| 1 |
! Copyright © 2025-2026 Dynare Team |
| 2 |
! |
| 3 |
! This file is part of Dynare. |
| 4 |
! |
| 5 |
! Dynare is free software: you can redistribute it and/or modify it under the terms of the |
| 6 |
! GNU General Public License as published by the Free Software Foundation, either version 3 of |
| 7 |
! the License, or (at your option) any later version. |
| 8 |
! |
| 9 |
! Dynare is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without |
| 10 |
! even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 11 |
! GNU General Public License for more details. |
| 12 |
! |
| 13 |
! You should have received a copy of the GNU General Public License along with Dynare. If not, |
| 14 |
! see <https://www.gnu.org/licenses/>. |
| 15 |
! |
| 16 |
! Original author: Normann Rion <[email protected]> |
| 17 |
! |
| 18 |
! Broyden quasi-Newton solver for heterogeneous-agent steady-state calibration |
| 19 |
! |
| 20 |
! Implements the Broyden method following the Python SSJ implementation |
| 21 |
! (sequence_jacobian/utilities/solvers.py). Uses LAPACK dgesv for linear solves. |
| 22 |
|
| 23 |
module broyden_solver |
| 24 |
use iso_fortran_env, only: real64, int32 |
| 25 |
use ieee_arithmetic, only: ieee_is_finite |
| 26 |
use blas, only: blint, matvecmul_add, rank_one_update |
| 27 |
use lapack, only: dgesv |
| 28 |
use calibration_types, only: broyden_workspace |
| 29 |
use matlab_mex, only: mexPrintf |
| 30 |
|
| 31 |
implicit none (type, external) |
| 32 |
|
| 33 |
private |
| 34 |
public :: broyden_solve |
| 35 |
|
| 36 |
! Default solver parameters |
| 37 |
real(real64), parameter :: DEFAULT_FD_STEP = 1.0e-5_real64 ! Finite difference step |
| 38 |
real(real64), parameter :: DEFAULT_BACKTRACK_C = 0.5_real64 ! Backtracking factor |
| 39 |
integer(int32), parameter :: DEFAULT_MAX_BACKTRACK = 30 ! Max backtracks per iteration |
| 40 |
|
| 41 |
! Return codes |
| 42 |
integer(int32), parameter, public :: BROYDEN_SUCCESS = 0 |
| 43 |
integer(int32), parameter, public :: BROYDEN_MAX_ITER = 1 |
| 44 |
integer(int32), parameter, public :: BROYDEN_SINGULAR_JAC = 2 |
| 45 |
integer(int32), parameter, public :: BROYDEN_MAX_BACKTRACK = 3 |
| 46 |
integer(int32), parameter, public :: BROYDEN_FCN_ERROR = 4 |
| 47 |
|
| 48 |
! Abstract interface for residual function callback |
| 49 |
abstract interface |
| 50 |
subroutine residual_fcn(n, x, fvec, iflag) |
| 51 |
use iso_fortran_env, only: real64 |
| 52 |
integer, intent(in) :: n |
| 53 |
real(real64), intent(in) :: x(n) |
| 54 |
real(real64), intent(out) :: fvec(n) |
| 55 |
integer, intent(inout) :: iflag |
| 56 |
end subroutine residual_fcn |
| 57 |
end interface |
| 58 |
|
| 59 |
contains |
| 60 |
|
| 61 |
!--------------------------------------------------------------------------- |
| 62 |
! Broyden quasi-Newton solver |
| 63 |
! |
| 64 |
! Solves f(x) = 0 using Broyden's method with finite-difference initial |
| 65 |
! Jacobian and rank-1 updates. |
| 66 |
! |
| 67 |
! Arguments: |
| 68 |
! ws : Pre-allocated Broyden workspace |
| 69 |
! x : On entry: initial guess; on exit: solution |
