xref: /petsc/src/tao/leastsquares/impls/pounders/pounders.c (revision 7efe37a1cedd385a2f501b843d47cdf14dfb49ea)

#include <../src/tao/leastsquares/impls/pounders/pounders.h>

static PetscErrorCode pounders_h(Tao subtao, Vec v, Mat H, Mat Hpre, void *ctx)
{
  PetscFunctionBegin;
  PetscFunctionReturn(0);
}

static PetscErrorCode pounders_fg(Tao subtao, Vec x, PetscReal *f, Vec g, void *ctx)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)ctx;
  PetscReal     d1, d2;

  PetscFunctionBegin;
  /* g = A*x  (add b later)*/
  PetscCall(MatMult(mfqP->subH, x, g));

  /* f = 1/2 * x'*(Ax) + b'*x  */
  PetscCall(VecDot(x, g, &d1));
  PetscCall(VecDot(mfqP->subb, x, &d2));
  *f = 0.5 * d1 + d2;

  /* now  g = g + b */
  PetscCall(VecAXPY(g, 1.0, mfqP->subb));
  PetscFunctionReturn(0);
}

static PetscErrorCode pounders_feval(Tao tao, Vec x, Vec F, PetscReal *fsum)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
  PetscInt      i, row, col;
  PetscReal     fr, fc;

  PetscFunctionBegin;
  PetscCall(TaoComputeResidual(tao, x, F));
  if (tao->res_weights_v) {
    PetscCall(VecPointwiseMult(mfqP->workfvec, tao->res_weights_v, F));
    PetscCall(VecDot(mfqP->workfvec, mfqP->workfvec, fsum));
  } else if (tao->res_weights_w) {
    *fsum = 0;
    for (i = 0; i < tao->res_weights_n; i++) {
      row = tao->res_weights_rows[i];
      col = tao->res_weights_cols[i];
      PetscCall(VecGetValues(F, 1, &row, &fr));
      PetscCall(VecGetValues(F, 1, &col, &fc));
      *fsum += tao->res_weights_w[i] * fc * fr;
    }
  } else {
    PetscCall(VecDot(F, F, fsum));
  }
  PetscCall(PetscInfo(tao, "Least-squares residual norm: %20.19e\n", (double)*fsum));
  PetscCheck(!PetscIsInfOrNanReal(*fsum), PETSC_COMM_SELF, PETSC_ERR_USER, "User provided compute function generated Inf or NaN");
  PetscFunctionReturn(0);
}

static PetscErrorCode gqtwrap(Tao tao, PetscReal *gnorm, PetscReal *qmin)
{
#if defined(PETSC_USE_REAL_SINGLE)
  PetscReal atol = 1.0e-5;
#else
  PetscReal atol = 1.0e-10;
#endif
  PetscInt      info, its;
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;

  PetscFunctionBegin;
  if (!mfqP->usegqt) {
    PetscReal maxval;
    PetscInt  i, j;

    PetscCall(VecSetValues(mfqP->subb, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
    PetscCall(VecAssemblyBegin(mfqP->subb));
    PetscCall(VecAssemblyEnd(mfqP->subb));

    PetscCall(VecSet(mfqP->subx, 0.0));

    PetscCall(VecSet(mfqP->subndel, -1.0));
    PetscCall(VecSet(mfqP->subpdel, +1.0));

    /* Complete the lower triangle of the Hessian matrix */
    for (i = 0; i < mfqP->n; i++) {
      for (j = i + 1; j < mfqP->n; j++) mfqP->Hres[j + mfqP->n * i] = mfqP->Hres[mfqP->n * j + i];
    }
    PetscCall(MatSetValues(mfqP->subH, mfqP->n, mfqP->indices, mfqP->n, mfqP->indices, mfqP->Hres, INSERT_VALUES));
    PetscCall(MatAssemblyBegin(mfqP->subH, MAT_FINAL_ASSEMBLY));
    PetscCall(MatAssemblyEnd(mfqP->subH, MAT_FINAL_ASSEMBLY));

    PetscCall(TaoResetStatistics(mfqP->subtao));
    /* PetscCall(TaoSetTolerances(mfqP->subtao,*gnorm,*gnorm,PETSC_DEFAULT)); */
    /* enforce bound constraints -- experimental */
    if (tao->XU && tao->XL) {
      PetscCall(VecCopy(tao->XU, mfqP->subxu));
      PetscCall(VecAXPY(mfqP->subxu, -1.0, tao->solution));
      PetscCall(VecScale(mfqP->subxu, 1.0 / mfqP->delta));
      PetscCall(VecCopy(tao->XL, mfqP->subxl));
      PetscCall(VecAXPY(mfqP->subxl, -1.0, tao->solution));
      PetscCall(VecScale(mfqP->subxl, 1.0 / mfqP->delta));

      PetscCall(VecPointwiseMin(mfqP->subxu, mfqP->subxu, mfqP->subpdel));
      PetscCall(VecPointwiseMax(mfqP->subxl, mfqP->subxl, mfqP->subndel));
    } else {
      PetscCall(VecCopy(mfqP->subpdel, mfqP->subxu));
      PetscCall(VecCopy(mfqP->subndel, mfqP->subxl));
    }
    /* Make sure xu > xl */
    PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
    PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
    PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
    PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "upper bound < lower bound in subproblem");
    /* Make sure xu > tao->solution > xl */
    PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
    PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
    PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
    PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess < lower bound in subproblem");

    PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
    PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
    PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
    PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess > upper bound in subproblem");

    PetscCall(TaoSolve(mfqP->subtao));
    PetscCall(TaoGetSolutionStatus(mfqP->subtao, NULL, qmin, NULL, NULL, NULL, NULL));

    /* test bounds post-solution*/
    PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
    PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
    PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
    if (maxval > 1e-5) {
      PetscCall(PetscInfo(tao, "subproblem solution < lower bound\n"));
      tao->reason = TAO_DIVERGED_TR_REDUCTION;
    }

    PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
    PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
    PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
    if (maxval > 1e-5) {
      PetscCall(PetscInfo(tao, "subproblem solution > upper bound\n"));
      tao->reason = TAO_DIVERGED_TR_REDUCTION;
    }
  } else {
    gqt(mfqP->n, mfqP->Hres, mfqP->n, mfqP->Gres, 1.0, mfqP->gqt_rtol, atol, mfqP->gqt_maxits, gnorm, qmin, mfqP->Xsubproblem, &info, &its, mfqP->work, mfqP->work2, mfqP->work3);
  }
  *qmin *= -1;
  PetscFunctionReturn(0);
}

static PetscErrorCode pounders_update_res(Tao tao)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
  PetscInt      i, row, col;
  PetscBLASInt  blasn = mfqP->n, blasn2 = blasn * blasn, blasm = mfqP->m, ione = 1;
  PetscReal     zero = 0.0, one = 1.0, wii, factor;

