IP_SSAHardavForwardProblem.cc

Go to the documentation of this file.

// Copyright (C) 2013, 2014, 2015, 2016, 2017, 2018, 2020, 2021, 2022, 2023, 2024, 2025  David Maxwell and Constantine Khroulev
//
// This file is part of PISM.
//
// PISM is free software; you can redistribute it and/or modify it under the
// terms of the GNU General Public License as published by the Free Software
// Foundation; either version 3 of the License, or (at your option) any later
// version.
//
// PISM is distributed in the hope that it will be useful, but WITHOUT ANY
// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
// FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
// details.
//
// You should have received a copy of the GNU General Public License
// along with PISM; if not, write to the Free Software
// Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
 
#include "pism/inverse/IP_SSAHardavForwardProblem.hh"
#include "pism/util/Grid.hh"
#include "pism/util/Config.hh"
#include "pism/util/Vars.hh"
#include "pism/util/error_handling.hh"
#include "pism/rheology/FlowLaw.hh"
#include "pism/geometry/Geometry.hh"
#include "pism/stressbalance/StressBalance.hh"
#include "pism/util/petscwrappers/DM.hh"
#include "pism/util/petscwrappers/Vec.hh"
#include "pism/util/fem/Quadrature.hh"
#include "pism/util/fem/DirichletData.hh"
#include "pism/util/Logger.hh"
 
namespace pism {
namespace inverse {
 
IP_SSAHardavForwardProblem::IP_SSAHardavForwardProblem(std::shared_ptr<const Grid> g,
                                                       IPDesignVariableParameterization &tp)
  : SSAFEM(g),
    m_stencil_width(1),
    m_zeta(NULL),
    m_dzeta_local(m_grid, "d_zeta_local"),
    m_fixed_design_locations(NULL),
    m_design_param(tp),
    m_du_global(m_grid, "linearization work vector (sans ghosts)"),
    m_du_local(m_grid, "linearization work vector (with ghosts)"),
    m_hardav(m_grid, "hardav"),
    m_element_index(*m_grid),
    m_element(*m_grid, fem::Q1Quadrature4()),
    m_rebuild_J_state(true)
{
 
  PetscErrorCode ierr;
 
  m_velocity_shared.reset(new array::Vector(m_grid, "dummy"));
  m_velocity_shared->metadata(0) = m_velocity.metadata(0);
  m_velocity_shared->metadata(1) = m_velocity.metadata(1);
 
  auto dm = m_velocity_global.dm();
 
  ierr = DMSetMatType(*dm, MATBAIJ);
  PISM_CHK(ierr, "DMSetMatType");
 
  ierr = DMCreateMatrix(*dm, m_J_state.rawptr());
  PISM_CHK(ierr, "DMCreateMatrix");
 
  ierr = KSPCreate(m_grid->com, m_ksp.rawptr());
  PISM_CHK(ierr, "KSPCreate");
 
  double ksp_rtol = 1e-12;
  ierr = KSPSetTolerances(m_ksp, ksp_rtol, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT);
  PISM_CHK(ierr, "KSPSetTolerances");
 
  PC pc;
  ierr = KSPGetPC(m_ksp, &pc);
  PISM_CHK(ierr, "KSPGetPC");
 
  ierr = PCSetType(pc, PCBJACOBI);
  PISM_CHK(ierr, "PCSetType");
 
  ierr = KSPSetFromOptions(m_ksp);
  PISM_CHK(ierr, "KSPSetFromOptions");
}
 
void IP_SSAHardavForwardProblem::init() {
 
  SSAFEM::init();
 
