The Number of Lattice Points in a k-Dimensional Hypersphere

Author(s): W. C. Mitchell
Source: Mathematics of Computation, Vol. 20, No. 94 (Apr., 1966), pp. 300-310
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/2003508
Accessed: 18-11-2015 11:34 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact [email protected].

American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend access to Mathematics of
Computation.

http://www.jstor.org

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 The Number of Lattice Points in a
k-dimensional Hypersphere*
By W. C. Mitchell

1. Introduction. One of the most interesting problems of analytic number theory

involves the difference between the number of lattice points in a k-dimensional
hypersphere and the "volume" of the hypersphere. Define the set Lk(x) as follows:

(1) Lk() = {(J I *J2 Jk) ?J2 <x}

where the Ji are integers. Let Ak(x) be the number of distinct points in Lk (X).
Thus Ak (X) is the number of lattice points in a k-dimensional hypersphere of radius
x/2 . Define Vk(x) as the "volume" of a k-dimensional hypersphere of radius xl2.
Vk/2 k/2 7[k/21 xk/2
V
(2)
(2) (~~~~~2
) 2 (2-1 1o 2)
where [z] is the integer part of z.
The problem of primary interest is to find the Greatest Lower Bound Ok of the
set of values 0 for which
(3) Pk(X) Ak(X) - Vk(X) = 0 (X).

Walfisz [1] gives the following general results:

Pk(X) = 0(X(kl)12), Pk(X) = (Xkl2-1

=
(4) P4(X) 0(X log2X) =
0(Xl+e) E > 0,

Pk(x) = 0(xk21), k > 5.

Thus for k ? 4 Ok = k/2 - 1.
The value of k which has received the greatest attention is k = 2, the number
of lattice points in a circle. Wilton [2] gives an account of the early work in this
problem. Since that time several results have been published establishing new
values of 0 for which P2(x) = 0(x0). One of the most recent is Chen Jing-ren's
proof [3] that P2(x) = 0(x'2137). Hardy (see [2]) has shown that P2(x) = (X14
It is a common conjecture that P2 (x) = 0(x14?E), E > 0, or 02 =4
There is less known for k = 3. From (4) we have 2 < 03 < 1. Fraser and Gotlieb
[4] conjectured on the basis of numerical evidence that .5 ? 03 ? .7. More recently
Chen Jing-ren [5] has shown that 2 _ 03 < 3
With the advent of high speed computers it has become possible to evaluate
Received April 8, 1965.
* This work began with the support of a National Science Foundation grant to the Digital
Computer Laboratory of the University of Illinois for the summer of 1964 and was continued
with some financial support for auxiliary computing time from the Computing Center of Miami
(Ohio) University and from the Department of Mathematics of the California Institute of
Technology.
300

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 NUMBER OF LATTICE POINTS IN k-DIMENSIONAL HYPERSPHERE 301

Pk(X) for "large" x in order to see if the calculated results are consistent with
theoretical results or if it is reasonable to make any. new conjectures concerning ok .
There have been at least three previous papers on this subject. Fraser and Gotlieb
[41calculated isolated values of P2(x) and P3(x) for x112 < 2000 on an IBM 650.
However their conclusions differ with the present paper for 02. Harry Mitchell [6]
calculated P2 (x) for xl2 < 200,000 on an IBM 7090, but the results for x1/2 > 3000
are incorrect, as pointed out by Keller and Swenson [7]. Keller and Swenson deter-
mined P2 (x) for many values of xl/2 < 260,000 on an IBM 7090 and their method
of interpretation leads them to suggest that 02 < .3. This seems unlikely from the
results of the present method of interpretation.
The problem of establishing that Ek(x) = O(x6) is equivalent to finding a
sequence ((Xi, Yi))t'=i such that

(5) Pk(x) I _ Yi + O(x0) for x _ Xi+, and lim sup Y' < m.

