the other hand, f is stable` if˘ and only` if˘ the` univariate˘ polynomial f|w2“1 has only real zeros. Thus, a Lorentzian polynomial need not be stable. For example, consider the cubic form 3 2 2 3 f “ 2w1 ` 12w1w2 ` 18w1w2 ` θw2, where θ is a real parameter. A straightforward computation shows that f is Lorentzian if and only if 0 ď θ ď 9, and f is stable if and only if 0 ď θ ď 8.

˚d d Example 2.4. Clearly, if f is in the closure of Ln in Hn, then f has nonnegative coefficients and α d´2 B f has at most one positive eigenvalue for every α P ∆n . 3 3 The bivariate cubic f “ w1 ` w2 shows that the converse fails. In this case, B1f and B2f are Lorentzian, but f is not Lorentzian.

˚2 We give alternative characterizations of Ln. Similar arguments were given in [Gre81] and [COSW04, Theorem 5.3].

2 Lemma 2.5. The following conditions are equivalent for any f P Pn. ˚2 (1) The Hessian of f has the Lorentzian signature p`, ´,..., ´q, that is, f P Ln. n T 2 T T n (2) For any nonzero u P Rě0, pu Hf vq ą pu Hf uqpv Hf vq for any v P R not parallel to u. n T 2 T T n (3) For some u P Rě0, pu Hf vq ą pu Hf uqpv Hf vq for any v P R not parallel to u. n (4) For any nonzero u P Rě0, the univariate polynomial fpxu ´ vq in x has two distinct real zeros for any v P Rn not parallel to u. n (5) For some u P Rě0, the univariate polynomial fpxu ´ vq in x has two distinct real zeros for any v P Rn not parallel to u.

It follows that a quadratic form with nonnegative coefficients is strictly Lorentzian if and only if it is strictly stable. Thus, a quadratic form with nonnegative coefficients is Lorentzian if and only if it is stable. LORENTZIAN POLYNOMIALS 9

T Proof. We prove p1q ñ p2q. Since all the entries of Hf are positive, u Hf u ą 0 for any nonzero n n u P Rě0. By Cauchy’s interlacing theorem, for any v P R not parallel to u, the restriction of Hf to the plane spanned by u, v has signature p`, ´q. It follows that

T T u Hf u u Hf v T T T 2 det T T “ pu Hf uqpv Hf vq ´ pu Hf vq ă 0. ˜ u Hf v v Hf v ¸

We prove p3q ñ p1q. Let u be the nonnegative vector in the statement p3q. Then Hf is negative n T 2 T definite on the hyperplane tv P R | u Hf v “ 0u. Since f P Pn, we have u Hf u ą 0, and hence Hf has the Lorentzian signature. 1 The remaining implications follows from the fact that the univariate polynomial 2 fpxu ´ vq T 2 T T has the discriminant pu Hf vq ´ pu Hf uqpv Hf vq. 

Matroid theory captures various combinatorial notions of independence. A matroid M on rns is a nonempty family of subsets B of rns, called the set of bases of M, that satisfies the exchange property:

For any B1,B2 P B and i P B1zB2, there is j P B2zB1 such that pB1ziq Y j P B.

We refer to [Oxl11] for background on matroid theory. More generally, following [Mur03], we define a subset J Ď Nn to be M-convex if it satisfies any one of the following equivalent condi- tions5:

– For any α, β P J and any index i satisfying αi ą βi, there is an index j satisfying

αj ă βj and α ´ ei ` ej P J.

– For any α, β P J and any index i satisfying αi ą βi, there is an index j satisfying

αj ă βj and α ´ ei ` ej P J and β ´ ej ` ei P J.

The first condition is called the exchange property for M-convex sets, and the second condition is called the symmetric exchange property for M-convex sets. A proof of the equivalence can be found in [Mur03, Chapter 4]. Note that any M-convex subset of Nn is necessarily contained d in the discrete simplex ∆n for some d. We refer to [Mur03] for a comprehensive treatment of M-convex sets.

Let f be a polynomial in Rrw1, . . . , wns. We write f in the normalized form n cα f “ wα, where α! “ α !. α! i αP n i“1 ÿN ź

5The class of M-convex sets is essentially identical to the class of generalized polymatroids in the sense of [Fuj05]. Some other notions in the literature that are equivalent to M-convex sets are integral polymatroids [Wel76], discrete polymatroids [HH03], and integral generalized permutohedras [Pos09]. We refer to [Mur03, Section 1.3] and [Mur03, Section 4.7] for more details. 10 PETTER BRAND¨ EN´ AND JUNE HUH

The support of the polynomial f is the subset of Nn defined by

n supppfq “ α P N | cα ‰ 0 . ! ) d We write Mn for the set of all degree d homogeneous polynomials in Rě0rw1, . . . , wns whose supports are M-convex. Note that, in our convention, the empty subset of Nn is an M-convex d d d´1 set. Thus, the zero polynomial belongs to Mn, and f P Mn implies Bif P Mn . It follows from d d [Bra07¨ , Theorem 3.2] that Sn Ď Mn.

0 0 1 1 2 2 Definition 2.6. We set Ln “ Sn, Ln “ Sn, and Ln “ Sn. For d larger than 2, we define

d d d´1 d α 2 d´2 Ln “ f P Mn | Bif P Ln for all i P rns “ f P Mn | B f P Ln for every α P ∆n . ! ) ! ) d ˚d d ˚d d Clearly, Ln contains Ln. In Theorem 2.25, we show that Ln is the closure of Ln in Hn. In other d words, Ln is exactly the set of degree d Lorentzian polynomials in n variables. In this section, ˚d d we show that Ln is contractible and its closure contains Ln. The following proposition plays a d central role in our analysis of Ln. Analogous statements, in the context of hyperbolic polyno- mials and stable polynomials, appear in [Nui68] and [LS81]. We fix a degree d homogeneous polynomial f in n variables and indices i, j in rns.

d d Proposition 2.7. If f P Ln, then 1 ` θwiBj f P Ln for every nonnegative real number θ. ` ˘ We prepare the proof of Proposition 2.7 with two lemmas.

d d Lemma 2.8. If f P Mn, then 1 ` θwiBj f P Mn for every nonnegative real number θ. ` ˘ Proof. We may suppose θ “ 1 and j “ n. We use two combinatorial lemmas from [KMT07].

Introduce a new variable wn`1, and set

d 1 gpw , . . . , w , w q “ fpw , . . . , w ` w q “ wk Bkfpw , . . . , w q. 1 n n`1 1 n n`1 k! n`1 n 1 n k“0 ÿ By [KMT07, Lemma 6], the support of g is M-convex. In terms of [KMT07], the support of g is obtained from the support of f by an elementary splitting, and the operation of splitting d preserves M-convexity. Therefore, g belongs to Mn`1. Since the intersection of an M-convex set with a cartesian product of intervals is M-convex, it follows that

d 1 ` wn`1Bn f P Mn`1.

By [KMT07, Lemma 9], the above displayed` inclusion˘ implies

d 1 ` wiBn f P Mn. ` ˘ In terms of [KMT07], the support of 1 ` wiBn f is obtained from the support of 1 ` wn`1Bn f M by an elementary aggregation, and the` operation˘ of aggregation preserves -convexity.` ˘ LORENTZIAN POLYNOMIALS 11

For stable polynomials f and g in Rrw1, . . . , wns, we define a relation f ă g by

f ă g ðñ g ` wn`1f is a stable polynomial in Rrw1, . . . , wn, wn`1s. If f and g are univariate polynomials with leading coefficients of the same sign, then f ă g if and only if the zeros of f interlace the zeros of g [BB10, Lemma 2.2]. In general, we have

n n f ă g ðñ fpxu ´ vq ă gpxu ´ vq for all u P Rą0 and v P R . For later use, we record here basic properties of stable polynomials and the relation ă.

Lemma 2.9. Let f, g1, g2, h1, h2 be stable polynomials satisfying h1 ă f ă g1 and h2 ă f ă g2.

(1) The derivative B1f is stable and B1f ă f.

(2) The diagonalization fpw1, w1, w3, . . . , wnq is stable.

n (3) The dilation fpa1w1, . . . , anwnq is stable for any a P Rě0.

(4) If f is not identically zero, then f ă θ1g1 ` θ2g2 for any θ1, θ2 ě 0.

(5) If f is not identically zero, then θ1h1 ` θ2h2 ă f for any θ1, θ2 ě 0.

The statement B1f ă f appears, for example, in [BBL09, Section 4]. It follows that, if f is stable, then p1 ` θwiBjqf is stable for every nonnegative real number θ. The remaining proof of Lemma 2.9 can be found in [Wag11, Section 2] and [BB10, Section 2].

Proof of Proposition 2.7. When d “ 2, Lemma 2.9 implies Proposition 2.7. Suppose d ě 3, and set

g “ 1 ` θwiBj f.

By Lemma 2.8, the support of g is M-convex.` Therefore,˘ it is enough to prove that Bαg is stable d´2 for all α P ∆n . We give separate arguments when αi “ 0 and αi ą 0. If αi “ 0, then

α α α`ej B g “B f ` θwiB f.

α α In this case, (1), (2), and (3) of Lemma 2.9 for B f show that B g is stable. If αi ą 0, then

α α α´ei`ej α`ej B g “B f ` θαiB f ` θwiB f

α´ei α´ei α´ei “Bi B f ` θαiBj B f ` θwiBiBj B f . ´ ¯ ´ ¯ ´ ¯ In this case, (1) of Lemma 2.9 applies to the stable polynomials Bαf and Bα´ei`ej f:

α´ei α´ei α´ei α´ei BiBj B f ă Bi B f and BiBj B f ă Bj B f . ´ ¯ ´ ¯ ´ ¯ ´ ¯ Therefore, unless Bα`ej f is identically zero, Bαg is stable by (2) and (4) of Lemma 2.9.

α`ej It remains to prove that, whenever αi is positive and B f is identically zero,

α´ei α´ei Bi B f ` φ Bj B f is stable for every nonnegative real number φ.

´ ¯α´ei ´ 3 ¯ Since the cubic form B f is in Ln, it is enough to prove the statement when d “ 3 and α “ ei. 12 PETTER BRAND¨ EN´ AND JUNE HUH

3 We show that, if f is in Ln and BiBjf is identically zero, then

Bif ` φ Bjf is stable for every nonnegative real number φ.

The statement is clear when pBifqpBjfq is identically zero. If otherwise, there are monomials of the form wiwi1 wi2 and wjwj1 wj2 in the support of f. We apply the symmetric exchange

property to the support of f, the monomials wiwi1 wi2 , wjwj1 wj2 , and the variable wi: We see

that the monomial wjwi1 wi2 must be in the support of f, since no monomial in the support of f

is divisible by wiwj. For a positive real parameter s, set

hs “ 1 ` swiBi1 f.

3 Since BiBi1 f is not identically zero, the argument` in the˘ first paragraph shows that hs is in Ln. Similarly, since BiBjhs is not identically zero, we have 3 1 ` φwiBj hs P Ln for every nonnegative real number φ. Since stability is a closed` condition,˘ it follows that

lim Bi hs ` φwiBjhs “Bif ` φ Bjf is stable for every nonnegative real number φ. sÑ0  ´ ¯ d We use Proposition 2.7 to show that any nonnegative linear change of variables preserves Ln.

d d Theorem 2.10. If fpwq P Ln, then fpAvq P Lm for any n ˆ m matrix A with nonnegative entries.

d Proof. Fix f “ fpw1, . . . , wnq in Ln. Note that Theorem 2.10 follows from its three special cases:

d (I) the elementary splitting fpw1, . . . , wn´1, wn ` wn`1q is in Ln`1, d (II) the dilation fpw1, . . . , wn´1, θwnq is in Ln for any θ ě 0, d (III) the diagonalization fpw1, . . . , wn´2, wn´1, wn´1q is in Ln´1, As observed in the proof of Lemma 2.8, an elementary splitting preserves M-convexity: d fpw1, . . . , wn´1, wn ` wn`1q P Mn`1. Therefore,6 the first statement follows from Proposition 2.7: k wn`1Bn d lim 1 ` f “ fpw1, . . . , wn´1, wn ` wn`1q P Ln`1. kÑ8 k ˆ ˙ For the second statement, note from the definition of M-convexity that d fpw1, . . . , wn´1, 0q P Mn. Thus the second statement for θ “ 0 follows from the case θ ą 0, which is trivial to verify. The proof of the third statement is similar to that of the first statement. As observed in the proof of Lemma 2.8, an elementary aggregation preserves M-convexity, and hence d fpw1, . . . , wn´1, wn´1 ` wnq P Mn.

6 d d It is necessary to check the inclusion in Mn`1 in advance because we have not yet proved that Ln`1 is closed. LORENTZIAN POLYNOMIALS 13

Therefore, Proposition 2.7 implies that k wn´1Bn d lim 1 ` f “ fpw1, . . . , wn´1, wn´1 ` wnq P Ln. kÑ8 k ˆ ˙ By the second statement, we may substitute wn in the displayed equation by zero. 

Theorem 2.10 can be used to show that taking directional derivatives in nonnegative direc- d d´1 tions takes polynomials in Ln to polynomials in Ln .

d n d´1 Corollary 2.11. If f P Ln, then i“1 aiBif P Ln for any a1, . . . , an ě 0.

ř n Proof. We apply Theorem 2.10 to f and the matrix with column vectors e1, . . . , en and i“1 aiei: d d´1 g :“ fpw1 ` a1wn`1, . . . , wn ` anwn`1q P Ln`1, and hence Bn`1g P Ln`1. ř

Applying Theorem 2.10 to Bn`1g and the matrix with column vectors e1, . . . , en and 0, we get n d´1 Bn`1g|wn`1“0 “ aiBif P Ln .  i“1 ÿ Let θ be a nonnegative real parameter. We define a linear operator Tnpθ, ´q by n´1 d Tnpθ, fq “ 1 ` θwiBn f. ˜ i“1 ¸ ź d `d ˘ d d By Proposition 2.7, if f P Ln, then Tnpθ, fq P Ln. In addition, if f P Pn, then Tnpθ, fq P Pn. Most importantly, the operator Tn satisfies the following Nuij-type homotopy lemma. For a similar argument in the context of hyperbolic polynomials, see the proof of the main theorem in [Nui68].

d d ˚d Lemma 2.12. If f P Ln X Pn, then Tnpθ, fq P Ln for every positive real number θ.

n n Proof. Let ei be the i-th standard unit vector in R , and let v be any vector in R not parallel to en. From here on, in this proof, all polynomials are restricted to the line xen ´ v and considered as univariate polynomials in x. d´2 Let α be an arbitrary element of ∆n . By Lemma 2.5, it is enough to show that the quadratic α polynomial B Tnpθ, fq has two distinct real zeros. Using Proposition 2.7, we can deduce the preceding statement from the following claims:

α α (I) If B f has two distinct real zeros, then B 1 ` θwiBn f has two distinct zeros.

α d (II) If vi is nonzero, then B 1 ` θwiBn f has` two distinct˘ real zeros.

α We first prove (I). Suppose B f `has two distinct˘ real zeros, and set g “ 1 ` θwiBn f. Note that

α α α´ei`en α`en B g “B f ` θαiB f ` θwiB f. ` ˘

Let c be the unique zero of Bα`en f. Since c strictly interlaces two distinct zeros of Bαf, we have α B f|x“c ă 0. 14 PETTER BRAND¨ EN´ AND JUNE HUH

Similarly, since Bα´ei`en f has only real zeros and Bα`en f ă Bα´ei`en f, we have

α´ei`en B f|x“c ď 0.

α α Thus B g|x“c ă 0, and hence B g has two distinct real zeros. This completes the proof of (I). Before proving (II), we strengthen (I) as follows:

(III) A multiple zero of Bαg is necessarily a multiple zero of Bαf.

Suppose Bαg has a multiple zero. Using (I), we know that Bαf has a multiple zero, say c. Clearly, c must be also a zero of Bα`en f. Since c interlaces the two (not necessarily distinct) zeros of Bα´ei`en f, we have

α α´ei`en B g|x“c “ θαiB f|x“c ď 0. Therefore, if c is not a zero of Bαg, then Bαg has two distinct zeros, contradicting the hypothesis that Bαg has a multiple zero. This completes the proof of (III). α d We prove (II). Suppose B 1 ` θwiBn f has a multiple zero, say c. Using (III), we know that

` ˘ α k the number c is a multiple zero of B 1 ` θwiBn f for all 0 ď k ď d.

Expanding the k-th power and using the linearity` of Bα, we deduce˘ that

α k k the number c is a zero of B wi Bn f for all 0 ď k ď d.

α αi`2 αi`2 However, since f has positive coefficients, the value of B wi Bn f at c is a positive multiple 2 of vi , and hence vi must be zero. This completes the proof of (II). 

We use Lemma 2.12 to prove the main result of this subsection.

