NAF campaign stages 1-2: LE load walks + the digit loop's arithmetic core

- `Proofs/DsmNafLoadSpec.lean` (generated) — the byte-to-word LE load of
  `non_adjacent_form`: four 8-peel inner walks (t |= bytes[8k+bi] << 8bi)
  and the outer 4-peel filling x_u64[0..3]; x_u64[4] stays 0 (the pad word
  the cross-word window reads at positions >= 251).

- `Proofs/DsmNafMath.lean` — the pure arithmetic of the w=5 digit loop:
  `nafSum`/`nafSum_set`; window-read lemmas `naf_window_single` /
  `naf_window_cross` (cross-word disjoint-OR read sees (V >> pos) mod 32);
  invariant steps `naf_even_step` / `naf_odd_step` (Nat.mod_mul telescope:
  digit + promoted carry reconstruct the consumed bits EXACTLY, in ZZ);
  carry-kill `naf_carry_even` / `naf_carry_odd` (V < 2^253 forces the
  carry dead before bit 256); `naf_exit` (nafSum naf 256 = V exactly).

The digit-loop walk (stage 3) composes these next; its invariant is
    nafSum naf 256 + carry*2^pos = V mod 2^pos
with digits at k >= pos all zero and carry = 1 -> pos <= 254.

CERTS += naf_load_spec, naf_window_cross, naf_exit (axiom-clean).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
mrwulf 2026-07-04 15:30:18 +02:00
parent 724feaeb8b
commit 417fb14b6a
3 changed files with 1731 additions and 0 deletions

