Hessian (second derivative) of error calculation for a logistic regression model.

LogisticRegressionHessian.java

//returns the hessian (gradient of gradient) of error with respect to weights 
//for a logistic regression with weights w on the given input and output
//output should be in the form 0 for negative, 1 for positive
public double[][] hessian(double w[]){
  heval++;//keep track of how many times this has been called
  double h[][] = new double[w.length][] ;
  //second derivative matrices are always symmetric so we only need triangular portion
  for(int j=0;j<h.length;j++){
    h[j] = new double [j+1] ;
  }
  for(int k=0;k<input.length;k++){
    //calculate coefficient{
    double dot = dot(w,input[k]) ;
    double ds = ds(dot) ;
    double coef = 2 * ( ds*ds + dds(dot) * (s(dot) - output[k])  ) ;
    for(int j=0;j<h.length;j++){// add x * x^t * coef to hessian
      for(int l=0;l<=j;l++){
        h[j][l] += input[k][j]*input[k][l] * coef ;
      }
    }
  }
  return h ;
}

//returns a numerically calculated hessian - approximation to above
//used only for unit testing hessian, not called in final version
public double[][] numericalHessian(double w[], double epsilon){
  double h[][] = new double[w.length][] ;
  for(int j=0;j<h.length;j++){
    w[j]+=epsilon ;
    h[j] = gradient(w) ;
    w[j] -= 2*epsilon ;
    h[j] = subtract(h[j], gradient(w)) ;
    w[j] +=epsilon ;
    for(int k=0;k<w.length;k++){
      h[j][k]/=2*epsilon ;
    }
  }
  return h ;
}