Compute locally constrained steps s
and t
, and phi(alpha)
.
Type | Intent | Optional | Attributes | Name | ||
---|---|---|---|---|---|---|
integer, | intent(in) | :: | n |
Number of observations. |
||
integer, | intent(in) | :: | m |
Number of columns of data in the explanatory variable. |
||
integer, | intent(in) | :: | np |
Number of function parameters. |
||
integer, | intent(in) | :: | q |
Number of responses per observation. |
||
integer, | intent(in) | :: | npp |
Number of function parameters being estimated. |
||
real(kind=wp), | intent(in) | :: | f(n,q) |
Weighted estimated values of |
||
real(kind=wp), | intent(in) | :: | fjacb(n,np,q) |
Jacobian with respect to |
||
real(kind=wp), | intent(in) | :: | fjacd(n,m,q) |
Jacobian with respect to |
||
real(kind=wp), | intent(in) | :: | wd(ldwd,ld2wd,m) |
Squared |
||
integer, | intent(in) | :: | ldwd |
Leading dimension of array |
||
integer, | intent(in) | :: | ld2wd |
Second dimension of array |
||
real(kind=wp), | intent(in) | :: | ss(np) |
Scaling values for the unfixed |
||
real(kind=wp), | intent(in) | :: | tt(ldtt,m) |
Scaling values for |
||
integer, | intent(in) | :: | ldtt |
Leading dimension of array |
||
real(kind=wp), | intent(in) | :: | delta(n,m) |
Estimated errors in the explanatory variables. |
||
real(kind=wp), | intent(in) | :: | alpha |
Levenberg-Marquardt parameter. |
||
real(kind=wp), | intent(in) | :: | epsfcn |
Function's precision. |
||
logical, | intent(in) | :: | isodr |
Variable designating whether the solution is by ODR ( |
||
real(kind=wp), | intent(out) | :: | tfjacb(n,q,np) |
Array |
||
real(kind=wp), | intent(out) | :: | omega(q,q) |
Array defined such that:
|
||
real(kind=wp), | intent(out) | :: | u(np) |
Approximate null vector for |
||
real(kind=wp), | intent(out) | :: | qraux(np) |
Array required to recover the orthogonal part of the Q-R decomposition. |
||
integer, | intent(out) | :: | kpvt(np) |
Pivot vector. |
||
real(kind=wp), | intent(out) | :: | s(np) |
Step for |
||
real(kind=wp), | intent(out) | :: | t(n,m) |
Step for |
||
real(kind=wp), | intent(out) | :: | phi |
Difference between the norm of the scaled step and the trust region diameter. |
||
integer, | intent(out) | :: | irank |
Rank deficiency of the Jacobian wrt |
||
real(kind=wp), | intent(out) | :: | rcond |
Approximate reciprocal condition number of |
||
logical, | intent(in) | :: | forvcv |
Variable designating whether this subroutine was called to set up for the
covariance matrix computations ( |
||
real(kind=wp), | intent(out) | :: | wrk1(n,q,m) |
A work array of |
||
real(kind=wp), | intent(out) | :: | wrk2(n,q) |
A work array of |
||
real(kind=wp), | intent(out) | :: | wrk3(np) |
A work array of |
||
real(kind=wp), | intent(out) | :: | wrk4(m,m) |
A work array of |
||
real(kind=wp), | intent(out) | :: | wrk5(m) |
A work array of |
||
real(kind=wp), | intent(out) | :: | wrk(lwrk) |
A work array of |
||
integer, | intent(in) | :: | lwrk |
Length of vector |
||
integer, | intent(inout) | :: | istopc |
Variable designating whether the computations were stopped due to a numerical
error within subroutine |
Type | Visibility | Attributes | Name | Initial | |||
---|---|---|---|---|---|---|---|
real(kind=wp), | public | :: | co | ||||
