Log-concavity in combinatorics and geometry
Omid Amini
CNRS - CMLS, ´Ecole Polytechnique
March 11th 2021
1 Log-concave sequences in combinatorics
2 Log-concavity versus convexity
3 Algebro-geometric log-concavity
4 Applications and further results
Log-concave sequences
Definition
A sequence of non-negative numbers(ai)∞i=0 is calledlog-concaveif for any i≥1
a2i ≥ai−1ai+1. Example (Binomial coefficients)
The sequence n0 , n1
,. . ., n−1n , nn
is log-concave.
Example (Newton’s inequality) Let t1, . . . , tn≥0. The sequence
Sk:= 1
n k
X
i1<···<ik
ti1ti2. . . tik
is log-concave.
Log-concave sequences
Definition
A sequence of non-negative numbers(ai)∞i=0 is calledlog-concaveif for any i≥1
a2i ≥ai−1ai+1. Example (Binomial coefficients)
The sequence n0 , n1
,. . ., n−1n , nn
is log-concave.
Example (Newton’s inequality) Let t1, . . . , tn≥0. The sequence
Sk:= 1
n k
X
i1<···<ik
ti1ti2. . . tik
is log-concave.
Log-concave sequences
Definition
A sequence of non-negative numbers(ai)∞i=0 is calledlog-concaveif for any i≥1
a2i ≥ai−1ai+1. Example (Binomial coefficients)
The sequence n0 , n1
,. . ., n−1n , nn
is log-concave.
Example (Newton’s inequality) Let t1, . . . , tn≥0. The sequence
Sk:= 1
n k
X
i1<···<ik
ti1ti2. . . tik
is log-concave.
Matchings in graphs
G= (V, E) a graph
A matchingin Gis a collection of edges which do not share any extremity.
Defineai := number of matchings of sizeiinG.
Theorem (Gruber-Kuntz 1976) The sequence ai is log-concave.
Proof.
The matching polynomial p(X) =P
iaiXi is real rooted.
Matchings in graphs
G= (V, E) a graph
A matchingin Gis a collection of edges which do not share any extremity.
Defineai := number of matchings of sizeiinG.
Theorem (Gruber-Kuntz 1976) The sequence ai is log-concave.
Proof.
The matching polynomial p(X) =P
iaiXi is real rooted.
Graph colorings
G= (V, E) a graph. Define
fG :N→N
fG(k) = number of proper coloringsofG bykcolors.
Graph colorings
G= (V, E) a graph. Define
fG :N→N
fG(k) = number of proper coloringsofG bykcolors.
Graph colorings
G= (V, E) a graph. Define
fG :N→N
fG(k) = number of proper coloringsofG bykcolors.
Graph colorings
For the Petersen graph:
fG(2) = 0.
fG(3) = 120.
fG(k) =k(k−1)(k−2) k7−12k6+67k5−230k4+529k3−814k2+775k−352 .
Graph colorings
Proposition
fG is a polynomial in k.
Proof.
Define
ar:= number of proper colorings ofGbyexactly kcolors.
Then
fG(k) =
k
X
r=0
ar k
r
.
Graph colorings
fG(X) =a0Xn−a1Xn−1+a2Xn−2+· · ·+ (−1)nan. fG(k)6= 0 iff Gis k-colorable. ForGplanar, fG(4)6= 0.
a0= 1.
ai are all non-negative.
Theorem (Huh 2010)
The sequence a0, a1, . . . , an is log-concave.
Conjectured by Read-Welsh 1968.
Spanning forests
G= (V, E) connected graph.
A spanning forestis a subgraph (V, F) for F ⊆E such that (V, F)does not have any cycle.
Spanning forests
G= (V, E) connected graph.
Define
ai:= number of spanning forests withiedge.
Theorem (Adiprasito-Huh-Katz 2015) The sequence a0, . . . , an−1 is log-concave Conjectured by Mason 1976.
Matroids
Matroid: a combinatorial structure axiomatising the linear dependency between finite collection of vectors by getting rid of the ambiant vector space.
Introduced by Hassler Withney 1935 in his study of graphs.
Question
How to recover Gfrom its cycles ?
Today, we see this as Torelli theorem for graphs.
Matroids arise naturally in the study of singularities in algebraic geometry.
Matroids
Given a collection of vectorsv1, . . . , vnin a vector spaceW, we can define the collection I of independent sets I ⊆ {1, . . . , n}
(i.e. such that vectors vi for i∈I are linearly independent)
Matroids
Example
v1 = (1,0,0),v2 = (0,1,0),v3= (0,0,1),v4 = (2,1,0).
