betaFeedback

WorksMathematics

Sums of three composite triangular numbers

Abstract

A triangular number T_k = k(k+1)/2 is composite for every k >= 3, the only non-composite ones being 0, 1 and 3. We ask which integers are sums of three composite triangular numbers, a restricted form of Gauss's Eureka theorem; equivalently, when 8n+3 is a sum of three squares none of which is 1, 9 or 25. We prove that every integer n > 194 is a sum of three composite triangular numbers, with at most three possible exceptions, all larger than 8.410^9. Such exceptions can occur only if Tatuzawa's form of Siegel's theorem, with eps = 1/20, has an exceptional discriminant -N, and then only for 8n+3 in {N*, 9N*, 25N*}. Up to 8.410^9 the integers that are not such sums are exactly 41 explicit values, the largest being 194. Under the Generalized Riemann Hypothesis these 41 values are the complete list. The proof combines an elementary lemma, a union bound for the representations that use a forbidden square, the class number formula for sums of three squares, Tatuzawa's theorem, and explicit bounds for a divisor-type function. The analytic range begins at 8n+3 = 3.456110^10 and overlaps the computed range by a factor 1.944. The computed range is certified by two independent algorithms, supplied as ancillary files together with a verifier for every constant.

Introduction

Gauss’s Eureka theorem states that every non-negative integer is a sum of three triangular numbers Tk=k(k+1)/2T_k = k(k+1)/2, k≥0k \ge0; it is equivalent to the representation of every N≡3(mod8)N \equiv3 \pmod{8} as a sum of three squares. Since Tk=k2(k+1)T_k = \frac{k}{2}(k+1) or k+12k\frac{k+1}{2}k is a product of two integers larger than 1 as soon as k≥3k \ge3, the only triangular numbers that are not composite are T0=0T_0 = 0, T1=1T_1 = 1 and T2=3T_2 = 3. It is natural to ask how much of Gauss’s theorem survives when these three are removed.

Definition 1.1. Let E\mathcal{E} be the set of integers n≥0n \ge0 that cannot be written n=Ta+Tb+Tcn = T_a + T_b + T_c with a,b,c≥3a,b,c \ge3.

Theorem 1.2. (i) The elements of E\mathcal{E} not exceeding 8.4⋅1098.4 \cdot10^9 are exactly

0,1,…,17,19,20,21,23,24,25,28,29,32,34,38,0, 1, \ldots, 17,\quad19, 20, 21, 23, 24, 25, 28, 29, 32, 34, 38,
39,43,47,50,54,60,68,69,73,95,113,194,39, 43, 47, 50, 54, 60, 68, 69, 73, 95, 113, 194,

forty-one values, the largest being 194.

(ii) Every element of E\mathcal{E} larger than 8.4⋅1098.4 \cdot10^9 is of the form (f2N∗−3)/8(f^2N_* - 3)/8 with f∈{1,3,5}f \in\{1,3,5\}, where −N∗-N_* is the exceptional discriminant allowed by Lemma 5.1 with ε=1/20\varepsilon= 1/20, if there is one. In particular E\mathcal{E} has at most 44 elements, and every n>194n > 194 is a sum of three composite triangular numbers with at most three possible exceptions.

The exceptional discriminant may well not exist; if it exists, it need not be ≡5(mod8)\equiv5 \pmod{8}, in which case part (ii) gives no exception at all. What the theorem cannot do is decide this, which is the familiar limitation of every effective form of Siegel’s theorem. Under the Generalized Riemann Hypothesis the question disappears.

Corollary 1.3. Assume the Generalized Riemann Hypothesis, in the form used in [8] and [9]. Then E\mathcal{E} consists exactly of the 41 values listed in Theorem 1.2; every integer n>194n > 194 is a sum of three composite triangular numbers.