  PetscFunctionBegin;
  for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] = 0;
  for (i = 0; i < mfqP->n * mfqP->n; i++) mfqP->Hres[i] = 0;

  /* Compute Gres= sum_ij[wij * (cjgi + cigj)] */
  if (tao->res_weights_v) {
    /* Vector(diagonal) weights: gres = sum_i(wii*ci*gi) */
    for (i = 0; i < mfqP->m; i++) {
      PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &factor));
      factor = factor * mfqP->C[i];
      PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * i], &ione, mfqP->Gres, &ione));
    }

    /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
    /* vector(diagonal weights) Hres = sum_i(wii*(ci*Hi + gi * gi')*/
    for (i = 0; i < mfqP->m; i++) {
      PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &wii));
      if (tao->niter > 1) {
        factor = wii * mfqP->C[i];
        /* add wii * ci * Hi */
        PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
      }
      /* add wii * gi * gi' */
      PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &wii, &mfqP->Fdiff[blasn * i], &blasn, &mfqP->Fdiff[blasn * i], &blasn, &one, mfqP->Hres, &blasn));
    }
  } else if (tao->res_weights_w) {
    /* General case: .5 * Gres= sum_ij[wij * (cjgi + cigj)] */
    for (i = 0; i < tao->res_weights_n; i++) {
      row = tao->res_weights_rows[i];
      col = tao->res_weights_cols[i];

      factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
      PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * row], &ione, mfqP->Gres, &ione));
      factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
      PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * col], &ione, mfqP->Gres, &ione));
    }

    /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
    /* .5 * sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
    for (i = 0; i < tao->res_weights_n; i++) {
      row    = tao->res_weights_rows[i];
      col    = tao->res_weights_cols[i];
      factor = tao->res_weights_w[i] / 2.0;
      /* add wij * gi gj' + wij * gj gi' */
      PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * row], &blasn, &mfqP->Fdiff[blasn * col], &blasn, &one, mfqP->Hres, &blasn));
      PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * col], &blasn, &mfqP->Fdiff[blasn * row], &blasn, &one, mfqP->Hres, &blasn));
    }
    if (tao->niter > 1) {
      for (i = 0; i < tao->res_weights_n; i++) {
        row = tao->res_weights_rows[i];
        col = tao->res_weights_cols[i];

        /* add  wij*cj*Hi */
        factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
        PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[row], &blasm, mfqP->Hres, &ione));

        /* add wij*ci*Hj */
        factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
        PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[col], &blasm, mfqP->Hres, &ione));
      }
    }
  } else {
    /* Default: Gres= sum_i[cigi] = G*c' */
    PetscCall(PetscInfo(tao, "Identity weights\n"));
    PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->C, &ione, &zero, mfqP->Gres, &ione));

    /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
    /*  Hres = G*G' + 0.5 sum {F(xkin,i)*H(:,:,i)}  */
    PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->Fdiff, &blasn, &zero, mfqP->Hres, &blasn));

    /* sum(F(xkin,i)*H(:,:,i)) */
    if (tao->niter > 1) {
      for (i = 0; i < mfqP->m; i++) {
        factor = mfqP->C[i];
        PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
      }
    }
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode phi2eval(PetscReal *x, PetscInt n, PetscReal *phi)
{
  /* Phi = .5*[x(1)^2  sqrt(2)*x(1)*x(2) ... sqrt(2)*x(1)*x(n) ... x(2)^2 sqrt(2)*x(2)*x(3) .. x(n)^2] */
  PetscInt  i, j, k;
  PetscReal sqrt2 = PetscSqrtReal(2.0);

  PetscFunctionBegin;
  j = 0;
  for (i = 0; i < n; i++) {
    phi[j] = 0.5 * x[i] * x[i];
    j++;
    for (k = i + 1; k < n; k++) {
      phi[j] = x[i] * x[k] / sqrt2;
      j++;
    }
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode getquadpounders(TAO_POUNDERS *mfqP)
{
  /* Computes the parameters of the quadratic Q(x) = c + g'*x + 0.5*x*G*x'
   that satisfies the interpolation conditions Q(X[:,j]) = f(j)
   for j=1,...,m and with a Hessian matrix of least Frobenius norm */

  /* NB --we are ignoring c */
  PetscInt     i, j, k, num, np = mfqP->nmodelpoints;
  PetscReal    one = 1.0, zero = 0.0, negone = -1.0;
  PetscBLASInt blasnpmax  = mfqP->npmax;
  PetscBLASInt blasnplus1 = mfqP->n + 1;
  PetscBLASInt blasnp     = np;
  PetscBLASInt blasint    = mfqP->n * (mfqP->n + 1) / 2;
  PetscBLASInt blasint2   = np - mfqP->n - 1;
  PetscBLASInt info, ione = 1;
  PetscReal    sqrt2 = PetscSqrtReal(2.0);

  PetscFunctionBegin;
  for (i = 0; i < mfqP->n * mfqP->m; i++) mfqP->Gdel[i] = 0;
  for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; i++) mfqP->Hdel[i] = 0;

  /* factor M */
  PetscCallBLAS("LAPACKgetrf", LAPACKgetrf_(&blasnplus1, &blasnp, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &info));
  PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine getrf returned with value %" PetscBLASInt_FMT, info);

  if (np == mfqP->n + 1) {
    for (i = 0; i < mfqP->npmax - mfqP->n - 1; i++) mfqP->omega[i] = 0.0;
    for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->beta[i] = 0.0;
  } else {
    /* Let Ltmp = (L'*L) */
    PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &blasint2, &blasint2, &blasint, &one, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, &zero, mfqP->L_tmp, &blasint));

    /* factor Ltmp */
    PetscCallBLAS("LAPACKpotrf", LAPACKpotrf_("L", &blasint2, mfqP->L_tmp, &blasint, &info));
    PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine potrf returned with value %" PetscBLASInt_FMT, info);
  }

  for (k = 0; k < mfqP->m; k++) {
    if (np != mfqP->n + 1) {
      /* Solve L'*L*Omega = Z' * RESk*/
      PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasnp, &blasint2, &one, mfqP->Z, &blasnpmax, &mfqP->RES[mfqP->npmax * k], &ione, &zero, mfqP->omega, &ione));
      PetscCallBLAS("LAPACKpotrs", LAPACKpotrs_("L", &blasint2, &ione, mfqP->L_tmp, &blasint, mfqP->omega, &blasint2, &info));
      PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine potrs returned with value %" PetscBLASInt_FMT, info);

      /* Beta = L*Omega */
      PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasint, &blasint2, &one, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, mfqP->omega, &ione, &zero, mfqP->beta, &ione));
    }

