Arxiv:2012.03523V3 [Math.NT] 16 May 2021 Favne Cetﬁ Optn Eerh(SR Spr Ft of Part (CM4)

Home , Adjugate matrix

arXiv:2012.03523v3 [math.NT] 16 May 2021 fAvne cetfi optn eerh(SR spr ft of part (CM4). as Materials (ASCR) of Research Computing Scientific Advanced of Keywords Date hsrsac a upre npr yteApidMathemat Applied the by part in supported was research This * 2020) (AMS Classification Subject 2. Introduction 1. ..Bokdaoaiaino nselWro on-shell of diagonalization Block 2.2. ..Wro 2.1. ..Atraierpeettoso eRa matrices Rham de of representations References Alternative diagrams Feynman to applications 4.3. moments Arithmetic Bessel on-shell for relations 4.2. Quadratic relations quadratic 4.1. moments Broadhurst–Roberts Bessel off-shell for relations 4. Quadratic 3.3. ..Smrlsfro-hl n f-hl eslmmns11 moments Bessel off-shell and on-shell for rules 3. Sum 2.3. ..Trsodbhvo fWro of behavior Threshold 3.2. ..Cfcoso Wro of Cofactors 3.1. WRO a 8 2021. 18, May : W W ln ihsm eeaiain.I h Wro the In generalizations. rel some quadratic the with of proof along analytic new a present we moments, ihDlgescnetrsfrciia auso motivic of values critical family for infinite conjectures an Deligne’s including with r diagrams, off-sh Feynman quadratic and on-shell Broadhurst–Roberts on-shell for by generated for ideal rules the sum From linear via diffe Vanhove’s pairing in section coefficients polynomial through pairing A BSTRACT SINAGBAADBODUS–OET UDAI RELATION QUADRATIC BROADHURST–ROBERTS AND ALGEBRA NSKIAN ⋆ ´ algebra eslmmns ena nerl,Wro integrals, Feynman moments, Bessel : algebra sinmtie n ahv operators Vanhove and matrices nskian ´ hog leri aiuain nWro on manipulations algebraic Through . nskians ´ 16,3C0 43 Piay 11,8T0(Secondary) 81T40 81T18, (Primary) 34M35 33C10, 11B68, : sincofactors nskian ´ otemmr fD.W Dr. of memory the To AU ZHOU YAJUN sinfaeok eritrrtted hmintersection Rham de the reinterpret we framework, nskian C ´ sinmtie,Brolinumbers Bernoulli matrices, nskian ´ sinmtie 8 matrices nskian ´ ONTENTS i sinmtie hs nre r euil oBessel to reducible are entries whose matrices nskian ´ eClaoaoyo ahmtc o eocpcModeling Mesoscopic for Mathematics on Collaboratory he L c rga ihnteDprmn fEeg DE Office (DOE) Energy of Department the within Program ics -functions. (1986–2020) eta prtr,adcmueteBtiinter- Betti the compute and operators, rential l ena igasa hehl momenta. threshold at diagrams Feynman ell toscnetrdb rahrtadRoberts, and Broadhurst by conjectured ations ltos edrv e o-iersmrules sum non-linear new derive we elations, fdtriatiette htaecompatible are that identities determinant of S 24 53 49 43 37 37 30 21 20 1 5 5 BROADHURST–ROBERTSQUADRATICRELATIONS 1

1. INTRODUCTION In perturbative quantum ﬁeld theory, one frequently encounters Feynman diagrams in the form of on-shell Bessel moments [33, 29, 3, 32, 34, 9, 12, 10, 13, 14, 25]

∞ a b n IKM(a,b;n) := [I0(t)] [K0(t)] t dt (1.1) Z0

= 1 π t cosθ = t coshu for certain non-negative integers a,b,n Z 0, where I0(t) π 0 e dθ and K0(t) 0∞ e− du are modiﬁed Bessel functions of zeroth order.∈ These≥ Bessel moments satisfy a wealth of algebraic relations, R R which had been discovered by Broadhurst and collaborators through numerical experimentations [9, 12, 10, 13, 14], before becoming formally proven truths [50, 48, 49, 53, 54, 25]. The Broadhurst–Mellit determinant formulae [9, Conjectures 4 and 7] were among the earliest veriﬁed [49] non-linear algebraic relations for Bessel moments: the k k matrices Mk := (IKM(a,2k +1 a;2b × − − 1))1 a,b k and Nk := (IKM(a,2k + 2 a;2b 1))1 a,b k have determinants ≤ ≤ − − ≤ ≤

k k j j (k+1)2/2 k+1 k+1 j (2 j) − π 2π (2 j 1) − detMk = , detNk = − , (1.2) ∏ 2 j+1 Γ + ∏ j j=1 (2 j + 1) ((k 1)/2) j=1 (2 j) p Γ = x 1 t where (x) : 0∞ t − e− dt for x > 0. = T These determinant formulae are refined by the Broadhurst–Roberts quadratic relations Bm PmDmPm R + b 1 a m 3 [10, 13, 14] for the period matrix Pm := (( 1) − π − 2 IKM(a,m + 2 a;2b 1)) m+1 , in which 1 a,b 2 m+1 m+1 − m−+1 m+−1 ≤ ≤ the Betti matrix Bm Q 2 × 2 and the de Rham matrix Dm Q 2 × 2 are filled with rational numbers. Fresán, Sabbah∈ and Yu have recently proved an equiv∈alent form [25, Theorem 1.4] of the Broadhurst–Roberts quadratic relations, with an algebro-geometric tour de force that draws on their profound insights into certain exponential motives [22], whose associated Galois representations are intimately related to moments of Kloosterman sums [23, 24] (analogs of on-shell Bessel moments in positive characterisitic). In this paper, we present an analytic proof of the Broadhurst–Roberts quadratic relations, by revisiting the Wronskian´ method in our verification of the Broadhurst–Mellit determinant formulae [49]. Our Wronskian-based´ proof also produces a new family of quadratic relations, which embody the Broadhurst–Roberts relations for on-shell Bessel moments as special cases. n n n 0 In §2, writing D f (u) := d f (u)/du for n Z>0, and D f (u) := f (u), we will consider a Wronskian´ i 1 ∈ matrix W[ f1(u),..., fm(u)] = (D − f j(u))1 i, j m whose entries are reducible to on-shell Bessel moments ≤ ≤ (more precisely, entries of Pm and Pm 2) as u 1−, and whose determinant detW := W[ f1(u),..., fm(u)] = i 1 − → det(D − f j(u))1 i, j m is explicitly computable through our proof of the Broadhurst–Mellit determinant ≤ ≤ ⋆ ⋆ formulae [49]. In §3, our key observation is the adjugate relation W W = WW = WIm, where the ⋆ ⋆ T m m matrix adjugate W := adjW of W[ f1(u),..., fm(u)] factorizes into W = SW U, with S Q × and m m T ∈ U/W Q(u) × . In §4, we explore the u 1− limit of the matrix equation SW UW = WIm, and establish∈ quadratic relations among on-shell Bessel→ moments. For a generic parameter u (0,1), the individual elements of the Wronskian´ matrix W[ f1(u),..., fm(u)] in §2 will be representable by∈ off-shell Bessel moments [49, Definition 2.1]

∞ a b n IKM(a + 1,b;n u) := I0(√ut)[I0(t)] [K0(t)] t dt, (1.3) | Z0 ∞ a b n eIKM(a,b + 1;n u) := K0(√ut)[I0(t)] [K0(t)] t dt, (1.4) | 0 Z e 2 YAJUN ZHOU and differentiated Bessel moments [49, Deﬁnition 2.3]

∞ a b n+1 IKM´ (a + 1,b;n u) := + I1(√ut)[I0(t)] [K0(t)] t dt, (1.5) | Z0 ∞ a b n+1 IKM´ (a,b + 1;n u) := K1(√ut)[I0(t)] [K0(t)] t dt, (1.6) | − Z0 where I1(x) = +d I0(x)/d x and K1(x) = d K0(x)/d x. Accordingly, our efforts in §3 will culminate in a quadratic relation involving these generalized− Bessel moments and some explicitly computable rational functions (in lieu of the rational numbers in the Betti matrix B and the de Rham matrix D of the Broadhurst–Roberts formulation), as stated below. 5 3( 1)m Theorem 1.1 (Wronskian´ algebra). For m Z>0,u 0, − 2− , normalize off-shell Feynman diagrams as ∈ ∈ IKM(1,m+1;1 u)+(m+1)IKM(1,m+1;1 u) = (|m+2)π(m+1)/2 | , j 1, (u) := eIKM( j,m+2 j;1 u) e m + (1.7) m, j  (m+1)/2+−1 j | , j Z 2, 2 1 , F π − ∈ ∩ IKM( j m ,m+2+ m j;1 u) e − 2 2 − | m + (m⌊+1)⌋/2+ m/2⌊+1 ⌋j , j Z 2 2,m , π ⌊ ⌋ − ∈ ∩  e and define a polynomial  + m 1 2 m(u) := u 2 ∏ (u n ), (1.8) L n Z [1,m+1] − n ∈m+∩1 (mod 2) ≡ 5 3( 1)m where x is the greatest integer less than or equal to x. For u 0, − − , the Wrońskian matrix ⌊ ⌋ ∈ 2 Wm(u) := W[ m,1(u),..., m,m(u)] has determinant F F m + 1 ( 1) m/4 [(m + 1)!]m 1 := = ⌊ ⌋ Wm(u) detWm(u) −m(m 1)/2 m+1 m/2 (1.9) m + 2 2 ∏ nn m(u) − n=1 |L | and satisfies a quadratic relation 1 T = [Vm(u)]− Wm(u)SmWm(u) (1.10) m(u) |L | for matrices V (u) = ( 1)m+1VT (u) Q(u)m m and S = ( 1)m+1ST Qm m. m − m ∈ × m − m ∈ × Explicitly, the Vanhove matrix V (u) Q(u)m m is upper-left triangular, with entries1 m ∈ × m n a b+1 = a+n+m+1 n a D − − ℓm,n(u) (Vm(u))a,b : ∑ ( 1) − , (1.11) n=a+b 1 − b 1 m(u) − − L 2 where the polynomials ℓm, j(u) Z[u] are defined through the Vanhove–Verrill relation (cf. [43, §3.6.1] and [44, §3.6.1]): ∈

m m +1 k 2 + j = m + 1 k m 1 α1 + αn αn+1 ∑ ℓm, j(u)D : uϑ ∑ u − ∑ (ϑ k) − ∏ αn(αn m 2)(ϑ k n) − , − − − − j=0 k=1 n Z [1,k],αk+1=1 n=1 ∈αn∩Z [1,m+1] b α∈n+1∩ αn 2 b b ≤ − (1.12)

1 n n! Throughout this article, we write := for the standard binomial coefficient with n Z 0,m Z [0,n]. By m m!(n m)! ∈ ≥ ∈ ∩ n = − B convention, we also define 0 when n Z 0,k Zr [0,n]. Empty sums like ∑n=A( ) with A > B are treated as zero. k ∈ ≥ ∈ ··· 2These polynomials can be constructed by a finite enumeration over integer parameters α ,...,α in (1.12). See Table I 1 n for a partial list of the Vanhove–Verrill polynomials. BROADHURST–ROBERTSQUADRATICRELATIONS 3

= m j TABLE I. A partial list of Vanhove’s operators Lm ∑ j=0 ℓm, j(u)D

m Lm e 1 (u 4)uD1 + (u 2)D0 − − 2 (eu 9)(u 1)uD2 + (3u2 20u + 9)D1 + (u 3)D0 − − − − 3 (u 16)(u 4)u2 D3 + 6u(u2 15u + 32)D2 + (7u2 68u + 64)D1 + (u 4)D0 − − − − − 4 (u 25)(u 9)(u 1)u2 D4 + 2u(5u3 140u2 + 777u 450)D3 + (25u3 518u2 + 1839u 450)D2 + (3u 5)(5u 57)D1 + (u 5)D0 − − − − − − − − − − 5 (u 36)(u 16)(u 4)u3 D5 + 5u2(3u3 140u2 + 1568u 3456)D4 + u(65u3 2408u2 + 19836u 27648)D3 − − − − − − − +6(15u3 406u2 + 2078u 1152)D2 + (31u2 516u + 1020)D1 + (u 6)D0 − − − − 6 (u 49)(u 25)(u 9)(u 1)u3 D6 + 3u2(7u4 504u3 + 9870u2 51664u + 33075)D5 − − − − − − +2u(70u4 4179u3 + 64749u2 247993u + 99225)D4 + 2(175u4 8232u3 + 92394u2 217012u + 33075)D3 − − − − +(301u3 10218u2 + 69561u 62020)D2 + 3(21u2 428u + 1071)D1 + (u 7)D0 − − − − 7 (u 64)(u 36)(u 16)(u 4)u4 D7 + 28u3(u4 105u3 + 3276u2 32800u + 73728)D6 − − − − − − +2u2(133u4 11928u3 + 307950u2 2432512u + 3981312)D5 + 30u(35u4 2590u3 + 52420u2 297856u + 294912)D4 − − − − +3(567u4 32820u3 + 474944u2 1616896u + 589824)D3 + 6(161u3 6645u2 + 57096u 70912)D2 − − − − +(127u2 3076u + 9312)D1 + (u 8)D0 − − 8 (u 81)(u 49)(u 25)(u 9)(u 1)u4 D8 + 12u3(3u5 440u4 + 20482u3 345620u2 + 1762035u 1190700)D7 − − − − − − − − +6u2(77u5 9856u4 + 391358u3 5455164u2 + 21884733u 10716300)D6 − − − +6u(441u5 48048u4 + 1569150u3 17075168u2 + 48951657u 14288400)D5 − − − +3(2317u5 207295u4 + 5264259u3 40651797u2 + 68940108u 7144200)D4 − − − +6(1295u4 89980u3 + 1614657u2 7325514u + 4441878)D3 + (3025u3 148126u2 + 1547883u 2474334)D2 − − − − +(255u2 7172u + 25461)D1 + (u 9)D0 − −

A B = 1 m m = Sm Sm 3 for (ϑ f )(u) D [uf (u)]; the matrix Sm Q × can be partitioned into Sm C D for ∈ Sm Sm m +2 a 2 − + m + b + + + + ⌊ ⌋ s m 2 a 2 1 s a+b+m+1 [1 (m 1)δa,1][1 (m 1)δb,1] ∑ ( 1) m+1 −+s ⌊ ⌋ − 1 + ( 1) = − 2 b 1 (SA ) := − s 1 − (1.13) m a,b 1 m a b 1+( 1)m m 2 − + − 1 ( 1) 2 2 − 4 −⌊ 2 ⌋− (a 1)!(m + 2 a)! − j k − − where a,b Z 1, m + 1 , ∈ ∩ 2 B m+1 C (S )a,b = (1) (S )b ,a m ′ − m ′ m +2 a 2 − + m + s m+1 +s + + ⌊ ⌋ s m 2 a 2 1 + b′ 2 a+b +m [1 (m 1)δa,1] ∑ ( 1) m+1 −+s ⌊ ⌋+ − ( 1) + 1 + ( 1) ′ = − 2 b′ 1 − b′ 1 := − s 1 (1.14) 1 m a 1+( 1)m b 1+(1)m h m+1 i 2 (a 1)m+ − + ′ + − − ( 1) − 2 − 4 2 4 − 2 (a 1)!(m + 2 a)! − j k j k − − m m 1 where a Z 1, + 1 ,b Z 1, − , and ∈ ∩ 2 ′ ∈ ∩ 2 + m+1 + + + m 1 2 1 + a′ + b′ m 1 m 1 D 1 ( 1) ( 4) − [1 ( 1) ][1 ( 1) ] 2 2 (S )a ,b := − − − − (1.15) m ′ ′ + 2 a′ + b′ + + 2 m 1 ! ( 1) 2 2 a′ 1 b′ 1 2 − m 1 where a′,b′ Z 1, − . ∈ ∩ 2 In §4, we push the theorem above to the u 1 limit, in the following form. → − Theorem 1.2 (Broadhurst–Roberts quadratic relations). Define the Bessel–Feynman moments ℓ (u) by Fm, j replacing all the occurrences of IKM( , ;1 u) [resp. IKM( , ;1 u)] in (1.7) with IKM( , ;2ℓ 1 u) [resp. · · | · · | · · − | IKM( , ;2ℓ 1 u)]. For each m Z>1, the Broadhurst–Roberts period matrix Pm := b· 1· b − | + ∈ (( 1) − m,a(1))1 a,b m 1 satisfiese a quadratic relatione e − F 2 e ≤ ≤ T = PmDmPm Bm, (1.16) + + + + = m+1 T m 1 m 1 = m+1 T m 1 m 1 for de Rham matrix Dm ( 1) Dm Q 2 × 2 and Betti matrix Bm ( 1) Bm Q 2 × 2 . − ∈ − ∈ 3 The Kronecker delta is defined as δm,n = 1 when m = n and δm,n = 0 when m = n. 6 4 YAJUN ZHOU

Explicitly, the de Rham matrix Dm 2 is upper-left triangular, with entries − 1 T 1 m(u) ((βm− ) Vm(u)βm− )a+ m 1 +1,b+ m 1 +1 = |L | 2− 2− (Dm 2)a,b : lim m 1 (1.17) − u 1− 4(m + 2)( 1) 2− → − m m computable via the Vanhove matrix Vm(u) [see (1.11)] and the Bessel matrix βm Z for ∈ × a 1 (a 1)! a 1 m+1 ( 4) − (b−a)! b−a , a Z 1, 2 , = − − − + + ∈ ∩ (βm)a,b : m+1 m 1 m 1 (1.18) a 1 2 a 2 1 ! a 2 m+1 ( 4) 2 − + − − + , a Z +1,m ; − − b a+ m 1 1 ! b a+ m 1 2  − − 2 − − 2 ∈ ∩ the Betti matrix Bm2 satisfies − 1 1 1 (Sm− )a+ m +1,b+ m +1 (Sm− )a+1,b+ m +1 (Sm− )a+ m +1,b+1 = 2 2 − 2 − 2 (Bm 2)a,b : ⌊ ⌋ ⌊ ⌋ ⌊m 1⌋ ⌊ ⌋ − 4(m + 2)( 1) 2− a+1 − ( 1) (m a)!(m b)! Bm+1 a b = − − , (1.19) − m+1 −+ − m a b 2 (m 1 a b)! ( 1) −2− − − − A B = Sm Sm = m+1 T where Sm C D ( 1) Sm is defined through (1.13)–(1.15), and the Bernoulli numbers Bn,n Sm Sm − ∈ Z 0 are generated by ≥ n t ∞ t := ∑ Bn , t < π. (1.20) et 1 n! | | − n=0 We note that Lee and Pomeransky [35, Appendix C] have recently produced some quadratic relations for off-shell Bessel moments, but not for differentiated Bessel moments. There are technical difficulties when one attempts to push their results to the on-shell limit [35, §6.4] and recover Broadhurst–Roberts quadratic relations therefrom. For each m Z>1, our Betti matrix Bm agrees with the original formulation of Broadhurst–Roberts [13, (5.4)–(5.6)],∈ up to reordering of rows and columns. Instead of considering Betti intersection pairing that inspired the conjecture of Broadhurst–Roberts [10, 13, 14] and the proof by Fresán–Sabbah–Yu [25, Theorem 1.4], we generate Bm by inverting Sm+2. The latter matrix Sm+2 draws on explicit sum rules (with rational coefficients) for on-shell and off-shell Bessel moments, which extend our earlier works [50, 53, 52] on the same topic. In addition to elucidating the relations between Betti matrices and sum rules for individual Bessel moments in §4.1, we will also establish a family of sum rules for minor determinants of Mk,k Z>1 in §4.2 (see Proposition 4.5 for details): ∈ + det(IKM(2a,2(k a) 1;2b 1))1 a,b k − − ≤ ≤ 2 + 2 ( 1)k 1 [(2k 1)!!] − − = + + k+1 k det(IKM(2a 1,2(k a 1);2b 1))1 a,b k 1 . (1.21) 1 ( 1)k − − − 2 π 2 2 p2 (k 1)!!(k!!) − − ≤ ≤ − Such sum rules may have some arithmetic interest: Deligne’s conjecture [19] for the motive M2k+1 (as Z defined in [25, §7.c]) predicts that the determinants on both sides of (1.21) are Q×π multiples of critical L-values, while these critical values for L(M2k+1, s) are related to each other by a functional equation [23, Theorem 1.2]. No such analogs appear to exist in minor determinants of Nk,k Z>1, for both algebraic and analytic reasons. ∈ Our algorithmic construction of the de Rham matrix Dm is based on derivatives of the polynomials ℓm+2, j(u) at u = 1, which are readily adaptable to systems different from Bessel moments (see §3.1). Within the same Wronskian´ framework, one can establish several other equivalent representations [see BROADHURST–ROBERTSQUADRATICRELATIONS 5

(4.76)–(4.79) for details] of Dm, along with new algebraic relations [see (4.80) for details] connecting the period matrix Pm to b 1 a = ( 1) − ∞ [I0(t)] logt δa,1 m+2 2b 1 Pm : −m+3 a [K0(t)] t − dt . (1.22) a [K (t)] m + 2 m+1 ◦ π 2 − 0 0 − 1 a,b Z ≤ ≤ 2 Outside our Wronskian´ framework, there are still other approaches to de Rham representations, such as the Broadhurst–Roberts combinatorial recursion formulae [13, (5.7)–(5.12)] and the Fresán–Sabbah–Yu self-duality pairing [25, §3.b]. A direct proof of the equivalence between Wronskian´ and non-Wronskian´ approaches to de Rham matrices will not appear in this article. Nevertheless, we will outline general principles that may reconcile our representation of de Rham matrices with those of Broadhurst–Roberts and Fresán–Sabbah–Yu, in §4.3. Acknowledgments The author thanks David Broadhurst and Javier Fresán for their feedback on the ﬁrst draft. The author is grateful to Pierre Vanhove for his invitation to Séminaire Motifs et intégrales de Feynman. The author is indebted to two anonymous referees for their suggestions on improving the presentation of this manuscript.

2. W ALGEBRA In §2.1, we open by recapitulating our previous work [49] on the Broadhurst–Mellit determinant formulae, with an extended discussion on the ordinary differential equations satisfied by off-shell Bessel moments m, j(u). In §2.2,F we elucidate the relation between Wronskian´ matrices and on-shell Bessel moments. We will achieve this by studying the Wronskian´ matrices Wm(u) := W[ m,1(u),..., m,m(u)] for m Z>0 in the F F ∈ u 1− limit, similar to procedures reported in [49]. →In §2.3, we establish, among other things, several families of Q-linear dependence relations concerning on-shell and off-shell Bessel moments, when the parameter u is a perfect square. This sets the stage ⋆ for the analysis of the adjugate matrices W (u) for m Z>0, which will unfold in §3. m ∈ 2.1. Wronskian´ matrices and Vanhove operators. For m Z>0, consider the functions m, j(u), j Z [1,m] introduced in (1.7), which are normalized versions∈ of the notations in [49, DefinitionF 4.1]. In∈ ∩ our previous work [49, §4], we have manipulated Wronskian´ matrices Wm(u) := W[ m,1(u),..., m,m(u)] F F of all sizes, and have shown that their determinants Wm(u) := detWm(u) [see (1.9)] satisfy (cf. [49, (4.36) and (4.62)]): (k 1)(k 2) k 1 − − Γ 2 2 k 2 2 = ( 1) 2 k[ (k/2)] (detNk 1) (2 j) − W2k 1(u) − 2 − + ∏ , u (0,4), (2.1) − πk uk(2k 1)/2(2k + 1) 2(k 1)(2k 1) 1 (2 j)2 u ∀ ∈ − − − j=1 − k(k 1) − 2 k+1 2 k ( 1) 2 (2k + 1)(detMk) (2 j 1) W2k(u) = − ∏ − , u (0,1), (2.2) πk(k+1) 2(2k 1)k+1uk2 (k + 1) (2 j 1)2 u ∀ ∈ − j=1 − − for each k Z 2. In particular, these collateral results for the Broadhurst–Mellit determinant formulae ∈ ≥ 5 3( 1)m (1.2) tell us that the Wronskian´ matrix Wm(u) is non-singular for u 0, − − , hence the linear ∈ 2 independence of the functions m,1(u),..., m,m(u) on the same interval. Furthermore, we may deduceF from [49,F Lemma 4.2] that the aforementioned linearly independent 4 functions form complete sets of fundamental solutions to certain ordinary differential equations: C∞ 0, m m 5 3( 1) ker L = span (u),u 0, 5 3( 1) j Z [1,m] , where Vanhove’s operator L as- − 2− m C m, j − 2− m sumes the∩ following form (omittingF lower∈ order terms) [42,∈ ∩ (9.11)–(9.12)]: e m m d m(u) m 1 e Lm = m(u)D + L D − + , (2.3) L 2 du ··· 4 + Using the same techniques of integratione by parts as in [49, Lemma 4.2], one can show that Lm annihilates IKM(a,m 2 a;1 u) for a Z [2,m + 1], when u belongs to a certain open interval that ensures the convergence of the respective integral.− | ∈ ∩ e e 6 YAJUN ZHOU with the polynomial m(u) defined in (1.8). The leading and sub-leading terms in (2.3) explain why L m/2 Wm(u) is a constant multipleof m(u) − [see (1.9)], according to standard manipulations of Wronskian´ determinants [31, §5.2, p. 119].|L | m j In general, Vanhove’s operator Lm,m Z>0 takes the form Lm = ∑ = ℓm, j(u)D , where ℓm, j(u) Z[u]. ∈ j 0 ∈ These polynomials ℓm, j(u) can be deduced from the closed-form representations in the next proposition, which in turn is amplified and extrapolatede from [43, §3.6.1] ande [44, §3.6.1].