| 70 |
! fvec : On exit: residual at solution |
| 71 |
! fcn : Residual function f(x) to solve for f(x) = 0 |
| 72 |
! tol : Convergence tolerance (max|fvec| < tol) |
| 73 |
! max_iter : Maximum number of Broyden iterations |
| 74 |
! nfev : (out) Number of function evaluations |
| 75 |
! fd_step : (optional) Finite difference step size |
| 76 |
! backtrack_c : (optional) Backtracking factor |
| 77 |
! verbosity : (optional) Verbosity level (2 = verbose, 1 = silent) |
| 78 |
! lower_bounds : (optional) Lower bounds for variables |
| 79 |
! upper_bounds : (optional) Upper bounds for variables |
| 80 |
! |
| 81 |
! Returns: status code (0 = success) |
| 82 |
!--------------------------------------------------------------------------- |
| 83 |
function broyden_solve(ws, x, fvec, fcn, tol, max_iter, nfev, fd_step, backtrack_c, verbosity, & |
| 84 |
lower_bounds, upper_bounds) result(info) |
| 85 |
type(broyden_workspace), intent(inout) :: ws |
| 86 |
real(real64), intent(inout) :: x(:) |
| 87 |
real(real64), intent(out) :: fvec(:) |
| 88 |
procedure(residual_fcn) :: fcn |
| 89 |
real(real64), intent(in) :: tol |
| 90 |
integer(int32), intent(in) :: max_iter |
| 91 |
integer(int32), intent(out) :: nfev |
| 92 |
real(real64), intent(in), optional :: fd_step, backtrack_c |
| 93 |
integer(int32), intent(in), optional :: verbosity |
| 94 |
real(real64), intent(in), optional :: lower_bounds(:), upper_bounds(:) |
| 95 |
integer(int32) :: info |
| 96 |
|
| 97 |
! Local variables |
| 98 |
integer(int32) :: n, iter, bcount, iflag |
| 99 |
real(real64) :: h, bc, max_residual |
| 100 |
integer(blint) :: n_bl, nrhs, lda, ldb, lapack_info |
| 101 |
real(real64), allocatable :: x_new(:) |
| 102 |
logical :: has_bounds |
| 103 |
|
| 104 |
n = ws%n_vars |
| 105 |
h = DEFAULT_FD_STEP |
| 106 |
bc = DEFAULT_BACKTRACK_C |
| 107 |
if (present(fd_step)) h = fd_step |
| 108 |
if (present(backtrack_c)) bc = backtrack_c |
| 109 |
|
| 110 |
! Check if bounds are provided |
| 111 |
has_bounds = present(lower_bounds) .and. present(upper_bounds) |
| 112 |
allocate(x_new(n)) |
| 113 |
|
| 114 |
nfev = 0 |
| 115 |
info = BROYDEN_SUCCESS |
| 116 |
|
| 117 |
! Step 1: Evaluate initial residual f(x0) |
| 118 |
iflag = 0 |
| 119 |
call fcn(n, x, fvec, iflag) |
| 120 |
nfev = nfev + 1 |
| 121 |
|
| 122 |
if (iflag < 0) then |
| 123 |
info = BROYDEN_FCN_ERROR |
| 124 |
return |
| 125 |
end if |
| 126 |
|
| 127 |
! Main iteration loop |
| 128 |
do iter = 1, max_iter |
| 129 |
! Check convergence |
| 130 |
max_residual = maxval(abs(fvec)) |
| 131 |
if (max_residual < tol) then |
| 132 |
info = BROYDEN_SUCCESS |
| 133 |
return |
| 134 |
end if |
| 135 |
|
| 136 |
! Step 2: Compute Jacobian (first iteration only) |
| 137 |
if (iter == 1) then |
| 138 |
if (present(verbosity)) then |
| 139 |
if (verbosity == 2) & |
| 140 |
call mexPrintf(' Computing initial Jacobian via finite differences...'//NEW_LINE('A')) |
| 141 |
end if |
| 142 |
call compute_fd_jacobian(ws, x, fvec, fcn, h, nfev, iflag, verbosity) |
| 143 |
if (iflag < 0) then |
| 144 |
info = BROYDEN_FCN_ERROR |
