Section IX.3

Rademacher Functions. The Khinchin Inequality

[31/90%]
Show
LaTeX
Problem IX.3.1
?
(a)

Prove that for any n∈Nn \in \mathbb {N} and any collection {εk}1⩽k⩽n\left\{ \varepsilon_{k}\right\}_{1 \leqslant k \leqslant n}, where the numbers εk\varepsilon_{k} take the values +1 or -1, there exists a point x0∈(0,1)x_{0} \in (0,1) such that rk(x0)=εkr_{k}\left(x_{0}\right)=\varepsilon_{k} for k=1,…,nk=1, \ldots , n. What is the measure of the set of such points?

(b)

Prove that r1(x+1)=r1(x),rn(x)=r1(2n−1x)r_{1}(x+1)=r_{1}(x), r_{n}(x)=r_{1}\left(2^{n-1} x\right).

(c)

Let an(x)a_{n}(x) be the nn-th digit of the binary expansion of the number x∈(0,1)x \in (0,1). Prove that an(x)=12(1−rn(x))a_{n}(x)=\frac{1}{2}\left(1-r_{n}(x)\right) almost everywhere on (0,1).

(d)

Find the sum of the series ∑2−nrn(x)(x∈(0,1))\sum 2^{-n} r_{n}(x)(x \in (0,1)).

Problem IX.3.2

Prove that

?
(a)

∫01φ(r1(x),…,rn(x))dx=2−n∑ε1,…,εn∈{−1,1}φ(ε1,…,εn)\int_{0}^{1} \varphi \left(r_{1}(x), \ldots , r_{n}(x)\right) d x=2^{-n} \sum_{\varepsilon_{1}, \ldots , \varepsilon_{n} \in \left\{ -1,1\right\} } \varphi \left(\varepsilon_{1}, \ldots , \varepsilon_{n}\right);

(b)

∫01(∏1⩽k⩽nφk(rk(x)))dx=∏1⩽k⩽n(∫01φk(r1(x))dx)\int_{0}^{1}\left(\prod_{1 \leqslant k \leqslant n} \varphi_{k}\left(r_{k}(x)\right)\right) d x=\prod_{1 \leqslant k \leqslant n}\left(\int_{0}^{1} \varphi_{k}\left(r_{1}(x)\right) d x\right).

Problem IX.3.3

Let m1,…,mnm_{1}, \ldots , m_{n} be pairwise distinct natural numbers. Prove that

?
(a)

∫01φ(rm1(x),rm2(x),…,rmn(x))dx=\int_{0}^{1} \varphi \left(r_{m_{1}}(x), r_{m_{2}}(x), \ldots , r_{m_{n}}(x)\right) d x=

=∫01φ(r1(x),r2(x),…,rn(x))dx =\int _{0}^{1} \varphi \left(r_{1}(x), r_{2}(x), \ldots , r_{n}(x)\right) d x
(b)

∫01rm1(x)rm2(x)…rmn(x)dx=0\int_{0}^{1} r_{m_{1}}(x) r_{m_{2}}(x) \ldots r_{m_{n}}(x) d x=0.

In particular, {rn}\left\{ r_{n}\right\} is an orthonormal system in L2(0,1)\mathscr {L}^{2}(0,1).

Problem IX.3.4

Evaluate the integrals

?
(a)

∫01eitΣ2−krk(x)dx\int_{0}^{1} e^{i t \Sigma 2^{-k} r_{k}(x)} d x;

(b)

∫01eitΣ3−krk(x)dx\int_{0}^{1} e^{i t \Sigma 3^{-k} r_{k}(x)} d x;

(c)

∫01ez(c1r1(x)+…+cnrn(x))dx(z,ck∈C)\int_{0}^{1} e^{z\left(c_{1} r_{1}(x)+\ldots +c_{n} r_{n}(x)\right)} d x\left(z, c_{k} \in \mathbb {C}\right).

