Skip to main content

gam_math/
jet_scalar.rs

1//! Order-specific Taylor-jet SCALAR algebras (#932 cutover, doc §A).
2//!
3//! [`crate::jet_tower::Tower4`] carries the full value/gradient/Hessian/`t3`/`t4`
4//! tensor stack: it answers EVERY channel a [`super::row_kernel::RowKernel`]
5//! consumer can ask for, but at `K = 9` that is a ~50 KiB per-row object whose
6//! by-value copies overflowed the stack and timed out the location-scale fit —
7//! which is exactly why `row_kernel_directional_supported()` /
8//! `row_kernel_joint_hessian_supported()` still `return false`. The cutover does
9//! NOT need the dense `Tower4<9>` per row; it needs, per consumer, only the one
10//! channel that consumer serves:
11//!
12//! | consumer | channel | scalar here | K=9 size |
13//! |---|---|---|---|
14//! | inner Newton / `row_kernel` | `(v, g, H)` | [`Order2`] | 728 B |
15//! | `row_third_contracted(dir)` | `Σ_c ℓ_{abc} dir_c` | [`OneSeed`] | 1.46 KiB |
16//! | `row_fourth_contracted(u, v)` | `Σ_{cd} ℓ_{abcd} u_c v_d` | [`TwoSeed`] | 2.8 KiB |
17//!
18//! Each is built on [`Order2`] (value/grad/Hessian), which is the production
19//! [`crate::jet_tower::Tower2`] re-expressed behind a generic interface: a row
20//! loss written ONCE against [`JetScalar`] re-instantiates at whatever order /
21//! representation a consumer needs, with the contraction folded INTO the
22//! differentiation (the nilpotent ε / δ directions), so `t3` / `t4` are never
23//! materialised. The single source of truth is the same one expression — the
24//! genus of #736 cross-block drift cannot reappear because there is no separate
25//! channel to forget.
26//!
27//! # Why each scalar is exact (doc §A.1–A.3)
28//!
29//! * [`Order2`] is the order-≤2 truncation of the Leibniz / Faà di Bruno rules.
30//!   Those order-2 terms read ONLY the order-≤2 channels of their inputs (see
31//!   [`crate::jet_tower::Tower4::mul`]: `out.h[i][j]` never touches `t3`/`t4`),
32//!   so its `(v, g, H)` is BIT-IDENTICAL to a full `Tower4<K>` — and identical
33//!   to [`crate::jet_tower::Tower2`], over which it is a thin newtype.
34//! * [`OneSeed`] carries an [`Order2`] base plus one nilpotent ε (`ε² = 0`)
35//!   holding another [`Order2`]. Seeding ε with the fixed direction `u` makes the
36//!   ε-component of the Hessian channel the contracted third `Σ_c ℓ_{abc} u_c`
37//!   (the nilpotent implements `d/dτ|₀` of `ℓ_{ab}(p + τu)` exactly).
38//! * [`TwoSeed`] carries an [`Order2`] base plus ε, δ (`ε² = δ² = 0`, `εδ`
39//!   retained) — four [`Order2`] parts. Seeding ε, δ with `u, v` makes the
40//!   εδ-component of the Hessian channel the contracted fourth
41//!   `Σ_{cd} ℓ_{abcd} u_c v_d` (the single mixed `∂_σ∂_ρ|₀` term, no `σ²`/`ρ²`
42//!   contamination).
43//!
44//! # Stability discipline
45//!
46//! As in [`crate::jet_tower`], humans own primitive stability and the algebra
47//! owns combinatorics: tail-critical special functions enter ONLY as
48//! hand-certified `[f64; 5]` derivative stacks through [`JetScalar::compose_unary`]
49//! (each scalar consumes the leading entries its order needs), never by
50//! differentiating an unstable primal.
51//!
52//! # Production scalars and the test-only all-channels oracle
53//!
54//! The `JetScalar` trait below is production: it is the bound on
55//! [`crate::jet_tower::RowNllProgramGeneric::row_nll_generic`], the seam a family
56//! row loss is written against. The order-specific scalars that *consume* it —
57//! [`Order2`] (value/grad/Hessian), [`OneSeed`] (contracted third) and
58//! [`TwoSeed`] (contracted fourth) — are production: the survival location-scale
59//! `RowKernel<9>` builds its joint Hessian / directional derivatives through them
60//! (`survival::location_scale::row_kernel`), paying only the small packed scalar
61//! per row instead of the ~50 KiB dense [`crate::jet_tower::Tower4`].
62//!
63//! The [`crate::jet_tower::Tower4`] all-channels `JetScalar` impl is test-only: it
64//! is the oracle that pins the contracted scalars against the dense
65//! value/grad/Hessian/`t3`/`t4` truth, so it lives in the `#[cfg(test)]` module.
66
67/// A truncated-Taylor scalar carrying derivatives in `K` primaries.
68///
69/// All concrete scalars here ([`Order2`], [`OneSeed`], [`TwoSeed`]) and the full
70/// [`crate::jet_tower::Tower4`] implement the SAME algebra; only the carried
71/// channel set differs. A row loss written once against this interface yields a
72/// different channel set per instantiation, all exact for the channel they serve
73/// (doc §A.0).
74pub trait JetScalar<const K: usize>: Copy {
75    /// A constant: value `c`, every derivative channel zero.
76    fn constant(c: f64) -> Self;
77
78    /// The seeded variable `p_axis` at value `x`: unit first derivative in slot
79    /// `axis`, all higher channels zero. (The nilpotent / cross channels of the
80    /// directional scalars are seeded zero — callers set ε/δ directions through
81    /// the scalar-specific [`OneSeed::seed_direction`] / [`TwoSeed::seed`].)
82    fn variable(x: f64, axis: usize) -> Self;
83
84    /// The value channel `ℓ(p)`.
85    fn value(&self) -> f64;
86
87    /// Exact truncated Leibniz sum `self + o`.
88    fn add(&self, o: &Self) -> Self;
89    /// Exact truncated Leibniz difference `self − o`.
90    fn sub(&self, o: &Self) -> Self;
91    /// Exact truncated Leibniz product `self · o`.
92    fn mul(&self, o: &Self) -> Self;
93    /// Negate every channel.
94    fn neg(&self) -> Self;
95    /// Multiply every channel by a plain scalar `s`.
96    fn scale(&self, s: f64) -> Self;
97
98    /// Exact multivariate Faà di Bruno composition `f ∘ self`, given the outer
99    /// derivative stack `d = [f(u), f′(u), f″(u), f‴(u), f⁗(u)]` at
100    /// `u = self.value()`.
101    ///
102    /// This is the SAME `[f64; 5]` stack shape [`crate::jet_tower::Tower4`] and
103    /// the families' `unary_derivatives_*` helpers (built on erfcx / log_ndtr)
104    /// already produce, so those stacks plug in directly. Each scalar consumes
105    /// only the leading entries its order needs (order-2 reads `d[0..=2]`; the
106    /// directional scalars read one / two beyond their base) — the fixed-length
107    /// array makes that windowing total, no length guard required.
108    fn compose_unary(&self, d: [f64; 5]) -> Self;
109
110    /// Compose with a unary special-function whose derivative STACK is built
111    /// from the scalar base value through `stack_fn` — the generic-over-`Lane`
112    /// seam that lets a single-sourced row program instantiate at BOTH the scalar
113    /// `f64` jets and the SIMD `f64x4` batch towers from ONE expression.
114    ///
115    /// On a scalar jet this evaluates `stack_fn(self.value())` ONCE and forwards
116    /// to [`compose_unary`](Self::compose_unary), so it is BIT-IDENTICAL to the
117    /// hand-written `self.compose_unary(stack_fn(self.value()))` (default body
118    /// below). The lever is that the SAME call shape exists on
119    /// [`crate::jet_tower::Tower3Lane`] / [`crate::jet_tower::Tower4Lane`], where
120    /// the four lanes carry FOUR DISTINCT base values, so the batch
121    /// implementation re-runs `stack_fn` per lane — a thing the old
122    /// `compose_unary(stack_from(self.value()))` shape could not express on a
123    /// batch type (it has no single scalar `.value()`). Writing a row program
124    /// against this method instead of the explicit two-step is what makes it
125    /// instantiate, unchanged, at `f64x4` for the 4-rows-per-pass batch path.
126    fn compose_unary_with(&self, stack_fn: impl Fn(f64) -> [f64; 5]) -> Self {
127        self.compose_unary(stack_fn(self.value()))
128    }
129
130    /// `e^self`. Convenience for tame arguments (see module stability note).
131    fn exp(&self) -> Self {
132        let e = self.value().exp();
133        self.compose_unary([e, e, e, e, e])
134    }
135
136    /// `√self`. Caller guarantees positivity.
137    fn sqrt(&self) -> Self {
138        let u = self.value();
139        let s = u.sqrt();
140        self.compose_unary([
141            s,
142            0.5 / s,
143            -0.25 / (u * s),
144            0.375 / (u * u * s),
145            -0.9375 / (u * u * u * s),
146        ])
147    }
148
149    /// `ln(self)`. Caller guarantees positivity. Same derivative stack
150    /// [`crate::jet_tower::Tower4::ln`] uses, so any program written over both
151    /// matches term-for-term.
152    fn ln(&self) -> Self {
153        let u = self.value();
154        let r = 1.0 / u;
155        self.compose_unary([u.ln(), r, -r * r, 2.0 * r * r * r, -6.0 * r * r * r * r])
156    }
157
158    /// `1/self`.
159    fn recip(&self) -> Self {
160        let r = 1.0 / self.value();
161        let r2 = r * r;
162        self.compose_unary([r, -r2, 2.0 * r2 * r, -6.0 * r2 * r2, 24.0 * r2 * r2 * r])
163    }
164
165    /// `self^a` for real exponent `a`. Caller guarantees a positive base.
166    /// Mirrors [`crate::jet_tower::Tower4::powf`] (falling-factorial stack).
167    fn powf(&self, a: f64) -> Self {
168        let u = self.value();
169        self.compose_unary([
170            u.powf(a),
171            a * u.powf(a - 1.0),
172            a * (a - 1.0) * u.powf(a - 2.0),
173            a * (a - 1.0) * (a - 2.0) * u.powf(a - 3.0),
174            a * (a - 1.0) * (a - 2.0) * (a - 3.0) * u.powf(a - 4.0),
175        ])
176    }
177
178    /// `ln Γ(self)`. Caller guarantees a positive argument. Uses the SAME
179    /// hand-certified derivative stack [`crate::jet_tower::Tower4::ln_gamma`]
180    /// consumes ([`crate::jet_tower::ln_gamma_derivative_stack`]), so any
181    /// program written over both matches term-for-term.
182    fn ln_gamma(&self) -> Self {
183        self.compose_unary(crate::jet_tower::ln_gamma_derivative_stack(self.value()))
184    }
185
186    /// `ψ(self) = d/dx ln Γ(x)` (digamma). Caller guarantees a positive
187    /// argument. Same hand-certified stack
188    /// [`crate::jet_tower::digamma_derivative_stack`].
189    fn digamma(&self) -> Self {
190        self.compose_unary(crate::jet_tower::digamma_derivative_stack(self.value()))
191    }
192}
193
194// ── Order2<K> ergonomic operator overloads (doc §A.1) ───────────────────
195//
196// The dispersion-family row NLLs are written with `+`/`-`/`*` operators over
197// the primaries (mirroring how they read as `Tower4` expressions). These
198// delegate channel-for-channel to the inner `Tower2` arithmetic (which has
199// `Add`/`Mul`; `Sub`/`Neg` are expressed as `+ (-1)·rhs` exactly as the
200// `JetScalar::sub` / `JetScalar::neg` impls do), so an `Order2` expression is
201// bit-identical to the same `Tower4` expression's order-≤2 channels.
202
203impl<const K: usize> std::ops::Add for Order2<K> {
204    type Output = Self;
205    #[inline]
206    fn add(self, o: Self) -> Self {
207        Order2(self.0 + o.0)
208    }
209}
210
211impl<const K: usize> std::ops::Add<f64> for Order2<K> {
212    type Output = Self;
213    #[inline]
214    fn add(self, c: f64) -> Self {
215        Order2(self.0 + c)
216    }
217}
218
219impl<const K: usize> std::ops::Sub for Order2<K> {
220    type Output = Self;
221    #[inline]
222    fn sub(self, o: Self) -> Self {
223        Order2(self.0 + o.0.scale(-1.0))
224    }
225}
226
227impl<const K: usize> std::ops::Sub<f64> for Order2<K> {
228    type Output = Self;
229    #[inline]
230    fn sub(self, c: f64) -> Self {
231        Order2(self.0 + (-c))
232    }
233}
234
235impl<const K: usize> std::ops::Mul for Order2<K> {
236    type Output = Self;
237    #[inline]
238    fn mul(self, o: Self) -> Self {
239        Order2(crate::jet_tower::Tower2::mul(&self.0, &o.0))
240    }
241}
242
243impl<const K: usize> std::ops::Mul<f64> for Order2<K> {
244    type Output = Self;
245    #[inline]
246    fn mul(self, c: f64) -> Self {
247        Order2(self.0.scale(c))
248    }
249}
250
251impl<const K: usize> std::ops::Neg for Order2<K> {
252    type Output = Self;
253    #[inline]
254    fn neg(self) -> Self {
255        Order2(self.0.scale(-1.0))
256    }
257}
258
259/// Filtered Hensel lift of a SCALAR implicit state `a(θ)` defined by the
260/// constraint `F(a, θ) = 0`, evaluated in ANY [`JetScalar`] algebra `S` (doc
261/// §11, "A generic implicit-lift operator for every production scalar").
262///
263/// This is the perf-respecting alternative to lifting through a dense
264/// `Tower4<K+1>` (which carries the implicit variable as an extra dense axis):
265/// the state `a` lives directly in the consumer's own `K`-primary algebra
266/// `S` — `Order2<K>` for value/gradient/Hessian, `Tower4<K>` for the full
267/// `t3`/`t4` — never paying for an extra variable.
268///
269/// **Method.** Fixed-Jacobian Newton in the nilpotent algebra. By the
270/// filtered-lift theorem (doc §11.1), if `F_a := ∂F/∂a(a₀, θ₀)` is the primal
271/// Jacobian at the base point and `inv_fa = 1/F_a`, then the iteration
272/// `A ← A − inv_fa · F(A, θ)` raises the filtration degree of the residual by
273/// at least one per step: each step kills exactly one graded layer. Starting
274/// from `A = const(a₀)` (whose residual lies in `F¹` because `θ − θ₀ ∈ 𝔫`),
275/// `iters` equal to the algebra's nilpotency order returns the *exact* lifted
276/// jet (`Order2`: 2, `OneSeed`: 3, `Tower4`/`TwoSeed`: 4). The value channel of
277/// `A` never moves — `F(A, θ).value() = F(a₀, θ₀) = 0` at the certified root —
278/// so a caller may precompute every primitive's derivative stack at the fixed
279/// base index once and let the cheap polynomial composition repeat per step.
280///
281/// `f` evaluates the constraint `F(a, θ)` in `S` (capturing the seeded
282/// parameter jets `θ`); `a0` is the certified scalar root `F(a₀, θ₀) ≈ 0`.
