コード例 #1
0
ファイル: lbfgs.cpp プロジェクト: huqinghao/T-LBFGS
 int line_search_backtracking_owlqn(
    int n,
    T *x,
    T *f,
    T *g,
    T *s,
    T *stp,
    const T* xp,
    const T* gp,
    T *wp,
    callback_data_t<T> *cd,
    const lbfgs_parameter_t *param
    )
{
    int i, count = 0;
    T width = 0.5, norm = 0.;
    T finit = *f, dgtest;

    /* Check the input parameters for errors. */
    if (*stp <= 0.) {
        return LBFGSERR_INVALIDPARAMETERS;
    }

    /* Choose the orthant for the new point. */
    for (i = 0;i < n;++i) {
        wp[i] = (xp[i] == 0.) ? -gp[i] : xp[i];
    }

    for (;;) {
        /* Update the current point. */
        veccpy(x, xp, n);
        vecadd(x, s, *stp, n);

        /* The current point is projected onto the orthant. */
        owlqn_project(x, wp, param->orthantwise_start, param->orthantwise_end);

        /* Evaluate the function and gradient values. */
        *f = cd->proc_evaluate(cd->instance, x, g, cd->n, *stp);

        /* Compute the L1 norm of the variables and add it to the object value. */
        norm = owlqn_x1norm(x, param->orthantwise_start, param->orthantwise_end);
        *f += norm * param->orthantwise_c;

        ++count;

        dgtest = 0.;
        for (i = 0;i < n;++i) {
            dgtest += (x[i] - xp[i]) * gp[i];
        }

        if (*f <= finit + param->ftol * dgtest) {
            /* The sufficient decrease condition. */
            return count;
        }

        if (*stp < param->min_step) {
            /* The step is the minimum value. */
            return LBFGSERR_MINIMUMSTEP;
        }
        if (*stp > param->max_step) {
            /* The step is the maximum value. */
            return LBFGSERR_MAXIMUMSTEP;
        }
        if (param->max_linesearch <= count) {
            /* Maximum number of iteration. */
            return LBFGSERR_MAXIMUMLINESEARCH;
        }

        (*stp) *= width;
    }
}
コード例 #2
0
ファイル: lbfgs.c プロジェクト: zhuoxiongzhao/ml-pack
static int line_search_backtracking_owlqn(
  int n,
  double* x,
  double* f,
  double* g,
  double* s,
  double* step,
  const double* xp,
  const double* gp,
  double* wp,
  callback_data_t* cd,
  const lbfgs_parameter_t* param
) {
  int i, count = 0;
  double width = 0.5, norm = 0.0;
  double finit = *f, dgtest;

  /* Check the input parameters for errors. */
  if (*step <= 0.0) {
    return LBFGSERR_INVALIDPARAMETERS;
  }

  /* Choose the orthant for the new point. */
  for (i = 0; i < n; i++) {
    wp[i] = (xp[i] == 0.0) ? -gp[i] : xp[i];
  }

  for (;;) {
    veccpy(x, xp, n);
    vecadd(x, s, *step, n);

    /* The current point is projected onto the orthant. */
    owlqn_project(x, wp, param->orthantwise_start, param->orthantwise_end);

    *f = cd->evaluate(cd->instance, cd->n, x, g, *step);
    count++;

    /* Compute the L1 norm of the variables and add it to the object value. */
    norm = owlqn_x1norm(x, param->orthantwise_start, param->orthantwise_end);
    *f += norm * param->orthantwise_c;

    dgtest = 0.0;
    for (i = 0; i < n; i++) {
      dgtest += (x[i] - xp[i]) * gp[i];
    }

    if (*f <= finit + param->ftol * dgtest) {
      /* The sufficient decrease condition. */
      return count;
    }

    if (*step < param->min_step) {
      /* The step is the minimum value. */
      return LBFGSERR_MINIMUMSTEP;
    }
    if (*step > param->max_step) {
      /* The step is the maximum value. */
      return LBFGSERR_MAXIMUMSTEP;
    }
    if (count >= param->max_linesearch) {
      /* Maximum number of iteration. */
      return LBFGSERR_MAXIMUMLINESEARCH;
    }

