Multinomial randomness

NEWS.md

@@ -14,12 +14,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 * introduce `rand(:HamiltonianMatrices)`
 * extend `rand` to also `rand!` for `HamiltonianMatrices`, `SymplecticMatrices` and `SymplecticStiefel`
 * implement `riemannian_gradient` conversion for `SymplecticMatrices` and `SymplecticGrassmann`
+* the new manifold of `MultinomialSymmetricPositiveDefinite` matrices
+* `rand!` for `MultinomialDoublyStochastic` and `MultinomialSymmetric`
 
 ### Deprecated
 
 * Rename `Symplectic` to `SimplecticMatrices` in order to have a `Symplectic` wrapper for such matrices as well in the future for the next breaking change.
 * Rename `SymplecticMatrix` to `SymplecticElement` to clarify that it is the special matrix ``J_{2n}`` and not an arbitrary symplectic matrix.
 
+### Fixed
+
+* a bug that cause `project` for tangent vectors to return wrong results on `MultinomialDoublyStochastic`
+
 ## [0.9.12] – 2024-01-21
 
 ### Fixed

docs/make.jl

@@ -119,6 +119,7 @@ makedocs(;
                 "Multinomial doubly stochastic matrices" => "manifolds/multinomialdoublystochastic.md",
                 "Multinomial matrices" => "manifolds/multinomial.md",
                 "Multinomial symmetric matrices" => "manifolds/multinomialsymmetric.md",
+                "Multinomial symmetric positive definite matrices" => "manifolds/multinomialsymmetricpositivedefinite.md",
                 "Oblique manifold" => "manifolds/oblique.md",
                 "Probability simplex" => "manifolds/probabilitysimplex.md",
                 "Positive numbers" => "manifolds/positivenumbers.md",

docs/src/manifolds/multinomialdoublystochastic.md

@@ -7,3 +7,8 @@ Order = [:type, :function]
 ```
 
 ## Literature
+
+```@bibliography
+Pages = ["multinomialdoublystochastic.md"]
+Canonical=false
+```

docs/src/manifolds/multinomialsymmetric.md

@@ -7,3 +7,8 @@ Order = [:type, :function]
 ```
 
 ## Literature
+
+```@bibliography
+Pages = ["multinomialsymmetric.md"]
+Canonical=false
+```

docs/src/manifolds/multinomialsymmetricpositivedefinite.md

@@ -0,0 +1,14 @@
+# Multinomial symmetric positive definite matrices
+
+```@autodocs
+Modules = [Manifolds]
+Pages = ["manifolds/MultinomialSymmetricPositiveDefinite.jl"]
+Order = [:type, :function]
+```
+
+## Literature
+
+```@bibliography
+Pages = ["multinomialsymmetricpositivedefinite.md"]
+Canonical=false
+```

docs/src/references.bib

@@ -275,6 +275,7 @@ @article{DouikHassibi:2019
     TITLE   = {Manifold Optimization Over the Set of Doubly Stochastic Matrices: A Second-Order Geometry},
     JOURNAL = {IEEE Transactions on Signal Processing}
 }
+
 #
 # E
 # ----------------------------------------------------------------------------------------

src/Manifolds.jl

@@ -252,6 +252,7 @@ using ManifoldsBase:
     IsEmbeddedSubmanifold,
     IsExplicitDecorator,
     LogarithmicInverseRetraction,
+    ManifoldDomainError,
     ManifoldsBase,
     NestedPowerRepresentation,
     NestedReplacingPowerRepresentation,
@@ -426,6 +427,7 @@ include("manifolds/GeneralizedStiefel.jl")
 include("manifolds/Hyperbolic.jl")
 include("manifolds/MultinomialDoublyStochastic.jl")
 include("manifolds/MultinomialSymmetric.jl")
+include("manifolds/MultinomialSymmetricPositiveDefinite.jl")
 include("manifolds/PositiveNumbers.jl")
 include("manifolds/ProjectiveSpace.jl")
 include("manifolds/SkewHermitian.jl")
@@ -646,6 +648,7 @@ export Euclidean,
     MultinomialDoubleStochastic,
     MultinomialMatrices,
     MultinomialSymmetric,
+    MultinomialSymmetricPositiveDefinite,
     Oblique,
     OrthogonalMatrices,
     PositiveArrays,
@@ -781,7 +784,7 @@ export CachedBasis,
     InducedBasis,
     ProjectedOrthonormalBasis
 # Errors on Manifolds
-export ComponentManifoldError, CompositeManifoldError
+export ComponentManifoldError, CompositeManifoldError, ManifoldDomainError
 # Functions on Manifolds
 export ×,
     ^,

