Riemannian quasi-Newton methods

Manopt.quasi_Newton — Function

quasi_Newton(M, f, grad_f, p; kwargs...)
quasi_Newton!(M, f, grad_f, p; kwargs...)

Perform a quasi Newton iteration to solve

\[\operatorname*{arg\,min}_{p ∈ \mathcal{M}} f(p)\]

with start point p. The iterations can be done in-place of p, which is used as $p^{(0)}$. The $k$th iteration consists of

Compute the search direction $η^{(k)} = -\mathcal{B}_k [\operatorname{grad}f (p^{(k)})]$ or solve $\mathcal{H}_k [η^{(k)}] = -\operatorname{grad}f (p^{(k)})$.
Determine a suitable stepsize $α_k$ along the curve $γ(α) = R_{p^{(k)}}(α η^{(k)})$, usually by using WolfePowellLinesearch.
Compute $p^{(k+1)} = R_{p^{(k)}}(α_k η^{(k)})$.
Define $s_k = \mathcal{T}_{p^{(k)}, α_k η^{(k)}}(α_k η^{(k)})$ and $y_k = β_k^{-1}\operatorname{grad}f(p^{(k+1)}) - \mathcal{T}_{p^{(k)}, α_k η^{(k)}}(\operatorname{grad}f(p^{(k)}))$, where $\mathcal{T}$ denotes a vector transport and $β_k = \frac{\lVert α_k η^{(k)} \rVert_{p^{(k)}}}{\lVert s_k \rVert_{p^{(k+1)}}}$ the locking condition scale.
Compute the new approximate Hessian $H_{k+1}$ or its inverse $B_{k+1}$.

Input

M::AbstractManifold: a Riemannian manifold $\mathcal{M}$
f: a cost function $f: \mathcal{M}→ ℝ$ implemented as (M, p) -> v
grad_f: the (Riemannian) gradient $\operatorname{grad}f: \mathcal{M} → T\mathcal{M}$ of f as a function (M, p) -> X or a function (M, X, p) -> X computing X in-place
p::P: a point on the manifold $\mathcal{M}$

Keyword arguments

basis::AbstractBasis=default_basis(M, typeof(p)): basis to use within each of the tangent spaces to represent the Hessian (inverse) for the cases where it is stored in full (matrix) form.
callbacks::D = Dict{Symbol,Function}(): provided callback functions given either as a single function (symbol, problem, state, k) called in every hook or as a (vector of) pairs :hook => function, which are processed by process_callbacks_arg. As key you can either pass single symbol or an array of symbols to indicate a callback should be added in multiple places
cautious_update::Bool=false: whether or not to use the QuasiNewtonCautiousDirectionUpdate which wraps the direction_update.
cautious_function=(x) -> x * 1e-4: a monotone increasing function for the cautious update that is zero at $x=0$ and strictly increasing at $0$
differential = missing: specify a specific function to evaluate the differential. By default, $Df(p)[X] = ⟨\operatorname{grad}f(p),X⟩$ is used
direction_update=InverseBFGS(): the AbstractQuasiNewtonUpdateRule to use.
evaluation::AbstractEvaluationType=AllocatingEvaluation(): specify whether the functions that return an array, for example a point or a tangent vector, work by allocating their result (AllocatingEvaluation) or whether they modify their input argument to return the result therein (InplaceEvaluation). Since usually the first argument is the manifold, the modified argument is the second. For example grad_f(M,p) allocates, but grad_f!(M, X, p) computes the result in-place of X.
initial_operator=Matrix{Float64}(I, n, n): initial matrix to use in case the Hessian (inverse) approximation is stored as a full matrix, that is n=manifold_dimension(M). This matrix is only allocated for the full matrix case. See also initial_scale.
initial_scale=1.0: the scale s in $\frac{s⟨s_k,y_k⟩_{p_k}}{\lVert y_k\rVert_{p_k}^2}$ of the initial Hessian approximation, see also initial_operator. Set to nothing by default when using a preconditioner.
memory_size::Int=min(manifold_dimension(M), 20): limited memory, number of $s_k, y_k$ to store. Set to a negative value to use a full memory (matrix) representation
nondescent_direction_behavior=:reinitialize_direction_update: specify how non-descent direction is handled. This can be
- :step_towards_negative_gradient: the direction is replaced with negative gradient, a message is stored.
- :ignore: the verification is not performed, so any computed direction is accepted. No message is stored.
- :reinitialize_direction_update: discards operator state stored in direction update rules.
- any other value performs the verification, keeps the direction but stores a message.
A stored message can be displayed using DebugMessages.
nonpositive_curvature_behavior=:ignore: how the limited memory update handles stored pairs $s_k, y_k$ with nonpositive inner product, either :ignore or :byrd, see QuasiNewtonLimitedMemoryDirectionUpdate.
preconditioner=missing specify a preconditioner, either
- the default missing does not activate a preconditioning
- a function of the form (M, p, X) -> Y or mutating (M, Y, p, X) -> Y depending on the evaluation
- a QuasiNewtonPreconditioner, see its docs for more details on the preconditioner.
Note that the preconditioner is applied to the gradient, i.e. the right hand side before solving the linear system.
project!=copyto!: for numerical stability it is possible to project onto the tangent space after every iteration. the function has to work inplace of Y, that is (M, Y, p, X) -> Y, where X and Y can be the same memory.
retraction_method::AbstractRetractionMethod=default_retraction_method(M, typeof(p)): a retraction $\operatorname{retr}$ to use, see the section on retractions
stepsize::Stepsize=WolfePowellLinesearch(; retraction_method=retraction_method, vector_transport_method=vector_transport_method, stop_when_stepsize_less=1e-10): a functor inheriting from Stepsize to determine a step size
stopping_criterion::StoppingCriterion=StopAfterIteration(max(1000, memory_size))|StopWhenGradientNormLess(1e-6): a functor indicating that the stopping criterion is fulfilled
sy_tol=1e-8: the tolerance used by the :byrd variant of nonpositive_curvature_behavior, see QuasiNewtonLimitedMemoryDirectionUpdate.
vector_transport_method::AbstractVectorTransportMethod=default_vector_transport_method(M, typeof(p)): a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports

All other keyword arguments are passed to decorate_state! for state decorators or decorate_objective! for objective decorators, respectively.