  // Get most of the inputs from Grid::variables() and fake the rest.
  //
  // I will need to fix this at some point.
  {
    Geometry geometry(m_grid);
    geometry.ice_thickness.copy_from(*m_grid->variables().get_2d_scalar("land_ice_thickness"));
    geometry.bed_elevation.copy_from(*m_grid->variables().get_2d_scalar("bedrock_altitude"));
    geometry.sea_level_elevation.set(0.0); // FIXME: this should be an input
 
    if (m_config->get_flag("geometry.part_grid.enabled")) {
      geometry.ice_area_specific_volume.copy_from(
          *m_grid->variables().get_2d_scalar("ice_area_specific_volume"));
    } else {
      geometry.ice_area_specific_volume.set(0.0);
    }
 
    geometry.ensure_consistency(m_config->get_number("stress_balance.ice_free_thickness_standard"));
 
    stressbalance::Inputs inputs;
 
    const auto &variables = m_grid->variables();
 
    const array::Scalar *vel_bc_mask = nullptr;
    if (variables.is_available("vel_bc_mask")) {
      vel_bc_mask = variables.get_2d_scalar("vel_bc_mask");
    }
 
    const array::Vector *vel_bc = nullptr;
    if (variables.is_available("vel_bc")) {
      vel_bc = variables.get_2d_vector("vel_bc");
    }
 
    inputs.geometry           = &geometry;
    inputs.basal_melt_rate    = NULL;
    inputs.basal_yield_stress = variables.get_2d_scalar("tauc");
    inputs.enthalpy           = variables.get_3d_scalar("enthalpy");
    inputs.age                = NULL;
    inputs.bc_mask            = vel_bc_mask;
    inputs.bc_values          = vel_bc;
 
    inputs.water_column_pressure = NULL;
 
    cache_inputs(inputs);
  }
}
 
//! Sets the current value of of the design paramter \f$\zeta\f$.
/*! This method sets \f$\zeta\f$ but does not solve the %SSA.
It it intended for inverse methods that simultaneously compute
the pair \f$u\f$ and \f$\zeta\f$ without ever solving the %SSA
directly.  Use this method in conjuction with
\ref assemble_jacobian_state and \ref apply_jacobian_design and their friends.
The vector \f$\zeta\f$ is not copied; a reference to the array::Array is
kept.
*/
void IP_SSAHardavForwardProblem::set_design(array::Scalar &new_zeta) {
 
  m_zeta = &new_zeta;
 
  // Convert zeta to hardav.
  m_design_param.convertToDesignVariable(*m_zeta, m_hardav);
 
  // Cache hardav at the quadrature points.
  array::AccessScope list{&m_coefficients, &m_hardav};
 
  for (auto p : m_grid->points_with_ghosts(1)) {
    const int i = p.i(), j = p.j();
    m_coefficients(i, j).hardness = m_hardav(i, j);
  }
 
  // Flag the state jacobian as needing rebuilding.
  m_rebuild_J_state = true;
}
 
//! Sets the current value of the design variable \f$\zeta\f$ and solves the %SSA to find the associated \f$u_{\rm SSA}\f$.
/* Use this method for inverse methods employing the reduced gradient. Use this method
in conjuction with apply_linearization and apply_linearization_transpose.*/
std::shared_ptr<TerminationReason> IP_SSAHardavForwardProblem::linearize_at(array::Scalar &zeta) {
  this->set_design(zeta);
  return this->solve_nocache();
}
 
//! Computes the residual function \f$\mathcal{R}(u, \zeta)\f$ as defined in the class-level documentation.
/* The value of \f$\zeta\f$ is set prior to this call via set_design or linearize_at. The value
of the residual is returned in \a RHS.*/
void IP_SSAHardavForwardProblem::assemble_residual(array::Vector &u, array::Vector &RHS) {
 
  array::AccessScope l{&u, &RHS};
 
  this->compute_local_function(u.array(), RHS.array());
}
 
//! Computes the residual function \f$\mathcal{R}(u, \zeta)\f$ defined in the class-level documentation.
/* The return value is specified via a Vec for the benefit of certain TAO routines.  Otherwise,
the method is identical to the assemble_residual returning values as a StateVec (an array::Vector).*/
void IP_SSAHardavForwardProblem::assemble_residual(array::Vector &u, Vec RHS) {
 
  array::AccessScope l{&u};
 