The closest result in the literature concerns sums of three nonzero triangular numbers, that is, the exclusion of T0T_0 alone. Kim, Lee and Oh [8] proved, under the Generalized Riemann Hypothesis, that every positive integer except 1, 2, 4, 6, 11, 20, 29 is such a sum. Excluding T0T_0, T1T_1 and T2T_2 amounts to excluding the three odd squares 1, 9, 25 instead of the single square 1. Their result enters our proof of Corollary 1.3, through Proposition 2.3. We are not aware of an earlier treatment of the present question. The recent historical survey of Aygin and Williams [1] is the natural place where one would be recorded; its abstract concerns formulas for the number of representations, and we have not been able to consult the full text.

Plan. Section 2 states the problem in terms of sums of three squares and treats every 8n+38n+3 divisible by the square of an odd q≥7q \ge7. Section 3 shows that the remaining cases follow from the inequality H(N)>3D((N−1)/2)H(N) > 3D((N-1)/2) between a class number and a divisor-type function. Section 4 bounds that function, and Section 5 recalls Tatuzawa’s lower bound for the class number. Section 6 closes the chain of ranges, Section 7 proves Corollary 1.3, and Section 8 describes the computation and the ancillary files. Section 9 states precisely what rests on what.

Reformulation and the elementary range

Lemma 2.1. Put N=8n+3N = 8n+3. Then n=Ta+Tb+Tcn = T_a + T_b + T_c with a,b,c≥0a,b,c \ge0 if and only if N=x2+y2+z2N = x^2 + y^2 + z^2 with x,y,zx,y,z odd, the correspondence being x=±(2a+1)x = \pm(2a+1), and similarly for y,zy,z. Every representation of NN as a sum of three squares has x,y,zx,y,z odd. Consequently n∉En \notin\mathcal{E} if and only if N=x2+y2+z2N = x^2 + y^2 + z^2 with ∣x∣,∣y∣,∣z∣≥7|x|,|y|,|z| \ge7, that is, with none of x2,y2,z2x^2,y^2,z^2 equal to 1, 9 or 25.

Proof. The identity 8Tk+1=(2k+1)28T_k+1 = (2k+1)^2 gives 8(Ta+Tb+Tc)+3=(2a+1)2+(2b+1)2+(2c+1)28(T_a+T_b+T_c)+3 = (2a+1)^2+(2b+1)^2+(2c+1)^2. Squares are 0, 1 or 4 modulo 8, and three of them sum to 3 modulo 8 only as 1+1+11+1+1, so x,y,zx,y,z are odd. An odd xx is ±(2a+1)\pm(2a+1) with a≥0a \ge0, and a≥3a \ge3 exactly when ∣x∣≥7|x| \ge7; the odd squares below 49 are 1, 9 and 25.

Proposition 2.2. If q2∣8n+3q^2 \mid8n+3 for some odd q≥7q \ge7, then n∉En \notin\mathcal{E}.

Proof. Every odd square is ≡1(mod8)\equiv1 \pmod{8}, so M=(8n+3)/q2≡3(mod8)M = (8n+3)/q^2 \equiv3 \pmod{8}. By Gauss’s theorem M=u2+v2+w2M = u^2+v^2+w^2, and by Lemma 2.1 u,v,wu,v,w are odd, hence non-zero. Then 8n+3=(qu)2+(qv)2+(qw)28n+3 = (qu)^2+(qv)^2+(qw)^2 with ∣qu∣,∣qv∣,∣qw∣≥q≥7|qu|,|qv|,|qw| \ge q \ge7.

Proposition 2.3. Let q∈{3,5}q \in\{3,5\} and suppose 8n+3=q2(8m+3)8n+3 = q^2(8m+3). If mm is a sum of three nonzero triangular numbers, then n∉En \notin\mathcal{E}.

Proof. If m=Ta+Tb+Tcm = T_a+T_b+T_c with a,b,c≥1a,b,c \ge1, then 8m+3=u2+v2+w28m+3 = u^2+v^2+w^2 with u=2a+1≥3u=2a+1 \ge3, and similarly for v,wv,w. Hence 8n+3=(qu)2+(qv)2+(qw)28n+3=(qu)^2+(qv)^2+(qw)^2 with qu≥3q≥9qu \ge3q \ge9.

