package jsat.distributions;
import jsat.linear.Vec;
import jsat.math.integration.Romberg;
import jsat.math.optimization.oned.GoldenSearch;
import jsat.math.rootfinding.Zeroin;
/**
* The ContinuousDistribution represents the contract for a continuous in one
* dimension.
*
* Many of the functions of a Continuous Distribution are implemented by default
* using numerical calculation and integration. For this reason, the base
* implementations may be slower or less accurate than desired - and could
* produce incorrect results for poorly behaved functions or large magnitude
* inputs. These base implementations are provided for easy completeness, but
* may not be appropriate for all methods. If needed, the implementer should
* check if these methods provide the needed level of accuracy and speed.
*
* @author Edward Raff
*/
public abstract class ContinuousDistribution extends Distribution
{
private static final long serialVersionUID = -5079392926462355615L;
/**
* Computes the log of the Probability Density Function. Note, that then the
* probability is zero, {@link Double#NEGATIVE_INFINITY} would be the true
* value. Instead, this method will always return the negative of
* {@link Double#MAX_VALUE}. This is to avoid propagating bad values through
* computation.
*
* @param x the value to get the log(PDF) of
* @return the value of log(PDF(x))
*/
public double logPdf(double x)
{
double pdf = pdf(x);
if(pdf <= 0)
return -Double.MAX_VALUE;
return Math.log(pdf);
}
/**
* Computes the value of the Probability Density Function (PDF) at the given point
* @param x the value to get the PDF
* @return the PDF(x)
*/
abstract public double pdf(double x);
@Override
public double cdf(double x)
{
double intMin = getIntegrationMin();
return Romberg.romb((z)->this.pdf(z), intMin, x);
}
@Override
public double invCdf(final double p)
{
if (p < 0 || p > 1)
throw new ArithmeticException("Value of p must be in the range [0,1], not " + p);
double a = getIntegrationMin();
double b = getIntegrationMax();
//default case, lets just do a root finding on the CDF for the specific value of p
return Zeroin.root(a, b, (x) -> this.cdf(x) - p);
}
@Override
public double mean()
{
double intMin = getIntegrationMin();
double intMax = getIntegrationMax();
return Romberg.romb((double x) -> x*pdf(x), intMin, intMax);
}
@Override
public double variance()
{
double intMin = getIntegrationMin();
double intMax = getIntegrationMax();
final double mean = mean();
return Romberg.romb((x)->Math.pow(x-mean, 2)*pdf(x), intMin, intMax);
}
@Override
public double skewness()
{
double intMin = getIntegrationMin();
double intMax = getIntegrationMax();
final double mean = mean();
return Romberg.romb((x)->Math.pow((x-mean), 3)*pdf(x), intMin, intMax)/Math.pow(variance(), 3.0/2);
}
@Override
public double mode()
{
double intMin = getIntegrationMin();
double intMax = getIntegrationMax();
return GoldenSearch.findMin(intMin, intMax, (x)->-pdf(x), 1e-6, 1000);
}
protected double getIntegrationMin()
{
double intMin = min();
if(Double.isInfinite(intMin))
{
intMin = -Double.MAX_VALUE/4;
//Lets find a suitbly small PDF starting value for this
//first, lets take big steps
for(int i = 0; i < 8; i++)
{
double sqrt = Math.sqrt(-intMin);
if(pdf(sqrt) < 1e-5)
intMin = -sqrt;
else
break;//no more big steps
}
//keep going until it looks like we should switch signs
while(pdf(intMin) < 1e-5 && intMin < -0.1)
{
intMin/=2;
}
if(pdf(intMin) < 1e-5)//still?
intMin *=-1;
//ok, search positive... keep multiplying till we get there
while(pdf(intMin) < 1e-5)
{
intMin *= 2;
}
}
return intMin;
}
protected double getIntegrationMax()
{
double intMax = max();
if(Double.isInfinite(intMax))
{
intMax = Double.MAX_VALUE / 4;
//Lets find a suitbly small PDF starting value for this
//first, lets take big steps
for (int i = 0; i < 8; i++)
{
double sqrt = Math.sqrt(intMax);
if (pdf(sqrt) < 1e-5)
intMax = sqrt;
else
break;//no more big steps
}
//keep going until it looks like we should switch signs
while (pdf(intMax) < 1e-5 && intMax > 0.1)
{
intMax /= 2;
}
if (pdf(intMax) < 1e-5)//still?
intMax *= -1;
//ok, search negative... keep multiplying till we get there
while (pdf(intMax) < 1e-5)
{
intMax *= 2;
}
}
return intMax;
}
/**
* The descriptive name of a distribution returns the name of the distribution, followed by the parameters of the distribution and their values.
* @return the name of the distribution that includes parameter values
*/
public String getDescriptiveName()
{
StringBuilder sb = new StringBuilder(getDistributionName());
sb.append("(");
String[] vars = getVariables();
double[] vals = getCurrentVariableValues();
sb.append(vars[0]).append(" = ").append(vals[0]);
for(int i = 1; i < vars.length; i++)
sb.append(", ").append(vars[i]).append(" = ").append(vals[i]);
sb.append(")");
return sb.toString();
}
/**
* Return the name of the distribution.
* @return the name of the distribution.
*/
abstract public String getDistributionName();
/**
* Returns an array, where each value contains the name of a parameter in the distribution.
* The order must always be the same, and match up with the values returned by {@link #getCurrentVariableValues() }
*
* @return a string of the variable names this distribution uses
*/
abstract public String[] getVariables();
/**
* Returns an array, where each value contains the value of a parameter in the distribution.
* The order must always be the same, and match up with the values returned by {@link #getVariables() }
* @return the current values of the parameters used by this distribution, in the same order as their names are returned by {@link #getVariables() }
*/
abstract public double[] getCurrentVariableValues();
/**
* Sets one of the variables of this distribution by the name.
* @param var the variable to set
* @param value the value to set
*/
abstract public void setVariable(String var, double value);
@Override
abstract public ContinuousDistribution clone();
/**
* Attempts to set the variables used by this distribution based on population sample data,
* assuming the sample data is from this type of distribution.
*
* @param data the data to use to attempt to fit against
*/
abstract public void setUsingData(Vec data);
@Override
public String toString()
{
return getDistributionName();
}
}
Ads help maintain this website.