McCrea's conjecture

What Lean checks

Complete proposed proof of the unique linear-density crossover for every modulus at least three. Positive coefficient asymptotics handle the floor, and the exponent difference has exactly one zero in the stated interval. Independent internal review and exact profile-word diagnostics passed.

Complete proposed proof, 21 November 2026. The complete written argument passed independent internal review. The priority audit records the sources checked and access limits. This work remains unrefereed. Coordination: issue 168.

1. Statement

For a nonempty integer partition π\pi, let its perimeter be its largest part plus its number of parts minus one. Let FDj,k(n)\mathrm{FD}_{j,k}(n) count partitions of perimeter nn having exactly jj distinct part sizes of multiplicity at least kk. Let FOj,k(n)\mathrm{FO}_{j,k}(n) count partitions of perimeter nn having exactly jj distinct occurring part sizes divisible by kk.

Marcus McCrea, Fixed-Perimeter Franklin Statistics and an Eventual Inequality, arXiv:2609.12005v2, dated 5 November 2026, Section 9, Conjecture 9.2, proposes the following statement. The full primary PDF was inspected.

Theorem. Fix an integer k≥3k\ge3, and put

ak=2−k−1,bk=14k.a_k=2^{-k-1},\qquad b_k=\frac1{4k}.

There exists a unique ck∈(ak,bk)c_k\in(a_k,b_k) such that, for every fixed x∈(ak,bk)\{ck}x\in(a_k,b_k)\setminus\{c_k\}, the following inequalities hold for all sufficiently large integers nn:

{FD⌊xn⌋,k(n)>FO⌊xn⌋,k(n),ak<x<ck,FD⌊xn⌋,k(n)<FO⌊xn⌋,k(n),ck<x<bk.\begin{cases} \mathrm{FD}_{\lfloor xn\rfloor,k}(n)> \mathrm{FO}_{\lfloor xn\rfloor,k}(n),&a_k<x<c_k,\\[2mm] \mathrm{FD}_{\lfloor xn\rfloor,k}(n)< \mathrm{FO}_{\lfloor xn\rfloor,k}(n),&c_k<x<b_k. \end{cases}

This is the full conjecture. It does not prescribe the comparison at x=ckx=c_k or outside the interval between the two means. The previously proved fixed-jj inequality is a different statement and is not counted here.

2. Positive generating functions

Fix kk throughout, and write

S(q)=q+q2+⋯+qk,H(q)=1−2q+qk+1=(1−q)(1−S(q)).S(q)=q+q^2+\cdots+q^k, \qquad H(q)=1-2q+q^{k+1}=(1-q)(1-S(q)).

The source's Proposition 3.1, equation (7), gives, for j≥1j\ge1,

∑n≥1FDj,k(n)qn=q(k+1)j−1(1−q)2H(q)j+1=AD(q)FD(q)j−1,\sum_{n\ge1}\mathrm{FD}_{j,k}(n)q^n =\frac{q^{(k+1)j-1}(1-q)^2}{H(q)^{j+1}} =A_D(q)F_D(q)^{j-1}, \tag{1}

where

FD(q)=qk+1(1−q)(1−S(q)),AD(q)=qk(1−S(q))2.F_D(q)=\frac{q^{k+1}}{(1-q)(1-S(q))}, \qquad A_D(q)=\frac{q^k}{(1-S(q))^2}. \tag{2}

For the other statistic, put

W(q)=q1−q,U(q)=∑i=1k−1W(q)i,V(q)=W(q)k,T(q)=qV(q),W(q)=\frac{q}{1-q},\quad U(q)=\sum_{i=1}^{k-1}W(q)^i,\quad V(q)=W(q)^k,\quad T(q)=qV(q),

g(q)=V(q)(1−q)=qk(1−q)k−1,K(q)=1−g(q).g(q)=V(q)(1-q)=\frac{q^k}{(1-q)^{k-1}}, \qquad K(q)=1-g(q).