(d)

Prove that if t,an∈R,an→0,∑an2=∞t, a_{n} \in \mathbb {R}, a_{n} \rightarrow 0, \sum a_{n}^{2}=\infty, then

∫01eit(a1r1(x)+…+anrn(x))dx⟶n→∞0 for t≠0 \int _{0}^{1} e^{i t\left(a_{1} r_{1}(x)+\ldots +a_{n} r_{n}(x)\right)} d x \underset {n \rightarrow \infty }{\longrightarrow } 0 \text{ for } t \neq 0
Problem IX.3.5

Prove that for any complex numbers c1,…,cnc_{1}, \ldots , c_{n} the following inequalities hold:

?
(a)

∫01∣∑1⩽k⩽nckrk(x)∣4dx⩽A(∑1⩽k⩽n∣ck∣2)2\int_{0}^{1}\left|\sum_{1 \leqslant k \leqslant n} c_{k} r_{k}(x)\right|^{4} d x \leqslant A\left(\sum_{1 \leqslant k \leqslant n}\left|c_{k}\right|^{2}\right)^{2};

(b)

(∑1⩽k⩽n∣ck∣2)1/2⩽B∫01∣∑1⩽k⩽nckrk(x)∣dx\left(\sum_{1 \leqslant k \leqslant n}\left|c_{k}\right|^{2}\right)^{1 / 2} \leqslant B \int_{0}^{1}\left|\sum_{1 \leqslant k \leqslant n} c_{k} r_{k}(x)\right| d x,

where A,BA, B are absolute constants.

Problem IX.3.6

Prove that for any complex numbers ckj(k,j∈N)c_{k j}(k, j \in \mathbb {N}) the following inequalities hold:

?
(a)

∫01∫01∣∑1⩽k,j⩽nckjrk(x)rj(y)∣4dxdy⩽A(∑1⩽k,j⩽n∣ckj∣2)2\int_{0}^{1} \int_{0}^{1}\left|\sum_{1 \leqslant k, j \leqslant n} c_{k j} r_{k}(x) r_{j}(y)\right|^{4} d x d y \leqslant A\left(\sum_{1 \leqslant k, j \leqslant n}\left|c_{k j}\right|^{2}\right)^{2};

(b)

(∑1⩽k,j⩽n∣ckj∣2)12⩽B∫01∫01∣∑1⩽k,j⩽nckjrk(x)rj(y)∣dxdy\left(\sum_{1 \leqslant k, j \leqslant n}\left|c_{k j}\right|^{2}\right)^{\frac{1}{2}} \leqslant B \int_{0}^{1} \int_{0}^{1}\left|\sum_{1 \leqslant k, j \leqslant n} c_{k j} r_{k}(x) r_{j}(y)\right| d x d y, where A,BA, B are absolute constants.

Problem IX.3.7

Prove that

?
(a)

{rjrk}j<k\left\{ r_{j} r_{k}\right\}_{j<k} is an orthonormal system in L2(0,1)\mathscr {L}^{2}(0,1);

(b)

for any complex numbers cjkc_{j k} the following inequalities hold:

∫01∣∑1⩽j<k⩽ncjkrj(x)rk(x)∣4dx⩽A(∑1⩽j<k⩽n∣cjk∣2)2(∑1⩽j<k⩽n∣cjk∣2)1/2⩽B∫01∣∑1⩽j<k⩽ncjkrj(x)rk(x)∣dx \begin{aligned} & \int _{0}^{1}\left|\sum _{1 \leqslant j<k \leqslant n} c_{j k} r_{j}(x) r_{k}(x)\right|^{4} d x \leqslant A\left(\sum _{1 \leqslant j<k \leqslant n}\left|c_{j k}\right|^{2}\right)^{2} \\ & \left(\sum _{1 \leqslant j<k \leqslant n}\left|c_{j k}\right|^{2}\right)^{1 / 2} \leqslant B \int _{0}^{1}\left|\sum _{1 \leqslant j<k \leqslant n} c_{j k} r_{j}(x) r_{k}(x)\right| d x \end{aligned}

where A,BA, B are absolute constants.

Problem IX.3.8

Prove that for any complex numbers c1,…,cnc_{1}, \ldots , c_{n} the following inequalities hold:

?
(a)

∫02π∣∑1⩽k⩽ncksin⁡2kx∣4dx⩽A(∑1⩽k⩽n∣ck∣2)2\int_{0}^{2 \pi }\left|\sum_{1 \leqslant k \leqslant n} c_{k} \sin 2^{k} x\right|^{4} d x \leqslant A\left(\sum_{1 \leqslant k \leqslant n}\left|c_{k}\right|^{2}\right)^{2};

(b)

(∑1⩽k⩽n∣ck∣2)1/2⩽B∫02π∣∑1⩽k⩽ncksin⁡2kx∣dx\left(\sum_{1 \leqslant k \leqslant n}\left|c_{k}\right|^{2}\right)^{1 / 2} \leqslant B \int_{0}^{2 \pi }\left|\sum_{1 \leqslant k \leqslant n} c_{k} \sin 2^{k} x\right| d x,

where A,BA, B are absolute constants.

