step-26.h

Go to the documentation of this file.

 
,
 *                                       const unsigned int component) const
 *     {
 *       (void)component;
 *       Assert(component == 0, ExcIndexRange(component, 0, 1));
 *       return 0;
 *     }
 *   
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-ThecodeHeatEquationcodeimplementation"></a> 
 * <h3>The <code>HeatEquation</code> implementation</h3>
 *   
 
 * 
 * It is time now for the implementation of the main class. Let's
 * start with the constructor which selects a linear element, a time
 * step constant at 1/500 (remember that one period of the source
 * on the right hand side was set to 0.2 above, so we resolve each
 * period with 100 time steps) and chooses the Crank Nicolson method
 * by setting @f$\theta=1/2@f$.
 * 
 * @code
 *     template <int dim>
 *     HeatEquation<dim>::HeatEquation()
 *       : fe(1)
 *       , dof_handler(triangulation)
 *       , time_step(1. / 500)
 *       , theta(0.5)
 *     {}
 *   
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-codeHeatEquationsetup_systemcode"></a> 
 * <h4><code>HeatEquation::setup_system</code></h4>
 *   
 
 * 
 * The next function is the one that sets up the DoFHandler object,
 * computes the constraints, and sets the linear algebra objects
 * to their correct sizes. We also compute the mass and Laplace
 * matrix here by simply calling two functions in the library.
 *   
 
 * 
 * Note that we do not take the hanging node constraints into account when
 * assembling the matrices (both functions have an AffineConstraints argument
 * that defaults to an empty object). This is because we are going to
 * condense the constraints in run() after combining the matrices for the
 * current time-step.
 * 
 * @code
 *     template <int dim>
 *     void HeatEquation<dim>::setup_system()
 *     {
 *       dof_handler.distribute_dofs(fe);
 *   
 *       std::cout << std::endl
 *                 << "===========================================" << std::endl
 *                 << "Number of active cells: " << triangulation.n_active_cells()
 *                 << std::endl
 *                 << "Number of degrees of freedom: " << dof_handler.n_dofs()
 *                 << std::endl
 *                 << std::endl;
 *   
 *       constraints.clear();
 *       DoFTools::make_hanging_node_constraints(dof_handler, constraints);
 *       constraints.close();
 *   
 *       DynamicSparsityPattern dsp(dof_handler.n_dofs());
 *       DoFTools::make_sparsity_pattern(dof_handler,
 *                                       dsp,
 *                                       constraints,
 *                                       /*keep_constrained_dofs = */ true);
 *       sparsity_pattern.copy_from(dsp);
 *   
 *       mass_matrix.reinit(sparsity_pattern);
 *       laplace_matrix.reinit(sparsity_pattern);
 *       system_matrix.reinit(sparsity_pattern);
 *   
 *       MatrixCreator::create_mass_matrix(dof_handler,
 *                                         QGauss<dim>(fe.degree + 1),
 *                                         mass_matrix);
 *       MatrixCreator::create_laplace_matrix(dof_handler,
 *                                            QGauss<dim>(fe.degree + 1),
 *                                            laplace_matrix);
 *   
 *       solution.reinit(dof_handler.n_dofs());
 *       old_solution.reinit(dof_handler.n_dofs());
 *       system_rhs.reinit(dof_handler.n_dofs());
 *     }
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-codeHeatEquationsolve_time_stepcode"></a> 
 * <h4><code>HeatEquation::solve_time_step</code></h4>
 *   
 
 * 
 * The next function is the one that solves the actual linear system
 * for a single time step. There is nothing surprising here:
 * 
 * @code
 *     template <int dim>
 *     void HeatEquation<dim>::solve_time_step()
 *     {
 *       SolverControl            solver_control(1000, 1e-8 * system_rhs.l2_norm());
 *       SolverCG<Vector<double>> cg(solver_control);
 *   
 *       PreconditionSSOR<SparseMatrix<double>> preconditioner;
 *       preconditioner.initialize(system_matrix, 1.0);
 *   
 *       cg.solve(system_matrix, solution, system_rhs, preconditioner);
 *   
 *       constraints.distribute(solution);
 *   
 *       std::cout << "     " << solver_control.last_step() << " CG iterations."
 *                 << std::endl;
 *     }
 *   
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-codeHeatEquationoutput_resultscode"></a> 
 * <h4><code>HeatEquation::output_results</code></h4>
 *   
 