The source's Proposition 3.2, equation (9), gives

∑n≥1FOj,k(n)qn=U(q)T(q)jK(q)j+1+V(q)T(q)j−1K(q)j=AO(q)FO(q)j−1,\begin{aligned} \sum_{n\ge1}\mathrm{FO}_{j,k}(n)q^n &=\frac{U(q)T(q)^j}{K(q)^{j+1}} +\frac{V(q)T(q)^{j-1}}{K(q)^j}\\ &=A_O(q)F_O(q)^{j-1}, \tag{3} \end{aligned}

where

FO(q)=T(q)K(q),AO(q)=U(q)T(q)K(q)2+V(q)K(q).F_O(q)=\frac{T(q)}{K(q)},\qquad A_O(q)=\frac{U(q)T(q)}{K(q)^2}+\frac{V(q)}{K(q)}. \tag{4}

All four series AD,FD,AO,FOA_D,F_D,A_O,F_O have nonnegative coefficients. Both FF series have positive coefficients at every power qℓq^\ell with ℓ≥k+1\ell\ge k+1, and no smaller powers. For FDF_D this follows from the factor 1/(1−q)1/(1-q) and the expansion of 1/(1−S)1/(1-S). For FOF_O, the numerator T=qk+1/(1−q)kT=q^{k+1}/(1-q)^k already has this support, and 1/K=∑h≥0gh1/K=\sum_{h\ge0}g^h has nonnegative coefficients and constant term one.

There is a unique RD∈(0,1)R_D\in(0,1) with S(RD)=1S(R_D)=1 and a unique RO∈(0,1)R_O\in(0,1) with g(RO)=1g(R_O)=1. Positivity of S′S' and g′g' makes these zeros simple. Each FiF_i is analytic on |q|<Ri|q|<R_i and has a simple pole at the positive real point RiR_i. Each AiA_i is analytic on the same open disk. Indeed, for |q|<RD|q|<R_D one has |S(q)|≤S(|q|)<1|S(q)|\le S(|q|)<1, and the analogous inequality holds for gg on |q|<RO|q|<R_O. Their remaining possible poles occur at one, beyond the stated radii.

3. A coefficient lemma, including the floor

We use an elementary large-power coefficient estimate and give its proof to make the local, rather than merely cumulative, comparison explicit.

Lemma. Let

F(q)=∑ℓ≥sfℓqℓ,fℓ>0(ℓ≥s),F(q)=\sum_{\ell\ge s}f_\ell q^\ell, \qquad f_\ell>0\quad(\ell\ge s),

have positive radius of convergence RR, and suppose FF has a simple pole at RR. Let AA be a nonzero power series with nonnegative coefficients, analytic on |q|<R|q|<R. For 0<x<1/s0<x<1/s, there exists a unique r=r(x)∈(0,R)r=r(x)\in(0,R) such that

M(r):=rF′(r)F(r)=1x.M(r):=\frac{rF'(r)}{F(r)}=\frac1x. \tag{5}

Set mn=⌊xn⌋−1m_n=\lfloor xn\rfloor-1. Then, as n→∞n\to\infty,

[qn]A(q)F(q)mn∼A(r)F(r)mnr−n2πmnσ(r)2,σ(r)2=rM′(r)>0.[q^n]A(q)F(q)^{m_n} \sim\frac{A(r)F(r)^{m_n}r^{-n}} {\sqrt{2\pi m_n\sigma(r)^2}}, \qquad \sigma(r)^2=rM'(r)>0. \tag{6}

In particular, these coefficients are positive for every sufficiently large nn, and

limn→∞1nlog([qn]AFmn)=Φ(x):=−logr(x)+xlogF(r(x)).\lim_{n\to\infty}\frac1n\log([q^n]AF^{m_n}) =\Phi(x):=-\log r(x)+x\log F(r(x)). \tag{7}

Proof. For 0<r<R0<r<R, define a random integer XrX_r by

ℙ(Xr=ℓ)=fℓrℓF(r).\mathbb P(X_r=\ell)=\frac{f_\ell r^\ell}{F(r)}.

Analyticity on a disk larger than rr gives finite exponential moments. Direct differentiation shows

𝔼Xr=M(r),VarXr=rM′(r).\mathbb E X_r=M(r),\qquad \mathop{\mathrm{Var}}X_r=rM'(r).

The variance is strictly positive because the support contains consecutive integers. Thus MM is strictly increasing. Its limit at zero is ss. Its limit as r↑Rr\uparrow R is infinity, because a simple positive pole gives

M(r)=rR−r+O(1).M(r)=\frac{r}{R-r}+O(1).

This proves existence and uniqueness in (5), and positivity of σ(r)2\sigma(r)^2.

