Go to the first, previous, next, last section, table of contents.
This chapter describes functions for multidimensional root-finding (solving nonlinear systems with n equations in n unknowns). The library provides low level components for a variety of iterative solvers and convergence tests. These can be combined by the user to achieve the desired solution, with full access to the intermediate steps of the iteration. Each class of methods uses the same framework, so that you can switch between solvers at runtime without needing to recompile your program. Each instance of a solver keeps track of its own state, allowing the solvers to be used in multi-threaded programs. The solvers are based on the original Fortran library MINPACK.
The header file `gsl_multiroots.h' contains prototypes for the multidimensional root finding functions and related declarations.
The problem of multidimensional root finding requires the simultaneous solution of n equations, f_i, in n variables, x_i,
f_i (x_1, ..., x_n) = 0 for i = 1 ... n.
In general there are no bracketing methods available for n dimensional systems, and no way of knowing whether any solutions exist. All algorithms proceed from an initial guess using a variant of the Newton iteration,
x -> x' = x - J^{-1} f(x)
where x, f are vector quantities and J is the Jacobian matrix J_{ij} = d f_i / d x_j. Additional strategies can be used to enlarge the region of convergence. These include requiring a decrease in the norm |f| on each step proposed by Newton's method, or taking steepest-descent steps in the direction of the negative gradient of |f|.
Several root-finding algorithms are available within a single framework. The user provides a high-level driver for the algorithms, and the library provides the individual functions necessary for each of the steps. There are three main phases of the iteration. The steps are,
The evaluation of the Jacobian matrix can be problematic, either because programming the derivatives is intractable or because computation of the n^2 terms of the matrix becomes too expensive. For these reasons the algorithms provided by the library are divided into two classes according to whether the derivatives are available or not.
The state for solvers with an analytic Jacobian matrix is held in a
gsl_multiroot_fdfsolver
struct. The updating procedure requires
both the function and its derivatives to be supplied by the user.
The state for solvers which do not use an analytic Jacobian matrix is
held in a gsl_multiroot_fsolver
struct. The updating procedure
uses only function evaluations (not derivatives). The algorithms
estimate the matrix J or
J^{-1} by approximate methods.
The following functions initialize a multidimensional solver, either with or without derivatives. The solver itself depends only on the dimension of the problem and the algorithm and can be reused for different problems.
const gsl_multiroot_fsolver_type * T = gsl_multiroot_fsolver_hybrid; gsl_multiroot_fsolver * s = gsl_multiroot_fsolver_alloc (T, 3);
If there is insufficient memory to create the solver then the function
returns a null pointer and the error handler is invoked with an error
code of GSL_ENOMEM
.
const gsl_multiroot_fdfsolver_type * T = gsl_multiroot_fdfsolver_newton; gsl_multiroot_fdfsolver * s = gsl_multiroot_fdfsolver_alloc (T, 2);
If there is insufficient memory to create the solver then the function
returns a null pointer and the error handler is invoked with an error
code of GSL_ENOMEM
.
printf ("s is a '%s' solver\n", gsl_multiroot_fdfsolver_name (s));
would print something like s is a 'newton' solver
.
