The SUNNonlinearSolver_FixedPoint implementation

This section describes the SUNNonlinSol implementation of a fixed point (functional) iteration with optional Anderson acceleration. To access the SUNNonlinSol_FixedPoint module, include the header file sunnonlinsol/sunnonlinsol_fixedpoint.h. We note that the SUNNonlinSol_FixedPoint module is accessible from SUNDIALS integrators without separately linking to the libsundials_sunnonlinsolfixedpoint module library.

SUNNonlinearSolver_FixedPoint description

To find the solution to

(1)\[G(y) = y \,\]

given an initial guess \(y^{(0)}\), the fixed point iteration computes a series of approximate solutions

(2)\[y^{(n+1)} = G(y^{(n)})\]

where \(n\) is the iteration index. The convergence of this iteration may be accelerated using Anderson’s method [A1965], [WN2011], [FS2009], [LWWY2012]. With Anderson acceleration using subspace size \(m\), the series of approximate solutions can be formulated as the linear combination

(3)\[y^{(n+1)} = \sum_{i=0}^{m_n} \alpha_i^{(n)} G(y^{(n-m_n+i)})\]

where \(m_n = \min{\{m,n\}}\) and the factors

\[\alpha^{(n)} =(\alpha_0^{(n)}, \ldots, \alpha_{m_n}^{(n)})\]

solve the minimization problem \(\min_\alpha \| F_n \alpha^T \|_2\) under the constraint that \(\sum_{i=0}^{m_n} \alpha_i = 1\) where

\[F_{n} = (f_{n-m_n}, \ldots, f_{n})\]

with \(f_i = G(y^{(i)}) - y^{(i)}\). Due to this constraint, in the limit of \(m=0\) the accelerated fixed point iteration formula (3) simplifies to the standard fixed point iteration (2).

Following the recommendations made in [WN2011], the SUNNonlinSol_FixedPoint implementation computes the series of approximate solutions as

(4)\[y^{(n+1)} = G(y^{(n)})-\sum_{i=0}^{m_n-1} \gamma_i^{(n)} \Delta g_{n-m_n+i}\]

with \(\Delta g_i = G(y^{(i+1)}) - G(y^{(i)})\) and where the factors

\[\gamma^{(n)} =(\gamma_0^{(n)}, \ldots, \gamma_{m_n-1}^{(n)})\]

solve the unconstrained minimization problem \(\min_\gamma \| f_n - \Delta F_n \gamma^T \|_2\) where

\[\Delta F_{n} = (\Delta f_{n-m_n}, \ldots, \Delta f_{n-1}),\]

with \(\Delta f_i = f_{i+1} - f_i\). The least-squares problem is solved by applying a QR factorization to \(\Delta F_n = Q_n R_n\) and solving \(R_n \gamma = Q_n^T f_n\).

The acceleration subspace size \(m\) is required when constructing the SUNNonlinSol_FixedPoint object. The default maximum number of iterations and the stopping criteria for the fixed point iteration are supplied by the SUNDIALS integrator when SUNNonlinSol_FixedPoint is attached to it. Both the maximum number of iterations and the convergence test function may be modified by the user by calling SUNNonlinSolSetMaxIters() and SUNNonlinSolSetConvTestFn() functions after attaching the SUNNonlinSol_FixedPoint object to the integrator.

SUNNonlinearSolver_FixedPoint functions

The SUNNonlinSol_FixedPoint module provides the following constructor for creating the SUNNonlinearSolver object.

SUNNonlinearSolver SUNNonlinSol_FixedPoint(N_Vector y, int m)

The function SUNNonlinSol_FixedPoint() creates a SUNNonlinearSolver object for use with SUNDIALS integrators to solve nonlinear systems of the form \(G(y) = y\).

Arguments:
  • y – a template for cloning vectors needed within the solver.
  • m – the number of acceleration vectors to use.

Return value: a SUNNonlinSol object if the constructor exits successfully, otherwise it will be NULL.

