uk.ac.gla.dcs.renaissance.iqir.qpr
Class EigenDecomposition

java.lang.Object
  uk.ac.gla.dcs.renaissance.iqir.qpr.EigenDecomposition

All Implemented Interfaces:

Serializable

Direct Known Subclasses:

Density, Subspace

public class EigenDecomposition
extends Object
implements Serializable

Common class shared by fuzzy subspaces and densities

The underlying density/subspace is represented by U (orthonormal) and S (diagonal) such as rho = U S^2 U^T

Author:

B. Piwowarski

See Also:

Serialized Form

Field Summary

it.unimi.dsi.fastutil.ints.Int2IntSortedMap	map Map from global id (index-based) to local ids - the map should be ordered (this is needed for fast intersections)
bpiwowar.maths.matrix.DiagonalDoubleMatrix	mS The singular values
bpiwowar.maths.matrix.DoubleMatrix2D	mU The basis of the subspace

Constructor Summary

EigenDecomposition(DensityTracker tracker, boolean deepCopy)
Construct a new eigen decomposition subspace/density based on a density tracker

EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix1D ray, boolean local)
Creates a one dimensional eigen-decomposition representation

EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix2D u, bpiwowar.maths.matrix.DiagonalDoubleMatrix s, it.unimi.dsi.fastutil.ints.Int2IntSortedMap map)

EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix2D mU, bpiwowar.maths.matrix.DiagonalDoubleMatrix mS, it.unimi.dsi.fastutil.ints.Int2IntSortedMap map, boolean deepCopy)

EigenDecomposition(bpiwowar.maths.svd.IncrementalSVD svd, it.unimi.dsi.fastutil.ints.Int2IntSortedMap map)

Method Summary

bpiwowar.maths.matrix.DoubleMatrix2D	getGlobalMatrix(int dimension) Get a matrix representation of this density (mostly to be used for debugging)
int[]	getLocalToGobalTermIdMap() Return the reverse map
bpiwowar.maths.matrix.DoubleMatrix2D	getMatrix() Get a matrix representation of this density (mostly to be used for debugging)
int	getRank()
bpiwowar.maths.matrix.DiagonalDoubleMatrix	getS()
double	normalise() Normalise with the L2 norm
double	normalise(boolean orthonormalU) Normalise the decomposition so that || U x S || = 1, i.e.
double	normalise(boolean orthonormalU, cern.colt.function.DoubleFunction g, cern.colt.function.DoubleDoubleFunction aggr, cern.colt.function.DoubleFunction f) Normalise the decomposition by a given norm normaliser = g(aggr_{i=1..n} f(x_i) )
void	trim(int newRank) Trim the eigenvalue decomposition to a lower rank

Methods inherited from class java.lang.Object

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

Field Detail

mS

public bpiwowar.maths.matrix.DiagonalDoubleMatrix mS

The singular values

mU

public bpiwowar.maths.matrix.DoubleMatrix2D mU

The basis of the subspace

map

public final it.unimi.dsi.fastutil.ints.Int2IntSortedMap map

Map from global id (index-based) to local ids - the map should be ordered (this is needed for fast intersections)

Constructor Detail

EigenDecomposition

public EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix2D u,
                          bpiwowar.maths.matrix.DiagonalDoubleMatrix s,
                          it.unimi.dsi.fastutil.ints.Int2IntSortedMap map)

EigenDecomposition

public EigenDecomposition(bpiwowar.maths.svd.IncrementalSVD svd,
                          it.unimi.dsi.fastutil.ints.Int2IntSortedMap map)

EigenDecomposition

public EigenDecomposition(DensityTracker tracker,
                          boolean deepCopy)

Construct a new eigen decomposition subspace/density based on a density tracker

Parameters:

tracker - The tracker from which to build this density

deepCopy - If all the structures underlying structures should be cloned (this is only for efficiency issues, note that at least the diagonal matrix will be cloned)

EigenDecomposition

public EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix2D mU,
                          bpiwowar.maths.matrix.DiagonalDoubleMatrix mS,
                          it.unimi.dsi.fastutil.ints.Int2IntSortedMap map,
                          boolean deepCopy)

EigenDecomposition

public EigenDecomposition(bpiwowar.maths.matrix.DoubleMatrix1D ray,
                          boolean local)

Creates a one dimensional eigen-decomposition representation

Parameters:

ray - The vector

Method Detail

normalise

public double normalise()

Normalise with the L2 norm

getS

public bpiwowar.maths.matrix.DiagonalDoubleMatrix getS()

normalise

public double normalise(boolean orthonormalU)

Normalise the decomposition so that || U x S || = 1, i.e. tr(U x S) = 1

Parameters:

orthonormalU - Should we expect U to be orthonormal (which should be the case normally)?

Returns:

The norm of the matrix

normalise

public double normalise(boolean orthonormalU,
                        cern.colt.function.DoubleFunction g,
                        cern.colt.function.DoubleDoubleFunction aggr,
                        cern.colt.function.DoubleFunction f)

Normalise the decomposition by a given norm normaliser = g(aggr_{i=1..n} f(x_i) )

Parameters:

orthonormalU - Should we expect U to be orthonormal (which should be the case normally)?

Returns:

The norm of the matrix

trim

public void trim(int newRank)

Trim the eigenvalue decomposition to a lower rank

Parameters:

newRank - The new rank of the subspace

getRank

public int getRank()

getLocalToGobalTermIdMap

public int[] getLocalToGobalTermIdMap()

Return the reverse map

Parameters:

tracker -

Returns:

getMatrix

public bpiwowar.maths.matrix.DoubleMatrix2D getMatrix()

Get a matrix representation of this density (mostly to be used for debugging)

Returns:

A matrix

getGlobalMatrix

public bpiwowar.maths.matrix.DoubleMatrix2D getGlobalMatrix(int dimension)

Get a matrix representation of this density (mostly to be used for debugging)

Returns:

A matrix