Professional Documents
Culture Documents
Gowers 1989 PHD 16056
Gowers 1989 PHD 16056
0, let us define » subspace U = U(8) of R", as follows, Let
rrr ———
13subspace U will be that generated by uy,...,te, For technical reasons, we shall
define kegs to bon-+1. Note that for1 0 and let U = U(é) bs as defined above,
Lemma 23, Let i] be a norm which is defined on U. Then there exists «
function [ : R™ + Ry such that any I-symmetrie norm [LI with the property
that [xI|= lll for every 2 € U satisfies also
fe) < I< 0+ 8)f(2)
for every 2 ER",
Proof. Let a= Sf" ey and let a) >... > dy 20. Let us define two vectors a!
and a" in U by
and
Blew ead 30 ya Ys
Our function f will be defined by f(a) = lla". Note that a” is dominated
"H< (lal < Es'b
We shall show that af <2 (1+ 6)a" and hence (by Lemma 22) that fa'l <
pointvvise by a, which is itself dominated pointwise by a. Thus
(+e
which will complete the proof. Let us write (a})$* and (a) forthe
yiSsén
coordinates ofa! and a" respectively. Then for a
1x) min( skies — 2}
“a
aod
Yooh = Dhow, ~ ena,)in(ssb}
But for each i as remarked above, kizs 1 § (1++6)h;, 50 clearly min{s, fuga 1}
(146) minfsyki). Hence a! <2 (1 + da” as stated. a
Lemma 2-4, Let 6! > 0 and let [jl] be any norm on R™ which is (2 — 8
symmetric at every a € U(6)- Then [i is (1+ 8)(1 — 26°)-*-semmetrie on R™
Proof. Define » orm fon RM by fal = 20 {qell:(n.0) € 8). Then, i
2€U, we have Jal| < [all < (1 +6) lal by assumption. Now, given any a €
define a ns in the proof of Lemma 28, We know that [all < fal. In the other '
direction, since a is dominated pointwise by a, 2a" — a is dominated pointwise
by al, so [2a — all < fa"J. But, since a” € U, we have
> 2a — f20" — al
> 21-6')fa"l- fal
= (1 26)f0"1 > (2-260 +8)"
where the last inequality follows from the proof of Lemma 2,
Hence lis (1+ 6)(1—26")-!-equivalent to the I-symmetsie norm I, which
roves the lemma. o
Lemma 2.5. Let [| bea norm on R™, and define a I-symmeteie norm [| on R™
by [af = max{llag oll: (9,2) € ¥}- Suppose that the set of vectors {a1,...,an)
forms a 6-net in J. ofthe [unit ball of U and that [li (14+ 8)-symmetrie at
every a. Then [ll is (1+ 6)(1 ~ 66)--symmetric on R™
Proof, By hypothesis, given any 1< i < N and any (n,0) € W, we have
Wadyoll > (1+6)-Mladle Let us pick a U with fa = 1, We shall show th
Ill is (1 ~ 36)-*-symmetrie at a. So pick any (n,2) € W and pick £ such that
15Jaa] <5. Wite b = aye and bf =(ai)qe- Then we have
(bE ~ by < Pb — bi + [PBT ~ [I] + Ib’ — |
< 641-0497 46.636
But [bE = fal] = 1, so the norm |} is indeed (1 ~ 36)-!symmetric at every
vector in U, But then, by Lemma 2:4, it is (1-4 4)(1 ~ 66)—!-symmetrie on A"
as stated. a
Proof of Proposition 2:1. The dimension of U(6) is at mest log, 44m, e0, by
Lemma 1.3, there is a S:net of the [-[-unit ball of U of cardinality at most (1-+
2/6}P8+0™. When 5 < 1/11 one can easily check that this is at most m*"#4(4""
5. o
Proposition 21 now follows immediately from Lemma
Although twill not be strictly necessary for our purposes, t sof some interest
tocstimate the size ofa 6.net ofthe decreasing pat ofthe unit ball ofan arbitrary
redimensional space with « 1-symmetri basis. We have not dane this in general,
but when the space is (2, we obtain an estimate of n(*/#1°(€) for some absolute
constants ¢ and ¢!. In other words, the size of a net can be about the same asthe
sive ofthe set obtained in Proposition 2.1. It seems very likely that the same is
true for any space with & symmetric basis,
Proposition 26. Let 1< p< oo, 0.<8 <1 and let KC 6 be the set
facg
ly, <1, 04 Bs. 2 09 2 0}
‘Then K contains a S-net A of cardinality N, where
Ng n(t/ia(144/3) 1065/0)
Proof. Let 8 = 6/3 and let a= (a))! € K. Ifa!