Since Lk(x) is composed only of k-tuples of integers, Ak(x) is piecewise con-

stant over [n, n + 1), where n is an integer. Thus for x E [n, n + 1)

(6) Lim Pk(n + #3) < Pk(x) _ Pk(n).

But
Lim Pk(n + 3) = Ak(n) - Vk(n + 1)

(7) = Pk(n) - (Vk(n + 1) - Vk(n))

= P (n) + 0 (xk21 )

However, by (4), Ok _ k/2 - 1. Therefore the sequence of "extreme" points

(Ni, | Pk(Ni) j), defined such that IPk(n) I < I Pk(NO) I for n < Ni+1, satisfies
the first requirement of (5) for all 0 of interest. This sequence is uniquely deter-
mined, given an initial element, and Ni+1 is the first integer for which IPk (Ni) I <
IPk(Ni+1) I.
For x too large to calculate Pk(x) conveniently for all integers, approximate
extreme points can be chosen in the same manner as the true extreme points but
from a more restricted set. These approximate extreme points are not necessarily
a subset of the true extreme points. This later method, used by both Fraser and
Gotlieb and Keller and Swenson (in a different context), is not so concise as the
former but allows one to consider a larger range of x.
The present calculations on an IBM 7094 include Pk (X) for k = 2, 3, 4, 5, 6 and
all integer x ? 250,000 (x112 < 500); some 250 isolated values of P2 (X) for x1/2 <
10,000,000; and about 20 values of P3(x) for xl/2 < 9000. The results of this work
show that the calculated values follow the theoretical limits quite closely. The
results for k = 2 fail to indicate that 02 is less than Chen Jing-ren's bound of
12/37 -.324. For k = 3 the most reasonable conclusion is .5 ? 03 < .6.
Efficient algorithms for various combinations of k and x are presented in Sec-
tion 2. Section 3 is composed of computing methods and Section 4 contains con-
clusions.

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
302 w. C. MITCHELL

2. Counting Algorithms. The most efficient method of evaluating Ak(x) depends

on the range of k and x and upon whether isolated values (for approximate extreme
points) or a large number of consecutive values (for true extreme points) is desired.
The following formula is similar to one given by Walfisz [1].
Al(0) = 1,
Al(x) = Al(x - 1) + 26(x)
where
for x a perfect square,
a(X) =1
IO otherwise,
[VW]
Ak(X) = Ak1(X) + 2ZAkl(X i)X-
i=1

Formula (8) provides the basic method of calculating Ak (X). For large values
of x, some terms of the above summation may be larger than the fixed-point single-
word capacity of the computer (236 - 1 on the IBM 7094). This difficulty can
often be remedied by defining Rk(x) as the number of points (J1, J2, ***, Jk)
such that
k

(9)(9) ZJ~2=x.
~~E
E Ji2 = X.

It is evident that

Rk(x) = JAk(X) - Ak(X 1), x an integer,

O. otherwise,
(10) [x]
Ak(X) = E Rk(i).
i=O

Also
R1(O) = 1,
(11) RR(x) =26(x)

Rk(x) = R-1 (x) + 2 E Rkl (x -i2).

The similarities of (8) and (11) are noticeable. By changing initial values the same
proceduremay be used for either Ak(X) or Rk(X).
The next formula makes use of the symmetries involved in the set Lk(x) re-
sulting from permutations and negatives of ordered k-tuples. Define
(12) Lk (X) = {(J1, J2, ,Jk) E Lk (X) < J1 < J2 < *.. < Jk}

and let M (J1, J2, * , Jk) be the number of distinct permutations and negations
of J1, J2, ,Jk. Then we have
2k-(0) k
Jk) = '
(13) M(J1, J2, _ _

n(p
P-0

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 NUMBER OF LATTICE POINTS IN k-DIMENSIONAL HYPERSPHERE 303

where n(p) is the number of i for which Ji = p. Thus, if Ym= (J1, J2 , Jm),

(14) Ak(X) = r 1 = E M(Yk)

YkE Lk(X) YkE Lk (X)

