Encoding Classes

All 15 encoding classes inherit from Encoding and share the same constructor signature and method set. See Encoding Base Class for the common interface.

This page documents the encoding-specific parameters and characteristics.

MAP family

Element type: float in [-1, 1], default dtype float32

Extra parameters (in addition to Encoding base parameters):

  • random_choice_range: float or None (rho): controls band randomization during bundling. Coordinates whose pre-aggregate |sum| falls within the band are replaced by independent Uniform[-1, 1] draws. Reduces systematic bias. Default: None (no band randomization).

Binding: ElementMultiplication (element-wise product, self-inverse)

Bundling: ElementAdditionCut (element-wise sum then clip to [-1, 1])

Similarity: CosineSimilarity → [-1, 1]

Unbind: Yes

Generator output type required: "floats"

Element type: int {-1, +1}, default dtype int32

Extra parameters: random_choice_range (same as MAP_C)

Binding: ElementMultiplication

Bundling: ElementAdditionCut

Similarity: CosineSimilarity

Unbind: Yes

Generator output type required: "bits"

Element type: bipolar {-1, +1}, saturating to a signed integer after bundling. Storage dtype auto-widens to the smallest signed type that holds the configured width (int8 / int16 / int32 / int64), default int32.

Extra parameters:

  • mask: int, default (2**32) - 1. Must have the form 2**k - 1 (contiguous low bits). Selects the signed saturation width k. A mask that is not 2**k - 1 raises ValueError (use bit_width instead). Example: mask=0xFF (= 2**8 - 1) gives 8-bit saturation in int8.

  • bit_width: int or None, default None. Explicit signed bit width k (1-64), overrides mask when set.

Binding: ElementMultiplication. Bipolar operands stay bipolar, but binding a bundled (saturated, non-bipolar) vector at a narrow width can overflow the storage dtype, since the product is not re-clipped. Bind before bundling, or use a wider width, if you need to bind sums.

Bundling: ElementAdditionBits (element-wise sum in a wide accumulator, then a single saturating clip to the signed k-bit range)

Similarity: CosineSimilarity

Unbind: Yes

Generator output type required: "words"

Element type: bipolar {-1, +1}, default dtype int8

Extra parameters: random_choice_range

Binding: ElementMultiplication

Bundling: ElementAdditionBipolarThreshold (element-wise sum, then sign to {-1, +1})

Similarity: CosineSimilarity

Unbind: Yes

Generator output type required: "bits"

HRR family

Element type: normal float, default dtype float32

Extra parameters: random_choice_range

Binding: CircularConvolution (via FFT)

Bundling: ElementAdditionNormalized (sum then L2 normalise)

Similarity: CosineSimilarity

Unbind: Yes (via CircularCorrelation)

Generator output type required: "floats"

Like HRR but bundling does not normalise the result. Vector magnitude grows with the number of bundled items.

Like HRR but bundling divides by \(\sqrt{M}\) rather than the L2 norm. Maintains constant expected magnitude.

Element type: angles in [-π, π), default dtype float32

Extra parameters: random_choice_range

Binding: ElementAngleAddition (modular angle addition)

Bundling: AnglesOfElementAddition (phasor resultant angle)

Similarity: AngleDistance → [-1, 1]

Unbind: Yes (ElementAngleSubtraction)

Generator output type required: "floats"

Matrix family

Element type: normal float, default dtype float32

Binding: VectorDerivedTransformation (matrix derived from the key vector)

Bundling: ElementAdditionNormalized

Similarity: CosineSimilarity

Unbind: Yes (transpose of the transformation matrix)

Generator output type required: "floats"

Element type: normal float, default dtype float32

Binding: MatrixMultiplication (random matrix)

Bundling: ElementAdditionNormalized

Similarity: CosineSimilarity

Unbind: Yes; requires get_metadata()["matrices"] from the bound result.

Note

MBAT stores the random binding matrices in the result’s metadata. You must preserve bound_hv.get_metadata() to perform unbinding later.

Generator output type required: "floats"

Binary family

Element type: binary {0, 1}, default dtype int8

Extra parameters: random_choice_range

Binding: ExclusiveOr (XOR; exactly self-inverse)

Bundling: ElementAdditionBinaryThreshold (majority vote)

Similarity: HammingDistance → [-1, 1]

Unbind: Yes (exact, not approximate)

Generator output type required: "bits"

Sparse binary family (BSDC)

All BSDC variants use sparse binary vectors (density ≈ 1–5%), bitwise OR for bundling, and Overlap similarity.

Binding: AdditiveContextDependentThinning

Bundling: Disjunction (OR; density grows without thinning)

Similarity: Overlap → [-1, 1]

Unbind: No; unbind() raises NotImplementedError

Binding: Shifting (circular shift by one position per bind step)

Bundling: Disjunction (OR; density grows)

Similarity: Overlap → [-1, 1]

Unbind: Yes (inverse circular shift)

Like BSDC_S but the circular shift is applied independently to each segment of the vector.

Added in v1.1.0

Binding: Shifting (same as BSDC_S)

Bundling: DisjunctionThinned (OR then random thinning to maintain density)

Similarity: Overlap → [-1, 1]

Unbind: Yes (inverse circular shift)

The recommended sparse binary default when you need stable density over many bundle operations.