# Invertible Eichler factor: all roots on the circle

10 November 2026, America/Los_Angeles.

[project, proved]  An elementary argument, not new deep mathematics, and not an RH claim.

## Statement

Let p be a prime and r a positive integer.  Write

    Q_r^{inv}(u) = 1 + p^r u^r + (p-1) sum_{j=1}^{r-1} p^{j-1} u^j

(for r=1 this is 1 + p u).  Every root of Q_r^{inv} lies on the circle |u|=1/p, and all roots are simple.  Equivalently, after the substitution u = w/p, every root of the palindromic polynomial

    P_r(w) = 1 + w^r + c sum_{j=1}^{r-1} w^j,    c = 1 - 1/p in (0,1),

lies on the unit circle |w|=1, and all roots are simple.  In particular every root of Q_r^{inv} has Re(s)=1/2.

The formula for Q_r^{inv} itself is proved separately, by an orbit/volume count, in `order-eichler-qinv-formula.md`.  Together these two notes prove part (1) of the Block 9 project conjecture.  The argument here treats the explicit polynomial; it does not need the source of the coefficients.

## Proof

The case r=1 is P_1(w)=1+w, with the single simple root w=-1.  Assume r>=2.

The polynomial is palindromic: w^r P_r(1/w)=P_r(w).  On the unit circle w=e^{i theta}, the function

    g(theta) = e^{-i r theta/2} P_r(e^{i theta})

is real for theta not a multiple of 2 pi, and equals

    g(theta) = 2 cos(r theta/2) + c sin((r-1) theta/2) / sin(theta/2).

(The form 2(1-c) cos(r theta/2) + c sin((r+1) theta/2)/sin(theta/2) is the same identity, by the sum-to-product formula for sines.)  Write x = cos(theta/2).  As theta runs through (0, 2 pi), the half-angle theta/2 runs through (0, pi), so x runs once through (-1, 1), from 1 down to -1.  In that coordinate,

    S(x) := 2 T_r(x) + c U_{r-2}(x)

equals g(theta), where T_r is the Chebyshev polynomial of the first kind (T_r(cos phi)=cos(r phi)) and U_{r-2} is the Chebyshev polynomial of the second kind (U_{n}(cos phi)=sin((n+1) phi)/sin phi, so U_{r-2}(cos(theta/2))=sin((r-1)theta/2)/sin(theta/2)).  The combination S is a real polynomial of degree r (T_r has degree r, and U_{r-2} has degree r-2).  If S has r distinct roots in (-1,1), they give r distinct theta in (0, 2 pi), hence r distinct unit-modulus roots of P_r, and there are no others.

Evaluate S at the endpoints and at the extrema of T_r.  The extrema occur at x_k = cos(pi k / r) for k=1,...,r-1, where T_r(x_k)=(-1)^k.  The second-kind value is

    U_{r-2}(x_k) = sin((r-1) pi k / r) / sin(pi k / r) = (-1)^{k+1},

because sin(pi k - pi k / r) = (-1)^{k+1} sin(pi k / r).  Therefore

    S(x_k) = 2(-1)^k + c(-1)^{k+1} = (-1)^k (2-c).

Since c is in (0,1), the factor 2-c is positive, and these values are nonzero with sign (-1)^k.  At the endpoints,

    S(1) = 2 T_r(1) + c U_{r-2}(1) = 2 + c(r-1) > 0,
    S(-1) = 2(-1)^r + c (-1)^{r-2}(r-1) = (-1)^r (2 + c(r-1)),

also nonzero.  Order the test points from left to right:

    -1 < x_{r-1} < x_{r-2} < ... < x_1 < 1.

The signs of S at these r+1 points are

    (-1)^r,  (-1)^{r-1},  (-1)^{r-2},  ...,  -1,  +1.

Adjacent signs are opposite, so the intermediate-value theorem supplies a root of S in each of the r open intervals

    (-1, x_{r-1}),  (x_{r-1}, x_{r-2}),  ...,  (x_2, x_1),  (x_1, 1).

