AFTD Auto-Formalizing Theoretical Domains

concave_power_sum_le

theoremverifiedformalization

For non-negative real numbers p and q such that p^2 + q^2 <= 1, and any real r with 0 < r <= 1, the sum of powers satisfies p^r + q^r <= 2^(1 - r / 2).

Statement

theorem concave_power_sum_le
    (p q r : ℝ) (hp : 0 ≤ p) (hq : 0 ≤ q) (hpq : p ^ 2 + q ^ 2 ≤ 1)
    (hr_pos : 0 < r) (hr_le : r ≤ 1) :
    p ^ r + q ^ r ≤ (2 : ℝ) ^ (1 - r / 2)
Formalizes
arXiv:2610.35838 Appendix E, Fact 2
Verified
01 Nov 2026
Axioms
Classical.choiceQuot.soundpropext
Used by 1

Read back from the Lean

For all real numbers $p$, $q$, and $r$, if $p \ge 0$, $q \ge 0$, $p^2 + q^2 \le 1$, $r > 0$, and $r \le 1$, then $p^r + q^r \le 2^{1 - r/2}$.

Written by a model that saw only the Lean, never the English above. If the two disagree, that is worth a challenge.

The proof, explained

Setting s = r / 2, the exponent satisfies 0 ≤ s ≤ 1, so the function t ↦ t^s is concave on [0, ∞). Applying concavity to p^2 and q^2 with equal weights 1/2 gives (p^r + q^r) / 2 = ((p^2)^s + (q^2)^s) / 2 ≤ ((p^2 + q^2) / 2)^s. Since p^2 + q^2 ≤ 1, monotonicity of power functions bounds this by (1/2)^s, which simplifies via exponent arithmetic to 2^(1 - r/2).

  1. Define s = r / 2 and show 0 < s, 0 ≤ s, and s ≤ 1.

    Set s = r / 2. Since r > 0, positivity shows 0 < s, which implies 0 ≤ s by le_of_lt. Nonlinear arithmetic establishes s = r / 2 ≤ 1 from the assumption r ≤ 1.

    lines 6–9

    uses le_of_lt

      set s := r / 2 with hs
      have hs_pos : 0 < s := by positivity
      have hs_nonneg : 0 ≤ s := le_of_lt hs_pos
      have hs_le : s ≤ 1 := by nlinarith
  2. (1 / 2) * (p^2)^s + (1 / 2) * (q^2)^s ≤ ((p^2 + q^2) / 2)^s

    Since 0 ≤ s ≤ 1, Real.concaveOn_rpow shows that x ↦ x^s is concave on [0, ∞). Evaluating concavity at the non-negative points p^2 and q^2 (using sq_nonneg) with equal convex weights a = 1/2 and b = 1/2 gives (1/2) * (p^2)^s + (1/2) * (q^2)^s ≤ ((1/2) * p^2 + (1/2) * q^2)^s. Ring simplification identifies the right-hand base with (p^2 + q^2) / 2.

    lines 10–18

    uses Real.concaveOn_rpow, sq_nonneg

      have hconc := Real.concaveOn_rpow (p := s) hs_nonneg hs_le
      obtain ⟨_, hconc'⟩ := hconc
      have hconc_apply := hconc' (x := p ^ 2) (by simpa using sq_nonneg p)
        (y := q ^ 2) (by simpa using sq_nonneg q)
        (a := (1 / 2 : ℝ)) (b := (1 / 2 : ℝ))
        (by norm_num) (by norm_num) (by norm_num : (1 : ℝ) / 2 + 1 / 2 = 1)
      norm_num at hconc_apply
      rw [show (1 / 2 : ℝ) * p ^ 2 + 1 / 2 * q ^ 2 = (p ^ 2 + q ^ 2) / 2 by ring]
        at hconc_apply
  3. p^r + q^r ≤ 2 * ((p^2 + q^2) / 2)^s