 * 
 * Neither is there anything new in generating graphical output other than the
 * fact that we tell the DataOut object what the current time and time step
 * number is, so that this can be written into the output file :
 * 
 * @code
 *     template <int dim>
 *     void HeatEquation<dim>::output_results() const
 *     {
 *       DataOut<dim> data_out;
 *   
 *       data_out.attach_dof_handler(dof_handler);
 *       data_out.add_data_vector(solution, "U");
 *   
 *       data_out.build_patches();
 *   
 *       data_out.set_flags(DataOutBase::VtkFlags(time, timestep_number));
 *   
 *       const std::string filename =
 *         "solution-" + Utilities::int_to_string(timestep_number, 3) + ".vtk";
 *       std::ofstream output(filename);
 *       data_out.write_vtk(output);
 *     }
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-codeHeatEquationrefine_meshcode"></a> 
 * <h4><code>HeatEquation::refine_mesh</code></h4>
 *   
 
 * 
 * This function is the interesting part of the program. It takes care of
 * the adaptive mesh refinement. The three tasks
 * this function performs is to first find out which cells to
 * refine/coarsen, then to actually do the refinement and eventually
 * transfer the solution vectors between the two different grids. The first
 * task is simply achieved by using the well-established Kelly error
 * estimator on the solution. The second task is to actually do the
 * remeshing. That involves only basic functions as well, such as the
 * <code>refine_and_coarsen_fixed_fraction</code> that refines those cells
 * with the largest estimated error that together make up 60 per cent of the
 * error, and coarsens those cells with the smallest error that make up for
 * a combined 40 per cent of the error. Note that for problems such as the
 * current one where the areas where something is going on are shifting
 * around, we want to aggressively coarsen so that we can move cells
 * around to where it is necessary.
 *   
 
 * 
 * As already discussed in the introduction, too small a mesh leads to
 * too small a time step, whereas too large a mesh leads to too little
 * resolution. Consequently, after the first two steps, we have two
 * loops that limit refinement and coarsening to an allowable range of
 * cells:
 * 
 * @code
 *     template <int dim>
 *     void HeatEquation<dim>::refine_mesh(const unsigned int min_grid_level,
 *                                         const unsigned int max_grid_level)
 *     {
 *       Vector<float> estimated_error_per_cell(triangulation.n_active_cells());
 *   
 *       KellyErrorEstimator<dim>::estimate(
 *         dof_handler,
 *         QGauss<dim - 1>(fe.degree + 1),
 *         std::map<types::boundary_id, const Function<dim> *>(),
 *         solution,
 *         estimated_error_per_cell);
 *   
 *       GridRefinement::refine_and_coarsen_fixed_fraction(triangulation,
 *                                                         estimated_error_per_cell,
 *                                                         0.6,
 *                                                         0.4);
 *   
 *       if (triangulation.n_levels() > max_grid_level)
 *         for (const auto &cell :
 *              triangulation.active_cell_iterators_on_level(max_grid_level))
 *           cell->clear_refine_flag();
 *       for (const auto &cell :
 *            triangulation.active_cell_iterators_on_level(min_grid_level))
 *         cell->clear_coarsen_flag();
 * @endcode
 * 
 * These two loops above are slightly different but this is easily
 * explained. In the first loop, instead of calling
 * <code>triangulation.end()</code> we may as well have called
 * <code>triangulation.end_active(max_grid_level)</code>. The two
 * calls should yield the same iterator since iterators are sorted
 * by level and there should not be any cells on levels higher than
 * on level <code>max_grid_level</code>. In fact, this very piece
 * of code makes sure that this is the case.
 * 
 