  petsc::DMDAVecArray rhs_a(m_velocity_global.dm(), RHS);
 
  this->compute_local_function(u.array(), (Vector2d**)rhs_a.get());
}
 
//! Assembles the state Jacobian matrix.
/* The matrix depends on the current value of the design variable \f$\zeta\f$ and the current
value of the state variable \f$u\f$.  The specification of \f$\zeta\f$ is done earlier
with set_design or linearize_at.  The value of \f$u\f$ is specified explicitly as an argument
to this method.
  @param[in] u Current state variable value.
  @param[out] J computed state Jacobian.
*/
void IP_SSAHardavForwardProblem::assemble_jacobian_state(array::Vector &u, Mat Jac) {
  array::AccessScope l{&u};
 
  this->compute_local_jacobian(u.array(), Jac);
}
 
//! Applies the design Jacobian matrix to a perturbation of the design variable.
/*! The return value uses a DesignVector (array::Vector), which can be ghostless. Ghosts (if present) are updated.
\overload
*/
void IP_SSAHardavForwardProblem::apply_jacobian_design(array::Vector &u,
                                                       array::Scalar &dzeta,
                                                       array::Vector &du) {
  array::AccessScope l{&du};
 
  this->apply_jacobian_design(u, dzeta, du.array());
}
 
//! Applies the design Jacobian matrix to a perturbation of the design variable.
/*! The return value is a Vec for the benefit of TAO. It is assumed to be ghostless; no communication is done.
\overload
*/
void IP_SSAHardavForwardProblem::apply_jacobian_design(array::Vector &u,
                                                       array::Scalar &dzeta,
                                                       Vec du) {
  petsc::DMDAVecArray du_a(m_velocity_global.dm(), du);
  this->apply_jacobian_design(u, dzeta, (Vector2d**)du_a.get());
}
 
//! @brief Applies the design Jacobian matrix to a perturbation of the
//! design variable.
 
/*! The matrix depends on the current value of the design variable
    \f$\zeta\f$ and the current value of the state variable \f$u\f$.
    The specification of \f$\zeta\f$ is done earlier with set_design
    or linearize_at. The value of \f$u\f$ is specified explicitly as
    an argument to this method.
 
  @param[in] u Current state variable value.
 
  @param[in] dzeta Perturbation of the design variable. Prefers
                   vectors with ghosts; will copy to a ghosted vector
                   if needed.
 
  @param[out] du_a Computed corresponding perturbation of the state
                   variable. The array \a du_a should be extracted
                   first from a Vec or an array::Array.
 
  Typically this method is called via one of its overloads.
*/
void IP_SSAHardavForwardProblem::apply_jacobian_design(array::Vector &u,
                                                       array::Scalar &dzeta,
                                                       Vector2d **du_a) {
 
  const unsigned int Nk     = fem::q1::n_chi;
  const unsigned int Nq     = m_element.n_pts();
  const unsigned int Nq_max = fem::MAX_QUADRATURE_SIZE;
 
  array::AccessScope list{&m_coefficients, m_zeta, &u};
 
  array::Scalar *dzeta_local;
  if (dzeta.stencil_width() > 0) {
    dzeta_local = &dzeta;
  } else {
    m_dzeta_local.copy_from(dzeta);
    dzeta_local = &m_dzeta_local;
  }
  list.add(*dzeta_local);
 
  // Zero out the portion of the function we are responsible for computing.
  for (auto p : m_grid->points()) {
    const int i = p.i(), j = p.j();
 
    du_a[j][i].u = 0.0;
    du_a[j][i].v = 0.0;
  }
 
  // Aliases to help with notation consistency below.
  const array::Scalar *dirichletLocations = &m_bc_mask;
  const array::Vector   *dirichletValues    = &m_bc_values;
  double                 dirichletWeight    = m_dirichletScale;
 
  Vector2d u_e[Nk];
  Vector2d U[Nq_max], U_x[Nq_max], U_y[Nq_max];
 