(af)9 € 9 is any vector such
that 0; < of < (1-4 8a for all 1 4 € m then la~ al], < AL} af)" € 8. So,
16iven a let us define a! to be the vector with
df= min{(1+6)-9":5 31,0467 9 > a}.
max{ol,(1-+8)~*), where k = 2 logy a(n).
Let a € 67 be defined by af
Note that & > logy 4(8-n!/*), so (1-40)-* < @n¥, Te follows easily that
Ja” —a'], <8, and therefore that lla" — ally < 20. If, given aay vector a, we can
find a vector b in such that [b— a] <6, then lb— al] <30 = 6, so then A.
will bee net, In other word, ti enough to approximate towithin 6 vectors of
the form a = Fait, where uw; = xu, for some sequence of possibly empty sete
Uy... Ue satisfying US;
Consider two vectors a= E¥aiu, and a! = 5 acu, where (wf and (vt
oI
Thaleresesdlve-vp)
inl and &y < hy whenever i 3/2 our construction fils heeause the basi it gives hes a large block
basis that is close to the unit vector basis of an approprinte (4,
‘Theorem 5:1. There exists an absolute constant C such that for any 1 [all
60(ii) if > Cn¥l-Y(logn)*/* then no block basis wy.....uy of the standard
) log:
Dasis is 2aymmetrie
Let ws begin by setting m = (C/2)n*logn)*, h = (2/Cn*(logn)-*!,
(Sn2V"*, We shall pick 27 fonetionals fis... fi
N = 6008/" Jog» and 2
independently at random from the set of all functionals on R* with 1-coordinates,
where this cet is endowed with the uniform probability measure. Given a vector
2 ER", we define «random norm by setting
lal
vax(llaly fila) s---sLfv(@)D)
Our aim isto show that there exists a choice of fi,..., 1 such that no block
basis of the standard basis of R* is 2-symmetric under the corresponding norm.
In fact, it is ensy to deduce from our proof tht, with high probity, random
choice will do. The other two conditions of Theorem 5.1 are trivially satiated by
ny notm of this kind,
As in the previous two chapters, we shall make one or two definitions and
"ns state (wo lemmas which will be enough to prove our main theorem.
Suppose we have a block basis uj,...,thy of the standard basis, and suppose
‘hat we can find some 1 nf. Then there i
block basis of (21)f with blocks of #1 coefficients and equal length, which is
(1+ e-symmetsic and has cardinality
m2 r(o(log nnd
where 7(e) = (€/3,000,000)(04(33/6))
‘The dependence on 1 in this theorem improves on the previously known de-
pendence of n@P"/9#" obtained by Amir and Milman in [4 Section 5.2). They
have also shown that, if X has ptype constant C, one can obtain a bound of
a{e.p,C)n'?-7)290490 (3, Theorem 2.4}. Our exponent of n of 2/p = 1 therefore
improves both these resus, with or without an assumption about type-conditions
We shall say a litle about the effect of type conditions at the end of this chapter
and more at the end of the next one
We sll begin with some notation.
Let W be the group {-1,1)" x Sn with multiplication given by ((n)f%.2) ©
eve R™, and
(nb .e!) = (nen). 202"), acting on R™ as follows. If a=
(12) € %9= (9, then
(yaa Bgl) =
Yonsei
Let 9 be the group (—1,1)" x Sy acting on X as follows. If b € X,
Tf bie, and (¢,x) € , then
(62) b a n(b) & Bow = Do birt
36‘We shall sometimes relabel the indices of (2))f. Let ry = euaaty (i
Auscosh), where im =n, and similarly let 6,
yney and
iy & a((i—h +3) for (6,7) €9.
We shall regard a block basis of (2, )P as a random embedding of R™ into X.
Let 6:R™ +X be the embedding defined by
Senn EE
am, Then let bie = Wen $5.
and write us = Shy ny for é
When 1 € i You might also like