283pub fn filtered_implicit_solve_scalar<const K: usize, S: JetScalar<K>>(
284    a0: f64,
285    inv_fa: f64,
286    iters: usize,
287    f: impl Fn(&S) -> S,
288) -> S {
289    let mut a = S::constant(a0);
290    for _ in 0..iters {
291        let residual = f(&a);
292        a = a.sub(&residual.scale(inv_fa));
293    }
294    a
295}
296
297// ── Order2<K>: value / gradient / Hessian (doc §A.1) ────────────────────
298
299/// Truncated SECOND-order scalar: value `v`, gradient `g_a`, Hessian `H_{ab}`.
300///
301/// This is a thin newtype over the production [`crate::jet_tower::Tower2`], so
302/// its `(v, g, H)` channels are obtained by the SAME formulas — and are
303/// therefore bit-identical to both [`crate::jet_tower::Tower2`] and the order-≤2
304/// channels of a full [`crate::jet_tower::Tower4`] (doc §A.1, "Bit-identity with
305/// the full tower"). The wrapper exists only to satisfy the generic
306/// [`JetScalar`] interface (the `compose_unary` / `add` / `sub` / `neg` /
307/// `recip` the trait demands, which `Tower2` does not expose by that shape) —
308/// every channel is delegated to `Tower2` arithmetic unchanged.
309#[derive(Clone, Copy, Debug)]
310pub struct Order2<const K: usize>(pub crate::jet_tower::Tower2<K>);
311
312impl<const K: usize> Order2<K> {
313    /// Read the gradient channel `g_a = ∂ℓ/∂p_a`.
314    #[inline]
315    pub fn g(&self) -> [f64; K] {
316        self.0.g
317    }
318
319    /// Read the Hessian channel.
320    #[inline]
321    pub fn h(&self) -> [[f64; K]; K] {
322        self.0.h
323    }
324}
325
326impl<const K: usize> JetScalar<K> for Order2<K> {
327    fn constant(c: f64) -> Self {
328        Order2(crate::jet_tower::Tower2::constant(c))
329    }
330    fn variable(x: f64, axis: usize) -> Self {
331        Order2(crate::jet_tower::Tower2::variable(x, axis))
332    }
333    fn value(&self) -> f64 {
334        self.0.v
335    }
336    fn add(&self, o: &Self) -> Self {
337        Order2(self.0 + o.0)
338    }
339    fn sub(&self, o: &Self) -> Self {
340        // Tower2 has no Sub op; subtract by adding the negation, matching
341        // Tower4::sub (self + o.scale(-1.0)).
342        Order2(self.0 + o.0.scale(-1.0))
343    }
344    fn mul(&self, o: &Self) -> Self {
345        Order2(crate::jet_tower::Tower2::mul(&self.0, &o.0))
346    }
347    fn neg(&self) -> Self {
348        Order2(self.0.scale(-1.0))
349    }
350    fn scale(&self, s: f64) -> Self {
351        Order2(self.0.scale(s))
352    }
353    fn compose_unary(&self, d: [f64; 5]) -> Self {
354        // Order-≤2 reads only [f, f', f''] of the stack.
355        Order2(self.0.compose_unary([d[0], d[1], d[2]]))
356    }
357}
358
359// ── Lane-batched Order-2 scalar: 4 rows per pass in SIMD lanes (perf) ────
360//
361// The hot per-row jet kernels evaluate ONE row's `(v, g, H)` tower at a time in
362// scalar `f64`. A hand-written scalar derivative does the same. The throughput
363// lever a jet has that scalar hand-code cannot is **row batching in SIMD
364// lanes**: the order-≤2 Leibniz product `Order2::mul` is `O(K²)` independent
365// per-channel float ops, and EVERY row runs the identical op graph on different
366// data — the textbook SPMD shape. Packing `LANES = 4` rows into a `wide::f64x4`
367// and running the algebra once per 4 rows replaces 4 scalar passes with one
368// vector pass: the `K²` Hessian channel updates become `K²` NEON `.2d` / SSE2
369// `pd` instructions covering 4 rows each, ~4× fewer FP instructions per row.
370//
371// The carried scalar field is abstracted by [`Lane`] so the SAME algebra body
372// instantiates at `f64` (1 row, used as the bit-identity oracle) or
373// [`wide::f64x4`] (4 rows). Bit-identity is structural, not approximate:
374//
375//   * Every arithmetic op is a plain lane-wise `+` / `-` / `*` (NEVER a fused
376//     `mul_add`), and IEEE-754 double `+`/`-`/`*`/`/` are correctly rounded and
377//     deterministic, so lane `i` of an `f64x4` op equals the scalar `f64` op on
378//     that lane's inputs bit-for-bit.
379//   * The transcendental derivative STACKS (`exp`/`ln`/`sqrt`/…) are produced
380//     **per lane by the identical scalar code** ([`Lane::unary3`] unpacks, runs
381//     the same `[f64; 3]` stack closure the scalar path runs, repacks), so the
382//     only thing vectorised is the cheap rational tensor composition — the
383//     library transcendental itself is the exact same `f64::exp` call per lane.
384//   * The op order mirrors [`crate::jet_tower::Tower2`] term-for-term, so
385//     [`Order2Lane<f64, K>`] is `to_bits`-identical to the production
386//     [`Order2<K>`] (= `Tower2<K>`), and [`Order2Lane<f64x4, K>`] lane `i` is
387//     `to_bits`-identical to that — proven by the `batch_tests` oracle below
388//     (≥2000 random 4-row batches across `K ∈ {2,3,4,9}`).
389
390/// The scalar field a [`Order2Lane`] carries: either a single `f64` (one row,
391/// the oracle) or a [`wide::f64x4`] (four rows evaluated in SIMD lanes). All ops
392/// are plain lane-wise IEEE arithmetic, so a vector op equals the scalar op on
393/// each lane bit-for-bit.
394pub trait Lane: Copy {
395    /// Broadcast a scalar to every lane.
396    fn splat(x: f64) -> Self;
397    /// Lane-wise `self + o`.
398    fn add(self, o: Self) -> Self;
399    /// Lane-wise `self - o`.
400    fn sub(self, o: Self) -> Self;
401    /// Lane-wise `self * o`.
402    fn mul(self, o: Self) -> Self;
403    /// The `f64` in lane `i` (`i < LANES`; `f64` ignores `i`).
404    fn lane(self, i: usize) -> f64;
405    /// Build the order-≤2 derivative stack `[f(u), f′(u), f″(u)]` **per lane**
406    /// from the lane value `u`, via the SAME scalar `stack` closure the
407    /// per-row path runs (so the transcendental/rational stack is bit-identical
408    /// to the scalar evaluation — only the subsequent tensor composition is
409    /// vectorised).
410    fn unary3(self, stack: impl Fn(f64) -> [f64; 3]) -> [Self; 3];
411    /// Build the order-≤4 derivative stack `[f, f′, f″, f‴, f⁗]` **per lane**
412    /// from the lane value `u`, via the SAME scalar `stack` closure the per-row
413    /// path runs. The one-/two-seed scalars ([`OneSeedLane`] / [`TwoSeedLane`])
414    /// need outer derivatives one / two orders beyond their order-2 base, so
415    /// they build their composition stack through this five-entry variant. As
416    /// with [`unary3`](Lane::unary3), only the transcendental/rational stack is
417    /// evaluated per lane (bit-identically to the scalar path); the subsequent
418    /// tensor composition is vectorised.
419    fn unary5(self, stack: impl Fn(f64) -> [f64; 5]) -> [Self; 5];
420    /// The general-`N` sibling of [`unary3`](Lane::unary3) / [`unary5`](Lane::unary5):
421    /// build an `N`-wide derivative stack **per lane** from the lane value, via
422    /// the SAME scalar `stack` closure the per-row path runs, then pack the `N`
423    /// columns lane-wise. This is the lane primitive the compose-with-stack seam
424    /// ([`crate::jet_tower::Tower4Lane::compose_unary_with`] and its `Tower3`
425    /// sibling) routes through: it evaluates `stack` once per lane at that lane's
426    /// OWN base value (each of the four rows in an `f64x4` carries a distinct
427    /// base), so lane `i` of the packed result equals the scalar `stack(value_i)`
428    /// bit-for-bit (only the cheap pack is vectorised; the closure body is the
429    /// identical scalar code). With `N = 3` / `N = 5` it is `to_bits`-identical to
430    /// [`unary3`](Lane::unary3) / [`unary5`](Lane::unary5).
431    fn unary_with<const N: usize>(self, stack: impl Fn(f64) -> [f64; N]) -> [Self; N];
432}
433
434impl Lane for f64 {
435    #[inline]
436    fn splat(x: f64) -> Self {
437        x
438    }
439    #[inline]
440    fn add(self, o: Self) -> Self {
441        self + o
442    }
443    #[inline]
444    fn sub(self, o: Self) -> Self {
445        self - o
446    }
447    #[inline]
448    fn mul(self, o: Self) -> Self {
449        self * o
450    }
451    #[inline]
452    fn lane(self, _: usize) -> f64 {
453        self
454    }
455    #[inline]
456    fn unary3(self, stack: impl Fn(f64) -> [f64; 3]) -> [Self; 3] {
457        stack(self)
458    }
459    #[inline]
460    fn unary5(self, stack: impl Fn(f64) -> [f64; 5]) -> [Self; 5] {
461        stack(self)
462    }
463    #[inline]
464    fn unary_with<const N: usize>(self, stack: impl Fn(f64) -> [f64; N]) -> [Self; N] {
465        // One row: the packed result IS the scalar stack ([Self; N] = [f64; N]).
466        stack(self)
467    }
468}
469
470impl Lane for wide::f64x4 {
471    #[inline]
472    fn splat(x: f64) -> Self {
473        wide::f64x4::splat(x)
474    }
475    #[inline]
476    fn add(self, o: Self) -> Self {
477        self + o
478    }
479    #[inline]
480    fn sub(self, o: Self) -> Self {
481        self - o
482    }
483    #[inline]
484    fn mul(self, o: Self) -> Self {
485        self * o
486    }
487    #[inline]
488    fn lane(self, i: usize) -> f64 {
489        self.to_array()[i]
490    }
491    #[inline]
492    fn unary3(self, stack: impl Fn(f64) -> [f64; 3]) -> [Self; 3] {
493        let a = self.to_array();
494        let mut d0 = [0.0_f64; 4];
495        let mut d1 = [0.0_f64; 4];
496        let mut d2 = [0.0_f64; 4];
497        for i in 0..4 {
498            let s = stack(a[i]);
499            d0[i] = s[0];
500            d1[i] = s[1];
501            d2[i] = s[2];
502        }
503        [
504            wide::f64x4::new(d0),
505            wide::f64x4::new(d1),
506            wide::f64x4::new(d2),
507        ]
508    }
509    #[inline]
510    fn unary5(self, stack: impl Fn(f64) -> [f64; 5]) -> [Self; 5] {
511        let a = self.to_array();
512        let mut d = [[0.0_f64; 4]; 5];
513        for i in 0..4 {
514            let s = stack(a[i]);
515            for (k, dk) in d.iter_mut().enumerate() {
516                dk[i] = s[k];
517            }
518        }
519        [
520            wide::f64x4::new(d[0]),
521            wide::f64x4::new(d[1]),
522            wide::f64x4::new(d[2]),
523            wide::f64x4::new(d[3]),
524            wide::f64x4::new(d[4]),
525        ]
526    }
527    #[inline]
528    fn unary_with<const N: usize>(self, stack: impl Fn(f64) -> [f64; N]) -> [Self; N] {
529        // Evaluate the scalar stack PER LANE at that lane's own base value, then
530        // pack the N derivative columns lane-wise (the same shape `unary5` uses,
531        // generalised to N). Lane `i` of column `k` is `stack(base_i)[k]`.
532        let a = self.to_array();
533        let mut cols = [[0.0_f64; 4]; N];
534        for (i, &base) in a.iter().enumerate() {
535            let s = stack(base);
536            for (k, sk) in s.iter().enumerate() {
537                cols[k][i] = *sk;
538            }
539        }
540        std::array::from_fn(|k| wide::f64x4::new(cols[k]))
541    }
542}
543
544/// A lane-batched order-≤2 Taylor scalar: value / gradient / Hessian carried in
545/// a SIMD field [`L: Lane`](Lane). With `L = f64x4` one instance carries FOUR
546/// rows at once, so the row loop processes 4 rows per vector pass instead of one
547/// per scalar pass.
548///
549/// The channel layout and every float op mirror [`crate::jet_tower::Tower2`]
550/// term-for-term, so `Order2Lane<f64, K>` is `to_bits`-identical to the
551/// production [`Order2<K>`] and `Order2Lane<f64x4, K>` lane `i` is
552/// `to_bits`-identical to that (see the module note and `batch_tests`).
553#[derive(Clone, Copy, Debug)]
554pub struct Order2Lane<L: Lane, const K: usize> {
555    /// Value channel `ℓ` (one entry per lane/row).
556    pub v: L,
557    /// Gradient channel `∂ℓ/∂p_a`.
558    pub g: [L; K],
559    /// Hessian channel `∂²ℓ/∂p_a∂p_b` (symmetric).
560    pub h: [[L; K]; K],
561}
562
563/// The 4-rows-per-pass batched order-≤2 scalar (`wide::f64x4` lanes).
564pub type Order2Batch<const K: usize> = Order2Lane<wide::f64x4, K>;
565
566impl<L: Lane, const K: usize> Order2Lane<L, K> {
567    /// A constant: value `c` in every channel-zero slot.
568    #[inline]
569    pub fn constant(c: L) -> Self {
570        Order2Lane {
571            v: c,
572            g: [L::splat(0.0); K],
573            h: [[L::splat(0.0); K]; K],
574        }
575    }
576
577    /// The seeded variable `p_axis` at (per-lane) value `value`: unit first
578    /// derivative in slot `axis`. With `L = f64x4`, `value` packs the four
579    /// rows' values of primary `axis`.
580    #[inline]
581    pub fn variable(value: L, axis: usize) -> Self {
582        let mut out = Self::constant(value);
583        out.g[axis] = L::splat(1.0);
584        out
585    }
586
587    /// Lane-wise `self + o` (mirrors `Tower2` Add: per-channel add).
588    #[inline]
589    pub fn add(&self, o: &Self) -> Self {
590        let mut out = *self;
591        out.v = self.v.add(o.v);
592        for i in 0..K {
593            out.g[i] = self.g[i].add(o.g[i]);
594            for j in 0..K {
595                out.h[i][j] = self.h[i][j].add(o.h[i][j]);
596            }
597        }
598        out
599    }
600
601    /// Multiply every channel by the plain scalar `s` (mirrors `Tower2::scale`).
602    #[inline]
603    pub fn scale(&self, s: f64) -> Self {
604        let sl = L::splat(s);
605        let mut out = *self;
606        out.v = self.v.mul(sl);
607        for i in 0..K {
608            out.g[i] = self.g[i].mul(sl);
609            for j in 0..K {
610                out.h[i][j] = self.h[i][j].mul(sl);
611            }
612        }
613        out
614    }
615
616    /// Lane-wise `self - o`, expressed as `self + o·(-1)` exactly as
617    /// [`Order2::sub`] / `Tower4::sub` do, so signed-zero handling matches.