By rearranging (12) and (14) it follows that

(15) Ak(X) = 1+ E Z, M(Yk-1, Jk)

Jk=1 Yk-jELk-1(X-Jk );Jk*1-Jk

Now, if Jk-l < Jk, then M(J1, J2, , Jk) = 2kM(J1, J2 *. Jk-l)., Simi-
larly, if Jk-i < Jki,+l = = Jk, then

M(J1) J2 *** Jk) = 2i ()M(Jl) J2 )** Jk-i)-

Thus we have
[Vx5- ki'15
(16) Ak(x) = 1 + E Z 2i (k) i=l
Z M(Yk-i).
Jkl_- Yk-iE Lk-i (X-jjk2 );Jk-i<Jk

Now define
Sm(Z, J) = Z M(Ym)
YmE Lm' (Z);JJm<J

Thus
k/]k
(17) Ak(x) = 1 + Z Z 2i () Sk-i(x _ ij2 J)
.J=i i=1 \It/

Sm (Z, J) can be defined recursivelyas follows:

J-1 m /
S (Z)J ) = 1 + E 2i(?) Sm-i(Z iJm2, Jm),
(18)
(~1) z 0
So(Z, J) =

It is convenient to note that

Sk(x, ??) = Ak(X),

(1 9) Sk ( oc, J) = (2J - 1)k.

Formula (18) used with (19) is the basic method for taking advantage of sym-
metries among the points of Lk (x). By algebraic reduction the following general
formula can be established:
m-1 MTNi /\
(20) Sm(Z, J) = (2N + 1)"n + E Y 2' () Sm-i(Z - iJm2, Jm)

where N = [ Z/m] and A1INM= min ([ Z/i], J - 1). For complete generality
Ak(X) must be defined as in (19). For k=2 this simplifies to

(21) A2(x) = 1 + 4[i] + 4[Vx/2]2 +8 Z [ X .J2].

Tif=oruawskont as].+1

This formula was known to Gauss.

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
304 W. C. MITCHELL

Formula (20) is of course ideally suited to programming for an algorithmic com-

piler; however, it can easily and efficiently be programmed in a machine-oriented
language. Consequently it was coded in SCATRE for the 7094 and used for isolated
values of Ak(x) for large x.
One extension of (21) should be very valuable for computing isolated values
of A2(x). It is possible, for certain x, to compute A2(x) for all u subject to
Iu I < 2x/- x1/4 in nearly the same time necessary to compute A2(x + u) alone.
For x = 1014, the largest argument used in this work, this would have made available
over 17,000 results in about twice the time required for the single value. Needless
to say, this is a significant improvement. Unfortunately, this method was not
known when the computations were done for this paper, but the method has been
used for moderate x (x 10,000).
Define the following:
x = r2 + 2r = (r + 1)2-1, ran integer,
U <2V,
U<2N2r
(22)
-U + 2 < u _ U.
[(22 2 =u
-U W ], where J is used in the context of (21).
The following theorem may be established by simple algebra:
___+__1, u > (W + 1)2- (x _

(23) + xU - J2]= W-1 u<W2 _(X j2))

tW. otherwise.

Thus, if the remainder of the integer square root routine is available, it is easy to
evaluate [V/x + u - J2] in the process of applying (21).
The true value of this method lies in computing A2(x + U) for all suitable u
simultaneously. Define
Q(0) = 0,
q(u) =0, -U + 2 ? u < U.
+ 1 to [V/x] = r, as in (21), do the following:
As J runsfrom [Nvxi72]
Q(0) = q(0) + [v/x - J2] = Q(0) + W.
(24) q(v) = q(v) + 1, where v = (W + 1)2- (x j2),

q(v) = q(v)-1, where v =W2 (x J2) -1.

Then, for u > 0,
u [v(x+u) /2 _
Q(u) =Q(0) + q(v) + [u-1]- [V+u-J2],
(25)
Q(-u) = Q(0) + E q(-v) + E [V/x u - - J2].
V=l '=[\/xu2]+l

And then, for all u,

(26) A2(x + u) = 1 + 4[ x + u] + 4[v"(x + u)/2]2 + 8Q(u).

