We prove lower bounds for central L-values of quadratic twists of an elliptic curve by discriminants with at most five prime factors. Under the rank-zero Birch–Swinnerton-Dyer formula, these bounds give large Tate–Shafarevich groups.
Introduction
The resonance method finds extreme values of L-functions by weighting them with suitable Dirichlet polynomials. Soundararajan [12] used it to study large values of the Riemann zeta-function and extreme central values in families of L-functions. We use the method to find large values in a sparse family of quadratic twists.
For quadratic twists of elliptic curves, the resonance method gives large central L-values [6]. Hua and Huang [7] proved such bounds for discriminants with at most W prime factors, for every fixed W≥20. Here we obtain bounds for W≥5.
Let E/Q be an elliptic curve of conductor N, and let Ed be its quadratic twist by d. We normalize the L-function so that its central point is s=1/2. Put N0=lcm(8,N). A class a (mod N0) is admissible if (a,N0)=1 and a≡1 or 5 (mod 8). For such a class, let Da− be the set of negative fundamental discriminants d≡a (mod N0). Write wa for the common root number of these twists. We fix an admissible class with wa=+1. For an integer W≥1, let Fa,W(X) be the set of d∈Da− with X≤∣d∣≤2X and ω(∣d∣)≤W, where ω(n) counts the distinct prime divisors of n. Put HX=logX/loglogX.
Theorem 1. Let E and a (mod N0) be as above. For every fixed integer W≥5 and every fixed 0<c<(7W−34)/(7W+1), the set Fa,W(X) is nonempty for all sufficiently large X, and
d∈Fa,W(X)maxL(1/2,Ed)≥exp(cHX).
For W=5, one may take any 0<c<1/6. The proof combines the resonance method with a weighted sieve.
Large central values also give large Tate–Shafarevich groups under the Birch–Swinnerton-Dyer formula. Goldfeld and Szpiro [4] studied bounds for their orders, and Radziwiłł and Soundararajan [11] studied their distribution in quadratic twist families. When L(1/2,Ed)=0, the group \Sha(E_d/\mathbb{Q}) is finite [5, 8]. The rank-zero formula relates its order to the central value, the real period, the rational torsion and the Tamagawa product. For the twists here, the period is proportional to ∣d∣−1/2, and the Tamagawa product is bounded when ω(∣d∣)≤W. Hua and Huang [7] used these facts to find large Tate–Shafarevich groups in almost-prime twist families. Theorem 1 gives the following result with at most five prime factors.
Corollary 1. Let E, a, W and c be as in the Theorem 1. Assume the rank-zero Birch–Swinnerton-Dyer formula for the twists Ed with d∈Da− and L(1/2,Ed)=0. For every sufficiently large X, there is d∈Fa,W(X) such that
#III(Ed/Q)≥Xexp(cHX).(1)
For elliptic curves with complex multiplication the conclusion holds unconditionally.
The first moment and the resonator
Write L(s,E)=∑n≥1a(n)n−s and χd(n)=(nd). The central values Ld=L(1/2,Ed) are nonnegative [11]. Fix a nonzero nonnegative Φ∈Cc∞((1,2)).