| 145 |
return |
| 146 |
end if |
| 147 |
end if |
| 148 |
|
| 149 |
! Step 3: Solve J * dx = -fvec |
| 150 |
! Copy J to J_copy (dgesv overwrites it) |
| 151 |
ws%J_copy = ws%J |
| 152 |
! Set dx = -fvec (right-hand side) |
| 153 |
ws%dx = -fvec |
| 154 |
|
| 155 |
! Call LAPACK dgesv |
| 156 |
n_bl = int(n, blint) |
| 157 |
nrhs = 1_blint |
| 158 |
lda = n_bl |
| 159 |
ldb = n_bl |
| 160 |
call dgesv(n_bl, nrhs, ws%J_copy, lda, ws%ipiv, ws%dx, ldb, lapack_info) |
| 161 |
|
| 162 |
if (lapack_info /= 0) then |
| 163 |
info = BROYDEN_SINGULAR_JAC |
| 164 |
return |
| 165 |
end if |
| 166 |
|
| 167 |
! Step 4: Backtracking line search |
| 168 |
do bcount = 1, DEFAULT_MAX_BACKTRACK |
| 169 |
! Compute new point: x + dx, then clamp to bounds if present |
| 170 |
x_new = x + ws%dx |
| 171 |
if (has_bounds) call clamp_to_bounds(x_new, lower_bounds, upper_bounds, n) |
| 172 |
|
| 173 |
! Try new point |
| 174 |
iflag = 0 |
| 175 |
call fcn(n, x_new, ws%fvec_new, iflag) |
| 176 |
nfev = nfev + 1 |
| 177 |
|
| 178 |
if (iflag < 0) then |
| 179 |
! Function evaluation failed, backtrack |
| 180 |
ws%dx = ws%dx * bc |
| 181 |
else |
| 182 |
! Success: update Jacobian with Broyden rank-1 update |
| 183 |
! df = fvec_new - fvec |
| 184 |
ws%df = ws%fvec_new - fvec |
| 185 |
|
| 186 |
! Compute actual dx (may differ from ws%dx due to clamping) |
| 187 |
ws%dx = x_new - x |
| 188 |
|
| 189 |
! Move to new point (must be done BEFORE broyden_update |
| 190 |
! because broyden_update overwrites ws%fvec_new as workspace) |
| 191 |
fvec = ws%fvec_new |
| 192 |
x = x_new |
| 193 |
|
| 194 |
! Broyden update: J = J + outer((df - J*dx) / ||dx||^2, dx) |
| 195 |
call broyden_update(ws) |
| 196 |
|
| 197 |
exit |
| 198 |
end if |
| 199 |
end do |
| 200 |
|
| 201 |
if (bcount > DEFAULT_MAX_BACKTRACK) then |
| 202 |
info = BROYDEN_MAX_BACKTRACK |
| 203 |
return |
| 204 |
end if |
| 205 |
end do |
| 206 |
|
| 207 |
! Maximum iterations exceeded |
| 208 |
info = BROYDEN_MAX_ITER |
| 209 |
|
| 210 |
end function broyden_solve |
| 211 |
|
| 212 |
!--------------------------------------------------------------------------- |
| 213 |
! Clamp variables to bounds |
| 214 |
! |
| 215 |
! For each variable, if the bound is finite (not -Inf or +Inf), clamp. |
| 216 |
!--------------------------------------------------------------------------- |
| 217 |
subroutine clamp_to_bounds(x, lb, ub, n) |
| 218 |
real(real64), intent(inout) :: x(:) |
| 219 |
real(real64), intent(in) :: lb(:), ub(:) |
| 220 |
integer(int32), intent(in) :: n |
| 221 |
|
| 222 |
integer :: i |
| 223 |
|
| 224 |
do i = 1, n |
| 225 |
if (ieee_is_finite(lb(i))) x(i) = max(x(i), lb(i)) |
| 226 |
if (ieee_is_finite(ub(i))) x(i) = min(x(i), ub(i)) |
| 227 |
end do |
| 228 |
end subroutine clamp_to_bounds |
| 229 |
|
| 230 |
!--------------------------------------------------------------------------- |
| 231 |
! Compute initial Jacobian via forward finite differences |
| 232 |
! |
| 233 |
! J(:,i) = (f(x + h*e_i) - f(x)) / h |
| 234 |
!--------------------------------------------------------------------------- |
