Cyclic Proximal Point
The Cyclic Proximal Point (CPP) algorithm is a Proximal Problem.
It aims to minimize
\[F(x) = \sum_{i=1}^c f_i(x)\]
assuming that the proximal maps $\operatorname{prox}_{λ f_i}(x)$ are given in closed form or can be computed efficiently (at least approximately).
The algorithm then cycles through these proximal maps, where the type of cycle might differ and the proximal parameter $λ_k$ changes after each cycle $k$.
For a convergence result on Hadamard manifolds see [Bačák, 2014].
Manopt.cyclic_proximal_point
— Functioncyclic_proximal_point(M, F, proxes, x)
perform a cyclic proximal point algorithm.
Input
M
– a manifold $\mathcal M$F
– a cost function $F:\mathcal M→ℝ$ to minimizeproxes
– an Array of proximal maps (Function
s)(λ,x) -> y
for the summands of $F$x
– an initial value $x ∈ \mathcal M$
Optional
the default values are given in brackets
evaluation
– (AllocatingEvaluation
) specify whether the proximal maps work by allocation (default) formprox(M, λ, x)
orMutatingEvaluation
in place, i.e. is of the formprox!(M, y, λ, x)
.evaluation_order
– (:Linear
) – whether to use a randomly permuted sequence (:FixedRandom
), a per cycle permuted sequence (:Random
) or the default linear one.λ
– (iter -> 1/iter
) a function returning the (square summable but not summable) sequence of λistopping_criterion
– (StopWhenAny
(
StopAfterIteration
(5000),
StopWhenChangeLess
(10.0^-8))
) aStoppingCriterion
.
and the ones that are passed to decorate_options
for decorators.
Output
the obtained (approximate) minimizer $x^*$, see get_solver_return
for details
Manopt.cyclic_proximal_point!
— Functioncyclic_proximal_point!(M, F, proxes, x)
perform a cyclic proximal point algorithm in place of x
.
Input
M
– a manifold $\mathcal M$F
– a cost function $F:\mathcal M→ℝ$ to minimizeproxes
– an Array of proximal maps (Function
s)(λ,x) -> y
for the summands of $F$x
– an initial value $x ∈ \mathcal M$
for all options, see cyclic_proximal_point
.
Options
Manopt.CyclicProximalPointOptions
— TypeCyclicProximalPointOptions <: Options
stores options for the cyclic_proximal_point
algorithm. These are the
Fields
x
– the current iteratestopping_criterion
– aStoppingCriterion
λ
– (@(iter) -> 1/iter) a function for the values of λ_k per iteration/cycleoder_type
– (:LinearOrder
) – whether to use a randomly permuted sequence (:FixedRandomOrder
), a per cycle permuted sequence (RandomOrder
) or the default linear one.
Constructor
CyclicProximalPointOptions(M, p)
Generate the options with the following keyword arguments
stopping_criterion
(StopAfterIteration(2000)
) – aStoppingCriterion
.λ
((iter) -> 1.0 / iter
) a function to compute the $λ_k, k ∈ \mathbb N$,evaluation_order
(:LinearOrder
) a Symbol indicating the order the proxes are applied.
See also
Debug Functions
Manopt.DebugProximalParameter
— TypeDebugProximalParameter <: DebugAction
print the current iterates proximal point algorithm parameter given by Options
s o.λ
.
Record Functions
Manopt.RecordProximalParameter
— TypeRecordProximalParameter <: RecordAction
recoed the current iterates proximal point algorithm parameter given by in Options
s o.λ
.
Literature
- [Bačák, 2014]
Bačák, M:
Computing Medians and Means in Hadamard Spaces. , SIAM Journal on Optimization, Volume 24, Number 3, pp. 1542–1566, doi: 10.1137/140953393, arxiv: 1210.2145.