Fix that rr. If A(q)=∑aℓqℓA(q)=\sum a_\ell q^\ell, let YY be an independent nonnegative integer with probabilities aℓrℓ/A(r)a_\ell r^\ell/A(r). Multiplying the probability generating functions gives the exact identity

[qn]A(q)F(q)m=A(r)F(r)mr−nℙ(Y+X1+⋯+Xm=n),[q^n]A(q)F(q)^m =A(r)F(r)^m r^{-n} \mathbb P\left(Y+X_1+\cdots+X_m=n\right), \tag{8}

where the XiX_i are independent copies of XrX_r.

Here is the local estimate needed for (8). Write μ=1/x\mu=1/x, σ2=σ(r)2\sigma^2=\sigma(r)^2, and let φ(θ)=𝔼eiθXr\varphi(\theta)=\mathbb E e^{i\theta X_r}, ψ(θ)=𝔼eiθY\psi(\theta)=\mathbb E e^{i\theta Y}. Fourier inversion for integer-valued random variables yields

ℙ(Y+X1+⋯+Xm=n)=12π∫−ππe−inθφ(θ)mψ(θ)dθ.\mathbb P(Y+X_1+\cdots+X_m=n) =\frac1{2\pi}\int_{-\pi}^{\pi} e^{-in\theta}\varphi(\theta)^m\psi(\theta)\,d\theta. \tag{9}

The consecutive support implies |φ(θ)|<1|\varphi(\theta)|<1 for 0<|θ|≤π0<|\theta|\le\pi: equality in the triangle inequality would force eiℓθe^{i\ell\theta} to have the same value for all consecutive ℓ\ell, hence eiθ=1e^{i\theta}=1. Continuity then gives a uniform bound strictly below one on every compact region away from zero. At zero, finite exponential moments give

logφ(θ)=iμθ−σ2θ22+O(|θ|3).\log\varphi(\theta)=i\mu\theta-\frac{\sigma^2\theta^2}{2} +O(|\theta|^3).

In particular, |φ(θ)|≤e−cθ2|\varphi(\theta)|\le e^{-c\theta^2} for some c>0c>0 and sufficiently small |θ||\theta|. Also |ψ|≤1|\psi|\le1 and ψ(θ)→1\psi(\theta)\to1 as θ→0\theta\to0.

For m=mnm=m_n,

n−mμ=1+{xn}x=O(1).n-m\mu=\frac{1+\{xn\}}x=O(1). \tag{10}

In the integral (9), put θ=t/m\theta=t/\sqrt m near zero. The integrand becomes

e−i(n−mμ)t/me−σ2t2/2+O(|t|3/m)ψ(t/m),e^{-i(n-m\mu)t/\sqrt m} e^{-\sigma^2t^2/2+O(|t|^3/\sqrt m)}\psi(t/\sqrt m),

which tends to e−σ2t2/2e^{-\sigma^2t^2/2} for each fixed tt. The bound e−ct2e^{-ct^2} permits dominated convergence on the expanding neighborhood; the remaining part of (9) is exponentially small by the uniform bound away from zero. Thus

mℙ(Y+X1+⋯+Xm=n)→12π∫ℝe−σ2t2/2dt=12πσ2>0.\sqrt m\,\mathbb P(Y+X_1+\cdots+X_m=n) \longrightarrow\frac1{2\pi}\int_{\mathbb R}e^{-\sigma^2t^2/2}\,dt =\frac1{\sqrt{2\pi\sigma^2}}>0.

This proves (6). Taking logarithms, using mn/n→xm_n/n\to x, proves (7). ∎

4. The two rate functions

Apply the lemma to (1) and (3), with s=k+1s=k+1. Both pairs satisfy all its assumptions. We obtain differentiable functions ΦD,ΦO\Phi_D,\Phi_O on 0<x<1/(k+1)0<x<1/(k+1) such that

limn→∞1nlogFD⌊xn⌋,k(n)=ΦD(x),limn→∞1nlogFO⌊xn⌋,k(n)=ΦO(x).\lim_{n\to\infty}\frac1n\log\mathrm{FD}_{\lfloor xn\rfloor,k}(n) =\Phi_D(x), \qquad \lim_{n\to\infty}\frac1n\log\mathrm{FO}_{\lfloor xn\rfloor,k}(n) =\Phi_O(x). \tag{11}

For either function, differentiation of (7) and use of (5) cancel the terms containing r′(x)r'(x):