 * 
 * As part of mesh refinement we need to transfer the solution vectors
 * from the old mesh to the new one. To this end we use the
 * SolutionTransfer class and we have to prepare the solution vectors that
 * should be transferred to the new grid (we will lose the old grid once
 * we have done the refinement so the transfer has to happen concurrently
 * with refinement). At the point where we call this function, we will
 * have just computed the solution, so we no longer need the old_solution
 * variable (it will be overwritten by the solution just after the mesh
 * may have been refined, i.e., at the end of the time step; see below).
 * In other words, we only need the one solution vector, and we copy it
 * to a temporary object where it is safe from being reset when we further
 * down below call <code>setup_system()</code>.
 *     
 
 * 
 * Consequently, we initialize a SolutionTransfer object by attaching
 * it to the old DoF handler. We then prepare the triangulation and the
 * data vector for refinement (in this order).
 * 
 * @code
 *       SolutionTransfer<dim> solution_trans(dof_handler);
 *   
 *       Vector<double> previous_solution;
 *       previous_solution = solution;
 *       triangulation.prepare_coarsening_and_refinement();
 *       solution_trans.prepare_for_coarsening_and_refinement(previous_solution);
 *   
 * @endcode
 * 
 * Now everything is ready, so do the refinement and recreate the DoF
 * structure on the new grid, and finally initialize the matrix structures
 * and the new vectors in the <code>setup_system</code> function. Next, we
 * actually perform the interpolation of the solution from old to new
 * grid. The final step is to apply the hanging node constraints to the
 * solution vector, i.e., to make sure that the values of degrees of
 * freedom located on hanging nodes are so that the solution is
 * continuous. This is necessary since SolutionTransfer only operates on
 * cells locally, without regard to the neighborhood.
 * 
 * @code
 *       triangulation.execute_coarsening_and_refinement();
 *       setup_system();
 *   
 *       solution_trans.interpolate(previous_solution, solution);
 *       constraints.distribute(solution);
 *     }
 *   
 *   
 *   
 * @endcode
 * 
 * 
 * <a name="step_26-codeHeatEquationruncode"></a> 
 * <h4><code>HeatEquation::run</code></h4>
 *   
 
 * 
 * This is the main driver of the program, where we loop over all
 * time steps. At the top of the function, we set the number of
 * initial global mesh refinements and the number of initial cycles of
 * adaptive mesh refinement by repeating the first time step a few
 * times. Then we create a mesh, initialize the various objects we will
 * work with, set a label for where we should start when re-running
 * the first time step, and interpolate the initial solution onto
 * out mesh (we choose the zero function here, which of course we could
 * do in a simpler way by just setting the solution vector to zero). We
 * also output the initial time step once.
 *   
 