    Since r = 2 * s, the power rule Real.rpow_natCast_mul rewrites p^r = (p^2)^s and q^r = (q^2)^s for the non-negative bases p and q. Multiplying the concavity inequality by 2 gives p^r + q^r ≤ 2 * ((p^2 + q^2) / 2)^s.

    lines 19–22

    uses Real.rpow_natCast_mul

      have hsum : p ^ r + q ^ r ≤ 2 * ((p ^ 2 + q ^ 2) / 2) ^ s := by
        rw [show r = ((2 : ℕ) : ℝ) * s by rw [hs]; norm_num; ring,
          Real.rpow_natCast_mul hp 2 s, Real.rpow_natCast_mul hq 2 s]
        linarith
  4. 2 * ((p^2 + q^2) / 2)^s ≤ 2 * (1 / 2)^s

    From p^2 + q^2 ≤ 1 and positivity, we have 0 ≤ (p^2 + q^2) / 2 ≤ 1 / 2. Because s ≥ 0, power monotonicity (Real.rpow_le_rpow) yields ((p^2 + q^2) / 2)^s ≤ (1 / 2)^s. Multiplying by 2 preserves the inequality.

    lines 23–28

    uses Real.rpow_le_rpow

      have hbase_nonneg : 0 ≤ (p ^ 2 + q ^ 2) / 2 := by positivity
      have hbase_le : (p ^ 2 + q ^ 2) / 2 ≤ 1 / 2 := by linarith
      have hmono : ((p ^ 2 + q ^ 2) / 2) ^ s ≤ (1 / 2 : ℝ) ^ s :=
        Real.rpow_le_rpow hbase_nonneg hbase_le hs_nonneg
      have hmul : 2 * ((p ^ 2 + q ^ 2) / 2) ^ s ≤ 2 * (1 / 2 : ℝ) ^ s := by
        nlinarith
  5. 2 * (1 / 2)^s = 2^(1 - s)

    Distributing the power over the fraction with Real.div_rpow gives (1 / 2)^s = 1^s / 2^s = (2^s)⁻¹, which is 2^(-s) by Real.rpow_neg. Rewriting 2 as 2^1 using Real.rpow_one and applying Real.rpow_add combines the factors into 2^(1 + -s) = 2^(1 - s).

    lines 29–35

    uses Real.div_rpow, Real.rpow_neg, Real.rpow_one, Real.rpow_add

      have htwo : 2 * (1 / 2 : ℝ) ^ s = 2 ^ (1 - s) := by
        rw [Real.div_rpow (by norm_num) (by norm_num)]
        norm_num
        rw [← Real.rpow_neg (by norm_num : (0 : ℝ) ≤ 2) s]
        nth_rewrite 1 [← Real.rpow_one (2 : ℝ)]
        rw [← Real.rpow_add (by norm_num : (0 : ℝ) < 2) 1 (-s)]
        ring_nf
  6. p^r + q^r ≤ 2^(1 - r / 2)

    A calculation chain combines the previous results: p^r + q^r ≤ 2 * ((p^2 + q^2) / 2)^s ≤ 2 * (1 / 2)^s = 2^(1 - s). Replacing s with r / 2 finishes the proof.

    lines 36–40

      calc
        p ^ r + q ^ r ≤ 2 * ((p ^ 2 + q ^ 2) / 2) ^ s := hsum
        _ ≤ 2 * (1 / 2 : ℝ) ^ s := hmul
        _ = 2 ^ (1 - s) := htwo
        _ = (2 : ℝ) ^ (1 - r / 2) := by rw [hs]
le_of_lt
In any preorder, a strict inequality a < b implies a ≤ b.
Real.concaveOn_rpow
For 0 ≤ p ≤ 1, the power function x ↦ x^p is concave on [0, ∞).
sq_nonneg
The square of any real number is non-negative.
Real.rpow_natCast_mul
For x ≥ 0, natural number n, and real z, x^(n * z) = (x^n)^z.
Real.rpow_le_rpow
For 0 ≤ x ≤ y and non-negative z, x^z ≤ y^z.
Real.div_rpow
For non-negative x and y, (x / y)^z = x^z / y^z.
Real.rpow_neg
For x ≥ 0, x^(-y) = (x^y)⁻¹.
Real.rpow_one
For any real x, x^1 = x.
Real.rpow_add
For x > 0, x^(y + z) = x^y * x^z.