    /* solve M'*Alpha = RESk - N'*Beta */
    PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasint, &blasnp, &negone, mfqP->N, &blasint, mfqP->beta, &ione, &one, &mfqP->RES[mfqP->npmax * k], &ione));
    PetscCallBLAS("LAPACKgetrs", LAPACKgetrs_("T", &blasnplus1, &ione, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &mfqP->RES[mfqP->npmax * k], &blasnplus1, &info));
    PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine getrs returned with value %" PetscBLASInt_FMT, info);

    /* Gdel(:,k) = Alpha(2:n+1) */
    for (i = 0; i < mfqP->n; i++) mfqP->Gdel[i + mfqP->n * k] = mfqP->RES[mfqP->npmax * k + i + 1];

    /* Set Hdels */
    num = 0;
    for (i = 0; i < mfqP->n; i++) {
      /* H[i,i,k] = Beta(num) */
      mfqP->Hdel[(i * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num];
      num++;
      for (j = i + 1; j < mfqP->n; j++) {
        /* H[i,j,k] = H[j,i,k] = Beta(num)/sqrt(2) */
        mfqP->Hdel[(j * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
        mfqP->Hdel[(i * mfqP->n + j) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
        num++;
      }
    }
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode morepoints(TAO_POUNDERS *mfqP)
{
  /* Assumes mfqP->model_indices[0]  is minimum index
   Finishes adding points to mfqP->model_indices (up to npmax)
   Computes L,Z,M,N
   np is actual number of points in model (should equal npmax?) */
  PetscInt         point, i, j, offset;
  PetscInt         reject;
  PetscBLASInt     blasn = mfqP->n, blasnpmax = mfqP->npmax, blasnplus1 = mfqP->n + 1, info, blasnmax = mfqP->nmax, blasint, blasint2, blasnp, blasmaxmn;
  const PetscReal *x;
  PetscReal        normd;

  PetscFunctionBegin;
  /* Initialize M,N */
  for (i = 0; i < mfqP->n + 1; i++) {
    PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
    mfqP->M[(mfqP->n + 1) * i] = 1.0;
    for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * i)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
    PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
    PetscCall(phi2eval(&mfqP->M[1 + ((mfqP->n + 1) * i)], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * i]));
  }

  /* Now we add points until we have npmax starting with the most recent ones */
  point              = mfqP->nHist - 1;
  mfqP->nmodelpoints = mfqP->n + 1;
  while (mfqP->nmodelpoints < mfqP->npmax && point >= 0) {
    /* Reject any points already in the model */
    reject = 0;
    for (j = 0; j < mfqP->n + 1; j++) {
      if (point == mfqP->model_indices[j]) {
        reject = 1;
        break;
      }
    }

    /* Reject if norm(d) >c2 */
    if (!reject) {
      PetscCall(VecCopy(mfqP->Xhist[point], mfqP->workxvec));
      PetscCall(VecAXPY(mfqP->workxvec, -1.0, mfqP->Xhist[mfqP->minindex]));
      PetscCall(VecNorm(mfqP->workxvec, NORM_2, &normd));
      normd /= mfqP->delta;
      if (normd > mfqP->c2) reject = 1;
    }
    if (reject) {
      point--;
      continue;
    }

    PetscCall(VecGetArrayRead(mfqP->Xhist[point], &x));
    mfqP->M[(mfqP->n + 1) * mfqP->nmodelpoints] = 1.0;
    for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * mfqP->nmodelpoints)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
    PetscCall(VecRestoreArrayRead(mfqP->Xhist[point], &x));
    PetscCall(phi2eval(&mfqP->M[1 + (mfqP->n + 1) * mfqP->nmodelpoints], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * (mfqP->nmodelpoints)]));

    /* Update QR factorization */
    /* Copy M' to Q_tmp */
    for (i = 0; i < mfqP->n + 1; i++) {
      for (j = 0; j < mfqP->npmax; j++) mfqP->Q_tmp[j + mfqP->npmax * i] = mfqP->M[i + (mfqP->n + 1) * j];
    }
    blasnp = mfqP->nmodelpoints + 1;
    /* Q_tmp,R = qr(M') */
    blasmaxmn = PetscMax(mfqP->m, mfqP->n + 1);
    PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->mwork, &blasmaxmn, &info));
    PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine geqrf returned with value %" PetscBLASInt_FMT, info);

    /* Reject if min(svd(N*Q(:,n+2:np+1)) <= theta2 */
    /* L = N*Qtmp */
    blasint2 = mfqP->n * (mfqP->n + 1) / 2;
    /* Copy N to L_tmp */
    for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2 * mfqP->npmax; i++) mfqP->L_tmp[i] = mfqP->N[i];
    /* Copy L_save to L_tmp */

    /* L_tmp = N*Qtmp' */
    PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasint2, &blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->L_tmp, &blasint2, mfqP->npmaxwork, &blasnmax, &info));
    PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine ormqr returned with value %" PetscBLASInt_FMT, info);

    /* Copy L_tmp to L_save */
    for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L_save[i] = mfqP->L_tmp[i];

    /* Get svd for L_tmp(:,n+2:np+1) (L_tmp is modified in process) */
    blasint = mfqP->nmodelpoints - mfqP->n;
    PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
    PetscCallBLAS("LAPACKgesvd", LAPACKgesvd_("N", "N", &blasint2, &blasint, &mfqP->L_tmp[(mfqP->n + 1) * blasint2], &blasint2, mfqP->beta, mfqP->work, &blasn, mfqP->work, &blasn, mfqP->npmaxwork, &blasnmax, &info));
    PetscCall(PetscFPTrapPop());
    PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine gesvd returned with value %" PetscBLASInt_FMT, info);

    if (mfqP->beta[PetscMin(blasint, blasint2) - 1] > mfqP->theta2) {
      /* accept point */
      mfqP->model_indices[mfqP->nmodelpoints] = point;
      /* Copy Q_tmp to Q */
      for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q[i] = mfqP->Q_tmp[i];
      for (i = 0; i < mfqP->npmax; i++) mfqP->tau[i] = mfqP->tau_tmp[i];
      mfqP->nmodelpoints++;
      blasnp = mfqP->nmodelpoints;

      /* Copy L_save to L */
      for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = mfqP->L_save[i];
    }
    point--;
  }

  blasnp = mfqP->nmodelpoints;
  /* Copy Q(:,n+2:np) to Z */
  /* First set Q_tmp to I */
  for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q_tmp[i] = 0.0;
  for (i = 0; i < mfqP->npmax; i++) mfqP->Q_tmp[i + mfqP->npmax * i] = 1.0;

  /* Q_tmp = I * Q */
  PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasnp, &blasnp, &blasnplus1, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));
  PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine ormqr returned with value %" PetscBLASInt_FMT, info);