 * 
 * @note If you're an experienced programmer, you may be surprised
 * that we use a <code>goto</code> statement in this piece of code!
 * <code>goto</code> statements are not particularly well liked any
 * more since Edsgar Dijkstra, one of the greats of computer science,
 * wrote a letter in 1968 called "Go To Statement considered harmful"
 * (see <a href="http://en.wikipedia.org/wiki/Considered_harmful">here</a>).
 * The author of this code subscribes to this notion whole-heartedly:
 * <code>goto</code> is hard to understand. In fact, deal.II contains
 * virtually no occurrences: excluding code that was essentially
 * transcribed from books and not counting duplicated code pieces,
 * there are 3 locations in about 600,000 lines of code at the time
 * this note is written; we also use it in 4 tutorial programs, in
 * exactly the same context as here. Instead of trying to justify
 * the occurrence here, let's first look at the code and we'll come
 * back to the issue at the end of function.
 * 
 * @code
 *     template <int dim>
 *     void HeatEquation<dim>::run()
 *     {
 *       const unsigned int initial_global_refinement       = 2;
 *       const unsigned int n_adaptive_pre_refinement_steps = 4;
 *   
 *       GridGenerator::hyper_L(triangulation);
 *       triangulation.refine_global(initial_global_refinement);
 *   
 *       setup_system();
 *   
 *       unsigned int pre_refinement_step = 0;
 *   
 *       Vector<double> tmp;
 *       Vector<double> forcing_terms;
 *   
 *     start_time_iteration:
 *   
 *       time            = 0.0;
 *       timestep_number = 0;
 *   
 *       tmp.reinit(solution.size());
 *       forcing_terms.reinit(solution.size());
 *   
 *   
 *       VectorTools::interpolate(dof_handler,
 *                                Functions::ZeroFunction<dim>(),
 *                                old_solution);
 *       solution = old_solution;
 *   
 *       output_results();
 *   
 * @endcode
 * 
 * Then we start the main loop until the computed time exceeds our
 * end time of 0.5. The first task is to build the right hand
 * side of the linear system we need to solve in each time step.
 * Recall that it contains the term @f$MU^{n-1}-(1-\theta)k_n AU^{n-1}@f$.
 * We put these terms into the variable system_rhs, with the
 * help of a temporary vector:
 * 
 * @code
 *       const double end_time = 0.5;
 *       while (time <= end_time)
 *         {
 *           time += time_step;
 *           ++timestep_number;
 *   
 *           std::cout << "Time step " << timestep_number << " at t=" << time
 *                     << std::endl;
 *   
 *           mass_matrix.vmult(system_rhs, old_solution);
 *   
 *           laplace_matrix.vmult(tmp, old_solution);
 *           system_rhs.add(-(1 - theta) * time_step, tmp);
 *   
 * @endcode
 * 
 * The second piece is to compute the contributions of the source
 * terms. This corresponds to the term @f$k_n
 * \left[ (1-\theta)F^{n-1} + \theta F^n \right]@f$. The following
 * code calls VectorTools::create_right_hand_side to compute the
 * vectors @f$F@f$, where we set the time of the right hand side
 * (source) function before we evaluate it. The result of this
 * all ends up in the forcing_terms variable:
 * 
 * @code
 *           RightHandSide<dim> rhs_function;
 *           rhs_function.set_time(time);
 *           VectorTools::create_right_hand_side(dof_handler,
 *                                               QGauss<dim>(fe.degree + 1),
 *                                               rhs_function,
 *                                               tmp);
 *           forcing_terms = tmp;
 *           forcing_terms *= time_step * theta;
 *   
 *           rhs_function.set_time(time - time_step);
 *           VectorTools::create_right_hand_side(dof_handler,
 *                                               QGauss<dim>(fe.degree + 1),
 *                                               rhs_function,
 *                                               tmp);
 *   
 *           forcing_terms.add(time_step * (1 - theta), tmp);
 *   
 * @endcode
 * 
 * Next, we add the forcing terms to the ones that
 * come from the time stepping, and also build the matrix
 * @f$M+k_n\theta A@f$ that we have to invert in each time step.
 * The final piece of these operations is to eliminate
 * hanging node constrained degrees of freedom from the
 * linear system:
 * 
 * @code
 *           system_rhs += forcing_terms;
 *   
 *           system_matrix.copy_from(mass_matrix);
 *           system_matrix.add(theta * time_step, laplace_matrix);
 *   
 *           constraints.condense(system_matrix, system_rhs);
 *   
 * @endcode
 * 
 * There is one more operation we need to do before we
 * can solve it: boundary values. To this end, we create
 * a boundary value object, set the proper time to the one
 * of the current time step, and evaluate it as we have
 * done many times before. The result is used to also
 * set the correct boundary values in the linear system:
 * 
 * @code
 *           {
 *             BoundaryValues<dim> boundary_values_function;
 *             boundary_values_function.set_time(time);
 *   
 *             std::map<types::global_dof_index, double> boundary_values;
 *             VectorTools::interpolate_boundary_values(dof_handler,
 *                                                      0,
 *                                                      boundary_values_function,
 *                                                      boundary_values);
 *   
 *             MatrixTools::apply_boundary_values(boundary_values,
 *                                                system_matrix,
 *                                                solution,
 *                                                system_rhs);
 *           }
 *   
 * @endcode
 * 
 * With this out of the way, all we have to do is solve the
 * system, generate graphical data, and...
 * 
 * @code
 *           solve_time_step();
 *   
 *           output_results();
 *   
 * @endcode
 * 
 * ...take care of mesh refinement. Here, what we want to do is
 * (i) refine the requested number of times at the very beginning
 * of the solution procedure, after which we jump to the top to
 * restart the time iteration, (ii) refine every fifth time
 * step after that.
 *         
 