Φ′(x)=logF(r(x)).\Phi'(x)=\log F(r(x)). \tag{12}

Since M′(r)=σ(r)2/rM'(r)=\sigma(r)^2/r, differentiating M(r(x))=1/xM(r(x))=1/x gives

r′(x)=−r(x)x2σ(r(x))2<0,Φ″(x)=−1x3σ(r(x))2<0.r'(x)=-\frac{r(x)}{x^2\sigma(r(x))^2}<0, \qquad \Phi''(x)=-\frac1{x^3\sigma(r(x))^2}<0. \tag{13}

Consequently each rate is strictly concave. Direct substitution gives

FD(1/2)=FO(1/2)=1,MD(1/2)=2k+1,MO(1/2)=4k.F_D(1/2)=F_O(1/2)=1, \qquad M_D(1/2)=2^{k+1},\quad M_O(1/2)=4k. \tag{14}

For clarity, the first logarithmic derivative is

MD(q)=k+1−qH′(q)H(q),M_D(q)=k+1-\frac{qH'(q)}{H(q)},

and H(1/2)=2−k−1H(1/2)=2^{-k-1}, H′(1/2)=−2+(k+1)2−kH'(1/2)=-2+(k+1)2^{-k} give the first mean in (14). On the other side, W(1/2)=1W(1/2)=1, W′(1/2)=4W'(1/2)=4, T(1/2)=K(1/2)=1/2T(1/2)=K(1/2)=1/2, and

T′(1/2)T(1/2)=4k+2,K′(1/2)K(1/2)=2−4k\frac{T'(1/2)}{T(1/2)}=4k+2, \qquad \frac{K'(1/2)}{K(1/2)}=2-4k

give MO(1/2)=4kM_O(1/2)=4k.

By (12)–(14), ΦD\Phi_D has its unique maximum at aka_k, while ΦO\Phi_O has its unique maximum at bkb_k. The two maximum values coincide:

ΦD(ak)=ΦO(bk)=log2.\Phi_D(a_k)=\Phi_O(b_k)=\log2. \tag{15}

Both means belong to the lemma's interval. Also ak<bka_k<b_k for k≥3k\ge3: 4k<2k+14k<2^{k+1} holds at k=3k=3 and persists by induction. Finally, bk<1/(k+1)b_k<1/(k+1) because k+1<4kk+1<4k.

5. Existence and uniqueness of the crossover

On [ak,bk][a_k,b_k], set Δk(x)=ΦD(x)−ΦO(x)\Delta_k(x)=\Phi_D(x)-\Phi_O(x). Strict maximality and (15) imply

Δk(ak)>0,Δk(bk)<0.\Delta_k(a_k)>0,\qquad \Delta_k(b_k)<0. \tag{16}

For ak<x<bka_k<x<b_k, equation (13) makes the DD saddle smaller than 1/21/2 and the OO saddle larger than 1/21/2. Each FF is strictly increasing, so (12) yields

ΦD′(x)<0<ΦO′(x),Δk′(x)<0.\Phi_D'(x)<0<\Phi_O'(x), \qquad \Delta_k'(x)<0. \tag{17}

Continuity, (16), and (17) give exactly one zero ck∈(ak,bk)c_k\in(a_k,b_k). This provides an implicit exact definition of the transition density, without a numerical approximation.

For fixed x≠ckx\ne c_k in that interval, (11) gives

limn→∞1nlogFD⌊xn⌋,k(n)FO⌊xn⌋,k(n)=Δk(x).\lim_{n\to\infty}\frac1n \log\frac{\mathrm{FD}_{\lfloor xn\rfloor,k}(n)} {\mathrm{FO}_{\lfloor xn\rfloor,k}(n)} =\Delta_k(x). \tag{18}

Both counts are eventually positive by the coefficient lemma. If x<ckx<c_k, the limit is positive, so the ratio tends to infinity. If x>ckx>c_k, the limit is negative, so the ratio tends to zero. This proves the two eventual strict inequalities for every allowed k,xk,x. Any other proposed transition density would give the opposite inequality at a point strictly between the two densities, so the density in the conjecture is unique. ∎

Scope and verification

The proof covers every integer k≥3k\ge3 and every fixed real density in the exact interval specified by the conjecture, apart from its expressly excluded crossover. The convergence threshold may depend on k,xk,x, as permitted by the source. It does not claim a single finite-nn sign change over all jj or a uniform threshold in xx. No numerical root or finite screen is a premise of the proof.

The positive generating-function rewrites, support, pole, mean, and floor conditions are all explicit. The coefficient lemma is proved directly rather than inferred from a large-deviation bound on cumulative probabilities. Source status and the priority audit are recorded separately.