Professional Documents
Culture Documents
Sarah Drewes 1
Fachbereich Mathematik
Technische Universität Darmstadt
Germany
2,3
Sebastian Pokutta
Fachbereich Mathematik
Technische Universität Darmstadt
Germany
February 5, 2010
Abstract
We analyze mixed 0/1 second order cone programs where the fractional and binary
variables are solely coupled via the conic constraints. For this special type of mixed-
integer second order cone programs we devise a cutting-plane framework based on
the generalized Benders cut. We show that the resulting cuts are very effective as
symmetric solutions are automatically cut off as well and each equivalence class of
0/1 solutions is visited at most once. We also present computational results showing
the effectiveness of our method and sketch an application in optimal pooling of
securities.
1
Email: drewes@mathematik.tu-darmstadt.de
2
Email: pokutta@mathematik.tu-darmstadt.de
3
Research partially supported by German Research Foundation (DFG) funded SFB 805.
1 Introduction
Mixed integer second order cone programming as one of the rather tractable
problem classes of mixed-integer nonlinear programming gained strong inter-
est in the recent years. Due to their structure and convexity, the underlying
relaxations can be solved in polynomial time to -optimality. Mixed integer
second order cone programs are usually solved via outer approximation algo-
rithms or within a branch-and-cut framework (see e.g., [5] for an overview).
Here we consider a restricted class of mixed integer second order cone pro-
grams where the integral variables are binary and the only coupling of binary
and fractional variables occurs in the conic constraints. For this problem class,
we derive cutting-planes based on the generalized Benders cuts (see e.g., [5]).
Applying an implicit Sherali-Adams step, we disentangle the binary variables
from the fractional variables in the conic constraints. As a result we obtain
a reformulation where the binary variables are aggregated in the linear parts
of the conic conditions. The resulting problem decomposes naturally and the
derived cuts account for symmetries arising from the weak coupling. We apply
these cuts and the resulting decomposition to solve a pooling problem arising
in portfolio optimization in finance.
Related work
Cutting-planes for mixed integer second order cone programs have been exten-
sively investigated. For example in [2], Gomory mixed-integer cuts for certain
classes of second order cone programs have been devised. A branch-and-cut
based method was outlined in [11] and in [8] lift-and-project based cutting-
planes have been used to solve mixed-integer second order cone programs.
Outline
In Section 2 we briefly recall the necessary notation, followed by the reformu-
lation of the considered problem class in Section 3. We then introduce the
cutting-plane framework based on the reformulation and the derived cutting-
planes in Section 4. The abstract concludes with briefly mentioning the pool-
ing problem in Section 5 and computational results in Section 6.
2 Preliminaries
We put [n] := {1, . . . , n} and for a set J ⊆ [n], we define J¯ := [n] \ J to be
the complement of J in [n]. All other notation is standard as to be found in
[1,8]. We will consider the following class of second order cone programs:
Definition 2.1 Let n ∈ N and J ⊆ [n]. Then a weakly-coupled 0/1 second
order cone program (wSOC) has the form:
min cJ xJ + cJ¯xJ¯
s.t. AJ xJ ≤ bJ
AJ¯xJ¯ ≤ bJ¯
(1)
(T` xJ , xJ¯ )
` ` 2
≤ κ` ∀ ` ∈ [k]
¯
xJ¯ ∈ RJ+
xJ ∈ {0, 1}|J| ,
with AJ ∈ RmJ ×|J| , AJ¯ ∈ RmJ¯×|J | , bJ ∈ RmJ , bJ¯ ∈ RmJ¯, c = (cJ , cJ¯) ∈ Rn ,
¯
The core of the classical outer approximation algorithm (see e.g., [5])
is to subsequently add linear approximations of the conic constraints. Let
(MIP-OA(ξ) ) denote the outer approximation problem (MIP-OA) after hav-
(ξ)
ing added ξ rounds of linearization cuts. Let xM denote its optimal solution
(ξ)
and let h` with ` ∈ [k] denote the associated h variables. Note that the latter
(ξ)
are indeed completely determined by (xM )J . In the following we will specify
the linearization cuts that we add in each round. Let
min cT x
s.t. AJ xJ ≤ bJ , AJ¯xJ¯ ≤ bJ¯,
(SOC(ξ) ) h` = κ2` − j∈J` ρ2j xj ≥ kxJ¯` k22 , ` ∈ [k],
P
(ξ)
xJ = (xM )J ,
be the second order cone program derived from (MISOC) by fixing the binary
(ξ) (ξ)
variables xJ to (xM )J . Let xSOC denote an optimal solution to (SOC(ξ) ) with
(ξ)
associated Lagrange multipliers µ̄` for ` ∈ [k]. Then, in round ξ of the
linearization phase, we add the following binary symmetric cut (BSC):
X (ξ) (ξ)
(ξ)
(BSC(ξ) ) µ̄` h` − h` ≥ cTJ¯ (xSOC )J¯ − cTJ¯ xJ¯
`∈[k]
The proposition above states that each coset is examined at most once;
a revisit allows for the construction of an optimal solution. As indicated
above, the (BSC) cuts can be integrated into e.g., an outer approximation or
branch-and-cut framework. While a detailed explanation will be contained in
the full-length paper, here we refer to e.g., [5].
