Symplectic Stiefel

SymplecticStiefel{n, k, 𝔽} <: AbstractEmbeddedManifold{𝔽, DefaultIsometricEmbeddingType}

The symplectic Stiefel manifold consists of all $2n × 2k, \; n \geq k$ matrices satisfying the requirement

\[\operatorname{SpSt}(2n, 2k, ℝ) = \bigl\{ p ∈ ℝ^{2n × 2n} \, \big| \, p^{\mathrm{T}}Q_{2n}p = Q_{2k} \bigr\},\]

where

\[Q_{2n} = \begin{bmatrix} 0_n & I_n \\ -I_n & 0_n \end{bmatrix}.\]

The symplectic Stiefel tangent space at $p$ can be parametrized as [BZ21]

\[ \begin{align*} T_p\operatorname{SpSt}(2n, 2k) = \{&X \in \mathbb{R}^{2n \times 2k} \;|\; p^{T}Q_{2n}X + X^{T}Q_{2n}p = 0 \}, \\ = \{&X = pΩ + p^sB \;|\; Ω ∈ ℝ^{2k × 2k}, Ω^+ = -Ω, \\ &\; p^s ∈ \operatorname{SpSt}(2n, 2(n- k)), B ∈ ℝ^{2(n-k) × 2k}, \}, \end{align*}\]

where $Ω \in \mathfrak{sp}(2n,F)$ is Hamiltonian and $p^s$ means the symplectic complement of $p$ s.t. $p^{+}p^{s} = 0$.

Constructor

SymplecticStiefel(2n::Int, 2k::Int, field::AbstractNumbers=ℝ)
    -> SymplecticStiefel{div(2n, 2), div(2k, 2), field}()

Generate the (real-valued) symplectic Stiefel manifold of $2n \times 2k$ matrices which span a $2k$ dimensional symplectic subspace of $ℝ^{2n \times 2n}$. The constructor for the SymplecticStiefel manifold accepts the even column dimension $2n$ and an even number of columns $2k$ for the real symplectic Stiefel manifold with elements $p \in ℝ^{2n × 2k}$.

exp(::SymplecticStiefel, p, X)
exp!(M::SymplecticStiefel, q, p, X)

Compute the exponential mapping

\[ \operatorname{exp}\colon T\operatorname{SpSt}(2n, 2k) \rightarrow \operatorname{SpSt}(2n, 2k)\]

at a point $p \in \operatorname{SpSt}(2n, 2k)$ in the direction of $X \in T_p\operatorname{SpSt}(2n, 2k)$.

The tangent vector $X$ can be written in the form $X = \bar{\Omega}p$ [BZ21], with

\[ \bar{\Omega} = X (p^{\mathrm{T}}p)^{-1}p^{\mathrm{T}} + Q_{2n}p(p^{\mathrm{T}}p)^{-1}X^{\mathrm{T}}(I_{2n} - Q_{2n}^{\mathrm{T}}p(p^{\mathrm{T}}p)^{-1}p^{\mathrm{T}}Q_{2n})Q_{2n} \in ℝ^{2n \times 2n},\]

where $Q_{2n}$ is the SymplecticMatrix. Using this expression for $X$, the exponential mapping can be computed as

\[ \operatorname{exp}_p(X) = \operatorname{Exp}([\bar{\Omega} - \bar{\Omega}^{\mathrm{T}}]) \operatorname{Exp}(\bar{\Omega}^{\mathrm{T}})p,\]

where $\operatorname{Exp}(\cdot)$ denotes the matrix exponential.

Computing the above mapping directly however, requires taking matrix exponentials of two $2n \times 2n$ matrices, which is computationally expensive when $n$ increases. Therefore we instead follow [BZ21] who express the above exponential mapping in a way which only requires taking matrix exponentials of an $8k \times 8k$ matrix and a $4k \times 4k$ matrix.