Problem IX.3.9

Relying on the results of problems 3.5−3.83.5-3.8, prove that for any t>0t>0 and any complex numbers ck,cjkc_{k}, c_{j k} the following inequalities hold:

?
(a)

λ{x∈(0,1)∣∣∑k=1nckrk(x)∣>t}⩽A1t4(∑k=1n∣ck∣2)2\lambda \left\{ x \in (0,1) \mid \left|\sum_{k=1}^{n} c_{k} r_{k}(x)\right|>t\right\} \leqslant \frac{A_{1}}{t^{4}}\left(\sum_{k=1}^{n}\left|c_{k}\right|^{2}\right)^{2};

(b)

λ{x∈(0,1)∣∣∑1⩽j<k⩽ncjkrj(x)rk(x)∣>t}⩽\lambda \left\{ x \in (0,1) \mid \left|\sum_{1 \leqslant j<k \leqslant n} c_{j k} r_{j}(x) r_{k}(x)\right|>t\right\} \leqslant

⩽A2t4(∑1⩽j<k⩽n∣cjk∣2)2; \leqslant \frac{A_{2}}{t^{4}}\left(\sum _{1 \leqslant j<k \leqslant n}\left|c_{j k}\right|^{2}\right)^{2} ;
(c)

λ{x∈(0,2π)∣∣∑k=1ncksin⁡2kx∣>t}⩽A3t4(∑k=1n∣ck∣2)2\lambda \left\{ x \in (0,2 \pi ) \mid \left|\sum_{k=1}^{n} c_{k} \sin 2^{k} x\right|>t\right\} \leqslant \frac{A_{3}}{t^{4}}\left(\sum_{k=1}^{n}\left|c_{k}\right|^{2}\right)^{2}, where A1,A2,A3A_{1}, A_{2}, A_{3} are absolute constants.

Problem IX.3.10

Let σn(x)=1n∑1⩽k⩽nrk(x)(x∈(0,1))\sigma_{n}(x)=\frac{1}{n} \sum_{1 \leqslant k \leqslant n} r_{k}(x)(x \in (0,1)). Prove that

?
(a)

∑λ{x∈(0,1)∣σn(x)⩾n−1/5}<∞\sum \lambda \left\{ x \in (0,1) \mid \sigma_{n}(x) \geqslant n^{-1 / 5}\right\} <\infty;

(b)

∑∣σn(x)∣4<∞\sum \left|\sigma_{n}(x)\right|^{4}<\infty almost everywhere on (0,1)(0,1).

Each of the statements a), b) implies the equality lim⁡n→∞σn(x)=0\lim_{n \rightarrow \infty } \sigma_{n}(x)=0 almost everywhere on (0,1), which is known as the strong law of large numbers. Using the connection between Rademacher functions and the digits of the binary expansion (see problem 3.1 c)), deduce from this that the binary expansion of almost every number contains "equally many" zeros and ones (see problem 2.14).

Problem IX.3.11

Let CC be an arbitrary square n×nn \times n matrix, AA the set of all vectors in Rn\mathbb {R}^{n} whose coordinates equal ±1\pm 1.

Prove that

?
(a)

∑x,y∈A(Cx,y)4⩽K1⋅2−2n(∑x,y∈A(Cx,y)2)2\sum_{x, y \in A}(C x, y)^{4} \leqslant K_{1} \cdot 2^{-2 n}\left(\sum_{x, y \in A}(C x, y)^{2}\right)^{2};

(b)

(∑x,y∈A(Cx,y)2)1/2⩽K2⋅2−n∑x,y∈A∣(Cx,y)∣\left(\sum_{x, y \in A}(C x, y)^{2}\right)^{1 / 2} \leqslant K_{2} \cdot 2^{-n} \sum_{x, y \in A}\left|(C x, y)\right|,

where K1,K2K_{1}, K_{2} are absolute constants.

Problem IX.3.12

Using the uniqueness theorem for the Fourier transform of finite Borel measures (see § 4), find the increasing distribution functions of the following functions:

?
(a)

f(x)=12+∑3−krk(x)f(x)=\frac{1}{2}+\sum 3^{-k} r_{k}(x);

(b)

f(x)=∑2−krmk(x)f(x)=\sum 2^{-k} r_{m_{k}}(x),

where x∈(0,1)x \in (0,1), and {mk}\left\{ m_{k}\right\} is an arbitrary (not necessarily increasing) sequence of pairwise distinct natural numbers.

