# General Chowla attempt: spectral factorization and Schur transforms ## Target and present status For a nonempty finite set `A` of distinct positive integers, put \[ P(z)=\sum_{a\in A}z^a,\quad f_A(x)=\Re P(e^{ix}),\quad K=-\min_x f_A(x),\quad N=|A|. \] The general target is a constant `C` independent of all frequencies and cardinalities such that `N<=C K^2`, or an admissible unbounded family with `K/sqrt(N)->0`. No such general proof or counterexample has been obtained in this pass. The statements below are exact reformulations, proved conditional inequalities, and explicit counterexamples to tempting weakened lemmas. They are not presented as solutions of Chowla's conjecture or as claims of mathematical novelty. ## 1. Spectral factorization, including signs and endpoint constraints Fejer--Riesz gives \[ h(x)=K+f_A(x)=|Q(e^{ix})|^2, \qquad Q(z)=\sum_{j=0}^{D}q_jz^j,\quad D=\max A. \] Write `r_d=sum_j q_(j+d) conjugate(q_j)`, with coefficients outside `[0,D]` set to zero. Then \[ r_0=\sum_j|q_j|^2=K,\qquad r_d=\begin{cases}1/2&d\in A,\\0&d>0,\ d\notin A.\end{cases} \] In particular `q_D conjugate(q_0)=1/2`, so both endpoints are nonzero and \[ K\ge|q_0|^2+|q_D|^2\ge2|q_0q_D|=1. \] The factor may be chosen **real**, but not necessarily nonnegative. Choose the unique outer factor with `Q(0)>0`. Conjugating its coefficients preserves its outer property and factors `h(-x)=h(x)`; uniqueness therefore forces real coefficients. Existence and uniqueness are covered by the classical scalar case of Dritschel--Rovnyak, *The Operator Fejer--Riesz Theorem*: . The real-coefficient conclusion is the argument just given. It is invalid to additionally suppose `q_j>=0`: already `A={1,3}`, `K=2` is a counterexample to that simplification. Nonnegative coefficients of a degree-3 factor would satisfy `q0*q3=1/2` and `q0*q2+q1*q3=r2=0`. Thus `q1=q2=0`, contradicting `r1=1/2`. ## 2. A proved bound for arbitrary complex factors with bounded multiplicity Let `B={j:q_j != 0}` and \[ R=\max_{d\in A}\#\{j:j,j+d\in B\}. \] For any complex coefficients, Cauchy--Schwarz gives, for each `d in A`, \[ \frac14=|r_d|^2 \le R\sum_j|q_{j+d}|^2|q_j|^2. \] Summing and using that positive differences partition unordered pairs, \[ \frac N4 \le\frac R2\left(K^2-\sum_j|q_j|^4\right), \] so \[ \boxed{N\le2R\left(K^2-\sum_j|q_j|^4\right)\le2R K^2.} \] The endpoint condition additionally gives `sum |q_j|^4>=1/2`, hence `N<=2R(K^2-1/2)`. Thus a bounded-multiplicity spectral factor establishes the desired square-root bound for that class, regardless of phases or signs. A general counterexample would force `R>=N/(2K^2)->infinity` for **every** factor. This is not a general proof: ordinary outer factors can have dense support and large `R`, and no frequency-independent bound on `R` was obtained. ## 3. Ultraflat factors point in the wrong direction Parseval gives the exact identity \[ \int|Q|^4=\int h^2=K^2+\frac N2, \qquad N=2\int(|Q|^2-K)^2. \] All circle integrals here use probability Haar measure. If `|| |Q|^2-K ||_infinity <= epsilon K`, then \[ N\le2\epsilon^2K^2. \] More generally, any uniform bound on `||Q||_4/||Q||_2` proves the desired bound in that class. A counterexample with `K=o(sqrt(N))` would instead require that norm ratio to diverge. Constructing ultraflat factors cannot provide such a counterexample. Removing signs from a Littlewood polynomial or replacing nonuniform autocorrelation coefficients by ones does not preserve the identity. ## 4. Positive idempotent