˚d d d Theorem 2.13. The closure of Ln in Hn contains Ln.

d Proof. Let f be a polynomial in Ln that is not identically zero, and let θ be a real parameter satisfying 0 ď θ ď 1. By Theorem 2.10, we have

1 d Spθ, fq :“ f p1 ´ θqw1 ` θpw1 ` ¨ ¨ ¨ ` wnq,..., p1 ´ θqwn ` θpw1 ` ¨ ¨ ¨ ` wnq P Ln, |f|1 ´ ¯ d where |f|1 is the sum of all coefficients of f. Since Spθ, fq belongs to Pn when 0 ă θ ď 1, Lemma 2.12 shows that we have a homotopy

˚d Tn θ, Spθ, fq P Ln, 0 ă θ ď 1, ´ ¯ d ˚d d that deforms f to the polynomial Tn 1, pw1 ` ¨ ¨ ¨ ` wnq . It follows that the closure of Ln in Hn Ld contains n. ` ˘ 

˚d d d We show in Theorem 2.25 that the closure of Ln in Hn is, in fact, equal to Ln. LORENTZIAN POLYNOMIALS 15

2.2. Hodge–Riemann relations for Lorentzian polynomials. Let f be a nonzero degree d ě 2

homogeneous polynomial with nonnegative coefficients in variables w1, . . . , wn. The following proposition may be seen as an analog of the Hodge–Riemann relations for homogeneous stable polynomials.7

d n Proposition 2.14. If f is in Snz0, then Hf pwq has exactly one positive eigenvalue for all w P Rą0. d n Moreover, if f is in the interior of Sn, then Hf pwq is nonsingular for all w P Rą0.

n Proof. Fix a vector w P Rą0. By Lemma 2.5, the Hessian of f has exactly one positive eigenvalue at w if and only if the following quadratic polynomial in z is stable:

T z Hf pwqz “ zizjBiBjfpwq. 1ďi,jďn ÿ The above is the quadratic part of the stable polynomial with nonnegative coefficients fpz ` wq, and hence is stable by [BBL09, Lemma 4.16]. d d Moreover, if f is strictly stable, then f “ f ˘pw1 `¨ ¨ ¨`wnq is stable for all sufficiently small positive . Therefore, by the result obtained in the previous paragraph, the matrix

d´2 d´2 Hf pwq “ Hf pwq ˘ dpd ´ 1q diagpw1 , . . . , wn q has exactly one positive eigenvalue for all sufficiently small positive , and hence Hf pwq is nonsingular. 

In Theorem 2.16, we extend the above result to Lorentzian polynomials.

n Lemma 2.15. If HBif pwq has exactly one positive eigenvalue for every i P rns and w P Rą0, then n n ker Hf pwq “ ker HBif pwq for every w P Rą0. i“1 č n Proof. We may suppose d ě 3. Fix w P Rą0, and write Hf for Hf pwq. We will use Euler’s formula for homogeneous functions: n df “ wi Bif. i“1 ÿ It follows that the Hessians of f and Bif satisfy the relation n

pd ´ 2q Hf “ wi HBif , i“1 ÿ

and hence the kernel of Hf contains the intersection of the kernels of HBif .

For the other inclusion, let z be a vector in the kernel of Hf . By Euler’s formula again,

T T pd ´ 2q ei Hf “ w HBif for every i P rns,

7We refer to [Huh18] for a survey of the Hodge–Riemann relations in combinatorial contexts. 16 PETTER BRAND¨ EN´ AND JUNE HUH

T T and hence w HBif z “ 0 for every i P rns. We have w HBif w ą 0 because Bif is nonzero and

has nonnegative coefficients. Since HBif pwq has exactly one positive eigenvalue, it follows that T HBif is negative semidefinite on the kernel of w HBif . In particular,

T z HBif z ď 0, with equality if and only if HBif z “ 0.

To conclude, we write zero as the positive linear combination n T T 0 “ pd ´ 2q z Hf z “ wi z HBif z . i“1 ´ ¯ ÿ ´ ¯ Since every summand in the right-hand side is non-positive by the previous analysis, we must T have z HBif z “ 0 for every i P rns, and hence HBif z “ 0 for every i P .rns. 

We now prove an analog of the Hodge–Riemann relation for Lorentzian polynomials. When f is the volume polynomial of a projective variety as defined in Section 4.2, then the one posi- tive eigenvalue condition for the Hessian of f at w is equivalent to the validity of the Hodge– Riemann relations on the space of divisor classes of the projective variety with respect to the polarization corresponding to w.

Theorem 2.16. Let f be a nonzero homogeneous polynomial in Rrw1, . . . , wns of degree d ě 2. ˚d n (1) If f is in Ln, then Hf pwq is nonsingular for all w P Rą0. d n (2) If f is in Ln, then Hf pwq has exactly one positive eigenvalue for all w P Rą0.

d ˚d Proof. By Theorem 2.13, Ln is in the closure of Ln. Note that, for any nonzero polynomial f of degree d ě 2 with nonnegative coefficients, Hf pwq has at least one positive eigenvalue for any n ˚d w P Rą0. Therefore, we may suppose f P Ln in (2). We prove (1) and (2) simultaneously by induction on d under this assumption. The base case d “ 2 is trivial. We suppose that d ě 3 and ˚d´1 that the theorem holds for Ln . ˚d That (1) holds for Ln follows from induction and Lemma 2.15. Using Proposition 2.14, we ˚d ˚d see that (2) holds for stable polynomials in Ln. Since Ln is connected by Theorem 2.13, the continuity of eigenvalues and the validity of (1) together implies (2). 

Theorem 2.16, when combined with the following proposition, shows that all polynomials in d Ln share a negative dependence property. The negative dependence property will be systemat- ically studied in the following section.

n Proposition 2.17. If Hf pwq has exactly one positive eigenvalue for all w P Rą0, then 1 fpwq B B fpwq ď 2 1 ´ B fpwq B fpwq for all w P n and i, j P rns. i j d i j Rě0 ´ ¯ n Proof. Fix w P Rą0, and write H for Hf pwq. By Euler’s formula for homogeneous functions,

T T w Hw “ dpd ´ 1qfpwq and w Hei “ pd ´ 1qBifpwq. LORENTZIAN POLYNOMIALS 17

Let t be a real parameter, and consider the restriction of H to the plane spanned by w and

vt “ ei ` tej. By Theorem 2.16, H has exactly one positive eigenvalue. Therefore, by Cauchy’s interlacing theorem, the restriction of H also has exactly one positive eigenvalue. In particular, the determinant of the restriction must be nonpositive:

T 2 T T w Hvt ´ w Hw ¨ vt Hvt ě 0 for all t P R.

In other words, for all t`P R, we˘ have` ˘ ` ˘ 2 2 2 2 2 pd ´ 1q pBif ` tBjfq ´ dpd ´ 1qfpBi f ` 2tBiBjf ` t Bj fq ě 0.

It follows that, for all t P R, we have 2 2 pd ´ 1q pBif ` tBjfq ´ 2tdpd ´ 1qfBiBjf ě 0.

Thus, the discriminant of the above quadratic polynomial in t should be nonpositive: 1 fB B f ´ 2 1 ´ B fB f ď 0. i j d i j ´ ¯ This completes the proof of Proposition 2.17. 

2.3. Independence and negative dependence. Let c be a fixed positive real number, and let f be a polynomial in Rrw1, . . . , wns. In this section, the polynomial f is not necessarily homogeneous. n As before, we write ei for the i-th standard unit vector in R .

Definition 2.18. We say that f is c-Rayleigh if f has nonnegative coefficients and

α α`ei`ej α`ei α`ej n n B fpwq B fpwq ď c B fpwq B fpwq for all i, j P rns, α P N , w P Rě0.

When f is the partition function of a discrete probability measure µ, the c-Rayleigh condition captures a negative dependence property of µ. More precisely, when f is multi-affine, that is, when f has degree at most one in each variable, the c-Rayleigh condition for f is equivalent to

n fpwq BiBjfpwq ď cBifpwq Bjfpwq for all distinct i, j P rns, and w P Rě0. Thus the 1-Rayleigh property of multi-affine polynomials is equivalent to the Rayleigh property for discrete probability measures studied in [Wag08] and [BBL09].

d 1 Proposition 2.19. Any polynomial in Ln is 2 1 ´ d -Rayleigh. ´ ¯ 1 Proof. The statement follows from Theorem 2.16 and Proposition 2.17 because 2 1 ´ d is an d increasing function of . ` ˘ 

The goal of this section is to show that the support of any homogeneous c-Rayleigh polyno- mial is M-convex (Theorem 2.23). The notion of M6-convexity will be useful for the proof: A subset J6 Ď Nn is said to be M6-convex if there is an M-convex set J in Nn`1 such that

6 J “ pα1, . . . , αnq | pα1, . . . , αn, αn`1q P J . ! ) 18 PETTER BRAND¨ EN´ AND JUNE HUH

The projection from J to J6 should be bijective for any such J, as the M-convexity of J implies d 6 that J is in ∆n for some d. We refer to [Mur03, Section 4.7] for more on M -convex sets. We prepare the proof of Theorem 2.23 with three lemmas. Verification of the first lemma is routine and will be omitted.

Lemma 2.20. The following polynomials are c-Rayleigh whenever f is c-Rayleigh:

(1) The contraction Bif of f.

(2) The deletion fzi of f, the polynomial obtained from f by evaluating wi “ 0.

n (4) The dilation fpa1w1, . . . , anwnq, for pa1, . . . , anq P Rě0. n (5) The translation fpa1 ` w1, . . . , an ` wnq, for pa1, . . . , anq P Rě0.

Remark. Lemma 2.20 in the previous version of this manuscript contained the following incor- rect statement:

(3) If fpw1, w2, w3, . . . , wnq is c-Rayleigh, then so is the diagonalization fpw1, w1, w3, . . . , wnq.

This implies that the rank sequence of any Rayleigh measure is strongly log-concave, which is known to be not true [BBL09, Section 7, Counterexample 1]. In fact, the rank sequence of a Rayleigh measure need not even be unimodal [KN10, Theorem 6]. Proof of Lemma 2.22 below is revised accordingly.

We introduce a partial order ď on Nn by setting

α ď β ðñ αi ď βi for all i P rns.

We say that a subset J6 of Nn is interval convex if the following implication holds:

α P J6, β P J6, α ď γ ď β ùñ γ P J6. ´ ¯ The augmentation property for J6 Ď Nn is the implication

6 6 6 α P J , β P J , |α|1 ă |β|1 ùñ αj ă βj and α ` ej P J for some j P rns . ´ ¯ ´ ¯ Lemma 2.21. Let J6 be an interval convex subset of Nn containing 0. Then J6 is M6-convex if and only if J6 satisfies the augmentation property.

Therefore, a nonempty interval convex subset of t0, 1un containing 0 is M6-convex if and only if it is the collection of independent sets of a matroid on rns.

Proof. Let d be any sufficiently large positive integer, and set

n`1 6 J “ pα1, . . . , αn, d ´ α1 ´ ¨ ¨ ¨ ´ αnq P N | pα1, . . . , αnq P J . ! ) The “only if” direction is straightforward: If J6 is M6-convex, then J is M-convex, and the aug- mentation property for J6 is a special case of the exchange property for J. LORENTZIAN POLYNOMIALS 19

We prove the “if” direction by checking the exchange property for J. Let α and β be elements of J, and let i be an index satisfying αi ą βi. We claim that there is an index j satisfying

αj ă βj and α ´ ei ` ej P J.

By the augmentation property for J6, it is enough to justify the claim when i ‰ n ` 1. When 6 αn`1 ă βn`1, then we may take j “ n ` 1, again by the augmentation property for M .

Suppose αn`1 ě βn`1. In this case, we consider the element γ “ α ´ ei ` en`1. The element γ 6 belongs to J, because J is an interval convex set containing 0. We have γn`1 ą βn`1, and hence the augmentation property for J6 gives an index j satisfying

γj ă βj and α ´ ei ` ej “ γ ´ en`1 ` ej P J.

This index j is necessarily different from i because αi ą βi. It follows that αj “ γj ă βj, and the M-convexity of J is proved. 

Lemma 2.22. Let f be a c-Rayleigh polynomial in Rrw1, . . . , wns.

(1) The support of f is interval convex. (2) If fp0q is nonzero, then supppfq is M6-convex.

Proof. Suppose that the polynomial f and the vectors α ď γ ď β constitute a minimal coun- terexample to (1) with respect to the degree and the number of variables of f. We may and will suppose that |β|1 is minimal among all such α ď γ ď β for the polynomial f. We have

αj “ 0 for all j, since otherwise some contraction Bjf is a smaller counterexample to (1). Similarly, we have

βj ą 0 for all j, since otherwise some deletion fzj is a smaller counterexample to (1). In addition, we may assume that γ is a unit vector, say

γ “ ei, since otherwise the contraction Bjf for any j satisfying γj ą 0 is a smaller counterexample to

(1). Suppose ej is in the support of f for some j. In this case, we should have

ei ` ej P supppfq, since otherwise Bjf is a smaller counterexample to (1). However, the above implies

Bifp0q “ 0 and fp0qBiBjfp0q ą 0, contradicting the c-Rayleigh property of f. Therefore, no ej is in the support of f, and hence

|δ|1 ě |β|1 for all δ P supppfq, 20 PETTER BRAND¨ EN´ AND JUNE HUH by the minimality of |β|1. Thus, for any indices k and l satisfying β ě ek ` el, we have

fp, . . . , q “ a1 ` higher order terms,

|β|1´1 Bkfp, . . . , q “ a2 ` higher order terms,

|β|1´1 Blfp, . . . , q “ a3 ` higher order terms,

|β|1´2 BkBlfp, . . . , q “ a4 ` higher order terms, for some positive constants a1, a2, a3, a4. This contradicts the c-Rayleigh property of f for suffi- ciently small positive , proving (1). Suppose f is a counterexample to (2) that is minimal with respect to the degree and the number of variables of f. By Lemma 2.21 and (1) of the current lemma, we know that the support of f fails to have the augmentation property. In other words, there are α and β in the support of f such that |α|1 ă |β|1 and, for all i,

αi ă βi ùñ α ` ei R supppfq.

We may and will suppose that |α|1 is minimal among all such α and β for the polynomial f. For any γ, write Spγq for the set of indices i such that γi ą 0. If i is in the intersection of Spαq and

Spβq, then Bif is a counterexample to (2) that has degree less than that of f, and hence

Spαq X Spβq “ ∅. Similarly, if i is not in the union of Spαq and Spβq, then fzi is a counterexample to (2) that involves less than n variables, and hence

Spαq Y Spβq “ rns.

Since supppfq is interval convex by (1), there is an index i in Spαq. In addition, since fp0q is nonzero, we have

α ´ ei P supppfq.

The vectors α ´ ei and β satisfy the augmentation property for supppfq, by the minimality of

|α|1. Therefore, there is an index j such that

pα ´ eiqj ă βj and α ´ ei ` ej P supppfq.

The first condition shows that the index j cannot be in Spαq, so it must be in Spβq. The vectors

α ´ ei ` ej and β satisfy the augmentation property for supppfq, since otherwise Bjf is a smaller counterexample to (2). Therefore, there is an index k such that

pα ´ ei ` ejqk ă βk and α ´ ei ` ej ` ek P supppfq.

The first condition shows that the index k cannot be in Spαq, so it must be in Spβq. On the other hand, since j and k are in Spβq and not in Spαq, the failure of the augmentation property for α and β implies

α ` ej R supppfq and α ` ek R supppfq. LORENTZIAN POLYNOMIALS 21

Note that the c-Rayleigh polynomial g – Bα´ei f satisfies

ei, ej ` ek P supppgq and ei ` ej, ei ` ek R supppgq.

The first pair of conditions shows that gpwqBjBkgpwq is a strictly increasing function of wi if

we set all the variables other than wi equal to 1. On the other hand, the second pair of condi-

tions shows that BjgpwqBkgpwq is independent of wi, by the interval convexity of supppgq. This contradicts the c-Rayleigh property of g, proving (2). 

Theorem 2.23. If f is homogeneous and c-Rayleigh, then the support of f is M-convex.

Proof. By (5) of Lemma 2.20 and (2) of Lemma 2.22, the support of the translation

gpw1, . . . , wnq “ fpw1 ` 1, . . . , wn ` 1q

is M6-convex. In other words, the support J of the homogenization of g is M-convex. Since the intersection of an M-convex set with a coordinate hyperplane is M-convex, this implies the M-convexity of the support of f. 

A multi-affine polynomial f is said to be strongly Rayleigh if

n fpwq BiBjfpwq ď Bifpwq Bjfpwq for all distinct i, j P rns, and w P R .

Clearly, any strongly Rayleigh multi-affine polynomial is 1-Rayleigh. Since a multi-affine poly- nomial is stable if and only if it is strongly Rayleigh [Bra07¨ , Theorem 5.6], Theorem 2.23 extends the following theorem of Choe et al. [COSW04, Theorem 7.1]: If f is a nonzero homogeneous stable multi-affine polynomial, then the support of f is the set of bases of a matroid. Lastly, we show that the bound in Proposition 2.19 is optimal.

d Proposition 2.24. When n ď 2, all polynomials in Ln are 1-Rayleigh. When n ě 3, we have 1 all polynomials in Ld are c-Rayleigh ùñ c ě 2 1 ´ . n d ´ ¯ ´ ¯ 1 d In other words, for any n ě 3 and any c ă 2 1 ´ d , there is f P Ln that is not c-Rayleigh. ´ ¯

Proof. We first show by induction that, for any homogeneous bivariate polynomial f “ fpw1, w2q with nonnegative coefficients, we have

2 fpwq B1B2fpwq ď B1fpwq B2fpwq for any w P Rě0. ´ ¯ ´ ¯´ ¯ We use the obvious fact that, for any homogeneous polynomial with nonnegative coefficients h,

n degphq ` 1 h ě 1 ` wiBi h for any i P rns and w P Rě0. ` ˘ ` ˘ 22 PETTER BRAND¨ EN´ AND JUNE HUH

d d Since f is bivariate, we may write f “ c1w1 ` c2w2 ` w1w2g. We have

2 d´1 d´1 B1fB2f ´ fB1B2f “ d c1c2w1 w2 d d ` dc1w1 p1 ` w2B2qg ´ c1w1 p1 ` w1B1qp1 ` w2B2qg d d ` dc2w2 p1 ` w1B1qg ´ c2w2 p1 ` w1B1qp1 ` w2B2qg

` w1w2pg ` w1B1gqpg ` w2B2gq ´ w1w2gp1 ` w1B1qp1 ` w2B2qg.

