gsl_multifit_nlin and finite difference Jacobian

[japelj]

The attached script test2.pyx returns "segmentation fault: 11" when called from the python script test.py. The script is a modification of the nonlinear fit gsl example code, the only change being:
f.df=NULL
f.fdf=NULL
The question is, whether the problem is in the cython source or in my paths. The .pyx code:

from cython_gsl cimport *
from libc.math cimport exp, sqrt, pow
from libc.stdio cimport *
from libc.stdlib cimport malloc,free
from gsl_multifit_nlin cimport *
from gsl cimport GSL_CONTINUE

ctypedef struct Data:
    size_t n
    double * y
    double * sigma

cdef int expb_f (const gsl_vector * x, void * data, gsl_vector * f) nogil:
    cdef Data * d = <Data *> data
    cdef size_t n = d.n
    cdef double * y = d.y
    cdef double * sigma = d.sigma

    cdef double A = gsl_vector_get (x, 0)
    cdef double lambd = gsl_vector_get (x, 1)
    cdef double b = gsl_vector_get (x, 2)

    cdef size_t i
    cdef double t, Yi

    for i from 0 <= i < n:
        # Model Yi = A * exp(-lambd * i) + b
        t = i
        Yi = A * exp (-lambd * t) + b
        gsl_vector_set (f, i, (Yi - y[i])/sigma[i])

    return GSL_SUCCESS

cdef void print_state (size_t iter, gsl_multifit_fdfsolver * s) nogil:
    printf ("iter: %3u x = % 15.8f % 15.8f % 15.8f |f(x)| = %g\n", iter, gsl_vector_get (s.x, 0), gsl_vector_get (s.x, 1), gsl_vector_get (s.x, 2), gsl_blas_dnrm2 (s.f))

def main_test2():
    cdef const gsl_multifit_fdfsolver_type * T
    cdef gsl_multifit_fdfsolver * s
    cdef int status = GSL_CONTINUE
    cdef unsigned int i
    cdef unsigned int iter = 0
    cdef size_t n = 40
    cdef size_t p = 3

    cdef gsl_matrix * covar = gsl_matrix_alloc (p, p)
    cdef double y[40]
    cdef double sigma[40]
    cdef Data d
    d.n = n
    d.y = y
    d.sigma = sigma

    cdef gsl_multifit_function_fdf f
    cdef double * x_init = [1.0, 0.0, 0.0]
    cdef gsl_vector_view x = gsl_vector_view_array (x_init, p)
    cdef const gsl_rng_type * type
    cdef gsl_rng * r

    gsl_rng_env_setup ()

    type = gsl_rng_default
    r = gsl_rng_alloc (type)

    f.f = &expb_f
    #f.df = &expb_df
    #f.fdf = &expb_fdf
    f.df=NULL
    f.fdf=NULL
    f.n = n
    f.p = p
    f.params = &d

    # This is the data to be fitted
    cdef double t

    for i from 0 <= i < n:
        t = i;
        y[i] = 1.0 + 5 * exp (-0.1 * t) + gsl_ran_gaussian (r, 0.1)
        sigma[i] = 0.1
        printf ("data: %u %g %g\n", i, y[i], sigma[i])


    T = gsl_multifit_fdfsolver_lmsder
    s = gsl_multifit_fdfsolver_alloc (T, n, p)
    gsl_multifit_fdfsolver_set (s, &f, &x.vector)

    print_state (iter, s)

    while (status == GSL_CONTINUE and iter < 500):

        iter += 1
        status = gsl_multifit_fdfsolver_iterate (s)

        printf ("status = %d\n", status)

        print_state (iter, s)

        if status:
            break

        status = gsl_multifit_test_delta (s.dx, s.x, 1e-4, 1e-4)

    gsl_multifit_covar (s.J, 0.0, covar)

    cdef double chi = gsl_blas_dnrm2 (s.f)
    cdef double dof = n - p
    cdef double c = GSL_MAX_DBL(1, chi / sqrt (dof))

    printf("chisq/dof = %g\n",  pow (chi, 2.0) / dof)

    cdef double fit, err

    fit = gsl_vector_get (s.x, 0)
    err = sqrt (gsl_matrix_get (covar, 0, 0))
    printf ("A      = %.5f +/- %.5f\n", fit, c*err)

    fit = gsl_vector_get (s.x, 1)
    err = sqrt (gsl_matrix_get (covar, 1, 1))
    printf ("lambda = %.5f +/- %.5f\n", fit, c*err)

    fit = gsl_vector_get (s.x, 2)
    err = sqrt (gsl_matrix_get (covar, 2, 2))
    printf ("b      = %.5f +/- %.5f\n", fit, c*err)

    printf ("status = %d\n", status)

    gsl_multifit_fdfsolver_free (s)
    gsl_matrix_free (covar)
    gsl_rng_free (r)

[twiecki]

I don't see it directly. Try using gdb to debug the segfault.