Purpose
This is a realistic example using
two levels of AD; see mul_level
.
The first level uses AD<double> to tape the solution of an
ordinary differential equation.
This solution is then differentiated with respect to a parameter vector.
The second level uses AD< AD<double> >
to take derivatives during the solution of the differential equation.
These derivatives are used in the application
of Taylor's method to the solution of the ODE.
ODE
For this example the function
@(@
y : \B{R} \times \B{R}^n \rightarrow \B{R}^n
@)@ is defined by
@(@
y(0, x) = 0
@)@ and
@(@
\partial_t y(t, x) = g(y, x)
@)@ where
@(@
g : \B{R}^n \times \B{R}^n \rightarrow \B{R}^n
@)@ is defined by
@[@
g(y, x) =
\left( \begin{array}{c}
x_0 \\
x_1 y_0 \\
\vdots \\
x_{n-1} y_{n-2}
\end{array} \right)
@]@
ODE Solution
The solution for this example 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
@[@
y(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)
@]@
Derivative of ODE Solution
Differentiating the solution above,
with respect to the parameter vector @(@
x
@)@,
we notice that
@[@
\partial_x y(t, x ) =
\left( \begin{array}{cccc}
y_0 (t,x) / x_0 & 0 & \cdots & 0 \\
y_1 (t,x) / x_0 & y_1 (t,x) / x_1 & 0 & \vdots \\
\vdots & \vdots & \ddots & 0 \\
y_{n-1} (t,x) / x_0 & y_{n-1} (t,x) / x_1 & \cdots & y_{n-1} (t,x) / x_{n-1}
\end{array} \right)
@]@
Taylor's Method Using AD
We define the function @(@
z(t, x)
@)@ by the equation
@[@
z ( t , x ) = g[ y ( t , x ) ] = h [ x , y( t , x ) ]
@]@
see taylor_ode
for the method used to compute the
Taylor coefficients w.r.t @(@
t
@)@ of @(@
y(t, x)
@)@.
# include <cppad/cppad.hpp>
// =========================================================================// define types for each levelnamespace { // BEGIN empty namespacetypedef CppAD::AD<double> a1double;
typedef CppAD::AD<a1double> a2double;
typedefCPPAD_TESTVECTOR(double) d_vector;
typedefCPPAD_TESTVECTOR(a1double) a1vector;
typedefCPPAD_TESTVECTOR(a2double) a2vector;
// -------------------------------------------------------------------------// class definition for C++ function object that defines ODEclass Ode {
private:
// copy of a that is set by constructor and used by g(y)
a1vector a1x_;
public:
// constructorOde(const a1vector& a1x) : a1x_(a1x)
{ }
// the function g(y) is evaluated with two levels of taping
a2vector operator()
( const a2vector& a2y) const
{ size_t n = a2y.size();
a2vector a2g(n);
size_t i;
a2g[0] = a1x_[0];
for(i = 1; i < n; i++)
a2g[i] = a1x_[i] * a2y[i-1];
return a2g;
}
};
// -------------------------------------------------------------------------// Routine that uses Taylor's method to solve ordinary differential equaitons// and allows for algorithmic differentiation of the solution.
a1vector taylor_ode(
Ode G , // function that defines the ODE
size_t order , // order of Taylor's method used
size_t nstep , // number of steps to takeconst a1double& a1dt , // Delta t for each stepconst a1vector& a1y_ini) // y(t) at the initial time
{
// some temporary indices
size_t i, k, ell;
// number of variables in the ODE
size_t n = a1y_ini.size();
// copies of x and g(y) with two levels of taping
a2vector a2y(n), a2z(n);
// y, y^{(k)} , z^{(k)}, and y^{(k+1)}
a1vector a1y(n), a1y_k(n), a1z_k(n), a1y_kp(n);
// initialize xfor(i = 0; i < n; i++)
a1y[i] = a1y_ini[i];
// loop with respect to each step of Taylors methodfor(ell = 0; ell < nstep; ell++)
{ // prepare to compute derivatives using a1doublefor(i = 0; i < n; i++)
a2y[i] = a1y[i];
CppAD::Independent(a2y);
// evaluate ODE in a2double
a2z = G(a2y);
// define differentiable version of a1g: y -> z// that computes its derivatives using a1double objects
CppAD::ADFun<a1double> a1g(a2y, a2z);
// Use Taylor's method to take a step
a1y_k = a1y; // initialize y^{(k)}
a1double a1dt_kp = a1dt; // initialize dt^(k+1)for(k = 0; k <= order; k++)
{ // evaluate k-th order Taylor coefficient of y
a1z_k = a1g.Forward(k, a1y_k);
for(i = 0; i < n; i++)
{ // convert to (k+1)-Taylor coefficient for x
a1y_kp[i] = a1z_k[i] / a1double(k + 1);
// add term for to this Taylor coefficient// to solution for y(t, x)
a1y[i] += a1y_kp[i] * a1dt_kp;
}
// next power of t
a1dt_kp *= a1dt;
// next Taylor coefficient
a1y_k = a1y_kp;
}
}
return a1y;
}
} // END empty namespace// ==========================================================================// Routine that tests alogirhtmic differentiation of solutions computed// by the routine taylor_ode.
bool mul_level_ode(void)
{ bool ok = true;
double eps = 100. * std::numeric_limits<double>::epsilon();
// number of components in differential equation
size_t n = 4;
// some temporary indices
size_t i, j;
// parameter vector in both double and a1double
d_vector x(n);
a1vector a1x(n);
for(i = 0; i < n; i++)
a1x[i] = x[i] = double(i + 1);
// declare the parameters as the independent variable
CppAD::Independent(a1x);
// arguments to taylor_ode
Ode G(a1x); // function that defines the ODE
size_t order = n; // order of Taylor's method used
size_t nstep = 2; // number of steps to take
a1double a1dt = double(1.); // Delta t for each step// value of y(t, x) at the initial time
a1vector a1y_ini(n);
for(i = 0; i < n; i++)
a1y_ini[i] = 0.;
// integrate the differential equation
a1vector a1y_final(n);
a1y_final = taylor_ode(G, order, nstep, a1dt, a1y_ini);
// define differentiable fucntion object f : x -> y_final// that computes its derivatives in double
CppAD::ADFun<double> f(a1x, a1y_final);
// check function values
double check = 1.;
double t = double(nstep) * Value(a1dt);
for(i = 0; i < n; i++)
{ check *= x[i] * t / double(i + 1);
ok &= CppAD::NearEqual(Value(a1y_final[i]), check, eps, eps);
}
// evaluate the Jacobian of h at a
d_vector jac ( f.Jacobian(x) );
// There appears to be a bug in g++ version 4.4.2 because it generates// a warning for the equivalent form// d_vector jac = f.Jacobian(x);// check Jacobianfor(i = 0; i < n; i++)
{ for(j = 0; j < n; j++)
{ double jac_ij = jac[i * n + j];
if( i < j )
check = 0.;
else
check = Value( a1y_final[i] ) / x[j];
ok &= CppAD::NearEqual(jac_ij, check, eps, eps);
}
}
return ok;
}