  Vector2d du_e[Nk];
 
  double dzeta_e[Nk];
 
  double zeta_e[Nk];
 
  double dB_e[Nk];
  double dB_q[Nq_max];
 
  fem::DirichletData_Vector dirichletBC(dirichletLocations, dirichletValues,
                                        dirichletWeight);
  fem::DirichletData_Scalar fixedZeta(m_fixed_design_locations, NULL);
 
  // Loop through all elements.
  const int
    xs = m_element_index.xs,
    xm = m_element_index.xm,
    ys = m_element_index.ys,
    ym = m_element_index.ym;
 
  ParallelSection loop(m_grid->com);
  try {
    for (int j =ys; j<ys+ym; j++) {
      for (int i =xs; i<xs+xm; i++) {
 
        // Zero out the element-local residual in prep for updating it.
        for (unsigned int k=0; k<Nk; k++) {
          du_e[k].u = 0;
          du_e[k].v = 0;
        }
 
        // Initialize the map from global to local degrees of freedom for this element.
        m_element.reset(i, j);
 
        // Obtain the value of the solution at the nodes adjacent to the element,
        // fix dirichlet values, and compute values at quad pts.
        m_element.nodal_values(u.array(), u_e);
        if (dirichletBC) {
          dirichletBC.constrain(m_element);
          dirichletBC.enforce(m_element, u_e);
        }
        m_element.evaluate(u_e, U, U_x, U_y);
 
        // Compute dzeta at the nodes
        m_element.nodal_values(dzeta_local->array(), dzeta_e);
        if (fixedZeta) {
          fixedZeta.enforce_homogeneous(m_element, dzeta_e);
        }
 
        // Compute the change in hardav with respect to zeta at the quad points.
        m_element.nodal_values(m_zeta->array(), zeta_e);
        for (unsigned int k=0; k<Nk; k++) {
          m_design_param.toDesignVariable(zeta_e[k], NULL, dB_e + k);
          dB_e[k]*=dzeta_e[k];
        }
        m_element.evaluate(dB_e, dB_q);
 
        double thickness[Nq_max];
        {
          Coefficients coeffs[Nk];
          int    mask[Nq_max];
          double tauc[Nq_max];
          double hardness[Nq_max];
 
          m_element.nodal_values(m_coefficients.array(), coeffs);
 
          quad_point_values(m_element, coeffs,
                            mask, thickness, tauc, hardness);
        }
 
        for (unsigned int q = 0; q < Nq; q++) {
          // Symmetric gradient at the quadrature point.
          double Duqq[3] = {U_x[q].u, U_y[q].v, 0.5 * (U_y[q].u + U_x[q].v)};
 
          double d_nuH = 0;
          if (thickness[q] >= strength_extension->get_min_thickness()) {
            m_flow_law->effective_viscosity(dB_q[q],
                                            secondInvariant_2D(U_x[q], U_y[q]),
                                            &d_nuH, NULL);
            d_nuH *= (2.0 * thickness[q]);
          }
 
          auto W = m_element.weight(q);
 
          for (unsigned int k = 0; k < Nk; k++) {
            const fem::Germ &testqk = m_element.chi(q, k);
            du_e[k].u += W*d_nuH*(testqk.dx*(2*Duqq[0] + Duqq[1]) + testqk.dy*Duqq[2]);
            du_e[k].v += W*d_nuH*(testqk.dy*(2*Duqq[1] + Duqq[0]) + testqk.dx*Duqq[2]);
          }
        } // q
        m_element.add_contribution(du_e, du_a);
      } // j
    } // i
  } catch (...) {
    loop.failed();
  }
  loop.check();
 
  if (dirichletBC) {
    dirichletBC.fix_residual_homogeneous(du_a);
  }
}
 
//! Applies the transpose of the design Jacobian matrix to a perturbation of the state variable.
/*! The return value uses a StateVector (array::Scalar) which can be ghostless; ghosts (if present) are updated.
\overload
*/
void IP_SSAHardavForwardProblem::apply_jacobian_design_transpose(array::Vector &u,
                                                                 array::Vector &du,
                                                                 array::Scalar &dzeta) {
  array::AccessScope l{&dzeta};
  this->apply_jacobian_design_transpose(u, du, dzeta.array());
}
 