    (*step) *= width;
  }
}
コード例 #3
0
ファイル: lbfgs.cpp プロジェクト: huqinghao/T-LBFGS
int lbfgs(
    int n,
    T *x,
    T *ptr_fx,
typename   FuncWrapper<T>::lbfgs_evaluate_t proc_evaluate,
    typename  FuncWrapper<T>::lbfgs_progress_t proc_progress,
    void *instance,
    lbfgs_parameter_t *_param
    )
{
    int ret;
    int i, j, k, ls, end, bound;
    T step;

    /* Constant parameters and their default values. */
    lbfgs_parameter_t param = (_param != NULL) ? (*_param) : _defparam;
    const int m = param.m;

    T *xp = NULL;
    T *g = NULL, *gp = NULL, *pg = NULL;
    T *d = NULL, *w = NULL, *pf = NULL;
    iteration_data_t<T> *lm = NULL;
    iteration_data_t<T>*it = NULL;
    T ys, yy;
    T xnorm, gnorm, beta;
    T fx = 0.;
    T rate = 0.;
typename    LineSearchWrapper<T>::line_search_proc linesearch = line_search_morethuente;

    /* Construct a callback data. */
    callback_data_t<T> cd;
    cd.n = n;
    cd.instance = instance;
    cd.proc_evaluate = proc_evaluate;
    cd.proc_progress = proc_progress;

#if     defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
    /* Round out the number of variables. */
    n = round_out_variables(n);
#endif/*defined(USE_SSE)*/

    /* Check the input parameters for errors. */
    if (n <= 0) {
        return LBFGSERR_INVALID_N;
    }
#if     defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
    if (n % 8 != 0) {
        return LBFGSERR_INVALID_N_SSE;
    }
    if ((uintptr_t)(const void*)x % 16 != 0) {
        return LBFGSERR_INVALID_X_SSE;
    }
#endif/*defined(USE_SSE)*/
    if (param.epsilon < 0.) {
        return LBFGSERR_INVALID_EPSILON;
    }
    if (param.past < 0) {
        return LBFGSERR_INVALID_TESTPERIOD;
    }
    if (param.delta < 0.) {
        return LBFGSERR_INVALID_DELTA;
    }
    if (param.min_step < 0.) {
        return LBFGSERR_INVALID_MINSTEP;
    }
    if (param.max_step < param.min_step) {
        return LBFGSERR_INVALID_MAXSTEP;
    }
    if (param.ftol < 0.) {
        return LBFGSERR_INVALID_FTOL;
    }
    if (param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_WOLFE ||
        param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
        if (param.wolfe <= param.ftol || 1. <= param.wolfe) {
            return LBFGSERR_INVALID_WOLFE;
        }
    }
    if (param.gtol < 0.) {
        return LBFGSERR_INVALID_GTOL;
    }
    if (param.xtol < 0.) {
        return LBFGSERR_INVALID_XTOL;
    }
    if (param.max_linesearch <= 0) {
        return LBFGSERR_INVALID_MAXLINESEARCH;
    }
    if (param.orthantwise_c < 0.) {
        return LBFGSERR_INVALID_ORTHANTWISE;
    }
    if (param.orthantwise_start < 0 || n < param.orthantwise_start) {
        return LBFGSERR_INVALID_ORTHANTWISE_START;
    }
    if (param.orthantwise_end < 0) {
        param.orthantwise_end = n;
    }
    if (n < param.orthantwise_end) {
        return LBFGSERR_INVALID_ORTHANTWISE_END;
    }
    if (param.orthantwise_c != 0.) {
        switch (param.linesearch) {
        case LBFGS_LINESEARCH_BACKTRACKING:
            linesearch = line_search_backtracking_owlqn;
            break;
        default:
            /* Only the backtracking method is available. */
            return LBFGSERR_INVALID_LINESEARCH;
        }
    } else {
        switch (param.linesearch) {
        case LBFGS_LINESEARCH_MORETHUENTE:
            linesearch = line_search_morethuente;
            break;
        case LBFGS_LINESEARCH_BACKTRACKING_ARMIJO:
        case LBFGS_LINESEARCH_BACKTRACKING_WOLFE:
        case LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE:
            linesearch = line_search_backtracking;
            break;
        default:
            return LBFGSERR_INVALID_LINESEARCH;
        }
    }