reformulation Since the nonzero nonconstant Fourier coefficients of `h` are exactly `1/2`, normalized circle convolution satisfies \[ h*h=\tfrac12h+(K^2-K/2). \] For the probability density `p=h/K` and `alpha=1/(2K)`, this becomes \[ p*p=\alpha p+(1-\alpha),\qquad \int p^2=1+2N\alpha^2. \] Consequently a universal bound on `int p^2` in this very particular class would prove Chowla. This identity alone does not supply that bound. An arbitrary nonnegative density, positive definite sequence, or finite cyclic idempotent does not satisfy the required finite, unit-coefficient structure on the integer group. ## 5. Finite, zero-free 0/1 polynomials do not suffice Let \[ P_N(z)=z+z^2+\cdots+z^N,\qquad W_N(z)=t+P_N(z),\quad t>1. \] Every `W_N` is zero-free even on the **closed** unit disk. Indeed, \[ (1-z)W_N(z)=t-(t-1)z-z^{N+1}. \] At a root with `|z|<=1`, equality in `t=|(t-1)z+z^(N+1)|<=t` forces `z=1`, where `W_N(1)=t+N>0`. Thus a fixed positive constant term, finite termination, root-freeness, and the 0/1 pattern do not bound the number of coefficients. Yet the negative excursion of `Re P_N` is linear. With `q=2N+1`, \[ f_N(3\pi/q)=-\frac1{2\sin(3\pi/(2q))}-\frac12 \le-\frac{q}{3\pi}-\frac12. \] The missing property is the full Schur condition `|P/(2K+P)|<=1`, which is equivalent to `Re P>=-K`. ## 6. Even the full H2 bound on the Schur transform is insufficient For the same finite `P_N`, put \[ \phi_N(z)=\frac{P_N(z)}{t+P_N(z)}=1-\frac t{W_N(z)}. \] Although `phi_N` is generally not a Schur function, its exact Hardy norm is \[ \boxed{\|\phi_N\|_{H^2}^2=\frac{N}{(t-1)(t+N)}.} \] In particular at `t=2`, corresponding formally to a proposed `K=1`, this norm is `N/(N+2)<1` for every `N`. Therefore **finite termination + stable denominator + the complete Parseval/H2 inequality** does not recover the Chowla bound. An argument that replaces Schur `H-infinity` contractivity by coefficient square summability has lost essential information. ### Exact derivation of the formula Define the real symmetric covariance sequence \[ \gamma_j=\int\frac{e^{ijx}}{|W_N(e^{ix})|^2}\,dx, \qquad\gamma_{-j}=\gamma_j. \] Root-freeness makes `1/W_N` analytic in a neighborhood of the closed disk. Fourier coefficients of `W_N/|W_N|^2=1/overline(W_N)` give \[ t\gamma_0+\sum_{k=1}^N\gamma_k=1/t, \] and, for `1<=j<=N`, \[ t\gamma_j+\gamma_0+ \sum_{r=1}^{j-1}\gamma_r+ \sum_{r=1}^{N-j}\gamma_r=0.\tag{*} \] The equations for `j` and `N+1-j` have the same two sums, so `gamma_j=gamma_(N+1-j)`. Subtracting adjacent equations then gives `(t-1)(gamma_(j+1)-gamma_j)=0`. Thus all `gamma_1,...,gamma_N` equal a common `g`. Solving the two remaining equations yields \[ g=-\frac{\gamma_0}{t+N-1},\qquad \gamma_0=\frac{t+N-1}{t(t-1)(t+N)}. \] Finally `int(1/W_N)=1/t` and therefore \[ \|1-t/W_N\|_2^2=t^2\gamma_0-1 =\frac{N}{(t-1)(t+N)}. \] ## 7. A valid primitive-frequency consequence of the Schur transform For an admissible `P`, `phi=P/(2K+P)` is Schur. Its formal Taylor series is \[ \phi=\sum_{m\ge1}(-1)^{m-1}\frac{P^m}{(2K)^m}. \] If `a in A` cannot be represented as a sum of two or more elements of `A`, the coefficient of `z^a` is exactly `1/(2K)`. If there are `s` such primitive frequencies, Parseval gives `s/(4K^2)<=1`, hence `s<=4K^2`. The gap is explicit: most elements of a general finite `A` can be decomposable, for example only `1` is primitive in `{1,...,N}`. Section 6 shows why replacing the missing information by an H2 argument alone cannot finish the general proof.