//! Applies the transpose of the design Jacobian matrix to a perturbation of the state variable.
/*! The return value uses a Vec for the benefit of TAO.  It is assumed to be ghostless; no communication is done.
\overload */
void IP_SSAHardavForwardProblem::apply_jacobian_design_transpose(array::Vector &u,
                                                                 array::Vector &du,
                                                                 Vec dzeta) {
 
  auto da2 = m_grid->get_dm(1, m_config->get_number("grid.max_stencil_width"));
  petsc::DMDAVecArray dzeta_a(da2, dzeta);
  this->apply_jacobian_design_transpose(u, du, (double**)dzeta_a.get());
}
 
//! @brief Applies the transpose of the design Jacobian matrix to a
//! perturbation of the state variable.
 
/*! The matrix depends on the current value of the design variable
    \f$\zeta\f$ and the current value of the state variable \f$u\f$.
    The specification of \f$\zeta\f$ is done earlier with set_design
    or linearize_at. The value of \f$u\f$ is specified explicitly as
    an argument to this method.
 
  @param[in] u Current state variable value.
 
  @param[in] du Perturbation of the state variable. Prefers vectors
                with ghosts; will copy to a ghosted vector if need be.
 
  @param[out] dzeta_a Computed corresponding perturbation of the
                      design variable. The array \a dzeta_a should be
                      extracted first from a Vec or an array::Array.
 
  Typically this method is called via one of its overloads.
*/
void IP_SSAHardavForwardProblem::apply_jacobian_design_transpose(array::Vector &u,
                                                                 array::Vector &du,
                                                                 double **dzeta_a) {
 
  const unsigned int Nk     = fem::q1::n_chi;
  const unsigned int Nq     = m_element.n_pts();
  const unsigned int Nq_max = fem::MAX_QUADRATURE_SIZE;
 
  array::AccessScope list{&m_coefficients, m_zeta, &u};
 
  array::Vector *du_local;
  if (du.stencil_width() > 0) {
    du_local = &du;
  } else {
    m_du_local.copy_from(du);
    du_local = &m_du_local;
  }
  list.add(*du_local);
 
  Vector2d u_e[Nk];
  Vector2d U[Nq_max], U_x[Nq_max], U_y[Nq_max];
 
  Vector2d du_e[Nk];
  Vector2d du_q[Nq_max];
  Vector2d du_dx_q[Nq_max];
  Vector2d du_dy_q[Nq_max];
 
  double dzeta_e[Nk];
 
  // Aliases to help with notation consistency.
  const array::Scalar *dirichletLocations = &m_bc_mask;
  const array::Vector   *dirichletValues    = &m_bc_values;
  double                 dirichletWeight    = m_dirichletScale;
 
  fem::DirichletData_Vector dirichletBC(dirichletLocations, dirichletValues,
                                        dirichletWeight);
 
  // Zero out the portion of the function we are responsible for computing.
  for (auto p : m_grid->points()) {
    const int i = p.i(), j = p.j();
 
    dzeta_a[j][i] = 0;
  }
 
  const int
    xs = m_element_index.xs,
    xm = m_element_index.xm,
    ys = m_element_index.ys,
    ym = m_element_index.ym;
 
  ParallelSection loop(m_grid->com);
  try {
    for (int j = ys; j < ys + ym; j++) {
      for (int i = xs; i < xs + xm; i++) {
        // Initialize the map from global to local degrees of freedom for this element.
        m_element.reset(i, j);
 