    /* Allocate working space. */
    xp = (T*)vecalloc(n * sizeof(T));
    g = (T*)vecalloc(n * sizeof(T));
    gp = (T*)vecalloc(n * sizeof(T));
    d = (T*)vecalloc(n * sizeof(T));
    w = (T*)vecalloc(n * sizeof(T));
    if (xp == NULL || g == NULL || gp == NULL || d == NULL || w == NULL) {
        ret = LBFGSERR_OUTOFMEMORY;
        goto lbfgs_exit;
    }

    if (param.orthantwise_c != 0.) {
        /* Allocate working space for OW-LQN. */
        pg = (T*)vecalloc(n * sizeof(T));
        if (pg == NULL) {
            ret = LBFGSERR_OUTOFMEMORY;
            goto lbfgs_exit;
        }
    }

    /* Allocate limited memory storage. */
    lm = (iteration_data_t<T>*)vecalloc(m * sizeof(iteration_data_t<T>));
    if (lm == NULL) {
        ret = LBFGSERR_OUTOFMEMORY;
        goto lbfgs_exit;
    }

    /* Initialize the limited memory. */
    for (i = 0;i < m;++i) {
        it = &lm[i];
        it->alpha = 0;
        it->ys = 0;
        it->s = (T*)vecalloc(n * sizeof(T));
        it->y = (T*)vecalloc(n * sizeof(T));
        if (it->s == NULL || it->y == NULL) {
            ret = LBFGSERR_OUTOFMEMORY;
            goto lbfgs_exit;
        }
    }

    /* Allocate an array for storing previous values of the objective function. */
    if (0 < param.past) {
        pf = (T*)vecalloc(param.past * sizeof(T));
    }

    /* Evaluate the function value and its gradient. */
    fx = cd.proc_evaluate(cd.instance, x, g, cd.n, 0);
    if (0. != param.orthantwise_c) {
        /* Compute the L1 norm of the variable and add it to the object value. */
        xnorm = owlqn_x1norm(x, param.orthantwise_start, param.orthantwise_end);
        fx += xnorm * param.orthantwise_c;
        owlqn_pseudo_gradient(
            pg, x, g, n,
            T(param.orthantwise_c), param.orthantwise_start, param.orthantwise_end
            );
    }

    /* Store the initial value of the objective function. */
    if (pf != NULL) {
        pf[0] = fx;
    }

    /*
        Compute the direction;
        we assume the initial hessian matrix H_0 as the identity matrix.
     */
    if (param.orthantwise_c == 0.) {
        vecncpy(d, g, n);
    } else {
        vecncpy(d, pg, n);
    }

    /*
       Make sure that the initial variables are not a minimizer.
     */
    vec2norm(&xnorm, x, n);
    if (param.orthantwise_c == 0.) {
        vec2norm(&gnorm, g, n);
    } else {
        vec2norm(&gnorm, pg, n);
    }
    if (xnorm < 1.0) xnorm = 1.0;
    if (gnorm / xnorm <= param.epsilon) {
        ret = LBFGS_ALREADY_MINIMIZED;
        goto lbfgs_exit;
    }

    /* Compute the initial step:
        step = 1.0 / sqrt(vecdot(d, d, n))
     */
    vec2norminv(&step, d, n);

    k = 1;
    end = 0;
    for (;;) {
        /* Store the current position and gradient vectors. */
        veccpy(xp, x, n);
        veccpy(gp, g, n);

        /* Search for an optimal step. */
        if (param.orthantwise_c == 0.) {
            ls = linesearch(n, x, &fx, g, d, &step, xp, gp, w, &cd, &param);
        } else {
            ls = linesearch(n, x, &fx, g, d, &step, xp, pg, w, &cd, &param);
            owlqn_pseudo_gradient(
                pg, x, g, n,
                T(param.orthantwise_c), param.orthantwise_start, param.orthantwise_end
                );
        }
        if (ls < 0) {
            /* Revert to the previous point. */
            veccpy(x, xp, n);
            veccpy(g, gp, n);
            ret = ls;
            goto lbfgs_exit;
        }

