Probab. Theory Relat.

Fields 109, 27 ± 37 (1997)

Random coin tossing

Matthew Harris, Michael Keane
Department of Mathematics and Informatics, Department of Statistics, Probability and Oper-
ations Research, Delft University of Technology, Mekelweg 4, 2628 CD Delft, The Netherlands
e-mail: [email protected]; [email protected]

Received: 20 November 1996 / In revised form: 20 February 1997

Summary. A sequence of heads and tails is produced by repeatedly selecting a

coin from two possible coins, and tossing it. The second coin is tossed at
renewal times in a renewal process, and the ®rst coin is tossed at all other
times. The ®rst coin is fair (Prob heads† ˆ 1=2), and the second coin is
known either to be fair, or to have known bias h 2 0; 1Š (Prob heads†
ˆ 12 1 ‡ h†). Letting uk :ˆ Prob (There is a renewal at time k), we show that if
P 1 2
kˆ0 uk ˆ 1, we can determine, using only the sequence
P of heads and tails
produced, if the second coin had bias h or 0. If 1 2 1
kˆ0 uk < 1 ‡ h2 , we show
that this is not possible.

Mathematics Subject Classi®cation (1991): 60K35, 60G30, 60G42

1 Introduction

The research in this paper developed from the study of problems concerning
random walks on scenery. We begin with a description of this to describe the
context for this paper.
Fix d  1, and de®ne a colouring of Zd to be a map n: Zd ! f0; 1g that
assigns a colour (either pink or purple, say) to each point in Zd . Let Sn be the
position at time n of a simple random walk on Zd starting from the origin at
time 0. We call n Sn † the colour seen by the random walker at time n, and we
call the sequence fn Sn †g1nˆ0 the colour record of the walk.
Consider now the situation where Zd is coloured with one of two known
colourings n or g, and the colour record obtained is either fn Sn †g1 nˆ0 or
fg Sn †g1
nˆ0 . One can ask when one can determine, using only the colour re-
cord, which of the colourings n or g was used, with zero probability of error.
If n 0† 6ˆ g 0†, then the colouring at hand can be determined by using
only the ®rst colour in the colour record. Similarly, if all the neighbours of
28 M. Harris, M. Keane

the origin are pink in n, and purple in g, then the colouring at hand can be
determined using only the second colour in the colour record. To discount
such trivial cases, we say n and g are distinguishable if one can determine the
colouring at hand (n or g) using only the arbitrary future of the colour
record.
It is not hard to show that colourings g obtained from n by translation
and/or re¯ection in a coordinate axis are never distinguishable. Benjamini,
and independently, Keane and den Hollander, conjecture that all pairs of
colourings n; g† not related as above are distinguishable.
Benjamini and Kesten (1996) study this problem when n and g are chosen
randomly from all the possible colourings of Zd , and give results almost sure
in the choice of n and g.
Howard (1995, 1996a, 1996b) studies the case of colourings on Z. He
shows that periodic colourings are distinguishable. He also shows that col-
ourings that are obtained from periodic colourings by altering the colour at
®nitely many locations are also distinguishable. The random walker returns
to the altered regions often enough for the ``deformities'' to be detected. One
can ask if the detection of such deformities is possible in higher dimensions.
This brings us to the related model studied in this paper.
A simple random walker on Zd (d  1† has two coins, and generates a
sequence of heads and tails by tossing a coin before taking each step. If at the
origin, the second coin is tossed, and if away from the origin, the ®rst coin is
tossed. The steps of the random walker are taken to be independent of the
results of the coin tosses, and are just used to determine which coin is to be
tossed. It is known that the ®rst coin is fair (Prob heads† ˆ 1=2), and the
second coin is either fair, or has a particular bias h 2 0; 1Š, (Prob heads†
ˆ 12 1 ‡ h†). We ask whether one can determine, just using the sequence of
heads and tails obtained, whether or not the second coin had bias h.
The second coin is only tossed when the random walker is at the origin. In
the colouring problem above, the ®nite regions where colourings n and g
di€er are only observed when the walker returns to these regions.
The simple random walk in three dimensions returns to the origin only
®nitely often almost surely, and thus the second coin is tossed only ®nitely
often. We could thus not hope to determine if the second coin had bias h. The
simple random walk in one or two dimensions does return to the origin
in®nitely often almost surely, and the second coin is tossed in®nitely often
almost surely. We prove that if d ˆ 1, one can determine whether or not the
second coin has bias h, but if d ˆ 2, one cannot.
The roÃle of the random walk above was just to determine the times that
the second coin is tossed. We generalise this to the case when the second coin
is tossed at renewal times in a renewal P process. If uk :ˆ Prob there is
a renewal at time k†, we show thatP if 1 2
kˆ0 k ˆ 1, then we can determine if
u
1
the second coin had bias h, and if kˆ0 uk < 1 ‡ h12 , then we cannot.
2

