Introduction
Throughout the paper all spaces, scalars and operators are real. We say that a normed space X X X has the Daugavet property , shortly X ∈ D P r X \in\mathrm{DPr} X ∈ DPr , if
∥ Id + T ∥ = 1 + ∥ T ∥ , \lVert\operatorname{Id} + T \rVert= 1 + \lVert T \rVert, ∥ Id + T ∥ = 1 + ∥ T ∥ ,
for every rank-one operator T : X → X T:X \to X T : X → X . The following characterization was obtained in [5 Lemma 2.2] .
Theorem 1.1. Let X X X be a normed space. The following assertions are equivalent:
(1) X X X has the Daugavet property .
(2) For every y ∈ S X y \in S_X y ∈ S X , x ∗ ∈ S X ∗ x^* \in S_{X^*} x ∗ ∈ S X ∗ , and every ε > 0 \varepsilon> 0 ε > 0 , there exists x ∈ S X x \in S_X x ∈ S X such that
x ∗ ( x ) ≥ 1 − ε and ∥ x + y ∥ ≥ 2 − ε . x^*(x) \ge1 - \varepsilon\quad\text{and} \quad\lVert x + y \rVert\ge2 - \varepsilon. x ∗ ( x ) ≥ 1 − ε and ∥ x + y ∥ ≥ 2 − ε . (3) For every ε > 0 \varepsilon> 0 ε > 0 and every y ∈ S X y \in S_X y ∈ S X , the closed convex hull of
{ u ∈ ( 1 + ε ) B X ∣ ∥ y + u ∥ ≥ 2 − ε } \{u \in(1+\varepsilon)B_X \mid\lVert y+u\rVert\ge2-\varepsilon\} { u ∈ ( 1 + ε ) B X ∣ ∥ y + u ∥ ≥ 2 − ε } contains $S_X.
Whether the Daugavet property passes from two factors to their injective or projective tensor product is a longstanding question; see [10 Section 6] . The version with only one Daugavet factor was answered negatively by Kadets, Kalton and Werner in the complex setting [4 Theorem 4.2 and Corollary 4.3] . Several positive results are known under additional assumptions. Rueda Zoca, Tradacete and Villanueva proved that the injective tensor product of two atomless L 1 L_1 L 1 -spaces has the Daugavet property, and that the projective tensor product of two L 1 L_1 L 1 -preduals with the Daugavet property also has this property [8 ] . Martín and Rueda Zoca introduced the weak operator Daugavet property and proved its stability under projective tensor products [6 Theorem 5.4] . For the injective tensor product, Rueda Zoca proved that it suffices to assume the weak operator Daugavet property of both dual spaces [7 Theorem 3.5] . We show that the Daugavet property need not be preserved under taking injective or projective tensor squares. In both cases, our approach is based on the standard technique of constructing a sequence of finite-dimensional extensions.
The last part of this paper will consider a short result regarding the isomorphic theory of Daugavet spaces. It is a known result [11 Corollary 8.4.12] that every separable Banach space containing an isomorphic copy of C [ 0 , 1 ] C[0,1] C [ 0 , 1 ] admits an equivalent norm under which it has the Daugavet property. Thus, although the Daugavet property is not invariant under equivalent renormings, the presence of an isomorphic copy of C [ 0 , 1 ] C[0,1] C [ 0 , 1 ] provides a sufficient condition for a separable Banach space to be renormed to have the Daugavet property. We will show that this is not true for nonseparable Banach spaces.
We recall the tensor notation. For u = ∑ i = 1 q x i ⊗ y i ∈ E ⊗ F u=\sum_{i=1}^{q}x_i\otimes y_i\in E\otimes F u = ∑ i = 1 q x i ⊗ y i ∈ E ⊗ F , put
∥ u ∥ ε = sup x ∗ ∈ B E ∗ , y ∗ ∈ B F ∗ ∣ ∑ i = 1 q x ∗ ( x i ) y ∗ ( y i ) ∣ , ∥ u ∥ π = inf u = ∑ i a i ⊗ b i ∑ i ∥ a i ∥ ∥ b i ∥ . \lVert u\rVert_{\varepsilon}=\sup_{x^{*}\in B_{E^{*}},\,y^{*}\in B_{F^{*}}}\left|\sum_{i=1}^{q}x^{*}(x_i)y^{*}(y_i)\right|,\qquad\lVert u\rVert_{\pi}=\inf_{u=\sum_{i}a_i\otimes b_i}\sum_i\lVert a_i\rVert\lVert b_i\rVert. ∥ u ∥ ε = x ∗ ∈ B E ∗ , y ∗ ∈ B F ∗ sup i = 1 ∑ q x ∗ ( x i ) y ∗ ( y i ) , ∥ u ∥ π = u = ∑ i a i ⊗ b i inf i ∑ ∥ a i ∥ ∥ b i ∥ .
The completions are denoted by E ⊗ ^ ε F E\widehat{\otimes}_{\varepsilon}F E ⊗ ε F and E ⊗ ^ π F E\widehat{\otimes}_{\pi}F E ⊗ π F . The tensor u u u defines an operator
E ∗ ⟶ F , x ∗ ⟼ ∑ i = 1 q x ∗ ( x i ) y i , E^{*}\longrightarrow F,\qquad x^{*}\longmapsto\sum_{i=1}^{q}x^{*}(x_i)y_i, E ∗ ⟶ F , x ∗ ⟼ i = 1 ∑ q x ∗ ( x i ) y i ,
whose operator norm equals ∥ u ∥ ε \lVert u\rVert_{\varepsilon} ∥ u ∥ ε . In finite dimensions this identifies E ⊗ ε F E\otimes_{\varepsilon}F E ⊗ ε F with all operators from E ∗ E^{*} E ∗ to F F F . Moreover, ( E ⊗ ^ π F ) ∗ (E\widehat{\otimes}_{\pi}F)^{*} ( E ⊗ π F ) ∗ is the space of bounded bilinear forms on E × F E\times F E × F , with its usual norm; see [9 ] .
Let H = ( R 2 , ∥ ⋅ ∥ 2 ) H=(\mathbb{R}^{2},\lVert\cdot\rVert_{2}) H = ( R 2 , ∥ ⋅ ∥ 2 ) , with standard basis e 1 , e 2 e_1,e_2 e 1 , e 2 , and identify H ∗ H^{*} H ∗ with H H H by the inner product. In particular, the map
H ⊗ ε H ⟶ ( M 2 ( R ) , ∥ ⋅ ∥ op ) , x ⊗ y ⟼ y x T , (1) H\otimes_{\varepsilon}H\longrightarrow(M_2(\mathbb{R}),\lVert\cdot\rVert_{\operatorname{op}}),\qquad x\otimes y\longmapsto yx^{\mathsf{T}}, \tag*{(1)} H ⊗ ε H ⟶ ( M 2 ( R ) , ∥ ⋅ ∥ op ) , x ⊗ y ⟼ y x T , (1)
is an onto linear isometry. Under this identification
z 0 = e 1 ⊗ e 1 + e 2 ⊗ e 2 z_0=e_1\otimes e_1+e_2\otimes e_2 z 0 = e 1 ⊗ e 1 + e 2 ⊗ e 2
corresponds to Id H \operatorname{Id}_{H} Id H and has injective norm one. We put ℓ 1 0 = ℓ ∞ 0 = { 0 } \ell_1^0=\ell_{\infty}^0=\{0\} ℓ 1 0 = ℓ ∞ 0 = { 0 } , and interpret empty sums as zero.
Theorem 1.2. There exist separable Banach spaces X X X and Y Y Y , both with the Daugavet property, such that neither
Z ε = X ⊗ ^ ε X nor Z π = Y ⊗ ^ π Y Z_{\varepsilon}=X\widehat{\otimes}_{\varepsilon}X\qquad\text{nor}\qquad Z_{\pi}=Y\widehat{\otimes}_{\pi}Y Z ε = X ⊗ ε X nor Z π = Y ⊗ π Y has the Daugavet property. More precisely, B Z ε B_{Z_{\varepsilon}} B Z ε has a strongly exposed point, and B Z π ∗ B_{Z_{\pi}^{*}} B Z π ∗ has a weak-star strongly exposed point. There are rank-one operators T ε T_{\varepsilon} T ε on Z ε Z_{\varepsilon} Z ε and T π T_{\pi} T π on Z π Z_{\pi} Z π such that, for ν ∈ { ε , π } \nu\in\{\varepsilon,\pi\} ν ∈ { ε , π } ,
∥ T ν ∥ = 1 , ∥ Id Z ν + T ν ∥ ≤ 199 100 < 2. \lVert T_{\nu}\rVert=1,\qquad\lVert\operatorname{Id}_{Z_{\nu}}+T_{\nu}\rVert\leq\frac{199}{100}<2. ∥ T ν ∥ = 1 , ∥ Id Z ν + T ν ∥ ≤ 100 199 < 2. The space X X X contains H H H isometrically as the range of a contractive projection, and Y Y Y is a closed subspace of X ∗ X^{*} X ∗ .
In § 2 we describe the finite extension of an arbitrary finite-dimensional space E E E . In § 3 we extend H H H infinitely many times to construct X X X . In § 4 , we show that the norms on the finite-dimensional spaces arising in our construction can be recovered from a family of linear maps into spaces of a particularly simple form. § 5 then establishes a quantitative estimate for operators on these spaces, with constants independent of their dimensions. The injective example is completed in § 6 . In § 7 and § 8 we construct the projective example Y Y Y . § 9 records the relation with the operator Daugavet properties. § 10 is independent from the rest and presents a result on the isomorphic theory of Daugavet spaces.
A finite-dimensional extension
Let E E E be a finite-dimensional normed space, let p ∈ S E p \in S_E p ∈ S E , and let m ≥ 2 m \ge2 m ≥ 2 be an integer. We define
G ( E , p , m ) = E ⊕ R m , ∥ ( x , a ) ∥ G = max 1 ≤ j ≤ m ( ∥ x − a j p ∥ E + ∣ a j ∣ ) . (2) G(E,p,m)=E\oplus\mathbb{R}^{m},\qquad\lVert(x,a)\rVert_G=\max_{1\le j\le m}\left(\lVert x-a_jp\rVert_E+\lvert a_j\rvert\right). \tag*{(2)} G ( E , p , m ) = E ⊕ R m , ∥( x , a ) ∥ G = 1 ≤ j ≤ m max ( ∥ x − a j p ∥ E + ∣ a j ∣ ) . (2)
We always identify E E E with E ⊕ { 0 } E\oplus\{0\} E ⊕ { 0 } .
Lemma 2.1. Formula (2) defines a norm. The inclusion E ↪ G E\hookrightarrow G E ↪ G is isometric, and the maps
R : G → E R:G\to E R : G → E , R ( x , a ) = x R(x,a)=x R ( x , a ) = x , A j : G → E ⊕ 1 R A_j:G\to E\oplus_1\mathbb{R} A j : G → E ⊕ 1 R , A j ( x , a ) = ( x − a j p , a j ) A_j(x,a)=(x-a_jp,a_j) A j ( x , a ) = ( x − a j p , a j )
are contractions. If u 1 , … , u m u_1,\ldots,u_m u 1 , … , u m is the standard basis of R m \mathbb{R}^m R m and h j = ( p , u j ) h_j=(p,u_j) h j = ( p , u j ) , then
∥ h j ∥ = 1 , ∥ x + h j ∥ = 2 ( x ∈ S E ) , ∥ 1 m ∑ j = 1 m h j − p ∥ = 2 m . (3) \lVert h_j\rVert=1,\qquad\lVert x+h_j\rVert=2\quad(x\in S_E),\qquad\left\lVert\frac{1}{m}\sum_{j=1}^{m}h_j-p\right\rVert=\frac{2}{m}. \tag*{(3)} ∥ h j ∥ = 1 , ∥ x + h j ∥ = 2 ( x ∈ S E ) , m 1 j = 1 ∑ m h j − p = m 2 . (3)
Proof. Homogeneity and the triangle inequality follow from the corresponding properties of the expressions in the maximum. If the maximum is zero, then every a j a_j a j is zero, and consequently x = 0 x=0 x = 0 . For a = 0 a=0 a = 0 the maximum equals ∥ x ∥ E \lVert x\rVert_E ∥ x ∥ E . Each A j A_j A j is contractive by the definition, and
∥ x ∥ E ≤ ∥ x − a j p ∥ E + ∣ a j ∣ ≤ ∥ ( x , a ) ∥ G \lVert x\rVert_E\le\lVert x-a_jp\rVert_E+\lvert a_j\rvert\le\lVert(x,a)\rVert_G ∥ x ∥ E ≤ ∥ x − a j p ∥ E + ∣ a j ∣ ≤ ∥( x , a ) ∥ G
proves the assertion about R R R .
For h j h_j h j , all expressions in the maximum equal one. For x + h j x+h_j x + h j , the j j j th expression equals ∥ x ∥ E + 1 = 2 \lVert x\rVert_E+1=2 ∥ x ∥ E + 1 = 2 , and the triangle inequality gives the reverse bound. Finally, the difference in the last part of (3) is ( 0 , m − 1 , … , m − 1 ) (0,m^{-1},\ldots,m^{-1}) ( 0 , m − 1 , … , m − 1 ) , whose norm is 2 / m 2/m 2/ m .