src/manifolds/Multinomial.jl

@@ -5,8 +5,11 @@ The multinomial manifold consists of `m` column vectors, where each column is of
 `n` and unit norm, i.e.
 
 ````math
-\mathcal{MN}(n,m) \coloneqq \bigl\{ p ∈ ℝ^{n×m}\ \big|\ p_{i,j} > 0 \text{ for all } i=1,…,n, j=1,…,m \text{ and } p^{\mathrm{T}}\mathbb{1}_m = \mathbb{1}_n\bigr\},
+\mathcal{MN}(n,m) \coloneqq \bigl\{
+    p ∈ ℝ^{n×m}\ \big|\ p_{i,j} > 0 \text{ for all } i=1,…,n, j=1,…,m
+    \text{ and } p^{\mathrm{T}}\mathbb{1}_m = \mathbb{1}_n\bigr\},
 ````
+
 where ``\mathbb{1}_k`` is the vector of length ``k`` containing ones.
 
 This yields exactly the same metric as
@@ -85,6 +88,28 @@ function power_dimensions(M::MultinomialMatrices)
     return (m,)
 end
 
+@doc raw"""
+    riemannian_gradient(M::MultinomialMatrices, p, Y; kwargs...)
+
+Let ``Y`` denote the Euclidean gradient of a function ``\tilde f`` defined in the
+embedding neighborhood of `M`, then the Riemannian gradient is given by
+Equation 5 of [DouikHassibi:2019](@cite) as
+
+```math
+  \operatorname{grad} f(p) = \proj_{T_p\mathcal M}(Y⊙p)
+```
+
+where ``⊙`` denotes the Hadamard or elementwise product.
+
+"""
+riemannian_gradient(M::MultinomialMatrices, p, Y; kwargs...)
+
+function riemannian_gradient!(M::MultinomialMatrices, X, p, Y; kwargs...)
+    X .= p .* Y
+    project!(M, X, p, X)
+    return X
+end
+
 representation_size(M::MultinomialMatrices) = get_parameter(M.size)
 
 function Base.show(io::IO, ::MultinomialMatrices{TypeParameter{Tuple{n,m}}}) where {n,m}

src/manifolds/MultinomialDoublyStochastic.jl

@@ -3,19 +3,57 @@
 
 A common type for manifolds that are doubly stochastic, for example by direct constraint
 [`MultinomialDoubleStochastic`](@ref) or by symmetry [`MultinomialSymmetric`](@ref),
+or additionally by symmetric positive definiteness [`MultinomialSymmetricPositiveDefinite`](@ref)
 as long as they are also modeled as [`IsIsometricEmbeddedManifold`](https://juliamanifolds.github.io/ManifoldsBase.jl/stable/decorator.html#ManifoldsBase.IsIsometricEmbeddedManifold).
+
+That way they share the inner product (just by restriction), and even the Riemannian gradient
 """
 abstract type AbstractMultinomialDoublyStochastic <: AbstractDecoratorManifold{ℝ} end
 
 function active_traits(f, ::AbstractMultinomialDoublyStochastic, args...)
     return merge_traits(IsIsometricEmbeddedManifold())
 end
 
