Proc. Nat. Acad. Sci.

USA
Vol. 71, No. 10, pp. 4082-4085, October 1974

An Analytic Generalization of the Rogers-Ramanujan Identities for Odd Moduli

(basic hypergeometric series/partitions/combinational identities/partition identities)

GEORGE E. ANDREWS
410 McAllister Building, Pennsylvania State University, University Park, Pa. 16802
Communicated by Mark Kac, July 12, 1974

ABSTRACT A (k - 1)-fold Eulerian series expansion is A large amount of work was done on identities of this
given for 11(1-qn)l, where the product runs over all posi-
tive integers n that Rre not congruent to 0i or - i modulo type, and a collection of some 130 such identities was
2k + 1. The Rogers-Rarnanujan identities are the cases published by L. J. Slater (5). In each of the identities
k =i 2andk 'i+1 =2. given by Slater, the modulus arising in the infinite
1. Introduction product is always < 64 and always has its prime factor
among {2,3,5,7}.
The Rogers-Ramanujan identities have a long and Recently (6) five double series identities were ob-
interesting history. In 1894, L. J. Rogers proved the tained related to the modulus 11; thus
following two identities which have subsequently be-
come known as the Rogers-Ramanujan or Rogers- H (1q~+q + q2mn +
q q ) E
Ramanujan-Schur identities (see ref. 1 -pp. 90-99, for .m1 nkO j0O
the early history of these identities): qft XJ+8J'+J(...)co;l~J=
Xq + 12 (1)'(,q)4n+2J§_ I
I

( -
n0. md5II (1 - q5)-1, [1.1] (q4jq4) n(q2 ;q2), (qua4jq4)2n+2j n -.
5md11
n , -4-2(mnod 5) X (1 - q5)1 [1 5]
and Unfortunately the methods used to prove [1.5] (as well
en +n at as those used by Rogers, Bailey, and Slater) seem
(1 - ql)-1, [1.2] limited to special cases and do not provide general
nZ?0- (q)-s H. theorems for an infinite family of moduli.
In attempting.to generalize fully [1.1 ] and [1.2], H. L.
where (a)o = 1, (a)n = (a;q)n = (1-a)(1 -aq)... Alder (7) was able to show that there exist polynomials
(-qq~) and Iq < 1. Gkj(n;q) such that
The search for identities of a similar character has
gone on intermittently ever since; however, no simple Gk(n;q) 0 (1 - qf)1; [1.6]
general identities of this nature have ever been dis- n(>0 q) n n0, 4i(mod 2k+ 1)
covered.
V. N. Singh's formula (ref. 2, Eq. 3.6) comes the however neither his work nor subsequent papers (refs.
closest to this objective; however it contains a rather 7; 8, Sec. 6; 9; 10; 2; 11; 12) on the Alder polynomials
complicated representation of the Alder polynomial. yielded analytic identities of the elegance and simplicity
L. J. Rogers (ref. 3, p. 331, Eq. 6) also discovered of [1.1] and [1.2].
further identities related to the modulus 7; for example, The object of this paper is to prove the following
he -proved that generalization of the Rogers-Ramanujan identities:
THEOREM 1. Let 1 _ i < k be integers; then
0I1 (1 + M ) n20 q2)n5 gNi2+N2'2+- * +N9JC1 +Ni+N;+. +Nk_1
at -i
nlun2, L * *,nk- 2
i0
- k1 (1 - q")-1. [1.3]
.
(q) n,(2\) ns -
(q) nlt
CD
n0,43(mod 7) fI
n=l
(1 -
qn)-1i [1.7]
In this same Yein, W. N. Bailey (ref. 4, p. 421) proved n*O,f4f-(mod 2k+l)
I

three identities related to the modulus 9; for example where N1 = nj + nj+l + * * * + nkl1.
co qVO
E(q)$m
We remark that this result reduces to [1.1] for k =
mn-1 mzO (q8;q')0(q3;q3)2m i = 2, and to [1.2] for k = 2, i = 1; otherwise, to our
knowledge, no special case of this result has appeared in
=H n=l
(1 -
q)-1. [1.4] the- literature. We should point out that the work in
n40, 4(mod 9) Section'3 here shows that V. N. Singh's formula (ref. 2,
4082
 Proc. Nat. Acad. Sci. USA 71 (1974) The Rogers-Ramanujan Identities 4083

