import Lean.Data.Json import Mathlib.Data.UInt import Mathlib.Tactic import Mathlib.Data.Int.Basic import Mathlib.Data.Nat.Basic set_option maxRecDepth 20000 namespace Semantics.FixedPoint open Lean /-! A proof-friendly fixed-point core. Design rule: * The semantic value is a bounded signed raw integer. * Saturation is performed by `ofRawInt`. * UInt bit-patterns are boundary/hardware artifacts, not the proof model. This removes the old proof debt caused by proving signed arithmetic facts directly against modular UInt32/UInt64 overflow behavior. -/ -- ═══════════════════════════════════════════════════════════════════════════ -- Q0.16 signed normalized fraction -- ═══════════════════════════════════════════════════════════════════════════ def q0_16MinRaw : Int := -32768 def q0_16MaxRaw : Int := 32767 def q0_16Scale : Int := 32767 /-- Q0.16 pure fraction representation. The canonical proof model stores the signed raw integer in [-32768, 32767]. Use boundary conversion functions when a UInt16 bit pattern is required. -/ abbrev Q0_16 := { x : Int // q0_16MinRaw ≤ x ∧ x ≤ q0_16MaxRaw } instance : Repr Q0_16 where reprPrec q _ := repr q.val instance : BEq Q0_16 where beq a b := a.val == b.val instance : Inhabited Q0_16 where default := ⟨0, by constructor <;> norm_num [q0_16MinRaw, q0_16MaxRaw]⟩ instance : ToJson Q0_16 where toJson q := Json.mkObj [("val", toJson q.val)] namespace Q0_16 @[ext] theorem ext {a b : Q0_16} (h : a.val = b.val) : a = b := Subtype.ext h @[inline] def toInt (q : Q0_16) : Int := q.val @[inline] def ofRawInt (raw : Int) : Q0_16 := if hhi : raw > q0_16MaxRaw then ⟨q0_16MaxRaw, by constructor <;> norm_num [q0_16MinRaw, q0_16MaxRaw]⟩ else if hlo : raw < q0_16MinRaw then ⟨q0_16MinRaw, by constructor <;> norm_num [q0_16MinRaw, q0_16MaxRaw]⟩ else ⟨raw, by constructor · dsimp [q0_16MinRaw, q0_16MaxRaw] at *; omega · dsimp [q0_16MinRaw, q0_16MaxRaw] at *; omega⟩ instance : FromJson Q0_16 where fromJson? j := do let raw : Int ← fromJson? (← j.getObjVal? "val") pure (ofRawInt raw) def zero : Q0_16 := ofRawInt 0 def one : Q0_16 := ofRawInt q0_16MaxRaw def half : Q0_16 := ofRawInt 16383 def neg (x : Q0_16) : Q0_16 := ofRawInt (-x.toInt) def add (a b : Q0_16) : Q0_16 := ofRawInt (a.toInt + b.toInt) def sub (a b : Q0_16) : Q0_16 := ofRawInt (a.toInt - b.toInt) def mul (a b : Q0_16) : Q0_16 := ofRawInt ((a.toInt * b.toInt) / q0_16Scale) def div (a b : Q0_16) : Q0_16 := if b.toInt = 0 then zero else ofRawInt ((a.toInt * q0_16Scale) / b.toInt) def abs (x : Q0_16) : Q0_16 := if x.toInt < 0 then neg x else x instance : Add Q0_16 where add := add instance : Sub Q0_16 where sub := sub instance : Mul Q0_16 where mul := mul instance : Div Q0_16 where div := div instance : Neg Q0_16 where neg := neg def lt (a b : Q0_16) : Bool := a.toInt < b.toInt def le (a b : Q0_16) : Bool := a.toInt ≤ b.toInt def gt (a b : Q0_16) : Bool := b.toInt < a.toInt def ge (a b : Q0_16) : Bool := b.toInt ≤ a.toInt def toFloat (q : Q0_16) : Float := Float.ofInt q.toInt / 32767.0 def ofFloat (f : Float) : Q0_16 := if f.isNaN then zero else if f ≥ 1.0 then one else if f ≤ -1.0 then neg one else if f < 0.0 then ofRawInt (-(Int.ofNat ((-f * 32767.0).round.toUInt16.toNat))) else ofRawInt (Int.ofNat ((f * 32767.0).round.toUInt16.toNat)) def log2 (q : Q0_16) : Q0_16 := if q.toInt = 0 then zero else let f := toFloat q if f ≤ 0.0 then zero else ofFloat (Float.log2 f) def min (a b : Q0_16) : Q0_16 := if a.toInt ≤ b.toInt then a else b end Q0_16 -- ═══════════════════════════════════════════════════════════════════════════ -- Q16.16 signed fixed-point -- ═══════════════════════════════════════════════════════════════════════════ def q16MinRaw : Int := -2147483648 def q16MaxRaw : Int := 2147483647 def q16Scale : Int := 65536 /-- Saturating clamp of a raw integer into [q16MinRaw, q16MaxRaw]. This is the pure-Int kernel of `Q16_16.ofRawInt`; all monotonicity