The inductive construction
Put E 0 = H E_0=H E 0 = H and N − 1 = ∅ N_{-1}=\varnothing N − 1 = ∅ . Suppose that E n E_n E n has been constructed. Choose a finite set
N n = { p n , 1 , … , p n , r n } ⊂ S E n , N n − 1 ⊂ N n , N_n=\{p_{n,1},\ldots,p_{n,r_n}\}\subset S_{E_n},\qquad N_{n-1}\subset N_n, N n = { p n , 1 , … , p n , r n } ⊂ S E n , N n − 1 ⊂ N n ,
such that every point of S E n S_{E_n} S E n has distance strictly less than ε n = 2 − n \varepsilon_n=2^{-n} ε n = 2 − n from N n N_n N n . This is possible: first take a finite open ε n \varepsilon_n ε n -net, and then add all points of N n − 1 N_{n-1} N n − 1 . These points still have norm one because the earlier inclusions are isometric. Set m n = 2 n + 3 m_n=2^{n+3} m n = 2 n + 3 and define successively
E n , 0 = E n , E n , j = G ( E n , j − 1 , p n , j , m n ) ( 1 ≤ j ≤ r n ) , E n + 1 = E n , r n . (4) E_{n,0}=E_n,\qquad E_{n,j}=G(E_{n,j-1},p_{n,j},m_n)\quad(1\le j\le r_n),\qquad E_{n+1}=E_{n,r_n}. \tag*{(4)} E n , 0 = E n , E n , j = G ( E n , j − 1 , p n , j , m n ) ( 1 ≤ j ≤ r n ) , E n + 1 = E n , r n . (4)
In particular, all p n , j p_{n,j} p n , j are chosen in the same space E n E_n E n , before the extensions in (4) are performed. Put
X = ⋃ n ≥ 0 E n . X=\bigcup_{n\ge0}E_n. X = n ≥ 0 ⋃ E n .
It will be convenient to number the elementary extensions as well. Let b 0 = 0 b_0=0 b 0 = 0 and b n + 1 = b n + r n b_{n+1}=b_n+r_n b n + 1 = b n + r n . For 0 ≤ j ≤ r n 0\le j\le r_n 0 ≤ j ≤ r n set V b n + j = E n , j V_{b_n+j}=E_{n,j} V b n + j = E n , j . Thus V 0 = H V_0=H V 0 = H , E n = V b n E_n=V_{b_n} E n = V b n , and
V s + 1 = G ( V s , p s , μ s ) , p b n + j − 1 = p n , j , μ b n + j − 1 = m n . V_{s+1}=G(V_s,p_s,\mu_s),\qquad p_{b_n+j-1}=p_{n,j},\qquad\mu_{b_n+j-1}=m_n. V s + 1 = G ( V s , p s , μ s ) , p b n + j − 1 = p n , j , μ b n + j − 1 = m n .
Here the notation p s p_s p s records an elementary extension point; it does not denote a new choice of a net.
Lemma 3.1. For every $s\ge 0 there is a contractive projection ρ s : X → V s \rho_s:X\to V_s ρ s : X → V s such that, when s ≤ t s\le t s ≤ t ,
ρ s ρ t = ρ s , ρ t ρ s = ρ s . \rho_s\rho_t=\rho_s,\qquad\rho_t\rho_s=\rho_s. ρ s ρ t = ρ s , ρ t ρ s = ρ s . Moreover, ρ s x → x \rho_sx\to x ρ s x → x for every x ∈ X x\in X x ∈ X .
Proof. For each elementary extension let
R s : V s + 1 → V s , R s ( x , a ) = x . R_s : V_{s+1} \to V_s,\qquad R_s(x,a)=x. R s : V s + 1 → V s , R s ( x , a ) = x .
By Lemma 2.1 , R s R_s R s is contractive and fixes V s V_s V s . For integers t > s t>s t > s define the map from V t V_t V t to V s V_s V s by
R s , t = R s ∘ R s + 1 ∘ ⋯ ∘ R t − 1 , R s , s = Id V s . R_{s,t}=R_s\circ R_{s+1}\circ\cdots\circ R_{t-1},\qquad R_{s,s}=\operatorname{Id}_{V_s}. R s , t = R s ∘ R s + 1 ∘ ⋯ ∘ R t − 1 , R s , s = Id V s .
The rightmost map R t − 1 R_{t-1} R t − 1 is applied first. Thus R s , t R_{s,t} R s , t is a contraction, and for s ≤ t ≤ u s\leq t\leq u s ≤ t ≤ u one has
R s , u = R s , t ∘ R t , u , R s , u ∣ V t = R s , t . R_{s,u}=R_{s,t}\circ R_{t,u},\qquad R_{s,u}|_{V_t}=R_{s,t}. R s , u = R s , t ∘ R t , u , R s , u ∣ V t = R s , t .
The second equality follows because R t , u R_{t,u} R t , u fixes every vector of V t V_t V t . Consequently the prescriptions ρ s x = R s , t x \rho_sx=R_{s,t}x ρ s x = R s , t x for x ∈ V t x\in V_t x ∈ V t , t ≥ s t\geq s t ≥ s , agree on intersections. They define a contraction on ⋃ t V t \bigcup_t V_t ⋃ t V t and hence extend uniquely to a contraction ρ s : X → V s \rho_s:X\to V_s ρ s : X → V s . Its restriction to V s V_s V s is the identity, so it is a projection. The displayed compatibility relations first hold on each V u V_u V u and then on X X X by continuity.
Given x ∈ X x\in X x ∈ X and η > 0 \eta>0 η > 0 , choose v ∈ V t v\in V_t v ∈ V t with ∥ x − v ∥ < η / 2 \lVert x-v\rVert<\eta/2 ∥ x − v ∥ < η /2 . For s ≥ t s\geq t s ≥ t we have ρ s v = v \rho_sv=v ρ s v = v , and hence
∥ ρ s x − x ∥ ≤ ∥ ρ s ( x − v ) ∥ + ∥ v − x ∥ ≤ 2 ∥ x − v ∥ < η . \lVert\rho_sx-x\rVert\leq\lVert\rho_s(x-v)\rVert+\lVert v-x\rVert\leq2\lVert x-v\rVert<\eta. ∥ ρ s x − x ∥ ≤ ∥ ρ s ( x − v )∥ + ∥ v − x ∥ ≤ 2 ∥ x − v ∥ < η .
This proves the convergence. □
Write P = ρ 0 : X → H P=\rho_0:X\to H P = ρ 0 : X → H and P s = P ∣ V s P_s=P|_{V_s} P s = P ∣ V s . We will repeatedly use P ρ s = P P\rho_s=P P ρ s = P . The nested choice of the nets gives the following two facts:
P = ⋃ n ≥ 0 N n is dense in S X , and every p ∈ P occurs as p s for arbitrarily large s . (5) \mathcal{P}=\bigcup_{n\geq0}N_n\text{ is dense in }S_X,\qquad\text{and every }p\in\mathcal{P}\text{ occurs as }p_s\text{ for arbitrarily large }s. \tag*{(5)} P = n ≥ 0 ⋃ N n is dense in S X , and every p ∈ P occurs as p s for arbitrarily large s . (5)
For the first assertion, choose a normalized vector v ∈ E N v\in E_N v ∈ E N close to a given point of S X S_X S X . For any n ≥ N n\geq N n ≥ N , this same vector belongs to S E n S_{E_n} S E n , and some p ∈ N n p\in N_n p ∈ N n satisfies ∥ v − p ∥ < ε n \lVert v-p\rVert<\varepsilon_n ∥ v − p ∥ < ε n . Letting n n n increase proves the assertion. The second follows from N n ⊂ N n + 1 N_n\subset N_{n+1} N n ⊂ N n + 1 and the fact that each point of N n N_n N n is used in one of the extensions leading from E n E_n E n to E n + 1 E_{n+1} E n + 1 .
Proposition 3.2. The separable Banach space X X X has the Daugavet property.
Proof. Fix f , g ∈ S X f,g\in S_X f , g ∈ S X and η > 0 \eta>0 η > 0 . Choose 0 < τ < η / 3 0<\tau<\eta/3 0 < τ < η /3 , and then choose N N N and f 0 , g 0 ∈ S E N f_0,g_0\in S_{E_N} f 0 , g 0 ∈ S E N such that ∥ f − f 0 ∥ < τ \lVert f-f_0\rVert<\tau ∥ f − f 0 ∥ < τ and ∥ g − g 0 ∥ < τ \lVert g-g_0\rVert<\tau ∥ g − g 0 ∥ < τ . Normalized approximants exist because ⋃ n E n \bigcup_n E_n ⋃ n E n is dense in X X X . Take n ≥ N n\geq N n ≥ N so large that
ε n + 2 m n < τ . \varepsilon_n+\frac{2}{m_n}<\tau. ε n + m n 2 < τ . The vector f 0 f_0 f 0 belongs to S E n S_{E_n} S E n . By the defining net property, there is j ∈ { 1 , … , r n } j\in\{1,\ldots,r_n\} j ∈ { 1 , … , r n } with ∥ f 0 − p n , j ∥ < ε n \lVert f_0-p_{n,j}\rVert<\varepsilon_n ∥ f 0 − p n , j ∥ < ε n . Apply Lemma 2.1 at the extension E n , j = G ( E n , j − 1 , p n , j , m n ) E_{n,j}=G(E_{n,j-1},p_{n,j},m_n) E n , j = G ( E n , j − 1 , p n , j , m n ) , and denote its new vectors by h 1 , … , h m n h_1,\ldots,h_{m_n} h 1 , … , h m n . Both g 0 g_0 g 0 and p n , j p_{n,j} p n , j belong to the preceding space E n , j − 1 E_{n,j-1} E n , j − 1 , and retain norm one there. Therefore
∥ h k ∥ = 1 , ∥ g 0 + h k ∥ = 2 , ∥ 1 m n ∑ k = 1 m n h k − p n , j ∥ = 2 m n . \lVert h_k\rVert=1,\qquad\lVert g_0+h_k\rVert=2,\qquad\left\lVert\frac{1}{m_n}\sum_{k=1}^{m_n}h_k-p_{n,j}\right\rVert=\frac{2}{m_n}. ∥ h k ∥ = 1 , ∥ g 0 + h k ∥ = 2 , m n 1 k = 1 ∑ m n h k − p n , j = m n 2 . These equalities remain true in X X X . It follows that
∥ 1 m n ∑ k = 1 m n h k − f ∥ ≤ ∥ 1 m n ∑ k = 1 m n h k − p n , j ∥ + ∥ p n , j − f 0 ∥ + ∥ f 0 − f ∥ \left\lVert\frac{1}{m_n}\sum_{k=1}^{m_n}h_k-f\right\rVert
\leq\left\lVert\frac{1}{m_n}\sum_{k=1}^{m_n}h_k-p_{n,j}\right\rVert+\lVert p_{n,j}-f_0\rVert+\lVert f_0-f\rVert m n 1 k = 1 ∑ m n h k − f ≤ m n 1 k = 1 ∑ m n h k − p n , j + ∥ p n , j − f 0 ∥ + ∥ f 0 − f ∥ < 2 m n + ε n + τ < 2 τ < η , <\frac{2}{m_n}+\varepsilon_n+\tau<2\tau<\eta, < m n 2 + ε n + τ < 2 τ < η , ∥ g + h k ∥ ≥ ∥ g 0 + h k ∥ − ∥ g − g 0 ∥ > 2 − τ > 2 − η . \lVert g+h_k\rVert\geq\lVert g_0+h_k\rVert-\lVert g-g_0\rVert>2-\tau>2-\eta. ∥ g + h k ∥ ≥ ∥ g 0 + h k ∥ − ∥ g − g 0 ∥ > 2 − τ > 2 − η . The Daugavet property follows now from Theorem 1.1 (3). □
Norming representations of the finite stages
For r ≥ 0 r \ge0 r ≥ 0 put F r = H ⊕ 1 ℓ 1 r F_r = H \oplus_1 \ell_1^r F r = H ⊕ 1 ℓ 1 r . If K : ℓ 1 r → H K : \ell_1^r \to H K : ℓ 1 r → H is a contraction, then so is
R K : F r → H , R K ( v , b ) = v + K b , R_K : F_r \to H,\qquad R_K(v,b) = v + Kb, R K : F r → H , R K ( v , b ) = v + K b ,
since ∥ v + K b ∥ 2 ≤ ∥ v ∥ 2 + ∥ b ∥ 1 \lVert v + Kb\rVert_2 \le\lVert v\rVert_2 + \lVert b\rVert_1 ∥ v + K b ∥ 2 ≤ ∥ v ∥ 2 + ∥ b ∥ 1 .
Lemma 4.1. For each s ≥ 0 s \ge0 s ≥ 0 there are a finite nonempty index set Γ s \Gamma_s Γ s , contractions J γ : V s → F s J_\gamma: V_s \to F_s J γ : V s → F s , γ ∈ Γ s \gamma\in\Gamma_s γ ∈ Γ s , and a contraction K s : ℓ 1 s → H K_s : \ell_1^s \to H K s : ℓ 1 s → H , independent of γ \gamma γ , such that
∥ x ∥ = max γ ∈ Γ s ∥ J γ x ∥ , J γ h = ( h , 0 ) ( h ∈ H ) , R K s J γ = P s . (6) \lVert x\rVert= \max_{\gamma\in\Gamma_s} \lVert J_\gamma x\rVert,\qquad J_\gamma h = (h,0)\ (h \in H),\qquad R_{K_s}J_\gamma= P_s. \tag*{(6)} ∥ x ∥ = γ ∈ Γ s max ∥ J γ x ∥ , J γ h = ( h , 0 ) ( h ∈ H ) , R K s J γ = P s . (6)
Proof. At s = 0 s = 0 s = 0 take a singleton index set, let K 0 : { 0 } → H K_0 : \{0\} \to H K 0 : { 0 } → H be zero, and define
J γ : H ⟶ H ⊕ 1 { 0 } , J γ h = ( h , 0 ) . J_\gamma: H \longrightarrow H \oplus_1 \{0\},\qquad J_\gamma h = (h,0). J γ : H ⟶ H ⊕ 1 { 0 } , J γ h = ( h , 0 ) .