File diff suppressed because it is too large Load diff

View file

@ -0,0 +1,239 @@
/- ──────────────────────────────────────────────────────────────────────────────
Proofs/DsmNafMath.lean — NAF campaign, stage 2: the pure arithmetic core
of the w=5 NAF digit loop (no extraction dependence beyond `nafDigit`).
The digit loop's state is (naf, pos, carry) with the exact invariant
nafSum naf 256 + carry·2^pos = V mod 2^pos
(digits at k ≥ pos all zero, carry ≤ 1, and carry = 1 → pos ≤ 254 given
V < 2^253 — the component that kills the carry at exit).
Step theorems the monadic walk plugs in:
· `nafSum_set` — writing a fresh digit adds d·2^pos to the sum.
· `div_pow_shift` / `mod32_absorb` / `naf_window_single` / `naf_window_cross`
— the 64-bit buffer read at bit position p+b sees (V >> (p+b)) mod 32
(single word when b + 5 ≤ 64; cross-word via disjoint-OR otherwise).
· `naf_even_step` — even window ⇒ pos+1, carry preserved (the parity
of V >> pos matches carry, so both sides absorb carry·2^(pos+1)).
· `naf_odd_step` — odd window ⇒ digit (window 32·carry'), pos+5:
the digit plus the new carry reconstruct the 5 consumed bits
(Nat.mod_mul telescoping).
· `naf_carry_even` / `naf_carry_odd` — carry = 1 → pos ≤ 254 propagation
from V < 2^253.
· `naf_exit` — pos ≥ 256 kills the carry: nafSum = V exactly.
────────────────────────────────────────────────────────────────────────────── -/
import Proofs.DsmLoopSpec
open Aeneas Aeneas.Std
open curve25519_dalek
set_option maxHeartbeats 4000000
set_option linter.unusedSimpArgs false
set_option maxRecDepth 8000
namespace CurveFieldProofs
/-! ### The signed digit sum -/
/-- Σ_{k<m} naf[k]·2^k over — the value a digit array denotes. -/
def nafSum (naf : Std.Array Std.I8 256#usize) (m : ) : :=
∑ k ∈ Finset.range m, nafDigit naf k * 2^k
/-- Setting entry `pos` (previously 0) adds d·2^pos to the full sum. -/
theorem nafSum_set (naf naf' : Std.Array Std.I8 256#usize) (pos : ) (d : Std.I8)
(hpos : pos < 256) (hold : nafDigit naf pos = 0)
(hset : (↑naf' : List Std.I8) = (↑naf : List Std.I8).set pos d) :
nafSum naf' 256 = nafSum naf 256 + d.val * 2^pos := by
have hlen : (↑naf : List Std.I8).length = 256 := by scalar_tac
have hlen' : ((↑naf : List Std.I8).set pos d).length = 256 := by
rw [List.length_set]; exact hlen
have hdig : ∀ k, k < 256 → nafDigit naf' k =
if k = pos then d.val else nafDigit naf k := by
intro k hk
unfold nafDigit
rw [hset]
by_cases h : k = pos
· subst h
rw [getElem!_pos ((↑naf : List Std.I8).set k d) k (by omega),
List.getElem_set_self]
simp
· rw [getElem!_pos ((↑naf : List Std.I8).set pos d) k (by omega),
List.getElem_set_ne (by omega),
← getElem!_pos (↑naf : List Std.I8) k (by omega)]
rw [if_neg h]
unfold nafSum
have hmem : pos ∈ Finset.range 256 := Finset.mem_range.mpr hpos
rw [← Finset.sum_erase_add _ _ hmem, ← Finset.sum_erase_add _ _ hmem]
have hcongr : ∑ k ∈ (Finset.range 256).erase pos, nafDigit naf' k * 2^k
= ∑ k ∈ (Finset.range 256).erase pos, nafDigit naf k * 2^k := by
apply Finset.sum_congr rfl
intro k hk
rw [hdig k (Finset.mem_range.mp (Finset.mem_of_mem_erase hk)),
if_neg (Finset.ne_of_mem_erase hk)]
rw [hcongr, hdig pos hpos, if_pos rfl, hold]
ring
/-! ### Window-read arithmetic -/
/-- Dividing the three-part value lo + 2^p·(w + 2^64·rest) by 2^(p+b)
(b ≤ 64, lo < 2^p) yields w >> b plus the rest shifted down. -/
theorem div_pow_shift (lo w rest p b : ) (hlo : lo < 2^p) (hb : b ≤ 64) :
(lo + 2^p * (w + 2^64 * rest)) / 2^(p+b) = w / 2^b + 2^(64-b) * rest := by
have hp : (0:) < 2^p := Nat.two_pow_pos p
have hbp : (0:) < 2^b := Nat.two_pow_pos b
rw [pow_add, ← Nat.div_div_eq_div_mul]
