Rate of Decay of the Bernstein Numbers

We show that if a Banach space X contains uniformly complemented l₂ⁿ 's then there exists a universal constant b = b(X) > 0 such that for each Banach space Y, and any sequence dn ↓ 0 there is a bounded linear operator T : X → Y with the Bernstein numbers bn(T) of T satisfying b⁻¹dn ≤ bn(T) ≤...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Datum:2013
1. Verfasser: Plichko, A.
Format: Artikel
Sprache:English
Veröffentlicht: Фізико-технічний інститут низьких температур ім. Б.І. Вєркіна НАН України 2013
Schriftenreihe:Журнал математической физики, анализа, геометрии
Online Zugang:https://nasplib.isofts.kiev.ua/handle/123456789/106737
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Назва журналу:Digital Library of Periodicals of National Academy of Sciences of Ukraine
Zitieren:Rate of Decay of the Bernstein Numbers / A. Plichko // Журнал математической физики, анализа, геометрии. — 2013. — Т. 9, № 1. — С. 59-72. — Бібліогр.: 26 назв. — англ.

Institution

Digital Library of Periodicals of National Academy of Sciences of Ukraine
id nasplib_isofts_kiev_ua-123456789-106737
record_format dspace
spelling nasplib_isofts_kiev_ua-123456789-1067372025-02-09T17:38:54Z Rate of Decay of the Bernstein Numbers Plichko, A. We show that if a Banach space X contains uniformly complemented l₂ⁿ 's then there exists a universal constant b = b(X) > 0 such that for each Banach space Y, and any sequence dn ↓ 0 there is a bounded linear operator T : X → Y with the Bernstein numbers bn(T) of T satisfying b⁻¹dn ≤ bn(T) ≤ bdn for all n. Показано, что для B-выпуклого сепарабельного пространства X, произвольного банахова пространства Y и любой последовательности dn ↓ 0 существует такой ограниченный линейный оператор T : X → Y и b > 0, что для всех чисел Бернштейна bn(T) оператора T имеем для любого n b⁻¹dn ≤ bn(T) ≤ bdn. The author express his thanks to T. Oikhberg and M. Popov for valuable consultations. 2013 Article Rate of Decay of the Bernstein Numbers / A. Plichko // Журнал математической физики, анализа, геометрии. — 2013. — Т. 9, № 1. — С. 59-72. — Бібліогр.: 26 назв. — англ. 1812-9471 https://nasplib.isofts.kiev.ua/handle/123456789/106737 en Журнал математической физики, анализа, геометрии application/pdf Фізико-технічний інститут низьких температур ім. Б.І. Вєркіна НАН України
institution Digital Library of Periodicals of National Academy of Sciences of Ukraine
collection DSpace DC
language English
description We show that if a Banach space X contains uniformly complemented l₂ⁿ 's then there exists a universal constant b = b(X) > 0 such that for each Banach space Y, and any sequence dn ↓ 0 there is a bounded linear operator T : X → Y with the Bernstein numbers bn(T) of T satisfying b⁻¹dn ≤ bn(T) ≤ bdn for all n.
format Article
author Plichko, A.
spellingShingle Plichko, A.
Rate of Decay of the Bernstein Numbers
Журнал математической физики, анализа, геометрии
author_facet Plichko, A.
author_sort Plichko, A.
title Rate of Decay of the Bernstein Numbers
title_short Rate of Decay of the Bernstein Numbers
title_full Rate of Decay of the Bernstein Numbers
title_fullStr Rate of Decay of the Bernstein Numbers
title_full_unstemmed Rate of Decay of the Bernstein Numbers
title_sort rate of decay of the bernstein numbers
publisher Фізико-технічний інститут низьких температур ім. Б.І. Вєркіна НАН України
publishDate 2013
url https://nasplib.isofts.kiev.ua/handle/123456789/106737
citation_txt Rate of Decay of the Bernstein Numbers / A. Plichko // Журнал математической физики, анализа, геометрии. — 2013. — Т. 9, № 1. — С. 59-72. — Бібліогр.: 26 назв. — англ.