Proposition 2.1 (Vanhove’s operators in explicit form). For each m Z>0, we normalize Verrill’s poly- m ∈ nomials as Vm,0(t) = t and

k αn 1 m t n − Vm,k(t) := t ∑ ∏ αn(αn m 2) − − − t n + 1 n Z [1,k] n=1 − αn ∈Z ∩[1,m+1] α∈ +∩ αn 2 n 1≤ − k m+1 α1 αn αn+1 = ∑ t − ∏ αn(αn m 2)(t n) − Z[t] (2.4) − − − ∈ n Z [1,k],αk+1=1 n=1 ∈αn∩Z [1,m+1] α∈ +∩ αn 2 n 1≤ − for k Z>0, with empty sums being 0 and empty products being 1. ∈ If we deﬁne (ϑ f )(u) = D1[uf (u)] and write x for the greatest integer less than or equal to x, then we have the following explicit representation for Vanhove’s⌊ ⌋ differential operator of m-th order: b m + 2 1 m 1 k Lm = ( 1) ∑ u − Vm,k(k ϑ). (2.5) − k=0 −

Proof. Using the arguments in [49,e Lemma 4.2], one can show thatb Vanhove’s operator Lm annihilates ∞ m+1 2 IKM(m + 1,1;1 u) := K0(√ut)[I0(t)] t dt, u ((m + 1) , ). (2.6) | ∈ ∞ e Z0 Thanks to a convergent Taylor series [7, (2.4)] e k 2 2 m+1 ∞ t k! = + + = [I0(t)] ∑ Wm 1(2k) k , where Wm 1(2k) ∑ , (2.7) = 2 k! a1!a2! am+1! k 0 a1,a2,...,am+1 Z 0 ··· a +a + +a ∈+ =≥k 1 2 ··· m 1 we may invoke the dominated convergence theorem, and integrate termwise, as done by Glasser–Montaldi [28, (A3)]: k 2 ∞ ∞ t ∞ Wm+1(2k) IKM(m + 1,1;1 u) = K0(√ut) ∑ Wm+1(2k) t dt = ∑ . (2.8) | 2kk! uk+1 Z0 "k=0 # k=0 e = m j We write the right-hand side of (2.5) as Rm ∑ j=0 rm, j(u)D , with rm, j(u) Q(u) being rational func- m +1 ∈ 2 V = tions in u. By virtue of Verrill’s recursion [45, Theorem 1] ∑k⌊=0⌋ m,k(n)Wm+1(2(n k)) 0, we know e 2 − that Rm annihilates IKM(m + 1,1;1 u),u ((m + 1) , ), so there exists a rational function ρ(u) Q(u) | ∈ ∞ m ∈ such that ρ(u)Lm = Rm. Moreover, we recover Vanhove’s leading term m(u)D in (2.3) from L m e e +1 k + m+1 2 m + m+1 = em 1 +e 2 2 1 k = 2 2 rm,m(u) u u ∑ u − ∑ ∏ αn(αn m 2) u ∏ (u n ) k=1 n Z [1,k] n=1 − − n Z [1,m+1] − αn ∈Z ∩[1,m+1] n ∈m+∩1 (mod 2) α∈n+1∩ αn 2 ≡ ≤ − (2.9) where the last combinatorial identity owes to Djakov–Mityagin (see [20, Theorem 4.1] or [21, Theorem 8]) and Zagier [7, Theorem 2.7]. Hence ρ(u) = 1 and m(u) = ℓm,m(u). L BROADHURST–ROBERTSQUADRATICRELATIONS 7 = 1 Remark For an equivalent formulation of Vanhove’s operator in the variable s u and its proof, see the recent work of Bönisch–Fischbach–Klemm–Nega–Safari [17, §3.1]. = m j By Vanhove’s original construction of Lm ∑ j=0 ℓm, j(u)D in [42, §9], we know that ℓm, j(u) is always a polynomial in u with integer coefﬁcients. In the next corollary, we will use the explicit formula (2.5) to show that such polynomials have certaine divisibility properties, thereby paving the way for the asymptotic analysis (to be elaborated in Proposition 2.10) of solutions to Vanhove’s differential equations in the u 0+ regime. → m j m+1 Corollary 2.2 (Regular singular point). For all m Z>0, we have u 2 ℓm, j(u) Z[u] for j Z ∈ − − ∈ ∈ ∩ [0,m]. As a consequence, for each m Z>0, the point u = 0 is a regular singular point for Vanhove’s ∈ differential equation Lm f (u) = 0, so that for a suitably chosen δm (0,1), every solution f (u),u (0,δm) is expressible as a convergent series in (fractional) powersofuand∈ logu. ∈

e m +1 j = m 2 j+1 kV Proof. By (2.5), we have Lmu ( 1) ∑k=0 u − m,k(k j 1) for all j R. Hence, we obtain + m+1 j m 1 j − − − ∈ limu 0+ u − − 2 Lmu = 0. m+1 m+1 → = m e j m+1 j j = m+1 j j n j From Lm ∑ j=0 ℓm, j(u)D , we have u − − 2 Lmu u − − 2 ∑n=0 ℓm,n(u)D u for j Z + ∈ ∩ e m+1 j m 1 [0,m]. In view of the output from the last paragraph, we can deduce limu 0+ u − − 2 ℓm, j(u) = 0 for all j eZ [0,m], beginning with the j = 0 case, ande proceeding with ﬁnitely→ many recursions. This ∈ ∩m j m+1 shows that u − − 2 ℓm, j(u) Z[u] for j Z [0,m], as claimed. m j ∈ ∈ ∩ Now that u ℓm, j(u)/ℓm,m(u) is holomorphic in a non-void C-neighborhood of u = 0 for every j − ∈ Z [0,m 1], we know that the differential equation Lm f (u) = 0,u (0,1) has a regular singular point at u =∩0, according− to the Fuchs condition [31, §15.3]. The existence of∈ generalized power series expansion then follows immediately. e n k For any differential operator O = ∑ = ak(u)D with ak C (a,b),k Z [0,n], its formal adjoint k 0 ∈ ∞ ∈ ∩ (or “formal dual”) O∗ acts on every smooth function f C∞(a,b) in the following manner: O∗ f (u) = n k k ∈ ∑ = ( 1) D [ak(u) f (u)]. This formalb adjoint is the unique operator satisfying Of,g a,b = f,O g a,b k 0 − h i( ) h ∗ i( ) for every smooth functionb f C∞(a,b) and every compactly supported smooth function g Cbc∞(a,b), = b ∈ ∈ where f,g (a,b) a f (u)g(u)du. b b We studyh i the formal adjoint of each Vanhove operator in the next proposition, which will become R useful later in Proposition 3.1.

m Proposition 2.3 (Parity of Vanhove operators). For every m Z>1, we have L = ( 1) Lm. ∈ m∗ − Proof. We compute each summand in e e m + 2 1 = m 1 kV Lm∗ ( 1) ∑ [u − m,k(k ϑ)]∗ (2.5∗) − k=0 − separately. e b m m m For k = 0, we have [uVm,0( ϑ)]∗ = ( 1) (uϑ )∗ = uϑ through integration by parts. − − 1 r For k Z>0, we operate on power functions u ,r R: ∈ − − ∈ b b b 1 ∏k α (α m 2) k 1 k n=1 n n α j α j+1 u − Vm,k(k ϑ) = ∑ − − ∏(r + k j) − , (2.10) − ur+1 ur+k − n Z [1,k],α0=m+1,αk+1=1 j=0 α∈n Z∩ [1,m+1],α + αn 2 b ∈ ∩ n 1≤ − 1 ∏k β (β m 2) k m 1 k n=1 n n β j β j+1 ( 1) [u − Vm,k(k ϑ)]∗ = ∑ − − ∏(r + j) − . (2.11) − − ur+1 ur+k n Z [1,k],β0=m+1,βk+1=1 j=0 β∈n Z∩ [1,m+1],β + βn 2 b ∈ ∩ n 1≤ − 8 YAJUN ZHOU

One can identify the left-hand sides of the last two equations, by pairing up the dummy indices βn = m + 2 αk+1 n for n Z [1,k]. − − ∈ ∩ 1 k Since the parameter r in the last paragraph is an arbitrary real number, we have [u − Vm,k(k ϑ)]∗ = m 1 kV = m − ( 1) u − m,k(k ϑ) for k Z 0, finally leading us to the claimed parity relation Lm∗ ( 1) Lm. − − ∈ ≥ − b An immediate consequence of the last proposition is a combinatorial constraint on the coefficients b m j e e ℓm, j(u) in Vanhove’s operator Lm = ∑ = ℓm, j(u)D for each m Z>1. j 0 ∈ Corollary 2.4 (Combinatorial constraint on ℓm, j(u)). For each m Z>1, we have e ∈ m m n n n j ( 1) ℓm, j(u) = ∑( 1) D − ℓm,n(u) (2.12) − − j n= j for j Z [0,m], which incorporates Vanhove’s sub-leading term [cf. (2.3)] ∈ ∩ m 1 ℓm,m 1(u) = D ℓm,m(u) (2.13) − 2 as a special case. Proof. With the product rule for differentiation, we compute m m n m n n n n n j j ( 1) Lm f (u) = L∗ f (u) = ∑( 1) D [ℓm,n(u) f (u)] = ∑( 1) ∑ [D − ℓm,n(u)]D f (u), (2.14) − m − − j n=0 n=0 j=0 whose coefficientse fore D j f (u) boil down to (2.12). m m 1 m 1 For j = m 1, the constraint (2.12) reduces to ( 1) ℓm,m 1(u) = ( 1) mD ℓm,m(u)+( 1) − ℓm,m 1(u), which is equivalent− to (2.13), a vital result for the− computa−tions of− Wronskian´ determinants.− − Remark We refer our readers to Corollary 3.3 in §3.1, for non-trivial applications of the generic recursion (2.12). During our proof of Proposition 2.1, we have considered solutions to Vanhove’s differential equations Lm f (u) = 0,u (m + 1, ). Later on in §§3.2–3.3, we will also find it helpful to solve Lm f (u) = 0,u 1+(∈ 1)m 2 ∞ m ∈ R r 2 j − j Z 1, + 1 for each m Z>0. To solve Vanhove’s differential equations − 2 ∈ ∩ 2 ∈ ine such scenarios, we need to extend our previous notations for the Bessel–Feynman momentse ℓ (u) Fm, j ℓ j (m+1)/2 m 1 m m m m to (u) = π 2 IKM j ,m + 2 + j;2ℓ 1 u for j Z + 2,m + + Fm, j − −⌊ ⌋− − 2 2 − − ∈ ∩ 2 2 1 . In addition, it is also appropriate to enlist the differentiated Bessel moments in (1.5)–(1.6), and define ´ ℓ (u) by updating eache occurrence of IKM( , ;2ℓ 1 u) [resp. IKM( , ;2ℓ 1 u)] in ℓ (u) Fm, j · · − | · · − | Fm, j with IKM´ ( , ;2ℓ 1 u) [resp. IKM´ ( , ;2ℓ 1 u)]. · · − | · · − | e e 2.2. Block diagonalization of on-shell Wronskian´ matrices. Before moving on, we need to reduce the higher-order derivatives in our formulation of Wronskian´ matrices into off-shell Bessel moments (1.3)–(1.4) and differentiated Bessel moments (1.5)–(1.6). As in [49, §4.1], we apply the Bessel differential operator iteratively, to derive ℓ (u) Fm, j = 2 + 1 ℓ 1 ℓ 1 (uD D ) − m, j(u), (2.15) 4 − F ´ ℓ (u) Fm, j = D1 uD2 + D1 ℓ 1 u . ℓ 1 2 ( ) − m, j( ) (2.16) 4 − √u F 2 1 ℓ 1 1 2 1 ℓ 1 In the next lemma, we will spell out the differential operators (uD + D ) − and D (uD + D ) − as linear differential operators with polynomial coefficients. BROADHURST–ROBERTSQUADRATICRELATIONS 9

Lemma 2.5 (Iterations of Bessel differential operators). For n Z 0, we have ∈ ≥ n n! n + (uD2 + D1)n = ∑ umDn m, (2.17) m! m m=0 n n! n + 1 + + D1(uD2 + D1)n = ∑ umDn m 1. (2.18) m! m + 1 m=0 Proof. It is clear that (2.17) is valid for n = 0. Suppose that it holds for n = N Z 0, then ∈ ≥ N + N! N + (uD2 + D1)N 1 f (u) = (uD2 + D1) ∑ umDN m f (u) m! m m=0 N N! N m 1 2 0 1 2 2 N+m = ∑ u − [m D + (2m + 1)uD + u D ]D f (u) m! m m=0 N + N! N N! N = uN 1D2N f (u) + ∑ (m + 1)2 + (2m + 1) (m + 1)! m + 1 m! m m=0 N! N + + + umDN 1 m f (u). (2.19) (m 1)! m 1 − − (N+1)! N+1 = + Since the last bracketed expression in equal to m! m , we know that (2.17) also holds for n N 1. Now that we have veriﬁed (2.17) by induction, we may differentiate its right-hand side termwise, to obtain n n! n + D1(uD2 + D1)n f (u) = D1 ∑ umDn m f (u) m! m m=0 n n! n m 1 0 1 n+m = ∑ u − (mD + uD )D f (u) m! m m=0 n n! n n! n + + = ∑ (m + 1) + umDn m 1 f (u). (2.20) (m + 1)! m + 1 m! m m=0 n! n+1 Here, the last bracketed expression can be identiﬁed with m! m+1 , hence our claim. Proposition 2.6 (Bessel moment representations of Wronskian´ matrices) . We have + m+1 m 1 1 2 1 2 1 2 ( (u),( 1) − (u),...,( 1) 2 − (u), Fm, j − Fm, j − Fm, j m ´ 1 2 1 ´ 2 k 2 ´ 2 T √u (u),( 1) − √u (u),...,( 1) − √u ⌊ ⌋(u)) Fm, j − Fm, j − Fm, j 0 m 1 T = βm(u)(D m, j(u),..., D − m, j(u)) , (2.21) F F m m where the (a,b)-th entries of the square matrix βm(u) Z[u] are given by ∈ × a 1 (a 1)! a 1 b a m+1 ( 4) − (b−a)! b−a u − , a Z 1, 2 , = − − − + + ∈ ∩ (βm(u))a,b m+1 m 1 m 1 m+1 (2.22) a 1 2 a 2 1 ! a 2 b a+ m+1 ( 4) 2 − + − − + u 2 , a Z +1,m . − − b a+ m 1 1 ! b a+ m 1 − 2  − − 2 − − 2 ∈ ∩ N = 1 = Here, it is understood that M 0 and ( N 1)! 0 if N Z 0 and M Z r [0, N]. − − ∈ ≥ ∈ Proof. This is essentially a transcription of the results from Lemma 2.5. [Note: we have abbreviated βm(1) as βm in (1.18), by slight abuse of notation.]

Remark It is not hard to show, through elementary manipulations of determinants, that detβm(u) = m(m 1)/2 m2/4 | | 2 − u . Therefore, the matrix βm(u) is invertible when u = 0. | | 6 10 YAJUN ZHOU

As we may recall from [49, §4], upon suitable column manipulations, some elements of the matrix βm(u)Wm(u) are expressible as on-shell Bessel moments when u 1−. Such reductions have led us to factorizations of Wronskian´ determinants into products of Broadhurst–Mellit→ determinants [49, §4]. In the next proposition, we transfer our earlier results onto the matrix βm(u)Wm(u) explicitly. (2k 1) (2k 1) Proposition 2.7 (Block tridiagonalization of Wronskian´ matrices). For k Z>0, deﬁne αk Z − × − as ∈ ∈

δa,b δa+k 1,b, a Z [2,k]; (αk)a,b := − − ∈ ∩ (2.23) δa,b, a Z ( 1 [k + 1,2k 1]), ( ∈ ∩ { } ∪ − 2k (2k 1) (2k 1) 2k and ψk Z ,̺k Z as ∈ × − ∈ − × δa,b(1 δa,k+1), a Z [1,k + 1]; (ψk)a,b := − ∈ ∩ (2.24) δa,b+1, a Z [k,2k], ( ∈ ∩ (̺k)a,b := δa,b, (2.25) 2k 2k (2k 1) (2k 1) so that the (k+1)-st column and the bottom row of X R × are dropped to form ̺kXψk R − × − . m+3 ∈ ∈ = b 1 a 2 + For period matrices Pm : (( 1) − π − IKM(a,m 2 a;2b 1))1 a,b m+1 of all sizes, we have − − − ≤ ≤ 2 PT = 2k 1 β2k 1(1)W2k 1(1)αk − (2.26) − − ´ T T P2k 1 P2k 3! − − − and T P2k ̺kβ2k(1) lim W2k(u)ψkαk = , (2.27) u 1 ´ T T → − P2k P2k 2! − − where the unwritten blocks in the top-right positions are filled with zeros, and P´ m := b 1 ´ b (( 1) − m,a(1))1 a m+1 ,1 b m 1 . − F ≤ ≤ 2 ≤ ≤ 2− Proof. Multiplying αk to the right of a (2k 1) (2k 1) matrix, we are effectively subtracting its 2nd to − × − ℓ ℓ = k-th columns from its last (k 1) columns. By definition, we have m, j(1) + m+1 (1) 0, j Z − F −Fm, j 2 1 ∀ ∈ ∩ m+1 − 2, 2 , accounting for the top-right blocks with vanishing entries in (2.26) and (2.27). Meanwhile, 1 ℓ ℓ the Wronskian´ relation I (t)K (t) + I (t)K (t) = leads us to an identity ´ (1) ´ + (1) = 1 0 0 1 t m, j m, j+ m 1 1 F − F 2 − ℓ (1), j Z 2, m+1 , accounting for the bottom-right blocks in (2.26) and (2.27). It is −Fm 2, j 1 ∀ ∈ ∩ 2 − − ´ T ´ T clear that the bottom-left blocks P2k 1 and P2k originate from the intact entries in β2k 1(1)W2k 1(1) and − − − ̺kβ2k(1)limu 1 W2k(u)ψk. → − The configurations in (2.26) and (2.27) were precisely the forms that we prepared [49, Propositions 4.3 and 4.6] for our proof of the Broadhurst–Mellit determinant formulae. To get ready for later devel- opments in §4.1 of the current article, we need to go one step further, eliminating the bottom-left blocks ´ T ´ T P2k 1 and P2k in the next proposition. − Proposition 2.8 (Block diagonalization of Wronskian´ matrices). Define square matrices ϑm, ϕm 2 m+1 1 2 m+1 1 ∈ Q 2 − × 2 − as follows: m+1 δa,b, a Z 1, 2 , = m+1 ∈ ∩ (ϑm)a,b : 2 a m+1 m+1 (2.28) + − 2 + + (δa,b m+2 δa m 1 ,b, a Z 2 1,2 2 1 ; − 2 ∈ ∩ − δ , a Z 1, m+1 , = a,b 2 (ϕm)a,b : b ∈ ∩ m+1 m+1 (2.29) + m+1 Z + (δa,b 1 m+2 δa +1,b, a 2 1,2 2 1 . − − 2 ∈ ∩ − BROADHURST–ROBERTSQUADRATICRELATIONS 11

For k Z>1, we have ∈ PT = 2k 1 ϑ2k 1β2k 1(1)W2k 1(1)αkϕ2k 1 − (2.30) − − − − T P2k 3! − − and

T P2k ϑ2k̺kβ2k(1) lim W2k(u)ψkαkϕ2k = . (2.31) u 1 T → − P2k 2! − −

Proof. Clearly, it would sufﬁce to check that the matrix elements of P´ m can be alternatively represented by

( 1)b 2b b (1), a = 1, − m+2 Fm,1 (P´ m)a,b := (2.32)  b 2b b a b m+1 ( 1) m+2 m,a(1) 1 m+2 m 2,a 1(1) , a Z 2, 2 ;  − F − − F − − ∈ ∩ h i Integrating by parts, we have

´ ℓ = 1 ∞ m 1 2ℓ m 1 ∞ m 2 2ℓ 1 m,1(1) : I1(t)[K0(t)] − t dt − K1(t)I0(t)[K0(t)] − t − dt F m 0 − m 0 Z Z 1 ∞ 2ℓ d m 1 2ℓ ℓ = t I0(t)[K0(t)] − dt = (1). (2.33) m dt{ } − m Fm,1 Z0 Likewise, for j Z 2, m+1 , we have ∈ ∩ 2 ℓ ∞ j 1 m j 2ℓ 1 ∞ m j 2ℓ d j ´ (1) := I1(t)[I0(t)] − [K0(t)] − t dt = [K0(t)] − t [I0(t)] dt Fm, j j dt Z0 Z0 2ℓ ℓ m j ∞ j m j 1 2ℓ = (1) + − K1(t)[I0(t)] [K0(t)] − − t dt − j Fm, j j Z0 = 2ℓ ℓ + m j ℓ ´ ℓ m, j(1) − m 2, j 1(1) m, j(1) . (2.34) − j F j F − − − F h i This can be rearranged into

j 2ℓ ´ ℓ (1) = 1 ℓ (1) ℓ (1), (2.35) Fm, j − m Fm 2, j 1 − m Fm, j − − which completes our task.

1+( 1)m 2 2.3. Sum rules for on-shell and off-shell Bessel moments. When u = 2 j − , j Z [1, m/2 + − 2 ∈ ∩ ⌊ ⌋ 1], certain functions among ℓ (u), ´ ℓ (u) for j Z [1,m + m/2 + 1] assume ﬁnite values, and are m, j m, j Q-linearly dependent. We referF to suchF Q-linear dependence∈ ∩ relations⌊ ⌋ (together with their modest generalizations) as sum rules for Bessel moments at thresholds. Here, we choose the word “threshold” to acknowledge the fact that certain functions among ℓ (u), ´ ℓ (u) for j Z [1,m + m/2 + 1] go Fm, j Fm, j ∈ ∩ ⌊ ⌋ 1+( 1)m 2 + + unbounded as u 2 j −2 0 , j Z [1, m/2 1]. (See Lemma 3.4 for quantitative characterization of such→ divergent− behavior± near∈ the threshold.)∩ ⌊ ⌋ 12 YAJUN ZHOU

These sum rules with rational coefﬁcients will be vital to the analysis of Wronskian´ cofactors in §3.2, which revolves around the identity

+ m 1 V m,1(u) ( 1) W m,1(u),..., m,r(u),..., m,m(u) F − F F . F . . .  .  Λ   + m m r V = S (u) . ( 1) W m,1(u),..., m,r(u),..., m,m(u) m 2 m  m,r  (2.36)  − F F F  2− F  .  m(u)  .   .  |L |  .      V    W m,1(u),..., m,r(u),..., m,m(u)  (u)  F F F   m,m    F  Here on the left-hand side, a caret indicates removal of the term underneath while evaluating the Wronskian´ determinant of (m 1) functions; on the right-hand side, we have [see (1.9)] − m + 1 ( 1) m/4 [(m + 1)!]m Λ := W (u) (u) m/2 = − ⌊ ⌋ (2.37) m m m + m(m 1)/2 m+1 n |L | m 2 2 − ∏n=1 n m m and Sm Q is a matrix filled with rational numbers. At the end of this section, we will illustrate the ∈ × usefulness of Q-linear sum rules with an explicit computation of S3 (Corollary 2.11), while deferring the treatment of generic Sm until the next section. To facilitate future discussion of sum rules, we introduce the following formal notations for m m 5 ∈ Z>0, s Z 1, + 1 : ∈ ∩ 2 m/2 +1 s ⌊ ⌋2 − m + ℓ ℓ j 2 1 s ℓ (u) := (m + 2) (u) + j ∑ k( 1) − + (u) Im,s Fm,1 − 2 j Fm,2 j 1 j=1 m/2 s ⌊ 2⌋− m + m j 2 1 s ℓ ( 1) i j ∑ k( 1) − + (u), (2.38) − − − 2 j + 1 Fm,2 j 2 j=0 (m+1)/2 1+s ⌊ ⌋− 2 m + 1 s m+1 + s ℓ = j k j 2 + 2 ℓ m,s(u) : ∑ ( 1) − m,2 j+ m +1(u) K − − 2 j + 1 2 j + 1 F 2 j=1 ⌊ ⌋ (m+1)/2 +s ⌊ ⌋ 2 m + 1 s m+1 + s m j k j 2 2 ℓ ( 1) i ∑ ( 1) − m,2 j+ m (u). (2.39) − − − 2 j − 2 j F 2 j=1 ⌊ ⌋ ℓ ℓ The domains of definition for the real-valued functions Re m,s(u), Im m,s(u) and so forth will be clear 6 ´ ℓ ´ ℓ I I ℓ from context. We also write m,s(u) [resp. m,s(u)] for expressions that trade in m,s(u) [resp. ℓ ´ ℓ =I ℓ ℓ K ´ ℓ = ´ ℓ ´ ℓ ℓ = Fℓ I ℓ ´ ℓ = m,s(u)] with , and set Fk,s 2k 1,s 2k 1,s, Fk,s 2k 1,s 2k 1,s, Gk,s 2k,s 2k,s, Gk,s K F I − − K − I − − K − I − K ´ ℓ ´ ℓ . I2k,s − K2k,s Proposition 2.9 (Sum rules at thresholds). In what follows, assume that k Z>1. k ℓ 2 ∈ k 1 (a) For s Z [1,k],ℓ Z 1, , we have F ((2s) ) = 0; for s Z [1,k],ℓ Z 1, − , we ∈ ∩ ∈ ∩ 2 k,s ∈ ∩ ∈ ∩ 2 have F´ ℓ s 2 = . k,s((2 ) ) 0 + k ℓ 2 = + (b) For s Z [1,k 1],ℓ Z 1, 2 , we have Gk,s((2s 1) ) 0; for s Z [1,k 1],ℓ Z k ∈1 ∩ ℓ ∈ ∩ 2 − ∈ ∩ ∈ ∩ 1, − , we have G´ ((2s 1) ) = 0. 2 k,s − 5 0 It is understood that expressions like ∑ = ( ) are empty sums, hence vanishing. j 1 ··· 6Note that Re ℓ (u) and Im ℓ (u) do not necessarily share their domains of definition, neither do Re ℓ (u) and Im,s Im,s Km,s Im ℓ (u). In view of this, the notations ℓ (u), ℓ (u) and so forth are convenient short-hands, instead of complex-valued Km,s Im,s Km,s functions in the usual sense. BROADHURST–ROBERTSQUADRATICRELATIONS 13

Proof. As in [53, §2.1, §3] and [52, Remark after Lemma 3.2], we will prove these sum rules by studying (1) (2) (1) (2) contour integrals involving the cylindrical Hankel functions H0 (z), H0 (z), H1 (z), H1 (z) for z C r ( ,0], whose asymptotic behavior reads ∈ −∞ (1) + nπ π = 2 i(z 2 4 ) + 1 Hn (z) πz e − − 1 O z | | (2.40) (2) nπ π  = q 2 i(z 2 4 ) h + 1 i Hn (z) πz e− − − 1 O z  | | for n 0,1 , as z , π < argz < π. Furthermore,q theseh Hankel i functions are related to modiﬁed ∈{ } | |→∞ −  Bessel functions I0(t), K0(t), I1(t), K1(t) for t > 0 in the following manner: (1) 2i1 n (2) n 2i1 n Hn (+it) = − Kn(t) Hn (+it) = 2i In(t) + − Kn(t) π + and + π (2.41) (1) = −2 2in 1 (2) = 2in 1 ( Hn ( it) n In(t) Kn(t) ( Hn ( it) Kn(t) − i − π − π where n 0,1 . ∈{ } (a) Armed with the aforementioned preparations, we use two methods to evaluate the contour integral i ∞ (2) (1) 2s (1) (2) k s 2ℓ+n 1 Hn (2sz)[H0 (z)] [H0 (z)H0 (z)] − z − dz, (2.42) i Z− ∞ k n which converges for k Z>1, s Z [1,k],n 0,1 ,ℓ Z 1, −2 . First, by the asymptotic behavior of cylindrical∈ Hankel functions∈ ∩ in (2.40),∈{ we} can∈ c∩lose the contour to the right, and show that the integral vanishes. Second, with the aid of (2.41), we may recognize the contour integral as a constant multiple of

∞ n 1 n k+s k s 2ℓ+n 1 [i πIn(2st) + i − Kn(2st)][ iK0(t)] [πI0(t) + iK0(t)] − t − dt − Z0 n ∞ n+1 k+s k s 2ℓ+n 1 ( 1) i Kn(2st)[πI0(t) iK0(t)] [iK0(t)] − t − dt (2.43) − − − Z0 For n = 0, the last vanishing integral reduces to a sum rule k s ∞ − k s j 2k j 2ℓ+n 1 [πI0(2st) + iK0(2st)] ∑ − [πI0(t)] [iK0(t)] − t − dt j Z0 j=0 k+s + ∞ k s j 2k j 2ℓ+n 1 iK0(2st) ∑ [πI0(t)] [ iK0(t)] − t − dt = 0, (2.44) − j − Z0 j=0 ℓ 2 = = ´ ℓ 2 = which is equivalent to Fk,s((2s) ) 0. For n 1, we can deduce Fk,s((2s) ) 0 instead. k n (b) When k Z>1, s Z [1,k + 1],n 0,1 ,ℓ Z 1, − , we can evaluate ∈ ∈ ∩ ∈{ } ∈ ∩ 2 i ∞ (2) (1) 2s 1 (1) (2) k+1 s 2ℓ+n 1 = Hn ((2s 1)z)[H0 (z)] − [H0 (z)H0 (z)] − z − dz 0 (2.45) i − Z− ∞ 2 by closing the contour to the right, while noting that the integrand has O( z − ) behavior for large z . This can be reinterpreted into the sum rules Gℓ ((2s 1)2) = 0 and G´ ℓ ((2| s| 1)2) = 0. || k,s − k,s − The sum rules in the last proposition will be essential to local analysis7 in §3.2, especially the determi- A B = m+1 C T nation of the blocks Sm [as given in (1.13)] and Sm ( 1) (Sm) [as given in (1.14)] for the partitioned A B − matrix S = Sm Sm that ﬁts into (2.36). To compute the block SD, and to complete the proof of Theo- m SC SD m m m + rem 1.1 in §3.3, we will need to investigate another threshold u 0 in the next proposition. → 7Here, by “local analysis”, we refer to asymptotic expansions for both sides of (2.36) (with appropriate real-analytic continuations if necessary) when u approaches a threshold. Corollary 2.11 illustrates such local analysis in the case where m = 3. 14 YAJUN ZHOU

In what follows, we deﬁne rational numbers Hk,n for k,n Z>0 through the following asymptotic expansion (cf. [53, (66)]): ∈

