vec_ad.cpp

View page source

AD Vectors that Record Index Operations: Example and Test

# include <cppad/cppad.hpp>
# include <cassert>

namespace {
   // return the vector x that solves the following linear system
   //    a[0] * x[0] + a[1] * x[1] = b[0]
   //    a[2] * x[0] + a[3] * x[1] = b[1]
   // in a way that will record pivot operations on the AD<double> tape
   typedef CPPAD_TESTVECTOR(CppAD::AD<double>) Vector;
   Vector Solve(const Vector &a , const Vector &b)
   {  using namespace CppAD;
      assert(a.size() == 4 && b.size() == 2);

      // copy the vector b into the VecAD object B
      VecAD<double> B(2);
      AD<double>    u;
      for(u = 0; u < 2; u += 1.)
         B[u] = b[ size_t( Integer(u) ) ];

      // copy the matrix a into the VecAD object A
      VecAD<double> A(4);
      for(u = 0; u < 4; u += 1.)
         A[u] = a [ size_t( Integer(u) ) ];

      // tape AD operation sequence that determines the row of A
      // with maximum absolute element in column zero
      AD<double> zero(0), one(1);
      AD<double> rmax = CondExpGt(fabs(a[0]), fabs(a[2]), zero, one);

      // divide row rmax by A(rmax, 0)
      A[rmax * 2 + 1]  = A[rmax * 2 + 1] / A[rmax * 2 + 0];
      B[rmax]          = B[rmax]         / A[rmax * 2 + 0];
      A[rmax * 2 + 0]  = one;

      // subtract A(other,0) times row A(rmax, *) from row A(other,*)
      AD<double> other   = one - rmax;
      A[other * 2 + 1]   = A[other * 2 + 1]
                           - A[other * 2 + 0] * A[rmax * 2 + 1];
      B[other]           = B[other]
                           - A[other * 2 + 0] * B[rmax];
      A[other * 2 + 0] = zero;

      // back substitute to compute the solution vector x.
      // Note that the columns of A correspond to rows of x.
      // Also note that A[rmax * 2 + 0] is equal to one.
      CPPAD_TESTVECTOR(AD<double>) x(2);
      x[1] = B[other] / A[other * 2 + 1];
      x[0] = B[rmax] - A[rmax * 2 + 1] * x[1];

      return x;
   }
}

bool vec_ad(void)
{  bool ok = true;

   using CppAD::AD;
   using CppAD::NearEqual;
   double eps99 = 99.0 * std::numeric_limits<double>::epsilon();

   // domain space vector
   size_t n = 4;
   CPPAD_TESTVECTOR(double)       x(n);
   CPPAD_TESTVECTOR(AD<double>) X(n);
   // 2 * identity matrix (rmax in Solve will be 0)
   X[0] = x[0] = 2.; X[1] = x[1] = 0.;
   X[2] = x[2] = 0.; X[3] = x[3] = 2.;

   // declare independent variables and start tape recording
   CppAD::Independent(X);

   // define the vector b
   CPPAD_TESTVECTOR(double)       b(2);
   CPPAD_TESTVECTOR(AD<double>) B(2);
   B[0] = b[0] = 0.;
   B[1] = b[1] = 1.;

   // range space vector solves X * Y = b
   size_t m = 2;
   CPPAD_TESTVECTOR(AD<double>) Y(m);
   Y = Solve(X, B);

   // create f: X -> Y and stop tape recording
   CppAD::ADFun<double> f(X, Y);

   // By Cramer's rule:
   // y[0] = [ b[0] * x[3] - x[1] * b[1] ] / [ x[0] * x[3] - x[1] * x[2] ]
   // y[1] = [ x[0] * b[1] - b[0] * x[2] ] / [ x[0] * x[3] - x[1] * x[2] ]

   double den   = x[0] * x[3] - x[1] * x[2];
   double dsq   = den * den;
   double num0  = b[0] * x[3] - x[1] * b[1];
   double num1  = x[0] * b[1] - b[0] * x[2];

   // check value
   ok &= NearEqual(Y[0] , num0 / den, eps99, eps99);
   ok &= NearEqual(Y[1] , num1 / den, eps99, eps99);

   // forward computation of partials w.r.t. x[0]
   CPPAD_TESTVECTOR(double) dx(n);
   CPPAD_TESTVECTOR(double) dy(m);
   dx[0] = 1.; dx[1] = 0.;
   dx[2] = 0.; dx[3] = 0.;
   dy    = f.Forward(1, dx);
   ok &= NearEqual(dy[0], 0.         - num0 * x[3] / dsq, eps99, eps99);
   ok &= NearEqual(dy[1], b[1] / den - num1 * x[3] / dsq, eps99, eps99);

   // compute the solution for a new x matrix such that pivioting
   // on the original rmax row would divide by zero
   CPPAD_TESTVECTOR(double) y(m);
   x[0] = 0.; x[1] = 2.;
   x[2] = 2.; x[3] = 0.;

   // new values for Cramer's rule
   den   = x[0] * x[3] - x[1] * x[2];
   dsq   = den * den;
   num0  = b[0] * x[3] - x[1] * b[1];
   num1  = x[0] * b[1] - b[0] * x[2];

   // check values
   y    = f.Forward(0, x);
   ok &= NearEqual(y[0] , num0 / den, eps99, eps99);
   ok &= NearEqual(y[1] , num1 / den, eps99, eps99);

   // forward computation of partials w.r.t. x[1]
   dx[0] = 0.; dx[1] = 1.;
   dx[2] = 0.; dx[3] = 0.;
   dy    = f.Forward(1, dx);
   ok   &= NearEqual(dy[0],-b[1] / den + num0 * x[2] / dsq, eps99, eps99);
   ok   &= NearEqual(dy[1], 0.         + num1 * x[2] / dsq, eps99, eps99);

   // reverse computation of derivative of y[0] w.r.t x
   CPPAD_TESTVECTOR(double) w(m);
   CPPAD_TESTVECTOR(double) dw(n);
   w[0] = 1.; w[1] = 0.;
   dw   = f.Reverse(1, w);
   ok  &= NearEqual(dw[0], 0.         - num0 * x[3] / dsq, eps99, eps99);
   ok  &= NearEqual(dw[1],-b[1] / den + num0 * x[2] / dsq, eps99, eps99);
   ok  &= NearEqual(dw[2], 0.         + num0 * x[1] / dsq, eps99, eps99);
   ok  &= NearEqual(dw[3], b[0] / den - num0 * x[0] / dsq, eps99, eps99);

   return ok;
}