You must provide n functions of n variables for the root finders to operate on. In order to allow for general parameters the functions are defined by the following data types:
int (* f) (const gsl_vector * x, void * params, gsl_vector * f)
size_t n
void * params
Here is an example using Powell's test function,
f_1(x) = A x_0 x_1 - 1, f_2(x) = exp(-x_0) + exp(-x_1) - (1 + 1/A)
with A = 10^4. The following code defines a
gsl_multiroot_function
system F
which you could pass to a
solver:
struct powell_params { double A; }; int powell (gsl_vector * x, void * p, gsl_vector * f) { struct powell_params * params = *(struct powell_params *)p; const double A = (params->A); const double x0 = gsl_vector_get(x,0); const double x1 = gsl_vector_get(x,1); gsl_vector_set (f, 0, A * x0 * x1 - 1); gsl_vector_set (f, 1, (exp(-x0) + exp(-x1) - (1.0 + 1.0/A))); return GSL_SUCCESS } gsl_multiroot_function F; struct powell_params params = { 10000.0 }; F.f = &powell; F.n = 2; F.params = ¶ms;
int (* f) (const gsl_vector * x, void * params, gsl_vector * f)
int (* df) (const gsl_vector * x, void * params, gsl_matrix * J)
int (* fdf) (const gsl_vector * x, void * params, gsl_vector * f, gsl_matrix * J)
size_t n
void * params
The example of Powell's test function defined above can be extended to include analytic derivatives using the following code,
int powell_df (gsl_vector * x, void * p, gsl_matrix * J) { struct powell_params * params = *(struct powell_params *)p; const double A = (params->A); const double x0 = gsl_vector_get(x,0); const double x1 = gsl_vector_get(x,1); gsl_matrix_set (J, 0, 0, A * x1); gsl_matrix_set (J, 0, 1, A * x0); gsl_matrix_set (J, 1, 0, -exp(-x0)); gsl_matrix_set (J, 1, 1, -exp(-x1)); return GSL_SUCCESS } int powell_fdf (gsl_vector * x, void * p, gsl_matrix * f, gsl_matrix * J) { struct powell_params * params = *(struct powell_params *)p; const double A = (params->A); const double x0 = gsl_vector_get(x,0); const double x1 = gsl_vector_get(x,1); const double u0 = exp(-x0); const double u1 = exp(-x1); gsl_vector_set (f, 0, A * x0 * x1 - 1); gsl_vector_set (f, 1, u0 + u1 - (1 + 1/A)); gsl_matrix_set (J, 0, 0, A * x1); gsl_matrix_set (J, 0, 1, A * x0); gsl_matrix_set (J, 1, 0, -u0); gsl_matrix_set (J, 1, 1, -u1); return GSL_SUCCESS } gsl_multiroot_function_fdf FDF; FDF.f = &powell_f; FDF.df = &powell_df; FDF.fdf = &powell_fdf; FDF.n = 2; FDF.params = 0;
Note that the function powell_fdf
is able to reuse existing terms
from the function when calculating the Jacobian, thus saving time.
The following functions drive the iteration of each algorithm. Each function performs one iteration to update the state of any solver of the corresponding type. The same functions work for all solvers so that different methods can be substituted at runtime without modifications to the code.
GSL_EBADFUNC
Inf
or NaN
.
GSL_ENOPROG
The solver maintains a current best estimate of the root at all times. This information can be accessed with the following auxiliary functions,
A root finding procedure should stop when one of the following conditions is true:
The handling of these conditions is under user control. The functions below allow the user to test the precision of the current result in several standard ways.
This function tests for the convergence of the sequence by comparing the
last step dx with the absolute error epsabs and relative
error epsrel to the current position x. The test returns
GSL_SUCCESS
if the following condition is achieved,
|dx_i| < epsabs + epsrel |x_i|
for each component of x and returns GSL_CONTINUE
otherwise.
GSL_SUCCESS
if the
following condition is achieved,
\sum_i |f_i| < epsabs
and returns GSL_CONTINUE
otherwise. This criterion is suitable
for situations where the precise location of the root, x, is
unimportant provided a value can be found where the residual is small
enough.
The root finding algorithms described in this section make use of both the function and its derivative. They require an initial guess for the location of the root, but there is no absolute guarantee of convergence -- the function must be suitable for this technique and the initial guess must be sufficiently close to the root for it to work. When the conditions are satisfied then convergence is quadratic.
The algorithm uses a generalized trust region to keep each step under control. In order to be accepted a proposed new position x' must satisfy the condition |D (x' - x)| < \delta, where D is a diagonal scaling matrix and \delta is the size of the trust region. The components of D are computed internally, using the column norms of the Jacobian to estimate the sensitivity of the residual to each component of x. This improves the behavior of the algorithm for badly scaled functions.