Output

The obtained approximate minimizer $p^*$. To obtain the whole final state of the solver, see get_solver_return for details, especially the return_state= keyword.

source

Manopt.quasi_Newton! — Function

quasi_Newton(M, f, grad_f, p; kwargs...)
quasi_Newton!(M, f, grad_f, p; kwargs...)

Perform a quasi Newton iteration to solve

\[\operatorname*{arg\,min}_{p ∈ \mathcal{M}} f(p)\]

with start point p. The iterations can be done in-place of p, which is used as $p^{(0)}$. The $k$th iteration consists of

Compute the search direction $η^{(k)} = -\mathcal{B}_k [\operatorname{grad}f (p^{(k)})]$ or solve $\mathcal{H}_k [η^{(k)}] = -\operatorname{grad}f (p^{(k)})$.
Determine a suitable stepsize $α_k$ along the curve $γ(α) = R_{p^{(k)}}(α η^{(k)})$, usually by using WolfePowellLinesearch.
Compute $p^{(k+1)} = R_{p^{(k)}}(α_k η^{(k)})$.
Define $s_k = \mathcal{T}_{p^{(k)}, α_k η^{(k)}}(α_k η^{(k)})$ and $y_k = β_k^{-1}\operatorname{grad}f(p^{(k+1)}) - \mathcal{T}_{p^{(k)}, α_k η^{(k)}}(\operatorname{grad}f(p^{(k)}))$, where $\mathcal{T}$ denotes a vector transport and $β_k = \frac{\lVert α_k η^{(k)} \rVert_{p^{(k)}}}{\lVert s_k \rVert_{p^{(k+1)}}}$ the locking condition scale.
Compute the new approximate Hessian $H_{k+1}$ or its inverse $B_{k+1}$.

Input

M::AbstractManifold: a Riemannian manifold $\mathcal{M}$
f: a cost function $f: \mathcal{M}→ ℝ$ implemented as (M, p) -> v
grad_f: the (Riemannian) gradient $\operatorname{grad}f: \mathcal{M} → T\mathcal{M}$ of f as a function (M, p) -> X or a function (M, X, p) -> X computing X in-place
p::P: a point on the manifold $\mathcal{M}$

Keyword arguments

basis::AbstractBasis=default_basis(M, typeof(p)): basis to use within each of the tangent spaces to represent the Hessian (inverse) for the cases where it is stored in full (matrix) form.
callbacks::D = Dict{Symbol,Function}(): provided callback functions given either as a single function (symbol, problem, state, k) called in every hook or as a (vector of) pairs :hook => function, which are processed by process_callbacks_arg. As key you can either pass single symbol or an array of symbols to indicate a callback should be added in multiple places
cautious_update::Bool=false: whether or not to use the QuasiNewtonCautiousDirectionUpdate which wraps the direction_update.
cautious_function=(x) -> x * 1e-4: a monotone increasing function for the cautious update that is zero at $x=0$ and strictly increasing at $0$
differential = missing: specify a specific function to evaluate the differential. By default, $Df(p)[X] = ⟨\operatorname{grad}f(p),X⟩$ is used
direction_update=InverseBFGS(): the AbstractQuasiNewtonUpdateRule to use.
evaluation::AbstractEvaluationType=AllocatingEvaluation(): specify whether the functions that return an array, for example a point or a tangent vector, work by allocating their result (AllocatingEvaluation) or whether they modify their input argument to return the result therein (InplaceEvaluation). Since usually the first argument is the manifold, the modified argument is the second. For example grad_f(M,p) allocates, but grad_f!(M, X, p) computes the result in-place of X.
initial_operator=Matrix{Float64}(I, n, n): initial matrix to use in case the Hessian (inverse) approximation is stored as a full matrix, that is n=manifold_dimension(M). This matrix is only allocated for the full matrix case. See also initial_scale.
initial_scale=1.0: the scale s in $\frac{s⟨s_k,y_k⟩_{p_k}}{\lVert y_k\rVert_{p_k}^2}$ of the initial Hessian approximation, see also initial_operator. Set to nothing by default when using a preconditioner.
memory_size::Int=min(manifold_dimension(M), 20): limited memory, number of $s_k, y_k$ to store. Set to a negative value to use a full memory (matrix) representation
nondescent_direction_behavior=:reinitialize_direction_update: specify how non-descent direction is handled. This can be
- :step_towards_negative_gradient: the direction is replaced with negative gradient, a message is stored.
- :ignore: the verification is not performed, so any computed direction is accepted. No message is stored.
- :reinitialize_direction_update: discards operator state stored in direction update rules.
- any other value performs the verification, keeps the direction but stores a message.
A stored message can be displayed using DebugMessages.
nonpositive_curvature_behavior=:ignore: how the limited memory update handles stored pairs $s_k, y_k$ with nonpositive inner product, either :ignore or :byrd, see QuasiNewtonLimitedMemoryDirectionUpdate.
preconditioner=missing specify a preconditioner, either
- the default missing does not activate a preconditioning
- a function of the form (M, p, X) -> Y or mutating (M, Y, p, X) -> Y depending on the evaluation
- a QuasiNewtonPreconditioner, see its docs for more details on the preconditioner.
Note that the preconditioner is applied to the gradient, i.e. the right hand side before solving the linear system.
project!=copyto!: for numerical stability it is possible to project onto the tangent space after every iteration. the function has to work inplace of Y, that is (M, Y, p, X) -> Y, where X and Y can be the same memory.
retraction_method::AbstractRetractionMethod=default_retraction_method(M, typeof(p)): a retraction $\operatorname{retr}$ to use, see the section on retractions
stepsize::Stepsize=WolfePowellLinesearch(; retraction_method=retraction_method, vector_transport_method=vector_transport_method, stop_when_stepsize_less=1e-10): a functor inheriting from Stepsize to determine a step size
stopping_criterion::StoppingCriterion=StopAfterIteration(max(1000, memory_size))|StopWhenGradientNormLess(1e-6): a functor indicating that the stopping criterion is fulfilled
sy_tol=1e-8: the tolerance used by the :byrd variant of nonpositive_curvature_behavior, see QuasiNewtonLimitedMemoryDirectionUpdate.
vector_transport_method::AbstractVectorTransportMethod=default_vector_transport_method(M, typeof(p)): a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports

All other keyword arguments are passed to decorate_state! for state decorators or decorate_objective! for objective decorators, respectively.

Output

The obtained approximate minimizer $p^*$. To obtain the whole final state of the solver, see get_solver_return for details, especially the return_state= keyword.

source

Background

The aim is to minimize a real-valued function on a Riemannian manifold, that is

\[\min f(p), \quad p ∈ \mathcal{M}.\]

Riemannian quasi-Newton methods are generalizations of their Euclidean counterparts and are Riemannian line search methods. These methods determine a search direction $η_k ∈ T_{p^{(k)}} \mathcal{M}$ at the current iterate $p^{(k)}$ and a suitable step size $α_k$ along $\gamma(α) = R_{p^{(k)}}(α η_k)$, where $R: T \mathcal{M} →\mathcal{M}$ is a retraction. The next iterate is obtained by

\[p^{(k+1)} = R_{p^{(k)}}(α_k η_k).\]

In quasi-Newton methods, the search direction is given by

\[η_k = -{\mathcal{H}_k}^{-1}[\operatorname{grad}f (p^{(k)})] = -\mathcal{B}_k [\operatorname{grad}f (p^{(k)})],\]

where $\mathcal{H}_k : T_{p^{(k)}} \mathcal{M} →T_{p^{(k)}} \mathcal{M}$ is a positive definite self-adjoint operator, which approximates the action of the Hessian $\operatorname{Hess} f (p^{(k)})[⋅]$ and $\mathcal{B}_k = {\mathcal{H}_k}^{-1}$. The idea of quasi-Newton methods is that, instead of creating a completely new approximation of the Hessian operator $\operatorname{Hess} f(p^{(k+1)})$ or its inverse at every iteration, the previous operator $\mathcal{H}_k$ or $\mathcal{B}_k$ is updated by a convenient formula using the obtained information about the curvature of the objective function during the iteration. The resulting operator $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$ acts on the tangent space $T_{p^{(k+1)}} \mathcal{M}$ of the freshly computed iterate $p^{(k+1)}$. In order to get a well-defined method, the following requirements are placed on the new operator $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$ that is created by an update. Since the Hessian $\operatorname{Hess} f(p^{(k+1)})$ is a self-adjoint operator on the tangent space $T_{p^{(k+1)}} \mathcal{M}$, and $\mathcal{H}_{k+1}$ approximates it, one requirement is that $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$ is also self-adjoint on $T_{p^{(k+1)}} \mathcal{M}$. In order to achieve a steady descent, the next requirement is that $η_k$ is a descent direction in each iteration. Hence a further requirement is that $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$ is a positive definite operator on $T_{p^{(k+1)}} \mathcal{M}$. In order to get information about the curvature of the objective function into the new operator $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$, the last requirement is a form of a Riemannian quasi-Newton equation:

\[\mathcal{H}_{k+1} [T_{p^{(k)} \rightarrow p^{(k+1)}}({R_{p^{(k)}}}^{-1}(p^{(k+1)}))] = \operatorname{grad}f(p^{(k+1)}) - T_{p^{(k)} \rightarrow p^{(k+1)}}(\operatorname{grad}f(p^{(k)}))\]

\[\mathcal{B}_{k+1} [\operatorname{grad}f(p^{(k+1)}) - T_{p^{(k)} \rightarrow p^{(k+1)}}(\operatorname{grad}f(p^{(k)}))] = T_{p^{(k)} \rightarrow p^{(k+1)}}({R_{p^{(k)}}}^{-1}(p^{(k+1)}))\]

where $T_{p^{(k)} \rightarrow p^{(k+1)}} : T_{p^{(k)}} \mathcal{M} →T_{p^{(k+1)}} \mathcal{M}$ and the chosen retraction $R$ is the associated retraction of $T$. Note that, of course, not all updates in all situations meet these conditions in every iteration. For specific quasi-Newton updates, the fulfillment of the Riemannian curvature condition, which requires that

\[g_{p^{(k+1)}}(s_k, y_k) > 0, \quad\text{where } s_k = T_{p^{(k)} \rightarrow p^{(k+1)}}(α_k η_k),\]

holds, is a requirement for the inheritance of the self-adjointness and positive definiteness of the $\mathcal{H}_k$ or $\mathcal{B}_k$ to the operator $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$. Unfortunately, the fulfillment of the Riemannian curvature condition is not given by a step size $\alpha_k > 0$ that satisfies the generalized Wolfe conditions. However, to create a positive definite operator $\mathcal{H}_{k+1}$ or $\mathcal{B}_{k+1}$ in each iteration, the so-called locking condition was introduced in [HGA15], which requires that the isometric vector transport $T^S$, which is used in the update formula, and its associated retraction $R$ fulfill

\[T^{S}_{p, ξ_p}(ξ_p) = β T^{R}_{p, ξ_p}(ξ_p), \quad β = \frac{\lVert ξ_p \rVert_p}{\lVert T^{R}_{p, ξ_p}(ξ_p) \rVert_{R_{p}(ξ_p)}},\]