I = n
∅,
{1},{2},{3},{4},
{1,2},{1,3},{1,4},{2,3},{2,4},{3,4},
{1,2,3},{1,3,4},{2,3,4}o .
Matroids
Given a collection of vectorsv1, . . . , vnin a vector spaceW, we can define the collection I of independent sets I ⊂ {1, . . . , n}.
the collection B of bases.
Example
v1 = (1,0,0),v2 = (0,1,0),v3= (0,0,1),v4 = (2,1,0).
B= n
{1,2,3},{1,3,4},{2,3,4}o .
Matroids
Given a collection of vectorsv1, . . . , vnin a vector spaceW, we can define the collection I of independent sets I ⊂ {1, . . . , n}.
the collection B of bases.
rank function r: 2{1,...,n} −→N∪ {0}.
Example
v1 = (1,0,0),v2 = (0,1,0),v3= (0,0,1),v4 = (2,1,0).
r({1,2,3}) = 3 r({1,2,4}) = 2.
Matroids
Given a collection of vectorsv1, . . . , vnin a vector spaceW, we can define the collection I of independent sets I ⊂ {1, . . . , n}
the collection B of bases
rank function r: 2{1,...,n} −→N∪ {0}.
Each of these satisfies some basic properties. E.g.
B satisfies the exchange axiom.
r is submodular:
∀A, B, r(A) +r(B)≥r(A∪B) +r(A∩B).
Matroids
A matroid M on the ground set E is defined by any of the following:
the collection I of independent sets I ⊂ {1, . . . , n}
the collection B of bases
rank function r: 2{1,...,n} −→N∪ {0}.
B satisfies the exchange axiom.
r is submodular:
∀A, B, r(A) +r(B)≥r(A∪B) +r(A∩B).
Matroids
A matroid M on the ground set E is defined by any of the following:
the collection I of independent sets I ⊂ {1, . . . , n}
the collection B of bases
rank function r: 2{1,...,n} −→N∪ {0}.
B satisfies the exchange axiom.
r is submodular:
∀A, B, r(A) +r(B)≥r(A∪B) +r(A∩B).
Matroids
Example
Any connected graph G= (V, E)gives a matroid MG:
Independent sets are I ⊂E such that(V, I) is a spanning forest.
Bases are all B ⊂E such that(V, B) is a spanning tree.
For any field K, a collection of vectors v1, . . . , vn in the vector spaceKr define a matroid. These are called representable.
The majority of matroids are non-representable.
Theorem (Nelson 2016)
Almost any matroid is not representable on any field.
Characteristic polynomial of matroids
M matroid of rank r on ground set E={1, . . . , n}
Define the characteristic polynomialof M by fM(x) = X
A⊆E
(−1)|I|xr−r(I).
Write fM(x) =a0xr−a1xr−1+· · ·+ (−1)rar.
Theorem (Adiprasito-Huh-Katz 2015) The sequence a0, . . . , ar is log-concave.
Conjectured by Rota 71, Heron 72 and Welsh 76.
Characteristic polynomial of matroids
M matroid of rank r on ground set E={1, . . . , n}
Define the characteristic polynomialof M by fM(x) = X
A⊆E
(−1)|I|xr−r(I).
Write fM(x) =a0xr−a1xr−1+· · ·+ (−1)rar. Theorem (Adiprasito-Huh-Katz 2015)
The sequence a0, . . . , ar is log-concave.
Conjectured by Rota 71, Heron 72 and Welsh 76.
1 Log-concave sequences in combinatorics
2 Log-concavity versus convexity
3 Algebro-geometric log-concavity
4 Applications and further results
Convex functions
Definition
f :Rn→Rcontinuous isconvex if for allx, y∈Rn, f(x+y
2 )≤ f(x) +f(y)
2 .
Example
f(x) =x21+· · ·+x2n is convex.
More generally, for symmetricA∈Mn(R), functionf(x) =xtAxis convex if and only ifA is positive semi-definite.
Mixed-determinant
A1, . . . , An symmetric matrices in Mn(R).
det(t1A1+· · ·+tnAn) is a polynomial int1, . . . , tn of degreen.
Definition (Mixed-determinant)
Definemdet(A1, . . . , An)as the coefficient of t1t2. . . tn in det(t1A1+· · ·+tnAn).
