7.6 Semidefinite Optimization¶
Semidefinite optimization is a generalization of conic optimization, allowing the use of matrix variables belonging to the convex cone of positive semidefinite matrices
where \(\Symm^r\) is the set of \(r \times r\) real-valued symmetric matrices.
MOSEK can solve semidefinite optimization problems stated in the primal form,
where the problem has \(p\) symmetric positive semidefinite variables \(\barX_j\in \PSD^{r_j}\) of dimension \(r_j\). The symmetric coefficient matrices \(\barC_j\in \Symm^{r_j}\) and \(\barA_{i,j}\in \Symm^{r_j}\) are used to specify PSD terms in the linear objective and the linear constraints, respectively. The symmetric coefficient matrices \(\barF_{i,j}\in \Symm^{r_j}\) are used to specify PSD terms in the affine conic constraints. Note that \(q\) ((7.13)) is the total dimension of all the cones, i.e. \(q=\text{dim}(\K_1 \times \ldots \times \K_k)\), given there are \(k\) ACCs. We use standard notation for the matrix inner product, i.e., for \(A,B\in \real^{m\times n}\) we have
In addition to the primal form presented above, semidefinite problems can be expressed in their dual form. Constraints in this form are usually called linear matrix inequalities (LMIs). LMIs can be easily specified in MOSEK using the vectorized positive semidefinite cone which is defined as:
Vectorized semidefinite domain:
\[\PSD^{d,\mathrm{vec}} = \left\{(x_1,\ldots,x_{d(d+1)/2})\in \real^n~:~ \mathrm{sMat}(x)\in\PSD^d\right\},\]where \(n=d(d+1)/2\) and,
\[\begin{split}\mathrm{sMat}(x) = \left[\begin{array}{cccc}x_1 & x_2/\sqrt{2} & \cdots & x_{d}/\sqrt{2} \\ x_2/\sqrt{2} & x_{d+1} & \cdots & x_{2d-1}/\sqrt{2} \\ \cdots & \cdots & \cdots & \cdots \\ x_{d}/\sqrt{2} & x_{2d-1}/\sqrt{2} & \cdots & x_{d(d+1)/2}\end{array}\right],\end{split}\]or equivalently
\[\PSD^{d,\mathrm{vec}} = \left\{\mathrm{sVec}(X)~:~X\in\PSD^d\right\},\]where
\[\mathrm{sVec}(X) = (X_{11},\sqrt{2}X_{21},\ldots,\sqrt{2}X_{d1},X_{22},\sqrt{2}X_{32},\ldots,X_{dd}).\]
In other words, the domain consists of vectorizations of the lower-triangular part of a positive semidefinite matrix, with the non-diagonal elements additionally rescaled. LMIs can be expressed by restricting appropriate affine expressions to this cone type.
For other types of cones supported by MOSEK, see Sec. 14.8 (Supported domains) and the other tutorials in this chapter. Different cone types can appear together in one optimization problem.
In Fusion the user can enter the linear expressions in a more convenient way, without having to cast the problem exactly in the above form.
We demonstrate the setup of semidefinite variables and their coefficient matrices in the following examples:
Sec. 7.6.1 (Example SDO1): A problem with one semidefinite variable and linear and conic constraints.
Sec. 7.6.2 (Example SDO2): A problem with two semidefinite variables with a linear constraint and bound.
Sec. 7.6.3 (Example SDO3): Shows how to efficiently set up many semidefinite variables of the same dimension.
7.6.1 Example SDO1¶
We consider the simple optimization problem with semidefinite and conic quadratic constraints:
The problem description contains a 3-dimensional symmetric semidefinite variable which can be written explicitly as:
and an affine conic constraint (ACC) \((x_0, x_1, x_2) \in \Q^3\). The objective is to minimize
subject to the two linear constraints
Our implementation in Fusion begins with creating a new model:
with Model("sdo1") as M:
We create a symmetric semidefinite variable \(\barX\) and another variable representing \(x\). For simplicity we immediately declare that \(x\) belongs to a quadratic cone
X = M.variable("X", Domain.inPSDCone(3))
x = M.variable("x", Domain.inQCone(3))
In this elementary example we are going to create an explicit matrix representation of the problem
and use it in the model via the dot product operation \(\langle\cdot,\cdot\rangle\) which applies to matrices as well as to vectors. This way we create each of the linear constraints and the objective as one expression.
