Integrality criterion for inverses of reciprocal binomial Hankel matrices

From papers

For a positive integer rr, let bk(r)=(k+r1r)b_k(r)=\binom{k+r-1}{r}, and let Rn(bk(r))R_n(b_k(r)) denote the corresponding reciprocal Hankel matrix. Integrality criterion conjecture. The inverse of Rn(bk(r))R_n(b_k(r)) has integer entries if and only if

n0(modq) or n1(modq)n\equiv 0 \pmod q \text{ or } n\equiv 1 \pmod q

for every prime power qq dividing rr. The source reports that the conjecture holds computationally for n20n\leq 20 and r10r\leq 10. Its general validity is left open.

Progress summary

Open

The conjecture remains open: computations support its rule, but no general proof or counterexample was found.

The conjecture concerns exactly when the inverse of a reciprocal binomial Hankel matrix has only integer entries, in terms of congruence conditions on nn and the prime powers dividing rr. The relevant 1999 source records the conjecture and leaves its general validity unresolved.

Known results

  • The 1999 paper gives explicit formulas for the inverse and reports computational verification for n20n\leq 20 and r10r\leq 10; it also records verification for n16n\leq 16 and r10r\leq 10.

Current status (as of August 2026): The criterion is computationally verified in the reported ranges, but no general proof, counterexample, or independently verified recent progress is recorded.

Sources
Sources & referencesView supporting material

Primary source

Thomas M. Richardson, “The Filbert Matrix”, arXiv:math/9905079 (1999).

Solutions 1

Proof

Proof

Use zero-based indices 0a,b<n0\leq a,b<n, and write H=Rn(bk(r))H=R_n(b_k(r)). Then

Hab=1(a+b+rr)=r01xa+b(1x)r1dx.(1)H_{ab} =\frac{1}{\binom{a+b+r}{r}} =r\int_0^1 x^{a+b}(1-x)^{r-1}\,dx. \qquad (1)

Thus HH is the moment matrix of the weight r(1x)r1r(1-x)^{r-1} on [0,1][0,1].

For k0k\geq0, define the shifted Jacobi polynomial

Qk(x)=Pk(r1,0)(2x1)=a=0k(1)k+a(ka)(k+r+a1a)xa.(2)Q_k(x)=P_k^{(r-1,0)}(2x-1) =\sum_{a=0}^k(-1)^{k+a} \binom{k}{a}\binom{k+r+a-1}{a}x^a. \qquad (2)

Every coefficient of QkQ_k is an integer. Its constant coefficient is (1)k(-1)^k, and its leading coefficient is

λk=(2k+r1k).(3)\lambda_k=\binom{2k+r-1}{k}. \qquad (3)

The corresponding Rodrigues formula is

Qk(x)=(1)kk!(1x)1rdkdxk(xk(1x)k+r1).Q_k(x)=\frac{(-1)^k}{k!}(1-x)^{1-r} \frac{d^k}{dx^k}\left(x^k(1-x)^{k+r-1}\right).

Integration by parts shows that QkQ_k is orthogonal to every polynomial of degree less than kk. All boundary terms vanish because of the factors xkx^k and (1x)k+r1(1-x)^{k+r-1}. Using Qk(k)=k!λkQ_k^{(k)}=k!\lambda_k also gives

r01(1x)r1Qk(x)2dx=r2k+r.(4)r\int_0^1(1-x)^{r-1}Q_k(x)^2\,dx =\frac{r}{2k+r}. \qquad (4)

Let Aka=[xa]Qk(x)A_{ka}=[x^a]Q_k(x). Equations (1) and (4) imply

AHAT=diag(rr,rr+2,,rr+2n2).AHA^{\mathsf T} =\operatorname{diag}\left( \frac r r,\frac r{r+2},\ldots,\frac r{r+2n-2} \right).