We see that this is almost a dichotomy. We P believe that this should

actually be a dichotomy, and conjecture that if 1 u
kˆ0 k
2
< 1, then it is not
possible to determine if the second coin has bias h for all h 2 0; 1Š. Our
Random coin tossing 29

ÿ ÿP 2
ÿ1=2

results only give this for small h (h 2 0; uk ÿ 1 . The extension of
this to h 2 0; 1Š remains an interesting open problem.
In all of the above, we have assumed that the second coin is either fair, or
has a particular bias h. One may ask, given the sequence of heads and tails
produced, if one can determine the value of h 2 ‰0; 1Š, given no prior
knowledge of its value. The methods of Howard (1995, 1996a, 1996b) give
that this is possible if the second coin is tossed at return times to the origin of
a simple random walk on Z.
In the next section (Section 2) we introduce the required notation,
making rigorous the above discussion, and we state the two main theorems
of the paper. In Section 3 we prove Theorem 1, and in Section 4 we give an
outline of Howard's methods that can be used to prove Theorem 2. In
Section 5 we conclude the paper by stating some open problems.

2 Notation and de®nitions

We represent the space of sequences of coin tosses by X :ˆ fÿ1; ‡1gN . We

let Xn : X ! fÿ1; ‡1g be the random variable de®ned by Xn x† :ˆ xn for
x ˆ x0 ; x1 ; . . .† 2 X , and let F :ˆ r X0 ; X1 ; . . .†. The random variable Xn
gives the n ‡ 1†st coordinate of x 2 X ± the result of the n ‡ 1†st coin toss.
We will de®ne measures on X ; F†. It will be sucient to de®ne the
measures on cylinder sets
\
n
‰x0 ; x1 ; . . . xn Š :ˆ ‰Xi ˆ xi Š 1†
iˆ0

for n ˆ 0; 1; . . .. This is because, for ®xed n, the cylinder sets above generate
the r-algebra Fn :ˆ r X0 ; . . . ; Xn †, and the sequence Fÿ1 :ˆ f;; X g;
F0 ; F1 ; . . . is a sequence of r-algebras increasing to F.
We will de®ne measures lh indexed by a parameter h 2 ‰0; 1Š, called the
bias of the second coin. As a ®rst step, we introduce measures lh;d on X ; F†,
for d ˆ d0 ; d1 ; . . .† 2 D :ˆ f0; 1gN . De®ne
Yn
1
lh;d ‰x0 ; . . . ; xn Š† :ˆ 1 ‡ hxk dk † : 2†
kˆ0
2
To understand this de®nition, we see that
lh;d ‰x0 ; . . . ; xn Š† ˆ lh;d ‰X0 ˆ x0 Š† . . . lh;d ‰Xn ˆ xn Š† 3†
is a product measure. If dk ˆ 0, then lh;d ‰Xk ˆ ‡1Š† ˆ lh;d ‰Xk ˆ ÿ1Š† ˆ 12,
and we say that the fair coin was used for the k ‡ 1†st toss. If dk ˆ 1, then
lh;d ‰Xk ˆ ‡1Š† ˆ 12 1 ‡ h† ˆ 1 ÿ lh;d ‰Xk ˆ ÿ1Š†, and we say that a coin with
bias h was used for the k ‡ 1†st toss. The sequence d 2 D determines the
times when the second coin with bias h was tossed.
To de®ne lh we wish to randomise d 2 D. We let G be the r-algebra on D
generated by cylinder sets on D (de®ned as for X ), and we de®ne a measure P
on D; G† by de®ning it on cylinder sets in G.
30 M. Harris, M. Keane