        /* Compute x and g norms. */
        vec2norm(&xnorm, x, n);
        if (param.orthantwise_c == 0.) {
            vec2norm(&gnorm, g, n);
        } else {
            vec2norm(&gnorm, pg, n);
        }

        /* Report the progress. */
        if (cd.proc_progress) {
            if ((ret = cd.proc_progress(cd.instance, x, g, fx, xnorm, gnorm, step, cd.n, k, ls))) {
                goto lbfgs_exit;
            }
        }

        /*
            Convergence test.
            The criterion is given by the following formula:
                |g(x)| / \max(1, |x|) < \epsilon
         */
        if (xnorm < 1.0) xnorm = 1.0;
        if (gnorm / xnorm <= param.epsilon) {
            /* Convergence. */
            ret = LBFGS_SUCCESS;
            break;
        }

        /*
            Test for stopping criterion.
            The criterion is given by the following formula:
                (f(past_x) - f(x)) / f(x) < \delta
         */
        if (pf != NULL) {
            /* We don't test the stopping criterion while k < past. */
            if (param.past <= k) {
                /* Compute the relative improvement from the past. */
                rate = (pf[k % param.past] - fx) / fx;

                /* The stopping criterion. */
                if (rate < param.delta) {
                    ret = LBFGS_STOP;
                    break;
                }
            }

            /* Store the current value of the objective function. */
            pf[k % param.past] = fx;
        }

        if (param.max_iterations != 0 && param.max_iterations < k+1) {
            /* Maximum number of iterations. */
            ret = LBFGSERR_MAXIMUMITERATION;
            break;
        }

        /*
            Update vectors s and y:
                s_{k+1} = x_{k+1} - x_{k} = \step * d_{k}.
                y_{k+1} = g_{k+1} - g_{k}.
         */
        it = &lm[end];
        vecdiff(it->s, x, xp, n);
        vecdiff(it->y, g, gp, n);

        /*
            Compute scalars ys and yy:
                ys = y^t \cdot s = 1 / \rho.
                yy = y^t \cdot y.
            Notice that yy is used for scaling the hessian matrix H_0 (Cholesky factor).
         */
        vecdot(&ys, it->y, it->s, n);
        vecdot(&yy, it->y, it->y, n);
        it->ys = ys;

        /*
            Recursive formula to compute dir = -(H \cdot g).
                This is described in page 779 of:
                Jorge Nocedal.
                Updating Quasi-Newton Matrices with Limited Storage.
                Mathematics of Computation, Vol. 35, No. 151,
                pp. 773--782, 1980.
         */
        bound = (m <= k) ? m : k;
        ++k;
        end = (end + 1) % m;

        /* Compute the steepest direction. */
        if (param.orthantwise_c == 0.) {
            /* Compute the negative of gradients. */
            vecncpy(d, g, n);
        } else {
            vecncpy(d, pg, n);
        }

        j = end;
        for (i = 0;i < bound;++i) {
            j = (j + m - 1) % m;    /* if (--j == -1) j = m-1; */
            it = &lm[j];
            /* \alpha_{j} = \rho_{j} s^{t}_{j} \cdot q_{k+1}. */
            vecdot(&it->alpha, it->s, d, n);
            it->alpha /= it->ys;
            /* q_{i} = q_{i+1} - \alpha_{i} y_{i}. */
            vecadd(d, it->y, -it->alpha, n);
        }

        vecscale(d, ys / yy, n);

        for (i = 0;i < bound;++i) {
            it = &lm[j];
            /* \beta_{j} = \rho_{j} y^t_{j} \cdot \gamma_{i}. */
            vecdot(&beta, it->y, d, n);
            beta /= it->ys;
            /* \gamma_{i+1} = \gamma_{i} + (\alpha_{j} - \beta_{j}) s_{j}. */
            vecadd(d, it->s, it->alpha - beta, n);
            j = (j + 1) % m;        /* if (++j == m) j = 0; */
        }

