Arxiv:1707.03332V3 [Math.CO] 7 Dec 2019 There Are Many Algorithms One Could Apply and Eﬃcient Shortcuts in Special Cases [Fuk04,FW05, GS93, MS91, Wei07]

LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS

BO LIN AND NGOC MAI TRAN

Abstract. Already for bivariate tropical polynomials, factorization is an NP- Complete problem. In this paper, we give an eﬃcient algorithm for factorization and rational factorization of a rich class of tropical polynomials in n variables. Special families of these polynomials have appeared in economics, discrete convex analysis, and combinatorics. Our theorems rely on an intrinsic characterization of regular mixed subdivisions of integral polytopes, and lead to many open problems of interest in discrete geometry. Keywords. Tropical polynomials, factorization, rational factorization, M-convex, L-convex, generalized permutohedra, Legendre transform, Minkowski sum 2010 MSC. 13P05, 14T05, 52B20

Consider the max-plus tropical algebra (R, , ⊕), a b = a+b, a⊕b = max(a, b). A unit f is a tropical polynomial inducing trivial regular subdivision of its Newton polytope Newt(f). For a set of lattice polytopes S in Rn, an S-unit f is a unit where Newt(f) is a translation of some polytope in S. We say that an n-variate polynomial f is S-factorizable if it equals a product of S-units. The set of S-factorizable polynomials N[S] is a monoid. Deciding if a given polynomial f is S-factorizable is an instance of the classic factorization problem in tropical geometry [MS15, SS09], which remains open except for univariate polynomials [Gri07, KR05]. By the Cayley trick [Stu94], this problem is equivalent to the following question on regular mixed subdivisions. Fix a set of lattice polytopes S in Zn, and let ∆ be a regular subdivision of a certain Newton polytope P . Is ∆ a mixed regular subdivision with respect to some sequence of polytopes in S? Computing and enumerating regular mixed subdivisions is a central problem polyhedral geometry [DLRS10], however, this problem too seems open. There are two layers of difficulties here. First, consider the special cases where ∆ is a trivial subdivision. Then ∆ is mixed if and only if P can be decomposed as the Minkowski sum of some sequence of polytopes in S. For n ≥ 2, Minkowski factorization of a polytope is an instance of subset sum, and thus is NP-Complete [GL01, Tiw08]. However, arXiv:1707.03332v3 [math.CO] 7 Dec 2019 there are many algorithms one could apply and efficient shortcuts in special cases [Fuk04,FW05, GS93, MS91, Wei07]. Now, consider the general case, when ∆ is not a trivial subdivision. Suppose that we have an oracle that can not only decide if P can be written as such a Minkowski sum, but also give us the decomposition when it is possible. A necessary condition for ∆ to be mixed is that each cell of ∆ must be factorizable as a Minkowski sum of faces of some sequence polytopes in S. We stress that this sequence could be non-unique, and it could differ from cell to cell. In this case, we say that ∆ is locally factorizable. Local factorization can be checked using the oracle. However, this is not sufficient to imply that ∆ is a regular mixed subdivision in general. Being a mixed subdivision is a global condition: it

1 2 BO LIN AND NGOC MAI TRAN requires that the cells must have a consistent mixed labels, in other words, the same sequence of polytopes and their facets must factorize all of the cells in ∆. This paper gives a large class of polytopes S such that if ∆ is locally factorizable, then it is a mixed subdivision to a unique sequence of polytopes in S, up to permu- tation. With this property, deciding if a given polynomial is S-factorizable reduces to a series of independent Minkowski decomposition problems. In particular, our result supplies a computational shortcut to verifying regular mixed subdivisions.

Theorem 1 (Local Factorization). Let S be a set of lattice polytopes in Rn. If S is a positive basis, then N[S] has unique and local factorization. In other words, if f is a tropical polynomial such that each cell in its regular subdivision ∆f is a Minkowski sum of some polytopes in S, then f admits a unique minimal factorization into a tropical product of S-unit polynomials.

The name ‘positive basis’ comes from the fact that S is associated with a basis of a certain vector space, with an extra orientability condition (cf. Definition 33). Positive bases are simple to construct and verify (cf. Section 5). As a result, Theorem 1 applies to a wide class of polynomials, with a number of interesting consequences. The second main result of our paper considers the problem of rational factorization. For n = 1, rational factorization of tropical polynomials have been studied in [Tsa12, KLT15], in connections to tropical meromorphic functions. Rational tropical polynomials form a rich class of functions that have appeared in a vari- ety of applications: unirational varieties [DF13], ultra discrete equations [KLT15], auction theory [BGK16] and topological data analysis [Ver16]. They are equivalent to deep neural networks with ReLU activation functions and integer weight coeffi- cients [ZNL18]. Unit tropical polynomials correspond to neural networks with linear decision functions, the simplest class. Rationally factorizable tropical polynomials is the subset of neural networks that can be expressed as the sum of independent linear decision functions. To compute the rational factorization of a tropical polynomial is to transform a neural network into a much simpler description. Knowing how and when we can do this is an important step towards understanding deep neural networks. Formally, say that f is rationally factorizable with respect to S, or S-rational, if f g ∈ N[S] for some g ∈ N[S0] for some set of lattice polytopes S0. Say that it is strong S-rational if f g ∈ N[S] for some g ∈ N[S]. The set of S-rationals E[S] and the set of strong S-rationals Z[S] are both monoids, and they are much richer than N[S]. In terms of cells of ∆f , the extension from factorization to rational factorization is the extension from Minkowski sums to signed Minkowski sums, a useful operation in combinatorial geometry [ABD10,GH99]. For f to be S-rational (resp. strongly S-rational), a necessary but not sufficient condition is that each cell of ∆f can be expressed a signed Minkowski sum where all of the positive terms (resp. all of the terms) lie in S. The following theorem states that if S is a positive basis, then this local condition is sufficient for rational factorization.

Theorem 2 (Local rational factorization). Suppose S is a positive basis. Then Z[S] has unique and local factorization.

The third main result of our paper gives a sufficient condition on S for E[S] to have unique and local factorization. We provide a large and important family of S LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 3 with this condition, and thereby generalizes a result in auction theory of Baldwin, Golberg and Klemperer [BGK16]. Theorem 3 (Local strong rational factorization). Suppose S is a positive basis. Then Z[S] = E[S] if and only if S is full. In this case, f ∈ Z[S] if and only if the edges of ∆f are parallel to integer multiples of edges (1-dimensional polytopes) in S, as vectors. Loosely speaking, being full means S is a maximal basis amongst all bases that have the same set of primitive edges (cf. Definition 24). One important example is the set SG which consists of all standard simplices indexed by cliques of a given graph G on n nodes (cf. Proposition 39). For the complete graph Kn, for instance, n SKn consists of the standard simplex in R and its faces. Tropical polynomials in N[SKn ] define tropical hyperplane arrangements. These play an important role in defining tropical polytopes and their generalizations [DS04, FR15, JL16], and have applications in economics and combinatorics [ABGJ15, AD09, BK15, CT16,

Jos16, Shi15]. In discrete convex analysis, the set E[SKn ] is the set of L-convex functions whose domain is all of Zn [Mur03]. Their Legendre transforms are M- convex functions with compact domains. These functions feature prominently in polymatroid theory and have many interesting properties and applications, see the monographs [Fuj05,Mur03] and references therein. Theorem 3 generalizes several statements known in the literature. The case for unit polynomials corresponds to the signed Minkowski decomposition of generalized permutohedra into standard simplices [PRW08, Pos09, FKPP14]. There are many papers devoted to their combinatorics and applications [ABD10, CL15, Dok11, MUWY16, MPS+09, PRW08,Pos09, POC13]. Baldwin, Golberg and Klem- perer [BGK16] showed that Z[SKn ] = E[SKn ] and gave an algorithm to produce a certificate of factorization. Since this case is of particular interest to auction theory, which may utilize different notations and languages, for ease of reference we restate their theorem in full here. Theorem 4 ( [BGK16]). Let f be a tropical polynomial. There exists g a product of linear polynomials such that f g is a product of linear polynomials if and only if the edges in ∆f are parallel to ei − ej for i, j ∈ {0, 1, . . . , n}, with the convention that e0 is the origin. Theorem 2 strengthens the above results in two directions. First, one has unique factorization. Second, only edges that appear in ∆f may appear in both the numerator and denominator of the rational factorization. Theorem 5. Let f be a tropical polynomial in n + 1 variables such that the edges of ∆f are parallel to ei − ej for i, j ∈ {0, 1, . . . , n}. Let G(f) be the graph on n + 1 nodes, where (i, j) ∈ G(f) whenever there exists an edge in ∆f parallel to ei − ej. Then f ∈ Z[SG(f)]. That is, there is a unique way to write f as a tropical rational function, the denominator and numerator are product of linear polynomials, such that no new edges are introduced. Our theorems are constructive. Given a set S of lattice polytopes in Zn, Algo- rithm 1 certifies if it is a positive basis, and if it is, for any f ∈ Z[S], Algorithm 4 produces the unique minimal g ∈ N[S] such that f g ∈ N[S], and Algorithm 3 produces the unique factorization of a polynomial in N[S]. Our algorithms have polynomial run time with respect to the number of polytopes in S, however, this 4 BO LIN AND NGOC MAI TRAN generally scales exponential in n. We implement these algorithms with the soft- wares Maple and polymake [GJ00]. Codes for the examples in this paper can be found at https://github.com/linbomath/TropPolyFactor. Organization. We collect background materials and discuss subtleties sur- rounding factorization in Section 1. We discuss factorization of units in Section 2, define positive bases and prove the main results in Section 3. Section 4 introduces two families of full positive bases and some examples. Section 5 and 6 show the various algorithms and their outputs on numerical examples. We conclude with open problems in Section 7. Notations. For a set of vectors B ∈ Rn, write NB for their span over N, ZB for their span over Z. Say that B0 is a basis of B if ZB = ZB0, and the vectors in B0 are linearly independent over Z. For polytopes P,Q ⊂ Rn, c ∈ N, let P + Q denote their Minkowski sum, c·P denote the dilation of P . Say that P is equivalent to Q, written P ≡ Q, if P = v + Q for some v ∈ Zn. If there exists a polytope S ⊂ Rn such that Q + S = P , say that Q is a Minkowski summand of P , and write Q ≤ P . Let N (P ) denote the normal fan of a polytope P . A face of P is either P itself, or any polytope obtained as the set of maximizers of some linear functional over P . A proper face of P is a face of P that is neither P nor one of its vertices. We denote the face of P supported by a vector v by facev(P ). For vectors v, w ∈ Rn, write v ·w for their inner product. For a matrix H ∈ Zr×n, write n r n ImZ(H) := {Hv : v ∈ Z } ⊂ Z for the image of Z under H.