π k N H 1 := (1) (2) k,n = hk,N(z) H0 (z)H0 (z) ∑ 2n+k 2 O 2N+k , (2.46) 2 − n=1 z − z h i | | where k, N Z>0. These rational numbers also turn up in the relations for generalized Crandall numbers (cf. [53, (65)]):∈

k + k k n+1 π ∞ [πI0(t) iK0(t)] [πI0(t) iK0(t)] k 2n+k 3 ( 1) Hk,n = − − [K0(t)] t − dt. (2.47) − 2 0 πi Z = π (1) (2) k By the convention that empty sums must vanish, we retroactively deﬁne hk,0(z) 2 H0 (z)H0 (z) and Hk,0 = 0. Furthermore, we introduce a function 1 k n+ k 1 2n+k 3 k/2 k 1 2 π u 2 t u t 1 ut := ∞ − − − − Lk(u,t) ∑ 2 2 1F2 k + k + . (2.48) = 2n 1 Γ + k ≡ k 1 Γ k + 2 1, 2 1 4 n 1 2 − n 2 2π − 2 1

+ √ut For each ﬁxed u (0,1), we have Lk(u,t) = O(1) as t 0 , and Lk(u,t) = O(e /√t) as t , so the dominated convergence∈ theorem allows us to conclude→ that either side of →∞

+ k k n+1 n+ k 1 ∞ [πI0(t) iK0(t)] [πI0(t) iK0(t)] k = ∞ π( 1) Hk,nu 2 − − − [K0(t)] Lk(u,t)dt ∑ − 2 (2.49) 0 πi = 2n+k 1 Γ + k Z n 1 2 − n 2 represents a continuous function in the variable u (0,1). ∈ + 8 Proposition 2.10 (“Sum rules” at threshold u 0 ). (a) For ℓ Z>0, we have → ∈ 1 1 ℓ (u) = O , √u ´ ℓ (u) = O , (2.50) F2k 1,2k 1 (√u)max 1,2ℓ k F2k 1,2k 1 (√u)max 1,2ℓ k − − { − } − − { − } as u 0+. → For each k 1 + 2Z 0 and ℓ Z>0, we have the following quantitative reﬁnements of (2.50) for u (0,1): ∈ ≥ ∈ ∈ k 1 + ( 1)b 1 k 2k 1 1 + ( 1)b k k Oℓ(u) := (2k + 1) ℓ (u) + − ℓ (u) + − − ℓ (u) k 2k 1,1 ∑ − b 1 2k 1,b ∑ − b k + 2k 1,b F − b=2 2( 1) −2 b 1 F − b=k+1 ( 1) −2 b k 1 F − − − − − + k k = ℓ 2 + 1 ℓ ∞ [πI0(t) iK0(t)] [πI0(t) iK0(t)] k 4 (uD D ) k−+1 − [K0(t)] Lk(u,t)dt; (2.51) 0 ( 1) 2 2kπi Z − k 1 + ( 1)b 1 k 2k 1 1 + ( 1)b k k O´ℓ(u) := (2k + 1) ´ ℓ (u) + − ´ ℓ (u) + − − ´ ℓ (u) k 2k 1,1 ∑ − b 1 2k 1,b ∑ − b k + 2k 1,b F − b=2 2( 1) −2 b 1 F − b=k+1 ( 1) −2 b k 1 F − − − − − + k k = ℓ 1 2 + 1 ℓ ∞ [πI0(t) iK0(t)] [πI0(t) iK0(t)] k 4 √uD (uD D ) k+−1 − [K0(t)] Lk(u,t)dt. (2.51′) 0 ( 1) 2 2k 1πi Z − −

8 ℓ ´ ℓ The asymptotic behavior of 2k 1,2k 1(u) and 2k 1,2k 1(u) was misstated as O(1/√u) (irrespective of ℓ) in [49, Propo- sition 4.4]. The corrections hereF − do not− affectF the− proof− of [49, Proposition 4.4], the latter of which only requires + k/2 ℓ = + k/2 ´ ℓ = limu 0 u 2k 1,2k 1(u)logu limu 0 u √u 2k 1,2k 1(u)logu 0 for ℓ Z [1,k 1] to work properly. → F − − → F − − ∈ ∩ − BROADHURST–ROBERTSQUADRATICRELATIONS 15

For each k 2Z>0 and ℓ Z>0, we have the following quantitative reﬁnements of (2.50) in the u 0+ regime:∈ ∈ →

2k 1 b k 2log √u ℓ = − 1 ( 1) − k ℓ + 2 + ℓ + Ek(u) : ∑ − − b k 2k 1,b(u) (2k 1) 2k 1,1(u) −2 b k + 1 F − π F − b=k+1 ( 1) − − 2k 1 1 + ( 1)b k k log2 √u k 1 + ( 1)b k + − − ℓ (u) 2 ℓ (u) ∑ − b k + 2k 1,b 2 ∑ − b 2k 1,b b=k+1 ( 1) −2 b k 1 F − # − π b=2 ( 1) 2 b 1 F − − − − − 2 22(ℓ n k)+1 ℓ n k ! = E ℓ + − − 2 k (u) ∑ − − k Hk,n, (2.52) n+k/2 ℓ n+1 n Z 1,max 0,ℓ k ( 1) πu − − 2 ∈ ∩ − 2 − 2k 1 b k 2log √u ´ℓ = − 1 ( 1) − k ´ ℓ + 2 + ´ ℓ + Ek(u) : ∑ − − b k 2k 1,b(u) (2k 1) 2k 1,1(u) −2 b k + 1 F − π F − b=k+1 ( 1) − − 2k 1 1 + ( 1)b k k log2 √u k 1 + ( 1)b k + − − ´ ℓ (u) 2 ´ ℓ (u) ∑ − b k + 2k 1,b 2 ∑ − b 2k 1,b b=k+1 ( 1) −2 b k 1 F − # − π b=2 ( 1) 2 b 1 F − − − − − E´ℓ ℓ n k+1 k k + k (u) 4 − − ℓ n 2 ! ℓ n 2 1 ! = + ∑ − − − − Hk,n, (2.52′) √u n 1+k/2 ℓ n+ 3 k n Z 1,max 0,ℓ k ( 1) − πu− 2 − 2 ∈ ∩ − 2 − E ℓ E´ℓ = where k (u) and k (u) are holomorphic in a non-void C-neighborhood of u 0. + Furthermore, if a function in spanC 2k 1,b(u),u (0,ε) b Z ( 1 [k 1,2k 1]) (where 0 < ε < 1) extends to a holomorphic function{F − in a non-void∈ C|-neighborhood∈ ∩ { } ∪ of u = 0−, then} it must be identically zero. + (b) If a function in spanC 2k,b(u),u (0,ε) b Z ( 1 [k 2,2k]) (where 0 < ε < 1) extends to a holomorphic function in{F a non-void∈ C-neighborhood| ∈ ∩ { } of ∪ u = 0, then} it must be identically zero.

ε Proof. (a) From sup t I0(t)K0(t) < for any ε (0,1], we can deduce that (cf. [49, (2.27), (2.30)] t>0 ∞ ∈

∞ k 2ℓ 1 ∞ 1 K (√ut)[I (t)K (t)] t − dt = O K (√ut)dt = O , (2.53) 0 0 0 0 √u Z0 Z0 ∞ k 2ℓ ∞ 1 √uK1(√ut)[I0(t)K0(t)] t dt = O √uK1(√ut)t dt = O , (2.54) √u Z0 Z0

n 1 when 2ℓ 1 k. Meanwhile, from Heaviside’s integral formula [46, §13.21(8)] for 0∞ K0(t)t − dt, we have − ≤ R

1 ∞ 2ℓ 1 k ∞ k 1 2ℓ 1 K0(√ut)t − − dt + K0(√ut) [I0(t)K0(t)] t − dt 2k − (2t)k Z0 Z0 Γ k 2 ℓ 2 ∞ 2ℓ 3 k = − + O K (√ut)t − − dt (2.55) 4k ℓ+1(√u)2ℓ k 0 − − Z0

when 2ℓ 1 > k, so the ﬁrst half of (2.50) follows from induction on ℓ. The second half of (2.50) founds on− a similar argument. 16 YAJUN ZHOU

= π (1) (2) k Recall the notation hk,N(z) for k, N Z>0 from (2.46) and set hk,0(z) 2 H0 (z)H0 (z) . The following contour integral ∈ k+1 ℓ 1 2 ∞ k k 2ℓ 1 i( 1) − K0(√ut)[K0(t)] [K0(t) iπI0(t)] t − dt − π 0 − Z 2(ℓ n)+1 k Γ + k 2 (2i) − − ℓ n 1 2 ∑ − − Hk,n − π(√u)2(ℓ n+1) k n Z 1,max 0, ℓ k 1 − − ∈ ∩ − −2 i ∞ (1) 2ℓ 1 = H (√uz)h k 1 (z)z − dz (2.56) 0 k,max 0, ℓ −2 Z0 − can be identiﬁed with ∞ (1) 2ℓ 1 H (√ux)h k 1 (x)x − d x, (2.57) 0 k,max 0, ℓ −2 Z0 − because Jordan’s lemma allows us to rotate the contour of integration 90◦ clockwise, from the positive Imz-axis to the positive Rez-axis, as in [51, §2]. When k is odd, we can relate the real parts of the last two displayed equations as follows: 2k 1 + b k k ℓ 1 + ℓ IKM(1,2k,2ℓ 1 u) + − 1 ( 1) − k ℓ 2 ( 1) − (2k 1) 2k 1,1(u) k − | ∑ − b k 2k 1,b(u) − F − π − b k + 1 F " − b=k+1 ( 1) 2 − # e − − n+ k 1 n+1 k 1 πu 2 ( 1) − ℓ 2 1 ℓ − ( 1) 2 4 (uD + D ) − Hk,n ∑ + 2 − − k 1 2n k 1 Γ + k n Z 1,max 0,ℓ − 2 − n 2 ∈ ∩ − 2 ∞ 2ℓ 1 = J0(√ux)h k 1 (x)x − d x, (2.58) k,max 0,ℓ −2 Z0 − = 1 (1) + (2) where J0(z) 2 [H0 (z) H0 (z)] is the Bessel function of the ﬁrst kind, satisfying J0(x) 1 for all 2ℓ 1 +| | ≤ real-valued x. Here, the function h k 1 (x)x is bounded in the x 0 regime, and has k,max 0,ℓ − − − 2 → O(x 2) behavior as x , so we can invoke the dominated convergence theorem to show that − →∞ ∞ 2ℓ 1 ∞ 2ℓ 1 √ k 1 = k 1 lim+ J0( ux)hk,max 0,ℓ (x)x − d x hk,max 0,ℓ (x)x − d x. (2.59) u 0 0 −2 0 −2 → Z − Z − After rotating the contour of integration 90◦ counterclockwise, we can identify the right-hand side of the last equation with i ∞ 2ℓ 1 Re h k 1 (z)z − dz k,max 0,ℓ −2 Z0 − k ℓ 2 ∞ k k 2ℓ 1 = ( 1) Re [K0(t)] [K0(t) iπI0(t)] t − dt − π − Z0 IKM(1,2k,2ℓ 1 u) k 1 + ( 1)b 1 k = k ℓ + − ℓ u . 2 ( 1) lim+ k − | ∑ − b 1 2k 1,b( ) (2.60) − u 0 π −2 b 1 F − → " b=2 2( 1) − # e − + ℓ ℓ = So far, we have already arrived at a precursor to (2.51), namely limu 0 [Ok(u) Ok(u)] 0, where ℓ → − Ok(u) denotes the right-hand side of (2.51). This precursor can be rewritten as 2 + 1 ℓ 1 1 1 = e lim (uD D ) − [Ok(u) Ok(u)] 0, (2.61) u 0+ − e → ℓ = ℓ 1 2 + 1 ℓ 1 1 for all ℓ Z>0, since we have Ok(u) 4 − (uD D ) − Ok(u) by the Bessel differential equation. ∈ 1 e Now that Ok(u),u (0,εk) is equal to a generalized power series (by Corollary 2.2) for a suitably ∈ 1 chosen εk (0,1), and so is O (u),u (0,1) (by its construction and the dominated convergence ∈ k ∈

e BROADHURST–ROBERTSQUADRATICRELATIONS 17

1 1 theorem), the vanishing derivatives in (2.61) bring us O (u) = O (u),u (0,εk). By real-analytic k k ∈ continuations and differentiations, we can establish (2.51) and (2.51′) in their entirety. When k is even, we can identify the imaginary parts of (2.56) and (2.57) as follows: e + 2k 1 b k k ℓ 1 2IKM(0,2k 1,2ℓ 1 u) − 1 ( 1) − k ℓ 2 ( 1) − k+1 − | ∑ − − b k 2k 1,b(u) − π − − b k + 1 F " b=k+1 ( 1) 2 − # e − − ℓ n 1 k k 2 ( 4) − 2 − ℓ n 2 ! ∑ − − − Hk,n − k/2 ℓ n+1 k n Z 1,max 0,ℓ k ( 1) πu − − 2 ∈ ∩ − 2 − ∞ 2ℓ 1 = Y0(√ux)h k (x)x − d x, (2.62) k,max 0,ℓ 2 Z0 − = 1 (1) (2) where Y0(z) 2i [H0 (z) H0 (z)] is the Bessel function of the second kind. Meanwhile, the real parts of (2.56) and (2.57)− leave us 2k 1 + b k k ℓ 1 + ℓ IKM(1,2k,2ℓ 1 u) + − 1 ( 1) − k ℓ 2 ( 1) − (2k 1) 2k 1,1(u) k − | ∑ − b k 2k 1,b(u) − F − − π −2 b k + 1 F − " b=k+1 ( 1) − # e − ∞ 2ℓ 1 = J0(√ux)h k (x)x − d x, (2.63) k,max 0,ℓ 2 Z0 − which is subtly different from (2.58). Here, we point out that √u IKM(0,2k + 1,2ℓ 1 u) + IKM(1,2k,2ℓ 1 u)log − | − | 2 √u = e∞ K (√ut) + I (√ut)loge [K (t)]2kt2ℓ 1 dt (2.64) 0 0 2 0 − Z0 equals its own Maclaurin series for u (0,(2k)2), because the dominated convergence theorem allows us to perform termwise integration on∈ the convergent series [46, §3.71(14)] (0) 2 m √ut ∞ ψ (m + 1) ut K0(√ut) + I0(√ut)log = ∑ , (2.65) 2 (m!)2 4 m=0 (0) + = 1 m x + 1 where ψ (m 1) : m! 0∞ x e− log xd x goes like logm O m as m . As we may recall from [46, §3.51], the following limit →∞ R 2 √u 2(γ0 + log x) lim Y0(√ux) J0(√ux)log = (2.66) u 0+ − π 2 π → (0) holds for all x > 0, where γ0 = ψ (1) is the Euler–Mascheroni constant. Meanwhile, we note that 2ℓ 1 = − 2ℓ 1 + the function hk,ℓ(x)x − : hk,max 0,ℓ k (x)x − Hk,max 0,ℓ+1 k h2,0(x)x is bounded in the x 0 2 − 2 → 3 − − regime, and has O(x− ) behavior as x , so the dominated convergence theorem allows us to conclude thate →∞

∞ 2 √u 2ℓ 1 = 2 ∞ 2ℓ 1 + lim+ Y0(√ux) J0(√ux)log hk,ℓ(x)x − d x hk,ℓ(x)z − (γ0 log x)d x (2.67) u 0 0 − π 2 π 0 → Z Z e e 2ℓ 1 2ℓ 1 is a ﬁnite real number. To compensate for the difference hk,max 0,ℓ k (x)x − hk,ℓ(x)x − , we need − 2 − the explicit knowledge that (cf. [52, Lemma 3.2]) 1 + 1 2 2 s e √ 4 i Γ s Γ s [Γ(s)] = 3 1 ∞ 3 3 108u − 3,1(u) / − − 1 d s F 20π1 2 2πi 1 i (4 u)Γ(1 s)Γ s + (4 u)3 Z 4 − ∞ − − 2 − 18 YAJUN ZHOU

u O u log (4 u)3 (4 u)3 = − + − , (2.68) − 10(4 u) 4 u − 1 + −1 1 2 3 s √ 4 i Γ s Γ s Γ s [Γ(s)] = 3 1 ∞ 3 2 3 108u − 3,3(u) 5/2 − − − 3 d s 3,2(u) F 8π 2πi 1 i 4 u (4 u) −F Z 4 − ∞ − − 2 + 2 u O u π 3log (4 u)3 (4 u)3 = − + − (2.69) 24π(4 u) 4 u − − hold in the u 0+ regime (where the logarithmic terms are attributed to residues at s = 0), so as to deduce →

∞ 2 √u Hk,max 0,ℓ+1 k Y0(√ux) J0(√ux)log h2,0(x)xd x 2 − π 2 − Z0 = 40 √u + Hk,max 0,ℓ+1 k 8 3,3(u) 3,1(u)log − 2 (− F − π F 2 2 3 √u + ∞ K (√ut) + I (√ut)log [K (t)]4t dt π 0 0 2 0 Z0 ) 2k( 1)ℓ log2 √u k + b = 2 1 ( 1) k ℓ + − 2 ∑ − b 2k 1,b(0) O(1). (2.70) π b=2 ( 1) 2 b 1 F − − − Here, in the last step, we have combined (2.47) with the generalized Bailey–Borwein–Broadhurst– Glasser sum rule [53, (5)] into the following form for k 2Z>0 and ℓ Z>0: ∈ ∈ k k k k [πI0(t) + iK0(t)] [πI0(t) iK0(t)] 2K0(t) ℓ 2 = ∞ 2ℓ 1 ( 1) − Hk,max 0,ℓ+1 k − − t − dt. (2.71) − 2 πi π − Z0 ℓ ℓ Let Ek(u) be the ﬁnite sum on the right-hand side of (2.52), then the efforts so far lead us to Ek(u) + − Eℓ(u) = O(1) as u 0 . Before proceeding further, we note that the sequence k → e 2n n 2 1 n √u 1 2 √u bn(u,t) := t 4 (uD + D ) K0(√ut)log + I0(√ut)log (2.72) e − 2 2 2 satisﬁes a recursion 2n 2 1 t √u 4(uD + D )bn(u,t) bn+1(u,t) = I0(√ut) + 2√ut I1(√ut)log K1(√ut) (2.73) − u 2 − with the initial condition b0(u,t) = 0, and [46, §3.71(15)] (0) (0) 2 m+1 √ut 1 ∞ ψ (m + 1) + ψ (m + 2) ut √ut I1(√ut)log K1(√ut) = 1 + ∑ , (2.74) 2 − − 2 m!(m + 1)! 4 m=0 so we can show inductively that n m 1 2 4 − [(m 1)!] 2(n m) = ∞ m bn(u,t) ∑ m− t − ∑ an,m(t)u (2.75) − m=1 u m=0

defines a holomorphic function in u C for each fixed t (0, ) and n Z 0. Therefore, we have ℓ ℓ = ℓ 1 2 + 1 ℓ 1 1∈ + +∈ ∞ ∈ ≥ Ek(u) Ek(u) 4 − (uD D ) − Ek (u) O(1) as u 0 . By Corollary 2.2, there exists a suitable − 1 → εk (0,1) such that the function E (u),u (0,εk) equals a generalized power series in (fractional) ∈ k ∈ powers ofe u and logu; such a series is in fact a Maclaurin series, the coefficients of which are com- + 2 + 1 ℓ 1 1 1 = E 1 putable from the finite numbers limu 0 (uD D ) − Ek (u) for ℓ Z>0. This reveals Ek (u) k (u) as a holomorphic function in a non-void→ C-neighborhood of u = ∈0. The full version of (2.52) then BROADHURST–ROBERTSQUADRATICRELATIONS 19

follows from repeated applications of the Bessel differentiation operator uD2 + D1 and invocations of the Laurent expansion for bn(u,t), as given in (2.75). To prove (2.52′), we differentiate (2.52) while noting that

1 √u 1 2 √u √u 1 2 √u 2uD K0(√ut)log + I0(√ut)log √ut K1(√ut)log + I1(√ut)log 2 2 2 − − 2 2 2 √u = K (√ut) + I (√ut)log (2.76) 0 0 2

extends to a holomorphic function of u C. ∈ + 2k √u + According to (2.65), we know that π 2k 1,1(u) 2k+1 2k 1,2(u)log 2 and π 2k 1,b(u) √u F − F − F − 2k 1,b k+2(u)log for b Z [k + 1,2k 2] are holomorphic in a non-void C-neighborhood of F − − 2 ∈ ∩ − u = 0. From (2.51) and (2.52), we know that the generalized power series expansion for 2k 1,2k 1(u) + 1 √u F − − Z 2 Z 2 additionally involve u or u log 2 terms, according as k is odd or even. k + 2k 1 2k b If a linear combination of these functions [say, c1π 2k 1,1(u) ∑b=−k+1 cbπ − 2k 1,b(u)] is holo- + 1F − √u F − Z 2 Z 2 = morphic near the origin, then it should not involve u or u log 2 terms, hence c2k 1 0. Further- k +− 2k 2 2k b more, we require that in the generalized power series expansion for c1π 2k 1,1(u) ∑b=−k+1 cbπ − − n √u F 2k 1,b(u), the coefﬁcients for u log ,n Z [0,k 2] should all vanish, that is, F − 2 ∈ ∩ −

k 1 2k 2 2kc1π − − cb 2k 2,b k+1(1) (1) + F − − = 0, ℓ Z [1,k 1]. (2.77) + 2k 2,1 ∑ b 2k+1 2k 1 F − b=k+1 π − ∈ ∩ −

= k+ 1 b ℓ = = + Since detNk 1 det(π 2 − 2k 2,b(1))1 b,ℓ k 1 0, we must have cb 0 for b Z ( 1 [k 1,2k 2]). − F − ≤ ≤ − 6 ∈ ∩ { }∪ − = + 2k+1 √u (b) Inanon-void C-neighborhood of u 0, we have holomorphic functions π 2k,1(u) 2k+2 2k,2(u)log 2 √u F F and π 2k,b(u)+ 2k,b k+1(u)log for b Z [k+2,2k]. The conclusion follows from the coefﬁcients F F − 2 ∈ ∩ n √u = k+1 b ℓ = for u log 2 ,n Z [0,k 1] and the fact that detMk det(π − 2k 1,b(1))1 b,ℓ k 0. ∈ ∩ − F − ≤ ≤ 6

Corollary 2.11 (Determination of S3). The matrix

25 = − 4 S3 3 4 (2.78)  −4    ﬁts into a special case of (2.36), namely

W 3,2(u), 3,3(u) 3,1(u) F F F √ 3(u) = |L | W 3,1(u), 3,3(u) S3 3,2(u) (2.79) Λ3 − F F  F  W (u), (u) 3,3(u)  3,1 3,2  F   F F    2 1 for 3(u) = u (4 u)(16 u),u (0,4) and Λ3 = . |L | − − ∈ 20 Proof. During the proof of [49, Proposition 5.5], we have already pointed out that the left-hand side of (2.79) is annihilated by L3 (which is a special case of Proposition 3.1 below), and have evaluated the ﬁrst row of S3. e 20 YAJUN ZHOU

2 π2 To compute the last row of S3, simply check that IKM(2,3;1 u) = IKM(1,3;1)[1 + O(u )] = [1 + | 16 O(u2)] [3, (55)] as well as e 32 D0 IKM(1,4;1 u) D0 IKM(2,3;1 u) lim 3(u) W 3,1(u), 3,2(u) = lim det | | u 0+ |L | F F u 0+ 5π3 uD1 IKM(1,4;1 u) uD1 IKM(2,3;1 u) → → | | p 2 2 e e 32 π logu + O(ulogu) π + O(u2) π 1 = lim det 32 16 = =e (0). e (2.80) + 3 − π2 3,2 u 0 5π + O(ulogu) O(u)! 80 5F → − 32 This accounts for the only non-vanishing element in the last row of S3, since any non-trivial linear = logu + = log2 u + combination of 3,1(u) 40 [1 O(u)] (cf. [3, (55)] or (2.68) above) and 3,3(u) 32π O(logu) (cf. [49, (5.43)] orF (2.69) above)− in the u 0+ regime is incompatible with suchF a ﬁnite limit. [To argue → for S2k 1,k Z>2, one will need sum rules (2.51) and (2.52) in Proposition 2.10, instead of the explicit formulae− like∈ (2.68) and (2.69).] = T To compute the second row of S3, we quote in advance the symmetry S3 S3 from Corollary 3.3, 1 which enables us to set up an equation √ 3(u) W 3,1(u), 3,3(u) = c 3,2(u) + 3,3(u) for a cer- − |L | F F F 5 F tain constant c. We observe that limu 4 √ 3(u) W 3,1(u), 3,3(u) = 0 (since W 3,1(u), 3,3(u) is → − |L | F F F F holomorphic in a neighborhood of u = 4) and that 3,2(4) 3 3,3(4) = 0 follows from the sum rule F − F ImF1 (4) = 0, a special case of Proposition 2.9(a). Thus, we must have c = 1 . 2,1 − 15

3. W⋆ ALGEBRA The main purpose of this section is to construct analogs of Corollary 2.11 for matrices of all sizes, and m m explore their consequences. In other words, for each m Z>2, we will build a matrix Sm Q that ﬁts ∈ ∈ × into (2.36)—the rational entries of Sm enable us to express the Wronskian´ determinant W[ f1(u),..., fm 1(u)] (2 m)/2 − as a Q-linear combination of m(u) m, j(u) j Z [1,m] , whenever fn n Z [1,m 1] is a {|L | − F | ∈ ∩ } { | ∈ ∩ − } subset of m, j j Z [1,m] . This master{F | plan∈ decomposes∩ } into several subtasks to be completed in the subsections to follow. In §3.1, we study algebraically all the cofactors of the Wronskian´ matrix W[h1(u),...,hm(u)], where = m j = m h1(u),...,hm(u) are annihilated by a differential operator Lm ∑ j=0 ℓm, j(u)D ( 1) Lm∗ , with smooth 9 − coefﬁcients ℓm, j(u). The algebraic properties of these cofactors account for not only the Vanhove matrix m+1 T Vm(u) in Theorem 1.1, but also the parity relation S = ( e1) S . e m − m In §3.2, we scrutinize the adjugate matrix for W[ m,1(u),..., m,m(u)] analytically, by revisiting some themes in §2, namely Bessel differential equations andF sum rulesF for Bessel moments. In particular, we will show that the local expansions of certain Wronskian´ cofactors and certain linear combinations of m,1(u),..., m,m(u) match each other, when u approaches a threshold value. F In §3.3, weF prove Theorem 1.1 in its entirety, by carefully handling real-analytic continuations of Wronskian´ cofactors across threshold values. In addition to proving the representation of Sm given by Theorem 1.1, we will also explore the recursive structures of the combinatorial sums therein. Such combinatorial analysis will prepare us for the inversion of Sm later in §4.1. Simply put, in §3.1 we settle the qualitative viability of (2.36) through differential equations, and Λ ⋆ = m T augment it to W (u) (m 2)/2 S W (u)V (u); in §3.2 we perform quantitative analysis on both m m(u) m m m |L | − sides of (2.36) near thresholds, preparing us for a complete calculation of Sm in §3.3, as well as an Λ ⋆ = m 1 off-shell quadratic relation that descends from W (u) m/2 [W (u)]− . m m(u) m |L |

9 Being indifferent to detailed structures of the polynomials ℓm, j(u), these algebraic mechanisms extend naturally to differential systems other than Vanhove’s operators. It is our hope that such extensions will enable us to construct other types of quadratic relations among periods of connections [24], going beyond the cases of Bessel moments treated in [25] and the current work. BROADHURST–ROBERTSQUADRATICRELATIONS 21

⋆ 3.1. Cofactors of Wronskians.´ The adjugate W of a Wronskian´ matrix W = W[ f1(u),..., fm(u)] is the transpose of its cofactor matrix cof W. Each entry in the last column of W⋆ = (cof W)T is expressible as a Wronskian´ determinant W[ fi1 (u),..., fim 1(u)], where i1,...,im 1 Z [1,m]. − { − } ⊂ ∩ m Proposition 3.1 (Vanhove duality). For any smooth differential operator Lm = m(u)D + (omitting m L ··· lower order terms) satisfying the parity relation L = ( 1) Lm, and fn(u),u I n Z [1,m 1] m∗ − { ∈ | ∈ ∩ − } ⊂ C (I) ker Lm for an open interval I, we have e ∞ ∩ (m 2)/2 e e Lm( m(u) − W[ f1(u),..., fm 1(u)]) = 0 (3.1) e |L | − on the same open interval. e In particular, this is true when Lm is the m-th order Vanhove operator (deﬁned in Proposition 2.1) with m parity ( 1) (proved in Proposition 2.3), and m(u) is the polynomial speciﬁed by (1.8). − L e Proof. Without loss of generality, we may assume that the functions f1,..., fm 1 are linearly independent. − We may further assume that there exists a function fm, such that span f j(u),u I j Z [1,m] = C{ ∈ | ∈ ∩ } C∞(I) ker Lm. Using∩ Ince’s notation [31, §5.21, p. 120] for Wronskian´ determinants

e ∆0 = 1; ∆r = W[ f1(u),..., fr(u)] where r Z [1,m], (3.2) ∈ ∩ we have 2 2 m Lm f (u) W[ f (u), f1(u),..., fm(u)] m ∆m d ∆m 1 d d ∆1 d ∆0 f (u) ( 1) = = ( 1) − . (3.3) − m(u) ∆m − ∆m 1 du ∆m∆m 2 du ··· du ∆2∆0 du ∆1 L − − e = m From the combinatorial constraint (2.13) that directly descends from the parity relation Lm∗ ( 1) Lm, = m + m d m(u) m 1 − we conclude that Lm m(u)D 2 Ldu D − [cf. (2.3)], up to sub-leading order, so ∆m is propor- m/2 L ··· m/2 tional to m(u) − . In view of the linear independenceof f1,..., fm, weknowthat ∆m = Ce m(u) − e,u |L | = m |L | ∈ I for a non-vanishinge constant C. Appealing again to the parity relation Lm∗ ( 1) Lm, we subsequently have − ∆2 ∆2 e e = C∆0 d 1 d d m 1 d m(u) f (u) Lm f (u) − L m/2 , (3.4) ∆1 du ∆2∆0 du ··· du ∆m∆m 2 du ∆m 1 m(u) − − |L | so (3.1) follows immediately.e Remark The Vanhove duality relations for Vanhove’s operators L3 and L4 (see Table I) have already appeared in [49, (5.34) and (5.67)]. In the next proposition, we show that all the cofactors of Wm aree expressiblee via linear combinations of fundamental solutions to Vanhove’s differential equation, together with their derivatives.

