Checkpointing an Extended ODE Solver: Example and Test

checkpoint_extended_ode.cpp

@(@\newcommand{\W}[1]{ \; #1 \; } \newcommand{\R}[1]{ {\rm #1} } \newcommand{\B}[1]{ {\bf #1} } \newcommand{\D}[2]{ \frac{\partial #1}{\partial #2} } \newcommand{\DD}[3]{ \frac{\partial^2 #1}{\partial #2 \partial #3} } \newcommand{\Dpow}[2]{ \frac{\partial^{#1}}{\partial {#2}^{#1}} } \newcommand{\dpow}[2]{ \frac{ {\rm d}^{#1}}{{\rm d}\, {#2}^{#1}} }@)@Checkpointing an Extended ODE Solver: Example and Test
See Also
checkpoint_ode.cpp , atomic_mul_level.cpp .

Discussion
Suppose that we wish to extend an ODE to include derivatives with respect to some parameter in the ODE. In addition, suppose we wish to differentiate a function that depends on these derivatives. Applying checkpointing to at the second level of AD would not work; see atomic_mul_level.cpp In this example we show how one can do this by checkpointing an extended ODE solver.

Problem
We consider the initial value problem with parameter @(@ x @)@ defined by, @(@ z(0, x) = z_0 (x) @)@, @[@ \partial_t z(t, x ) = h [ x , z(t, x) ] @]@ Note that if @(@ t @)@ needs to be in the equation, one can define the first component of @(@ z(t, x) @)@ to be equal to @(@ t @)@.

ODE Solver
For this example, we consider the Fourth order Runge-Kutta ODE solver. Given an approximation solution at time @(@ t_k @)@ denoted by @(@ \tilde{z}_k (x) @)@, and @(@ \Delta t = t_{k+1} - t_k @)@, it defines the approximation solution @(@ \tilde{z}_{k+1} (x) @)@ at time @(@ t_{k+1} @)@ by @[@ \begin{array}{rcl} h_1 & = & h [ x , \tilde{z}_k (x) ] \\ h_2 & = & h [ x , \tilde{z}_k (x) + \Delta t \; h_1 / 2 ] \\ h_3 & = & h [ x , \tilde{z}_k (x) + \Delta t \; h_2 / 2 ] \\ h_4 & = & h [ x , \tilde{z}_k (x) + \Delta t \; h_3 ] \\ \tilde{z}_{k+1} (x) & = & \tilde{z}_k (x) + \Delta t \; ( h_1 + 2 h_2 + 2 h_3 + h_4 ) / 6 \end{array} @]@ If @(@ \tilde{z}_k (x) = z_k (x) @)@, then @(@ \tilde{z}_{k+1} (x) = z_{k+1} (x) + O( \Delta t^5 ) @)@, then Other ODE solvers can use a similar method to the one used below.

ODE
For this example the ODE is defined by @(@ z(0, x) = 0 @)@ and @[@ h[ x, z(t, x) ] = \left( \begin{array}{c} x_0 \\ x_1 z_0 (t, x) \\ \vdots \\ x_{n-1} z_{n-2} (t, x) \end{array} \right) = \left( \begin{array}{c} \partial_t z_0 (t , x) \\ \partial_t z_1 (t , x) \\ \vdots \\ \partial_t z_{n-1} (t , x) \end{array} \right) @]@

Solution
The solution of the ODE for this example, which is used to check the results, can be calculated by starting with the first row and then using the solution for the first row to solve the second and so on. Doing this we obtain @[@ z(t, x) = \left( \begin{array}{c} x_0 t \\ x_1 x_0 t^2 / 2 \\ \vdots \\ x_{n-1} x_{n-2} \ldots x_0 t^n / n ! \end{array} \right) @]@



# include <cppad/cppad.hpp>

namespace {
     using CppAD::AD;
     typedef AD<double>                     a1double;
     typedef AD<a1double>                   a2double;
     //
     typedef CPPAD_TESTVECTOR(   double )   a0vector;
     typedef CPPAD_TESTVECTOR( a1double )   a1vector;
     typedef CPPAD_TESTVECTOR( a2double )   a2vector;
     //
     // set once by main and kept that way
     double delta_t_ = std::numeric_limits<double>::quiet_NaN();
     size_t n_       = 0;
     //
     // The function h( x , y)
     template <class FloatVector>
     FloatVector h(const FloatVector& x, const FloatVector& y)
     {     assert( size_t( x.size() ) == n_ );
          assert( size_t( y.size() ) == n_ );
          FloatVector result(n_);
          result[0] = x[0];
          for(size_t i = 1; i < n_; i++)
               result[i] = x[i] * y[i-1];
          return result;
     }