# Objective
M.objective(ObjectiveSense.Minimize, Expr.add(Expr.dot(C, X), x.index(0)))
# Constraints
M.constraint("c1", Expr.add(Expr.dot(A1, X), x.index(0)), Domain.equalsTo(1.0))
M.constraint("c2", Expr.add(Expr.dot(A2, X), Expr.sum(x.slice(1,3))), Domain.equalsTo(0.5))
Now it remains to solve the problem with Model.solve
.
import mosek
from mosek.fusion import *
def main(args):
with Model("sdo1") as M:
# Setting up the variables
X = M.variable("X", Domain.inPSDCone(3))
x = M.variable("x", Domain.inQCone(3))
# Setting up constant coefficient matrices
C = Matrix.dense ( [[2.,1.,0.],[1.,2.,1.],[0.,1.,2.]] )
A1 = Matrix.eye(3)
A2 = Matrix.ones(3,3)
# Objective
M.objective(ObjectiveSense.Minimize, Expr.add(Expr.dot(C, X), x.index(0)))
# Constraints
M.constraint("c1", Expr.add(Expr.dot(A1, X), x.index(0)), Domain.equalsTo(1.0))
M.constraint("c2", Expr.add(Expr.dot(A2, X), Expr.sum(x.slice(1,3))), Domain.equalsTo(0.5))
M.solve()
print(X.level())
print(x.level())
7.6.2 Example SDO2¶
We now demonstrate how to define more than one semidefinite variable using the following problem with two matrix variables and two types of constraints:
In our example \(\dim(\barX_1)=3\), \(\dim(\barX_2)=4\), \(b=23\), \(k=-3\) and
are constant symmetric matrices.
Note that this problem does not contain any scalar variables, but they could be added in the same fashion as in Sec. 7.6.1 (Example SDO1).
The code representing the above problem is shown below.
# Convert input data into Fusion sparse matrices
C1 = Matrix.sparse(3, 3, C1_k, C1_l, C1_v)
C2 = Matrix.sparse(4, 4, C2_k, C2_l, C2_v)
A1 = Matrix.sparse(3, 3, A1_k, A1_l, A1_v)
A2 = Matrix.sparse(4, 4, A2_k, A2_l, A2_v)
# Define the model
with Model('sdo2') as M:
# Two semidefinite variables
X1 = M.variable(Domain.inPSDCone(3))
X2 = M.variable(Domain.inPSDCone(4))
# Objective
M.objective(ObjectiveSense.Minimize, Expr.add(Expr.dot(C1,X1), Expr.dot(C2,X2)))
# Equality constraint
M.constraint(Expr.add([Expr.dot(A1,X1), Expr.dot(A2,X2)]), Domain.equalsTo(b))
# Inequality constraint
M.constraint(X2.index([0,1]), Domain.lessThan(k))
# Solve
M.setLogHandler(sys.stdout)
M.solve()
# Retrieve result
print("X1:\n{0}".format(np.reshape(X1.level(), (3,3))))
print("X2:\n{0}".format(np.reshape(X2.level(), (4,4))))
7.6.3 Example SDO3¶
Here we demonstrate how to use the facilities provided in Fusion to set up a model with many semidefinite variables of the same dimension more efficiently than via looping. We consider a problem with \(n\) semidefinite variables of dimension \(d\) and \(k\) constraints:
with symmetric data matrices \(A_{ij}\).
The key construction is:
X = M.variable(Domain.inPSDCone(d, n))
It creates \(n\) symmetric, semidefinite matrix variables of dimension \(d\) arranged in a single variable object X
of shape \((n,d,d)\). Individual matrix variables can be accessed as slices from \((i,0,0)\) to \((i+1,d,d)\) (reshaped into shape \((d,d)\) if necessary). It is also possible to operate on the full variable X
when constructing expressions that involve entries of all the semidefinite matrices in a natural way. The source code example illustrates both these approaches.
# Sample input data
n = 100
d = 4
k = 3
b = [9,10,11]
A = list(map(lambda x: x+np.transpose(x),
[np.random.normal(0, 5, size=(d, d)) for _ in range(k*n)]))
# Create a model with n semidefinite variables od dimension d x d
with Model("sdo3") as M:
X = M.variable(Domain.inPSDCone(d, n))
# Pick indexes of diagonal entries for the objective
alldiag = [[j,s,s] for j in range(n) for s in range(d)]
M.objective(ObjectiveSense.Minimize, Expr.sum( X.pick(alldiag) ))
# Each constraint is a sum of inner products
# Each semidefinite variable is a slice of X
for i in range(k):
M.constraint(Expr.add([Expr.dot(A[i*n+j], X[j:j+1, :, :].reshape([d,d]))
for j in range(n)]) >= b[i])
# Solve
M.setLogHandler(sys.stdout) # Add logging
M.writeTask("sdo3.ptf") # Save problem in readable format
M.solve()
# Get results. Each variable is a slice of X
print("Contributing variables:")
for j in range(n):
Xj = X.slice([j,0,0],[j+1,d,d]).level()
if any(Xj[s]>1e-6 for s in range(d*d)):
print("X{0}=\n{1}".format(j, np.reshape(Xj,(d,d))))