have h1 : (lo + 2^p * (w + 2^64 * rest)) / 2^p = w + 2^64 * rest := by
rw [Nat.add_mul_div_left _ _ hp, Nat.div_eq_of_lt hlo, Nat.zero_add]
rw [h1]
have h2 : (2:)^64 = 2^b * 2^(64-b) := by
rw [← pow_add]; congr 1; omega
rw [h2, Nat.mul_assoc, Nat.add_mul_div_left _ _ hbp]
/-- Multiples of 2^m (m ≥ 5) vanish mod 32. -/
theorem mod32_absorb (x y m : ) (hm : 5 ≤ m) :
(x + 2^m * y) % 32 = x % 32 := by
have h : (2:)^m = 32 * 2^(m-5) := by
rw [show (32:) = 2^5 by norm_num, ← pow_add]; congr 1; omega
rw [h, Nat.mul_assoc, Nat.add_mul_mod_self_left]
/-- Single-word window read: when the 5-bit window at bit b fits inside
word w (b + 5 ≤ 64), (w >> b) mod 32 is the value's window at p+b. -/
theorem naf_window_single (V lo w rest p b : )
(hV : V = lo + 2^p * (w + 2^64 * rest)) (hlo : lo < 2^p) (hb : b + 5 ≤ 64) :
(w / 2^b) % 32 = (V / 2^(p+b)) % 32 := by
rw [hV, div_pow_shift lo w rest p b hlo (by omega)]
exact (mod32_absorb _ _ _ (by omega)).symm
/-- Cross-word window read: when the window at bit b straddles into the next
word w' (b ≥ 60), the extracted read (w >> b) ||| ((w' << (64b)) mod 2^64)
still sees the value's window at p+b, mod 32. -/
theorem naf_window_cross (V lo w w' rest p b : )
(hV : V = lo + 2^p * (w + 2^64 * (w' + 2^64 * rest)))
(hlo : lo < 2^p) (hw : w < 2^64) (hb : b < 64) :
((w / 2^b) ||| (w' <<< (64 - b)) % 2^64) % 32 = (V / 2^(p+b)) % 32 := by
have hp64 : (2:)^(64-b) * 2^b = 2^64 := by
rw [← pow_add]; congr 1; omega
have hp128 : (2:)^(64-b) * 2^64 = 2^(128-b) := by
rw [← pow_add]; congr 1; omega
-- the truncated shift: (w' << (64b)) mod 2^64 = 2^(64b)·(w' mod 2^b)
have hsh : (w' <<< (64 - b)) % 2^64 = 2^(64-b) * (w' % 2^b) := by
rw [Nat.shiftLeft_eq, Nat.mul_comm w' _, ← hp64, Nat.mul_mod_mul_left]
-- the OR is disjoint: w >> b < 2^(64b)
have hdl : w / 2^b < 2^(64-b) := by
rw [Nat.div_lt_iff_lt_mul (Nat.two_pow_pos b)]
calc w < 2^64 := hw
_ = 2^(64-b) * 2^b := hp64.symm
have hor := Nat.two_pow_add_eq_or_of_lt (b := w / 2^b) (i := 64-b) hdl (w' % 2^b)
rw [hsh, Nat.lor_comm, ← hor]
-- the value side
rw [hV, div_pow_shift lo w _ p b hlo (by omega)]
-- both sides are (w/2^b + 2^(64b)·(w' mod 2^b)) mod 32 after absorbing
-- the 2^64-multiples: w' = w' mod 2^b + 2^b·(w'/2^b)
have hw' : w' = w' % 2^b + 2^b * (w' / 2^b) := (Nat.mod_add_div _ _).symm
have e1 : w / 2^b + 2^(64-b) * (w' + 2^64 * rest)
= (2^(64-b) * (w' % 2^b) + w / 2^b)
+ 2^64 * (w' / 2^b + 2^(64-b) * rest) := by
conv_lhs => rw [hw']
rw [Nat.mul_add, Nat.mul_add, Nat.mul_add, ← Nat.mul_assoc, hp64,
← Nat.mul_assoc, hp128]
have hp128' : (2:)^(128-b) = 2^64 * 2^(64-b) := by
rw [← pow_add]; congr 1; omega
rw [hp128']
ring
rw [e1]
have := mod32_absorb (2^(64-b) * (w' % 2^b) + w / 2^b)
(w' / 2^b + 2^(64-b) * rest) 64 (by omega)
rw [this]
/-! ### Invariant step theorems -/
/-- Even window: digit 0, position advances by 1, carry preserved.
The parity of V >> pos equals the carry, so the invariant extends. -/
theorem naf_even_step (V pos : ) (carry : ) (S : )
(hc : carry ≤ 1)
(hinv : S + carry * 2^pos = ((V % 2^pos : ) : ))
(heven : (carry + (V / 2^pos) % 32) % 2 = 0) :
S + carry * 2^(pos+1) = ((V % 2^(pos+1) : ) : ) := by
have hmm : V % 2^(pos+1) = V % 2^pos + 2^pos * ((V / 2^pos) % 2) := by
rw [pow_succ, Nat.mod_mul]
have hpar : (V / 2^pos) % 2 = carry := by omega
rw [hmm, hpar]
push_cast at hinv ⊢
linear_combination hinv
/-- Odd window: digit window 32·carry', position advances by 5.
The digit plus the promoted carry reconstruct the 5 consumed bits. -/
theorem naf_odd_step (V pos : ) (carry carry' : ) (S d : )
(hinv : S + carry * 2^pos = ((V % 2^pos : ) : ))
