Rademacher functions

Jordan Bell
jordan.bell@gmail.com
Department of Mathematics, University of Toronto
July 16, 2014

Binary expansions

Define S : {0, 1}N [0, 1] by

S() =

X
k
k=1

2k

{0, 1}N .

For example, for 1 = 0 and 2 = 1, 3 = 1, . . .,

S() =

0 1 1
1
+ + + = ;
2 4 8
2

for 1 = 1 and 2 = 0, 3 = 0, . . .,
S() =

1 0 0
1
+ + + = .
2 4 8
2

Let {0, 1}N . If there is some n

k n + 1, then defining

k
k = 1

0
we have
S() =

n1
X
k=1

N such that n = 0 and k = 1 for all

k n1
k=n
k n,

n1
X
X k
k
1
1
+
=
+ n = S( ).
2k
2k
2k
2
k=n+1

k=1

One proves that if either (i) there is some n N such that n = 0 and k = 1
for all k n + 1 or (ii) there is some n N such that n = 1 and k = 0 for all
k n + 1, then S 1 (S()) contains exactly two elements, and that otherwise
S 1 (S()) contains exactly one element.
In words, except for the sequence whose terms are only 0 or the sequence
whose terms are only 1, S 1 (S()) contains exactly two elements when is
1

eventually 0 or eventually 1, and S 1 (S()) contains exactly one element otherwise.

We define  : [0, 1] {0, 1}N by taking (t) to be the unique element of
1
S (t) if S 1 (t) contains exactly one element, and to be the element of S 1 (t)
that is eventually 0 if S 1 (t) contains exactly two elements. For k N we define
k : [0, 1] {0, 1} by
k (t) = (t)k ,
t [0, 1].
Then, for all t [0, 1],
t = S((t)) =

X
k (t)
k=1

2k

(1)

which we call the binary expansion of t.

Rademacher functions

For k N, the kth Rademacher function rk : [0, 1] {1, +1} is defined by

rk (t) = 1 2k (t),

t [0, 1].

We can rewrite the binary expansion of t [0, 1] in (1) as

X
rk (t)
k=1

2k




X
X
1
k (t)
k (t)

= 1 2t.
=
1

2
2k
2k
2k

k=1

(2)

k=1

Define r : R {1, +1} by

r(x) = (1)[x] ,
where [x] denotes the greatest integer x. Thus, for 0 x < 1 we have
r(x) = 1, for 1 x < 2 we have r(x) = 1, and r has period 2.
Lemma 1. For any n N,
rn (t) = (1)[2

t]

= r(2n t),

t [0, 1]

In the following theorem we use the Rademacher functions to prove an identity for trigonometric functions.1
Theorem 2. For any nonzero real x,

Y
k=1
1 Mark

cos

sin x
x
.
=
2k
x

Kac, Statistical Independence in Probability, Analysis and Number Theory, p. 4,

3.

Proof. Let n N and let c1 , . . . , cn R. The function

n
X

ck rk

k=1

is constant on each of the intervals



s s+1
,
,
2n 2n

0 s 2n 1.

(3)

There is a bijection between n = {1, +1}n and the collection of intervals (3).
Without explicitly describing this bijection, we have
!
!
n
Z 1
2X
1 Z (s+1)2n
n
n
X
X
ck rk (t) dt
exp i
ck rk (t) dt =
exp i
0

s2n

s=0

k=1

k=1

X 1
exp i
2n

!
k ck

k=1

n
X Y
eik ck
2

n
X

n k=1
n
ick
Y

k=1

+ eick
,
2

giving
1

exp i
0

We have
Z 1
e

ix(12t)

n
X

!
ck rk (t) dt =

k=1

dt = e

cos ck .

(4)

k=1

2ixt 1


 2ix

1
sin x
 = eix e
+
=
.
2ix 0
2ix
2ix
x

ix e

n
Y

(5)

Pn
Using (2) we check that the sequence of functions k=1 rk2(t)
converges unik
formly on [0, 1] to 1 2t, and hence using (5) we get
!
Z 1
Z 1
n
X
rk (t)
sin x
exp ix
dt

eix(12t )dt =
k
2
x
0
0
k=1

as n . Combining this with (4), which we apply with ck =

n
Y
k=1

cos

x
sin x

2k
x

as n , proving the claim.

x
,
2k

we get

We now give an explicit formula for the measure of those t for which exactly
l of r1 (t), . . . , rn (t) are equal to 1.2 We denote by Lebesgue measure on R.
We can interpret the following formula as stating the probability that out of n
tosses of a coin, exactly l of the outcomes are heads.
Theorem 3. For n N and 0 l n,
 
1 n
.
{t [0, 1] : r1 (t) + + rn (t) = 2l n} = n
l
2

Proof. Define : [0, 1] R by

2

1
(t) =
2

Pn

eix((2ln)+

But for m Z,
1
2
hence

k=1

rk (t))

dx

imx

dx = m,0 =

(
1
(t) =
0

Therefore


n
X
t [0, 1] :
rk (t) = 2l n

1
0

m=0
m 6= 0,

(6)

Pn
rk (t) = 2l n
Pk=1
n
k=1 rk (t) 6= 2l n.