On each iteration the algorithm first determines the standard Newton step by solving the system J dx = - f. If this step falls inside the trust region it is used as a trial step in the next stage. If not, the algorithm uses the linear combination of the Newton and gradient directions which is predicted to minimize the norm of the function while staying inside the trust region.
dx = - \alpha J^{-1} f(x) - \beta \nabla |f(x)|^2
This combination of Newton and gradient directions is referred to as a dogleg step.
The proposed step is now tested by evaluating the function at the resulting point, x'. If the step reduces the norm of the function sufficiently then it is accepted and size of the trust region is increased. If the proposed step fails to improve the solution then the size of the trust region is decreased and another trial step is computed.
The speed of the algorithm is increased by computing the changes to the Jacobian approximately, using a rank-1 update. If two successive attempts fail to reduce the residual then the full Jacobian is recomputed. The algorithm also monitors the progress of the solution and returns an error if several steps fail to make any improvement,
GSL_ENOPROG
GSL_ENOPROGJ
hybridsj
. The steps are
controlled by a spherical trust region |x' - x| < \delta, instead
of a generalized region. This can be useful if the generalized region
estimated by hybridsj
is inappropriate.
Newton's Method is the standard root-polishing algorithm. The algorithm begins with an initial guess for the location of the solution. On each iteration a linear approximation to the function F is used to estimate the step which will zero all the components of the residual. The iteration is defined by the following sequence,
x -> x' = x - J^{-1} f(x)
where the Jacobian matrix J is computed from the derivative functions provided by f. The step dx is obtained by solving the linear system,
J dx = - f(x)
using LU decomposition.
t = (\sqrt(1 + 6 r) - 1) / (3 r)
is proposed, with r being the ratio of norms |f(x')|^2/|f(x)|^2. This procedure is repeated until a suitable step size is found.
The algorithms described in this section do not require any derivative information to be supplied by the user. Any derivatives needed are approximated from by finite difference.
gsl_multiroots_fdjac
with a relative step size of GSL_SQRT_DBL_EPSILON
.
The discrete Newton algorithm is the simplest method of solving a multidimensional system. It uses the Newton iteration
x -> x - J^{-1} f(x)
where the Jacobian matrix J is approximated by taking finite differences of the function f. The approximation scheme used by this implementation is,
J_{ij} = (f_i(x + \delta_j) - f_i(x)) / \delta_j
where \delta_j is a step of size \sqrt\epsilon |x_j| with \epsilon being the machine precision ( \epsilon \approx 2.22 \times 10^-16). The order of convergence of Newton's algorithm is quadratic, but the finite differences require n^2 function evaluations on each iteration. The algorithm may become unstable if the finite differences are not a good approximation to the true derivatives.
The Broyden algorithm is a version of the discrete Newton algorithm which attempts to avoids the expensive update of the Jacobian matrix on each iteration. The changes to the Jacobian are also approximated, using a rank-1 update,
J^{-1} \to J^{-1} - (J^{-1} df - dx) dx^T J^{-1} / dx^T J^{-1} df
where the vectors dx and df are the changes in x and f. On the first iteration the inverse Jacobian is estimated using finite differences, as in the discrete Newton algorithm. This approximation gives a fast update but is unreliable if the changes are not small, and the estimate of the inverse Jacobian becomes worse as time passes. The algorithm has a tendency to become unstable unless it starts close to the root. The Jacobian is refreshed if this instability is detected (consult the source for details).
This algorithm is not recommended and is included only for demonstration purposes.
The multidimensional solvers are used in a similar way to the
one-dimensional root finding algorithms. This first example
demonstrates the hybrids
scaled-hybrid algorithm, which does not
require derivatives. The program solves the Rosenbrock system of equations,
f_1 (x, y) = a (1 - x) f_2 (x, y) = b (y - x^2)
with a = 1, b = 10. The solution of this system lies at (x,y) = (1,1) in a narrow valley.