(hd : d = (carry : ) + ((V / 2^pos) % 32 : ) - 32 * carry') :
(S + d * 2^pos) + carry' * 2^(pos+5) = ((V % 2^(pos+5) : ) : ) := by
have hmm : V % 2^(pos+5) = V % 2^pos + 2^pos * ((V / 2^pos) % 32) := by
have h : (2:)^(pos+5) = 2^pos * 32 := by rw [pow_add]; norm_num
rw [h, Nat.mod_mul]
rw [hmm, hd]
push_cast at hinv ⊢
linear_combination hinv
/-- Carry propagation, even step: with V < 2^253, an even step that keeps
carry = 1 must be reading a set bit, so pos ≤ 252 and pos+1 ≤ 254. -/
theorem naf_carry_even (V pos : ) (carry : ) (hV : V < 2^253)
(hc : carry ≤ 1)
(heven : (carry + (V / 2^pos) % 32) % 2 = 0) :
carry = 1 → pos + 1 ≤ 254 := by
intro h1
subst h1
have h3 : 1 ≤ (V / 2^pos) % 32 := by
generalize (V / 2^pos) % 32 = r at heven ⊢
omega
have hge : 1 ≤ V / 2^pos := le_trans h3 (Nat.mod_le _ _)
have hle : 2^pos ≤ V := by
have h5 := (Nat.le_div_iff_mul_le (Nat.two_pow_pos pos)).mp hge
simpa using h5
have hpb : pos ≤ 252 := by
by_contra h
have h253 : (2:)^253 ≤ 2^pos := Nat.pow_le_pow_right (by norm_num) (by omega)
exact absurd (lt_of_le_of_lt (le_trans h253 hle) hV) (lt_irrefl _)
omega
/-- Carry propagation, odd step: producing carry' = 1 needs window ≥ 16, so
V >> pos ≥ 15, forcing pos ≤ 249 (V < 2^253) and pos+5 ≤ 254. -/
theorem naf_carry_odd (V pos : ) (carry carry' : ) (hV : V < 2^253)
(hc : carry ≤ 1)
(hcw : (carry + (V / 2^pos) % 32 < 16 ∧ carry' = 0)
(16 ≤ carry + (V / 2^pos) % 32 ∧ carry' = 1)) :
carry' = 1 → pos + 5 ≤ 254 := by
intro h1
rcases hcw with ⟨-, h0⟩ | ⟨hge, -⟩
· omega
· have h15 : 15 ≤ (V / 2^pos) % 32 := by
generalize (V / 2^pos) % 32 = r at hge ⊢
omega
have hge15 : 15 ≤ V / 2^pos := le_trans h15 (Nat.mod_le _ _)
have hmul : 15 * 2^pos ≤ V :=
(Nat.le_div_iff_mul_le (Nat.two_pow_pos pos)).mp hge15
have hpb : pos ≤ 249 := by
by_contra h
have h250 : (2:)^250 ≤ 2^pos := Nat.pow_le_pow_right (by norm_num) (by omega)
have hbig : (2:)^253 < 15 * 2^250 := by norm_num
have hmono : 15 * 2^250 ≤ 15 * 2^pos := Nat.mul_le_mul_left 15 h250
exact absurd (lt_of_le_of_lt (le_trans hmono hmul) hV) (not_lt.mpr hbig.le)
omega
/-- Exit: at pos ≥ 256 the carry must be dead (carry = 1 forces pos ≤ 254),
and V mod 2^pos = V, so the digit sum equals V exactly. -/
theorem naf_exit (V pos : ) (carry : ) (S : ) (hV : V < 2^253)
(hpos : 256 ≤ pos) (hc : carry ≤ 1) (hcp : carry = 1 → pos ≤ 254)
(hinv : S + carry * 2^pos = ((V % 2^pos : ) : )) : S = V := by
have hc0 : carry = 0 := by
by_contra h
have h1 : carry = 1 := by omega
have := hcp h1
omega
subst hc0
have hmod : V % 2^pos = V := by
apply Nat.mod_eq_of_lt
calc V < 2^253 := hV
_ ≤ 2^pos := Nat.pow_le_pow_right (by norm_num) (by omega)
rw [hmod] at hinv
push_cast at hinv
linarith
end CurveFieldProofs

View file

@ -52,16 +52,35 @@ PROOFS=(
EdAddAffNiels
EdConvert
EdMain
DsmTableSpec
DsmStepSpec
DsmLoopSpec
DsmNafLoadSpec
DsmNafMath
)
# Fully-qualified certificate names; each must be axiom-clean.
CERTS=(
CurveFieldProofs.fieldImplementation
CurveFieldProofs.edwardsImplementation
CurveFieldProofs.naf_table_spec
CurveFieldProofs.naf_select_spec
CurveFieldProofs.proj_double_law
CurveFieldProofs.compl_as_projective_law
CurveFieldProofs.dsm_step_p_law
CurveFieldProofs.dsm_step_b_law
CurveFieldProofs.dsm_loop_spec
CurveFieldProofs.naf_load_spec
CurveFieldProofs.naf_exit
)
# Imports needed so every certificate in CERTS is in scope for the audit.
AUDIT_IMPORTS=(
Proofs.FieldMain
Proofs.EdMain
Proofs.DsmTableSpec
Proofs.DsmStepSpec
Proofs.DsmLoopSpec
Proofs.DsmNafLoadSpec
Proofs.DsmNafMath
)
# ── Phase 0: resource + integrity guards ────────────────────────────────────