Problem IX.3.13

Let Rn=a1r1+…+anrn(a1,…,an∈R),ΔR_{n}=a_{1} r_{1}+\ldots +a_{n} r_{n}\left(a_{1}, \ldots , a_{n} \in \mathbb {R}\right), \Delta be one of the intervals Δm,j\Delta_{m, j} (see the definition of the Rademacher functions). Prove that for m⩽nm \leqslant n the following inequalities hold:

?
(a)

∫Δ∣Rm(x)∣dx⩽∫Δ∣Rn(x)∣dx;\int_{\Delta }\left|R_{m}(x)\right| d x \leqslant \int_{\Delta }\left|R_{n}(x)\right| d x ;

(b)

∫ΔeRm(x)dx⩽∫ΔeRn(x)dx\int_{\Delta } e^{R_{m}(x)} d x \leqslant \int_{\Delta } e^{R_{n}(x)} d x.

(c)

Prove that inequalities a), b) remain valid if Δ\Delta is replaced by the set E={x∈(0,1)∣∣∑1⩽k⩽makrk(x)∣>t}E=\left\{ x \in (0,1) \mid \left|\sum_{1 \leqslant k \leqslant m} a_{k} r_{k}(x)\right|>t\right\}.

Problem IX.3.14

Let Rn=a1r1+…+anrn,A2=∣a1∣2+…+∣an∣2(a1,a2,…,an∈R)R_{n}=a_{1} r_{1}+\ldots +a_{n} r_{n}, A^{2}=\left|a_{1}\right|^{2}+\ldots +\left|a_{n}\right|^{2}\left(a_{1}, a_{2}, \ldots , a_{n} \in \mathbb {R}\right). Using the inequality ch⁡u⩽eu2/2(u∈R)\operatorname {ch} u \leqslant e^{u^{2} / 2}(u \in \mathbb {R}), prove that for t>0t>0

?
(a)

λ{x∈(0,1)∣∣Rn(x)∣>t}⩽2e−t2/(2A2)\lambda \left\{ x \in (0,1) \mid \left|R_{n}(x)\right|>t\right\} \leqslant 2 e^{-t^{2} /\left(2 A^{2}\right)};

(b)

if E={x∈(0,1)∣max⁡{∣R1(x)∣,…,∣Rn(x)∣}>t}E=\left\{ x \in (0,1) \mid \max \left\{ \left|R_{1}(x)\right|, \ldots ,\left|R_{n}(x)\right|\right\} >t\right\}, then λ(E)⩽At\lambda (E) \leqslant \frac{A}{t}, λ(E)⩽2e−t2/(2A2)\lambda (E) \leqslant 2 e^{-t^{2} /\left(2 A^{2}\right)}.

Verify that for complex aka_{k} inequality a) is replaced by λ{…}⩽4e−t2/(4A2)\lambda \left\{ \ldots \right\} \leqslant 4 e^{-t^{2} /\left(4 A^{2}\right)}.

Problem IX.3.15

Let ∑∣ak∣2<∞\sum \left|a_{k}\right|^{2}<\infty. Prove that

?
(a)

the series ∑akrk(x)\sum a_{k} r_{k}(x) converges almost everywhere;

(b)

if M(x)=sup⁡n⩾1∣a1r1(x)+…+anrn(x)∣M(x)=\sup_{n \geqslant 1}\left|a_{1} r_{1}(x)+\ldots +a_{n} r_{n}(x)\right|, then ∫01esM2(x)dx<∞\int_{0}^{1} e^{s M^{2}(x)} d x<\infty for any number s>0s>0.

Problem IX.3.16

Prove that if the series ∑akrk\sum a_{k} r_{k} converges in measure on some set EE of positive measure, then ∑∣ak∣2<∞\sum \left|a_{k}\right|^{2}<\infty.

?
Problem IX.3.17

Prove that the following statements are equivalent:

  1. ∑∣ak∣2<∞\sum \left|a_{k}\right|^{2}<\infty;

  2. the series ∑akrk\sum a_{k} r_{k} converges in Lp(0,1)\mathscr {L}^{p}(0,1) for some p∈(0,2)p \in (0,2);

  3. the series ∑akrk\sum a_{k} r_{k} converges in measure on the interval (0,1)(0,1);

  4. the series ∑akrk\sum a_{k} r_{k} converges almost everywhere on (0,1)(0,1).