Written by a model (gemini-3.8-flash) from the Lean proof and the goal states Lean reported at each step. Lean verified the proof; this explanation is a reading aid and is not itself machine-checked.

Lean source view module on GitHub

/-- Fact 2: Power sum inequality p^r + q^r <= 2^(1 - r/2) from concavity of t |-> t^(r/2). -/theorem concave_power_sum_le    (p q r : ℝ) (hp : 0 ≤ p) (hq : 0 ≤ q) (hpq : p ^ 2 + q ^ 2 ≤ 1)    (hr_pos : 0 < r) (hr_le : r ≤ 1) :    p ^ r + q ^ r ≤ (2 : ℝ) ^ (1 - r / 2) := by  set s := r / 2 with hs  have hs_pos : 0 < s := by positivity  have hs_nonneg : 0 ≤ s := le_of_lt hs_pos  have hs_le : s ≤ 1 := by nlinarith  have hconc := Real.concaveOn_rpow (p := s) hs_nonneg hs_le  obtain ⟨_, hconc'⟩ := hconc  have hconc_apply := hconc' (x := p ^ 2) (by simpa using sq_nonneg p)    (y := q ^ 2) (by simpa using sq_nonneg q)    (a := (1 / 2 : ℝ)) (b := (1 / 2 : ℝ))    (by norm_num) (by norm_num) (by norm_num : (1 : ℝ) / 2 + 1 / 2 = 1)  norm_num at hconc_apply  rw [show (1 / 2 : ℝ) * p ^ 2 + 1 / 2 * q ^ 2 = (p ^ 2 + q ^ 2) / 2 by ring]    at hconc_apply  have hsum : p ^ r + q ^ r ≤ 2 * ((p ^ 2 + q ^ 2) / 2) ^ s := by    rw [show r = ((2 : ℕ) : ℝ) * s by rw [hs]; norm_num; ring,      Real.rpow_natCast_mul hp 2 s, Real.rpow_natCast_mul hq 2 s]    linarith  have hbase_nonneg : 0 ≤ (p ^ 2 + q ^ 2) / 2 := by positivity  have hbase_le : (p ^ 2 + q ^ 2) / 2 ≤ 1 / 2 := by linarith  have hmono : ((p ^ 2 + q ^ 2) / 2) ^ s ≤ (1 / 2 : ℝ) ^ s :=    Real.rpow_le_rpow hbase_nonneg hbase_le hs_nonneg  have hmul : 2 * ((p ^ 2 + q ^ 2) / 2) ^ s ≤ 2 * (1 / 2 : ℝ) ^ s := by    nlinarith  have htwo : 2 * (1 / 2 : ℝ) ^ s = 2 ^ (1 - s) := by    rw [Real.div_rpow (by norm_num) (by norm_num)]    norm_num    rw [← Real.rpow_neg (by norm_num : (0 : ℝ) ≤ 2) s]    nth_rewrite 1 [← Real.rpow_one (2 : ℝ)]    rw [← Real.rpow_add (by norm_num : (0 : ℝ) < 2) 1 (-s)]    ring_nf  calc    p ^ r + q ^ r ≤ 2 * ((p ^ 2 + q ^ 2) / 2) ^ s := hsum    _ ≤ 2 * (1 / 2 : ℝ) ^ s := hmul    _ = 2 ^ (1 - s) := htwo    _ = (2 : ℝ) ^ (1 - r / 2) := by rw [hs]
See it in the graphDiscuss this resultChallenge it