Let f ˆ f1 ; f2 ; . . .† be a probability vector, and de®ne the sequence

u ˆ u0 ; u1 ; . . .† as follows:

u0 :ˆ1
Xn
un :ˆ fk unÿk n1 : 4†
kˆ1

We take f to be the probability distribution for the ®rst renewal time in a

renewal process, and then un is then the probability that there is a renewal at
time n. If, for example, the renewal times are return times to the origin of a
simple random walk on Z, then
  2n
2n 1
u2n ˆ
n 2
f2n ˆ u2n ÿ u2nÿ2
u2nÿ1 ˆ f2nÿ1 ˆ 0 5†
for n  1.
We now de®ne the measure P on D; G† as follows: given a set
n1 < n2 < . . . of positive integers,
P ‰D0 ˆ Dn1 ˆ . . . Dnk ˆ 1Š† :ˆ un1 un2 ÿn1 . . . unk ÿnkÿ1 : 6†
The cylinder set in (6) is the event that there are renewals at times 0; n1 ; . . . ; nk
(and possibly at other times too).
We now let lh be the measure lh;d where d is chosen from D randomly
according to P . More precisely, we de®ne
Z
lh ‰x0 ; . . . ; xn Š† :ˆ lh;d ‰x0 ; . . . ; xn Š† dP d† 7†
D
Z Y n
1
ˆ 1 ‡ hxk Dk † dP : 8†
D kˆ0 2

A sequence of heads and tails generated as described in the introduction

using a second coin with bias h > 0 corresponds to drawing an x 2 X ac-
cording to measure lh . Using a fair (h ˆ 0) second coin corresponds to
drawing an x 2 X according to the measure l0 . For h > 0 ®xed, we say that
the second coin can be determined if the measures lh and l0 are singular with
respect to each other, written lh ? l0 . There then exists an A 2 F such that
lh A† ˆ 0 and l0 Ac † ˆ 0, (Ac is the complement of A) and the generating
measure of x 2 X can be determined with zero probability of error by de-
termining whether or not x is in A.
If, for all B 2 F, l0 B† ˆ 0 ) lh B† ˆ 0, we say that lh is absolutely
continuous with respect to l0 , and we write lh  l0 . If lh  l0 , then a set A
as above cannot exist, and the second coin can not be determined with zero
probability of error.
We can now state the main theorem of the paper.
Random coin tossing 31

Theorem 1 For h 2 0; 1Š,

P
1. If P1 2
kˆ0 uk ˆ 1, then lh ? l0 .
1
2. If kˆ0 u2k < 1 ‡ h12 , then lh  l0 .
P  P 
Note that if 1 kˆ0 ku 2
< 1, then l h  l 0 for all h 2 0; u2
k ÿ 1† ÿ1=2
,
and the second coin can not be determined P with zero probability of error if
its bias is too small. Also, note that if 1 2
kˆ0 uk < 2, then lh  l0 for all
h 2 0; 1Š.
Theorem 1 can be compared to the Kakutani (1948) dichotomy con-
cerning two measures a and b on X that make the coordinates Xn indepen-
dent. In this case either a ? b or a  b. Under lh , the coordinates Xn are
dependent, and Theorem 1 can be thought of as being a relative of Kaku-
tani's result.
If the renewal times are taken to be return
P times to the origin of a simple
random walk on Z2 , then we will see that 1 u
kˆ0 k
2
< 2, and Theorem 1 gives
us
Corollary 1 If the renewals are returns to the origin of a simple random walk on
Zd , then, for all h 2 0; 1Š:

1. If d ˆ 1, then lh ? l0 ,
2. If d  2 then lh  l0 .
Given x 2 X drawn according to either lh (h known), or l0 , Theorem 1
tells us when the underlying measure can be determined from x with zero
probability of error. One may ask for a sequence x 2 X drawn according to
the measure lh for h 2 ‰0; 1Š unknown, can the value of h be determined from
x with zero probability of error? We answer this for the case when the
renewal times are return times to the origin of a simple random walk on Zd .
From Corollary 1, we see that the answer to the above question is clearly `no'
for the case when d  2. Theorem 2 tells us that the answer is `yes' for d ˆ 1.
Theorem 2 If the renewals are returns to the origin of a simple random walk on
Z, then there exists an f : X ! ‰0; 1Š, such that for all h 2 ‰0; 1Š, f is lh
measurable, and lh ‰f ˆ hŠ† ˆ 1.
Theorem 2 can be proved using methods in Howard (1995, 1996a, 1996b),
and we only outline the proof in this paper.