618    #[inline]
619    pub fn sub(&self, o: &Self) -> Self {
620        self.add(&o.scale(-1.0))
621    }
622
623    /// Negate every channel (= `scale(-1.0)`, matching [`Order2::neg`]).
624    #[inline]
625    pub fn neg(&self) -> Self {
626        self.scale(-1.0)
627    }
628
629    /// Exact order-≤2 Leibniz product, term-for-term identical to
630    /// [`crate::jet_tower::Tower2::mul`] (same factor order, no `mul_add`).
631    ///
632    /// The Hessian channel is symmetric under `i ↔ j` (see
633    /// [`crate::jet_tower::Tower2::mul`] for why the invariant always holds), so
634    /// we compute the upper triangle `j ≥ i` and mirror it — `K(K+1)/2` lane
635    /// entry-chains instead of `K²`. Because each lane entry is already a full
636    /// SIMD op (no cross-`j` lane packing to lose), halving the entry count is a
637    /// direct throughput win (~18 % on `Order2Lane<f64x4, 9>`, the survival batch
638    /// kernel, and ~2× on the `f64` oracle). The upper triangle uses the EXACT
639    /// term order of `Tower2::mul`, so `Order2Lane<f64>` stays `to_bits`-identical
640    /// to `Order2` (= `Tower2`) and `Order2Lane<f64x4>` lane `i` stays
641    /// `to_bits`-identical to that; the mirror makes the batch Hessian exactly
642    /// symmetric, matching the scalar `Tower2::mul` (which mirrors identically).
643    #[inline]
644    pub fn mul(&self, o: &Self) -> Self {
645        let a = self;
646        let b = o;
647        let mut out = Self::constant(a.v.mul(b.v));
648        for i in 0..K {
649            // a.v*b.g[i] + a.g[i]*b.v
650            out.g[i] = a.v.mul(b.g[i]).add(a.g[i].mul(b.v));
651        }
652        for i in 0..K {
653            for j in i..K {
654                // a.v*b.h + a.g[i]*b.g[j] + a.g[j]*b.g[i] + a.h*b.v
655                let hij = a
656                    .v
657                    .mul(b.h[i][j])
658                    .add(a.g[i].mul(b.g[j]))
659                    .add(a.g[j].mul(b.g[i]))
660                    .add(a.h[i][j].mul(b.v));
661                out.h[i][j] = hij;
662                out.h[j][i] = hij;
663            }
664        }
665        out
666    }
667
668    /// Exact order-≤2 Faà di Bruno composition `f ∘ self`, given the per-lane
669    /// derivative stack `d = [f(u), f′(u), f″(u)]`. Mirrors
670    /// [`crate::jet_tower::Tower2::compose_unary`] term-for-term (`acc` starts at
671    /// `0` then accumulates, so signed-zero collapses identically).
672    #[inline]
673    pub fn compose_unary(&self, d: [L; 3]) -> Self {
674        let mut out = Self::constant(d[0]);
675        for i in 0..K {
676            let mut acc = L::splat(0.0);
677            acc = acc.add(d[1].mul(self.g[i]));
678            out.g[i] = acc;
679        }
680        for i in 0..K {
681            for j in 0..K {
682                let mut acc = L::splat(0.0);
683                acc = acc.add(d[1].mul(self.h[i][j]));
684                acc = acc.add(d[2].mul(self.g[i]).mul(self.g[j]));
685                out.h[i][j] = acc;
686            }
687        }
688        out
689    }
690
691    /// `e^self`, per-lane stack `[e, e, e]` (matches the [`JetScalar::exp`]
692    /// default forwarded through `Order2`).
693    #[inline]
694    pub fn exp(&self) -> Self {
695        let d = self.v.unary3(|u| {
696            let e = u.exp();
697            [e, e, e]
698        });
699        self.compose_unary(d)
700    }
701
702    /// `ln(self)`; caller guarantees positivity. Per-lane stack
703    /// `[ln u, 1/u, -1/u²]` (matches [`JetScalar::ln`] truncated to order 2).
704    #[inline]
705    pub fn ln(&self) -> Self {
706        let d = self.v.unary3(|u| {
707            let r = 1.0 / u;
708            [u.ln(), r, -r * r]
709        });
710        self.compose_unary(d)
711    }
712
713    /// `√self`; caller guarantees positivity. Per-lane stack
714    /// `[s, 0.5/s, -0.25/(u·s)]` (matches [`JetScalar::sqrt`]).
715    #[inline]
716    pub fn sqrt(&self) -> Self {
717        let d = self.v.unary3(|u| {
718            let s = u.sqrt();
719            [s, 0.5 / s, -0.25 / (u * s)]
720        });
721        self.compose_unary(d)
722    }
723
724    /// `1/self`. Per-lane stack `[r, -r², 2r³]` (matches [`JetScalar::recip`]).
725    #[inline]
726    pub fn recip(&self) -> Self {
727        let d = self.v.unary3(|u| {
728            let r = 1.0 / u;
729            let r2 = r * r;
730            [r, -r2, 2.0 * r2 * r]
731        });
732        self.compose_unary(d)
733    }
734
735    /// `self^a` for real `a`; caller guarantees a positive base. Per-lane
736    /// falling-factorial stack (matches [`JetScalar::powf`]).
737    #[inline]
738    pub fn powf(&self, a: f64) -> Self {
739        let d = self.v.unary3(|u| {
740            [
741                u.powf(a),
742                a * u.powf(a - 1.0),
743                a * (a - 1.0) * u.powf(a - 2.0),
744            ]
745        });
746        self.compose_unary(d)
747    }
748}
749
750impl<const K: usize> Order2Batch<K> {
751    /// Extract lane `i`'s `(v, g, H)` as a production [`Order2<K>`] scalar.
752    /// Lane `i` is `to_bits`-identical to evaluating the same program at
753    /// [`Order2<K>`] on row `i` (see `batch_tests`).
754    #[inline]
755    #[must_use]
756    pub fn lane(&self, i: usize) -> Order2<K> {
757        let mut t = crate::jet_tower::Tower2::<K>::constant(self.v.lane(i));
758        for a in 0..K {
759            t.g[a] = self.g[a].lane(i);
760            for b in 0..K {
761                t.h[a][b] = self.h[a][b].lane(i);
762            }
763        }
764        Order2(t)
765    }
766}
767
768// ── Order1<K>: value / gradient only (doc §A.1, first-order prune) ──────
769
770/// Truncated FIRST-order scalar: value `v` and gradient `g_a` only — NO Hessian.
771///
772/// This is [`Order2`] with the K×K Hessian channel deleted. Its value and
773/// gradient are computed by the SAME order-≤1 truncation of the Leibniz / Faà
774/// di Bruno rules that [`Order2`] uses for those two channels, with the float
775/// operations applied in the identical order — so its `(v, g)` is BIT-IDENTICAL
776/// to both [`Order2`]'s and a full [`crate::jet_tower::Tower4`]'s order-≤1
777/// channels. Use it at a consumer that reads ONLY value + gradient (the SAE
778/// β-border channel: the reconstruction is linear in β, so the Hessian-in-β
779/// vanishes and the dense K×K Hessian product `Tower2::mul` would build is pure
780/// discarded work). Order-≤1 value/gradient never read any input's Hessian, so
781/// dropping that channel changes neither result nor float-op order — it only
782/// removes the `K²` arithmetic that produced an unread tensor.
783#[derive(Clone, Copy, Debug)]
784pub struct Order1<const K: usize> {
785    /// Value ℓ.
786    pub v: f64,
787    /// Gradient ∂ℓ/∂p_a.
788    pub g: [f64; K],
789}
790
791impl<const K: usize> Order1<K> {
792    /// Read the gradient channel `g_a = ∂ℓ/∂p_a`.
793    #[inline]
794    pub fn g(&self) -> [f64; K] {
795        self.g
796    }
797}
798
799impl<const K: usize> JetScalar<K> for Order1<K> {
800    fn constant(c: f64) -> Self {
801        // Order2::constant -> Tower2::constant: value c, all derivatives zero.
802        Order1 { v: c, g: [0.0; K] }
803    }
804    fn variable(x: f64, axis: usize) -> Self {
805        // Order2::variable -> Tower2::variable: unit first derivative in `axis`.
806        let mut g = [0.0; K];
807        g[axis] = 1.0;
808        Order1 { v: x, g }
809    }
810    fn value(&self) -> f64 {
811        self.v
812    }
813    fn add(&self, o: &Self) -> Self {
814        // Tower2 Add: out.v += o.v; out.g[i] += o.g[i] (same float order).
815        let mut g = self.g;
816        for i in 0..K {
817            g[i] += o.g[i];
818        }
819        Order1 { v: self.v + o.v, g }
820    }
821    fn sub(&self, o: &Self) -> Self {
822        // Mirror Order2::sub == self + o.scale(-1.0) exactly: scale then add.
823        self.add(&o.scale(-1.0))
824    }
825    fn mul(&self, o: &Self) -> Self {
826        // Tower2::mul value/grad terms, identical float order:
827        //   v = a.v*b.v;  g[i] = a.v*b.g[i] + a.g[i]*b.v.
828        // (The Hessian loop `a.v*b.h + a.g*b.g + ... + a.h*b.v` is the discarded
829        //  work this type exists to skip; it never feeds v or g.)
830        let a = self;
831        let b = o;
832        let mut g = [0.0; K];
833        for i in 0..K {
834            g[i] = a.v * b.g[i] + a.g[i] * b.v;
835        }
836        Order1 { v: a.v * b.v, g }
837    }
838    fn neg(&self) -> Self {
839        // Order2::neg == self.0.scale(-1.0).
840        self.scale(-1.0)
841    }
842    fn scale(&self, s: f64) -> Self {
843        // Tower2::scale: out.v *= s; out.g[i] *= s (same float order).
844        let mut g = self.g;
845        for i in 0..K {
846            g[i] *= s;
847        }
848        Order1 { v: self.v * s, g }
849    }
850    fn compose_unary(&self, d: [f64; 5]) -> Self {
851        // Faà di Bruno truncated to order ≤ 1 (matches `faa_di_bruno` /
852        // `Tower2::compose_unary` for the value and gradient channels):
853        //   value channel (m=0): d[0].
854        //   grad channel (positions=[i], single partition {{0}}): d[1]·g[i].
855        // Order-≤1 reads only d[0], d[1]; trailing stack entries are unused.
856        let mut g = [0.0; K];
857        for i in 0..K {
858            g[i] = d[1] * self.g[i];
859        }
860        Order1 { v: d[0], g }
861    }
862}
863
864// ── OneSeed<K>: one-seed directional, contracted third (doc §A.2) ───────
865
866/// One-seed directional scalar: an [`Order2`] base plus ONE nilpotent ε
867/// (`ε² = 0`) whose coefficient is itself an [`Order2`].
868///
869/// A scalar is `s = base + ε·eps`. Arithmetic is the `ε² = 0` truncation of the
870/// product (doc §A.2): the base parts multiply as ordinary [`Order2`] products,
871/// and the ε-coefficient picks up `a.base·b.eps + a.eps·b.base`. Composition
872/// pushes ε through one extra outer derivative.
873///
874/// Seed each primary with [`seed_direction`](Self::seed_direction): the base is
875/// the usual seeded variable (carrying `e_a` for the Hessian channel) and the
876/// ε-coefficient is the FIXED contraction direction `u_a` (a constant). Then the
877/// ε-component of the evaluated Hessian channel is the contracted third
878/// `[eps.h][a][b] = Σ_c ℓ_{abc} u_c` — exactly `row_third_contracted(dir = u)`,
879/// without materialising `t3`.
880#[derive(Clone, Copy, Debug)]
881pub struct OneSeed<const K: usize> {
882    /// The `ε⁰` part: value / gradient / Hessian of `ℓ`.
883    pub base: Order2<K>,
884    /// The `ε¹` part: value / gradient / Hessian of the ε-coefficient. After a
885    /// `seed_direction(u)` evaluation, `eps.h[a][b] = Σ_c ℓ_{abc} u_c`.
886    pub eps: Order2<K>,
887}
888
889impl<const K: usize> OneSeed<K> {
890    /// Seed primary `axis` at value `x` with ε-direction component `u_axis`:
891    /// `p_axis = p_axis⁰ + x-seed + ε·u_axis`, i.e. base = `variable(x, axis)`
892    /// and eps = `constant(u_axis)` (doc §A.2 "Seeding").
893    pub fn seed_direction(x: f64, axis: usize, u_axis: f64) -> Self {
894        OneSeed {
895            base: Order2::variable(x, axis),
896            eps: Order2::constant(u_axis),
897        }
898    }
899
900    /// The contracted-third channel after a `seed_direction(u)` evaluation:
901    /// `out[a][b] = Σ_c ℓ_{abc} u_c`, i.e. the ε-coefficient's Hessian (doc §A.2).
902    pub fn contracted_third(&self) -> [[f64; K]; K] {
903        self.eps.h()
904    }
905}
906
907impl<const K: usize> JetScalar<K> for OneSeed<K> {
908    fn constant(c: f64) -> Self {
909        OneSeed {
910            base: Order2::constant(c),
911            eps: Order2::constant(0.0),
912        }
913    }
914    fn variable(x: f64, axis: usize) -> Self {
915        // No ε-direction unless seeded via `seed_direction`.
916        OneSeed {
917            base: Order2::variable(x, axis),
918            eps: Order2::constant(0.0),
919        }
920    }
921    fn value(&self) -> f64 {
922        self.base.value()
923    }
924    fn add(&self, o: &Self) -> Self {
925        OneSeed {
926            base: self.base.add(&o.base),
927            eps: self.eps.add(&o.eps),
928        }
929    }
930    fn sub(&self, o: &Self) -> Self {
931        OneSeed {
932            base: self.base.sub(&o.base),
933            eps: self.eps.sub(&o.eps),
934        }
935    }
936    fn mul(&self, o: &Self) -> Self {
937        // (a.base + ε a.eps)(b.base + ε b.eps), dropping ε².
938        OneSeed {
939            base: self.base.mul(&o.base),
940            eps: self.base.mul(&o.eps).add(&self.eps.mul(&o.base)),
941        }
942    }
943    fn neg(&self) -> Self {
944        OneSeed {
945            base: self.base.neg(),
946            eps: self.eps.neg(),
947        }
948    }
949    fn scale(&self, s: f64) -> Self {
950        OneSeed {
951            base: self.base.scale(s),
952            eps: self.eps.scale(s),
953        }
954    }
955    fn compose_unary(&self, d: [f64; 5]) -> Self {
956        // f(base + ε eps) = f(base) + ε · f'(base)·eps  (ε² = 0). Each factor is
957        // an Order2 composition: the base composes with the f-stack, and the
958        // ε-coefficient is the Order2 of the SHIFTED stack (the chain rule
959        // `f'(base)` as an Order2) times eps. Order2 reads only the leading
960        // three entries of whatever stack it is handed, so the trailing slots
961        // are unused padding (the fixed-length array makes the windowing total).
962        let base = self.base.compose_unary([d[0], d[1], d[2], d[3], d[4]]);
963        // f'(base) as an Order2 (consumes [f', f'', f''']).
