public static void assertEigen(
Matrix eigens,
VectorIterable corpus,
int numEigensToCheck,
double errorMargin,
boolean isSymmetric) {
for (int i = 0; i < numEigensToCheck; i++) {
Vector e = eigens.getRow(i);
if (e.getLengthSquared() == 0) {
continue;
}
Vector afterMultiply = isSymmetric ? corpus.times(e) : corpus.timesSquared(e);
double dot = afterMultiply.dot(e);
double afterNorm = afterMultiply.getLengthSquared();
double error = 1 - dot / Math.sqrt(afterNorm * e.getLengthSquared());
assertTrue(
"Error margin: {" + error + " too high! (for eigen " + i + ')',
Math.abs(error) < errorMargin);
}
}
/**
* For the distributed case, the best guess at a useful initialization state for Lanczos we'll
* chose to be uniform over all input dimensions, L_2 normalized.
*/
public static Vector getInitialVector(VectorIterable corpus) {
Vector initialVector = new DenseVector(corpus.numCols());
initialVector.assign(1.0 / Math.sqrt(corpus.numCols()));
return initialVector;
}
/**
* Solves the system Ax = b, where A is a linear operator and b is a vector. Uses the specified
* preconditioner to improve numeric stability and possibly speed convergence. This version of
* solve() allows control over the termination and iteration parameters.
*
* @param a The matrix A.
* @param b The vector b.
* @param preconditioner The preconditioner to apply.
* @param maxIterations The maximum number of iterations to run.
* @param maxError The maximum amount of residual error to tolerate. The algorithm will run until
* the residual falls below this value or until maxIterations are completed.
* @return The result x of solving the system.
* @throws IllegalArgumentException if the matrix is not square, if the size of b is not equal to
* the number of columns of A, if maxError is less than zero, or if maxIterations is not
* positive.
*/
public Vector solve(
VectorIterable a,
Vector b,
Preconditioner preconditioner,
int maxIterations,
double maxError) {
if (a.numRows() != a.numCols()) {
throw new IllegalArgumentException("Matrix must be square, symmetric and positive definite.");
}
if (a.numCols() != b.size()) {
throw new CardinalityException(a.numCols(), b.size());
}
if (maxIterations <= 0) {
throw new IllegalArgumentException("Max iterations must be positive.");
}
if (maxError < 0.0) {
throw new IllegalArgumentException("Max error must be non-negative.");
}
Vector x = new DenseVector(b.size());
iterations = 0;
Vector residual = b.minus(a.times(x));
residualNormSquared = residual.dot(residual);
log.info("Conjugate gradient initial residual norm = {}", Math.sqrt(residualNormSquared));
double previousConditionedNormSqr = 0.0;
Vector updateDirection = null;
while (Math.sqrt(residualNormSquared) > maxError && iterations < maxIterations) {
Vector conditionedResidual;
double conditionedNormSqr;
if (preconditioner == null) {
conditionedResidual = residual;
conditionedNormSqr = residualNormSquared;
} else {
conditionedResidual = preconditioner.precondition(residual);
conditionedNormSqr = residual.dot(conditionedResidual);
}
++iterations;
if (iterations == 1) {
updateDirection = new DenseVector(conditionedResidual);
} else {
double beta = conditionedNormSqr / previousConditionedNormSqr;
// updateDirection = residual + beta * updateDirection
updateDirection.assign(Functions.MULT, beta);
updateDirection.assign(conditionedResidual, Functions.PLUS);
}
Vector aTimesUpdate = a.times(updateDirection);
double alpha = conditionedNormSqr / updateDirection.dot(aTimesUpdate);
// x = x + alpha * updateDirection
PLUS_MULT.setMultiplicator(alpha);
x.assign(updateDirection, PLUS_MULT);
// residual = residual - alpha * A * updateDirection
PLUS_MULT.setMultiplicator(-alpha);
residual.assign(aTimesUpdate, PLUS_MULT);
previousConditionedNormSqr = conditionedNormSqr;
residualNormSquared = residual.dot(residual);
log.info(
"Conjugate gradient iteration {} residual norm = {}",
iterations,
Math.sqrt(residualNormSquared));
}
return x;
}