1. Background 1.1. Background on tropical polynomials. A tropical polynomial in n variables n is a piecewise linear, convex function f : R → R such that there exists ca ∈ R, a ∈ A ⊂ Rn where n ! M a X n (1) f(x) = ca x = max ca + aixi for all x ∈ R . a∈A a∈A i=1 The convex hull of A is called the Newton polytope of f, denoted as Newt(f). Points a ∈ A are said to be lifted by the height function a 7→ ca. The Legendre transform of f is the function f ∗ : Rn → R ∪ {+∞}, given by n ! ∗ X n f (y) = sup yixi − f(x) for all y ∈ R . n x∈R i=1 The Legendre transform f ∗ has a particularly simple interpretation: f ∗(y) 6= +∞ if and only if y is in Newt(f), and on this set, the graph of f ∗ equals the lower convex hull of the points {(a, −ca): a ∈ A}. The projection of this graph onto Newt(f) is called the regular subdivision of Newt(f) induced by f, denoted as ∆f . A regular subdivision ∆f is called mixed with respect to a sequence of polytopes Pr (F1,F2,...,Fr) if each cell in ∆f equals to a Minkowski sum i=1 Bi, where Bi is a face of Fi for each i, and such that this representation intersects properly as Pr 0 Pr 0 0 a sum, meaning that if σ = i=1 Bi, and σ = i=1 Bi for faces Bi,Bi of Fi, 0 then the intersection of Bi and Bi is a face of both, for each i ∈ {1, . . . , r}. The polyhedral version of the Cayley trick [Stu94] can be restated in the language of tropical factorization as follows. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 5

Theorem 6 (Cayley trick). Let S be a set of polytopes. Then f ∈ N[S] if and only if ∆f is a regular mixed subdivision of Newt(f) with respect to a sequence of possibly repeated polytopes in S. The tropical hypersurface T (f) defined by f is the set of points in Rn where the graph of f are not differentiable. A tropical hypersurface defines a balanced, weighted polyhedral complex, pure of dimension n − 1 in Rn, and a converse of this statement also holds, see [MS15, Proposition 3.3.10]. A straight-forward definition chase from this result gives Corollary 8, a characterization of when a unit can be factorized off a given tropical polynomial. 0 0 Definition 7. Let σ, σ be maximal cells in ∆f . Say that σ is a neighbor of σ in 0 0 direction v if facev(σ) is a maximal face of σ, and facev(σ) = face−v(σ ). Say that σ is in direction v from σ if there exists a sequence of cells σ1, σ2, . . . , σk in ∆f where 0 σ1 = σ, σk = σ , and σi+1 is a neighbor of σi in direction v, for i = 1, . . . , k − 1. Corollary 8. Let h be a tropical polynomial. Then T (h) = T (f) ∪ T (h0) for some unit f with Newt(f) = F and some tropical polynomial h0 if and only if there exists a cell σ ∈ T (h) where F ≤ σ, and for each maximal face facev(σ) of σ, all cells of ∆h in direction v from σ has facev(F ) as a Minkowski summand. 1.2. What counts as factorization. There are at least three natural notions of ‘equality’ for two tropical polynomials f and g in n variables x1, . . . , xn. n (1) As algebraic polynomials: f =1 g if and only if ca(f) = ca(g) for all a ∈ Z . n (2) As functions: f =2 g if and only if f(x) = g(x) for all x ∈ R . (3) As balanced polyhedral complexes: f =3 g if and only if T (f) = T (g) as sets and as balanced weighted polyhedral complexes.

One can check that f =1 g implies f =2 g, and f =2 g implies f =3 g, but the converses are not true. Often equality as functions is taken to be the deﬁnition of equality in factorization problems [MS15, SS09]. As with classical factorization, it is more natural to consider this equality up to multiplication by constants and monomials. On the surface this seems to be a fourth notion of equality. However, we show that this is exactly =3, and this is the notion of equality for tropical polynomials used throughout this paper. v n Lemma 9. We have f =3 g ⇔ f =2 a x g for some v ∈ Z , a ∈ R. v Proof. Suppose f =2 a x g. Then T (f) = T (g). Conversely, suppose T (f) = T (g). The weighted polyhedral complex T (f) uniquely determines ∆f up n to a translation, thus, ∆f = ∆g + v for some v ∈ Z . Furthermore, a face σ lifted in the graph of f ∗ is supported by the same set of vectors as the face σ + v lifted in the graph of g∗. Thus f ∗(x) = g∗(x) + a for some a ∈ R. Since the function f is uniquely determined by its Legendre transform, rewriting in polynomial terms v gives f =2 a x g. Remark 10. In several papers [GM07, Izh08, Tsa12] one associates a function f ¯ ¯ with the unique polynomial f where f =2 f and all lattice points in ∆f¯ are lifted, including interior points. This gives the stronger equivalence between =3 and =1 ¯ v f =3 g ⇔ f =1 a x g.¯ We do not take this approach here, as tropical multiplication does not commute with taking this canonical element. That is, for general polynomials f and g, ¯ ¯ (2) f g¯ 6=1 f g =1 f g.¯ 6 BO LIN AND NGOC MAI TRAN

¯ Characterizing when f g¯ =1 f g is the problem of finding competitive equilibrium in the product-mix auctions pioneered in [BK15]. For connections to integer programming and toric geometry, see [TY15]. 1.3. Signed Minkowski sums as vector additions. Definition 11. For non-empty polytopes P,Q ⊂ Rn, if there exists another nonempty polytope R ⊂ Rn such that P = Q + R, then the signed Minkowski sum P − Q is defined as R. According to Lemma 12, such R must be unique, so it is well-defined. Lemma 12. For non-empty polytopes P, Q, R ⊂ Rn, if P = Q + R, then R = {x ∈ Rn | x + Q ⊂ P }. Proof. For any x ∈ R and q ∈ Q, we have x + q ∈ P , hence x + Q ⊂ P . Conversely, suppose there exists a point y ∈ Rn such that y + Q ⊂ P while y∈ / R. By Farkas Lemma, there exists a hyperplane such that y and R are separated by it. In other words, there exists a linear function l defined on Rn such that l(y) > 0 while l(r) < 0 for all r ∈ R. Since Q is closed, l attains its maximum on Q at some point q1. Then l(y +q1) = l(y)+l(q1) > l(q1). Note that y +q1 ∈ P = Q+R, so there exists r ∈ R and q2 ∈ Q such that y+q1 = r+q2. But l(r+q2) = l(r)+l(q2) < 0+l(q1) = l(q1), a contradiction! Hence such point y does not exist and Lemma 12 is proved. In general, P + (−Q) 6= P − Q. For instance, if P = Q and P is a symmetric polytope around the origin, so P = −P , then P + (−P ) = 2 · P . In contrast, P − P = {0}. Lemma 13 (Signed Minkowski sum operations). Let P, Q, R, S ⊂ Rn be non-empty lattice polytopes. (i) If P − Q is well-defined, then (P − Q) + Q = P . (ii) If P − Q and (P − Q) − R are well-defined, then so is P − (Q + R) and it equals to (P − Q) − R. (iii) If P − Q is well-defined, then so is (P + R) − Q and it equals to (P − Q) + R. (iv) If both P − Q and R − S are well-defined, then so is (P + R) − (Q + S) and it equals to (P − Q) + (R − S). (v) If P − Q is well-defined, then it is a convex lattice polytope. Proof. Statements (i), (ii), (iii) and (iv) follow directly from Definition 11. State- ments (v) can be found in [Pos09, Lemma 11.1]. For statement (v), P − Q being lattice appears in [Pos09, Lemma 11.1]. As for being a lattice polytope, let S = P − Q. Let V (S),V (P ),V (Q) be the set of vertices of S, P and Q, respectively. Note that V (P ) ⊆ V (Q) + V (S). Take a vertex s ∈ V (S). If there exists a vertex q ∈ V (Q) with q + s ∈ V (P ), then s ∈ Zn. If there is no such vertex q ∈ v(Q), then s + Q ⊂ P \V (P ). Since s + Q is a closed polytope, there exists a direction w ∈ Rn and a small > 0 such that s+[−, ]·w +Q ⊂ P . So s cannot be n a vertex of S, a contradiction. Therefore, all vertices of S are in Z , as claimed. n Definition 14. Let P1,P2, ··· ,Pm be non-empty polytopes in R , c1, c2, ··· , cm ∈ Z with at least one being positive. If there exists a polytope P 0 such that

X 0 X (3) (−ci)Pi + P = ciPi,

ci<0 ci>0 Pm 0 then the signed Minkowski sum i=1 ciPi is deﬁned to be P . Throughout this Pm work, when we write i=1 ciPi, we mean the signed Minkowski sum. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 7

Remark 15. (1) By Lemma 13(i)-(iv), a signed Minkowski sum is independent of the order of its summands. (2) For general polytopes P,Q, P − Q is defined as {x ∈ Rn | x + Q ⊂ P } [Pos09]. However, for this definition P − Q could be empty and Lemma 13(iii) may not hold. As a result, we follow the authors of [ABD10] in order to define signed Minkowski sums.