2 The summand in the second line is nonnegative on Rě0 by the mentioned fact for p1 ` w2B2qg. 2 The summand in the third line is nonnegative on Rě0 by the mentioned fact for p1`w1B1qg. The 2 summand in the fourth line is nonnegative on Rě0 by the induction hypothesis applied to g.

We next show that, for any bivariate Lorentzian polynomial f “ fpw1, w2q, we have

2 fpwq B1B1fpwq ď B1fpwq B1fpwq for any w P Rě0. ´ ¯ ´ ¯´ ¯ Since f is homogeneous, it is enough to prove the inequality when w2 “ 1. In this case, the

inequality follows from the concavity of the function log f restricted to the line w2 “ 1. This completes the proof that any bivariate Lorentzian polynomial is 1-Rayleigh. To see the second statement, consider the polynomial 1 f “ 2 1 ´ wd ` wd´1w ` wd´1w ` wd´2w w . d 1 1 2 1 3 1 2 3 ´ ¯ d n It is straightforward to check that f is in Ln. If f is c-Rayleigh, then, for any w P Rě0, 1 w2d´4 2 1 ´ w2 ` w w ` w w ` w w ď cw2d´4 w ` w w ` w . 1 d 1 1 2 1 3 2 3 1 1 2 1 3 ´ ´ ¯ ¯ ´ ¯´ ¯ The desired lower bound for c is obtained by setting w1 “ 1, w2 “ 0, w3 “ 0. 

2.4. Characterizations of Lorentzian polynomials. We may now give a complete and useful d description of the space of Lorentzian polynomials. As before, we write Hn for the space of degree d homogeneous polynomials in n variables.

˚d d d d d Theorem 2.25. The closure of Ln in Hn is Ln. In particular, Ln is a closed subset of Hn.

˚d d Proof. By Theorem 2.13, the closure of Ln contains Ln. Since any limit of c-Rayleigh polynomials must be c-Rayleigh, the other inclusion follows from Theorem 2.23 and Proposition 2.19. 

Therefore, a degree d homogeneous polynomial f with nonnegative coefficients is Lorentzian if and only if the support of f is M-convex and Bαf has at most one positive eigenvalue for every d´2 α P ∆n . In other words, Definitions 2.1 and 2.6 define the same class of polynomials.

Example 2.26. A sequence of nonnegative numbers a0, a1, . . . , ad is said to be ultra log-concave if

2 a ak 1 ak 1 k ě ´ ` for all 0 ă k ă d. d 2 d d k k´1 k`1 ` ˘ ` ˘ ` ˘ LORENTZIAN POLYNOMIALS 23

The sequence is said to have no internal zeros if

ak1 ak3 ą 0 ùñ ak2 ą 0 for all 0 ď k1 ă k2 ă k3 ď d. d k d´k Recall from Example 2.3 that a bivariate homogeneous polynomial k“0 akw1 w2 is strictly Lorentzian if and only if the sequence a is positive and strictly ultra log-concave. Theorem 2.25 k ř d k d´k says that, in this case, the polynomial k“0 akw1 w2 is Lorentzian if and only if the sequence a is nonnegative, ultra log-concave, and has no internal zeros. k ř Example 2.27. Using Theorem 2.25, it is straightforward to check that elementary symmetric polynomials are Lorentzian. In fact, one can show more generally that all normalized Schur polynomials are Lorentzian [HMMS, Theorem 3]. Any elementary symmetric polynomial is stable [COSW04, Theorem 9.1], but a normalized Schur polynomial need not be stable [HMMS, Example 9].

d d d Let PHn be the projectivization of Hn equipped with the quotient topology. The image PLn d d of Ln in the projective space is homeomorphic to the intersection of Ln with the unit sphere in d Hn for the Euclidean norm on the coefficients.

d Theorem 2.28. The space PLn is compact and contractible.

d d Proof. Since PHn is compact, Theorem 2.25 implies that PLn is compact. A deformation retract d of PLn can be constructed using Theorem 2.10. 

d We conjecture that PLn is homeomorphic to a familiar space.

d Conjecture 2.29. The space PLn is homeomorphic to a closed Euclidean ball.

For other appearances of stratified Euclidean balls in the interface of analysis of combina- torics, see [GKL, GKL19] and references therein. Prominent examples are the totally nonnega- tive parts of Grassmannian and other partial flag varieties. Let f be a polynomial in n variables with nonnegative coefficients. In [Gur09], Gurvits de- fines f to be strongly log-concave if, for all α P Nn, α α n B f is identically zero or logpB fq is concave on Rą0. In [AOVI], Anari et al. define f to be completely log-concave if, for all m P N and any m ˆ n matrix paijq with nonnegative entries, m m n Di f is identically zero or log Di f is concave on Rą0, ´ i“1 ¯ ´´ i“1 ¯ ¯ ź n ź where Di is the differential operator j“1 aijBj. We show that the two notions agree with each other and with the Lorentzian property for homogeneous polynomials.8 ř Theorem 2.30. The following conditions are equivalent for any homogeneous polynomial f.

8An implication similar to p3q ñ p1q of Theorem 2.30 can be found in [ALOVIII, Theorem 3.2]. 24 PETTER BRAND¨ EN´ AND JUNE HUH

(1) f is completely log-concave. (2) f is strongly log-concave. (3) f is Lorentzian.

The support of any Lorentzian polynomial is M-convex by Theorem 2.25. Thus, by Theorem 2.30, the same holds for any strongly log-concave homogeneous polynomial. This answers a question of Gurvits [Gur09, Section 4.5 (iii)].

Corollary 2.31. The support of any strongly log-concave homogeneous polynomial is M-convex.

Similarly, we can use Theorem 2.30 to show that the class of strongly log-concave homoge- neous polynomials is closed under multiplication. This answers another question of Gurvits [Gur09, Section 4.5 (iv)] for homogeneous polynomials.

Corollary 2.32. The product of strongly log-concave homogeneous polynomials is strongly log- concave.

d e Proof. Let fpwq be an element of Ln, and let gpwq be an element of Ln. It is straightforward to d`e check that fpwqgpuq is an element of Ln`n, where u is a set of variables disjoint from w. It follows d`e that fpwqgpwq is an element of Ln , since setting u “ w preserves the Lorentzian property by Theorem 2.10. 

Corollary 2.32 extends the following theorem of Liggett [Lig97, Theorem 2]: The convolu- tion product of two ultra log-concave sequences with no internal zeros is an ultra log-concave sequence with no internal zeros. To prove Theorem 2.30, we use the following elementary observation. Let f be a homoge- neous polynomial in n ě 2 variables of degree d ě 2.

Proposition 2.33. The following are equivalent for any w P Rn satisfying fpwq ą 0. (1) The Hessian of f 1{d is negative semidefinite at w. (2) The Hessian of log f is negative semidefinite at w. (3) The Hessian of f has exactly one positive eigenvalue at w.

The equivalence of (2) and (3) appears in [AOVI].

Proof. We fix w throughout the proof. For n ˆ n symmetric matrices A and B, we write A ă B to mean the following interlacing relationship between the eigenvalues of A and B:

λ1pAq ď λ1pBq ď λ2pAq ď λ2pBq ď ¨ ¨ ¨ ď λnpAq ď λnpBq.

1{d Let H1, H2, and H3 for the Hessians of f , log f, and f, respectively. We have 1 df ´1{dH “ H ` f ´2pgradfqpgradfqT and H “ f ´1H ´ f ´2pgradfqpgradfqT . 1 2 d 2 3 LORENTZIAN POLYNOMIALS 25

Since pgradfqpgradfqT is positive semidefinite of rank one, Weyl’s inequalities for Hermitian matrices [Ser10, Theorem 6.3] show that

H2 ă H1 and H2 ă H3 and H1 ă H3.

T Since w H3w “ dpd ´ 1qf, H3 has at least one positive eigenvalue, and hence (1) ñ (2) ñ (3).

For (3) ñ (1), suppose that H3 has exactly one positive eigenvalue. We introduce a positive real parameter  and consider the polynomial

d d f “ f ´ pw1 ` ¨ ¨ ¨ ` wnq.

1{d We write H3, for the Hessian of f, and write H1, for the Hessian of f .

Note that H3, is nonsingular and has exactly one positive eigenvalue for all sufficiently small positive . In addition, we have H1, ă H3,, and hence H1, has at most one nonnegative eigenvalue for all sufficiently small positive . However, by Euler’s formula for homogeneous functions, we have

H1,w “ 0, so that 0 is the only nonnegative eigenvalue of H1, for any such . The implication (3) ñ (1) now follows by limiting  to 0. 

It follows that, for any nonzero degree d ě 2 homogeneous polynomial f with nonnegative coefficients, the following conditions are equivalent:

1{d n – The function f is concave on Rą0. n – The function log f is concave on Rą0. n – The Hessian of f has exactly one positive eigenvalue on Rą0.

Proof of Theorem 2.30. We may suppose that f has degree d ě 2. Clearly, completely log-concave polynomials are strongly log-concave. Suppose f is a strongly log-concave homogeneous polynomial of degree d. By Proposition 2.33, either Bαf is identically zero or the Hessian of Bαf has exactly one positive eigenvalue on n n 1 Rą0 for all α P N . By Proposition 2.17, f is 2 1 ´ d -Rayleigh, and hence, by Theorem 2.23, the f M f support of is -convex. Therefore, by Theorem` 2.25˘ , is Lorentzian. Suppose f is a nonzero Lorentzian polynomial. Theorem 2.16 and Proposition 2.33 together n n show that log f is concave on Rą0. Therefore, it is enough to prove that i“1 aiBi f is Lorentzian for any nonnegative numbers a , . . . , a . This is a direct consequence of Theorem 2.25 and 1 n ` ř ˘ Corollary 2.11.  26 PETTER BRAND¨ EN´ AND JUNE HUH

3. ADVANCED THEORY

3.1. Linear operators preserving Lorentzian polynomials. We describe a large class of linear operators that preserve the Lorentzian property. An analog was achieved for the class of sta- ble polynomials in [BB09, Theorem 2.2], where the linear operators preserving stability were characterized. For an element κ of Nn, we set

Rκrwis “ polynomials in Rrwis1ďiďn of degree at most κi in wi for every i ,

a ! ) Rκrwijs “ multi-affine polynomials in Rrwijs1ďiďn,1ďjďκi .

! Ó a ) The projection operator Πκ : Rκrwijs Ñ Rκrwis is the linear map that substitutes each wij by wi: Ó Πκpgq “ g|wij “wi .

Ò a α The polarization operator Πκ : Rκrwis Ñ Rκrwijs is the linear map that sends w to the product 1 n κ elementary symmetric polynomial of degree αi in the variables twiju1ďjďκi , α i“1 ź ` ˘ where` ˘κ stands for the product of binomial coefficients n κi . Note that α i“1 αi Ó Ò – for every` ˘ f, we have Πκ ˝ Πκpfq “ f, and ś ` ˘ Ò – for every f and every i, the polynomial Πκpfq is symmetric in the variables twiju1ďjďκi .

Ò a The above properties characterize Πκ among the linear operators from Rκrwis to Rκrwijs.

Ó Ò Proposition 3.1. The operators Πκ and Πκ preserve the Lorentzian property.

Ò In other words, Πκpfq is a Lorentzian polynomial for any Lorentzian polynomial f P Rκrwis, Ó a and Πκpgq is a Lorentzian polynomial for any Lorentzian polynomial g P Rκrwijs.

Ó Ò Proof. The statement for Πκ follows from Theorem 2.10. We prove the statement for Πκ. It is Ò ˚d enough to prove that Πκpfq is Lorentzian when f P Ln X Rκrwis for d ě 2. k ˚d Set k “ |κ|1, and identify N with the set of all monomials in wij. Since f P Ln, we have k supp ΠÒ pfq “ , κ d „  which is clearly M-convex. Therefore, by` Theorem˘ 2.25, it remains to show that the quadratic β Ò k form B Πκpfq is stable for any β P d´2 . Ó β α Define α by the equality Πκpw q“ “ w‰ . Note that, after renaming the variables if necessary, Ò the β-th partial derivative of Πκpfq is a positive multiple of a polarization of the α-th partial derivative of f: pκ ´ αq! BβΠÒ pfq “ ΠÒ pBαfq. κ κ! κ´α Ò Since the operator Πκ´α preserves stability [BB09, Proposition 3.4], the conclusion follows from α the stability of the quadratic form B f.  LORENTZIAN POLYNOMIALS 27

n m Let κ be an element of N , let γ be an element of N , and set k “ |κ|1. In the remainder of this section, we fix a linear operator

T : Rκrwis Ñ Rγ rwis, and suppose that the linear operator T is homogeneous of degree ` for some ` P Z: 0 ď α ď κ and T pwαq ‰ 0 ùñ deg T pwαq “ deg wα ` `.

The symbol of T is a´ homogeneous polynomial¯ of degree k ` ` in m ` n variables defined by κ sym pw, uq “ T pwαquκ´α. T α 0ďαďκ ÿ ˆ ˙ We show that the homogeneous operator T preserves the Lorentzian property if its symbol

symT is Lorentzian.

k`` d d`` Theorem 3.2. If symT P Lm`n and f P Ln X Rκrwis, then T pfq P Lm .

When n “ 2, Theorem 3.2 provides a large class of linear operators that preserve the ultra log- concavity of sequences of nonnegative numbers with no internal zeros. We prepare the proof of Theorem 3.2 with a special case.

Lemma 3.3. Let T “ Tw1,w2 : Rp1,...,1qrwis Ñ Rp1,...,1qrwis be the linear operator defined by wSz1 if 1 P S and 2 P S, wSz1 if 1 P S and 2 R S, T wS $ S n p q “ Sz2 for all Ď r s. ’ w if 1 R S and 2 P S, &’ 0 if 1 R S and 2 R S, ’ Then T preserves the Lorentzian%’ property.

d ˚d`1 Proof. It is enough to prove that T pfq P Ln when f P Ln X Rp1,...,1qrwis for d ě 2. In this case, supp T pfq “ d-element subsets of rns not containing 1 , which is clearly M-convex.` Therefore,˘ by Theorem 2.25, it suffices to show( that the quadratic S n form B T pfq is stable for any S P d´2 not containing 1. We write h for the Lorentzian polyno- f| f mial w1“0. Since is multi-affine,“ we‰ have

f “ h ` w1B1f and T pfq “ B2h ` B1f. We give separate arguments when 2 P S and 2 R S. If S contains 2, then BST pfq “ BSY1f, and hence BST pfq is stable. If S does not contain 2, then

S SY1Y2 ˚d`1 – the linear form B B1B2f “B f is not identically zero, because f P Ln , – we have BSY1Y2f ă BSY2h, because BSY2f is stable, and – we have BSY1Y2f ă BSY1f, by Lemma 2.9 (1). 28 PETTER BRAND¨ EN´ AND JUNE HUH

S SY2 SY1 Therefore, by Lemma 2.9 (4), the quadratic form B T pfq “ B h ` B f is stable. 

Ò Proof of Theorem 3.2. The polarization of T : Rκrwis Ñ Rγ rwis is the operator Π pT q defined by Ò Ò Ó Π pT q “ Πγ ˝ T ˝ Πκ.

We write γ ‘ κ for the concatenation of γ and κ in Nm`n. By [BB09, Lemma 3.5], the symbol of the polarization is the polarization of the symbol9: Ò symΠÒpT q “ Πγ ‘κpsymT q. Therefore, by Proposition 3.1, the proof reduces to the case κ “ p1,..., 1q and γ “ p1,..., 1q. d Suppose fpvq is a multi-affine polynomial in Ln and symT pw, uq is a multi-affine polynomial ``n in Lm`n. Since the product of Lorentzian polynomials is Lorentzian by Corollary 2.32, we have S rnszS d```n symT pw, uqfpvq “ T pw qu fpvq P Lm`n`n. S n Ďrÿ s Applying the operator in Lemma 3.3 for the pair of variables pui, viq for i “ 1, . . . , n, we have n S S d`` Tui,vi symT pw, uqfpvq “ T pw qpB fqpvq P Lm`n. i“1 SĎrns ź ` ˘ ÿ We substitute every vi by zero in the displayed equation to get

T pwSqpBSfqpvq “ T fpwq . v“0 SĎrns ” ÿ ı ` ˘ d`` Theorem 2.10 shows that the right-hand side belongs to Lm , completing the proof. 

We remark that there are homogeneous linear operators T preserving the Lorentzian property whose symbols are not Lorentzian. This contrasts the analog of Theorem 3.2 for stable polyno-

mials [BB09, Theorem 2.2]. As an example, consider the linear operator T : Rp1,1qrw1, w2s Ñ Rp1,1qrw1, w2s defined by

T p1q “ 0,T pw1q “ w1,T pw2q “ w2,T pw1w2q “ w1w2. The symbol of T is not Lorentzian because its support is not M-convex. The operator T preserves Lorentzian polynomials but does not preserve (non-homogeneous) stable polynomials.

Theorem 3.4. If T is a homogeneous linear operator that preserves stable polynomials and poly- nomials with nonnegative coefficients, then T preserves Lorentzian polynomials.

Proof. According to [BB09, Theorem 2.2], T preserves stable polynomials if and only if either

(I) the rank of T is not greater than two and T is of the form T pfq “ αpfqP ` βpfqQ,

where α, β are linear functionals and P,Q are stable polynomials satisfying P ă Q,

9The statement was proved in [BB09, Lemma 3.5] when m “ n. Clearly, this special case implies the general case. LORENTZIAN POLYNOMIALS 29

(II) the polynomial symT pw, uq is stable, or

(III) the polynomial symT pw, ´uq is stable. Suppose one of the three conditions, and suppose in addition that T preserves polynomials with nonnegative coefficients. Suppose (I) holds. In this case, the image of T is contained in the set of stable polynomials [BB10, Theorem 1.6]. By Proposition 2.2, homogeneous stable polynomials with nonnegative coefficients are Lorentzian. Since T preserves polynomials with nonnegative coefficients, T pfq is Lorentzian whenever f is a homogeneous polynomial with nonnegative coefficients.

