Separation oracle

A separation oracle (also called a cutting-plane oracle) is a concept in the matematical theory of convex optimization. It is a method to describe a convex set that is given as an input to an optimization algorithm. Separation oracles are used as input to ellipsoid methods.

Definition

Let K be a convex and compact set in Rⁿ. A strong separation oracle for K is an oracle (black box) that, given a vector y in Rⁿ, returns one of the following:

• Assert that y is in K.
• Find a hyperplane that separates y from K: a vector c in Rⁿ, such that $c\cdot y>c\cdot x$ for all x in K.

A strong separation oracle is completely accurate, and thus may be hard to construct. For practical reasons, a weaker version is considered, which allows for small errors in the boundary of K and the inequalities. Given a small error tolerance d>0, we say that:

• A vector y is d-near K if its Euclidean distance from K is at most d;
• A vector y is d-deep in K if it is in K, and its Euclidean distance from any point in outside K is at least d.

The weak version also considers rational numbers, which have a representation of finite length, rather than arbitrary real numbers. A weak separation oracle for K is an oracle that, given a vector y in Qⁿ and a rational number d>0, returns one of the following::

• Assert that y is d-near K;
• Find a vector c in Qⁿ, normalized such that its maximum element is 1, such that $c\cdot y+d\geq c\cdot x$ for all x that are d-deep in K.

Implementation

A special case of a convex set is a set represented by linear inequalities: $K=\{x|Ax\leq b\}$. Such a set is called a convex polytope. A strong separation oracle for a convex polytope can be implemented, but its run-time depends on the input format.

Representation by inequalities

If the matrix A and the vector b are given as input, so that $K=\{x|Ax\leq b\}$, then a strong separation oracle can be implemented as follows. Given a point y, compute $Ay$:

• If the outcome is at most $b$, then y is in K by definition;
• Otherwise, there is at least one row $c$ of A, such that $c\cdot y$ is larger than the corresponding value in $b$; this row $c$ gives us the separating hyperplane, as $c\cdot y>b\geq c\cdot x$ for all x in K.

This oracle runs in polynomial time as long as the number of constraints is polynomial.

Representation by vertices

Suppose the set of vertices of K is given as an input, so that $K={\text{conv}}(v_{1},\ldots ,v_{k})=$ the convex hull of its vertices. Then, deciding whether y is in K requires to check whether y is a convex combination of the input vectors, that is, whether there exist coefficients z₁,...,z_k such that:

• $z_{1}\cdot v_{1}+\cdots +z_{k}\cdot v_{k}=y$;
• $0\leq z_{i}\leq 1$ for all i in 1,...,k.

This is a linear program with k variables and n equality constraints (one for each element of y). If y is not in K, then the above program has no solution, and the separation oracle needs to find a vector c such that

• $c\cdot y>c\cdot v_{i}$ for all i in 1,...,k.

Note that the two above represenations can be very different in size: it is possible that a polytope can be represented by a small number of inequalities, but has exponentially many vertices (for example, an n-dimensional cube). Conversely, it is possible that a polytope has a small number of vertices, but requires exponentially many inequalities (for example, the convex hull of the 2n vectors of the form (0,...,±1,...,0).

Problem-specific representation

In some linear optimization problems, even though the number of constraints is exponential, one can still write a custom separation oracle that works in polynomial time. Two examples are:

• The minimum-cost arborescence problem: given a weighted directed graph and a vertex r in it, find a subgraph of minimum cost that contains a directed path from r to any other vertex. The problem can be presented as an LP with a constraint for each subset of vertices, which is an exponential number of constraints. However, a separation oracle can be implemented using n-1 applications of the minimum cut procedure.
• The maximum independent set problem. It can be approximated by an LP with a constraint for every odd-length cycle. While there are exponentially-many such cycles, a separation oracle that works in polynomial time can be implemented by just finding an odd cycle of minimum length.

Non-linear sets

Let f be a convex function on Rⁿ. The set $K=\{(x,t)|f(x)\leq t\}$ is a convex set in Rⁿ⁺¹. Given an evaluation oracle for f (a black box that returns the value of f for every given point), one can easily check whether a vector (y,t) is in K. In order to get a separation oracle, we need also an oracle to evaluate the subgradient of f. Suppose some vector (y,s) is not in K, so f(y) > s. Let g be the subgradient of f at y (g is a vector in Rⁿ). Denote $c:=(g,-1)$.Then, $c\cdot (y,s)=g\cdot y-s>g\cdot y-f(y)$, and for all (x,t) in K: $c\cdot (x,t)=g\cdot x-t\leq g\cdot x-f(x)$. By definition of a subgradient: $f(x)\geq f(y)+g\cdot (x-y)$ for all x in Rⁿ. Therefore, $g\cdot y-f(y)\geq g\cdot x-f(x)$, so $c\cdot (y,s)>c\cdot (x,t)$ , and c represents a separating hyperplane.

Converting a weak oracle to a strong oracle

Given a weak separation oracle for a polyhedron, it is possible to construct a strong separation oracle by a careful method of rounding, or by diophantine approximations.

Relation to other oracles

Membership oracle

A membership oracle is a weaker variant of the separation oracle, which does not require the separating hyperplane.

Formally, a strong membership oracle for K is an oracle that, given a vector y in Rⁿ, returns one of the following:

• Assert that y is in K.
• Assert that y is in not in K.

Obviously, a strong separation oracle implies a strong membership oracle.

A weak membership oracle for K is an oracle that, given a vector y in Qⁿ and a rational number d>0, returns one of the following:

• Assert that y is d-near K;
• Assert that y is not d-deep in K.

Using a weak separation oracle (WSO), one can construct a weak membership oracle as follows.

• Run WSO(K,y,d). If it returns "yes", then return "yes".
• Otherwise, run WSO(K,y,d/3). If it returns "yes", then return "yes".
• Otherwise, return "no"; see for proof of correctness.

Optimization oracle

An optimization oracle is an oracle which finds an optimal point in a given convex set.

Formally, a strong optimization oracle for K is an oracle that, given a vector c in Rⁿ, returns one of the following:

• Assert that K is empty.
• Find a vector y in K that maximizes $c\cdot y$ (that is, $c\cdot y\geq c\cdot x$. for all x in K).

A weak optimization oracle for K is an oracle that, given a vector c in Qⁿ and a rational number d>0, returns one of the following:

• Assert that no point is d-deep in K;
• Find a vector y that is d-near K, such that $c\cdot y+d\geq c\cdot x$ for all x that are d-deep in K.

Violation oracle

A strong violation oracle for K is an oracle that, given a vector c in Rⁿ and a real number r, returns one of the following:

• Assert that $c\cdot x\leq r$ for for all x in K;
• Find a vector y in K with $c\cdot y>r$.

A weak violation oracle for K is an oracle that, given a vector c in Qⁿ, a rational number r, and a rational number d>0, returns one of the following:

• Assert that $c\cdot x\leq r+d$ for all x that are d-deep in K;
• Find a vector y that is d-near K, such that $c\cdot y+d\geq r$.

A major result in convex optimization is that, for any "well-described" polyhedron, the strong-separation problem, the strong-optimization problem and the strong-violation problem are equivalent

A major result in convex optimization is that, for any "well-described" polyhedron, the strong-separation problem, the strong-optimization problem and the strong-violation problem are polynomial-time-equivalent. That is, given an oracle for any one of these three problems, the other two problems can be solved in polynomial time. A polyhedron is called "well-described" if the input contains n (the number of dimensions of the space it lies in), and contains a number p such K can be defined by linear inequalities with encoding-length at most p. The result is proved using the ellipsoid method.