Deﬁnition 16. For matrix H ∈ Zr×n whose rows are primitive vectors, and a r vector b ∈ R , let PH,b denote the possibly empty polytope given by n PH,b = {x ∈ R | Hx ≤ b}. T Suppose H = h1 h2 ··· hr , where hi is the i-th row vector of H. For any polytope P , let T v(H,P ) = maxx∈P h1 · x maxx∈P h2 · x ··· maxx∈P hr · x . r And let b(H) = {b ∈ R | PH,b 6= ∅ and v(H,PH,b) = b}. Remark 17. A polytope P could be obtained by diﬀerent pairs of (H, b). If an H is given, v(H,P ) is the smallest vector b such that P = PH,b. Lemma 18. Let H ∈ Zr×n be a matrix whose rows are primitive vectors, and P ⊆ Rn be a lattice polytope such that for every facet F of P , F is contained in a hyperplane {x ∈ Rn | h · x = c} where c ∈ R and h is a row vector of H. Then PH,v(H,P ) = P and v(H,P ) ∈ b(H).

Proof. For 1 ≤ i ≤ r let hi be the i-th row of H. For any x ∈ P , we have hi·x ≤ maxy∈P hi · y = v(H,P )i, hence H·x ≤ v(H,P ), by deﬁnition x ∈ PH,v(H,P ). Therefore P ⊆ PH,v(H,P ). Suppose there exists a point z ∈ PH,v(H,P ) − P . Then z and P are separated by at least one facet F of P . So there exists 1 ≤ i ≤ r and c ∈ R such that hi · x ≤ c for all x ∈ P but hi · z > c. Then c ≥ maxx∈P hi · x = v(H,P )i. However, since z ∈ PH,v(H,P ), we have hi · z ≤ v(H,P )i ≤ c, a contradiction! So PH,v(H,P ) = P . By deﬁnition of b(H), we have v(H,P ) ∈ b(H). The following lemma states that when H is appropriately chosen, then Minkowski addition of polytopes is equivalent to vector addition. Lemma 19. Let H be a matrix whose row vectors are all distinct primitive normal vectors of a polytope P and b = v(H,P ) ∈ b(H). If b1, . . . , bm ∈ b(H) such b = Pm i=1 bi, then m X PH,b = PH,bi . i=1 In this case, for any constants c1, . . . , cm > 0, m X P Pm = c P . H, i=1 cibi i H,bi i=1 Pm Pm Proof. By [Sch93, Theorem 1.7.5], for any x ∈ i=1 PH,bi we have Hx ≤ i=1 bi = Pm b, so x ∈ PH,b. Hence i=1 PH,bi ⊆ PH,b. Conversely, suppose there exists a Pm Pm point y ∈ PH,b − i=1 PH,bi . Then y and i=1 PH,bi are separated by some hyperplane. In other words, their exists a vector l ∈ Rn such that l · y > l · x for all Pm r x ∈ i=1 PH,bi . Now consider the following optimization problem: given a ∈ R , for x ∈ Rn, maximize l · x subject to Hx ≤ a. Since all constraints are linear, the 8 BO LIN AND NGOC MAI TRAN solution could be obtain by Fourier-Motzkin elimination and the solution must be Pn in the form j=1 wjaj, where wj’s are positive constants only depending on H and l. Then we have

n X l · y ≤ wjbj. j=1 Pn Now for each 1 ≤ i ≤ r, choose xi ∈ PH,bi such that l · xi = j=1 wj(bi)j. Then Pr Pm x = i=1 ∈ i=1 PH,bi and

r r  n  n r ! n X X X X X X l · x = l · xi =  wj(bi)j = wj (bi)j = wjbj, i=1 i=1 j=1 j=1 i=1 j=1 Pm a contradiction! So i=1 PH,bi = PH,b. The second claim follows from replacing P with P Pm and b with c b . H, i=1 cibi i i i

Remark 20. Lemma 19 is no longer true if some bi does not belong to b(H). Here is an example. Let n = 2, r = 4 and P be the convex hull of (0, 0), (6, 0), (1, 5), (0, 5), which is a right-angled trapezoid. Then

−1 0   0 −1 T H =   , b = (0, 0, 5, 6) .  0 1  1 1 Now let T T b1 = (0, 0, 2, 4) , b2 = (0, 0, 3, 2) .

Then b = b1+b2 and PH,b1 , PH,b2 are the convex hulls of {(0, 0), (4, 0), (2, 2), (0, 2)} T and {(0, 0), (2, 0), (0, 2)}. Here note that v(H,PH,b2 ) = (0, 0, 2, 2) , so b2 ∈/ b(H). While

PH,b1 + PH,b2 = conv (0, 0), (6, 0), (2, 4), (0, 4) 6= P. Proposition 21 below generalizes this result to signed Minkowski addition. This proposition serves two purposes. First, it is a vehicle to prove unique factorization. Second, it gives an algorithm to decompose a polytope into a signed Minkowski sum with respect to some given set of polytopes.

n Proposition 21. For i = 1, . . . , m, let PH,bi ⊂ Z be non-empty lattice polytopes where H ∈ Zr×n is the set of primitive normal vectors of their Minkowski Pm + − m sum P := i=1 PH,bi . For y , y ∈ N , suppose the signed Minkowski sum Pm + − i=1 (yi − yi )PH,bi is well-deﬁned. Then m X + − (yi − yi )PH,bi = PH,b i=1 where m X + − r (4) b = (yi − yi )bi ∈ Z . i=1 LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 9

Pm + − n Proof. Suppose i=1(yi − yi )PH,bi = P , where P ⊂ R is a nonempty polytope. By Deﬁnition 14,

X + − X − + (5) (yi − yi )PH,bi = (yi − yi )PH,bi + P. + − + − yi >yi yi

Since P ≤ P + , all primitive normal vectors of P belong to H, hence there H,bi exists a vector b ∈ b(H) such that P = PH,b. By Lemma 19, the LHS of (5) is P + , H,bi + + − − + P + − P + − − where b = (y − y )bi. In addition, (y − y )PH,bi = P , i yi >yi i i yi

m + − X + − X − + X + − r b = bi − bi = (yi − yi )bi − (yi − yi )bi = (yi − yi )bi ∈ Z . + − + − i=1 yi >yi yi

2. Unit Polynomials and Bases We now characterize and give conditions for unique factorizations for the set of units in N[S], Z[S] and E[S]. Recall that f is a unit if ∆f is the trivial subdivision of its Newton polytope Newt(f). By Corollary 8, (rational) factorization of a unit is equivalent to (signed) Minkowski decomposition of its Newton polytope. Proposition 21 converts this problem to vector addition. In particular, unique factorization is possible if and only if the set of initial vectors forms a basis over N and Z for factorization and rational factorization, respectively. Throughout this section let S be a finite set of lattice polytopes in Rn. Let H(S) ∈ Zr×n be a matrix whose row vectors are all distinct primitive normal P vectors of the polytope S∈S S, with coordinate-wise lexicographic order. Then H(S) is uniquely defined. Define r B(S) = {b ∈ Z ∩ b(H(S)) | PH(S),b ∈ S}, r n B(S) = {b ∈ Z ∩ b(H(S)) | PH(S),b ⊂ Z is a non-empty lattice polytope}. Recall that NB denote the free module over N, which is    X r  NB(S) = cb · b ∈ Z : cb ∈ N for all b ∈ B(S) , b∈B(S)  and ZB(S) is defined analogously. r Remark 22. Note that even if b ∈ Z ∩b(H(S)), PH(S),b could still not be a lattice polytope. Here is an example. Let S = {conv{(0, 0), (1, 2)}, conv{(0, 0), (−1, 2)}}. Then X S = conv{(0, 0), (1, 2), (0, 4), (−1, 2)} S∈S 10 BO LIN AND NGOC MAI TRAN and −2 −1 −2 1  H(S) =   .  2 −1 2 1

Now b = (−1, 2, 0, 3)T ∈ Z4 ∩ b(H(S)), but

PH(S),b = conv{(0.25, 0.5), (0.75, 1.5), (0.25, 2.5), (−0.25, 1.5)} is not a lattice polytope.

Proposition 23. Let f be a unit tropical polynomial in Rn, P its Newton polytope. Then

(i) f ∈ E[S] if and only if P = PH(S),b for some b ∈ B(S) (ii) f ∈ N[S] if and only if P = PH(S),b for some b ∈ NB(S). (iii) f ∈ Z[S] if and only if P = PH(S),b for some b ∈ B(S) ∩ ZB(S).