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 NUMBER OF LATTICE POINTS IN k-DIMENSIONAL HYPERSPHERE 305

3. Computer Methods and Numerical Results. The time-consuming part of

computing Ak(x) by any of the methods mentioned in this paper is evaluating
[+/x]. However, from the logical order of summation successive arguments often
happen to be close together. Furthermore, while most square root routines are
floating-point, exact fixed-point results are necessary for this work. Thus the effi-
ciency of the square root routine may be improved by using fixed-point operations
and by using the previous result as a first approximation for the current argument.
Using this and the identity
(N - 1)2 = N2 ? 2N + 1,

Ak(x) may be calculated on a binary machine with no multiplications and no

numbers larger than xl/2 except the sum. This procedure was developed inde-
pendently by Keller and Swenson [7] and the present author. It is particularly
useful for employing (26). Keller and Swenson present the necessary algorithm
and a basic derivation of the process.
For the current paper two methods were used for determining Ak (x). For cal-
culating isolated values, (20) and the above method were used. Special square root
routines were used throughout. The time to compute Ak(X) was on the order of
(k-l) /2
T(k, x)ax
k!
When a large quantity of consecutive values was desired, (8) and (11) were
used. An IBM 1301 Disc File was available for additional storage. This Disc File
is particularly desirable in allowing the use of one portion of core-storage for com-
putation while data is moved between the Disc File and another portion of core.
For the present problem this effectively created a million words of core-storage.
The time required to compute Ak(X) for 2 ? k ? K and all integer x ? X is
T(K, X)a(k -1 )X312.

Using this method 3' hours were required for T(6, 250000). Formula (11) was used
with the assumption that Rk (X) < 236. This assumption was violated near
R6(40,000).
Integer arithmetic was used exclusively for Ak (X) in all programs, and it is
expected that all values are correct. Complete agreement was noted for all values
published in [7]. Similar agreement existed with [4] except for A3(18002), the largest
argument published in that paper. This value was calculated twice for this paper,
each calculation requiring 3 minutes.

4. Conclusions. Table 1 gives the first fifty true extreme points for k = 2, 3.
The number of extreme points for x < 250,000 (xl/2 < 500) is

k number of
extreme points
2 76
(27) 3 80
4 170
5 434
6 474 (x < 40,000).

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
306 W. C. MITCHELL

TABLE 1
First 50 extremepoints for k = 2, 3
Xi A2(Xi) P2(XX) Xi A3(Xi) P3(Xi)

1 ~~~5 2 1 7 3
2 9 3 2 19 7
5 21 5 5 57 10
10 37 6 6 81 19
20 69 6 14 251 32
24 69 - 6 21 437 34
26 89 7 29 691 37
41 137 8 30 739 51
53 177 10 54 1743 81
130 421 13 90 3695 119
149 481 13 134 6619 122
205 657 13 155 8217 134
234 749 14 174 9771 157
287 885 -17 230 14771 160
340 1085 17 234 15155 161
410 1305 17 251 16831 174
425 1353 18 270 18805 221
480 1489 -19 342 26745 252
586 1861 20 374 30551 254
840 2617 -22 461 41755 294
850 2693 23 494 46297 305
986 3125 27 550 54339 309
1680 5249 -29 666 72359 364
1843 5761 -29 750 86407 371
2260 7129 29 810 96969 405
2591 8109 -31 990 131059 580
3023 9465 -32 1890 344859 682
3024 9465 -35 2070 395231 734
3400 10717 36 2486 519963 756
3959 12401 -37 2757 607141 763
3960 12401 -40 2966 677397 776
5182 16237 -43 3150 741509 959
5183 16237 -46 3566 893019 1028
7920 24833 -48 3630 917217 1105
9796 30725 -50 4554 1288415 1120
11233 35237 -53 4829 1406811 1170
14883 46701 - 55 5670 1789599 1205
15119 47441 -57 5750 1827927 1550
15120 47441 -60 8154 3085785 1570
19593 61493 -60 8382 3216051 1576
21600 67797 -61 8774 3444439 1851
21603 67805 -63 8910 3524869 1930
21604 67805 -66 10350 4412643 2028
22177 69605 -66 10710 4645127 2404
28559 89653 - 68 15734 8269399 2411
28560 89653 -71 15750 8282167 2565
31679 99449 -74 16302 8721339 2675
31680 99449 -77 17550 9741669 2895
38015 119349 -79 23310 14910309 2905
38016 119349 -82 23894 15474065 2940
38017 119349 -85 24174 15746999 3133