        // Obtain the value of the solution at the nodes adjacent to the element.
        // Compute the solution values and symmetric gradient at the quadrature points.
        m_element.nodal_values(du.array(), du_e);
        if (dirichletBC) {
          dirichletBC.enforce_homogeneous(m_element, du_e);
        }
        m_element.evaluate(du_e, du_q, du_dx_q, du_dy_q);
 
        m_element.nodal_values(u.array(), u_e);
        if (dirichletBC) {
          dirichletBC.enforce(m_element, u_e);
        }
        m_element.evaluate(u_e, U, U_x, U_y);
 
        // Zero out the element-local residual in prep for updating it.
        for (unsigned int k = 0; k < Nk; k++) {
          dzeta_e[k] = 0;
        }
 
        double thickness[Nq_max];
        {
          Coefficients coeffs[Nk];
          int    mask[Nq_max];
          double tauc[Nq_max];
          double hardness[Nq_max];
 
          m_element.nodal_values(m_coefficients.array(), coeffs);
 
          quad_point_values(m_element, coeffs,
                            mask, thickness, tauc, hardness);
        }
 
        for (unsigned int q = 0; q < Nq; q++) {
          // Symmetric gradient at the quadrature point.
          double Duqq[3] = {U_x[q].u, U_y[q].v, 0.5 * (U_y[q].u + U_x[q].v)};
 
          // Determine "d_nuH / dB" at the quadrature point
          double d_nuH_dB = 0;
          if (thickness[q] >= strength_extension->get_min_thickness()) {
            m_flow_law->effective_viscosity(1.0,
                                            secondInvariant_2D(U_x[q], U_y[q]),
                                            &d_nuH_dB, NULL);
            d_nuH_dB *= (2.0 * thickness[q]);
          }
 
          auto W = m_element.weight(q);
 
          for (unsigned int k = 0; k < Nk; k++) {
            dzeta_e[k] += W*d_nuH_dB*m_element.chi(q, k).val*((du_dx_q[q].u*(2*Duqq[0] + Duqq[1]) +
                                                               du_dy_q[q].u*Duqq[2]) +
                                                              (du_dy_q[q].v*(2*Duqq[1] + Duqq[0]) +
                                                               du_dx_q[q].v*Duqq[2]));
          }
        } // q
 
        m_element.add_contribution(dzeta_e, dzeta_a);
      } // j
    } // i
  } catch (...) {
    loop.failed();
  }
  loop.check();
 
  for (auto p : m_grid->points()) {
    const int i = p.i(), j = p.j();
 
    double dB_dzeta;
    m_design_param.toDesignVariable((*m_zeta)(i, j), NULL, &dB_dzeta);
    dzeta_a[j][i] *= dB_dzeta;
  }
 
  if (m_fixed_design_locations != nullptr) {
    fem::DirichletData_Scalar fixedZeta(m_fixed_design_locations, NULL);
    fixedZeta.fix_residual_homogeneous(dzeta_a);
  }
}
 
/*!\brief Applies the linearization of the forward map (i.e. the reduced gradient \f$DF\f$ described in
the class-level documentation.) */
/*! As described previously,
\f[
Df = J_{\rm State}^{-1} J_{\rm Design}.
\f]
Applying the linearization then involves the solution of a linear equation.
The matrices \f$J_{\rm State}\f$ and \f$J_{\rm Design}\f$ both depend on the value of the
design variable \f$\zeta\f$ and the value of the corresponding state variable \f$u=F(\zeta)\f$.
These are established by first calling linearize_at.
  @param[in]   dzeta     Perturbation of the design variable
  @param[out]  du        Computed corresponding perturbation of the state variable; ghosts (if present) are updated.
*/
void IP_SSAHardavForwardProblem::apply_linearization(array::Scalar &dzeta, array::Vector &du) {
 
  PetscErrorCode ierr;
 
  if (m_rebuild_J_state) {
    this->assemble_jacobian_state(m_velocity, m_J_state);
    m_rebuild_J_state = false;
  }
 
  this->apply_jacobian_design(m_velocity, dzeta, m_du_global);
  m_du_global.scale(-1);
 