Write N=8n+3=f2N0N = 8n+3 = f^2N_0 with N0N_0 squarefree and f≥1f \ge1; ff is odd. If f∉{1,3,5}f \notin\{1,3,5\}, then f≥7f \ge7 and Proposition 2.2 applies with q=fq=f. Since N0≡N≡3(mod8)N_0 \equiv N \equiv3 \pmod{8}, the number −N0≡5(mod8)-N_0 \equiv5 \pmod{8} is a fundamental discriminant. From now on f∈{1,3,5}f \in\{1,3,5\}.

The counting inequality

Let t3(n)t_3(n) be the number of ordered triples (a,b,c)(a,b,c) with a,b,c≥0a,b,c \ge0 and Ta+Tb+Tc=nT_a+T_b+T_c=n, and ρ2(m)\rho_2(m) the number of ordered pairs (b,c)(b,c) with b,c≥0b,c \ge0 and Tb+Tc=mT_b+T_c=m (with ρ2(m)=0\rho_2(m)=0 for m<0m<0).

Lemma 3.1. Let N=8n+3=f2N0N=8n+3=f^2N_0 as above, with N0>3N_0>3. Then

t3(n)=18r3(N)=3H(N),H(N)=∑d∣fh(−d2N0).t_3(n)=\frac{1}{8}r_3(N)=3H(N),\qquad H(N)=\sum_{d\mid f}h(-d^2N_0).

where r3r_3 counts representations as a sum of three squares and h(−d2N0)h(-d^2N_0) is the class number of primitive positive definite binary quadratic forms of discriminant $-d^2N_0. In particular

H(N)=h(−N0)⋅{1if f=1,4−χ(3)if f=3,6−χ(5)if f=5,H(N)=h(-N_0)\cdot \begin{cases} 1 & \text{if } f=1,\\ 4-\chi(3) & \text{if } f=3,\\ 6-\chi(5) & \text{if } f=5, \end{cases}

where χ\chi is the Kronecker character of −N0-N_0; hence H(N)≥fh(−N0)H(N)\ge f h(-N_0).

Proof. By Lemma 2.1 each triple (a,b,c)(a,b,c) corresponds to exactly 8 representations of NN, one for each choice of signs, so t3(n)=r3(N)/8t_3(n)=r_3(N)/8. For N≡3(mod8)N\equiv3\pmod8, Gauss’s theorem on primitive representations [4 art. 291–292], see also [5], gives 24h(−N)24h(-N) primitive representations of N>3N>3; summing over the squares e2∣Ne^2\mid N, that is over N/e2=d2N0N/e^2=d^2N_0 with d∣fd\mid f, gives r3(N)=24H(N)r_3(N)=24H(N). The class number of the order of conductor dd is h(−d2N0)=h(−N0)d∏p∣d(1−χ(p)/p)h(-d^2N_0)=h(-N_0)d\prod_{p\mid d}(1-\chi(p)/p), the unit index being 1 since N0>3N_0>3 [2 Section 7]; for d=3d=3 and d=5d=5 this is h(−N0)(3−χ(3))h(-N_0)(3-\chi(3)) and h(−N0)(5−χ(5))h(-N_0)(5-\chi(5)).

:::

Lemma 3.2. For m≥0m\ge0, ρ2(m)=d1(4m+1)−d3(4m+1)\rho_2(m)=d_1(4m+1)-d_3(4m+1), where di(M)d_i(M) is the number of divisors of MM congruent to ii modulo 4. Consequently

ρ2(m)≤∏pep∣4m+1p≡1(mod4)(ep+1).\rho_2(m)\le\prod_{\substack{p^{e_p}\mid4m+1\\p\equiv1\pmod4}}(e_p+1).