Eq. 3.6) can be transformed into Theorem 1 in the case define Qti(x) = E C(J,j(MN)xMq, then by [2.11
M,N2O
k = i by suitable rearrangements of series. For example, and [2.2]: (i) Cto(MN) = 0 for all M and N, (ii)
with k = i = 3, we obtain CQi(MN) = 0 if either M or N is nonpositive and
q2%2+2*m+mi co
M2 + N2 # 0, (iii) Ck,t(0,0) = 1 for 1 < i < k, (iy)
co
q = II (1 - q*)-1 [1.8]
n,mLO- (q)n(q)m n=1,+(od7 CU-.j(MN) -Ct,- (M N) = Ck.-l+l(M - i + 1, N -
instead of [1.3]. M) for 1 . i < k. Now these four conditions uniqiely
It is important to note that the technique we shall determine the Ct,,(M,N); this is easily established by
use to prove Theorem 1 is applicable to a number of a double induction first on N and then on i. Hence we
problems in the theory of partitions, and we hope in time see that the Qkt(x) (O < i . k) are the only functions
to apply it to some of the questions raised in (ref. 13; analytic in x and q around (0,0) that satisfy the func-
Sec. 3) concerning partition identities. tional equation [2.2] for 1 < i < k and the boundary
After proving Theorem i in Section 2, we shall utilize conditions Qt.o(x) - 0, Qkf(O) = 1, 1 < i < k. To prove
our results to obtain a new formula for the Alder poly- [2.41 then, we need only show that the right hand side
nomials. of this identity also fulfills these defining conditions.
We define for k 2,0 < i <
2. Proof of Theorem 1
XR(=(1)*q(k`1)U"'+(Ac-i)n
We begin with a function that was originally introduced Rs,tj(x) = n20 -Qk-lit(xq2*).
by L. J. Rogers (14) and was considered independently (q)n
by A. Selberg (15). Immediately we see that Rko(x) = 0 and Rk,(0) =1
for 1 < i < k. Furthermore for 1 < i < k
QG,(x) = E
n20O X~t-l)sq(k -1)ft2+('t-i)n
(- 1)nz>uq1/2(2t+D)5+1)*t-(1 - Xiq(20+l) ) [2.1] Rki,(x) Rk f.1(x) = E n20:O (q)3
X (Qk-ij(xq2l) - q*Qj-i (q2"))
where (a)..= (a;q). = lim(a;q),. = (k-1)nq(kl-)*'+(k-tO*
For completeness we include a proof of the following n>0
X(A-J)Aq
A;-~s"(A;i~x(Qc-1,,t-1 (xq21)
-(q)x
functional equation that was independently discovered + (Xq2M+1) 'lQk-lk-t(xq2s+1) - q*Qk-l~t-l(xq2"))
by Rogers (14) and Selberg (15).
Qk,.t(x) Qk.t -(x) = (xq)t-Qk,..k-{+1(xq). [2.2] n>O ~(q)3 (1 - q*)Qk-l,t-i(zq2*)
This follows from the fact that
q/2(2k+1)n(n+l)
Qat,t(x) -Q -I(X) =
E
nO (q)nj(xqn+') (q-10 xiq('&+I)i -

q(-'+')'$ + xi-iq(lg+')(i-l))
_(_1)nxknql/2(2t+1)n(n+l) 1)St/2(2A:+l)X(X+I)-in
6X_~
n>0 (q)n(xqf+l). (q-t1(l -
qM) + xz-lq(s+l)(l-1)(1 - xqn+l)) n +l