real(kind=wp), | public | :: | si | ||||
real(kind=wp), | public | :: | temp | ||||
integer, | public | :: | i | ||||
integer, | public | :: | imax | ||||
integer, | public | :: | inf | ||||
integer, | public | :: | ipvt | ||||
integer, | public | :: | j | ||||
integer, | public | :: | k | ||||
integer, | public | :: | k1 | ||||
integer, | public | :: | k2 | ||||
integer, | public | :: | kp | ||||
integer, | public | :: | l | ||||
logical, | public | :: | elim | ||||
real(kind=wp), | public | :: | dum(2) |
subroutine lcstep & (n, m, np, q, npp, & f, fjacb, fjacd, & wd, ldwd, ld2wd, ss, tt, ldtt, delta, & alpha, epsfcn, isodr, & tfjacb, omega, u, qraux, kpvt, & s, t, phi, irank, rcond, forvcv, & wrk1, wrk2, wrk3, wrk4, wrk5, wrk, lwrk, istopc) !! Compute locally constrained steps `s` and `t`, and `phi(alpha)`. ! @note: This is one of the most time-consuming subroutines in ODRPACK (~25% of total). use odrpack_kinds, only: zero, one use linpack, only: dchex, dqrdc, dqrsl, dtrco, dtrsl use blas_interfaces, only: drot, drotg integer, intent(in) :: n !! Number of observations. integer, intent(in) :: m !! Number of columns of data in the explanatory variable. integer, intent(in) :: np !! Number of function parameters. integer, intent(in) :: q !! Number of responses per observation. integer, intent(in) :: npp !! Number of function parameters being estimated. real(wp), intent(in) :: f(n, q) !! Weighted estimated values of `epsilon`. real(wp), intent(in) :: fjacb(n, np, q) !! Jacobian with respect to `beta`. real(wp), intent(in) :: fjacd(n, m, q) !! Jacobian with respect to `delta`. real(wp), intent(in) :: wd(ldwd, ld2wd, m) !! Squared `delta` weights. integer, intent(in) :: ldwd !! Leading dimension of array `wd`. integer, intent(in) :: ld2wd !! Second dimension of array `wd`. real(wp), intent(in) :: ss(np) !! Scaling values for the unfixed `beta`s. real(wp), intent(in) :: tt(ldtt, m) !! Scaling values for `delta`. integer, intent(in) :: ldtt !! Leading dimension of array `tt`. real(wp), intent(in) :: delta(n, m) !! Estimated errors in the explanatory variables. real(wp), intent(in) :: alpha !! Levenberg-Marquardt parameter. real(wp), intent(in) :: epsfcn !! Function's precision. logical, intent(in) :: isodr !! Variable designating whether the solution is by ODR (`.true.`) or by OLS (`.false.`). real(wp), intent(out) :: tfjacb(n, q, np) !! Array `omega*fjacb`. real(wp), intent(out) :: omega(q, q) !! Array defined such that: !! `omega*trans(omega) = inv(I + fjacd*inv(e)*trans(fjacd)) !! = (I - fjacd*inv(p)*trans(fjacd))` !! where `e = d**2 + alpha*tt**2` and !! `p = trans(fjacd)*fjacd + d**2 + alpha*tt**2`. real(wp), intent(out) :: u(np) !! Approximate null vector for `tfjacb`. real(wp), intent(out) :: qraux(np) !! Array required to recover the orthogonal part of the Q-R decomposition. integer, intent(out) :: kpvt(np) !! Pivot vector. real(wp), intent(out) :: s(np) !! Step for `beta`. real(wp), intent(out) :: t(n, m) !! Step for `delta`. real(wp), intent(out) :: phi !! Difference between the norm of the scaled step and the trust region diameter. integer, intent(out) :: irank !! Rank deficiency of the Jacobian wrt `beta`. real(wp), intent(out) :: rcond !! Approximate reciprocal condition number of `tfjacb`. logical, intent(in) :: forvcv !! Variable designating whether this subroutine was called to set up for the !! covariance matrix computations (`forvcv = .true.