Proof. Tb+Tc=mT_b+T_c=m is (2b+1)2+(2c+1)2=8m+2(2b+1)^2+(2c+1)^2=8m+2. Every representation of 8m+28m+2 as a sum of two squares has both terms odd, so the pairs (b,c)(b,c) are the representations with both terms positive, that is r2(8m+2)/4r_2(8m+2)/4. By Jacobi’s formula [6 Thm. 278], r2(8m+2)=4(d1(8m+2)−d3(8m+2))r_2(8m+2)=4(d_1(8m+2)-d_3(8m+2)), and the odd divisors of 8m+28m+2 are those of 4m+14m+1. The last assertion is the usual evaluation of d1−d3d_1-d_3 as a multiplicative function: it vanishes unless every prime ≡3(mod4)\equiv3\pmod4 occurs to an even power, and then equals the product.

Proposition 3.3. If H(N)>ρ2(n)+ρ2(n−1)+ρ2(n−3)H(N)>\rho_2(n)+\rho_2(n-1)+\rho_2(n-3), then n∉En\notin\mathcal{E}.

Proof. For i∈{1,2,3}i\in\{1,2,3\} and j∈{0,1,2}j\in\{0,1,2\}, let A(i,j)A(i,j) be the set of triples counted by t3(n)t_3(n) whose ii-th entry is jj. Then ∣A(i,j)∣=ρ2(n−Tj)|A(i,j)|=\rho_2(n-T_j), and T0,T1,T2=0,1,3T_0,T_1,T_2=0,1,3. A triple outside ⋃A(i,j)\bigcup A(i,j) has all entries ≥3\ge3. Such a triple exists as soon as

t3(n)=3H(N)>∑i,j∣A(i,j)∣=3[ρ2(n)+ρ2(n−1)+ρ2(n−3)].t_3(n)=3H(N)>\sum_{i,j}|A(i,j)|=3\bigl[\rho_2(n)+\rho_2(n-1)+\rho_2(n-3)\bigr].

:::

Definition 3.4. For Y≥1Y\ge1 let

D(Y)=max⁡1≤M≤Y∏p∣M, p≡1(mod4)(ep+1),pep∥M.D(Y)=\max_{1\le M\le Y}\prod_{\substack{p\mid M,\ p\equiv1\pmod4}}(e_p+1),\qquad p^{e_p}\parallel M.

The three integers 4n+14n+1, 4n−34n-3, 4n−114n-11 attached by Lemma 3.2 to nn, n−1n-1, n−3n-3 are at most 4n+1=(N−1)/24n+1=(N-1)/2. Since DD is non-decreasing, Proposition 3.3 gives:

:::`

Corollary 3.5. If H(N)>3D((N−1)/2)H(N)>3D((N-1)/2), then n∉En\notin\mathcal{E}.

Remark 3.6. The union bound in Proposition 3.3 is crude: a triple with two forbidden entries is counted twice. Inclusion–exclusion would improve the constants of Section 6, but the chain closes without it, and the cruder form is easier to check.

The function DD

Let 5=q1<q2<q3<⋯5=q_1<q_2<q_3<\cdots be the primes ≡1(mod4)\equiv1\pmod{4}, and write v(M)=∏p∣M, p≡1(4)(ep+1)v(M)=\prod_{p\mid M,\ p\equiv1(4)}(e_p+1).

Lemma 4.1. For every YY, D(Y)D(Y) is the maximum of v(M)v(M) over the integers

M=q1e1q2e2⋯qrer≤Y,e1≥e2≥⋯≥er≥1.M=q_1^{e_1}q_2^{e_2}\cdots q_r^{e_r}\le Y,\qquad e_1\ge e_2\ge\cdots\ge e_r\ge1.

Proof. Given M≤YM\le Y, delete its prime factors ≢1(mod4)\not\equiv1\pmod{4}, then assign the exponents of the remaining primes, in non-increasing order, to q1,q2,…q_1,q_2,\ldots. The value vv does not change and, by the rearrangement inequality, the integer does not increase.

The integers of Lemma 4.1 up to 101310^{13} involve at most eight primes, since q1⋯q9>1013q_1\cdots q_9>10^{13}, and are enumerated exhaustively by the ancillary verifier. The records that matter are in Table 1; the enumeration is checked against a brute-force scan of all M≤106M\le10^6.