3 Proof of Theorem 1

We prove Theorem 1 using a martingale argument. De®ne

lh ‰x0 ; . . . ; xn Š†
qn x† :ˆ 9†
l0 ‰x0 ; . . . ; xn Š†
for X 3 x ˆ x0 ; x1 ; . . .†, so that
32 M. Harris, M. Keane

Z
lh A† ˆ qn x† dl0 10†
A

for A 2 Fn . We see that qn is just the Radon-NikodyÂm derivative of lh with

respect to l0 restricted to Fn . Now qn is a positive l0 martingale with
®ltration Fn and has a ®nite limit q l0 almost surely. To de®ne q on the
whole of X , de®ne
q x† :ˆ lim sup qn x† : 11†
n!1

The Lebesgue decomposition of lh with respect to l0 can be written

lh A† ˆ lch A† ‡ l?
h A† 12†
for A 2 F, with
Z
lch A† :ˆ q dl0 13†
A
l?
h A† :ˆlh A \ ‰q ˆ 1Š† : 14†

It is clear that lch  l0 and, as l0 ‰q ˆ 1Š† ˆ 0, we have l?

h ? l0 .
We prove the ®rst part of Theorem 1 by showing that l0 ‰q ˆ 0Š† ˆ 1. If
this is the case, lh ˆ l?h and lh ? l0 .
To prove the second part of Theorem 1, we show that qn is uniformly
integrable. If this is the case, then
Z Z
lh A† ˆ lim qn dl0 ˆ q dl0 15†
n!1 A A
S
for A in the p-system n Fn , and then for all A 2 F. This means that lh ˆ lch
and lh  l0 .

3.1 Frequent renewals

In this section we prove the ®rst part of Theorem 1. We begin by proving the
following preliminary lemma.
Lemma 1 l0 ‰q ˆ 0Š† ˆ 0 or 1:
Proof To prove Lemma 1 we show that the event ‰q ˆ 0Š is a tail event, and
use the Kolmogorov 0-1 Law. De®ne

F0m :ˆ r Xm ; Xm‡1 ; . . .† 16†

and let
\
1
T :ˆ F0m 17†
mˆ0

be the tail r-algebra. We show that the event ‰q ˆ 0Š 2 T. Consider

x ˆ x0 ; x1 ; . . . ; xnÿ1 ; xn ; xn‡1 ; . . .† 2 ‰q ˆ 0Š
Random coin tossing 33

and de®ne
y :ˆ x0 ; x1 ; . . . ; xnÿ1 ; ÿxn ; xn‡1 ; . . .† :
Then
Z Y
m Z Y
m
1ÿh
lim sup 1 ‡ hxk Dk † dP  lim sup 1 ‡ hyk Dk † dP
1 ‡ h m!1 D kˆ0 m!1 D kˆ0
Z Y
m
1‡h
 lim sup 1 ‡ hxk Dk † dP 18†
1 ÿ h m!1 D kˆ0

Thus, q y† ˆ 0 if and only if q x† ˆ 0, and the eventT‰q ˆ 0Š is independent

of the nth coordinate. This means that ‰q ˆ 0Š 2 1 0
mˆ0 Fm ˆ T. As X0 ;
X1 ; . . . are independent under l0 , the Kolmogorov 0-1 Law proves the
lemma. (
Now consider the integral A n†:
Z
A n† :ˆ qn x†qn ÿx† dl0 x†
X
Z Z Z Y n
ˆ 1 ‡ hxk Dk † 1 ÿ hxk D0k † dP 0 dP dl0 x†
X D D0 kˆ0
Z Z Y
n Z
ˆ 1 ‡ hxk Dk ÿ hxk D0k ÿ h2 Dk D0k † dl0 x† dP 0 dP
D D0 kˆ0 X
Z Z Y n
ˆ 1 ÿ h2 Dk D0k † dP 0 dP : 19†
D D0 kˆ0

The product space D  D; G  G; P  P † is represented as D; D0 †; G; G0 †;

