How to Control Density in Sparse Binary Encodings

Sparse binary encodings (BSDC family) work best when the fraction of 1-bits (density) stays well below 0.5. The sections below show how to measure density and keep it bounded during bundling.

Measuring density

Density is simply the mean of a binary hypervector’s data array:

import pyhdc

enc = pyhdc.BSDC_S(dimension=10_000)
hv  = enc.generate()
print(f"density = {hv.data.mean():.4f}")   # ~= 0.01

The density growth problem

BSDC uses bitwise OR for bundling. OR can only turn bits on, so repeated bundling drives density toward 1.0:

enc    = pyhdc.BSDC_S(dimension=10_000)
result = enc.generate()
print(f"step  0: density = {result.data.mean():.4f}")

for i in range(1, 21):
    result = result.bundle(enc.generate())
    if i % 5 == 0:
        print(f"step {i:2d}: density = {result.data.mean():.4f}")

# density increases with each step

Solving it with BSDC_THIN

BSDC_THIN applies random thinning after each OR step; bits are randomly cleared until the density returns to the initial level:

enc    = pyhdc.BSDC_THIN(dimension=10_000)
result = enc.generate()
print(f"step  0: density = {result.data.mean():.4f}")

for i in range(1, 21):
    result = result.bundle(enc.generate())
    if i % 5 == 0:
        print(f"step {i:2d}: density = {result.data.mean():.4f}")

# density stays stable throughout

Using DisjunctionThinned directly

If you are building a custom pipeline with the components submodule, you can access the thinned OR operation directly:

from pyhdc.components.bundling import DisjunctionThinned
import numpy as np

a = enc.generate().data
b = enc.generate().data

result = DisjunctionThinned(a, b, density=0.05)

Density guidelines

Density	Dimension	Notes
≥ 0.1	Any	Dense binary; more like BSC; less sparse-binary advantage
0.01–0.05	≥ 1,000	Recommended BSDC range; good balance of capacity and stability
≤ 0.01	≥ 10,000	Very sparse; high capacity; need large D for enough 1s per vector

Lower density means greater orthogonality between random vectors and higher theoretical capacity, but too-low density at small dimensions means each vector has very few 1s, making similarity estimates noisy.