For p∤N0, write Qp=1−(a(p)2−2)/p+p−2 and Tp=1+p−1+Qp/p. Define g multiplicatively by g(1)=1, g(p2r−1)=Tp−1 and g(p2r)=(1+p−1)Tp−1 for r≥1. Put δ(p)=1−g(p2)=Qp/(pTp), and extend δ multiplicatively to squarefree integers coprime to N0. By [6 Lemma 2.1, r = 0], for (t,N0)=1 and t=t1t22 with t1 squarefree,
Fix θ>0, 0<ρ<1, and put M=Xθ, L=ρlogMloglogM, P−=L2, and P+=exp((logL)2). Define a real multiplicative function b supported on squarefree integers by b(1)=1. For primes P−≤p≤P+ with p∤N0, set b(p)=a(p)L/(plogp); set b(p)=0 for all other primes. Let R(d)=∑m≤Mb(m)χd(m), A0(p)=1+b(p)2p/(p+1) and A1(p)=1+b(p)2g(p2)+2a(p)b(p)g(p)/p. Let Bi=∏p∤N0Ai(p) for i=0,1. These products are finite. Rankin–Selberg theory and partial summation give
Here logp≥log2, and the harmonic sum bounds the exponent sum by 4(1+logP+) for large M. Also L(1+(logL)2)=o(logM). For the second assertion write m1=qr, m2=qs, with q,r,s pairwise coprime and squarefree. The summand becomes
b(q)2g(q2)ra(r)b(r)g(r)sa(s)b(s)g(s),
which is nonnegative. The unrestricted sum is the product of A1(p) over p∤ℓN0. Put ν=(logL)−3. Rankin’s trick bounds the relative contribution of qr>M or qs>M by
On the support, pν−1=νlogp(1+o(1)) uniformly. By (2), the quadratic term with the factor logp sums to (ρ2+o(1))logM. The linear term sums to O(LloglogL)=o(logM). Omitting primes dividing ℓ decreases these sums. The exponent is at most −(1−ρ2+o(1))νlogM.
Lemma 2. We have log(B1/B0)=(2ρθ+o(1))HX.
Proof. On the support of b, put xp=2a(p)b(p)g(p)/p and yp=b(p)2g(p2). Then
Proof. For a fundamental discriminant d, 1ℓ∣d=∑q∣ℓμ(q)χd(q2). Apply [6 Lemma 2.1, r = 0] to each insertion uq2. Since (u,ℓ)=1, all squarefree parts are u1 and g(uq2)=g(u)g(q2). As ℓ is squarefree, ∑q∣ℓμ(q)g(q2)=δ(ℓ). Thus ∑q∣ℓμ(q)g(uq2)=g(u)δ(ℓ). The errors sum to O(X1/2+εu1/2+ε∑q∣ℓq1+2ε). Use ∑q∣ℓq1+2ε≪εℓ1+3ε and choose smaller intermediate exponents. □
Lemma 4. Let M=Xθ, D=Xγ and 4θ+4γ<1. Put X=C1XB1, G(p)=δ(p)/A1(p) for p∤N0, and G(p)=0 for p∣N0. Extend G multiplicatively to squarefree integers. For squarefree ℓ coprime to N0, write
Set rℓ=0 for all other ℓ. Then, for every fixed A>0,
ℓ≤Dμ(ℓ)2=1(ℓ,N0)=1∑∣rℓ∣≪AX(logX)−A.
Proof. Expand R(d)2. Terms with (m1m2,ℓ)>1 vanish. Apply Lemma 3 to the other terms. The truncation formula (4) gives the main term XG(ℓ) and a relative error O(τM). The coefficient bound (3) gives
∣rℓ∣≪XG(ℓ)τM+X1/2+εM2+o(1)ℓ1+ε.
We have 0≤G(p)<1/p. To see this, set x=1/p and q=Qp. Hasse’s bound gives
(1−x)2≤q≤(1+x)2,Tp−q=(1−x)((1+x)2−q)+x2(1+x)>0.
Moreover g(p),g(p2)>0 and a(p)b(p)≥0, so A1(p)≥1. It follows that 0≤G(p)≤δ(p)<1/p when p∤N0; for p∣N0 this is immediate from G(p)=0. In particular, B1≥1, and hence X≥C1X. Thus
ℓ≤D∑∣rℓ∣≪XτMlogD+X1/2+εM2+o(1)D2+ε.
Indeed, for each sufficiently small fixed ε>0 the exponent of X in the second term is at most 1/2+2θ+2γ+ε(1+γ)+o(1), which is strictly less than 1 by 4θ+4γ<1. Since X≥C1X, this gives a fixed power saving relative to X. The first term is smaller than X(logX)−A for every fixed A. □
The weighted sieve
Proof of Theorem 1. Choose fixed ρ∈(0,1) and θ,γ>0 with c<2ρθ. We also require 35/(4(7W+1))<γ<1/4−θ. Take the resonator of § 2, and put D=Xγ, z=D1/5, P(z)=∏p<zp∤N0p and V(z)=∏p∣P(z)(1−G(p)). The inequality G(p)<1/p gives the uniform dimension-one condition and V(z)≫E1/logz by Mertens’ theorem.