Proposition 3.2 (Vanhove matrix Vm(u)). Fix an integer m Z>1. For a smooth differential operator = m j = m ∈ Lm ∑ j=0 ℓm, j(u)D ( 1) Lm∗ where ℓm, j C∞(I), consider f1,..., fm 1 C∞(I) ker Lm. Deﬁne − ∈ − ∈ ∩ m 2 − i 1 Fr(u) := ℓm,m(u) 2 det(D − f j(u))i Z ([1,m 1 r] [m+1 r,m]), j Z [1,m 1] (3.5) e | e | ∈ ∩ − − ∪ − ∈ ∩ − e for r Z [0,m 1], where all the row indices are arranged in increasing order (even when they are not consecutive∈ ∩ integers).− We have m 1 r (( 1) − Fm 1(u),...,( 1) Fr(u),..., F0(u)) − − − m+1 0 m r m 1 T = ( 1) (D F0(u),..., D − F0(u),..., D − F0(u))V (u) (3.6) − m where m n a b+1 = a+n+m+1 n a D − − ℓm,n(u) (Vm(u))a,b : ∑ ( 1) − (3.7) n=a+b 1 − b 1 ℓm,m(u) − − for a,b Z [1,m]. ∈ ∩ 22 YAJUN ZHOU

In particular, when Lm is the m-th order Vanhove operator (defined in Proposition 2.1) with parity ( 1)m (proved in Proposition 2.3), this explains the origin of the Vanhove matrix V (u) defined in (1.11). − m e Proof. Set m(u) := ℓm,m(u). If f1,..., fm 1 C∞(I) ker Lm, then the first-order derivative of L − ∈ ∩ m 2 m 2 = 2− i 1 = 2− F0(u) : m(u) det(D − f j(u))i, j Z [1,m 1] m(ue) W[ f1(u),..., fm 1(u)] C∞(I) ker Lm |L | ∈ ∩ − |L | − ∈ ∩ (3.8) reads e

m 2 m 2 F (u) 1 − i 1 0 1 D F0(u) = m(u) 2 det(D − f j(u))i Z ([1,m 2] m ), j Z [1,m 1] + − D m(u) |L | ∈ ∩ − ∪{ } ∈ ∩ − 2 m(u) L L m 2 F0(u) 1 = F1(u) + − D m(u). (3.9) 2 m(u) L L For higher order derivatives, we have the following recursion:

Fr(u) = i 1 m 2 : det(D − f j(u))i Z ([1,m 1 r] [m+1 r,m]), j Z [1,m 1] m(u) 2− ∈ ∩ − − ∪ − ∈ ∩ − |L | 1 i 1 = D det(D − f j(u))i Z ([1,m r] [m+2 r,m]), j Z [1,m 1] ∈ ∩ − ∪ − ∈ ∩ − i 1 det(D − f j(u))i Z ([1,m r] [m+2 r,m 1] m+1 ), j Z [1,m 1] (3.10) − ∈ ∩ − ∪ − − ∪{ } ∈ ∩ − according to fundamental properties of determinants. Upon invoking the differential equations Lm f j(u) = 0, j Z [1,m 1], we can further convert this recursion into ∈ ∩ − r e 1 Fr 1(u) 1 ( 1) ℓm,m r(u) Fr(u) = D Fr 1(u) + − D m(u) + − − F0(u). (3.11) − m(u) L m(u) L L Such a recursion is actually retroactively compatible with the r = 1 case in (3.9), since ℓm,m 1(u) = m d = m j = m − 2 du m(u) is a consequence of the parity relation Lm ∑ j=0 ℓm, j(u)D ( 1) Lm∗ . In other words, to buildL a linear combination − e q e Fr(u) = ∑ϕr,q(u)D F0(u) (3.12) q for r Z [0,m 1], it would sufﬁce to set ϕr,q(u) 0 for q Z r [0,r], and ∈ ∩ − ≡ ∈ ϕ0,0(u) 1, ≡ 1 r D [ m(u)ϕr 1,q(u)] ( 1) ℓm,m r(u) (3.13)  ϕr,q(u) = ϕr 1,q 1(u) + L − + − − ϕ0,q(u).  − − m(u) m(u) L L It is then not hard to verify that m + r+p m q p+1 r p m D − − ℓm,p(u) ϕr,q(u) = ∑ ( 1) − (3.14) p=m r+q − q m(u) − L for q Z [0,r] is the unique solution to the aforementioned recursion. This explains the origin of (3.7). ∈ ∩

Suppose that the Wronskian´ matrix Wm(u) = W[ f1(u),..., fm(u)] of our concern has determinant m/2 Λm m(u) − , where Λm = 0. The statement of Proposition 3.1 guarantees the existence of a constant |L | 6 ⋆ = ⋆ matrix Sm such that the last column of W Wm(u) satisﬁes [cf. (2.36)] m 2 − m(u) 2 ⋆ ⋆ T T |L | ((W )1,m,...,(W )m,m) = Sm( f1(u),..., fm(u)) . (3.15) Λm BROADHURST–ROBERTSQUADRATICRELATIONS 23

m+1Λ ⋆ = ( 1) m T T T = The result in Proposition 3.2 then tells us that Wm(u) − (m 2)/2 SmWm(u)Vm(u), where detVm(u) m(u) − ⋆ |L | m/2 | | 1, so the adjugate equation WW = det(W)I for W = Wm(u) and det(W) =Λm m(u) translates into |L |− 1 T 1 T 1 m+1 [Wm(u)]− [Vm(u)]− [Wm(u)]− Sm = ( 1) . (3.16) − m(u) |L | Corollary 3.3 ((Skew) symmetries of certain matrices). Under the same assumptions as Proposition 3.2, we have the relation V (u) = ( 1)m+1VT (u), which entails S = ( 1)m+1ST . m − m m − m = Proof. First we show that (V2k 1(u))a,b (V2k 1(u))b,a for k Z>0. − − = ∈ a 1 = Along the main anti-diagonal, we have (V2k 1(u))a,2k a ( 1) − (V2k 1(u))2k a,a for a Z [1,2k 1]. As we march away from the main anti-diagonal,− and− work− with a + b− Z (2− ,2k),a ∈Z ∩[1,2k − 2],b Z [2,2k 2], we have the following recursion: ∈ ∩ ∈ ∩ − ∈ ∩ − 2k 1 n a b+1 + = − a+n n a n a 1 D − − ℓ2k 1,n(u) (V2k 1(u))a,b (V2k 1(u))a+1,b 1 ∑ ( 1) − − − − − − − n=a+b 1 − b 1 − b 2 2k 1(u) − − − L − 2k 1 n a b+1 − a+n n a 1 D − − ℓ2k 1,n(u) = ∑ ( 1) − − − = + − b 1 2k 1(u) n a b − L − 1 D [ 2k 1(u)(V2k 1(u))a+1,b] = L − − . (3.17) − 2k 1(u) L − = In view of this recursion, in order to show that (V2k 1(u))a,2k 1 a (V2k 1(u))2k 1 a,a for all a − = − − − − − ∈ Z [1,2k 2], it would sufﬁce to verify that (V2k 1(u))1,2k 2 (V2k 1(u))2k 2,1. Direct computations reveal∩ that− − − − −

2k 1 + n 2k+2 = − 1+n n 1 + n 2k 2 D − ℓ2k 1,n(u) (V2k 1(u))1,2k 2 (V2k 1(u))2k 2,1 ∑ ( 1) − − − − − − − − n=2k 2 − 2k 3 0 2k 1(u) − − L − 1 1 = [(2k 1)D ℓ2k 1,2k 1(u) 2ℓ2k 1,2k 2(u)] = 0, (3.18) 2k 1(u) − − − − − − L − according to (2.13). = The efforts in the last paragraph amount to a veriﬁcation that (V2k 1(u))a,b (V2k 1(u))b,a holds for a + b = 2k 1. To show that the same identity holds true for smaller− values of a + b,− we exploit (3.17) − = again, and build up on the “boundary conditions” (V2k 1(u))1,b (V2k 1(u))b,1 for b Z [2,2k 3]. By (3.7) and elementary combinatorics, we have − − ∈ ∩ −

2k 1 n b = − 1+n n 1 b+n n b D − ℓ2k 1,n(u) (V2k 1(u))1,b (V2k 1(u))b,1 ∑ ( 1) − ( 1) − − − − − = − b 1 − − 0 2k 1(u) n b − L − 2k 1 n b − 1+n n n n 1 b+n D − ℓ2k 1,n(u) = ∑ ( 1) + ( 1) − ( 1) − . (3.19) = − b − b − − 2k 1(u) n b L − Referring back to the combinatorial constraint in (2.12), which says

2k 1 − n n n b ℓ2k 1,b(u) = ∑ ( 1) D − ℓ2k 1,n(u), (3.20) − − − b − n=b 24 YAJUN ZHOU we may further reduce (3.19) into

(V2k 1(u))1,b (V2k 1(u))b,1 − − − 2k 1 n b 1+1 ℓ2k 1,b(u) ℓ2k 1,b(u) − 1+n n 1 b+1+n D − − ℓ2k 1,n(u) = − − − ∑ ( 1) − ( 1) − 2k 1(u) − = + − b + 1 1 − − 2k 1(u) L − n b 1 − L − 1 D 2k 1(u)[(V2k 1(u))1,b+1 (V2k 1(u))b+1,1] = {L − − − − }. (3.19′) − 2k 1(u) L − = = Thus, it is clear that (V2k 1(u))1,b (V2k 1(u))b,1 indeed holds for b Z [2,2k 3], so V2k 1(u) T − − ∈ ∩ − − V2k 1(u) is true. − T 1 1 = Since [W2k 1(u)]− is the transpose of [W2k 1(u)]− , the relation (3.16) allows us to deduce S2k 1 T − = T − − S2k 1 from V2k 1(u) V2k 1(u). − − − = The aforementioned techniques also lead us to (V2k(u))1,b (V2k(u))b,1 for b Z [2,2k], and con- = + + − ∈ ∩ sequently (V2k(u))a,b (V2k(u))b,a for a b Z [3,2k 1]. What remains to be investigated is the case where a = b = 1. By− direct computation, we∈ have∩

2k n 1 2k n 1 = n n 1 D − ℓ2k,n(u) = n D − ℓ2k,n(u) (V2k(u))1,1 ∑( 1) − ∑( 1) , (3.21) − 0 2k(u) − 2k(u) n=1 L n=1 L 2k n 2 2k n 2 = n+1 n 2 D − ℓ2k,n(u) = n+1 D − ℓ2k,n(u) (V2k(u))2,1 ∑( 1) − ∑( 1) , (3.22) − 0 2k(u) − 2k(u) n=2 L n=2 L 2k n 2 2k n 2 = n n 1 D − ℓ2k,n(u) = n D − ℓ2k,n(u) (V2k(u))1,2 ∑( 1) − ∑( 1) (n 1) , (3.23) − 1 2k(u) − − 2k(u) n=2 L n=2 L so another invocation of (2.12) brings us

1 V + V = + D 2k(u)[( 2k(u))2,1 ( 2k(u))1,2] 2(V2k(u))1,1 2(V2k(u))1,1 {L } k(u) L2 2ℓ (u) 2k n Dn 1ℓ (u) = 2k,1 + ∑( 1)n − 2k,n − k(u) − 1 k(u) L2 n=2 L2 2ℓ2k,1(u) + ℓ2k,1(u) + ℓ2k,1(u) = − = 0, (3.24) k(u) L2 as expected. Now that we have V (u) = VT (u), the skew symmetry S = ST follows from (3.16). 2k − 2k 2k − 2k 3.2. Threshold behavior of Wronskian´ cofactors. Naturally, our next goal is to seek explicit formulae (in terms of fundamental solutions to Vanhove’s differential equations) for the cofactor Wronskian´ determinants

r+m ⋆ V ( 1) (W (u))r,m = W m,1(u),..., m,r(u),..., m,m(u) − m F F F = i 1 det( D − m, j(u))i Z [1,m 1], j Z ([1,r 1] [r+1,m]), (3.25) F ∈ ∩ − ∈ ∩ − ∪ for all r Z [1,m]. This will amount to the complete characterizations of the matrix Sm that ﬁts into the identity∈ (2.36)∩ for u (0,1). Even though the (skew)∈ symmetry relations in Corollary 3.3 may nearly halve our workload on the elements of Sm, we still need somewhat extensive preparations in Lemmata 3.4 and 3.5 below, before carrying out our computations for k Z>2 in Proposition 3.6. ∈ BROADHURST–ROBERTSQUADRATICRELATIONS 25

Let Cω(a,b) be the totality of real-analytic functions10 on the interval (a,b). The following patterns

k 3 (u) 2 2k 1 − V ω |L − | W 2k 1,1(u),..., 2k 1,k(u),..., 2k 1,2k 1(u) C (0,16), (3.26) k 3 4 u 2 F − F − F − − ∈ | − | − (u)k 1 2k − V ω |L | W 2k,1(u),..., 2k,k+1(u),..., 2k,2k(u)] C (0,9) (3.27) 1 u k 1 F F F ∈ | − | − motivate us to search for certain linear combinations of Bessel moments with suitable orders of vanishing and analytic behavior as u approaches a threshold, in the next lemma. Lemma 3.4 (Local behavior near thresholds). In what follows, the exact value of a positive number ε may vary from context to context. ω 2 2 (a) For each s Z [1,k], there exists a real-analytic function fk,s C ((2s) ε,(2s) +ε) such that we have ∈ ∩ ∈ − k s 1 k − +(k 1)(k s) 1 ( 4s) − ( 1) 2 − − 2 k 3 2 ReF (u) = − − u (2s) − 2 1 + [u (2s) ]fk,s(u) , (3.28) k,s (2k 3)!! 2 | − | − − for ((2s)2 u)( 1)k s [0,ε), and − − − ∈ k s 1 1 k − − +(k 1)(k s 1) 1 ( 4s) − ( 1) 2 − − − 2 k 3 2 ImF (u) = − − u (2s) − 2 1 + [u (2s) ]fk,s(u) , (3.29) k,s (2k 3)!! 2 | − | − − 2 k s 1 for ((2s) u)( 1) − − [0,ε). − − ∈ ω 2 2 (b) For each s Z [1,k + 1], there exists a real-analytic function gk,s C ((2s 1) ε,(2s 1) + ε) such that we∈ have∩ ∈ − − − k s − +k s ( 1) 2 − π 2 k 1 2 Gk,s(u) := − [u (2s 1) ] − 1 + [u (2s 1) ]gk,s(u) 22k(k 1)! (2s 1)2k 1 − − − − − r − − ReG1 (u), ( 1)k s = +1, = k,s − − (3.30) ImG1 (u), ( 1)k s = 1, ( k,s − − − and the expression

Lk,s(u) k s 1 1 + ( 1) − − iG 2 iπ 2 2 2k,s(u) − π(1 i) − k,s(u) log u (2s 1) 2 , u ((2s 1) ε,(2s 1) ), := I k −s 1 | − − | − ∈ − − − (3.31) 1 + ( 1) − − iG 2 + iπ 2 2 + ( 2k,s(u) − π(1 i) − k,s(u) log u (2s 1) 2 , u ((2s 1) ,(2s 1) ε) K − | − − | ∈ − − extends to a real-analytic function in Cω((2s 1)2 ε,(2s 1)2 + ε). − − − Proof. (a) If k s is even, then ReF1 (u) [see (2.38) and (2.39) for explicit formulae] is well-deﬁned − k,s 2 2 k s and real-analytic for u ((2s 2) ,(2s) ), and the function ( 1) −2 (1 + 2kδs,k) k s + (u) = 2k 1,2 −2 1 k s ∈ − − F − ( 1) −2 (1 + 2kδs,k) 2k 1,k s+1(u) is the only addend that fails to be real-analytic in a neighb or- − − − = 2 F 1 = hood of u (2s) . Moreover, the function ReFk,s(u) has vanishing derivatives up to order k 2 k k 1 − max 2 2 1 ,2 −2 1 , according to Proposition 2.9(a). We claim that the leading order be- − 1 − 2 k 3 havior for ReF (u) is A[(2s) u] 2 [1 + o(1)], where the constant A can be determined through k,s − −

n 10 (u u0) n For each u0 (a,b), there exists a positive number ε, such that the Taylor series ∑n∞=0 −n! D f (u0) of a real-analytic ω ∈ function f C (a,b) converges to f (u) for all u (u0 ε,u0 + ε). As a slight abuse of notations, we sometimes use the same symbol for∈ a real-analytic function and its real-analytic∈ − continuation to a larger domain of deﬁnition. 26 YAJUN ZHOU

l’Hôpital’s rule. Concretely speaking, when k is odd, we have

1 k 1 1 ReF (u) D − ReF (u) A = lim k,s = lim k,s + 3 + 3 u (2s)2 0 2 k 2 u (2s)2 0 k 1 2 k 2 → − [(2s) u] − → − D − [(2s) u] − −k s { − } + − (1 2kδs,k)( 1) 2 k 1 = − lim (2s)2 uD + (u), (3.32) 1 k + − 2k 1,k s 1 ( 2) − (2k 3)!! u (2s)2 0 − F − − − − → − p where

2 k 1 lim (2s) uD − 2k 1,k s+1(u) u (2s)2 0+ − F − − → − 1 p k+1 = lim (2s)2 u 2 (u) k 1 + 2k 1,k s+1 (4s) − u (2s)2 0 − F − − → − + p √ut t k s k s 1 (2s)2 u e e − π = lim − ∞ e t tk dt k 1 + s − (4s) u (2s)2 0 π (1 + 2kδs,k) 0 2π√ut √2πt 2t − → − p Z r 22 3k s1 k = − − p (3.33) 1 + 2kδs,k follows from the asymptotic behavior [46, §7.23] t e 1 π t 1 I0(t) = 1 + O and K0(t) = e− 1 + O , (3.34) √2πt t 2t t r 2 + 1 = t2 2 + 1 = t2 along with Bessel’s differential equations (uD D )I0(t) 4 I0(t) and (uD D )K0(t) 4 K0(t). k s 1 k −2 = ( 4s) − ( 1) This shows that A (2−k 3)!! − j2 k when k is odd. Based on the asymptotic behavior [46, §7.23] − t e 1 π t 1 I1(t) = 1 + O and K1(t) = e− 1 + O , (3.35) √2πt t 2t t r one can perform a similar analysis when k is even. 1 2 + 2 If k s is odd, then ReFk,s(u) is well-deﬁned and real-analytic for u ((2s) ,(2s 2) ), and the − k+s 1 k+s 1 ∈ − ℓ − ℓ function ( 1) 2 + (u) = ( 1) 2 (u) is the only addend that fails to be , k s 1 + 2k 1,3k+s 1 − F2k 1 2 2− k − F − − − = 2 real-analytic in a neighborhood of u (2s) . One can repeat the remaining procedures in the last paragraph to ﬁnd the leading order behavior in such scenarios. 2 By the Fuchs condition [31, §15.3], the operator L2k 1 has a regular singular point at u = (2s) for each s Z [1,k], so the left-hand sides of (3.28) and− (3.29) should be equal to convergent gener- ∈ ∩ alized power series. According to the Frobenius methode [31, §16.1], such series may (in principle) 2 Z 2 Z+ 1 involve both integer powers [u (2s) ] and half-integer powers u (2s) 2 . Our next task is to show that integer powers [u (2−s)2]Z actually will not appear in these| − scenarios.| Without loss of generality,− we only elaborate on the cases where u < (2s)2, k s is even, and k is − odd. We start from the following asymptotic expansions for M Z 0,t : ∈ ≥ →∞ + k s + k s k+2M k+s [πI0(t) iK0(t)] − [πI0(t) iK0(t)] − Fk,s,M(u,t)t := I0(√ut)[ iK0(t)] − − 2 + k s k s k+s [πI0(t) iK0(t)] − [πI0(t) iK0(t)] − + iK0(√ut)[ iK0(t)] − − − 2 + k+s k+s k+2M k s [πI0(t) iK0(t)] [πI0(t) iK0(t)] t +iK0(√ut)[ iK0(t)] − − − − 2 πk+1 BROADHURST–ROBERTSQUADRATICRELATIONS 27

e(√u 2s)t 2M f (u) e(√u 2s)t − ∑ k,s,n = O − , (3.36) − √t tn 2M t 3/2 n=0 − | | ! 2 where fk,s,n(u) is holomorphic in a C-neighborhood of u = (2s) . Integrating the left-hand side of the equation above over t (0, ), while invoking the dominated convergence theorem, we obtain ∈ ∞ k + 2M f u Γ M n + 1 2 M k,s,n( ) 2 2 ∞ 2 k+2M lim ReF (u) ∑ − = Fk,s,M((2s) ,t)t dt. (3.37) 2 + k,s 2M n+ 1 u (2s) 0 " − n=0 √u 2s 2 # 0 → − | − | − Z The right-hand side of the equation above vanishes, because it is a constant multiple of [cf. (2.42)]

i 2M ∞ (2) (1) 2s (1) (2) k s k+2M 1 hk,s,n Re H0 (2sz)[H0 (z)] [H0 (z)H0 (z)] − z ∑ n 2M dz, (3.38) i − √z = z − Z− ∞ ( n 0 ) 3/2 π where the integrand has O( z − ) asymptotics for z , argz 2 , and the contour closes to | | | |→∞ | | ≤ k + 2 M the right. So far, we know that the generalized power series expansion for ReFk,s (u) contains no Z 0 [√u 2s] ≤ terms, for each M Z 0. Likewise, one can argue that the same property applies to −k +M ∈ ≥ ´ 2 1 ReFk,s (u), M Z 0. This proves that the generalized power series expansion for Fk,s(u) contains ∈ ≥ no [√u 2s]Z terms, and equivalently, no [u (2s)2]Z terms, as claimed. (b) Without− loss of generality, we will focus on− the scenarios where k s is even, and k is odd. Under such assumptions, one can show (through direct examination of the− integral representations for off- shell Bessel moments) that ReG1 (u) is real-analytic for u ((max 0,2s 3 )2,(2s + 1)2). At the k,s ∈ { − } threshold u = (2s 1)2, this function has vanishing derivatives up to (k 2)nd order, according to Proposition 2.9(b).− − As a variation on our proof [53, Lemma 2] of the integral formulae for generalized Crandall numbers (2.47), we consider (2) s iε iT H ((2s 1)z) 2 − + + 0 (1) 2s 1 (1) (2) k+1 s k ( 1) k−+1 [H0 (z)] − [H0 (z)H0 (z)] − z − dz (3.39) iT Cε iε ( (2/π) − √2s 1z) Z− Z Z − where Cε is a semi-circular arc in the right half-plane, joining iε to iε. For each fixed ε > 0, we can close the contour to the right, as T . Noting that dz = π−i, we may read off the imaginary part →∞ Cε z k+1 s 2 2 = R( 1) 2 π of the last displayed equation as ReGk,s ((2s 1) ) −2k+j1 k 2s 1 . Falling back on l’Hôpital’s rule, k s − − 1 = ( 1) −2 π 2pk 1 + 2 we can deduce ReGk,s(u) 2−2k(kj 1)!k (2s 1)2k 1 [u (2s 1) ] − [1 o(1)] as u (2s 1) . Here, the − − − − − → − 2 ω 2 2 o(1) term can be rewritten as [u (2sq 1) ]gk,s(u) with gk,s C ((2s 1) ε,(2s 1) + ε), because 1 − − 2 ∈ − − − ReGk,s(u) is real-analytic as u approaches (2s 1) . − n 2 One can represent the Taylor coefficients gk,s,n = n!D gk,s((2s 1) ),n Z 0 through the off- k+1 +M k+1 +M − ∈ ≥ 2 2 ´ 2 2 shell moments ReGk,s ((2s 1) ) and ReGk,s ((2s 1) ) for M Z 0. Here, one can evaluate + − − ∈ ≥ k 1 +M ReG 2 ((2s 1)2) via a contour integral over k,s − (2) 2M+1 H0 ((2s 1)z) (1) 2s 1 (1) (2) k+1 s k+2M ηk,s,n 1 − [H (z)] − [H (z)H (z)] − z ∑ = O , (3.40) (2/π)k+1 0 0 0 − zn 2M z2 n=1 − k+1 +M √π 2 2 = M + akin to our treatment of (3.39). It then follows that ReGk,s ((2s 1) ) 2k+1 ( 1) ηk,s,2M 1. There + − − k 1 +M ´ 2 2 is a similar relation that expresses ReGk,s ((2s 1) ) through coefficients of asymptotic expansions. − 28 YAJUN ZHOU