To this end, first define

\[\bar{A} = Q_{2k}p^{\mathrm{T}}X(p^{\mathrm{T}}p)^{-1}Q_{2k} + (p^{\mathrm{T}}p)^{-1}X^{\mathrm{T}}(p - Q_{2n}^{\mathrm{T}}p(p^{\mathrm{T}}p)^{-1}Q_{2k}) \in ℝ^{2k \times 2k},\]

and

\[\bar{H} = (I_{2n} - pp^+)Q_{2n}X(p^{\mathrm{T}}p)^{-1}Q_{2k} \in ℝ^{2n \times 2k}.\]

We then let $\bar{\Delta} = p\bar{A} + \bar{H}$, and define the matrices

\[ γ = \left[\left(I_{2n} - \frac{1}{2}pp^+\right)\bar{\Delta} \quad -p \right] \in ℝ^{2n \times 4k},\]

and

\[ λ = \left[Q_{2n}^{\mathrm{T}}pQ_{2k} \quad \left(\bar{\Delta}^+\left(I_{2n} - \frac{1}{2}pp^+\right)\right)^{\mathrm{T}}\right] \in ℝ^{2n \times 4k}.\]

With the above defined matrices it holds that $\bar{\Omega} = λγ^{\mathrm{T}}$. As a last preliminary step, concatenate $γ$ and $λ$ to define the matrices $Γ = [λ \quad -γ] \in ℝ^{2n \times 8k}$ and $Λ = [γ \quad λ] \in ℝ^{2n \times 8k}$.

With these matrix constructions done, we can compute the exponential mapping as

\[ \operatorname{exp}_p(X) = Γ \operatorname{Exp}(ΛΓ^{\mathrm{T}}) \begin{bmatrix} 0_{4k} \\ I_{4k} \end{bmatrix} \operatorname{Exp}(λγ^{\mathrm{T}}) \begin{bmatrix} 0_{2k} \\ I_{2k} \end{bmatrix}.\]

which only requires computing the matrix exponentials of $ΛΓ^{\mathrm{T}} \in ℝ^{8k \times 8k}$ and $λγ^{\mathrm{T}} \in ℝ^{4k \times 4k}$.

inv(::SymplecticStiefel{n, k}, A)
inv!(::SymplecticStiefel{n, k}, q, p)

Compute the symplectic inverse $A^+$ of matrix $A ∈ ℝ^{2n × 2k}$. Given a matrix

\[A ∈ ℝ^{2n × 2k},\quad A = \begin{bmatrix} A_{1, 1} & A_{1, 2} \\ A_{2, 1} & A_{2, 2} \end{bmatrix},\; A_{i, j} \in ℝ^{2n × 2k}\]

the symplectic inverse is defined as:

\[A^{+} := Q_{2k}^{\mathrm{T}} A^{\mathrm{T}} Q_{2n},\]

where

\[Q_{2n} = \begin{bmatrix} 0_n & I_n \\ -I_n & 0_n \end{bmatrix}.\]

For any $p \in \operatorname{SpSt}(2n, 2k)$ we have that $p^{+}p = I_{2k}$.

The symplectic inverse of a matrix A can be expressed explicitly as:

\[A^{+} = \begin{bmatrix} A_{2, 2}^{\mathrm{T}} & -A_{1, 2}^{\mathrm{T}} \\[1.2mm] -A_{2, 1}^{\mathrm{T}} & A_{1, 1}^{\mathrm{T}} \end{bmatrix}.\]

rand(M::SymplecticStiefel; vector_at=nothing,
    hamiltonian_norm=(vector_at === nothing ? 1/2 : 1.0))

Generate a random point $p \in \operatorname{SpSt}(2n, 2k)$ or a random tangent vector $X \in T_p\operatorname{SpSt}(2n, 2k)$ if vector_at is set to a point $p \in \operatorname{Sp}(2n)$.

A random point on $\operatorname{SpSt}(2n, 2k)$ is found by first generating a random point on the symplectic manifold $\operatorname{Sp}(2n)$, and then projecting onto the Symplectic Stiefel manifold using the canonical_project $π_{\operatorname{SpSt}(2n, 2k)}$. That is, $p = π_{\operatorname{SpSt}(2n, 2k)}(p_{\operatorname{Sp}})$.

To generate a random tangent vector in $T_p\operatorname{SpSt}(2n, 2k)$ this code exploits the second tangent vector space parametrization of SymplecticStiefel, showing that any $X \in T_p\operatorname{SpSt}(2n, 2k)$ can be written as $X = pΩ_X + p^sB_X$. To generate random tangent vectors at $p$ then, this function sets $B_X = 0$ and generates a random Hamiltonian matrix $Ω_X \in \mathfrak{sp}(2n,F)$ with Frobenius norm of hamiltonian_norm before returning $X = pΩ_X$.