+@doc raw"""
+    representation_size(M::AbstractMultinomialDoublyStochastic)
+
+return the representation size of doubly stochastic matrices, whic are embedded
+in the ``ℝ^{n×n}`` matrices and hence the answer here is ``
+"""
+function representation_size(M::AbstractMultinomialDoublyStochastic)
+    n = get_parameter(M.size)[1]
+    return (n, n)
+end
+
+@doc raw"""
+    riemannian_gradient(M::AbstractMultinomialDoublyStochastic, p, Y; kwargs...)
+
+Let ``Y`` denote the Euclidean gradient of a function ``\tilde f`` defined in the
+embedding neighborhood of `M`, then the Riemannian gradient is given by
+Lemma 1 [DouikHassibi:2019](@cite) as
+
+```math
+  \operatorname{grad} f(p) = \proj_{T_p\mathcal M}(Y⊙p)
+```
+
+where ``⊙`` denotes the Hadamard or elementwise product, and the projection
+is the projection onto the tangent space of the corresponding manifold.
+
+"""
+riemannian_gradient(M::AbstractMultinomialDoublyStochastic, p, Y; kwargs...)
+
+function riemannian_gradient!(M::AbstractMultinomialDoublyStochastic, X, p, Y; kwargs...)
+    X .= p .* Y
+    project!(M, X, p, X)
+    return X
+end
+
 @doc raw"""
     MultinomialDoublyStochastic{T} <: AbstractMultinomialDoublyStochastic
 
 The set of doubly stochastic multinomial matrices consists of all ``n×n`` matrices with
 stochastic columns and rows, i.e.
+
 ````math
 \begin{aligned}
 \mathcal{DP}(n) \coloneqq \bigl\{p ∈ ℝ^{n×n}\ \big|\ &p_{i,j} > 0 \text{ for all } i=1,…,n, j=1,…,m,\\
@@ -60,44 +98,27 @@ end
 Checks whether `p` is a valid point on the [`MultinomialDoubleStochastic`](@ref)`(n)` `M`,
 i.e. is a  matrix with positive entries whose rows and columns sum to one.
 """
-function check_point(
-    M::MultinomialDoubleStochastic,
-    p::T;
-    atol::Real=sqrt(prod(representation_size(M))) * eps(real(float(number_eltype(T)))),
-    kwargs...,
-) where {T}
+function check_point(M::MultinomialDoubleStochastic, p; kwargs...)
     n = get_parameter(M.size)[1]
-    r = sum(p, dims=2)
-    if !isapprox(norm(r - ones(n, 1)), 0.0; atol=atol, kwargs...)
-        return DomainError(
-            r,
-            "The point $(p) does not lie on $M, since its rows do not sum up to one.",
-        )
-    end
-    return nothing
+    s = check_point(MultinomialMatrices(n, n), p; kwargs...)
+    !isnothing(s) && return s
+    s2 = check_point(MultinomialMatrices(n, n), p'; kwargs...)
+    return s2
 end
+
 @doc raw"""
     check_vector(M::MultinomialDoubleStochastic p, X; kwargs...)
 
 Checks whether `X` is a valid tangent vector to `p` on the [`MultinomialDoubleStochastic`](@ref) `M`.
 This means, that `p` is valid, that `X` is of correct dimension and sums to zero along any
 column or row.
 """
-function check_vector(
-    M::MultinomialDoubleStochastic,
-    p,
-    X::T;
-    atol::Real=sqrt(prod(representation_size(M))) * eps(real(float(number_eltype(T)))),
-    kwargs...,
-) where {T}
-    r = sum(X, dims=2) # check for stochastic rows
-    if !isapprox(norm(r), 0.0; atol=atol, kwargs...)
-        return DomainError(
-            r,
-            "The matrix $(X) is not a tangent vector to $(p) on $(M), since its rows do not sum up to zero.",
-        )
-    end
-    return nothing
+function check_vector(M::MultinomialDoubleStochastic, p, X; kwargs...)
+    n = get_parameter(M.size)[1]
+    s = check_vector(MultinomialMatrices(n, n), p, X; kwargs...)
+    !isnothing(s) && return s
+    s2 = check_vector(MultinomialMatrices(n, n), p, X'; kwargs...)
+    return s2
 end
 
 function get_embedding(::MultinomialDoubleStochastic{TypeParameter{Tuple{n}}}) where {n}
@@ -134,11 +155,14 @@ end
 
 Project `Y` onto the tangent space at `p` on the [`MultinomialDoubleStochastic`](@ref) `M`, return the result in `X`.
 The formula reads
+
 ````math
     \operatorname{proj}_p(Y) = Y - (α\mathbf{1}_n^{\mathrm{T}} + \mathbf{1}_nβ^{\mathrm{T}}) ⊙ p,
 ````
+
 where ``⊙`` denotes the Hadamard or elementwise product and ``\mathbb{1}_n`` is the vector of length ``n`` containing ones.
 The two vectors ``α,β ∈ ℝ^{n×n}`` are computed as a solution (typically using the left pseudo inverse) of
+
 ````math
     \begin{pmatrix} I_n & p\\p^{\mathrm{T}} & I_n \end{pmatrix}
     \begin{pmatrix} α\\ β\end{pmatrix}
@@ -152,8 +176,8 @@ project(::MultinomialDoubleStochastic, ::Any, ::Any)
 