Now, for M Z 0, we consider ∈ ≥ (2) + H (√uz) + + H (u,z)zk 2M := 0 [H(1)(z)]2s 1[H(1)(z)H(2)(z)]k 1 szk 2M k,s,M (2/π)k+1 0 − 0 0 − 2M η + (u) η (u) e i[√u (2s 1)]z k,s,2M 1 + k,s,n , (3.41) − − − + ∑ n 2M − " z 1 n=1 z − # 2 where the functions ηk,s,n(u) are holomorphicin a C-neighborhood of u = (2s 1) , and Hk,s,M(u,z) = i[√u (2s 1)]z 2 k 2M π π − 2 2 O(e− − − /z − − ) for ﬁxed u and z , 2 argz 2 . For u ((2s 1) ε,(2s 1) ), we can invoke Jordan’s lemma in the form| of|→∞ − ≤ ≤ ∈ − − − i ∞ k+2M ∞ k+2M Hk,s,M(u,z)z dz = Hk,s,M(u, x)x d x. (3.42) Z0 Z0 In view of (2.41), we rewrite the left-hand side of (3.42) as k+1 2 M ∞ k+s k+1 s k+2M ( 1) [iπI0(√ut) K0(√ut)][K0(t)] [iπI0(t) K0(t)] − t dt πi − − − Z0 i i[√u (2s 1)]z 2M ∞ e− − − ηk,s,n(u)(2M n)! ηk,s,2M+1(u) dz ∑ − + , (3.43) − 0 z + 1 − i[√u (2s 1)] 2M 1 n Z n=1 { − − } − + k+1 k 1 + 2 2 M 2 M 2 where ηk,s,2M+1((2s 1) ) = ( 1) ReG ((2s 1) ) and − √π − k,s − i i[√u (2s 1)]z ∞ e− − − iπ dz = log 2s 1 √u γ0 + + o(1) (3.44) z + 1 − | − − | − 2 Z0 as u (2s 1)2 0+. Meanwhile, the right-hand side of (3.42) is continuous across the threshold u = (2→s 1)−2, and− admits analytic continuation to u ((2s 1)2,(2s 1)2 + ε), in the following form: − ∈ − − i ∞ k+2M − ∞ k+2M Hk,s,M(u, x)x d x = Hk,s,M(u,z)z dz Z0 Z0 k+1 2 M ∞ k+1 s k+s k+2M = ( 1) K0(√ut)[K0(t)] − [iπI0(t) + K0(t)] t dt πi − Z0 i i[√u (2s 1)]z 2M − ∞ e− − − ηk,s,n(u)(2M n)! ηk,s,2M+1(u) dz ∑ − + , (3.45) − 0 z + 1 − i[√u (2s 1)] 2M 1 n Z n=1 { − − } − where i i[√u (2s 1)]z − ∞ e− − − iπ dz = log 2s 1 √u γ0 + o(1) (3.46) 0 z + 1 − | − − | − − 2 Z 2 + M 2M n 1 ω as u (2s 1) +0 . Therefore, with h (u) := ∑ = ηk,s,n(u)(2M n)! i[√u (2s 1)] C ((2s → − k,s n 1 − { − − } − ∈ − 1)2 ε,(2s 1)2 + ε) and the understanding that h0 (u) 0, we have the following continuity condi- − − k,s ≡ tion for every M Z 0: ∈ ≥ ∞ k+s k+1 s k+2M lim [iπI0(√ut) K0(√ut)][K0(t)] [iπI0(t) K0(t)] − t dt+ u (2s 1)2 0+ 0 − − → − − Z k+1 k+1 + M πi 2 iπ k 1 +M h (u) + log 2s 1 √u ReG 2 ((2s 1)2) k,s 2 √π | − − | − 2 k,s − −[√u (2s 1)]2M − − #) ∞ k+1 s k+s k+2M = lim K0(√ut)[K0(t)] − [iπI0(t) + K0(t)] t dt+ u (2s 1)2+0+ 0 → − Z BROADHURST–ROBERTSQUADRATICRELATIONS 29

k+1 k+1 + M πi 2 iπ k 1 +M h (u) + log 2s 1 √u + ReG 2 ((2s 1)2) k,s . (3.47) 2 √π | − − | 2 k,s − −[√u (2s 1)]2M − − #) There is a similar continuity condition relating derivatives of these off-shell Bessel moments (with k+1 +M ´ 2 2 respect to u) to ReGk,s ((2s 1) ). − 1 i G Combining the input from the last two paragraphs, one sees that both 2k,s(u) π k,s(u) log u 2 iπ 2 2 1 i I −2 iπ | − (2s 1) ,u ((2s 1) ε,(2s 1) ) and (u) Gk,s(u) log u (2s 1) + ,u ((2s − | − 2 ∈ − − − K2k,s − π | − − | 2 ∈ − 1)2,(2s 1)2 +ε) are equal to their respective Taylor series in powers of [u (2s 1)2]Z 0 , and all the ≥ corresponding− Taylor coefficients agree. − − The next lemma will provide us with special values of the determinants occurring in (3.26) and (3.27), among other things. m + Lemma 3.5 (Some special determinants). Fix an integer m Z>1. For every s Z 1, 2 1 , define 1+( 1)m 2 ∈ ∈ ∩ thresholds θm,s := 2s − and θm, m/2 +2 = , along with Wrońskian determinants − 2 ⌊ ⌋ ∞ i 1 Wm,s(u) := det(D − m, j(u))i Z [1,m], j Z ([1, m/2 +1 s] [ m/2 +2,m+s]), (3.48) F ∈ ∩ ∈ ∩ ⌊ ⌋ − ∪ ⌊ ⌋ i 1 wm,s(u) := det(D − m, j(u))i Z [1,m 1], j Z ([1, m/2 +1 s] [ m/2 +2,m 1+s]), (3.49) F ∈ ∩ − ∈ ∩ ⌊ ⌋ − ∪ ⌊ ⌋ − so that we have

C∞(θm,s,θm,s+1) ker Lm ∩ = m + m + + spanC m, j(u),u (θm,s,θm,s+1) j Z 1, 1 s 2,m s (3.50) F ∈ e ∈ ∩ 2 − ∪ 2 and

1 + (m + 1)δs, m/2 +1 Λm W (u) = ⌊ ⌋ , u (θ ,θ + ), (3.51) m,s s(s+1) m + m/2 m,s m,s 1 ( 1)s m(u) ∈ ( 1) 2 − ⌊ 2 ⌋ |L | − + m m 1 m 1 1 ( 1) m+ 4− − 1 + (m + 1)δs, m/2 +1 2 2− θm,s ( √π) 2 Λm wm,s(θm,s) = + ⌊ ⌋ − , (3.52) s(s 1) m +1 s m/2 ( 1) 2 ( 2 ) (m 2)!! ′m(θm,s) − − ⌊ ⌋ − |L | where Λm is deﬁned in (2.37) and m(u) in (1.8). The expression L + m 1 + m = 2 1 ( 1) 2 m′ (θm,s) θm,s ∏ [θm,s (2 j −2 ) ] L j Z ([1,s 1] [s+1, m/2 +1]) − − ∈ ∩ − ∪ ⌊ ⌋ m m 1 +1 s m 2− m m+1 = ( 1) 2 − 2 θm,s ( + 1 s)!( + s)! (3.53) − ⌊ ⌋ 2 − 2 1 evaluates the derivative (u) = D (u) at u = θm,s. Lm′ Lm Proof. In view of the arguments in [49, Lemma 4.2], one can show that each off-shell Bessel moment m, j(u) listed in (3.50) indeed resides in the kernel space of Vanhove’s operator of order m. What remains F to be elucidated is the mechanism by which Wm,s 1(u),u (θm,s 1,θm,s) and Wm,s(u),u (θm,s,θm,s+1) are relayed to each other [see (3.51)], as well as the rationale− ∈ behind− (3.52). ∈ Without loss of generality, we assume that m = 2k 1 is an odd number. By cofactor expansion, one can compute − 2k 2 W2k,s 1(u) k s 2 D − 2k 1,k+1 s(u) lim − = ( 1) w ((2s) ) lim − − , + 1 − 2k 1,s + F 1 u (2s)2 0 2 2 k − − u (2s)2 0 2 2 k → − (2s) u − → − (2s) u − | − | 2k 2 | − | W2k,s(u) 2 D − 2k 1,2k 1+s(u) lim = w2k 1,s((2s) ) lim F − − , (3.54) 2 + 1 k 2 + 1 k u (2s) +0 (2s)2 u 2 − u (2s) +0 (2s)2 u 2 → | − | − → | − | − 30 YAJUN ZHOU

2 which involve a common factor w2k 1,s((2s) ). When s Z [1,k), by a procedure akin to the proof of (3.33), one can show that − ∈ ∩

2 k 1 2k 2 lim (2s) u − 2 D − 2k 1,k+1 s(u) u (2s)2 0+ | − | F − − → − 1 2 k 1 k = lim (2s) u 2 + (u) 2k 2 + − 2k 1,k 1 s (4s) u (2s)2 0 | − | F − − − → − + 2 k 1 √ut t k s k s (2s) u 2 e e − π = lim | − | − ∞ e t t2k 1 dt + s 2k 2 − − u (2s)2 0 π (4s) 0 2π√ut √2πt 2t → − − Z r Γ 1 k 2 2 3k 1 k p 3 4k 1 k = − 2 − s − = (2k 3)!!2 − s − (3.55) √π − 2 k 1 2k 2 and that limu (2s)2+0+ (2s) u − 2 D − 2k 1,2k 1+s(u) evaluates to exactly the same number. This | − | F − − shows that the→ asymptotic behavior of our Wronskian´ determinants match at u (2s)2 0+, up to a trackable sign. Therefore, we can verify (3.51) for s Z [1,k) by successively applying→ such± a matching procedure to (1.9). When s = k, one must exert extra∈ care,∩ in that

2 k 1 2k 2 1 (2k) u − 2 D − IKM(1,2k;1 u) 2 k 2 2k 2 = lim (2k) u − D − 2k 1,1(u) lim | −k | | (3.56) u (2k)2 0+ | − | F − u (2k)2 0+ π 2k + 1 → − → − e 1 carries a factor of 2k+1 . 2 As a by-product, the common factor w2k 1,s((2s) ) during this matching procedure can be identiﬁed with the right-hand side of (3.52). −

3.3. Quadratic relations for off-shell Bessel moments. With the preparations in the last two lemmata, we can compute Sm for all m Z>0. ∈ Proposition 3.6 (Explicit formulae for Sm). (a) If a Z [1,k], then ∈ ∩ (a 1)!(2k + 1 a)!(S2k 1)a,b − − − 2k 1 a k + 2 − ( 1) 2 − (1 2kδa,1) − + 1+( 1)a 1 2k a + = −2 k− (2k 1), b 1, − + k a+1 1+( 1)a b − s 2k+1 a k s = − b ∑ ( 1) k+s− b−1 , b Z [2,k],  2( 1) 2 s=1 − − ∈ ∩ (3.57)  − j k +  a 1+ b k k a 1 + + ( 1) − ( 1) − − s 2k 1 a k s + b k k s + − b−k ∑ ( 1) k+s− b −k+1 ( 1) − b k+1 , b Z [k 1,2k 1], 2( 1) −2 s=1 − − − − ∈ ∩ −  − j k h i  where the ﬁrst row can be rewritten as 2k + 1, b = 1, + 2 1+( 1)b 1 b k (k!) (S2k 1)1,b − + , b Z [2,k], − =  b 2k 1 b b (3.58) k 1 + ( 1) 2 − ∈ ∩ ( 4) − (2k 1)  +− j bkk −  1 ( 1) − k +  − b k b k+1 , b Z [k 1,2k 1], ( 1) −2 − ∈ ∩ − − j k  which entails  k 1 + a k + b k = ( 4) − [1 ( 1) − ][1 ( 1) − ] k k (S2k 1)a,b − 2 − a k b −k (3.59) − (k!) − + − a k + 1 b k + 1 ( 1) 2 2 − − − for a,b Z [k + 1,2k 1]. ∈ ∩ − BROADHURST–ROBERTSQUADRATICRELATIONS 31

(b) If a Z [1,k], then ∈ ∩ (a 1)!(2k + 2 a)!(S2k)a,b 2k−1 k − 2 ( 1) [1 + (2k + 1)δa,1] − − + + k+2 a 1+( 1)a b 1 − s 2k+2 a k+1 s − [1 + (2k + 1)δ , ] ∑ ( 1) − − , b Z [1,k + 1], a + b 1 1 b 1 k+s b 1 ( 1) 2 −2 − s=1 − − ∈ ∩ =  − j k j k + (3.60) + a+b k 1 k 2 a + + +  1 ( 1) − − ∑− ( 1)s 2k 2 a k 1 s + ( 1)b k 1 k s , b Z [k + 2,2k],  −a 1 + b k k+s− b −k − − b k ( 1) −2 −2 s=1 − − − − ∈ ∩ − j k j k h i  where the ﬁrst row can be rewritten as k+1 22kb 1+( 1)b ( ) − b , b Z [1,k + 1] (k!)2 b 1 +k 2k+2 b (S2k)1,b = ( 1) −2 − ∈ ∩ (3.61)  − j k 0, b Z [k + 2,2k].  ∈ ∩ If a,b Z [k + 2,2k], then (S2k)a,b = 0. ∈ ∩  Proof. In the proof below, we will repeatedly use the Wronskian´ determinant wm,s(u) introduced in (3.49). (a) Theinputsfrom Lemma3.4(a) and Lemma3.5 (m = 2k 1) allow us to set up the followingidentities: − 2 3k 2 1 k 2 k 3 4 − ( s ) − 2′ k 1((2s) ) 2k 1(u) − 2 |L | − w (u) − s(s+1) +− |L Λ | 2k 1,s ks 2 1 2kδs,k 2k 1 − ( 1) − − − k s + − (k 1)(k s) 1 k s = + = ( 1) 2 − − ReFk,s(u), ( 1) − 1, − k s 1 + − (3.62) − − (k 1)(k s 1) 1 k s = (( 1) 2 − − − ImFk,s(u), ( 1) − 1, − − − when u ((2s 2)2,(2s)2), and ∈ − 2 3k 2 1 k 2 k 3 4 − ( s ) − 2′ k 1((2s) ) 2k 1(u) − 2 |L | − w (u) − s(s+1) +− |L Λ | 2k 1,s ks 2 1 2kδs,k 2k 1 − ( 1) − − − k s + − (k 1)(k s) 1 k s = = ( 1) 2 − − ReFk,s(u), ( 1) − 1, − k s 1 + − − (3.63) − − (k 1)(k s 1) 1 k s = + (( 1) 2 − − − ImFk,s(u), ( 1) − 1, − − when u ((2s)2,(2s + 2)2). Here, the left-hand sides of (3.62) and (3.63) behave like a constant ∈ 2 2 multiple of 1 + [u (2s) ]fk,s(u) as u (2s) , where fk,s(u) is the real-analytic function mentioned in Lemma 3.4(a). Thus,− one can establish→ equalities in (3.62) and (3.63) after comparing leading order asymptotic behavior. Without loss of generality, we momentarily consider the cases where k is odd. We begin with the kth row of S2k 1 [cf. (2.36)], which concerns −

V W 2k 1,1(u),..., 2k 1,k(u),..., 2k 1,2k 1(u) = w2k 1,1(u). (3.64) F − F − F − − − = = 4k 3 + Setting s 1 in (3.62), while simplifying 2′ k 1(4) 2 − (k 1)(k 1)!k! and + k s +(k 1)(k s)+(k 1)+ks s(s 1) k 1 +k 1 |L − | − ( 1) −2 2 s=1 = ( 1) −2 , we may spell out − − − − − | − − 3 k 1 k 2k 1 − 1 2k 1(u) − 2 k 2 − ( 1) 2 − 1 − = |L | ( 1) w2k 1,1(u) − ReFk,1(u) Λ2k 1 − − (k 1)!(k + 1)! − − k 1 k 1 2k 1 − 1 −2 = 2 − ( 1) 2 − + ℓ + j k 1 ℓ − (2k 1) 2k 1,1(u) ∑( 1) − 2k 1,2 j+1(u) (k 1)!(k + 1)!  F − = − 2 j F − −  j 1  32 YAJUN ZHOU

k 2 + + j k 1 + k 1 ℓ ∑( 1) − 2k 1,2 j+k(u) , u (0,16), (3.65) = − 2 j + 1 2 j + 1 F −  ∈ j 1  which explains why  2k + 1, b = 1, + 1 ( 1)b k 1 (k 1)!(k 1)!(S2k 1)k,b − − b b−1 , b Z [2,k], − − =  2 ∈ ∩ (3.66) k 1 1 2( 1) − 22k 1( 1) −2  − j kb + − −  1 ( 1) k 1 + k 1 , b Z [k + 1,2k 1], −  − −b k b −k+1 b k+1 2( 1) −2 − − ∈ ∩ −  − j k h i when k is odd. Then we proceed to the (k 1)-st row. We observe that  − k 1 V ( 1) − W 2k 1,1(u),..., 2k 1,k 1(u),..., 2k 1,2k 1(u) − F − F − − F − − 3 kΛ 2k 1(u) 2 − 2k 1 |L − | − W 2k 1,1(u),..., 2k 1,k 2(u), 2k 1,k(u),..., 2k 1,2k 1(u) = F − F − − F − F − − 3 kΛ 2k 1(u) 2 − 2k 1 |L − | 1 − W 2k 1,1(u),..., 2k 1,k 2(u),ReFk,1(u), 2k 1,k+1(u),..., 2k 1,2k 1(u) = − − − − − − F F k 1 3F F (3.67) − k ( 1) 2 2k 1(u) 2 − Λ2k 1 − − − |L | can be developed into a Taylor series centered at u0 = 4 for u (4 ε,4), by virtue of (3.28). Such a function admits real-analytic continuation to ∈ − 1 W 2k 1,1(u),..., 2k 1,k 2(u),ImFk,1(u), 2k 1,k+1(u),..., 2k 1,2k 1(u) F − F − − F − F − − (3.68) k 1 3 k − ( 1) −2 2k 1(u) 2 − Λ2k 1 − |L − | − for u (4,16), by a comparison between (3.28) and (3.29). Since (2.38) and (2.39) tell us that ∈ k k+1 1 2 1 + 2 ImFk,1(u) ( 1) − (k 1) 2k 1,k 1(u) ( 1) 2k 1,2k(u) − − − F − − − F − + spanR 2k 1n, j(u),u (4,16) j Z ([1,k 2] [k 1,2k 1]) , o (3.69) ∈ {F − ∈ | ∈ ∩ − ∪ − } k 3 we can convert (3.68) into [cf. (3.62) and (3.63)] a Q-linear combination of 2k 1(u) − 2 w2k 1,1(u) k 3 |L − | − and 2k 1(u) − 2 w2k 1,2(u), in the following manner: |L − | − k 3 2k 1(u) − 2 |L − | W 2k 1,1(u),..., 2k 1,k 2(u),(k 1) 2k 1,k 1(u), 2k 1,k+1(u),..., 2k 1,2k 1(u) Λ2k 1 F − F − − − F − − F − F − − − + W 2k 1,1(u),..., 2k 1,k 2(u), 2k 1,2k(u), 2k 1,k+1(u),..., 2k 1,2k 1(u) F − F − − F − F − F − − 2k 1 k 1 1 2 ( 1) −2 k 1 = − − − (k 1)ImF1 (u) − ImF1 (u) (k 1)!(k+ 1)! − k,1 − k + 2 k,2 − k 1 k 1 2k 1 − 1 2 = 2 − ( 1) 2 − − j + k 1 k 2 ℓ − ∑ ( 1) (k 2) − − 2k 1,2 j+2(u) (k 2)!(k + 2)!  = − 2 j + 1 − 2 j + 1 F − −  j 0 k 1 −2 + + j + k 1 + k 1 k 2 + k 2 ℓ ∑( 1) (k 2) − (k 2) − 2k 1,2 j+k 1(u) . (3.70) − = − 2 j − 2 j − 2 j 2 j F − −  j 1  1 1 1 On the right-hand side of this expression, the last summands of ImFk,1(u) and k+2 ImFk,2(u) which k+1 k+1 k+1 k+2 2 ℓ 2 1 ℓ are ( 1) k+1 2k 1,2k(u) and ( 1) k+2 k+1 2k 1,2k(u), respectively cancel each other. Thanks to this− cancelationF of− terms undeﬁned− for u (0F,4),− the right-hand side of (3.70) real-analytically ∈ continues to u (0,16), and is compatible with our statement about (S2k 1)k 1,b in (3.57). ∈ − − BROADHURST–ROBERTSQUADRATICRELATIONS 33

Proceeding as the last paragraph, we obtain the following real-analytic continuations of (3.62):11 k 3 i 1 (a 1)!(2k + 1 a)! 2k 1(u) − 2 det(D − 2k 1, j(u))i Z [1,2k 2], j Z ([1,a 1] [a+1,k t] [k+1,2k 1+t]) − − |L − | F − ∈ ∩ − ∈ ∩ − ∪ − ∪ − 2k 1 + Λ kt t(t 1) a k + 2 − (1 2kδa,1) 2k 1 ( 1) −2 ( 1) 2 ( 1)a 1 − − − − − − k t + a+b k+1 a + − 1 ( 1) − s 2k 1 a k s = ∑ (1 + 2kδb,1) − ∑ ( 1) − − 2k 1,b(u) b − k + s b 1 F − b=1 " 2( 1) 2 s=t+1 − # − 2k 1+t ( 1)a 1 + ( 1)b+k k+1 a 2k + 1 a + − − − ( 1)s ∑ − b−k ∑ + − b=k+1 ( 2( 1) −2 s=t+1 − k s × − + k s b k+1 k s − ( 1) − 2k 1,b(u), (3.62*) × b k + 1 − − b k + 1 F − − − ) where t Z [0,k 1],a = k 1 t,u ((2t)2,(2t + 2)2). [If we set t Z [0,k 1],a = k t,u ((2t)2,(2∈t + 2)∩2) in− (3.62*), we− recover− ∈ the original (3.62).] Recursively∈ p∩erforming− real-analytic− ∈ continuations with the aid of (3.63), down-stepping √u by an amount of 2 every time, we can further extend the validity of (3.62*) to t Z [0,k 1],a Z [1,k t],u ((2t)2,(2t + 2)2). In particular, ∈ ∩ − ∈ ∩ − ∈ setting t = 0 in (3.62*), one can verify the expression (S2k 1)a,b in (3.57), for each a Z [1,k]. One can readily generalize the procedures in the last two− paragraphs [which essentially∈ ∩ revolve around real-analytic continuations of w2k 1,s(u)] to even numbers k, thus establishing (3.57) in its entirety. − Plugging two combinatorial identities k 2k k s 2k 2k 1 k 1 ∑( 1)s − = − − , (3.71) − k + s n − 2k n k n s=1 − k 2k k + s 2k 2k 1 k ∑( 1)s = − (3.72) − k + s n − 2k n k n s=1 − into (3.57) for the expression (S2k 1)1,b, we obtain (3.58). − Let k Z 2 and r Z [k + 1,2k 1]. Through elementary manipulations of rows and columns in ∈ ≥ ∈ ∩ − (k 1)(2k 3) (k 1)2 a determinant (bearing in mind that detβ2k 2(u) = 2 − − u − ), one can identify | − | | | 3 k r+1 V 2k 1(u) − 2 ( 1) W 2k 1,1(u),..., 2k 1,r(u),..., 2k 1,2k 1(u) |L − | ( − F − F − F − − 1 + ( 1)r k k − V − r k W 2k 1,1(u),..., 2k 1,r(u),..., 2k 1,2k 1(u) (3.73) − − r k + 1 F − F − F − − ( 1) 2 − ) − for u (0,4) with a constant multiple of ∈ k 3 k − 2 2 ιk 1(u) ι´k 1(u) ∏[(2 j) u] det − − , (3.74) = − κk 1,r(u) κ´k 1,r(u) j 1 ! − − where the blocks are designed as follows: = b = ´ b (1) We have ιk 1(u) ( 2k 1,a+1(u))1 a,b k 1 and ι´k 1(u) (√u 2k 1,a+1(u))1 a,b k 1; − F − ≤ ≤ − − F − ≤ ≤ − = b + 2k b √u (2) The first row of κk 1,r(u) [resp. κ´k 1,r(u)] is given by (κk 1,r(u))1,b π 2k 1,1(u) 2k+1 2k 1,2(u)log 2 − − − F − F − = ´ b + 2k ´ b √u resp. (κ´k 1,r(u))1,b √u π 2k 1,1(u) 2k+1 2k 1,2(u)log 2 ; − F − F − h i 11Similar to Proposition 3.2, here we require that all the column indices be arranged in increasing order (even when they are not consecutive integers). 34 YAJUN ZHOU b + b (3) The next k 3 rows in κk 1,r(u) [resp. κ´k 1,r(u)] form a subset of π 2k 1,a(u) 2k 1,a k+2(u) − − − F − F − − √u + ´ b + ´ b √u log 2 1 b k 1 a Z [k 1,2k 2] resp. √u π 2k 1,a(u) 2k 1,a k+2(u)log 2 1 b k 1 a ≤ ≤ − ∈ ∩ − F − F − − ≤ ≤ − ∈ Z [k + 1,2k 2] ; h ∩ − (4) The last row of κk i1,r(u) is specified by [cf. (2.51) and (2.52)] − (k 2)/2 b + u − Ok(u), k 1 2Z>0, (k 2)/2 b ∈ = (κk 1,r(u))k 1,b = u − Ek (u), k 2Z>0,r 2k 1, (3.75) − −  ∈ 6 −  (k 2)/2 b + b √u = u − π 2k 1,2k 2(u) 2k 1,k(u)log 2 , k 2Z>0,r 2k 1, F − − F − ∈ − while the last row of κ´k 1h,r(u) is specified by [cf. (2.51′) andi (2.52′)]  − (k 1)/2 ´b + u − Ok(u), k 1 2Z>0, (k 1)/2 ´b ∈ = (κ´k 1,r(u))k 1,b = u − Ek (u), k 2Z>0,r 2k 1, (3.75′) − −  ∈ 6 −  (k 1)/2 ´ b + ´ b √u = u − π 2k 1,2k 2(u) 2k 1,k(u)log 2 , k 2Z>0,r 2k 1. F − − F − ∈ − = h i [For k 2, the instructions in items (3) and (4) must be ignored, whereupon we are reduced to the case already treated in Corollary 2.11.] According to the arguments in Proposition 2.10(a), the expression in (3.74) defines a holomorphic function in a non-void C-neighborhood of u = 0, hence a member of spanC 2k 1,a(u) a Z [2,k] . This completes the proof of (3.59). (b) By (3.31), the expression{F − | ∈ ∩ } k s 1 2 k 1 n 1 ( 1) − − i n 2 iπ n [u (2s 1) ] − D (u) + − − [D Gk,s(u)] log u (2s 1) D Lk,s(u) (3.76) − − I2k,s π(1 i) | − − | − 2 − − 2 Z 0 for n Z [0,2k 1] is equal to a convergent Taylor series in powers of [u (2s 1) ] ≥ for u ((2s ∈1)2 ∩ε,(2s −1)2), where ε > 0. Accordingly, for each s Z [1,k], there− exists− a non-vanishing∈ − − − ∈ ∩ constant ck,s that allows us to identify the following formula k 1 [ 2k(u)] − W[ 2k,1(u),..., 2k,k s(u), 2k,k+2 s(u), L F F − F − + 2 2k,k+2(u),..., 2k,2k 1+s(u)] ck,sw2k,s(u)log u (2s 1) (3.77) F F − | − − | 2 Z 0 2 2 with another convergent Taylor series in powers of [u (2s 1) ] ≥ for u ((2 s 1) ε,(2s 1) ). k 1 − − ∈ − − − In the light of this, the function [ 2k(u)] − w2k,s(u) must be a constant multiple of Gk,s(u), and the constant in question can be determinedL through Lemma 3.5 (m = 2k). In fact, the arguments in the last paragraph lead us to k 1 4k 2k 1 k(u) 2 (2s 1) Gk,s(u) |L2 | − w (u) = − − (3.78) Λ 2k,s k s + (2k s)(s+1) 2 2k −( 1) −2 − 2 k′ ((2s 1) ) − L − for u ((max 0,2s 3 )2,(2s + 1)2), s Z [1,k]. Setting s = 1 in the formula above, we recover ∈ { − } ∈ ∩ the expression for (S2k)k+1,b,b Z [1,2k], as claimed in (3.60). Setting s = 2 in the last dis- ∈ ∩ played formula, and performing real-analytic continuation of Lk,1(u) [which allows us to trade k, k+ (u),u (1,9) for k,k+ (u),u (0,1), up to some trailing terms contributed by Gk, (u)], F2 2 1 ∈ F2 1 ∈ 1 we can verify the formula for (S2k)k,b,b Z [1,2k], as claimed in (3.60). Subsequently, after finitely many steps of real-analytic continuations,∈ ∩ we can check the combinatorial expression for (S2k)a,b,a Z [2,k],b Z [1,2k]. ∈ ∩ ∈ ∩ We can directly compute (S k) ,b,b Z [2,k] from our knowledge of (S k)b, ,b Z [2,k]. 2 1 ∈ ∩ − 2 1 ∈ ∩ We can deduce (S2k)a,b = 0 for a,b Z ( 1 [k + 2,2k]) from Proposition 2.10(b). ∈ ∩ { } ∪ At this point, we have verified all the statements in Theorem 1.1, effectively proving the following matrix inversion for u (0,1): ∈ 1 T [W (u)]− = m(u) S W (u)V (u), (3.79) m |L | m m m BROADHURST–ROBERTSQUADRATICRELATIONS 35 which in turn, can be recast into the following alternative formulation of quadratic relations for off-shell Bessel moments: 1 T = Sm− Wm(u)Vm(u)Wm(u) . (3.80) m(u) |L | 1 To prepare for an inductive computation of Sm− in Proposition 4.2, we wrap up this section with a A B = Sm Sm combinatorial analysis of Sm C D in the proposition below. To ease computation, we comb Sm Sm Sm m m with a reshuffling-rescaling matrix ξm Q × whose elements are ∈ m δa,b 1, a Z 1, , − ∈ ∩ 2 (ξ ) := m+2 δ , a = m + 1, (3.81) m a,b  m+1 1,b 2 m δa,b, a Z + 2,m . ∈ ∩ 2 SA SB As a result, the block partition of S := (ξ 1)TS ξ 1 = m m has the following structure: m m− m m− S◦ C S◦ D ◦ m m ◦ ◦ A = A m (Sm)a,b (Sm)a+1,b+1, a,b Z 1, 2 , ◦ B = B ∈ ∩ m m+1 (Sm)a,b (Sm)a+1,b 1, a Z 1, ,b Z 1, , (3.82)  − ∈ ∩ 2 ∈ ∩ 2  ◦ D = D m+1 (Sm)a,b (Sm)a 1,b 1, a,b Z 1, 2 , ◦ − − ∈ ∩ B = m+1 C with parity (Sm)a,b ( 1) (Sm)b,a, where extensions of (1.13)–(1.15) lead us to ◦ − ◦ m + 2 1 a m − s m+1 a +1 s ⌊ ⌋∑ ( 1) m+1 2 − + a+b+m+1 −+s ⌊ ⌋b A 1 ( 1) s=1 − 2 (S )a,b = − , (1.13◦) m 1 m + a+1 + b+1 1+( 1)m m ◦ 2 − a!(m 1 a)! − 1 − ( 1) 2 2 − 4 −⌊ 2 ⌋− − j k m + 2 1 a + − s m+1 a m +1 s b m 1 +s ⌊ ⌋∑ ( 1) m+1 2 − ( 1) 2 + a+b+m −+s ⌊ ⌋b b B m+1 C 1 ( 1) s=1 − 2 − − (S )a,b = ( 1) (S )b,a = − , m m 1 m + a+1 1+( 1)hm b 1 1+( 1)m m+1 i ◦ − ◦ 2 a!(m 1 a)! am+ − + + − − − ( 1) 2 − 4 −2 4 − 2 j k j k − (1.14◦) and + m+1 m 1 1 a b m+1 m+1 1 + ( 1) ( 4) 2 [1 ( 1) ][1 ( 1) ] (SD) = − 2 2 (1.15 ) m a,b − − + 2 − − a 1 +−b −1 ◦ ◦ 2 m 1 −2 −2 a b 2 ! ( 1) − for a,b Z 0. (It is still understood that empty sums vanish and certain combinatorial expressions are zero.) ∈ ≥