Proof. (i) By Corollary 8, f ∈ E[S] if and only if there exist lattice polytopes Q, R such that P + Q = R and R is a Minkowski sum of polytopes in S. Suppose such Q and R exist. Let b = v(H(S),P ). By [Zie95, Proposition 7.12], the normal fan of R is a refinement of the normal fan of P . Hence H(S) and P satisfies the condition of Lemma 18, and thus P = PH(S),b and b ∈ b(H(S)). Hence b ∈ B(S). Conversely, suppose P = PH(S),b for some b ∈ B(S). Let S be the Minkowski sum of all polytopes in S. Since P = PH(S),b, every facet of P is parallel to some facet of S, hence the normal fan of S is a refinement of the normal fan of P . By [Grü67,Page 318-319], there exists λ > 0 such that P is a Minkowski summand of λS. Since any λ0 > λ works too, we may assume λ ∈ N. Hence there exists another convex polytope Q such that P + Q = λS. Since both λS and P are lattice polytopes, so is Q. In addition, λS is a Minkowski sum of polytopes in S. Therefore we can take R = λS. (ii) By Corollary 8, f ∈ N[S] if and only if P is a Minkowski sum of polytopes in Pk S. Suppose P = i=1 Pi where each Pi ∈ S. Since Pi ∈ S, all facets of Pi belong to hyperplanes cut out by rows of H(S). We let bi = v(H(S),Pi), by Lemma 18, Pk Pi = PH(S),bi and bi ∈ b(H(S)), thus bi ∈ B(S). Now let b = i=1 bi ∈ NB(S). Note that b = v(H(S),P ), by Lemma 18 again we have P = PH(S),b. Conversely, if P = P (H(S), b) for some b ∈ NB(S). Then there exists b1, ·, bk ∈ Pk B(S) such that b = i=1 bi. Since bi ∈ b(H(S)) for 1 ≤ i ≤ k, by Lemma 19, we Pk have P = i=1 PH(S),bi is a Minkowski sum of polytopes in S. (iii) By Corollary 8, f ∈ Z[S] if and only if there exist lattice polytopes Q, R such that P + Q = R and both Q and R are Minkowski sum of polytopes in S. Suppose such Q and R exist. Then we can write

k l X X Q = PH(S),bi ,R = PH(S),cj , i=1 j=1

where all bi, cj ∈ BS. Now R − Q = P is a well-deﬁned polytope, by Proposition Pl Pk 21, P = P (H(S), b), where b = j=1 cj − i=1 bi. So b ∈ ZB(S). In addition, since P is a non-empty lattice polytope, b ∈ B(S). LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 11

Conversely, suppose P = PH,b for some b ∈ B(S) ∩ ZB(S). Since b ∈ ZB(S), there exists k, l ∈ N and vectors bi, cj ∈ BS for 1 ≤ i ≤ k and 1 ≤ j ≤ l such that Pl Pk Pl Pk b = j=1 cj − i=1 bi. Let a = j=1 cj = b + i=1 bi. By Lemma 19, we have

l k X X PH(S),a = PH(S),cj = PH(S),b + PH(S),bi . j=1 i=1

Since bi, cj ∈ BS, by deﬁnition, PH(S),bi ,PH(S),cj ∈ S. Hence we can let Q = Pk Pl i=1 PH(S),bi and R = j=1 PH(S),cj .

Now we consider the problem of unique factorization. Note that both B and B are integral vectors in Rr for some finite r. Thus, their Z-modules ZB and ZB are each isomorphic to some subgroup of Zr, so each must be finitely generated over Z. Definition 24 (Basis, Full Basis). Say that S is a basis if B(S) is a basis over Z for ZB(S). Say that S is a full basis if B(S) is a basis over Z for B(S). Proposition 25 (Unique Factorization). Let S be a set of polytopes. Then N[S] has unique factorization if and only if S is a basis.

Proof. Since N[S] contains the set of units polynomials, by Proposition 23, S is a basis is a necessary condition for N[S] to have unique factorization. Now we prove suﬃciency. Suppose S is a basis. Let f ∈ N[S]. Suppose there are two factorizations of f. Multiply f by a constant and a monomial if necessary, one can assume 0 0 f = f1 · · · fr = f1 · · · fr0 0 0 0 for some units fi, fj with Newt(fi), Newt(fj) ∈ S, i ∈ {1, . . . , r}, j ∈ {1, . . . , r }. 0 Pr Pr 0 By the Cayley trick, i=1 Newt(fi) and j=1 Newt(fj) are both equal to the support of ∆f . Since S is a basis, the sequence (Newt(fi)) must equal the sequence 0 (Newt(fj)), counting multiplicity. By the bijection given in the Cayley trick, the factorization of f is uniquely determined by the sequence of polytopes to which ∆f is a regular mixed subdivision. So f has a unique factorization.

Example 26. For d = 2, let S = {S1,...,S10} be the ten lattice polytopes shown in Figure 1. Up to translation, this set S contains six primitive edges corresponding to the following vectors (6) S1 = {(0, 1), (1, 0), (1, 1), (1, −2), (−2, 1), (1, −1)}. The matrix H(S) has 12 row vectors, which are these six and their negatives (7) (1, 0), (0, 1), (1, 1), (1, −1), (1, 2), (2, 1). The set B(S) consists of all lattice polytopes whose primitive eges are in S1. If P ∈ B(S) has r edges, then one can list its consecutive edges, so that up to translation P can be represented as a sequence of pairs ((w1, s1),..., (wr, sr)), where |wi| is 1 the length of the i-th edge of P which is parallel to some si ∈ S . Conversely, any Pr P such sequence with i=1 wisi = (0, 0) and i∈I wisi 6= (0, 0) for all I ( [r] deﬁnes a polytope in B(S) up to translation. By simple geometric arguments, one can derive the H-representation of P from its edge sum sequence ((wi, si)), and thus prove that P ∈ B(S) if and only if P ∈ ZS. So S is a full basis. 12 BO LIN AND NGOC MAI TRAN

5 6 8 2 9 7 10 1 3 4 Figure 1. A full basis S.

Example 27. Figure 2 shows another full basis S0 for the edge set S1 in (6). Any polygon in Z2 with edges parallel to those in S1 has a unique decomposition in ZS as well as ZS0. The decomposition with respect to one basis can be simpler. For example, let P be the second polytope from the left of S0 in Figure 1. It has a trivial decomposition in ZS0, while its decomposition in ZS is shown in Figure 3.

Figure 2. Another full basis S0.

Figure 3. Decomposition of a polytope in ZS with respect to two diﬀerent full bases S and S0.

3. Positive bases and local factorization Fix a finite set S of lattice polytopes in Rn. In this section, we define the key concept of positive bases and prove Theorems 1 and 2. Since unique local factorization is a very strong criterion, the definition of positive bases is somewhat intricate. Throughout this section, we shall analyze the set of lattice polytopes S in Example 26 as the running example. Definition 28. Say that S is hierarchical if S ∈ S implies all proper faces of S are in NS.

Example 29. Let S be the ten lattice polytopes in Figure 1. The edges of S1 are integer multiples of S5,S7 and S8. Similarly, edges of S2 correspond to S5,S6,S10, those of S3 correspond to S5,S6,S9, and those of S4 correspond to S5,S6,S8. So S is hierarchical. Note that S is still hierarchical if any polytopes from S1 to S4 were removed, although the remaining set of polytopes would no longer be a basis.

Lemma 30. Suppose N[S] has unique and local factorization. Then S is a hierarchical basis. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 13

Proof. By Proposition 25, S must be a basis. For each S ∈ S, let f be a unit with Newt(f) = S. Then trivially f ∈ N[S]. Since N[S] has local factorization, each lower-dimensional cell of ∆f is in N[S], therefore proper faces of S must be in NS. So S is hierarchical. Positive bases are hierarchical bases with an extra orientability condition. To define this, first we need the observation that the rows of the H-matrix of a hierarchical basis S come in pairs. Lemma 31. Suppose S is hierarchical. If v is a row vector of the H-matrix of S, then −v is also. Proof. For each S ∈ S, each edge e of S is a proper face of dimension 1. Thus e ∈ NS, which means e is a positive integer multiple of some edge w ∈ S as vector. Let Σ1 be the sum of all such shortest edges w ∈ S. The H-matrix of S is that of the P 1 polytope S∈S S, whose normal fan equals to the normal fan of Σ , which equals a hyperplane arrangement. So its H-matrix has the form stated in the lemma. Definition 32 (Orientation). Suppose S is hierarchical. Let H be its H-matrix. An orientation τ is a map from row vectors of H to {±1}, such that τ(v) = −τ(−v). τ τ Given an orientation τ, let H− = {v : τ(v) = −1} and H+ = {v : τ(v) = 1}. Say τ that S is positive with orientation τ if for each v ∈ H− and S ∈ S, facev(S) is either a vertex of S, or is S. Definition 33 (Positive basis). Say that S is a positive basis if it is a hierarchical basis and there exists some orientation τ such that it is positive with respect to. Example 34. Let S be the full basis from Example 26, Figure 1. Orient the vectors of H(S) as −, −, +, −, +, + in the order that they are listed in (7), and orient their negatives with opposite signs. The six edges of S are translations of the six edges listed in (6). This shows that S is a positive basis. The full basis S0 in Example 27 is positive with respect to orientation τ 0 which has signs −, −, +, +, −, −. Both S and S0 are thus positive bases associated to the same H-matrix.

The positive orientation restricts when two polytopes in NS can share a face. In particular, regular mixed subdivisions constructible from a sequence of polytopes in S must have a particular structure. The following gives an equivalent characterization in terms of pairs of polytopes in S, without reference to an orientation τ, and thus is easy to verify in speciﬁc examples. This result underpins Algorithm 1 for verifying whether a given set of lattice polytopes S is a positive basis. Proposition 35 (Characterization of positive basis). Suppose S is a hierarchical basis with H-matrix H. Then S is a positive basis if and only if for each row vector v of H, there are no polytopes S, S0 ∈ S, not necessarily distinct, such that both 0 0 facev(S) and face−v(S ) are proper faces of S and S of dimension at least one, respectively. Proof. Suppose S is a positive basis with orientation τ. For any pair of polytopes 0 S, S ∈ S, either facev(S) or face−v(S) must be a proper face. So the criterion holds. Conversely, suppose the criterion holds. This means for each tuple (v, −v, S), exactly one of the following cases hold

(i) facev(S) is a proper face and face−v(S) is a vertex (ii) face−v(S) is a proper face and facev(S) is a vertex 14 BO LIN AND NGOC MAI TRAN