series Журнал математической физики, анализа, геометрии
work_keys_str_mv AT plichkoa rateofdecayofthebernsteinnumbers
first_indexed 2025-11-28T19:35:12Z
last_indexed 2025-11-28T19:35:12Z
_version_ 1850064024113250304
fulltext Journal of Mathematical Physics, Analysis, Geometry 2013, vol. 9, No. 1, pp. 59–72 Rate of Decay of the Bernstein Numbers A. Plichko Department of Mathematics, Cracow University of Technology Cracow, Poland E-mail: aplichko@pk.edu.pl Received August 2, 2012 We show that if a Banach space X contains uniformly complemented `n 2 ’s then there exists a universal constant b = b(X) > 0 such that for each Banach space Y , and any sequence dn ↓ 0 there is a bounded linear operator T : X → Y with the Bernstein numbers bn(T ) of T satisfying b−1dn ≤ bn(T ) ≤ bdn for all n. Key words: B-convex space, Bernstein numbers, Bernstein pair, uni- formly complemented `n 2 , superstrictly singular operator. Mathematics Subject Classification 2010: 47B06, 47B10. To the memory of M.I. Kadets 1. Introduction and Main Result Let X, Y be Banach spaces and let L(X,Y ) be the space of all bounded linear operators from X to Y . Notationally, all spaces are infinite dimensional real Banach spaces unless otherwise specified. Definition 1. An operator T ∈ L(X, Y ) is called superstrictly singular (SSS for short; finitely strictly singular in other terminology) if there are no number c > 0 and no sequence of subspaces En ⊂ X, dimEn = n, such that ‖Tx‖ ≥ c‖x‖ for all x in ∪n En . (1) Put for an operator T bn(T ) = sup min x∈SE ‖Tx‖ , (2) where supremum is taken over all n-dimensional subspaces E ⊂ X and SE is the unit sphere of E. Evidently, ‖T‖ = b1(T ) ≥ b2(T ) ≥ · · · ≥ 0 , c© A. Plichko, 2013 A. Plichko T is SSS if and only if bn(T ) → 0 as n →∞ and the greatest constant c for which (1) is satisfied, is equal to limn→∞ bn(T ) . Obviously, every compact operator is SSS and T has finite rank if and only if bn(T ) = 0 beginning with some integer n. Observe, that if T has infinite rank then for each n the set In(T ) of all n-dimensional subspaces E such that T |E are injective, is dense in the set of all n-dimensional subspaces. Then the formula (2) turns into the following one bn(T ) = sup E∈In(T ) 1 ‖(T |E)−1‖ . (3) The bn(T ), which are called the Bernstein numbers, were considered in Ap- proximation and Operator Theory. The constants bn(T ) show how small is the T -image of the unit sphere SX . For a compact operator T in a Hilbert space H they coincide with s-numbers which are defined as eigenvalues of the operator (T ∗T )1/2. There are several generalizations of s-numbers to Banach spaces (see below for details). The Bernstein numbers take origin (see Whitley [26]) in the following classical inequalities: If pn is a polynomial of degree at most n, then for its derivative ‖p′n‖ ≤ n2‖pn‖ , the norm being the supremum norm on [−1, 1] (Markov [13]). If qn is a complex trigonometric polynomial of degree at most n, then ‖q′n‖ ≤ n‖qn‖ , the norm being the supremum norm on the unite circle (Bernstein [2]). Both of these inequalities have the same form: A Banach space, a derivation operator D and an (n + 1)-dimensional subspace F are given. The conclusion estimates the value of ‖D|F ‖. From this point of view it is natural to ask to what extent the norm depend on F . In particular, what improvement is possible, i.e. what is the best possible constant inf{‖D|F ‖ : dimF = n} ? It appears that this constant is equal to n [26]. Considering the inverse of D we arrive to the notion of the Bernstein numbers. We find bn(T ) as far as in (Krein/Krasnoselskĭı/Milman [11]). After (Mitiagin/Henkin [16]), SSS operators 60 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers were introduced implicitly by Mitiagin and PeÃlczyński [17] and explicitly, under the name “operators of the class C∗ 0”, by Milman [15]. The important role has been played by Pietsch’s paper [19] where systematic theory of abstract s-numbers in Banach spaces was developed (see also [20]). In particular, Pietsch noted the importance of duality and of the principle of local reflexivity. The term “superstrictly singular operator” was introduced in (Hin- richs/Pietsch [7]), where this class was investigated by machinery of superideals, and by Mascioni [14]. For further progress in the theory of SSS operators in general Banach spaces see e.g. (Plichko [24]) and (Flores/Hernández/Raynaud [6]). As we noted, an operator T is SSS if and only if bn(T ) ↓ 0. One can pose an “inverse” problem. Let X, Y be Banach spaces and dn ↓ 0. Does there exist T ∈ L(X,Y ) such that bn(T ) = dn for every n? We have a little chance to obtain a positive answer. So, we will consider a weaker question which is natural in a more general setting. According to Pietsch [21], a map s which assigns to each bounded linear ope- rator T between Banach spaces a unique sequence (sn(T )), is called an s-function if for all Banach spaces W,X, Y, Z: 1. ‖T‖ = s1(T ) ≥ s2(T ) ≥ · · · ≥ 0 for all T ∈ L(X, Y ). 2. sn(S + T ) ≤ sn(S) + ‖T‖ for all S, T ∈ L(X,Y ) and all n. 3. sn(RST )≤ ‖R‖sn(S)‖T‖ for all T ∈ L(X,Y ), S∈ L(Y,Z) and R∈ L(Z, W ). 4. If T ∈ L(X,Y ) and rank T < n, then sn(T ) = 0. 5. sn(I) = 1 for all n, where I is the identity map of `n 2 . The scalar sn(T ) is called the nth s-number of the operator T . The Bernstein numbers are s-numbers. Another example of s-numbers are the approximation numbers defined by the formula an(T ) = inf{‖T − L‖ : L ∈ L(X,Y ), rank L < n}. These numbers are connected with the well known approximation property of Banach spaces and characterize the ideal of approximable operators: an(T ) → 0 if and only if T is approximable. The approximation numbers are the largest s-numbers [19]. Aksoy and Lewicki [1] have introduced the following general concept. Definition 2. Banach spaces X and Y are said to form a Bernstein pair with respect to s-numbers sn if for any sequence dn ↓ 0, there exists T ∈ L(X, Y ) such Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 61 A. Plichko that (sn(T )) is equivalent to (dn), i.e. there is a constant b depending only on T such that for every n b−1dn ≤ sn(T ) ≤ bdn . This definition was motivated by well known Bernstein’s “lethargy” theorem [4] and is a generalization of Bernstein pair with respect to the approximation numbers (see Hutton/Morell/Retherforsd [8, 9]). Note that Hutton, Morell and Retherforsd implicitly refereed Bernstein’s lethargy theorem to [3]. In [8, 9] it was proved that many pairs of classical Banach spaces form the Bernstein pair with respect to the approximation numbers. The authors advanced a hypothesis that all couples of Banach spaces form Bernstein pairs (with respect to approximation numbers). Aksoy and Lewicki [1] showed that many classical Banach spaces form Bernstein pairs with respect to all s-numbers. Detailed investigations of “rate of decay” of many s-numbers (Kolmogorov, Gelfand, Weyl, Hilbert,. . . numbers) was