 * 
 * The time loop and, indeed, the main part of the program ends
 * with starting into the next time step by setting old_solution
 * to the solution we have just computed.
 * 
 * @code
 *           if ((timestep_number == 1) &&
 *               (pre_refinement_step < n_adaptive_pre_refinement_steps))
 *             {
 *               refine_mesh(initial_global_refinement,
 *                           initial_global_refinement +
 *                             n_adaptive_pre_refinement_steps);
 *               ++pre_refinement_step;
 *   
 *               std::cout << std::endl;
 *   
 *               goto start_time_iteration;
 *             }
 *           else if ((timestep_number > 0) && (timestep_number % 5 == 0))
 *             {
 *               refine_mesh(initial_global_refinement,
 *                           initial_global_refinement +
 *                             n_adaptive_pre_refinement_steps);
 *               tmp.reinit(solution.size());
 *               forcing_terms.reinit(solution.size());
 *             }
 *   
 *           old_solution = solution;
 *         }
 *     }
 *   } // namespace Step26
 * @endcode
 * 
 * Now that you have seen what the function does, let us come back to the issue
 * of the <code>goto</code>. In essence, what the code does is
 * something like this:
 * <div class=CodeFragmentInTutorialComment>
 * @code
 *   void run ()
 *   {
 *     initialize;
 *   start_time_iteration:
 *     for (timestep=1...)
 *     {
 *        solve timestep;
 *        if (timestep==1 && not happy with the result)
 *        {
 *          adjust some data structures;
 *          goto start_time_iteration; // simply try again
 *        }
 *        postprocess;
 *     }
 *   }
 * @endcode
 * </div>
 * Here, the condition "happy with the result" is whether we'd like to keep
 * the current mesh or would rather refine the mesh and start over on the
 * new mesh. We could of course replace the use of the <code>goto</code>
 * by the following:
 * <div class=CodeFragmentInTutorialComment>
 * @code
 *   void run ()
 *   {
 *     initialize;
 *     while (true)
 *     {
 *        solve timestep;
 *        if (not happy with the result)
 *           adjust some data structures;
 *        else
 *           break;
 *     }
 *     postprocess;
 *
 
 *     for (timestep=2...)
 *     {
 *        solve timestep;
 *        postprocess;
 *     }
 *   }
 * @endcode
 * </div>
 * This has the advantage of getting rid of the <code>goto</code>
 * but the disadvantage of having to duplicate the code that implements
 * the "solve timestep" and "postprocess" operations in two different
 * places. This could be countered by putting these parts of the code
 * (sizable chunks in the actual implementation above) into their
 * own functions, but a <code>while(true)</code> loop with a
 * <code>break</code> statement is not really all that much easier
 * to read or understand than a <code>goto</code>.
 * 
 
 * 
 * In the end, one might simply agree that <i>in general</i>
 * <code>goto</code> statements are a bad idea but be pragmatic and
 * state that there may be occasions where they can help avoid code
 * duplication and awkward control flow. This may be one of these
 * places, and it matches the position Steve McConnell takes in his
 * excellent book "Code Complete" @cite CodeComplete about good
 * programming practices (see the mention of this book in the
 * introduction of @ref step_1 "step-1") that spends a surprising ten pages on the
 * question of <code>goto</code> in general.
 * 
 