+ (xq)t1 E
I-1)u(n+1) +n(i-1) ( )7.ql/2(2k+l)U(X+l) +(2t+l-i)s
n20O (q) n (x- 1l+2)
I
= -Xkq2k+1-+(E
n.0 (q) 1-in
.(xqlg+2)OD
+ + (xq)'' )

E ( -l1)x(xq)ksql/2(2Ak+1)X(n+1) -(k-i+1)u(1
(_l)*Xt(ql/2(2k+lDn(n+l)+n(t if: (q ( - (qk'I

n20 (q) .(xqg 2)a

(xq)t-I (E.(xq)xq)
1) (xq)
;
-
- = ~~~~~~~~~~~(xq)
f lQt,k-i+,(xq)-
Since Qk,o(x) = 0 by [2.1], we see that by [2.2]
Ql,(x) = Q1,i(xq) = Ql,1(xq2) = + E
(xq) i-I n>O (q)X
= lim Ql,(xq*) = Ql,(O) = 1. [2.3]
x (QJC-l k--+l(x* (xq2 +2) IQJ_, ,t-(xq2*+2))
-

The next formula is new and is the key to the proof of

Theorem 1. nO (q)n
X Q1-,.1-,(xq2M+2) + (xq) I- RtAt+L(xq)
Qkti(X) = E )--
(q)n
Qk-l,t(xq2*), [2.4] x( - 1) #+k -lq(,t- 1) 1" +(3- -2^) +(2k i-1)
-
n2k: 0
for 1 < i- < k, k > 2. nO (q)3
To establish [2.4] we begin by observing that if we X Qt-_,t-l(Xeq2"2) = (xq) '-R*,,lt~kf+,(xq).
AnQA
JLF%-T--JL Mathematics: Andrews Proc. Nat. Acad. Sci. USA 71 (1974)

Thus, since the Rkj(x) fulfill all the defining conditions, defined by
we see that Rt, (x) = Q, j(x) for 0 _ i . k, k _ 2.
Hence [2.4] is established Qk, ( =)- Gki(n;q)x [3.1]
Next we wish to show that for 1 _ i . k, k > 2, nZO (q)n

XNI+N2+. . . +N- qN'12+N2 2+... +N2^_1+N,+..-*+N. _

IE
QA;t(X) = nlni2,"...,n-12tO
(q)Sl(q)X2. (q)" l- .
'[2.5J.
where Nj = nj + n1+l + - + n"-i. -

When k = 2, we see by [2.4] and [2.3] that THEOREM 2. For 1 < i < k,

which is just [2.5] for k

Q2. (X) =

=
E
nO

2, i
(q)n

= 1. Furthermore by
Gk,i(n;q) =
nt+2n2+ * - -
+ (k-1)nt-1 = n Ll-2J
ni, *.,nk-.20
*

[2.2]
gk]. .. 1](q"l+')N2+ -+NV^_l
x~qS2
Q2,2(X) = Q2,1(xq-1) =
.qN 2 +- -* * +2k _ 1+N +- - * +Nk_ [3.21]
n0o (q) n

which is [2.5] for k = i = 2. Hence [2.5] is valid when

k = 2. where Nj = *-- + nt-1, and
nj + n1+l +
We now assume [2.5] is valid for k- 1. Then for (q)(q) n-'(q) .,-, the Gaussian polynomial.
1 _ i <k-, weseeby [2.4]that
x(Q-())
lq(-= (Xq2s-l),Vl+ + . - -

+k-2qn2+p2+ -
*. +w^-2+vi+ . . . +p-k2
Qt,i W = E
nk-I 20- (q)sn*_ n, * ** ,fI-22o (q)xl(q)x2 ... (q)xh-1
where = n, + n,+l + - -
* + n,_2. If we write N, =
Proof. Comparing [3.1 ] with [2.5 ], we see that
Vf + nk-1, Nk-, = n*_-, then