reasoning is proved once here and reused by higher lemmas. -/ def q16Clamp (i : Int) : Int := if i > q16MaxRaw then q16MaxRaw else if i < q16MinRaw then q16MinRaw else i /-- `q16Clamp` is monotone: a ≤ b → q16Clamp a ≤ q16Clamp b. -/ theorem q16Clamp_monotone (a b : Int) (h : a ≤ b) : q16Clamp a ≤ q16Clamp b := by unfold q16Clamp by_cases ha_hi : a > q16MaxRaw · by_cases hb_hi : b > q16MaxRaw · simp [ha_hi, hb_hi] · by_cases hb_lo : b < q16MinRaw · simp [ha_hi, hb_hi, hb_lo]; dsimp [q16MinRaw, q16MaxRaw] at *; omega · simp [ha_hi, hb_hi, hb_lo]; dsimp [q16MaxRaw] at *; omega · by_cases ha_lo : a < q16MinRaw · by_cases hb_hi : b > q16MaxRaw · simp [ha_hi, ha_lo, hb_hi]; dsimp [q16MinRaw, q16MaxRaw] at *; omega · by_cases hb_lo : b < q16MinRaw · simp [ha_hi, ha_lo, hb_hi, hb_lo] · simp [ha_hi, ha_lo, hb_hi, hb_lo]; dsimp [q16MinRaw] at *; omega · by_cases hb_hi : b > q16MaxRaw · simp [ha_hi, ha_lo, hb_hi]; dsimp [q16MaxRaw] at *; omega · by_cases hb_lo : b < q16MinRaw · simp [ha_hi, ha_lo, hb_hi, hb_lo]; dsimp [q16MinRaw] at *; omega · simp [ha_hi, ha_lo, hb_hi, hb_lo]; exact h /-- `q16Clamp` is idempotent on in-range values. -/ theorem q16Clamp_id_of_inRange (i : Int) (hlo : q16MinRaw ≤ i) (hhi : i ≤ q16MaxRaw) : q16Clamp i = i := by unfold q16Clamp simp [show ¬ i > q16MaxRaw from by omega, show ¬ i < q16MinRaw from by omega] lemma q16Clamp_lower (x : Int) : q16MinRaw ≤ q16Clamp x := by unfold q16Clamp q16MinRaw q16MaxRaw; split_ifs <;> omega lemma q16Clamp_upper (x : Int) : q16Clamp x ≤ q16MaxRaw := by unfold q16Clamp q16MinRaw q16MaxRaw; split_ifs <;> omega lemma q16Clamp_nonneg_of_nonneg {x : Int} (hx : 0 ≤ x) : 0 ≤ q16Clamp x := by unfold q16Clamp q16MinRaw q16MaxRaw split_ifs <;> omega lemma q16Clamp_idem (x : Int) : q16Clamp (q16Clamp x) = q16Clamp x := by have h_upper := q16Clamp_upper x have h_lower := q16Clamp_lower x unfold q16Clamp q16MinRaw q16MaxRaw split_ifs <;> omega /-- Q16.16 fixed-point representation. The canonical proof model stores the signed raw integer in [-2147483648, 2147483647]. Hardware/serialization UInt32 bit patterns should enter through `ofBits` and leave through `toBits`. All semantic proofs use `toInt`. -/ abbrev Q16_16 := { x : Int // q16MinRaw ≤ x ∧ x ≤ q16MaxRaw } instance : Repr Q16_16 where reprPrec q _ := repr q.val instance : BEq Q16_16 where beq a b := a.val == b.val instance : Inhabited Q16_16 where default := ⟨0, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ instance : ToJson Q16_16 where toJson q := Json.mkObj [("val", toJson q.val)] namespace Q16_16 @[ext] theorem ext {a b : Q16_16} (h : a.val = b.val) : a = b := Subtype.ext h @[inline] def toInt (q : Q16_16) : Int := q.val @[inline] def ofRawInt (raw : Int) : Q16_16 := if hhi : raw > q16MaxRaw then ⟨q16MaxRaw, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ else if hlo : raw < q16MinRaw then ⟨q16MinRaw, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ else ⟨raw, by constructor · dsimp [q16MinRaw, q16MaxRaw] at *; omega · dsimp [q16MinRaw, q16MaxRaw] at *; omega⟩ instance : FromJson Q16_16 where fromJson? j := do let raw : Int ← fromJson? (← j.getObjVal? "val") pure (ofRawInt raw) /-- Decode a UInt32 two's-complement hardware bit pattern into the signed model. -/ @[inline] def ofBits (u : UInt32) : Q16_16 := let n := u.toNat if n ≥ 2147483648 then ofRawInt ((n : Int) - 4294967296) else ofRawInt (n : Int) /-- Encode the signed model as a UInt32 two's-complement hardware bit pattern. -/ @[inline] def toBits (q : Q16_16) : UInt32 := UInt32.ofInt q.toInt def zero : Q16_16 := ⟨0, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ def one : Q16_16 := ⟨q16Scale, by constructor <;> norm_num [q16MinRaw, q16MaxRaw, q16Scale]⟩ def negOne : Q16_16 := ⟨-q16Scale, by constructor <;> norm_num [q16MinRaw, q16MaxRaw, q16Scale]⟩ def epsilon : Q16_16 := ⟨1, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ def two : Q16_16 := ⟨2 * q16Scale, by constructor <;> norm_num [q16MinRaw, q16MaxRaw, q16Scale]⟩ def maxVal : Q16_16 := ⟨q16MaxRaw, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ def minVal : Q16_16 := ⟨q16MinRaw, by constructor <;> norm_num [q16MinRaw, q16MaxRaw]⟩ /-- Saturating infinity/illegal sentinel. If you need the old 0xFFFFFFFF bit sentinel, use `ofRawInt (-1)` or `ofBits 0xFFFFFFFF` explicitly. -/ def infinity : Q16_16 := maxVal def scale : Nat := 65536 def ofNat (n : Nat) : Q16_16 := ofRawInt ((n : Int) * q16Scale) def satFromNat (n : Nat) : Q16_16 := ofRawInt ((n : Int) * q16Scale) def ofRatio (num : Nat) (den : Nat) : Q16_16 := if den = 0 then zero else ofRawInt (Int.ofNat (num * scale / den)) instance : OfNat Q16_16 n where ofNat := ofNat n @[inline] def ofInt (n : Int) : Q16_16 := ofRawInt (n * q16Scale) /-- Saturating addition. -/ @[inline] def add (a b : Q16_16) : Q16_16 := ofRawInt (a.toInt + b.toInt) /-- Saturating subtraction. -/ @[inline] def sub (a b : Q16_16) : Q16_16 := ofRawInt (a.toInt - b.toInt) /-- Saturating Q16.16 multiplication: raw result is `(a*b)/65536`. -/ @[inline] def mul (a b : Q16_16) : Q16_16 := ofRawInt ((a.toInt * b.toInt) / q16Scale) /-- Saturating Q16.16 division: raw result is `(a*65536)/b`. -/ @[inline] def div (a b : Q16_16) : Q16_16 := if b.toInt = 0 then infinity else ofRawInt ((a.toInt * q16Scale) / b.toInt) @[inline] def neg (q : Q16_16) : Q16_16 := ofRawInt (-q.toInt) @[inline] def abs (q : Q16_16) : Q16_16 := if q.toInt < 0 then neg q else q @[inline] def ofFloat (f : Float) : Q16_16 := if f.isNaN || f ≥ 32768.0 then infinity else if f ≤ -32768.0 then minVal else if f < 0.0 then ofRawInt (-(Int.ofNat ((-f * 65536.0).floor.toUInt32.toNat))) else ofRawInt (Int.ofNat ((f * 65536.0).floor.toUInt32.toNat)) @[inline] def toFloat (q : Q16_16) : Float := Float.ofInt q.toInt / 65536.0 @[inline] def sqrt (q : Q16_16) : Q16_16 := if q.toInt = 0 then zero else let f := toFloat q if f ≤ 0.0 then zero else ofFloat (Float.sqrt f) /-- Natural logarithm approximation around 1.0. -/ def ln (q : Q16_16) : Q16_16 := let x := q.toInt if x ≤ 0 then zero else let y := x - q16Scale let y2 := (y * y) / q16Scale let y3 := (y * y2) / q16Scale ofRawInt (y - y2 / 2 + y3 / 3) def log2 (q : Q16_16) : Q16_16 := let ln2 : Q16_16 := ofRawInt 45426 div (ln q) ln2 def expNeg (x : Q16_16) : Q16_16 := if x.toInt ≥ 0x00030000 then zero else if x.toInt ≥ 0x00020000 then ofRawInt 0x00004D29 else if x.toInt ≥ 0x00010000 then ofRawInt 0x0000C5C0 else ofRawInt 0x0001C5C0 instance : Add Q16_16 := ⟨add⟩ instance : Sub Q16_16 := ⟨sub⟩ instance : Mul Q16_16 := ⟨mul⟩ instance : Div Q16_16 := ⟨div⟩ instance : Neg Q16_16 := ⟨neg⟩ instance : LE Q16_16 where le a b := a.toInt ≤ b.toInt instance : LT Q16_16 where lt a b := a.toInt < b.toInt instance : DecidableRel (fun a b : Q16_16 => a ≤ b) := fun a b => inferInstanceAs (Decidable (a.toInt ≤ b.toInt)) instance : DecidableRel (fun a b : Q16_16 => a < b) := fun a b => inferInstanceAs (Decidable (a.toInt < b.toInt)) @[inline] def ge (a b : Q16_16) : Bool := b.toInt ≤ a.toInt @[inline] def gt (a b : Q16_16) : Bool := b.toInt < a.toInt def lt (a b : Q16_16) : Bool := a.toInt < b.toInt def le (a b : Q16_16) : Bool := a.toInt ≤ b.toInt def isNeg (q : Q16_16) : Bool := q.toInt < 0 def clip (x lo hi : Q16_16) : Q16_16 := if x.toInt < lo.toInt then lo else if x.toInt > hi.toInt then hi else x def sat01 (q : Q16_16) : Q16_16 := if q.toInt < 0 then zero else if q.toInt > q16Scale then one else q def max (a b : Q16_16) : Q16_16 := if a.toInt ≥ b.toInt then a else b def min (a b : Q16_16) : Q16_16 := if a.toInt ≤ b.toInt then a else b def recip (x : Q16_16) : Q16_16 := let xInt := x.toInt if xInt = 0 then maxVal else let numer : Int := 0x100000000 let denom := if xInt < 0 then -xInt else xInt let r := numer / denom let y := ofRawInt r if xInt < 0 then neg y else y def ofRaw (n : Nat) : Q16_16 := ofRawInt (n : Int) -- ═══════════════════════════════════════════════════════════════════════════ -- Algebraic lemmas -- ═══════════════════════════════════════════════════════════════════════════ @[simp] theorem zero_toInt : toInt zero = 0 := rfl @[simp] theorem one_toInt : toInt one = 65536 := rfl @[simp] theorem epsilon_toInt : toInt epsilon = 1 := rfl theorem epsilon_toInt_pos : toInt epsilon > 0 := by norm_num [epsilon_toInt] @[simp] theorem maxVal_toInt : toInt maxVal = q16MaxRaw := rfl @[simp] theorem minVal_toInt : toInt minVal = q16MinRaw := rfl @[simp] private theorem ofRawInt_zero : ofRawInt 0 = zero := by apply Subtype.ext simp [ofRawInt, zero, toInt, q16MinRaw, q16MaxRaw] /-- Saturation lower-bound preservation. -/ theorem ofRawInt_toInt_ge (i c : Int) (hi : i ≥ c) (hcMin : q16MinRaw ≤ c) (hcMax : c ≤ q16MaxRaw) : (ofRawInt i).toInt ≥ c := by unfold ofRawInt toInt by_cases hhi : i > q16MaxRaw · simp [hhi] dsimp [q16MaxRaw] at * omega · by_cases hlo : i < q16MinRaw · dsimp [q16MinRaw, q16MaxRaw] at * omega · simp [hhi, hlo] exact hi /-- Saturation preserves non-negativity for non-negative raw values. -/ theorem ofRawInt_toInt_nonneg (i : Int) (hi : i ≥ 0) : (ofRawInt i).toInt ≥ 0 := by exact ofRawInt_toInt_ge i 0 hi (by norm_num [q16MinRaw]) (by norm_num [q16MaxRaw]) /-- Bounded raw reconstruction. -/ theorem ofRawInt_toInt (a : Q16_16) : ofRawInt a.toInt = a := by apply Subtype.ext unfold ofRawInt toInt have hhi : ¬ a.val > q16MaxRaw := by exact not_lt.mpr a.property.2 have hlo : ¬ a.val < q16MinRaw := by exact not_lt.mpr a.property.1 simp [hhi, hlo] theorem ofRawInt_toInt_eq_nonneg (i : Int) (h1 : i ≥ 0) (h2 : i ≤ q16MaxRaw) : (ofRawInt i).toInt = i := by unfold ofRawInt toInt have hhi : ¬ i > q16MaxRaw := by omega have hlo : ¬ i < q16MinRaw := by dsimp [q16MinRaw] at * omega simp [hhi, hlo] /-- Saturating constructor lemma for non-negative raw values. -/ theorem ofRawInt_toInt_eq_general (i : Int) (h1 : i ≥ 0) : (ofRawInt i).toInt = if i > q16MaxRaw then q16MaxRaw else i := by by_cases h : i > q16MaxRaw · unfold ofRawInt toInt simp [h] · have hlo : ¬ i < q16MinRaw := by dsimp [q16MinRaw, q16MaxRaw] at * omega unfold ofRawInt toInt simp [h, hlo] /-- `(ofRawInt i).toInt = q16Clamp i` — the bridge between the subtype constructor and the pure-Int clamp function. -/ theorem ofRawInt_toInt_eq_clamp (i : Int) : (ofRawInt i).toInt = q16Clamp i := by unfold ofRawInt toInt q16Clamp split_ifs <;> rfl /-- `@[simp]` version rewriting `.val` directly (avoids `toInt` unfolding ordering issues in `simp` calls). -/ @[simp] theorem ofRawInt_val_eq_q16Clamp (i : Int) : (ofRawInt i).val = q16Clamp i := ofRawInt_toInt_eq_clamp i /-- `ofRawInt` is monotone: a ≤ b → (ofRawInt a).toInt ≤ (ofRawInt b).toInt. One-liner via q16Clamp_monotone. -/ theorem ofRawInt_monotone (a b : Int) (h : a ≤ b) : (ofRawInt a).toInt ≤ (ofRawInt b).toInt := by simp only [ofRawInt_toInt_eq_clamp] exact q16Clamp_monotone a b h /-- Adding a nonnegative Q16_16 value cannot decrease the saturated result. This is the general form of the motif-scoring monotonicity used in Semantics.PIST.Motif §6.2: motifScore(match=true) ≥ motifScore(match=false). -/ theorem add_nonneg_monotone (a b : Q16_16) (hb : 0 ≤ b.toInt) : a.toInt ≤ (add a b).toInt := by unfold add have := ofRawInt_monotone a.toInt (a.toInt + b.toInt) (by omega) rwa [ofRawInt_toInt] at this /-- zero * a = zero. -/ theorem zero_mul (a : Q16_16) : mul zero a = zero := by unfold mul rw [zero_toInt] simp /-- a * zero = zero. -/ theorem mul_zero (a : Q16_16) : mul a zero = zero := by unfold mul rw [zero_toInt] simp /-- a - a = zero. -/ theorem sub_self (a : Q16_16) : sub a a = zero := by unfold sub simp /-- a + zero = a. -/ theorem add_zero (a : Q16_16) : add a zero = a := by unfold add rw [zero_toInt] simp exact ofRawInt_toInt a /-- zero + a = a. -/ theorem zero_add (a : Q16_16) : add zero