gam-sae 0.3.155

Sparse-autoencoder latent-manifold terms for the gam penalized-likelihood engine
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214

//! Cross-replicate circle-coordinate stability modulo the exact `O(2)` gauge.
//!
//! Reconstruction quality cannot certify a latent angle (#2260). This module
//! compares replicate coordinate vectors after quotienting only the circle's
//! rotation/reflection gauge. It also certifies that every replicate spans both
//! circle-embedding axes, so two collapsed coordinate vectors cannot receive a
//! vacuous perfect score merely because a rotation can align their constants.

use ndarray::ArrayView2;

#[derive(Clone, Debug, PartialEq)]
pub struct CircularReplicateCoverage {
    pub replicate: usize,
    pub minimum_embedding_eigenvalue: f64,
    pub maximum_embedding_eigenvalue: f64,
    pub isotropic_coverage: f64,
    pub rank_resolution: f64,
    pub well_posed: bool,
}

#[derive(Clone, Debug, PartialEq)]
pub struct CircularPairConcordance {
    pub left: usize,
    pub right: usize,
    pub rotation_score: Option<f64>,
    pub reflection_score: Option<f64>,
    pub aligned_score: Option<f64>,
    pub reflected: Option<bool>,
    pub phase_shift: Option<f64>,
}

#[derive(Clone, Debug, PartialEq)]
pub struct CircularConcordanceReport {
    pub n_replicates: usize,
    pub n_rows: usize,
    pub period: f64,
    pub coverage: Vec<CircularReplicateCoverage>,
    pub pairs: Vec<CircularPairConcordance>,
    pub minimum_aligned_score: Option<f64>,
    pub mean_aligned_score: Option<f64>,
}

fn embedding_coverage(cos: &[f64], sin: &[f64], replicate: usize) -> CircularReplicateCoverage {
    let n = cos.len();
    let mut g00 = 0.0;
    let mut g01 = 0.0;
    let mut g11 = 0.0;
    for row in 0..n {
        g00 += cos[row] * cos[row];
        g01 += cos[row] * sin[row];
        g11 += sin[row] * sin[row];
    }
    let trace = g00 + g11;
    let discriminant = ((g00 - g11).powi(2) + 4.0 * g01 * g01).sqrt();
    let minimum = (0.5 * (trace - discriminant)).max(0.0);
    let maximum = 0.5 * (trace + discriminant);
    // Backward-error resolution for a sum of `n` two-dimensional outer
    // products. This is derived from machine epsilon and problem size, not a
    // user-facing agreement threshold.
    let rank_resolution = f64::EPSILON * 2.0 * n as f64;
    CircularReplicateCoverage {
        replicate,
        minimum_embedding_eigenvalue: minimum,
        maximum_embedding_eigenvalue: maximum,
        isotropic_coverage: (2.0 * minimum / trace).clamp(0.0, 1.0),
        rank_resolution,
        well_posed: minimum > rank_resolution,
    }
}

/// Compare replicate circle coordinates modulo rotation/reflection.
///
/// `coordinates` is `(replicates, rows)` with corresponding rows referring to
/// the same observations in every replicate. `period` is the coordinate period.
/// For pair `(a, b)`, the rotation score is
/// `|mean exp(i(theta_a-theta_b))|`; the reflection score replaces the minus by
/// plus. Their maximum is the exact `O(2)`-aligned circular concordance. Scores
/// are withheld when either embedding is numerically rank deficient.
pub fn circular_concordance(
    coordinates: ArrayView2<'_, f64>,
    period: f64,
) -> Result<CircularConcordanceReport, String> {
    let (n_replicates, n_rows) = coordinates.dim();
    if n_replicates < 2 || n_rows < 2 {
        return Err(format!(
            "circular_concordance requires at least two replicates and two aligned rows; got {n_replicates}x{n_rows}"
        ));
    }
    if !(period.is_finite() && period > 0.0) {
        return Err(format!(
            "circular_concordance period must be finite and positive, got {period}"
        ));
    }
    if coordinates.iter().any(|value| !value.is_finite()) {
        return Err("circular_concordance coordinates must be finite".to_string());
    }

    let mut cos = vec![vec![0.0; n_rows]; n_replicates];
    let mut sin = vec![vec![0.0; n_rows]; n_replicates];
    for replicate in 0..n_replicates {
        for row in 0..n_rows {
            let phase =
                std::f64::consts::TAU * coordinates[[replicate, row]].rem_euclid(period) / period;
            let (sin_phase, cos_phase) = phase.sin_cos();
            sin[replicate][row] = sin_phase;
            cos[replicate][row] = cos_phase;
        }
    }

