Professional Documents
Culture Documents
Goemans - Approximating Submodular
Goemans - Approximating Submodular
Goemans - Approximating Submodular
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
As Published http://www.siam.org/proceedings/soda/2009/soda09.php
Again, if K is polyhedral, we only need to write the 3 Algorithm for Axis-Aligned Convex Bodies
constraint kck2A−1 ≤ 1 for c such that cT x ≤ 1 defines In this section, we consider the question of constructing
a facet of K. ellipsoidal approximations efficiently, we show how to
exploit symmetries of the convex body, and we relate
John’s theorem. John’s theorem, well-known in the
ellipsoidal approximations to the problem of approxi-
theory of Banach spaces, says that K is contained in
√ mating a submodular function everywhere.
n·E(A), where E(A) is the maximum√volume ellipsoid
We say that E(A) is a λ-ellipsoidal approximation
inscribed in K; in other words, kxkA ≤ n for all x ∈ K.
to K if E(A) ⊆ √ K and K ⊆ λE(A). The John ellipsoid
In terms of Banach spaces, this says that the (Banach-
is therefore a n-ellipsoidal
√ approximation to a convex
Mazur) distance between any n-dimensional Banach
body K, and so is 1/ n times the Löwner ellipsoid.
space (whose unit ball is √ K) and the n-dimensional
These are existential results. Algorithmically, the sit-
Hilbert space l2n is at most n.
uation very much depends on how the convex body is
John’s theorem
given. If it is a polyhedral set given explicitly as the
can be proved in sev- z
intersection of halfspaces then the convex program for
eral ways. See, for E(A) the John’s ellipsoid given above has one constraint for
example, Ball [4] or −z
each given inequality and can be solved approximately,
Matoušek [24, §13.4].
to within any desired accuracy. This gives an alter-
We adopt a more algo-
nate way to derive the result of Grötschel,
√ Lovász and
rithmic argument.
√ Suppose there is an element z ∈ K Schrijver giving in polynomial-time a n + 1-ellipsoidal
with kzkA > n. Then the following lemma gives an
approximation to a symmetric convex body K given ex-
explicit construction of an ellipsoid of strictly larger
plicitly by a system of linear inequalities. However, if K
volume that is contained in the convex hull of E(A),
is given by a separation oracle and comes with the as-
z and −z, as illustrated in the figure. The resulting
sumption of being well-bounded2 then the best (known)
ellipsoid is larger since kn (l) > 1 for l > n. This proves
algorithmic
p result is a polynomial-time algorithm giving
John’s theorem.
only a n(n + 1)-ellipsoidal approximation (Grötschel,
Lemma 2. For A Â 0 and z ∈ Rn with l = kzk2A ≥ n, Lovász, Schrijver [12], Theorem 4.6.3), and this will be
let too weak for our purpose. In fact, as was pointed out
µ ¶ to us by José Soto, no algorithm, even randomized, can
n l−1 n l−1 produce an approximation better than Õ(n) for general
L(A, z) = A+ 2 1− Azz T A.
l n−1 l n−1 centrally symmetric convex bodies.
The proof given above of John’s theorem can be
Then L(A, z) is positive definite, the ellipsoid made algorithmic if we have an α-approximation algo-
E(L(A, z)) is contained in conv{E(A), {z, −z}}, rithm (α ≤ 1) for maximizing√kxkA over x ∈ K and
and its volume vol(E(L(A, z)) equals kn (l) · vol(E(A)) we are willing to settle for a n + 1/α-ellipsoidal ap-
where sµ ¶ µ proximation. In fact, we only need an α-approximate
n ¶n−1
l n−1 decision procedure which, given A Â 0 with
kn (l) = . √ E(A) ⊆ K,
n l−1 either returns an x ∈ K with kxkA > n +√ 1 or guar-
antees that every x ∈ K satisfies kxkA ≤ n + 1/α.