where $T^R$ is the vector transport by differentiated retraction. With the requirement that the isometric vector transport $T^S$ and its associated retraction $R$ satisfies the locking condition and using the tangent vector

\[y_k = {β_k}^{-1} \operatorname{grad}f(p^{(k+1)}) - T^{S}_{p^{(k)}, α_k η_k}(\operatorname{grad}f(p^{(k)})),\]

where

\[β_k = \frac{\lVert α_k η_k \rVert_{p^{(k)}}}{\lVert T^{R}_{p^{(k)}, α_k η_k}(α_k η_k) \rVert_{p^{(k+1)}}},\]

in the update, it can be shown that choosing a step size $α_k > 0$ that satisfies the Riemannian Wolfe conditions leads to the fulfillment of the Riemannian curvature condition, which in turn implies that the operator generated by the updates is positive definite. In the following the specific operators are denoted in matrix notation and hence use $H_k$ and $B_k$, respectively.

Direction updates

There are different ways to compute the search direction from a fixed AbstractQuasiNewtonUpdateRule. These are represented by

Manopt.AbstractQuasiNewtonDirectionUpdate — Type

AbstractQuasiNewtonDirectionUpdate

An abstract representation of a Quasi Newton update rule to determine the next direction given current QuasiNewtonState.

All subtypes should be functors as well, callable as H(mp, st) and in-place as H(η, mp, st), given an AbstractManoptProblem mp and a QuasiNewtonState st, to compute a new update direction.

source

Manopt.QuasiNewtonMatrixDirectionUpdate — Type

QuasiNewtonMatrixDirectionUpdate <: AbstractQuasiNewtonDirectionUpdate

The QuasiNewtonMatrixDirectionUpdate represent a quasi-Newton update rule, where the operator is stored as a matrix. A distinction is made between the update of the approximation of the Hessian, $H_k ↦ H_{k+1}$, and the update of the approximation of the Hessian inverse, $B_k ↦ B_{k+1}$. For the first case, the coordinates of the search direction $η_k$ with respect to a basis $\{b_i\}_{i=1}^{n}$ are determined by solving a linear system of equations

\[\text{Solve}\quad H_k \hat{η_k} = - \widehat{\operatorname{grad}f(p_k)},\]

where $H_k$ is the matrix representing the operator with respect to the basis $\{b_i\}_{i=1}^{n}$ and $\widehat{\operatorname{grad} f(p_k)}$ represents the coordinates of the gradient of the objective function $f$ in $p_k$ with respect to the basis $\{b_i\}_{i=1}^{n}$. If a method is chosen where Hessian inverse is approximated, the coordinates of the search direction $η_k$ with respect to a basis $\{b_i\}_{i=1}^{n}$ are obtained simply by matrix-vector multiplication

\[\hat{η_k} = - B_k \widehat{\operatorname{grad}f(p_k)},\]

where $B_k$ is the matrix representing the operator with respect to the basis $\{b_i\}_{i=1}^{n}$ and $\widehat{\operatorname{grad} f(p_k)}$ again represents the coordinates of the gradient of $f$ in $p_k$. In the end, the search direction $η_k$ is generated from the coordinates $\hat{η}_k$ and the vectors of the basis $\{b_i\}_{i=1}^{n}$ in both variants. The AbstractQuasiNewtonUpdateRule indicates which quasi-Newton update rule is used. In all of them, the Euclidean update formula is used to generate the matrix $H_{k+1}$ and $B_{k+1}$, and the basis $\{b_i\}_{i=1}^{n}$ is transported into the upcoming tangent space $T_{p_{k+1}}\mathcal{M}$, preferably with an isometric vector transport, or generated there.

Provided functors

(mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction
(η, mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction in-place of η

Fields

basis: an AbstractBasis to use in the tangent spaces
matrix: the matrix which represents the approximating operator.
initial_scale: at the first update the stored matrix of the InverseBFGS, BFGS, InverseDFP and DFP rules is multiplied once by this factor times $⟨s_1,y_1⟩/‖y_1‖^2$, $‖y_1‖^2/⟨s_1,y_1⟩$, $‖s_1‖^2/⟨s_1,y_1⟩$ and $⟨s_1,y_1⟩/‖s_1‖^2$, respectively
update: a AbstractQuasiNewtonUpdateRule.
vector_transport_method::AbstractVectorTransportMethod: a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports

Constructor

QuasiNewtonMatrixDirectionUpdate(    M::AbstractManifold,    update,    basis::B=default_basis(M),    m=Matrix{Float64}(I, manifold_dimension(M), manifold_dimension(M));    kwargs...)

Keyword arguments

initial_scale=1.0 – this can also be deactivated by passing nothing.
vector_transport_method::AbstractVectorTransportMethod=default_vector_transport_method(M, typeof(p)): a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports

Generate the Update rule with defaults from a manifold and the names corresponding to the fields.

source

Manopt.QuasiNewtonLimitedMemoryDirectionUpdate — Type

QuasiNewtonLimitedMemoryDirectionUpdate <: AbstractQuasiNewtonDirectionUpdate

This AbstractQuasiNewtonDirectionUpdate represents the limited-memory Riemannian BFGS update.

