1
0
mirror of https://github.com/kalmarek/PropertyT.jl.git synced 2024-11-19 15:25:29 +01:00
PropertyT.jl/property(T).jl

367 lines
12 KiB
Julia
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

using JuMP
import MathProgBase: AbstractMathProgSolver
import Base: rationalize
using GroupAlgebras
using ProgressMeter
using ValidatedNumerics
function create_product_matrix(basis, limit)
product_matrix = zeros(Int, (limit,limit))
basis_dict = Dict{Array, Int}(x => i
for (i,x) in enumerate(basis))
for i in 1:limit
x_inv::eltype(basis) = inv(basis[i])
for j in 1:limit
w = x_inv*basis[j]
product_matrix[i,j] = basis_dict[w]
# index = findfirst(basis, w)
# index ≠ 0 || throw(ArgumentError("Product is not supported on basis: $w"))
# product_matrix[i,j] = index
end
end
return product_matrix
end
function constraints_from_pm(pm, total_length=maximum(pm))
n = size(pm,1)
constraints = constraints = [Array{Int,1}[] for x in 1:total_length]
for j in 1:n
Threads.@threads for i in 1:n
idx = pm[i,j]
push!(constraints[idx], [i,j])
end
end
return constraints
end
function splaplacian_coeff(S, basis, n=length(basis))
result = spzeros(n)
result[1] = float(length(S))
for s in S
ind = findfirst(basis, s)
result[ind] += -1.0
end
return result
end
function laplacian_coeff(S, basis)
return full(splaplacian_coeff(S,basis))
end
function create_SDP_problem(matrix_constraints, Δ::GroupAlgebraElement; upper_bound=Inf)
N = size(Δ.product_matrix,1)
const Δ² = Δ*Δ
@assert length(Δ) == length(matrix_constraints)
m = JuMP.Model();
JuMP.@variable(m, A[1:N, 1:N], SDP)
JuMP.@SDconstraint(m, A >= 0)
JuMP.@constraint(m, sum(A[i] for i in eachindex(A)) == 0)
JuMP.@variable(m, κ >= 0.0)
if upper_bound < Inf
JuMP.@constraint(m, κ <= upper_bound)
end
JuMP.@objective(m, Max, κ)
for (pairs, δ², δ) in zip(matrix_constraints, Δ².coefficients, Δ.coefficients)
JuMP.@constraint(m, sum(A[i,j] for (i,j) in pairs) == δ² - κ*δ)
end
return m
end
function solve_SDP(SDP_problem, solver)
@show SDP_problem
@show solver
JuMP.setsolver(SDP_problem, solver);
# @time MathProgBase.writeproblem(SDP_problem, "/tmp/SDP_problem")
solution_status = JuMP.solve(SDP_problem);
if solution_status != :Optimal
warn("The solver did not solve the problem successfully!")
end
@show solution_status
κ = JuMP.getvalue(JuMP.getvariable(SDP_problem, ))
A = JuMP.getvalue(JuMP.getvariable(SDP_problem, :A))
return κ, A
end
function EOI{T<:Number}(Δ::GroupAlgebraElement{T}, κ::T)
return Δ*Δ - κ*Δ
end
function square_as_elt(vector, elt)
zzz = zeros(elt.coefficients)
zzz[1:length(vector)] = vector
# new_base_elt = GroupAlgebraElement(zzz, elt.product_matrix)
# return (new_base_elt*new_base_elt).coefficients
return GroupAlgebras.algebra_multiplication(zzz, zzz, elt.product_matrix)
end
function compute_SOS{T<:Number}(sqrt_matrix::Array{T,2},
elt::GroupAlgebraElement{T})
n = size(sqrt_matrix,2)
result = zeros(T, length(elt.coefficients))
p = Progress(n, 1, "Checking SOS decomposition...", 50)
for i in 1:n
result .+= square_as_elt(sqrt_matrix[:,i], elt)
next!(p)
end
return GroupAlgebraElement{T}(result, elt.product_matrix)
end
function correct_to_augmentation_ideal{T<:Rational}(sqrt_matrix::Array{T,2})
sqrt_corrected = similar(sqrt_matrix)
l = size(sqrt_matrix,2)
for i in 1:l
col = view(sqrt_matrix,:,i)
sqrt_corrected[:,i] = col - sum(col)//l
# @assert sum(sqrt_corrected[:,i]) == 0
end
return sqrt_corrected
end
function check_solution{T<:Number}(κ::T, sqrt_matrix::Array{T,2}, Δ::GroupAlgebraElement{T}; verbose=true, augmented=false)
result = compute_SOS(sqrt_matrix, Δ)
if augmented
epsilon = GroupAlgebras.ɛ(result)
if isa(epsilon, Interval)
@assert 0 in epsilon
elseif isa(epsilon, Rational)
@assert epsilon == 0//1
else
warn("Does checking for augmentation has meaning for $(typeof(epsilon))?")