    let coverage = (0..n_replicates)
        .map(|replicate| embedding_coverage(&cos[replicate], &sin[replicate], replicate))
        .collect::<Vec<_>>();
    let mut pairs = Vec::with_capacity(n_replicates * (n_replicates - 1) / 2);
    let mut aggregate = Vec::with_capacity(pairs.capacity());
    for left in 0..n_replicates {
        for right in (left + 1)..n_replicates {
            if !(coverage[left].well_posed && coverage[right].well_posed) {
                pairs.push(CircularPairConcordance {
                    left,
                    right,
                    rotation_score: None,
                    reflection_score: None,
                    aligned_score: None,
                    reflected: None,
                    phase_shift: None,
                });
                continue;
            }
            let mut rotation_cos = 0.0;
            let mut rotation_sin = 0.0;
            let mut reflection_cos = 0.0;
            let mut reflection_sin = 0.0;
            for row in 0..n_rows {
                let ca = cos[left][row];
                let sa = sin[left][row];
                let cb = cos[right][row];
                let sb = sin[right][row];
                rotation_cos += ca * cb + sa * sb;
                rotation_sin += sa * cb - ca * sb;
                reflection_cos += ca * cb - sa * sb;
                reflection_sin += sa * cb + ca * sb;
            }
            let denominator = n_rows as f64;
            let rotation_score = rotation_cos.hypot(rotation_sin) / denominator;
            let reflection_score = reflection_cos.hypot(reflection_sin) / denominator;
            let reflected = reflection_score > rotation_score;
            let (aligned_score, shift_radians) = if reflected {
                (reflection_score, reflection_sin.atan2(reflection_cos))
            } else {
                (rotation_score, rotation_sin.atan2(rotation_cos))
            };
            let aligned_score = aligned_score.clamp(0.0, 1.0);
            aggregate.push(aligned_score);
            pairs.push(CircularPairConcordance {
                left,
                right,
                rotation_score: Some(rotation_score.clamp(0.0, 1.0)),
                reflection_score: Some(reflection_score.clamp(0.0, 1.0)),
                aligned_score: Some(aligned_score),
                reflected: Some(reflected),
                phase_shift: Some(
                    (shift_radians / std::f64::consts::TAU * period).rem_euclid(period),
                ),
            });
        }
    }

    let all_pairs_well_posed = aggregate.len() == pairs.len();
    let minimum_aligned_score = all_pairs_well_posed
        .then(|| aggregate.iter().copied().reduce(f64::min))
        .flatten();
    let mean_aligned_score =
        all_pairs_well_posed.then(|| aggregate.iter().sum::<f64>() / aggregate.len() as f64);
    Ok(CircularConcordanceReport {
        n_replicates,
        n_rows,
        period,
        coverage,
        pairs,
        minimum_aligned_score,
        mean_aligned_score,
    })
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::Array2;

    #[test]
    fn quotient_recovers_rotation_and_reflection_without_accepting_collapse() {
        let n = 37;
        let mut coordinates = Array2::<f64>::zeros((3, n));
        for row in 0..n {
            let base = ((row * row + 3 * row + 1) as f64 / n as f64).rem_euclid(1.0);
            coordinates[[0, row]] = base;
            coordinates[[1, row]] = (base + 0.23).rem_euclid(1.0);
            coordinates[[2, row]] = (0.41 - base).rem_euclid(1.0);
        }
        let report = circular_concordance(coordinates.view(), 1.0).expect("report");
        assert!(report.coverage.iter().all(|entry| entry.well_posed));
        let resolution = f64::EPSILON.sqrt();
        assert!(report.pairs[0].aligned_score.unwrap() >= 1.0 - resolution);
        assert_eq!(report.pairs[0].reflected, Some(false));
        assert!(report.pairs[1].aligned_score.unwrap() >= 1.0 - resolution);
        assert_eq!(report.pairs[1].reflected, Some(true));

        coordinates.row_mut(2).fill(0.0);
        let collapsed = circular_concordance(coordinates.view(), 1.0).expect("report");
        assert!(!collapsed.coverage[2].well_posed);
        assert_eq!(collapsed.pairs[1].aligned_score, None);
        assert_eq!(collapsed.mean_aligned_score, None);
    }
}