Advanced Features

DCP warnings

When an expression is created which is not of DCP form, a warning is emitted. For example,

x = Variable()
y = Variable()
x*y

To disable this, run

Convex.emit_dcp_warnings() = false

to redefine the method. See Convex.emit_dcp_warnings for more details.

Dual Variables

Convex.jl also returns the optimal dual variables for a problem. These are stored in the dual field associated with each constraint.

using Convex, SCS

x = Variable()
constraint = x >= 0
p = minimize(x, constraint)
solve!(p, SCS.Optimizer)

# Get the dual value for the constraint
p.constraints[1].dual
# or
constraint.dual

Warmstarting

If you're solving the same problem many times with different values of a parameter, Convex.jl can initialize many solvers with the solution to the previous problem, which sometimes speeds up the solution time. This is called a warm start.

To use this feature, pass the optional argument warmstart=true to the solve! method.

# initialize data
n = 1000
y = rand(n)
x = Variable(n)

# first solve
lambda = Variable(Positive())
fix!(lambda, 100)
problem = minimize(sumsquares(y - x) + lambda * sumsquares(x - 10))
@time solve!(problem, SCS.Optimizer)

# now warmstart
# if the solver takes advantage of warmstarts,
# this run will be faster
fix!(lambda, 105)
@time solve!(problem, SCS.Optimizer, warmstart=true)

Fixing and freeing variables

Convex.jl allows you to fix a variable x to a value by calling the fix! method. Fixing the variable essentially turns it into a constant. Fixed variables are sometimes also called parameters.

fix!(x, v) fixes the variable x to the value v.

fix!(x) fixes x to its current value, which might be the value obtained by solving another problem involving the variable x.

To allow the variable x to vary again, call free!(x).

Fixing and freeing variables can be particularly useful as a tool for performing alternating minimization on nonconvex problems. For example, we can find an approximate solution to a nonnegative matrix factorization problem with alternating minimization as follows. We use warmstarts to speed up the solution.

# initialize nonconvex problem
n, k = 10, 1
A = rand(n, k) * rand(k, n)
x = Variable(n, k)
y = Variable(k, n)
problem = minimize(sum_squares(A - x*y), x>=0, y>=0)

# initialize value of y
set_value!(y, rand(k, n))
# we'll do 10 iterations of alternating minimization
for i=1:10
    # first solve for x
    # with y fixed, the problem is convex
    fix!(y)
    solve!(problem, SCS.Optimizer, warmstart = i > 1 ? true : false)
    free!(y)

    # now solve for y with x fixed at the previous solution
    fix!(x)
    solve!(problem, SCS.Optimizer, warmstart = true)
    free!(x)
end

Custom Variable Types

By making subtypes of Convex.AbstractVariable that conform to the appropriate interface (see the Convex.AbstractVariable docstring for details), one can easily provide custom variable types for specific constructions. These aren't always necessary though; for example, one can define the following function probabilityvector:

using Convex

function probabilityvector(d::Int)
    x = Variable(d, Positive())
    add_constraint!(x, sum(x) == 1)
    return x
end

probabilityvector (generic function with 1 method)

and then use, say, p = probabilityvector(3) in any Convex.jl problem. The constraints that the entries of p are non-negative and sum to 1 will be automatically added to any problem p is used in.

Custom types are necessary when one wants to dispatch on custom variables, use them as callable types, or provide a different implementation. Continuing with the probability vector example, let's say we often use probability vectors variables in taking expectation values, and we want to use function notation for this. To do so, we define

using Convex
mutable struct ProbabilityVector <: Convex.AbstractVariable
    head::Symbol
    id_hash::UInt64
    size::Tuple{Int, Int}
    value::Convex.ValueOrNothing
    vexity::Convex.Vexity
    function ProbabilityVector(d)
        this = new(:ProbabilityVector, 0, (d,1), nothing, Convex.AffineVexity())
        this.id_hash = objectid(this)
        this
    end
end

Convex.constraints(p::ProbabilityVector) = [ sum(p) == 1 ]
Convex.sign(::ProbabilityVector) = Convex.Positive()
Convex.vartype(::ProbabilityVector) = Convex.ContVar

(p::ProbabilityVector)(x) = dot(p, x)

Then one can call p = ProbabilityVector(3) to construct a our custom variable which can be used in Convex, which already encodes the appropriate constraints (non-negative and sums to 1), and which can act on constants via p(x). For example,

using SCS
p = ProbabilityVector(3)
x = [1.0, 2.0, 3.0]
prob = minimize( p(x) )
solve!(prob, SCS.Optimizer)
evaluate(p) # [1.0, 0.0, 0.0]

3-element Vector{Float64}:
 0.9999992771781011
 5.586492647192877e-7
 1.6194983054806818e-7

Subtypes of AbstractVariable must have the fields head, id_hash, and size, and id_hash must be populated as shown in the example. Then they must also

• either have a field value, or implement Convex._value and Convex.set_value!
• either have a field vexity, or implement Convex.vexity and Convex.vexity! (though the latter is only necessary if you wish to support Convex.fix! and Convex.free!
• have a field constraints or implement Convex.constraints (optionally, implement Convex.add_constraint! to be able to add constraints to your variable after its creation),
• either have a field sign or implement Convex.sign, and
• either have a field vartype, or implement Convex.vartype (optionally, implement Convex.vartype! to be able to change a variables' vartype after construction.)

