bitpolar 0.3.3

BitPolar: near-optimal vector quantization with zero training overhead — 3-bit precision, provably unbiased inner products (ICLR 2026)
Documentation 
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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236

//! Haar-distributed orthogonal rotation matrix.
//!
//! `StoredRotation` pre-computes and stores a random orthogonal matrix generated
//! via QR decomposition of a Gaussian matrix, producing a Haar-distributed
//! (uniformly random) rotation. This spreads energy uniformly across all
//! coordinates before polar encoding, improving quantization quality.

use crate::error::{Result, TurboQuantError};

/// A stored, pre-computed Haar-distributed orthogonal rotation matrix.
///
/// Constructed once from `(dim, seed)` and reused for all encode/decode
/// operations. The matrix is deterministic: same `(dim, seed)` always
/// produces the same rotation.
#[derive(Debug, Clone)]
#[cfg_attr(feature = "serde-support", derive(serde::Serialize, serde::Deserialize))]
pub struct StoredRotation {
    /// Vector dimension
    dim: usize,
    /// Random seed used to generate this rotation
    seed: u64,
    /// Flat row-major storage of the dim×dim orthogonal matrix
    data: Vec<f32>,
}

impl StoredRotation {
    /// Create a new Haar-distributed rotation for the given dimension and seed.
    ///
    /// # Errors
    /// - `ZeroDimension` if `dim == 0`
    pub fn new(dim: usize, seed: u64) -> Result<Self> {
        if dim == 0 {
            return Err(TurboQuantError::ZeroDimension);
        }
        // Generate a Haar-distributed rotation via QR decomposition.
        // We use a simple construction: fill with Gaussian random values,
        // then perform Gram-Schmidt orthogonalisation (or use nalgebra QR).
        let data = Self::generate_rotation(dim, seed);

        Ok(Self { dim, seed, data })
    }

    /// Return the dimension this rotation operates on.
    pub fn dim(&self) -> usize {
        self.dim
    }

    /// Return the seed used to generate this rotation.
    pub fn seed(&self) -> u64 {
        self.seed
    }

    /// Apply the rotation `R * v` in-place, writing the result into `out`.
    ///
    /// `slice` must have length `self.dim()`.
    #[inline]
    pub fn apply_slice(&self, slice: &[f32], out: &mut Vec<f32>) {
        let d = self.dim;
        debug_assert_eq!(slice.len(), d, "apply_slice: slice.len() must equal self.dim");
        debug_assert_eq!(self.data.len(), d * d, "apply_slice: data.len() must equal dim*dim");
        // Defensive: clear output to zeros on dimension mismatch (prevents stale data)
        out.clear();
        out.resize(d, 0.0);
        if slice.len() != d {
            return;
        }
        for (i, o) in out.iter_mut().enumerate().take(d) {
            *o = self.data[i * d..i * d + d]
                .iter()
                .zip(slice.iter())
                .map(|(m, v)| m * v)
                .sum();
        }
    }

    /// Apply the inverse rotation `R^T * v` (transpose = inverse for orthogonal).
    ///
    /// `slice` must have length `self.dim()`.
    #[inline]
    pub fn apply_inverse_slice(&self, slice: &[f32], out: &mut Vec<f32>) {
        let d = self.dim;
        debug_assert_eq!(slice.len(), d, "apply_inverse_slice: slice.len() must equal self.dim");
        debug_assert_eq!(self.data.len(), d * d, "apply_inverse_slice: data.len() must equal dim*dim");
        // Defensive: clear output to zeros on dimension mismatch (prevents stale data)
        out.clear();
        out.resize(d, 0.0);
        if slice.len() != d {
            return;
        }
        // Transpose multiplication: out[i] = sum_j data[j*d+i] * slice[j]
        for (i, o) in out.iter_mut().enumerate().take(d) {
            *o = slice
                .iter()
                .enumerate()
                .take(d)
                .map(|(j, &s)| self.data[j * d + i] * s)
                .sum();
        }
    }

    /// Approximate heap bytes consumed by this rotation matrix.
    pub fn size_bytes(&self) -> usize {
        self.data.len() * core::mem::size_of::<f32>()
    }

    /// Return the rotation as a nalgebra `DMatrix<f32>` (requires `std` feature).
    #[cfg(feature = "std")]
    pub fn matrix(&self) -> nalgebra::DMatrix<f32> {
        nalgebra::DMatrix::from_row_slice(self.dim, self.dim, &self.data)
    }