        /*
            Constrain the search direction for orthant-wise updates.
         */
        if (param.orthantwise_c != 0.) {
            for (i = param.orthantwise_start;i < param.orthantwise_end;++i) {
                if (d[i] * pg[i] >= 0) {
                    d[i] = 0;
                }
            }
        }

        /*
            Now the search direction d is ready. We try step = 1 first.
         */
        step = 1.0;
    }

lbfgs_exit:
    /* Return the final value of the objective function. */
    if (ptr_fx != NULL) {
        *ptr_fx = fx;
    }

    vecfree(pf);

    /* Free memory blocks used by this function. */
    if (lm != NULL) {
        for (i = 0;i < m;++i) {
            vecfree(lm[i].s);
            vecfree(lm[i].y);
        }
        vecfree(lm);
    }
    vecfree(pg);
    vecfree(w);
    vecfree(d);
    vecfree(gp);
    vecfree(g);
    vecfree(xp);

    return ret;
}
コード例 #4
0
ファイル: lbfgs.c プロジェクト: zhuoxiongzhao/ml-pack
int lbfgs(
  int n,
  double* x,
  double* pfx,
  lbfgs_evaluate_t evaluate,
  lbfgs_progress_t progress,
  void* instance,
  const lbfgs_parameter_t* _param
) {
  int ret;
  int i, j, k, ls, end, bound, n_evaluate = 0;
  int enalbe_owlqn;
  double step;
  lbfgs_parameter_t param = (_param) ? (*_param) : default_param;
  const int m = param.m;
  double* xp;
  double* g, *gp, *pg = 0;
  double* d, *w, *pf = 0;
  iteration_data_t* lm = 0, *it = 0;
  double ys, yy;
  double xnorm, gnorm, rate, beta;
  double fx;
  line_search_proc_t linesearch = line_search_morethuente;

  callback_data_t cd;
  cd.n = n;
  cd.instance = instance;
  cd.evaluate = evaluate;
  cd.progress = (progress) ? progress : default_lbfgs_progress;

  /* Check the input parameters for errors. */
  if (n <= 0) {
    return LBFGSERR_INVALID_N;
  }
  if (param.epsilon < 0.0) {
    return LBFGSERR_INVALID_EPSILON;
  }
  if (param.past < 0) {
    return LBFGSERR_INVALID_TESTPERIOD;
  }
  if (param.delta < 0.0) {
    return LBFGSERR_INVALID_DELTA;
  }
  if (param.min_step < 0.0) {
    return LBFGSERR_INVALID_MINSTEP;
  }
  if (param.max_step < param.min_step) {
    return LBFGSERR_INVALID_MAXSTEP;
  }
  if (param.ftol < 0.0) {
    return LBFGSERR_INVALID_FTOL;
  }
  if (param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_WOLFE ||
      param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
    if (param.wolfe <= param.ftol || 1. <= param.wolfe) {
      return LBFGSERR_INVALID_WOLFE;
    }
  }
  if (param.gtol < 0.0) {
    return LBFGSERR_INVALID_GTOL;
  }
  if (param.xtol < 0.0) {
    return LBFGSERR_INVALID_XTOL;
  }
  if (param.max_linesearch <= 0) {
    return LBFGSERR_INVALID_MAXLINESEARCH;
  }
  if (param.orthantwise_c < 0.0) {
    return LBFGSERR_INVALID_ORTHANTWISE;
  }
  if (param.orthantwise_start < 0 || param.orthantwise_start > n) {
    return LBFGSERR_INVALID_ORTHANTWISE_START;
  }
  if (param.orthantwise_end < 0) {
    param.orthantwise_end = n;
  }
  if (param.orthantwise_end > n) {
    return LBFGSERR_INVALID_ORTHANTWISE_END;
  }

