Kovalev's conjecture

What Lean checks

Complete proposed analytic disproof of the lower Hardy-duality comparison by a uniform positive diagonal matrix family in every order at least 500001. Independent mathematical and root reviews passed, with two exact arithmetic replays and a separate priority audit. The upper inequality is not resolved. No Lean formalization is claimed.

Statement and proposed disproof

Conjecture 6.1 of Leonid V. Kovalev, Hardy-type norms of matrices, arXiv:2607.17373v1, printed page 9, proposes

cn∥A∥H4≤∥A∥H*1≤∥A∥H4(n>2),c_n\|A\|_{H^4}\leq\|A\|_{H^1_*}\leq\|A\|_{H^4} \qquad(n>2), \tag{1}

for every real matrix of order nn. The definitions are

∥A∥Hp=(𝔼|AU|p)1/p,∥A∥H*1=sup∥B∥H1≤11ntr(ATB),\|A\|_{H^p}=\left(\mathbb E|AU|^p\right)^{1/p}, \qquad \|A\|_{H^1_*} =\sup_{\|B\|_{H^1}\leq1}\frac1n\mathop{\mathrm{tr}}(A^TB),

where UU is uniformly distributed on the Euclidean unit sphere. The conjectured constant, determined by a rank-two projection in the source, is

cn=25/4πΓ((n+1)/2)Γ(n/2)(n+2)1/4n3/4.c_n=\frac{2^{5/4}}{\sqrt\pi} \frac{\Gamma((n+1)/2)}{\Gamma(n/2)} \frac{(n+2)^{1/4}}{n^{3/4}}. \tag{2}

We give positive diagonal counterexamples to the first inequality in (1) in every order at least 500001. This refutes the full conjecture. The second inequality in (1) is not decided here. The dimension threshold is a convenient uniform estimate, not a claim about the first possible counterexample.

All comparisons in the proof are analytic or rational. The supplementary standard-library script verify.py repeats the finite rational arithmetic; it does not numerically integrate, search for matrices, or test a finite list of dimensions.

1. A Gaussian description of a dual unit vector

Let Z=(Z0,…,Zn−1)Z=(Z_0,\ldots,Z_{n-1}) be a standard real Gaussian vector, and set

mn=𝔼|Z|=2Γ((n+1)/2)Γ(n/2),N(B)=𝔼|BZ|.m_n=\mathbb E|Z| =\sqrt2\frac{\Gamma((n+1)/2)}{\Gamma(n/2)}, \qquad N(B)=\mathbb E|BZ|.

Gaussian radius and direction are independent, so

N(B)=mn∥B∥H1.N(B)=m_n\|B\|_{H^1}. \tag{3}

For any positive diagonal matrix B=diag(b0,…,bn−1)B=\mathop{\mathrm{diag}}(b_0,\ldots,b_{n-1}), define

gi=𝔼biZi2|BZ|,D=diag(g0,…,gn−1).g_i=\mathbb E\frac{b_iZ_i^2}{|BZ|}, \qquad D=\mathop{\mathrm{diag}}(g_0,\ldots,g_{n-1}).

For every real matrix CC, pointwise Cauchy–Schwarz gives

𝔼|CZ|≥𝔼(BZ)T(CZ)|BZ|=∑igiCii=tr(DTC),tr(DTB)=N(B).\mathbb E|CZ| \geq\mathbb E\frac{(BZ)^T(CZ)}{|BZ|} =\sum_i g_iC_{ii} =\mathop{\mathrm{tr}}(D^TC), \qquad \mathop{\mathrm{tr}}(D^TB)=N(B).

In the middle equality, sign symmetry makes every off-diagonal expectation zero. All expectations are finite, because the absolute value of the inner product is bounded by |CZ||CZ|. Equality holds when C=BC=B. This proves the needed supporting-functional identity for arbitrary matrices, without restricting the dual supremum to diagonal matrices.

Using (3), with equality attained at B/∥B∥H1B/\|B\|_{H^1}, yields

∥D∥H*1=mnn.\|D\|_{H^1_*}=\frac{m_n}{n}.