?
Problem IX.3.18

Prove that if the series ∑aksin⁡2kx\sum a_{k} \sin 2^{k} x converges in measure on some set of positive measure, then ∑∣ak∣2<∞\sum \left|a_{k}\right|^{2}<\infty.

?
Problem IX.3.19

Let ∑∣ak∣2<∞,Δ\sum \left|a_{k}\right|^{2}<\infty , \Delta be a nonempty interval. Prove that if vrai sup ∣∑akrk(x)∣<∞\left|\sum a_{k} r_{k}(x)\right|<\infty, then ∑∣ak∣<∞\sum \left|a_{k}\right|<\infty (cf. problem 4.15). x∈Δx \in \Delta

?
Problem IX.3.20
?
(a)

Let {φn}n⩾1\left\{ \varphi_{n}\right\}_{n \geqslant 1} be an orthonormal system in the space L2(0,1)\mathscr {L}^{2}(0,1). Prove that

∑n⩾1∣∫eφn(t)dt∣2⩽λ(e) for e⊂(0,1). \sum _{n \geqslant 1}\left|\int _{e} \varphi _{n}(t) d t\right|^{2} \leqslant \lambda (e) \quad \text{ for } \quad e \subset (0,1) .
(b)

Sharpen this inequality for the Rademacher system.

Problem IX.3.21

The Haar system of functions {hm}m⩾0\left\{ h_{m}\right\}_{m \geqslant 0} is defined as follows. Every natural number mm can be uniquely represented in the form m=2n+km=2^{n}+k, where n,kn, k are nonnegative integers, 0⩽k<2n0 \leqslant k<2^{n}. By definition we set h0(x)=1h_{0}(x)=1 for any x∈(0,1)x \in (0,1), and for m∈Nm \in \mathbb {N}

hm(x)={2n/2, for 2−nk<x<2−n(k+12),−2n/2, for 2−n(k+12)<x<2−n(k+1),0, for all other values of x∈(0,1). h_{m}(x)= \begin{cases} 2^{n / 2}, & \text{ for } 2^{-n} k<x<2^{-n}\left(k+\frac{1}{2}\right), \\ -2^{n / 2}, & \text{ for } 2^{-n}\left(k+\frac{1}{2}\right)<x<2^{-n}(k+1), \\ 0, & \text{ for all other values of } x \in (0,1) . \end{cases}

Prove that

?
(a)

{hm}m⩾0\left\{ h_{m}\right\}_{m \geqslant 0} is an orthonormal system in L2(0,1)\mathscr {L}^{2}(0,1);

(b)

if E={x∈(0,1)∣∣∑0⩽k⩽makhk(x)∣>t}E=\left\{ x \in (0,1) \mid \left|\sum_{0 \leqslant k \leqslant m} a_{k} h_{k}(x)\right|>t\right\}, then for n⩾mn \geqslant m

∫E∣∑0⩽k⩽makhk(x)∣dx⩽∫E∣∑0⩽k⩽nakhk(x)∣dx; \int _{E}\left|\sum _{0 \leqslant k \leqslant m} a_{k} h_{k}(x)\right| d x \leqslant \int _{E}\left|\sum _{0 \leqslant k \leqslant n} a_{k} h_{k}(x)\right| d x ;
(c)

if E={x∈(0,1)∣max⁡0⩽m⩽n∣∑0⩽k⩽makhk(x)∣>t}E=\left\{ x \in (0,1) \mid \max_{0 \leqslant m \leqslant n}\left|\sum_{0 \leqslant k \leqslant m} a_{k} h_{k}(x)\right|>t\right\}, then

λ(E)⩽At−1, where A2=∣a1∣2+…+∣an∣2; \lambda (E) \leqslant A t^{-1}, \quad \text{ where } \quad A^{2}=\left|a_{1}\right|^{2}+\ldots +\left|a_{n}\right|^{2} ;
(d)

if ∑∣ak∣2<∞\sum \left|a_{k}\right|^{2}<\infty, then the series ∑akhk(x)\sum a_{k} h_{k}(x) converges almost everywhere on (0,1);

(e)

the partial sum S2nS_{2^{n}} of the Fourier series of a function ff from L1(0,1)\mathscr {L}^{1}(0,1) with respect to the Haar system is the result of averaging ff over the intervals Δn,k=(k2n,k+12n),k=0,1,…,2n−1\Delta_{n, k}=\left(\frac{k}{2^{n}}, \frac{k+1}{2^{n}}\right), k=0,1, \ldots , 2^{n}-1;

(f)

deduce from e) that ∥S2n∥1⩽∥f∥1\left\| S_{2^{n}}\right\|_{1} \leqslant \left\| f\right\|_{1}.