964        let fprime = self.base.compose_unary([d[1], d[2], d[3], d[4], d[4]]);
965        let eps = fprime.mul(&self.eps);
966        OneSeed { base, eps }
967    }
968}
969
970// ── OneSeedLane<L, K>: lane-batched one-seed directional (doc §A.2) ──────
971
972/// Lane-batched [`OneSeed`]: the same one-seed directional scalar with its two
973/// [`Order2`] parts re-typed to [`Order2Lane<L, K>`], so one `L = f64x4`
974/// instance carries FOUR rows' contracted-third evaluations per vector pass.
975///
976/// Every operation (`add`/`sub`/`mul`/`neg`/`scale`/`compose_unary` and the
977/// transcendentals) is a term-for-term structural re-type of the scalar
978/// [`OneSeed`] ops onto the lane-implemented [`Order2Lane`] algebra. With
979/// `L = f64`, `OneSeedLane<f64, K>` is `to_bits`-identical to [`OneSeed<K>`];
980/// with `L = f64x4`, lane `i` is `to_bits`-identical to that (see `batch_tests`).
981#[derive(Clone, Copy, Debug)]
982pub struct OneSeedLane<L: Lane, const K: usize> {
983    /// The `ε⁰` part (lane-batched value / gradient / Hessian of `ℓ`).
984    pub base: Order2Lane<L, K>,
985    /// The `ε¹` part. After a `seed_direction(u)` evaluation,
986    /// `eps.h[a][b]` lane `i` is row `i`'s `Σ_c ℓ_{abc} u_c`.
987    pub eps: Order2Lane<L, K>,
988}
989
990/// The 4-rows-per-pass batched one-seed scalar (`wide::f64x4` lanes).
991pub type OneSeedBatch<const K: usize> = OneSeedLane<wide::f64x4, K>;
992
993impl<L: Lane, const K: usize> OneSeedLane<L, K> {
994    /// A constant: base = `constant(c)`, ε-part zero (mirrors [`OneSeed::constant`]).
995    #[inline]
996    pub fn constant(c: L) -> Self {
997        OneSeedLane {
998            base: Order2Lane::constant(c),
999            eps: Order2Lane::constant(L::splat(0.0)),
1000        }
1001    }
1002
1003    /// The seeded variable `p_axis` at (per-lane) value `value`, no ε-direction
1004    /// (mirrors [`OneSeed::variable`]).
1005    #[inline]
1006    pub fn variable(value: L, axis: usize) -> Self {
1007        OneSeedLane {
1008            base: Order2Lane::variable(value, axis),
1009            eps: Order2Lane::constant(L::splat(0.0)),
1010        }
1011    }
1012
1013    /// Seed primary `axis` at (per-lane) value `value` with ε-direction
1014    /// component `u_axis`: base = `variable(value, axis)`, eps = `constant(u_axis)`
1015    /// (mirrors [`OneSeed::seed_direction`]). With `L = f64x4`, `value` / `u_axis`
1016    /// pack the four rows' values / directions of primary `axis`.
1017    #[inline]
1018    pub fn seed_direction(value: L, axis: usize, u_axis: L) -> Self {
1019        OneSeedLane {
1020            base: Order2Lane::variable(value, axis),
1021            eps: Order2Lane::constant(u_axis),
1022        }
1023    }
1024
1025    /// The contracted-third channel after a `seed_direction(u)` evaluation:
1026    /// `out[a][b]` lane `i` is row `i`'s `Σ_c ℓ_{abc} u_c` (the ε-part Hessian).
1027    #[inline]
1028    #[must_use]
1029    pub fn contracted_third(&self) -> [[L; K]; K] {
1030        self.eps.h
1031    }
1032
1033    /// Lane-wise `self + o` (mirrors [`OneSeed::add`]).
1034    #[inline]
1035    pub fn add(&self, o: &Self) -> Self {
1036        OneSeedLane {
1037            base: self.base.add(&o.base),
1038            eps: self.eps.add(&o.eps),
1039        }
1040    }
1041
1042    /// Lane-wise `self - o` (mirrors [`OneSeed::sub`]).
1043    #[inline]
1044    pub fn sub(&self, o: &Self) -> Self {
1045        OneSeedLane {
1046            base: self.base.sub(&o.base),
1047            eps: self.eps.sub(&o.eps),
1048        }
1049    }
1050
1051    /// Lane-wise `self · o`, ε² = 0 truncation (mirrors [`OneSeed::mul`]).
1052    #[inline]
1053    pub fn mul(&self, o: &Self) -> Self {
1054        OneSeedLane {
1055            base: self.base.mul(&o.base),
1056            eps: self.base.mul(&o.eps).add(&self.eps.mul(&o.base)),
1057        }
1058    }
1059
1060    /// Negate every part (mirrors [`OneSeed::neg`]).
1061    #[inline]
1062    pub fn neg(&self) -> Self {
1063        OneSeedLane {
1064            base: self.base.neg(),
1065            eps: self.eps.neg(),
1066        }
1067    }
1068
1069    /// Multiply every part by the plain scalar `s` (mirrors [`OneSeed::scale`]).
1070    #[inline]
1071    pub fn scale(&self, s: f64) -> Self {
1072        OneSeedLane {
1073            base: self.base.scale(s),
1074            eps: self.eps.scale(s),
1075        }
1076    }
1077
1078    /// Exact order-≤2-per-part Faà di Bruno composition `f ∘ self`, given the
1079    /// per-lane outer-derivative stack `d = [f, f′, f″, f‴, f⁗]`. Term-for-term
1080    /// identical to [`OneSeed::compose_unary`]: the base reads `d[0..=2]` and the
1081    /// ε-coefficient is `f′(base)` (reads `d[1..=3]`) times `eps`.
1082    #[inline]
1083    pub fn compose_unary(&self, d: [L; 5]) -> Self {
1084        let base = self.base.compose_unary([d[0], d[1], d[2]]);
1085        let fprime = self.base.compose_unary([d[1], d[2], d[3]]);
1086        let eps = fprime.mul(&self.eps);
1087        OneSeedLane { base, eps }
1088    }
1089
1090    /// `e^self`, per-lane stack `[e, e, e, e, e]` (matches [`JetScalar::exp`]).
1091    #[inline]
1092    pub fn exp(&self) -> Self {
1093        let d = self.base.v.unary5(|u| {
1094            let e = u.exp();
1095            [e, e, e, e, e]
1096        });
1097        self.compose_unary(d)
1098    }
1099
1100    /// `ln(self)`; caller guarantees positivity (matches [`JetScalar::ln`]).
1101    #[inline]
1102    pub fn ln(&self) -> Self {
1103        let d = self.base.v.unary5(|u| {
1104            let r = 1.0 / u;
1105            [u.ln(), r, -r * r, 2.0 * r * r * r, -6.0 * r * r * r * r]
1106        });
1107        self.compose_unary(d)
1108    }
1109
1110    /// `√self`; caller guarantees positivity (matches [`JetScalar::sqrt`]).
1111    #[inline]
1112    pub fn sqrt(&self) -> Self {
1113        let d = self.base.v.unary5(|u| {
1114            let s = u.sqrt();
1115            [
1116                s,
1117                0.5 / s,
1118                -0.25 / (u * s),
1119                0.375 / (u * u * s),
1120                -0.9375 / (u * u * u * s),
1121            ]
1122        });
1123        self.compose_unary(d)
1124    }
1125
1126    /// `1/self` (matches [`JetScalar::recip`]).
1127    #[inline]
1128    pub fn recip(&self) -> Self {
1129        let d = self.base.v.unary5(|u| {
1130            let r = 1.0 / u;
1131            let r2 = r * r;
1132            [r, -r2, 2.0 * r2 * r, -6.0 * r2 * r2, 24.0 * r2 * r2 * r]
1133        });
1134        self.compose_unary(d)
1135    }
1136
1137    /// `self^a` for real `a`; caller guarantees a positive base (matches
1138    /// [`JetScalar::powf`]).
1139    #[inline]
1140    pub fn powf(&self, a: f64) -> Self {
1141        let d = self.base.v.unary5(|u| {
1142            [
1143                u.powf(a),
1144                a * u.powf(a - 1.0),
1145                a * (a - 1.0) * u.powf(a - 2.0),
1146                a * (a - 1.0) * (a - 2.0) * u.powf(a - 3.0),
1147                a * (a - 1.0) * (a - 2.0) * (a - 3.0) * u.powf(a - 4.0),
1148            ]
1149        });
1150        self.compose_unary(d)
1151    }
1152
1153    /// `ln Γ(self)`; caller guarantees positivity (matches [`JetScalar::ln_gamma`],
1154    /// same hand-certified stack).
1155    #[inline]
1156    pub fn ln_gamma(&self) -> Self {
1157        let d = self
1158            .base
1159            .v
1160            .unary5(crate::jet_tower::ln_gamma_derivative_stack);
1161        self.compose_unary(d)
1162    }
1163
1164    /// `ψ(self)` digamma; caller guarantees positivity (matches
1165    /// [`JetScalar::digamma`], same hand-certified stack).
1166    #[inline]
1167    pub fn digamma(&self) -> Self {
1168        let d = self
1169            .base
1170            .v
1171            .unary5(crate::jet_tower::digamma_derivative_stack);
1172        self.compose_unary(d)
1173    }
1174}
1175
1176impl<const K: usize> OneSeedBatch<K> {
1177    /// Extract lane `i`'s parts as a production [`OneSeed<K>`]. Lane `i` is
1178    /// `to_bits`-identical to evaluating the same program at [`OneSeed<K>`] on
1179    /// row `i` (see `batch_tests`).
1180    #[inline]
1181    #[must_use]
1182    pub fn lane(&self, i: usize) -> OneSeed<K> {
1183        OneSeed {
1184            base: self.base.lane(i),
1185            eps: self.eps.lane(i),
1186        }
1187    }
1188}
1189
1190// ── TwoSeed<K>: two-seed, contracted fourth (doc §A.3) ──────────────────
1191
1192/// Two-seed scalar: an [`Order2`] base plus TWO nilpotents ε, δ
1193/// (`ε² = δ² = 0`, `εδ` retained) — four [`Order2`] parts
1194/// `s = base + ε·eps + δ·del + εδ·eps_del`.
1195///
1196/// Product truncates `ε² = δ² = 0` (doc §A.3): each part is built from
1197/// [`Order2`] products of the four input parts. Composition picks up
1198/// successively higher outer derivatives, the cross part carrying the second
1199/// Faà di Bruno term `f''·eps·del + f'·eps_del`.
1200///
1201/// Seed each primary with [`seed`](Self::seed): base = `variable(x, axis)`,
1202/// eps = `constant(u_axis)`, del = `constant(v_axis)`, eps_del = `constant(0)`.
1203/// Then the εδ-component of the evaluated Hessian channel is the contracted
1204/// fourth `[eps_del.h][a][b] = Σ_{cd} ℓ_{abcd} u_c v_d` — exactly
1205/// `row_fourth_contracted(u, v)`, without materialising `t4`.
1206#[derive(Clone, Copy, Debug)]
1207pub struct TwoSeed<const K: usize> {
1208    /// The `ε⁰δ⁰` part: value / grad / Hessian of `ℓ`.
1209    pub base: Order2<K>,
1210    /// The `ε¹δ⁰` part.
1211    pub eps: Order2<K>,
1212    /// The `ε⁰δ¹` part.
1213    pub del: Order2<K>,
1214    /// The `ε¹δ¹` part. After a `seed(u, v)` evaluation,
1215    /// `eps_del.h[a][b] = Σ_{cd} ℓ_{abcd} u_c v_d`.
1216    pub eps_del: Order2<K>,
1217}
1218
1219impl<const K: usize> TwoSeed<K> {
1220    /// Seed primary `axis` at value `x` with ε-direction `u_axis` and
1221    /// δ-direction `v_axis`:
1222    /// `p_axis = p_axis⁰ + x-seed + ε·u_axis + δ·v_axis` (doc §A.3 "Seeding").
1223    pub fn seed(x: f64, axis: usize, u_axis: f64, v_axis: f64) -> Self {
1224        TwoSeed {
1225            base: Order2::variable(x, axis),
1226            eps: Order2::constant(u_axis),
1227            del: Order2::constant(v_axis),
1228            eps_del: Order2::constant(0.0),
1229        }
1230    }
1231
1232    /// The contracted-fourth channel after a `seed(u, v)` evaluation:
1233    /// `out[a][b] = Σ_{cd} ℓ_{abcd} u_c v_d`, i.e. the εδ-coefficient's Hessian.
1234    pub fn contracted_fourth(&self) -> [[f64; K]; K] {
1235        self.eps_del.h()
1236    }
1237}
1238
1239impl<const K: usize> JetScalar<K> for TwoSeed<K> {
1240    fn constant(c: f64) -> Self {
1241        TwoSeed {
1242            base: Order2::constant(c),
1243            eps: Order2::constant(0.0),
1244            del: Order2::constant(0.0),
1245            eps_del: Order2::constant(0.0),
1246        }
1247    }
1248    fn variable(x: f64, axis: usize) -> Self {
1249        TwoSeed {
1250            base: Order2::variable(x, axis),
1251            eps: Order2::constant(0.0),
1252            del: Order2::constant(0.0),
1253            eps_del: Order2::constant(0.0),
1254        }
1255    }
1256    fn value(&self) -> f64 {
1257        self.base.value()
1258    }
1259    fn add(&self, o: &Self) -> Self {
1260        TwoSeed {
1261            base: self.base.add(&o.base),
1262            eps: self.eps.add(&o.eps),
1263            del: self.del.add(&o.del),
1264            eps_del: self.eps_del.add(&o.eps_del),
1265        }
1266    }
1267    fn sub(&self, o: &Self) -> Self {
1268        TwoSeed {
1269            base: self.base.sub(&o.base),
1270            eps: self.eps.sub(&o.eps),
1271            del: self.del.sub(&o.del),
1272            eps_del: self.eps_del.sub(&o.eps_del),
1273        }
1274    }
1275    fn mul(&self, o: &Self) -> Self {
1276        let a = self;
1277        let b = o;
1278        // Truncate ε² = δ² = 0 (doc §A.3 product table).
1279        let base = a.base.mul(&b.base);
1280        let eps = a.base.mul(&b.eps).add(&a.eps.mul(&b.base));
1281        let del = a.base.mul(&b.del).add(&a.del.mul(&b.base));
1282        let eps_del = a
1283            .base
1284            .mul(&b.eps_del)
1285            .add(&a.eps.mul(&b.del))
1286            .add(&a.del.mul(&b.eps))
1287            .add(&a.eps_del.mul(&b.base));
1288        TwoSeed {
1289            base,
1290            eps,
1291            del,
1292            eps_del,
1293        }
1294    }
1295    fn neg(&self) -> Self {
1296        TwoSeed {
1297            base: self.base.neg(),
1298            eps: self.eps.neg(),
1299            del: self.del.neg(),
1300            eps_del: self.eps_del.neg(),
1301        }
1302    }
1303    fn scale(&self, s: f64) -> Self {
1304        TwoSeed {
1305            base: self.base.scale(s),
1306            eps: self.eps.scale(s),
1307            del: self.del.scale(s),
1308            eps_del: self.eps_del.scale(s),
1309        }
1310    }
1311    fn compose_unary(&self, d: [f64; 5]) -> Self {
1312        // f(s) with s = base + ε eps + δ del + εδ eps_del, ε²=δ²=0:
1313        //   f(s) = f(base)
1314        //        + ε · f'(base)·eps
1315        //        + δ · f'(base)·del
1316        //        + εδ · ( f''(base)·eps·del + f'(base)·eps_del ).