DDsmallest MMfactorisation
96243 061 32552⋅13⋅17⋅29⋅37⋅415^2\cdot13\cdot17\cdot29\cdot37\cdot41
1281 215 306 62553⋅13⋅17⋅29⋅37⋅415^3\cdot13\cdot17\cdot29\cdot37\cdot41
1443 159 797 22552⋅132⋅17⋅29⋅37⋅415^2\cdot13^2\cdot17\cdot29\cdot37\cdot41
1606 076 533 12554⋅13⋅17⋅29⋅37⋅415^4\cdot13\cdot17\cdot29\cdot37\cdot41
19212 882 250 22552⋅13⋅17⋅29⋅37⋅41⋅535^2\cdot13\cdot17\cdot29\cdot37\cdot41\cdot53
21653 716 552 82552⋅132⋅172⋅29⋅37⋅415^2\cdot13^2\cdot17^2\cdot29\cdot37\cdot41

Table 1. Records of DD: D(Y)D(Y) first reaches the value in the first column at Y=MY=M.

In particular

D(Y)≤192for Y<53 716 552 825.(1)D(Y)\le192\qquad\text{for }Y<53\,716\,552\,825. \tag*{(1)}

Lemma 4.2. For t>0t>0 put g(t)=∏p≡1(4), pt<2max⁡e≥0(e+1)p−etg(t)=\prod_{p\equiv1(4),\ p^t<2}\max_{e\ge0}(e+1)p^{-et}. Then D(Y)≤g(t)YtD(Y)\le g(t)Y^t for all Y≥1Y\ge1. Moreover g(0.2)<2.18407g(0.2)<2.18407, the maxima being attained at e=2,1,1,1e=2,1,1,1 for p=5,13,17,29p=5,13,17,29.

Proof. For M≤YM\le Y, v(M)=∏p(ep+1)p−ept⋅∏ppept≤∏p(ep+1)p−ept⋅Mtv(M)=\prod_p(e_p+1)p^{-e_pt}\cdot\prod_p p^{e_pt}\le\prod_p(e_p+1)p^{-e_pt}\cdot M^t, the products running over p≡1(mod4)p\equiv1\pmod{4}. If pt≥2p^t\ge2 then (e+1)p−et≤(e+1)2−e≤1(e+1)p^{-et}\le(e+1)2^{-e}\le1. The numerical value is computed by the verifier.

Tatuzawa's theorem

Lemma 5.1 (Tatuzawa). Let 0<ε<1/20<\varepsilon<1/2. There is at most one fundamental discriminant D<0D<0 with ∣D∣≥max⁡(e1/ε,e11.2)|D|\ge\max(e^{1/\varepsilon},e^{11.2}) such that

L(1,χD)≤0.655ε∣D∣−ε.L(1,\chi_D)\le0.655\varepsilon|D|^{-\varepsilon}.

This is [10 Theorem 2], proved there from Lemma 9; it is restated by Elsenhans, Klüners and Nicolae [3 Lemma 2] and used by Hopkins and Stopple [7]. We take ε=1/20\varepsilon= 1/20, so that max⁡(e20,e11.2)=e20=4.8517⋅108\max(e^{20},e^{11.2}) = e^{20} = 4.8517 \cdot10^{8}. If −N0≠−N∗-N_0 \ne-N_* is a fundamental discriminant with N0≥e20N_0 \ge e^{20}, then, by the class number formula h(−N0)=N0L(1,χ)/πh(-N_0) = \sqrt{N_0}L(1,\chi)/\pi,

h(−N0)>AN00.45,A=0.65520π=0.01042464877…(2)h(-N_0) > A N_0^{0.45}, \qquad A = \frac{0.655}{20\pi} = 0.01042464877\ldots \tag*{(2)}

Here and below, −N∗-N_* denotes the exceptional discriminant of Lemma 5.1 for ε=1/20\varepsilon= 1/20, if there is one.