The first stage of the program is to define the system of equations,
#include <stdlib.h> #include <stdio.h> #include <gsl/gsl_vector.h> #include <gsl/gsl_multiroots.h> struct rparams { double a; double b; }; int rosenbrock_f (const gsl_vector * x, void *params, gsl_vector * f) { double a = ((struct rparams *) params)->a; double b = ((struct rparams *) params)->b; const double x0 = gsl_vector_get (x, 0); const double x1 = gsl_vector_get (x, 1); const double y0 = a * (1 - x0); const double y1 = b * (x1 - x0 * x0); gsl_vector_set (f, 0, y0); gsl_vector_set (f, 1, y1); return GSL_SUCCESS; }
The main program begins by creating the function object f
, with
the arguments (x,y)
and parameters (a,b)
. The solver
s
is initialized to use this function, with the hybrids
method.
int main (void) { const gsl_multiroot_fsolver_type *T; gsl_multiroot_fsolver *s; int status; size_t i, iter = 0; const size_t n = 2; struct rparams p = {1.0, 10.0}; gsl_multiroot_function f = {&rosenbrock_f, n, &p}; double x_init[2] = {-10.0, -5.0}; gsl_vector *x = gsl_vector_alloc (n); gsl_vector_set (x, 0, x_init[0]); gsl_vector_set (x, 1, x_init[1]); T = gsl_multiroot_fsolver_hybrids; s = gsl_multiroot_fsolver_alloc (T, 2); gsl_multiroot_fsolver_set (s, &f, x); print_state (iter, s); do { iter++; status = gsl_multiroot_fsolver_iterate (s); print_state (iter, s); if (status) /* check if solver is stuck */ break; status = gsl_multiroot_test_residual (s->f, 1e-7); } while (status == GSL_CONTINUE && iter < 1000); printf ("status = %s\n", gsl_strerror (status)); gsl_multiroot_fsolver_free (s); gsl_vector_free (x); return 0; }
Note that it is important to check the return status of each solver step, in case the algorithm becomes stuck. If an error condition is detected, indicating that the algorithm cannot proceed, then the error can be reported to the user, a new starting point chosen or a different algorithm used.
The intermediate state of the solution is displayed by the following
function. The solver state contains the vector s->x
which is the
current position, and the vector s->f
with corresponding function
values.
int print_state (size_t iter, gsl_multiroot_fsolver * s) { printf ("iter = %3u x = % .3f % .3f " "f(x) = % .3e % .3e\n", iter, gsl_vector_get (s->x, 0), gsl_vector_get (s->x, 1), gsl_vector_get (s->f, 0), gsl_vector_get (s->f, 1)); }
Here are the results of running the program. The algorithm is started at (-10,-5) far from the solution. Since the solution is hidden in a narrow valley the earliest steps follow the gradient of the function downhill, in an attempt to reduce the large value of the residual. Once the root has been approximately located, on iteration 8, the Newton behavior takes over and convergence is very rapid.
iter = 0 x = -10.000 -5.000 f(x) = 1.100e+01 -1.050e+03 iter = 1 x = -10.000 -5.000 f(x) = 1.100e+01 -1.050e+03 iter = 2 x = -3.976 24.827 f(x) = 4.976e+00 9.020e+01 iter = 3 x = -3.976 24.827 f(x) = 4.976e+00 9.020e+01 iter = 4 x = -3.976 24.827 f(x) = 4.976e+00 9.020e+01 iter = 5 x = -1.274 -5.680 f(x) = 2.274e+00 -7.302e+01 iter = 6 x = -1.274 -5.680 f(x) = 2.274e+00 -7.302e+01 iter = 7 x = 0.249 0.298 f(x) = 7.511e-01 2.359e+00 iter = 8 x = 0.249 0.298 f(x) = 7.511e-01 2.359e+00 iter = 9 x = 1.000 0.878 f(x) = 1.268e-10 -1.218e+00 iter = 10 x = 1.000 0.989 f(x) = 1.124e-11 -1.080e-01 iter = 11 x = 1.000 1.000 f(x) = 0.000e+00 0.000e+00 status = success
Note that the algorithm does not update the location on every iteration. Some iterations are used to adjust the trust-region parameter, after trying a step which was found to be divergent, or to recompute the Jacobian, when poor convergence behavior is detected.