In this extended abstract, most proofs are deferred Assume we are given an ellipsoid E0 ⊆ K such that
to the full version. Actually, the lemma also follows K ⊆ pE0 (p is for example R/r in the definition of well-
from existing results by considering the polar statement, boundedness, and for our application, we will be able
which says there exists an ellipsoid E(B −1 ) containing to use p = n). Iteratively, we find larger and larger
© ª (multiplicatively in volume) ellipsoids guaranteed to be
(2.2) E(A−1 ) ∩ x : −1 ≤ z T x ≤ 1
within K. Given an ellipsoid Ej = E(Aj ) at iteration j,
−1 −1
such that
√ vol(E(B )) < vol(E(A )), assuming
kzkA > n. See, for example, Grötschel, Lovász 2 As part of the input of this centrally symmetric convex body,
and Schrijver [12, p72], Bland, Goldfarb and Todd we get R ≥ r > 0 such that B(0, r) ⊆ K ⊆ B(0, R), and the
[5, p1056], and Todd [33]. In fact, Todd derives running time can thus involve log(R/r) polynomially.
with f (∅) = 0, its polymatroid Pf ⊆ Rn is defined by: (which shows that the maximum volume ellipsoid is axis
( ) aligned). We need some notation. For a vector a ∈ Rn ,
x(S) ≤ f (S), ∀S ⊆ [n] let Diag(a) be the diagonal matrix with main diagonal
P (f ) = a; for a matrix A ∈ Rn×n , let diag(A) ∈ Rn be its main
x ≥ 0
diagonal.
P
where x(S) = i∈S xi . To make it centrally symmetric,
let S(Q) = { x ∈ Rn : |x| ∈ Q }, where |x| denotes Proposition 3.1. Let K be an axis-aligned convex
component-wise absolute value. It is easy to see that, if body, and let E(A) be an ellipsoid inscribed in K.
f ({i}) > 0 for all i then S(Pf ) is a centrally symmetric Then the ellipsoid E(B)
−1 −1
defined by the diagonal matrix
convex body. (If there exists an index i with f ({i}) = B = (Diag(diag(A ))) satisfies (i) E(B) ⊆ K and
0, we can simply get rid of it as monotonicity and (ii) vol(E(B)) ≥ vol(E(A)).
submodularity imply that f (S) = f (S − i) for all S
with i ∈ S.) Suppose now that E(A) is a λ-ellipsoidal (ii) is a restatement of Hadamard’s inequality (ap-
approximation to S(Pf ). This implies that, for any plied to A−1 ) whichQsays that for a positive definite ma-
n
c ∈ Rn , trix C, det(C) ≤ i=1 ii . To prove (i), one can show
c
that E(B) ⊆ conv{T (E(A)) : T ∈ Aut(K)}.
kckA−1 = max{cT x : x ∈ E(A)} Proposition 3.1 shows that, for an axis-aligned
convex body such as S(Pf ), we can maintain throughout
≤ max{cT x : x ∈ S(Pf )}
the algorithm axis-aligned ellipsoids. This has two
≤ λ max{cT x : x ∈ E(A)} = λkckA−1 . important consequences. First, this means that we only
need an α-approximate decision procedure for the case
In particular, taking c = 1S (the indicator vector for S)
when A is diagonal. To emphasize this, we rename
for any S ⊆ [n], we get that
A by D. Recall that such a procedure, when given a
k1S kA−1 ≤ f (S) ≤ λk1S kA−1 , D Â 0 with E(D) ⊆ S(Pf√), either outputs a vector
x ∈ S(Pf√) with kxkD > n + 1 or guarantees that
where we have used the fact that max{1TS x : x ∈ Pf } = kxkD ≤ n + 1/α for all x ∈ S(Pf ). In section 4, we
f (S). Thus the function fˆ defined by fˆ(S) = k1S kA−1 show that, for rank functions of matroids, max{kxkD :
√
Figure 1: The algorithm for constructing a function fˆ which is a n + 1/α-approximation to f .
