/// <summary>
/// Solves the linear system A * x = b by solving the preconditioned system inv(P) * A * inv(P)^T * y = inv(P) * b,
/// where A = <paramref name="matrix"/>, b = <paramref name="rhsVector"/>, x is the solution, y = P^T * x,
/// P*P^T = <paramref name="preconditioner"/>.
/// Initially x = <paramref name="initialGuess"/> and then it converges to the solution.
/// </summary>
/// <param name="matrix">
/// Represents the matrix A of the linear system A * x = b, which must be symmetric positive definite.
/// </param>
/// <param name="rhs">
/// The right hand side vector b of the linear system A * x = b. Constraints:
/// <paramref name="rhs"/>.<see cref="IIndexable1D.Length"/>
/// == <paramref name="matrix"/>.<see cref="IIndexable2D.NumRows"/>.
/// </param>
/// <param name="preconditioner">
/// A preconditioner matrix that is also symmetric positive definite and has the same dimensions as A.
/// </param>
/// <param name="solution">
/// The vector from which to start refining the solution vector x. Constraints:
/// <paramref name="solution"/>.<see cref="IIndexable1D.Length"/>
/// == <paramref name="matrix"/>.<see cref="IIndexable2D.NumColumns"/>.
/// </param>
/// <param name="initialGuessIsZero">
/// If <paramref name="solution"/> is 0, then set <paramref name="initialGuessIsZero"/> to true to avoid performing the
/// operation b-A*0 before starting.
/// </param>
/// <exception cref="NonMatchingDimensionsException">
/// Thrown if <paramref name="rhs"/> or <paramref name="solution"/> violate the described constraints.
/// </exception>
public IterativeStatistics Solve(ILinearTransformation matrix, IPreconditioner preconditioner, IVectorView rhs,
IVector solution, bool initialGuessIsZero, Func <IVector> zeroVectorInitializer)
{
//TODO: find a better way to handle optimizations for the case x0=0, than using an initialGuessIsZero flag
Preconditions.CheckMultiplicationDimensions(matrix.NumColumns, solution.Length);
Preconditions.CheckSystemSolutionDimensions(matrix.NumRows, rhs.Length);
this.Matrix = matrix;
this.Preconditioner = preconditioner;
this.Rhs = rhs;
this.solution = solution;
// r = b - A * x
if (initialGuessIsZero)
{
residual = rhs.Copy();
}
else
{
residual = ExactResidual.Calculate(matrix, rhs, solution);
}
return(SolveInternal(maxIterationsProvider.GetMaxIterations(matrix.NumColumns), zeroVectorInitializer));
}
internal IterativeStatistics Solve(IInterfaceFlexibilityMatrix matrix, IFetiPreconditioner preconditioner,
IInterfaceProjection projector, Vector rhs, Vector lagrangeMultipliers)
{
int n = matrix.Order;
int maxIterations = maxIterationsProvider.GetMaxIterations(matrix.Order);
// r0 = d - F * λ0
var residual = matrix.Multiply(lagrangeMultipliers);
residual.LinearCombinationIntoThis(-1.0, rhs, 1.0);
// Other allocations
var w = Vector.CreateZero(n);
var y = Vector.CreateZero(n);
var z = Vector.CreateZero(n);
var direction = Vector.CreateZero(n);
var matrixTimesDirection = Vector.CreateZero(n);
double residualDotProductPrevious = double.NaN;
double residualNormRatio = double.NaN;
double beta = double.NaN;
for (int iter = 1; iter <= maxIterations; ++iter)
{
// w(m-1) = P * r(m-1)
projector.ProjectVector(residual, w, false);
// z(m-1) = preconditioner * w(m-1)
preconditioner.SolveLinearSystem(w, z);
// Check convergence: usually if ||z|| / ||f|| < tolerance
residualNormRatio = convergence.EstimateResidualNormRatio(lagrangeMultipliers, z);
if (residualNormRatio <= residualTolerance)
{
//TODO: is it correct to check for convergence here? How many iterations should I return?
return(new IterativeStatistics
{
AlgorithmName = name,
HasConverged = true,
NumIterationsRequired = iter - 1, //TODO: not sure about this.
ResidualNormRatioEstimation = residualNormRatio
});
}
// y(m-1) = P * z(m-1)
projector.ProjectVector(z, y, false);
double residualDotProductCurrent = y.DotProduct(w);
if (iter == 1)
{
// β(1) = 0
beta = 0;
// p(1) = y0
direction.CopyFrom(y);
}
else
{
// β(m) = (y(m-1) * w(m-1)) / (y(m-2) * w(m-2))
beta = residualDotProductCurrent / residualDotProductPrevious;
// p(m) = y(m-1) + β(m) * p(m-1), if m > 1
direction.LinearCombinationIntoThis(beta, y, 1.0);
}
residualDotProductPrevious = residualDotProductCurrent;
// γ(m) = (y(m-1) * w(m-1)) / (p(m) * F * p(m))
matrix.Multiply(direction, matrixTimesDirection);
double stepSize = (y * w) / (direction * matrixTimesDirection);
// λ(m) = λ(m-1) + γ(m) * p(m)
lagrangeMultipliers.AxpyIntoThis(direction, stepSize);
// r(m) = r(m-1) -γ(m) * F * p(m)
residual.AxpyIntoThis(matrixTimesDirection, -stepSize);
}
return(new IterativeStatistics
{
AlgorithmName = name,
HasConverged = false,
NumIterationsRequired = maxIterations,
ResidualNormRatioEstimation = residualNormRatio
});
}