Second order cone constraints naturally appear in the context of portfolio opti-
mization problems in finance: Limiting the risk of a portfolio can be expressed
as
kQxk2 ≤ σ
where σ limits the risk of the portfolio, x ∈ [0, 1]n with i∈[n] xi = 1 is the dis-
P
tribution of securities, and Q is the positive semidefinite square root (Cholesky
decomposition) of the covariance matrix of the securities’ returns. Portfolio
optimization problems using risk constraints have been studied extensively
(see e.g., [12,4,9,3,6] and the references contained therein). We study a slightly
different problem of optimal pooling: Consider a set of securities where some
of those can be split (almost) arbitrarily into parts and others cannot. The
latter are assumed to be uncorrelated and therefore possess only idiosyncratic
risks, e.g., cat-linked securities (see e.g., [7]). We search for a pooling of these
securities maximizing the return of each pool such that a predefined risk level
for each pool is not exceeded. We further allow enforcing combinatorial con-
straints on integral securities; it could be sensible to prevent two cat securities
with identical trigger from being in the same pool. The resulting optimization
problem constitutes a (wSOC). A detailed investigation will be included in
the full-length paper.
OA (BSC) only (BSC)+OA integer fractional rows obj.
# iterations # iterations # iterations # vars # vars val.
6 Computational results
In this section we present computational results for solving instances of the
problem described in Section 5 of various sizes; for each size we generated two
instances that are already reformulated as described above. We compare the
classical outer approximation algorithm to our algorithm in two variants; one
that solely uses (BSC) cuts and one that also includes outer approximation
cuts. Whereas for smaller instances the difference is negligible, for larger
instances the advantage of adding (BSC) cuts becomes apparent. The results
are summarized in Table 1. Instances marked with ∗ resulted in infeasible
sub-NLPs. In this case the algorithms use the same cuts.
References
[5] P. Bonami, M. Kilinc, and J. Linderoth. Algorithms and Software for Convex
Mixed Integer Nonlinear Programs. 2009.
[6] S. Ceria and R.A. Stubbs. Incorporating estimation errors into portfolio
selection: Robust portfolio construction. Journal of Asset Management,
7(2):109–127, 2006.
[7] J.D. Cummins. CAT Bonds and Other Risk-Linked Securities: State of the
Market and Recent Developments. Risk Management & Insurance Review,
11(1):23–47, 2008.
[8] S. Drewes. Mixed Integer Second Order Cone Programming. PhD thesis,
Technische Universität Darmstadt, 2009.
[9] M.S. Lobo, M. Fazel, and S. Boyd. Portfolio optimization with linear and fixed
transaction costs. Annals of Operations Research, 152(1):341–365, 2007.
[11] R.A. Stubbs and S. Mehrotra. A branch-and-cut method for 0-1 mixed convex
programming. Mathematical Programming, 86(3):515–532, 1999.
[12] J.P. Vielma, S. Ahmed, and G.L. Nemhauser. A Lifted Linear Programming
Branch-and-Bound Algorithm for Mixed-Integer Conic Quadratic Programs.
INFORMS Journal on Computing, 20(3):438–450, 2008.