  /* Copy Q_tmp(:,n+2:np) to Z) */
  offset = mfqP->npmax * (mfqP->n + 1);
  for (i = offset; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Z[i - offset] = mfqP->Q_tmp[i];

  if (mfqP->nmodelpoints == mfqP->n + 1) {
    /* Set L to I_{n+1} */
    for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = 0.0;
    for (i = 0; i < mfqP->n; i++) mfqP->L[(mfqP->n * (mfqP->n + 1) / 2) * i + i] = 1.0;
  }
  PetscFunctionReturn(0);
}

/* Only call from modelimprove, addpoint() needs ->Q_tmp and ->work to be set */
static PetscErrorCode addpoint(Tao tao, TAO_POUNDERS *mfqP, PetscInt index)
{
  PetscFunctionBegin;
  /* Create new vector in history: X[newidx] = X[mfqP->index] + delta*X[index]*/
  PetscCall(VecDuplicate(mfqP->Xhist[0], &mfqP->Xhist[mfqP->nHist]));
  PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, &mfqP->Q_tmp[index * mfqP->npmax], INSERT_VALUES));
  PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
  PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));
  PetscCall(VecAYPX(mfqP->Xhist[mfqP->nHist], mfqP->delta, mfqP->Xhist[mfqP->minindex]));

  /* Project into feasible region */
  if (tao->XU && tao->XL) PetscCall(VecMedian(mfqP->Xhist[mfqP->nHist], tao->XL, tao->XU, mfqP->Xhist[mfqP->nHist]));

  /* Compute value of new vector */
  PetscCall(VecDuplicate(mfqP->Fhist[0], &mfqP->Fhist[mfqP->nHist]));
  CHKMEMQ;
  PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));

  /* Add new vector to model */
  mfqP->model_indices[mfqP->nmodelpoints] = mfqP->nHist;
  mfqP->nmodelpoints++;
  mfqP->nHist++;
  PetscFunctionReturn(0);
}

static PetscErrorCode modelimprove(Tao tao, TAO_POUNDERS *mfqP, PetscInt addallpoints)
{
  /* modeld = Q(:,np+1:n)' */
  PetscInt     i, j, minindex = 0;
  PetscReal    dp, half = 0.5, one = 1.0, minvalue = PETSC_INFINITY;
  PetscBLASInt blasn = mfqP->n, blasnpmax = mfqP->npmax, blask, info;
  PetscBLASInt blas1 = 1, blasnmax = mfqP->nmax;

  PetscFunctionBegin;
  blask = mfqP->nmodelpoints;
  /* Qtmp = I(n x n) */
  for (i = 0; i < mfqP->n; i++) {
    for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i + mfqP->npmax * j] = 0.0;
  }
  for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[j + mfqP->npmax * j] = 1.0;

  /* Qtmp = Q * I */
  PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasn, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));

  for (i = mfqP->nmodelpoints; i < mfqP->n; i++) {
    PetscCallBLAS("BLASdot", dp = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->Gres, &blas1));
    if (dp > 0.0) { /* Model says use the other direction! */
      for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i * mfqP->npmax + j] *= -1;
    }
    /* mfqP->work[i] = Cres+Modeld(i,:)*(Gres+.5*Hres*Modeld(i,:)') */
    for (j = 0; j < mfqP->n; j++) mfqP->work2[j] = mfqP->Gres[j];
    PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &half, mfqP->Hres, &blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, &one, mfqP->work2, &blas1));
    PetscCallBLAS("BLASdot", mfqP->work[i] = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->work2, &blas1));
    if (i == mfqP->nmodelpoints || mfqP->work[i] < minvalue) {
      minindex = i;
      minvalue = mfqP->work[i];
    }
    if (addallpoints != 0) PetscCall(addpoint(tao, mfqP, i));
  }
  if (!addallpoints) PetscCall(addpoint(tao, mfqP, minindex));
  PetscFunctionReturn(0);
}

static PetscErrorCode affpoints(TAO_POUNDERS *mfqP, PetscReal *xmin, PetscReal c)
{
  PetscInt         i, j;
  PetscBLASInt     blasm = mfqP->m, blasj, blask, blasn = mfqP->n, ione = 1, info;
  PetscBLASInt     blasnpmax = mfqP->npmax, blasmaxmn;
  PetscReal        proj, normd;
  const PetscReal *x;

  PetscFunctionBegin;
  for (i = mfqP->nHist - 1; i >= 0; i--) {
    PetscCall(VecGetArrayRead(mfqP->Xhist[i], &x));
    for (j = 0; j < mfqP->n; j++) mfqP->work[j] = (x[j] - xmin[j]) / mfqP->delta;
    PetscCall(VecRestoreArrayRead(mfqP->Xhist[i], &x));
    PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, mfqP->work2, &ione));
    PetscCallBLAS("BLASnrm2", normd = BLASnrm2_(&blasn, mfqP->work, &ione));
    if (normd <= c) {
      blasj = PetscMax((mfqP->n - mfqP->nmodelpoints), 0);
      if (!mfqP->q_is_I) {
        /* project D onto null */
        blask = (mfqP->nmodelpoints);
        PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &ione, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->work2, &ione, mfqP->mwork, &blasm, &info));
        PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "ormqr returned value %" PetscBLASInt_FMT, info);
      }
      PetscCallBLAS("BLASnrm2", proj = BLASnrm2_(&blasj, &mfqP->work2[mfqP->nmodelpoints], &ione));

      if (proj >= mfqP->theta1) { /* add this index to model */
        mfqP->model_indices[mfqP->nmodelpoints] = i;
        mfqP->nmodelpoints++;
        PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, &mfqP->Q_tmp[mfqP->npmax * (mfqP->nmodelpoints - 1)], &ione));
        blask = mfqP->npmax * (mfqP->nmodelpoints);
        PetscCallBLAS("BLAScopy", BLAScopy_(&blask, mfqP->Q_tmp, &ione, mfqP->Q, &ione));
        blask     = mfqP->nmodelpoints;
        blasmaxmn = PetscMax(mfqP->m, mfqP->n);
        PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->mwork, &blasmaxmn, &info));
        PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "geqrf returned value %" PetscBLASInt_FMT, info);
        mfqP->q_is_I = 0;
      }
      if (mfqP->nmodelpoints == mfqP->n) break;
    }
  }

  PetscFunctionReturn(0);
}

static PetscErrorCode TaoSolve_POUNDERS(Tao tao)
{
  TAO_POUNDERS    *mfqP = (TAO_POUNDERS *)tao->data;
  PetscInt         i, ii, j, k, l;
  PetscReal        step = 1.0;
  PetscInt         low, high;
  PetscReal        minnorm;
  PetscReal       *x, *f;
  const PetscReal *xmint, *fmin;
  PetscReal        deltaold;
  PetscReal        gnorm;
  PetscBLASInt     info, ione = 1, iblas;
  PetscBool        valid, same;
  PetscReal        mdec, rho, normxsp;
  PetscReal        one = 1.0, zero = 0.0, ratio;
  PetscBLASInt     blasm, blasn, blasncopy, blasnpmax;
  static PetscBool set = PETSC_FALSE;