Let F and f be the upper and lower linear sieve functions [3 Theorem 11.13], and let C=2eγEM, where γEM is Euler’s constant. The sieve formulas give F(t)=C/t for 1<t≤3 and f(4)=Clog3/4. For 3≤r≤4, the delay equation gives
rF(r)=C(1+∫3rt−1log(t−2)dt)≤3Cr,
since log(t−2)≤t−3≤(t−1)/3 on this interval. It also gives
5f(5)=4f(4)+∫34F(t)dt≥Clog4,f(5)≥5Clog4>f(4).
For 1<v≤4, put I(v)=∫1v(1−t/v)F(5−t)dt/t. Splitting at t=2 and using the preceding bounds gives
I(4)≤C(158log2−121−201log3)<76f(4).
Thus I(v)≤I(4)<(6/7)f(5) for v≤4. Fix η=6/7 and choose 1<v<4 close enough to 4 that γ>5/(v(W+1/7)). Put y=zv and define the logarithmic weight
w(d)=1−η1z≤p<yp∣d∑(1−logylogp).
For (d,P(z))=1, all prime factors of ∣d∣ are at least z. If ω(∣d∣)≥W+1, then
z≤p<yp∣d∑(1−logylogp)≥W+1−logylog(2X)>η
for large X. Thus w(d)≤0 on these twists, and w(d)≤1 on all sifted twists. Write ad=L(1/2,Ed)R(d)2Φ(∣d∣/X)≥0, and put
S(z)=(d,P(z))=1∑ad,Sp(z)=(d,P(z))=1p∣d∑ad.
All sums over d below are over Dα−. The lower linear sieve at level D gives S(z)≥XV(z)(f(5)+o(1)). For z≤p<y, apply the upper linear sieve to the sequence {ad1p∣d} at level D/p. Its divisor sums are XG(p)G(ℓ)+rpℓ for ℓ∣P(z), and its sieve parameter is 5−logp/logz∈[5−v,4]. Consequently,
The o(1) is uniform in p since 5−v>1 is fixed. Each pℓ in this sum has exactly one prime factor at least z, so Lemma 4 gives
z≤p<y∑ℓ≤D/pℓ∣P(z)∑∣rpℓ∣≤q≤D∑∣rq∣=o(XV(z)).
Here logL=O(loglogX), so logP+=o(logX). Since logz=(γ/5)logX, we have P+<z for large X. Also z>N0 eventually. Consequently, for every prime p≥z, b(p)=0, p∤N0, and G(p)=δ(p); the formulas above and Hasse’s bound give the uniform estimate ∣G(p)−1/p∣≤6/p2. The function F is bounded on the fixed compact interval [5−v,4], and the logarithmic weight lies in [0,1]. Hence replacing G(p) by 1/p in the following sum changes it by at most Ov(∑n≥zn−2)=o(1). The prime number theorem and partial summation then give
Since T>0, some term has w(d)>0 and ω(∣d∣)≤W. Discard the terms with w(d)≤0. For the remaining terms, w(d)≤1, so
T≤d∈Fa,W(X)maxL(1/2,Ed)d∈Da−∑R(d)2Φ(∣d∣/X).
The sum on the right is ≪XB0 by [6 Proposition 3.1], since θ<1/4. Therefore
d∈Fa,W(X)maxL(1/2,Ed)≫B0logXB1.