carried out by Oikhberg [18]. We consider a similar question for the Bernstein numbers. Ideal properties of the Bernstein numbers was considered by Samarskĭı [25] and Pietsch [21]. First, we present simple examples of pairs (X,Y ) which are not Bernstein with respect to the Bernstein numbers. They are, in fact, well known (see e.g. Mitiagin/PeÃlczyński [17]). For a subspace E of a Banach space X denote by λ(E, X) the relative projec- tion constant λ(E, X) = inf ‖P‖ , where inf is taken over all projections P of X onto E. Given a Banach space X put pn(X) = inf{λ(E,X) : E ⊂ X , dimE = n}. Note that one can take infimum here only over a dense subset of all n-dimensional subspaces. Proposition 1. Let T ∈ L(X,H), where H is a Hilbert space and dimT (X) = ∞. Then for every n bn(T ) ≤ 1 pn(X) ‖T‖ . P r o o f. Let b > 1 and Eb ∈ In(T ) be such that ‖(T |Eb )−1‖ < bbn(T ) (see (3)). Take the orthogonal projection Q of H onto T (Eb). Then P = (T |Eb )−1QT is a projection of X onto Eb. So λ(Eb, X) ≤ ‖P‖ ≤ ‖(T |Eb )−1‖ · ‖Q‖ · ‖T‖ < bbn(T )−1‖T‖. Hence pn(X) = inf dimE=n λ(E,X) ≤ λ(Eb, X) ≤ bbn(T )−1‖T‖. Since b > 1 is arbitrary, this implies Proposition 1. 62 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers Corollary 1. Let an operator T ∈ L(X, Y ) can be factored through a Hilbert space H : T = RS, R ∈ L(X, H), S ∈ L(H,Y ) and dimT (X) = ∞. Then for every n bn(T ) ≤ 1 pn(X) ‖R‖‖S‖ . P r o o f. Indeed, by Proposition 1, bn(T ) ≤ bn(R)‖S‖ ≤ 1 pn(X) ‖R‖‖S‖ . For operators, factored through Hilbert spaces see [12]. Definition 3. We say that a Banach space X contains no uniformly comple- mented finite-dimensional subspaces if pn(X) →∞ as n →∞. The well known Pisier space P [22, 23] contains no uniformly complemented finite-dimensional subspaces. Moreover, there exists λ > 0 such that pn(P) ≥ λ √ n for all n. Corollary 2. Every operator from a Banach space X, containing no uniformly complemented finite-dimensional subspaces, into a Hilbert space H is SSS. Corollary 3. There is λ > 0 such that for every operator T ∈ L(P,H) and every n bn(T ) ≤ 1 λ √ n ‖T‖ . R e m a r k 1. Since every n-dimensional subspace E ⊂ X is a range of a projection P : X → E with ‖P‖ ≤ √ n (Kadets/Snobar [10]), one cannot obtain a better estimation of bn(T ) with using of projections. A similar estimation for operators from C(K) into H, but with constants 4 √ n instead of √ n, was noted in [17]. Proposition 1 implies Corollary 4. Assume X contains no uniformly complemented finite-dimensional subspaces and H is a Hilbert space. Then the pair (X,H) is not Bernstein with respect to the Bernstein numbers. P r o o f. Indeed, by Proposition 1, for every T ∈ L(X, H) the sequence bn(T ) cannot go to 0 “more slowly” than 1/pn(c). Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 63 A. Plichko An n-dimensional normed space (E, ‖ ‖) is said to be a-isomorphic to `n 2 (write E a∼ `n 2 ), a > 1, if there exists an Euclidean norm ‖ ‖2 on E such that for every e ∈ E a−1‖e‖ ≤ ‖e‖2 ≤ a‖e‖. If in this definition the constants a and n are inessential, we say simply about almost Euclidean subspaces. R e m a r k 2. If E a∼ `n 2 then for every subspace F ⊂ E there is a projection P : E → F with ‖P‖ ≤ a2. If E a∼ `n 2 then it have an a-orthonormal basis, i.e. a system (ei)n 1 such that ‖ei‖ = 1 for all i and for all scalars (ai) a−1 (∑n 1 a2 i )1/2 ≤ ∥∥∥ ∑n 1 aiei ∥∥∥ ≤ a (∑n 1 a2 i )1/2 . R e m a r k 3. For an a-orthonormal basis, the norm of each