     // The 4-th Order Runge-Kutta Step
     template <class FloatVector>
     FloatVector Runge4(const FloatVector& x, const FloatVector& z0
     )
     {     assert( size_t( x.size() ) == n_ );
          assert( size_t( z0.size() ) == n_ );
          //
          typedef typename FloatVector::value_type Float;
          //
          Float  dt = Float(delta_t_);
          size_t m  = z0.size();
          //
          FloatVector h1(m), h2(m), h3(m), h4(m), result(m);
          h1 = h( x, z0 );
          //
          for(size_t i = 0; i < m; i++)
               h2[i] = z0[i] + dt * h1[i] / 2.0;
          h2 = h( x, h2 );
          //
          for(size_t i = 0; i < m; i++)
               h3[i] = z0[i] + dt * h2[i] / 2.0;
          h3 = h( x, h3 );
          //
          for(size_t i = 0; i < m; i++)
               h4[i] = z0[i] + dt * h3[i];
          h4 = h( x, h4 );
          //
          for(size_t i = 0; i < m; i++)
          {     Float dz = dt * ( h1[i] + 2.0*h2[i] + 2.0*h3[i] + h4[i] ) / 6.0;
               result[i] = z0[i] + dz;
          }
          return result;
     }

     // Derivative of 4-th Order Runge-Kutta Step w.r.t x
     a1vector Runge4_x(const a1vector& x, const a1vector& z0)
     {     assert( size_t( x.size() ) == n_ );
          assert( size_t( z0.size() ) == n_ );
          //
          a2vector ax(n_);
          for(size_t j = 0; j < n_; j++)
               ax[j] = x[j];
          //
          a2vector az0(n_);
          for(size_t i = 0; i < n_; i++)
               az0[i] = z0[i];
          //
          CppAD::Independent(ax);
          a2vector az(n_);
          az = Runge4(ax, az0);
          CppAD::ADFun<a1double> f(ax, az);
          //
          a1vector result =  f.Jacobian(x);
          //
          return result;
     }

     // Derivative of 4-th Order Runge-Kutta Step w.r.t z0
     a1vector Runge4_z0(const a1vector& x, const a1vector& z0)
     {     assert( size_t( x.size()  ) == n_ );
          assert( size_t( z0.size() ) == n_ );
          //
          a2vector ax(n_);
          for(size_t j = 0; j < n_; j++)
               ax[j] = x[j];
          //
          a2vector az0(n_);
          for(size_t i = 0; i < n_; i++)
               az0[i] = z0[i];
          //
          CppAD::Independent(az0);
          a2vector az(n_);
          az = Runge4(ax, az0);
          CppAD::ADFun<a1double> f(az0, az);
          //
          a1vector result =  f.Jacobian(z0);
          //
          return result;
     }

     // pack an extended ode vector
     template <class FloatVector>
     void pack(
          FloatVector&         extended_ode ,
          const FloatVector&   x            ,
          const FloatVector&   z            ,
          const FloatVector&   z_x          )
     {     assert( size_t( extended_ode.size() ) == n_ + n_ + n_ * n_ );
          assert( size_t( x.size()            ) == n_                );
          assert( size_t( z.size()            ) == n_                );
          assert( size_t( z_x.size()          ) == n_ * n_           );
          //
          size_t offset = 0;
          for(size_t i = 0; i < n_; i++)
               extended_ode[offset + i] = x[i];
          offset += n_;
          for(size_t i = 0; i < n_; i++)
               extended_ode[offset + i] = z[i];
          offset += n_;
          for(size_t i = 0; i < n_; i++)
          {     for(size_t j = 0; j < n_; j++)
               {     // partial of z_i (t , x ) w.r.t x_j
                    extended_ode[offset + i * n_ + j] = z_x[i * n_ + j];
               }
          }
     }

     // unpack an extended ode vector
     template <class FloatVector>
     void unpack(
          const FloatVector&         extended_ode ,
          FloatVector&               x            ,
          FloatVector&               z            ,
          FloatVector&               z_x          )
     {     assert( size_t( extended_ode.size() ) == n_ + n_ + n_ * n_ );
          assert( size_t( x.size()            ) == n_                );
          assert( size_t( z.size()            ) == n_                );
          assert( size_t( z_x.size()          ) == n_ * n_           );
          //
          size_t offset = 0;
          for(size_t i = 0; i < n_; i++)
               x[i] = extended_ode[offset + i];
          offset += n_;
          for(size_t i = 0; i < n_; i++)
               z[i] = extended_ode[offset + i];
          offset += n_;
          for(size_t i = 0; i < n_; i++)
          {     for(size_t j = 0; j < n_; j++)
               {     // partial of z_i (t , x ) w.r.t x_j
                    z_x[i * n_ + j] = extended_ode[offset + i * n_ + j];
               }
          }
     }