X = riemannian_gradient(::SymplecticStiefel, f, p, Y; embedding_metric::EuclideanMetric=EuclideanMetric())
riemannian_gradient!(::SymplecticStiefel, f, X, p, Y; embedding_metric::EuclideanMetric=EuclideanMetric())

Compute the riemannian gradient X of f on SymplecticStiefel at a point p, provided that the gradient of the function $\tilde f$, which is f continued into the embedding is given by Y. The metric in the embedding is the Euclidean metric.

The manifold gradient X is computed from Y as

\[ X = Yp^{\mathrm{T}}p + Q_{2n}pY^{\mathrm{T}}Q_{2n}p,\]

where $Q_{2n}$ is the SymplecticMatrix.

canonical_project(::SymplecticStiefel, p_Sp)
canonical_project!(::SymplecticStiefel{n,k}, p, p_Sp)

Define the canonical projection from $\operatorname{Sp}(2n, 2n)$ onto $\operatorname{SpSt}(2n, 2k)$, by projecting onto the first $k$ columns and the $n + 1$'th onto the $n + k$'th columns [BZ21].

It is assumed that the point $p$ is on $\operatorname{Sp}(2n, 2n)$.

get_total_space(::SymplecticStiefel)

Return the total space of the SymplecticStiefel manifold, which is the corresponding Symplectic manifold.

symplectic_inverse_times(::SymplecticStiefel, p, q)
symplectic_inverse_times!(::SymplecticStiefel, A, p, q)

Directly compute the symplectic inverse of $p \in \operatorname{SpSt}(2n, 2k)$, multiplied with $q \in \operatorname{SpSt}(2n, 2k)$. That is, this function efficiently computes $p^+q = (Q_{2k}p^{\mathrm{T}}Q_{2n})q \in ℝ^{2k \times 2k}$, where $Q_{2n}, Q_{2k}$ are the SymplecticMatrix of sizes $2n \times 2n$ and $2k \times 2k$ respectively.

This function performs this common operation without allocating more than a $2k \times 2k$ matrix to store the result in, or in the case of the in-place function, without allocating memory at all.

check_point(M::SymplecticStiefel, p; kwargs...)

Check whether p is a valid point on the SymplecticStiefel, $\operatorname{SpSt}(2n, 2k)$ manifold. That is, the point has the right AbstractNumbers type and $p^{+}p$ is (approximately) the identity, where for $A \in \mathbb{R}^{2n \times 2k}$, $A^{+} = Q_{2k}^{\mathrm{T}}A^{\mathrm{T}}Q_{2n}$ is the symplectic inverse, with

\[Q_{2n} = \begin{bmatrix} 0_n & I_n \\ -I_n & 0_n \end{bmatrix}.\]

The tolerance can be set with kwargs... (e.g. atol = 1.0e-14).

check_vector(M::Symplectic, p, X; kwargs...)

Checks whether X is a valid tangent vector at p on the SymplecticStiefel, $\operatorname{SpSt}(2n, 2k)$ manifold. First recall the definition of the symplectic inverse for $A \in \mathbb{R}^{2n \times 2k}$, $A^{+} = Q_{2k}^{\mathrm{T}}A^{\mathrm{T}}Q_{2n}$ is the symplectic inverse, with

\[ Q_{2n} = \begin{bmatrix} 0_n & I_n \\ -I_n & 0_n \end{bmatrix}.\]

The we check that $H = p^{+}X \in 𝔤_{2k}$, where $𝔤$ is the Lie Algebra of the symplectic group $\operatorname{Sp}(2k)$, characterized as [BZ21],

\[ 𝔤_{2k} = \{H \in ℝ^{2k \times 2k} \;|\; H^+ = -H \}.\]

The tolerance can be set with kwargs... (e.g. atol = 1.0e-14).

inner(M::SymplecticStiefel{n, k}, p, X. Y)

Compute the Riemannian inner product $g^{\operatorname{SpSt}}$ at $p \in \operatorname{SpSt}$ between tangent vectors $X, X \in T_p\operatorname{SpSt}$. Given by Proposition 3.10 in [BZ21].

\[g^{\operatorname{SpSt}}_p(X, Y) = \operatorname{tr}\left(X^{\mathrm{T}}\left(I_{2n} - \frac{1}{2}Q_{2n}^{\mathrm{T}}p(p^{\mathrm{T}}p)^{-1}p^{\mathrm{T}}Q_{2n}\right)Y(p^{\mathrm{T}}p)^{-1}\right).\]

inverse_retract(::SymplecticStiefel, p, q, ::CayleyInverseRetraction)
inverse_retract!(::SymplecticStiefel, q, p, X, ::CayleyInverseRetraction)

Compute the Cayley Inverse Retraction $X = \mathcal{L}_p^{\operatorname{SpSt}}(q)$ such that the Cayley Retraction from $p$ along $X$ lands at $q$, i.e. $\mathcal{R}_p(X) = q$ [BZ21].