(iii) face−v(S) and facev(S) are both equal to S or are both vertices of S. Define a local partial orientation ν :(v, S) 7→ {±1, 0} as follows.   +1 if (i) ν(v, S) = −1 if (ii)  0 else. Note that ν(−v, S) = −ν(v, S). By the hypothesis, for each fixed v, there are no two polytopes S, S0 ∈ S such that ν(v, S) = +1 and ν(v, S0) = −1. Thus, one can define a global partial orientation τ 0 : v 7→ {±1, 0} such that τ 0(v) = +1, τ 0(−v) = −1 whenever v supports a proper face of some S ∈ S, and τ 0(v) = τ 0(−v) = 0 if v and −v never support a proper face of S for all S ∈ S. Set τ : v 7→ {±1} by τ(v) = τ 0(v) if τ 0(v) 6= 0, otherwise choose τ(v) = +1, τ(−v) = −1 at random. τ Now take v ∈ H+, and S ∈ S. Only cases (i) and (iii) can happen. That is, face−v(S) is not a proper face of S. By definition, S is a positive basis with respect to orientation τ. Corollary 36. If S is a positive basis of ZS and S0 ⊆ S is hierarchical, then S0 is a positive basis of ZS0 ⊆ ZS. Proof of Theorem 1. Let S be a positive basis, and f be a tropical polynomial such that each cell of ∆f is a Minkowski sum of some polytopes in S. As S is a basis, by Proposition 23, this decomposition is unique. Let S(f) denote the sequence of polytopes in S that appear as Minkowski summands of the maximal cells of ∆f , with multiplicity. Note that |S(f)| is finite. We shall do induction on |S(f)|. If |S(f)| = 1, then ∆f is the trivial subdivision of a single polytope in S, so we are done. If |S(f)| > 1, pick F ∈ S(f) of maximal dimension. We shall use Corollary 8 to show that T (f) = T (F ) ∪ T (f 0) for some polynomial f 0. We then argue that 0 0 each cell of ∆f 0 is still in NS, and S(f ) ⊂ S(f), so |S(f )| < |S(f)|. This would complete the induction step. Let us prove the first claim that the condition of Corollary 8— holds for ∆f . By the setup, there exists a cell σ ∈ ∆f with F ≤ σ. Since σ ∈ NS, maximal faces of σ are supported by vectors in H(S). Let v ∈ H(S) be such a vector. We need to show that all cells in direction v from σ has facev(F ) as a Minkowski summand (recall Definition 7). By induction on the number of neighbors, it suffices to show that the immediate neighbor of σ in direction v has this property. If σ does not have a neighbor in direction v, then we are done. 0 0 Otherwise, let σ be this neighbor, that is, facev(σ) = face−v(σ ). If facev(F ) is a 0 0 point, then this is trivial. If F ≤ σ , then trivially facev(F ) ≤ facev(σ ). Therefore, 0 we are left with the case that facev(F ) is not a vertex, and F 6≤ σ . Write 0 X X σ = ySS + yT T S∈V ⊥ T ∈V for unique yS, yT ∈ N, where ⊥ V = {S ∈ S(f) : facev(S) is a proper face of S, S 6= F } and

V = {T ∈ S(f) : facev(T ) = T = face−v(T ),T 6= F }. By Proposition 21, F ≤ σ implies

0 X X facev(F ) ≤ facev(σ) = face−v(σ ) = yS face−v(S) + yT T. S∈V ⊥ T ∈V LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 15

P 0 P We now argue that facev(F ) ≤ T ∈V yT T . If this holds, then facev(σ ) ≥ T yT T 0 so facev(F ) ≤ facev(σ ) as needed. Suppose for contradiction that this does not ⊥ hold. There is at least one S ∈ V such that face−v(S) is a proper face of S, and c · facev(F ) ≥ face−v(S) for some c ∈ N. If facev(F ) is also a proper face of F , then S cannot be a positive basis by Proposition 35. So we must have facev(F ) = F . As S is a hierarchical basis, F ∈ S and face−v(S) ∈ NS, c · F ≥ face−v(S) implies face−v(S) = F . But this means S has dimension strictly larger than that of F , which is not possible as F is maximal amongst those in S(f). So we obtain the desired contradiction. This proves the ﬁrst claim. For the second claim on |S(f 0)|, note that cells of ∆f 0 are either equivalent to some cells of ∆f , or they have the form 0 τ ≡ τ −ω for some τ ∈ ∆f and some face ω of F . Since S is hierarchical, τ, ω ∈ NS. 0 0 Since S is a basis, τ ∈ NS. So all cells of ∆f are in NS, and S(f ) ⊆ S(f). But F ∈ S(f) and F/∈ S(f 0), so S(f 0) ⊂ S(f), and thus |S(f 0)| < |S(f)|. This concludes the proof. Proof of Theorem 2. Suppose f ∈ Z[S]. Let g ∈ N[S] be such that h = g f ∈ N[S]. For each cell σf of ∆f , there exists cells σg of ∆g and σh of ∆h such that

σf + σg = σh.

By the Cayley trick, σg, σh ∈ NS, so σf ∈ ZS. Conversely, suppose all cells of ∆f are in ZS. Compute the signed Minkowski sum decomposition of each cell of ∆f with respect to S. Let S−(f) be the sequence of polytopes in S that appear with negative signs, with multiplicity. Similar to the proof of Theorem 1, we shall do an − − induction on |S (f)|. If S (f) = ∅ then all cells of ∆f are in NS. By Theorem 1, f ∈ N[S], so we are done. If not, for S a polytope of maximal dimension in S−(f), let σS be the cell of ∆f where −ySS appears in its signed Minkowsi decomposition ∗ for some yS > 0. Define a unit g(S) such that Newt(g(S)) = S, and that g is a ∗ ∗ 0 yS classical linear function such that g = f restricted to σS. Let f := f (g(S)) . Since S is hierarchical, S ∈ NS implies that its faces are in NS. So cells of ∆f 0 are in ZS, and S−(f 0) ( S−(f). As S−(f) is a finite sequence, by induction we are done. So f ∈ Z[S], which proves that Z[S] has local factorization. Note that our proof produces a polynomial g ∈ N[S] such that f g ∈ N[S]. For uniqueness of this g, it is sufficient to show that this g does not depend on the order amongst polytopes of maximal dimension in S−(f). Indeed, note that if S, S0 ∈ NS are two polytopes of the same dimension, and ω is a proper face of S, then ω 6= S0. Therefore, if S, S0 are two maximal dimensional polytopes in S−(f), S 6= S0, then S0 ∈ S−(f g(S)). So the g produced by the proof is unique. Furthermore, any otherg ˜ ∈ N[S] such that f g˜ ∈ N[S] must contain enough units to bring all cells of ∆f from ZS\NS to NS, and therefore must contain g in its factorization. So the g produced is the minimal denominator. Finally, let us prove the assertion on full positive basis. Suppose Z[S] = E[S]. In particular, ZS = ES, so S is a full basis by Definition 24. Conversely, suppose S is a full positive basis. For f ∈ E[S], let g be a product of units such that f g ∈ N[S]. Then edges in ∆f must be parallel to 1 integer multiples of primitive edges in ∆f g, which are contained in S . Therefore, each cell of ∆f is in ES. But S is a full basis, so each cell of ∆f is also in ZS. As S is a positive basis, Z[S] has local factorization so f ∈ Z[S]. Example 37 (Factorization into tropical plane curves of degree 2). Let S be the positive basis of Example 26. For concreteness, we fix an ordering on the rows of the H-matrix, so that it is the transpose of the following 3 × 14 matrix: 16 BO LIN AND NGOC MAI TRAN

1 −1 0 0 1 −1 1 −1 1 −1 2 −2 1 −1 (8) 0 0 1 −1 1 −1 −1 1 2 −2 1 −1 1 −1 . 0000000000001 −1

Let fq(x1, x2, x3) = max(2x1 + 2x2, x1 + 3x2 − 2, x1 + x2 + 2x3 − 3, 3x1 + x3 −

1, x1 + 2x2 + x3 − 4, 4x1 − 3). Its regular subdivision ∆fq consist of 3 maximal cells C1,C2,C3 shown in Figure 4. Note that these cells are in ZS, with signed Minkowski decomposition

C1 = P1 − P2 + P3 − P4 + P10 + (1, 0, 1),

C2 = −P1 + 2P4 + P7 + (1, 1, −2),

C3 = −P3 + 2P4 + P9 + (2, 0, −2).

By Theorem 2, fq admits a rational factorization hq = fq gq, where gq, hq ∈ N[S], that is, they are products of tropical quadratics in three variables. Indeed, Algorithm 4 outputs

gq(x) = max(2x1, 2x3 − 10/3, x2 + x3 − 2)

+ max(x1 + x3, 2x3 − 5/3, x2 + x3 − 1/3)

+ max(2x3, 2x2 − 1, x1 + x3 − 2) + max(2x1, 2x3 − 5, x1 + x2 − 2).

The product hq = fq gq has degree 12. Algorithm 3 shows it is factorizable as product of seven S-units. The decompositions of Newt(hq) are shown in Figure 4. One can readily check from the ﬁgure that gq is the minimal polynomial in N[S] such that fq gq ∈ N[S]. 5 h (x) = x − 3x + 2 max(2x − , x + x , x + x − 2) q 1 3 3 2 1 3 2 3 8 + max(2x − , x + x − 1, 2x ) + max(x + x − 2, 2x ) 3 3 1 3 2 2 3 1 1 10 1 + 2 max(2x , x + x − 2, x + x − ) + max(2x − , x + x − , 2x ). 3 1 3 2 3 2 3 3 1 2 3 1

Figure 4. An illustration of Theorem 2. Two ways to decompose

∆hq in Example 37: by writing hq as a product of units, or by writing hq = fq gq. In particular, this shows that fq is rationally factorizable. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 17

4. Two families of full positive bases

Lemma 38. Let S = {∆I : I ⊆ [n], |I| ≥ 2} be the set of the standard simplex in Rn and its proper faces. Then S is a full positive basis. P Proof. The H-matrix of S consists of vectors of the form vI = i∈I ei where I ⊆ [n] and their negatives. It is straight forward to verify that S satisﬁes the criterion of Proposition 35, so S is a positive basis. By [ABD10, Proposition 2.4], P has edges parallel to ei − ej if and only if P ∈ ZS. Thus S is full.

Proposition 39 (Graphical bases). Given a graph G on n nodes, let SG consist of simplices ∆I , where I ⊆ [n] runs over all cliques in G. Then SG is a full positive basis.

