/// <include file='FactorDocs.xml' path='factor_docs/message_op_class[@name="InnerProductOpBase"]/message_doc[@name="InnerProductAverageLogarithm(DenseVector, PositiveDefiniteMatrix, DenseVector, PositiveDefiniteMatrix)"]/*'/>
public static Gaussian InnerProductAverageLogarithm(DenseVector AMean, PositiveDefiniteMatrix AVariance, DenseVector BMean, PositiveDefiniteMatrix BVariance)
{
Gaussian result = new Gaussian();
// p(x|a,b) = N(E[a]'*E[b], E[b]'*var(a)*E[b] + E[a]'*var(b)*E[a] + trace(var(a)*var(b)))
// Uses John Winn's rule for deterministic factors.
// Strict variational inference would set the variance to 0.
result.SetMeanAndVariance(AMean.Inner(BMean), AVariance.QuadraticForm(BMean) + BVariance.QuadraticForm(AMean) + AVariance.Inner(BVariance));
return(result);
}
/// <summary>
/// Create a COM-Poisson distribution with the given sufficient statistics.
/// </summary>
/// <param name="mean">E[x]</param>
/// <param name="meanLogFactorial">E[log(x!)]</param>
/// <returns>A new COM-Poisson distribution</returns>
/// <remarks>
/// This routine implements maximum-likelihood estimation of a COM-Poisson distribution, as described by
/// Thomas P. Minka, Galit Shmueli, Joseph B. Kadane, Sharad Borle, and Peter Boatwright,
/// "Computing with the COM-Poisson distribution", CMU Tech Report 776.
/// http://www.stat.cmu.edu/tr/tr776/tr776.html
/// </remarks>
public static Poisson FromMeanAndMeanLogFactorial(double mean, double meanLogFactorial)
{
DenseVector theta = DenseVector.Zero(2);
theta[0] = Math.Log(mean);
theta[1] = 1;
DenseVector newTheta = DenseVector.Zero(2);
DenseVector gradient2 = DenseVector.Zero(2);
DenseVector gradient = DenseVector.Zero(2);
PositiveDefiniteMatrix hessian = new PositiveDefiniteMatrix(2, 2);
double stepsize = 1;
const int numIter = 1000;
double rate = Math.Exp(theta[0]);
double precision = theta[1];
double logZ = GetLogNormalizer(rate, precision);
double meanEst = Math.Exp(GetLogPowerSum(rate, precision, 1) - logZ);
double Z = Math.Exp(logZ);
double meanLogFactorialEst = GetSumLogFactorial(rate, precision) / Z;
for (int iter = 0; iter < numIter; iter++)
{
double varEst = Math.Exp(GetLogPowerSum(rate, precision, 2) - logZ) - meanEst * meanEst;
double varLogFactorialEst = GetSumLogFactorial2(rate, precision) / Z - meanLogFactorialEst * meanLogFactorialEst;
double covXLogFactorialEst = GetSumXLogFactorial(rate, precision) / Z - meanEst * meanLogFactorialEst;
gradient[0] = mean - meanEst;
gradient[1] = meanLogFactorialEst - meanLogFactorial;
double sqrGrad = gradient.Inner(gradient);
// the Hessian is guaranteed to be positive definite
hessian[0, 0] = varEst;
hessian[0, 1] = -covXLogFactorialEst;
hessian[1, 0] = hessian[0, 1];
hessian[1, 1] = varLogFactorialEst;
gradient.PredivideBy(hessian);
// line search to reduce sqrGrad
// we use sqrGrad instead of the likelihood since the point of zero gradient may not exactly match the likelihood mode due to numerical inaccuracy
while (true)
{
newTheta.SetToSum(1.0, theta, stepsize, gradient);
double sqrGrad2 = sqrGrad;
if (newTheta[1] >= 0)
{
rate = Math.Exp(newTheta[0]);
precision = newTheta[1];
logZ = GetLogNormalizer(rate, precision);
meanEst = Math.Exp(GetLogPowerSum(rate, precision, 1) - logZ);
Z = Math.Exp(logZ);
meanLogFactorialEst = GetSumLogFactorial(rate, precision) / Z;
gradient2[0] = mean - meanEst;
gradient2[1] = meanLogFactorialEst - meanLogFactorial;
sqrGrad2 = gradient2.Inner(gradient2);
}
if (sqrGrad2 < sqrGrad)
{
theta.SetTo(newTheta);
gradient.SetTo(gradient2);
stepsize *= 2;
break;
}
else
{
stepsize /= 2;
if (stepsize == 0)
{
break;
}
}
}
//Console.WriteLine("{0}: {1}", iter, theta);
double maxGradient = Math.Max(Math.Abs(gradient[0]), Math.Abs(gradient[1]));
if (maxGradient < 1e-10)
{
break;
}
if (stepsize == 0)
{
break;
}
if (iter == numIter - 1)
{
throw new Exception("not converging");
}
}
return(new Poisson(Math.Exp(theta[0]), theta[1]));
}