All assertions in (4.1) then hold.
Suppose they hold for V s V_s V s , and write V s + 1 = G ( V s , p s , μ s ) V_{s+1} = G(V_s,p_s,\mu_s) V s + 1 = G ( V s , p s , μ s ) . Set Γ s + 1 = Γ s × { 1 , … , μ s } \Gamma_{s+1} = \Gamma_s \times\{1,\ldots,\mu_s\} Γ s + 1 = Γ s × { 1 , … , μ s } . Under the natural identification F s + 1 = F s ⊕ 1 R F_{s+1} = F_s \oplus_1 \mathbb{R} F s + 1 = F s ⊕ 1 R , define
J γ , j ′ ( x , a ) = ( J γ ( x − a j p s ) , a j ) , K s + 1 ( b , t ) = K s b + t P s p s . J'_{\gamma,j}(x,a) = \bigl(J_\gamma(x-a_jp_s),a_j\bigr),\qquad K_{s+1}(b,t) = K_sb + tP_sp_s. J γ , j ′ ( x , a ) = ( J γ ( x − a j p s ) , a j ) , K s + 1 ( b , t ) = K s b + t P s p s .
Since ∥ P s p s ∥ ≤ 1 \lVert P_sp_s\rVert\le1 ∥ P s p s ∥ ≤ 1 , one has ∥ K s + 1 ( b , t ) ∥ ≤ ∥ b ∥ 1 + ∣ t ∣ \lVert K_{s+1}(b,t)\rVert\le\lVert b\rVert_1 + |t| ∥ K s + 1 ( b , t )∥ ≤ ∥ b ∥ 1 + ∣ t ∣ . Moreover,
max γ , j ∥ J γ , j ′ ( x , a ) ∥ = max j ( max γ ∥ J γ ( x − a j p s ) ∥ + ∣ a j ∣ ) = max j ( ∥ x − a j p s ∥ + ∣ a j ∣ ) = ∥ ( x , a ) ∥ V s + 1 . \max_{\gamma,j}\lVert J'_{\gamma,j}(x,a)\rVert
= \max_j\left(\max_\gamma\lVert J_\gamma(x-a_jp_s)\rVert+ |a_j|\right)
= \max_j\left(\lVert x-a_jp_s\rVert+ |a_j|\right)
= \lVert(x,a)\rVert_{V_{s+1}}. γ , j max ∥ J γ , j ′ ( x , a )∥ = j max ( γ max ∥ J γ ( x − a j p s )∥ + ∣ a j ∣ ) = j max ( ∥ x − a j p s ∥ + ∣ a j ∣ ) = ∥( x , a ) ∥ V s + 1 .
In particular each J γ , j ′ J'_{\gamma,j} J γ , j ′ is contractive. For h ∈ H h \in H h ∈ H , its new coordinates vanish, so J γ , j ′ h = ( ( h , 0 ) , 0 ) = ( h , 0 ) J'_{\gamma,j}h=((h,0),0)=(h,0) J γ , j ′ h = (( h , 0 ) , 0 ) = ( h , 0 ) . Finally, if J γ ( x − a j p s ) = ( v , b ) J_\gamma(x-a_jp_s)=(v,b) J γ ( x − a j p s ) = ( v , b ) , then
R K s + 1 J γ , j ′ ( x , a ) = v + K s b + a j P s p s = P s ( x − a j p s ) + a j P s p s = P s x = P s + 1 ( x , a ) . R_{K_{s+1}}J'_{\gamma,j}(x,a)
= v + K_sb + a_jP_sp_s
= P_s(x-a_jp_s) + a_jP_sp_s
= P_sx = P_{s+1}(x,a). R K s + 1 J γ , j ′ ( x , a ) = v + K s b + a j P s p s = P s ( x − a j p s ) + a j P s p s = P s x = P s + 1 ( x , a ) .
This proves the induction. Notice that K s + 1 K_{s+1} K s + 1 does not depend on either γ \gamma γ or j j j . □
Lemma 4.2. Suppose that J γ : E → U γ J_\gamma: E \to U_\gamma J γ : E → U γ and L η : F → W η L_\eta: F \to W_\eta L η : F → W η are finite families of linear maps satisfying
∥ x ∥ = max γ ∥ J γ x ∥ , ∥ y ∥ = max η ∥ L η y ∥ . \lVert x\rVert= \max_\gamma\lVert J_\gamma x\rVert,\qquad\lVert y\rVert= \max_\eta\lVert L_\eta y\rVert. ∥ x ∥ = γ max ∥ J γ x ∥ , ∥ y ∥ = η max ∥ L η y ∥ . Then, for every u ∈ E ⊗ F u \in E \otimes F u ∈ E ⊗ F ,
∥ u ∥ ε = max γ , η ∥ ( J γ ⊗ L η ) u ∥ ε . (7) \lVert u\rVert_\varepsilon= \max_{\gamma,\eta}\lVert(J_\gamma\otimes L_\eta)u\rVert_\varepsilon. \tag*{(7)} ∥ u ∥ ε = γ , η max ∥( J γ ⊗ L η ) u ∥ ε . (7)
Proof. Let U = ( ⨁ γ U γ ) ∞ \mathcal{U} = \left(\bigoplus_\gamma U_\gamma\right)_\infty U = ( ⨁ γ U γ ) ∞ and W = ( ⨁ η W η ) ∞ \mathcal{W} = \left(\bigoplus_\eta W_\eta\right)_\infty W = ( ⨁ η W η ) ∞ . The maps J x = ( J γ x ) γ Jx=(J_\gamma x)_\gamma J x = ( J γ x ) γ and L y = ( L η y ) η Ly=(L_\eta y)_\eta L y = ( L η y ) η are linear isometries.
Let u = ∑ i = 1 n x i ⊗ y i ∈ E ⊗ F u=\sum_{i=1}^n x_i\otimes y_i\in E\otimes F u = ∑ i = 1 n x i ⊗ y i ∈ E ⊗ F . By the definition of the injective tensor norm,
∥ ( J ⊗ L ) u ∥ ε = sup φ ∈ B U ∗ ψ ∈ B W ∗ ∣ ∑ i = 1 n φ ( J x i ) ψ ( L y i ) ∣ . \lVert(J\otimes L)u\rVert_\varepsilon
= \sup_{\substack{\varphi\in B_{\mathcal{U}^*}\\ \psi\in B_{\mathcal{W}^*}}}
\left|\sum_{i=1}^n \varphi(Jx_i)\psi(Ly_i)\right|. ∥( J ⊗ L ) u ∥ ε = φ ∈ B U ∗ ψ ∈ B W ∗ sup i = 1 ∑ n φ ( J x i ) ψ ( L y i ) .
Since J J J and L L L are isometries, the Hahn–Banach theorem gives
J ∗ ( B U ∗ ) = B E ∗ , L ∗ ( B W ∗ ) = B F ∗ . J^*(B_{\mathcal{U}^*})=B_{E^*},\qquad L^*(B_{\mathcal{W}^*})=B_{F^*}. J ∗ ( B U ∗ ) = B E ∗ , L ∗ ( B W ∗ ) = B F ∗ .
Indeed, every f ∈ B E ∗ f\in B_{E^*} f ∈ B E ∗ induces a functional on J ( E ) J(E) J ( E ) of the same norm by J x ↦ f ( x ) Jx\mapsto f(x) J x ↦ f ( x ) , and this functional extends to U \mathcal{U} U without increasing its norm. The same argument applies to L . g o z L. goz L . g oz Consequently,
∥ ( J ⊗ L ) u ∥ ε = sup f ∈ B E ∗ g ∈ B F ∗ ∣ ∑ i = 1 n f ( x i ) g ( y i ) ∣ = ∥ u ∥ ε . \begin{aligned}
\|(J \otimes L)u\|_{\varepsilon}=\sup_{\substack{f\in B_{E^*}\\g\in B_{F^*}}}\left|\sum_{i=1}^{n}f(x_i)g(y_i)\right| \\
&=\|u\|_{\varepsilon}.
\end{aligned} ∥ ( J ⊗ L ) u ∥ ε = f ∈ B E ∗ g ∈ B F ∗ sup i = 1 ∑ n f ( x i ) g ( y i ) = ∥ u ∥ ε .
It remains to compute the tensor norm in the two finite sums. Their duals are the corresponding ℓ 1 \ell_1 ℓ 1 -sums. If v = ∑ i = 1 n x i ⊗ y i ∈ U ⊗ W v=\sum_{i=1}^{n}x_i\otimes y_i\in U\otimes W v = ∑ i = 1 n x i ⊗ y i ∈ U ⊗ W has components v γ , η = ∑ i = 1 n x i , γ ⊗ y i , η v_{\gamma,\eta}=\sum_{i=1}^{n}x_{i,\gamma}\otimes y_{i,\eta} v γ , η = ∑ i = 1 n x i , γ ⊗ y i , η for x i = ( x i , γ ) γ x_i=(x_{i,\gamma})_\gamma x i = ( x i , γ ) γ and y i = ( y i , η ) η y_i=(y_{i,\eta})_\eta y i = ( y i , η ) η , then functionals a ∗ = ( a γ ∗ ) a^*=(a^*_\gamma) a ∗ = ( a γ ∗ ) and b ∗ = ( b η ∗ ) b^*=(b^*_\eta) b ∗ = ( b η ∗ ) with ∑ γ ∥ a γ ∗ ∥ ≤ 1 \sum_\gamma\|a^*_\gamma\|\leq1 ∑ γ ∥ a γ ∗ ∥ ≤ 1 and ∑ η ∥ b η ∗ ∥ ≤ 1 \sum_\eta\|b^*_\eta\|\leq1 ∑ η ∥ b η ∗ ∥ ≤ 1 satisfy
( a ∗ ⊗ b ∗ ) ( v ) = ∑ i = 1 n a ∗ ( x i ) b ∗ ( y i ) = ∑ i = 1 n ( ∑ γ a γ ∗ ( x i , γ ) ) ( ∑ η b η ∗ ( y i , η ) ) = ∑ γ , η ( a γ ∗ ⊗ b η ∗ ) ( v γ , η ) . \begin{aligned}
(a^*\otimes b^*)(v)=\sum_{i=1}^{n}a^*(x_i)b^*(y_i) \\
&=\sum_{i=1}^{n}\left(\sum_\gamma a^*_\gamma(x_{i,\gamma})\right)\left(\sum_\eta b^*_\eta(y_{i,\eta})\right) \\
&=\sum_{\gamma,\eta}(a^*_\gamma\otimes b^*_\eta)(v_{\gamma,\eta}).
\end{aligned} ( a ∗ ⊗ b ∗ ) ( v ) = i = 1 ∑ n a ∗ ( x i ) b ∗ ( y i ) = i = 1 ∑ n ( γ ∑ a γ ∗ ( x i , γ ) ) ( η ∑ b η ∗ ( y i , η ) ) = γ , η ∑ ( a γ ∗ ⊗ b η ∗ ) ( v γ , η ) .
By the triangle inequality and the definition of the injective tensor norm,
∣ ( a ∗ ⊗ b ∗ ) ( v ) ∣ ≤ ∑ γ , η ∣ ( a γ ∗ ⊗ b η ∗ ) ( v γ , η ) ∣ ≤ ∑ γ , η ∥ a γ ∗ ∥ ∥ b η ∗ ∥ ∥ v γ , η ∥ ε ≤ ( max γ , η ∥ v γ , η ∥ ε ) ( ∑ γ ∥ a γ ∗ ∥ ) ( ∑ η ∥ b η ∗ ∥ ) ≤ max γ , η ∥ v γ , η ∥ ε . \begin{aligned}
\left|(a^*\otimes b^*)(v)\right|\leq\sum_{\gamma,\eta}\left|(a^*_\gamma\otimes b^*_\eta)(v_{\gamma,\eta})\right| \\
&\leq\sum_{\gamma,\eta}\|a^*_\gamma\|\|b^*_\eta\|\|v_{\gamma,\eta}\|_\varepsilon\\
&\leq\left(\max_{\gamma,\eta}\|v_{\gamma,\eta}\|_\varepsilon\right)\left(\sum_\gamma\|a^*_\gamma\|\right)\left(\sum_\eta\|b^*_\eta\|\right) \\
&\leq\max_{\gamma,\eta}\|v_{\gamma,\eta}\|_\varepsilon.
\end{aligned} ∣ ( a ∗ ⊗ b ∗ ) ( v ) ∣ ≤ γ , η ∑ ( a γ ∗ ⊗ b η ∗ ) ( v γ , η ) ≤ γ , η ∑ ∥ a γ ∗ ∥∥ b η ∗ ∥∥ v γ , η ∥ ε ≤ ( γ , η max ∥ v γ , η ∥ ε ) ( γ ∑ ∥ a γ ∗ ∥ ) ( η ∑ ∥ b η ∗ ∥ ) ≤ γ , η max ∥ v γ , η ∥ ε .
Taking the supremum over all a ∗ ∈ B U ∗ a^*\in B_{U^*} a ∗ ∈ B U ∗ and b ∗ ∈ B W ∗ b^*\in B_{W^*} b ∗ ∈ B W ∗ gives
∥ v ∥ ε ≤ max γ , η ∥ v γ , η ∥ ε . \|v\|_\varepsilon\leq\max_{\gamma,\eta}\|v_{\gamma,\eta}\|_\varepsilon. ∥ v ∥ ε ≤ γ , η max ∥ v γ , η ∥ ε .
