Levenberg-Marquardt

Manopt.LevenbergMarquardtFunction
LevenbergMarquardt(M, F, jacF, x, num_components=-1)

Solve an optimization problem of the form

\[\operatorname{arg\,min}_{p ∈ \mathcal M} \frac{1}{2} \lVert F(p) \rVert^2,\]

where $F\colon\mathcal M \to ℝ^d$ is a continuously differentiable function, using the Riemannian Levenberg-Marquardt algorithm [Peeters1993]. The implementation follows Algorithm 1[Adachi2022].

Input

  • M – a manifold $\mathcal M$
  • F – a cost function $F: \mathcal M→ℝ^d$
  • jacF – the Jacobian of $F$. jacF is supposed to accept a keyword argument basis_domain which specifies basis of the tangent space at a given point in which the Jacobian is to be calculated. By default it should be the DefaultOrthonormalBasis.
  • x – an initial value $x ∈ \mathcal M$
  • num_components – length of the vector returned by the cost function (d). By default its value is -1 which means that it will be determined automatically by calling F one additional time. Only possible when evaluation is AllocatingEvaluation, for mutating evaluation this must be explicitly specified.

Optional

... and the ones that are passed to decorate_state! for decorators.

Output

the obtained (approximate) minimizer $x^*$, see get_solver_return for details

References

source

Options

Manopt.LevenbergMarquardtStateType
LevenbergMarquardtState{P,T} <: AbstractGradientSolverState

Describes a Gradient based descent algorithm, with

Fields

A default value is given in brackets if a parameter can be left out in initialization.

  • x – a point (of type P) on a manifold as starting point
  • stop – (StopAfterIteration(200) | StopWhenGradientNormLess(1e-12) | StopWhenStepsizeLess(1e-12)) a StoppingCriterion
  • retraction_method – (default_retraction_method(M, typeof(p))) the retraction to use, defaults to the default set for your manifold.
  • residual_values – value of $F$ calculated in the solver setup or the previous iteration
  • residual_values_temp – value of $F$ for the current proposal point
  • jacF – the current Jacobian of $F$
  • gradient – the current gradient of $F$
  • step_vector – the tangent vector at x that is used to move to the next point
  • last_stepsize – length of step_vector
  • η – parameter of the algorithm, the higher it is the more likely the algorithm will be to reject new proposal points
  • damping_term – current value of the damping term
  • damping_term_min – initial (and also minimal) value of the damping term
  • β – parameter by which the damping term is multiplied when the current new point is rejected
  • expect_zero_residual – (false) if true, the algorithm expects that the value of residual (objective) at mimimum is equal to 0.

Constructor

LevenbergMarquardtState(M, initialX, initial_residual_values, initial_jacF; initial_vector), kwargs...)

Generate Levenberg-Marquardt options.

See also

gradient_descent, LevenbergMarquardt

source
  • Adachi2022

    S. Adachi, T. Okuno, and A. Takeda, “Riemannian Levenberg-Marquardt Method with Global and Local Convergence Properties.” arXiv, Oct. 01, 2022. arXiv: 2210.00253.

  • Peeters1993

    R. L. M. Peeters, “On a Riemannian version of the Levenberg-Marquardt algorithm,” VU University Amsterdam, Faculty of Economics, Business Administration and Econometrics, Serie Research Memoranda 0011, 1993. link: https://econpapers.repec.org/paper/vuawpaper/1993-11.htm.