Fix inference with empty cone

src/sos_polynomial.jl

@@ -27,6 +27,10 @@ function vectorized_matrix(
     )
 end
 
+function gram_matrix_type(B::Type, matrix_cone_type, T::Type)
+    return GramMatrix{T,B,T}
+end
+
 # Need these two methods to avoid ambiguity
 function build_gram_matrix(
     q::Vector,
@@ -70,21 +74,28 @@ end
 #    return build_gram_matrix(C, basis, T, SymmetricVectorized())
 #end
 
+# Need to annotate the element type in case `bases`
+# is empty
+function _build_matrix_eltype(::Type{T}, bases, f) where {T}
+    return Base._return_type(f, Tuple{Vector{T},eltype(bases)})
+end
+
 function build_matrix(
     Q::Function,
     bases::Vector{<:AbstractPolynomialBasis},
     f::Function,
-)
-    return map(eachindex(bases)) do i
-        return f(Q(i), bases[i])
-    end
+    ::Type{T},
+) where {T}
+    @show T
+    return T[f(Q(i), bases[i]) for i in eachindex(bases)]
 end
 function build_matrix(
     Q::Function,
     bases::Vector{<:Vector{<:AbstractPolynomialBasis}},
     f::Function,
-)
-    return [
+    ::Type{T},
+) where {T}
+    return T[
         f(Q(i), bases[i][j]) for i in eachindex(bases) for
         j in eachindex(bases[i])
     ]
@@ -96,13 +107,14 @@ function build_gram_matrix(Q::Function, bases::Vector, matrix_cone_type, T)
             Q,
             bases,
             (Q, b) -> build_gram_matrix(Q, b, matrix_cone_type, T),
+            gram_matrix_type(eltype(bases), matrix_cone_type, T),
         ),
     )
 end
 
 function build_moment_matrix(Q::Function, bases::Vector)
     return MultivariateMoments.BlockDiagonalMomentMatrix(
-        build_matrix(Q, bases, build_moment_matrix),
+        build_matrix(Q, bases, build_moment_matrix, T),
     )
 end
 
@@ -132,10 +144,10 @@ _first(b::AbstractPolynomialBasis) = b
 _first(b::Vector) = first(b)
 function add_gram_matrix(
     model::MOI.ModelLike,
-    matrix_cone_type::Type,
+    ::Type{matrix_cone_type},
     bases::Vector,
-    T::Type,
-)
+    ::Type{T},
+) where {matrix_cone_type,T}
     Qs = Vector{Vector{MOI.VariableIndex}}(undef, length(bases))
     cQs = Vector{union_constraint_indices_types(matrix_cone_type)}(
         undef,

test/issue_344.jl

@@ -0,0 +1,38 @@
+using JuMP
+using SumOfSquares
+using DynamicPolynomials
+function sos_min(sparsity, d, obj, dom)
+    model =
+        Model(() -> MOI.Utilities.MockOptimizer(MOI.Utilities.Model{Float64}()))
+    @variable(model, t)
+    @objective(model, Min, t)
+    con_ref = @constraint(
+        model,
+        obj - t in SOSCone(),
+        sparsity = sparsity,
+        maxdegree = d,
+        domain = dom
+    )
+    return optimize!(model)
+    #return moment_matrix(con_ref)
+end
+@polyvar x[1:3] l[1:3]
+my_game_obj = x[1]^2 + x[2]^2 + 1.0 * x[3]^2
+my_eq_list = [
+    -1.0 + l[3] - 3l[2] - 3x[2]^2 - x[1] * x[2],
+    3x[2] * l[2] - x[2]^2 * l[2] - x[1]^2 * l[2],
+    l[3] - x[3] * l[3],
+    -2 + x[3] + 2x[1] + 2x[1] * l[1],
+    -2 - x[3] + 2x[2] + 2x[2] * l[1],
+    2l[1] - x[3] * l[1] - x[2]^2 * l[1] - x[1]^2 * l[1],
+]
+my_ineq_list = [
+    3x[3] - x[2]^2 - x[1]^2,
+    1 - x[3],
+    l[2],
+    l[3],
+    2 - x[3] - x[2]^2 - x[1]^2,
+    l[1],
+]
+my_Dom = basicsemialgebraicset(algebraicset(my_eq_list), my_ineq_list)
+ν = sos_min(Sparsity.Monomial(ChordalCompletion()), 4, my_game_obj, my_Dom)