  /* n = # of parameters
     m = dimension (components) of function  */
  PetscFunctionBegin;
  PetscCall(PetscCitationsRegister("@article{UNEDF0,\n"
                                   "title = {Nuclear energy density optimization},\n"
                                   "author = {Kortelainen, M.  and Lesinski, T.  and Mor\'e, J.  and Nazarewicz, W.\n"
                                   "          and Sarich, J.  and Schunck, N.  and Stoitsov, M. V. and Wild, S. },\n"
                                   "journal = {Phys. Rev. C},\n"
                                   "volume = {82},\n"
                                   "number = {2},\n"
                                   "pages = {024313},\n"
                                   "numpages = {18},\n"
                                   "year = {2010},\n"
                                   "month = {Aug},\n"
                                   "doi = {10.1103/PhysRevC.82.024313}\n}\n",
                                   &set));
  tao->niter = 0;
  if (tao->XL && tao->XU) {
    /* Check x0 <= XU */
    PetscReal val;

    PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));
    PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->XU));
    PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
    PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 > upper bound");

    /* Check x0 >= xl */
    PetscCall(VecCopy(tao->XL, mfqP->Xhist[0]));
    PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->solution));
    PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
    PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 < lower bound");

    /* Check x0 + delta < XU  -- should be able to get around this eventually */

    PetscCall(VecSet(mfqP->Xhist[0], mfqP->delta));
    PetscCall(VecAXPY(mfqP->Xhist[0], 1.0, tao->solution));
    PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->XU));
    PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
    PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 + delta > upper bound");
  }

  blasm     = mfqP->m;
  blasn     = mfqP->n;
  blasnpmax = mfqP->npmax;
  for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; ++i) mfqP->H[i] = 0;

  PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));

  /* This provides enough information to approximate the gradient of the objective */
  /* using a forward difference scheme. */

  PetscCall(PetscInfo(tao, "Initialize simplex; delta = %10.9e\n", (double)mfqP->delta));
  PetscCall(pounders_feval(tao, mfqP->Xhist[0], mfqP->Fhist[0], &mfqP->Fres[0]));
  mfqP->minindex = 0;
  minnorm        = mfqP->Fres[0];

  PetscCall(VecGetOwnershipRange(mfqP->Xhist[0], &low, &high));
  for (i = 1; i < mfqP->n + 1; ++i) {
    PetscCall(VecCopy(mfqP->Xhist[0], mfqP->Xhist[i]));

    if (i - 1 >= low && i - 1 < high) {
      PetscCall(VecGetArray(mfqP->Xhist[i], &x));
      x[i - 1 - low] += mfqP->delta;
      PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));
    }
    CHKMEMQ;
    PetscCall(pounders_feval(tao, mfqP->Xhist[i], mfqP->Fhist[i], &mfqP->Fres[i]));
    if (mfqP->Fres[i] < minnorm) {
      mfqP->minindex = i;
      minnorm        = mfqP->Fres[i];
    }
  }
  PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
  PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
  PetscCall(PetscInfo(tao, "Finalize simplex; minnorm = %10.9e\n", (double)minnorm));

  /* Gather mpi vecs to one big local vec */

  /* Begin serial code */

  /* Disp[i] = Xi-xmin, i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
  /* Fdiff[i] = (Fi-Fmin)', i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
  /* (Column oriented for blas calls) */
  ii = 0;

  PetscCall(PetscInfo(tao, "Build matrix: %" PetscInt_FMT "\n", (PetscInt)mfqP->size));
  if (1 == mfqP->size) {
    PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
    for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
    PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
    PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
    for (i = 0; i < mfqP->n + 1; i++) {
      if (i == mfqP->minindex) continue;

      PetscCall(VecGetArray(mfqP->Xhist[i], &x));
      for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
      PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));

      PetscCall(VecGetArray(mfqP->Fhist[i], &f));
      for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
      PetscCall(VecRestoreArray(mfqP->Fhist[i], &f));

      mfqP->model_indices[ii++] = i;
    }
    for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
    PetscCall(VecRestoreArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
  } else {
    PetscCall(VecSet(mfqP->localxmin, 0));
    PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));
    PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));

    PetscCall(VecGetArrayRead(mfqP->localxmin, &xmint));
    for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
    PetscCall(VecRestoreArrayRead(mfqP->localxmin, &xmint));

    PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
    PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
    PetscCall(VecGetArrayRead(mfqP->localfmin, &fmin));
    for (i = 0; i < mfqP->n + 1; i++) {
      if (i == mfqP->minindex) continue;

      PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
      PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
      PetscCall(VecGetArray(mfqP->localx, &x));
      for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
      PetscCall(VecRestoreArray(mfqP->localx, &x));

      PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
      PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
      PetscCall(VecGetArray(mfqP->localf, &f));
      for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
      PetscCall(VecRestoreArray(mfqP->localf, &f));

      mfqP->model_indices[ii++] = i;
    }
    for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
    PetscCall(VecRestoreArrayRead(mfqP->localfmin, &fmin));
  }

  /* Determine the initial quadratic models */
  /* G = D(ModelIn,:) \ (F(ModelIn,1:m)-repmat(F(xkin,1:m),n,1)); */
  /* D (nxn) Fdiff (nxm)  => G (nxm) */
  blasncopy = blasn;
  PetscCallBLAS("LAPACKgesv", LAPACKgesv_(&blasn, &blasm, mfqP->Disp, &blasnpmax, mfqP->iwork, mfqP->Fdiff, &blasncopy, &info));
  PetscCall(PetscInfo(tao, "Linear solve return: %" PetscInt_FMT "\n", (PetscInt)info));

  PetscCall(pounders_update_res(tao));

  valid = PETSC_TRUE;

  PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
  PetscCall(VecAssemblyBegin(tao->gradient));
  PetscCall(VecAssemblyEnd(tao->gradient));
  PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
  gnorm *= mfqP->delta;
  PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));

  tao->reason = TAO_CONTINUE_ITERATING;
  PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
  PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
  PetscUseTypeMethod(tao, convergencetest, tao->cnvP);

  mfqP->nHist        = mfqP->n + 1;
  mfqP->nmodelpoints = mfqP->n + 1;
  PetscCall(PetscInfo(tao, "Initial gradient: %20.19e\n", (double)gnorm));

  while (tao->reason == TAO_CONTINUE_ITERATING) {
    PetscReal gnm = 1e-4;
    /* Call general purpose update function */
    PetscTryTypeMethod(tao, update, tao->niter, tao->user_update);
    tao->niter++;
    /* Solve the subproblem min{Q(s): ||s|| <= 1.0} */
    PetscCall(gqtwrap(tao, &gnm, &mdec));
    /* Evaluate the function at the new point */

    for (i = 0; i < mfqP->n; i++) mfqP->work[i] = mfqP->Xsubproblem[i] * mfqP->delta + mfqP->xmin[i];
    PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[mfqP->nHist]));
    PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[mfqP->nHist]));
    PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, mfqP->work, INSERT_VALUES));
    PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
    PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));

    PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));

    rho = (mfqP->Fres[mfqP->minindex] - mfqP->Fres[mfqP->nHist]) / mdec;
    mfqP->nHist++;

    /* Update the center */
    if ((rho >= mfqP->eta1) || (rho > mfqP->eta0 && valid == PETSC_TRUE)) {
      /* Update model to reflect new base point */
      for (i = 0; i < mfqP->n; i++) mfqP->work[i] = (mfqP->work[i] - mfqP->xmin[i]) / mfqP->delta;
      for (j = 0; j < mfqP->m; j++) {
        /* C(j) = C(j) + work*G(:,j) + .5*work*H(:,:,j)*work';
         G(:,j) = G(:,j) + H(:,:,j)*work' */
        for (k = 0; k < mfqP->n; k++) {
          mfqP->work2[k] = 0.0;
          for (l = 0; l < mfqP->n; l++) mfqP->work2[k] += mfqP->H[j + mfqP->m * (k + l * mfqP->n)] * mfqP->work[l];
        }
        for (i = 0; i < mfqP->n; i++) {
          mfqP->C[j] += mfqP->work[i] * (mfqP->Fdiff[i + mfqP->n * j] + 0.5 * mfqP->work2[i]);
          mfqP->Fdiff[i + mfqP->n * j] += mfqP->work2[i];
        }
      }
      /* Cres += work*Gres + .5*work*Hres*work';
       Gres += Hres*work'; */

      PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &one, mfqP->Hres, &blasn, mfqP->work, &ione, &zero, mfqP->work2, &ione));
      for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] += mfqP->work2[i];
      mfqP->minindex = mfqP->nHist - 1;
      minnorm        = mfqP->Fres[mfqP->minindex];
      PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
      /* Change current center */
      PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
      for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
      PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
    }

    /* Evaluate at a model-improving point if necessary */
    if (valid == PETSC_FALSE) {
      mfqP->q_is_I       = 1;
      mfqP->nmodelpoints = 0;
      PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
      if (mfqP->nmodelpoints < mfqP->n) {
        PetscCall(PetscInfo(tao, "Model not valid -- model-improving\n"));
        PetscCall(modelimprove(tao, mfqP, 1));
      }
    }

    /* Update the trust region radius */
    deltaold = mfqP->delta;
    normxsp  = 0;
    for (i = 0; i < mfqP->n; i++) normxsp += mfqP->Xsubproblem[i] * mfqP->Xsubproblem[i];
    normxsp = PetscSqrtReal(normxsp);
    if (rho >= mfqP->eta1 && normxsp > 0.5 * mfqP->delta) {
      mfqP->delta = PetscMin(mfqP->delta * mfqP->gamma1, mfqP->deltamax);
    } else if (valid == PETSC_TRUE) {
      mfqP->delta = PetscMax(mfqP->delta * mfqP->gamma0, mfqP->deltamin);
    }

    /* Compute the next interpolation set */
    mfqP->q_is_I       = 1;
    mfqP->nmodelpoints = 0;
    PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c1 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c1));
    PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
    if (mfqP->nmodelpoints == mfqP->n) {
      valid = PETSC_TRUE;
    } else {
      valid = PETSC_FALSE;
      PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c2 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c2));
      PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c2));
      if (mfqP->n > mfqP->nmodelpoints) {
        PetscCall(PetscInfo(tao, "Model not valid -- adding geometry points\n"));
        PetscCall(modelimprove(tao, mfqP, mfqP->n - mfqP->nmodelpoints));
      }
    }
    for (i = mfqP->nmodelpoints; i > 0; i--) mfqP->model_indices[i] = mfqP->model_indices[i - 1];
    mfqP->nmodelpoints++;
    mfqP->model_indices[0] = mfqP->minindex;
    PetscCall(morepoints(mfqP));
    for (i = 0; i < mfqP->nmodelpoints; i++) {
      PetscCall(VecGetArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
      for (j = 0; j < mfqP->n; j++) mfqP->Disp[i + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / deltaold;
      PetscCall(VecRestoreArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
      PetscCall(VecGetArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
      for (j = 0; j < mfqP->m; j++) {
        for (k = 0; k < mfqP->n; k++) {
          mfqP->work[k] = 0.0;
          for (l = 0; l < mfqP->n; l++) mfqP->work[k] += mfqP->H[j + mfqP->m * (k + mfqP->n * l)] * mfqP->Disp[i + mfqP->npmax * l];
        }
        PetscCallBLAS("BLASdot", mfqP->RES[j * mfqP->npmax + i] = -mfqP->C[j] - BLASdot_(&blasn, &mfqP->Fdiff[j * mfqP->n], &ione, &mfqP->Disp[i], &blasnpmax) - 0.5 * BLASdot_(&blasn, mfqP->work, &ione, &mfqP->Disp[i], &blasnpmax) + f[j]);
      }
      PetscCall(VecRestoreArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
    }

    /* Update the quadratic model */
    PetscCall(PetscInfo(tao, "Get Quad, size: %" PetscInt_FMT ", points: %" PetscInt_FMT "\n", mfqP->n, mfqP->nmodelpoints));
    PetscCall(getquadpounders(mfqP));
    PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
    PetscCallBLAS("BLAScopy", BLAScopy_(&blasm, fmin, &ione, mfqP->C, &ione));
    /* G = G*(delta/deltaold) + Gdel */
    ratio = mfqP->delta / deltaold;
    iblas = blasm * blasn;
    PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->Fdiff, &ione));
    PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Gdel, &ione, mfqP->Fdiff, &ione));
    /* H = H*(delta/deltaold)^2 + Hdel */
    iblas = blasm * blasn * blasn;
    ratio *= ratio;
    PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->H, &ione));
    PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Hdel, &ione, mfqP->H, &ione));

    /* Get residuals */
    PetscCall(pounders_update_res(tao));