 * 
 * 
 
 * 
 * 
 * <a name="step_26-Thecodemaincodefunction"></a> 
 * <h3>The <code>main</code> function</h3>
 * 
 
 * 
 * Having made it this far,  there is, again, nothing
 * much to discuss for the main function of this
 * program: it looks like all such functions since @ref step_6 "step-6".
 * 
 * @code
 *   int main()
 *   {
 *     try
 *       {
 *         using namespace Step26;
 *   
 *         HeatEquation<2> heat_equation_solver;
 *         heat_equation_solver.run();
 *       }
 *     catch (std::exception &exc)
 *       {
 *         std::cerr << std::endl
 *                   << std::endl
 *                   << "----------------------------------------------------"
 *                   << std::endl;
 *         std::cerr << "Exception on processing: " << std::endl
 *                   << exc.what() << std::endl
 *                   << "Aborting!" << std::endl
 *                   << "----------------------------------------------------"
 *                   << std::endl;
 *   
 *         return 1;
 *       }
 *     catch (...)
 *       {
 *         std::cerr << std::endl
 *                   << std::endl
 *                   << "----------------------------------------------------"
 *                   << std::endl;
 *         std::cerr << "Unknown exception!" << std::endl
 *                   << "Aborting!" << std::endl
 *                   << "----------------------------------------------------"
 *                   << std::endl;
 *         return 1;
 *       }
 *   
 *     return 0;
 *   }
 * @endcode
<a name="step_26-Results"></a><h1>Results</h1>
 
 
As in many of the tutorials, the actual output of the program matters less
than how we arrived there. Nonetheless, here it is:
@code
===========================================
Number of active cells: 48
Number of degrees of freedom: 65
 
Time step 1 at t=0.002
     7 CG iterations.
 
===========================================
Number of active cells: 60
Number of degrees of freedom: 81
 
 
Time step 1 at t=0.002
     7 CG iterations.
 
===========================================
Number of active cells: 105
Number of degrees of freedom: 136
 
 
Time step 1 at t=0.002
     7 CG iterations.
 
[...]
 
Time step 249 at t=0.498
     13 CG iterations.
Time step 250 at t=0.5
     14 CG iterations.
 
===========================================
Number of active cells: 1803
Number of degrees of freedom: 2109
@endcode
 
Maybe of more interest is a visualization of the solution and the mesh on which
it was computed:
 
<img src="https://www.dealii.org/images/steps/developer/step-26.movie.gif" alt="Animation of the solution of step 26.">
 
The movie shows how the two sources switch on and off and how the mesh reacts
to this. It is quite obvious that the mesh as is is probably not the best we
could come up with. We'll get back to this in the next section.
 
 
<a name="step-26-extensions"></a>
<a name="step_26-Possibilitiesforextensions"></a><h3>Possibilities for extensions</h3>
 
 
There are at least two areas where one can improve this program significantly:
adaptive time stepping and a better choice of the mesh.
 
<a name="step_26-Adaptivetimestepping"></a><h4>Adaptive time stepping</h4>
 
 
Having chosen an implicit time stepping scheme, we are not bound by any
CFL-like condition on the time step. Furthermore, because the time scales on
which change happens on a given cell in the heat equation are not bound to the
cells diameter (unlike the case with the wave equation, where we had a fixed
speed of information transport that couples the temporal and spatial scales),
we can choose the time step as we please. Or, better, choose it as we deem
necessary for accuracy.
 
Looking at the solution, it is clear that the action does not happen uniformly
over time: a lot is changing around the times when we switch on a source, things
become less dramatic once a source is on for a little while, and we enter a
long phase of decline when both sources are off. During these times, we could
surely get away with a larger time step than before without sacrificing too
much accuracy.
 
The literature has many suggestions on how to choose the time step size
adaptively. Much can be learned, for example, from the way ODE solvers choose
their time steps. One can also be inspired by a posteriori error estimators
that can, ideally, be written in a way that they consist of a temporal and a
spatial contribution to the overall error. If the temporal one is too large,
we should choose a smaller time step. Ideas in this direction can be found,
for example, in the PhD thesis of a former principal developer of deal.II,
Ralf Hartmann, published by the University of Heidelberg, Germany, in 2002
(see @cite Har02).
 