  enalbe_owlqn = (param.orthantwise_c != 0.0);
  if (enalbe_owlqn) {
    switch (param.linesearch) {
    case LBFGS_LINESEARCH_BACKTRACKING_WOLFE:
      linesearch = line_search_backtracking_owlqn;
      break;
    default:
      /* Only the backtracking method is available. */
      return LBFGSERR_INVALID_LINESEARCH;
    }
  } else {
    switch (param.linesearch) {
    case LBFGS_LINESEARCH_MORETHUENTE:
      linesearch = line_search_morethuente;
      break;
    case LBFGS_LINESEARCH_BACKTRACKING_ARMIJO:
    case LBFGS_LINESEARCH_BACKTRACKING_WOLFE:
    case LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE:
      linesearch = line_search_backtracking;
      break;
    default:
      return LBFGSERR_INVALID_LINESEARCH;
    }
  }

  /* Allocate working space. */
  xp = vecalloc(n);
  g = vecalloc(n);
  gp = vecalloc(n);
  d = vecalloc(n);
  w = vecalloc(n);

  /* Allocate pseudo gradient. */
  if (enalbe_owlqn) {
    pg = vecalloc(n);
  }

  /* Allocate and initialize the limited memory storage. */
  lm = (iteration_data_t*)xalloc(m * sizeof(iteration_data_t));
  for (i = 0; i < m; i++) {
    it = &lm[i];
    it->alpha = 0.0;
    it->s = vecalloc(n);
    it->y = vecalloc(n);
    it->ys = 0.0;
  }

  /* Allocate an array for storing previous values of the objective function. */
  if (param.past > 0) {
    pf = vecalloc((size_t)param.past);
  }

  fx = cd.evaluate(cd.instance, cd.n, x, g, 0);
  n_evaluate++;

  if (enalbe_owlqn) {
    xnorm = owlqn_x1norm(x, param.orthantwise_start, param.orthantwise_end);
    fx += xnorm * param.orthantwise_c;
    owlqn_pseudo_gradient(
      pg, x, g, n,
      param.orthantwise_c, param.orthantwise_start, param.orthantwise_end);
  }

  /* Store the initial value of the objective function. */
  if (pf) {
    pf[0] = fx;
  }

  /**
  * Compute the direction.
  * we assume the initial hessian matrix H_0 as the identity matrix.
  */
  if (!enalbe_owlqn) {
    vecncpy(d, g, n);
  } else {
    vecncpy(d, pg, n);
  }

  /**
  * Make sure that the initial variables are not a minimizer.
  */
  vec2norm(&xnorm, x, n);
  if (!enalbe_owlqn) {
    vec2norm(&gnorm, g, n);
  } else {
    vec2norm(&gnorm, pg, n);
  }
  if (xnorm < 1.0) {
    xnorm = 1.0;
  }
  if (gnorm / xnorm <= param.epsilon) {
    ret = LBFGS_ALREADY_MINIMIZED;
    goto lbfgs_exit;
  }

  /**
  * Compute the initial step:
  * step = 1.0 / ||d||
  */
  vec2norminv(&step, d, n);

  k = 1;
  end = 0;
  for (;;) {
    /* Store the current position and gradient vectors. */
    veccpy(xp, x, n);
    veccpy(gp, g, n);

    /* Search for an optimal step. */
    if (!enalbe_owlqn) {
      ls = linesearch(n, x, &fx, g, d, &step, xp, gp, w, &cd, &param);
    } else {
      ls = linesearch(n, x, &fx, g, d, &step, xp, pg, w, &cd, &param);
      owlqn_pseudo_gradient(
        pg, x, g, n,
        param.orthantwise_c, param.orthantwise_start, param.orthantwise_end
      );
    }

    if (ls < 0) {
      /* Revert to the previous point. */
      veccpy(x, xp, n);
      veccpy(g, gp, n);
      ret = ls;
      break;
    }

    n_evaluate += ls;

    /* Compute x and g norms. */
    vec2norm(&xnorm, x, n);
    if (!enalbe_owlqn) {
      vec2norm(&gnorm, g, n);
    } else {
      vec2norm(&gnorm, pg, n);
    }