`) or not (`forvcv = .false.`). real(wp), intent(out) :: wrk1(n, q, m) !! A work array of `(n, q, m)` elements. real(wp), intent(out) :: wrk2(n, q) !! A work array of `(n, q)` elements. real(wp), intent(out) :: wrk3(np) !! A work array of `(np)` elements. real(wp), intent(out) :: wrk4(m, m) !! A work array of `(m, m)` elements. real(wp), intent(out) :: wrk5(m) !! A work array of `(m)` elements. real(wp), intent(out) :: wrk(lwrk) !! A work array of `(lwrk)` elements, _equivalenced_ to `wrk1` and `wrk2`. integer, intent(in) :: lwrk !! Length of vector `wrk`. integer, intent(inout) :: istopc !! Variable designating whether the computations were stopped due to a numerical !! error within subroutine `dodstp`. ! Local scalars real(wp) :: co, si, temp integer :: i, imax, inf, ipvt, j, k, k1, k2, kp, l logical :: elim ! Local arrays real(wp) :: dum(2) ! Variable definitions (alphabetically) ! CO: The cosine from the plane rotation. ! DUM: A dummy array. ! ELIM: The variable designating whether columns of the Jacobian ! I: An indexing variable. ! IMAX: The index of the element of U having the largest absolute value. ! INF: The return code from LINPACK routines. ! IPVT: The variable designating whether pivoting is to be done. ! J: An indexing variable. ! K: An indexing variable. ! K1: An indexing variable. ! K2: An indexing variable. ! KP: The rank of the Jacobian wrt BETA. ! L: An indexing variable. ! SI: The sine from the plane rotation. ! TEMP: A temporary storage LOCATION. ! Compute loop parameters which depend on weight structure ! Set up KPVT if ALPHA = 0 if (alpha == zero) then kp = npp do k = 1, np kpvt(k) = k end do else if (npp >= 1) then kp = npp - irank else kp = npp end if end if if (isodr) then ! T = WD * DELTA = D*G2 call scale_mat(n, m, wd, ldwd, ld2wd, delta, t) do i = 1, n ! Compute WRK4, such that TRANS(WRK4)*WRK4 = E = (D**2 + ALPHA*TT**2) call esubi(n, m, wd, ldwd, ld2wd, alpha, tt, ldtt, i, wrk4) call fctr(.false., wrk4, m, m, inf) if (inf /= 0) then istopc = 60000 return end if ! Compute OMEGA, such that ! trans(OMEGA)*OMEGA = I+FJACD*inv(E)*trans(FJACD) ! inv(trans(OMEGA)*OMEGA) = I-FJACD*inv(P)*trans(FJACD) call vevtr(m, q, i, fjacd, n, m, wrk4, m, wrk1, n, q, omega, q, wrk5) do l = 1, q omega(l, l) = one + omega(l, l) end do call fctr(.false., omega, q, q, inf) if (inf /= 0) then istopc = 60000 return end if ! Compute WRK1 = trans(FJACD)*(I-FJACD*inv(P)*trans(JFACD)) ! = trans(FJACD)*inv(trans(OMEGA)*OMEGA) do j = 1, m wrk1(i, :, j) = fjacd(i, j, :) call solve_trl(q, omega, q, wrk1(i, 1:q, j), 4) call solve_trl(q, omega, q, wrk1(i, 1:q, j), 2) end do ! Compute WRK5 = inv(E)*D*G2 wrk5 = t(i, :) call solve_trl(m, wrk4, m, wrk5, 4) call solve_trl(m, wrk4, m, wrk5, 2) ! Compute TFJACB = inv(trans(OMEGA))*FJACB do k = 1, kp tfjacb(i, :, k) = fjacb(i, kpvt(k), :) call solve_trl(q, omega, q, tfjacb(i, 1:q, k), 4) if (ss(1) > zero) then tfjacb(i, :, k) = tfjacb(i, :, k)/ss(kpvt(k)) else tfjacb(i, :, k) = tfjacb(i, :, k)/abs(ss(1)) end if end do ! Compute WRK2 = (V*inv(E)*D**2*G2 - G1) do l = 1, q