Printing and the tree structure

A Convex problem is structured as a tree, with the root being the problem object, with branches to the objective and the set of constraints. The objective is an AbstractExpr which itself is a tree, with each atom being a node and having children which are other atoms, variables, or constants. Convex provides children methods from AbstractTrees.jl so that the tree-traversal functions of that package can be used with Convex.jl problems and structures. This is what allows powers the printing of problems, expressions, and constraints. The depth to which the tree corresponding to a problem, expression, or constraint is printed is controlled by the global variable Convex.MAXDEPTH, which defaults to 3. This can be changed by e.g. setting

Convex.MAXDEPTH[] = 5

Likewise, Convex.MAXWIDTH, which defaults to 15, controls the "width" of the printed tree. For example, when printing a problem with 20 constraints, only the first MAXWIDTH of the constraints will be printed. Vertical dots, "⋮", will be printed indicating that some constraints were omitted in the printing.

A related setting is Convex.MAXDIGITS, which controls printing the internal IDs of atoms: if the string representation of an ID is longer than double the value of MAXDIGITS, then it is shortened by printing only the first and last MAXDIGITS characters.

The AbstractTrees methods can also be used to analyze the structure of a Convex.jl problem. For example,

using Convex, AbstractTrees
x = Variable()
p = maximize( log(x), x >= 1, x <= 3 )
for leaf in AbstractTrees.Leaves(p)
    println("Here's a leaf: $(summary(leaf))")
end

Here's a leaf: real variable (id: 161…695)
Here's a leaf: real variable (id: 161…695)
Here's a leaf: constant (constant; positive)
Here's a leaf: real variable (id: 161…695)
Here's a leaf: constant (constant; positive)

We can also iterate over the problem in various orders. The following descriptions are taken from the AbstractTrees.jl docstrings, which have more information.

PostOrderDFS

Iterator to visit the nodes of a tree, guaranteeing that children will be visited before their parents.

for (i, node) in enumerate(AbstractTrees.PostOrderDFS(p))
    println("Here's node $i via PostOrderDFS: $(summary(node))")
end

Here's node 1 via PostOrderDFS: real variable (id: 161…695)
Here's node 2 via PostOrderDFS: log (concave; real)
Here's node 3 via PostOrderDFS: real variable (id: 161…695)
Here's node 4 via PostOrderDFS: constant (constant; positive)
Here's node 5 via PostOrderDFS: >= constraint (affine)
Here's node 6 via PostOrderDFS: real variable (id: 161…695)
Here's node 7 via PostOrderDFS: constant (constant; positive)
Here's node 8 via PostOrderDFS: <= constraint (affine)
Here's node 9 via PostOrderDFS: 2-element Vector{Constraint}
Here's node 10 via PostOrderDFS: Problem{Float64}

PreOrderDFS

Iterator to visit the nodes of a tree, guaranteeing that parents will be visited before their children.

for (i, node) in enumerate(AbstractTrees.PreOrderDFS(p))
    println("Here's node $i via PreOrderDFS: $(summary(node))")
end

Here's node 1 via PreOrderDFS: Problem{Float64}
Here's node 2 via PreOrderDFS: log (concave; real)
Here's node 3 via PreOrderDFS: real variable (id: 161…695)
Here's node 4 via PreOrderDFS: 2-element Vector{Constraint}
Here's node 5 via PreOrderDFS: >= constraint (affine)
Here's node 6 via PreOrderDFS: real variable (id: 161…695)
Here's node 7 via PreOrderDFS: constant (constant; positive)
Here's node 8 via PreOrderDFS: <= constraint (affine)
Here's node 9 via PreOrderDFS: real variable (id: 161…695)
Here's node 10 via PreOrderDFS: constant (constant; positive)

StatelessBFS

Iterator to visit the nodes of a tree, guaranteeing that all nodes of a level will be visited before their children.

for (i, node) in enumerate(AbstractTrees.StatelessBFS(p))
    println("Here's node $i via StatelessBFS: $(summary(node))")
end

Here's node 1 via StatelessBFS: Problem{Float64}
Here's node 2 via StatelessBFS: log (concave; real)
Here's node 3 via StatelessBFS: 2-element Vector{Constraint}
Here's node 4 via StatelessBFS: real variable (id: 161…695)
Here's node 5 via StatelessBFS: >= constraint (affine)
Here's node 6 via StatelessBFS: <= constraint (affine)
Here's node 7 via StatelessBFS: real variable (id: 161…695)
Here's node 8 via StatelessBFS: constant (constant; positive)
Here's node 9 via StatelessBFS: real variable (id: 161…695)
Here's node 10 via StatelessBFS: constant (constant; positive)

Reference

MAXDEPTH

Convex.MAXDEPTH[] = 5

MAXWIDTH

Convex.MAXWIDTH[] = 15

MAXDIGITS

Convex.MAXDIGITS[] = 3