x ∈ S(Pf )} can be solved exactly (thus α = 1) and For a matroid rank function f , the problem
efficiently (in polynomial time and with polynomially max{kxkD : x ∈ Pf } can be solved exactly in
many oracle calls), while in Section 5, we describe polynomial-time and with a polynomial number of or-
an efficient 1/O(log n)-decision procedure for general acle calls, when D is a positive definite, diagonal ma-
monotone submodular functions. Secondly, the function trix. Indeed, maximizing ©kxk P D is2 equivalentª to max-
fˆ we construct based on an ellipsoidal approximation imizing its square: max i di xi : x ∈ Pf , where
takes a particularly simple form when the ellipsoid d = diag(D). This is the maximization of a convex
E(D) is given by a diagonal matrix D. In this case, function over a polyhedral set, and therefore the maxi-
fˆ(S) = k1S kD−1 reduces to: mum is attained at one of the vertices. But any ver-
sX tex x of Pf is a 0 − 1 vector [7] and thus satisfies
ˆ
f (S) = pi , x2i = xi . The problem
P is thus equivalent to maximizing
i∈S
the linear function i di xi over Pf which can be solved
in polynomial-time by the greedy algorithm for find-
where pi = 1/Dii for i ∈ [n]. Observe that this ing a maximum weight independent set in a matroid.
approximation fˆ is actually submodular (while this was Therefore,
√ Axis-Aligned-Ellipsoidal-Approx gives
not necessarily the case for non axis-aligned ellipsoids). a n + 1-approximation everywhere for rank functions
Summarizing, of matroids.
√ Figure 1 gives our algorithm for We should emphasize that the simple approach of
constructing a n + √ 1/α-ellipsoidal approximation of linearizing x2 by x would have failed if our ellipsoids
S(Pf ) and thus a n + 1/α-approximation to f ev- i i
erywhere, given an α-approximate decision procedure were not axis aligned, i.e., if D were not diagonal. In
Max-Norm(D) for maximizing kxkD over S(Pf ) (or fact, the quadratic spanning tree problem, defined as
equivalently over Pf , by symmetry) for a positive defi- max{kxkD : x ∈ Pf } where Pf is a graphic matroid
nite diagonal matrix D (i.e. dii > 0). polytope and D is a symmetric, non-diagonal matrix, is
One can easily check that the ellipsoid E0 = E(D0 ) NP-hard as it includes the Hamiltonian path problem
given in the algorithm is an n-ellipsoidal approximation: as a special case [3]. We remark that NP-hardness holds
it satisfies E0 ⊆ S(Pf ) and S(Pf ) ⊆ nE0 . even if D is positive definite.
giving us the required approximation guarantee. (6.4) β + |S ∩ R̄| < min {|S|, α} .
The lemmas below give a closed form expression
for g(·) and h(·); their proofs are used in the proof of Case 1: |S| ≤ α. Eq. (6.4) holds iff β + |S ∩ R̄| < |S|,
Theorem 6. They follow from the fact that the greedy which holds iff β < |S ∩ R|. That inequality together
algorithm can be used to maximize a linear function with |S| ≤ α implies that |S ∩ R̄| < α − β.
over a polymatroid. Both lemmas apply to any set S Case 2: |S| > α. Eq. (6.4) holds iff β + |S ∩ R̄| < α.
after renumbering its indices. For any i and j, we define That inequality implies that |S ∩R| > β +(|S|−α) > β.
[i, j] = {k ∈ N : i ≤ k ≤ j} and f (i, j) = f ([i, j]). Observe Our family of monotone functions is
that f (i, j) = 0 for i > j.
F = { rMR : R ⊆ [n], |R| = α } ∪ {rU } .
Lemma 7. For S = [k] with c1 ≤ c2 ≤ · · · ≤ ck , we
Pk
have g(S) = i=1 ci [f (i, k) − f (i + 1, k)] . Our family of non-monotone functions is
A union bound shows that none of these events oc- to the non-uniform submodular load balancing problem.
cur with high probability, and thus the algorithm fails
to distinguish rMR from rU . The approximation ra- 7.2 Submodular Max-Min Fair Allocation
tio√of the algorithm is at most f 0 (R)/f (R) = α/β = Consider m buyers and a ground set [n] of items. Let
Ω( n/ log k). This lower bound also applies to ran- f1 , . . . , fm be monotone submodular functions on the
domized algorithms, by the same reasoning as in the ground set [n], and let fj be the valuation function
non-monotone case. Since k = nO(1) , this proves the of buyer j. The submodular max-min fair allocation
desired result. problem is