Problem IX.3.22

The Haar system of functions {hm}m⩾0\left\{ h_{m}\right\}_{m \geqslant 0} is defined as follows. Every natural number mm can be uniquely represented in the form m=2n+km=2^{n}+k, where n,kn, k are nonnegative integers, 0⩽k<2n0 \leqslant k<2^{n}. By definition we set h0(x)=1h_{0}(x)=1 for any x∈(0,1)x \in (0,1), and for m∈Nm \in \mathbb {N}

hm(x)={2n/2, for 2−nk<x<2−n(k+12),−2n/2, for 2−n(k+12)<x<2−n(k+1),0, for all other values of x∈(0,1). h_{m}(x)= \begin{cases} 2^{n / 2}, & \text{ for } 2^{-n} k<x<2^{-n}\left(k+\frac{1}{2}\right), \\ -2^{n / 2}, & \text{ for } 2^{-n}\left(k+\frac{1}{2}\right)<x<2^{-n}(k+1), \\ 0, & \text{ for all other values of } x \in (0,1) . \end{cases}

It is known that the Fourier series with respect to the Haar system of any summable function converges to it ∗){ }^{*)} in L1(0,1)\mathscr {L}^{1}(0,1). Relying on this, prove that

?
(a)

this series converges almost everywhere on (0,1)(0,1);

(b)

if E={x∈(0,1)∣sup⁡n∣Sn(x)∣>t}E=\left\{ x \in (0,1) \mid \sup_{n}\left|S_{n}(x)\right|>t\right\}, where SnS_{n} is the nn-th partial sum of the Fourier series of the function ff with respect to the Haar system, then λ(E)⩽∥f∥1t\lambda (E) \leqslant \frac{\left\| f\right\|_{1}}{t}.

Remark.
?

∗){ }^{*)} A reader familiar with the elements of functional analysis will notice that this is easily derived from the Banach-Steinhaus theorem.

Problem IX.3.23
?
(a)

Using the result of Problem 3.14, prove that for any p>0p>0 and any a1,…,an∈Ra_{1}, \ldots , a_{n} \in \mathbb {R} the Khintchine inequality holds:

Ap(∑1⩽k⩽nak2)1/2⩽∥∑1⩽k⩽nakrk∥p⩽Bp(∑1⩽k⩽nak2)1/2, A_{p}\left(\sum _{1 \leqslant k \leqslant n} a_{k}^{2}\right)^{1 / 2} \leqslant \left\| \sum _{1 \leqslant k \leqslant n} a_{k} r_{k}\right\| _{p} \leqslant B_{p}\left(\sum _{1 \leqslant k \leqslant n} a_{k}^{2}\right)^{1 / 2},

where Ap,BpA_{p}, B_{p} are positive constants depending only on pp, and lim‾⁡p→+∞Bp/p⩽1/e\varlimsup_{p \rightarrow +\infty } B_{p} / \sqrt{p} \leqslant 1 / \sqrt{e}.

Note that the lower bound for p⩾2p \geqslant 2 is obvious (with coefficient Ap=1A_{p}=1, which cannot be increased), while the upper bound is obvious for p⩽2p \leqslant 2 (with coefficient Bp=1B_{p}=1, which cannot be decreased).

(b)

Prove that Ap⩽αp=212−1pmin⁡{1,γp}\quad A_{p} \leqslant \alpha_{p}=2^{\frac{1}{2}-\frac{1}{p}} \min \left\{ 1, \gamma_{p}\right\} \quad for p∈(0,2]\quad p \in (0,2], Bp⩾βp=212−1pγpB_{p} \geqslant \beta_{p}=2^{\frac{1}{2}-\frac{1}{p}} \gamma_{p} for p⩾2p \geqslant 2, where γp=(Γ(1+p2)Γ(32))1p(p>0)\gamma_{p}=\left(\frac{\Gamma \left(\frac{1+p}{2}\right)}{\Gamma \left(\frac{3}{2}\right)}\right)^{\frac{1}{p}}(p>0) (it is proved in [Haag] that αp,βp\alpha_{p}, \beta_{p} are the best values of the coefficients ApA_{p} and BpB_{p}). In particular, B2m⩾((2m−1)!!)1/(2m)B_{2 m} \geqslant ((2 m-1)!!)^{1 /(2 m)} for all m∈Nm \in \mathbb {N}.