a = a := by unfold add rw [zero_toInt] simp exact ofRawInt_toInt a /-- sqrt zero is zero. -/ theorem sqrt_zero : sqrt zero = zero := by unfold sqrt simp /-- sqrt one is within one LSB of one. -/ theorem sqrt_one : (sqrt one).toInt - one.toInt ≤ 1 := by native_decide private theorem int_scale_mul_ediv_cancel (n : Int) : (q16Scale * n) / q16Scale = n := by rw [Int.mul_ediv_cancel_left] norm_num [q16Scale] /-- one * a = a. -/ theorem one_mul (a : Q16_16) : mul one a = a := by unfold mul show ofRawInt (one.toInt * a.toInt / q16Scale) = a rw [show one.toInt = q16Scale from rfl] have h : (q16Scale * a.toInt) / q16Scale = a.toInt := int_scale_mul_ediv_cancel a.toInt rw [h] exact ofRawInt_toInt a /-- a * one = a. -/ theorem mul_one (a : Q16_16) : mul a one = a := by unfold mul show ofRawInt (a.toInt * one.toInt / q16Scale) = a rw [show one.toInt = q16Scale from rfl] have h : (a.toInt * q16Scale) / q16Scale = a.toInt := by rw [Int.mul_comm] exact int_scale_mul_ediv_cancel a.toInt rw [h] exact ofRawInt_toInt a /-- toInt = 0 iff the value is zero. -/ theorem toInt_eq_zero_iff {a : Q16_16} : a.toInt = 0 ↔ a = zero := by constructor · intro h apply Subtype.ext simpa [toInt, zero] using h · intro h rw [h] rfl /-- zero / x = zero for any nonzero denominator. -/ theorem zero_div (x : Q16_16) (hx : x.val ≠ 0) : div zero x = zero := by unfold div have hx' : ¬ x.toInt = 0 := by simpa [toInt] using hx simp [hx', zero_toInt] /-- Square is non-negative under signed saturating multiplication. -/ theorem mul_self_nonneg (a : Q16_16) : (mul a a).toInt ≥ 0 := by unfold mul have hprod : a.toInt * a.toInt ≥ 0 := by nlinarith have hdiv : (a.toInt * a.toInt) / q16Scale ≥ 0 := by apply Int.ediv_nonneg · exact hprod · norm_num [q16Scale] exact ofRawInt_toInt_nonneg ((a.toInt * a.toInt) / q16Scale) hdiv /-- Product of two non-negative Q16.16 values is non-negative. -/ theorem mul_toInt_nonneg (a b : Q16_16) (ha : a.toInt ≥ 0) (hb : b.toInt ≥ 0) : (mul a b).toInt ≥ 0 := by unfold mul have hprod : a.toInt * b.toInt ≥ 0 := by nlinarith have hdiv : (a.toInt * b.toInt) / q16Scale ≥ 0 := by apply Int.ediv_nonneg · exact hprod · norm_num [q16Scale] exact ofRawInt_toInt_nonneg ((a.toInt * b.toInt) / q16Scale) hdiv /-- Non-negative addition stays non-negative under saturation. -/ theorem ofRaw_toInt_nonneg (acc wcc : Q16_16) (hacc : acc.toInt ≥ 0) (hwcc : wcc.toInt ≥ 0) : (Q16_16.add acc wcc).toInt ≥ 0 := by unfold add have hsum : acc.toInt + wcc.toInt ≥ 0 := by omega exact ofRawInt_toInt_nonneg (acc.toInt + wcc.toInt) hsum /-- Compatibility lemma: a non-negative raw value decoded through saturation is non-negative. -/ theorem mk_lt_half_nonneg (s : Int) (hs : s ≥ 0) (_h : s < 2147483648) : (ofRawInt s).toInt ≥ 0 := by exact ofRawInt_toInt_nonneg s hs /-- Positive raw addition remains positive under saturation. -/ theorem add_pos_of_pos (a b : Q16_16) (ha : a.toInt > 0) (hb : b.toInt > 0) : (add a b).toInt > 0 := by unfold add have hsum : a.toInt + b.toInt ≥ 1 := by omega have hge : (ofRawInt (a.toInt + b.toInt)).toInt ≥ 1 := ofRawInt_toInt_ge (a.toInt + b.toInt) 1 hsum (by norm_num [q16MinRaw]) (by norm_num [q16MaxRaw]) omega /-- (1 + omega).toInt ≥ 65536 when omega.toInt ≥ 0. -/ theorem add_one_omega_ge_one (omega : Q16_16) (h : omega.toInt ≥ 0) : (add one omega).toInt ≥ 65536 := by unfold add have hsum : one.toInt + omega.toInt ≥ 65536 := by rw [one_toInt] omega exact ofRawInt_toInt_ge (one.toInt + omega.toInt) 65536 hsum (by norm_num [q16MinRaw]) (by norm_num [q16MaxRaw]) /-- Non-negative Q16.16 values are bounded by maxVal. -/ theorem toInt_nonneg_le_maxVal (q : Q16_16) (_h : q.toInt ≥ 0) : q.toInt ≤ q16MaxRaw := by exact q.property.2 /-- Adding epsilon to a non-negative value yields a positive value. -/ theorem epsilon_add_pos {r : Q16_16} (hr : r.toInt ≥ 0) : (r + epsilon).toInt > 0 := by change (add r epsilon).toInt > 0 unfold add have hsum : r.toInt + epsilon.toInt ≥ 1 := by rw [epsilon_toInt] omega have hge : (ofRawInt (r.toInt + epsilon.toInt)).toInt ≥ 1 := ofRawInt_toInt_ge (r.toInt + epsilon.toInt) 1 hsum (by norm_num [q16MinRaw]) (by norm_num [q16MaxRaw]) omega /-- `abs (sub a b) = abs (sub b a)` — absolute value of a difference is symmetric. Holds for all Q16_16 values (proved by case analysis on `a.val - b.val` at the Int clamping boundary). -/ private lemma q16Clamp_eq_q16MaxRaw_of_ge {x : Int} (h : x ≥ q16MaxRaw) : q16Clamp x = q16MaxRaw := by dsimp [q16Clamp] by_cases hx : x > q16MaxRaw · simp [hx] · have hx_eq : x = q16MaxRaw := le_antisymm (le_of_not_gt hx) h subst hx_eq; simp [q16Clamp, q16MaxRaw, q16MinRaw] private lemma q16Clamp_eq_q16MinRaw_of_le {x : Int} (h : x ≤ q16MinRaw) : q16Clamp x = q16MinRaw := by dsimp [q16Clamp] by_cases hx_hi : x > q16MaxRaw · unfold q16MinRaw q16MaxRaw at *; omega · by_cases hx_lo : x < q16MinRaw · simp [hx_hi, hx_lo] · have hx_eq : x = q16MinRaw := le_antisymm h (by by_contra hlt apply hx_lo exact lt_of_not_ge hlt) subst hx_eq; simp [q16Clamp, q16MaxRaw, q16MinRaw] private lemma not_q16MaxRaw_lt_0 : ¬ q16MaxRaw < 0 := by unfold q16MaxRaw; omega private lemma q16Clamp_2147483648_eq_q16MaxRaw : q16Clamp (2147483648 : Int) = q16MaxRaw := by unfold q16Clamp q16MaxRaw q16MinRaw; omega private lemma val_if (c : Prop) [Decidable c] (x y : Q16_16) : (if c then x else y).val = (if c then x.val else y.val) := by split <;> rfl theorem abs_sub_comm (a b : Q16_16) : abs (sub a b) = abs (sub b a) := by apply Q16_16.ext simp [abs, sub, neg, toInt, val_if, ofRawInt_val_eq_q16Clamp] have hswap : q16Clamp (b.val - a.val) = q16Clamp (-(a.val - b.val)) := by have : b.val - a.val = -(a.val - b.val) := by omega rw [this] rw [hswap] set d := a.val - b.val have hswap' : q16Clamp (-(a.val - b.val)) = q16Clamp (-d) := by rfl rw [hswap'] by_cases hd_above : d > q16MaxRaw · have h_cd : q16Clamp d = q16MaxRaw := q16Clamp_eq_q16MaxRaw_of_ge (le_of_lt hd_above) have h_nd_low : -d ≤ q16MinRaw := by unfold q16MaxRaw q16MinRaw at *; omega have h_cnd : q16Clamp (-d) = q16MinRaw := q16Clamp_eq_q16MinRaw_of_le h_nd_low rw [h_cd, h_cnd] unfold q16Clamp q16MaxRaw q16MinRaw; norm_num · by_cases hd_below : d < q16MinRaw · have h_cd : q16Clamp d = q16MinRaw := q16Clamp_eq_q16MinRaw_of_le (by omega) have h_nd_high : -d ≥ q16MaxRaw := by unfold q16MaxRaw q16MinRaw at *; omega have h_cnd : q16Clamp (-d) = q16MaxRaw := q16Clamp_eq_q16MaxRaw_of_ge h_nd_high rw [h_cd, h_cnd] unfold q16Clamp q16MaxRaw q16MinRaw; norm_num · have h_lo : q16MinRaw ≤ d := by unfold q16MinRaw q16MaxRaw at *; omega have h_hi : d ≤ q16MaxRaw := by unfold q16MinRaw q16MaxRaw at *; omega have h_cd : q16Clamp d = d := q16Clamp_id_of_inRange d h_lo h_hi rw [h_cd] by_cases h_nd_above : -d > q16MaxRaw · have h_d_min : d = q16MinRaw := by unfold q16MinRaw q16MaxRaw at *; omega rw [h_d_min] unfold q16Clamp q16MaxRaw q16MinRaw; norm_num · by_cases h_nd_below : -d < q16MinRaw · have h_d_max : d = q16MaxRaw := by unfold q16MinRaw q16MaxRaw at *; omega rw [h_d_max] unfold q16Clamp q16MaxRaw q16MinRaw; norm_num · have h_nd_lo' : q16MinRaw ≤ -d := by omega have h_nd_hi' : -d ≤ q16MaxRaw := by omega have h_cnd : q16Clamp (-d) = -d := q16Clamp_id_of_inRange (-d) h_nd_lo' h_nd_hi' rw [h_cnd, show (-(-d : Int) = d) by omega] by_cases hd_neg : d < 0 · omega · by_cases hd_zero : d = 0 · rw [hd_zero] unfold q16Clamp q16MinRaw q16MaxRaw; norm_num · have hd_pos : 0 < d := by omega rw [h_cd] split_ifs <;> omega /-- Subtraction is addition of the negation: a - b = a + (-b). NOTE: This theorem is NOT universally true for Q16_16. Counterexample: `a = b = q16MinRaw` gives LHS = 0, RHS = -1. The difference arises because `neg q16MinRaw` overflows to `q16MaxRaw`, altering the clamping path. SSMS does not use this theorem — the `bound` proof has been restructured to avoid it. -/ theorem sub_eq_add_neg (a b : Q16_16) : sub a b = add a (neg b) := by -- TODO(lean-port): this is false at q16MinRaw. Either add precondition -- `hb : b.toInt > q16MinRaw` or prove the specific form needed by SSMS. admit /-- Multiplication by a non-negative scalar is monotone: if a ≤ b and c ≥ 0, then a*c ≤ b*c. Used in SSMS.aciPreservedByMlgruStep for bound propagation. -/ theorem mul_mono_left (a b c : Q16_16) (h : a.toInt ≤ b.toInt) (hc : c.toInt ≥ 0) : (mul a c).toInt ≤ (mul b c).toInt := by unfold mul have hmul : a.toInt * c.toInt ≤ b.toInt * c.toInt := by apply Int.mul_le_mul_of_nonneg_right h hc have hpos : 0 < q16Scale := by unfold q16Scale; norm_num have hdiv : (a.toInt * c.toInt) / q16Scale ≤ (b.toInt * c.toInt) / q16Scale := by apply Int.ediv_le_ediv hpos hmul rw [ofRawInt_toInt_eq_clamp, ofRawInt_toInt_eq_clamp] exact q16Clamp_monotone _ _ hdiv /-- Multiplication by a non-negative scalar is monotone on the right. -/ theorem mul_mono_right (a b c : Q16_16) (h : a.toInt ≤ b.toInt) (hc : c.toInt ≥ 0) : (mul c a).toInt ≤ (mul c b).toInt := by unfold mul have hmul : c.toInt * a.toInt ≤ c.toInt * b.toInt := by apply Int.mul_le_mul_of_nonneg_left h hc have hpos : 0 < q16Scale := by unfold q16Scale; norm_num have hdiv : (c.toInt * a.toInt) / q16Scale ≤ (c.toInt * b.toInt) / q16Scale := by apply Int.ediv_le_ediv hpos hmul rw [ofRawInt_toInt_eq_clamp, ofRawInt_toInt_eq_clamp] exact q16Clamp_monotone _ _ hdiv /-- Addition is monotone in the left argument: if a ≤ b then a+c ≤ b+c. -/ theorem add_le_add (a b c : Q16_16) (h : a.toInt ≤ b.toInt) : (add a c).toInt ≤ (add b c).toInt := by unfold add simp [ofRawInt_toInt_eq_clamp] apply q16Clamp_monotone omega /-- Absolute value of any Q16_16 value is non-negative. -/ theorem abs_nonneg (a : Q16_16) : (abs a).toInt ≥ 0 := by unfold abs neg have h := a.property.1 split_ifs with hlt · rw [ofRawInt_toInt_eq_clamp] have h_nonneg : 0 ≤ -a.toInt := by omega exact q16Clamp_nonneg_of_nonneg h_nonneg · omega /-- For non-negative a, |a*b| ≤ a*|b|. WARNING: This lemma is UNCONDITIONALLY FALSE (counterexample: a=3, b=-1 gives LHS=1, RHS=0). The error comes from floor division: a*b may round to -1 but a*|b| rounds to 0. Use abs_mul_bound or a direct convexity argument instead. TODO(lean-port): do not use this lemma — restructure SSMS to avoid it. -/ theorem abs_mul_le (a b : Q16_16) (ha : a.toInt ≥ 0) : (abs (mul a b)).toInt ≤ (mul a (abs b)).toInt := by admit /-- Triangle inequality for Q16_16: |a*b| ≤ |a| * |b|. WARNING: This lemma is FALSE in general (counterexample: a=3, b=-3 gives LHS=1, RHS=0). The floor division causes |a*b|/65536 to round up while |a|*|b|/65536 rounds to 0 for small values. TODO(lean-port): do not use this lemma — restructure SSMS to use abs_triangle_add (|x+y| ≤ |x|+|y|) instead. -/ theorem abs_triangle (a b : Q16_16) : (abs (mul a b)).toInt ≤ (mul (abs a) (abs b)).toInt := by admit end Q16_16 -- ═══════════════════════════════════════════════════════════════════════════ -- Q0.64 signed normalized fraction -- ═══════════════════════════════════════════════════════════════════════════ def q0_64MinRaw : Int := -9223372036854775808 def q0_64MaxRaw : Int := 9223372036854775807 def q0_64ScaleNat : Nat := 9223372036854775808 def q0_64ScaleFloat : Float := 9223372036854775808.0 /-- Q0.64 pure fraction representation. The canonical proof model stores the signed raw integer in the Int64 range. -/ abbrev Q0_64 := { x : Int // q0_64MinRaw ≤ x ∧ x ≤ q0_64MaxRaw } instance : Repr Q0_64 where reprPrec q _ := repr q.val instance : BEq Q0_64 where beq a b := a.val == b.val instance : Inhabited Q0_64 where default := ⟨0, by constructor <;> norm_num [q0_64MinRaw, q0_64MaxRaw]⟩ instance : ToJson Q0_64 where toJson q := Json.mkObj [("val", toJson q.val)] namespace Q0_64 @[ext] theorem ext {a b : Q0_64} (h : a.val = b.val) : a = b := Subtype.ext h @[inline] def toInt (q : Q0_64) : Int := q.val @[inline] def