Example A1=
1 2 2 5
,A2=
2 −1
−1 1
,t1A1+t2A2=
t1+ 2t2 2t1−t2 2t1−t2 5t1+t2
. det(t1A1+t2A2) =t1t2+ 10t1t2−2t1t2−2t1t2+. . . This gives mdet(A1, A2) = 7
Mixed-determinant
Example
IfA1=· · ·=An=A, then
det(t1A1+. . . tnAn) = (t1+· · ·+tn)ndet(A)
=n! det(A)t1t2. . . tn+. . . and so
mdet(A, . . . , A) =n! det(A).
Mixed-determinant
A, B two positive-definite matrices inMn(R). Define ai = det(A, . . . , A
| {z }
i times
, B, . . . , B
| {z }
(n−i) times
).
Theorem (Alexandrov 1938)
The sequence a0, a1, . . . , an is log-concave.
Example A=1 2
2 5
,B = 2 −1
−1 1
.
a0 = 2, a1 = 7, a2 = 2
Mixed-determinant
A, B two positive-definite matrices inMn(R). Define ai = det(A, . . . , A
| {z }
i times
, B, . . . , B
| {z }
(n−i) times
).
Theorem (Alexandrov 1938)
The sequence a0, a1, . . . , an is log-concave.
Example A=1 2
2 5
,B = 2 −1
−1 1
.
a0 = 2, a1 = 7, a2 = 2
Mixed-determinant
A, B two positive-definite matrices inMn(R). Define ai = det(A, . . . , A
| {z }
i times
, B, . . . , B
| {z }
(n−i) times
).
Theorem (Alexandrov 1938)
The sequence a0, a1, . . . , an is log-concave.
Example A=1 2
2 5
,B = 2 −1
−1 1
.
a0 = 2, a1 = 7, a2 = 2
Mixed-volume
P1, . . . , Pn convex bodies inRn.
vol(t1P1+· · ·+tnPn) is a polynomial int1, . . . , tn of degreen.
Definition (Mixed-volume)
Definemvol(P1, . . . , Pn) as the coefficient oft1. . . tn in vol(t1P1+· · ·+tnPn).
IfP1 =· · ·=Pn=P, then t1P1+· · ·+tnPn= (t1+· · ·+tn)P, and so vol(P, . . . , P) = (t1+· · ·+tn)nvol(P).
This gives mvol(P, . . . , P) =n!vol(P).
Mixed-volume
ConsiderP, Q two convex bodies inRn. Define ai := mvol(P, . . . , P
| {z }
itimes
, Q, . . . , Q
| {z }
(n−i) times
).
Theorem (Alexandrov-Fenschel 1937) The sequence a0, . . . , an is log-concave.
Slogan: log-concavity is related to convexity via mixed-volumes.
1 Log-concave sequences in combinatorics
2 Log-concavity versus convexity
3 Algebro-geometric log-concavity
4 Applications and further results
Combinatorial log-concavity
fG(k) = number of proper colorings of Gbyk colors
=a0xn−a1xn−1+a2xn−2+· · ·+ (−1)nan. More generally, for a matroid M,
fM(x) =a0xr−a1xr−1+· · ·+ (−1)rar
the characteristic polynomial of M.
Theorem (Adiprasito-Huh-Katz 2015) The sequence a0, . . . , ar is log-concave.
Algebraic-topological context
X: atopological object.
Suppose we can associate to X an algebraA(X) decomposed as A(X) =A0(X)⊕A1(X)⊕ · · · ⊕Ad(X) such that
eachAi(X) is a real vector space of finite dimension.
the multiplication map sends (a, b)∈Ai×Aj to ab∈Ai+j. Example
X compact smooth manifold, andAi(X)theith cohomology ofX.
X=S2,A=R⊕(0)⊕R.
X=T torus, then A=R⊕R2⊕R. X=Sn, then A=R⊕(0)⊕ · · · ⊕(0)⊕R.
Algebraic-topological context
The decomposition
A(X) =A0(X)⊕A1(X)⊕ · · · ⊕Ad(X) usually has nice properties.
A0(X)'Ad(X)'R.
(Poincar´e duality) The multiplication
Ai(X)×Ad−i(X)→R=Ad(X) is non-degenerate.
Example
X anorientedcompact smooth manifold, Ai(X) theith cohomology group of X,A satisfies the Poincar´e duality.
Algebraic-topological context
The decomposition
A(X) =A0(X)⊕A1(X)⊕ · · · ⊕Ad(X) usually has nice properties.