projection Pi , i < n, of E onto lin(ej)i 1 along to lin(ej)n i+1 is not greater than a2. Definition 4. (see e.g. [22, p. 215]). A Banach space X contains uniformly complemented `n 2 ’s if there is a constant d such that for every ε > 0 and for each n there is a subspace E ⊂ X and a projection P : X → E such that E 1+ε∼ `n 2 and ‖P‖ < d. Note that by Dvoretzky’s theorem, if this holds for some ε, then it automat- ically holds for all ε. We will show that uniformly complemented almost Euclidean subspaces play a crucial role in constructing of Bernstein pairs. Theorem 1. If a Banach space X contains uniformly complemented `n 2 ’s then there exists a universal constant b = b(X) > 0 such that for each Banach space Y , and any sequence dn ↓ 0 there exist a bounded linear operator T : X → Y such that for all n b−1dn ≤ bn(T ) ≤ bdn . Corollary 5. Let a Banach space X contain uniformly complemented `n 2 ’s. Then for every Banach space Y the pair (X, Y ) is Bernstein with respect to the Bernstein numbers. A Banach space X is B-convex if it does not contain `n 1 ’s uniformly. Since every B-convex Banach space contains uniformly complemented `n 2 ’s (see e.g. [22, pp. 208, 215]), we have 64 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers Corollary 6. Let X be a B-convex Banach space. Then for every Banach space Y the pair (X, Y ) is Bernstein with respect to the Bernstein numbers. This corollary recalls us the well known Davis–Johnson compact non-nuclear operator in a B-convex Banach space [5]. Problem. Does (X, X) form a Bernstein pair with respect to the Bernstein numbers for every Banach space X? 2. Proof of the Main Result To prove Theorem 1 we construct a “bounded minimal system” consisting of almost Euclidean subspaces of arbitrary large dimensions in an arbitrary Banach space containing uniformly complemented `n 2 ’s. Lemma 1. Let X contain uniformly complemented `n 2 ’s, with corresponding ε and d and let d′ > (1 + ε)4d. Then for each finite codimensional subspace X ′ ⊂ X, each finite dimensional subspace E ⊂ X and each m there exists a subspace E′ ⊂ X ′ , E′ 1+ε∼ `m 2 and a projection P ′ : X → E′ with ‖P ′‖ < d′ and kerP ′ ⊃ E. P r o o f. By definition, one can find an almost Euclidean subspace E0 ⊂ X, dimE0 > m + dimE + dimX/X0 and a projection P0 : X → E0 with ‖P0‖ < d. Since E0 is almost Euclidean, by Remark 2, there exists a projection Q0 : E0 → E1 := E0 ∩ X0 with ‖Q0‖ ≤ (1 + ε)2. Obviously, dimE1 ≥ m + dimE. Put P1 = Q0P0. Then P1 is a projection of X onto E1 and ‖P1‖ ≤ (1 + ε)2d. Since E1 is almost Euclidean, by Remark 2, there exists a subspace E′ ⊂ E1 , dimE′ = m, and a projection Q1 : E1 → E′ with ‖Q1‖ ≤ (1 + ε)2 and kerQ1 ⊃ P (E). Then P ′ = Q1P1 is the desired projection. Lemma 2. Let X contain uniformly complemented `n 2 ’s, with corresponding ε and d. Then for any subsequence (mk)∞k=1 of integers there are subspaces Ek ⊂ X, each Ek 1+ε∼ `mk 2 , with projections Pk : X → Ek , ‖Pk‖ ≤ d, such that each Ei, i 6= k, belongs to kerPk. P r o o f. Of course, one must write ‖Pk‖ ≤ d′, where d′ is from the previous lemma, but the exact value of the constant d is non-essential here. We present a construction only. Take, by definition, a subspace E1 ⊂ X, E1 1+ε∼ `m1 2 , and a projection P1 : X → E1 with ‖P1‖ ≤ d. Then take, by Lemma 1, a subspace E2 ⊂ kerP1, E2 1+ε∼ `m2 2 , and a projection P2 : X → E2 with ‖P2‖ ≤ d and kerP2 ⊃ E1. Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 65 A. Plichko Next take, by Lemma 1, a subspace E3 ⊂ kerP1 ∩ kerP2, E3 1+ε∼ `m3 2 , and a projection P3 : X → E3 with ‖P3‖ ≤ d and kerP3 ⊃ (E1 ∪ E2), and so on. R e m a r k 4. Let (Ek) be subspaces from Lemma 2. Then for every k ≥ 1 X = E1 ⊕ E2 ⊕ · · ·Ek ⊕ (∩k i=1 kerPi). Next, using the Dvoretzky theorem, we construct in an arbitrary Banach space a subspace with “bounded minimal system” consisting of almost Euclidean subspaces of arbitrary large dimensions. Denote by [A] the closed linear span of the set A. Lemma 3. Let Y be a Banach space, ε > 0, and (mk)∞k=1 be a sequence of integers. Then there exist subspaces Fk ⊂ Y , each Fk 1+ε∼ `mk 2 , and projections Qk : [Fi]∞1 → lin(Fi)k 1 along [Fi]∞k+1 with ‖Qk‖ ≤ 1 + ε. P r o o f. Lemma 3 is a standard combination of the Dvoretzky and Mazur theorems. We present a construction only. Recall that a subset Φ ⊂ Y ∗ λ-norms a subspace F ⊂ Y if for every y ∈ SF there is ϕ ∈ Φ such that ϕ(y) ≥ λ. For each finite-dimensional subspace F ⊂ Y and 0 < λ < 1 there is a finite set Φ ⊂ SY ∗ which λ-norms F . So, take a subspace F1 ⊂ Y , F1 1+ε∼ `m1 2 , and a finite subset Φ1 ⊂ SX∗ which (1 + ε)−1-norms F1. Then take a subspace F2 ⊂ Φ>1 := {y ∈ Y : ϕ(y) = 0 for all ϕ ∈ Φ1}, F2 1+ε∼ `m2 2 , and a finite subset Φ2 ⊂ SX∗ which (1 + ε)−1-norms F1 + F2. Next, take a subspace F3 ⊂ Φ>2 , F3 1+ε∼ `m3 2 , and a finite subset Φ3 ⊂ SX∗ which (1 + ε)−1-norms F1 + F2 + F3, and so on. In the proof we use diagonal operators in Euclidean spaces whose Bernstein numbers are well known. Definition 5. Let E and F be linear spaces with bases (en)m 1 and (fn)m 1 . Let (dn)m 1 be scalars. A map D (∑m 1 anen ) = ∑m 1 dnanfn is called the diagonal operator corresponding to (en), (fn) and (dn). 66 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers Proposition 2. (sf. [19, Th. 7.1]). Let (en)m 1 be the standard basis of `m 2 , d1 ≥ d2 ≥ · · · ≥ dm ≥ 0 and D be the diagonal operator in `m 2 corresponding to (en)m 1 and (dn)m 1 . Then for all n ≤ m min{‖Dx‖ : x ∈ lin(ej)n 1 , ‖x‖ = 1} = dn and max{‖Dx‖ : x ∈ lin(ej)m n , ‖x‖ = 1} = dn . Corollary 7. Assume m-dimensional normed spaces E and F have a-orthonormal bases (en)m 1 and (fn)m 1 , d1 ≥ d2 ≥ · · · ≥ dm ≥ 0 and D is the diagonal operator corresponding to (en), (fn), (dn). Then there is c > 1, depending only on a, such that for all n ≤ m min{‖Dx‖ : x ∈ lin(ej)n 1 , ‖x‖ = 1} ≥ dn c and max{‖Dx‖ : x ∈ lin(ej)m n , ‖x‖ = 1} ≤ cdn . P r o o f of Theorem 1. Let dn ↓ 0. Take a subsequence (nk)∞k=1 of integers which approach to ∞ so quickly that for all k ≥ 1 dnk+1 < 1 4 dnk . (4) Hence, for every k ≥ 1 ∞∑ i=k+1 dni < 1 2 dnk . (5) Let 0 < ε < 1 , Ek be subspaces from Lemma 2 and Fk , k ≥ 1, be subspaces from Lemma 3 with mk := nk − nk−1 (and n0 = 0). Take in each Ek and each Fk some (1 + ε)-orthonormal bases. Rearrange these bases in the natural way, putting first the basis e1, . . . , en1 of E1, then the basis en1+1, . . . , en2 of E2 and so on; and similarly for Y . We obtain systems (en)∞1 in X and (fn)∞1 in Y . Put Nk = {n : nk−1 < n ≤ nk}. Using Corollary 7, (with c from this corollary) we construct for every k ≥ 1 the diagonal operator Dk : Ek → Fk corresponding to the bases (en), (fn) and scalars (dn) , n ∈ Nk, such that for all n ∈ Nk min { ‖Dkx‖ : x ∈ [ej ]nnk−1+1 , ‖x‖ = 1 } ≥ dn c and (6) max {‖Dkx‖ : x ∈ [ej ]nk n , ‖x‖ = 1} ≤ cdn . (7) Let Pk be the projections from Lemma 2. For every x ∈ X put Tx = ∞∑ i=1 DiPix (8) (bellow we will show that the series (8) converges for each x ∈ X). Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 67 A. Plichko We make forth estimations. Let d be from Lemma 2 and c be from Corollary 7. 1. For every k ≥ 1 and x ∈ X, ‖x‖ = 1, ∞∑ i=k+1 ‖DiPix‖ < 2cddnk . Indeed, Pix ∈ Ei and ‖Pix‖ ≤ ‖Pi‖‖x‖ ≤ d for all i, so ∞∑ i=k+1 ‖DiPix‖ ≤ by (7) ≤ cddnk+1 + ∞∑ i=k+2 cddni+1 ≤ by (5) ≤ cddnk + c 2 ddnk+1 < 2cddnk . In particular, this inequality shows that series (8) converges for each x ∈ X, so T is well defined. 2. For every k ≥ 1 and n ∈ Nk sup { ‖Tx‖ : x ∈ [ej ]nk n ⊕ ∩k i=1 kerPi , ‖x‖ = 1 } ≤ 3cddn (by Remark 4, the sum here is direct). Indeed, take x ∈ [ej ]nk n ⊕ ∩k i=1 kerPi , ‖x‖ = 1. Then, by definition of Pi, Tx = ∑∞ i=k DiPix, so ‖Tx‖ ≤ ‖DkPkx‖+ ∞∑ i=k+1 ‖DiPix‖ ≤ (by 1) ≤ ‖DkPkx‖+ 2cddnk ≤ (since ‖Pk‖ ≤ d, by (7)) ≤ cddn + 2cddn = 3cddn . 3. For every k ≥ 1 and x ∈ lin(Ei)k 1 , ‖x‖ = 1, ‖Tx‖ ≥ 1 4(1 + ε) · dnk c . We prove estimation 3 by induction. For k = 1, Tx = D1x, so 3 is followed from (6) if we take in (6) n = n1. Suppose k > 1, estimation 3 is proved for k−1, and x ∈ lin(Ei)k 1 , ‖x‖ = 1. Then x = x1 + x2 , x1 ∈ lin(Ei)k−1 1 , x2 ∈ Ek , and, by the construction of Pi, Tx1 = k−1∑ i=1 DiPix1 and Tx2 = DkPkx2. 68 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers Hence, by the construction of Di, Tx1 ∈ lin(Fi)k−1 1 and Tx2 ∈ Fk . By the construction of projections Qi from Lemma 3, Qk−1Tx = Qk−1Tx1 + Qk−1Tx2 = Tx1 and (Qk −Qk−1)Tx = (Qk −Qk−1)Tx1 + (Qk −Qk−1)Tx2 = Tx2 . Since ‖Qi‖ ≤ 1 + ε, hence ‖Qi −Qi−1‖ ≤ 2(1 + ε). So, ‖Tx‖ ≥ 1 1 + ε ‖Qk−1Tx‖ = 1 1 + ε ‖Tx1‖ (9) and ‖Tx‖ ≥ 1 2(1 + ε) ‖(Qk −Qk−1)Tx‖ = 1 2(1 + ε) ‖Tx2‖. (10) Since ‖x‖ = 1, we have that either ‖x1‖ ≥ 1 2 or ‖x2‖ ≥ 1 2 . If ‖x1‖ ≥ 1 2 , then by the induction assumption ‖Tx‖ by (9) ≥ 1 1 + ε ‖Tx1‖ ≥ 1 1 + ε · 1 2 · 1 4(1 + ε) · dnk−1 c by (4) ≥ 1 2(1 + ε)2 · 1 4 · 4dnk c since ε < 1≥ 1 4(1 + ε) · dnk c . If ‖x2‖ ≥ 1 2 , then ‖Tx‖ by (10) ≥ 1 2(1 + ε) ‖Tx2‖ by (6) ≥ 1 2(1 + ε) · 1 2 · dnk c = 1 4(1 + ε) · dnk c . Therefore, 3 is proved. 4. For every k ≥ 1 and n ∈ Nk min{‖Tx‖ : x ∈ lin(ej)n 1 , ‖x‖ = 1} ≥ 1 4(1 + ε) · dn c . Indeed, take x ∈ lin(ej)n 1 , ‖x‖ = 1, where n ∈ Nk. Then, as in 3, x = x1+x2, x1 ∈ lin(Ei)k−1 1 , x2 ∈ Ek; either ‖x1‖ ≥ 1 2 or ‖x2‖ ≥ 1 2 ; Tx1 ∈ lin(Fi)k−1 1 , Tx2 ∈ Fk, and the inequalities (9), (10) hold. Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 69 A. Plichko If ‖x1‖ ≥ 1 2 , then ‖Tx‖ ≥ by 3 ≥ 1 4(1 + ε) · dnk−1 c ≥ 1 4(1 + ε) · dn c . If ‖x2‖ ≥ 1 2 , then ‖Tx‖ by (10) ≥ 1 2(1 + ε) ‖Tx2‖ = 1 2(1 + ε) ‖Dkx2‖ by (6) ≥ 1 4(1 + ε) · dn c . Therefore, 4 is proved. Put b = max{3cd, 4(1 + ε)c}. Inequality 4 shows that for all n bn(T ) ≥ b−1dn. Let G ⊂ X be an n-dimensional subspace and n ∈ Nk. Then, by Remark 4, G ∩ ( [ej ]nk n ⊕ ∩k i=1 kerPi ) 6= 0. So, the inequality 4 confirms that for all n min x∈SG ‖Tx‖ ≤ bdn , i.e. bn(T ) ≤ bdn . Acknowledgements. The author express his thanks to T. Oikhberg and M. Popov for valuable consultations. References [1] A.G. Aksoy and G. Lewicki, Diagonal Operators, s-Numbers and Bernstein Pairs. — Note Mat. 17 (1999), 209–216. [2] S. Bernstein, Sur l’ordre de la Meilleure Approximation des Fonctions Continues par des Polynômes de Degré Donné. — Mém. Acad. Roy. Belgique 4 (1912), 1–104. [3] S. Bernstein, Leçons sur les Propriétés Extrémales et la Meilleure Approximation des Fonctions Analytiques d’une Variable Réelle. Paris, Gauthier–Villars, X, 1926, 207 pp., Collection de monographies sur la théorie des fonctions, publ. par E. Borel. [4] S.N. Bernstein, Sur le Problème Inverse de la Théorie de la Meilleure Approximation des Fonctions Continues. — Comp. Rend. Paris 206 (1938), 1520–1523. 