Suppose (II) holds. Since T preserves polynomials with nonnegative coefficients, symT pw, uq is Lorentzian by Proposition 2.2. Therefore, by Theorem 3.2, T pfq is Lorentzian whenever f is Lorentzian. Suppose (III) holds. Since all the nonzero coefficients of a homogeneous stable polynomial have the same sign [COSW04, Theorem 6.1], we have

symT pw, ´vq “ symT pw, vq or symT pw, ´vq “ ´symT pw, vq.

In both cases, symT pw, vq is stable and has nonnegative coefficients. Thus symT pw, vq is Lorentzian, and the conclusion follows from Theorem 3.2. 

In the remainder of this section, we record some useful operators that preserves the Lorentzian α α property. The multi-affine part of a polynomial n cαw is the polynomial n cαw . αPN αPt0,1u Corollary 3.5. The multi-affine part of any Lorentzianř polynomial is a Lorentzianř polynomial.

Proof. Clearly, taking the multi-affine part is a homogeneous linear operator that preserves poly- nomials with nonnegative coefficients. Since this operator also preserves stable polynomials [COSW04, Proposition 4.17], the proof follows from Theorem 3.4.  d Remark 3.6. Corollary 3.5 can be used to obtain a multi-affine analog of Theorem 2.28. Write Hn d for the space of multi-affine degree d homogeneous polynomials in n variables, and write Ln for d the corresponding set of multi-affine Lorentzian polynomials. Let PHn be the projectivization d d of the vector space Hn, and let LB be the set of polynomials in Ln with support B. We identify n d a rank d matroid M on rns with its set of bases B Ď d . Writing PLn and PLB for the images of Ld 0 L Hd nz and B in P n respectively, we have “ ‰ d PLn “ PLB, B ž d where the union is over all rank d matroids on rns. By Theorem 2.10 and Corollary 3.5, PLn d is a compact contractible subset of PHn. By Theorem 3.10, PLB is nonempty for every matroid n n B Ď d . In addition, by Proposition 3.25, PLB is contractible for every matroid B Ď d .

“ ‰ α wα “ ‰ Let N be the linear operator defined by the condition Npw q “ α! . The normalization oper- ator N turns generating functions into exponential generating functions. 30 PETTER BRAND¨ EN´ AND JUNE HUH

Corollary 3.7. If f is a Lorentzian polynomial, then Npfq is a Lorentzian polynomial.

It is shown in [HMMS, Theorem 3] that the normalized Schur polynomial Npsλpw1, . . . , wnqq is Lorentzian for any partition λ. Note that the converse of Corollary 3.7 fails in general. For example, complete symmetric polynomials, which are special cases of Schur polynomials, need not be Lorentzian.

Proof. Let κ be any element of Nn. By Theorem 3.2, it suffices to show that the symbol

α n αj κ w κj wj κ α sym pw, uq “ uκ´α “ u j ´ j N α α! α α ! j 0ďαďκ j“1 ˜ 0ďα ďκ j j ¸ ÿ ˆ ˙ ź ÿj j ˆ ˙ is a Lorentzian polynomial. Since the product of Lorentzian polynomials is Lorentzian by Corol- lary 2.32, the proof is reduced to the case when the symbol is bivariate. In this case, using the characterization of bivariate Lorentzian polynomials in Example 2.26, we get the Lorentzian property from the log-concavity of the sequence 1{k!. 

Corollary 3.8 below extends the classical fact that the convolution product of two log-concave sequences with no internal zeros is a log-concave sequence with no internal zeros. For early proofs of the classical fact, see [Kar68, Chapter 8] and [Men69].

Corollary 3.8. If Npfq and Npgq are Lorentzian polynomials, then Npfgq is a Lorentzian poly- nomial.

Note that the analogous statement for stable polynomials fails to hold in general. For exam- ple, when f “ x3 ` x2y ` xy2 ` y3, the polynomial Npfq is stable but Npf 2q is not.

Proof. Suppose that f and g belong to Rκrwis. We consider the linear operator

T : Rκrwis ÝÑ Rrwis,Nphq ÞÝÑ Nphgq. By Theorem 3.2, it is enough to show that its symbol uκ´α sym pw, uq “ κ! Npwαgq T pκ ´ αq! 0ďαďκ ÿ is a Lorentzian polynomial in 2n variables. For this, we consider the linear operator κ´α α u S : Rκrwis ÝÑ Rrwi, uis,Nphq ÞÝÑ Npw hq . pκ ´ αq! 0ďαďκ ÿ By Theorem 3.2, it is enough to show that its symbol wα`β uκ´α vκ´β sym pw, u, vq “ κ! S pα ` βq! pκ ´ αq! pκ ´ βq! 0ďβďκ 0ďαďκ ÿ ÿ is a Lorentzian polynomial in 3n variables. The statement is straightforward to check using Theorem 2.25. See Theorem 3.10 below for a more general statement.  LORENTZIAN POLYNOMIALS 31

The symmetric exclusion process is one of the main models considered in interacting particle systems. It is a continuous time Markov chain which models particles that jump symmetrically between sites, where each site may be occupied by at most one particle [Lig10]. A problem that has attracted much attention is to find negative dependence properties that are preserved under the symmetric exclusion process. In [BBL09, Theorem 4.20], it was proved that strongly Rayleigh measures are preserved under the symmetric exclusion process. In other words, if f “ fpw1, w2, . . . , wnq is a stable multi-affine polynomial with nonnegative coefficients, then the 1,2 multi-affine polynomial Φθ pfq defined by

1,2 Φθ pfq “ p1 ´ θqfpw1, w2, w3, . . . , wnq ` θfpw2, w1, w3, . . . , wnq

is stable for all 0 ď θ ď 1. We prove an analog for Lorentzian polynomials.

Corollary 3.9. Let f “ fpw1, w2, . . . , wnq be a multi-affine polynomial with nonnegative coef- ficients. If the homogenization of f is a Lorentzian polynomial, then the homogenization of 1,2 Φθ pfq is a Lorentzian polynomial for all 0 ď θ ď 1.

Proof. Recall that a polynomial with nonnegative coefficients is stable if and only if its homoge- 1,2 nization is stable [BBL09, Theorem 4.5]. Clearly, Φθ is homogeneous and preserves polynomi- 1,2 als with nonnegative coefficients. Since Φθ preserves stability of multi-affine polynomials by [BBL09, Theorem 4.20], the statement follows from Theorem 3.4. 

3.2. Matroids, M-convex sets, and Lorentzian polynomials. The generating function of a subset J Ď Nn is, by definition, wα n f “ , where α! “ α !. J α! i αPJ i“1 ÿ ź We characterize matroids and M-convex sets in terms of their generating functions.

Theorem 3.10. The following are equivalent for any nonempty J Ď Nn.

(1) There is a Lorentzian polynomial whose support is J. (2) There is a homogeneous 2-Rayleigh polynomial whose support is J. (3) There is a homogeneous c-Rayleigh polynomial whose support is J for some c ą 0.

(4) The generating function fJ is a Lorentzian polynomial.

(5) The generating function fJ is a homogeneous 2-Rayleigh polynomial.

(6) The generating function fJ is a homogeneous c-Rayleigh polynomial for some c ą 0. (7) J is M-convex.

When J Ď t0, 1un, any one of the above conditions is equivalent to

(8) J is the set of bases of a matroid on rns. 32 PETTER BRAND¨ EN´ AND JUNE HUH

The statement that the basis generating polynomial fJ is log-concave on the positive orthant

can be found in [AOVI, Theorem 4.2]. An equivalent statement that the Hessian fJ has ex- actly one positive eigenvalue on the positive orthant has been noted earlier in [HW17, Remark 15]. The equivalence of the conditions (4) and (7) will be generalized to M-convex functions in Theorem 3.14. We prepare the proof of Theorem 3.10 with an analysis of the quadratic case.

Lemma 3.11. The following conditions are equivalent for any n ˆ n symmetric matrix A with entries in t0, 1u.

(1) The quadratic polynomial wT Aw is Lorentzian. (2) The support of the quadratic polynomial wT Aw is M-convex.

Proof. Theorem 2.25 implies (1) ñ (2). We prove (2) ñ (1). We may and will suppose that no column of A is zero. Let J be the M-convex support of wT Aw, and set

S “ i P rns | 2ei P J . ! ) The exchange property for J shows that ei ` ej P J for every i P S and j P rns. In addition, again by the exchange property for J,

B :“ ei ` ej P J | i R S and j R S ! ) is the set of bases of a rank 2 matroid on rnszS without loops. Writing S1 Y ¨ ¨ ¨ Y Sk for the decomposition of rnszS into parallel classes in the matroid, we have 2 2 2 T w Aw “ wj ´ wj ´ ¨ ¨ ¨ ´ wj , j n jPS jPS ´ ÿPr s ¯ ´ ÿ1 ¯ ´ ÿk ¯ T and hence w Aw is a Lorentzian polynomial. 

Proof of Theorem 3.10. Theorem 2.23, Theorem 2.25, and Proposition 2.19 show that

p1q ñ p2q ñ p3q ñ p7q and p4q ñ p5q ñ p6q ñ p7q.

Since (4) ñ (1), we only need to prove (7) ñ (4). n If J is an M-convex subset of N , then fJ is a homogeneous polynomial of some degree d. d´2 α Suppose d ě 2, and let α be an element of ∆n . Note that, in general, the support of B fJ is α M-convex whenever the support of fJ is M-convex. Therefore, B fJ is Lorentzian by Lemma 3.11, and hence fJ is Lorentzian by Theorem 2.25. 

Let J be the set of bases of a matroid M on rns. If M is regular [FM92], if M is representable over the finite fields F3 and F4 [COSW04], if the rank of M is at most 3 [Wag05], or if the number of elements n is at most 7 [Wag05], then fJ is 1-Rayleigh. Seymour and Welsh found the first example of a matroid whose basis generating function is not 1-Rayleigh [SW75]. We propose the following improvement of Theorem 3.10. LORENTZIAN POLYNOMIALS 33

Conjecture 3.12. The following conditions are equivalent for any nonempty J Ď t0, 1un.

(1) J is the set of bases of a matroid on rns.

8 (2) The generating function fJ is a homogeneous 7 -Rayleigh polynomial.

8 8 The constant 7 is best possible: For any positive real number c ă 7 , there is a matroid whose basis generating function is not c-Rayleigh [HSW, Theorem 7].

3.3. Valuated matroids, M-convex functions, and Lorentzian polynomials. Let ν be a function from Nn to R Y t8u. The effective domain of ν is, by definition,

n dompνq “ α P N | νpαq ă 8 . ! ) The function ν is said to be M-convex if satisfies the symmetric exchange property:

(1) For any α, β P dompνq and any i satisfying αi ą βi, there is j satisfying

αj ă βj and νpαq ` νpβq ě νpα ´ ei ` ejq ` νpβ ´ ej ` eiq.

Note that the effective domain of an M-convex function on Nn is an M-convex subset of Nn. In n d particular, the effective domain of an M-convex function on N is contained in ∆n for some d. In d this case, we identify ν with its restriction to ∆n. When the effective domain of ν is is M-convex, the symmetric exchange property for ν is equivalent to the following local exchange property:

(2) For any α, β P dompνq with |α ´ β|1 “ 4, there are i and j satisfying

αi ą βi, αj ă βj and νpαq ` νpβq ě νpα ´ ei ` ejq ` νpβ ´ ej ` eiq.

A proof of the equivalence of the two exchange properties can be found in [Mur03, Section 6.2].

n n Example 3.13. The indicator function of J Ď N is the function νJ : N Ñ R Y t8u defined by 0 if α P J, νJpαq “ # 8 if α R J. n Clearly, J Ď N is M-convex if and only if the indicator function νJ is M-convex.

A function ν : Nn Ñ R Y t´8u is said to be M-concave if ´ν is M-convex. The effective domain of an M-concave function is

n dompνq “ α P N | νpαq ą ´8 . ! ) A valuated matroid on rns is an M-concave function on Nn whose effective domain is a nonempty subset of t0, 1un . The effective domain of a valuated matroid ν on rns is the set of bases of a matroid on rns, the underlying matroid of ν. In this section, we prove that the class of tropicalized Lorentzian polynomials coincides with the class of M-convex functions. The tropical connection is used to produce Lorentzian polyno- mials from M-convex functions. First, we state a classical version of the result. For any function 34 PETTER BRAND¨ EN´ AND JUNE HUH

d ν : ∆n Ñ R Y t8u and a positive real number q, we define qνpαq δ f ν pwq “ wα and gν pwq “ qνpαqwα, q α! q α α ν α ν Pdomÿ p q Pdomÿ p q ˆ ˙ where δ “ pd, . . . , dq and δ is the product of binomial coefficients n d . When ν is the α i“1 αi indicator function of J Ď n, the polynomial f ν is independent of q and equal to the generating N` ˘ q ś ` ˘ function fJ considered in Section 3.2.

d Theorem 3.14. The following conditions are equivalent for ν : ∆n Ñ R Y t8u. (1) The function ν is M-convex.

ν (2) The polynomial fq pwq is Lorentzian for all 0 ă q ď 1. ν (3) The polynomial gq pwq is Lorentzian for all 0 ă q ď 1.

The proof of Theorem 3.14, which relies on the theory of phylogenetic trees and the problem of isometric embeddings of finite metric spaces in Euclidean spaces, will be given at the end of this subsection.

Example 3.15. A function µ from Nn to R Y t8u is said to be M6-convex if, for some positive integer d, the function ν from Nn`1 to R Y t8u defined by

d µpα1, . . . , αnq if α P ∆n`1, νpα0, α1, . . . , αnq “ d # 8 if α R ∆n`1, is M-convex. The condition does not depend on d, and M6-concave functions are defined simi- larly. We refer to [Mur03, Chapter 6] for more on M6-convex and M6-concave functions. n It can be shown that every matroid rank function rkM, viewed as a function on N with the effective domain t0, 1un, is M6-concave. See [Shi12, Section 3] for an elementary proof and other related results. Thus, by Theorem 3.14, the normalized rank generating function 1 ´rkMpAq A cpAq q w w , where w “ pw1, . . . , wnq and cpAq “ n ´ |A|, cpAq! 0 A n ÿĎr s is Lorentzian for all 0 ă q ď 1. We will obtain a sharper result on rkM in Section 4.3.

Theorem 3.14 provides a useful sufficient condition for a homogeneous polynomial to be Lorentzian. Let f be an arbitrary homogeneous polynomial with nonnegative real coefficients written in the normalized form c f “ α wα. α! α ∆d ÿP n We define a discrete function νf using natural logarithms of the normalized coefficients:

d νf : ∆n ÝÑ R Y t´8u, α ÞÝÑ logpcαq.

Corollary 3.16. If νf is an M-concave function, then f is a Lorentzian polynomial. LORENTZIAN POLYNOMIALS 35

´νf pαq q wα q “ e´1 Proof. By Theorem 3.14, the polynomial αPdompνf q α! is Lorentzian when .  ř We note that the converse of Corollary 3.16 does not hold. For example, the polynomial n´1 f “ pwi ` wnq i“1 ź is Lorentzian, being a product of Lorentzian polynomials. However, νf fails to be M-concave when n ą 2.

We formulate a tropical counterpart of Theorem 3.14. Let Cpptqqconv be the field of Laurent series with complex coefficients that have a positive radius of convergence around 0. By defini-

tion, any nonzero element of Cpptqqconv is a series of the form

a1 a2 a3 sptq “ c1t ` c2t ` c3t ` ¨ ¨ ¨ ,

where c1, c2,... are nonzero complex numbers and a1 ă a2 ă ¨ ¨ ¨ are integers, that converges

on a punctured open disk centered at 0. Let Rpptqqconv be the subfield of elements that have real coefficients. We define the fields of real and complex convergent Puiseux series10 by

1{k 1{k K “ Rppt qqconv and K “ Cppt qqconv. kě1 kě1 ď ď Any nonzero element of K is a series of the form

a1 a2 a3 sptq “ c1t ` c2t ` c3t ` ¨ ¨ ¨ ,

where c1, c2,... are nonzero complex numbers and a1 ă a2 ă ¨ ¨ ¨ are rational numbers that

have a common denominator. The leading coefficient of sptq is c1, and the leading exponent of sptq is a1. A nonzero element of K is positive if its leading coefficient is positive. The valuation map is the function val : K ÝÑ R Y t8u, that takes the zero element to 8 and a nonzero element to its leading exponent. For a nonzero element sptq P K, we have

val sptq “ lim logt sptq . tÑ0` ` ˘ ` ˘ The field K is algebraically closed, and the field K is real closed. See [Spe05, Section 1.5] and references therein. Since the theory of real closed fields has quantifier elimination [Mar02,

Section 3.3], for any first-order formula ϕpx1, . . . , xmq in the language of ordered fields and any s1ptq, . . . , smptq P K, we have

ϕps1ptq, . . . , smptqq holds in K ðñ ´ ¯ ϕps1pqq, . . . , smpqqq holds in R for all sufficiently small positive real numbers q . ´ ¯ In particular, Tarski’s principle holds for K: A first-order sentence in the language of ordered fields holds in K if and only if it holds in R.

10The main statements in this section are valid over the field of formal Puiseux series as well. 36 PETTER BRAND¨ EN´ AND JUNE HUH

α Definition 3.17. Let ft “ d sαptqw be a nonzero homogeneous polynomial with coeffi- αP∆n cients in . The tropicalization of f is the discrete function defined by Kě0 ř t d troppftq : ∆n ÝÑ R Y t8u, α ÞÝÑ val sαptq . n n We say that ft is log-concave on Ką0 if the function logpfqq is concave` on˘ Rą0 for all sufficiently small positive real numbers q.