 function project!(M::MultinomialDoubleStochastic, X, p, Y)
     n = get_parameter(M.size)[1]
-    ζ = [I p; p I] \ [sum(Y, dims=2); sum(Y, dims=1)'] # Formula (25) from 1802.02628
-    return X .= Y .- (repeat(ζ[1:n], 1, 3) .+ repeat(ζ[(n + 1):end]', 3, 1)) .* p
+    ζ = [I p; p' I] \ [sum(Y, dims=2); sum(Y, dims=1)'] # Formula (25) from 1802.02628
+    return X .= Y .- (repeat(ζ[1:n], 1, n) .+ repeat(ζ[(n + 1):end]', n, 1)) .* p
 end
 
 @doc raw"""
@@ -201,15 +225,45 @@ function project!(
     return q
 end
 
-function representation_size(M::MultinomialDoubleStochastic)
-    n = get_parameter(M.size)[1]
-    return (n, n)
+@doc raw"""
+    rand(::MultinomialDoubleStochastic; vector_at=nothing, σ::Real=1.0, kwargs...)
+
+Generate random points on the [`MultinomialDoubleStochastic`](@ref) manifold
+or tangent vectors at the point `vector_at` if that is not `nothing`.
+
+Let ``n×n`` denote the matrix dimension of the [`MultinomialDoubleStochastic`](@ref).
+
+When `vector_at` is nothing, this is done by generating a random matrix`rand(n,n)`
+with positive entries and projecting it onto the manifold. The `kwargs...` are
+passed to this projection.
+
+When `vector_at` is not `nothing`, a random matrix in the ambient space is generated
+and projected onto the tangent space
+"""
+rand(::MultinomialDoubleStochastic; σ::Real=1.0)
+
+function Random.rand!(
+    rng::AbstractRNG,
+    M::MultinomialDoubleStochastic,
+    pX;
+    vector_at=nothing,
+    σ::Real=one(real(eltype(pX))),
+    kwargs...,
+)
+    rand!(rng, pX)
+    pX .*= σ
+    if vector_at === nothing
+        project!(M, pX, pX; kwargs...)
+    else
+        project!(M, pX, vector_at, pX)
+    end
+    return pX
 end
 
 @doc raw"""
     retract(M::MultinomialDoubleStochastic, p, X, ::ProjectionRetraction)
 
-compute a projection based retraction by projecting $p\odot\exp(X⨸p)$ back onto the manifold,
+compute a projection based retraction by projecting ``p\odot\exp(X⨸p)`` back onto the manifold,
 where ``⊙,⨸`` are elementwise multiplication and division, respectively. Similarly, ``\exp``
 refers to the elementwise exponentiation.
 """