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 NUMBER OF LATTICE POINTS IN k-DIMENSIONAL HYPERSPHERE 307

This reflects the increasing values of Ok and perhaps more "regularity" for the
higher values of k.
The problem of showing

Lim sup Nti) I < +00

j--j,0 N

is equivalent to showing
Lim sup (log I Pk(Ni) I-0 log Nj) < + 00.

Graphically this corresponds to finding a straight line with slope 0 which

majorizes the points (log Ni , log IPk (Ni) 1). Figure 1 shows the sequence of ex-
treme points for x < 250,000. Only a sample of the points for k = 5, 6 are shown.

8~~~~~~ I 1._6 ~~~~~~~

0~~~~~~8 2

LOGX
FIGURE 1. Extreme points for x S 250,000 and k =2, 3, 4, 5, 6.

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 308 w. C. MITCHELL

TABLE 2
xl/2 A2(x) P2(x) xl/2 A3(x) P3(x)

1000000 3141592649625 -3965 1000 4188781437 -8768

1500000 7068583465945 -4632 1200 7238202017 -27457
2000000 12566370610285 -4074 1400 11494026189 - 14133
2500000 19634954076697 -8239 1600 17157266213 - 18466
3000000 28274333873841 -8467 1800 24428982249 -42225
3500000 38484509999277 - 7198 2000 33510290993 -30645
4000000 50265482451357 -6080 2500 65449818205 -28745
4500000 63617251226505 -8688 3000 113097275709 -59820
5000000 78539816333093 -6652 3500 179594325465 -54565
5500000 95033177762429 -8662 4000 268082474393 -98713
6000000 113097335520185 -9048 4500 381703453381 -54030
6500000 132732289606241 -7928 5000 523598707861 -67737
7000000 153938040012805 -13095 5500 696909887157 -83164
7500000 176714586754401 - 10025 6000 904778525345 - 158889
8000000 201061929820913 -8834 6500 1150346427953 -82036
8500000 226980069212125 -9738 7000 1436754948853 -91389
9000000 254469004930845 -9928 7500 1767145772565 -95079
9500000 283528736973257 -13222 8000 2144660422929 - 161922
9600000 289529178944573 - 10262 8500 2572440705977 -78537
9700000 295592452772029 -4235 9000 3053627854381 -204908
9800000 301718558438929 -11835
9900000 307907495964805 -13531
10000000 314159265350589 -8390

The lines drawn represent the minimum slopes which appear to parallel the extreme
points. In addition, for k = 2, 3, 4 the theoretical minima for Ok (see (4)) are shown.
If Okis estimated from these points, the results are
02 .324 = 12/37,
03 .60 = 3/5,
(28) 04 1.06,
0. 1.52,
06 2.00.
The accuracy of visual estimation limits this method to a precision of at most
?4.01. For instance in Figure 1, for k = 2 a line with slope 12/37 would be indis-
tinguishable from one with slope 1/3. The way which the results for k = 4, 5, 6
approach the known values suggests that this method is valuable for the range
of x used.
In addition to the true extreme points for x1/2 < 500 a number of approximate
extreme points were calculated from isolated values of P2(x) and P3(x). Some of
these are shown in Table 2. In Figure 2, the values of P2(x) for x1/2 < 10,000,000
are shown with the true extreme points for xl/2 < 500. If only the approximate
extreme points are considered, one is led to agree with Fraser and Gotlieb [4] that
"02 = 4 is not inconsistent with observed results." But when the distribution of
approximate extreme points is considered independently of sampling distribution,