Note that the support of ft is the effective domain of the tropicalization of ft. We write d MnpKq for the set of all degree d homogeneous polynomials in Kě0rw1, . . . , wns whose support is M-convex. 0 0 1 1 Definition 3.18 (Lorentzian polynomials over K). We set LnpKq “ MnpKq, LnpKq “ MnpKq, and 2 2 LnpKq “ ft P MnpKq | The Hessian of ft has at most one eigenvalue in Ką0 .

! d ) For d ě 3, we define LnpKq by setting d d α 2 d´2 LnpKq “ ft P MnpKq | B ft P LnpKq for all α P ∆n .

d ! ) The polynomials in LnpKq will be called Lorentzian.

By Proposition 2.33, the log-concavity of homogeneous polynomials can be expressed in the first-order language of ordered fields. It follows that the analog of Theorem 2.30 holds for any homogeneous polynomial ft with coefficients in Kě0.

Theorem 3.19. The following conditions are equivalent for ft.

(1) For any m P N and any m ˆ n matrix paijq with entries in Kě0, m m n Di ft is identically zero or Di ft is log-concave on Ką0, ´ i“1 ¯ ´ i“1 ¯ ź n ź where Di is the differential operator j“1 aijBj. n α n (2) For any α P N , the polynomial B ft řis identically zero or log-concave on Ką0.

(3) The polynomial ft is Lorentzian.

The field K is real closed, and the field K is algebraically closed [Spe05, Section 1.5]. Any element sptq of K can be written as a sum sptq “ pptq ` i qptq, where pptq P K is the real part of sptq and qptq P K is the imaginary part of sptq. The open upper half plane in K is the set of elements in K with positive imaginary parts. A polynomial ft in n Krw1, . . . , wns is stable if ft is non-vanishing on H or identically zero, where H is the open K K upper half plane in K. According to [Bra10¨ , Theorem 4], tropicalizations of homogeneous stable polynomials over K are M-convex functions.11 Here we prove that tropicalizations of Lorentzian

11 In [Bra10¨ ], the field of formal Puiseux series with real exponents Rttu containing K was used. The tropicalization used in [Bra10¨ ] differs from ours by a sign. LORENTZIAN POLYNOMIALS 37

polynomials over K are M-convex, and that all M-convex functions are limits of tropicalizations of Lorentzian polynomials over K.12

d Theorem 3.20. The following conditions are equivalent for any function ν : ∆n Ñ Q Y t8u. (i) The function ν is M-convex.

(ii) There is a Lorentzian polynomial in Krw1, . . . , wns whose tropicalization is ν.

Let M be a matroid with the set of bases B. The Dressian of M, denoted DrpMq, is the tropical variety in RB obtained by intersecting the tropical hypersurfaces of the Plucker¨ relations in RB [MS15, Section 4.4]. Since DrpMq is a rational polyhedral fan whose points bijectively correspond to the valuated matroids with underlying matroid M, Theorem 3.20 shows that

DrpMq “ closure ´ troppftq | ft is a Lorentzian polynomial with supppftq “ B . We note that the corresponding! statement for stable polynomials fails to hold. For example,) when M is the Fano plane, there is no stable polynomial whose support is B [Bra07¨ , Section 6]. We prove Theorems 3.14 and 3.20 together after reviewing the needed results on the space of phylogenetic trees and the isometric embeddings of finite metric spaces in Euclidean spaces. A phylogenetic tree with n leaves is a tree with n labelled leaves and no vertices of degree 2.A n function d : 2 Ñ R is a tree distance if there is a phylogenetic tree τ with n leaves and edge ` weights e P R“ ‰such that

dpi, jq “ the sum of all `e along the unique path in τ joining the leaves i and j . ´ pnq ¯ The space of phylogenetic trees Tn is the set of all tree distances in R 2 . The Fundamental Theorem of Phylogenetics shows that

Tn “ Drp2, nq, where Drp2, nq is the Dressian of the rank 2 uniform matroid on rns [MS15, Section 4.3]. n We give a spectral characterization of tree distances. For any function d : 2 Ñ R and any q n n H d positive real number , we define an ˆ symmetric matrix qp q by “ ‰ 0 if i “ j, H d qp qij “ dpi,jq # q if i ‰ j. We say that an n ˆ n symmetric matrix H is conditionally negative definite if p1,..., 1qw “ 0 ùñ wT Hw ď 0. Basic properties of conditionally negative definite matrices are collected in [BR97, Chapter 4].

n Lemma 3.21. The following conditions are equivalent for any function d : 2 Ñ R.

12 “ ‰ If Rttu is used instead of K, then all M-convex functions are tropicalizations of Lorentzian polynomials. More precisely, a discrete function ν with values in R Y t8u is M-convex if and only if there is a Lorentzian polynomial over Rttu whose tropicalization is ν. In this setting, the Dressian of a matroid M can be identified with the set of tropicalized Lorentzian polynomials ft with supppftq “ B, where B is the set of bases of M. 38 PETTER BRAND¨ EN´ AND JUNE HUH

(1) The matrix Hqpdq is conditionally negative semidefinite for all q ě 1.

(2) The matrix Hqpdq has exactly one positive eigenvalue for all q ě 1. (3) The function d is a tree distance.

Lemma 3.21 is closely linked to the problem of isometric embeddings of ultrametric spaces in Hilbert spaces. Let d be a metric on rns. Since dpi, iq “ 0 and dpi, jq “ dpj, iq for all i, we may n identify d with a function 2 Ñ R. We define an n ˆ n symmetric matrix Epdq by 2 “ ‰ Epdqij “ dpi, jq .

We say that d admits an isometric embedding into Rm if there is φ : rns Ñ Rm such that

dpi, jq “ |φpiq ´ φpjq|2 for all i, j P rns,

m where |¨|2 is the standard Euclidean norm on R . The following theorem of Schoenberg [Sch38] characterizes metrics on rns that admit an isometric embedding into some Rm.

Theorem 3.22. A metric d on rns admits an isometric embedding into some Rm if and only if the matrix Epdq is conditionally negative semidefinite.

Recall that an ultrametric on rns is a metric d on rns such that

dpi, jq ď max dpi, kq, dpj, kq for any i, j, k P rns.

Equivalently, d is an ultrametric if the! maximum of d)pi, jq, dpi, kq, dpj, kq is attained at least twice for any i, j, k P rns. Any ultrametric is a tree distance given by a phylogenetic tree [MS15, Section 4.3]. In [TV83], Timan and Vestfrid proved that any separable ultrametric space is isometric to a

subspace of `2. We use the following special case.

Theorem 3.23. Any ultrametric on rns admits an isometric embedding into Rn´1.

Proof of Lemma 3.21. Cauchy’s interlacing theorem shows (1) ñ (2). We prove (2) ñ (3). We may T suppose that d takes rational values. If (2) holds, then the quadratic polynomial w Hqpdqw is stable for all q ě 1. Therefore, by the quantifier elimination for the theory of real closed fields, the quadratic form ´dpi,jq t wiwj P Krw1, . . . , wns iăj ÿ is stable. By [Bra10¨ , Theorem 4], tropicalizations of stable polynomials are M-convex13, and hence the function ´d is M-convex. In other words, we have

d P Drp2, nq “ Tn.

For (3) ñ (1), we first consider the special case when d is an ultrametric on rns. In this case, d q is also an ultrametric on rns for all q ě 1. It follows from Theorems 3.22 and 3.23 that Hqpdq is

conditionally negative semidefinite for all q ě 1. In the general case, we use that Tn is the sum

13The tropicalization used in [Bra10¨ ] differs from ours by a sign. LORENTZIAN POLYNOMIALS 39

of its linearity space with the space of ultrametrics on rns [MS15, Lemma 4.3.9]. Thus, for any

tree distance d on rns, there is an ultrametric d on rns and real numbers c1, . . . , cn such that n pnq d “ d ` ci eij P R 2 . i“1 i‰j ÿ ´ ÿ ¯ Therefore, the symmetric matrix Hqpdq is congruent to Hqpdq, and the conclusion follows from the case of ultrametrics. 

We start the proof of Theorems 3.14 and 3.20 with a linear algebraic lemma. Let paijq be an n ˆ n symmetric matrix with entries in Rą0.

p Lemma 3.24. If paijq has exactly one positive eigenvalue, then paijq has exactly one positive eigenvalue for 0 ď p ď 1.

aij Proof. If pviq is the Perron eigenvector of paijq, then p q is conditionally negative definite vivj p aij [BR97, Lemma 4.4.1]. Therefore, p p p q is conditionally negative definite [BCR84, Corollary vi vj 2.10], and the conclusion follows. 

Let f be a degree d homogeneous polynomial written in the normalized form c f “ α wα. α! α f Psuppÿ p q For any nonnegative real number p, we define c p R pfq “ α wα. p α! α f Psuppÿ p q We use Lemma 3.24 to construct a homotopy from any Lorentzian polynomial to the generating function of its support. The following proposition was proved in [ALOVII] for strongly log- concave multi-affine polynomials.

Proposition 3.25. If f is Lorentzian, then Rppfq is Lorentzian for all 0 ď p ď 1.

Proof. Using the characterization of Lorentzian polynomials in Theorem 2.25, the proof reduces to the case of quadratic polynomials. Using Theorem 2.10, the proof further reduces to the case 2 f P Pn. In this case, the assertion is Lemma 3.24. 

d d Set m “ nd, and let ν : ∆n Ñ R Y t8u and µ : ∆m Ñ R Y t8u be arbitrary functions. Write m eij for the standard unit vectors in R with 1 ď i ď n and 1 ď j ď d, and let φ be the linear map m n φ : R ÝÑ R , eij ÞÝÑ ei.

Ò d We define the polarization of ν to be the function Π ν : ∆m Ñ R Y t8u satisfying m m dom ΠÒν Ď and ΠÒν “ ν ˝ φ on . d d „  „  ` ˘ 40 PETTER BRAND¨ EN´ AND JUNE HUH

Ó d We define the projection of µ to be the function Π µ : ∆n Ñ R Y t8u satisfying ΠÓµpαq “ min µpβq | φpβq “ α . ! ) It is straightforward to check the symmetric exchange properties of ΠÒν and ΠÓµ from the sym- metric exchange properties of ν and µ.14

d d Lemma 3.26. Let ν : ∆n Ñ R Y t8u and µ : ∆m Ñ R Y t8u be arbitrary functions. (1) If ν is an M-convex function, then ΠÒν is an M-convex function. (2) If µ is an M-convex function, then ΠÓµ is an M-convex function.

As a final preparation for the proof of Theorems 3.14 and 3.20, we show that any M-convex d d function on ∆n can be approximated by M-convex functions whose effective domain is ∆n.

d Lemma 3.27. For any M-convex function ν : ∆n Ñ R Y t8u, there is a sequence of M-convex d functions νk : ∆n Ñ R such that d lim νkpαq “ νpαq for all α P ∆n. kÑ8

The sequence νk can be chosen so that νk “ ν in dompνq and νk ă νk`1 outside dompνq.

Proof. It is enough to prove the case when ν is not the constant function 8. Write eij for the n2 d d d d standard unit vectors in R . Let ϕ : ∆n2 Ñ ∆n and ψ : ∆n2 Ñ ∆n be the restrictions of the 2 linear maps from Rn to Rn given by

ϕpeijq “ ei and ψpeijq “ ej.

d ˚ d For any function µ : ∆n Ñ R Y t8u, we define the function ϕ µ : ∆n2 Ñ R Y t8u by ϕ˚µpβq “ µ ϕpβq .

d ´ ¯ d For any function µ : ∆n2 Ñ R Y t8u, we define the function ψ˚µ : ∆n Ñ R Y t8u by

ψ˚µpαq “ min µpβq | ψpβq “ α . Recall that the operations of splitting [KMT07!, Section 4] and) aggregation [KMT07, Section 5] ˚ preserve M-convexity of discrete functions. Therefore, ϕ and ψ˚ preserve M-convexity. Now, given ν, set ˚ νk “ ψ˚p`k ` ϕ νq, n2 where `k is the restriction of the linear function on R defined by 0 if i “ j, `kpeijq “ # k if i ‰ j. The existence theorem for nonnegative matrices with given row and column sums shows that the restriction of ψ to any fiber of ϕ is surjective [Bru06, Corollary 1.4.2]. Thus, the assumption

14 m In the language of [KMT07], the polarization of ν is obtained from ν by splitting of variables and restricting to d , and the projection of µ is obtained from µ by aggregation of variables. “ ‰ LORENTZIAN POLYNOMIALS 41 that ν is not identically 8 implies that νk ă 8 for every k. It is straightforward to check that the sequence νk has the other required properties for large enough k. 

d Proof of Theorem 3.20, (ii) ñ (i). Let ft be a polynomial in LnpKq whose tropicalization is ν. We show the M-convexity of ν by checking the local exchange property: For any α, β P dompνq with

|α ´ β|1 “ 4, there are i and j satisfying

αi ą βi, αj ă βj and νpαq ` νpβq ě νpα ´ ei ` ejq ` νpβ ´ ej ` eiq.

d´2 Since |α ´ β|1 “ 4, we can find γ in ∆n and indices p, q, r, s in rns such that such that

α “ γ ` ep ` eq and β “ γ ` er ` es and tp, qu X tr, su “ H.

γ γ Since B ft is stable, the tropicalization of B ft is M-convex by [Bra10¨ , Theorem 4]. The conclu- γ sion follows from the local exchange property for the tropicalization of B ft. 

Proof of Theorem 3.14. We prove (1) ñ (3). We first show the implication in the special case n dompνq “ . d „  α ν Since dompνq is M-convex, it is enough to prove that B gq is has exactly one positive eigenvalue n for all α P d´2 and all 0 ă q ď 1. Since Tn “ Drp2, nq by [MS15, Theorem 4.3.5] and [MS15, Definition 4.4.1],“ ‰ the desired statement follows from Lemma 3.21. This proves the first special case. Now consider the second special case

d dompνq “ ∆n.

Ò nd By Lemma 3.26, the polarization Π ν is an M-convex function with effective domain d , and ΠÒν hence we may apply the known implication (1) ñ (3) for . Therefore, “ ‰ 1 Ò ΠÒpgν q “ gΠ ν is a Lorentzian polynomial for 0 ă q ď 1, δ q dd q ν where δ “ pd, . . . , dq. Thus, by Proposition 3.1, the polynomial gq is Lorentzian for all 0 ă q ď 1, and the second special case is proved. Next consider the third special case

dompνq is an arbitrary M-convex set and q “ 1.

By Lemma 3.26, the effective domain of ΠÒν is an M-convex set. Therefore, by Theorem 3.10,

1 Ò ΠÒpgν q “ gΠ ν is a Lorentzian polynomial. δ 1 dd 1 ν Thus, by Proposition 3.1, the polynomial g1 is Lorentzian, and the the third special case is proved. In the remaining case when q ă 1 and the effective domain of ν is arbitrary, we ex- d press ν as the limit of M-convex functions νk with effective domain ∆n using Lemma 3.27. Since q ă 1, we have

ν νk gq “ lim gq . kÑ8 νk Thus the conclusion follows from the second special case applied to each gq . 42 PETTER BRAND¨ EN´ AND JUNE HUH

We prove (1) ñ (2). Introduce a positive real number p, and consider the M-convex function ν ν{p p . Applying the known implication (1) ñ (3), we see that the polynomial gq is Lorentzian for all 0 ă q ď 1. Therefore, by Proposition 3.25, δ p qνpαq R pgν{pq “ pα!qp wα p q α α! α ν Pdomÿ p q ˆ ˙ is Lorentzian for all 0 ă p ď 1. Taking the limit p to zero, we have (2). We prove (2) ñ (1) and (3) ñ (1). By the quantifier elimination for the theory of real closed ν fields, the polynomial ft with coefficients in K is Lorentzian if (2) holds. Similarly, the polyno- ν mial gt is Lorentzian if (3) holds. Since

ν ν ν “ troppft q “ troppgt q, the conclusion follows from (ii) ñ (i) of Theorem 3.20. 

ν Proof of Theorem 3.20, (i) ñ (ii). By Theorem 3.14, fq is Lorentzian for all sufficiently small posi- tive real numbers q. Therefore, by the quantifier elimination for the theory of real closed fields, ν ν the polynomial ft is Lorentzian over K. Clearly, the tropicalization of ft is ν. 

Corollary 3.28. Tropicalizations of Lorentzian polynomials over K are M-convex, and all M- convex functions are limits of tropicalizations of Lorentzian polynomials over K.

d Proof. By Theorem 3.20, it is enough to show that any M-convex function ν : ∆n Ñ R Y t8u is d a limit of M-convex functions νk : ∆n Ñ Q Y t8u. By Lemma 3.27, we may suppose that

d dompνq “ ∆n.

In this case, by Lemma 3.26, the polarization ΠÒν is M-convex function satisfying nd dom ΠÒν “ . d „  ` ˘ In other words, ´ΠÒν is a valuated matroid whose underlying matroid is uniform of rank d on nd elements. Since the Dressian of the matroid is a rational polyhedral fan [MS15, Section 4.4], d there are M-convex functions µk : ∆nd Ñ Q Y t8u satisfying

Ò Π ν “ lim µk. kÑ8 Ó Ò Ó d By Lemma 3.26, ν “ Π Π ν is the limit of M-convex functions Π µk : ∆n Ñ Q Y t8u. 