P ; P 0 ††, so the probability space D0 ; G0 ; P 0 † above is just a copy of D; G; P †.
The third line above follows from the independence of the x coordinates
under the l0 measure and Rthe fact that x2k ˆ 1 for all k. The last line is
obtained by observing that X xk dl0 x† ˆ 0.
We observe that the simultaneous renewal times of two independent re-
newal processes are renewal times of a new renewal process, and then
 ÿP1
P
2 ÿ1
0
P  P ‰Dk Dk ˆ 0 8k > 0Š† ˆ kˆ0 uk if P u2k < 1 20†
0 if u2k ˆ 1 :
P
(See, for example, Kingman (1972), Theorem 1.5.) If 1 2
kˆ0 uk ˆ 1, we see
0
P  P ‰Dk Dk ˆ 1 for some k > 0Š† ˆ 1, and as P  P represents a renewal
process,
P  P ‰Dk D0k ˆ 1 for infinitely many kŠ† ˆ 1 : 21†
and
" #!
Y
n
2
P P 1ÿh Dk D0k † !0 ˆ1 : 22†
kˆ0
34 M. Harris, M. Keane

Q
As nkˆ0 1 ÿ h2 Dk D0k † is bounded from above by 1, it follows that A n† ! 0.
Now, from Fatou's Lemma, we obtain
Z
0 ˆ lim inf qn x†qn ÿx† dl0 x†
n X
Z
 lim inf qn x†qn ÿx† dl0 x†
n
ZX
ˆ q x†q ÿx† dl0 x† : 23†
X

The last equality holds as we know that qn ! q l0 almost surely. As qn is a

positive martingale, this implies
Z
q x†q ÿx† dl0 x† ˆ 0 : 24†
X

It is now clearly not the case that l0 ‰q ˆ 0Š† ˆ 0, so, by Lemma 1, we must
have that l0 ‰q ˆ 0Š† ˆ 1, and lh ? l0 . (
Remark 1 The fact that the P originates from a renewal process has not been
used. All that was used was

P  P ‰Dk D0k ˆ 1 for infinitely many kŠ† ˆ 1 : 25†

If P is a measure on D; G† that satis®es (25), then lh ? l0 .

3.2 Infrequent renewals

In
P1this 2section 1we prove the second part of Theorem 1. We show that if
kˆ0 uk < 1 ‡ h2 , then qn is an L2 -bounded martingale, and therefore uni-
formly integrable. This, as already mentioned, will give us that lh  l0 .
Let Rn be the L2 -norm of qn . Analogous to (19), we obtain
Z
Rn :ˆ q2n x† dl0 x†
X
Z Z Y n
ˆ 1 ‡ h2 Dk D0k † dP 0 dP : 26†
D D0 kˆ0

Note that the terms in thePabove product have value either 1 or 1 ‡ h2 , the
0 1 2
latter when Dk Dk ˆ 1. If kˆ0 uk < 1, then (20) gives that
!ÿ1
X1
0 2
a :ˆ P  P ‰Dk Dk ˆ 1 for some k > 0Š† ˆ 1 ÿ uk <1 ; 27†
kˆ0

and, as P  P represents a renewal process, we see that

P  P ‰Dk D0k ˆ 1 for exactly m values of k > 0Š† ˆ am 1 ÿ a† : 28†
Random coin tossing 35

Clearly
Z Z Y
1
R1 :ˆ 1 ‡ h2 Dk D0k † dP 0 dP 29†
D D0 kˆ0

is an upper bound for Rn , n  0, and

X
1
R1 ˆ 1 ‡ h2 †m‡1 P  P Dk D0k ˆ 1 for exactly m values of k > 0†
mˆ0
X
1
ˆ 1 ‡ h2 †m am 1 ÿ a† 1 ‡ h2 †
mˆ0
<1 30†
2 2
if a 1 ‡ h † < 1. We note the extra factor 1 ‡ h † in the above is due to the
fact that D0 D00 ˆ 1 always. From the de®nition of a (equation (28)), wePsee
that the condition that a 1 ‡ h2 † < 1 is equivalent to the condition on u2k
given in the theorem. We have thus shown that in this case, qn is an L2 -
bounded martingale, and hence lh  l0 . (