Remark 5.2. The proof requires the form of Lemma 5.1, with the constant 0.655ε0.655\varepsilon, as given in [10 Theorem 2].

Proof of Theorem 1.2

Let N=8n+3=f2N0N = 8n+3 = f^2N_0 with f∈{1,3,5}f \in\{1,3,5\}, N0N_0 squarefree, and suppose N0≠N∗N_0 \ne N_*. By Lemma 3.1 and (2), as long as N0≥e20N_0 \ge e^{20},

H(N)≥fh(−N0)>f⋅f−0.9AN00.45≥AN00.45,H(N) \ge f h(-N_0) > f \cdot f^{-0.9} A N_0^{0.45} \ge A N_0^{0.45},

since 30.1=1.116…3^{0.1} = 1.116\ldots and 50.1=1.174…5^{0.1} = 1.174\ldots. By Corollary 3.5 it is therefore enough to have

AN00.45≥3D((N−1)/2).(3)A N_0^{0.45} \ge3D((N-1)/2). \tag*{(3)}

The condition N0≥e20N_0 \ge e^{20} holds throughout the analytic ranges below, which start at N>3.4⋅1010N > 3.4 \cdot10^{10}, so that N0≥N/25>1.38⋅109>e20N_0 \ge N/25 > 1.38 \cdot10^9 > e^{20}.

Range 1: N>8.9635⋅1010N > 8.9635 \cdot10^{10}. By Lemma 4.2 with t=0.2t = 0.2, 3D((N−1)/2)≤3g(0.2)(N/2)0.23D((N-1)/2) \le3g(0.2)(N/2)^{0.2}, so (3) holds as soon as

N0.25≥3g(0.2)A⋅20.2,N^{0.25} \ge\frac{3g(0.2)}{A \cdot2^{0.2}},

that is N≥8.9635⋅1010N \ge8.9635 \cdot10^{10}.

Range 2: 3.4561⋅1010<N≤1.0743⋅10113.4561 \cdot10^{10} < N \le1.0743 \cdot10^{11}. Here (N−1)/2<53 716 552 825(N-1)/2 < 53\,716\,552\,825, so D((N−1)/2)≤192D((N-1)/2) \le192 by (1), and (3) holds as soon as AN00.45≥576A N_0^{0.45} \ge576, that is

N≥(576A)1/0.45=3.456051⋅1010.N \ge\left(\frac{576}{A}\right)^{1/0.45} = 3.456051 \cdot10^{10}.

Range 2 extends beyond the start of Range 1, so Ranges 1 and 2 together cover every N>3.4561⋅1010N > 3.4561 \cdot10^{10}, that is every n>4.3201⋅109n > 4.3201 \cdot10^9.

Range 3: N≤6.72⋅1010N \le6.72 \cdot10^{10}. By direct computation (Section 8), the elements of E\mathcal{E} up to 8.4⋅1098.4 \cdot10^9, that is with N≤8⋅8.4⋅109+3=67 200 000 003N \le8 \cdot8.4 \cdot10^9 + 3 = 67\,200\,000\,003, are exactly the 41 values of part (i).

Closure. Ranges 2 and 3 overlap, by the factor

67 200 000 0033.456051⋅1010=1.9444.\frac{67\,200\,000\,003}{3.456051 \cdot10^{10}} = 1.9444.

Hence every n>8.4⋅109n > 8.4 \cdot10^9 with N0≠N∗N_0 \ne N_* lies outside E\mathcal{E}. The remaining integers have N∈{N∗,9N∗,25N∗}N \in\{N_*,9N_*,25N_*\}, which is part (ii). Part (i) is the computation.

Remark 6.1. The factor 1.9444 is the whole of the slack between the analytic and the computed ranges: a factor 2 lost in (3) would separate them. Every constant entering the chain is recomputed in the ancillary verifier in 50-digit decimal arithmetic, with π\pi replaced by an upper bound so that (2) is never overstated, and the closure of the chain is checked by assertions.