     // Algorithm that advances the partial of z(t, x) w.r.t x
     void ext_ode_algo(const a1vector& ext_ode_in, a1vector& ext_ode_out)
     {     assert( size_t( ext_ode_in.size()  ) == n_ + n_ + n_ * n_ );
          assert( size_t( ext_ode_out.size() ) == n_ + n_ + n_ * n_ );
          //
          // initial extended ode information
          a1vector x(n_), z0(n_), z0_x(n_ * n_);
          unpack(ext_ode_in, x, z0, z0_x);
          //
          // advance z(t, x)
          a1vector z1 = Runge4(x, z0);
          //
          // partial of z1 w.r.t. x
          a1vector z1_x = Runge4_x(x, z0);
          //
          // partial of z1 w.r.t. z0
          a1vector z1_z0 = Runge4_z0(x, z0);
          //
          // total derivative of z1 w.r.t x
          for(size_t i = 0; i < n_; i++)
          {     for(size_t j = 0; j < n_; j++)
               {     a1double sum = 0.0;
                    for(size_t k = 0; k < n_; k++)
                         sum += z1_z0 [ i * n_ + k ] * z0_x [ k * n_ + j ];
                    z1_x[ i * n_ + j] += sum;
               }
          }
          //
          // final extended ode information
          pack(ext_ode_out, x, z1, z1_x);
          //
          return;
     }
}
//
bool extended_ode(void)
{     bool ok = true;
     using CppAD::NearEqual;
     double eps = std::numeric_limits<double>::epsilon();
     //
     // number of terms in the differential equation
     n_ = 6;
     //
     // step size for the differentiail equation
     size_t n_step = 10;
     double T      = 1.0;
     delta_t_ = T / double(n_step);
     //
     // set parameter value and initial value of the extended ode
     a1vector ax(n_), az0(n_), az0_x(n_ * n_);
     for(size_t i = 0; i < n_; i++)
     {     ax[i]  = a1double(i + 1);
          az0[i] = a1double(0);
          for(size_t j = 0; j < n_; j++)
               az0_x[ i * n_ + j ] = 0.0;
     }
     //
     // pack into extended ode information input vector
     size_t n_ext = n_ + n_ + n_ * n_;
     a1vector aext_ode_in(n_ext);
     pack(aext_ode_in, ax, az0, az0_x);
     //
     // create checkpoint version of the algorithm
     a1vector aext_ode_out(n_ext);
     CppAD::checkpoint<double> ext_ode_check(
          "ext_ode", ext_ode_algo, aext_ode_in, aext_ode_out
     );
     //
     // set the independent variables for recording
     CppAD::Independent( ax );
     //
     // repack to get dependence on ax
     pack(aext_ode_in, ax, az0, az0_x);
     //
     // Now run the checkpoint algorithm n_step times
     for(size_t k = 0; k < n_step; k++)
     {     ext_ode_check(aext_ode_in, aext_ode_out);
          aext_ode_in = aext_ode_out;
     }
     //
     // Unpack the results (must use ax1 so do not overwrite ax)
     a1vector ax1(n_), az1(n_), az1_x(n_ * n_);
     unpack(aext_ode_out, ax1, az1, az1_x);
     //
     // We could record a complicated funciton of x and z_x(T, x) in f,
     // but make this example simpler we record x -> z_x(T, x).
     CppAD::ADFun<double> f(ax, az1_x);
     //
     // check function values
     a0vector x(n_), z1(n_), z1_x(n_ * n_);
     for(size_t j = 0; j < n_; j++)
          x[j] = double(j + 1);
     z1_x = f.Forward(0, x);
     //
     // use z(t, x) for checking solution
     z1[0] = x[0] * T;
     for(size_t i = 1; i < n_; i++)
          z1[i] = x[i] * T * z1[i-1] / double(i+1);
     //
     // expected accuracy for each component of of z(t, x)
     a0vector acc(n_);
     for(size_t i = 0; i < n_; i++)
     {     if( i < 4 )
          {     // Runge-Kutta methos is exact for this case
               acc[i] = 10. * eps;
          }
          else
          {     acc[i] = 1.0;
               for(size_t k = 0; k < 5; k++)
                         acc[i] *= x[k] * delta_t_;
          }
     }
     // check z1(T, x)
     for(size_t i = 0; i < n_; i++)
     {     for(size_t j = 0; j < n_; j++)
          {     // check partial of z1_i w.r.t x_j
               double check = 0.0;
               if( j <= i )
                    check = z1[i] / x[j];
               ok &= NearEqual(z1_x[ i * n_ + j ] , check, acc[i], acc[i]);
          }
     }
     //
     // Now use f to compute a derivative. For this 'simple' example it is
     // the derivative with respect to x of the
     // parital with respect to x[n-1] of z_{n-1} (t , x)
     a0vector w(n_ * n_), dw(n_);
     for(size_t i = 0; i < n_; i++)
     {     for(size_t j = 0; j < n_; j++)
          {     w[ i * n_ + j ] = 0.0;
               if( i == n_ - 1 && j == n_ - 1 )
                    w[ i * n_ + j ] = 1.0;
          }
     }
     dw = f.Reverse(1, w);
     for(size_t j = 0; j < n_; j++)
     {     double check = 0.0;
          if( j < n_ - 1 )
               check = z1[n_ - 1] / ( x[n_ - 1] * x[j] );
          ok &= NearEqual(dw[j] , check, acc[n_-1], acc[n_-1]);
     }
     //
     return ok;
}

Input File: example/atomic/extended_ode.cpp