  // call PETSc to solve linear system by iterative method.
  ierr = KSPSetOperators(m_ksp, m_J_state, m_J_state);
  PISM_CHK(ierr, "KSPSetOperators");
 
  ierr = KSPSolve(m_ksp, m_du_global.vec(), m_du_global.vec());
  PISM_CHK(ierr, "KSPSolve"); // SOLVE
 
  KSPConvergedReason reason;
  ierr = KSPGetConvergedReason(m_ksp, &reason);
  PISM_CHK(ierr, "KSPGetConvergedReason");
 
  if (reason < 0) {
    throw RuntimeError::formatted(PISM_ERROR_LOCATION, "IP_SSAHardavForwardProblem::apply_linearization solve"
                                  " failed to converge (KSP reason %s)",
                                  KSPConvergedReasons[reason]);
  }
 
  m_log->message(4,
                 "IP_SSAHardavForwardProblem::apply_linearization converged"
                 " (KSP reason %s)\n",
                 KSPConvergedReasons[reason]);
 
  du.copy_from(m_du_global);
}
 
/*! \brief Applies the transpose of the linearization of the forward map
 (i.e. the transpose of the reduced gradient \f$DF\f$ described in the class-level documentation.) */
/*!  As described previously,
\f[
Df = J_{\rm State}^{-1} J_{\rm Design}.
\f]
so
\f[
Df^t = J_{\rm Design}^t \; (J_{\rm State}^t)^{-1} .
\f]
Applying the transpose of the linearization then involves the solution of a linear equation.
The matrices \f$J_{\rm State}\f$ and \f$J_{\rm Design}\f$ both depend on the value of the
design variable \f$\zeta\f$ and the value of the corresponding state variable \f$u=F(\zeta)\f$.
These are established by first calling linearize_at.
  @param[in]   du     Perturbation of the state variable
  @param[out]  dzeta  Computed corresponding perturbation of the design variable; ghosts (if present) are updated.
*/
void IP_SSAHardavForwardProblem::apply_linearization_transpose(array::Vector &du,
                                                               array::Scalar &dzeta) {
 
  PetscErrorCode ierr;
 
  if (m_rebuild_J_state) {
    this->assemble_jacobian_state(m_velocity, m_J_state);
    m_rebuild_J_state = false;
  }
 
  // Aliases to help with notation consistency below.
  const array::Scalar *dirichletLocations = &m_bc_mask;
  const array::Vector   *dirichletValues    = &m_bc_values;
  double                 dirichletWeight    = m_dirichletScale;
 
  m_du_global.copy_from(du);
 
  fem::DirichletData_Vector dirichletBC(dirichletLocations, dirichletValues,
                                        dirichletWeight);
  if (dirichletBC) {
    array::AccessScope list{&m_du_global};
    dirichletBC.fix_residual_homogeneous(m_du_global.array());
  }
 
  // call PETSc to solve linear system by iterative method.
  ierr = KSPSetOperators(m_ksp, m_J_state, m_J_state);
  PISM_CHK(ierr, "KSPSetOperators");
 
  ierr = KSPSolve(m_ksp, m_du_global.vec(), m_du_global.vec());
  PISM_CHK(ierr, "KSPSolve"); // SOLVE
 
  KSPConvergedReason  reason;
  ierr = KSPGetConvergedReason(m_ksp, &reason);
  PISM_CHK(ierr, "KSPGetConvergedReason");
 
  if (reason < 0) {
    throw RuntimeError::formatted(PISM_ERROR_LOCATION, "IP_SSAHardavForwardProblem::apply_linearization solve failed to converge (KSP reason %s)",
                                  KSPConvergedReasons[reason]);
  }
 
  m_log->message(4, "IP_SSAHardavForwardProblem::apply_linearization converged (KSP reason %s)\n",
                 KSPConvergedReasons[reason]);
 
  this->apply_jacobian_design_transpose(m_velocity, m_du_global, dzeta);
  dzeta.scale(-1);
 
  if (dzeta.stencil_width() > 0) {
    dzeta.update_ghosts();
  }
}
 
} // end of namespace inverse
} // end of namespace pism