Proof of Corollary 1.3

By Theorem 1.2 it suffices to show that, under GRH, no n>8.4⋅109n > 8.4 \cdot10^{9} with 8n+3∈{N∗,9N∗,25N∗}8n + 3 \in\{N_{*}, 9N_{*}, 25N_{*}\} lies in E\mathcal{E}.

If 8n+3=q2N∗8n + 3 = q^{2}N_{*} with q∈{3,5}q \in\{3,5\}, then N∗≡3(mod8)N_{*} \equiv3 \pmod{8}, say N∗=8m+3N_{*} = 8m + 3, and m>(e20−3)/8>29m > (e^{20} - 3)/8 > 29. By the theorem of Kim, Lee and Oh [8 Cor. 3.11], valid under GRH, mm is a sum of three nonzero triangular numbers, and Proposition 2.3 gives n∉En \notin\mathcal{E}.

If 8n+3=N∗8n + 3 = N_{*}, then N∗>6.72⋅1010N_{*} > 6.72 \cdot10^{10}. Under GRH, Lamzouri, Li and Soundararajan [9 Thm. 1.5] prove, for every primitive character χ\chi modulo q≥1010q \ge10^{10},

1∣L(1,χ)∣≤12eγπ2(log⁡log⁡q−log⁡2+12+1log⁡log⁡q+14log⁡log⁡qlog⁡q).\frac{1}{\lvert L(1,\chi)\rvert} \le\frac{12e^{\gamma}}{\pi^{2}}\left(\log\log q-\log2+\frac{1}{2}+\frac{1}{\log\log q}+\frac{14\log\log q}{\log q}\right).

At q=6.72⋅1010q = 6.72 \cdot10^{10} this gives L(1,χ)≥0.0898L(1,\chi) \ge0.0898, while the exceptional inequality of Lemma 5.1 would require L(1,χ)≤0.655⋅120q−1/20=0.00942L(1,\chi) \le0.655 \cdot\frac{1}{20}q^{-1/20} = 0.00942; the ratio between the two bounds is at least 9.54 for every q≥6.72⋅1010q \ge6.72 \cdot10^{10}, since the first decreases only like 1/log⁡log⁡q1/\log\log q while the second decreases like q−1/20q^{-1/20}. Hence N∗N_{*} does not exist in this range.

Computation

The ancillary file VERIFY_range.py recomputes E∩[0,8.4⋅109]\mathcal{E} \cap[0, 8.4 \cdot10^{9}] by two algorithms that share no code except an exact test used on rare survivors. Both work on segments of 2262^{26} consecutive integers.

Method A. On each segment, the set of sums Ta+TbT_a + T_b with 3≤a≤b3 \le a \le b is built on a window reaching just below the segment, and nn is declared covered as soon as n−Tkn - T_k lies in it for one of the 64 composite triangular numbers T3,…,T66T_3,\ldots,T_{66}.

Method B. The set of sums Ta+TbT_a + T_b with 3≤a≤b3 \le a \le b is built once on [0,4⋅106][0, 4 \cdot10^{6}], and nn is declared covered as soon as n−Tcn - T_c lies in it for some cc. For n≤4⋅106n \le4 \cdot10^{6} this test is exhaustive.

Every integer that is not declared covered is passed to an exact test, which tries every c≥3c \ge3 and decides whether n−Tcn - T_c is a sum of two composite triangular numbers; the exceptions are therefore certified by exhaustion. As a guard against a marking error, 24 integers declared covered are drawn at random in each segment and re-tested exactly. Up to 8.4⋅1098.4 \cdot10^{9}, method A passes 368 integers to the exact test and method B passes 16 341; both return exactly the 41 values of Theorem 1.2 (i). The two runs together take about six and a half minutes on two cores.