Problem IX.3.24

Prove that if for p>2p>2 and some BpB_{p} the inequality

∥∑1⩽k⩽nakrk∥p⩽Bp(∑1⩽k⩽n∣ak∣2)1/2 \left\| \sum _{1 \leqslant k \leqslant n} a_{k} r_{k}\right\| _{p} \leqslant B_{p}\left(\sum _{1 \leqslant k \leqslant n}\left|a_{k}\right|^{2}\right)^{1 / 2}

holds for any real numbers a1,…,ana_{1}, \ldots , a_{n}, then it also holds for any complex a1,a2,…,ana_{1}, a_{2}, \ldots , a_{n}.

For 0<p<20<p<2 an analogous statement holds for the inequality

∥∑1⩽k⩽nakrk∥p⩾Ap(∑1⩽k⩽n∣ak∣2)1/2. \left\| \sum _{1 \leqslant k \leqslant n} a_{k} r_{k}\right\| _{p} \geqslant A_{p}\left(\sum _{1 \leqslant k \leqslant n}\left|a_{k}\right|^{2}\right)^{1 / 2} .
?
Problem IX.3.25

Let p>0,n∈Np>0, n \in \mathbb {N} and a1,…,an∈Ra_{1}, \ldots , a_{n} \in \mathbb {R}. Prove that

?
(a)

∥∑k=1nakrk∥p+2p+2=(p+1)∑j=1naj2∫01∥tajrj+∑k=1k≠jnakrk∥ppdt\left\| \sum_{k=1}^{n} a_{k} r_{k}\right\|_{p+2}^{p+2}=(p+1) \sum_{j=1}^{n} a_{j}^{2} \int_{0}^{1}\left\| t a_{j} r_{j}+\sum_{\substack {k=1 \\ k \neq j}}^{n} a_{k} r_{k}\right\|_{p}^{p} d t;

(b)

the best constants BpB_{p} in the Khintchine inequality satisfy the inequality Bp+2p+2⩽(p+1)BppB_{p+2}^{p+2} \leqslant (p+1) B_{p}^{p}; moreover, B2m2m=(2m−1)!!B_{2 m}^{2 m}=(2 m-1)!!.

Problem IX.3.26

Let n∈Nn \in \mathbb {N} and a1,…,an∈Ra_{1}, \ldots , a_{n} \in \mathbb {R}. Prove that

?
(a)

∫01∣a1r1(x)+…+anrn(x)∣dx=2π∫0∞(1−∏k=1ncos⁡akt)dtt2\int_{0}^{1}\left|a_{1} r_{1}(x)+\ldots +a_{n} r_{n}(x)\right| d x=\frac{2}{\pi } \int_{0}^{\infty }\left(1-\prod_{k=1}^{n} \cos a_{k} t\right) \frac{d t}{t^{2}};

(b)

∫01∣a1r1(x)+…+anrn(x)∣dx⩾12(a12+…+an2)1/2\int_{0}^{1}\left|a_{1} r_{1}(x)+\ldots +a_{n} r_{n}(x)\right| d x \geqslant \frac{1}{\sqrt{2}}\left(a_{1}^{2}+\ldots +a_{n}^{2}\right)^{1 / 2};

(c)

for p∈(0,1]p \in (0,1] the best constant ApA_{p} in the Khintchine inequality equals 212−1p2^{\frac{1}{2}-\frac{1}{p}}.

Problem IX.3.27

Prove that statements a)--d) of Problem 3.17 are equivalent to the statement: e) the series ∑akrk\sum a_{k} r_{k} converges in Lp(0,1)\mathscr {L}^{p}(0,1) for every pp in (0,∞)(0, \infty ).

?
Problem IX.3.28
?
(a)

Prove that for any trigonometric polynomial ∑∣k∣⩽nckeikx\sum_{\left|k\right| \leqslant n} c_{k} e^{i k x} one can choose numbers εk=±1(∣k∣⩽n)\varepsilon_{k}= \pm 1 \quad (\left|k\right| \leqslant n) such that

∫−ππ∣∑∣k∣⩽nεkckeikx∣dx⩾c(∑∣k∣⩽n∣ck∣2)1/2, \int _{-\pi }^{\pi }\left|\sum _{\left|k\right| \leqslant n} \varepsilon _{k} c_{k} e^{i k x}\right| d x \geqslant c\left(\sum _{\left|k\right| \leqslant n}\left|c_{k}\right|^{2}\right)^{1 / 2},

where c>0c>0 is an absolute constant.