1317        // Each f^{(r)}(base) is the Order2 composition of base with the stack
1318        // shifted r entries (doc §A.3 composition). Order2 reads only the
1319        // leading three entries of whatever stack it is handed, so the trailing
1320        // padding slots are unused (the fixed-length array makes this total).
1321        let base = self.base.compose_unary([d[0], d[1], d[2], d[3], d[4]]);
1322        let fprime = self.base.compose_unary([d[1], d[2], d[3], d[4], d[4]]); // f'(base) as Order2
1323        let fsecond = self.base.compose_unary([d[2], d[3], d[4], d[4], d[4]]); // f''(base) as Order2
1324        let eps = fprime.mul(&self.eps);
1325        let del = fprime.mul(&self.del);
1326        let eps_del = fsecond
1327            .mul(&self.eps)
1328            .mul(&self.del)
1329            .add(&fprime.mul(&self.eps_del));
1330        TwoSeed {
1331            base,
1332            eps,
1333            del,
1334            eps_del,
1335        }
1336    }
1337}
1338
1339// ── TwoSeedLane<L, K>: lane-batched two-seed, contracted fourth (doc §A.3) ─
1340
1341/// Lane-batched [`TwoSeed`]: the same two-seed scalar with its four [`Order2`]
1342/// parts re-typed to [`Order2Lane<L, K>`], so one `L = f64x4` instance carries
1343/// FOUR rows' contracted-fourth evaluations per vector pass.
1344///
1345/// Every operation is a term-for-term structural re-type of the scalar
1346/// [`TwoSeed`] ops onto the lane-implemented [`Order2Lane`] algebra. With
1347/// `L = f64`, `TwoSeedLane<f64, K>` is `to_bits`-identical to [`TwoSeed<K>`];
1348/// with `L = f64x4`, lane `i` is `to_bits`-identical to that (see `batch_tests`).
1349#[derive(Clone, Copy, Debug)]
1350pub struct TwoSeedLane<L: Lane, const K: usize> {
1351    /// The `ε⁰δ⁰` part.
1352    pub base: Order2Lane<L, K>,
1353    /// The `ε¹δ⁰` part.
1354    pub eps: Order2Lane<L, K>,
1355    /// The `ε⁰δ¹` part.
1356    pub del: Order2Lane<L, K>,
1357    /// The `ε¹δ¹` part. After a `seed(u, v)` evaluation, `eps_del.h[a][b]`
1358    /// lane `i` is row `i`'s `Σ_{cd} ℓ_{abcd} u_c v_d`.
1359    pub eps_del: Order2Lane<L, K>,
1360}
1361
1362/// The 4-rows-per-pass batched two-seed scalar (`wide::f64x4` lanes).
1363pub type TwoSeedBatch<const K: usize> = TwoSeedLane<wide::f64x4, K>;
1364
1365impl<L: Lane, const K: usize> TwoSeedLane<L, K> {
1366    /// A constant: base = `constant(c)`, all seed parts zero (mirrors
1367    /// [`TwoSeed::constant`]).
1368    #[inline]
1369    pub fn constant(c: L) -> Self {
1370        let z = Order2Lane::constant(L::splat(0.0));
1371        TwoSeedLane {
1372            base: Order2Lane::constant(c),
1373            eps: z,
1374            del: z,
1375            eps_del: z,
1376        }
1377    }
1378
1379    /// The seeded variable `p_axis` at (per-lane) value `value`, no ε/δ direction
1380    /// (mirrors [`TwoSeed::variable`]).
1381    #[inline]
1382    pub fn variable(value: L, axis: usize) -> Self {
1383        let z = Order2Lane::constant(L::splat(0.0));
1384        TwoSeedLane {
1385            base: Order2Lane::variable(value, axis),
1386            eps: z,
1387            del: z,
1388            eps_del: z,
1389        }
1390    }
1391
1392    /// Seed primary `axis` at (per-lane) value `value` with ε-direction `u_axis`
1393    /// and δ-direction `v_axis` (mirrors [`TwoSeed::seed`]). With `L = f64x4`,
1394    /// each argument packs the four rows' values for primary `axis`.
1395    #[inline]
1396    pub fn seed(value: L, axis: usize, u_axis: L, v_axis: L) -> Self {
1397        TwoSeedLane {
1398            base: Order2Lane::variable(value, axis),
1399            eps: Order2Lane::constant(u_axis),
1400            del: Order2Lane::constant(v_axis),
1401            eps_del: Order2Lane::constant(L::splat(0.0)),
1402        }
1403    }
1404
1405    /// The contracted-fourth channel after a `seed(u, v)` evaluation:
1406    /// `out[a][b]` lane `i` is row `i`'s `Σ_{cd} ℓ_{abcd} u_c v_d`
1407    /// (the εδ-part Hessian).
1408    #[inline]
1409    #[must_use]
1410    pub fn contracted_fourth(&self) -> [[L; K]; K] {
1411        self.eps_del.h
1412    }
1413
1414    /// Lane-wise `self + o` (mirrors [`TwoSeed::add`]).
1415    #[inline]
1416    pub fn add(&self, o: &Self) -> Self {
1417        TwoSeedLane {
1418            base: self.base.add(&o.base),
1419            eps: self.eps.add(&o.eps),
1420            del: self.del.add(&o.del),
1421            eps_del: self.eps_del.add(&o.eps_del),
1422        }
1423    }
1424
1425    /// Lane-wise `self - o` (mirrors [`TwoSeed::sub`]).
1426    #[inline]
1427    pub fn sub(&self, o: &Self) -> Self {
1428        TwoSeedLane {
1429            base: self.base.sub(&o.base),
1430            eps: self.eps.sub(&o.eps),
1431            del: self.del.sub(&o.del),
1432            eps_del: self.eps_del.sub(&o.eps_del),
1433        }
1434    }
1435
1436    /// Lane-wise `self · o`, ε² = δ² = 0 truncation (mirrors [`TwoSeed::mul`]).
1437    #[inline]
1438    pub fn mul(&self, o: &Self) -> Self {
1439        let a = self;
1440        let b = o;
1441        let base = a.base.mul(&b.base);
1442        let eps = a.base.mul(&b.eps).add(&a.eps.mul(&b.base));
1443        let del = a.base.mul(&b.del).add(&a.del.mul(&b.base));
1444        let eps_del = a
1445            .base
1446            .mul(&b.eps_del)
1447            .add(&a.eps.mul(&b.del))
1448            .add(&a.del.mul(&b.eps))
1449            .add(&a.eps_del.mul(&b.base));
1450        TwoSeedLane {
1451            base,
1452            eps,
1453            del,
1454            eps_del,
1455        }
1456    }
1457
1458    /// Negate every part (mirrors [`TwoSeed::neg`]).
1459    #[inline]
1460    pub fn neg(&self) -> Self {
1461        TwoSeedLane {
1462            base: self.base.neg(),
1463            eps: self.eps.neg(),
1464            del: self.del.neg(),
1465            eps_del: self.eps_del.neg(),
1466        }
1467    }
1468
1469    /// Multiply every part by the plain scalar `s` (mirrors [`TwoSeed::scale`]).
1470    #[inline]
1471    pub fn scale(&self, s: f64) -> Self {
1472        TwoSeedLane {
1473            base: self.base.scale(s),
1474            eps: self.eps.scale(s),
1475            del: self.del.scale(s),
1476            eps_del: self.eps_del.scale(s),
1477        }
1478    }
1479
1480    /// Exact composition `f ∘ self`, given the per-lane outer-derivative stack
1481    /// `d = [f, f′, f″, f‴, f⁗]`. Term-for-term identical to
1482    /// [`TwoSeed::compose_unary`]: base reads `d[0..=2]`, `f′(base)` reads
1483    /// `d[1..=3]`, `f″(base)` reads `d[2..=4]`, and the cross part carries
1484    /// `f″·eps·del + f′·eps_del`.
1485    #[inline]
1486    pub fn compose_unary(&self, d: [L; 5]) -> Self {
1487        let base = self.base.compose_unary([d[0], d[1], d[2]]);
1488        let fprime = self.base.compose_unary([d[1], d[2], d[3]]);
1489        let fsecond = self.base.compose_unary([d[2], d[3], d[4]]);
1490        let eps = fprime.mul(&self.eps);
1491        let del = fprime.mul(&self.del);
1492        let eps_del = fsecond
1493            .mul(&self.eps)
1494            .mul(&self.del)
1495            .add(&fprime.mul(&self.eps_del));
1496        TwoSeedLane {
1497            base,
1498            eps,
1499            del,
1500            eps_del,
1501        }
1502    }
1503
1504    /// `e^self`, per-lane stack `[e; 5]` (matches [`JetScalar::exp`]).
1505    #[inline]
1506    pub fn exp(&self) -> Self {
1507        let d = self.base.v.unary5(|u| {
1508            let e = u.exp();
1509            [e, e, e, e, e]
1510        });
1511        self.compose_unary(d)
1512    }
1513
1514    /// `ln(self)`; caller guarantees positivity (matches [`JetScalar::ln`]).
1515    #[inline]
1516    pub fn ln(&self) -> Self {
1517        let d = self.base.v.unary5(|u| {
1518            let r = 1.0 / u;
1519            [u.ln(), r, -r * r, 2.0 * r * r * r, -6.0 * r * r * r * r]
1520        });
1521        self.compose_unary(d)
1522    }
1523
1524    /// `√self`; caller guarantees positivity (matches [`JetScalar::sqrt`]).
1525    #[inline]
1526    pub fn sqrt(&self) -> Self {
1527        let d = self.base.v.unary5(|u| {
1528            let s = u.sqrt();
1529            [
1530                s,
1531                0.5 / s,
1532                -0.25 / (u * s),
1533                0.375 / (u * u * s),
1534                -0.9375 / (u * u * u * s),
1535            ]
1536        });
1537        self.compose_unary(d)
1538    }
1539
1540    /// `1/self` (matches [`JetScalar::recip`]).
1541    #[inline]
1542    pub fn recip(&self) -> Self {
1543        let d = self.base.v.unary5(|u| {
1544            let r = 1.0 / u;
1545            let r2 = r * r;
1546            [r, -r2, 2.0 * r2 * r, -6.0 * r2 * r2, 24.0 * r2 * r2 * r]
1547        });
1548        self.compose_unary(d)
1549    }
1550
1551    /// `self^a` for real `a`; caller guarantees a positive base (matches
1552    /// [`JetScalar::powf`]).
1553    #[inline]
1554    pub fn powf(&self, a: f64) -> Self {
1555        let d = self.base.v.unary5(|u| {
1556            [
1557                u.powf(a),
1558                a * u.powf(a - 1.0),
1559                a * (a - 1.0) * u.powf(a - 2.0),
1560                a * (a - 1.0) * (a - 2.0) * u.powf(a - 3.0),
1561                a * (a - 1.0) * (a - 2.0) * (a - 3.0) * u.powf(a - 4.0),
1562            ]
1563        });
1564        self.compose_unary(d)
1565    }
1566
1567    /// `ln Γ(self)`; caller guarantees positivity (matches [`JetScalar::ln_gamma`]).
1568    #[inline]
1569    pub fn ln_gamma(&self) -> Self {
1570        let d = self
1571            .base
1572            .v
1573            .unary5(crate::jet_tower::ln_gamma_derivative_stack);
1574        self.compose_unary(d)
1575    }
1576
1577    /// `ψ(self)` digamma; caller guarantees positivity (matches
1578    /// [`JetScalar::digamma`]).
1579    #[inline]
1580    pub fn digamma(&self) -> Self {
1581        let d = self
1582            .base
1583            .v
1584            .unary5(crate::jet_tower::digamma_derivative_stack);
1585        self.compose_unary(d)
1586    }
1587}
1588
1589impl<const K: usize> TwoSeedBatch<K> {
1590    /// Extract lane `i`'s parts as a production [`TwoSeed<K>`]. Lane `i` is
1591    /// `to_bits`-identical to evaluating the same program at [`TwoSeed<K>`] on
1592    /// row `i` (see `batch_tests`).
1593    #[inline]
1594    #[must_use]
1595    pub fn lane(&self, i: usize) -> TwoSeed<K> {
1596        TwoSeed {
1597            base: self.base.lane(i),
1598            eps: self.eps.lane(i),
1599            del: self.del.lane(i),
1600            eps_del: self.eps_del.lane(i),
1601        }
1602    }
1603}
1604
1605// ── Tower3<K>: value / gradient / Hessian / third tensor ────────────────
1606
1607/// The order-≤3 [`crate::jet_tower::Tower3`] is also a [`JetScalar`]. It serves
1608/// consumers that read `.t3` but never `.t4`, avoiding the fourth-tensor
1609/// product/composition work while preserving the lower channels
1610/// bit-for-bit against [`crate::jet_tower::Tower4`].
1611impl<const K: usize> JetScalar<K> for crate::jet_tower::Tower3<K> {
1612    fn constant(c: f64) -> Self {
1613        crate::jet_tower::Tower3::constant(c)
1614    }
1615    fn variable(x: f64, axis: usize) -> Self {
1616        crate::jet_tower::Tower3::variable(x, axis)
1617    }
1618    fn value(&self) -> f64 {
1619        self.v
1620    }
1621    fn add(&self, o: &Self) -> Self {
1622        *self + *o
1623    }
1624    fn sub(&self, o: &Self) -> Self {
1625        *self + o.scale(-1.0)
1626    }
1627    fn mul(&self, o: &Self) -> Self {
1628        crate::jet_tower::Tower3::mul(self, o)
1629    }
1630    fn neg(&self) -> Self {
1631        self.scale(-1.0)
1632    }
1633    fn scale(&self, s: f64) -> Self {
1634        crate::jet_tower::Tower3::scale(self, s)
1635    }
1636    fn compose_unary(&self, d: [f64; 5]) -> Self {
1637        crate::jet_tower::Tower3::compose_unary(self, [d[0], d[1], d[2], d[3]])
1638    }
1639}
1640
1641// ── Tower4<K>: full dense tower as a JetScalar (the all-channels scalar) ─
1642
1643/// The full dense [`crate::jet_tower::Tower4`] is itself a [`JetScalar`]: it
1644/// carries EVERY channel, so a row expression written ONCE against [`JetScalar`]
1645/// can be evaluated at `Tower4` to obtain the full `(v, g, H, t3, t4)` in one
1646/// pass. This is BOTH the #932 oracle ground truth the packed [`Order2`] /
1647/// [`OneSeed`] / [`TwoSeed`] scalars are pinned against, AND a production scalar:
1648/// a family whose uncontracted third / fourth derivative tensors are needed
1649/// (the BMS rigid `third_full` / `fourth_full` caches) evaluates the SAME
1650/// generic row-NLL expression at `Tower4` and reads `.t3` / `.t4` off the
1651/// result — so the dense tensors come from the single source of truth, not a
1652/// separately hand-written jet. The packed scalars serve the consumers that
1653/// need only `(v, g, H)` (`Order2`) or one / two contractions
1654/// (`OneSeed` / `TwoSeed`) without paying for the dense tensors.