The file VERIFY_composite_triangular.py recomputes the exceptional set up to 2⋅1072 \cdot10^{7} by a third implementation, the records of DD (Table 1) by the enumeration of Lemma 4.1 checked against a brute-force scan up to 10610^{6}, the constants AA, g(t)g(t), the thresholds and the overlap of § 6, and the identities of Lemma 3.1 for f=1f = 1 (11≤N<200011 \le N < 2000), f=3f = 3 (11≤N0<40011 \le N_{0} < 400) and f=5f = 5 (11≤N0<20011 \le N_{0} < 200), with class numbers computed by counting reduced forms. The files CHECK_classnumbers.magma and CHECK_exceptions.magma repeat the class number identities and the exceptional set up to 2⋅1052 \cdot10^{5} under Magma, with Magma’s own class number routine. Both Python files require only Python 3 and numpy, and take no input.

As consistency checks, none of the 41 exceptions has 8n+38n + 3 divisible by q2q^{2} with q≥7q \ge7 odd, as Proposition 2.2 requires; and the six exceptions with 9∣8n+39 \mid8n + 3 or 25∣8n+325 \mid8n + 3, namely n=3,9,12,21,34,39n = 3, 9, 12, 21, 34, 39, give m=0,0,1,2,1,4m = 0, 0, 1, 2, 1, 4 in Proposition 2.3, all of them on the list of Kim, Lee and Oh.

Status of the argument

Proved here: Lemmas 2.1, 3.2, 4.1, 4.2, Propositions 2.2, 2.3, 3.3, and the deduction of Theorem 1.2 (ii) and Corollary 1.3 from the inputs below.

Verified by computation: the 41 exceptions up to 8.4⋅1098.4 \cdot10^9; the records of DD; the constants and the closure of the chain; the identities of Lemma 3.1 on the ranges stated.

Quoted: Gauss’s formula for r3r_3 and the class number of orders (Lemma 3.1); Jacobi’s formula for r2r_2; Tatuzawa’s theorem (Lemma 5.1), checked against the original [10 Theorem 2]; and, for Corollary 1.3 only, [8 Cor. 3.11] and [9 Thm. 1.5], both conditional on GRH.

Open: whether the exceptional discriminant can be excluded unconditionally, which would make the list of 41 values complete without any reserve.

Acknowledgements. Technical assistance: Claude (Anthropic).

References

  1. [1]Z. S. Aygin and K. S. Williams, Historical survey of sums of three squares, Expo. Math. 44 (2026), no. 3, article 125776. doi:10.1016/j.exmath.2026.125776.DOI
  2. [2]D. A. Cox, Primes of the Form x² + ny²: Fermat, Class Field Theory, and Complex Multiplication, Wiley, New York, 1989.
  3. [3]A.-S. Elsenhans, J. Klüners and F. Nicolae, Imaginary quadratic number fields with class groups of small exponent, Acta Arith. 193 (2020), 217–233. doi:10.4064/aa180220-20-3.DOI
  4. [4]C. F. Gauss, Disquisitiones Arithmeticae, Leipzig, 1801.
  5. [5]E. Grosswald, Representations of Integers as Sums of Squares, Springer, New York, 1985. doi:10.1007/978-1-4613-8566-0.DOI
  6. [6]G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, Oxford University Press.
  7. [7]K. Hopkins and J. Stopple, Lower bounds for the principal genus of definite binary quadratic forms, Integers 10 (2010). doi:10.1515/integ.2010.021.DOI
  8. [8]D. Kim, J. Lee and B.-K. Oh, A sum of three nonzero triangular numbers, Int. J. Number Theory 17 (2021), no. 10, 2279–2300. doi:10.1142/S1793042121500883.DOI
  9. [9]Y. Lamzouri, X. Li and K. Soundararajan, Conditional bounds for the least quadratic non-residue and related problems, Math. Comp. 84 (2015), no. 295; Corrigendum, Math. Comp. 86 (2017), 2551–2554. Corrected version: arXiv:1309.3595v3.arxiv.org/abs/1309.3595
  10. [10]T. Tatuzawa, On a theorem of Siegel, Japan. J. Math. 21 (1951), 163–178. doi:10.4099/jjml1924.21.0_163.DOI

Paper details

vlnc:W_ub2vtkwc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Anthropic Claude Opus

View statement
Looking for related papers…
Contents