Since AA is lower triangular with nonzero diagonal, it is invertible. Therefore

(H1)ab=1rk=0n1(2k+r)AkaAkb.(5)(H^{-1})_{ab} =\frac1r\sum_{k=0}^{n-1}(2k+r)A_{ka}A_{kb}. \qquad (5)

Define the integer polynomial

Kn(x,y)=k=0n1(2k+r)Qk(x)Qk(y).K_n(x,y)=\sum_{k=0}^{n-1}(2k+r)Q_k(x)Q_k(y).

Equation (5) says that H1H^{-1} is the coefficient matrix of Kn/rK_n/r. Hence H1H^{-1} is integral exactly when every coefficient of KnK_n is divisible by rr.

The Christoffel-Darboux identity, using the leading coefficients (3) and norms (4), gives

Kn(x,y)=n(n+r1)2n+r1Qn(x)Qn1(y)Qn1(x)Qn(y)xy.(6)K_n(x,y) =\frac{n(n+r-1)}{2n+r-1} \frac{ Q_n(x)Q_{n-1}(y)-Q_{n-1}(x)Q_n(y) }{x-y}. \qquad (6)

Indeed, the required scalar follows from

λn1λn2n+r2r=n(n+r1)r(2n+r1).\frac{\lambda_{n-1}}{\lambda_n}\frac{2n+r-2}{r} =\frac{n(n+r-1)}{r(2n+r-1)}.

Put

Ln(x,y)=Qn(x)Qn1(y)Qn1(x)Qn(y)xy.L_n(x,y)= \frac{ Q_n(x)Q_{n-1}(y)-Q_{n-1}(x)Q_n(y) }{x-y}.

The numerator is an integer polynomial that vanishes when x=yx=y. Since xyx-y is monic, LnL_n belongs to Z[x,y]\mathbb Z[x,y]. Thus (6) is equivalently

(2n+r1)Kn=n(n+r1)Ln.(7)(2n+r-1)K_n=n(n+r-1)L_n. \qquad (7)

Suppose first that

rn(n1).r\mid n(n-1).

Then rn(n+r1)r\mid n(n+r-1), since

n(n+r1)=n(n1)+nr.n(n+r-1)=n(n-1)+nr.

Moreover,

gcd(r,2n+r1)=1.(8)\gcd(r,2n+r-1)=1. \qquad (8)

To see this, any common divisor divides both 2n12n-1 and n(n1)n(n-1). But

(2n1)2=4n(n1)+1,(2n-1)^2=4n(n-1)+1,

so that common divisor must divide 11.

Taking coefficients in (7), the right-hand side is divisible by rr. Equation (8) allows cancellation of 2n+r12n+r-1, proving that every coefficient of KnK_n is divisible by rr. Hence every entry of H1H^{-1} is an integer.

Conversely, the constant coefficient of each Qk(x)Qk(y)Q_k(x)Q_k(y) is 11. Therefore

[x0y0]Kn=k=0n1(2k+r)=n(n+r1).(9)[x^0y^0]K_n =\sum_{k=0}^{n-1}(2k+r) =n(n+r-1). \qquad (9)

If H1H^{-1} is integral, equations (5) and (9) imply

rn(n+r1).r\mid n(n+r-1).

Subtracting the term nrnr gives

rn(n1).r\mid n(n-1).

We have therefore proved

H1 is integralrn(n1).(10)H^{-1}\text{ is integral} \quad\Longleftrightarrow\quad r\mid n(n-1). \qquad (10)

Finally, write r=ppepr=\prod_p p^{e_p}. Since gcd(n,n1)=1\gcd(n,n-1)=1,

pepn(n1)p^{e_p}\mid n(n-1)

holds exactly when

n0(modpep)orn1(modpep).n\equiv0\pmod{p^{e_p}} \quad\text{or}\quad n\equiv1\pmod{p^{e_p}}.

This is equivalent to the stated condition for every prime power dividing rr. Thus (10) proves Conjecture 5.1.

Original source: https://arxiv.org/abs/math/9905079

0 endorsements
Samuel Schlesinger ·