For the reverse inequality, fix one pair ( γ , η ) (\gamma,\eta) ( γ , η ) and choose functionals supported only in these coordinates, with evaluations arbitrarily close to ∥ v γ , η ∥ ε \|v_{\gamma,\eta}\|_\varepsilon ∥ v γ , η ∥ ε . Thus ∥ v ∥ ε = max γ , η ∥ v γ , η ∥ ε \|v\|_\varepsilon=\max_{\gamma,\eta}\|v_{\gamma,\eta}\|_\varepsilon ∥ v ∥ ε = max γ , η ∥ v γ , η ∥ ε . Apply this identity to v = ( J ⊗ L ) u v=(J\otimes L)u v = ( J ⊗ L ) u . □ \square □
A uniform operator estimate
Lemma 5.1. Let U = H ⊕ 1 ℓ 1 k U=H\oplus_1\ell^k_1 U = H ⊕ 1 ℓ 1 k and V = H ⊕ 1 ℓ 1 l V=H\oplus_1\ell^l_1 V = H ⊕ 1 ℓ 1 l , and let K : ℓ 1 k → H K:\ell^k_1\to H K : ℓ 1 k → H and L : ℓ 1 l → H L:\ell^l_1\to H L : ℓ 1 l → H be contractions. Suppose that W : U ∗ → V W:U^*\to V W : U ∗ → V is linear and ∥ W ∥ ≤ 1 \|W\|\leq1 ∥ W ∥ ≤ 1 . Put
M = R L W R K ∗ : H → H , δ = 1 − 1 2 tr M , I 0 ( h , t ) = ( h , 0 ) ( ( h , t ) ∈ U ∗ ) . M=R_LWR_K^*:H\to H,\qquad\delta=1-\frac{1}{2}\operatorname{tr}M,\qquad I_0(h,t)=(h,0)\quad((h,t)\in U^*). M = R L W R K ∗ : H → H , δ = 1 − 2 1 tr M , I 0 ( h , t ) = ( h , 0 ) (( h , t ) ∈ U ∗ ) . Then 0 ≤ δ ≤ 2 0\leq\delta\leq2 0 ≤ δ ≤ 2 and
∥ W − I 0 ∥ ≤ 2 δ + 122 δ . (8) \|W-I_0\|\leq2\sqrt{\delta}+122\delta. \tag*{(8)} ∥ W − I 0 ∥ ≤ 2 δ + 122 δ . (8)
Proof. Using U ∗ = H ⊕ ∞ ℓ ∞ k U^*=H\oplus_\infty\ell^k_\infty U ∗ = H ⊕ ∞ ℓ ∞ k , write
W ( h , t ) = ( A h + B t , C h + D t ) , W(h,t)=(Ah+Bt,Ch+Dt), W ( h , t ) = ( A h + B t , C h + D t ) ,
where the four bounded linear operators are
A : H → H , B : ℓ ∞ k → H , C : H → ℓ 1 l , D : ℓ ∞ k → ℓ 1 l . A:H\to H,\qquad B:\ell^k_\infty\to H,\qquad C:H\to\ell^l_1,\qquad D:\ell^k_\infty\to\ell^l_1. A : H → H , B : ℓ ∞ k → H , C : H → ℓ 1 l , D : ℓ ∞ k → ℓ 1 l .
The contraction assumption means that
∥ A h + B t ∥ 2 + ∥ C h + D t ∥ 1 ≤ 1 ( ∥ h ∥ 2 ≤ 1 , ∥ t ∥ ∞ ≤ 1 ) . (9) \lVert Ah + Bt\rVert_2 + \lVert Ch + Dt\rVert_1 \le1 \quad(\lVert h\rVert_2 \le1,\ \lVert t\rVert_\infty\le1). \tag*{(9)} ∥ A h + B t ∥ 2 + ∥ C h + D t ∥ 1 ≤ 1 (∥ h ∥ 2 ≤ 1 , ∥ t ∥ ∞ ≤ 1 ) . (9)
Since R K R_K R K and R L R_L R L are contractions, ∥ M ∥ ≤ 1 \lVert M\rVert\le1 ∥ M ∥ ≤ 1 . In particular, − 2 ≤ tr M ≤ 2 -2 \le\operatorname{tr} M \le2 − 2 ≤ tr M ≤ 2 , proving the assertion about δ \delta δ .
Set
A 0 = A + B K ∗ , N = C + D K ∗ . (10) A_0 = A + BK^*, \qquad N = C + DK^*. \tag*{(10)} A 0 = A + B K ∗ , N = C + D K ∗ . (10)
Then R K ∗ h = ( h , K ∗ h ) R_K^*h = (h,K^*h) R K ∗ h = ( h , K ∗ h ) and
M = A 0 + L N , ∥ A 0 h ∥ 2 + ∥ N h ∥ 1 ≤ 1 ( h ∈ S H ) . (11) M = A_0 + LN,\qquad\lVert A_0h\rVert_2 + \lVert Nh\rVert_1 \le1 \quad(h \in S_H). \tag*{(11)} M = A 0 + L N , ∥ A 0 h ∥ 2 + ∥ N h ∥ 1 ≤ 1 ( h ∈ S H ) . (11)
Indeed, for h ∈ H h \in H h ∈ H :
M h = R L W R K ∗ h = R L W ( h , K ∗ h ) = R L ( A h + B K ∗ h , C h + D K ∗ h ) = A h + B K ∗ h + L ( C h + D K ∗ h ) = ( A + B K ∗ ) h + L ( C + D K ∗ ) h = A 0 h + L N h . \begin{aligned}
Mh &= R_LWR_K^*h = R_LW(h,K^*h)\\
&= R_L(Ah+BK^*h, Ch+DK^*h)\\
&= Ah+BK^*h+L(Ch+DK^*h)\\
&= (A+BK^*)h+L(C+DK^*)h\\
&= A_0h+LNh.
\end{aligned} M h = R L W R K ∗ h = R L W ( h , K ∗ h ) = R L ( A h + B K ∗ h , C h + D K ∗ h ) = A h + B K ∗ h + L ( C h + D K ∗ h ) = ( A + B K ∗ ) h + L ( C + D K ∗ ) h = A 0 h + L N h .
Moreover, if h ∈ S H h \in S_H h ∈ S H , then
∥ K ∗ h ∥ ∞ ≤ ∥ K ∗ ∥ ∥ h ∥ 2 ≤ 1. \lVert K^*h\rVert_\infty\le\lVert K^*\rVert\lVert h\rVert_2 \le1. ∥ K ∗ h ∥ ∞ ≤ ∥ K ∗ ∥ ∥ h ∥ 2 ≤ 1.
Thus ( h , K ∗ h ) (h,K^*h) ( h , K ∗ h ) belongs to the unit ball of U ∗ = H ⊕ ∞ ℓ ∞ k U^* = H \oplus_\infty\ell_\infty^k U ∗ = H ⊕ ∞ ℓ ∞ k . Applying the contraction estimate (5.2) with t = K ∗ h t=K^*h t = K ∗ h , we obtain
1 ≥ ∥ A h + B K ∗ h ∥ 2 + ∥ C h + D K ∗ h ∥ 1 = ∥ A 0 h ∥ 2 + ∥ N h ∥ 1 . \begin{aligned}
1 &\ge\lVert Ah+BK^*h\rVert_2+\lVert Ch+DK^*h\rVert_1\\
&= \lVert A_0h\rVert_2+\lVert Nh\rVert_1.
\end{aligned} 1 ≥ ∥ A h + B K ∗ h ∥ 2 + ∥ C h + D K ∗ h ∥ 1 = ∥ A 0 h ∥ 2 + ∥ N h ∥ 1 .
Let E \mathbb{E} E denote integration with respect to normalized arc length on S H S_H S H . Parametrizing h = ( cos θ , sin θ ) h=(\cos\theta,\sin\theta) h = ( cos θ , sin θ ) gives
E ( h h T ) = 1 2 I d H , E ∣ ⟨ v , h ⟩ ∣ = 2 π ∥ v ∥ 2 ( v ∈ H ) . (12) \mathbb{E}(hh^{\mathsf T})=\frac{1}{2}I_d{}_H,\qquad\mathbb{E}|\langle v,h\rangle|=\frac{2}{\pi}\lVert v\rVert_2\quad(v\in H). \tag*{(12)} E ( h h T ) = 2 1 I d H , E ∣ ⟨ v , h ⟩ ∣ = π 2 ∥ v ∥ 2 ( v ∈ H ) . (12)
To verify these identities, write
h ( t ) = cos t e 1 + sin t e 2 , 0 ≤ t ≤ 2 π . h(t)=\cos t\,e_1+\sin t\,e_2,\qquad0\le t\le2\pi. h ( t ) = cos t e 1 + sin t e 2 , 0 ≤ t ≤ 2 π .
Recall that integration with respect to the normalized uniform measure on S H S_H S H is given by
E f ( h ) = 1 2 π ∫ 0 2 π f ( h ( t ) ) d t . \mathbb{E}f(h)=\frac{1}{2\pi}\int_0^{2\pi}f(h(t))\,dt. E f ( h ) = 2 π 1 ∫ 0 2 π f ( h ( t )) d t .
For the first identity, we have
h ( t ) h ( t ) T = cos 2 t e 1 e 1 T + sin 2 t e 2 e 2 T + cos t sin t ( e 1 e 2 T + e 2 e 1 T ) . h(t)h(t)^{\mathsf T}=\cos^2t\,e_1e_1^{\mathsf T}+\sin^2t\,e_2e_2^{\mathsf T}+\cos t\sin t\left(e_1e_2^{\mathsf T}+e_2e_1^{\mathsf T}\right). h ( t ) h ( t ) T = cos 2 t e 1 e 1 T + sin 2 t e 2 e 2 T + cos t sin t ( e 1 e 2 T + e 2 e 1 T ) .
Since
∫ 0 2 π cos 2 t d t = ∫ 0 2 π sin 2 t d t = π , \int_0^{2\pi}\cos^2t\,dt=\int_0^{2\pi}\sin^2t\,dt=\pi, ∫ 0 2 π cos 2 t d t = ∫ 0 2 π sin 2 t d t = π ,
and
∫ 0 2 π cos t sin t d t = 0 , \int_0^{2\pi}\cos t\sin t\,dt=0, ∫ 0 2 π cos t sin t d t = 0 ,
it follows that
E ( h h T ) = 1 2 π ∫ 0 2 π h ( t ) h ( t ) T d t = 1 2 ( e 1 e 1 T + e 2 e 2 T ) = 1 2 I d H . \begin{aligned}
\mathbb{E}(hh^{\mathsf T})&=\frac{1}{2\pi}\int_0^{2\pi}h(t)h(t)^{\mathsf T}\,dt\\
&=\frac{1}{2}\left(e_1e_1^{\mathsf T}+e_2e_2^{\mathsf T}\right)\\
&=\frac{1}{2}I_d{}_H.
\end{aligned} E ( h h T ) = 2 π 1 ∫ 0 2 π h ( t ) h ( t ) T d t = 2 1 ( e 1 e 1 T + e 2 e 2 T ) = 2 1 I d H .
For the second identity, the case v = 0 v=0 v = 0 is immediate. If v ≠ 0 v\ne0 v = 0 , choose θ ∈ R \theta\in\mathbb{R} θ ∈ R such that v = ∥ v ∥ 2 ( cos θ e 1 + sin θ e 2 ) v=\lVert v\rVert_2(\cos\theta\,e_1+\sin\theta\,e_2) v = ∥ v ∥ 2 ( cos θ e 1 + sin θ e 2 ) .
Then
E ∣ ⟨ v , h ⟩ ∣ = 1 2 π ∫ 0 2 π ∣ ⟨ v , h ( t ) ⟩ ∣ d t = ∥ v ∥ 2 2 π ∫ 0 2 π ∣ cos θ cos t + sin θ sin t ∣ d t = ∥ v ∥ 2 2 π ∫ 0 2 π ∣ cos ( t − θ ) ∣ d t = ∥ v ∥ 2 2 π ∫ 0 2 π ∣ cos t ∣ d t = 4 2 π ∥ v ∥ 2 = 2 π ∥ v ∥ 2 . \mathbb{E}\lvert\langle v,h\rangle\rvert=\frac{1}{2\pi}\int_{0}^{2\pi}\lvert\langle v,h(t)\rangle\rvert\,dt
=\frac{\lVert v\rVert_2}{2\pi}\int_{0}^{2\pi}\lvert\cos\theta\cos t+\sin\theta\sin t\rvert\,dt
=\frac{\lVert v\rVert_2}{2\pi}\int_{0}^{2\pi}\lvert\cos(t-\theta)\rvert\,dt
=\frac{\lVert v\rVert_2}{2\pi}\int_{0}^{2\pi}\lvert\cos t\rvert\,dt
=\frac{4}{2\pi}\lVert v\rVert_2
=\frac{2}{\pi}\lVert v\rVert_2. E ∣⟨ v , h ⟩∣ = 2 π 1 ∫ 0 2 π ∣⟨ v , h ( t )⟩∣ d t = 2 π ∥ v ∥ 2 ∫ 0 2 π ∣ cos θ cos t + sin θ sin t ∣ d t = 2 π ∥ v ∥ 2 ∫ 0 2 π ∣ cos ( t − θ )∣ d t = 2 π ∥ v ∥ 2 ∫ 0 2 π ∣ cos t ∣ d t = 2 π 4 ∥ v ∥ 2 = π 2 ∥ v ∥ 2 .