Proof of Proposition 39. By Lemma 38, SKn is a full positive basis. Now, G is a subgraph of Kn, so SG ⊆ SKn . Clearly SG is hierarchical, so by Corollary 36, SG is a positive basis. It remains to show that it is full. It is suﬃcient to show that if P is a lattice polytope whose edges are parallel to ei − ej for (i, j) ∈ e(G), then P ∈ ZSG. Suppose P is a lattice polytope with such edge directions. By Lemma 38, P ∈ ZS(Kn), so X P = yI (P )∆I ∈ ZS(Kn) I⊆[n] for unique constants yI (P ) ∈ Z, I ⊆ [n]. Let

I = {I ⊆ [n] : ∆I ∈/ SG, yI (P ) 6= 0}. If I = ∅, then we are done. Otherwise, for contradiction, consider two cases.

• There exists some I ∈ I such that yI (P ) > 0. Then there is some edge (i, j) ∈/ e(G) such that i, j ∈ I. But P must contain an edge parallel to ei − ej, a contradiction. • For all I ∈ I, yI (P ) < 0. Let 0 X P := P + (−yI (P ))∆I . I∈I 0 P 0 Then P = L∈S yL(P )∆L, so edges in P are parallel to ei − ej for G (i, j) ∈ e(G). On the other hand, since I 6= ∅, there exists some I ∈ I 0 0 such that ∆I ≤ P , so P must contain an edge parallel to ei − ej for some (i, j) ∈/ e(G), a contradiction.

Therefore, one must have P ∈ ZSG. So SG is a full basis. Next we show that there are full positive bases in Z2 starting from any given set of primitive edges S1. As there are many full positive bases for a given set S1, we deliberately present a non-constructive proof. In speciﬁc examples, it is not diﬃcult to construct a given full positive basis in Z2, see Example 37. Proposition 40. Let S1 be a set of primitive edges in Z2. There exists a full positive basis S of ES1. Proof. If S1 has cardinality one or two, take S = S1 and the result holds trivially. Now suppose S1 consists of at least three edges. The row vectors of H(S1) consists of primitive vectors in Z2 which are orthogonal to those in S1. Since S1 has at least three edges, one can choose an orientation τ such that span (Hτ ) = 2. Let R + R Pτ be the set of all full-dimensional lattice polygons whose outer normal vectors 18 BO LIN AND NGOC MAI TRAN

τ 0 τ 0 are nonnegative integer multiples of those in H+. Choose S ⊆ P so that S is a basis for ZPτ . Set S = S0 ∪ S1. Then S is a hierarchical basis. We claim that S 1 1 in fact generates ES . Indeed, let P be a polygon in ES . Let v−(P ) be the set of τ outer normal vectors of P that are positive integer multiples of vectors in H−. If τ v−(P ) = ∅, then P ∈ P , so P ∈ ZS. Otherwise, for each cv · (−v) ∈ v−(P ) with τ τ −v ∈ H−, there is a polygon Q(v) ∈ P such that cv · v is an outer normal vector of Q(v). Then

X 0 X (9) P + Q(v) = P + cv · e(v),

v∈v−(P ) v∈v−(P ) where e(v) ∈ S1 is the edge orthogonal to v, and P 0 is some polytope whose outer τ normal vectors are all in H+. Thus the RHS of (9) is in ZS, and each Q(v) is in 1 ZS, so P ∈ ZS. Thus S is a full positive basis of ES . Corollary 41 (Rational Factorization for Bivariates). Any bivariate tropical polynomial is rationally factorizable into a product of aﬃne monomials of the form a b (x, y) 7→ c0 ⊕ x y , a, b ∈ N, c0 ∈ R.

Proof of Corollary 41. Let e(f) be the set of primitive edges in ∆f , Σ be their Minkowski sum. By Proposition 40, there exists a full positive basis S such that S1 = e(f), so f ∈ Z[S] = E[S]. Note that by deﬁnition, E[S1] = E[S]. Thus f ∈ E[S1], that is, f is rationally factorizable into a product of aﬃne monomials 1 whose Newton polygon are integer multiples of the edges in S .

5. Algorithms In this section we discuss various algorithms for factorization, rational factorization and their implementations. Without loss generality, we assume that the tropical polynomials of interest are homogeneous. Given such a tropical polynomial f and a ﬁnite set of lattice polytopes S, we supply algorithms to do the following (1) Decide whether S is a positive basis (Algorithm 1). (2) Given S a positive basis, decide whether f ∈ Z[S] \ N[S], f ∈ N[S], or neither (Algorithm 2). (3) If f ∈ N[S], produce the unique factorization for f (Algorithm 3). (4) If f ∈ Z[S] \ N[S], produce a g ∈ N[S] such that f g ∈ N[S] (Algorithm 4). There are a few subroutines of polytopes used in our algorithms. All these subroutines can be done using the software Polymake [GJ00]. They include (all input P are in V-representation) • Input a polytope P , return its H-representation. In particular, its H-matrix H(P ). • Input a polytope P , return the set of its faces F aces(P ), each in V- representation. • Input a polynomial f, output its Newton polytope Newt(f). • Input a tropical polynomial f with a weight at each vertex of P , return the maximal polytopes in its regular subdivision with respect to the weights, denoted ∆f . • Input a polytope P and a nonzero vector v, return facev(P ). LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 19

Algorithm 1 is based on Proposition 35. As a by-product, it also computes the H-matrix of S, which is needed for subsequent algorithms. Computing H(S) is in fact the most intensive part, as this matrix could have exponentially many row vectors. The other algorithms are based on Proposition 23 and the proofs of Theorems 1 and 2. A crucial diﬀerence to the proof is that these algorithms use the stricter notion of =2 instead of =3 for equality of two tropical polynomials. This means we need to keep track of translations of the regular subdivisions and translations of the Legendre transform. A major part of the algorithms is spelling out the details of this step.

Algorithm 1 Is Positive Basis 1: function IsPositiveBasis(S) 2: Input: S, a ﬁnite set of lattice polytopes, each in V-representation 3: Output: True if S is a positive basis; False otherwise 4: for F in F aces(S) do 5: if not F ∈ S then P 6: S ← P ∈S P 7: H ← H(S) 8: if rank(H) 6= number of rows in H then return False 9: for row vector v in H do 10: f(v) ← ∅ 11: for P in S do 12: if facev(P ) ∈ F aces(P ) and facev(P ) 6= P then 13: f(v) ← f(v) ∪ {P } 14: for row vector v in H do 15: if f(v) ∩ f(−v) 6= ∅ then return False 16: return True

Algorithm 2 Membership test for Z[S] and N[S] 1: function Membership(S, f) 2: Input: A tropical polynomial f and a positive basis S 3: Output: f∈ / Z[S]; f ∈ Z[S] / N[S]; f ∈ N[S] with decomposition of all polytopes in ∆f as Minkowski sums of polytopes in S 4: H ← H(S) 5: status ← 1 6: for σ ∈ ∆f do 7: b(σ) ← v(H, σ) 8: if b(σ) ∈/ ZB(S) then return f∈ / Z[S] 9: else if b(σ) ∈/ NB(S) then 10: status ← 0 11: else P 12: Write b(σ) = P ∈S σP · P , where each σP ∈ N 13: if status = 0 then 14: return f ∈ Z[S] / N[S] 15: else 16: return σP for all σ ∈ ∆f and P ∈ S 20 BO LIN AND NGOC MAI TRAN

Algorithm 3 Factorization for N[S] 1: function FactorN(S, h) 2: Input: a positive basis S and a tropical polynomial h ∈ N[S] mi 3: Output: S-units hi with multiplicities mi such that h =2 ihi . 4: H ← H(S)

5: (σP )σ∈∆h,P ∈S ← Membership(S, h) 6: s ← ∅ 7: for vertex V ∈ σ do Pn 8: s ← s ∪ { i=1 Vi · σi = h(V )} 9: Solve s for the unique solution (σ1, . . . , σn) . uniqueness guaranteed by the fact that σ has maximal dimension) n Pn 10: lσ ← linear function R → R with lσ(x1, . . . , xn) = i=1 σi · xi 11: O ← ∅ . output of factors with multiplicity 12: for P ∈ S do 13: for σ ∈ ∆h do 14: aσ,P ← σP

15: while ∃aσ,P > 0 do 16: S ← an element of {P ∈ S | ∃σ ∈ ∆h such that aσ,P > 0} with maximal dimension 17: η ← a polytope such that aη,S > 0 18: Add the following S-unit to O:

x 7→ max (v · x + lη(v)) vectex v∈η

19: for σ ∈ ∆h do 20: JS,σ ← conv argmaxvectex v∈η (lη(v) − lσ(v)) 21: bS,σ ← v(H,JS,σ) P 22: Write bS,σ = P ∈S (JS,σ)P · P . bS,σ ∈ NB(S) 23: for P ∈ S do 24: aσ,P ← aσ,P − (JS,σ)P 25: Return O

Remark 42. When S = SKn or more generally SG, there are two major computation shortcuts for Algorithms 3 and 4. Firstly, as S is a full positive basis, it is easy to check if a polytope is a signed Minkowski sum of polytopes in S. Secondly, the H-matrix is highly symmetric, and in this case, bH (P ) for some polytope P can be computed by M¨obiusinversion [ABD10, Proposition 2.4].