(b)

Consider the trigonometric polynomial Tn(φ)=∑k=1nakeikφT_{n}(\varphi )=\sum_{k=1}^{n} a_{k} e^{i k \varphi }, whose coefficients a1,…,ana_{1}, \ldots , a_{n} form a permutation of the numbers 1,2,…,n1,2, \ldots , n. Prove that the trivial estimate

∥Tn∥1=∫−ππ∣Tn(φ)∣dφ⩽2π∥Tn∥2=2π12+…+n2⩽const⁡n3/2, \left\| T_{n}\right\| _{1}=\int _{-\pi }^{\pi }\left|T_{n}(\varphi )\right| d \varphi \leqslant \sqrt{2 \pi }\left\| T_{n}\right\| _{2}=2 \pi \sqrt{1^{2}+\ldots +n^{2}} \leqslant \operatorname {const} n^{3 / 2},

cannot, in general, be improved in order: for every n∈Nn \in \mathbb {N} there exists a permutation a1,…,ana_{1}, \ldots , a_{n} of the numbers 1,…,n1, \ldots , n such that ∥Tn∥1⩾cn3/2(c>0\left\| T_{n}\right\|_{1} \geqslant c n^{3 / 2}(c>0 and independent of n)n).

Problem IX.3.29

Prove that for any trigonometric polynomial ∑∣k∣⩽nckeikx(n⩾2)\sum_{\left|k\right| \leqslant n} c_{k} e^{i k x}(n \geqslant 2) one can choose numbers εk=±1(∣k∣⩽n)\varepsilon_{k}= \pm 1(\left|k\right| \leqslant n) such that

sup⁡x∈R∣∑∣k∣⩽nεkckeikx∣⩽10ln⁡n(∑∣k∣⩽n∣ck∣2)1/2. \sup _{x \in \mathbb {R}}\left|\sum _{\left|k\right| \leqslant n} \varepsilon _{k} c_{k} e^{i k x}\right| \leqslant 10 \sqrt{\ln n}\left(\sum _{\left|k\right| \leqslant n}\left|c_{k}\right|^{2}\right)^{1 / 2} .
?
Problem IX.3.30

Prove that for any a>2a>2 there exist numbers qn>0q_{n}>0 such that the series

∑n⩾2(∣r1(x)+…+rn(x)∣anln⁡n)qn \sum _{n \geqslant 2}\left(\frac{\left|r_{1}(x)+\ldots +r_{n}(x)\right|}{\sqrt{a n \ln n}}\right)^{q_{n}}

converges almost everywhere. Deduce from this that almost everywhere

lim‾⁡n→∞∣r1(x)+…+rn(x)∣2nln⁡n⩽1. \varlimsup _{n \rightarrow \infty } \frac{\left|r_{1}(x)+\ldots +r_{n}(x)\right|}{\sqrt{2 n \ln n}} \leqslant 1 .
?
Problem IX.3.31

Let {nj}\left\{ n_{j}\right\} be an arbitrary increasing sequence of natural numbers. Prove that

lim‾⁡j→+∞∣r1(x)+r2(x)+…+rnj(x)∣2njln⁡j⩽1 \varlimsup _{j \rightarrow +\infty } \frac{\left|r_{1}(x)+r_{2}(x)+\ldots +r_{n_{j}}(x)\right|}{\sqrt{2 n_{j} \ln j}} \leqslant 1

almost everywhere on (0,1)(0,1). If j=O(ln⁡nj)j=O\left(\ln n_{j}\right), then we obtain from this a rather weak version of the law of the iterated logarithm:

lim‾⁡j→+∞∣r1(x)+r2(x)+…+rnj(x)∣2njln⁡ln⁡nj⩽1 almost everywhere on (0,1).  \varlimsup _{j \rightarrow +\infty } \frac{\left|r_{1}(x)+r_{2}(x)+\ldots +r_{n_{j}}(x)\right|}{\sqrt{2 n_{j} \ln \ln n_{j}}} \leqslant 1 \text{ almost everywhere on }(0,1) \text{. }
?