    /* Export solution and gradient residual to TAO */
    PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
    PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
    PetscCall(VecAssemblyBegin(tao->gradient));
    PetscCall(VecAssemblyEnd(tao->gradient));
    PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
    gnorm *= mfqP->delta;
    /*  final criticality test */
    PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
    PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
    PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
    /* test for repeated model */
    if (mfqP->nmodelpoints == mfqP->last_nmodelpoints) {
      same = PETSC_TRUE;
    } else {
      same = PETSC_FALSE;
    }
    for (i = 0; i < mfqP->nmodelpoints; i++) {
      if (same) {
        if (mfqP->model_indices[i] == mfqP->last_model_indices[i]) {
          same = PETSC_TRUE;
        } else {
          same = PETSC_FALSE;
        }
      }
      mfqP->last_model_indices[i] = mfqP->model_indices[i];
    }
    mfqP->last_nmodelpoints = mfqP->nmodelpoints;
    if (same && mfqP->delta == deltaold) {
      PetscCall(PetscInfo(tao, "Identical model used in successive iterations\n"));
      tao->reason = TAO_CONVERGED_STEPTOL;
    }
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode TaoSetUp_POUNDERS(Tao tao)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
  PetscInt      i, j;
  IS            isfloc, isfglob, isxloc, isxglob;

  PetscFunctionBegin;
  if (!tao->gradient) PetscCall(VecDuplicate(tao->solution, &tao->gradient));
  if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution, &tao->stepdirection));
  PetscCall(VecGetSize(tao->solution, &mfqP->n));
  PetscCall(VecGetSize(tao->ls_res, &mfqP->m));
  mfqP->c1 = PetscSqrtReal((PetscReal)mfqP->n);
  if (mfqP->npmax == PETSC_DEFAULT) mfqP->npmax = 2 * mfqP->n + 1;
  mfqP->npmax = PetscMin((mfqP->n + 1) * (mfqP->n + 2) / 2, mfqP->npmax);
  mfqP->npmax = PetscMax(mfqP->npmax, mfqP->n + 2);

  PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Xhist));
  PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fhist));
  for (i = 0; i < mfqP->n + 1; i++) {
    PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[i]));
    PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[i]));
  }
  PetscCall(VecDuplicate(tao->solution, &mfqP->workxvec));
  PetscCall(VecDuplicate(tao->ls_res, &mfqP->workfvec));
  mfqP->nHist = 0;

  PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fres));
  PetscCall(PetscMalloc1(mfqP->npmax * mfqP->m, &mfqP->RES));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->work));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->work2));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->work3));
  PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n + 1), &mfqP->mwork));
  PetscCall(PetscMalloc1(mfqP->npmax - mfqP->n - 1, &mfqP->omega));
  PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2, &mfqP->beta));
  PetscCall(PetscMalloc1(mfqP->n + 1, &mfqP->alpha));

  PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->H));
  PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q));
  PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q_tmp));
  PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L));
  PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_tmp));
  PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_save));
  PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->N));
  PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->n + 1), &mfqP->M));
  PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->npmax - mfqP->n - 1), &mfqP->Z));
  PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau));
  PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau_tmp));
  mfqP->nmax = PetscMax(5 * mfqP->npmax, mfqP->n * (mfqP->n + 1) / 2);
  PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxwork));
  PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxiwork));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->xmin));
  PetscCall(PetscMalloc1(mfqP->m, &mfqP->C));
  PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Fdiff));
  PetscCall(PetscMalloc1(mfqP->npmax * mfqP->n, &mfqP->Disp));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->Gres));
  PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Hres));
  PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Gpoints));
  PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->model_indices));
  PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->last_model_indices));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->Xsubproblem));
  PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Gdel));
  PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->Hdel));
  PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n), &mfqP->indices));
  PetscCall(PetscMalloc1(mfqP->n, &mfqP->iwork));
  PetscCall(PetscMalloc1(mfqP->m * mfqP->m, &mfqP->w));
  for (i = 0; i < mfqP->m; i++) {
    for (j = 0; j < mfqP->m; j++) {
      if (i == j) {
        mfqP->w[i + mfqP->m * j] = 1.0;
      } else {
        mfqP->w[i + mfqP->m * j] = 0.0;
      }
    }
  }
  for (i = 0; i < PetscMax(mfqP->m, mfqP->n); i++) mfqP->indices[i] = i;
  PetscCallMPI(MPI_Comm_size(((PetscObject)tao)->comm, &mfqP->size));
  if (mfqP->size > 1) {
    PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localx));
    PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localxmin));
    PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localf));
    PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localfmin));
    PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxloc));
    PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxglob));
    PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfloc));
    PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfglob));

    PetscCall(VecScatterCreate(tao->solution, isxglob, mfqP->localx, isxloc, &mfqP->scatterx));
    PetscCall(VecScatterCreate(tao->ls_res, isfglob, mfqP->localf, isfloc, &mfqP->scatterf));

    PetscCall(ISDestroy(&isxloc));
    PetscCall(ISDestroy(&isxglob));
    PetscCall(ISDestroy(&isfloc));
    PetscCall(ISDestroy(&isfglob));
  }

  if (!mfqP->usegqt) {
    KSP ksp;
    PC  pc;
    PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Xsubproblem, &mfqP->subx));
    PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->subxl));
    PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subb));
    PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subxu));
    PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subpdel));
    PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subndel));
    PetscCall(TaoCreate(PETSC_COMM_SELF, &mfqP->subtao));
    PetscCall(PetscObjectIncrementTabLevel((PetscObject)mfqP->subtao, (PetscObject)tao, 1));
    PetscCall(TaoSetType(mfqP->subtao, TAOBNTR));
    PetscCall(TaoSetOptionsPrefix(mfqP->subtao, "pounders_subsolver_"));
    PetscCall(TaoSetSolution(mfqP->subtao, mfqP->subx));
    PetscCall(TaoSetObjectiveAndGradient(mfqP->subtao, NULL, pounders_fg, (void *)mfqP));
    PetscCall(TaoSetMaximumIterations(mfqP->subtao, mfqP->gqt_maxits));
    PetscCall(TaoSetFromOptions(mfqP->subtao));
    PetscCall(TaoGetKSP(mfqP->subtao, &ksp));
    if (ksp) {
      PetscCall(KSPGetPC(ksp, &pc));
      PetscCall(PCSetType(pc, PCNONE));
    }
    PetscCall(TaoSetVariableBounds(mfqP->subtao, mfqP->subxl, mfqP->subxu));
    PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Hres, &mfqP->subH));
    PetscCall(TaoSetHessian(mfqP->subtao, mfqP->subH, mfqP->subH, pounders_h, (void *)mfqP));
  }
  PetscFunctionReturn(0);
}

static PetscErrorCode TaoDestroy_POUNDERS(Tao tao)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
  PetscInt      i;