4. EXAMPLESANDAPPLICATIONS

4.1. Convex bodies and Lorentzian polynomials. For any collection of convex bodies K “ d pK1,..., Knq in R , consider the function

n volK : Rě0 ÝÑ R, w ÞÝÑ volpw1K1 ` ¨ ¨ ¨ ` wnKnq, LORENTZIAN POLYNOMIALS 43

where w1K1 ` ¨ ¨ ¨ ` wnKn is the Minkowski sum and vol is the Euclidean volume. Minkowski noticed that the function volK is a degree d homogeneous polynomial in w “ pw1, . . . , wnq with nonnegative coefficients. We may write d! vol pwq “ V pK ,..., K qw ¨ ¨ ¨ w “ V pKqwα, K i1 id i1 id α! α 1ďi1,...,idďn α ∆d ÿ ÿP n where VαpKq is, by definition, the mixed volume 1 V pKq “ V pK ,..., K ,..., K ,..., K q – Bαvol . α 1 1 n n d! K α1 αn d For any convex bodies C0, C1,...,looooomooooonCd in R , the mixedlooooomooooon volume V pC1, C2,..., Cdq is symmetric in its arguments and satisfies the relation

V pC0 ` C1, C2,..., Cdq “ V pC0, C2,..., Cdq ` V pC1, C2,..., Cdq. We refer to [Sch14] for background on mixed volumes.

Theorem 4.1. The volume polynomial volK is a Lorentzian polynomial for any K “ pK1,..., Knq.

When combined with Theorem 2.25, Theorem 4.1 implies the following statement.

Corollary 4.2. The support of volK is an M-convex for any K “ pK1,..., Knq.

d In other words, the set of all α P ∆n satisfying the non-vanishing condition

V pK1,..., K1,..., Kn,..., Knq ‰ 0

α1 αn d is M-convex for any convex bodieslooooomooooonK1,..., Kn inlooooomooooonR .

Remark 4.3. The mixed volume V pC1,..., Cdq is positive precisely when there are line segments

`i Ď Ci with linearly independent directions [Sch14, Theorem 5.1.8]. Thus, when K consists of n line segments in Rd, Corollary 4.2 states the familiar fact that, for any configuration of n vectors A Ď Rd, the collection of linearly independent d-subsets of A is the set of bases of a matroid. The same reasoning shows that, in fact, the basis generating polynomial of a matroid on rns is the volume polynomial of n convex bodies precisely when the matroid is regular. In particular, not every Lorentzian polynomial is a volume polynomial of convex bodies. For example, the elementary symmetric polynomial

w1w2 ` w1w3 ` w1w4 ` w2w3 ` w2w4 ` w3w4 is not the volume polynomial of four convex bodies in the plane. By the compactness theorem of Shephard for the affine equivalence classes of n convex bodies [She60, Theorem 1], the image d of the set of volume polynomials of convex bodies in PLn is compact. Thus, the displayed elementary symmetric polynomial is not even the limit of volume polynomials of convex bodies. n On the other hand, a collection J Ď d is the support of a volume polynomial of n convex d J d n bodies in R if and only if is the set of“ basis‰ of a rank matroid on r s that is representable 44 PETTER BRAND¨ EN´ AND JUNE HUH over R. For example, there are no seven convex bodies in R3 whose volume polynomial has the support given by the set of bases of the Fano matroid.

Proof of Theorem 4.1. By continuity of the volume functional [Sch14, Theorem 1.8.20], we may

suppose that every convex body in K is d-dimensional. In this case, every coefficient of volK α is positive. Thus, by Theorem 2.25, it is enough to show that B volH is Lorentzian for every d´2 α P ∆n . For this we use a special case of the Brunn-Minkowski theorem [Sch14, Theorem d 7.4.5]: For any convex bodies C3,..., Cd in R , the function

n n 1{2 w ÞÝÑ V wiKi, wiKi, C3,..., Cd ˜ i“1 i“1 ¸ ÿ ÿ n is concave on Rą0. In particular, the function 1{2 2! 1{2 n n Bαvol pwq “ V w K , w K , K ,..., K ,..., K ,..., K d! K i i i i 1 1 n n ˜ i“1 i“1 ¸ ´ ¯ ÿ ÿ α1 αn n d´2 is concave on Rą0 for every α P ∆n . The conclusionlooooomooooon follows from Propositionlooooomooooon 2.33. 

The Alexandrov–Fenchel inequality [Sch14, Section 7.3] states that

2 V pC1, C2, C3,..., Cdq ě V pC1, C1, C3,..., CdqV pC2, C2, C3,..., Cdq. We show that an analog holds for any Lorentzian polynomial.

cα α Proposition 4.4. If f “ d w is a Lorentzian polynomial, then αP∆n α! 2 d cα ěřcα`ei´ej cα´ei`ej for any i, j P rns and any α P ∆n.

α´ei´ej Proof. Consider the Lorentzian polynomial B f. Substituting wk by zero for all k other than i and j, we get the bivariate quadratic polynomial 1 1 c w2 ` c w w c w2. 2 α`ei´ej i α i ` 2 α´ei`ej j 2 The displayed polynomial is Lorentzian by Theorem 2.10, and hence cα ě cα`ei´ej cα´ei`ej . 

We may reformulate Proposition 4.4 as follows. Let f be a homogeneous polynomial of de- n d gree d in n variables. The complete homogeneous form of f is the multi-linear function Ff : pR q Ñ R defined by 1 B B Ff pv1, . . . , vdq “ ¨ ¨ ¨ fpx1v1 ` ¨ ¨ ¨ ` xdvdq. d! Bx1 Bxd Note that the complete homogeneous form of f is symmetric in its arguments. By Euler’s for- mula for homogeneous functions, we have

Ff pw, w, ¨ ¨ ¨ , wq “ fpwq.

n n Proposition 4.5. If f is Lorentzian, then, for any v1 P R and v2, . . . , vd P Rě0, 2 Ff pv1, v2, v3, . . . , vdq ě Ff pv1, v1, v3, . . . , vdqFf pv2, v2, v3, . . . , vdq. LORENTZIAN POLYNOMIALS 45

Proof. For every k “ 1, . . . , d, we write vk “ pvk1, vk2, . . . , vknq, and set B B B Dk “ vk1 ` vk2 ` ¨ ¨ ¨ ` vkn . Bw1 Bw2 Bwn

By Corollary 2.11, the quadratic polynomial D3 ¨ ¨ ¨ Ddf is Lorentzian. We may suppose that the T Hessian H of the quadratic polynomial is not identically zero and v2 Hv2 ą 0. Note that T vi Hvj “ DiDjD3 ¨ ¨ ¨ Ddf “ d!Ff pvi, vj, v3, . . . , vdq for any i and j. Since H has exactly one positive eigenvalue, the conclusion follows from Cauchy’s interlacing theorem. 

4.2. Projective varieties and Lorentzian polynomials. Let Y be a d-dimensional irreducible projective variety over an algebraically closed field F. If D1,..., Dd are Cartier divisors on Y , the intersection product pD1 ¨ ... ¨ Ddq is an integer defined by the following properties:

– the product pD1 ¨ ... ¨ Ddq is symmetric and multilinear as a function of its arguments,

– the product pD1 ¨ ... ¨ Ddq depends only on the linear equivalence classes of the Di, and

– if D1,..., Dn are effective divisors meeting transversely at smooth points of Y , then

pD1 ¨ ... ¨ Ddq “ #D1 X ... X Dd.

Given an irreducible subvariety X Ď Y of dimension k, the intersection product

pD1 ¨ ... ¨ Dk ¨ Xq is then defined by replacing each divisor Di with a linearly equivalent Cartier divisor whose support does not contain X and intersecting the restrictions of Di in X. The definition of the in- tersection product linearly extends to Q-linear combination of Cartier divisors, called Q-divisors [Laz04, Section 1.3]. If D is a Q-divisor on Y , we write pDqd for the self-intersection pD ¨ ... ¨ Dq. For a gentle introduction to Cartier divisors and their intersection products, we refer to [Laz04, Section 1.1]. See [Ful98] for a comprehensive study.

Let H “ pH1,..., Hnq be a collection of Q-divisors on Y . We define the volume polynomial of H by d! vol pwq “ pw H ` ¨ ¨ ¨ ` w H qd “ V pHqwα, H 1 1 n n α! α α ∆d ÿP n where VαpHq is the intersection product 1 V pHq “ pH ¨ ... ¨ H ¨ ... ¨ H ¨ ... ¨ H q “ Bαvol . α 1 1 n n d! H α1 αn A Q-divisor D on Y is said to belooooomooooonnef if pD ¨ Cq ěloooooomoooooon0 for every irreducible curve C in Y [Laz04, Section 1.4].

Theorem 4.6. If H1,..., Hn are nef divisors on Y , then volHpwq is a Lorentzian polynomial.

When combined with Theorem 2.25, Theorem 4.6 implies the statement. 46 PETTER BRAND¨ EN´ AND JUNE HUH

Corollary 4.7. If H1,..., Hn are nef divisors on Y , then the support of volHpwq is M-convex.

d In other words, the set of all α P ∆n satisfying the non-vanishing condition

pH1 ¨ ... ¨ H1 ¨ ... ¨ Hn ¨ ... ¨ Hnq ‰ 0

α1 αn is M-convex for any d-dimensionallooooomooooon projective varietyloooooomoooooonY and any nef divisors H1,..., Hn on Y . Corollary 4.7 implies a result of Castillo et al. [CRLMZ, Proposition 5.4], which says that the support of the multidegree of any irreducible mutiprojective variety is a discrete polymatroid.

d Remark 4.8. Let A “ tv1, . . . , vnu be a collection of n vectors in F . In [HW17, Section 4], one can find a d-dimensional projective variety YA and nef divisors H1,..., Hn on YA such that

α volHpwq “ cαw , αP n ÿrds where cα “ 1 if α corresponds to a linearly independent subset of A and cα “ 0 if otherwise. Thus, in this case, Corollary 4.7 states the familiar fact that the collection of linearly independent d-subsets of A Ď Fd is the set of bases of a matroid.

Proof of Theorem 4.6. By Kleiman’s theorem [Laz04, Section 1.4], every nef divisor is a limit of ample divisors, and we may suppose that every divisor in H is very ample. In this case, ev- α ery coefficient of volH is positive. Thus, by Theorem 2.25, it is enough to show that B volH is d´2 Lorentzian for every α P ∆n . Note that 2! n n Bαvol pwq “ w H ¨ w H ¨ H ¨ ... ¨ H ¨ ... ¨ H ¨ ... ¨ H . d! H i i i i 1 1 n n ˜ i“1 i“1 ¸ ÿ ÿ α1 αn By Bertini’s theorem [Laz04, Section 3.3], there is anlooooomooooon irreducible surfaceloooooomoooooonS Ď Y such that 2! n n Bαvol pwq “ w H ¨ w H ¨ S . d! H i i i i ˜ i“1 i“1 ¸ ÿ ÿ If S is smooth, then the Hodge index theorem [Har77, Theorem V.1.9] shows that the displayed quadratic form has exactly one positive eigenvalue. In general, the Hodge index theorem ap- plied to any resolution of singularities of S implies the one positive eigenvalue condition, by the projection formula [Ful98, Example 2.4.3]. 

The theory of toric varieties shows that the volume polynomial of any set of convex bodies is the limit of a sequence of volume polynomials of nef divisors on projective varieties [Ful93, Section 5.4]. How large is the set of volume polynomials of projective varieties within the set d of Lorentzian polynomials? We formulate various precise versions of this question. Let VnpFq be the set of volume polynomials of n nef divisors on a d-dimensional projective variety over F, and let Vd “ Vd p q, where the union is over all algebraically closed fields. n F n F Question 4.9.ŤFix any algebraically closed field F. LORENTZIAN POLYNOMIALS 47

d d (1) Is there a polynomial in Ln that is not in the closure of Vn? d d (2) Is there a polynomial in Ln that is not in the closure of VnpFq? d d (3) Is there a polynomial in Ln X Qrws that is not in Vn? d d (4) Is there a polynomial in Ln X Qrws that is not in VnpFq?

d Shephard’s construction in [She60, Section 3] shows that every polynomial in L2 XQrws is the volume polynomial of a pair of d-dimensional convex polytopes with rational vertices. Thus, by [Ful93, Section 5.4], we have

d d d L2 X Qrws “ V2 “ V2pFq for any d and any F. A similar reasoning based on the construction of [Hei38, Section I] shows that

2 2 2 L3 X Qrws “ V3 “ V3pFq for any F. When n ě 4, not every Lorentzian polynomial is the limit of a sequence of volume polynomials of rational convex polytopes (Remark 4.3), and we do not know how to answer any of the above questions.

4.3. Potts model partition functions and Lorentzian polynomials. The q-state Potts model, or the random-cluster model, of a graph is a much studied class of measures introduced by Fortuin and Kasteleyn [FK72]. We refer to [Gri06] for a comprehensive introduction to random-cluster models.

Let M be a matroid on rns, and let rkM be the rank function of M. For a nonnegative integer k and a positive real parameter q, consider the degree k homogeneous polynomial in n variables

k ´rkMpAq A Zq,Mpwq “ q w , w “ pw1, . . . , wnq. AP n ÿrks We define the homogeneous multivariate Tutte polynomial of M by n k n´k Zq,Mpw0, w1, . . . , wnq “ Zq,Mpwq w0 , k“0 ÿ which is a homogeneous polynomial of degree n in n`1 variables. When M is the cycle matroid of a graph G, the polynomial obtained from Zq,M by setting w0 “ 1 is the partition function of the q-state Potts model associated to G [Sok05]. Since the rank function of a matroid is M6-concave, the normalized rank generating function of M is Lorentzian when the parameter q satisfies 0 ă q ď 1, see Example 3.15. In this subsection, we prove the following refinement.

Theorem 4.10. For any matroid M and 0 ă q ď 1, the polynomial Zq,M is Lorentzian.

We prepare the proof with two simple lemmas.

Lemma 4.11. The support of Zq,M is M-convex for all 0 ă q ď 1. 48 PETTER BRAND¨ EN´ AND JUNE HUH

6 Proof. Writing Zq,M for the polynomial obtained from Zq,M by setting w0 “ 1, we have 6 n supp Zq,M “ 0, 1 . n It is straightforward to verify the augmentation` ˘ property ( in Lemma 2.21 for t0, 1u . 

For a nonnegative integer k and a subset S Ď rns, we define a degree k homogeneous poly- k nomial eSpwq by the equation n k A eSpwq “ w . k“0 AĎS k ÿ ÿ In other words, eSpwq is the k-th elementary symmetric polynomial in the variables twiuiPS.

Lemma 4.12. If S1 \ ... \ Sm is a partition of rns into m nonempty parts, then 1 e1 pwq2 ď e1 pwq2 ` ¨ ¨ ¨ ` e1 pwq2 for all w P n. n rns S1 Sm R

Proof. Since m ď n, it is enough to prove the statement when m “ n. In this case, we have

2 2 2 pw1 ` ¨ ¨ ¨ ` wnq ď npw1 ` ¨ ¨ ¨ ` wnq,

n by the Cauchy-Schwarz inequality for the vectors p1,..., 1q and pw1, . . . , wnq in R . 

n´2 Proof of Theorem 4.10. Let α be an element of ∆n`1. By Theorem 2.25 and Lemma 4.11, the proof α reduces to the statement that the quadratic form B Zq,M is stable. We prove the statement by induction on n. The assertion is clear when n “ 1, so suppose n ě 2. When i ‰ 0, we have

´rkMpiq BiZq,M “ q Zq,M{i, where M{i is the contraction of M by i [Oxl11, Chapter 3]. Thus, it is enough to prove that the following quadratic form is stable: n! w2 ` pn ´ 1q!Z1 pwqw ` pn ´ 2q!Z2 pwq. 2 0 q,M 0 q,M Recall that a homogeneous polynomial f with nonnegative coefficients in n ` 1 variables is stable if and only if the univariate polynomial fpxu ´ vq has only real zeros for all v P Rn`1 for n`1 some u P Rě0 satisfying fpuq ą 0. Therefore, it suffices to show that the discriminant of the displayed quadratic form with respect to w0 is nonnegative:

1 2 n 2 n Zq,Mpwq ě 2 Zq,Mpwq for all w P R . n ´ 1 We prove the inequality after making the change of variables

wi if i is a loop in M, wi ÞÝÑ # qwi if i is not a loop in M.

Write L Ď rns for the set of loops and P1,...,P` Ď rnszL for the parallel classes in M [Oxl11, Section 1.1]. The above change of variables gives Z1 w e1 w Z2 w e2 w 1 q e2 w e2 w . q,Mp q “ rnsp q and q,Mp q “ rnsp q ´ p ´ q P1 p q ` ¨ ¨ ¨ ` P` p q ` ˘ LORENTZIAN POLYNOMIALS 49

When q “ 1, the desired inequality directly follows from the case m “ n of Lemma 4.12. There- fore, when proving the desired inequality for an arbitrary 0 ă q ď 1, we may assume that e2 w e2 w 0. P1 p q ` ¨ ¨ ¨ ` P` p q ă 2 Therefore, exploiting the monotonicity of Zq,M in q, the desired inequality reduces to n 1 e1 w 2 2n e2 w e2 w e2 w 0. p ´ q rnsp q ´ rnsp q ´ P1 p q ´ ¨ ¨ ¨ ´ P` p q ě Note that the left-hand side of the above´ inequality simplifies to ¯ n e1 w 2 e1 w 2 w2 e1 w 2. P1 p q ` ¨ ¨ ¨ ` P` p q ` i ´ rnsp q iPL ´ ÿ ¯ The conclusion now follows from Lemma 4.12. 