By Lemma 2, log(B1/B0)=(2ρθ+o(1))HX. As loglogX=o(HX) and c<2ρθ, the asserted bound follows. □
Large Tate–Shafarevich groups
For an elliptic curve A/Q, our normalization gives L(1/2,A)=LHW(1,A), where LHW is the classical Hasse–Weil L-function. Let ωA be a minimal Néron differential and put Ω(A)=∫A(R)∣ωA∣ and Tam(A)=∏p<∞cp(A). The integral includes all c∞(A)=#π0(A(R)) connected components. Write H1(A(C),Z)± for the eigensubgroups of complex conjugation, and define Ω−(A)=∫γAωA using a generator γA of the minus subgroup. This period is purely imaginary and determined up to sign. Put
For any fixed W≥1, we have B(Ed)≫E,W∣d∣L(1/2,Ed) if ω(∣d∣)≤W.
Proof. By [10 Theorem 3.2], Ω(Ed)=u~(d)c∞(Ed)∣Ω−(E)∣/∣d∣, where u~(d) is the minimal-model correction. Here d is odd, squarefree, and congruent to 1 modulo 4. It is coprime to the minimal discriminant of E, so u~(d)=1 by [10 Proposition 2.5 and Corollary 2.6]. Quadratic twisting preserves the sign of the discriminant, so c∞(Ed)=c∞(E). This proves (7).
Choose the twisting isomorphism ϕ:Ed→E over C so that ϕ∗ωE=dωEd; see [10 Lemma 3.1]. Complex conjugation satisfies σE∘ϕ=[−1]∘ϕ∘σEd. Thus ϕ∗ exchanges the two integral homology eigensubgroups and maps γEd to a generator of H1(E(C),Z)+. The absolute period of this generator is Ω(E)/c∞(E); see [10 Lemma 5.1]. Hence
∣d∣∣Ω−(Ed)∣=∫ϕ∗γEdωE=c∞(E)Ω(E).
The complementary formula and the product identity follow.
We have cp(Ed)=1 for p∤dN. If p∣d, then E has good reduction at p and Ed has type I0∗, with cp(Ed)=#E(Fp)[2]∈{1,2,4}; see [1 Theorem 4.1 and Table 4]. For p∣2N, the local twist depends only on the squareclass of d in Qp×. There are finitely many such classes, so their Tamagawa numbers are bounded in terms of E. This gives (8). Since #Ed(Q)tors≥1, the last bound follows from (7), (8) and (5). □
Proof of Corollary 1. Take c′ such that c<c′<(7W−34)/(7W+1). Theorem 1 gives d∈Fa,W(X) with L(1/2,Ed)≥exp(c′HX). Lemma 5 gives B(Ed)≫E,WXexp(c′HX). Since c′>c, we have B(Ed)≥Xexp(cHX) for all sufficiently large X.
The nonzero central value implies that Ed(Q) and \Sha(E_d/\mathbb{Q}) are finite by the Gross–Zagier–Kolyvagin theorem [5, 8]. Under (6), this proves (1). If E has complex multiplication, then so does Ed. By [2 Corollary 2 and the paragraph following its proof, p. 362], the full rank-zero BSD formula holds for every CM elliptic curve over Q with nonzero central value. This gives the unconditional assertion for CM curves. □
Remark 1 (The Tamagawa denominator). For d∈Da−, the local squareclass of d is fixed at every p∣2N. The number BE,a=∏p∣2Ncp(Ed) is therefore constant, and Tam(Ed)=BE,a∏p∣d#E(Fp)[2]. Frobenius acting on E[2] determines each factor; see also [11 §5]. We have
For the upper bound, use Tam(Ed)≤BE,a4ω(∣d∣). If r=ω(n), then n≥(r+1)!. Stirling’s formula gives ω(n)≤(1+o(1))logX/loglogX uniformly for n≤2X.