Q~cs:N)
= +N29+ **+N2_'l+N,+Nj,++.--+N..

ni, *-* *,nl-k O (q) 3l(q) 32 . .() h-,

which is [2.5] for 1 _ i < k. When i = k, by [2.2]

Gki(n;q) = (q)n
Qk.k(x) = Qc,,(xq-1) qNl' +-.
q +N2,;_ +N +- +Nk-_
NAI+.- * *+NtlqNi2+N22+ . * +N2'-1
= E
ni, * * * ,nm-12o (q)nl(q)n2 ... (q) nA- I
Ni+ **+Nk-1=n
ni,* * *,n& _ 0
(q) n1(q)n2... (q)nk_
E ~~~(q)n -2+nk-I
which is [2.5] for i = k. nl+2n2+ + (k l)n- =In
-
(q) n3-2(q) nlk-1
We now easily deduce Theorem 1 from [2.1 ] and [2.5]. ni, * * * ,nk-iO

co (q)nj;_s+njt_2+nj;_l (q) n**l-+ni;-

H (1 -q (q) nk 2+nk-l (q) n_. (q) n+-- *+nA (q)n,
n*O,-4--i(mod 2k+l) (qnl +- * *+nk-_l+)n2+2ns+--*+(Xt-2)ni;-l
a7

E (_ j)%ql12(12+1)n(n+l)-fS(1 - q(2*+i) 9 Nli2+. +10k-,+N;+ +Nt-,

n=O0
(q)X rNk-2lFNki31 [Nl
(by Jacobi's identity (ref. 16, pp. 169-170) ni+2nt+ - - -
+(k -1)n-i n Ln42 JLnk3J Ln. J
..0
**
ni, ,nt.
= Qx,(l) (by [2.1]) (qN +)Nr+---+Nh-l . qNfl2 + +Nf2k-I+Nf;+ +N1f 1
[3.31
-"qN+N22 +. +N.h- INi + vs4.1+"~+Nhf l - -

as desired.
nl,.*** ,n-i 2.0 (q) xl(q) x2 - - -
(q Of all the previous representations of the Alder poly-
nomials, [3.2] most closely resembles that of V. N.
(by [2.5]). Singh (ref. 12, Eq. 3.7). This is probably due to the
fact that both approaches use an iterative process. In
3. The Alder polynomials fact, we may, by the following series of algebraic
Equation [2.5] allows us to deduce easily tble following manipulations, transform Singh's representation into
representation for the Alder polynomials Gki(n;q), ours:
 Proc. Nat. Acad. Sci. USA 71 (1974) The Rogers-Ramanujan Identities 4085

Gkk(n;q) qn2 E
(q -32£..+1)21k 2
(q"2 )21+(qS1 52+1)2**
q 22(n- t) -2(1-5)--2 k-2(tk3-3k-2)

Stl ,-z**
(q2-2 (qk1 'i+l) -2 t2+1) * .* (q '-4-2 Sk-3+1)gk2
=2 ) (q~
-) ( q____
.. . q-2_1__(__-__1)_-2__
______ -3 q ( g (
1-1) -2) 13(2 -3) -2t5k2(tk-- t-2)
-q2l,.
**-2_o (q) n-2t, (q)51-22 (q) tk3-24-2(q) I(q) .
t. (q) Ik-. (ql-2 5i+1) 12(q t,-2 t2 +1) 3 . . . (qik-4-2 tk3 +1)52
(q)-qnqJ * -2 tk-2( tk-3 '2)
_ qn 52
i (n-'1) -2 12( t-12)
n-2
E
t1, ** ,t-20o (q) n-251+g,(q) 51-2j2-+-S* (9) tk4-25k.3+5gk2(q) tk3-2512(q) Sk-2
[3.4]
Now we let tV-2 = nk1, tk nk-2 + 2nk1, "What finite linear q-difference equations with poly-
tk-j-l = nk-j + 2nk-j+l + *-- + jnk- for 1 < j <
7c - 2. Thus
IC-2. Thus tk-2
t~2 = n~1,
= nk,
atndt~3 -2tk2 ~t+ -1+-
= nk-2,
2tA~..2 = nomial coefficients ... have solutions that can be
represented by 'higher dimentional' q-series?"
l
t