(t)dt
0

k=1

Z 2
Pn
1
eix((2ln)+ k=1 rk (t)) dxdt
0 2 0
Z 2
Z 1
Pn
1
eix(2ln)
eix k=1 rk (t) dtdx
2 0
0
Z 2
1
eix(2ln) cosn xdx;
2 0

Z
=
=
=

the last equality uses (4) with c1 = x, . . . , cn = x. Furthermore, writing

n

cos x = 2

(e

ix

+e

ix

)=2

n  
X
n ix(2kn)
e
,
k

k=0
2 Mark

Kac, Statistical Independence in Probability, Analysis and Number Theory, pp. 89.

we calculate using (6) that

1
2

ix(2ln)

cos xdx

Z 1
n  
X
n 1
= 2
eix(2ln) eix(2kn) dx
k 2 0
k=0
Z 1
n  
X
n 1
eix(2k2l) dx
= 2n
k 2 0
k=0
n  
X
n
= 2n
2k2l,0
k
k=0
n  
X
n
n
= 2
k,l
k
k=0
 
n n
= 2
,
l
n

proving the claim.

We now prove that the expected value of a product of distinct Rademacher
functions is equal to the product of their expected values.3
Theorem 4. If k1 , . . . , kn are positive integers and k1 < < kn , then
Z 1
rk1 (t) rkn (t)dt = 0.
0

Proof. Write J =

R1
0

rk1 (t) rkn (t)dt and define

(x) =

n
Y

r(2ks k1 x),

x R,

s=2

which satisfies
(x + 1) =

n
Y

r(2ks k1 x + 2ks k1 ) =

s=2
3 Masayoshi

n
Y

r(2ks k1 x) = (x).

s=2

Hata, Problems and Solutions in Real Analysis, p. 185, Solution 13.2.

Hence, as has period 1 and r has period 2,

Z 1
J =
rk1 (t)(2k1 t)dt
0
1

r(2k1 t)(2k1 t)dt

=
0

=
=

=
=

1
2k1
1
2k1

2k1

r(x)(x)dx
0
1
2k1X
1 Z 2j+2

1
2k1X
1 Z 2
1
r(x)(x)dx
2k1 j=0
0
Z
1 2
r(x)(x)dx.
2 0

But, as has period 1,

Z 2
Z 1
Z
r(x)(x)dx =
(x)dx
0

r(x)(x)dx

2j

j=0

Z
(x)dx

(x)dx =
0

(x)dx = 0,
0

hence J = 0, proving the claim.

For each n N, if f is a function defined on the integers we define
!
Z 1
n
X
In (f ) =
f
rk (t) dt.
0

k=1

Lemma 5. For any n N,

In (x2 ) = n,

In (x4 ) = 3n2 2n.

Proof. Using Theorem 4 we get

2

In (x )

Z
=
0

Z
=

n
X

!2
rk (t)

dt

rk (t)2 +

k=1
n
1X

0 k=1
Z 1X
n

j6=k

rk (t)2 dt

0 k=1

= n.

rj (t)rk (t)dt

Using Theorem 4 we get, since rj (t)4 = rj (t)2 = 1 and rj (t)3 = rj (t),

!4
Z 1 X
n
In (x4 ) =
rk (t) dt
0

Z
=

k=1
n
1X

rk (t)4 +

0 k=1

 X
 X
n X
n X
4
4
rj (t)3 rk (t) +
rj (t)2 rk (t)2
3 j=1
2 j=1
k6=j

k6=j

 X
n
X
4
+
rj (t)2 rk (t)rl (t)
2 j=1
j, k, l all distinct
X
+
rj (t)rk (t)rl (t)rm (t)dt
j, k, l, m all distinct

= n+

 
4
n(n 1).
2

Our proof of the next identity follows Hata.4

Lemma 6. For any n N,
In (|x|) =

1 cosn x
dx.
x2

Proof. For n N and c1 , . . . , cn R,

!
!
Z 1
Z 1
Z
n
n
X
X
ck rk (t) dt =
exp i
cos
ck rk (t) dt+i
0

k=1

k=1

sin

n
X

!
ck rk (t) dt,

k=1

and since (4) tells us that the left-hand side of the above is real, it follows that
we can write (4) as
!
Z 1
n
n
X
Y
cos
ck rk (t) dt =
cos ck .
(7)
0

k=1

k=1

Suppose that is a positive real number. Using t = x and doing integration

by parts,
Z
Z
1 cos x
1 cos t
dx
=

dt
2
x
t2
0
0

Z
1 cos t 
sin t
=
dt
+

t 0
t
0
Z
sin t
=
dt
t
0

= .
2
4 Masayoshi

Hata, Problems and Solutions in Real Analysis, p. 188, Solution 13.6.