1655impl<const K: usize> JetScalar<K> for crate::jet_tower::Tower4<K> {
1656    fn constant(c: f64) -> Self {
1657        crate::jet_tower::Tower4::constant(c)
1658    }
1659    fn variable(x: f64, axis: usize) -> Self {
1660        crate::jet_tower::Tower4::variable(x, axis)
1661    }
1662    fn value(&self) -> f64 {
1663        self.v
1664    }
1665    fn add(&self, o: &Self) -> Self {
1666        *self + *o
1667    }
1668    fn sub(&self, o: &Self) -> Self {
1669        *self - *o
1670    }
1671    fn mul(&self, o: &Self) -> Self {
1672        crate::jet_tower::Tower4::mul(self, o)
1673    }
1674    fn neg(&self) -> Self {
1675        self.scale(-1.0)
1676    }
1677    fn scale(&self, s: f64) -> Self {
1678        crate::jet_tower::Tower4::scale(self, s)
1679    }
1680    fn compose_unary(&self, d: [f64; 5]) -> Self {
1681        crate::jet_tower::Tower4::compose_unary(self, d)
1682    }
1683}
1684
1685#[cfg(test)]
1686mod tests {
1687    use super::*;
1688    use crate::jet_tower::{RowNllProgram, Tower4, evaluate_program};
1689
1690    /// A small polynomial-plus-unary row expression written ONCE, generically
1691    /// over `S: JetScalar<2>`, so it can be evaluated against every scalar:
1692    /// `ℓ = (e^{p0·p1} + 2) · √(p0·p0 + 1) − p1·p1·0.5`.
1693    /// Exercises mul, add/sub, scale, exp, sqrt — every algebra op.
1694    fn row_expr<S: JetScalar<2>>(p: &[S; 2]) -> S {
1695        let g = p[0].mul(&p[1]).exp();
1696        let inner = g.add(&S::constant(2.0));
1697        let radic = p[0].mul(&p[0]).add(&S::constant(1.0)).sqrt();
1698        inner.mul(&radic).sub(&p[1].mul(&p[1]).scale(0.5))
1699    }
1700
1701    /// The same expression as a Tower4 `RowNllProgram`, the ground-truth tower.
1702    struct ExprProgram {
1703        p: [f64; 2],
1704    }
1705    impl RowNllProgram<2> for ExprProgram {
1706        fn n_rows(&self) -> usize {
1707            1
1708        }
1709        fn primaries(&self, row: usize) -> Result<[f64; 2], String> {
1710            if row >= self.n_rows() {
1711                return Err(format!("ExprProgram: row {row} out of range"));
1712            }
1713            Ok(self.p)
1714        }
1715        fn row_nll(&self, row: usize, p: &[Tower4<2>; 2]) -> Result<Tower4<2>, String> {
1716            if row >= self.n_rows() {
1717                return Err(format!("ExprProgram: row {row} out of range"));
1718            }
1719            Ok(row_expr(p))
1720        }
1721    }
1722
1723    const SEED: [f64; 2] = [0.37, -0.81];
1724    const U: [f64; 2] = [0.6, -0.2];
1725    const V: [f64; 2] = [-0.4, 1.1];
1726    const TOL: f64 = 1e-10;
1727
1728    fn close(a: f64, b: f64, label: &str) {
1729        let band = TOL + TOL * a.abs().max(b.abs());
1730        assert!(
1731            (a - b).abs() <= band,
1732            "{label}: {a:+.15e} vs {b:+.15e} (band {band:.3e})"
1733        );
1734    }
1735
1736    fn tower() -> Tower4<2> {
1737        evaluate_program(&ExprProgram { p: SEED }, 0).expect("tower")
1738    }
1739
1740    /// Order2 reproduces Tower4's value/grad/Hessian channels exactly.
1741    #[test]
1742    fn order2_matches_tower_value_grad_hessian() {
1743        let t = tower();
1744        let vars: [Order2<2>; 2] = std::array::from_fn(|a| Order2::variable(SEED[a], a));
1745        let s = row_expr(&vars);
1746        close(s.value(), t.v, "value");
1747        for a in 0..2 {
1748            close(s.0.g[a], t.g[a], &format!("grad[{a}]"));
1749            for b in 0..2 {
1750                close(s.h()[a][b], t.h[a][b], &format!("hess[{a}][{b}]"));
1751            }
1752        }
1753    }
1754
1755    /// The `compose_unary_with` seam on a scalar jet is `to_bits`-identical to
1756    /// the explicit `compose_unary(stack_fn(value))` — the contract the batch
1757    /// arm (`Tower{3,4}Lane::compose_unary_with`) lane-matches. Exercised on
1758    /// [`Order2`] across `K ∈ {2,3,4,9}`, ≥ 4000 random seeded inputs.
1759    #[test]
1760    fn compose_unary_with_scalar_seam_bit_identical() {
1761        fn rand_unit(state: &mut u64) -> f64 {
1762            let mut x = *state;
1763            x ^= x << 13;
1764            x ^= x >> 7;
1765            x ^= x << 17;
1766            *state = x;
1767            2.0 * ((x >> 11) as f64 / ((1u64 << 53) as f64)) - 1.0
1768        }
1769        // A base-value-dependent finite stack standing in for a family stack.
1770        fn stack(u: f64) -> [f64; 5] {
1771            [u.sin(), u.cos(), (2.0 * u).sin(), (0.5 * u).cos(), u * u - 0.3]
1772        }
1773        fn run<const K: usize>(state: &mut u64, n: usize) -> usize {
1774            for _ in 0..n {
1775                // A non-trivial Order2<K> jet: a seeded variable pushed through a
1776                // couple of algebra ops so g/h are dense, then exercise the seam.
1777                let base = rand_unit(state);
1778                let mut s = Order2::<K>::variable(base, 0);
1779                for a in 1..K {
1780                    s = JetScalar::mul(&s, &Order2::<K>::variable(rand_unit(state), a));
1781                }
1782                let with = s.compose_unary_with(stack);
1783                let explicit = s.compose_unary(stack(s.value()));
1784                assert_eq!(with.value().to_bits(), explicit.value().to_bits(), "value");
1785                for a in 0..K {
1786                    assert_eq!(with.g()[a].to_bits(), explicit.g()[a].to_bits(), "g[{a}]");
1787                    for b in 0..K {
1788                        assert_eq!(
1789                            with.h()[a][b].to_bits(),
1790                            explicit.h()[a][b].to_bits(),
1791                            "h[{a}][{b}]"
1792                        );
1793                    }
1794                }
1795            }
1796            n
1797        }
1798        let mut st = 0x9e37_79b9_7f4a_7c15u64;
1799        let total =
1800            run::<2>(&mut st, 1100) + run::<3>(&mut st, 1100) + run::<4>(&mut st, 1100) + run::<9>(&mut st, 1100);
1801        assert_eq!(total, 4400);
1802    }
1803
1804    /// OneSeed's ε-Hessian is the contracted third Σ_c ℓ_{abc} u_c, matching
1805    /// `Tower4::third_contracted(u)`. Base channels also match the tower.
1806    #[test]
1807    fn one_seed_matches_tower_third_contracted() {
1808        let t = tower();
1809        let truth = t.third_contracted(&U);
1810        let vars: [OneSeed<2>; 2] =
1811            std::array::from_fn(|a| OneSeed::seed_direction(SEED[a], a, U[a]));
1812        let s = row_expr(&vars);
1813        // Base channels are the plain (v, g, H).
1814        close(s.value(), t.v, "value");
1815        for a in 0..2 {
1816            for b in 0..2 {
1817                close(s.base.h()[a][b], t.h[a][b], &format!("base hess[{a}][{b}]"));
1818            }
1819        }
1820        let third = s.contracted_third();
1821        for a in 0..2 {
1822            for b in 0..2 {
1823                close(third[a][b], truth[a][b], &format!("third[{a}][{b}]"));
1824            }
1825        }
1826    }
1827
1828    /// TwoSeed's εδ-Hessian is the contracted fourth Σ_{cd} ℓ_{abcd} u_c v_d,
1829    /// matching `Tower4::fourth_contracted(u, v)`. The ε / δ single-seed parts
1830    /// reproduce the two third contractions Σ_c ℓ_{abc} u_c and …v_d.
1831    #[test]
1832    fn two_seed_matches_tower_fourth_contracted() {
1833        let t = tower();
1834        let truth4 = t.fourth_contracted(&U, &V);
1835        let truth3_u = t.third_contracted(&U);
1836        let truth3_v = t.third_contracted(&V);
1837        let vars: [TwoSeed<2>; 2] = std::array::from_fn(|a| TwoSeed::seed(SEED[a], a, U[a], V[a]));
1838        let s = row_expr(&vars);
1839        close(s.value(), t.v, "value");
1840        for a in 0..2 {
1841            close(s.base.0.g[a], t.g[a], &format!("grad[{a}]"));
1842            for b in 0..2 {
1843                close(s.base.h()[a][b], t.h[a][b], &format!("base hess[{a}][{b}]"));
1844                close(
1845                    s.eps.h()[a][b],
1846                    truth3_u[a][b],
1847                    &format!("eps third_u[{a}][{b}]"),
1848                );
1849                close(
1850                    s.del.h()[a][b],
1851                    truth3_v[a][b],
1852                    &format!("del third_v[{a}][{b}]"),
1853                );
1854            }
1855        }
1856        let fourth = s.contracted_fourth();
1857        for a in 0..2 {
1858            for b in 0..2 {
1859                close(fourth[a][b], truth4[a][b], &format!("fourth[{a}][{b}]"));
1860            }
1861        }
1862    }
1863
1864    /// The generic `row_nll_generic` seam (added to Tower4's program trait
1865    /// surface) evaluates the SAME expression on each scalar and extracts the
1866    /// channel a consumer asks for, agreeing with the direct Tower4 contraction.
1867    #[test]
1868    fn generic_program_seam_matches_tower_for_every_channel() {
1869        let t = tower();
1870        // Order2 via generic seam.
1871        let o2: [Order2<2>; 2] = std::array::from_fn(|a| Order2::variable(SEED[a], a));
1872        let so2 = row_expr(&o2);
1873        close(so2.value(), t.v, "seam order2 value");
1874        // OneSeed third.
1875        let os: [OneSeed<2>; 2] =
1876            std::array::from_fn(|a| OneSeed::seed_direction(SEED[a], a, U[a]));
1877        let third = row_expr(&os).contracted_third();
1878        let truth3 = t.third_contracted(&U);
1879        for a in 0..2 {
1880            for b in 0..2 {
1881                close(third[a][b], truth3[a][b], &format!("seam third[{a}][{b}]"));
1882            }
1883        }
1884        // TwoSeed fourth.
1885        let ts: [TwoSeed<2>; 2] = std::array::from_fn(|a| TwoSeed::seed(SEED[a], a, U[a], V[a]));
1886        let fourth = row_expr(&ts).contracted_fourth();
1887        let truth4 = t.fourth_contracted(&U, &V);
1888        for a in 0..2 {
1889            for b in 0..2 {
1890                close(
1891                    fourth[a][b],
1892                    truth4[a][b],
1893                    &format!("seam fourth[{a}][{b}]"),
1894                );
1895            }
1896        }
1897    }
1898
1899    /// The (test-only) `Tower4: JetScalar` impl is the all-channels oracle scalar:
1900    /// evaluating the SAME generic `row_expr` at `S = Tower4` (through the
1901    /// `JetScalar` trait ops) must reproduce, channel-for-channel, the `Tower4`
1902    /// obtained from the `RowNllProgram` / inherent-operator path
1903    /// (`evaluate_program`). This pins that the trait impl delegates faithfully to
1904    /// the inherent `Tower4` arithmetic (so the contracted-scalar oracles above,
1905    /// which compare against `evaluate_program`'s tower, are comparing against the
1906    /// same algebra the `JetScalar` interface exposes).
1907    #[test]
1908    fn tower4_as_jetscalar_matches_program_tower_all_channels() {
1909        let t = tower();
1910        let vars: [Tower4<2>; 2] = std::array::from_fn(|a| Tower4::variable(SEED[a], a));
1911        let s = row_expr(&vars);
1912        close(s.v, t.v, "tower-jetscalar value");
1913        for a in 0..2 {
1914            close(s.g[a], t.g[a], &format!("tower-jetscalar grad[{a}]"));
1915            for b in 0..2 {
1916                close(
1917                    s.h[a][b],
1918                    t.h[a][b],
1919                    &format!("tower-jetscalar hess[{a}][{b}]"),
1920                );
1921                for c in 0..2 {
1922                    close(
1923                        s.t3[a][b][c],
1924                        t.t3[a][b][c],
1925                        &format!("tower-jetscalar t3[{a}][{b}][{c}]"),
1926                    );
1927                    for d in 0..2 {
1928                        close(
1929                            s.t4[a][b][c][d],
1930                            t.t4[a][b][c][d],
1931                            &format!("tower-jetscalar t4[{a}][{b}][{c}][{d}]"),
1932                        );
1933                    }
1934                }
1935            }
1936        }
1937    }
1938}
1939
1940#[cfg(test)]
1941mod batch_tests {
1942    //! SIMD row-batching oracle: prove [`Order2Batch<K>`] (4 rows in
1943    //! `wide::f64x4` lanes) is `to_bits`-identical, on every value/gradient/
1944    //! Hessian channel, to the production [`Order2<K>`] evaluated per row — and
1945    //! that the new scalar field [`Order2Lane<f64, K>`] is too. Composing the two
1946    //! claims, batch lane `i` reproduces the production scalar for row `i` bit
1947    //! for bit, so the 4× throughput is a free lunch (no result change).
1948
1949    use super::{
1950        JetScalar, Lane, OneSeed, OneSeedBatch, OneSeedLane, Order2, Order2Batch, Order2Lane,
1951        TwoSeed, TwoSeedBatch, TwoSeedLane,
1952    };
1953
1954    /// The ops the witness row expression needs, so ONE generic body evaluates
1955    /// at the production [`Order2<K>`], the new scalar [`Order2Lane<f64, K>`],
1956    /// and the batched [`Order2Batch<K>`].