end
end
SOS_diff = EOI(Δ, κ) - result
eoi_SOS_L₁_dist = norm(SOS_diff,1)
if verbose
@show κ
if augmented
println("ɛ(Δ² - κΔ - ∑ξᵢ*ξᵢ) = ", GroupAlgebras.ɛ(SOS_diff))
else
ɛ_dist = GroupAlgebras.ɛ(SOS_diff)
if typeof(ɛ_dist) <: Interval
ɛ_dist = ɛ_dist.lo
end
@printf("ɛ(Δ² - κΔ - ∑ξᵢ*ξᵢ) ≈ %.10f\n", ɛ_dist)
end
L₁_dist = eoi_SOS_L₁_dist
if typeof(L₁_dist) <: Interval
L₁_dist = L₁_dist.lo
end
@printf("‖Δ² - κΔ - ∑ξᵢ*ξᵢ‖₁ ≈ %.10f\n", L₁_dist)
end
distance_to_cone = κ - 2^3*eoi_SOS_L₁_dist
return distance_to_cone
end
import ValidatedNumerics
function (±)(X::AbstractArray, tol::Real)
r{T}(x::T) = ( x==zero(T) ? @interval(x) : x ± tol)
return r.(X)
end
(±)(X::GroupAlgebraElement, tol::Real) = GroupAlgebraElement(X.coefficients ± tol, X.product_matrix)
function Base.rationalize{T<:Integer, S<:Real}(::Type{T},
X::AbstractArray{S}; tol::Real=eps(eltype(X)))
r(x) = rationalize(T, x, tol=tol)
return r.(X)
end
(x, tol::Real) = rationalize(BigInt, x, tol=tol)
function _distance_to_positive_cone(Δ::GroupAlgebraElement, κ, A;
tol=1e-7, verbose=true, rational=false)
isapprox(eigvals(A), abs(eigvals(A)), atol=tol) ||
warn("The solution matrix doesn't seem to be positive definite!")
@assert A == Symmetric(A)
A_sqrt = real(sqrtm(A))
# println("")
# println("Checking in floating-point arithmetic...")
# @time fp_distance = check_solution(κ, A_sqrt, Δ, verbose=verbose)
# println("Floating point distance (to positive cone) ≈ $(Float64(trunc(fp_distance,8)))")
# println("-------------------------------------------------------------")
# println("")
#
# if fp_distance ≤ 0
# return fp_distance
# end
println("Checking in interval arithmetic...")
A_sqrtᴵ = A_sqrt ± tol
κᴵ = κ ± tol
Δᴵ = Δ ± tol
@time Interval_distance = check_solution(κᴵ, A_sqrtᴵ, Δᴵ, verbose=verbose)
# @assert isa(_distance, Rational)
println("The actual distance (to positive cone) is contained in $Interval_distance")
println("-------------------------------------------------------------")
println("")
if Interval_distance.lo 0
return Interval_distance.lo
end
println("Projecting columns of A_sqrt to the augmentation ideal...")
A_sqrt_ = (A_sqrt, tol)
A_sqrt__aug = correct_to_augmentation_ideal(A_sqrt_)
κ_ = (κ, tol)
Δ_ = (Δ, tol)
A_sqrt__augᴵ = A_sqrt__aug ± tol
κᴵ = κ_ ± tol
Δᴵ = Δ_ ± tol
@time Interval_dist_to_Σ² = check_solution(κᴵ, A_sqrt__augᴵ, Δᴵ, verbose=verbose, augmented=true)
println("The Augmentation-projected actual distance (to positive cone) is contained in $Interval_dist_to_Σ²")
println("-------------------------------------------------------------")
println("")
if Interval_dist_to_Σ².lo 0 || !rational
return Interval_dist_to_Σ².lo
else
println("Checking Projected SOS decomposition in exact rational arithmetic...")