Thereo,
nk_>-j-eN -j. Therefore, This
tion was partially supported by National Science Founda-
workGP-23774.
Grant
n2 - 2t1(n - ti) - 2t2(t - t2) - *-- 1. Hardy, G. H. (1940) Ramanujan (Cambridge University
Press, London) (Reprinted: Chelsea, New York).
- 2tk-2(tk - tk_2) 2. Singh, V. N. (1957) "Certain generalized hypergeometric
identities of the Rogers-Ramanujan type," Pac. J. Math.
- -2(N2 + ... + Nk1)(n -N2- *-- -Nk1) 7, 1011-1014.
3. Rogers, L. J. (1917) "On two theorems of combinatory
- 2(N3 + *-- + Nk-l)N2 analysis and some allied identities," Proc. London Math.
Soc. 16, 315-336.
- 2(N4 + + Nkl)N3 4. Bailey, W. N. (1947) "Some identities in combinatory
analysis," Proc. London Math. Soc. 49, 421-435.
5. Slater, L. J. (1952) "Further identities of the Rogers-Rama-
nujan type," Proc. London Math. Soc. 54, 147-167.
- 2Nk-1 * Nk-2 6. Andrews, G. E. (1974) "On Rogers-Ramanujan type identi-
ties related to the modulus 11," Proc. London Math. Soc.,
= n2 - 2(N2+± - + Nk-l)n + 2N22 + *-- inpress.
7. Alder, H. L. (1954) "Generalizations of the Rogers-Rama-
+ 2N2k1 + 2 i NjNj nujan identities," Pac. J. Math. 4, 161-168.
2<i<j:ck- 1 8. Alder, H. L. (1969) "Partition identities-from Euler to the
present," Amer. Math. Monthly 76, 733-746.
- (n - N2 - * * * - Nk)2 + N22 + *** + N2k-l. 9. Andrews, G. E. (1974) "On the Alder polynomials and a new
generalization of the Rogers-Ramanujan identities," Trans.
Hence the above shifts in summation (ile., tk-j-1 = Amer. Math. Soc., in press.
3 10. Carlitz, L. (1960) "Note on Alder's polynomials," Pac. J.
Iin k-i-l+s) transform [3.4] into [3.3]. Math. 10, 517-519.
11. Singh, V. N. (1957) "Certain generalized hypergeometric
identities of the Rogers-Ramanujan type (II)," Pac. J.
Math. 7, 1691-1699.
4. Conclusion 12. Singh, V. N. (1959) "A note on the computation of Alder's
Apart from the fact that we now have a reasonable 13. polynomials," Pac. J. Math. 9,271-275.
13. Andrews, G. E. (1974) "A general theory of identities of the
4analytic
I generalization of the Rogers-Ramanujan Rogers-Ramanujan type," Bull. Amer. Math. Soc., in
iidentities, there is at least one other significant feature press.
Irelated the present work. 14. Ramanujan, S. & Rogers, L. J. (1919) "Proof of certain
totohepreenwok.Naelyw
related Namely,
we now have noidentities in combinatory analysis," Proc. Cambridge Phil.
Igood reason to hope that the iterative use of q-dif- Soc. 19, 211-216.
Iference equations (such as was used to prove [2.4]) 15. Selberg, A. (1936) "t8ber einige arithmetische Identitaten,"
*ill be instrumental in providing N. Vidensk. Akad. Oslo Mat. Naturvidensk. Kl. 8.
Iwill providing at partial
at least aa partial 16. Avh.
Andrews, G. E. (1971) Number Theory (W. B. Saunders,
Ianswer to Question 2 of (ref. 13, Sec. 3): Philadelphia, Pa.).