wrk2(i, l) = dot_product(fjacd(i, :, l), wrk5) - f(i, l) end do ! Compute WRK2 = inv(trans(OMEGA))*(V*inv(E)*D**2*G2 - G1) call solve_trl(q, omega, q, wrk2(i, 1:q), 4) end do else do l = 1, q do i = 1, n do k = 1, kp tfjacb(i, l, k) = fjacb(i, kpvt(k), l) if (ss(1) > zero) then tfjacb(i, l, k) = tfjacb(i, l, k)/ss(kpvt(k)) else tfjacb(i, l, k) = tfjacb(i, l, k)/abs(ss(1)) end if end do wrk2(i, l) = -f(i, l) end do end do end if ! Compute S ! Do QR factorization (with column pivoting of TFJACB if ALPHA = 0) if (alpha == zero) then ipvt = 1 kpvt = 0 else ipvt = 0 end if call dqrdc(tfjacb, n*q, n*q, kp, qraux, kpvt, wrk3, ipvt) call dqrsl(tfjacb, n*q, n*q, kp, qraux, wrk2, dum, wrk2, dum, dum, dum, 1000, inf) if (inf /= 0) then istopc = 60000 return end if ! Eliminate alpha part using givens rotations if (alpha /= zero) then s(1:npp) = zero do k1 = 1, kp wrk3(1:kp) = zero wrk3(k1) = sqrt(alpha) do k2 = k1, kp call drotg(tfjacb(k2, 1, k2), wrk3(k2), co, si) if (kp - k2 >= 1) then call drot(kp - k2, tfjacb(k2, 1, k2 + 1), n*q, wrk3(k2 + 1), 1, co, si) end if temp = co*wrk2(k2, 1) + si*s(kpvt(k1)) s(kpvt(k1)) = -si*wrk2(k2, 1) + co*s(kpvt(k1)) wrk2(k2, 1) = temp end do end do end if ! Compute solution - eliminate variables if necessary if (npp >= 1) then if (alpha == zero) then kp = npp elim = .true. do while (elim .and. kp >= 1) ! Estimate RCOND - U will contain approx null vector call dtrco(tfjacb, n*q, kp, rcond, u, 1) if (rcond <= epsfcn) then elim = .true. imax = maxloc(u(1:kp), dim=1) ! IMAX is the column to remove - use DCHEX and fix KPVT if (imax /= kp) then call dchex(tfjacb, n*q, kp, imax, kp, wrk2, n*q, 1, qraux, wrk3, 2) k = kpvt(imax) do i = imax, kp - 1 kpvt(i) = kpvt(i + 1) end do kpvt(kp) = k end if kp = kp - 1 else elim = .false. end if end do irank = npp - kp end if end if if (forvcv) return ! Backsolve and unscramble if (npp >= 1) then do i = kp + 1, npp wrk2(i, 1) = zero end do if (kp >= 1) then call dtrsl(tfjacb, n*q, kp, wrk2, 01, inf) if (inf /= 0) then istopc = 60000 return end if end if do i = 1, npp if (ss(1) > zero) then s(kpvt(i)) = wrk2(i, 1)/ss(kpvt(i)) else s(kpvt(i)) = wrk2(i, 1)/abs(ss(1)) end if end do end if if (isodr) then ! NOTE: T and WRK1 have been initialized above, ! where T = WD * DELTA = D*G2 ! WRK1 = trans(FJACD)*(I-FJACD*inv(P)*trans(JFACD)) do i = 1, n ! Compute WRK4, such that trans(WRK4)*WRK4 = E = (D**2 + ALPHA*TT**2) call esubi(n, m, wd, ldwd, ld2wd, alpha, tt, ldtt, i, wrk4) call fctr(.false., wrk4, m, m, inf) if (inf /= 0) then istopc = 60000 return end if ! Compute WRK5 = inv(E)*D*G2 wrk5 = t(i, :) call solve_trl(m, wrk4, m, wrk5, 4) call solve_trl(m, wrk4, m, wrk5, 2) do l = 1, q wrk2(i, l) = f(i, l) & + dot_product(fjacb(i, 1:npp, l), s(1:npp)) & - dot_product(fjacd(i, :, l), wrk5) end do do j = 1, m wrk5(j) = dot_product(wrk1(i, :, j), wrk2(i, :)) t(i, j) = -(wrk5(j) + t(i, j)) end do call solve_trl(m, wrk4, m, t(i, 1:m), 4) call solve_trl(m, wrk4, m, t(i, 1:m), 2) end do end if ! Compute PHI(ALPHA) from scaled S and T call scale_mat(npp, 1, ss, npp, 1, s, wrk(1:npp)) if (isodr) then call scale_mat(n, m, tt, ldtt, 1, t, wrk(npp + 1:npp + 1 + n*m - 1)) phi = norm2(wrk(1:npp + n*m)) else phi = norm2(wrk(1:npp)) end if end subroutine lcstep