The approximating operator is represented by $m$ stored pairs of tangent vectors $\{\widetilde{s}_i\}_{i=k-m}^{k-1}$ and $\{\widetilde{y}_i\}_{i=k-m}^{k-1}$ in the $k$-th iteration. For the calculation of the search direction $X_k$, the generalization of the two-loop recursion is used (see [HGA15]), since it only requires inner products and linear combinations of tangent vectors in $T_{p_k}\mathcal{M}$. For that the stored pairs of tangent vectors $s_i, y_i$, the gradient $\operatorname{grad} f(p_k)$ of the objective function $f$ in $p_k$ and the positive definite self-adjoint operator

\[\mathcal{B}_k^{(0)}[⋅] = \frac{⟨s_{k-1},y_{k-1}⟩_{p_k}}{⟨y_{k-1},y_{k-1}⟩_{p_k}}\mathrm{Id}_{T_{p}\mathcal{M}}[⋅]\]

are used. The two-loop recursion can be understood as that the InverseBFGS update is executed $m$ times in a row on $\mathcal{B}^{(0)}_k[⋅]$ using the tangent vectors $\widehat{s}_i,\widehat{y}_i$, and in the same time the resulting operator $\mathcal{B}^{\mathrm{LRBFGS}}_k [⋅]$ is directly applied on $\operatorname{grad}f(p_k)$. When updating there are two cases: if there is still free memory, $k < m$, the previously stored vector pairs $\widehat{s}_i,\widehat{y}_i$ have to be transported into the upcoming tangent space $T_{p_{k+1}}\mathcal{M}$. If there is no free memory, the oldest pair $\widehat{s}_i,\widehat{y}_i$ has to be discarded and then all the remaining vector pairs $\widehat{s}_i,\widehat{y}_i$ are transported into the tangent space $T_{p_{k+1}}\mathcal{M}$. After that the new values $s_k = \widehat{s}_k = T^{S}_{p_k, α_k η_k}(α_k η_k)$ and $y_k = \widehat{y}_k$ are appended at the end. This process ensures that new information about the objective function is always included and the old, probably no longer relevant, information is discarded.

Provided functors

(mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction
(η, mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction in-place of η

Fields

memory_s: the set of the stored (and transported) search directions times step size $\{\widehat{s}_i\}_{i=k-m}^{k-1}$.
memory_y: set of the stored gradient differences $\{\widehat{y}_i\}_{i=k-m}^{k-1}$.
ξ: a variable used in the two-loop recursion.
ρ: a variable used in the two-loop recursion.
initial_scale: initial scaling of the Hessian, deactivate (e.g. when using a preconditioner) by passing nothing
vector_transport_method::AbstractVectorTransportMethod: a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports
message: a string containing a potential warning that might have appeared
project!: a function to stabilize the update by projecting on the tangent space
nonpositive_curvature_behavior: how non-positive-definite pairs (s, y) are detected and handled in vector transport. Allowed values are: - :ignore (default): pairs whose inner product is zero are omitted from the current Hessian approximation but are retained in memory for further iterations. This may lead to non-positive-definite Hessians and non-descent directions being selected and thus needs to be handled elsewhere. - :byrd: pairs such that inner(M, p, X_s, Y_s) <= sy_tol * norm(M, p, Y_s)^2 are removed from memory (see [BLNZ95], Eq. (3.9) and its discussion).
sy_tol: tolerance for detecting non-positive-definite pairs (Xs, Ys). The pairs may lose positive-definiteness after vector transport.

Constructor

QuasiNewtonLimitedMemoryDirectionUpdate(    M::AbstractManifold,    p,    update::AbstractQuasiNewtonUpdateRule,    memory_size::Int;    initial_vector=zero_vector(M, p),    initial_scale::Real=1.0,    project!=copyto!,    vector_transport_method=default_vector_transport_method(M, typeof(p)),    nonpositive_curvature_behavior::Symbol=:ignore,    sy_tol::Real=1.0e-8,)

source

Manopt.QuasiNewtonCautiousDirectionUpdate — Type

QuasiNewtonCautiousDirectionUpdate <: AbstractQuasiNewtonDirectionUpdate

These AbstractQuasiNewtonDirectionUpdates represent any quasi-Newton update rule, which are based on the idea of a so-called cautious update. The search direction is calculated as given in QuasiNewtonMatrixDirectionUpdate or QuasiNewtonLimitedMemoryDirectionUpdate, but the update then is only executed if

\[\frac{g_{p_{k+1}}(y_k,s_k)}{\lVert s_k \rVert_{p_{k+1}}^{2}} ≥ θ\bigl(\lVert \operatorname{grad}f(p_k) \rVert_{p_k}\bigr),\]

is satisfied, where $θ$ is a monotone increasing function satisfying $θ(0) = 0$ and $θ$ is strictly increasing at $0$. If this is not the case, the corresponding update is skipped, which means that for QuasiNewtonMatrixDirectionUpdate the matrix $H_k$ or $B_k$ is not updated. The basis $\{b_i\}_{i=1}^{n}$ is nevertheless transported into the upcoming tangent space $T_{p_{k+1}} \mathcal{M}$, and for QuasiNewtonLimitedMemoryDirectionUpdate neither the oldest vector pair $\widetilde{s}_{k-m}$, $\widetilde{y}_{k-m}$ is discarded nor the newest vector pair $\widetilde{s}_k, \widetilde{y}_k$ is added into storage, but all stored vector pairs $\{\widetilde{s}_i, \widetilde{y}_i\}_{i=k-m}^{k-1}$ are transported into the tangent space $T_{p_{k+1}} \mathcal{M}$. If InverseBFGS is chosen as update — either within a QuasiNewtonMatrixDirectionUpdate or a QuasiNewtonLimitedMemoryDirectionUpdate — then the resulting method follows the method of [HAG18], taking into account that the corresponding step size is chosen.

Provided functors

(mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction
(η, mp::AbstractManoptProblem, st::QuasiNewtonState) -> η to compute the update direction in-place of η

Fields

update: an AbstractQuasiNewtonDirectionUpdate
θ: a monotone increasing function satisfying $θ(0) = 0$ and $θ$ is strictly increasing at $0$.