For the lower bound, let K=Q(E[2]) and δ=[K:Q]−1. Let P be the primes p∤2N that split completely in K, and put πP(y)=#{p≤y:p∈P}. Each such prime satisfies #E(Fp)[2]=4. Chebotarev gives
πP(y)∼logyδy,p≤yp∈P∑logp∼δy;
see [9 Lemma 2.4]. Fix 0<ε<1, and put y=(1−ε)logX/δ and m=∏p≤yp∈Pp. Then logm=(1−ε+o(1))logX and Y=X/m=Xε+o(1). The prime number theorem in fixed arithmetic progressions gives a prime q∈[Y,2Y] with q≡−am−1(modN0). For large X, we have q>y and q∤m. The integer d=−qm is squarefree, belongs to Da−, and satisfies X≤∣d∣≤2X. Therefore
If E(Q)[2]=0, the local product gives Tam(Ed)≥BE,a2ω(∣d∣). If all 2-torsion points are rational, then Tam(Ed)=BE,a4ω(∣d∣). For these curves, a uniform bound for Tam(Ed) also bounds ω(∣d∣). The almost prime condition controls this denominator for every fixed elliptic curve. More generally, the condition logTam(Ed)=o(HX) on the selected twists preserves the central-value gain.
Acknowledgements
S.H. would like to thank Chung Pang Mok for the constant encouragement. S.H. is supported by the National Natural Science Foundation of China (No. 12601008) and the China Postdoctoral Science Foundation (No. 2026M793348).
AI Use Disclosure
The author introduced sieve weights into the argument, using calculations based on earlier work, and first reduced the number of prime factors to six. ChatGPT introduced Richert’s logarithmic weighted sieve to reduce the bound to five and prepared the draft.
References
Shanghai Institute for Mathematics and Interdisciplinary Sciences (SIMIS), Shanghai 200433, China
Research Institute of Intelligent Complex Systems, Fudan University, Shanghai, 200433, China
[1]Barrios, A. J.; Roy, M.; Sahajpal, N.; Tallana, D.; Tobin, B.; Wiersema, H. Local data of elliptic curves under quadratic twist. Res. Number Theory 11 (2025), no. 3, Paper No. 75, 39 pp.DOI
[2]Burungale, Ashay A.; Flach, Matthias. The conjecture of Birch and Swinnerton-Dyer for certain elliptic curves with complex multiplication. Cambridge J. Math. 12 (2024), no. 2, 357–415.arxiv.org/abs/2206.09874
[3]Friedlander, John B.; Iwaniec, Henryk. Opera de cribro. American Mathematical Society Colloquium Publications 57, American Mathematical Society, 2010.
[4]Goldfeld, Dorian; Szpiro, Lucien. Bounds for the order of the Tate-Shafarevich group. Compositio Math. 97 (1995), no. 1–2, 71–87.
[5]Gross, Benedict H.; Zagier, Don B. Heegner points and derivatives of L-series. Invent. Math. 84 (1986), no. 2, 225–320.
[6]Hamdar, Mohammad H.; Leung, Sun-Kai. Extreme values of central L-derivatives and heights of Heegner points. (Preprint). arXiv:2609.26786v1 (2026).arxiv.org/abs/2609.26786
[7]Hua, Shenghao; Huang, Bingrong. Extreme central L-values of almost prime quadratic twists of elliptic curves. Sci. China Math. 66 (2023), 2755–2766.DOI
[8]Kolyvagin, V. A. Finiteness of E(Q) and Ш(E,Q) for a subclass of Weil curves. Math. USSR-Izv. 32 (1989), no. 3, 523–541.DOI
[9]Loughran, Daniel; Matthiesen, Lilian. Frobenian multiplicative functions and rational points in fibrations. J. Eur. Math. Soc. (JEMS) 26 (2024), no. 12, 4779–4830.arxiv.org/abs/1904.12845
[10]Pal, Vivek. Periods of quadratic twists of elliptic curves. With an appendix by A. Agashe. Proc. Amer. Math. Soc. 140 (2012), no. 5, 1513–1525.arxiv.org/abs/1012.0094
[11]Radziwiłł, Maksym; Soundararajan, Kannan. Moments and distribution of central L-values of quadratic twists of elliptic curves. Invent. Math. 202 (2015), 1029–1068.DOI
[12]Soundararajan, Kannan. Extreme values of zeta and L-functions. Math. Ann. 342 (2008), no. 2, 467–486.arxiv.org/abs/0708.3990