edu.cmu.sphinx.util

Class LogMath

java.lang.Object

edu.cmu.sphinx.util.LogMath

public final class LogMath
extends java.lang.Object

Provides a set of methods for performing simple math in the log domain. The logarithmic base can be set by the property: edu.cmu.sphinx.util.LogMath.logBase

Field Summary

Fields

Modifier and Type	Field and Description
static float	LOG_ONE
static float	LOG_ZERO

Method Summary

Modifier and Type	Method and Description
float	addAsLinear(float logVal1, float logVal2) Returns the summation of two numbers when the arguments and the result are in log.
float	getLogBase()
static LogMath	getLogMath()
boolean	isUseTable()
float	linearToLog(double linearValue) Converts the value from linear scale to log scale.
void	linearToLog(float[] vector) Converts a vector from linear domain to log domain using a given LogMath-instance for conversion.
float	lnToLog(float logSource) Converts the source, which is a number in base Math.E, to a log value which base is the LogBase of this LogMath.
static float	log10(float value) Returns the log (base 10) of value
float	log10ToLog(float logSource) Converts the source, which is a number in base 10, to a log value which base is the LogBase of this LogMath.
double	logToLinear(float logValue) Converts the value from log scale to linear scale.
void	logToLinear(float[] vector, float[] out) Converts a vector from log to linear domain using a given LogMath-instance for conversion.
float	logToLn(float logSource) Converts the source, whose base is the LogBase of this LogMath, to a log value which is a number in base Math.E.
static float	logToLog(float logSource, float sourceBase, float resultBase) Converts the source, which is assumed to be a log value whose base is sourceBase, to a log value whose base is resultBase.
static void	setLogBase(float logBase) Sets log base.
static void	setUseTable(boolean useTable) The property that controls whether we use the old, slow (but correct) method of performing the LogMath.add by doing the actual computation.
float	subtractAsLinear(float logMinuend, float logSubtrahend) Returns the difference between two numbers when the arguments and the result are in log.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

LOG_ZERO

public static final float LOG_ZERO

See Also:

Constant Field Values

LOG_ONE

public static final float LOG_ONE

See Also:

Constant Field Values

Method Detail

getLogMath

public static LogMath getLogMath()

setLogBase

public static void setLogBase(float logBase)

Sets log base.

According to forum discussions a value between 1.00001 and 1.0004 should be used for speech recognition. Going above 1.0005 will probably hurt.

Parameters:

logBase - Log base

setUseTable

public static void setUseTable(boolean useTable)

The property that controls whether we use the old, slow (but correct) method of performing the LogMath.add by doing the actual computation.

Parameters:

useTable - to configure table lookups

addAsLinear

public final float addAsLinear(float logVal1,
                               float logVal2)

Returns the summation of two numbers when the arguments and the result are in log.

That is, it returns log(a + b) given log(a) and log(b)

This method makes use of the equality:

log(a + b) = log(a) + log (1 + exp(log(b) - log(a)))

which is derived from:

a + b = a * (1 + (b / a))

which in turns makes use of:

b / a = exp (log(b) - log(a))

Important to notice that subtractAsLinear(a, b) is *not* the same as addAsLinear(a, -b), since we're in the log domain, and -b is in fact the inverse.

No underflow/overflow check is performed.

Parameters:

logVal1 - value in log domain (i.e. log(val1)) to add

logVal2 - value in log domain (i.e. log(val2)) to add

Returns:

sum of val1 and val2 in the log domain

subtractAsLinear

public final float subtractAsLinear(float logMinuend,
                                    float logSubtrahend)
                             throws java.lang.IllegalArgumentException

Returns the difference between two numbers when the arguments and the result are in log.

That is, it returns log(a - b) given log(a) and log(b)

Implementation is less efficient than add(), since we're less likely to use this function, provided for completeness. Notice however that the result only makes sense if the minuend is higher than the subtrahend. Otherwise, we should return the log of a negative number.

It implements the subtraction as:

log(a - b) = log(a) + log(1 - exp(log(b) - log(a)))

No need to check for underflow/overflow.

Parameters:

logMinuend - value in log domain (i.e. log(minuend)) to be subtracted from

logSubtrahend - value in log domain (i.e. log(subtrahend)) that is being subtracted

Returns:

difference between minuend and the subtrahend in the log domain

Throws:

java.lang.IllegalArgumentException -

This is a very slow way to do this, but this method should rarely be used.

logToLog

public static float logToLog(float logSource,
                             float sourceBase,
                             float resultBase)
                      throws java.lang.IllegalArgumentException

Converts the source, which is assumed to be a log value whose base is sourceBase, to a log value whose base is resultBase. Possible values for both the source and result bases include Math.E, 10.0, LogMath.getLogBase(). If a source or result base is not supported, an IllegalArgumentException will be thrown.

It takes advantage of the relation:

log_a(b) = log_c(b) / lob_c(a)

or:

log_a(b) = log_c(b) * lob_a(c)

where log_a(b) is logarithm of b base a etc.

Parameters:

logSource - log value whose base is sourceBase

sourceBase - the base of the log the source

resultBase - the base to convert the source log to

Returns:

converted value

Throws:

java.lang.IllegalArgumentException - if arguments out of bounds

lnToLog

public final float lnToLog(float logSource)

Converts the source, which is a number in base Math.E, to a log value which base is the LogBase of this LogMath.

Parameters:

logSource - the number in base Math.E to convert

Returns:

converted value

log10ToLog

public final float log10ToLog(float logSource)

Converts the source, which is a number in base 10, to a log value which base is the LogBase of this LogMath.

Parameters:

logSource - the number in base Math.E to convert

Returns:

converted value

logToLn

public final float logToLn(float logSource)

Converts the source, whose base is the LogBase of this LogMath, to a log value which is a number in base Math.E.

Parameters:

logSource - the number to convert to base Math.E

Returns:

converted value

linearToLog

public final float linearToLog(double linearValue)
                        throws java.lang.IllegalArgumentException

Converts the value from linear scale to log scale. The log scale numbers are limited by the range of the type float. The linear scale numbers can be any double value.

Parameters:

linearValue - the value to be converted to log scale

Returns:

the value in log scale

Throws:

java.lang.IllegalArgumentException - if value out of range

logToLinear

public final double logToLinear(float logValue)

Converts the value from log scale to linear scale.

Parameters:

logValue - the value to be converted to the linear scale

Returns:

the value in the linear scale

getLogBase

public final float getLogBase()

Returns:

the actual log base.

isUseTable

public boolean isUseTable()

log10

public static float log10(float value)

Returns the log (base 10) of value

Parameters:

value - the value to take the log of

Returns:

the log (base 10) of value

linearToLog

public void linearToLog(float[] vector)

Converts a vector from linear domain to log domain using a given LogMath-instance for conversion.

Parameters:

vector - to convert in-place

logToLinear

public void logToLinear(float[] vector,
                        float[] out)

Converts a vector from log to linear domain using a given LogMath-instance for conversion.

Parameters:

vector - to convert

out - result