Constructor

QuasiNewtonCautiousDirectionUpdate(U::QuasiNewtonMatrixDirectionUpdate; θ = identity)QuasiNewtonCautiousDirectionUpdate(U::QuasiNewtonLimitedMemoryDirectionUpdate; θ = identity)

Generate a cautious update for either a matrix based or a limited memory based update rule.

source

Manopt.initialize_update! — Function

initialize_update!(s::AbstractQuasiNewtonDirectionUpdate)

Initialize direction update. By default no change is made.

source

initialize_update!(d::QuasiNewtonLimitedMemoryDirectionUpdate)

Initialize the limited memory direction update by emptying the memory buffers.

source

Manopt.QuasiNewtonPreconditioner — Type

QuasiNewtonPreconditioner{F}

A preconditioner for the quasi-Newton direction updates: it is applied to the gradient before the quasi-Newton operator is applied (or the linear system is solved), see quasi_Newton.

Fields

preconditioner!::F: the preconditioner function

Constructors

QuasiNewtonPreconditioner(preconditioner)

Wrap preconditioner for use in a QuasiNewtonState.

Input

preconditioner: preconditioner function, either as a (M, p, X) -> Y allocating or (M, Y, p, X) -> Y mutating function

Keyword arguments

evaluation::AbstractEvaluationType=AllocatingEvaluation(): specify whether the functions that return an array, for example a point or a tangent vector, work by allocating their result (AllocatingEvaluation) or whether they modify their input argument to return the result therein (InplaceEvaluation). Since usually the first argument is the manifold, the modified argument is the second.

source

Manopt.QuasiNewtonLimitedMemoryBoxDirectionUpdate — Type

QuasiNewtonLimitedMemoryBoxDirectionUpdate <: AbstractQuasiNewtonDirectionUpdate

A limited memory quasi-Newton direction update with support for box constraints.

It stores an approximation of the Hessian of a scalar function in the compact form $B_k = θ_k I - W_k M_k W_k^{\mathrm{T}}$, where $θ_k > 0$ is the current scaling factor stored in current_scale; while the memory is empty, $B_k = θ^{-1} I$ holds for the initial scaling guess $θ > 0$. Matrix $M_k = \left(\begin{smallmatrix}M₁₁ & M₂₁^{\mathrm{T}}\\ M₂₁ & M₂₂\end{smallmatrix}\right)$ is stored using its blocks. Blocks $W_k$ are (implicitly) composed from memory_y and memory_s stored in qn_du of type QuasiNewtonLimitedMemoryDirectionUpdate.

Initial scale $θ$ is the initial_scale of the wrapped update qn_du; if the memory is not empty, the current scale is set to $\frac{\|y_k\|^2}{⟨s_k, y_k⟩ θ}$, where $k$ is the most recent index for which $⟨s_k, y_k⟩$ is not equal to 0.

last_gcd_result and last_gcd_stepsize store the status and the maximal step size returned by the last generalized Cauchy direction search, see find_generalized_cauchy_direction!.

Constructor

QuasiNewtonLimitedMemoryBoxDirectionUpdate(qn_du::QuasiNewtonLimitedMemoryDirectionUpdate)

See [BNS94] for details.

source

Hessian update rules

Using

Manopt.update_hessian! — Function

update_hessian!(gh::QuasiNewtonLimitedMemoryBoxDirectionUpdate, mp, st, p_old, k)

Update the Hessian approximation gh by moving it from the previous point p_old to the current iterate and updating the stored s and y vectors, respectively.

source

update_hessian!(d::AbstractQuasiNewtonDirectionUpdate, amp, st, p_old, k)

update the Hessian within the QuasiNewtonState st given an AbstractManoptProblem amp as well as an AbstractQuasiNewtonDirectionUpdate d and the last iterate p_old. Note that the current (kth) iterate is already stored in get_iterate(st).

See also AbstractQuasiNewtonUpdateRule and its subtypes for the different rules that are available within d.

source

the following update formulae for either $H_{k+1}$ or $B_{k+1}$ are available.

Manopt.AbstractQuasiNewtonUpdateRule — Type

AbstractQuasiNewtonUpdateRule

Specify a type for the different AbstractQuasiNewtonDirectionUpdates.

For a QuasiNewtonMatrixDirectionUpdate there are several different updates to the matrix, while for the QuasiNewtonLimitedMemoryDirectionUpdate the default and most prominent one is InverseBFGS.

source

Manopt.BFGS — Type

BFGS <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the Riemannian BFGS update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{H}_k^\mathrm{BFGS}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[H^\mathrm{BFGS}_{k+1} = \widetilde{H}^\mathrm{BFGS}_k + \frac{y_k y^{\mathrm{T}}_k }{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{H}^\mathrm{BFGS}_k s_k s^{\mathrm{T}}_k \widetilde{H}^\mathrm{BFGS}_k }{ s^{\mathrm{T}}_k \widetilde{H}^\mathrm{BFGS}_k s_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{ and }\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

source

Manopt.DFP — Type

DFP <: AbstractQuasiNewtonUpdateRule

indicates in an AbstractQuasiNewtonDirectionUpdate that the Riemannian DFP update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{H}_k^\mathrm{DFP}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[H^\mathrm{DFP}_{k+1} = \Bigl( \mathrm{Id}_{T_{p_{k+1}} \mathcal{M}} - \frac{y_k s^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k} \Bigr) \widetilde{H}^\mathrm{DFP}_k \Bigl( \mathrm{Id}_{T_{p_{k+1}} \mathcal{M}} - \frac{s_k y^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k} \Bigr) + \frac{y_k y^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

source

Manopt.Broyden — Type

Broyden <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the Riemannian Broyden update is used in the Riemannian quasi-Newton method, which is a convex combination of BFGS and DFP.

