Vector.NormalizeLogWeights<T> Method

public static Vector<T> NormalizeLogWeights<T>(
	Vector<T> logWeights,
	out T logNormalizer
)

Public Shared Function NormalizeLogWeights(Of T) ( 
	logWeights As Vector(Of T),
	<OutAttribute> ByRef logNormalizer As T
) As Vector(Of T)

public:
generic<typename T>
static Vector<T>^ NormalizeLogWeights(
	Vector<T>^ logWeights, 
	[OutAttribute] T% logNormalizer
)

static member NormalizeLogWeights : 
        logWeights : Vector<'T> * 
        logNormalizer : 'T byref -> Vector<'T> 

Parameters

logWeights  Vector<T>

The unnormalized log weights in logical order.

logNormalizer  T

The logarithm of the sum of the exponentials of logWeights. A finite value indicates valid normalized weights.

Type Parameters

T

A supported real floating-point element type.

Return Value

When logNormalizer is finite, the returned weights are finite, nonnegative, and sum to one within floating-point roundoff. Negative-infinite scores receive zero weight. The calculation shares a shifted exponential sum between both outputs, avoiding direct exponentiation of large positive or negative scores.

Any NaN score gives a NaN normalizer and all-NaN weights. Otherwise, positive infinities give a positive-infinite normalizer and share the probability mass equally; other entries receive zero. All-negative-infinite scores give a negative-infinite normalizer and all-NaN weights. Empty input gives an empty vector and a negative-infinite normalizer. These cases can be detected afterwards by testing the normalizer for finiteness.

Every logical score is inspected, even when the input is missing-aware. NaNs are never skipped. Input storage is neither modified nor retained. In the Pro edition, the result retains the input's row index and name, using ordinary numeric missing-value semantics rather than inheriting the input's missing-value settings.

Reference