First, recall the definition the standard symplectic matrix

\[Q = \begin{bmatrix} 0 & I \\ -I & 0 \end{bmatrix}\]

as well as the symplectic inverse of a matrix $A$, $A^{+} = Q^{\mathrm{T}} A^{\mathrm{T}} Q$.

For $p, q ∈ \operatorname{SpSt}(2n, 2k, ℝ)$ then, we can define the inverse cayley retraction as long as the following matrices exist.

\[ U = (I + p^+ q)^{-1} \in ℝ^{2k \times 2k}, \quad V = (I + q^+ p)^{-1} \in ℝ^{2k \times 2k}.\]

If that is the case, the inverse cayley retration at $p$ applied to $q$ is

\[\mathcal{L}_p^{\operatorname{Sp}}(q) = 2p\bigl(V - U\bigr) + 2\bigl((p + q)U - p\bigr) ∈ T_p\operatorname{Sp}(2n).\]

is_flat(::SymplecticStiefel)

Return false. SymplecticStiefel is not a flat manifold.

manifold_dimension(::SymplecticStiefel{n, k})

Returns the dimension of the symplectic Stiefel manifold embedded in $ℝ^{2n \times 2k}$, i.e. [BZ21]

\[ \operatorname{dim}(\operatorname{SpSt}(2n, 2k)) = (4n - 2k + 1)k.\]

project(::SymplecticStiefel, p, A)
project!(::SymplecticStiefel, Y, p, A)

Given a point $p \in \operatorname{SpSt}(2n, 2k)$, project an element $A \in \mathbb{R}^{2n \times 2k}$ onto the tangent space $T_p\operatorname{SpSt}(2n, 2k)$ relative to the euclidean metric of the embedding $\mathbb{R}^{2n \times 2k}$.

That is, we find the element $X \in T_p\operatorname{SpSt}(2n, 2k)$ which solves the constrained optimization problem

\[ \operatorname{min}_{X \in \mathbb{R}^{2n \times 2k}} \frac{1}{2}||X - A||^2, \quad \text{s.t.}\; h(X)\colon= X^{\mathrm{T}} Q p + p^{\mathrm{T}} Q X = 0,\]

where $h : \mathbb{R}^{2n \times 2k} \rightarrow \operatorname{skew}(2k)$ defines the restriction of $X$ onto the tangent space $T_p\operatorname{SpSt}(2n, 2k)$.

retract(::SymplecticStiefel, p, X, ::CayleyRetraction)
retract!(::SymplecticStiefel, q, p, X, ::CayleyRetraction)

Compute the Cayley retraction on the Symplectic Stiefel manifold, computed inplace of q from p along X.

Given a point $p \in \operatorname{SpSt}(2n, 2k)$, every tangent vector $X \in T_p\operatorname{SpSt}(2n, 2k)$ is of the form $X = \tilde{\Omega}p$, with

\[ \tilde{\Omega} = \left(I_{2n} - \frac{1}{2}pp^+\right)Xp^+ - pX^+\left(I_{2n} - \frac{1}{2}pp^+\right) \in ℝ^{2n \times 2n},\]

as shown in Proposition 3.5 of [BZ21]. Using this representation of $X$, the Cayley retraction on $\operatorname{SpSt}(2n, 2k)$ is defined pointwise as

\[ \mathcal{R}_p(X) = \operatorname{cay}\left(\frac{1}{2}\tilde{\Omega}\right)p.\]

The operator $\operatorname{cay}(A) = (I - A)^{-1}(I + A)$ is the Cayley transform.

However, the computation of an $2n \times 2n$ matrix inverse in the expression above can be reduced down to inverting a $2k \times 2k$ matrix due to Proposition 5.2 of [BZ21].

Let $A = p^+X$ and $H = X - pA$. Then an equivalent expression for the Cayley retraction defined pointwise above is

\[ \mathcal{R}_p(X) = -p + (H + 2p)(H^+H/4 - A/2 + I_{2k})^{-1}.\]

It is this expression we compute inplace of q.