Lodestar.Decomposition 0.2.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

TruncatedSvdOptions

What TruncatedSvd.Fit is allowed to vary. Every property has an initialiser, so new TruncatedSvdOptions() is scikit-learn's own default configuration and you set only what you are changing.

Properties

Property	Default	What it does
Oversampling	10	Extra columns drawn beyond the rank asked for. More is more accurate and slower; scikit-learn's default is the same 10.
PowerIterations	5	How many times the probe block is pushed through A and Aᵀ. This is the knob that matters when the singular values decay slowly.
Normalizer	Auto	What happens to the block between the two products — see PowerIterationNormalizer.
Seed	0	Seeds this package's generator when RandomMatrix is null. It reproduces a run of Lodestar, never a run of NumPy.
RandomMatrix	null	Ω itself, row-major and FeatureCount × (componentCount + Oversampling). Given, it replaces the draw entirely.

Seed and RandomMatrix answer two different questions. Seed makes a run repeatable on your machine; the block it draws comes from a SplitMix64 generator this package owns, because System.Random changed algorithm in .NET 6 and this package ships to runtimes on both sides of that. RandomMatrix makes a run repeatable across implementations — hand it the block NumPy drew and the components come back equal to scikit-learn's entry by entry, which is how this package's conformance is proved. See ADR 0072.

Oversampling and PowerIterations are both refused when negative, and so is an Oversampling too large to add to the rank asked for — an int that wrapped would be a block width nobody asked for rather than an error anybody could read. Oversampling = 0 is allowed and means the probe block is exactly as wide as the rank, which is the fastest and the least accurate the method gets.