Proposition 3.7 (Recursions for S ). (a) For k Z>1 and a,b Z [1,k], we have m ∈ ∈ ∩ + ◦ A = + A A a(2k 1 b)(S2k)a,b 2(2k 1)(S2k 1)a 1,b 4(S2k 2)a 1,b 1, (3.83) − − − − − − − + ◦ B = + ◦ B ◦ B a(2k 1 b)(S2k)a,b 2(2k 1)(S2k 1)a 1,b 4(S2k 2)a 1,b 1. (3.84) − ◦ ◦ − − − ◦ − − − (b) For k Z>1, we have ∈ 2k+1 B 1 1 1 T 1 T = 2 S2k ϕ2−k 1αk− S2k 1(αk− ) (ϕ2−k 1) ◦ 2 B (3.85) − − − 2k+1 S2k 2 − ◦ − in block diagonal form, along with a recursion B + a k+1 B 4(S2k)a,b B 1 ( 1) a+1 (S2k+2)1,b+1 = (S ) + + (3.86) + ◦ + 2k+2 a 1,b 1 − a ◦+ (2k 2 a)(2k 2 b) ◦ − ( 1) 2 2k 2 a − − − − for a,b Z [1,k]. ∈ ∩ 36 YAJUN ZHOU

Proof. (a) Define + m + m 1 a 2 1 s m+1 −+s ⌊ ⌋ − S A = 2 b = S A m,a,b,s : m,b,a,s, (3.87) a!(m +1 a)! − m+1 + b 2 s B A ( 1) b S := S − (3.88) m,a,b,s m,a,b,s − a!(m + 1 a)! − for all the integers m,a,b, s that make these expressions meaningful. By direct computation, we obtain + S A = + S A + S A a(2k 1 b) 2k,a,b,s (2k 1) 2k 1,a 1,b,s 2k 2,a 1,b 1,s, (3.89) − − − − − − for integers k,a,b, s that make each summand well-defined. Multiplying through ( 1)s and summing − over s Z>1, we arrive at the recursion (3.83) that involves ∈ ∑∞ sS A + a+b+m+1 ( 1) m,a,b,s A 1 ( 1) s=1 − (S )a,b = − . (3.90) m 1 m a+1 + b+1 1+( 1)m m ◦ 2 − − 1 ( 1) 2 2 − 4 −⌊ 2 ⌋− − j k Here, the infinite series over s truncates after finitely many terms, in accordance with the previous A definition of (Sm)a,b in (1.13◦). Likewise, one◦ can sum over + S B = + S B + S B a(2k 1 b) 2k,a,b,s (2k 1) 2k 1,a 1,b,s 2k 2,a 1,b 1,s (3.91) − − − − − − to arrive at (3.84). SA SB TA TB (b) We partition the matrix α 1S (α 1)T = α 1 2k 1 2k 1 (α 1)T = 2k 1 2k 1 into four blocks. k− 2k 1 k− k− SC − SD − k− TC − TD − − 2k 1 2k 1 2k 1 2k 1 A − − 1 1 − 1− T 1 T Here, the matrix T2k 1 also forms the top-left k k block of ϕ2−k 1αk− S2k 1(αk− ) (ϕ2−k 1) . − D × = D − − − By definition of αk [see (2.23)], we have T2k 1 S2k 1. Thanks to (3.58), we can evaluate the first− row and− the first column in the top-left k k block A A = A = 2k+×1 B T2k 1 in closed form. This allows us to verify the relations (T2k 1)1,1 : (S2k 1)1,1 2 (S2k)1,1 − A A = A + B = 2k+1 B − − ◦ and (T2k 1)1,b (T2k 1)b,1 : (S2k 1)1,b (S2k 1)1,b 1 2 (S2k)1,b,b Z [2,k] explicitly. In gen- − ≡ − − A +− B − = a◦ B ∈ ∩ eral, we may construct a sum rule (S2k 1)a,b (S2k 1)a,b 1 2 (S2k)a,b for a,b Z [2,k], using the k s + k s = −k+1 s − − ◦ ∈ ∩ combinatorial identity b−1 −b b− . − A = A + B + C + By definition, for a,b Z [2,k], we have (T2k 1)a,b : (S2k 1)a,b (S2k 1)a,b 1 (S2k 1)a 1,b D C ∈ ∩ = C + −D − − − −A − = (S2k 1)a 1,b 1 and (T2k 1)a 1,b : (S2k 1)a 1,b (S2k 1)a 1,b 1, which together imply (T2k 1)a,b a B− −+ − C − − − − − − − 2k+1 B − 2 (S2k)a,b (T2k 1)a 1,b. In order to identify the last expression with 2 (S2k)a,b [as claimed in ◦ − − C = 2k+1◦ a B the right-hand side of (3.85)], we will prove the relation (T2k 1)a 1,b 2 − (S2k)a,b in the next paragraph. − − ◦ We aim for an identity that is slightly stronger than what has been just claimed, namely + B + D = 2k 1 b B (S2k 1)a 1,b (S2k 1)a,b − (S2k)a,b (3.92) ◦ − − ◦ − 2 ◦ for all the possible values of k Z>0 and a,b Z that make both sides well-defined. For k = 1, we ∈ ∈ can check (3.92) by direct computation. The same is true for a,b Z 0 and arbitrary k Z>0. For larger integer values of k, we will verify (3.92) by recursion. Summing∈ ≤ over ∈ + S A = + S A + S A b(2k 2 a) 2k+1,a,b,s (2k 2) 2k,a,b 1,s+1 2k 1,a 1,b 1,s, (3.93) − − − − − while taking care of contributions from the boundary value (for s = 1) and a combinatorial identity m+1 + m+1 a +1 m 1 m+1 − 2 ∞ + 1 b +1 b ( 1)s(S B S A ) = 2 − 2 , (3.94) ∑ m,a,b,s m,a,b,s + b + = − − m 1 a b ( 1) a!(m 1 a)! s 1 − − − − BROADHURST–ROBERTSQUADRATICRELATIONS 37

we reach a result + B + B + B b(2k 2 a)(S2k+1)a,b 4(k 1)(S2k)a,b 1 4(S2k 1)a 1,b 1 − − − − − − = + D ◦ D ◦ ◦ (a 1)b(S2k+1)a+1,b 4(S2k 1)a,b 1. (3.95) ◦ − ◦ − − We then combine the equation above with (3.84) into a recursion + B D 2k 3 b B (a + 1)b (S + )a,b + (S + )a+1,b − (S + )a+1,b 2k 1 2k 1 − 2 2k 2 ◦ ◦ ◦ + = B + D 2k 2 b B 4 (S2k 1)a 1,b 1 (S2k 1)a,b 1 − (S2k)a,b 1 , (3.96) − − − − 2 − ◦ − ◦ − ◦ which proves (3.92). 2k+1 a B = C = 2k+1 a A The transpose of (3.92) brings us 2 − (S2k)a,b (T2k 1)a 1,b 2k+−1 (T2k 1)a,b for a,b Z C = 2k+1 a A ◦ − − − ∈ ∩ [2,k], while (T2k 1)a 1,1 2k+−1 (T2k 1)a,1 also holds for a Z [2,k]. Thus, the top-left, top-right − − 1 1 − 1 T 1 T ∈ ∩ and bottom-left blocks of ϕ2−k 1αk− S2k 1(αk− ) (ϕ2−k 1) assume the forms described in the right- hand side of (3.85). − − − a 1 1 ( 1)a B D = −2 Eliminating (S2k 1)a 1,b from (3.84) and (3.92), while noting that (S2k 1)a,b ( 1) − 2− 2k+1 b k B ◦ − − ◦ − − 2k+−1 a (S2k)1,b, we arrive at (3.86). ◦ 1 1 With the aid of (3.86), one can verify that the bottom-right (k 1) (k 1) block of ϕ2−k 1αk− S2k 1 1 T 1 T 2 B − × − − − (αk− ) (ϕ2−k 1) is indeed equal to 2k+1 S2k 2. − − ◦ −

4. BROADHURST–ROBERTS QUADRATIC RELATIONS 1 In §4.1, we characterize Sm− in terms of Bernoulli numbers, thereby connecting sum rules of Feynman diagrams to the Betti matrix Bm 2 in the on-shell limit. In the meantime, we will also represent the de − Rham matrix Dm 2 for Broadhurst–Roberts quadratic relations in the form declared in Theorem 1.2. − + + BR = m 1 m 1 T In §4.2, we explore the Broadhurst–Roberts varieties Vm+2 : X m+1 C 2 × 2 X m+1 DmX m+1 2 ∈ 2 2 − = n Bm 0 , replacing the Broadhurst–Roberts period matrices Pm with formal variables. We demonstrate BR algebraicallyo that V2k+1 imposes a constraint on certain minor determinants of Xk, in the form of a re- ﬂection formula (1.21). The purpose of §4.3 is twofold. First, we display the diversity of de Rham representations in the Wronskian´ framework, along with formal proofs. Second, we conclude with an open-ended discussion on other representations of the de Rham matrix Dm outside the Wronskian´ framework.

4.1. Quadratic relations for on-shell Bessel moments. It is now clear from Proposition 2.7 that T T = T 1 2k 1(1) αk W2k 1(1)V2k 1(1)W2k 1(1)αk αk S2−k 1αk (4.1) |L − | − − − − can be rewritten as T 1 ´ PT αk S2−k 1αk P2k 1 P2k 1 1 T 1 2k 1 = − − (β ) V (1)β − , (4.2) − 2−k 1 2k 1 2−k 1 T T 2k 1(1) P2k 3 − − − P´ P |L − | ! 2k 1 2k 3! − − − − − whose bottom-right (k 1) (k 1) block is partly responsible for Theorem 1.2 [m = 2k 1 in (1.20) and (1.22)]. Proposition 2.8− then× brings− us a further reﬁnement: −

T T 1 T ϕ2k 1αk S2−k 1αkϕ2k 1 P2k 1 1 T 1 T 1 1 P2k 1 = − (ϑ ) (β ) V (1)β ϑ − . − −(1) − 2−k 1 2−k 1 2k 1 2−k 1 2−k 1 T 2k 1 P2k 3! − − − − − P2k 1! |L − | − − − − (4.2∗) 38 YAJUN ZHOU

T 1 1 T Here, the bottom right (k 1) (k 1) block in αk S2−k 1αk [resp. (β2−k 1) V2k 1(1)β2k 1] is unaffected −T × − 1 T − − − − 1 by left multiplication of ϕ2k 1 [resp. (ϑ2−k 1) ] and right multiplication of ϕ2k 1 (resp. ϑ2−k 1). To obtain similar block (tri)diagonalization− − for Theorem 1.2 (m = 2k), we need− to expend− a little more effort, in the lemma below.

Lemma 4.1 (Margin behavior). For each given k Z>0, the 2k-th row and the 2k-th column of ∈ 1 T 1 lim 2k(u) (β2−k ) V2k(u)β2−k (4.3) u 1 |L | → − are filled with zeros. Proof. We first note that 0 2k 1 T 1 2 k 1 2k 2 (D I0(√ut),..., D − I0(√ut)) = [β2k(u)]− (I0(√ut), I0(√ut)t ...,( 1) − I0(√ut)t − , − − 3 k 1 2k 1 T √uI1(√ut)t, √uI1(√ut)t ...,( 1) − √uI1(√ut)t − ) (4.4) − − 1 entails a characterization of the last column in β2−k : ( 1)k 1 β 1 = − (β2−k )a,2k −2k 1 δa,2k. (4.5) 2 − k 1 k 1 k 1 β 1 = ∑2k ( 1) − = ( 1) − = ( 1) − Then we observe that (V2k(u)β2−k )a,2k t=1(V2k(u))a,t −22k 1 δt,2k −22k 1 (V2k(u))a,2k −22k 1 δa,1 and k 1 − − − β 1 T β = ( 1) − β 1 T β = ((β2−k ) V2k(u)β2k)a,2k −22k 1 δa,1 follow immediately. By skew symmetry, we have ((β2−k ) V2k(u)β2k)2k,b k − ( 1) δ , so the right and bottom margins of the matrix lim (u) (β 1)TV (u)β are indeed filled 2−2k 1 1,b u 1− 2k 2−k 2k 2k with− zeros. → |L | According to the result from the last lemma, we have (cf. Proposition 2.7) T T 1 αk ψk S2−k ψkαk = T T T lim 2k(u) αk ψk W2k(u)V2k(u)W2k(u)ψkαk u 1 |L | → − P P´ PT = 2k 2k 1 T 1 T 2k lim 2k(u) ̺k (β2−k ) V2k(u)β2−k ̺k , (4.6) u 1 |L | ´ T T P2k 2! → − P2k P2k 2! − − − − which is partly responsible for Theorem 1.2 [m = 2k in (1.20) and (1.22)]. Its block diagonalized counterpart reads (cf. Proposition 2.8) T T T 1 ϕ2kαk ψk S2−k ψkαkϕ2k P PT = 2k 1 T 1 T 1 T 1 2k (ϑ2−k ) lim 2k(u) ̺k (β2−k ) V2k(u)β2−k ̺k ϑ2−k . (4.6∗) P u 1 |L | T 2k 2! → − P2k 2! − − − − 1 To complete the proof of Theorem 1.2, we need to express the entries of Sm− via Bernoulli numbers, in the next proposition. 1 Proposition 4.2 (Bernoulli representation of Sm− ). Let Bn be the Bernoulli numbers generated by (1.20). A B BA B B = Sm Sm B = m m 12 For all m Z>0, the inverse of Sm C D is given by m BC BD for ∈ Sm Sm m m + A (m 1)! δmin a,b ,1Bm+3 a b B = { } − − ( m)a,b : m+2 a b 1+a ( 1)m(1 a) m+1 (4.7) m + 2 m 1 − − + − − − + 2 − ( 1) 2 2 2 − 12 = B = 1 = A = B 1 = B A 1 = 0 8 It is understood that (S1)1,1 9,( 1)1,1 9 , and S2 S2 ( 2)− ( 2 )− 8− 0 . BROADHURST–ROBERTSQUADRATICRELATIONS 39 where a,b Z 1, m + 1 , ∈ ∩ 2 a+b 1 ′ δ B + (m + 2 a)!(m b )! m+2 a,1 m 2 a b′ B B = m+1 BC := ′ − − − ( m)a,b′ ( 1) ( m)b′,a − − + (4.8) + m 1a b′ + + m − (m 2 a b′)! m 1 −2 − b′ 2 − − 2 − ( 1) ⌊ ⌋ − j k m m 1 where a Z 1, + 1 ,b Z 1, − , and ∈ ∩ 2 ′ ∈ ∩ 2 D (m a′)!(m b′)! (m a′ b′)Bm+1 a b B := − ′− ′ ( m)a′,b′ − − − − (4.9) + m a′ b′ + + m 1 (m 1 a′ b′)! m 1 − 2− a′ 2− − − 2 − ( 1) − j k m 1 where a′,b′ Z 1, 2− . Consequently,∈ ∩ both (4.2) and (4.2 ) [resp. (4.6) and (4.6 )] account for the representations of the de ∗ ∗ Rham matrix D2k 3 [resp. D2k 2] per (1.17) and the Betti matrix B2k 3 [resp. B2k 2] per (1.19). − − − − Proof. By direct computation, one can check that

4 k 1 2 ( 1) − B ϕT αTB α ϕ = 2−k+1 2k 1 , (4.10) 2k 1 k 2k 1 k 2k 1 − 2 + k 1 − − − 2 (2k 1)( 1) − B2k 3! − − and 4 k 1 2 ( 1) − B ϕT αTψTB ψ α ϕ = 2−k+2 2k , (4.11) 2k k k 2k k k 2k 2 + k 1 2 (2k 2)( 1) − B2k 2! − − m 1 m 1 where the entries of Bm 2 Q 2− × 2− are given by the last line of (1.19). (This is essentially the − original form of Betti matrix∈ conjectured by Broadhurst–Roberts [13, (5.4)], except that all its rows and 1 = B columns are placed in reverse order.) Therefore, so long as one can verify that Sm− m for m Z>1, one T = ∈ can deduce all the Broadhurst–Roberts quadratic relations PmDmPm Bm from the bottom-right blocks of (4.2∗) and (4.6∗). B = B T = Recall the reshuffling-rescaling matrix ξm from (3.81). In the block partition of m : ξm mξm B B ◦ m m m B B = BC B◦ D , the top-left 2 2 block is filled with zeros, while the remaining blocks are ( m)a,b m m × ◦ ◦ m+◦1 B C = B B m m+1 ( 1) ( m)b,a ( m)a+1,b 1fora Z 1, 2 ,b Z 1, 2 and − ◦ − ∈ ∩ ∈ ∩ + B D = B D + δa,1 δb,1 ( m)a,b ( m)a 1,b 1 1 (4.12) − − m + 2 a b ◦ − − m+1 B B B D B D for a,b Z 1, 2 . Here, expressions like ( m)a+1,0, ( m)0,b 1 and ( m)a 1,0 are defined by the right-hand∈ sides∩ of (4.8) and (4.9). − − In what follows, we will demonstrate inductively that SBBC SAB B + SBB D S B = m m m m m m m m ◦ DB◦ C ◦ CB◦ B + ◦ DB◦ D (4.13) ◦ ◦ Sm m Sm m Sm m ◦ ◦ ◦ ◦ ◦ ◦ is the m m identity matrix Im, for all positive integers m. Suppose that we have verified the aforemen- × tioned claim up to a certain odd number m = 2k 1. We will show that the same is true for S2kB2k and − ◦ ◦ S2k+1B2k+1. ◦ ◦ Inductive evaluation of S2kB2k D ◦ ◦ D BC We note that S2k is filled with zeros. So must be the case for S2k 2k, which forms the bottom-left ◦ ◦ ◦ block of S2kB2k. ◦ ◦ 1 = B B 1 = Paraphrasing our induction hypothesis S2−k 1 2k 1 with (3.85) and (4.10), we have (S2k)− 3 k 1 BC = 3 k 1 − − B B◦ C = 2 ( 1) − B2k 1. Since 2k 2 ( 1) − B2k 1, we can check that the top-left block satisfies S2k 2k Ik − − ◦ − − C B B + D B D = BC B T = ◦ ◦ and the bottom-right block has the same behavior: S2k 2k S2k 2k ( 2kS2k) Ik. ◦ ◦ ◦ ◦ ◦ ◦ 40 YAJUN ZHOU A B B + B B D In the next three paragraphs, we show that the top-right block S2k 2k S2k 2k has vanishing entries. This task can be simplified by a similarity transformation ◦ ◦ ◦ ◦ A B A + B BC A B B + B B D T B T 1 = B = S2k 2k S2k 2k S2k 2k S2k 2k ξ2kS2k 2k(ξ2k)− S2k 2k . (4.14) Ik ◦ ◦ Here, the only possible exceptions to (S2kB2k)p,q = δp,q occur at positions p Z [2,k + 1],q Z + A B A + B BC = 2k+1 A B B + B B∈D ∩ ∈ ∩ ( 1 [k 2,2k]), where we have (S2k 2k S2k 2k)a+1,1 2k+2 (S2k 2k S2k 2k)a,1 for a Z [1,k] { } ∪ A B B + B B D = A B B + B B D ◦ ◦ ◦ ◦ ∈ ∩ and (S2k 2k S2k 2k)a+1,b 1 (S2k 2k S2k 2k)a,b for a Z [1,k],b Z [2,k]. A −B B +◦ B ◦B D ◦ ◦ = ∈ ∩ ∈ ∩ First, we show that (S2k 2k S2k 2k)a+1,b 1 0 for a Z [1,k] and b Z [3,k]. Contract- − B 2k+1 n 1+2kδn,0 B ∈ ∩ ∈ ∩ ing the recursions (B ) + , = − (B ) + , for n Z [0,k],b Z [3,k] and 2k n 1 b 1 2 1+(2k+1)δn,0 2k 1 n 1 b 2 + − + − − ∈ ∩ ∈ ∩ B B = [2k (1 b)(1 δn,1)](2k 1 n) B B ( 2k)n+1,b 1 − −4 − ( 2k 2)n,b 2 for n Z [1,k],b Z [3,k] with a variation on (3.83), namely− − − ∈ ∩ ∈ ∩ + A + A A a(2k 1 n)(S2k)a+1,n+1 2(2k 1)(S2k 1)a,n+1 4(S2k 2)a,n − = − − (3.83′) 1 + (2k + 1)δn,0 (1 + 2kδa,1)(1 + 2kδn,0) − [1 + (2k 1)δa,1][1 + (2k 1)δn,1] − − for a Z [1,k],n Z [0,k], we compute ∈ ∩ ∈ ∩ + (SA B B ) + (SA B B ) A B B = 2k 1 2k 1 2k 1 a,b 2 2k 1 b 2k 2 2k 2 a,b 2 (S2k 2k)a+1,b 1 − − − − − − − − a 1 + 2kδa,1 − a 1 + (2k 1)δa,1 − A B B 2k b (S2k 2)a,1( 2k 2)1,b 2 + − − − − (4.15) a 1 + (2k 1)δa,1 − A = A = for a Z [1,k],b Z [3,k], where we have exploited the facts that (S2k 1)a,k+1 (S2k 2)a,0 0 [by extensions∈ ∩ of (1.13)].∈ Likewise,∩ we can argue that − − + (SB B D ) + (SB B D ) B B D = 2k 1 2k 1 2k 1 a,b 2 2k 1 b 2k 2 2k 2 a,b 2 (S2k 2k)a+1,b 1 − − − − − − − − a 1 + 2kδa,1 − a 1 + (2k 1)δa,1 − + B B D 2k 1 b (S2k 2)a,0( 2k 2)0,b 2 − − − − , (4.16) − a 1 + (2k 1)δa,1 − for a Z [1,k],b Z [3,k]. Adding up (4.15) and (4.16), while falling back on our induction hy- ∈ ∩ 1 ∈= B∩ 1 = B A B B + B B D = A B B + pothesis that S2−k 1 2k 1 and S2−k 2 2k 2, we arrive at (S2k 2k S2k 2k)a+1,b 1 (S2k 2k B B D = 2−k+1 A−B B + B −B D − 2k+1 b A B B + B B D − ◦ = ◦ S2k 2k)a,b + (S2k 2k S2k 2k)a,b 2 + − (S2k 2 2k 2 S2k 2 2k 2)a,b 2 0 for a(1 2kδa,1) − − a[1 (2k 1)δa,1] − a◦ Z◦ [1,k],b Z [3,k]. − − − − − ∈ ∩ ∈ ∩ A B B + B B D = = = Second, we prove that (S2k 2k S2k 2k)2,1 0. Adding up analogs of (4.15) and (4.16) for a 1,b 2, we obtain k A B B + B B D = A B B + B B D (S2k 2k S2k 2k)2,1 ∑[(S2k 1)1,n( 2k 1)n,0 (S2k 1)1,n( 2k 1)n,0] n=1 − − − − k 2k 1 A B B + B B D − ∑[(S2k 2)1,n( 2k 2)n,0 (S2k 2)1,n( 2k 2)n,0]. (4.17) − 2k n=2 − − − − B = A B B + Through the explicit formula in (3.58), one can check that (S2k 1)1,k 0 and (S2k 1)1,n( 2k 1)n,0 B B D = B BC − − − (S2k 1)1,n 1( 2k 1)n 1,0(1 δn,1) (S2k)1,n( 2k)n,1 for n Z [1,k], so the first displayed sum in (4.17) − − B B− C − = − ◦ ◦ ∈ ∩ B = A B B = evaluates to (S2k 2k)1,1 1; referring back to (3.61), one sees that (S2k 2)1,n 0 and (S2k 2)1,n( 2k 2)n,0 2k B ◦ B◦ C − − − 2k 2k 1 (S2k 2)1,n 1( 2k 1)n 1,1 for n Z [2,k], so the second displayed sum in (4.17) evaluates to 2k 1 −B ◦B−C − =◦ 2k− − ∈ ∩ − (S2k 2 2k 2)1,1 2k 1 . Thus, the claimed identity is a consequence of our induction hypothesis. ◦ − ◦ − − A B B + B B D = Third, we demonstrate that (S2k 2k S2k 2k)a,b 0 for a Z [1,k],b 1,2 . If the integers a ◦ A◦ B◦ B ◦ = B B D∈ ∩= ∈{ } and b have the same parity, then (S2k)a,n( 2k)n,b (S2k)a,n( 2k)n,b 0 for all n Z [1,k]. Thus, we A B A + B BC = 2k+1 ◦ A B B ◦+ B B D ◦ = ◦ ∈ ∩ B = have (S2k 2k S2k 2k)2,1 2k+2 (S2k 2k S2k 2k)1,1 0. Now we have a matrix partition S2k 2k ◦ ◦ ◦ ◦ BROADHURST–ROBERTSQUADRATICRELATIONS 41

I2 U Ik 1 V where all the possibly non-vanishing entries of U and V are confined to their first columns. − Ik 1 − I 1 2 Such a matrix partition is invertible, and S = B U Ik 1 V is skew-symmetric. With a matrix 2−k 2k − − − Ik 1 − B A′ (BA′′ )T (BC′ )T 2k − 2k − 2k A ♭ T ♭ (k 1) (k 1) partition B2k = B ′′ (B ) where B Q is taken from the last (k 1) columns 2k − k k − × − BC′ B♭ BD ∈ − 2k k 2k ! BA′ + BA′′ T BA′′ T BC′ T+ BA′′ T 2k ( 2k ) U ( 2k ) ( 2k ) ( 2k ) V C 1 − − of B , we will spell out the skew symmetry of S = BA′′ (B♭)T . Here, 2k 2−k 2k − k BC′ B♭ B♭ BD B♭ 2k kU k 2k kV ! B♭ B♭ − − the matrix kV is equal to kv followed by the (k 2) identically vanishing columns to the right, − B♭ where v is the first column of V. In order to guarantee the skew symmetry of kV, one has to have B♭v = 0 Rk 1. We claim that detB♭ = 0, which can be seen from the relation between k ∈ − k 6 2 2k 1 1 B + detB♭ = − n! det a b (4.18) k (k 1)(2k 1) ∏ + | | 2 − − n=k+1 ! (a b)! 1 a,b k 1 ≤ ≤ −

and 2 2k 2 1 B + detBC = − n! det a b , (4.19) 2k 2 (k 1)(2k 3) ∏ + | ◦ − | 2 − − n=k ! (a b)! 1 a,b k 1 ≤ ≤ −