Denote by $\widetilde{H}_k^\mathrm{Br}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[H^{\mathrm{Br}}_{k+1} = \widetilde{H}^{\mathrm{Br}}_k - \frac{\widetilde{H}^{\mathrm{Br}}_k s_k s^{\mathrm{T}}_k \widetilde{H}^{\mathrm{Br}}_k}{s^{\mathrm{T}}_k \widetilde{H}^{\mathrm{Br}}_k s_k} + \frac{y_k y^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k} + φ_k s^{\mathrm{T}}_k \widetilde{H}^{\mathrm{Br}}_k s_k \Bigl( \frac{y_k}{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{H}^{\mathrm{Br}}_k s_k}{s^{\mathrm{T}}_k \widetilde{H}^{\mathrm{Br}}_k s_k} \Bigr) \Bigl( \frac{y_k}{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{H}^{\mathrm{Br}}_k s_k}{s^{\mathrm{T}}_k \widetilde{H}^{\mathrm{Br}}_k s_k} \Bigr)^{\mathrm{T}}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively, and $φ_k$ is the Broyden factor which is :constant by default but can also be set to :Davidon.

Constructor

Broyden(φ, update_rule::Symbol = :constant)

source

Manopt.SR1 — Type

SR1 <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the Riemannian SR1 update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{H}_k^{\mathrm{SR1}}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[H^{\mathrm{SR1}}_{k+1} = \widetilde{H}^{\mathrm{SR1}}_k + \frac{(y_k - \widetilde{H}^{\mathrm{SR1}}_k s_k) (y_k - \widetilde{H}^{\mathrm{SR1}}_k s_k)^{\mathrm{T}}}{(y_k - \widetilde{H}^{\mathrm{SR1}}_k s_k)^{\mathrm{T}} s_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

This method can be stabilized by only performing the update if the absolute value of the denominator is at least $r\lVert s_k \rVert_{p_{k+1}}\lVert y_k - \widetilde{H}^{\mathrm{SR1}}_k s_k \rVert_{p_{k+1}}$ for some $r>0$. For more details, see Section 6.2 in [NW06].

Constructor

SR1(r::Float64=-1.0)

Generate the SR1 update.

source

Manopt.InverseBFGS — Type

InverseBFGS <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the inverse Riemannian BFGS update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{B}_k^\mathrm{BFGS}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[B^\mathrm{BFGS}_{k+1} = \Bigl( \mathrm{Id}_{T_{p_{k+1}} \mathcal{M}} - \frac{s_k y^{\mathrm{T}}_k }{s^{\mathrm{T}}_k y_k} \Bigr) \widetilde{B}^\mathrm{BFGS}_k \Bigl( \mathrm{Id}_{T_{p_{k+1}} \mathcal{M}} - \frac{y_k s^{\mathrm{T}}_k }{s^{\mathrm{T}}_k y_k} \Bigr) + \frac{s_k s^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

source

Manopt.InverseDFP — Type

InverseDFP <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the inverse Riemannian DFP update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{B}_k^\mathrm{DFP}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[B^\mathrm{DFP}_{k+1} = \widetilde{B}^\mathrm{DFP}_k + \frac{s_k s^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{B}^\mathrm{DFP}_k y_k y^{\mathrm{T}}_k \widetilde{B}^\mathrm{DFP}_k}{y^{\mathrm{T}}_k \widetilde{B}^\mathrm{DFP}_k y_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

source

Manopt.InverseBroyden — Type

InverseBroyden <: AbstractQuasiNewtonUpdateRule

Indicates in AbstractQuasiNewtonDirectionUpdate that the Riemannian Broyden update is used in the Riemannian quasi-Newton method, which is a convex combination of InverseBFGS and InverseDFP.

Denote by $\widetilde{B}_k^{\mathrm{Br}}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[B^{\mathrm{Br}}_{k+1} = \widetilde{B}^{\mathrm{Br}}_k - \frac{\widetilde{B}^{\mathrm{Br}}_k y_k y^{\mathrm{T}}_k \widetilde{B}^{\mathrm{Br}}_k}{y^{\mathrm{T}}_k \widetilde{B}^{\mathrm{Br}}_k y_k} + \frac{s_k s^{\mathrm{T}}_k}{s^{\mathrm{T}}_k y_k} + φ_k y^{\mathrm{T}}_k \widetilde{B}^{\mathrm{Br}}_k y_k \Bigl( \frac{s_k}{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{B}^{\mathrm{Br}}_k y_k}{y^{\mathrm{T}}_k \widetilde{B}^{\mathrm{Br}}_k y_k} \Bigr) \Bigl( \frac{s_k}{s^{\mathrm{T}}_k y_k} - \frac{\widetilde{B}^{\mathrm{Br}}_k y_k}{y^{\mathrm{T}}_k \widetilde{B}^{\mathrm{Br}}_k y_k} \Bigr)^{\mathrm{T}}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively, and $φ_k$ is the Broyden factor which is :constant by default but can also be set to :InverseDavidon.

Constructor

InverseBroyden(φ, update_rule::Symbol = :constant)

source

Manopt.InverseSR1 — Type

InverseSR1 <: AbstractQuasiNewtonUpdateRule

indicates in AbstractQuasiNewtonDirectionUpdate that the inverse Riemannian SR1 update is used in the Riemannian quasi-Newton method.

Denote by $\widetilde{B}_k^{\mathrm{SR1}}$ the operator concatenated with a vector transport and its inverse before and after to act on $p_{k+1} = R_{p_k}(α_k η_k)$. Then the update formula reads

\[B^{\mathrm{SR1}}_{k+1} = \widetilde{B}^{\mathrm{SR1}}_k + \frac{(s_k - \widetilde{B}^{\mathrm{SR1}}_k y_k) (s_k - \widetilde{B}^{\mathrm{SR1}}_k y_k)^{\mathrm{T}}}{(s_k - \widetilde{B}^{\mathrm{SR1}}_k y_k)^{\mathrm{T}} y_k}\]

where $s_k$ and $y_k$ are the coordinate vectors with respect to the current basis (from QuasiNewtonState) of

\[T^{S}_{p_k, α_k η_k}(α_k η_k) \quad\text{and}\quad \operatorname{grad}f(p_{k+1}) - T^{S}_{p_k, α_k η_k}(\operatorname{grad}f(p_k)) ∈ T_{p_{k+1}} \mathcal{M},\]

respectively.

