# Derivation — exact capped moments and the deleted-base-cutoff correction Companion to `report.md` (route 157). Notation as in the engine's `u = t/A` coordinates: support radius `U = 1 + ε`, base cutoff `ℓ = 1 − ε`, rough threshold `d`, positive caps `c_m`. All quantities rational. ## §1 Dirichlet / Beta moments (Lemma 4.19) For `n ≥ 0`, `V > 0`, `b ≥ 0` and an exponent partition `α` with `ℓ(α) ≤ n`, ∫_{s(t) ≤ V} (V − s(t))^b m_α(t) dt = V^{n+b+|α|} · b! · emb_n(α) · ∏_{q∈α} q! / (n+b+|α|)! , emb_n(α) = n! / ((n − ℓ(α))! ∏_q m_q(α)!) . `uncapped_integral(n, e, α, V)` implements this. Verified exactly against the `EvenEngine` exact `M₁` for `k = 4, 5, 6` (`tests/test_capped_forms.py`). ## §2 Capped moment (Eqs. (76)–(77)) Target: I(e, ν) = ∫_T (U − s(t))^e m_ν(t) dt , T = { t ≥ 0 : s(t) < U, Σ_{tᵢ>d} tᵢ ≤ c_{r(t)} } . Split by `r = r(t)`. By symmetry, fix the rough set as the first `r` coordinates, shift `tᵢ = d + yᵢ` (`i ≤ r`), keep `vᵢ = t_{r+i} ∈ [0,d]`. Then, writing `Y = Σ yᵢ`, `V = Σ vᵢ`, domain: Y ≤ c_r − r d , Y + V ≤ U − r d , slack: U − r d − Y − V . The coefficient counting pairs (rough set, exponent embedding) with `u_q` occurrences of part `q` placed rough is (`u = Σ u_q`) emb_n(ν) · C(n − ℓ(ν), r − u) · ∏_q C(m_q(ν), u_q) (76) This is the number of (rough set, embedding) pairs; summing over `u_q` gives `emb_n(ν) · C(n, r)`, consistent with `C(n,r)` choices of rough set. Each rough occurrence contributes `(d + y)^q`, each small one `(d + v)^q`; both are expanded. The small box `[0,d]^{n−r}` is removed by inclusion–exclusion: for a set of `j` small coordinates shifted by `d`, with sign `(−1)^j`, the residual budget becomes `R′ = U − (r+j)d`. A coordinate with exponent `0` contributes the factor `(1 − t)`. After the small Dirichlet integral (dimension `m = n − r`, exponents `h′`): ∫ (R′ − Y − V)^e v^{h′} dv = e! · ∏_i h′_i! / p! · (R′ − Y)^p , p = e + m + |h′| . The rough integral, with `C = min(c_r − r d, R′)` and rough exponents `g`, `|g| = Σ gᵢ`, is (`r ≥ 1`) ∏_i gᵢ! · Σ_{q=0}^{p} C(p,q) (R′ − C)^{p−q} · q! · C^{r+|g|+q} / (r+|g|+q)! (77) and for `r = 0` the expression is evaluated at `Y = 0` (value `(R′)^p`). Empty strata contribute zero; `r d > c_r` terminates the `r` sum. `capped_integral` implements §2. Controls: §1 recovery when the cap is inert (`d = U`, `c ≡ 0`), Monte-Carlo agreement at `n = 3`, monotonicity in `c`. ## §3 Product expansion `m_λ m_μ = Σ_ν c^ν_{λμ} m_ν`. To fix one monomial of type `ν`, choose for each of its `ℓ(ν)` coordinates a pair `(p,q) ∈ (supp λ ∪ {0}) × (supp μ ∪ {0})` with `p + q = νᵢ`, using each part of `λ` and `μ` with its prescribed multiplicity. Enforcing a non-increasing exponent sequence counts each monomial once and gives `c^ν_{λμ}` directly (`capped_forms.mul_m`). Dimension-independent for `ℓ(ν) ≤ n`. ## §4 Cap constants and the base-cutoff gap Candidate (`Althoefer`, H₁ ≤ 216): A = 2583/10000 , ε_s = 3/400 , δ = 3/250 , U = A + ε_s = 1329/5000 = 0.2658 , ℓ = A − ε_s = 627/2500 = 0.2508 , d = δ/A = 40/861 , c₁ = c₂ = (3/20)/A = 500/861 , c_m = (4/25)/A = 1600/2583 (m ≥ 3) , 1/A = 10000/2583 . Source eprint 2026/1893, Appendix B: `u = 2605/10000`, `ℓ = 2455/10000`, `d = 182/10000`, `b = 1899/10000`; hence `u − d = 2423 < 2455 = ℓ`. The gap separating the two: u − d − ℓ = 2 ε_s − δ . For the source `2 ε_s = 150 < d = 182`, so the gap is negative (`= −32/10000`). For the candidate `2 ε_s = 3/200 > δ = 3/250`, so the gap is positive (`= 3/1000`). Therefore on the candidate the source's inclusion `C_r, D_r ⊆ {s ≤ ℓ}` fails, and the correction domains must be intersected with the base cutoff. In the shifted coordinates the missing condition is (r + j) d + Y + Z ≤ ℓ ⟺ Y + Z ≤ ℓ − (r + j) d . This is the only structural change needed in Appendix B for the candidate; the integrable form of `C_r, D_r, E_r` is otherwise unchanged, because `R_r = r d + Y` and `V_r = j d + Z` make every boundary axis-aligned in `(Y, Z)`. Machine-checked in `lean/CapBaseCutoff.lean` and `tests/test_cap_base_cutoff.py`.