6. Numerical Examples Example 43 (Rational polynomials from spanning trees). This example is adapted from the family of M-convex functions given in [Mur03, Example 6.27]. Let G be the edge-weighted graph on n = 5 edges shown below. 2 3.5 1 3

4 LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 21

Algorithm 4 Factorization for Z[S] 1: function FactorZ(S, f) 2: Input: a positive basis S and a tropical polynomial f ∈ Z[S] 3: Output: a tropical polynomial g ∈ N[S] such that f g ∈ N[S]. 4: H ← H(S) 5: m ← Membership(S, f) 6: if m = f ∈ Z[S] / N[S] then 7: for σ ∈ ∆f do 8: b(σ) ← v(H, σ) P 9: Write b(σ) = P ∈S σP · P , where each σP ∈ Z 10: L ← ∅ 11: for σ ∈ ∆f do 12: for P ∈ S do 13: if σP < 0 then 14: s ← ∅ 15: for vertex V ∈ σ do Pn 16: s ← s ∪ { i=1 Vi · σi = f(V )} 17: Solve s for the unique solution (σ1, . . . , σn) . uniqueness guaranteed by the fact that σ has maximal dimension) n 18: lσ ← linear function R → R with n X lσ(x1, . . . , xn) = σi · xi i=1 19: IsNew ← 1 20: for (l, mt) ∈ L do 21: if lσ =2 l then 22: IsNew ← 0 23: if −σP > mt then 24: Replace (l, mt) by (l, −σP ) in L 25: if IsNew = 1 then 26: Append (lσ, −σP ) to L 27: K g ← l mt (l,mt)∈L 28: else 29: g ← 0

Let Θ be the set of spanning trees of G. Deﬁne the following tropical polynomial n fG : R → R

   M K K (10) fG(x) = − wi xi.

T ∈Θ ei∈E(T ) ei∈E(T ) 22 BO LIN AND NGOC MAI TRAN

Explicitly, fG(x) is the maximum of the following

x1 + x2 + x3 − 6, x1 + x2 + x4 − 7, x1 + x3 + x4 − 8,

x2 + x3 + x4 − 9, x1 + x2 + x5 − 6.5, x1 + x3 + x5 − 7.5,

x2 + x4 + x5 − 9.5, x3 + x4 + x5 − 10.5.

We ﬁnd that fG is a unit, with

Newt(fG) = ∆{1,4} + ∆{2,3} + ∆{1,2,5} + ∆{1,3,5} + ∆{2,4,5} (11) + ∆{3,4,5} − ∆{1,2,3,5} − ∆{1,2,4,5} − ∆{1,3,4,5} − ∆{2,3,4,5} + ∆{1,2,3,4,5}.

Algorithm 4 outputs the following

gG(x) = max(x1, x2 − 1, x3 − 2, x5 − 2.5) + max(x1, x2 − 1, x4 − 3, x5 − 2.5)

+ max(x1, x3 − 2, x4 − 3, x5 − 2.5) + max(x2, x3 − 1, x4 − 2, x5 − 1.5).

One can check that fG gG is still a unit, and Newt(fG gG) is a Minkowski sum of the 7 simplices with positive coeﬃcients amongst those in (11). By Algorithm 3, we get its factorization as follows:

5 f g (x) = max(x , x − 1, x − 2, x − 3, x − ) G G 1 2 3 4 5 2 5 5 + max(x , x − 1, x − ) + max(x , x − 2, x − ) 1 2 5 2 1 3 5 2 3 + max(x , x − 2, x − ) + max(x , x − 1, x − 1/2) 2 4 5 2 3 4 5 + max(x1, x4 − 3) + max(x2, x3 − 1).

Next we present a non-unit tropical polynomial f ∈ Z[SK3 ]\N[SK3 ]. This example comes from a quadratic M-convex function in [Mur03, Example 2.10].

Example 44. Let f : R3 → R be a homogeneous quadric tropical polynomial that is the maximum of the following

3x1 − 18, 3x2 − 45, 3x3 − 54, 3x4 − 81, x1 + 2x2 − 34, x1 + 2x3 − 34,

x1 + 2x4 − 42, 2x1 + x2 − 25, 2x1 + x3 − 22, 2x1 + x4 − 21, x2 + 2x3 − 45,

x2 + 2x4 − 53, 2x2 + x3 − 42, 2x2 + x4 − 41, x3 + 2x4 − 54, 2x3 + x4 − 45,

x1 + x2 + x3 − 31, x1 + x2 + x4 − 30, x1 + x3 + x4 − 29, x2 + x3 + x4 − 40.

The Newton polytope of f is 3 times the standard simplex R4. Its regular subdivision consists of 14 maximal cells, which are all in ZS4. Their signed Minkowski LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 23 sum decompositions are

1 · ∆{1} + 1 · ∆{1,2} + 1 · ∆{2,3,4},

1 · ∆{1} + 1 · ∆{4} + 1 · ∆{1,2,3,4},

1 · ∆{1,2} + 1 · ∆{2,3} + 1 · ∆{2,4} + 1 · ∆{3,4} − 1 · ∆{2,3,4},

1 · ∆{3} + 1 · ∆{1,2} + 1 · ∆{2,3,4},

1 · ∆{3} + 1 · ∆{3,4} + 1 · ∆{1,2,3} + 1 · ∆{1,2,4} − 1 · ∆{1,2,3,4},

2 · ∆{3} + 1 · ∆{1,2,3,4},

2 · ∆{4} + 1 · ∆{1,2,3,4},

1 · ∆{3} + 1 · ∆{4} + 1 · ∆{1,2,3,4},

1 · ∆{4} + 1 · ∆{3,4} + 1 · ∆{1,2,3} + 1 · ∆{1,2,4} − 1 · ∆{1,2,3,4},

1 · ∆{4} + 1 · ∆{1,2} + 1 · ∆{2,3,4},

1 · ∆{2} + 1 · ∆{1,2} + 1 · ∆{2,3,4},

1 · ∆{1} + 1 · ∆{3} + 1 · ∆{1,2,3,4},

1 · ∆{1} + 1 · ∆{3,4} + 1 · ∆{1,2,3} + 1 · ∆{1,2,4} − 1 · ∆{1,2,3,4},

2 · ∆{1} + 1 · ∆{1,2,3,4}. There are four terms with negative coeﬃcients. Algorithm 4 outputs g a product of four units

g(x) = max(x2, x3 − 11, x4 − 10) + max(x1, x2 − 11, x3 − 15, x4 − 11)

+ max(x1, x2 − 11, x3 − 12, x4 − 25) + max(x1, x2 − 9, x3 − 8, x4 − 7). Algorithm 3 gives the following factorization

(f g)(x)

= max(x1, x2 − 9, x3 − 12, x4 − 7) + max(x1, x2 − 9, x3 − 8, x4 − 21)

+ max(x1, x2 − 11, x3 − 12, x4 − 39) + max(x1, x2 − 11, x3 − 16, x4 − 25)

+ max(x1, x2 − 11, x3 − 20, x4 − 11) + max(x1, x2 − 7, x3 − 4, x4 − 3)

+ max(x2 − 2, x3 − 1, x4).

7. Summary and open questions In this work, we showed that if a finite set of lattice polytopes S is a positive basis, then one has an efficient algorithm to decide if a given tropical polynomial is S-factorizable or strong S-rational. Furthermore, when S is a full positive basis, then one has an even better description of strong S-rationals. The tropical rational factorization solved in this paper is part of the recent efforts on generaliz- ing Minkowski sum algorithms to signed Minkowski sums [EFG16]. We close with a number of interesting open questions in polyhedral computations and tropical geometry. A major open problem in our paper is to find full positive bases. It is easy to construct and verify positive bases, or construct a full basis. However, to construct one that is simultaneously full and positive is more difficult. It is not clear whether a full positive basis always exists for any set of primitive edge directions. We conjecture this to be true. 24 BO LIN AND NGOC MAI TRAN

Conjecture 45 (The Full Positive Basis Conjecture). Let E be a set of primitive lattice edges in Rn. There is a full positive basis S such that S1 = E. If this conjecture holds, then for a given tropical polynomial f, let e(f) be the set of primitive edges parallel to those in ∆f . Let S be a full positive basis such that S1 = e(f). Then Theorem 2 says that f must be rationally factorizable, where both the numerator and denominator are S-units. In other words, the conjecture implies that the following is true. Conjecture 46 (Conjecture for Rational Factorization). Any tropical polynomial in any number of variables is rationally factorizable into a product of affine monomials of the form n M ai (x1, . . . , xn) 7→ ci ⊕ x , i=0 for some ai ∈ N It is easy to show that these two conjectures are true for n = 2 (cf. Proposition 40 and Corollary 41). However, we do not know if either of them hold for n ≥ 3. A sub- problem is the following the edge skeleton variant of the Minkowski reconstruction problem [GH99]. Open question 47. Characterize all lattice polytopes that can be constructed from a given set of edge directions. To the best of our knowledge, the solution for n = 3 is known [MM19] but not in higher dimensions. Another major question is to generalize beyond the positive basis condition. As noted in Lemma 30, a necessary condition for unique and local factorization is that S is a hierarchical basis. One could either ask for a tighter condition, that is, an if and only if characterization of unique and local factorization for N[S]. Alternatively, one could sacrifice uniqueness and local factorization in favor of computational efficiency. This leads to the following concrete problem. Open question 48. What other families of polytopes S where the membership problem for N[S] and Z[S] can be efficiently solved? The above question can be taken further: if tropical factorization into units is a special case of the general tropical factorization problem of Speyer and Sturmfels [SS09], what are other special cases where irreducibility of tropical polynomials with respect to a given class can efficiently be determined?

