Stiefel

CaleyRetraction <: AbstractRetractionMethod

A retraction based on the Caley transform, which is realized by using the PadeRetraction{1}.

CaleyVectorTransport <: AbstractVectorTransportMethod

A vector transport that one obtains by differentiating a CaleyRetraction.

PadeRetraction{m} <: AbstractRetractionMethod

A retraction based on the Padé approximation of order $m$

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

The Stiefel manifold consists of all $n × k$, $n ≥ k$ unitary matrices, i.e.

where $𝔽 ∈ \{ℝ, ℂ\}$, $\cdot^{\mathrm{H}}$ denotes the complex conjugate transpose or Hermitian, and $I_k ∈ ℝ^{k × k}$ denotes the $k × k$ identity matrix.

The tangent space at a point $p ∈ \mathcal M$ is given by

where $0_k$ is the $k × k$ zero matrix and $\overline{\cdot}$ the (elementwise) complex conjugate.

This manifold is modeled as an embedded manifold to the Euclidean, i.e. several functions like the inner product and the zero_tangent_vector are inherited from the embedding.

The manifold is named after Eduard L. Stiefel (1909–1978).

Constructor

Stiefel(n, k, field = ℝ)

Generate the (real-valued) Stiefel manifold of $n × k$ dimensional orthonormal matrices.

exp(M::Stiefel, p, X)

Compute the exponential map on the Stiefel{n,k,𝔽}() manifold M emanating from p in tangent direction X.

where $\operatorname{Exp}$ denotes matrix exponential, $\cdot^{\mathrm{H}}$ denotes the complex conjugate transpose or Hermitian, and $I_k$ and $0_k$ are the identity matrix and the zero matrix of dimension $k × k$, respectively.

uniform_distribution(M::Stiefel{n,k,ℝ}, p)

Uniform distribution on given (real-valued) Stiefel M. Specifically, this is the normalized Haar and Hausdorff measure on M. Generated points will be of similar type as p.

The implementation is based on Section 2.5.1 in ^{[Chikuse2003]}; see also Theorem 2.2.1(iii) in ^{[Chikuse2003]}.

check_manifold_point(M::Stiefel, p; kwargs...)

Check whether p is a valid point on the Stiefel M=$\operatorname{St}(n,k)$, i.e. that it has the right AbstractNumbers type and $p^{\mathrm{H}}p$ is (approximately) the identity, where $\cdot^{\mathrm{H}}$ is the complex conjugate transpose. The settings for approximately can be set with kwargs....

check_tangent_vector(M::Stiefel, p, X; check_base_point = true, kwargs...)

Checks whether X is a valid tangent vector at p on the Stiefel M=$\operatorname{St}(n,k)$, i.e. the AbstractNumbers fits and it (approximately) holds that $p^{\mathrm{H}}X + \overline{X^{\mathrm{H}}p} = 0$, where $\cdot^{\mathrm{H}}$ denotes the Hermitian and $\overline{\cdot}$ the (elementwise) complex conjugate. The optional parameter check_base_point indicates, whether to call check_manifold_point for p. The settings for approximately can be set with kwargs....

inverse_retract(M::Stiefel, p, q, ::PolarInverseRetraction)

Compute the inverse retraction based on a singular value decomposition for two points p, q on the Stiefel manifold M. This follows the folloing approach: From the Polar retraction we know that

if such a symmetric positive definite $k × k$ matrix exists. Since $qs - t$ is also a tangent vector at $p$ we obtain

which can either be solved by a Lyapunov approach or a continuous-time algebraic Riccati equation.

This implementation follows the Lyapunov approach.

inverse_retract(M::Stiefel, p, q, ::QRInverseRetraction)

Compute the inverse retraction based on a qr decomposition for two points p, q on the Stiefel manifold M and return the resulting tangent vector in X. The computation follows Algorithm 1 in ^{[KanekoFioriTanaka2013]}.

manifold_dimension(M::Stiefel)

Return the dimension of the Stiefel manifold M=$\operatorname{St}(n,k,𝔽)$. The dimension is given by

project(M::Stiefel,p)

Projects p from the embedding onto the Stiefel M, i.e. compute q as the polar decomposition of $p$ such that $q^{\mathrm{H}q$ is the identity, where $\cdot^{\mathrm{H}}$ denotes the hermitian, i.e. complex conjugate transposed.

project(M::Stiefel, p, X)

Project X onto the tangent space of p to the Stiefel manifold M. The formula reads

where $\operatorname{Sym}(q)$ is the symmetrization of $q$, e.g. by $\operatorname{Sym}(q) = \frac{q^{\mathrm{H}}+q}{2}$.

representation_size(M::Stiefel)

Returns the representation size of the Stiefel M=$\operatorname{St}(n,k)$, i.e. (n,k), which is the matrix dimensions.

retract(::Stiefel, p, X, ::CaleyRetraction)

Compute the retraction on the Stiefel that is based on the Caley transform^[Zhu2016]. Using

the formula reads

It is implemented as the case $m=1$ of the PadeRetraction.

retract(M::Stiefel, p, X, ::PadeRetraction{m})

Compute the retraction on the Stiefel manifold M based on the Padé approximation of order $m$^{[ZhuDuan2018]}. Let $p_m$ and $q_m$ be defined for any matrix $A ∈ ℝ^{n×x}$ as

and

respectively. Then the Padé approximation (of the matrix exponential $\exp(A)$) reads

Defining further

the retraction reads

retract(M::Stiefel, p, X, ::PolarRetraction)

Compute the SVD-based retraction PolarRetraction on the Stiefel manifold M. With $USV = p + X$ the retraction reads

retract(M::Stiefel, p, X, ::QRRetraction)

Compute the QR-based retraction QRRetraction on the Stiefel manifold M. With $QR = p + X$ the retraction reads

where $D$ is a $n × k$ matrix with

where $\operatorname{sgn}(p) = \begin{cases} 1 & \text{ for } p > 0,\\ 0 & \text{ for } p = 0,\\ -1& \text{ for } p < 0. \end{cases}$

vector_transport_direction(::Stiefel, p, X, d, ::CaleyVectorTransport)

Compute the vector transport given by the differentiated retraction of the CaleyRetraction, cf. ^[Zhu2016] Equation (17).

The formula reads

with

Since this is the differentiated retraction as a vector transport, the result will be in the tangent space at $q=\operatorname{retr}_p(d)$ using the CaleyRetraction.