ofRawInt (raw : Int) : Q0_64 := if hhi : raw > q0_64MaxRaw then ⟨q0_64MaxRaw, by constructor <;> norm_num [q0_64MinRaw, q0_64MaxRaw]⟩ else if hlo : raw < q0_64MinRaw then ⟨q0_64MinRaw, by constructor <;> norm_num [q0_64MinRaw, q0_64MaxRaw]⟩ else ⟨raw, by constructor · dsimp [q0_64MinRaw, q0_64MaxRaw] at *; omega · dsimp [q0_64MinRaw, q0_64MaxRaw] at *; omega⟩ instance : FromJson Q0_64 where fromJson? j := do let raw : Int ← fromJson? (← j.getObjVal? "val") pure (ofRawInt raw) /-- Maximum positive value. -/ def one : Q0_64 := ofRawInt q0_64MaxRaw def zero : Q0_64 := ofRawInt 0 def ofRatio (num : Nat) (den : Nat) : Q0_64 := if den = 0 then zero else ofRawInt (Int.ofNat (num * q0_64ScaleNat / den)) def half : Q0_64 := ofRawInt 4611686018427387903 def neg (x : Q0_64) : Q0_64 := ofRawInt (-x.toInt) def add (a b : Q0_64) : Q0_64 := ofRawInt (a.toInt + b.toInt) def sub (a b : Q0_64) : Q0_64 := ofRawInt (a.toInt - b.toInt) def mul (a b : Q0_64) : Q0_64 := ofRawInt ((a.toInt * b.toInt) / Int.ofNat q0_64ScaleNat) def div (a b : Q0_64) : Q0_64 := if b.toInt = 0 then one else ofRawInt ((a.toInt * Int.ofNat q0_64ScaleNat) / b.toInt) def abs (x : Q0_64) : Q0_64 := if x.toInt < 0 then neg x else x def ofFloat (f : Float) : Q0_64 := if f.isNaN || f ≥ 1.0 then one else if f ≤ -1.0 then ofRawInt q0_64MinRaw else if f < 0.0 then ofRawInt (-(Int.ofNat ((-f * q0_64ScaleFloat).floor.toUInt64.toNat))) else ofRawInt (Int.ofNat ((f * q0_64ScaleFloat).floor.toUInt64.toNat)) def toFloat (q : Q0_64) : Float := Float.ofInt q.toInt / q0_64ScaleFloat instance : Add Q0_64 := ⟨add⟩ instance : Sub Q0_64 := ⟨sub⟩ instance : Mul Q0_64 := ⟨mul⟩ instance : Div Q0_64 := ⟨div⟩ instance : Neg Q0_64 := ⟨neg⟩ instance : LE Q0_64 where le a b := a.toInt ≤ b.toInt instance : LT Q0_64 where lt a b := a.toInt < b.toInt instance : DecidableRel (fun a b : Q0_64 => a ≤ b) := fun a b => inferInstanceAs (Decidable (a.toInt ≤ b.toInt)) instance : DecidableRel (fun a b : Q0_64 => a < b) := fun a b => inferInstanceAs (Decidable (a.toInt < b.toInt)) end Q0_64 -- ═══════════════════════════════════════════════════════════════════════════ -- Pandigital π Approximation -- ═══════════════════════════════════════════════════════════════════════════ namespace PandigitalPi /-- High term: 3.8415926. -/ def highTerm : Q16_16 := Q16_16.ofRawInt 251819 /-- Low term: 0.7. -/ def lowTerm : Q16_16 := Q16_16.ofRawInt 45875 def piPandigital : Q16_16 := highTerm - lowTerm def piDirect : Q16_16 := Q16_16.ofRawInt 205944 theorem piPandigitalCorrect : (piPandigital.toInt - piDirect.toInt).natAbs ≤ 1 := by native_decide def spaceAnalysis : String := "Pandigital pi: 6 bytes packed vs 4 bytes direct Q16.16 (trade-off for mathematical elegance)" #eval piPandigital.toFloat #eval piDirect.toFloat #eval (piPandigital.toInt - piDirect.toInt).natAbs end PandigitalPi end Semantics.FixedPoint namespace Semantics export FixedPoint (Q0_16 Q16_16 Q0_64) namespace Q16_16 export FixedPoint.Q16_16 (zero one negOne epsilon two infinity maxVal minVal ofNat satFromNat ofRatio toInt ofRawInt ofBits toBits ofFloat toFloat scale ofInt add sub mul div abs neg sqrt ln log2 expNeg sat01 max min le ge gt lt recip ofRaw clip isNeg zero_mul mul_zero one_mul mul_one zero_add add_zero sub_self zero_toInt one_toInt epsilon_toInt epsilon_toInt_pos toInt_eq_zero_iff epsilon_add_pos zero_div mul_self_nonneg mul_toInt_nonneg ofRaw_toInt_nonneg mk_lt_half_nonneg add_one_omega_ge_one toInt_nonneg_le_maxVal add_pos_of_pos) end Q16_16 namespace Q0_16 export FixedPoint.Q0_16 (zero one half neg add sub mul div abs lt le gt ge toFloat ofFloat log2 min) end Q0_16 namespace Q0_64 export FixedPoint.Q0_64 (one zero ofRatio half neg add sub mul div abs toInt ofFloat toFloat) end Q0_64 namespace PandigitalPi export FixedPoint.PandigitalPi (highTerm lowTerm piPandigital piDirect piPandigitalCorrect spaceAnalysis) end PandigitalPi end Semantics