1957    trait RowAlg<const K: usize>: Copy {
1958        fn constant(c: f64) -> Self;
1959        fn add(&self, o: &Self) -> Self;
1960        fn sub(&self, o: &Self) -> Self;
1961        fn mul(&self, o: &Self) -> Self;
1962        fn scale(&self, s: f64) -> Self;
1963        fn exp(&self) -> Self;
1964        fn sqrt(&self) -> Self;
1965        fn recip(&self) -> Self;
1966    }
1967
1968    impl<const K: usize> RowAlg<K> for Order2<K> {
1969        fn constant(c: f64) -> Self {
1970            <Self as JetScalar<K>>::constant(c)
1971        }
1972        fn add(&self, o: &Self) -> Self {
1973            JetScalar::add(self, o)
1974        }
1975        fn sub(&self, o: &Self) -> Self {
1976            JetScalar::sub(self, o)
1977        }
1978        fn mul(&self, o: &Self) -> Self {
1979            JetScalar::mul(self, o)
1980        }
1981        fn scale(&self, s: f64) -> Self {
1982            JetScalar::scale(self, s)
1983        }
1984        fn exp(&self) -> Self {
1985            JetScalar::exp(self)
1986        }
1987        fn sqrt(&self) -> Self {
1988            JetScalar::sqrt(self)
1989        }
1990        fn recip(&self) -> Self {
1991            JetScalar::recip(self)
1992        }
1993    }
1994
1995    impl<L: Lane, const K: usize> RowAlg<K> for Order2Lane<L, K> {
1996        fn constant(c: f64) -> Self {
1997            Order2Lane::constant(L::splat(c))
1998        }
1999        fn add(&self, o: &Self) -> Self {
2000            Order2Lane::add(self, o)
2001        }
2002        fn sub(&self, o: &Self) -> Self {
2003            Order2Lane::sub(self, o)
2004        }
2005        fn mul(&self, o: &Self) -> Self {
2006            Order2Lane::mul(self, o)
2007        }
2008        fn scale(&self, s: f64) -> Self {
2009            Order2Lane::scale(self, s)
2010        }
2011        fn exp(&self) -> Self {
2012            Order2Lane::exp(self)
2013        }
2014        fn sqrt(&self) -> Self {
2015            Order2Lane::sqrt(self)
2016        }
2017        fn recip(&self) -> Self {
2018            Order2Lane::recip(self)
2019        }
2020    }
2021
2022    /// A dense witness row expression touching every algebra op (mul, add, sub,
2023    /// scale, exp, sqrt, recip) over ALL `K` primaries, so the gradient and the
2024    /// full `K×K` Hessian are dense (no trivially-zero channel). All transcend.
2025    /// arguments are kept finite/positive: `sqrt(s²+1) > 0`, `recip(exp+2) > 0`.
2026    fn row_expr<const K: usize, A: RowAlg<K>>(p: &[A; K]) -> A {
2027        let mut s = A::constant(0.3);
2028        for a in 0..K {
2029            let b = (a + 1) % K;
2030            s = s.add(&p[a].mul(&p[b]).scale(0.1 + 0.05 * a as f64));
2031        }
2032        let e = s.exp();
2033        let r = s.mul(&s).add(&A::constant(1.0)).sqrt();
2034        let denom = e.add(&A::constant(2.0));
2035        e.mul(&r).sub(&s.scale(0.5)).mul(&denom.recip())
2036    }
2037
2038    /// xorshift64 → `f64` in `[-1, 1)`.
2039    fn rand_unit(state: &mut u64) -> f64 {
2040        let mut x = *state;
2041        x ^= x << 13;
2042        x ^= x >> 7;
2043        x ^= x << 17;
2044        *state = x;
2045        let u = (x >> 11) as f64 / ((1u64 << 53) as f64); // [0, 1)
2046        2.0 * u - 1.0
2047    }
2048
2049    /// Returns the number of (batch, row) pairs whose every channel was
2050    /// verified bit-identical, so the caller can assert the expected total ran.
2051    fn check_k<const K: usize>(state: &mut u64, batches: usize) -> usize {
2052        let mut verified_rows = 0usize;
2053        for _ in 0..batches {
2054            // Four independent rows of K primary values.
2055            let rows: [[f64; K]; 4] =
2056                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2057
2058            // Production ground truth, evaluated per row at Order2<K>.
2059            let prod: [Order2<K>; 4] = std::array::from_fn(|r| {
2060                let p: [Order2<K>; K] = std::array::from_fn(|a| Order2::variable(rows[r][a], a));
2061                row_expr(&p)
2062            });
2063
2064            // New scalar field (Order2Lane<f64>), per row.
2065            let scal: [Order2Lane<f64, K>; 4] = std::array::from_fn(|r| {
2066                let p: [Order2Lane<f64, K>; K] =
2067                    std::array::from_fn(|a| Order2Lane::variable(rows[r][a], a));
2068                row_expr(&p)
2069            });
2070
2071            // Batched: 4 rows packed into f64x4 lanes, ONE vector pass.
2072            let pbatch: [Order2Batch<K>; K] = std::array::from_fn(|a| {
2073                let packed =
2074                    wide::f64x4::new([rows[0][a], rows[1][a], rows[2][a], rows[3][a]]);
2075                Order2Batch::variable(packed, a)
2076            });
2077            let batch = row_expr(&pbatch);
2078
2079            for r in 0..4 {
2080                let g = prod[r].0;
2081                // Order2Lane<f64> == Order2<K> (bit-identical scalar field).
2082                assert_eq!(scal[r].v.to_bits(), g.v.to_bits(), "K={K} scalar v");
2083                // Batch lane r == Order2<K> for row r.
2084                let lr = batch.lane(r).0;
2085                assert_eq!(lr.v.to_bits(), g.v.to_bits(), "K={K} batch lane {r} v");
2086                for a in 0..K {
2087                    assert_eq!(
2088                        scal[r].g[a].to_bits(),
2089                        g.g[a].to_bits(),
2090                        "K={K} scalar g[{a}]"
2091                    );
2092                    assert_eq!(
2093                        lr.g[a].to_bits(),
2094                        g.g[a].to_bits(),
2095                        "K={K} batch lane {r} g[{a}]"
2096                    );
2097                    for b in 0..K {
2098                        assert_eq!(
2099                            scal[r].h[a][b].to_bits(),
2100                            g.h[a][b].to_bits(),
2101                            "K={K} scalar h[{a}][{b}]"
2102                        );
2103                        assert_eq!(
2104                            lr.h[a][b].to_bits(),
2105                            g.h[a][b].to_bits(),
2106                            "K={K} batch lane {r} h[{a}][{b}]"
2107                        );
2108                    }
2109                }
2110                verified_rows += 1;
2111            }
2112        }
2113        verified_rows
2114    }
2115
2116    /// ≥2000 random 4-row batches per K, across K ∈ {2,3,4,9}: every channel of
2117    /// every lane is `to_bits`-identical to the production scalar per row.
2118    #[test]
2119    fn batch_lanes_bit_identical_to_scalar_per_row() {
2120        let mut state = 0x9E37_79B9_7F4A_7C15_u64;
2121        let mut verified = 0usize;
2122        verified += check_k::<2>(&mut state, 2000);
2123        verified += check_k::<3>(&mut state, 2000);
2124        verified += check_k::<4>(&mut state, 2000);
2125        verified += check_k::<9>(&mut state, 2000);
2126        // 4 K-values × 2000 batches × 4 packed rows each, all bit-identical.
2127        assert_eq!(verified, 4 * 2000 * 4, "every batch row must be verified");
2128    }
2129
2130    // ── One-/two-seed lane oracles ──────────────────────────────────────────
2131    //
2132    // The same dense `row_expr` witness program runs over the SEEDED directional
2133    // scalars: the scalar `OneSeed`/`TwoSeed` per row, the `f64`-lane re-type
2134    // (`*SeedLane<f64>`), and the 4-rows-per-pass batch (`*SeedBatch`). The
2135    // headline claim is that the contracted-third / contracted-fourth channel of
2136    // every lane is `to_bits`-identical to the production scalar's per row.
2137
2138    impl<const K: usize> RowAlg<K> for OneSeed<K> {
2139        fn constant(c: f64) -> Self {
2140            <Self as JetScalar<K>>::constant(c)
2141        }
2142        fn add(&self, o: &Self) -> Self {
2143            JetScalar::add(self, o)
2144        }
2145        fn sub(&self, o: &Self) -> Self {
2146            JetScalar::sub(self, o)
2147        }
2148        fn mul(&self, o: &Self) -> Self {
2149            JetScalar::mul(self, o)
2150        }
2151        fn scale(&self, s: f64) -> Self {
2152            JetScalar::scale(self, s)
2153        }
2154        fn exp(&self) -> Self {
2155            JetScalar::exp(self)
2156        }
2157        fn sqrt(&self) -> Self {
2158            JetScalar::sqrt(self)
2159        }
2160        fn recip(&self) -> Self {
2161            JetScalar::recip(self)
2162        }
2163    }
2164
2165    impl<L: Lane, const K: usize> RowAlg<K> for OneSeedLane<L, K> {
2166        fn constant(c: f64) -> Self {
2167            OneSeedLane::constant(L::splat(c))
2168        }
2169        fn add(&self, o: &Self) -> Self {
2170            OneSeedLane::add(self, o)
2171        }
2172        fn sub(&self, o: &Self) -> Self {
2173            OneSeedLane::sub(self, o)
2174        }
2175        fn mul(&self, o: &Self) -> Self {
2176            OneSeedLane::mul(self, o)
2177        }
2178        fn scale(&self, s: f64) -> Self {
2179            OneSeedLane::scale(self, s)
2180        }
2181        fn exp(&self) -> Self {
2182            OneSeedLane::exp(self)
2183        }
2184        fn sqrt(&self) -> Self {
2185            OneSeedLane::sqrt(self)
2186        }
2187        fn recip(&self) -> Self {
2188            OneSeedLane::recip(self)
2189        }
2190    }
2191
2192    impl<const K: usize> RowAlg<K> for TwoSeed<K> {
2193        fn constant(c: f64) -> Self {
2194            <Self as JetScalar<K>>::constant(c)
2195        }
2196        fn add(&self, o: &Self) -> Self {
2197            JetScalar::add(self, o)
2198        }
2199        fn sub(&self, o: &Self) -> Self {
2200            JetScalar::sub(self, o)
2201        }
2202        fn mul(&self, o: &Self) -> Self {
2203            JetScalar::mul(self, o)
2204        }
2205        fn scale(&self, s: f64) -> Self {
2206            JetScalar::scale(self, s)
2207        }
2208        fn exp(&self) -> Self {
2209            JetScalar::exp(self)
2210        }
2211        fn sqrt(&self) -> Self {
2212            JetScalar::sqrt(self)
2213        }
2214        fn recip(&self) -> Self {
2215            JetScalar::recip(self)
2216        }
2217    }
2218
2219    impl<L: Lane, const K: usize> RowAlg<K> for TwoSeedLane<L, K> {
2220        fn constant(c: f64) -> Self {
2221            TwoSeedLane::constant(L::splat(c))
2222        }
2223        fn add(&self, o: &Self) -> Self {
2224            TwoSeedLane::add(self, o)
2225        }
2226        fn sub(&self, o: &Self) -> Self {
2227            TwoSeedLane::sub(self, o)
2228        }
2229        fn mul(&self, o: &Self) -> Self {
2230            TwoSeedLane::mul(self, o)
2231        }
2232        fn scale(&self, s: f64) -> Self {
2233            TwoSeedLane::scale(self, s)
2234        }
2235        fn exp(&self) -> Self {
2236            TwoSeedLane::exp(self)
2237        }
2238        fn sqrt(&self) -> Self {
2239            TwoSeedLane::sqrt(self)
2240        }
2241        fn recip(&self) -> Self {
2242            TwoSeedLane::recip(self)
2243        }
2244    }
2245
2246    fn check_oneseed<const K: usize>(state: &mut u64, batches: usize) -> usize {
2247        let mut rows_checked = 0;
2248        for _ in 0..batches {
2249            let rows: [[f64; K]; 4] =
2250                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2251            // Per-row ε-direction.
2252            let u: [[f64; K]; 4] =
2253                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2254
2255            // Production ground truth (scalar OneSeed per row).
2256            let prod: [OneSeed<K>; 4] = std::array::from_fn(|r| {
2257                let p: [OneSeed<K>; K] =
2258                    std::array::from_fn(|a| OneSeed::seed_direction(rows[r][a], a, u[r][a]));
2259                row_expr(&p)
2260            });
2261
2262            // f64-lane re-type per row.
2263            let scal: [OneSeedLane<f64, K>; 4] = std::array::from_fn(|r| {
2264                let p: [OneSeedLane<f64, K>; K] =
2265                    std::array::from_fn(|a| OneSeedLane::seed_direction(rows[r][a], a, u[r][a]));
2266                row_expr(&p)
2267            });
2268
2269            // 4-rows-per-pass batch.
2270            let pbatch: [OneSeedBatch<K>; K] = std::array::from_fn(|a| {
2271                let val = wide::f64x4::new([rows[0][a], rows[1][a], rows[2][a], rows[3][a]]);
2272                let uu = wide::f64x4::new([u[0][a], u[1][a], u[2][a], u[3][a]]);
2273                OneSeedBatch::seed_direction(val, a, uu)
2274            });
2275            let batch = row_expr(&pbatch);
2276
2277            for r in 0..4 {
2278                let want = prod[r].contracted_third();
2279                let got_scal = scal[r].contracted_third();
2280                let got_batch = batch.lane(r).contracted_third();
2281                // Value channel too (sanity that the base program agrees).
2282                assert_eq!(
2283                    scal[r].base.v.to_bits(),
2284                    prod[r].base.value().to_bits(),
2285                    "OneSeed K={K} scalar value"
2286                );
2287                assert_eq!(
2288                    batch.lane(r).base.value().to_bits(),
2289                    prod[r].base.value().to_bits(),
2290                    "OneSeed K={K} batch lane {r} value"
2291                );
2292                for a in 0..K {
2293                    for b in 0..K {
2294                        assert_eq!(
2295                            got_scal[a][b].to_bits(),
2296                            want[a][b].to_bits(),
2297                            "OneSeed K={K} scalar third[{a}][{b}]"
2298                        );
2299                        assert_eq!(
2300                            got_batch[a][b].to_bits(),
2301                            want[a][b].to_bits(),
2302                            "OneSeed K={K} batch lane {r} third[{a}][{b}]"
2303                        );
2304                    }
2305                }
2306                rows_checked += 1;
2307            }
2308        }
2309        rows_checked
2310    }
2311
2312    fn check_twoseed<const K: usize>(state: &mut u64, batches: usize) -> usize {
2313        let mut rows_checked = 0;
2314        for _ in 0..batches {
2315            let rows: [[f64; K]; 4] =
2316                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2317            let u: [[f64; K]; 4] =
2318                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2319            let v: [[f64; K]; 4] =
2320                std::array::from_fn(|_| std::array::from_fn(|_| rand_unit(state)));
2321
2322            let prod: [TwoSeed<K>; 4] = std::array::from_fn(|r| {
2323                let p: [TwoSeed<K>; K] =
2324                    std::array::from_fn(|a| TwoSeed::seed(rows[r][a], a, u[r][a], v[r][a]));
2325                row_expr(&p)
2326            });
2327
2328            let scal: [TwoSeedLane<f64, K>; 4] = std::array::from_fn(|r| {
2329                let p: [TwoSeedLane<f64, K>; K] =
2330                    std::array::from_fn(|a| TwoSeedLane::seed(rows[r][a], a, u[r][a], v[r][a]));
2331                row_expr(&p)
2332            });
2333
2334            let pbatch: [TwoSeedBatch<K>; K] = std::array::from_fn(|a| {
2335                let val = wide::f64x4::new([rows[0][a], rows[1][a], rows[2][a], rows[3][a]]);
2336                let uu = wide::f64x4::new([u[0][a], u[1][a], u[2][a], u[3][a]]);
2337                let vv = wide::f64x4::new([v[0][a], v[1][a], v[2][a], v[3][a]]);
2338                TwoSeedBatch::seed(val, a, uu, vv)
2339            });
2340            let batch = row_expr(&pbatch);
2341
2342            for r in 0..4 {
2343                let want = prod[r].contracted_fourth();
2344                let got_scal = scal[r].contracted_fourth();
2345                let got_batch = batch.lane(r).contracted_fourth();
2346                assert_eq!(
2347                    scal[r].base.v.to_bits(),
2348                    prod[r].base.value().to_bits(),
2349                    "TwoSeed K={K} scalar value"
2350                );
2351                assert_eq!(
2352                    batch.lane(r).base.value().to_bits(),
2353                    prod[r].base.value().to_bits(),
2354                    "TwoSeed K={K} batch lane {r} value"
2355                );
2356                for a in 0..K {
2357                    for b in 0..K {
2358                        assert_eq!(
2359                            got_scal[a][b].to_bits(),
2360                            want[a][b].to_bits(),
2361                            "TwoSeed K={K} scalar fourth[{a}][{b}]"
2362                        );
2363                        assert_eq!(
2364                            got_batch[a][b].to_bits(),
2365                            want[a][b].to_bits(),
2366                            "TwoSeed K={K} batch lane {r} fourth[{a}][{b}]"
2367                        );
2368                    }
2369                }
2370                rows_checked += 1;
2371            }
2372        }
2373        rows_checked
2374    }
2375
2376    /// ≥2000 random 4-row batches per K, across K ∈ {2,3,4,9}: the
2377    /// contracted-third channel of every `OneSeedLane` lane is `to_bits`-identical
2378    /// to the production [`OneSeed`] per row.