Consequently

A=nmnDsatisfies∥A∥H*1=1.A=\frac{n}{m_n}D \quad\text{satisfies}\quad \|A\|_{H^1_*}=1. \tag{4}

For a diagonal vector gg, the fourth spherical moment identity is

∥D∥H44=V(g)n(n+2),V(g)=(∑igi2)2+2∑igi4.\|D\|_{H^4}^4 =\frac{V(g)}{n(n+2)}, \qquad V(g)=\left(\sum_i g_i^2\right)^2+2\sum_i g_i^4. \tag{5}

For completeness, independent Gaussian moments give 𝔼|DZ|4=(∑igi2)2+2∑igi4\mathbb E|DZ|^4=(\sum_i g_i^2)^2+2\sum_i g_i^4; dividing by 𝔼|Z|4=n(n+2)\mathbb E|Z|^4=n(n+2) proves (5). Formula (2) implies

cn4=8(n+2)mn4π2n3.c_n^4=\frac{8(n+2)m_n^4}{\pi^2n^3}.

Combining this with (4) and (5) gives the exact identity

(cn∥A∥H4)4=8V(g)π2.\left(c_n\|A\|_{H^4}\right)^4=\frac{8V(g)}{\pi^2}. \tag{6}

Thus it suffices to construct BB for which V(g)>π2/8V(g)>\pi^2/8.

2. One large coordinate and many small coordinates

Fix

c=140.c=\frac1{40}.

For every positive integer kk, take n=k+1n=k+1 and

Bk=diag(1,c/k,…,c/k).B_k=\mathop{\mathrm{diag}} \left(1,\sqrt{c/k},\ldots,\sqrt{c/k}\right).

Let

Xk=1k∑i=1kZi2,Rk=Z02+cXk,uk=𝔼Z02Rk,vk=𝔼XkRk.X_k=\frac1k\sum_{i=1}^kZ_i^2, \qquad R_k=\sqrt{Z_0^2+cX_k}, \qquad u_k=\mathbb E\frac{Z_0^2}{R_k}, \qquad v_k=\mathbb E\frac{X_k}{R_k}.

By permutation symmetry of the last kk coordinates, the gradient vector is

g=(uk,c/kvk,…,c/kvk).g=\left(u_k,\sqrt{c/k}\,v_k,\ldots,\sqrt{c/k}\,v_k\right).

Therefore

V(g)=(uk2+cvk2)2+2uk4+2c2kvk4.V(g)=(u_k^2+cv_k^2)^2+2u_k^4+\frac{2c^2}{k}v_k^4. \tag{7}

Introduce two fixed one-dimensional expectations, with GG standard normal:

a=𝔼G2G2+c,b=𝔼1G2+c.a=\mathbb E\frac{G^2}{\sqrt{G^2+c}}, \qquad b=\mathbb E\frac1{\sqrt{G^2+c}}. \tag{8}

Since 𝔼Xk=1\mathbb EX_k=1, convexity of x↦G2/G2+cxx\mapsto G^2/\sqrt{G^2+cx} and conditional Jensen give

uk≥a.u_k\geq a. \tag{9}

For every nonnegative uu and positive xx,

xu+cx≥min(x,1)u+c.\frac{x}{\sqrt{u+cx}} \geq\frac{\min(x,1)}{\sqrt{u+c}}. \tag{10}

If x≤1x\leq1, this follows by decreasing the denominator; if x≥1x\geq1, squaring reduces it to (x−1)((x+1)u+cx)≥0(x-1)((x+1)u+cx)\geq0. Independence of XkX_k and Z0Z_0 gives

vk≥b𝔼min(Xk,1).v_k\geq b\,\mathbb E\min(X_k,1).

Moreover, centering and Cauchy–Schwarz yield

𝔼(1−Xk)+=12𝔼|Xk−1|≤12𝔼(Xk−1)2=12k.\mathbb E(1-X_k)_+ =\frac12\mathbb E|X_k-1| \leq\frac12\sqrt{\mathbb E(X_k-1)^2} =\frac1{\sqrt{2k}}.

Hence, for every k≥500000k\geq500000,

vk≥9991000b.v_k\geq\frac{999}{1000}b. \tag{11}

No limit in the matrix size has been used. Estimates (9) and (11) hold for each integer in this entire range.