Here we used the periodicity of ∣ cos t ∣ \lvert\cos t\rvert ∣ cos t ∣ and the identity
∫ 0 2 π ∣ cos t ∣ d t = 4 ∫ 0 π / 2 cos t d t = 4. \int_{0}^{2\pi}\lvert\cos t\rvert\,dt=4\int_{0}^{\pi/2}\cos t\,dt=4. ∫ 0 2 π ∣ cos t ∣ d t = 4 ∫ 0 π /2 cos t d t = 4.
Write ( N h ) r = ⟨ n r , h ⟩ (Nh)_r=\langle n_r,h\rangle ( N h ) r = ⟨ n r , h ⟩ for n r ∈ H n_r\in H n r ∈ H , put l r = L u r l_r=Lu_r l r = L u r , and set S = ∑ r = 1 l ∥ n r ∥ 2 S=\sum_{r=1}^{l}\lVert n_r\rVert_2 S = ∑ r = 1 l ∥ n r ∥ 2 . The sum is finite, so
E ∥ N h ∥ 1 = ∑ r = 1 l E ∣ ⟨ n r , h ⟩ ∣ = 2 π ∑ r = 1 l ∥ n r ∥ 2 = 2 π S . \mathbb{E}\lVert Nh\rVert_1=\sum_{r=1}^{l}\mathbb{E}\lvert\langle n_r,h\rangle\rvert=\frac{2}{\pi}\sum_{r=1}^{l}\lVert n_r\rVert_2=\frac{2}{\pi}S. E ∥ N h ∥ 1 = r = 1 ∑ l E ∣⟨ n r , h ⟩∣ = π 2 r = 1 ∑ l ∥ n r ∥ 2 = π 2 S .
Integrating (11) therefore yields
E ∥ A 0 h ∥ 2 + 2 π S ≤ 1. (13) \mathbb{E}\lVert A_0h\rVert_2+\frac{2}{\pi}S\le1. \tag*{(13)} E ∥ A 0 h ∥ 2 + π 2 S ≤ 1. (13)
On the other hand, the left equation in (12) implies E ⟨ h , M h ⟩ = 1 2 tr M \mathbb{E}\langle h,Mh\rangle=\frac{1}{2}\operatorname{tr}M E ⟨ h , M h ⟩ = 2 1 tr M . Indeed,
⟨ h , M h ⟩ = tr ( M h h T ) \langle h,Mh\rangle=\operatorname{tr}(Mhh^{\mathsf T}) ⟨ h , M h ⟩ = tr ( M h h T )
is elementary, and integration commutes with both M M M and the trace. Since L N h = ∑ r = 1 l ⟨ n r , h ⟩ l r LNh=\sum_{r=1}^{l}\langle n_r,h\rangle l_r L N h = ∑ r = 1 l ⟨ n r , h ⟩ l r , we obtain together with (11)
1 − δ = 1 2 tr M = E ⟨ h , M h ⟩ = E ⟨ h , A 0 h ⟩ + ∑ r = 1 l E ( ⟨ n r , h ⟩ ⟨ h , l r ⟩ ) = E ⟨ h , A 0 h ⟩ + 1 2 ∑ r = 1 l ⟨ n r , l r ⟩ ≤ E ∥ A 0 h ∥ 2 + 1 2 ∑ r = 1 l ∥ n r ∥ 2 ∥ l r ∥ 2 ≤ E ∥ A 0 h ∥ 2 + 1 2 S . (14) \begin{aligned}
1-\delta=\frac{1}{2}\operatorname{tr}M=\mathbb{E}\langle h,Mh\rangle=\mathbb{E}\langle h,A_0h\rangle+\sum_{r=1}^{l}\mathbb{E}(\langle n_r,h\rangle\langle h,l_r\rangle) \\
&=\mathbb{E}\langle h,A_0h\rangle+\frac{1}{2}\sum_{r=1}^{l}\langle n_r,l_r\rangle\le\mathbb{E}\lVert A_0h\rVert_2+\frac{1}{2}\sum_{r=1}^{l}\lVert n_r\rVert_2\lVert l_r\rVert_2 \\
&\le\mathbb{E}\lVert A_0h\rVert_2+\frac{1}{2}S. \tag*{(14)}
\end{aligned} 1 − δ = 2 1 tr M = E ⟨ h , M h ⟩ = E ⟨ h , A 0 h ⟩ + r = 1 ∑ l E (⟨ n r , h ⟩ ⟨ h , l r ⟩) = E ⟨ h , A 0 h ⟩ + 2 1 r = 1 ∑ l ⟨ n r , l r ⟩ ≤ E ∥ A 0 h ∥ 2 + 2 1 r = 1 ∑ l ∥ n r ∥ 2 ∥ l r ∥ 2 ≤ E ∥ A 0 h ∥ 2 + 2 1 S . (14)
Here ∥ h ∥ 2 = 1 \lVert h\rVert_2=1 ∥ h ∥ 2 = 1 and ∥ l r ∥ 2 = ∥ L u r ∥ 2 ≤ 1 \lVert l_r\rVert_2=\lVert Lu_r\rVert_2\le1 ∥ l r ∥ 2 = ∥ L u r ∥ 2 ≤ 1 . Combining (13) and (14) , we get
1 − δ ≤ 1 − ( 2 π − 1 2 ) S . 1-\delta\le1-\left(\frac{2}{\pi}-\frac{1}{2}\right)S. 1 − δ ≤ 1 − ( π 2 − 2 1 ) S .
As 2 / π − 1 / 2 > 1 / 8 2/\pi-1/2>1/8 2/ π − 1/2 > 1/8 , it follows that S ≤ 8 δ S\le8\delta S ≤ 8 δ . Also (11) gives ∥ A 0 h ∥ 2 ≤ 1 \lVert A_0h\rVert_2\le1 ∥ A 0 h ∥ 2 ≤ 1 , whereas (14) gives E ∥ A 0 h ∥ 2 ≥ 1 − δ − S / 2 \mathbb{E}\lVert A_0h\rVert_2\ge1-\delta-S/2 E ∥ A 0 h ∥ 2 ≥ 1 − δ − S /2 . Consequently
S ≤ 8 δ , 0 ≤ 1 − E ∥ A 0 h ∥ 2 ≤ δ + 1 2 S ≤ 5 δ . (15) S\le8\delta,\qquad0\le1-\mathbb{E}\lVert A_0h\rVert_2\le\delta+\frac{1}{2}S\le5\delta. \tag*{(15)} S ≤ 8 δ , 0 ≤ 1 − E ∥ A 0 h ∥ 2 ≤ δ + 2 1 S ≤ 5 δ . (15)
In addition, ∥ N ∥ 2 → 1 ≤ S \lVert N\rVert_{2\to1}\le S ∥ N ∥ 2 → 1 ≤ S , because ∑ r ∣ ⟨ n r , h ⟩ ∣ ≤ S ∥ h ∥ 2 \sum_r\lvert\langle n_r,h\rangle\rvert\le S\lVert h\rVert_2 ∑ r ∣⟨ n r , h ⟩∣ ≤ S ∥ h ∥ 2 .
We next estimate D D D . Assume first that k , l > 0 k,l>0 k , l > 0 . Denote the columns of K K K by k i = K u i k_i=Ku_i k i = K u i , and put
a = 1 − 2 π > 1 3 , d i = E ( 1 − ∣ ⟨ k i , h ⟩ ∣ ) = 1 − 2 π ∥ k i ∥ 2 ≥ a . a=1-\frac{2}{\pi}>\frac{1}{3},\qquad d_i=\mathbb{E}(1-\lvert\langle k_i,h\rangle\rvert)=1-\frac{2}{\pi}\lVert k_i\rVert_2\geq a. a = 1 − π 2 > 3 1 , d i = E ( 1 − ∣⟨ k i , h ⟩∣) = 1 − π 2 ∥ k i ∥ 2 ≥ a .
For a fixed t ∈ B ℓ ∞ k t\in B_{\ell_\infty}^k t ∈ B ℓ ∞ k and h ∈ S H h\in S_H h ∈ S H , define
( t h ) i = a d i t i ( 1 − ∣ ⟨ k i , h ⟩ ∣ ) . (t_h)_i=\frac{a}{d_i}t_i(1-\lvert\langle k_i,h\rangle\rvert). ( t h ) i = d i a t i ( 1 − ∣⟨ k i , h ⟩∣) .
For each coordinate,
∣ ( K ∗ h ) i ± ( t h ) i ∣ ≤ ∣ ⟨ k i , h ⟩ ∣ + a d i ∣ t i ∣ ( 1 − ∣ ⟨ k i , h ⟩ ∣ ) ≤ 1. \lvert(K^*h)_i\pm(t_h)_i\rvert\leq\lvert\langle k_i,h\rangle\rvert+\frac{a}{d_i}\lvert t_i\rvert(1-\lvert\langle k_i,h\rangle\rvert)\leq1. ∣( K ∗ h ) i ± ( t h ) i ∣ ≤ ∣⟨ k i , h ⟩∣ + d i a ∣ t i ∣ ( 1 − ∣⟨ k i , h ⟩∣) ≤ 1.
Thus K ∗ h ± t h ∈ B ℓ ∞ k K^*h\pm t_h\in B_{\ell_\infty}^k K ∗ h ± t h ∈ B ℓ ∞ k . The definition of d i d_i d i also gives E t h = a t \mathbb{E}t_h=at E t h = a t coordinate by coordinate. Applying (9) with K ∗ h + t h K^*h+t_h K ∗ h + t h and K ∗ h − t h K^*h-t_h K ∗ h − t h gives
∥ A 0 h + B t h ∥ 2 + ∥ N h + D t h ∥ 1 ≤ 1 , ∥ A 0 h − B t h ∥ 2 + ∥ N h − D t h ∥ 1 ≤ 1. \lVert A_0h+Bt_h\rVert_2+\lVert Nh+Dt_h\rVert_1\leq1,\qquad
\lVert A_0h-Bt_h\rVert_2+\lVert Nh-Dt_h\rVert_1\leq1. ∥ A 0 h + B t h ∥ 2 + ∥ N h + D t h ∥ 1 ≤ 1 , ∥ A 0 h − B t h ∥ 2 + ∥ N h − D t h ∥ 1 ≤ 1.
The sum of the two Hilbert-space terms is at least 2 ∥ A 0 h ∥ 2 2\lVert A_0h\rVert_2 2 ∥ A 0 h ∥ 2 , and ∥ N h + D t h ∥ 1 + ∥ N h − D t h ∥ 1 ≥ 2 ∥ D t h ∥ 1 \lVert Nh+Dt_h\rVert_1+\lVert Nh-Dt_h\rVert_1\geq2\lVert Dt_h\rVert_1 ∥ N h + D t h ∥ 1 + ∥ N h − D t h ∥ 1 ≥ 2 ∥ D t h ∥ 1 . Adding the inequalities and dividing by two therefore yields
∥ A 0 h ∥ 2 + ∥ D t h ∥ 1 ≤ 1. \lVert A_0h\rVert_2+\lVert Dt_h\rVert_1\leq1. ∥ A 0 h ∥ 2 + ∥ D t h ∥ 1 ≤ 1.
By linearity, the triangle inequality for the integral, and (15) ,
a ∥ D t ∥ 1 = ∥ D ( E t h ) ∥ 1 = ∥ E ( D t h ) ∥ 1 ≤ E ∥ D t h ∥ 1 ≤ 1 − E ∥ A 0 h ∥ 2 ≤ 5 δ . a\lVert Dt\rVert_1=\lVert D(\mathbb{E}t_h)\rVert_1=\lVert\mathbb{E}(Dt_h)\rVert_1\leq\mathbb{E}\lVert Dt_h\rVert_1\leq1-\mathbb{E}\lVert A_0h\rVert_2\leq5\delta. a ∥ D t ∥ 1 = ∥ D ( E t h ) ∥ 1 = ∥ E ( D t h ) ∥ 1 ≤ E ∥ D t h ∥ 1 ≤ 1 − E ∥ A 0 h ∥ 2 ≤ 5 δ .
Taking the supremum over t ∈ B ℓ ∞ k t\in B_{\ell_\infty}^k t ∈ B ℓ ∞ k gives ∥ D ∥ ∞ → 1 ≤ 15 δ \lVert D\rVert_{\infty\to1}\leq15\delta ∥ D ∥ ∞ → 1 ≤ 15 δ . If k = 0 k=0 k = 0 or l = 0 l=0 l = 0 , the same bound holds because D = 0 D=0 D = 0 . Since C = N − D K ∗ C=N-DK^* C = N − D K ∗ and ∥ K ∗ ∥ ≤ 1 \lVert K^*\rVert\leq1 ∥ K ∗ ∥ ≤ 1 , we have proved
∥ D ∥ ∞ → 1 ≤ 15 δ , ∥ C ∥ 2 → 1 ≤ ∥ N ∥ 2 → 1 + ∥ D ∥ ∞ → 1 ≤ 23 δ . (16) \lVert D\rVert_{\infty\to1}\leq15\delta,\qquad
\lVert C\rVert_{2\to1}\leq\lVert N\rVert_{2\to1}+\lVert D\rVert_{\infty\to1}\leq23\delta. \tag*{(16)} ∥ D ∥ ∞ → 1 ≤ 15 δ , ∥ C ∥ 2 → 1 ≤ ∥ N ∥ 2 → 1 + ∥ D ∥ ∞ → 1 ≤ 23 δ . (16)
All spaces in this lemma are finite-dimensional. The adjoint operator W ∗ : V ∗ → U W^*:V^*\to U W ∗ : V ∗ → U is consequently a contraction of the same form, with block matrix
W ∗ = ( A ∗ C ∗ B ∗ D ∗ ) . W^*=
\begin{pmatrix}
A^* & C^*\\
B^* & D^*
\end{pmatrix}. W ∗ = ( A ∗ B ∗ C ∗ D ∗ ) .