These r roots are simple (the values at the test points are nonzero, so no root is a test point, and a multiple root would not produce a genuine sign change in every adjacent pair while accounting for only r roots in total: more directly, a degree-r polynomial with r distinct roots is a product of distinct linear factors).  Hence every root of S is real, simple, and in (-1,1), and every root of P_r is simple and of modulus 1.

When r is odd, w=-1 is one of those roots: it corresponds to theta=pi, hence to x=0, and S(0)=0 because both T_r(0) and U_{r-2}(0) vanish.  The root is already counted in whichever interval contains 0, and it is simple: P_r'(-1)= r - c(r-1)/2 > 0, since c<1 implies c(r-1)/2 < r.  In all cases P_r(1)=2+c(r-1) is nonzero, so w=1 is never a root.

This proves the claim for every prime p and every r>=1.

## Literature, and what is not claimed

The conclusion is an elementary fact about a classical one-parameter family of palindromic polynomials.  It is not a new theorem in the geometry of polynomials, and it is not a claim about the Riemann hypothesis.

Lakatos, "On zeros of reciprocal polynomials", Publ. Math. Debrecen 2002, gives the sufficient coefficient test |A_r| >= sum_{k=1}^{r-1} |A_k - A_r|.  For P_r this is 1 >= (r-1)/p, i.e. only the range r <= p+1.  That misses, for instance, the already-computed case p=2, r=4.  Among the 900 pairs with 1<=r<=60 and p a prime at most 50, the test covers 343 and misses 557.

Lakatos-Losonczi, "Self-inversive polynomials whose zeros are on the unit circle", Publ. Math. Debrecen 2004, give |A_r| >= (1/2) sum_{k=1}^{r-1} |A_k|, which for this family is (1-1/p)(r-1) <= 2.  That covers only 48 of the same 900 pairs (all r<=3, together with (r,p)=(4,2), (4,3), and (5,2)), and is often worse than the 2002 test.  Schinzel's generalisation of Lakatos's bound to complex self-inversive polynomials is of the same sufficient-coefficient type.  A later odd-degree tightening, replacing the 2002 right-hand side by a cosine factor, still does not reach large r at p=2.

Cohn's theorem (Cohn, Math. Z. 14, 1922, as formulated by Bonsall-Marden, Proc. Amer. Math. Soc. 1952) is a necessary and sufficient criterion: a self-inversive polynomial has all its zeros on the unit circle if and only if all zeros of its derivative lie in the closed unit disk.  Chen has another necessary and sufficient condition, in terms of an auxiliary polynomial.  Kim-Park (2012) give a necessary and sufficient condition related to canonical systems.  These are not coefficient tests, and they do not by themselves place the roots of this explicit P_r without further work.  Other sufficient coefficient tests (Vieira, Kwon, and related work) were not found to cover the whole range c in (0,1) for every r.

The exact family 1 + a(z+...+z^{n-1}) + z^n is discussed on Mathematics Stack Exchange question 3544959, which states a necessary and sufficient range for the real parameter a: for n even, -2/(n-1) <= a <= 2; for n odd, -2/(n-1) <= a <= 2n/(n-1).  Our c lies strictly inside both intervals.  That argument proceeds by analysing a sine ratio and reading off the range from endpoint limits (and, at the endpoints of the a-interval, from a vanishing derivative).  It is not a journal article, and the interior of the range is presented with a plot-aided count of solutions.  The Chebyshev-node argument above is a complete elementary proof for the subrange c in (0,1) that this project needs; it does not attempt to recover the full necessary and sufficient interval.

## Numerical check

The sign pattern of S at the r+1 test points, and a bisection of each of the r sign-change intervals followed by a Horner evaluation of |P_r| on the corresponding point of the unit circle, was checked for every r from 1 to 60 and every prime p at most 50 (900 pairs).  All 900 sign patterns match the proof, and all 900 bisections produce r roots with |P_r| at most 2.72e-12 (the maximum occurring at r=60, p=43).  As a second method, numpy.roots of P_r for r<=40 (600 pairs) places every root on |w|=1 to at most 5.11e-15 (the maximum occurring at r=39, p=19).  An earlier draft evaluated a different family of test angles with a wrong predicted sign and is discarded.

Record: `order-eichler-unitcircle-check.json`.  Figure: `figures/order-eichler-unitcircle`.
