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
//! FALSIFY contract tests for GLM (Generalized Linear Models)
//!
//! Tests fundamental mathematical invariants of GLM predictions:
//! - Poisson predictions are non-negative (exp of linear predictor)
//! - Binomial predictions are in (0, 1) (logistic link)
//! - All predictions are finite (no NaN/Inf)
//! - Predictions are deterministic
// CONTRACT: glm-v1.yaml
// HASH: sha256:c7d8e9f0a1b23456
// Generated by: pv probar --binding
// DO NOT EDIT — regenerate with `pv probar --binding`
use aprender::glm::{Family, GLM};
use aprender::primitives::{Matrix, Vector};
use proptest::prelude::*;
/// Strategy: generate Poisson-suitable training data.
/// y values are non-negative (required by Poisson family validation).
/// Features in [-2, 2] to keep predictions in a reasonable range.
fn poisson_training_strategy() -> impl Strategy<Value = (Matrix<f32>, Vector<f32>)> {
(20usize..=30, 2usize..=3)
.prop_flat_map(|(n, d)| {
let features = proptest::collection::vec(-2.0f32..2.0f32, n * d);
// Poisson requires non-negative counts: abs(val) + 0.1 avoids zero
let targets = proptest::collection::vec(0.1f32..10.0f32, n);
let n_val = Just(n);
let d_val = Just(d);
(features, targets, n_val, d_val)
})
.prop_map(|(features, targets, n, d)| {
let x = Matrix::from_vec(n, d, features).expect("valid matrix dimensions");
let y = Vector::from_vec(targets);
(x, y)
})
}
/// Strategy: generate Binomial-suitable training data.
/// y values are in [0, 1] (required by Binomial family validation).
fn binomial_training_strategy() -> impl Strategy<Value = (Matrix<f32>, Vector<f32>)> {
(20usize..=30, 2usize..=3)
.prop_flat_map(|(n, d)| {
let features = proptest::collection::vec(-2.0f32..2.0f32, n * d);
// Binomial requires values in [0, 1]
let targets = proptest::collection::vec(0.01f32..0.99f32, n);
let n_val = Just(n);
let d_val = Just(d);
(features, targets, n_val, d_val)
})
.prop_map(|(features, targets, n, d)| {
let x = Matrix::from_vec(n, d, features).expect("valid matrix dimensions");
let y = Vector::from_vec(targets);
(x, y)
})
}
proptest! {
#![proptest_config(ProptestConfig::with_cases(128))]
// ──────────────────────────────────────────────────────────
// FALSIFY-GLM-001: Poisson GLM predictions are non-negative
// Formal: ∀ i, predict(X)[i] >= 0 (since μ = exp(η) > 0)
// ──────────────────────────────────────────────────────────
/// Obligation: Poisson GLM with log link produces non-negative predictions.
/// The inverse link is exp(η), which is strictly positive for all finite η.
#[test]
fn prop_poisson_predictions_non_negative(
(x, y) in poisson_training_strategy(),
) {
let mut model = GLM::new(Family::Poisson);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let predictions = model.predict(&x);
prop_assume!(predictions.is_ok());
let preds = predictions.expect("predict succeeded after assume");
for i in 0..preds.len() {
prop_assert!(
preds[i] >= 0.0,
"FALSIFY-GLM-001: Poisson prediction[{}] = {}, expected >= 0.0 (exp of linear predictor)",
i, preds[i]
);
}
}
// ──────────────────────────────────────────────────────────
// FALSIFY-GLM-002: Binomial GLM predictions in (0, 1)
// Formal: ∀ i, 0 < predict(X)[i] < 1 (logistic sigmoid range)
// ──────────────────────────────────────────────────────────
/// Obligation: Binomial GLM with logit link produces predictions in (0, 1).
/// The inverse link is σ(η) = 1/(1+exp(-η)), bounded in (0, 1).
#[test]
fn prop_binomial_predictions_in_unit_interval(
(x, y) in binomial_training_strategy(),
) {
let mut model = GLM::new(Family::Binomial);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let predictions = model.predict(&x);
prop_assume!(predictions.is_ok());
let preds = predictions.expect("predict succeeded after assume");
for i in 0..preds.len() {
prop_assert!(
preds[i] > 0.0 && preds[i] < 1.0,
"FALSIFY-GLM-002: Binomial prediction[{}] = {}, expected in (0, 1) (logistic link)",
i, preds[i]
);
}
}
// ──────────────────────────────────────────────────────────
// FALSIFY-GLM-003: Predictions are finite (no NaN/Inf)
// Formal: ∀ family ∈ {Poisson, Binomial}, ∀ i, predict(X)[i].is_finite()
// ──────────────────────────────────────────────────────────
/// Obligation: GLM predictions contain no NaN or Inf values.
/// IRLS with bounded features should always produce finite coefficients.
#[test]
fn prop_predictions_finite_poisson(
(x, y) in poisson_training_strategy(),
) {
let mut model = GLM::new(Family::Poisson);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let predictions = model.predict(&x);
prop_assume!(predictions.is_ok());
let preds = predictions.expect("predict succeeded after assume");
for i in 0..preds.len() {
prop_assert!(
preds[i].is_finite(),
"FALSIFY-GLM-003: Poisson prediction[{}] = {}, expected finite (no NaN/Inf)",
i, preds[i]
);
}
}
#[test]
fn prop_predictions_finite_binomial(
(x, y) in binomial_training_strategy(),
) {
let mut model = GLM::new(Family::Binomial);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let predictions = model.predict(&x);
prop_assume!(predictions.is_ok());
let preds = predictions.expect("predict succeeded after assume");
for i in 0..preds.len() {
prop_assert!(
preds[i].is_finite(),
"FALSIFY-GLM-003: Binomial prediction[{}] = {}, expected finite (no NaN/Inf)",
i, preds[i]
);
}
}
// ──────────────────────────────────────────────────────────
// FALSIFY-GLM-004: Predictions are deterministic
// Formal: predict(X) == predict(X) for same fitted model and input
// ──────────────────────────────────────────────────────────
/// Obligation: Calling predict twice on the same input produces identical results.
#[test]
fn prop_predictions_deterministic_poisson(
(x, y) in poisson_training_strategy(),
) {
let mut model = GLM::new(Family::Poisson);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let pred1 = model.predict(&x);
let pred2 = model.predict(&x);
prop_assume!(pred1.is_ok() && pred2.is_ok());
let p1 = pred1.expect("predict succeeded (1st)");
let p2 = pred2.expect("predict succeeded (2nd)");
for i in 0..p1.len() {
prop_assert!(
p1[i] == p2[i],
"FALSIFY-GLM-004: Poisson prediction[{}] differs: {} vs {} (must be deterministic)",
i, p1[i], p2[i]
);
}
}
#[test]
fn prop_predictions_deterministic_binomial(
(x, y) in binomial_training_strategy(),
) {
let mut model = GLM::new(Family::Binomial);
let fit_result = model.fit(&x, &y);
prop_assume!(fit_result.is_ok());
let pred1 = model.predict(&x);
let pred2 = model.predict(&x);
prop_assume!(pred1.is_ok() && pred2.is_ok());
let p1 = pred1.expect("predict succeeded (1st)");
let p2 = pred2.expect("predict succeeded (2nd)");
for i in 0..p1.len() {
prop_assert!(
p1[i] == p2[i],
"FALSIFY-GLM-004: Binomial prediction[{}] differs: {} vs {} (must be deterministic)",
i, p1[i], p2[i]
);
}
}
}