It is thus apparent that for any real ,

Z
1 cos x

dx = || .
2
x
2
0
Pn
For any n N, applying the above with = k=1 rk (t) we get

Z  n

2 1 X

rk (t) dt
In (|x|) =


2
0
k=1
Pn
Z Z
2 1 1 cos (x k=1 rk (t))
dxdt
=
0 0
x2
!
Z
Z 1
n
X
2 1
=
1 cos x
rk (t) dtdx
0 x2 0
k=1
Z
2 1
(1 In (cos x)) dx.
=
0 x2
Applying (7) with ck = x for each k, this is equal to
!
Z
Z
n
Y
2 1
2 1 cosn x
cos
x
dx
=
1

dx,
0 x2
0
x2
k=1

completing the proof.

We use the above formula for In (|x|) to obtain an asymptotic formula for
In (|x|).5
Theorem 7.

r
In (|x|)

2
n.

Proof. By Lemma 6,
Z
2 1 cosn x
In (|x|) =
dx.
0
x2



For 0  < 1, define  : 0, 2 R by
 (x) =

x2
+ log cos x.
2(1 )

We also define

 = arccos 1 ,

 =


1 cosn x
dx,
x2

5 Mark Kac, Statistical Independence in Probability, Analysis and Number Theory, p. 12,
Masayoshi Hata, Problems and Solutions in Real Analysis, p. 188, Solution 13.6.

and for > 0,



Z
K, =



2
1 exp nx

dx.

x2

Let 0 <  < 1. Until the end of the proof, at which point
we take
 0, we
shall keep  fixed. For 0 < x <  we have, using arccos 1  ,
0 (x) =

x2


1
+ log cos x < + log 1  = + log(1 ) < 0,
2
2
2 2

hence

 2
x
.
cos x < exp
2

On the other hand,

x
tan x,
1

0 (x) =

1
sec2 x,
1

00 (x) =

so  (0) = 0 (0) = 0 and 00 (t) > 0 for all 0 t <  , giving
 (x) > 0,
and hence

x2
exp
2(1 )



< cos x.

Collecting what we have established so far, for 0 < x <  we have



 2
x2
x
exp
< cos x < exp
.
2(1 )
2
This shows that
Z

K,2(1) =



nx2
1 exp 2(1)
x2

Z
dx
0

and therefore
K,2(1) + 

1 cosn x
dx,
x2

In (|x|).
2

On the other hand,

Z
K,2 =



2
1 exp nx2
x2

dx
0

1 cosn x
dx,
x2

so

In (|x|).
2
Now summarizing what we have obtained, we have
K,2 + 

K,2 + 

In (|x|) K,2(1) +  .
2
9

(8)

p
For > 0, doing the change of variable t = n x,


r Z n 
Z  1 exp nx2
2

1 et
n
dx
=
dt.
K, =
x2
0
t2
0
As n , the right-hand side of this is asymptotic to
r Z
r
2
n
1 et
n
.
dt =
0
t2

Dividing (8) by n and taking the limsup then gives

r
In (|x|)

lim sup

,
2(1 )
n
n 2
or
In (|x|)

lim sup
n
n
indeed  depends on n, but  <
 0 yields

2
 ,

2
;
(1 )

which does not depend on n. Taking

r

In (|x|)
lim sup

n
n

2
.

On the other hand, taking the liminf of (8) divided by

r
In (|x|)

lim inf
,

n 2
2
n
or
In (|x|)
lim inf

n
n

n gives

2
.

Combining the limsup and the liminf inequalities proves the claim.
Lemma 8. For any R and n N,
In (e|x| ) < In (ex ) + In (ex ) = 2(cosh )n .

We will use the following theorem to establish an estimate similar to but

weaker than the law of the iterated logarithm.6
Theorem 9. For any  > 0, for almost all t [0, 1],

!
r

n

X
1
2 log n X

exp
r
(t)

 dt < .
k

n2+
n 
n=1

6 Masayoshi

k=1

Hata, Problems and Solutions in Real Analysis, p. 189, Solution 13.7.

10

Proof. Define fn : [0, 1] (0, ) by

fn (t) =

1
n2+

Applying Lemma 8 with =

fn (t)dt
0

r
exp

 n
!