  PetscFunctionBegin;
  if (!mfqP->usegqt) {
    PetscCall(TaoDestroy(&mfqP->subtao));
    PetscCall(VecDestroy(&mfqP->subx));
    PetscCall(VecDestroy(&mfqP->subxl));
    PetscCall(VecDestroy(&mfqP->subxu));
    PetscCall(VecDestroy(&mfqP->subb));
    PetscCall(VecDestroy(&mfqP->subpdel));
    PetscCall(VecDestroy(&mfqP->subndel));
    PetscCall(MatDestroy(&mfqP->subH));
  }
  PetscCall(PetscFree(mfqP->Fres));
  PetscCall(PetscFree(mfqP->RES));
  PetscCall(PetscFree(mfqP->work));
  PetscCall(PetscFree(mfqP->work2));
  PetscCall(PetscFree(mfqP->work3));
  PetscCall(PetscFree(mfqP->mwork));
  PetscCall(PetscFree(mfqP->omega));
  PetscCall(PetscFree(mfqP->beta));
  PetscCall(PetscFree(mfqP->alpha));
  PetscCall(PetscFree(mfqP->H));
  PetscCall(PetscFree(mfqP->Q));
  PetscCall(PetscFree(mfqP->Q_tmp));
  PetscCall(PetscFree(mfqP->L));
  PetscCall(PetscFree(mfqP->L_tmp));
  PetscCall(PetscFree(mfqP->L_save));
  PetscCall(PetscFree(mfqP->N));
  PetscCall(PetscFree(mfqP->M));
  PetscCall(PetscFree(mfqP->Z));
  PetscCall(PetscFree(mfqP->tau));
  PetscCall(PetscFree(mfqP->tau_tmp));
  PetscCall(PetscFree(mfqP->npmaxwork));
  PetscCall(PetscFree(mfqP->npmaxiwork));
  PetscCall(PetscFree(mfqP->xmin));
  PetscCall(PetscFree(mfqP->C));
  PetscCall(PetscFree(mfqP->Fdiff));
  PetscCall(PetscFree(mfqP->Disp));
  PetscCall(PetscFree(mfqP->Gres));
  PetscCall(PetscFree(mfqP->Hres));
  PetscCall(PetscFree(mfqP->Gpoints));
  PetscCall(PetscFree(mfqP->model_indices));
  PetscCall(PetscFree(mfqP->last_model_indices));
  PetscCall(PetscFree(mfqP->Xsubproblem));
  PetscCall(PetscFree(mfqP->Gdel));
  PetscCall(PetscFree(mfqP->Hdel));
  PetscCall(PetscFree(mfqP->indices));
  PetscCall(PetscFree(mfqP->iwork));
  PetscCall(PetscFree(mfqP->w));
  for (i = 0; i < mfqP->nHist; i++) {
    PetscCall(VecDestroy(&mfqP->Xhist[i]));
    PetscCall(VecDestroy(&mfqP->Fhist[i]));
  }
  PetscCall(VecDestroy(&mfqP->workxvec));
  PetscCall(VecDestroy(&mfqP->workfvec));
  PetscCall(PetscFree(mfqP->Xhist));
  PetscCall(PetscFree(mfqP->Fhist));

  if (mfqP->size > 1) {
    PetscCall(VecDestroy(&mfqP->localx));
    PetscCall(VecDestroy(&mfqP->localxmin));
    PetscCall(VecDestroy(&mfqP->localf));
    PetscCall(VecDestroy(&mfqP->localfmin));
  }
  PetscCall(PetscFree(tao->data));
  PetscFunctionReturn(0);
}

static PetscErrorCode TaoSetFromOptions_POUNDERS(Tao tao, PetscOptionItems *PetscOptionsObject)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;

  PetscFunctionBegin;
  PetscOptionsHeadBegin(PetscOptionsObject, "POUNDERS method for least-squares optimization");
  PetscCall(PetscOptionsReal("-tao_pounders_delta", "initial delta", "", mfqP->delta, &mfqP->delta0, NULL));
  mfqP->delta = mfqP->delta0;
  PetscCall(PetscOptionsInt("-tao_pounders_npmax", "max number of points in model", "", mfqP->npmax, &mfqP->npmax, NULL));
  PetscCall(PetscOptionsBool("-tao_pounders_gqt", "use gqt algorithm for subproblem", "", mfqP->usegqt, &mfqP->usegqt, NULL));
  PetscOptionsHeadEnd();
  PetscFunctionReturn(0);
}

static PetscErrorCode TaoView_POUNDERS(Tao tao, PetscViewer viewer)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
  PetscBool     isascii;

  PetscFunctionBegin;
  PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
  if (isascii) {
    PetscCall(PetscViewerASCIIPrintf(viewer, "initial delta: %g\n", (double)mfqP->delta0));
    PetscCall(PetscViewerASCIIPrintf(viewer, "final delta: %g\n", (double)mfqP->delta));
    PetscCall(PetscViewerASCIIPrintf(viewer, "model points: %" PetscInt_FMT "\n", mfqP->nmodelpoints));
    if (mfqP->usegqt) {
      PetscCall(PetscViewerASCIIPrintf(viewer, "subproblem solver: gqt\n"));
    } else {
      PetscCall(TaoView(mfqP->subtao, viewer));
    }
  }
  PetscFunctionReturn(0);
}
/*MC
  TAOPOUNDERS - POUNDERS derivate-free model-based algorithm for nonlinear least squares

  Options Database Keys:
+ -tao_pounders_delta - initial step length
. -tao_pounders_npmax - maximum number of points in model
- -tao_pounders_gqt - use gqt algorithm for subproblem instead of TRON

  Level: beginner

M*/

PETSC_EXTERN PetscErrorCode TaoCreate_POUNDERS(Tao tao)
{
  TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;

  PetscFunctionBegin;
  tao->ops->setup          = TaoSetUp_POUNDERS;
  tao->ops->solve          = TaoSolve_POUNDERS;
  tao->ops->view           = TaoView_POUNDERS;
  tao->ops->setfromoptions = TaoSetFromOptions_POUNDERS;
  tao->ops->destroy        = TaoDestroy_POUNDERS;

  PetscCall(PetscNew(&mfqP));
  tao->data = (void *)mfqP;
  /* Override default settings (unless already changed) */
  if (!tao->max_it_changed) tao->max_it = 2000;
  if (!tao->max_funcs_changed) tao->max_funcs = 4000;
  mfqP->npmax      = PETSC_DEFAULT;
  mfqP->delta0     = 0.1;
  mfqP->delta      = 0.1;
  mfqP->deltamax   = 1e3;
  mfqP->deltamin   = 1e-6;
  mfqP->c2         = 10.0;
  mfqP->theta1     = 1e-5;
  mfqP->theta2     = 1e-4;
  mfqP->gamma0     = 0.5;
  mfqP->gamma1     = 2.0;
  mfqP->eta0       = 0.0;
  mfqP->eta1       = 0.1;
  mfqP->usegqt     = PETSC_FALSE;
  mfqP->gqt_rtol   = 0.001;
  mfqP->gqt_maxits = 50;
  mfqP->workxvec   = NULL;
  PetscFunctionReturn(0);
}