| 235 |
subroutine compute_fd_jacobian(ws, x, fvec, fcn, h, nfev, iflag, verbosity) |
| 236 |
type(broyden_workspace), intent(inout) :: ws |
| 237 |
real(real64), intent(in) :: x(:), fvec(:) |
| 238 |
procedure(residual_fcn) :: fcn |
| 239 |
real(real64), intent(in) :: h |
| 240 |
integer(int32), intent(inout) :: nfev |
| 241 |
integer, intent(out) :: iflag |
| 242 |
integer(int32), intent(in), optional :: verbosity |
| 243 |
|
| 244 |
integer(int32) :: i, n |
| 245 |
real(real64) :: x_orig |
| 246 |
character(len=256) :: msg |
| 247 |
|
| 248 |
n = ws%n_vars |
| 249 |
iflag = 0 |
| 250 |
|
| 251 |
do i = 1, n |
| 252 |
! Perturb x(i) |
| 253 |
x_orig = x(i) |
| 254 |
ws%fvec_new(1:n) = x ! Use fvec_new as temporary x_pert |
| 255 |
ws%fvec_new(i) = x_orig + h |
| 256 |
|
| 257 |
! Evaluate f(x + h*e_i) |
| 258 |
call fcn(n, ws%fvec_new, ws%df, iflag) |
| 259 |
nfev = nfev + 1 |
| 260 |
|
| 261 |
if (iflag < 0) return |
| 262 |
|
| 263 |
! Forward difference: J(:,i) = (f_pert - f) / h |
| 264 |
ws%J(:, i) = (ws%df - fvec) / h |
| 265 |
|
| 266 |
! Debug output |
| 267 |
if (present(verbosity)) then |
| 268 |
if (verbosity == 2) then |
| 269 |
write(msg, '(A,I2,A,ES12.5,A,ES12.5,A,ES12.5,A,ES12.5)') & |
| 270 |
' FD[', i, ']: x=', x_orig, ', h=', h, & |
| 271 |
', f=', fvec(1), ', f_pert=', ws%df(1) |
| 272 |
call mexPrintf(trim(msg)//NEW_LINE('A')) |
| 273 |
write(msg, '(A,ES12.5)') ' J=', ws%J(1, i) |
| 274 |
call mexPrintf(trim(msg)//NEW_LINE('A')) |
| 275 |
end if |
| 276 |
end if |
| 277 |
end do |
| 278 |
|
| 279 |
end subroutine compute_fd_jacobian |
| 280 |
|
| 281 |
!--------------------------------------------------------------------------- |
| 282 |
! Broyden rank-1 update |
| 283 |
! |
| 284 |
! J = J + outer((df - J*dx) / ||dx||^2, dx) |
| 285 |
! |
| 286 |
! where df = fvec_new - fvec (already computed in ws%df) |
| 287 |
! dx = Newton step (in ws%dx) |
| 288 |
!--------------------------------------------------------------------------- |
| 289 |
subroutine broyden_update(ws) |
| 290 |
type(broyden_workspace), intent(inout) :: ws |
| 291 |
|
| 292 |
integer(int32) :: n |
| 293 |
real(real64) :: dx_norm_sq, scale |
| 294 |
|
| 295 |
n = ws%n_vars |
| 296 |
|
| 297 |
! Compute ||dx||^2 |
| 298 |
dx_norm_sq = dot_product(ws%dx, ws%dx) |
| 299 |
|
| 300 |
! Skip update if dx is too small to avoid division by zero |
| 301 |
if (dx_norm_sq < 1.0e-30_real64) return |
| 302 |
|
| 303 |
! Compute correction = J * dx using BLAS (store in fvec_new as temp) |
| 304 |
ws%fvec_new = 0.0_real64 |
| 305 |
call matvecmul_add('N', 1.0_real64, ws%J, ws%dx, 0.0_real64, ws%fvec_new) |
| 306 |
|
| 307 |
! fvec_new = (df - J*dx) / ||dx||^2 |
| 308 |
scale = 1.0_real64 / dx_norm_sq |
| 309 |
ws%fvec_new = (ws%df - ws%fvec_new) * scale |
| 310 |
|
| 311 |
! Rank-1 update: J = J + outer(fvec_new, dx) using BLAS dger |
| 312 |
! dger: A = alpha * x * y^T + A |
| 313 |
call rank_one_update(n, n, 1.0_real64, ws%fvec_new, ws%dx, ws%J) |
| 314 |
|
| 315 |
end subroutine broyden_update |
| 316 |
|
| 317 |
end module broyden_solver |