 
<a name="step_26-Bettertimesteppingmethods"></a><h4>Better time stepping methods</h4>
 
 
We here use one of the simpler time stepping methods, namely the second order
in time Crank-Nicolson method. This is surely already better than the even
more widely used (and even less accurate) implicit (=backward) Euler method,
but many other, more accurate methods such as BDF or
Runge-Kutta methods are available and should be used as they do not represent
much additional effort. It is not difficult to implement this for the current
program, if one wanted; a more systematic treatment is also given in @ref step_52 "step-52".
 
As a general rule, however, one should not be implementing time stepping
methods by hand, as we do here, for problems that do not require
exploiting special properties of the equation and consequently require
specialized time stepping methods. (The heat equation does not fall into
this category, and "standard" time stepping methods are all we need here.)
Rather, one should use one of the available
high-quality libraries for time stepping, for the same reasons as one should
not be implementing finite element methods by hand but use deal.II instead.
Indeed, deal.II has interfaces to two such time stepping library,
[SUNDIALS](https://computing.llnl.gov/projects/sundials) and the
[PETSc TS package](https://petsc.org/release/manualpages/TS/TS/), already available.
In particular, the SUNDIALS::ARKode class would make for a great starting
point for the use of much better (and much more accurate) time steppers,
as would PETScWrappers::TimeStepper;
the methods one would then get also implement the automatic time step
control mentioned above. To see how this works, take a look at
@ref step_86 "step-86".
 
 
<a name="step_26-Betterrefinementcriteria"></a><h4>Better refinement criteria</h4>
 
 
If you look at the meshes in the movie above, it is clear that they are not
particularly well suited to the task at hand. In fact, they look rather
random.
 
There are two factors at play. First, there are some islands where cells
have been refined but that are surrounded by non-refined cells (and there
are probably also a few occasional coarsened islands). These are not terrible,
as they most of the time do not affect the approximation quality of the mesh,
but they also don't help because so many of their additional degrees of
freedom are in fact constrained by hanging node constraints. That said,
this is easy to fix: the Triangulation class takes an argument to its
constructor indicating a level of "mesh smoothing". Passing one of many
possible flags, this instructs the triangulation to refine some additional
cells, or not to refine some cells, so that the resulting mesh does not have
these artifacts.
 
The second problem is more severe: the mesh appears to lag the solution.
The underlying reason is that we only adapt the mesh once every fifth
time step, and only allow for a single refinement in these cases. Whenever a
source switches on, the solution had been very smooth in this area before and
the mesh was consequently rather coarse. This implies that the next time step
when we refine the mesh, we will get one refinement level more in this area,
and five time steps later another level, etc. But this is not enough: first,
we should refine immediately when a source switches on (after all, in the
current context we at least know what the right hand side is), and we should
allow for more than one refinement level. Of course, all of this can be done
using deal.II, it just requires a bit of algorithmic thinking in how to make
this work!
 
 
<a name="step_26-Positivitypreservation"></a><h4>Positivity preservation</h4>
 
 
To increase the accuracy and resolution of your simulation in time, one
typically decreases the time step size @f$k_n@f$. If you start playing around
with the time step in this particular example, you will notice that the
solution becomes partly negative, if @f$k_n@f$ is below a certain threshold.
This is not what we would expect to happen (in nature).
 