    // -----------------------------------------------------------------------
    // Private helpers
    // -----------------------------------------------------------------------

    /// Generate a dim×dim Haar-distributed orthogonal matrix stored row-major.
    ///
    /// Uses a seeded ChaCha RNG for reproducibility. On `std` builds we use
    /// nalgebra's QR decomposition; on `no_std` we fall back to a simple
    /// random-sign diagonal (rotation quality is reduced but still valid).
    fn generate_rotation(dim: usize, seed: u64) -> Vec<f32> {
        #[cfg(feature = "std")]
        {
            use nalgebra::DMatrix;
            use rand::SeedableRng;
            use rand_chacha::ChaCha8Rng;
            use rand_distr::StandardNormal;

            let mut rng = ChaCha8Rng::seed_from_u64(seed);
            let dist = StandardNormal;

            // Fill dim×dim matrix with N(0,1) entries.
            let raw: Vec<f32> = (0..dim * dim)
                .map(|_| <StandardNormal as rand_distr::Distribution<f64>>::sample(&dist, &mut rng) as f32)
                .collect();

            let m = DMatrix::from_row_slice(dim, dim, &raw);
            // QR decomposition gives a Haar-distributed rotation Q.
            let qr = m.qr();
            let q = qr.q();

            // Ensure positive determinant (QR can produce reflections).
            let det_sign = if q.determinant() >= 0.0 { 1.0_f32 } else { -1.0_f32 };
            let mut data = vec![0.0_f32; dim * dim];
            for i in 0..dim {
                for j in 0..dim {
                    data[i * dim + j] = q[(i, j)] * det_sign;
                }
            }
            data
        }

        #[cfg(not(feature = "std"))]
        {
            // Fallback: random-sign diagonal matrix (not truly Haar but valid).
            use rand::SeedableRng;
            use rand_chacha::ChaCha8Rng;
            use rand::Rng;

            let mut rng = ChaCha8Rng::seed_from_u64(seed);
            let mut data = vec![0.0_f32; dim * dim];
            for i in 0..dim {
                let sign: f32 = if rng.gen::<bool>() { 1.0 } else { -1.0 };
                data[i * dim + i] = sign;
            }
            data
        }
    }
}

impl PartialEq for StoredRotation {
    fn eq(&self, other: &Self) -> bool {
        self.dim == other.dim && self.seed == other.seed
    }
}

// ---------------------------------------------------------------------------
// RotationStrategy impl
// ---------------------------------------------------------------------------

impl crate::traits::RotationStrategy for StoredRotation {
    fn rotate(&self, vector: &[f32]) -> Vec<f32> {
        let mut out = Vec::with_capacity(self.dim);
        self.apply_slice(vector, &mut out);
        out
    }

    fn rotate_inverse(&self, vector: &[f32]) -> Vec<f32> {
        let mut out = Vec::with_capacity(self.dim);
        self.apply_inverse_slice(vector, &mut out);
        out
    }

    fn dim(&self) -> usize {
        self.dim
    }
}

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

    #[test]
    fn test_zero_dimension_error() {
        assert!(matches!(StoredRotation::new(0, 42), Err(TurboQuantError::ZeroDimension)));
    }

    #[test]
    fn test_basic_construction() {
        let r = StoredRotation::new(4, 42).unwrap();
        assert_eq!(r.dim(), 4);
        assert_eq!(r.seed(), 42);
    }

    #[test]
    fn test_partial_eq() {
        let a = StoredRotation::new(4, 42).unwrap();
        let b = StoredRotation::new(4, 42).unwrap();
        let c = StoredRotation::new(4, 99).unwrap();
        assert_eq!(a, b);
        assert_ne!(a, c);
    }

    #[test]
    fn test_round_trip_identity() {
        let r = StoredRotation::new(4, 7).unwrap();
        let v = vec![1.0_f32, 2.0, 3.0, 4.0];
        let mut rotated = Vec::new();
        let mut reconstructed = Vec::new();
        r.apply_slice(&v, &mut rotated);
        r.apply_inverse_slice(&rotated, &mut reconstructed);
        for (a, b) in v.iter().zip(reconstructed.iter()) {
            assert!((a - b).abs() < 1e-4, "round-trip failed: {a} vs {b}");
        }
    }
}