symmetry-validation-suite（对称性验证套件）

Symmetry Validation Suite

Wrong symmetry assumptions hurt model performance -- too much symmetry over-constrains, while missing symmetry wastes capacity. Validate before committing to equivariant architecture.

Workflow

Copy this checklist and track your progress:

Symmetry Validation Progress:
- [ ] Step 1: List symmetry hypotheses to test
- [ ] Step 2: Design transformation test sets
- [ ] Step 3: Run invariance/equivariance tests
- [ ] Step 4: Verify group structure
- [ ] Step 5: Analyze data distribution under transforms
- [ ] Step 6: Document validation results

Step 1: List symmetry hypotheses to test

Gather candidate symmetries from previous discovery work. For each, document: the transformation type, whether invariance or equivariance is expected, and confidence level. Prioritize testing low-confidence hypotheses. If no hypotheses exist, work with user through domain analysis to identify candidate symmetries first.

Step 2: Design transformation test sets

For each symmetry, create test protocol: Sample representative inputs from data distribution. Define transformation sampling strategy (random rotations, all permutations, etc.). Determine appropriate sample sizes for statistical significance. Consider edge cases and boundary conditions. See Transformation Sampling for guidance. For detailed methodology, consult Methodology Details.

Step 3: Run invariance/equivariance tests

For invariance testing: Apply transformation T to input x, compute outputs f(x) and f(T(x)), measure error ||f(T(x)) - f(x)||. For equivariance testing: Compute f(T(x)) and T'(f(x)) where T' is the output transformation, measure error ||f(T(x)) - T'(f(x))||. Use Testing Protocols for implementation details. Aggregate across samples and compute statistics. For complete code examples, see Test Implementation Examples.

Step 4: Verify group structure

Check that claimed transformations form a valid group: Test closure (composition of two transforms is a transform). Test associativity. Verify identity element exists. Verify inverses exist. For Lie groups, check that generators close under commutator. See Group Structure Tests.

Step 5: Analyze data distribution under transforms

Check if transformed data stays in-distribution: Apply transforms to training data. Compare statistics of original vs transformed data. Check for distributional shift that might break assumptions. Identify transformation ranges that maintain validity. This catches "approximate symmetry" cases where symmetry holds only within bounds.

Step 6: Document validation results

Create validation report using Output Template. For each symmetry: state hypothesis, test methodology, quantitative results, pass/fail decision. Recommend whether to use hard equivariance constraint, soft constraint (regularization), data augmentation, or no symmetry at all. Quality criteria for this output are defined in Quality Rubric.

Testing Protocols

Invariance Test Protocol

def test_invariance(model, data_samples, transform_fn, n_transforms=100):
    """
    Test if model output is invariant to transformations.

    Returns:
        mean_error: Average ||f(T(x)) - f(x)||
        max_error: Maximum error observed
        pass_rate: Fraction with error < threshold
    """
    errors = []
    for x in data_samples:
        y_orig = model(x)
        for _ in range(n_transforms):
            x_transformed = transform_fn(x)
            y_transformed = model(x_transformed)
            error = norm(y_transformed - y_orig)
            errors.append(error)

    return {
        'mean_error': mean(errors),
        'max_error': max(errors),
        'std_error': std(errors),
        'pass_rate': sum(e < threshold for e in errors) / len(errors)
    }

Equivariance Test Protocol

def test_equivariance(model, data_samples, input_transform, output_transform):
    """
    Test if f(T(x)) = T'(f(x)) for equivariance.

    Returns:
        mean_error: Average ||f(T(x)) - T'(f(x))||
        relative_error: Error normalized by output magnitude
    """
    errors = []
    for x in data_samples:
        # Method 1: Transform then model
        x_T = input_transform(x)
        y1 = model(x_T)

        # Method 2: Model then transform
        y = model(x)
        y2 = output_transform(y)

        error = norm(y1 - y2)
        relative = error / (norm(y2) + eps)
        errors.append({'absolute': error, 'relative': relative})

    return aggregate_stats(errors)

Statistical Significance

For reliable results:

• Use at least 100 data samples
• Test at least 50 random transformations per sample
• Report mean, std, and percentiles (95th, 99th)
• Set threshold based on numerical precision expectations
• Use hypothesis testing if comparing methods

Transformation Sampling

Continuous Groups

Discrete Groups

Group Structure Tests

Closure Test

For random g₁, g₂ ∈ G:
  Compute g₃ = g₁ · g₂
  Verify g₃ ∈ G (within numerical tolerance)

Associativity Test

For random g₁, g₂, g₃ ∈ G:
  Compute (g₁ · g₂) · g₃
  Compute g₁ · (g₂ · g₃)
  Verify equality (within tolerance)

Identity and Inverse Test

For random g ∈ G:
  Verify g · e = e · g = g
  Find g⁻¹ and verify g · g⁻¹ = e

Interpretation Guide

Error Thresholds

Decision Matrix

Output Template

SYMMETRY VALIDATION REPORT
==========================

Tested Symmetries:

1. [Transformation]: [Invariance/Equivariance]
   - Sample size: [N samples × M transforms]
   - Mean error: [value]
   - Max error: [value]
   - Pass rate: [%] at threshold [value]
   - RESULT: [PASS/FAIL/PARTIAL]
   - Recommendation: [Hard constraint/Soft/Augmentation/None]

2. [Transformation]: [Invariance/Equivariance]
   ...

Group Structure:
- Closure: [PASS/FAIL]
- Associativity: [PASS/FAIL]
- Identity/Inverse: [PASS/FAIL]

Distribution Analysis:
- Transform range where symmetry holds: [bounds]
- Detected breaking factors: [list]

SUMMARY:
- Confirmed symmetries: [list]
- Rejected symmetries: [list]
- Proceed to architecture design with: [group specification]