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
 NUMBER OF LATTICE POINTS IN k-DIMENSIONAL HYPERSPHERE 309

(9
0

LOGX
FIGuRE 2. Distribution of true and approximate extreme points for k = 2.

there is no reason to believe that the true extreme points do not continue near
1/2 12/37 000
slope _ xl/2were
for000 > 500. Thus it is not reasonable
0 to conjecture from these results
that 02 is appreciably less than Chen Jing-ren's bound of 12/37. A logical conjecture
based upon these results is 03 ? .3.
The approximate extreme points for Jo= 3 suggest that .5 ? 03 ? .6, but there
are too few points from which to extrapolate with assurance. For instance half of
the isolated values qualify as approximate extreme points. The time required to
calculate more values of P3(x) would be prohibitive.
An additional matter of interest is the sign of Pk (x ), Keller and Swenson reported
that, while most of the values of P2(x) for integer values of xl/2 ? 260,000 were
negative, the sign distribution for noninteger xl/ 'was abu uniform or perhaps
even slightly biased in favor of positive values." In this experiment all of the true
extreme values for 3400 < x ? 250,000 and all of the isolated values for integer
x1/ < 10,0000wr negative.
For k = 3, 95%O of the true extreme values were positive while the larger isolated
values were negative. The four negative extreme values were among the larger ex-
treme points.
For k = 4, 5, 6 all of the true extreme values were positive.
Another question is whether or not noninteger values of x would provide different
extreme points than the integer values used thus far. From (6) we need only con-
sider Limjso1-- Pk(n + f3). This question is of little interest for Jo - 2 because
P2(n + ,3) = P2(n) - r. However for Jo= 3, for x < 1000, the P3 (n + fi) values
were of the same magnitudeas the P3 (n) values. For the larger values of Joand
necessarily smaller x, in accordance with the greater density of extreme points as
in (27), there is an alternation of extreme points for small x and a random assort-
ment for larger x.

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions
310 W. C. MITCHELL

5. Acknowledgements. The author would like to express his appreciation to

Stephen Czuchlewski of Manhattan College (now at Yale University), J. R. Ehrman
of the University of Illinois, and T. M. Apostol of the California Institute of Tech-
nology for assistance in this work.
Department of Computer Science
University of Illinois
Urbana, Illinois
Now at
California Institute of Technology
Pasadena, California
1. A. WALFISZ, Gitterpunktein mehrdimensionalenKuglen, Monografie Matematyczne,
Vol. 33, Pa-'stwowe Wydawnictwo Naukowe, Warsaw, 1957. MR 20 #3826.
2. J. R. WILTON, "The lattice points of a circle: an historical account of the problem,"
Messengerof Mathematics,v. 58, 1929, pp. 67-80.
3. CHEN JING-REN, Science Abstractsof China, v. 2, 1964, Number 2, abstract 271.
4. W. FRASER & C. C. GOTLIEB, "A calculation of the number of lattice points in the circle
and sphere," Math. Comp.,v. 16, 1962, pp. 282-290. MR 27 #5722.
5. CHEN JING-REN, Science Abstractsof China, v. 2, 1964, Number 1, abstract 213.
6. H. L. MITCHELL, III, Numerical Experimentson theNumberof Lattice Points in theCircle,
Apple.Math. and Statist. Labs., Stanford Univ., Stanford, California, 1961.
7. H. B. KELLER & J. R. SWENSON, "Experiments on the lattice problem of Gauss,"
Math. Comp., v. 17, 1963, pp. 223-230. MR 29 #3445.

This content downloaded from 193.0.65.67 on Wed, 18 Nov 2015 11:34:41 UTC
All use subject to JSTOR Terms and Conditions