Acknowledgment The authors thank Bernd Sturmfels for suggesting the collaboration. They also thank Elizabeth Baldwin and Paul Klemperer for stimulating discussions and Kazuo Murota for enlightening them the intricate and interesting properties of the M- convex functions. Bo Lin is supported by the Max-Planck Institute of Mathematics in the Sciences (Leipzig, Germany) and the 2017 summer grant for doctoral students provided by the Graduate Division of University of California, Berkeley. Ngoc Mai Tran is supported by the Bonn Junior Fellowship of the Hausdorﬀ Center for Mathematics. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 25

References [ABD10] Federico Ardila, Carolina Benedetti, and Jeffrey Doker. Matroid polytopes and their volumes. Discrete Comput. Geom., 43(4):841–854, 2010. [ABGJ15] Xavier Allamigeon, Pascal Benchimol, StéphaneGaubert, and Michael Joswig. Trop- icalizing the simplex algorithm. SIAM Journal on Discrete Mathematics, 29(2):751– 795, 2015. [AD09] Federico Ardila and Mike Develin. Tropical hyperplane arrangements and oriented matroids. Mathematische Zeitschrift, 262(4):795–816, 2009. [BGK16] Elizabeth Baldwin, Paul Goldberg, and Paul Klemperer. Tropical intersections and equilibrium (day 2 slides). Hausdorff School on Tropical Geometry and Economics, 2016. [BK15] Elizabeth Baldwin and Paul Klemperer. Understanding preferences: de- mand types, and the existence of equilibrium with indivisibilities. Preprint http://www.nuff.ox.ac.uk/users/klemperer/demandtypes.pdf, 2015. [BML87] Roswitha Blind and Peter Mani-Levitska. Puzzles and polytope isomorphisms. Ae- quationes Mathematicae, 34(2-3):287–297, 1987. [CL15] Federico Castillo and Fu Liu. Ehrhart positivity for generalized permutohedra. In 27th Formal Power Series and Algebraic Combinatorics, 2015. [CT16] Robert Alexander Crowell and Ngoc Mai Tran. Tropical geometry and mechanism design. arXiv preprint arXiv:1606.04880, 2016. [DF13] Jan Draisma and Bart Frenk. Tropically unirational varieties. In Algebraic and combinatorial aspects of tropical geometry, volume 589 of Contemp. Math., pages 109–123. Amer. Math. Soc., Providence, RI, 2013. [DLRS10] JesúsA. De Loera, Jörg Rambau, and Francisco Santos. Triangulations, volume 25 of Algorithms and Computation in Mathematics. Springer-Verlag, Berlin, 2010. Struc- tures for algorithms and applications. [Dok11] Jeffrey Samuel Doker. Geometry of generalized permutohedra. PhD thesis, University of California, Berkeley, 2011. [DS04] Mike Develin and Bernd Sturmfels. Tropical convexity. Doc. Math, 9(1-27):7–8, 2004. [EFG16] Ioannis Z Emiris, Vissarion Fisikopoulos, and Bernd Gärtner.Efficient edge-skeleton computation for polytopes defined by oracles. Journal of symbolic computation, 73:139–152, 2016. [FKPP14] AndrásFrank, TamásKirály, JúliaPap, and David Pritchard. Characterizing and recognizing generalized polymatroids. Mathematical Programming, 146(1-2):245–273, 2014. [FR15] Alex Fink and Felipe Rincón.Stiefel tropical linear spaces. J. Combin. Theory Ser. A, 135:291–331, 2015. [Fuj05] Satoru Fujishige. Submodular functions and optimization, volume 58 of Annals of Discrete Mathematics. Elsevier B. V., Amsterdam, second edition, 2005. [Fuk04] Komei Fukuda. From the zonotope construction to the Minkowski addition of convex polytopes. Journal of Symbolic Computation, 38(4):1261–1272, 2004. [FW05] Komei Fukuda and Christophe Weibel. Computing all faces of the Minkowski sum of V-polytopes. In Proceedings of the 17th Canadian Conference on Computational geometry, number ROSO-CONF-2005-001, pages 253–256, 2005. [Gao01] Shuhong Gao. Absolute irreducibility of polynomials via Newton polytopes. J. Alge- bra, 237(2):501–520, 2001. [GH99] Peter Gritzmann and Alexander Hufnagel. On the algorithmic complexity of minkowski’s reconstruction theorem. Journal of the London Mathematical Society, 59(3):1081–1100, 1999. [GJ00] Ewgenij Gawrilow and Michael Joswig. polymake: a framework for analyzing convex polytopes. In Polytopes—combinatorics and computation (Oberwolfach, 1997), volume 29 of DMV Sem., pages 43–73. Birkhäuser,Basel, 2000. [GL01] Shuhong Gao and Alan G.B. Lauder. Decomposition of polytopes and polynomials. Discrete & Computational Geometry, 26(1):89–104, 2001. [GM07] Nathan Grigg and Nathan Manwaring. An elementary proof of the fundamental theorem of tropical algebra. arXiv preprint arXiv:0707.2591, 2007. 26 BO LIN AND NGOC MAI TRAN

[Gri07] Nathan Grigg. Factorization of Tropical Polynomials. PhD thesis, Brigham Young University, 2007. [Grü67] Branko Grünbaum. Convex polytopes. Pure and applied mathematics. Interscience, 1967. [GS93] Peter Gritzmann and Bernd Sturmfels. Minkowski addition of polytopes: computational complexity and applications to Gröbnerbases. SIAM Journal on Discrete Mathematics, 6(2):246–269, 1993. [HRS00] Birkett Huber, JörgRambau, and Francisco Santos. The Cayley trick, lifting subdivisions and the Bohne-Dress theorem on zonotopal tilings. Journal of the European Mathematical Society, 2(2):179–198, 2000. [Izh08] Zur Izhakian. Tropical algebraic sets, ideals and an algebraic Nullstellensatz. Inter- national Journal of Algebra and Computation, 18(06):1067–1098, 2008. [JKK02] Michael Joswig, Volker Kaibel, and Friederike Körner.On thek-systems of a simple polytope. Israel Journal of Mathematics, 129(1):109–117, 2002. [JL16] Michael Joswig and Georg Loho. Weighted digraphs and tropical cones. Linear Alge- bra and its Applications, 501:304–343, 2016. [Jos16] Michael Joswig. The Cayley trick for tropical hypersurfaces with a view toward Ri- cardian economics. arXiv preprint arXiv:1606.09165, 2016. [Kal88] Gil Kalai. A simple way to tell a simple polytope from its graph. Journal of combinatorial theory, Series A, 49(2):381–383, 1988. [KLT15] Risto Korhonen, Ilpo Laine, and Kazuya Tohge. Tropical value distribution theory and ultra-discrete equations. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2015. [KR05] Ki Hang Kim and Fred William Roush. Factorization of polynomials in one variable over the tropical semiring. arXiv preprint math/0501167, 2005. [MM19] Yves Martinez-Maure. Existence and uniqueness theorem for a 3-dimensional polytope of r3ˆ with prescribed directions and perimeters of the facets. 2019. [MPS+09] Jason Morton, Lior Pachter, Anne Shiu, Bernd Sturmfels, and Oliver Wienand. Convex rank tests and semigraphoids. SIAM Journal on Discrete Mathematics, 23(3):1117–1134, 2009. [MS91] David M. Mount and Ruth Silverman. Combinatorial and computational aspects of Minkowski decompositions. In Vision geometry (Hoboken, NJ, 1989), volume 119 of Contemp. Math., pages 107–124. Amer. Math. Soc., Providence, RI, 1991. [MS15] Diane Maclagan and Bernd Sturmfels. Introduction to tropical geometry, volume 161 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2015. [Mur03] Kazuo Murota. Discrete convex analysis. SIAM Monographs on Discrete Mathematics and Applications. Society for Industrial and Applied Mathematics (SIAM), Philadel- phia, PA, 2003. [MUWY16] Fatemeh Mohammadi, Caroline Uhler, Charles Wang, and Josephine Yu. Generalized permutohedra from probabilistic graphical models. arXiv preprint arXiv:1606.01814, 2016. [MV99] Tom Michiels and Jan Verschelde. Enumerating regular mixed-cell configurations. Discrete Comput. Geom., 21(4):569–579, 1999. [POC13] Alexander Postnikov, Suho Oh, and Dorian Croitoru. Poset vectors and generalized permutohedra. Discrete Mathematics & Theoretical Computer Science, 2013. [Pos09] Alexander Postnikov. Permutohedra, associahedra, and beyond. Int. Math. Res. Not. IMRN, (6):1026–1106, 2009. [PRW08] Alex Postnikov, Victor Reiner, and Lauren Williams. Faces of generalized permutohedra. Doc. Math, 13(51):207–273, 2008. [Sch93] Rolf Schneider. Convex Bodies: The Brunn-Minkowski Theory. Encyclopedia of Mathematics and its Applications. Cambridge University Press, 1993. [She63] Geoffrey C. Shephard. Decomposable convex polyhedra. Mathematika, 10:89–95, 1963. [Shi15] Yoshinori Shiozawa. International trade theory and exotic algebra. Evolutionary and Institutional Economics Review, 12(1):177–212, 2015. [SS09] David Speyer and Bernd Sturmfels. Tropical mathematics. Math. Mag., 82(3):163– 173, 2009. LINEAR AND RATIONAL FACTORIZATION OF TROPICAL POLYNOMIALS 27

[Stu94] Bernd Sturmfels. On the Newton polytope of the resultant. J. Algebraic Combin., 3(2):207–236, 1994. [Tiw08] Hans Raj Tiwary. On the hardness of computing intersection, union and Minkowski sum of polytopes. Discrete Comput. Geom., 40(3):469–479, 2008. [Tsa12] Yen-Lung Tsai. Working with tropical meromorphic functions of one variable. Tai- wanese Journal of Mathematics, 16(2):pp–691, 2012. [TY15] Ngoc Mai Tran and Josephine Yu. Product-mix auctions and tropical geometry. arXiv preprint arXiv:1505.05737, 2015. [Ver16] Sara Kalisnik Verovsek. Tropical coordinates on the space of persistence barcodes. arXiv preprint arXiv:1604.00113, 2016. [Wei07] Christophe Weibel. Minkowski sums of polytopes. PhD thesis, EPFL, 2007. [Zie95] G¨unter M Ziegler. Lectures on polytopes, volume 152 of Graduate Texts in Mathe- matics. Springer-Verlag, New York, 1995. [ZNL18] Liwen Zhang, Gregory Naitzat, and Lek-Heng Lim. Tropical geometry of deep neural networks. arXiv preprint arXiv:1805.07091, 2018.

Department of Mathematics, University of Texas at Austin, TX 78712, USA, and School of Mathematics, Georgia Institute of Technology, GA 30332, USA E-mail address: [email protected]

Department of Mathematics, University of Texas at Austin, TX 78712, USA, and the Hausdorff Center for Mathematics, Bonn 53115, Germany E-mail address: [email protected]