Interchanging K K K and L L L , its compressed operator is R K W ∗ R L ∗ = M ∗ R_KW^*R_L^*=M^* R K W ∗ R L ∗ = M ∗ , whose trace equals tr M \operatorname{tr}M tr M . Apply the bound for the lower left block in (16) to this operator. Since that block is B ∗ : H → ℓ 1 k B^*:H\to\ell_1^k B ∗ : H → ℓ 1 k , we obtain
∥ B ∥ ∞ → 2 = ∥ B ∗ ∥ 2 → 1 ≤ 23 δ . (17) \lVert B\rVert_{\infty\to2}=\lVert B^*\rVert_{2\to1}\leq23\delta. \tag*{(17)} ∥ B ∥ ∞ → 2 = ∥ B ∗ ∥ 2 → 1 ≤ 23 δ . (17)
Let ∥ ⋅ ∥ H S \lVert\cdot\rVert_{\mathrm{HS}} ∥ ⋅ ∥ HS be the Hilbert–Schmidt norm on operators on H H H . Since ∥ M ∥ ≤ 1 \lVert M\rVert\leq1 ∥ M ∥ ≤ 1 and dim H = 2 \dim H=2 dim H = 2 ,
∥ M − Id H ∥ 2 ≤ ∥ M − Id H ∥ H S 2 = ∑ i = 1 2 ∥ ( M − Id H ) e i ∥ 2 2 = ∑ i = 1 2 ∥ M e i − e i ∥ 2 2 = ∑ i = 1 2 ( ∥ M e i ∥ 2 2 + ∥ e i ∥ 2 2 − 2 ⟨ M e i , e i ⟩ ) ≤ ∑ i = 1 2 ( 2 − 2 ⟨ M e i , e i ⟩ ) = 4 − 2 ∑ i = 1 2 ⟨ M e i , e i ⟩ = 4 − 2 tr M = 4 − 4 ( 1 − δ ) = 4 δ . \begin{aligned}
\lVert M-\operatorname{Id}_H\rVert^2
&\leq\lVert M-\operatorname{Id}_H\rVert_{\mathrm{HS}}^2
=\sum_{i=1}^{2}\lVert(M-\operatorname{Id}_H)e_i\rVert_2^2\\
&=\sum_{i=1}^{2}\lVert Me_i-e_i\rVert_2^2
=\sum_{i=1}^{2}\left(\lVert Me_i\rVert_2^2+\lVert e_i\rVert_2^2-2\langle Me_i,e_i\rangle\right)\\
&\leq\sum_{i=1}^{2}(2-2\langle Me_i,e_i\rangle)
=4-2\sum_{i=1}^{2}\langle Me_i,e_i\rangle\\
&=4-2\operatorname{tr}M=4-4(1-\delta)=4\delta.
\end{aligned} ∥ M − Id H ∥ 2 ≤ ∥ M − Id H ∥ HS 2 = i = 1 ∑ 2 ∥( M − Id H ) e i ∥ 2 2 = i = 1 ∑ 2 ∥ M e i − e i ∥ 2 2 = i = 1 ∑ 2 ( ∥ M e i ∥ 2 2 + ∥ e i ∥ 2 2 − 2 ⟨ M e i , e i ⟩ ) ≤ i = 1 ∑ 2 ( 2 − 2 ⟨ M e i , e i ⟩) = 4 − 2 i = 1 ∑ 2 ⟨ M e i , e i ⟩ = 4 − 2 tr M = 4 − 4 ( 1 − δ ) = 4 δ .
Equations (10) and (11) give A = M − B K ∗ − L C − L D K ∗ A=M-BK^*-LC-LDK^* A = M − B K ∗ − L C − L D K ∗ . Hence (16) and (17) yield
∥ A − Id H ∥ ≤ 2 δ + ( 23 + 23 + 15 ) δ = 2 δ + 61 δ . \lVert A-\operatorname{Id}_H\rVert\leq2\sqrt{\delta}+(23+23+15)\delta=2\sqrt{\delta}+61\delta. ∥ A − Id H ∥ ≤ 2 δ + ( 23 + 23 + 15 ) δ = 2 δ + 61 δ .
Finally, for ∥ h ∥ 2 , ∥ t ∥ ∞ ≤ 1 \|h\|_{2},\|t\|_{\infty}\leq1 ∥ h ∥ 2 , ∥ t ∥ ∞ ≤ 1 ,
∥ ( W − I 0 ) ( h , t ) ∥ = ∥ ( A − Id H ) h + B t ∥ 2 + ∥ C h + D t ∥ 1 ≤ ∥ A − Id H ∥ + ∥ B ∥ + ∥ C ∥ + ∥ D ∥ \begin{aligned}
\|(W-I_{0})(h,t)\|=\|(A-\operatorname{Id}_{H})h+Bt\|_{2}+\|Ch+Dt\|_{1} \\
&\leq\|A-\operatorname{Id}_{H}\|+\|B\|+\|C\|+\|D\|
\end{aligned} ∥ ( W − I 0 ) ( h , t ) ∥ = ∥ ( A − Id H ) h + B t ∥ 2 + ∥ C h + D t ∥ 1 ≤ ∥ A − Id H ∥ + ∥ B ∥ + ∥ C ∥ + ∥ D ∥
≤ 2 δ + 122 δ . \leq2\sqrt{\delta}+122\delta. ≤ 2 δ + 122 δ .
This proves (8) . □
The injective tensor square
Recall that P s = P ∣ V s P_{s}=P|_{V_{s}} P s = P ∣ V s and z 0 = e 1 ⊗ e 1 + e 2 ⊗ e 2 z_{0}=e_{1}\otimes e_{1}+e_{2}\otimes e_{2} z 0 = e 1 ⊗ e 1 + e 2 ⊗ e 2 . For w ∈ V s ⊗ V s w\in V_{s}\otimes V_{s} w ∈ V s ⊗ V s put
Φ s ( w ) = 1 2 tr ( ( P s ⊗ P s ) w ) , \Phi_{s}(w)=\frac{1}{2}\operatorname{tr}((P_{s}\otimes P_{s})w), Φ s ( w ) = 2 1 tr (( P s ⊗ P s ) w ) ,
using the matrix identification (1) .
Proposition 6.1. For every s ≥ 0 s\geq0 s ≥ 0 and every w ∈ V s ⊗ V s w\in V_{s}\otimes V_{s} w ∈ V s ⊗ V s with ∥ w ∥ ε ≤ 1 \|w\|_{\varepsilon}\leq1 ∥ w ∥ ε ≤ 1 ,
∥ w − z 0 ∥ ε ≤ 2 1 − Φ s ( w ) + 122 ( 1 − Φ s ( w ) ) . (18) \|w-z_{0}\|_{\varepsilon}\leq2\sqrt{1-\Phi_{s}(w)}+122(1-\Phi_{s}(w)). \tag*{(18)} ∥ w − z 0 ∥ ε ≤ 2 1 − Φ s ( w ) + 122 ( 1 − Φ s ( w )) . (18)
Proof. Apply Lemma 4.1 to V s V_{s} V s . For a fixed pair γ , η ∈ Γ s \gamma,\eta\in\Gamma_{s} γ , η ∈ Γ s , regard
W γ , η = ( J γ ⊗ J η ) w W_{\gamma,\eta}=(J_{\gamma}\otimes J_{\eta})w W γ , η = ( J γ ⊗ J η ) w
as an operator F s ∗ → F s F_{s}^{*}\to F_{s} F s ∗ → F s . Its norm is at most one because J γ J_{\gamma} J γ and J η J_{\eta} J η are contractions. We apply Lemma 5.1 with
U = V = F s , k = l = s , K = L = K s , W = W γ , η . U=V=F_{s},\qquad k=l=s,\qquad K=L=K_{s},\qquad W=W_{\gamma,\eta}. U = V = F s , k = l = s , K = L = K s , W = W γ , η .
By the factorization R K s J γ = P s R_{K_{s}}J_{\gamma}=P_{s} R K s J γ = P s in Lemma 4.1 , the corresponding compressed operator is
R K s W γ , η R K s ∗ = ( R K s J γ ⊗ R K s J η ) w = ( P s ⊗ P s ) w . \begin{aligned}
R_{K_{s}}W_{\gamma,\eta}R_{K_{s}}^{*}=(R_{K_{s}}J_{\gamma}\otimes R_{K_{s}}J_{\eta})w \\
&=(P_{s}\otimes P_{s})w.
\end{aligned} R K s W γ , η R K s ∗ = ( R K s J γ ⊗ R K s J η ) w = ( P s ⊗ P s ) w .
The first equality is read using the tensor-to-operator identification: applying maps to the first and second tensor factors corresponds to composition on the right by the adjoint and on the left, respectively. Thus every pair has the same trace defect δ = 1 − Φ s ( w ) \delta=1-\Phi_{s}(w) δ = 1 − Φ s ( w ) . The identities J γ e i = ( e i , 0 ) J_{\gamma}e_{i}=(e_{i},0) J γ e i = ( e i , 0 ) imply that ( J γ ⊗ J η ) z 0 (J_{\gamma}\otimes J_{\eta})z_{0} ( J γ ⊗ J η ) z 0 is the operator I 0 ( h , t ) = ( h , 0 ) I_{0}(h,t)=(h,0) I 0 ( h , t ) = ( h , 0 ) . Lemma 5.1 therefore gives, for every pair,
∥ ( J γ ⊗ J η ) ( w − z 0 ) ∥ ε ≤ 2 δ + 122 δ . \|(J_{\gamma}\otimes J_{\eta})(w-z_{0})\|_{\varepsilon}\leq2\sqrt{\delta}+122\delta. ∥ ( J γ ⊗ J η ) ( w − z 0 ) ∥ ε ≤ 2 δ + 122 δ .
Taking the maximum over γ , η \gamma,\eta γ , η and using Lemma 4.2 proves (18) . □
Put Z ε = X ⊗ ^ ε X Z_{\varepsilon}=X\widehat{\otimes}_{\varepsilon}X Z ε = X ⊗ ε X and define
Φ ( z ) = 1 2 tr ( ( P ⊗ P ) z ) ( z ∈ Z ε ) . (19) \Phi(z)=\frac{1}{2}\operatorname{tr}((P\otimes P)z)\qquad(z\in Z_{\varepsilon}). \tag*{(19)} Φ ( z ) = 2 1 tr (( P ⊗ P ) z ) ( z ∈ Z ε ) . (19)
The map P ⊗ P P\otimes P P ⊗ P is contractive, and ∣ tr A ∣ / 2 ≤ ∥ A ∥ o p |\operatorname{tr}A|/2\leq\|A\|_{\mathrm{op}} ∣ tr A ∣/2 ≤ ∥ A ∥ op on M 2 ( R ) M_{2}(\mathbb{R}) M 2 ( R ) . Moreover, ∥ z 0 ∥ ε = 1 \|z_{0}\|_{\varepsilon}=1 ∥ z 0 ∥ ε = 1 by the injectivity of the tensor norm and (1) , and Φ ( z 0 ) = 1 \Phi(z_{0})=1 Φ ( z 0 ) = 1 . Consequently ∥ Φ ∥ = 1 \|\Phi\|=1 ∥Φ∥ = 1 .
Proposition 6.2. For every z ∈ B Z ε z\in B_{Z_{\varepsilon}} z ∈ B Z ε ,
∥ z − z 0 ∥ ε ≤ 2 1 − Φ ( z ) + 122 ( 1 − Φ ( z ) ) . (20) \|z-z_{0}\|_{\varepsilon}\leq2\sqrt{1-\Phi(z)}+122(1-\Phi(z)). \tag*{(20)} ∥ z − z 0 ∥ ε ≤ 2 1 − Φ ( z ) + 122 ( 1 − Φ ( z )) . (20) In particular, Φ \Phi Φ strongly exposes z 0 z_{0} z 0 in B Z ε B_{Z_{\varepsilon}} B Z ε
Proof. By Lemma 3.1 , the operators Q s = ρ s ⊗ ρ s Q_s=\rho_s\otimes\rho_s Q s = ρ s ⊗ ρ s extend to contractions on Z ε Z_\varepsilon Z ε , with range in V s ⊗ V s V_s\otimes V_s V s ⊗ V s . For an elementary tensor,
∥ Q s ( x ⊗ y ) − x ⊗ y ∥ ε ≤ ∥ ρ s x − x ∥ ∥ ρ s y ∥ + ∥ x ∥ ∥ ρ s y − y ∥ ⟶ 0. \left\|Q_s(x\otimes y)-x\otimes y\right\|_\varepsilon\leq\left\|\rho_sx-x\right\|\left\|\rho_sy\right\|+\left\|x\right\|\left\|\rho_sy-y\right\|\longrightarrow0. ∥ Q s ( x ⊗ y ) − x ⊗ y ∥ ε ≤ ∥ ρ s x − x ∥ ∥ ρ s y ∥ + ∥ x ∥ ∥ ρ s y − y ∥ ⟶ 0.
The same holds for finite sums. By density and ∥ Q s ∥ ≤ 1 \|Q_s\|\leq1 ∥ Q s ∥ ≤ 1 , it holds for every z ∈ Z ε z\in Z_\varepsilon z ∈ Z ε . Furthermore, P ρ s = P P\rho_s=P P ρ s = P , so
Φ s ( Q s z ) = Φ ( z ) . \Phi_s(Q_sz)=\Phi(z). Φ s ( Q s z ) = Φ ( z ) .