3.3 The simple random walk special case

In this section we prove Corollary 1. For un as de®ned in (5) (the renewal

times are taken to
P be return times to the origin of a simple random walk on
Z), we see that 1 u
kˆ0 k
2
ˆ 1, and Theorem 1 proves the ®rst part of Cor-
ollary 1.
If the renewal times are taken to be return times to the origin of the
simple random walk on Z2 , it is standard (see Feller (1957) page 328) that
  2n !2
2n 1
u2n ˆ
n 2
u2n‡1 ˆ 0 31†
for n  0. Then
X
1 X1   2n !4
2n 1
u2k ˆ
kˆ0 kˆ0 n 2
X1
1
1‡
kˆ1
p n2
2

7
ˆ : 32†
6
The inequality is obtained fromP bounds on n! obtained by Robbins (1955)
(see also Feller (1957)). Thus 1 u
kˆ0 k
2
< 2, and as noted
P after2 the statement
of Theorem 0, this implies that lh  l0 . The value of 1 kˆ0 uk is clearly not
larger than 7=6 for the cases when the renewal times are taken to return times
36 M. Harris, M. Keane

to the origin of the simple random walk on Zd for d  3, completing the

proof of the corollary. (

4 Calculation of h

In this section we sketch the proof of Theorem 2. The methods used are those
of Howard (1995, 1996a, 1996b).
We consider the case when the renewals are return times to the origin of a
simple random walk on Z. Thus un is as in (5). We de®ne
X
n
Un :ˆ uk 33†
kˆ1

to be the expected number of returns to the origin from time 1 up to and

including time n. This is also the expected number of times that the second
coin is tossed between these times. Thus
1 X n
Nn x† :ˆ xk 34†
Un kˆ1

has expectation h under measure lh . This is true for all h 2 ‰0; 1Š.
Un
It can be shown that lim infn!1 p n
> 0, a property of the un in (5), implies
that Nn x† is almost independent of Nm x† for n much larger than m, and that
Nn x† has bounded variance under lh , uniform in n and h. A dependent weak
law of large numbers gives the existence of a sequence ak , a1 < a2 < . . ., such
that
1X n
Mn x† :ˆ Na x† ! h 35†
n iˆ1 i

in probability. The same sequence ai works for all h 2 ‰0; 1Š. Taking a further
subsequence gives almost sure convergence. P1 2 (
Un
As lim infn!1 p n
> 0 does not follow from u
nˆ0 n ˆ 1, we have not
shownP that h can be determined with zero probability of error more generally
when 1 2
nˆ0 un ˆ 1. It remains an interesting question as to whether or not
this can be done.

5 Questions

TheoremP 1 is unsatisfactory in the sense that the situation where

1 ‡ h12  1 2
kˆ0 uk < 1 is not covered. We believe that Theorem 1 should
actually be a dichotomy, and the second part should be replaced by
X
1
if u2k < 1; then lh  l0 : 36†
kˆ0
Random coin tossing 37

To show that lh  l0 , we showed that qn is an L2 -bounded

P martingale. It is
conceivable that for h 2 0; 1Š such that 1 ‡ h12  1 2
kˆ0 uk < 1, qn is L1‡
bounded for some  2 0; 1Š, and it would also follow that lh  l0 . We have
not been able to show this.
We could consider the situation in which we know the second coin has
bias either h or u 2 0; 1Š. We may then ask when lh  lu , and when
lh ? lu . More generally, we could ask when the bias of the second coin can
be determined, with zero probability of error, from the sequence of heads
and tails with no prior knowledge about its bias. This would generalise
Theorem 2.

6 References

[1] Benjamini, I., Kesten, H.: Distinguishing sceneries by observing the scenery along a random
walk path. J. Anal. Math., 69, 97±135 (1996)
[2] Feller, W.: An introduction to Probability: Theory and Examples. New York: Wiley (1957)
[3] Howard, D.: The orthogonality of measures induced by random walks with scenery. PhD
thesis, Courant Inst. Math. Sciences, New York, NY. (1995)
[4] Howard, D.: Detecting defects in periodic scenery by random walks on Z, Random
Structures Algorithms, 8 no 1, 59±74 (1996a)
[5] Howard, D.: Orthogonality of measures induced by random walks with scenery Combin.
Probab. Comput., 5 no 3, 247±256 (1996b)
[6] Kakutani, S.: On equivalence of in®nite product measures. Ann. of Math. II. Ser 49, 214±
224 (1948)
[7] Kingman, J.: Regenerative Phenomena, John Wiley and Sons (1972)
[8] Robbins, H.: A remark on Stirling's formula, American Mathematical Monthly, 62, 26±29
(1955)