70 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 Rate of Decay of the Bernstein Numbers [5] W.J. Davis and W.B. Johnson, Compact, Non-Nuclear Operators. — Studia Math. 51 (1974), 81–85. [6] J. Flores, F.L. Hernández, and Y. Raynaud, Super Strictly Singular and Cosingular Operators and Related Classes. — J. Operator Theory 67 (2012), 121–152. [7] A. Hinrichs and A. Pietsch, Closed Ideals and Dvoretzky’s Theorem for Operators. (unpublished). [8] C.V. Hutton, J.S. Morrell, and J.R. Retherford, Approximation Numbers and Kol- mogoroff Diameters of Bounded Linear Operators. — Bull. Amer. Math. Soc. 80 (1974), 462–466. [9] C.V. Hutton, J.S. Morrell, and J.R. Retherford, Diagonal Operators, Approxima- tion Numbers and Kolmogoroff Diameters. — J. Approx. Theory 16 (1976), 48–80. [10] M.I. Kadets and M.G. Snobar, Certain Functionals on the Minkowski Compactum. — Math. Notes 10 (1971), 694–696. [11] M.G. Krein, M.A. Krasnoselskĭı, and D.P. Milman, On the Defect Numbers of Linear Operators in Banach Space and on Some Geometric Problems. — Sb. Trud. Inst. Mat. Akad. Nauk USSR 11 (1948), 97–112. (Russian) [12] J. Lindenstrauss and A. PeÃlczyński, Absolutely Summing Operators on Lp-Spaces and their Applications. — Studia Math. 29 (1968), 275–328. [13] A.A. Markov, On a Question of D.I. Mendeleev. — Zapiski Imperatorskŏı Akademii Nauk 62 (1889), 1–24. (Russian). [14] V. Mascioni, A Restriction-Extension Property for Operators on Banach Spaces. — Rocky Mountain J. Math. 24 (1994), 1497–1507. [15] V.D. Milman, Operators of the Class C0 and C∗0 . — Teor. Funktsii, Funkts. Anal. i Prilozen. 10 (1970), 15–26. (Russian) [16] B.S. Mitiagin and G.M. Henkin, Inequalities Between Various n-Widths. — Trudy Sem. Funkts. Anal. (Voronezh) 7 (1963), 97–103. (Russian) [17] B.S. Mitiagin and A. PeÃlczyński, Nuclear Operators and Approximative Dimension. Proceedings of International Congress of Mathematicians. Moscow, 1966, 366–372. [18] T. Oikhberg, Rate of Decay of s-Numbers. — J. Approx. Theory 163 (2011), 311– 327. [19] A. Pietsch, s-Numbers of Operators in Banach Spaces. — Studia Math. 51 (1974), 201–223. [20] A. Pietsch, Eigenvalues and s-Numbers. Greest and Portig, Leipzig, and Cambridge University Press, Cambridge, 1987. [21] A. Pietsch, Bad Properties of the Bernstein Numbers. — Studia Math. 184 (2008), 263–269. Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1 71 A. Plichko [22] G. Pisier, Probability Methods in the Geometry of Banach Spaces. — Lect. Notes Math. 1206 (1986), 167–241. [23] G. Pisier, Counterexamples to a Conjecture of Grothendieck. — Acta Math. 151 (1983), 181–208. [24] A. Plichko, Superstrictly Singular and Superstrictly Cosingular Operators. — In: Functional Analysis and its Applications. Proc. of the Stefan Banach Conf., (V. Kadets and W. Zelazko, Eds.) Elsevier Sci. Publ., North-Holland Math. Stud. 197 (2004), 239–256. [25] V.G. Samarskĭı, The Construction of an Operator Ideal by Means of Bernštĕın s-Numbers. — Vestnik Leningrad. Univ. Mat. Mekh. Astronom. 1978, vyp. 1, 77– 80. (Russian) [26] R.J. Whitley, Markov and Bernstein’s Inequalities and Compact and Strictly Sin- gular Operators. — J. Approx. Theory 34 (1982), 277–285. 72 Journal of Mathematical Physics, Analysis, Geometry, 2013, vol. 9, No. 1