= B♭ 1 = the latter of which is equal to 1/ detS2k 2 > 0, by induction. Therefore, we conclude that v ( k)− 0 k 1 | ◦ − | 0 R − . Now that V is filled with zeros, the skew symmetry concerning the bottom-left and the top- ∈ 1 B♭ right blocks of S−2k further brings us a matrix kU filled with two columns and (k 1) rows of zeros, hence an identically vanishing U. − 1 = B = Thus far, we have verified Sm− m up to an even number m 2k. Inductive evaluation of S2k+1B2k+1 ◦ ◦ B BC = In view of (3.85) and (4.10), we only need to check that S2k+2 2k+2 Ik+1. B = ◦BC ◦ = 4 BC Noting that (S2k)a,0 0 by definition in (1.14◦), and that ( 2k)n,b (2k+2 n)(2k+2 b) ( 2k+2)n+1,b+1, we can deduce from◦ (3.86) the following result for a,b Z [1,◦k]: − − ◦ ∈ ∩ + k 2k 2 b B C δa,b = − ∑(S )a,n(B )n,b 2k + 2 a 2k 2k − n=1 ◦ ◦ + a k+1 B BC B C 1 ( 1) a+1 (S2k+2 2k+2)1,b+1 = (S + B + )a+1,b+1 − ◦ ◦ , (4.20) 2k 2 2k 2 − a 2k + 2 a ◦ ◦ ( 1) 2 − − B BC upon invoking our induction hypothesis that S2k 2k is the k k identity matrix. The last displayed ◦ ◦ B B×C = B BC = identity alleviates our workload to a demonstration that (S2k+2 2k+2)1,b+1 (S2k+2 2k+2)b+1,1 δb,0 for all b Z [0,k]. This can be achieved in three steps, in the◦ paragraphs◦ to follow.◦ ◦ ∈ ∩ B BC = First, we show that (S2k+2 2k+2)1,b+1 0 for b Z [1,k]. If b is an odd number in Z [1,k], then B BC =◦ ◦ + ∈ ∩ B BC = ∩ (S2k+2)1,n( 2k+2)n,b+1 0 for all n Z [1,k 1], so we have (S2k+2 2k+2)1,b+1 0. If b is an even number◦ in Z◦ [2,k], then we have a∈ recursion∩ ◦ ◦ ∩ + B BC B BC k(k 1)(S2k+2)1,n( 2k+2)n,b+1 (b 2)(S2k)1,n( 2k)n,b 1 = − − (2k◦+ 3)(2k + 2◦ b) 2◦ k + 2 ◦b − − + A B B + B B D (1 δn,1)(S2k)1,n 1( 2k)n 1,b (S2k)1,n( 2k)n,b, (4.21) − ◦ − ◦ − ◦ ◦ 42 YAJUN ZHOU + B BC = B = whose sum over n Z [1,k 1] results in (S2k+2 2k+2)1,b+1 0. This is because (S2k)1,k+1 0 by ∈ ∩ ◦ B ◦ B B◦C definition, and our induction hypothesis on S2k 2k leaves us an identity matrix S2k 2k along with a A B B + B B D ◦ ◦ ◦ ◦ k k block S2k 2k S2k 2k that is filled with zeros. × ◦ ◦ ◦ ◦ B BC = Second, we verify the identity (S2k+2 2k+2)1,1 1 by directly computing ◦ ◦ k k + k+1 k+1 B C B C 2( 1) (2k 3)! k+1 n B2k+3 n (S + B + )1,1 = ∑(S + )1,n+1(B + )n+1,1 = − ∑ − − , (4.22) 2k 2 2k 2 2k 2 2k 2 [(k + 1)!]2 2k + 3 n ◦ ◦ n=0 ◦ ◦ n=0 − and enlisting the help from a special case of the Saalschütz–Nielsen–Gelfand reciprocity relation [39, 40, 36, 27] revisited by Agoh–Dilcher [2]: α β α B + + β B + + β+1 j β 1 j + α+1 j α 1 j = α!β! ( 1) ∑ + + ( 1) ∑ + + + + , (4.23) − j=0 β 1 j − j=0 α 1 j (α β 1)! where α = β = k + 1, j = k + 1 n. B −BC = Third, we prove that (S2k+2 2k+2)b+1,1 0 for b Z [1,k]. In other words, we will show that B ◦ C ◦ 1 ∈ ∩ k in the block partition S + B + = , the column vector x R is in fact a zero vector. Taking 2k 2 2k 2 x Ik ∈ determinants on both sides◦ of◦ the block partition, we know that the symmetric matrix SB is invertible, 2k+2 B 1 = BC 1 ◦ BC = and its inverse (S2k+2)− 2k+2 x Ik is also symmetric. Taking another block partition 2k+2 T ◦ ◦ − T ◦ T bk+1 bk+1 C B 1 bk+1 bk+1x bk+1 where bk+1 = (B + )1,1, we may exploit the symmetry of (S + ) = − b + B♭ 2k 2 2k 2 − b B♭ x B♭ k 1 k+1 ◦ ◦ k+1 k+1 k+1 B♭ = k − to arrive at the following relation: k+1x 0 R . In view of (4.18) and (4.19), we can deduce the B♭ ∈ BC = B 1 = B♭ 1 = k invertibility of k+1 from the non-singular matrix 2k (S2k)− , so x ( k+1)− 0 0 R . This completes the induction. ◦ ◦ ∈ As a by-product of the proof above, we can quickly recover part of [25, Proposition 4.9(1)] without using any combinatorial identities concerning Bernoulli numbers. Corollary 4.3 (Determinants of Betti matrices). The following formula holds true: k 1 = ( 1) − (2k 1)!Λ detB2k 1 − 5k 1 − 2k 1. (4.24) − 2 − − When k is odd, we have detB2k = 0. Λ2 = Proof. Taking determinants on both sides of (3.80), we arrive at 2k 1 1/detS2k 1. Taking determi- − k 1 − 1 3k nants on both sides of (4.10), we obtain detB2k 1 detB2k 3 detS2k 1 = ( 1) − (2k + 1)4 − . Hence, we − − = 1 − − = k 1 + can build the formula for detB2k 1 inductively on detB1 24 3 and detB2k 1 detB2k 3 ( 1) − (2k 1 3kΛ2 − · − − − 1)4 − 2k 1. − When k is odd, the odd-numbered rows of B2k have non-zero elements in the even-numbered columns. k k+1 = k + Since there are 2 such columns and 2 2 1 such rows, all the rows in question must be linearly dependent. Hence detB2k = 0. + + o k 1 k 1 e k k For k Z>0, define minors B Q 2 2 and B Q 2 2 as follows: ∈ 2k 1 ∈ × 2k 1 ∈ × o −= e− = (B2k 1)a,b (B2k 1)2a 1,2b 1, (B2k 1)a,b (B2k 1)2a,2b. (4.25) − − − − − − e = e = = o e By convention, we set B1 ∅ and detB1 1. Clearly, we have detB2k 1 detB2k 1 detB2k 1 for all e − − − positive integers k. In the next corollary, we evaluate detB2k 1, a result that will be used later in the proof of Proposition 4.5. − Corollary 4.4 (Determinants of Betti minors). We have

3 ( 1)k k + − − + 2 2 k k j e = [(2k 1)!!] 2 (2k 1)!! (2 j) − detB2k 1 k + ∏ + (4.26) 1 ( 1)k k 1 + j 1 − ( 2) 2 (k 1)!!(k!!) 2 j=1 (2 j 1) − − − − for each positive integer k. BROADHURST–ROBERTSQUADRATICRELATIONS 43

Proof. By rescaling and rearranging columns and rows of determinants, we have the following results for each m Z>0: ∈ 2 2m 1 o 1 − B2(a+b 1) detB = (2n)! det − , (4.27) 4m 3 4m2 ∏ + − 2 " n=m # [2(a b 1)]! 1 a,b m − ≤ ≤ 2(m 1) 2 1 B + detBe = − (2n + 1)! det 2(a b) , (4.28) 4m 3 4m(m 1) ∏ + − 2 − " n=m # [2(a b)]! 1 a,b m 1 ≤ ≤ − 2 2m ( 1)m B + detBo = (2n)! det 2(a b) , (4.29) 4m 1 2m−(2m+1) ∏ + − 2 "n=m+1 # [2(a b)]! 1 a,b m ≤ ≤ 2 m 2m e ( 1) B2(a+b 1) detB = (2n 1)! det − . (4.30) 4m 1 2m−(2m+1) ∏ + − 2 "n=m+1 − # [2(a b 1)]! 1 a,b m − ≤ ≤ Thus, we have a recursion Λ o e (4m 1)(4m 2) 4m 1 detB4m 1 detB4m 1 detB4m 1 − = − = − 5 − Λ o − e − − 2 4m 3 detB4m 3 detB4m 3 detB4m 3 − − − − B2(a+b) 2 det [2(a+b)]! = 1 (4m)!(4m 1)! 1 a,b m − B ≤ ≤ , (4.31) 28m (2m)! 2(a+b) det + [2(a b)]! 1 a,b m 1 ≤ ≤ − namely

B2(a+b) 2 det [2(a+b)]! 1 (2m)! = 1 a,b m B ≤ ≤ , (4.32) −4m + 1 (4m)! 2(a+b) det + [2(a b)]! 1 a,b m 1 ≤ ≤ − from which we can solve m 2 B2(a+b) = m [(2n)!] det + ( 1) ∏ + . (4.33) [2(a b)]! 1 a,b m − n=1 (4n)!(4n 1)! ≤ ≤ Substituting back into (4.28) and (4.29), we may verify m+1 (5 4m)m 2m 1 e = ( 1) 2 − − detB4m 3 − ∏ (2n 1)!, (4.34) − (2m 1)!!(4m 2)!! − − − n=1 m(4m+3) 2m o = 2− detB4m 1 + ∏(2n)!, (4.35) − (2m)!!(4m 1)!! n=1 where the last equation also implies m 2m e detB4m 1 ( 1) (2m)!! = − = detB4m 1 o m−(4m+1) ∏(2n 1)!. (4.36) − detB4m 1 2 (4m)!! n=1 − − Through elementary manipulations of the factors, we may convert (4.34) and (4.36) into our statement in (4.26).

4.2. Arithmetic applications to Feynman diagrams. The Broadhurst–Roberts quadratic relations in- ject new insights into the arithmetic properties of on-shell Bessel moments, an important class of Feyn- man diagrams that respect classical equations of motion. First, we examine period matrices P2k 1 for the (2k + 1)-Bessel problems, where k Z>1. − ∈ 44 YAJUN ZHOU

TABLE II. Some Betti matrices and de Rham matrices in Broadhurst–Roberts quadratic relations

m Bm Dm 1 13 32 52 24 5 23 2·6 3 · 3 32 52 − 26 2·6 ! 1 2 32 4 25 − · 1 2 32 25 −3 5 3 · 3 17 3 863 32 52 72 25· 7 25 2· 4 ·25 ·28· 5 · 5 3 863 32 52 72 26  ·25 ·28·   3 − 3  32 52 72 5 4 2 2  ·28·   3 5   5 2 6 2 2·6 2 3 2 3 3 5 3 5 5 2 − · − · 6 ·6 ·6 2 3 − 2 − 2   ·  3 5 26 32 2·6   · 3 2 6 2 2  3 7 3 5  3 7 3 20173 3 5 97 3 5 7 2·4 2·4 2·2 · 26 ·2·8 ·210· 3 5 7 3 5 7 3 20173 32 11833 36 52 72 ·27· ·27· · 6 · 7 · 10·  − 2 −   2 2 2  7 3 5 3 5 33 5 972 36 52 72 2·4 26· ·2·8 ·210·  3 5 7 3 52   36 52 72   7    ·2 · 2· 6 ·210·  − 3 − 3    3 7 35  35 47  5 3 2 25· 25· ·2 3 59 191 2 3 5 3 2 3 7 3 5 7 35 47 − − ·5 2· 2 − · · 5· · 5· 2 3 5 8  2 2   2·2  − 3 5 7 − 32 5 7 5 2 2 − · · ·25· 2·7· 3 59 191 2 3 5  3 2   · · · ·   3 5 3 5 7   25 33 52   25· 2·7·   · ·  − −   

From the work of Bailey–Borwein–Broadhurst–Glasser [3], Bloch–Kerr–Vanhove [5], Samart [41], and the present author [48, 49, 54], we know that

2 1 IKM(1,4;1) 1 IKM(1,4;3) C 2 13C 1 P := π2 π2 = 15 10C , (4.37) 3 1 − 1 √15 √15− 2 2 − 1 π IKM(2,3;1) π IKM(2,3;3) C 13C + ! − 2 − 2 15 10C where Γ 1 Γ 2 Γ 4 Γ 8 C := 15 15 15 15 (4.38) 2 240√5π is the “Bologna constant” attributed to Broadhurst [8, 3] and Laporta [32]. Previously, each individual = b 1 b entry of P3 (( 1) − 3,a(1))1 a,b 2 was computed with sophisticated arguments. Now, a quadratic T−= F ≤ ≤ relation P3D3P3 B3 (see Table II for non-vanishing entries of B3 and D3) produces three independent quadratic equations involving Bessel moments, which immediately allow us to express all the matrix 1 = elements [cf. (4.37)] in terms of the top-left entry 3,1(1) C. As for the 7-Bessel problem, numerical evidenceF has motivated the following conjecture by Broad- hurst and Mellit (see [12, (6.8)] or [9, (129)]): 2 ? 5π IKM(2,5;1) = ζ (2). (4.39) 24 7,1 Here, we have an L-function associated with the 7th symmetric power moments of Kloosterman sums 1 := G 7 ζ7,1(s) ∏ Lp( m,Fp ,Sym Kl2, s) ∏ s , (4.40) p ≡ p Z7(p, p− )

7 whose local factors at primes p are written as Lp(Gm,Fp ,Sym Kl2, s) in the Fu–Wan notation [26, §0] 1 and s in the Broadhurst notation [9, §2]. Meanwhile, according to the Fresán–Sabbah–Yu state- Z7(p,p− ) ment [25, §7.c] of Deligne’s conjecture for the motive M7 whose period realization is isomorphic to BROADHURST–ROBERTSQUADRATICRELATIONS 45

1 7 mid 7 mid (HdR,mid(Gm,Sym Kl2),H1 (Gm,Sym Kl2),P7 ), one has 4 IKM(1,6;1) IKM(1,6;3) ? π det = ζ7,1(2). (4.41) IKM(3,4;1) IKM(3,4;3) 14√105 In transcribing the conjectural identity above, we have already exploited a relation between ζ7,1(2) and ζ7,1(3), which results from a functional equation conjectured by Broadhurst–Mellit (see [12, (6.7)] or [9, (128)]) and veriﬁed by Fresán–Sabbah–Yu [23, Theorem 1.2]: 105 s/2 s 1 s s + 1 Λ7(s) := Γ − Γ Γ ζ7,1(s) =Λ7(5 s). (4.42) π3 2 2 2 − The conjectures (4.39) and (4.41) are actually equivalent to each other, since we have 1 53 73 1 (1) 2 (1) 1 (1) = · det F5,1 F5,1 (4.43) F5,2 −22 3 1 (1) 2 (1) r F5,3 F5,3 cof P = 5 1 by reading off the element in the second row and the ﬁrst column from both sides of P5 B5 D5− , detP5 4 which is a reformulation of the quadratic relation P D PT = B using the determinant detP = 2 5 5 5 5 5 − √335577 = T 1 and the cofactor matrix cof P5 (detP5)(P5 )− . The sum rule in (4.43) is in fact a continuation of a pattern attested in P3 [see (4.37)]: √3 5 1 (1) = · 1 (1), (4.44) F3,2 2 F3,1 which in turn, is compatible with proven instances [38, 41] of Deligne’s conjecture for the motive M5:

1 1 1 3 (1) = ζ , (1), (1) = ζ , (2). (4.45) F3,1 5 5 1 F3,2 4π 5 1 Here, s 5 1 ∞ (2πy) f3,15(iy)dy ζ5,1(s) := ∏ Lp(Gm,F ,Sym Kl2, s) ∏ = (4.46) p ≡ Z (p, p s) Γ(s) y p p 5 − Z0 3 3 is the L-function associated with the cusp form f3,15(z) := [η(3z)η(5z)] + [η(z)η(15z)] of weight 3 and level 15 [37], satisfying a functional equation [9, (95)]: 15 s/2 s s + 1 Λ (s) := Γ Γ ζ , (s) =Λ (3 s). (4.47) 5 π2 2 2 5 1 5 − cof P7 1 Trading each element in a minor determinant of P7 for its counterpart in B7 D7− , we can also detP7 construct another sum rule 1 (1) 2 (1) √3 5 7 1 (1) 2 (1) det F7,2 F7,2 = · · det F7,1 F7,1 , (4.48) 1 (1) 2 (1) 22 1 (1) 2 (1) F7,4 F7,4 F7,3 F7,3 which is compatible with Deligne’s conjecture [25, §7.c] for the motive M9 and the corresponding functional equation [23, Theorem 1.2] for ζ9,1(s). For an arbitrarily large k, there is a similar identity connecting two minor determinants of P2k 1, as established in the proposition below. − Proposition 4.5 (Reﬂection formula for minor determinants of P2k 1). For k Z>1, we have an identity that is equivalent to (1.21): − ∈ b 1 b det(( 1) − 2k 1,2a(1))1 a,b k − F − ≤ ≤ 2 k + + 2 ( 1) k 1 + 1 k [(2k 1)!!] − − b 1 b = ( 1) 4 2 2 det(( 1) (1)) k+1 . (4.49) k k − 2k 1,2a 1 1 a,b − 2 1 ( 1) − F − − 2 2 p(k 1)!!(k!!) − − ≤ ≤ − 46 YAJUN ZHOU

Here, the leading factor on the right-hand side can be rewritten as b 1 δ k k+1 + 1 k det(( 1) − a,b)1 a,b 1, k 3 (mod4), ( 1) 4 2 2 = − ≤ ≤ 2 = − ≡ (4.50) b 1 + − det(( 1) − δa,b)1 a,b k 1 (1, otherwise. − ≤ ≤ 2 = A B Proof. Suppose that we partition an invertible matrix into four blocks M C D , where the top-left block A is also invertible. The Banachiewicz formula [4] provides us with a partition 1 1 1 = I A− B A− I M− − 1 1 1 I (D CA B) CA− I − − − − A 1 + A 1B(D CA 1B) 1CA 1 A 1B(D CA 1B) 1 = − − − − − − − − , (4.51) (D − CA 1B) 1CA 1 − (D − CA 1B) 1 − − − − − − − − where I stands for an identity matrix of appropriate size. Taking determinants on the first equality in 1 (4.51), we have det(M) = det(A)det(D CA− B). Define − b 1 b ⌢ (( 1) − 2k 1,2a 1(1)) a k+1 , b k = 1 2 1 P2k 1 − b 1F b − − ≤ ≤ ≤ ≤ (4.52) − (( 1) − 2k 1,2a(1))1 a k ,1 b k ! − F − ≤ ≤ 2 ≤ ≤ by reshuffling rows of P2k 1. The Broadhurst–Roberts quadratic relation can be reformulated into − ⌢ o ⌢ = B2k 1 T 1 P2k 1D2k 1 − e P2k 1 − , (4.53) − − B2k 1 − − o e where B2k 1 and B2k 1 are defined in (4.25). − − b 1 b = Suppose that det(( 1) − 2k 1,2a 1(1))1 a,b k+1 0, and take a matrix partition − F − − ≤ ≤ 2 6 ⌢ ⌢ b 1 b ⌢ A := (( 1) − (1)) k+1 B = 2k 1,2a 1 1 a,b 2 P2k 1 ⌢ − F − − ≤ ≤ ⌢ . (4.54) − C D! k k We may read off the bottom-right 2 2 block from both sides of (4.53), while referring back to the Banachiewicz formula (4.51), as follows:×

∠ ⌢ ⌢ ⌢ ⌢ 1 b 1 b | = e T T 1 T T − (( 1) − 2k 1,2a(1))1 a,b k D2k 1 B2k 1 D B A− C . (4.55) − F − ≤ ≤ 2 − − −

Here, no matter k is even or odd, the matrix D| ∠ is extracted from the top-right k k block of 2k 1 ∠ 2 2 − | × D2k 1. We may quote in advance from Corollary 4.7 that D2k 1 is upper-left triangular, with all its − 2 − (2k+1)!! anti-diagonal elements being 2k+1 . Since

h i ⌢ T 1 − b 1 b + ⌢ ⌢ ⌢ ⌢ 1 det P2k 1 det(( 1) − 2k 1,2a 1(1))1 a,b k 1 T T 1 T T − = − = − F − − ≤ ≤ 2 det D B A− C ⌢ + ⌢ , (4.56) T 1 k 1 − det A − ( 1) 4 detP2k 1 − − we can deduce + k 1 e b b ( 1) 4 detB b b

1 = 2k 1 1 + det(( 1) − 2k 1,2a(1))1 a,b k − ⌢ ∠ − det(( 1) − 2k 1,2a 1(1))1 a,b k 1 (4.57) − F − ≤ ≤ 2 | − F − − ≤ ≤ 2 detP2k 1 detD2k 1 − − ⌢ k (2 j)k j = 2 k − from (4.55). Simplifying the right-hand side of the equation abovewith detP2k 1 ( 1) ∏ j=1 + , − − √(2 j+1)2 j 1

k ∠ 2 1 k (2k+1)!! 2 | = 2 2 detD2k 1 ( 1) 2k+1 and Corollary 4.4, we reach (4.49). − − h i BROADHURST–ROBERTSQUADRATICRELATIONS 47 b 1 b = Starting from an alternative assumption that det(( 1) − 2k 1,2a(1))1 a,b k 0, while switching the − F 2 6 b 1 b b 1 b− ≤ ≤ rôles of (( 1) − 2k 1,2a 1(1))1 a k+1 ,1 b k and (( 1) − 2k 1,2a(1))1 a k ,1 b k, one can repeat the − F − − ≤ ≤ 2 ≤ ≤ − F − ≤ ≤ 2 ≤ ≤ procedures above to arrive at the same identity in (4.49). b 1 b b 1 b If both det(( 1) − 2k 1,2a 1(1))1 a,b k+1 and det(( 1) − 2k 1,2a(1))1 a,b k vanish, then (4.49) is − F − − ≤ ≤ 2 − F − ≤ ≤ 2 trivially true. Remark The reﬂection formula (4.49) applies retroactively to the case where k = 1, so long as we adopt the convention that det∅ = 1. In this case, the classical evaluation IKM(1,2;1) = π [3, (23)] is 3√3 recovered.

The cases of P2k,k Z>1 involve additional subtleties. ∈ T = The quadratic relation P4D4P4 B4 (see Table II for non-vanishing entries of B4 and D4) produces = 1 only one independent quadratic equation for Bessel moments, which is equivalent to detP4 2632 . Thus, we do not have news for the 6-Bessel problem. Especially, on the algebraic side, there seem− to be no quadratic mechanisms to relate the ratio [ 1 (1)]2/[ 1 (1)]2 to a rational number; on the analytic F4,2 F4,1 side, the proven instances [48, §4] of Deligne’s conjecture for the motive M6 bring us two evaluations 5/2 1 = π2 3/2 1 = π2 = 3 π 4,1(1) 2 ζ6,1(2) and π 4,2(1) 2 ζ6,1(1) 2 ζ6,1(3), which are not connected to each other by the functionalF equation [9, (106)]F 6 s/2 s s + 1 Λ (s) := Γ Γ ζ , (s) =Λ (4 s) (4.58) 6 π2 2 2 6 1 6 − 2 for ζ6,1(s), the L-function associated with the cusp form f4,6(z) := [η(z)η(2z)η(3z)η(6z)] of weight 4 and level 6 [30]. Unlike the situations discussed so far, the L-function for the motive M8 needs an additional correction factor, in the form of [9, §7.6] s 4 s ∞ (2πy) f6,6(iy)dy Z4(p, p − ) L( f6,6, s) := = ∏ , (4.59) Γ(s) y Z (p, p s) Z0 p 8 − where the associated cusp form [47] [η(2z)η(3z)]9 [η(z)η(6z)]9 f (z) := + (4.60) 6,6 [η(z)η(6z)]3 [η(2z)η(3z)]3 has weight 6 and level 6. The Fresán–Sabbah–Yu statement [25, §7.c] of Deligne’s conjecture holds for the motive M8, according to the following identities conjectured by Broadhurst [9, §7.6] and proved 7/2 1 7/2 1 π4 by the present author (see [48, §5] as well as [54, §3.1]): π (1) = π (1) = L( f6,6,2), F6,1 F6,3 4 5/2 1 = π4 = 9π2 T = π 6,2(1) 8 L( f6,6,1) 14 L( f6,6,3). The quadratic relation P6D6P6 B6 produces three indepen- dentF quadratic equations concerning on-shell Bessel moments, which in turn, can be reduced (say, by computation of Gröbner basis) to the following sum rules: 1 (1) = 1 (1), (4.61) F6,1 F6,3 2 (1) + 2 3 (1) = 2 (1) + 2 3 (1), (4.62) F6,1 F6,1 F6,3 F6,3 1 (1) 2 (1) + 2 3 (1) 5 det F6,2 F6,2 F6,2 = . (4.63) 1 (1) 2 (1) + 2 3 (1) 211 3 F6,3 F6,3 F6,3 · All these three identities have been reported earlier [53, (98)]. Moreover, we point out that the quadratic T = relation P6D6P6 B6 does not exhaust algebraic relations [over Q(π)] for the entries of P6. Noticeably, such a relation does not entail the following identities [54, Theorem 1.3]: 7√π 1 (1) + 23 32 2 (1) = , (4.64) F6,1 · F6,1 24 3 · 48 YAJUN ZHOU

1 (1) + 23 32 2 (1) = 0, (4.65) F6,2 · F6,2 5√π 1 (1) + 23 32 2 (1) = , (4.66) F6,3 · F6,3 − 22 3 · 1 (1) 3 (1) 32 5 28 33 det F6,1 F6,1 + · = 7√π 1 (1), (4.67) · 1 (1) 3 (1) 24 F6,2 F6,2 F6,2 which are all provable by analytic properties of the cusp form f6,6(z) [54, §3]. We say that a sum rule P( 1 (1), 2 (1),..., 4 (1)) = 0 for the 10-Bessel problem arises from a F8,1 F8,1 F8,4 “trivial factor” if P(x1,1, x1,2,..., x4,4) Z[x1,1, x1,2,..., x4,4] divides the following polynomial in 16 for- ∈ mal variables belonging to the matrix X4 = (xa,b)1 a,b 4: ≤ ≤ 4 4 T 1 cof X4 ∏ X D X B ∏ ∏ B− X D . (4.68) 4 8 4 − 8 a,b 8 4 8 − detP 1 a

T = T X4D8X4 B8 : ∑ X4D8X4 B8 a,b pa,b pa,b C[x1,1, x1,2,..., x4,4] , (4.69) − ( 1 a

BROADHURST–ROBERTSQUADRATICRELATIONS 49 preliminary experiments in this direction have not yet produced any new Q-linear sum rules [like (4.65)] that go beyond the relations for generalized Crandall numbers (2.47). T = Broadhurst and Roberts originally speculated that the quadratic equation PmDmPm Bm delivers a complete set of algebraic relations among entries of Pm, unless (m 2) is an integer multiple of 4. Fresán, Sabbah and Yu provided heuristic justiﬁcation (see comments− after [25, Theorem 1.4]) of this speculation, citing Grothendieck’s period conjecture that equates the transcendental degree of Q(Pm) with the dimension of the motivic Galois group for Pm. The aberrant sum rule (4.65) serves as a reminder that there may exist very surprising algebraic relations when m 2 (mod 4). In view of the foregoing applications, we can say that the≡ Broadhurst–Roberts quadratic relations convey more detailed messages about on-shell Bessel moments than the Broadhurst–Mellit determinant formulae. Perhaps paradoxically, the algebraic mechanism underlying the former relations is essentially identical to what we have explored and exploited during the proof [49] of the latter, namely Wronskian´ algebra for homogeneous solutions to Vanhove’s differential equations. Our off-shell generalizations (Theorem 1.1) to the Broadhurst–Roberts quadratic relations also have some interesting consequences. ˙ ℓ = ´ ℓ = 1 ℓ If we deﬁne m, j(u) : √u m, j(u) 2uD m, j(u), then the Gröbner basis of the off-shell ideal FT F F (u)W (u)S W (u)V (u) + I4 offers us a relation hL4 4 4 4 4 i 1 (u) 1 (u) 1 (u) 1 (u) = 2 + F4,2 F4,3 + F4,2 F4,3 0 2u(3u 70u 259)det 2 2 8(u 13)(u 1)det ˙ 1 ˙ 1 − 4,2(u) 4,3(u) − − 4,2(u) 4,3(u) F F F F 1 (u) 1 (u) 2 (u) 2 (u) + F4,2 F4,3 + F4,2 F4,3 (u 25)(u 9)(u 1) det ˙ 2 ˙ 2 det ˙ 1 ˙ 1 (4.72) − − − 4,2(u) 4,3(u) 4,2(u) 4,3(u) F F F F for u (0,1), which extends to u ( ,1) by analytic continuation. The last displayed identity reveals equivalence∈ between two conjectural∈ −∞ determinant formulae of Broadhurst (see [11, §3] or [1, §5.3]) 1 1 1 2 3 ( 7) + 4 ˙ ( 7) ( 7) + 28 ( 7) ? 3 det F4,3 − F4,3 − F4,3 − F4,3 − = , (4.73) 3 1 ( 7) + 4 ˙ 1 ( 7) 1 ( 7) + 28 2 ( 7) − 32 F4,2 − F4,2 − F4,2 − F4,2 − 1 1 2 2 ( 7) 39 ˙ ( 7) 427 ( 7) 112 ˙ ( 7) ? 3 det F4,3 − F4,3 − − F4,3 − − F4,3 − = , (4.74) 1 ( 7) 39 ˙ 1 ( 7) 427 2 ( 7) 112 ˙ 2 ( 7) 32 F4,2 − F4,2 − − F4,2 − − F4,2 − which in turn, are inspired by the recent work of Candelas, de la Ossa, Elmi, and van Straten [18]. Simi- larly, for u = (√17 4)2, one can demonstrate that two determinant formulae conjectured by Broadhurst (see [11, §4] or [1,− §5.4]) are equivalent to each other. It is perhaps appropriate to ask whether there is an algebro-geometric interpretation of our off-shell quadratic relations, in the spirit of Fresán–Jossen [22] and Fresán–Sabbah–Yu [23, 24, 25]. We would like to see such an extension of the motivic method to certain algebraic values of the off-shell parameter 1 1 1 2 1 2 u. In the special cases of 4,2( 7), 4,3( 7), 4,2((√17 4) ) and 4,3((√17 4) ), we hope that powerful tools in algebraic geometryF may− eventuallyF − F lead to proofs− of Broadhurst’sF − recent conjectures (see [11, §§3–4] or [1, §§5.3–5.4]) that evaluate these off-shell Bessel moments in terms of special L-values.