Mason [Mas72] offered the following three conjectures of increasing strength. Several au- thors studied correlations in matroid theory partly in pursuit of these conjectures [SW75, Wag08, BBL09, KN10, KN11]. Conjecture 4.13. For any matroid M on rns and any positive integer k,

2 (1) IkpMq ě Ik´1pMqIk`1pMq, 2 k`1 (2) IkpMq ě k Ik´1pMqIk`1pMq, 2 k`1 n´k`1 (3) IkpMq ě k n´k Ik´1pMqIk`1pMq, where IkpMq is the number of k-element independent sets of M.

Conjecture 4.13(1) was proved in [AHK18], and Conjecture 4.13(2) was proved in [HSW]. Note that Conjecture 4.13(3) may be written I pMq2 I pMq I pMq k ě k`1 k´1 , n 2 n n k k`1 k´1 and the equality holds when all pk``˘ 1q-subsets` of˘ rns` are˘ independent in M. Conjecture 4.13 (3) is known to hold when n is at most 11 or k is at most 5 [KN11]. See [Sey75, Dow80, Mah85, Zha85, HK12, HS89, Len13] for other partial results. Theorem 4.14. For any matroid M on rns and any positive integer k, I pMq2 I pMq I pMq k ě k`1 k´1 , n 2 n n k k`1 k´1 where IkpMq is the number of k-element` ˘ independent` ˘ sets` of˘ M.

In [BH], direct proofs of Theorems 4.10 and 4.14 were given.15 Here we deduce Theorem 4.14 from the Lorentzian property of A n´|A| fMpw0, w1, . . . , wnq “ w w0 , w “ pw1, . . . , wnq, A I M Pÿp q

15An independent proof of 4.14 was given by Anari et al. in [ALOVIII]. 50 PETTER BRAND¨ EN´ AND JUNE HUH where IpMq is the collection of independent sets of M.

Proof of Theorem 4.14. The polynomial fM is Lorentzian by Theorem 4.10 and the identity

fMpw0, w1, . . . , wnq “ lim Zq,Mpw0, qw1, . . . , qwnq. qÑ0

Therefore, by Theorem 2.10, the bivariate polynomial obtained from fM by setting w1 “ ¨ ¨ ¨ “ wn is Lorentzian. The conclusion follows from the fact that a bivariate homogeneous polynomial with nonnegative coefficients is Lorentzian if and only if the sequence of coefficients form an ultra log-concave sequence with no internal zeros. 

The Tutte polynomial of a matroid M on rns is the bivariate polynomial

rkMprnsq´rkMpAq |A|´rkMpAq TMpx, yq “ px ´ 1q py ´ 1q . A n ÿĎr s Theorem 4.10 reveals several nontrivial inequalities satisfied by the coefficients of the Tutte poly- nomial. For example, if we write

q n n wrkMprnsqT 1 ` , 1 ` w “ qrkMprnsq´rkMpAq wk “ ckpMqwk, M w q k“0 AP n k“0 ´ ¯ ÿ ´ ÿrks ¯ ÿ k then the sequence cq pMq is ultra log-concave whenever 0 ď q ď 1. This and other results in this subsection are recently extended to flag matroids in [EH20].

4.4. M-matrices and Lorentzian polynomials. We write In for the n ˆ n identity matrix, Jn for the n ˆ n matrix all of whose entries are 1, and 1n for the n ˆ 1 matrix all of whose entries are

1. Let A “ paijq be an n ˆ n matrix with real entries. The following conditions are equivalent if aij ď 0 for all i ‰ j [BP94, Chapter 6]:

– The real part of each nonzero eigenvalue of A is positive. – The real part of each eigenvalue of A is nonnegative. – All the principal minors of A are nonnegative. – Every real eigenvalue of A is nonnegative.

– The matrix A `  In is nonsingular for every  ą 0.

– The univariate polynomial detpxIn ` Aq has nonnegative coefficients.

The matrix A is an M-matrix if aij ď 0 for all i ‰ j and if it satisfies any one of the above conditions. One can find 50 different characterizations of nonsingular M-matrices in [BP94, Chapter 6]. We will use the 29-th condition: There are positive diagonal matrices D and D1 such that DAD1 has all diagonal entries 1 and all row sums positive. For a discussion of M-matrices in the context of ultrametrics and potentials of finite Markov chains, see [DMS14]. LORENTZIAN POLYNOMIALS 51

We define the multivariate characteristic polynomial of A by the equation

pApw0, w1, . . . , wnq “ det w0In ` diagpw1, . . . , wnqA . ´ ¯ In [Hol05, Theorem 4], Holtz proved that the coefficients of the characteristic polynomial of an M-matrix form an ultra log-concave sequence. We will strengthen this result and prove that the multivariate characteristic polynomial of an M-matrix is Lorentzian.

Theorem 4.15. If A is an M-matrix, then pA is a Lorentzian polynomial.

Using Example 2.26, we may recover the theorem of Holtz by setting w1 “ ¨ ¨ ¨ “ wn.

Corollary 4.16. If A is an M-matrix, then the support of pA is M-convex.

Since every M-matrix is a limit of nonsingular M-matrices, it is enough to prove Theorem 4.15 for nonsingular M-matrices.

Lemma 4.17. If A is a nonsingular M-matrix, the support of pA is M-convex.

6 6 Proof. It is enough to prove that the support of pA is M -convex, where 6 pApw1, . . . , wnq “ pAp1, w1, . . . , wnq. If A is a nonsingular M-matrix, then all the principal minors of A are positive, and hence

6 n suppppAq “ t0, 1u .

n It is straightforward to verify the augmentation property in Lemma 2.21 for t0, 1u . 

We prepare the proof of Theorem 4.15 with a proposition on doubly sub-stochastic matrices.

Recall that an n ˆ n matrix B “ pbijq with nonnegative entries is said to be doubly sub-stochastic if n n bij ď 1 for every i and bij ď 1 for every j. j“1 i“1 ÿ ÿ A partial permutation matrix is a zero-one matrix with at most one nonzero entry in each row and column. We use Mirsky’s analog of the Birkhoff-von Neumann theorem for doubly sub- stochastic matrices [HJ94, Theorem 3.2.6]: The set of n ˆ n doubly sub-stochastic matrix is equal to the convex hull of the n ˆ n partial permutation matrices.

Lemma 4.18. For n ě 2, define n ˆ n matrices Mn and Nn by

2 1 0 ¨ ¨ ¨ 0 1 2 1 0 ¨ ¨ ¨ 0 0 1 2 1 ¨ ¨ ¨ 0 0 1 2 1 ¨ ¨ ¨ 0 0 ¨ ˛ ¨ ˛ 0 1 2 ¨ ¨ ¨ 0 0 0 1 2 ¨ ¨ ¨ 0 0 Mn “ ˚ ‹ , Nn “ ˚ ‹ . ˚ ...... ‹ ˚ ...... ‹ ˚ ...... ‹ ˚ ...... ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 ‹ ˚ ‹ ˚ ‹ ˚ 1 0 0 ¨ ¨ ¨ 1 2 ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 2 ‹ ˚ ‹ ˚ ‹ ˝ ‚ ˝ ‚ 52 PETTER BRAND¨ EN´ AND JUNE HUH

2 2 Then the matrices Mn ´ n Jn and Nn ´ n Jn are positive semidefinite. Equivalently,

Mn 1n Nn 1n Mn`1 – T n , Nn`1 – T n ˜ 1n 2 ¸ ˜ 1n 2 ¸ are positive semidefinite.

Proof. We define symmetric matrices Ln`1 and Kn`1 by

1 1 1 1 0 ¨ ¨ ¨ 0 0 2 1 1 0 ¨ ¨ ¨ 0 0 2 1 2 1 ¨ ¨ ¨ 0 0 1 1 2 1 ¨ ¨ ¨ 0 0 1 ¨ ˛ ¨ ˛ 0 1 2 ¨ ¨ ¨ 0 0 1 0 1 2 ¨ ¨ ¨ 0 0 1 ˚ ...... ‹ ˚ ...... ‹ Ln 1 “ ˚ ...... ‹ , Kn 1 “ ˚ ...... ‹ . ` ˚ ...... ‹ ` ˚ ...... ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 1 ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 2 1 ‹ ˚ 2 ‹ ˚ ‹ ˚ 1 1 1 ¨ ¨ ¨ 1 1 n ‹ ˚ 1 1 1 ¨ ¨ ¨ 1 1 n ‹ ˚ 2 2 2 ‹ ˚ 2 2 ‹ ˝ ‚ ˝ ‚ As before, the subscript indicates the size of the matrix. We show, by induction on n, that the

matrices Ln`1 and Kn`1 are positive semidefinite. It is straightforward to check that L3 and

K3 are positive semidefinite. Perform the symmetric row and column elimination of Ln`1 and

Kn`1 based on their 1 ˆ 1 entries, and notice that

1 0 0 ¨ ¨ ¨ 0 0 0 1 0 0 ¨ ¨ ¨ 0 0 0 0 1 1 ¨ ¨ ¨ 0 0 1 0 1 1 ¨ ¨ ¨ 0 0 1 ¨ 2 ˛ ¨ 2 ˛ 0 1 2 ¨ ¨ ¨ 0 0 1 0 1 2 ¨ ¨ ¨ 0 0 1 ˚ ...... ‹ ˚ ...... ‹ Ln 1 » ˚ ...... ‹ , Kn 1 » ˚ ...... ‹ , ` ˚ ...... ‹ ` ˚ ...... ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 1 ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 1 1 ‹ ˚ 2 ‹ ˚ ‹ ˚ 0 1 1 ¨ ¨ ¨ 1 1 n ´ 1 ‹ ˚ 0 1 1 ¨ ¨ ¨ 1 1 n ´ 1 ‹ ˚ 2 2 2 4 ‹ ˚ 2 2 4 ‹ ˝ ‚ ˝ ‚ where the symbol » stands for the congruence relation for symmetric matrices. Since Ln is

positive semidefinite, Ln`1 is congruent to the sum of positive semidefinite matrices, and hence

Ln`1 is positive semidefinite. Similarly, since Kn is positive semidefinite, Kn`1 is congruent to

the sum of positive semidefinite matrices, and hence Kn`1 is positive semidefinite.

We now prove that the symmetric matrices Mn`1 and Nn`1 are positive semidefinite. Per- form the symmetric row and column elimination of Mn`1 and Nn`1 based on their 1 ˆ 1 entries, and notice that 2 0 0 ¨ ¨ ¨ 0 0 0 2 0 0 ¨ ¨ ¨ 0 0 0 0 3 1 ¨ ¨ ¨ 0 ´ 1 1 0 3 1 ¨ ¨ ¨ 0 0 1 ¨ 2 2 2 ˛ ¨ 2 2 ˛ 0 1 2 ¨ ¨ ¨ 0 0 1 0 1 2 ¨ ¨ ¨ 0 0 1 ˚ ...... ‹ ˚ ...... ‹ M » ˚ ...... ‹ , N » ˚ ...... ‹ . n`1 ˚ ...... ‹ n`1 ˚ ...... ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 2 1 1 ‹ ˚ 1 3 1 ‹ ˚ ‹ ˚ 0 ´ 0 ¨ ¨ ¨ 1 ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 2 1 ‹ ˚ 2 2 2 ‹ ˚ ‹ ˚ 0 1 1 ¨ ¨ ¨ 1 1 n´1 ‹ ˚ 0 1 1 ¨ ¨ ¨ 1 1 n´1 ‹ ˚ 2 2 2 ‹ ˚ 2 2 ‹ ˝ ‚ ˝ ‚ LORENTZIAN POLYNOMIALS 53

Since Ln is positive semidefinite, Mn`1 is congruent to the sum of two positive semidefinite

matrices, and hence Mn`1 is positive semidefinite. Similarly, since Kn is positive semidefinite,

Nn`1 is congruent to the sum of two positive semidefinite matrices, and hence Nn`1 is positive semidefinite. 

T 2 Proposition 4.19. If B is an n ˆ n doubly sub-stochastic matrix, then 2In ` B ` B ´ n Jn is positive semidefinite.

T Proof. Let Cn be the symmetric matrix 2In ` B ` B , and let Cn`1 be the symmetric matrix

Cn 1n Cn`1 – T n . ˜ 1n 2 ¸

It is enough to prove that Cn`1 is positive semidefinite. Since the convex hull of the partial per- mutation matrices is the set of doubly sub-stochastic matrix, the proof reduces to the case when B is a partial permutation matrix. We use the following extension of the cycle decomposition for partial permutations: For any partial permutation matrix B, there is a permutation matrix P such that PBPT is a block diagonal matrix, where each block diagonal is either zero, identity,

0 0 0 ¨ ¨ ¨ 0 1 0 0 0 ¨ ¨ ¨ 0 0 1 0 0 ¨ ¨ ¨ 0 0 1 0 0 ¨ ¨ ¨ 0 0 ¨ ˛ ¨ ˛ 0 1 0 ¨ ¨ ¨ 0 0 0 1 0 ¨ ¨ ¨ 0 0 ˚ ‹ or ˚ ‹ . ˚ ...... ‹ ˚ ...... ‹ ˚ ...... ‹ ˚ ...... ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 0 0 ‹ ˚ 0 0 0 ¨ ¨ ¨ 0 0 ‹ ˚ ‹ ˚ ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 0 ‹ ˚ 0 0 0 ¨ ¨ ¨ 1 0 ‹ ˚ ‹ ˚ ‹ ˝ ‚ ˝ ‚ Using the cyclic decomposition for B, we can express the matrix Cn`1 as a sum, where each summand is positive semidefinite by Lemma 4.18. 

The remaining part of the proof of Theorem 4.15 parallels that of Theorem 4.10.

Proof of Theorem 4.15. Since every M-matrix is a limit of nonsingular M-matrices, we may sup- pose that A is a nonsingular M-matrix. For k “ 0, 1, . . . , n, we set

k α pApwq “ Aαw , w “ pw1, . . . , wnq, αP n ÿrks

where Aα is the principal minor of A corresponding to α, so that n k n´k pApw0, w1, . . . , wnq “ pApwqw0 . k“0 ÿ Lemma 4.17 shows that the support of pA is M-convex. Therefore, by Theorem 2.25, it is enough

to prove that BippAq is Lorentzian for i “ 0, 1, . . . , n. We prove this statement by induction on n. The assertion is clear when n “ 1, so suppose n ě 2. When i ‰ 0, we have

BippAq “ pA{i, 54 PETTER BRAND¨ EN´ AND JUNE HUH where A{i is the M-matrix obtained from A by deleting the i-th row and column. Thus, it is enough to prove that the following quadratic form is stable: n! w2 ` pn ´ 1q! p1 pwqw ` pn ´ 2q! p2 pwq. 2 0 A 0 A As in the proof of Theorem 4.10 it suffices to show that the discriminant of the displayed qua-

dratic form with respect to w0 is nonnegative:

1 2 2n 2 n pApwq ě pApwq for all w P R . n ´ 1 In terms of the entries of A, the displayed inequality is equivalent to the statement that the 1 matrix aijaji ´ n aiiajj is positive semidefinite. According to the 29-th characterization of nonsingular´ M-matrices¯ in [BP94, Chapter 6], there are positive diagonal matrices D and D1 such that DAD1 has all diagonal entries 1 and all row sums positive. Therefore, we may suppose that A has all diagonal entries 1 and all the row sums of A are positive. Under this assumption, 1 1 a a ´ a a “ I ´ B ´ J , ij ji n ii jj n n n where ´B is a symmetric doubly´ sub-stochastic¯ matrix all of whose diagonal entries are zero. The conclusion follows from Proposition 4.19. 

4.5. Lorentzian probability measures. There are numerous important examples of negatively dependent “repelling” random variables in probability theory, combinatorics, stochastic pro- cesses, and statistical mechanics. See, for example, [Pem00]. A theory of negative dependence for strongly Rayleigh measures was developed in [BBL09], but the theory is too restrictive for several applications. Here we introduce a broader class of discrete probability measures using the Lorentzian property. A discrete probability measure µ on t0, 1un is a probability measure on t0, 1un such that all subsets of t0, 1un are measurable. The partition function of µ is the polynomial

Zµpwq “ µ tSu wi. SĎrns iPS ÿ ` ˘ ź The following notions capture various aspects of negative dependence:

– The measure µ is pairwise negatively correlated (PNC) if for all distinct i and j in rns,

µpEi X Ejq ď µpEiqµpEjq,

where Ei is the collection of all subsets of rns containing i. – The measure µ is ultra log-concave (ULC) if for every positive integer k ă n, 2 n n n µ k µ k´1 µ k`1 ě . ´ n 2¯ ´ n ¯ ´ n ¯ “k ‰ “k´1 ‰ “k`1 ‰ – The measure µ is strongly Rayleigh` ˘if for all distinct` ˘ i and` j˘in rns, n Zµpwq BiBjZµpwq ď BiZµpwq BjZµpwq for all w P R . LORENTZIAN POLYNOMIALS 55

Let P be a property of discrete probability measures. We say that µ has property P if, for every n n x P Rą0, the discrete probability measure on t0, 1u with the partition function

Zµpx1w1, . . . , xnwnq{Zµpx1, . . . , xnq

has property P. The new discrete probability measure is said to be obtained from µ by applying n the external field x P Rą0. For example, the property PNC for µ is equivalent to the 1-Rayleigh property

n Zµpwq BiBjZµpwq ď BiZµpwq BjZµpwq for all distinct i, j in rns and all w P Rą0. More generally, for a positive real number c, we say that µ is c-Rayleigh if

n Zµpwq BiBjZµpwq ď c BiZµpwq BjZµpwq for all distinct i, j in rns and all w P Rą0.

Definition 4.20. A discrete probability measure µ on t0, 1un is Lorentzian if the homogenization n of the partition function w0 Zµpw1{w0, . . . , wn{w0q is a Lorentzian polynomial.