This method can be stabilized by only performing the update if the absolute value of the denominator is at least $r\lVert y_k \rVert_{p_{k+1}}\lVert s_k - \widetilde{B}^{\mathrm{SR1}}_k y_k \rVert_{p_{k+1}}$ for some $r>0$. For more details, see Section 6.2 in [NW06].

Constructor

InverseSR1(r::Float64=-1.0)

Generate the InverseSR1.

source

State

The quasi-Newton algorithm is based on a DefaultManoptProblem.

Manopt.QuasiNewtonState — Type

QuasiNewtonState <: AbstractManoptSolverState

The AbstractManoptSolverState represents any quasi-Newton based method and stores all necessary fields.

Fields

callbacks::D: provided callback functions given as a dictionary with symbols as keys
direction_update: an AbstractQuasiNewtonDirectionUpdate rule.
η: the current update direction
nondescent_direction_behavior: a Symbol to specify how to handle directions that are not descent ones.
nondescent_direction_value: the value from the last inner product from checking for descent directions
p::P: a point on the manifold $\mathcal{M}$ storing the current iterate
p_old: the last iterate
preconditioner: a QuasiNewtonPreconditioner
sk: the current step
retraction_method::AbstractRetractionMethod: a retraction $\operatorname{retr}$ to use, see the section on retractions
stepsize::Stepsize: a functor inheriting from Stepsize to determine a step size
stop::StoppingCriterion: a functor indicating that the stopping criterion is fulfilled
X::T: a tangent vector at the point $p$ on the manifold $\mathcal{M}$ storing the gradient at the current iterate
vector_transport_method::AbstractVectorTransportMethod: a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports
X_old: the last gradient
yk: the current gradient difference

Constructor

QuasiNewtonState(M::AbstractManifold; kwargs...)

Generate the Quasi Newton state on the manifold M.

Keyword arguments

callbacks::D = Dict{Symbol,Function}(): provided callback functions given as a dictionary with symbols as keys
direction_update=QuasiNewtonLimitedMemoryDirectionUpdate(M, p, InverseBFGS(), memory_size; vector_transport_method=vector_transport_method, initial_scale=initial_scale)
stopping_criterion::StoppingCriterion=StopAfterIteration(1000)|StopWhenGradientNormLess(1e-6): a functor indicating that the stopping criterion is fulfilled
initial_scale=1.0: a relative initial scale. By default deactivated when using a preconditioner.
memory_size=20: a shortcut to set the memory in the default direction update
nondescent_direction_behavior=:reinitialize_direction_update: how to handle a direction that is not a descent direction, see quasi_Newton
p::P =rand(M): a point on the manifold $\mathcal{M}$ to specify the initial value
preconditioner::Union{QuasiNewtonPreconditioner, Missing} = missing specify a preconditioner or deactivate by passing missing.
retraction_method::AbstractRetractionMethod=default_retraction_method(M, typeof(p)): a retraction $\operatorname{retr}$ to use, see the section on retractions
stepsize::Stepsize=default_stepsize(M,QuasiNewtonState; retraction_method=retraction_method, vector_transport_method=vector_transport_method): a functor inheriting from Stepsize to determine a step size
vector_transport_method::AbstractVectorTransportMethod=default_vector_transport_method(M, typeof(p)): a vector transport $\mathcal T_{⋅←⋅}$ to use, see the section on vector transports
X::T =zero_vector(M, p): a tangent vector at the point $p$ on the manifold $\mathcal{M}$ to specify the representation of a tangent vector

See also

quasi_Newton

source

Technical details

The quasi_Newton solver requires the following functions of a manifold to be available

A retract!(M, q, p, X); it is recommended to set the default_retraction_method to a favorite retraction. If this default is set, a retraction_method= does not have to be specified.
A vector_transport_to!(M, Y, p, X, q); it is recommended to set the default_vector_transport_method to a favorite vector transport. If this default is set, a vector_transport_method= does not have to be specified.
By default quasi-Newton uses WolfePowellLinesearch which requires max_stepsize(M) to be set and an implementation of inner(M, p, X, Y).
the norm as well, to stop when the norm of the gradient is small, but if you implemented inner, the norm is provided already.
A copyto!(M, q, p) and copy(M,p) for points and similarly copy(M, p, X) for tangent vectors.
By default the tangent vector storing the gradient is initialized calling zero_vector(M,p).
If you do not provide a memory_size=, a manifold_dimension is required for its default; the full-matrix variant (memory_size < 0) needs it for the default initial_operator= as well.

Most Hessian approximations further require get_coordinates(M, p, X, b) and get_vector!(M, X, p, c, b) with respect to the AbstractBasis b provided, which is default_basis(M, typeof(p)) by default from the basis= keyword.

Literature

[BLNZ95]: R. H. Byrd, P. Lu, J. Nocedal and C. Zhu. A Limited Memory Algorithm for Bound Constrained Optimization. SIAM Journal on Scientific Computing 16, 1190–1208 (1995). Publisher: Society for Industrial and Applied Mathematics.
[BNS94]: R. H. Byrd, J. Nocedal and R. B. Schnabel. Representations of quasi-Newton matrices and their use in limited memory methods. Mathematical Programming 63, 129–156 (1994).
[HAG18]: W. Huang, P.-A. Absil and K. A. Gallivan. A Riemannian BFGS method without differentiated retraction for nonconvex optimization problems. SIAM Journal on Optimization 28, 470–495 (2018).
[HGA15]: W. Huang, K. A. Gallivan and P.-A. Absil. A Broyden class of quasi-Newton methods for Riemannian optimization. SIAM Journal on Optimization 25, 1660–1685 (2015).
[NW06]: J. Nocedal and S. J. Wright. Numerical Optimization. 2 Edition (Springer, New York, 2006).