A0(X)'Ad(X)'R.
(Poincar´e duality) The multiplication
Ai(X)×Ad−i(X)→R=Ad(X) is non-degenerate.
Example
X anorientedcompact smooth manifold, Ai(X) theith cohomology group of X,A satisfies the Poincar´e duality.
Algebro-geometric context
Additionally, in nice geometricsituations, one can definepositiveelements ω ∈A1(X).
(Lefschetz property) multiplication byωd−2i gives an isomorphism ωd−2i :Ai(X)−'→Ad−i(X).
In particular,
ωd:A0(X)−→' Ad(X) ωd−2:A1(X)−→' Ad−1(X) (Hodge-index) the bilinear form
A1(X)×A1(X)−→R (α, β)→αβωd−2 has precisely one positive eigenvalue.
Algebro-geometric context
Additionally, in nice geometricsituations, one can definepositiveelements ω ∈A1(X).
(Lefschetz property) multiplication byωd−2i gives an isomorphism ωd−2i :Ai(X)−'→Ad−i(X).
In particular,
ωd:A0(X)−→' Ad(X) ωd−2:A1(X)−→' Ad−1(X) (Hodge-index) the bilinear form
A1(X)×A1(X)−→R (α, β)→αβωd−2 has precisely one positive eigenvalue.
Algebro-geometric context
Additionally, in nice geometricsituations, one can definepositiveelements ω ∈A1(X).
(Lefschetz property) multiplication byωd−2i gives an isomorphism ωd−2i :Ai(X)−'→Ad−i(X).
In particular,
ωd:A0(X)−→' Ad(X) ωd−2:A1(X)−→' Ad−1(X) (Hodge-index) the bilinear form
A1(X)×A1(X)−→R (α, β)→αβωd−2 has precisely one positive eigenvalue.
Algebro-geometric context
A(X) =A0(X)⊕A1(X)⊕ · · · ⊕Ad(X).
(1) A0(X)'Ad(X)'R.
(2) (Poincar´e duality) multiplication Ai(X)×Ad−i(X)→Ris non-degenerate.
(3) (Lefschetz property) for positive ω∈A1(X),ωd−2i :Ai −→' Ad−i. (4) (Hodge-index) the bilinear form
(. , .) :A1(X)×A1(X)−→R, (α, β)→αβωd−2 has a unique positive eigenvalue.
Corollary
For positiveα, β ∈A1(X), we have det
(α, α) (α, β) (α, β) (β, β)
≤0.
Algebro-geometric context
A(X) =A0(X)⊕A1(X)⊕ · · · ⊕Ad(X).
(1) A0(X)'Ad(X)'R.
(2) (Poincar´e duality) multiplication Ai(X)×Ad−i(X)→Ris non-degenerate.
(3) (Lefschetz property) for positive ω∈A1(X),ωd−2i :Ai −→' Ad−i. (4) (Hodge-index) the bilinear form
(. , .) :A1(X)×A1(X)−→R, (α, β)→αβωd−2 has a unique positive eigenvalue.
Corollary
For positiveα, β ∈A1(X), we have det
(α, α) (α, β) (α, β) (β, β)
≤0.
Log-concavity in the algebro-geometric setting
For α, β two positive elements in A1(X), define ai :=αiβd−i ∈R'Ad(X).
Theorem (Khovanski-Teissier 80s) The sequence a0, . . . , adis log-concave.
Example
X projective complex manifold of dimensiond.
Ai(X) =H2i(X)the 2ith cohomology group of X.
X satisfies all the above properties.
For positive classes α, β∈H2(X), we get a log-concave sequence βd, αβd−1, . . . , αd−1β, αd.
Theorem (Okounkov 1996, Kaveh-Khovanski, Lazarsfeld-Mustata 2013)
There are convex bodies P andQ in Rd associated to α andβ such that
αiβd−i= mvol(P, . . . , P
| {z }
itimes
, Q, . . . , Q
| {z }
(d−i)times
).
The convex bodyP is called the Newton-Okounkov body ofα.
Example
X projective complex manifold of dimensiond.
Ai(X) =H2i(X)the 2ith cohomology group of X.
X satisfies all the above properties.
For positive classes α, β∈H2(X), we get a log-concave sequence βd, αβd−1, . . . , αd−1β, αd.
Theorem (Okounkov 1996, Kaveh-Khovanski, Lazarsfeld-Mustata 2013)
There are convex bodies P andQ in Rd associated to α andβ such that αiβd−i= mvol(P, . . . , P
| {z }
itimes
, Q, . . . , Q
| {z }
(d−i)times
).