3. Rational lower bounds for the two expectations

We prove

a>776710000,b>20831000.a>\frac{7767}{10000}, \qquad b>\frac{2083}{1000}. \tag{12}

Put

z=c4=1160,w=z24=1102400,L=log320−γ,z=\frac c4=\frac1{160}, \qquad w=\frac{z^2}{4}=\frac1{102400}, \qquad L=\log320-\gamma,

where γ\gamma is Euler's constant. The elementary constant estimates proved in §5 are

519100<L<649125,2π<250663100000,π<355113.\frac{519}{100}<L<\frac{649}{125}, \qquad \sqrt{2\pi}<\frac{250663}{100000}, \qquad \pi<\frac{355}{113}. \tag{13}

In the integrals in (8), substitute G=csinhtG=\sqrt c\sinh t on the positive half-axis. The standard integral

Kν(z)=∫0∞e−zcoshtcosh(νt)dt(z>0)K_\nu(z)=\int_0^\infty e^{-z\cosh t}\cosh(\nu t)\,dt \qquad(z>0)

from DLMF 10.32.9, after also replacing 2t2t by the integration variable, gives

b=ezK0(z)2π,a=cez22π(K1(z)−K0(z)).b=\frac{e^zK_0(z)}{\sqrt{2\pi}}, \qquad a=\frac{ce^z}{2\sqrt{2\pi}}\bigl(K_1(z)-K_0(z)\bigr). \tag{14}

The positive-axis expansions DLMF 10.25.2 and 10.31.2 are

I0(z)=∑j=0∞wj(j!)2,K0(z)=LI0(z)+S(z),S(z)=∑j=1∞Hjwj(j!)2,I_0(z)=\sum_{j=0}^\infty\frac{w^j}{(j!)^2}, \qquad K_0(z)=L I_0(z)+S(z), \qquad S(z)=\sum_{j=1}^\infty H_j\frac{w^j}{(j!)^2}, \tag{15}

where Hj=∑i=1j1/iH_j=\sum_{i=1}^j1/i. All terms relevant here are positive. Since L>0L>0, (15) gives K0(z)>LK_0(z)>L. Using ez>1+ze^z>1+z and (13) in (14) yields

b>(161/160)(519/100)250663/100000>20831000.b> \frac{(161/160)(519/100)}{250663/100000} >\frac{2083}{1000}. \tag{16}

To bound the difference in (14), use Hj≤jH_j\leq j and 1/(j!)2≤11/(j!)^2\leq1. Termwise differentiation of the convergent series gives

1≤I0(z)≤11−w,I0′(z)≤2wz(1−w)2,1\leq I_0(z)\leq\frac1{1-w}, \qquad I_0'(z)\leq\frac{2w}{z(1-w)^2},

0<S(z)≤w(1−w)2,S′(z)≤2w(1+w)z(1−w)3.0<S(z)\leq\frac{w}{(1-w)^2}, \qquad S'(z)\leq\frac{2w(1+w)}{z(1-w)^3}. \tag{17}

Before evaluating at z=1/160z=1/160, regard the coefficient in (15) as the function L(z)=log(2/z)−γL(z)=\log(2/z)-\gamma. Thus L′(z)=−1/zL'(z)=-1/z. Using K0′=−K1K_0'=-K_1 (DLMF 10.29.3), differentiating (15) gives

K1(z)−K0(z)=I0(z)z−L(I0(z)+I0′(z))−S(z)−S′(z).K_1(z)-K_0(z) =\frac{I_0(z)}z-L\bigl(I_0(z)+I_0'(z)\bigr)-S(z)-S'(z).

Consequently this difference is greater than the positive rational number

T=1z−649125(11−w+2wz(1−w)2)−w(1−w)2−2w(1+w)z(1−w)3.T=\frac1z -\frac{649}{125}\left(\frac1{1-w}+\frac{2w}{z(1-w)^2}\right) -\frac{w}{(1-w)^2}-\frac{2w(1+w)}{z(1-w)^3}. \tag{18}

Substitution of the rational values of z,wz,w gives

(1/40)(161/160)T2(250663/100000)>776710000.\frac{(1/40)(161/160)T}{2(250663/100000)} >\frac{7767}{10000}. \tag{19}

Equations (14), (18), and (19) prove the other part of (12). Both rational comparisons (16) and (19) are explicitly reproduced in verify.py.