2 log n X

rk (t) .


n 
k=1

2 log n
n ,

1
n2+

cosh

2 log n
n

!n
.

It is not obvious, but we take as given the asymptotic expansion

!n
r
8
1
(log n)3 + 18
(log n)4
1
2 log n
cosh
= n (log n)2 + 45
+ O(n3/2 ),
n
3
n
and using this,
1
n2+
Thus

r
2

cosh

2 log n
n

Z
X
n=1

!n
=

2
n1+

+O

(log n)2
n2+


=

2
+ O(n2 ).
n1+



X
2
2
fn (t)dt =
+ O(n ) < .
n1+
n=1

Because each fn is nonnegative, using this with the monotone convergence

theorem gives the claim.
Theorem 10. For almost all t [0, 1],
Pn
|
rk (t)|
2.
lim sup k=1
n log n
n
Proof. Let  > 0. By Theorem 9, for almost all t [0, 1] there is some nt such
that n nt implies that fn (t) < 1, where we are talking about the functions fn
defined in the proof of that theorem; certainly the terms of a convergent series
are eventually less than 1. That is, for almost all t [0, 1] there is some nt such
that n nt implies that (taking logarithms),


r
n

2 log n X

(2 ) log n +
rk (t) < 0,


n 
k=1

and rearranging,

Pn

rk (t)|

k=1
< 2 + = 2 + 0 .
n log n
2
For each s N, let Es be those t [0, 1] such that
Pn
| k=1 rk (t)|
1
lim sup
> 2+ .
s
n
log
n
n
|

11

For each s, taking 0 < 0 < 1s we get that almost all t [0, 1] do not belong to
Es . That is, for each s, the set Es has measure 0. Therefore

E=

Es

s=1

has measure 0. That is, for almost all t [0, 1], for all s N we have t 6 Es ,
i.e.
Pn
|
rk (t)|
1
lim sup k=1
2+ ,
s
n log n
n
and this holding for all s N yields
Pn
| k=1 rk (t)|
lim sup
2,
n log n
n
completing the proof.

Hypercubes

Let mn be Lebesgue measure on Rn , and let Qn = [0, 1]n .7

Theorem 11. If f C([0, 1]), then


 
Z
x1 + + xn
1
f
lim
dmn (x) = f
.
n Q
n
2
n
Proof. Define Xn : Qn R by
Xn =

x1 + + xn
,
n

We have

Z
Xn dmn (x) =
Qn
7 Masayoshi

1X
n

k=1

x Qn .
n

xk 1dxk =
0

1X1
1
= ,
n
2
2
k=1

Hata, Problems and Solutions in Real Analysis, p. 161, Solution 11.1.

12

and we define


Z
Vn

Xn

=
Qn

Z
=

=
=
=

n
X
x2

1
2

2
dmn (x)

1
Xn + dmn (x)
4
Qn k=1
j6=k
Z
Z 1
n Z 1
n
X
1
1 XX 1
1 1
2
x
dx
+
x
dx
xk dxk +
k
j
j
k
2
n2
n
2
4
0
0
0
j=1
1
n2

k
n2

k=1
n
X
k=1

X xj xk
n2

k6=j

1
1
+ 2
3 n

n X
X
j=1 k6=j

1 1

4 4

n1 1
1
+

3n
4n
4
n1
.
12

Suppose that cn is a sequence of positive real numbers tending to 0, and

define Jn = Jn (c) to be those x Qn such that




Xn (x) 1  cn .

2
Then
2
1
dmn (x)
2
Qn
2
Z 
1

Xn
dmn (x)
2
Jn
Z
c2n dmn (x)

Vn

Xn

Jn
2
cn mn (Jn ),

so
mn (Jn )

Vn
n1
=
.
c2n
12c2n

Take cn = n1/3 , giving

mn (Jn )

n1/3
.
12

Let  > 0. Because f is continuous, there is some > 0 such that |t 12 | <
implies that |f (t) f ( 12 )| < ; furthermore, we take such that
kf k
< .
6
13

Set N > 3 . For n N and x Qn \ Jn ,





Xn (x) 1  < cn = n1/3 N 1/3 < ,

2
and so


 


(f (Xn (x)) f 1  < .

2 

This gives us
Z
 

1 

f (Xn (x))dmn (x) f
=

2 
Qn

Z

 


1

f (Xn (x)) f
dmn (x)

2
Qn
 
Z 


f (Xn (x)) f 1  dmn (x)



2
Jn

 
Z

1 

+
f (Xn (x)) f 2  dmn (x)
Qn \Jn
Z
Z

2 kf k dmn (x) +
dmn (x)
Qn \Jn

Jn

2 kf k mn (Jn ) + 
2 kf k

n1/3
+
12

kf k
+
6
< 2,

<

which proves the claim.

14