The convex bodyP is called the Newton-Okounkov body ofα.
1 Log-concave sequences in combinatorics
2 Log-concavity versus convexity
3 Algebro-geometric log-concavity
4 Applications and further results
Combinatorial log-concavity via convex bodies
Conjecture
There exist convex bodies Pα and Qβ so that log-concavity of matroids is explained by log-concavity of mixed volumes.
Other log-concavity conjectures
Welsh conjecture on log-concavity of the values of the Tutte polynomial on the diagonal.
Log-concavity conjecture of Amelunxen-B¨urgisser for intrinsic volumes.
Hodge theory for tropical varieties
Smooth tropical fans provide local charts for smooth tropical varieties.
Theorem (A.-Piquerez)
Any smooth projective tropical variety verifies all the nice properties.
Computational complexity of counting bases of matroids
Counting number of spanning trees in a graph is easy (Matrix-Tree theorem). What about counting forests ?
Theorem (Anari-Gharan-Vinzant 2018)
The number of bases a matroidM of rank r can be approximated within a factor of O(2r)in polynomial time.
The proof is based on log-concavity of thebasis partition function on basis. Letxe for e∈E be variables. Define
P(x1, . . . , xn) = X
B∈B
Y
e∈B
xe.
Theorem (Anari-Gharan-Vinzant 2018) The application P :Rn+→R+ is log-concave.
Computational complexity of counting bases of matroids
Counting number of spanning trees in a graph is easy (Matrix-Tree theorem). What about counting forests ?
Theorem (Anari-Gharan-Vinzant 2018)
The number of bases a matroidM of rank r can be approximated within a factor of O(2r)in polynomial time.
The proof is based on log-concavity of thebasis partition function on basis. Letxe for e∈E be variables. Define
P(x1, . . . , xn) = X
B∈B
Y
e∈B
xe.
Theorem (Anari-Gharan-Vinzant 2018) The application P :Rn+→R+ is log-concave.
Computational complexity of counting bases of matroids
Theorem (Anari-Liu-Gharan-Vinzant 2018)
The number of bases a matroidM of rank r can be approximated within a factor of (1±) with probability larger than1−δ in polynomial time.
Proof based on log-concavity of the basis partition function and mixing properties of random walk on higher dimensional simplicial complexes.
Negative dependance for random forests
Conjecture (Kahn 2000 - Grimmet-Winkler 2004) For all pair of edges e, f ∈E,
P(e, f ∈F)≤P(e∈F)P(f ∈F) on uniform random forests F.
Theorem (Br¨anden-Huh 2019)
P(e, f ∈F)≤2P(e∈F)P(f ∈F).
Some References
Adiprasito-Huh-Katz,Hodge theory for combinatorial geometries, Ann. Math. 2018.
Amelunxen-B¨urgisser,Probabilistic analysis of the grassman condition number, Found. Comp. Math. 2015.
Amini-Piquerez,Homology of tropical fans, preprint 2021.
Amini-Piquerez,Hodge theory for tropical varieties, preprint 2020.
Anari-Gharan-Vinzant,Log-concave polynomials I, FOCS 2018.
Anari-Gharan-Liu-Vinzant,Log-concave polynomials I, STOCS 2019.
Borcea-Br¨anden-Ligget,Negative dependence and the geometry of polynomials, JAMS 2009.
Br¨anden-Hu,Lorenzian polynomials, Ann. Math. 2020.
Bobkov,Isoprimetric inequality for log-concave probability measures, Ann. Prob. 1999.
Gruber-Kuntz,General properties of polymer systems, Comm. Math. Physics 1971.
Huh,Milnor numbers and chromatic polynomial of graphs, JAMS 2012.
Kahn-Neiman,Negative correlation and log-concavity, Random structures and algorithms, 2010.
Kaveh-Khovanskii,Newton-Okounkov bodies, Ann. Math. 2012.
Lazarsfeld-Mustata,Convex bodies associated to linear series, Ann. de l’ENS 2009.
Milman-Rotem,Mixed integrals and related inequalities, J. Func. Anal. 2013.
Nelson,Almost all matroids are non-representable, Bulltin London Math. Soc. 2018.
Okounkov,Brunn–Minkowski inequality for multiplicities, Inventiones 1996.
Saumard-Wellner,Log-concavity and strong log-concavity, Statis. Surv. 2014.