4. Completion of the counterfamily

Set

a0=776710000,b0=20831000,d=9991000.a_0=\frac{7767}{10000},\qquad b_0=\frac{2083}{1000},\qquad d=\frac{999}{1000}.

For every k≥500000k\geq500000, equations (7), (9), (11), and (12) imply

V(g)>(a02+cd2b02)2+2a04=19745767844560848456032703211600000000000000000000000000.V(g)>(a_0^2+c d^2b_0^2)^2+2a_0^4 =\frac{1974576784456084845603270321} {1600000000000000000000000000}. \tag{20}

The difference between the last rational number and (355/113)2/8(355/113)^2/8 is exactly

837096071974739350815872884920430400000000000000000000000000>0.\frac{8370960719747393508158728849} {20430400000000000000000000000000}>0. \tag{21}

It follows from (13) that V(g)>π2/8V(g)>\pi^2/8. The positive diagonal matrix

Ak=k+1mk+1diag(uk,140kvk,…,140kvk)A_k=\frac{k+1}{m_{k+1}} \mathop{\mathrm{diag}} \left(u_k,\sqrt{\frac1{40k}}v_k,\ldots, \sqrt{\frac1{40k}}v_k\right)

therefore satisfies

∥Ak∥H*1=1<ck+1∥Ak∥H4for every integer k≥500000,\|A_k\|_{H^1_*}=1<c_{k+1}\|A_k\|_{H^4} \qquad\text{for every integer }k\geq500000,

by (4) and (6). These are full-rank real matrices. This proves the claimed disproof of (1).

5. Elementary certificates for the constants

Here are explicit finite bounds proving (13), so no decimal approximation of a special function is a premise.

For 0<q<10<q<1, define

PM(q)=2∑j=0M−1q2j+12j+1,EM(q)=2q2M+1(2M+1)(1−q2).P_M(q)=2\sum_{j=0}^{M-1}\frac{q^{2j+1}}{2j+1}, \qquad E_M(q)=\frac{2q^{2M+1}}{(2M+1)(1-q^2)}.

The positive logarithmic series gives

PM(q)<log1+q1−q<PM(q)+EM(q).P_M(q)<\log\frac{1+q}{1-q}<P_M(q)+E_M(q). \tag{22}

Take ten terms at q=1/3,1/9,1/31q=1/3,1/9,1/31. The identities

log320=8log2+log(5/4),log30=5log2−log(16/15)\log320=8\log2+\log(5/4), \qquad \log30=5\log2-\log(16/15)

then produce rational enclosures. Direct substitution in (22) shows

721125<log320<57691000.\frac{721}{125}<\log320<\frac{5769}{1000}. \tag{23}

To enclose Euler's constant without a tabulated value, write

HN−logN−γ=∑j=N∞∫011−t(j+t)(j+1)dt.H_N-\log N-\gamma =\sum_{j=N}^\infty \int_0^1\frac{1-t}{(j+t)(j+1)}\,dt.

Each summand is between 1/[2(j+1)2]1/[2(j+1)^2] and 1/[2j(j+1)]1/[2j(j+1)]. A telescoping upper sum and an integral lower estimate give

12(N+1)<HN−logN−γ<12N.\frac1{2(N+1)}<H_N-\log N-\gamma<\frac1{2N}. \tag{24}

At N=30N=30, use the rational enclosure of log30\log30 from (22) and the exact finite sum H30H_{30}. The resulting rational inequalities are

5771000<γ<289500.\frac{577}{1000}<\gamma<\frac{289}{500}. \tag{25}

Subtracting (25) from (23) proves the two stated bounds for LL.

Finally, Machin's identity

π=16arctan(1/5)−4arctan(1/239)\pi=16\arctan(1/5)-4\arctan(1/239)

follows from the tangent addition formula, with the resulting angle lying in the first quadrant. Enclose each arctangent by ten alternating series terms and its first omitted term. The resulting rational upper bound is less than 355/113355/113. Squaring positive numbers then gives

2π<710113<(250663100000)2,2\pi<\frac{710}{113} <\left(\frac{250663}{100000}\right)^2,

which proves the remaining part of (13). The exact checker reproduces (23), (25), this comparison, (16), (19), and the final positive fraction (21).