    /* Report the progress. */
    if ((ret = cd.progress(cd.instance, cd.n, x, g, fx, xnorm, gnorm, step, k, n_evaluate)) != 0) {
      ret = LBFGSERR_CANCELED;
      break;
    }

    /* Convergence test. */
    if (xnorm < 1.0) {
      xnorm = 1.0;
    }
    if (gnorm / xnorm <= param.epsilon) {
      ret = LBFGS_CONVERGENCE;
      break;
    }

    /* Stopping criterion test. */
    if (pf) {
      /* We don't test the stopping criterion while k < past. */
      if (param.past <= k) {
        /* Compute the relative improvement from the past. */
        rate = (pf[k % param.past] - fx) / fx;

        /* The stopping criterion. */
        if (rate < param.delta) {
          ret = LBFGS_CONVERGENCE_DELTA;
          break;
        }
      }

      /* Store the current value of the objective function. */
      pf[k % param.past] = fx;
    }

    if (param.max_iterations != 0 && param.max_iterations < k + 1) {
      ret = LBFGSERR_MAXIMUMITERATION;
      break;
    }

    /**
    * Update s and y:
    * s_{k+1} = x_{k+1} - x_{k} = step * d_{k}
    * y_{k+1} = g_{k+1} - g_{k}
    */
    it = &lm[end];
    vecdiff(it->s, x, xp, n);
    vecdiff(it->y, g, gp, n);

    /**
    * Compute scalars ys and yy:
    * ys = y^t s = 1 / \rho
    * yy = y^t y
    * Notice that yy is used for scaling the hessian matrix H_0 (Cholesky factor).
    */
    vecdot(&ys, it->y, it->s, n);
    vecdot(&yy, it->y, it->y, n);
    it->ys = ys;

    /**
    * Recursive formula to compute d = -(H g).
    * This is described in page 779 of:
    * Jorge Nocedal.
    * Updating Quasi-Newton Matrices with Limited Storage.
    * Mathematics of Computation, Vol. 35, No. 151,
    * pp. 773--782, 1980.
    */
    bound = (m <= k) ? m : k;
    k++;
    end = (end + 1) % m;

    /* Compute the steepest direction. */
    /* Compute the negative of (pseudo) gradient. */
    if (!enalbe_owlqn) {
      vecncpy(d, g, n);
    } else {
      vecncpy(d, pg, n);
    }

    j = end;
    for (i = 0; i < bound; i++) {
      j = (j + m - 1) % m; /* if (--j == -1) j = m-1; */
      it = &lm[j];
      /* \alpha_{j} = \rho_{j} s^{t}_{j} q_{k+1} */
      vecdot(&it->alpha, it->s, d, n);
      it->alpha /= it->ys;
      /* q_{i} = q_{i+1} - \alpha_{i} y_{i} */
      vecadd(d, it->y, -it->alpha, n);
    }

    vecscale(d, ys / yy, n);

    for (i = 0; i < bound; i++) {
      it = &lm[j];
      /* \beta_{j} = \rho_{j} y^t_{j} \gamma_{i} */
      vecdot(&beta, it->y, d, n);
      beta /= it->ys;
      /* \gamma_{i+1} = \gamma_{i} + (\alpha_{j} - \beta_{j}) s_{j} */
      vecadd(d, it->s, it->alpha - beta, n);
      j = (j + 1) % m; /* if (++j == m) j = 0; */
    }

    /* Constrain the search direction for orthant-wise updates. */
    if (enalbe_owlqn) {
      owlqn_contrain_line_search(d, pg, param.orthantwise_start, param.orthantwise_end);
    }

    /* Now the search direction d is ready. We try step = 1 first. */
    step = 1.0;
  }

lbfgs_exit:
  /* Return the final value of the objective function. */
  if (pfx) {
    *pfx = fx;
  }

  vecfree(pf);
  if (lm != 0) {
    for (i = 0; i < m; i++) {
      vecfree(lm[i].s);
      vecfree(lm[i].y);
    }
    xfree(lm);
  }
  vecfree(pg);
  vecfree(w);
  vecfree(d);
  vecfree(gp);
  vecfree(g);
  vecfree(xp);
  return ret;
}