The next example program adds derivative information, in order to
accelerate the solution. There are two derivative functions
rosenbrock_df
and rosenbrock_fdf
. The latter computes both
the function and its derivative simultaneously. This allows the
optimization of any common terms. For simplicity we substitute calls to
the separate f
and df
functions at this point in the code
below.
int rosenbrock_df (const gsl_vector * x, void *params, gsl_matrix * J) { const double a = ((struct rparams *) params)->a; const double b = ((struct rparams *) params)->b; const double x0 = gsl_vector_get (x, 0); const double df00 = -a; const double df01 = 0; const double df10 = -2 * b * x0; const double df11 = b; gsl_matrix_set (J, 0, 0, df00); gsl_matrix_set (J, 0, 1, df01); gsl_matrix_set (J, 1, 0, df10); gsl_matrix_set (J, 1, 1, df11); return GSL_SUCCESS; } int rosenbrock_fdf (const gsl_vector * x, void *params, gsl_vector * f, gsl_matrix * J) { rosenbrock_f (x, params, f); rosenbrock_df (x, params, J); return GSL_SUCCESS; }
The main program now makes calls to the corresponding fdfsolver
versions of the functions,
int main (void) { const gsl_multiroot_fdfsolver_type *T; gsl_multiroot_fdfsolver *s; int status; size_t i, iter = 0; const size_t n = 2; struct rparams p = {1.0, 10.0}; gsl_multiroot_function_fdf f = {&rosenbrock_f, &rosenbrock_df, &rosenbrock_fdf, n, &p}; double x_init[2] = {-10.0, -5.0}; gsl_vector *x = gsl_vector_alloc (n); gsl_vector_set (x, 0, x_init[0]); gsl_vector_set (x, 1, x_init[1]); T = gsl_multiroot_fdfsolver_gnewton; s = gsl_multiroot_fdfsolver_alloc (T, n); gsl_multiroot_fdfsolver_set (s, &f, x); print_state (iter, s); do { iter++; status = gsl_multiroot_fdfsolver_iterate (s); print_state (iter, s); if (status) break; status = gsl_multiroot_test_residual (s->f, 1e-7); } while (status == GSL_CONTINUE && iter < 1000); printf ("status = %s\n", gsl_strerror (status)); gsl_multiroot_fdfsolver_free (s); gsl_vector_free (x); return 0; }
The addition of derivative information to the hybrids
solver does
not make any significant difference to its behavior, since it able to
approximate the Jacobian numerically with sufficient accuracy. To
illustrate the behavior of a different derivative solver we switch to
gnewton
. This is a traditional Newton solver with the constraint
that it scales back its step if the full step would lead "uphill". Here
is the output for the gnewton
algorithm,
iter = 0 x = -10.000 -5.000 f(x) = 1.100e+01 -1.050e+03 iter = 1 x = -4.231 -65.317 f(x) = 5.231e+00 -8.321e+02 iter = 2 x = 1.000 -26.358 f(x) = -8.882e-16 -2.736e+02 iter = 3 x = 1.000 1.000 f(x) = -2.220e-16 -4.441e-15 status = success
The convergence is much more rapid, but takes a wide excursion out to the point (-4.23,-65.3). This could cause the algorithm to go astray in a realistic application. The hybrid algorithm follows the downhill path to the solution more reliably.
The original version of the Hybrid method is described in the following articles by Powell,
The following papers are also relevant to the algorithms described in this section,
Go to the first, previous, next, last section, table of contents.