For example, if A is an M-matrix of size n, the probability measure on t0, 1un given by

µ tSu 9 the principal minor of A corresponding to S ,S Ď rns,

is Lorentzian by` Theorem˘ ´ 4.15. Results from the previous sections¯ reveal basic features of Lorentzian measures, some of which may be interpreted as negative dependence properties.

Proposition 4.21. If µ is Lorentzian, then µ is 2-Rayleigh.

1 Proof. Lemma 2.20 and Proposition 2.19 show that Zµ is a 2 1 ´ n -Rayleigh polynomial.  ´ ¯ Proposition 4.22. If µ is Lorentzian, then µ is ULC.

Proof. Since any probability measure obtained from a Lorentzian probability measure by ap- plying an external field is Lorentzian, it suffices to prove that µ is ULC. By Theorem 2.10, the n bivariate homogeneous polynomial w0 Zµpw1{w0, . . . , w1{w0q is Lorentzian. Therefore, by Ex- ample 2.26, its sequence of coefficients must be ultra log-concave. 

Proposition 4.23. The class of Lorentzian measures is preserved under the symmetric exclusion process.

Proof. The statement is Corollary 3.9 for homogenized partition functions of Lorentzian proba- bility measures. 

Proposition 4.24. If µ is strongly Rayleigh, then µ is Lorentzian.

Proof. A multi-affine polynomial is stable if and only if it is strongly Rayleigh [Bra07¨ , Theorem 5.6], and a polynomial with nonnegative coefficients is stable if and only if its homogeniza- tion is stable [BBL09, Theorem 4.5]. By Proposition 2.2, homogeneous stable polynomials with nonnegative coefficients are Lorentzian.  56 PETTER BRAND¨ EN´ AND JUNE HUH

n For a matroid M on rns, we define probability measures µM and νM on t0, 1u by

n µM “ the uniform measure on t0, 1u concentrated on the independent sets of M, n νM “ the uniform measure on t0, 1u concentrated on the bases of M.

Proposition 4.25. For any matroid M on rns, the measures µM and νM are Lorentzian.

Proof. Note that the homogenized partition function fM of µM satisfies

fMpw0, w1, . . . , wnq “ lim Zq,Mpw0, qw1, . . . , qwnq. qÑ0

Since a limit of Lorentzian polynomials is Lorentzian, µM is Lorentzian by Theorem 4.10. The partition function of νM is Lorentzian by Theorem 3.10. 

Let G be an arbitrary finite graph and let i and j be any distinct edges of G. A conjecture of Kahn [Kah00] and Grimmett–Winkler [GW04] states that, if F is a forest in G chosen uniformly at random, then

PrpF contains i and jq ď PrpF contains iq PrpF contains jq.

The conjecture is equivalent to the statement that µM is 1-Rayleigh for any graphic matroid M.

Propositions 4.21 and 4.25 show that µM is 2-Rayleigh for any matroid M.

REFERENCES

[AHK18] Karim Adiprasito, June Huh, and Eric Katz, Hodge theory for combinatorial geometries. Ann. of Math. (2) 188 (2018), no. 2, 381–452.5, 49 [AOVI] Nima Anari, Shayan Oveis Gharan, and Cynthia Vinzant, Log-Concave Polynomials I: Entropy and a Deterministic Approximation Algorithm for Counting Bases of Matroids. arXiv:1807.00929.3,5, 23, 24, 32 [ALOVII] Nima Anari, Kuikui Liu, Shayan Oveis Gharan, and Cynthia Vinzant, Log-Concave Polynomials II: High- dimensional walks and an FPRAS for counting bases of a matroid. arXiv:1811.0181.5, 39 [ALOVIII] Nima Anari, Kuikui Liu, Shayan Oveis Gharan, and Cynthia Vinzant, Log-Concave Polynomials III: Mason’s Ultra-Log-Concavity Conjecture for Independent Sets of Matroids. arXiv:1811.01600.5, 23, 49 [BR97] Ravindra Bapat and Tirukkannamangai Raghavan, Nonnegative matrices and applications. Encyclopedia of Math- ematics and its Applications 64. Cambridge University Press, Cambridge, 1997. 37, 39 [BCR84] Christian Berg, Jens Christensen, and Paul Ressel, Harmonic analysis on semigroups. Theory of positive definite and related functions. Graduate Texts in Mathematics 100. Springer-Verlag, New York, 1984. 39 [BP94] Abraham Berman, and Robert Plemmons, Nonnegative matrices in the mathematical sciences. Revised reprint of the 1979 original. Classics in Applied Mathematics 9. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 19945, 50, 54 [BB08] Julius Borcea and Petter Brand¨ en,´ Applications of stable polynomials to mixed determinants: Johnson’s conjectures, unimodality, and symmetrized Fischer products, Duke Math. J. 143 (2008), 205–223.8 [BB09] Julius Borcea and Petter Brand¨ en,´ The Lee–Yang and P´olya–Schurprograms. I. Linear operators preserving stability. Invent. Math. 177 (2009), 541–569.3, 26, 28 [BB10] Julius Borcea and Petter Brand¨ en,´ Multivariate P´olya–Schurclassification problems in the Weyl algebra. Proc. Lond. Math. Soc. 101 (2010), no. 1, 73–104. 11, 29 [BBL09] Julius Borcea, Petter Brand¨ en,´ and Thomas Liggett, Negative dependence and the geometry of polynomials. J. Amer. Math. Soc. 22 (2009), no. 2, 521–567.5,7, 11, 15, 17, 18, 31, 49, 54, 55 LORENTZIAN POLYNOMIALS 57

[Bra07]¨ Petter Brand¨ en,´ Polynomials with the half-plane property and matroid theory. Adv. Math. 216 (2007), no. 1, 302–320. 3, 10, 21, 37, 55 [Bra10]¨ Petter Brand¨ en,´ Discrete concavity and the half-plane property, Siam J. Discrete Math. 24 (2010), 921–933.4, 36, 38, 41 [BH] Petter Brand¨ en´ and June Huh, Hodge–Riemann relations for Potts model partition functions. arXiv:1811.01696.5, 49 [Bru06] Richard Brualdi, Combinatorial matrix classes. Encyclopedia of Mathematics and its Applications 108. Cambridge University Press, Cambridge, 2006. 40 [CRLMZ] Federico Castillo, Yairon Cid Ruiz, Binglin Li, Jonathan Montano,˜ and Naizhen Zhang, When are multidegrees positive? arXiv:2005.07808. 46 [COSW04] Youngbin Choe, James Oxley, Alan Sokal, and David Wagner, Homogeneous multivariate polynomials with the half-plane property. Adv. in Appl. Math. 32 (2004), no. 1-2, 88–187.3,7,8, 21, 23, 29, 32 [DMS14] Claude Dellacherie, Servet Martinez, and Jaime San Martin, Inverse M-matrices and ultrametric matrices. Lecture Notes in Mathematics 2118. Springer, Cham, 2014. 50 [Dow80] Thomas Dowling, On the independent set numbers of a finite matroid. Combinatorics 79 (Proc. Colloq., Univ. Montral,´ Montreal, Que., 1979), Part I. Ann. Discrete Math. 8 (1980), 21–28. 49 [EH20] Christopher Eur and June Huh, Logarithmic concavity for morphisms of matroids. Adv. in Math. 367 (2020), 107094. 50 [FM92] Tomas´ Feder and Milena Mihail, Balanced matroids, Proceedings of the 24th Annual ACM Symposium on Theory of Computing, 26–38, ACM Press, 1992. 32 [FK72] Cornelius Fortuin and Pieter Kasteleyn, On the random-cluster model. I. Introduction and relation to other models. Physica 57 (1972), 536–564. 47 [Fuj05] Satoru Fujishige, Submodular functions and optimization. Second edition. Annals of Discrete Mathematics 58. Elsevier, Amsterdam, 2005.9 [Ful93] William Fulton, Introduction to toric varieties. Annals of Mathematics Studies 131. The William H. Roever Lec- tures in Geometry. Princeton University Press, Princeton, NJ, 1993. 46, 47 [Ful98] William Fulton, Intersection theory. Second edition. Ergebnisse der Mathematik und ihrer Grenzgebiete 3,A Series of Modern Surveys in Mathematics 2. Springer-Verlag, Berlin, 1998. 45, 46 [GKL] Pavel Galashin, Steven N. Karp, and Thomas Lam, The totally nonnegative Grassmannian is a ball. arXiv:1707.02010. 23 [GKL19] Pavel Galashin, Steven N. Karp, and Thomas Lam, The totally nonnegative part of G{P is a ball. Adv. in Math. 351 (2019), 614–620. 23 [Gre81] Jirˇ´ı Gregor, On quadratic Hurwitz forms. I. Apl. Mat. 26 (1981), no. 2, 142–153.8 [Gri06] Geoffrey Grimmett, The random-cluster model. Springer-Verlag, Berlin, 2006. 47 [GW04] Geoffrey Grimmett and Stephan Winkler, Negative association in uniform forests and connected graphs. Random Structures Algorithms 24 (2004), no. 4, 444–460.5, 56 [Gur09] Leonid Gurvits, On multivariate Newton-like inequalities. Advances in combinatorial mathematics, 61–78, Springer, Berlin, 2009.3, 23, 24 [HS89] Yahya Ould Hamidoune and Isabelle Salaun,¨ On the independence numbers of a matroid. J. Combin. Theory Ser. B 47 (1989), no. 2, 146–152. 49 [Har77] Robin Hartshorne, Algebraic geometry. Graduate Texts in Mathematics 52. Springer-Verlag, New York- Heidelberg, 1977. 46 [Hei38] Rudolf Heine, Der Wertvorrat der gemischten Inhalte von zwei, drei und vier ebenen Eibereichen. Math. Ann. 115 (1938), no. 1, 115–129. 47 [HH03]J urgen¨ Herzog and Takayuki Hibi, Discrete polymatroids. J. Algebraic Combin. 16 (2002), no. 3, 239–268.9 [Hol05] Olga Holtz, M-matrices satisfy Newton’s inequalities. Proc. Amer. Math. Soc. 133 (2005), no. 3, 711–717.5, 51 [HJ94] Roger Horn and Charles Johnson, Topics in matrix analysis. Cambridge University Press, Cambridge, 1994. 51 58 PETTER BRAND¨ EN´ AND JUNE HUH

[Huh18] June Huh, Combinatorial applications of the Hodge–Riemann relations. Proceedings of the International Congress of Mathematicians 3 (2018), 3079-3098. 15 [HK12] June Huh and Eric Katz, Log-concavity of characteristic polynomials and the Bergman fan of matroids. Math. Ann. 354 (2012), 1103–1116. 49 [HSW] June Huh, Benjamin Schroter¨ and Botong Wang, Correlation bounds for fields and matroids. arXiv:1806.02675. 33, 49 [HW17] June Huh and Botong Wang, Enumeration of points, lines, planes, etc. Acta Math. 218 (2017), no. 2, 297–317.5, 32, 46 [HMMS] June Huh, Jacob P. Matherne, Karola Mesz´ aros,´ and Avery St. Dizier, Logarithmic concavity of Schur and related polynomials. arXiv:1906.09633. 23, 30 [Kah00] Jeff Kahn, A normal law for matchings. Combinatorica 20 (2000), no. 3, 339–391.5, 56 [KN10] Jeff Kahn and Michael Neiman, Negative correlation and log-concavity. Random Structures Algorithms 37 (2010), no. 3, 367–388. 18, 49 [KN11] Jeff Kahn and Michael Neiman, A strong log-concavity property for measures on Boolean algebras. J. Combin. Theory Ser. A 118 (2011), no. 6, 1749–1760. 49 [Kar68] Samuel Karlin, Total positivity. Vol. I. Stanford University Press, Stanford, 1968. 30 [KMT07] Yusuke Kobayashi, Kazuo Murota, and Ken’ichiro Tanaka, Operations on M-convex functions on jump systems. SIAM. J. Discrete Math. 21 (2007), no.1, 107–129. 10, 40 [Laz04] Robert Lazarsfeld, Positivity in algebraic geometry. I. Classical setting: line bundles and linear series. Ergebnisse der Mathematik und ihrer Grenzgebiete 3. A Series of Modern Surveys in Mathematics 48. Springer-Verlag, Berlin, 2004. 45, 46 [Len13] Matthias Lenz, The f-vector of a representable-matroid complex is log-concave. Adv. in Appl. Math. 51 (2013), no. 5, 543–545. 49 [LS81] Elliott Lieb and Alan Sokal, A general Lee–Yang theorem for one-component and multicomponent ferromagnets. Comm. Math. Phys. 80 (1981), no. 2, 153–179. 10 [Lig97] Thomas Liggett, Ultra log-concave sequences and negative dependence. J. Comb. Th. A 79 (1997), 315–325. 24 [Lig10] Thomas Liggett, Continuous time Markov processes. An introduction. Graduate Studies in Mathematics 113. Amer- ican Mathematical Society, Providence, RI, 2010. 31 [MS15] Diane Maclagan and Bernd Sturmfels, Introduction to tropical geometry. Graduate Studies in Mathematics 161. American Mathematical Society, Providence, RI, 2015. 37, 38, 39, 41, 42 [Mah85] Carolyn Mahoney, On the unimodality of the independent set numbers of a class of matroids. J. Combin. Theory Ser. B 39 (1985), no. 1, 77–85. 49 [Mar02] David Marker, Model theory: An introduction. Graduate Texts in Mathematics 217. Springer, New York, 2002. 35 [Mas72] John Mason, Matroids: unimodal conjectures and Motzkin’s theorem. Combinatorics (Proc. Conf. Combinatorial Math., Math. Inst., Oxford, 1972), 207–220, Inst. Math. Appl., Southend-on-Sea, 1972.5, 49 [Men69] Korandattil Venugopalan Menon, On the convolution of logarithmically concave sequences. Proc. Amer. Math. Soc. 23 (1969), 439–441. 30 [Mur03] Kazuo Murota, Discrete convex analysis. SIAM Monographs on Discrete Mathematics and Applications. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, 2003.2,4,9, 18, 33, 34 [Nui68] Wim Nuij, A note on hyperbolic polynomials. Math. Scand. 23 (1968), 69–72.7, 10, 13 [Oxl11] James Oxley, Matroid theory. Second edition. Oxford Graduate Texts in Mathematics 21. Oxford University Press, Oxford, 2011.9, 48 [Pem00] Robin Pemantle, Towards a theory of negative dependence. J. Math. Phys., 41, No. 3, (2000), 1371–1390. 54 [Pem12] Robin Pemantle, Hyperbolicity and stable polynomials in combinatorics and probability. Current developments in mathematics, 2011, 57–123, Int. Press, Somerville, MA, 2012.7 [Pos09] Alexander Postnikov, Permutohedra, associahedra, and beyond. Int. Math. Res. Not. IMRN 2009, no. 6, 1026–1106. 9 LORENTZIAN POLYNOMIALS 59

[Sch14] Rolf Schneider, Convex bodies: the Brunn–Minkowski theory. Second expanded edition. Encyclopedia of Mathe- matics and its Applications 151. Cambridge University Press, Cambridge, 2014. 43, 44 [Sch38] Isaac Jacob Schoenberg, Metric spaces and positive definite functions. Trans. Amer. Math. Soc. 44 (1938), 522–536. 38 [Ser10] Denis Serre, Matrices. Theory and applications. Graduate Texts in Mathematics 216. Springer, New York, 2010. 25 [Sey75] Paul Seymour, Matroids, hypergraphs, and the max-flow min-cut theorem. Thesis, University of Oxford, 1975. 49 [SW75] Paul Seymour and Dominic Welsh, Combinatorial applications of an inequality from statistical mechanics. Math. Proc. Cambridge Philos. Soc. 77 (1975), 485–495. 32, 49 [She60] Geoffrey Shephard, Inequalities between mixed volumes of convex sets. Mathematika 7 (1960), 125–138. 43, 47 [Shi12] Akiyoshi Shioura, Matroid rank functions and discrete concavity. Jpn. J. Ind. Appl. Math. 29 (2012), no. 3, 535–546. 34 [Sok05] Alan Sokal, The multivariate Tutte polynomial (alias Potts model) for graphs and matroids. Surveys in combinatorics 2005, 173–226, London Math. Soc. Lecture Note Ser. 327, Cambridge University Press, Cambridge, 2005. 47 [Spe05] David Speyer, Horn’s problem, Vinnikov curves and hives. Duke Math. J. 127 (2005), 395–428. 35, 36

[TV83] Aleksandr Timan and Igor Vestfrid, Any separable ultrametric space is isometrically embeddable in `2. Funktsional. Anal. i Prilozhen. 17 (1983), no. 1, 85–86. 38 [Wag05] David Wagner, Rank-three matroids are Rayleigh. Electron. J. Combin. 12 (2005), Note 8, 11 pp. 32 [Wag08] David Wagner, Negatively correlated random variables and Mason’s conjecture for independent sets in matroids. Ann. Comb. 12 (2008), no. 2, 211–239. 17, 49 [Wag11] David Wagner, Multivariate stable polynomials: theory and applications. Bull. Amer. Math. Soc. (N.S.) 48 (2011), no. 1, 53–84.7, 11 [Wel76] Dominic Welsh, Matroid theory. London Mathematical Society Monographs 8. Academic Press, London-New York, 1976.9 [Zha85] Cui Kui Zhao, A conjecture on matroids. Neimenggu Daxue Xuebao 16 (1985), no. 3, 321–326. 49

DEPARTMENT OF MATHEMATICS,KTH,ROYAL INSTITUTEOF TECHNOLOGY,STOCKHOLM,SWEDEN. Email address: [email protected]

INSTITUTEFOR ADVANCED STUDYAND PRINCETON UNIVERSITY,PRINCETON,NJ,USA.

KOREA INSTITUTEFOR ADVANCED STUDY,SEOUL,KOREA. Email address: [email protected]