Equivariance Tolerance Matrix¶
Specification matrix — instantiates Symmetry-Commuting Transformation Design
Tabulates, per transformation, the required exactness class and numeric tolerance so each symmetry gets a declared standard rather than an implicit one.
Not every symmetry demands the same fidelity: some transformations must commute exactly, others only within a numeric band, others only in a statistical sense. An Equivariance Tolerance Matrix is the planning artifact that pins this down before anyone tests anything — a grid whose rows are the transformations in scope and whose cells declare, for each, the required exactness class (exact, approximate-within-ε, statistical, or procedural) and the tolerance that goes with it. It is a declaration, not a measurement: it runs no inputs and renders no verdict. Its defining role is to set the standards that a test suite or probe will later be judged against, so "how close is close enough" is a written specification rather than a number someone picks at test-writing time.
Example¶
A team building an audio feature-extraction library needs to say, transformation by transformation, how tightly its analysis (F) must respect each one. They lay out the matrix. A time-shift of the waveform must move detected note onsets by exactly the same offset — exact equivariance, zero tolerance. A pitch-shift should move estimated pitch correspondingly, but interpolation makes perfect agreement impossible, so the cell reads approximate, within a few cents. A channel swap on stereo input must swap the per-channel outputs exactly. Added background noise cannot preserve features exactly, so its cell reads statistical: 95% of samples within a bounded feature distance. A format re-encode (lossless to lossy) is declared procedural — signed off by review rather than checked numerically, because bit-comparison is meaningless across codecs.
The matrix does nothing to the audio. But it converts a pile of implicit assumptions into a single reviewable table: everyone can see that pitch-shift is held to a looser standard than time-shift on purpose, and the downstream test suite now has an authoritative source for each transformation's pass criterion instead of inventing one per test.
How it works¶
- Enumerate the transformation rows. List every transformation the map is meant to respect — the scope — as the matrix's row axis.
- Assign an exactness class per row. Mark each as exact, approximate, statistical, or procedural, based on what the domain actually permits.
- Set the tolerance in each cell. Attach the concrete band — an ε, a confidence level, a sign-off requirement — that operationalizes the class.
- Publish as the coverage-and-standard grid. The completed matrix is the single reference the executing mechanisms read tolerances from; a blank cell is an admission that a transformation's standard is undecided.
Tuning parameters¶
- Exactness granularity — how many classes the matrix distinguishes (a simple exact/approximate split vs a four-way scheme with a statistical tier). More tiers fit reality better but make the matrix heavier to maintain.
- Tolerance stringency — how tight each band is set. Tighter bands demand more of the map and flag more failures; looser bands risk waving through real drift.
- Row resolution — whether a transformation family is one row or split into regimes (small vs large pitch-shifts). Splitting captures regime-dependent standards but grows the grid.
- Revision cadence — how often the declared tolerances are revisited as the map or requirements change, so the standard doesn't ossify while the system moves.
When it helps, and when it misleads¶
Its strength is making the standard an explicit, reviewable object rather than an accident of whoever wrote the first test. It forces the team to decide, in the open, which symmetries are exact and which are merely approximate — a decision that otherwise gets made implicitly and inconsistently across a test suite. It is the equivariance analog of a numerical tolerance specification, the rtol/atol pair that turns "approximately equal" into a defined acceptance criterion.[n1]
Its failure mode is that a matrix is only as good as the judgment behind its cells. A tolerance set too loose blesses a map that is drifting; set arbitrarily, it lends a spurious air of rigor to numbers nobody justified. And because it measures nothing, a beautifully filled matrix can coexist with a map that violates every cell — the matrix states the target, it does not hit it. The classic misuse is treating a completed matrix as evidence of compliance rather than as the criterion compliance will be judged against. The guarding discipline is to justify each tolerance from domain requirements, and to keep the matrix strictly upstream — it names the bar; a separate executing mechanism must clear it.
How it implements the components¶
Equivariance Tolerance Matrix fills the standard-declaration slice of the archetype:
transformation_coverage_matrix— the grid itself is the coverage matrix: every in-scope transformation gets a row, and an empty row is a visible gap in the plan.approximation_tolerance_band— each cell's declared exactness class and numeric band is the tolerance the downstream checkers must satisfy.transformation_scope— the row enumeration fixes which transformations are in scope to hold a standard at all.
It executes nothing: the equivariance_test_oracle that actually runs inputs and returns pass/fail against these tolerances is Transformation-Pair Test Suite; this matrix declares the bar, the suite clears it.
Related¶
- Instantiates: Symmetry-Commuting Transformation Design — this matrix supplies the exactness standards the pattern's tests are judged by.
- Sibling mechanisms: Commutative Diagram Review · Transformation-Pair Test Suite · Coordinate-Frame Consistency Check · Permutation Equivariance Audit · Data-Augmentation Equivariance Probe · Schema and Label Relabeling Harness · Symmetry Exception Register
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Equivariance Tolerance Matrix operates as a non-executable information artifact that externalizes static or prospective structure because it tabulates, per transformation, the required exactness class and numeric tolerance so each symmetry gets a declared standard rather than an implicit one.
Independent corroboration: The frozen evidence defines Equivariance Tolerance Matrix as 'Tabulates, per transformation, the required exactness class and numeric tolerance so each symmetry gets a declared standard rather than an implicit one', so its operative form is Representation, Specification & Plan.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Mathematics
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Group theory and geometry supply equivariance as preservation of structure under declared transformations.
Related originating lineages:
- Computer Science & Software Engineering — Machine-learning evaluation supplies transformation-specific tests of approximate equivariance.
- Engineering & Design — Tolerance engineering supplies explicit numeric allowances instead of an implicit exactness claim.
Review resolution: The current reviewers agree that mathematics is primary. For the reported differences (reported_ambiguity, alternate_origin_disagreement, encyclopedia_synthesis_disagreement), the evidence supports cross_disciplinary_synthesis, specialized, and computer_science, engineering_design; these choices preserve materially formative origins without conflating later domain reach.
Attribution caveat: The matrix is a synthesized verification artifact around a formal symmetry concept.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Numerical libraries define "approximately equal" through an explicit relative and absolute tolerance (e.g. rtol/atol in a floating-point comparison), rather than trusting bit-equality. The tolerance matrix generalizes that idea to equivariance: each transformation gets its own declared acceptance band instead of an unstated one. ↩