For z ∈ B Z ε z\in B_{Z_\varepsilon} z ∈ B Z ε , Proposition 6.1 applies to Q s z Q_sz Q s z without any normalization and gives
∥ Q s z − z 0 ∥ ε ≤ 2 1 − Φ ( z ) + 122 ( 1 − Φ ( z ) ) . \left\|Q_sz-z_0\right\|_\varepsilon\leq2\sqrt{1-\Phi(z)}+122(1-\Phi(z)). ∥ Q s z − z 0 ∥ ε ≤ 2 1 − Φ ( z ) + 122 ( 1 − Φ ( z )) .
Passing to the limit proves (6.3). Its right-hand side tends to zero when Φ ( z ) \Phi(z) Φ ( z ) tends to one, which is precisely the asserted strong exposure. □
Corollary 6.3. The operator T ε z = − Φ ( z ) z 0 T_\varepsilon z=-\Phi(z)z_0 T ε z = − Φ ( z ) z 0 on Z ε Z_\varepsilon Z ε has rank one, ∥ T ε ∥ = 1 \|T_\varepsilon\|=1 ∥ T ε ∥ = 1 , and ∥ Id Z ε + T ε ∥ ≤ 199 / 100 \|\operatorname{Id}_{Z_\varepsilon}+T_\varepsilon\|\leq199/100 ∥ Id Z ε + T ε ∥ ≤ 199/100 .
Proof. The norm and rank assertions follow from ∥ z 0 ∥ = ∥ Φ ∥ = Φ ( z 0 ) = 1 \|z_0\|=\|\Phi\|=\Phi(z_0)=1 ∥ z 0 ∥ = ∥Φ∥ = Φ ( z 0 ) = 1 . Let z ∈ B Z ε z\in B_{Z_\varepsilon} z ∈ B Z ε . If ∣ Φ ( z ) ∣ ≤ 99 / 100 |\Phi(z)|\leq99/100 ∣Φ ( z ) ∣ ≤ 99/100 , then
∥ z − Φ ( z ) z 0 ∥ ≤ 1 + ∣ Φ ( z ) ∣ ≤ 199 / 100. \left\|z-\Phi(z)z_0\right\|\leq1+|\Phi(z)|\leq199/100. ∥ z − Φ ( z ) z 0 ∥ ≤ 1 + ∣Φ ( z ) ∣ ≤ 199/100.
Otherwise let w = sign ( Φ ( z ) ) z w=\operatorname{sign}(\Phi(z))z w = sign ( Φ ( z )) z , so that Φ ( w ) = ∣ Φ ( z ) ∣ > 99 / 100 \Phi(w)=|\Phi(z)|>99/100 Φ ( w ) = ∣Φ ( z ) ∣ > 99/100 . Proposition 6.2 gives
∥ w − z 0 ∥ ≤ 2 1 / 100 + 122 / 100 = 71 / 50. \left\|w-z_0\right\|\leq2\sqrt{1/100}+122/100=71/50. ∥ w − z 0 ∥ ≤ 2 1/100 + 122/100 = 71/50.
Therefore
∥ z − Φ ( z ) z 0 ∥ = ∥ w − Φ ( w ) z 0 ∥ ≤ ∥ w − z 0 ∥ + 1 − Φ ( w ) ≤ 143 / 100. \left\|z-\Phi(z)z_0\right\|=\left\|w-\Phi(w)z_0\right\|\leq\left\|w-z_0\right\|+1-\Phi(w)\leq143/100. ∥ z − Φ ( z ) z 0 ∥ = ∥ w − Φ ( w ) z 0 ∥ ≤ ∥ w − z 0 ∥ + 1 − Φ ( w ) ≤ 143/100.
Taking the supremum over z z z proves the result. □
The coordinate dual space
For each s ≥ 0 s\geq0 s ≥ 0 , define
Y s = ρ s ∗ ( V s ∗ ) ⊂ X ∗ , Y = ⋃ s ≥ 0 Y s ‾ ∥ ⋅ ∥ X ∗ . (21) Y_s=\rho_s^*(V_s^*)\subset X^*,\qquad Y=\overline{\bigcup_{s\geq0}Y_s}^{\|\cdot\|_{X^*}}. \tag*{(21)} Y s = ρ s ∗ ( V s ∗ ) ⊂ X ∗ , Y = s ≥ 0 ⋃ Y s ∥ ⋅ ∥ X ∗ . (21)
The adjoint map ρ s ∗ : V s ∗ → X ∗ \rho_s^*:V_s^*\to X^* ρ s ∗ : V s ∗ → X ∗ is an isometry. Indeed, ∥ ρ s ∗ f ∥ ≤ ∥ f ∥ \|\rho_s^*f\|\leq\|f\| ∥ ρ s ∗ f ∥ ≤ ∥ f ∥ by contractivity, while evaluating on B V s ⊂ B X B_{V_s}\subset B_X B V s ⊂ B X gives the reverse inequality. The compatibility of the projections also gives Y s ⊂ Y t Y_s\subset Y_t Y s ⊂ Y t for s ≤ t s\leq t s ≤ t : an element f ∘ ρ s f\circ\rho_s f ∘ ρ s equals ( f ∘ R s , t ) ∘ ρ t (f\circ R_{s,t})\circ\rho_t ( f ∘ R s , t ) ∘ ρ t . In particular, Y Y Y is separable. An element of Y s Y_s Y s vanishes on the new coordinate directions of every later extension. We use the inherited norm on Y Y Y ; no identification of Y ∗ Y^* Y ∗ with X X X is required.
Lemma 7.1. Let p ∈ S V M p\in S_{V_M} p ∈ S V M and suppose that M ≤ s 1 < ⋯ < s r M\leq s_1<\cdots<s_r M ≤ s 1 < ⋯ < s r satisfy p s k = p p_{s_k}=p p s k = p . At the extension V s k + 1 = G ( V s k , p , μ s k ) V_{s_k+1}=G(V_{s_k},p,\mu_{s_k}) V s k + 1 = G ( V s k , p , μ s k ) , let c k : V s k + 1 → R c_k:V_{s_k+1}\to\mathbb{R} c k : V s k + 1 → R , c k ( x , a ) = a 1 c_k(x,a)=a_1 c k ( x , a ) = a 1 , and put λ k = c k ∘ ρ s k + 1 ∈ Y s k + 1 \lambda_k=c_k\circ\rho_{s_k+1}\in Y_{s_k+1} λ k = c k ∘ ρ s k + 1 ∈ Y s k + 1 . Then
∑ k = 1 r ∣ λ k ( x ) ∣ ≤ ∥ x ∥ ( x ∈ X ) . (22) \sum_{k=1}^{r}|\lambda_k(x)|\leq\|x\|\qquad(x\in X). \tag*{(22)} k = 1 ∑ r ∣ λ k ( x ) ∣ ≤ ∥ x ∥ ( x ∈ X ) . (22) In particular, ∥ ∑ k = 1 r λ k ∥ ≤ 1 \left\|\sum_{k=1}^{r}\lambda_k\right\|\leq1 ∥ ∑ k = 1 r λ k ∥ ≤ 1 .
Proof. We construct contractions
J t : V t ⟶ V M ⊕ 1 ℓ 1 d ( t ) , d ( t ) = # { k : s k < t } , M ≤ t ≤ s r + 1 , J_t:V_t\longrightarrow V_M\oplus_1\ell_1^{d(t)},\qquad d(t)=\#\{k:s_k<t\},\qquad M\leq t\leq s_r+1, J t : V t ⟶ V M ⊕ 1 ℓ 1 d ( t ) , d ( t ) = # { k : s k < t } , M ≤ t ≤ s r + 1 ,
such that J t v = ( v , 0 ) J_t v=(v,0) J t v = ( v , 0 ) for v ∈ V M v\in V_M v ∈ V M . Start with J M v = ( v , 0 ) J_Mv=(v,0) J M v = ( v , 0 ) . If t t t is not one of the selected indices, i.e., d ( t ) = d ( t + 1 ) d(t)=d(t+1) d ( t ) = d ( t + 1 ) , define J t + 1 = J t R t J_{t+1}=J_tR_t J t + 1 = J t R t , where R t : V t + 1 → V t R_t:V_{t+1}\to V_t R t : V t + 1 → V t is the canonical retraction given by R t ( x , a ) = x R_t(x,a)=x R t ( x , a ) = x . If t = s k t=s_k t = s k , i.e., d ( t + 1 ) = d ( t ) + 1 d(t+1)=d(t)+1 d ( t + 1 ) = d ( t ) + 1 , define
J t + 1 ( x , a ) = ( J t ( x − a 1 p ) , a 1 ) , J_{t+1}(x,a)=(J_t(x-a_1p),a_1), J t + 1 ( x , a ) = ( J t ( x − a 1 p ) , a 1 ) ,
where the last coordinate is appended to the ℓ 1 \ell_1 ℓ 1 -part. In the second case,
∥ J t + 1 ( x , a ) ∥ ≤ ∥ x − a 1 p ∥ V t + ∣ a 1 ∣ ≤ ∥ ( x , a ) ∥ V t + 1 ; \lVert J_{t+1}(x,a)\rVert\leq\lVert x-a_1p\rVert_{V_t}+|a_1|\leq\lVert(x,a)\rVert_{V_{t+1}}; ∥ J t + 1 ( x , a )∥ ≤ ∥ x − a 1 p ∥ V t + ∣ a 1 ∣ ≤ ∥( x , a ) ∥ V t + 1 ;
in the first case contractivity follows from that of J t J_t J t and R t R_t R t . Both definitions preserve the stated action on V M V_M V M .
More explicitly, induction gives, for every x ∈ V t x\in V_t x ∈ V t ,
J t x = ( ρ M x − p ∑ s k < t λ k ( x ) , ( λ k ( x ) ) s k < t ) . (23) J_tx=\left(\rho_Mx-p\sum_{s_k<t}\lambda_k(x),(\lambda_k(x))_{s_k<t}\right). \tag*{(23)} J t x = ( ρ M x − p s k < t ∑ λ k ( x ) , ( λ k ( x ) ) s k < t ) . (23)
The case t = M t=M t = M is trivial. Suppose that t ∉ { s 1 , … , s r } t\notin\{s_1,\ldots,s_r\} t ∈ / { s 1 , … , s r } . By definition, we have
J t + 1 = J t R t , d ( t + 1 ) = d ( t ) . J_{t+1}=J_tR_t,\qquad d(t+1)=d(t). J t + 1 = J t R t , d ( t + 1 ) = d ( t ) .
Let z = ( x , a ) ∈ V t + 1 z=(x,a)\in V_{t+1} z = ( x , a ) ∈ V t + 1 . The canonical retractions are compatible, so
ρ M z = ρ M ( R t z ) = ρ M x . \rho_Mz=\rho_M(R_tz)=\rho_Mx. ρ M z = ρ M ( R t z ) = ρ M x .
Therefore the formula (7.3) also holds for J t + 1 J_{t+1} J t + 1 if it holds for J t J_t J t . Suppose that t = s j t=s_j t = s j and again let z = ( x , a ) ∈ V t + 1 z=(x,a)\in V_{t+1} z = ( x , a ) ∈ V t + 1 . Then
J t + 1 z = ( J t ( x − a 1 p ) , a 1 ) = ( ρ M ( x − a 1 p ) − p ∑ k < j λ k ( x − a 1 p ) , ( λ k ( x − a 1 p ) ) k < j , a 1 ) = ( ρ M x − a 1 p − p ∑ k < j ( λ k ( x ) − a 1 λ k ( p ) ) , ( λ k ( x ) − a 1 λ k ( p ) ) k < j , a 1 ) = ( ρ M x − a 1 p − p ∑ k < j λ k ( x ) , ( λ k ( x ) ) k < j , a 1 ) = ( ρ M z − p ( ∑ k < j λ k ( z ) + λ j ( z ) ) , ( λ k ( z ) ) k ≤ j ) = ( ρ M z − p ∑ s k < t + 1 λ k ( z ) , ( λ k ( z ) ) s k < t + 1 ) . \begin{aligned}
J_{t+1}z
&=(J_t(x-a_1p),a_1)\\
&=\left(\rho_M(x-a_1p)-p\sum_{k<j}\lambda_k(x-a_1p),(\lambda_k(x-a_1p))_{k<j},a_1\right)\\
&=\left(\rho_Mx-a_1p-p\sum_{k<j}(\lambda_k(x)-a_1\lambda_k(p)),(\lambda_k(x)-a_1\lambda_k(p))_{k<j},a_1\right)\\
&=\left(\rho_Mx-a_1p-p\sum_{k<j}\lambda_k(x),(\lambda_k(x))_{k<j},a_1\right)\\
&=\left(\rho_Mz-p\left(\sum_{k<j}\lambda_k(z)+\lambda_j(z)\right),(\lambda_k(z))_{k\leq j}\right)\\
&=\left(\rho_Mz-p\sum_{s_k<t+1}\lambda_k(z),(\lambda_k(z))_{s_k<t+1}\right).
