How to Choose the Right Encoding

PyHDC provides 15 encoding classes. The decision tree and notes below narrow the choice for most use cases.

Quick decision guide

Start here and follow the branches:

• Need GPU / PyTorch integration? → all encodings support both backends

• Need binary output (1-bit elements)?

  • Dense binary → BSC or MAP_B

  • Sparse binary (low density) → BSDC_THIN (best default) or BSDC_S / BSDC_SEG / BSDC_CDT

• Need complex / phase-based values? → FHRR

• Need matrix-style binding? → VTB or MBAT

• Otherwise (continuous float vectors)?

  • General purpose → MAP_C (best default)

  • Normalized bundling theory → HRR

  • Fixed-width integers → MAP_I or MAP_I_Bits

Full comparison table

Encoding	Element type	Default dtype	Binding	Bundling	Similarity	Unbind	Notes
MAP_C	float [-1,1]	float32	ElementMultiply	Add + cut	Cosine	Yes	Best all-round default
MAP_I	int {-1,1}	int32	ElementMultiply	Add + cut	Cosine	Yes	Integer; needs bit generator
MAP_I_Bits	int (custom width)	int32	ElementMultiply	Add (clipped)	Cosine	Yes	mask sets bit width
MAP_B	binary {0,1}	int8	ElementMultiply	Add (clipped)	Cosine	Yes	Binary MAP
HRR	float (normal)	float32	CircularConv	Add + normalise	Cosine	Yes	Theoretically clean
HRR_NoNorm	float (normal)	float32	CircularConv	Add (no norm)	Cosine	Yes	Faster than HRR
HRR_ConstNorm	float (normal)	float32	CircularConv	Add / √M	Cosine	Yes	Constant-norm bundles
FHRR	angle [-π, π)	float32	AngleAdd	AngleAdd	AngleDist	Yes	Phase/periodic signals
VTB	float (normal)	float32	VDTransform	Add + normalise	Cosine	Yes	Matrix derived from key
MBAT	float (normal)	float32	MatrixMult	Add + normalise	Cosine	Yes (+ metadata)	Random matrix; save get_metadata()
BSC	binary {0,1}	int8	XOR	Majority vote	Hamming	Yes (exact)	Dense binary; XOR is exact inverse
BSDC_CDT	sparse {0,1}	int8	CDThinning	OR	Overlap	No	Context-dependent thinning
BSDC_S	sparse {0,1}	int8	CircShift	OR	Overlap	Yes	Shift-based; good for sequences
BSDC_SEG	sparse segmented	int8	SegShift	OR	Overlap	Yes	Per-segment shift
BSDC_THIN	sparse {0,1}	int8	CircShift	OR + thin	Overlap	Yes	Best sparse default (v1.1.0+)

Side-by-side example

The encoding API is identical across families; only the constructor call changes:

import pyhdc

for EncClass in [pyhdc.MAP_C, pyhdc.HRR, pyhdc.BSC, pyhdc.BSDC_THIN]:
    enc = EncClass(dimension=10_000)
    a   = enc.generate()
    b   = enc.generate()
    c   = a.bind(b)
    print(f"{EncClass.__name__:12s}  "
          f"sim(a,a)={a.similarity(a):.2f}  "
          f"sim(a,bind(a,b))={a.similarity(c):.2f}")

Note

Choosing an encoding picks the algebra (bind / bundle / similarity). To turn raw data (scalars, periodic values, or feature vectors) into hypervectors in that encoding, use a data encoder. See How to Encode Data into Hypervectors. A few encoders are restricted by family: Thermometer and Density need a discrete family (MAP_I/MAP_B/BSC/BSDC), Projection needs a family with a normalize step (MAP/HRR/VTB/MBAT/FHRR), and FractionalPower is defined only for the FHRR and the HRR families.