@time _dist_to_Σ² = check_solution(κ_, A_sqrt__aug, Δ_, verbose=verbose, augmented=true)
@assert isa(_dist_to_Σ², Rational)
println("Augmentation-projected rational distance (to positive cone) ≥ $(Float64(trunc(_dist_to_Σ²,8)))")
println("-------------------------------------------------------------")
return _dist_to_Σ²
end
end
function pmΔfilenames(name::String)
if !isdir(name)
mkdir(name)
end
prefix = name
pm_filename = joinpath(prefix, "product_matrix.jld")
Δ_coeff_filename = joinpath(prefix, "delta.coefficients.jld")
return pm_filename, Δ_coeff_filename
end
function κSDPfilenames(name::String)
if !isdir(name)
mkdir(name)
end
prefix = name
κ_filename = joinpath(prefix, "kappa.jld")
SDP_filename = joinpath(prefix, "SDPmatrixA.jld")
return κ_filename, SDP_filename
end
function ΔandSDPconstraints(name::String)
pm_fname, Δ_fname = pmΔfilenames(name)
f₁ = isfile(pm_fname)
f₂ = isfile(Δ_fname)
if f₁ && f₂
println("Loading precomputed pm, Δ, sdp_constraints...")
product_matrix = load(pm_fname, "pm")
L = load(Δ_fname, "Δ")[:, 1]
Δ = GroupAlgebraElement(L, Array{Int,2}(product_matrix))
sdp_constraints = constraints_from_pm(product_matrix)
else
throw(ArgumentError("You need to precompute pm and Δ to load it!"))
end
return Δ, sdp_constraints
end
function ΔandSDPconstraints(name::String, ID, generating_func::Function)
pm_fname, Δ_fname = pmΔfilenames(name)
Δ, sdp_constraints = ΔandSDPconstraints(ID, generating_func())
save(pm_fname, "pm", Δ.product_matrix)
save(Δ_fname, "Δ", Δ.coefficients)
return Δ, sdp_constraints
end
function κandA(name::String)
κ_fname, SDP_fname = κSDPfilenames(name)
f₁ = isfile(κ_fname)
f₂ = isfile(SDP_fname)
if f₁ && f₂
println("Loading precomputed κ, A...")
κ = load(κ_fname, "κ")
A = load(SDP_fname, "A")
else
throw(ArgumentError("You need to precompute κ and SDP matrix A to load it!"))
end
return κ, A
end
function κandA(name::String, sdp_constraints, Δ::GroupAlgebraElement, solver::AbstractMathProgSolver; upper_bound=Inf)
println("Creating SDP problem...")
@time SDP_problem = create_SDP_problem(sdp_constraints, Δ; upper_bound=upper_bound)
println("Solving SDP problem maximizing κ...")
κ, A = solve_SDP(SDP_problem, solver)
κ_fname, A_fname = κSDPfilenames(name)
if κ > 0
save(κ_fname, "κ", κ)
save(A_fname, "A", A)
else
throw(ErrorException("Solver $solver did not produce a valid solution!: κ = "))
end
return κ, A
end
function check_property_T(name::String, ID, generate_B₄::Function;
verbose=true, tol=1e-6, upper_bound=Inf)
# solver = MosekSolver(INTPNT_CO_TOL_REL_GAP=tol, QUIET=!verbose)
solver = SCSSolver(eps=tol, max_iters=100000, verbose=verbose)
@show name
@show verbose
@show tol
Δ, sdp_constraints = try
ΔandSDPconstraints(name)
catch err
if isa(err, ArgumentError)
ΔandSDPconstraints(name, ID, generate_B₄)
else
throw(err)
end
end
println("|S| = $(countnz(Δ.coefficients) -1)")
@show length(Δ)
@show size(Δ.product_matrix)
κ, A = try
κandA(name)
catch err
if isa(err, ArgumentError)
κandA(name, sdp_constraints, Δ, solver; upper_bound=upper_bound)
else
throw(err)
end
end
@show κ
@show sum(A)
@show maximum(A)
@show minimum(A)
if κ > 0
true_kappa = _distance_to_positive_cone(Δ, κ, A, tol=tol, verbose=verbose, rational=true)
true_kappa = Float64(trunc(true_kappa,12))
if true_kappa > 0
println("κ($name, S) ≥ $true_kappa: Group HAS property (T)!")
else
println("κ($name, S) ≥ $true_kappa: Group may NOT HAVE property (T)!")
end
else
println("κ($name, S) ≥ < 0: Tells us nothing about property (T)")
end
end