Since the accelerated fixed point iteration (2) does not require the setup or solution of any linear systems, the SUNNonlinSol_FixedPoint module implements all of the functions defined in sections SUNNonlinearSolver core functions through SUNNonlinearSolver get functions except for the SUNNonlinSolSetup(), SUNNonlinSolSetLSetupFn(), and SUNNonlinSolSetLSolveFn() functions, that are set to NULL. The SUNNonlinSol_FixedPoint functions have the same names as those defined by the generic SUNNonlinSol API with _FixedPoint appended to the function name. Unless using the SUNNonlinSol_FixedPoint module as a standalone nonlinear solver the generic functions defined in sections SUNNonlinearSolver core functions through SUNNonlinearSolver get functions should be called in favor of the SUNNonlinSol_FixedPoint-specific implementations.

The SUNNonlinSol_FixedPoint module also defines the following additional user-callable function.

int SUNNonlinSolGetSysFn_FixedPoint(SUNNonlinearSolver NLS, SUNNonlinSolSysFn *SysFn)

The function SUNNonlinSolGetSysFn_FixedPoint() returns the fixed-point function that defines the nonlinear system.

Arguments:
  • NLS – a SUNNonlinSol object
  • SysFn – the function defining the nonlinear system.

Return value: the return value should be zero for a successful call, and a negative value for a failure.

Notes: This function is intended for users that wish to evaluate the fixed-point function in a custom convergence test function for the SUNNonlinSol_FixedPoint module. We note that SUNNonlinSol_FixedPoint will not leverage the results from any user calls to SysFn.

SUNNonlinearSolver_FixedPoint content

The content field of the SUNNonlinSol_FixedPoint module is the following structure.

struct _SUNNonlinearSolverContent_FixedPoint {

  SUNNonlinSolSysFn      Sys;
  SUNNonlinSolConvTestFn CTest;

  int       m;
  int      *imap;
  realtype *R;
  realtype *gamma;
  realtype *cvals;
  N_Vector *df;
  N_Vector *dg;
  N_Vector *q;
  N_Vector *Xvecs;
  N_Vector  yprev;
  N_Vector  gy;
  N_Vector  fold;
  N_Vector  gold;
  N_Vector  delta;
  int       curiter;
  int       maxiters;
  long int  niters;
  long int  nconvfails;
};

The following entries of the content field are always allocated:

  • Sys – function for evaluating the nonlinear system,
  • CTest – function for checking convergence of the fixed point iteration,
  • yprevN_Vector used to store previous fixed-point iterate,
  • gyN_Vector used to store \(G(y)\) in fixed-point algorithm,
  • deltaN_Vector used to store difference between successive fixed-point iterates,
  • curiter – the current number of iterations in the solve attempt,
  • maxiters – the maximum number of fixed-point iterations allowed in a solve, and
  • niters – the total number of nonlinear iterations across all solves.
  • nconvfails – the total number of nonlinear convergence failures across all solves.
  • m – number of acceleration vectors,

If Anderson acceleration is requested (i.e., \(m>0\) in the call to SUNNonlinSol_FixedPoint()), then the following items are also allocated within the content field:

  • imap – index array used in acceleration algorithm (length m)
  • R – small matrix used in acceleration algorithm (length m*m)
  • gamma – small vector used in acceleration algorithm (length m)
  • cvals – small vector used in acceleration algorithm (length m+1)
  • df – array of N_Vectors used in acceleration algorithm (length m)
  • dg – array of N_Vectors used in acceleration algorithm (length m)
  • q – array of N_Vectors used in acceleration algorithm (length m)
  • XvecsN_Vector pointer array used in acceleration algorithm (length m+1)
  • foldN_Vector used in acceleration algorithm
  • goldN_Vector used in acceleration algorithm

SUNNonlinearSolver_FixedPoint Fortran interface

For SUNDIALS integrators that include a Fortran interface, the SUNNonlinSol_FixedPoint module also includes a Fortran-callable function for creating a SUNNonlinearSolver object.

subroutine FSUNFixedPointInit(CODE, M, IER)

The function FSUNFixedPointInit() can be called for Fortran programs to create a SUNNonlinearSolver object for use with SUNDIALS integrators to solve nonlinear systems of the form \(G(y) = y\).

This routine must be called after the N_Vector object has been initialized.

Arguments:
  • CODE (int, input) – flag denoting the SUNDIALS solver this matrix will be used for: CVODE=1, IDA=2, ARKode=4.
  • M (int, input) – the number of acceleration vectors.
  • IER (int, output) – return flag (0 success, -1 for failure). See printed message for details in case of failure.