To get an idea of this behavior mathematically, let us consider a general,
fully discrete problem:
@f{align*}{
  A u^{n} = B u^{n-1}.
@f}
The general form of the @f$i@f$th equation then reads:
@f{align*}{
  a_{ii} u^{n}_i &= b_{ii} u^{n-1}_i +
  \sum\limits_{j \in S_i} \left( b_{ij} u^{n-1}_j - a_{ij} u^{n}_j \right),
@f}
where @f$S_i@f$ is the set of degrees of freedom that DoF @f$i@f$ couples with (i.e.,
for which either the matrix @f$A@f$ or matrix @f$B@f$ has a nonzero entry at position
@f$(i,j)@f$). If all coefficients
fulfill the following conditions:
@f{align*}{
  a_{ii} &> 0, & b_{ii} &\geq 0, & a_{ij} &\leq 0, & b_{ij} &\geq 0,
  &
  \forall j &\in S_i,
@f}
all solutions @f$u^{n}@f$ keep their sign from the previous ones @f$u^{n-1}@f$, and
consequently from the initial values @f$u^0@f$. See e.g.
<a href="http://bookstore.siam.org/cs14/">Kuzmin, H&auml;m&auml;l&auml;inen</a>
for more information on positivity preservation.
 
Depending on the PDE to solve and the time integration scheme used, one is
able to deduce conditions for the time step @f$k_n@f$. For the heat equation with
the Crank-Nicolson scheme,
<a href="https://doi.org/10.2478/cmam-2010-0025">Schatz et. al.</a> have
translated it to the following ones:
@f{align*}{
  (1 - \theta) k a_{ii} &\leq m_{ii},\qquad \forall i,
  &
  \theta k \left| a_{ij} \right| &\geq m_{ij},\qquad j \neq i,
@f}
where @f$M = m_{ij}@f$ denotes the @ref GlossMassMatrix "mass matrix" and @f$A = a_{ij}@f$ the stiffness
matrix with @f$a_{ij} \leq 0@f$ for @f$j \neq i@f$, respectively. With
@f$a_{ij} \leq 0@f$, we can formulate bounds for the global time step @f$k@f$ as
follows:
@f{align*}{
  k_{\text{max}} &= \frac{ 1 }{ 1 - \theta }
  \min\left( \frac{ m_{ii} }{ a_{ii} } \right),~ \forall i,
  &
  k_{\text{min}} &= \frac{ 1 }{ \theta  }
  \max\left( \frac{ m_{ij} }{ \left|a_{ij}\right| } \right),~ j \neq i.
@f}
In other words, the time step is constrained by <i>both a lower
and upper bound</i> in case of a Crank-Nicolson scheme. These bounds should be
considered along with the CFL condition to ensure significance of the performed
simulations.
 
Being unable to make the time step as small as we want to get more
accuracy without losing the positivity property is annoying. It raises
the question of whether we can at least <i>compute</i> the minimal time step
we can choose  to ensure positivity preservation in this particular tutorial.
Indeed, we can use
the SparseMatrix objects for both mass and stiffness that are created via
the MatrixCreator functions. Iterating through each entry via SparseMatrixIterators
lets us check for diagonal and off-diagonal entries to set a proper time step
dynamically. For quadratic matrices, the diagonal element is stored as the
first member of a row (see SparseMatrix documentation). An exemplary code
snippet on how to grab the entries of interest from the <code>mass_matrix</code>
is shown below.
 
@code
Assert (mass_matrix.m() == mass_matrix.n(), ExcNotQuadratic());
const unsigned int num_rows = mass_matrix.m();
double mass_matrix_min_diag    = std::numeric_limits<double>::max(),
       mass_matrix_max_offdiag = 0.;
 
SparseMatrixIterators::Iterator<double,true> row_it (&mass_matrix, 0);
 
for(unsigned int m = 0; m<num_rows; ++m)
{
  // check the diagonal element
  row_it = mass_matrix.begin(m);
  mass_matrix_min_diag = std::min(row_it->value(), mass_matrix_min_diag);
  ++row_it;
 
  // check the off-diagonal elements
  for(; row_it != mass_matrix.end(m); ++row_it)
    mass_matrix_max_offdiag = std::max(row_it->value(), mass_matrix_max_offdiag);
}
@endcode
 
Using the information so computed, we can bound the time step via the formulas
above.
 *
 *
<a name="step_26-PlainProg"></a>
<h1> The plain program</h1>
@include "step-26.cc"
*/