zerofold 0.1.4

Deterministic compute cache for SVD and PCA — lossless, 100-300x faster on repeated calls (same matrix)

ZeroFold

100–300× faster on repeated calls (same matrix) — lossless, deterministic, zero bit difference.

pip install zerofold

from zerofold import svd, pca

# First call: computes at standard speed (NumPy/SciPy)
result = svd(weight_matrix, n_components=64)

# Every subsequent call: O(1) retrieval — bitwise identical output
result = svd(weight_matrix, n_components=64)  # microseconds, not seconds

---

What this is

A deterministic compute cache for expensive linear algebra operations.

Call	Cost	Output
First	Standard NumPy/SciPy speed	Exact result, stored
Subsequent	O(1) retrieval	Bitwise identical to first call

No approximation. No tolerance. Zero bit difference between calls.

Important: Speedups occur only when the same matrix is reused. First-time computations run at standard speed. If every matrix you compute on is unique, this tool is not for you.

---

When this is useful

Use case	Why it helps
Neural network inference	Same weight matrices queried every batch → 99%+ hit rate
Repeated analytics pipelines	Same dataset processed repeatedly
Scientific computing	Same Laplacian/Hamiltonian, different parameters
Feature engineering / PCA reuse	Common in production ML pipelines

When this has zero value

• One-off computations on unique matrices
• Streaming data where every matrix is different
• Workloads with no matrix reuse

---

Benchmark (SEED=42 — run it yourself, get the same correctness results)

python -X utf8 benchmark.py

Test 1 — Same matrix repeated calls: first call vs retrieval

n	First call	Retrieval	Speedup
128	~10 ms	~120 µs	~80×
512	~280 ms	~1.6 ms	~175×
1024	~1.6 s	~6.5 ms	~245×
2048	~5.9 s	~22 ms	~270×

Timing varies by hardware. Correctness results are identical on every machine.

Test 3 — Neural network weights (fixed per batch)

Metric	Result
Weight matrix hit rate	99.5%
Bit difference on retrieval	0.00e+00

Test 4 — Lossless verification

[PASS] n= 64  S_diff=0.00e+00  Vt_diff=0.00e+00  U_diff=0.00e+00
[PASS] n=128  S_diff=0.00e+00  Vt_diff=0.00e+00  U_diff=0.00e+00
[PASS] n=256  S_diff=0.00e+00  Vt_diff=0.00e+00  U_diff=0.00e+00
[PASS] n=512  S_diff=0.00e+00  Vt_diff=0.00e+00  U_diff=0.00e+00
5/5 PASS — all diffs exactly 0

---

How it works

Role classification routes first-time computation to the fastest correct algorithm, then stores the result indexed by the matrix's structural signature:

Role	Matrix type	First-call algorithm
Completion	Near-identity	Diagonal shortcut — O(n), exact
Prime	Symmetric	scipy.eigh — faster for symmetric, exact
Composite	General	numpy.linalg.svd — full precision

After the first call, every subsequent call is O(1) retrieval regardless of role. The returned result is the stored value — not recomputed, not approximated.

---

API

from zerofold import svd, pca, ZeroSubstrate

# Drop-in functions (global shared substrate)
r = svd(X, n_components=64)
r.U             # (m, k) left singular vectors
r.S             # (k,)   singular values
r.Vt            # (k, n) right singular vectors
r.from_receipt  # True if returned from cache
r.algorithm     # "receipt" | "completion_exact" | "prime_exact" | "composite_exact"

r = pca(X, n_components=50)
r.components            # (k, n_features)
r.explained_var_ratio   # (k,)
r.transform(X_new)      # project new data
r.inverse_transform(Z)  # reconstruct

# Explicit substrate (isolated cache, useful for namespacing)
substrate = ZeroSubstrate(max_receipts=10_000)
r = substrate.svd(X, n_components=64)
print(substrate.stats())
# {'hits': 8, 'misses': 2, 'hit_rate': 0.8, 'receipts_stored': 2}

substrate.clear()  # evict all cached results

# Disk persistence — survives restarts, shared across workers
substrate = ZeroSubstrate(cache_dir="/tmp/zerofold_cache")
r = substrate.svd(X, n_components=64)  # computed + saved to disk
# restart process, new worker, same cache_dir:
substrate2 = ZeroSubstrate(cache_dir="/tmp/zerofold_cache")
r2 = substrate2.svd(X, n_components=64)  # loaded from disk, from_receipt=True

---

Real-world value

If your ML inference pipeline recomputes SVD on the same weight matrices:

• n=512 weight matrix → ~280ms → ~1.6ms after first call
• 1000 batches/day → saves ~278 seconds/day per matrix
• At scale: the savings compound across every layer, every model, every deployment

"We reduced inference cost by 30–70% on fixed-weight workloads." That is where the acquisition conversations start.

---

License

Business Source License 1.1. Free for individuals, researchers, and startups under $1M revenue. Converts to Apache 2.0 on 2027-01-01. Commercial license available — contact [your email].