2379    #[test]
2380    fn oneseed_lanes_contracted_third_bit_identical() {
2381        let mut state = 0x1234_5678_9ABC_DEF0_u64;
2382        let batches = 2000;
2383        let rows_checked = check_oneseed::<2>(&mut state, batches)
2384            + check_oneseed::<3>(&mut state, batches)
2385            + check_oneseed::<4>(&mut state, batches)
2386            + check_oneseed::<9>(&mut state, batches);
2387        // 4 widths × `batches` batches × 4 rows each: a silently empty inner
2388        // loop would leave this at zero instead of passing as a no-op.
2389        assert_eq!(rows_checked, 4 * batches * 4);
2390    }
2391
2392    /// ≥2000 random 4-row batches per K, across K ∈ {2,3,4,9}: the
2393    /// contracted-fourth channel of every `TwoSeedLane` lane is `to_bits`-identical
2394    /// to the production [`TwoSeed`] per row.
2395    #[test]
2396    fn twoseed_lanes_contracted_fourth_bit_identical() {
2397        let mut state = 0x0FED_CBA9_8765_4321_u64;
2398        let batches = 2000;
2399        let rows_checked = check_twoseed::<2>(&mut state, batches)
2400            + check_twoseed::<3>(&mut state, batches)
2401            + check_twoseed::<4>(&mut state, batches)
2402            + check_twoseed::<9>(&mut state, batches);
2403        // 4 widths × `batches` batches × 4 rows each: a silently empty inner
2404        // loop would leave this at zero instead of passing as a no-op.
2405        assert_eq!(rows_checked, 4 * batches * 4);
2406    }
2407}
2408
2409#[cfg(test)]
2410mod unit_tests {
2411    use super::{JetScalar, Order1, Order2, filtered_implicit_solve_scalar};
2412
2413    // ── Order2 direct property tests ─────────────────────────────────────────
2414
2415    /// `Order2::constant(c)` carries value `c` and zero everywhere else.
2416    #[test]
2417    fn order2_constant_has_zero_derivatives() {
2418        let s = Order2::<3>::constant(7.5);
2419        assert_eq!(s.value(), 7.5);
2420        for a in 0..3 {
2421            assert_eq!(s.g()[a], 0.0, "grad[{a}] should be zero");
2422            for b in 0..3 {
2423                assert_eq!(s.h()[a][b], 0.0, "hess[{a}][{b}] should be zero");
2424            }
2425        }
2426    }
2427
2428    /// `Order2::variable(x, axis)` has unit gradient in slot `axis` and zero Hessian.
2429    #[test]
2430    fn order2_variable_has_unit_gradient_in_seeded_slot() {
2431        let x = -2.5_f64;
2432        let s = Order2::<4>::variable(x, 2);
2433        assert_eq!(s.value(), x);
2434        for a in 0..4 {
2435            let expected_g = if a == 2 { 1.0 } else { 0.0 };
2436            assert_eq!(s.g()[a], expected_g, "grad[{a}]");
2437            for b in 0..4 {
2438                assert_eq!(s.h()[a][b], 0.0, "hess[{a}][{b}] should be zero");
2439            }
2440        }
2441    }
2442
2443    /// `Order2::add` sums gradient channels; `sub` is the inverse on gradients.
2444    /// Uses integer-valued primaries so the value roundtrip is also exact.
2445    #[test]
2446    fn order2_add_sub_roundtrip() {
2447        let p = Order2::<2>::variable(3.0, 0);
2448        let q = Order2::<2>::variable(2.0, 1);
2449        let pq = JetScalar::add(&p, &q);
2450        // value = 3 + 2 = 5
2451        assert_eq!(pq.value(), 5.0, "add value");
2452        let back = JetScalar::sub(&pq, &q);
2453        // (p + q) - q gradient should equal p's gradient exactly
2454        for a in 0..2 {
2455            assert_eq!(back.g()[a], p.g()[a], "grad[{a}] roundtrip");
2456        }
2457    }
2458
2459    /// `Order2::mul` of two variables satisfies the Leibniz product rule:
2460    ///   ∂(p·q)/∂p = q,  ∂(p·q)/∂q = p,  ∂²(p·q)/∂p∂q = 1.
2461    #[test]
2462    fn order2_mul_satisfies_leibniz_rule() {
2463        let pv = 3.0_f64;
2464        let qv = -2.0_f64;
2465        let p = Order2::<2>::variable(pv, 0);
2466        let q = Order2::<2>::variable(qv, 1);
2467        let pq = JetScalar::mul(&p, &q);
2468        assert_eq!(pq.value(), pv * qv, "value = p·q");
2469        assert_eq!(pq.g()[0], qv, "∂(p·q)/∂p = q");
2470        assert_eq!(pq.g()[1], pv, "∂(p·q)/∂q = p");
2471        assert_eq!(pq.h()[0][1], 1.0, "∂²(p·q)/∂p∂q = 1");
2472        assert_eq!(pq.h()[1][0], 1.0, "∂²(p·q)/∂q∂p = 1 (symmetric)");
2473        assert_eq!(pq.h()[0][0], 0.0, "∂²(p·q)/∂p² = 0");
2474        assert_eq!(pq.h()[1][1], 0.0, "∂²(p·q)/∂q² = 0");
2475    }
2476
2477    /// `Order2::scale(s)` multiplies every channel by `s`.
2478    #[test]
2479    fn order2_scale_multiplies_all_channels() {
2480        let p = Order2::<2>::variable(4.0, 0);
2481        let s = 2.5_f64;
2482        let ps = JetScalar::scale(&p, s);
2483        assert_eq!(ps.value(), 4.0 * s);
2484        assert_eq!(ps.g()[0], 1.0 * s);
2485        assert_eq!(ps.g()[1], 0.0);
2486    }
2487
2488    /// `Order2::exp` at a constant has value `e^c`, gradient `e^c * g`, Hessian `e^c * (g⊗g + H)`.
2489    /// At a seeded variable `p₀`, the first derivative is `e^{p₀}` and second is `e^{p₀}`.
2490    #[test]
2491    fn order2_exp_derivative_stack_correct() {
2492        let p0 = 1.0_f64;
2493        let p = Order2::<1>::variable(p0, 0);
2494        let ep = JetScalar::exp(&p);
2495        let e = p0.exp();
2496        assert!((ep.value() - e).abs() < 1e-15, "exp value");
2497        assert!((ep.g()[0] - e).abs() < 1e-15, "d/dp exp(p) = exp(p)");
2498        assert!((ep.h()[0][0] - e).abs() < 1e-15, "d²/dp² exp(p) = exp(p)");
2499    }
2500
2501    /// `Order2::ln` at a seeded variable: d/dp ln(p) = 1/p, d²/dp² ln(p) = -1/p².
2502    #[test]
2503    fn order2_ln_derivative_stack_correct() {
2504        let p0 = 2.0_f64;
2505        let p = Order2::<1>::variable(p0, 0);
2506        let lnp = JetScalar::ln(&p);
2507        assert!((lnp.value() - p0.ln()).abs() < 1e-15, "ln value");
2508        assert!((lnp.g()[0] - 1.0 / p0).abs() < 1e-15, "d/dp ln(p) = 1/p");
2509        assert!((lnp.h()[0][0] - (-1.0 / (p0 * p0))).abs() < 1e-15, "d²/dp² ln(p) = -1/p²");
2510    }
2511
2512    /// `exp` and `ln` are mutual inverses: `ln(exp(p)).value() == p` at the scalar.
2513    #[test]
2514    fn order2_exp_ln_roundtrip_at_value() {
2515        let p0 = 0.8_f64;
2516        let p = Order2::<1>::variable(p0, 0);
2517        let roundtrip = JetScalar::ln(&JetScalar::exp(&p));
2518        assert!((roundtrip.value() - p0).abs() < 1e-14, "ln(exp(p)) ≈ p");
2519    }
2520
2521    // ── Order1 tests ─────────────────────────────────────────────────────────
2522
2523    /// `Order1::constant` carries the correct value with all-zero gradient.
2524    #[test]
2525    fn order1_constant_has_zero_gradient() {
2526        let s = Order1::<3>::constant(-5.0);
2527        assert_eq!(s.value(), -5.0);
2528        for a in 0..3 {
2529            assert_eq!(s.g()[a], 0.0, "g[{a}] should be zero");
2530        }
2531    }
2532
2533    /// `Order1::variable(x, axis)` has unit gradient only in `axis`.
2534    #[test]
2535    fn order1_variable_has_unit_gradient_in_seeded_slot() {
2536        let s = Order1::<3>::variable(2.0, 1);
2537        assert_eq!(s.value(), 2.0);
2538        assert_eq!(s.g()[0], 0.0);
2539        assert_eq!(s.g()[1], 1.0);
2540        assert_eq!(s.g()[2], 0.0);
2541    }
2542
2543    /// `Order1::mul` satisfies the product rule (value and gradient, no Hessian).
2544    #[test]
2545    fn order1_mul_satisfies_product_rule() {
2546        let pv = 3.0_f64;
2547        let qv = -2.0_f64;
2548        let p = Order1::<2>::variable(pv, 0);
2549        let q = Order1::<2>::variable(qv, 1);
2550        let pq = JetScalar::mul(&p, &q);
2551        assert_eq!(pq.value(), pv * qv);
2552        assert_eq!(pq.g()[0], qv, "∂(p·q)/∂p = q");
2553        assert_eq!(pq.g()[1], pv, "∂(p·q)/∂q = p");
2554    }
2555
2556    /// `Order1::exp` carries the correct value and gradient `e^{p₀}`.
2557    #[test]
2558    fn order1_exp_has_correct_value_and_gradient() {
2559        let p0 = 0.5_f64;
2560        let p = Order1::<2>::variable(p0, 0);
2561        let ep = JetScalar::exp(&p);
2562        let e = p0.exp();
2563        assert!((ep.value() - e).abs() < 1e-15, "exp value");
2564        assert!((ep.g()[0] - e).abs() < 1e-15, "d/dp exp(p)");
2565        assert_eq!(ep.g()[1], 0.0, "irrelevant gradient slot is zero");
2566    }
2567
2568    /// `Order1` and `Order2` agree on value and gradient for the same expression.
2569    #[test]
2570    fn order1_and_order2_agree_on_value_and_gradient() {
2571        let p0 = 1.3_f64;
2572        let q0 = -0.7_f64;
2573        // evaluate (p * q + p).exp() at (p0, q0)
2574        let p1 = Order1::<2>::variable(p0, 0);
2575        let q1 = Order1::<2>::variable(q0, 1);
2576        let expr1 = JetScalar::exp(&JetScalar::add(&JetScalar::mul(&p1, &q1), &p1));
2577
2578        let p2 = Order2::<2>::variable(p0, 0);
2579        let q2 = Order2::<2>::variable(q0, 1);
2580        let expr2 = JetScalar::exp(&JetScalar::add(&JetScalar::mul(&p2, &q2), &p2));
2581
2582        assert!((expr1.value() - expr2.value()).abs() < 1e-14, "value mismatch");
2583        for a in 0..2 {
2584            assert!(
2585                (expr1.g()[a] - expr2.g()[a]).abs() < 1e-14,
2586                "gradient[{a}] mismatch"
2587            );
2588        }
2589    }
2590
2591    // ── filtered_implicit_solve_scalar ────────────────────────────────────────
2592
2593    /// Lift the trivial linear constraint F(a, θ) = a - θ = 0 through `Order2<1>`.
2594    /// The exact lifted jet is a(θ) = θ, so value=θ₀, gradient=1.
2595    #[test]
2596    fn filtered_implicit_solve_linear_constraint_gives_exact_jet() {
2597        let theta0 = 3.0_f64;
2598        let theta = Order2::<1>::variable(theta0, 0);
2599        // a0 = theta0, F_a = 1, inv_fa = 1; 2 iters suffice for Order2.
2600        let a = filtered_implicit_solve_scalar::<1, Order2<1>>(
2601            theta0,
2602            1.0,
2603            2,
2604            |a_jet| JetScalar::sub(a_jet, &theta),
2605        );
2606        assert!((a.value() - theta0).abs() < 1e-14, "value = theta0");
2607        // da/dtheta = 1 (identity)
2608        assert!((a.g()[0] - 1.0).abs() < 1e-14, "gradient = 1");
2609        // d²a/dtheta² = 0 (linear)
2610        assert!(a.h()[0][0].abs() < 1e-14, "hessian = 0");
2611    }
2612
2613    /// `filtered_implicit_solve_scalar` on a quadratic constraint F(a,θ)=a²-θ=0
2614    /// with primal root a₀=√θ₀, giving da/dθ = 1/(2√θ₀), d²a/dθ² = -1/(4θ₀^{3/2}).
2615    #[test]
2616    fn filtered_implicit_solve_quadratic_constraint_matches_analytic_derivatives() {
2617        let theta0 = 4.0_f64;
2618        let a0 = theta0.sqrt();
2619        let inv_fa = 1.0 / (2.0 * a0);
2620        let theta = Order2::<1>::variable(theta0, 0);
2621        // F(a,theta) = a*a - theta
2622        let a = filtered_implicit_solve_scalar::<1, Order2<1>>(a0, inv_fa, 2, |a_jet| {
2623            let aa = JetScalar::mul(a_jet, a_jet);
2624            JetScalar::sub(&aa, &theta)
2625        });
2626        let tol = 1e-12;
2627        assert!((a.value() - a0).abs() < tol, "value = sqrt(theta0)");
2628        let expected_g = 0.5 / a0;
2629        assert!((a.g()[0] - expected_g).abs() < tol, "da/dtheta = 1/(2*sqrt)");
2630        let expected_h = -0.25 / (theta0 * a0);
2631        assert!((a.h()[0][0] - expected_h).abs() < tol, "d2a/dtheta2 = -1/(4*theta^1.5)");
2632    }
2633}