4.3. Alternative representations of de Rham matrices. From the invertibility of Pm guaranteed by (1.2), T = we know that the de Rham matrix Dm satisfying the Broadhurst–Roberts quadratic relation PmDmPm Bm is unique. Nonetheless, there exist several different expressions for these de Rham matrices, whose mutual equivalence is not self-evident (absent the uniqueness argument).

Proposition 4.6 (Representations of Dm). Let βm(u) be the Bessel matrix deﬁned in Proposition 2.6, and abbreviate βm(1) as βm. Pick the matrix ̺k as given in Proposition 2.7, together with the matrix ϑm as (2k 1) (2k 2) described in Proposition 2.8. We further introduce a matrix Ψk Z with entries ∈ − × − δa,b(1 δa,k), a Z [1,k], (Ψk)a,b = − ∈ ∩ (4.75) δa,b+1, a Z [k + 1,2k 1]. ( ∈ ∩ − 50 YAJUN ZHOU

The following representations of de Rham matrices hold for k Z>1: ∈ 1 T 1 T 1 1 2 2 2k 1(1) (ϑ2−k 1) (β2−k 1) V2k 1(1)β2−k 1ϑ2−k 1 D |L − | = 2k+1 2k 1 , (4.76) − 2 + − k −1 − − − 2 (2k 1)( 1) − D2k 3 − − 1 T 1 T 1 T 1 2 2k(u) (ϑ− ) ̺ (β− ) V (u)β− ̺ ϑ− 1 D lim |L | 2k k 2k 2k 2k k 2k = k+1 2k , (4.77) 3 + k 1 u 1− 2 (k 1)( 1) − D2k 2 → − − T 1 1 2k(u) [β2k(u)]− V2k(u)[β2k(u)]− D2k 1 lim |L | = − − , (4.78) u 0+ 23( 1)k D2k 1 D2k 1 → − − ◦ − T T 1 1 2k 1(u) Ψk [β2k 1(u)]− V2k 1(u)[β2k 1(u)]− Ψk D2k 2 lim |L − | − − − = − − , (4.79) u 0+ 23( 1)k D2k 2 D2k 2 → − − ◦ − m + = 3 2 1 B D where Dm and (Bm)a,b 2− ( 1)⌊ ⌋ ( m+1)a,b [deﬁned in (4.12)] play rôles in the following relation ◦ ◦ − ◦ for Pm [deﬁned in (1.22)]: ◦ T = + T T PmDmPm πBm PmDmPm PmDmPm. (4.80) ◦ ◦ ◦ − ◦ Proof. According to Proposition 4.2, the left-hand of (4.76) [resp. (4.77)] is equal to

1 1 2 T Pm − 2 B Pm − m+2 m (4.81) T Pm 2! Bm 2 Pm 2! − − − − − for m = 2k 1 (resp. m = 2k), hence expressible in terms of de Rham matrices. In the u − 0+ regime, we have → 2 I0(√ut) = 1 + O(ut ), (4.82) √ut K (√ut) + γ + log I (√ut) = O(ut2), (4.83) 0 0 2 0 2 √utI1(√ut) = O(ut ), (4.84)

√ut 2 √ut K1(√ut) γ0 + log I1(√ut) = 1 + O(ut ), (4.85) − 2 where γ0 is the Euler–Mascheroni constant. 2k 2k With a reshufﬂing-rescaling matrix ξ2k Q × deﬁned by (3.81), we may paraphrase [49, Proposition 4.7] into ∈

√πPT 1 PT T T = 2k 1 − √π 2k 1 lim β2k(u)W2k(u)ξ2kγ2k(u) − ◦ − , (4.86) u 0+ 1 PT √π 2k 1! → − −

Ik where γ (u) = 1 √u , and P is described in (1.22). Meanwhile, by direct computation, 2k γ0+log Ik Ik 2k 1 π 2 ◦ − we have 

k 3 B2k 1 = 1 T = 1 T T ( 1) 2 − ξ2kS2−k (u)ξ2k γ2k(u)ξ2kS2−k (u)ξ2kγ2k(u) − B2k 1 B2k 1 − − ◦ − T T T = k(u) γ (u)ξ W (u)V (u)W (u)ξ γ (u). (4.87) |L2 | 2k 2k 2k 2k 2k 2k 2k BROADHURST–ROBERTSQUADRATICRELATIONS 51

Therefore, the limit in (4.78) exists, and is equal to 1 1 √πP − √πPT 1 PT − 2k 1 B2k 1 2k 1 √π 2k 1 − − − ◦ − 1 1 − 1 P P B2k 1 B2k 1 PT √π 2k 1 √π 2k 1! √π 2k 1! − ◦ − − − − − ◦ − − − D2k 1 = − − 1 + 1 T 1 T T 1 . (4.88) D2k 1 P2−k 1 πB2k 1 P2k 1P2−k 1B2k 1 B2k 1 P2k 1 − P2k 1 P2k 1 − ! − − ◦ − ◦ − − − − − − ◦ − − h T = i Recalling the Broadhurst–Roberts quadratic relation P2k 1D2k 1P2k 1 B2k 1, we may also rewrite − 1 T 1 T = T − − − T P2k 1P2−k 1B2k 1 B2k 1 P2k 1 − P2k 1 P2k 1D2k 1P2k 1 P2k 1D2k 1P2k 1, thereby conﬁrming (4.80) when◦ − m =− 2k −1− is an− odd number.− ◦ − ◦ − − − − − − ◦ − When m = 2−k is an even number, the ideas behind (4.80) are essentially similar, except that one needs a + T 1 variation on Lemma 4.1 to show that the k-th row and the k-th column of limu 0 2k 1(u) [β2k 1(u)]− 1 → |L − | − V2k 1(u)[β2k 1(u)]− are ﬁlled with zeros. We leave the details to diligent readers. − − Remark Equating the different representations of de Rham matrix Dm, we may deduce recursive rela- n n tions connecting a party of integers D ℓm, j(1), j Z [0,m] to another party D ℓm 2, j(1), j Z [0,m 2] ∈ ∩ n − ∈ ∩ − [see (4.76) and (4.77)], as well as relations connecting both parties to D ℓm 1, j(0), j Z [0,m 1] [see n − ∈ ∩ − (4.78) and (4.79)]. These recursions on D ℓm, j are effectively just restatements of the following facts un- + = 1 + = + = der suitably chosen bases: I0(t)K1(t) I1(t)K0(t) t , limu 0 I0(√ut) 1 and limu 0 √utK1(√ut) 1. Direct combinatorial proofs of such recursions may appear→ hard, though not infeasible.→

Corollary 4.7 (Anti-diagonal structure of D2k 1). The de Rham matrix D2k 1 is upper-left triangular, − 2 − (2k+1)!! with its main anti-diagonal being occupied by 2k+1 . Proof. After inspecting the highest order derivativesh andi highest order moments in 2k equations gener- T 1 ated from rows of (4.4), we immediately know that [β2k(u)]− has the following structure: 0, a = 1,b Z [2,2k], ∈ ∩ ( 4u)1 a, a Z [1,k],b = 2a 1,  − T 1 = − ∈ ∩ − ([β2k(u)]− )a,b 0, a Z [2,k],b Z [1,2a 1), (4.89)  1 k+1 a ∈ ∩ ∈ ∩ −  ( 4u) − , a Z [k + 1,2k],b = 2(a k), 2u − ∈ ∩ − 0, a Z [k + 1,2k],b Z [1,2(a k)).  ∈ ∩ ∈ ∩ −  a 1 Since V2k(u) is upper-left triangular, with anti-diagonal elements (V2k(u))a,2k+1 a = ( 1) − , we can show that − − ( 4u)1 a, a Z [1,k],b = 2(k a), − − ∈ ∩ − 0, a Z [2,k],b Z (2(k a),2k], ([βT (u)] 1V (u)) =  (4.90) 2k − 2k a,b 1 k+1 a ∈ ∩ + ∈ ∩= − +  ( 4u) − , a Z [k 1,2k],b 2(2k a) 1, − 2u − ∈ ∩ − 0, a Z [k + 1,2k],b Z (2(2k a) + 1,2k]. ∈ ∩ ∈ ∩ −   T 1 1 U k k This further leads us to a block partition [β (u)] V (u)[β (u)] = T∗ where U Q(u) is an 2k − 2k 2k − U ∈ × − k 1 ∗ ( 1) − upper-left triangular matrix whose main anti-diagonal is occupied by −2k 1 k . (Here, an asterisk denotes 2 − u a block with irrelevant details.) Referring back to (4.78), we know that D2k 1 is an upper-left triangular matrix whose main anti-diagonal is ﬁlled with − k 1 + 2 ( 1) − 2k(u) = (2k 1)!! lim − |L | + , (4.91) u 0+ 22k 1uk 23( 1)k 1 2k 1 → − − − as claimed. 52 YAJUN ZHOU

Remark In [25, Theorem 1.4], Fresán, Sabbah and Yu have characterized (effectively) the inverse de 1 1 Rham matrices D2−k 1,k Z>0 and D2−k ,k 2Z>0, in the form of lower-right triangular matrices, with explicit formulae for− elements∈ on the main anti-diagonal.∈ Paraphrasing their result with matrix inversions and suitable normalizations, we can also conﬁrm the anti-diagonal behavior of D2k 1. − Extending the proof the proposition above, one can show that the de Rham matrix D2k is always upper-left triangular. When k is odd, the center of the main anti-diagonal in the skew-symmetric matrix = 1 T 1 = D2k P2−k B2k P2k − must be zero, hence detD2k 0. (Alternatively, one may deduce this fact from detB2k = 0, which has been proved in Corollary 4.3.) To avoid inverting a singular de Rham matrix, Fresán–Sabbah–Yu [25, §3.b] constructed a quadratic T = relation that was different from P2kD2kP2k B2k for odd k, without involving the last row in P2k. Accord- ingly, the Gröbner basis generating their quadratic relation does not contain linear sum rules derivable from generalized Crandall numbers (see (2.47), which recapitulates [53, (65)]).

The uniqueness of de Rham matrices gives an indirect proof that our constructions of D2k 1,k Z>0 − ∈ and D2k,k 2Z>0 are equivalent to Fresán–Sabbah–Yu [25, §3.b]. Without resorting to a uniqueness argument, we∈ may interpret such equivalence directly through the following identities (cf. [49, (4.22)] or [54, (2.10]):

m 2 + 1 I0(√ut) = ( 1) I0(√ut) Lm(uD D ) t 2−m+2 Lm∗ +2 t , m (4.92) 2 + 1 K0(√ut) = ( 1) K0(√ut) ( Lm(uD D ) t 2−m+2 Lm∗ +2 t . e Here, the operator L is formal adjoint to the Borwein–Salvy operator L + [6, Lemma 3.3], which m∗ +2 e m 2 + d 2 2 d 0 is the (m 1)-st symmetric power of the Bessel differential operator t dt t t dt , annihilating each + − member in the set [I (t)] j[K (t)]m 1 j j Z [0,m+1] . The Borwein–Salvy operator L + = L + + 0 0 − m 2 m 2,m 2 is constructed recursively{ by the Bronstein–Mulders–Weil| ∈ ∩ } algorithm [15, Theorem 1]:

d 0 d L + = t ,L + = t , m 2,0 dt m 2,1 dt (4.93) = d + 2 + (Lm+2,k+1 t Lm+2,k k(m 2 k)t Lm+2,k 1, k Z [1,m 1]. dt − − − ∀ ∈ ∩ To ﬁnd out the de Rham matrix that ﬁts into the relation P D = B cof Pm , we have differentiated off- m m m detPm shell Bessel moments with respect to u, and have evaluated these derivatives in the u 1− limit. To attain the same goal, Fresán, Sabbah and Yu [25, Proposition 3.4] have integrated by parts→ with respect to t, relating Bessel moments IKM(a,b;n) of different orders n by a recursive algorithm [25, (3.9)] that emulates Bronstein–Mulders–Weil [15, Theorem 1] and Borwein–Salvy [6, Lemma 3.3]. In other words, the equivalence between our Wronskian´ approach to de Rham matrices and the non-Wronskian´ approach of Fresán–Sabbah–Yu [25, §3.b] ultimately roots from the duality relation (4.92) between Vanhove and Borwein–Salvy operators, at u = 1. We will not further elaborate on such a duality, as it is not essential to the theme of the current work. Broadhurst and Roberts have proposed a set of combinatorial formulae [13, (5.7)–(5.12)] that conjec- turally generate the de Rham matrices Dm,m Z>0 through a recursion on m. The uniqueness argument alone does not allow us to settle their conjecture.∈ What we currently know is that various representations of Dm in Proposition 4.6 do provide us with recursive relations concerning special values of the polyno- n mials D ℓm, j with three consecutive values of m, which are reminiscent of the three-term recursion for the Broadhurst–Roberts bivariate polynomials [13, (5.10)]. Perhaps the former three-term relations can be translated into the Broadhurst–Roberts recursions for Dm, through some combinatorial efforts. It is our hope that the recursive structures of Dm, once completely understood, will shed new light BR = T on the algebraic and arithmetic nature of the Broadhurst–Roberts ideal Im+2 X (m+1)/2 DmX (m+1)/2 ⌊+ +⌋ ⌊ ⌋ − BR = m 1 m 1 Bm and the corresponding Broadhurst–Roberts variety Vm+2 : X (m+1)/2 C 2 × 2 X (m+1)/2 Dm ⌊ ⌋ ∈ ⌊ ⌋ T = n X ( m+1)/2 Bm 0 . ⌊ ⌋ − o BROADHURST–ROBERTSQUADRATICRELATIONS 53

REFERENCES [1] Kevin Acres and David Broadhurst. Empirical determinations of Feynman integrals using integer relation algorithms. arXiv:2103.06345v1 [hep-ph], 2021. [2] Takashi Agoh and Karl Dilcher. Reciprocity relations for Bernoulli numbers. Amer. Math. Monthly, 115(3):237–244, 2008. [3] David H. Bailey, Jonathan M. Borwein, David Broadhurst, and M. L. Glasser. Elliptic integral evaluations of Bessel moments and applications. J. Phys. A, 41(20):205203 (46pp), 2008. arXiv:0801.0891v2 [hep-th]. [4] Tadeusz Banachiewicz. Zur Berechnung der Determinanten, wie auch der Inversen, und zur darauf basierten Auflösung der Systeme linearer Gleichungen. Acta Astronomica, Series C, 3:41–67, 1937. [5] Spencer Bloch, Matt Kerr, and Pierre Vanhove. A Feynman integral via higher normal functions. Compos. Math., 151(12):2329–2375, 2015. arXiv:1406.2664v3 [hep-th]. [6] Jonathan M. Borwein and Bruno Salvy. A proof of a recurrence for Bessel moments. Experiment. Math., 17(2):223–230, 2008. arXiv:0706.1409v2 [cs.SC]. [7] Jonathan M. Borwein, Armin Straub, James Wan, and Wadim Zudilin. Densities of short uniform random walks. Canad. J. Math., 64(5):961–990, 2012. (With an appendix by Don Zagier) arXiv:1103.2995 [math.CA]. [8] David Broadhurst. Reciprocal PSLQ and the tiny nome of Bologna. In In- ternational Workshop “Frontiers in Perturbative Quantum Field Theory”, Biele- feld, Germany, June 14, 2007. Zentrum für interdisziplinäre Forschung in Bielefeld. https://www.researchgate.net/publication/266404058_Reciprocal_PSLQ_and_the_tiny_Nome_of_Bologna. [9] David Broadhurst. Feynman integrals, L-series and Kloosterman moments. Commun. Number Theory Phys., 10(3):527–569, 2016. arXiv:1604.03057v1 [physics.gen-ph]. [10] David Broadhurst. L-series from Feynman diagrams with up to 22 loops. In Workshop on Multi-loop Calculations: Methods and Applications, Paris, France, June 7, 2017. Séminaires Internationaux de Recherche de Sorbonne Universités. https://multi-loop-2017.sciencesconf.org/data/program/Broadhurst.pdf. [11] David Broadhurst. Feynman integrals, quasi-periods and black holes. In Elliptics and Beyond (MITP Virtual Workshop), Mainz, Germany, September 10, 2020. Mainz Institute for Theoretical Physics. [12] David Broadhurst and Anton Mellit. Perturbative quantum field theory informs algebraic geometry. In Loops and Legs in Quantum Field Theory. PoS (LL2016) 079, 2016. https://pos.sissa.it/archive/conferences/260/079/LL2016_079.pdf. [13] David Broadhurst and David Roberts. Quadratic relations between Feynman integrals. In Loops and Legs in Quantum Field Theory. PoS (LL2018) 053, 2018. https://pos.sissa.it/303/053/pdf. [14] David Broadhurst and David P. Roberts. L-series and Feynman integrals. In David R. Wood, Jan de Gier, Cheryl E. Praeger, and Terence Tao, editors, 2017 MATRIX Annals, volume 2 of MATRIX Book Series, pages 401–403. Springer, Cham, Switzerland, 2019. [15] Manuel Bronstein, Thom Mulders, and Jacques-Arthur Weil. On symmetric powers of differential operators. In Proceedings of the 1997 International Symposium on Symbolic and Algebraic Computation (Kihei, HI), pages 156–163. ACM, New York, 1997. [16] Bruno Buchberger. An algorithm for finding the basis elements of the residue class ring of a zero dimensional polynomial ideal. J. Symbolic Comput., 41(3-4):475–511, 2006. (Translated from the 1965 German original by Michael P. Abramson). [17] Kilian Bönisch, Fabian Fischbach, Albrecht Klemm, Christoph Nega, and Reza Safari. Analytic structure of all loop banana amplitudes. arXiv:2008.10574 [hep-th], 2020. [18] Philip Candelas, Xenia de la Ossa, Mohamed Elmi, and Duco van Straten. A one parameter family of Calabi-Yau manifolds with attractor points of rank two. J. High Energy Phys., 2020(10), 2020. 54 YAJUN ZHOU

arXiv:1912.06146v1 [math.NT]. [19] P. Deligne. Valeurs de fonctions L et périodes d’intégrales. In Automorphic forms, representations and L-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 2, Proc. Sympos. Pure Math., XXXIII, pages 313–346. Amer. Math. Soc., Providence, R.I., 1979. (With an appendix by N. Koblitz and A. Ogus). [20] Plamen Djakov and Boris Mityagin. Asymptotics of instability zones of Hill operators with a two term potential. C. R. Math. Acad. Sci. Paris, 339(5):351–354, 2004. [21] Plamen Djakov and Boris Mityagin. Asymptotics of instability zones of the Hill operator with a two term potential. J. Funct. Anal., 242(1):157–194, 2007. arXiv:math-ph/0509034v1. [22] Javier Fresán and Peter Jossen. Exponential motives. javier.fresan.perso.math.cnrs.fr/expmot.pdf, 2020. [23] Javier Fresán, Claude Sabbah, and Jeng-Daw Yu. Hodge theory of Kloosterman connections. arXiv:1810.0654v3 [math.AG], 2020. [24] Javier Fresán, Claude Sabbah, and Jeng-Daw Yu. Quadratic relations between periods of connections. arXiv:2005.11525v2 [math.AG], 2020. [25] Javier Fresán, Claude Sabbah, and Jeng-Daw Yu. Quadratic relations between Bessel moments. arXiv:2006.02702v1 [math.AG], 2020. [26] Lei Fu and Daqing Wan. L-functions of symmetric products of the Kloosterman sheaf over Z. Math. Ann., 342(2):387–404, 2008. [27] M. B. Gel’fand. A note on a certain relation among Bernoulli numbers. Baškir. Gos. Univ. Uˇcen. Zap. Vyp., 31(Ser. Mat. 3):215–216, 1968. [28] M. L. Glasser and E. Montaldi. Staircase polygons and recurrent lattice walks. Phys. Rev. E, 48:R2339–R2342, 1993. [29] S. Groote, J. G. Körner, and A. A. Privovarov. On the evaluation of a certain class of Feynman diagrams in x-space: Sunrise-type topologies at any loop order. Ann. Phys., 322:2374–2445, 2007. arXiv:hep-ph/0506286v1. [30] K. Hulek, J. Spandaw, B. van Geemen, and D. van Straten. The modularity of the Barth-Nieto quintic and its relatives. Adv. Geom., 1(3):263–289, 2001. arXiv:math/0010049v1 [math.AG]. [31] E. L. Ince. Ordinary differential equations. Dover Publications, New York, NY, 1956. [32] S. Laporta. Analytical expressions of three- and four-loop sunrise Feynman integrals and four-dimensional lattice integrals. Internat. J. Modern Phys. A, 23(31):5007–5020, 2008. arXiv:0803.1007v4 [hep-ph]. [33] S. Laporta and E. Remiddi. The analytical value of the electron (g 2) at order α3 in QED. Physics Letters B, 379(1):283–291, 1996. arXiv:hep-ph/9602417. − [34] Stefano Laporta. High-precision calculation of the 4-loop contribution to the electron g 2 in QED. Physics Letters B, 772(Supplement C):232–238, 2017. arXiv:1704.06996 [hep-th]. − [35] Roman N. Lee and Andrei A. Pomeransky. Differential equations, recurrence relations, and quadratic constraints for L-loop two-point massive tadpoles and propagators. J. High Energy Phys., (8):027, 26, 2019. arXiv:1904.12496v1 [hep-ph]. [36] Niels Nielsen. Traité élémentaire des nombres de Bernoulli. Gauthier-Villars, Paris, France, 1923. [37] C. Peters, J. Top, and M. van der Vlugt. The Hasse zeta function of a K3 surface related to the number of words of weight 5 in the Melas codes. J. Reine Angew. Math., 432:151–176, 1992. [38] M. Rogers, J. G. Wan, and I. J. Zucker. Moments of elliptic integrals and critical L-values. Ra- manujan J., 37(1):113–130, 2015. [39] Louis Saalschütz. Verkürzte Recursionsformeln für die Bernoullischen Zahlen. Zeit. für Math. und Phys., 37:374–378, 1892. [40] Louis Saalschütz. Vorlesungen über die Bernoullischen Zahlen, ihren Zusammenhang mit den Secanten-Coefﬁcienten und ihre wichtigeren Anwendungen. Springer, Berlin, Germany, 1893. [41] Detchat Samart. Feynman integrals and critical modular L-values. Commun. Number Theory Phys., 10(1):133–156, 2016. arXiv:1511.07947v2 [math.NT]. BROADHURST–ROBERTSQUADRATICRELATIONS 55

[42] Pierre Vanhove. The physics and the mixed Hodge structure of Feynman integrals. In String-Math 2013, volume 88 of Proc. Sympos. Pure Math., pages 161–194. Amer. Math. Soc., Providence, RI, 2014. arXiv:1401.6438 [hep-th]. [43] Pierre Vanhove. Feynman integrals, toric geometry and mirror symmetry. In Johannes Blümlein, Carsten Schneider, and Peter Paule, editors, Elliptic integrals, elliptic functions and modular forms in quantum ﬁeld theory, Texts & Monographs in Symbolic Computation, chapter 17, pages 415– 458. Springer, Cham, Switzerland, 2019. arXiv:1807.11466v2 [hep-th]. [44] Pierre Vanhove. Feynman integrals and mirror symmetry. In Valery A. Gritsenko and Vyacheslav P. Spiridonov, editors, Partition Functions and Automorphic Forms, Moscow Lectures 5, pages 319– 367. Springer, Cham, Switzerland, 2020. [45] H. A. Verrill. Sums of squares of binomial coefﬁcients, with applications to Picard–Fuchs equations. arXiv:math/0407327v1 [math.CO], 2004. [46] G. N. Watson. A Treatise on the Theory of Bessel Functions. Cambridge University Press, Cam- bridge, UK, 2nd edition, 1944. [47] Zhiwei Yun. Galois representations attached to moments of Kloosterman sums and conjectures of Evans. Compos. Math., 151(1):68–120, 2015. (Appendix B by Christelle Vincent) arXiv:1308.3920v2 [math.NT]. [48] Yajun Zhou. Wick rotations, Eichler integrals, and multi-loop Feynman diagrams. Commun. Num- ber Theory Phys., 12(1):127–192, 2018. arXiv:1706.08308v4 [math.NT]. [49] Yajun Zhou. Wronskian´ factorizations and Broadhurst–Mellit determinant formulae. Commun. Number Theory Phys., 12(2):355–407, 2018. arXiv:1711.01829v2 [math.CA]. [50] Yajun Zhou. Hilbert transforms and sum rules of Bessel moments. Ramanujan J., 48(1):159–172, 2019. arXiv:1706.01068v2 [math.CA]. [51] Yajun Zhou. On Borwein’s conjectures for planar uniform random walks. J. Aust. Math. Soc., 107(3):392–411, 2019. arXiv:1708.02857v2 [math.CA]. [52] Yajun Zhou. On Laporta’s 4-loop sunrise formulae. Ramanujan J., 50(3):465–503, 2019. arXiv:1801.02182v1 [math.CA]. [53] Yajun Zhou. Some algebraic and arithmetic properties of Feynman diagrams. In Johannes Blüm- lein, Carsten Schneider, and Peter Paule, editors, Elliptic Integrals, Elliptic Functions and Modular Forms in Quantum Field Theory, Texts & Monographs in Symbolic Computation, chapter 19, pages 485–509. Springer, Cham, Switzerland, 2019. arXiv:1801.05555v2 [math.NT]. [54] Yajun Zhou. Q-linear dependence of certain Bessel moments. Ramanujan J., 2021. doi:10.1007/s11139-021-00416-9 (to appear) arXiv:1911.04141v3 [math.NT].

PROGRAM IN APPLIED AND COMPUTATIONAL MATHEMATICS (PACM), PRINCETON UNIVERSITY, PRINCETON, NJ 08544 Email address: [email protected] Current address: ACADEMY OF ADVANCED INTERDISCIPLINARY STUDIES (AAIS), PEKING UNIVERSITY, BEIJING 100871, P. R. CHINA Email address: [email protected]