\end{aligned} J t + 1 z = ( J t ( x − a 1 p ) , a 1 ) = ρ M ( x − a 1 p ) − p k < j ∑ λ k ( x − a 1 p ) , ( λ k ( x − a 1 p ) ) k < j , a 1 = ρ M x − a 1 p − p k < j ∑ ( λ k ( x ) − a 1 λ k ( p )) , ( λ k ( x ) − a 1 λ k ( p ) ) k < j , a 1 = ρ M x − a 1 p − p k < j ∑ λ k ( x ) , ( λ k ( x ) ) k < j , a 1 = ρ M z − p k < j ∑ λ k ( z ) + λ j ( z ) , ( λ k ( z ) ) k ≤ j = ( ρ M z − p s k < t + 1 ∑ λ k ( z ) , ( λ k ( z ) ) s k < t + 1 ) .
since λ k ( p ) = 0 \lambda_k(p)=0 λ k ( p ) = 0 . At t = s r + 1 t=s_r+1 t = s r + 1 , contractivity and (7.3) imply
∥ ρ M x − p ∑ k = 1 r λ k ( x ) ∥ + ∑ k = 1 r ∣ λ k ( x ) ∣ ≤ ∥ x ∥ ( x ∈ V s r + 1 ) . \left\lVert\rho_Mx-p\sum_{k=1}^{r}\lambda_k(x)\right\rVert+\sum_{k=1}^{r}|\lambda_k(x)|\leq\lVert x\rVert\qquad(x\in V_{s_r+1}). ρ M x − p k = 1 ∑ r λ k ( x ) + k = 1 ∑ r ∣ λ k ( x ) ∣ ≤ ∥ x ∥ ( x ∈ V s r + 1 ) .
Discarding the first term gives (7.2) on this stage. For x ∈ X x\in X x ∈ X , apply that inequality to ρ s r + 1 x \rho_{s_r+1}x ρ s r + 1 x . Compatibility gives λ k ( ρ s r + 1 x ) = λ k ( x ) \lambda_k(\rho_{s_r+1}x)=\lambda_k(x) λ k ( ρ s r + 1 x ) = λ k ( x ) , and ∥ ρ s r + 1 x ∥ ≤ ∥ x ∥ \lVert\rho_{s_r+1}x\rVert\leq\lVert x\rVert ∥ ρ s r + 1 x ∥ ≤ ∥ x ∥ . The final assertion follows by the triangle inequality for scalar values. □
Proposition 7.2. The space Y Y Y has the Daugavet property.
Proof. First fix f , g ∈ S Y N f,g \in S_{Y_N} f , g ∈ S Y N and η > 0 \eta>0 η > 0 . Since P \mathcal{P} P is dense in S X S_X S X by (3.2) and ∥ g ∥ X ∗ = 1 \|g\|_{X^*}=1 ∥ g ∥ X ∗ = 1 , choose p ∈ P p\in\mathcal{P} p ∈ P with g ( p ) > 1 − η g(p)>1-\eta g ( p ) > 1 − η . Choose M ≥ N M\ge N M ≥ N with p ∈ V M p\in V_M p ∈ V M . The restriction of f f f to V M V_M V M still has norm one: it is at most one, and already has norm one on V N V_N V N by the isometry of ρ N ∗ \rho_N^* ρ N ∗ .
For any positive integer r r r , the recurrence in (3.2) gives indices M ≤ s 1 < ⋯ < s r M\le s_1<\cdots<s_r M ≤ s 1 < ⋯ < s r with p s k = p p_{s_k}=p p s k = p . Let λ k \lambda_k λ k be as in Lemma 7.1 , and put
t = 1 − f ( p ) ∈ [ 0 , 2 ] , h k = f + t λ k ∈ Y . t=1-f(p)\in[0,2],\qquad h_k=f+t\lambda_k\in Y. t = 1 − f ( p ) ∈ [ 0 , 2 ] , h k = f + t λ k ∈ Y .
On V s k + 1 = G ( V s k , p , μ s k ) V_{s_k+1}=G(V_{s_k},p,\mu_{s_k}) V s k + 1 = G ( V s k , p , μ s k ) , the old functional f f f ignores the new coordinates. Therefore
h k ( x , a ) = f ( x ) + ( 1 − f ( p ) ) a 1 = f ( x − a 1 p ) + a 1 , h_k(x,a)=f(x)+(1-f(p))a_1=f(x-a_1p)+a_1, h k ( x , a ) = f ( x ) + ( 1 − f ( p )) a 1 = f ( x − a 1 p ) + a 1 ,
and hence
∣ h k ( x , a ) ∣ ≤ ∥ x − a 1 p ∥ + ∣ a 1 ∣ ≤ ∥ ( x , a ) ∥ V s k + 1 . |h_k(x,a)|\le\|x-a_1p\|+|a_1|\le\|(x,a)\|_{V_{s_k+1}}. ∣ h k ( x , a ) ∣ ≤ ∥ x − a 1 p ∥ + ∣ a 1 ∣ ≤ ∥ ( x , a ) ∥ V s k + 1 .
Both f f f and λ k \lambda_k λ k factor through ρ s k + 1 \rho_{s_k+1} ρ s k + 1 , therefore also h k h_k h k , so this inequality gives ∥ h k ∥ X ∗ ≤ 1 \|h_k\|_{X^*}\le1 ∥ h k ∥ X ∗ ≤ 1 . On V M V_M V M , one has h k = f h_k=f h k = f , because λ k \lambda_k λ k vanishes there. Thus ∥ h k ∥ = 1 \|h_k\|=1 ∥ h k ∥ = 1 .
Let q k = ( p , u 1 ) ∈ V s k + 1 q_k=(p,u_1)\in V_{s_k+1} q k = ( p , u 1 ) ∈ V s k + 1 . Lemma 2.1 gives ∥ q k ∥ = 1 \|q_k\|=1 ∥ q k ∥ = 1 . Since f , g ∈ Y N f,g\in Y_N f , g ∈ Y N and N ≤ s k N\le s_k N ≤ s k ,
h k ( q k ) = f ( p ) + t = 1 , g ( q k ) = g ( p ) > 1 − η . h_k(q_k)=f(p)+t=1,\qquad g(q_k)=g(p)>1-\eta. h k ( q k ) = f ( p ) + t = 1 , g ( q k ) = g ( p ) > 1 − η .
It follows that
∥ g + h k ∥ ≥ ( g + h k ) ( q k ) > 2 − η . \|g+h_k\|\ge(g+h_k)(q_k)>2-\eta. ∥ g + h k ∥ ≥ ( g + h k ) ( q k ) > 2 − η .
On the other hand, Lemma 7.1 gives
∥ 1 r ∑ k = 1 r h k − f ∥ = t r ∥ ∑ k = 1 r λ k ∥ ≤ 2 r . (24) \left\|\frac{1}{r}\sum_{k=1}^{r}h_k-f\right\|=\frac{t}{r}\left\|\sum_{k=1}^{r}\lambda_k\right\|\le\frac{2}{r}. \tag*{(24)} r 1 k = 1 ∑ r h k − f = r t k = 1 ∑ r λ k ≤ r 2 . (24)
Thus these two estimates hold with arbitrarily small errors for normalized vectors in a common finite dual stage.
For arbitrary f , g ∈ S Y f,g\in S_Y f , g ∈ S Y and η > 0 \eta>0 η > 0 , choose 0 < τ < η / 3 0<\tau<\eta/3 0 < τ < η /3 and f 0 , g 0 ∈ S Y N f_0,g_0\in S_{Y_N} f 0 , g 0 ∈ S Y N with ∥ f − f 0 ∥ < τ \|f-f_0\|<\tau ∥ f − f 0 ∥ < τ and ∥ g − g 0 ∥ < τ \|g-g_0\|<\tau ∥ g − g 0 ∥ < τ . This is possible by (7.1). Apply the preceding construction to f 0 , g 0 f_0,g_0 f 0 , g 0 with tolerance τ \tau τ and an integer r r r such that 2 / r < τ 2/r<\tau 2/ r < τ . Then
∥ 1 r ∑ k = 1 r h k − f ∥ ≤ ∥ 1 r ∑ k = 1 r h k − f 0 ∥ + ∥ f 0 − f ∥ < 2 τ < η , \left\|\frac{1}{r}\sum_{k=1}^{r}h_k-f\right\|\le\left\|\frac{1}{r}\sum_{k=1}^{r}h_k-f_0\right\|+\|f_0-f\|<2\tau<\eta, r 1 k = 1 ∑ r h k − f ≤ r 1 k = 1 ∑ r h k − f 0 + ∥ f 0 − f ∥ < 2 τ < η ,
while
∥ g + h k ∥ ≥ ∥ g 0 + h k ∥ − ∥ g − g 0 ∥ > 2 − 2 τ > 2 − η . \|g+h_k\|\ge\|g_0+h_k\|-\|g-g_0\|>2-2\tau>2-\eta. ∥ g + h k ∥ ≥ ∥ g 0 + h k ∥ − ∥ g − g 0 ∥ > 2 − 2 τ > 2 − η .
The proof is therefore finished by Theorem 1.1 (3). □
The projective tensor square
Put Z π = Y ⊗ ^ π Y Z_\pi=Y\widehat{\otimes}_\pi Y Z π = Y ⊗ π Y and regard Z π ∗ Z_\pi^* Z π ∗ as the space of bounded bilinear forms on Y × Y Y\times Y Y × Y . Let e i ∗ ∈ H ∗ e_i^*\in H^* e i ∗ ∈ H ∗ be the coordinate functionals and define
u i = ρ 0 ∗ e i ∗ ∈ Y 0 , F = 1 2 ( u 1 ⊗ u 1 + u 2 ⊗ u 2 ) , B 0 ( f , g ) = ∑ i = 1 2 f ( e i ) g ( e i ) . (25) u_i=\rho_0^*e_i^*\in Y_0,\qquad F=\frac{1}{2}(u_1\otimes u_1+u_2\otimes u_2),\qquad B_0(f,g)=\sum_{i=1}^{2}f(e_i)g(e_i). \tag*{(25)} u i = ρ 0 ∗ e i ∗ ∈ Y 0 , F = 2 1 ( u 1 ⊗ u 1 + u 2 ⊗ u 2 ) , B 0 ( f , g ) = i = 1 ∑ 2 f ( e i ) g ( e i ) . (25)
Restriction f ↦ f ∣ H f \mapsto f|_H f ↦ f ∣ H is a contraction from Y Y Y to H ∗ H^* H ∗ . The Cauchy–Schwarz inequality therefore gives ∥ B 0 ∥ ≤ 1 \|B_0\| \le1 ∥ B 0 ∥ ≤ 1 . Also ∥ u i ∥ = 1 \|u_i\|=1 ∥ u i ∥ = 1 and u i ( e j ) = δ i j u_i(e_j)=\delta_{ij} u i ( e j ) = δ ij , so B 0 ( u i , u i ) = 1 B_0(u_i,u_i)=1 B 0 ( u i , u i ) = 1 and therefore
B 0 ( F ) = B 0 ( 1 2 ( u 1 ⊗ u 1 + u 2 ⊗ u 2 ) ) = 1 2 ( B 0 ( u 1 , u 1 ) + B 0 ( u 2 , u 2 ) ) = 1 2 ∑ i = 1 2 ∑ j = 1 2 u i ( e j ) 2 = 1 2 ∑ i = 1 2 ∑ j = 1 2 δ i j 2 = 1 2 ( 1 + 1 ) = 1 , B_0(F)=B_0\left(\frac{1}{2}(u_1\otimes u_1+u_2\otimes u_2)\right)
= \frac{1}{2}\left(B_0(u_1,u_1)+B_0(u_2,u_2)\right)
= \frac{1}{2}\sum_{i=1}^{2}\sum_{j=1}^{2}u_i(e_j)^2
= \frac{1}{2}\sum_{i=1}^{2}\sum_{j=1}^{2}\delta_{ij}^{2}
= \frac{1}{2}(1+1)=1, B 0 ( F ) = B 0 ( 2 1 ( u 1 ⊗ u 1 + u 2 ⊗ u 2 ) ) = 2 1 ( B 0 ( u 1 , u 1 ) + B 0 ( u 2 , u 2 ) ) = 2 1 i = 1 ∑ 2 j = 1 ∑ 2 u i ( e j ) 2 = 2 1 i = 1 ∑ 2 j = 1 ∑ 2 δ ij 2 = 2 1 ( 1 + 1 ) = 1 ,
by the identification of Z π ∗ Z_\pi^* Z π ∗ with the bounded bilinear forms on Y × Y Y\times Y Y × Y . The representation of F F F in (8.1) gives ∥ F ∥ π ≤ 1 \|F\|_\pi\le1 ∥ F ∥ π ≤ 1 , while evaluation by B 0 B_0 B 0 gives the reverse inequality. Thus
∥ B 0 ∥ = ∥ F ∥ π = B 0 ( F ) = 1. (26) \|B_0\|=\|F\|_\pi=B_0(F)=1. \tag*{(26)} ∥ B 0 ∥ = ∥ F ∥ π = B 0 ( F ) = 1. (26)
Proposition 8.1. For every B ∈ B Z π ∗ B\in B_{Z_\pi^*} B ∈ B Z π ∗ ,
∥ B − B 0 ∥ ≤ 2 1 − B ( F ) + 122 ( 1 − B ( F ) ) . (27) \|B-B_0\|\le2\sqrt{1-B(F)}+122(1-B(F)). \tag*{(27)} ∥ B − B 0 ∥ ≤ 2 1 − B ( F ) + 122 ( 1 − B ( F )) . (27) In particular, evaluation at F F F weak-star strongly exposes B 0 B_0 B 0 in B Z π ∗ B_{Z_\pi^*} B Z π ∗ .
Proof. Fix s ≥ 0 s\ge0 s ≥ 0 . Via the isometry ρ s ∗ : V s ∗ → Y s \rho_s^*:V_s^*\to Y_s ρ s ∗ : V s ∗ → Y s , the restriction of B B B to Y s × Y s Y_s\times Y_s Y s × Y s becomes the bilinear form
B ~ s ( a ∗ , b ∗ ) = B ( ρ s ∗ a ∗ , ρ s ∗ b ∗ ) ( a ∗ , b ∗ ∈ V s ∗ ) . \widetilde{B}_s(a^*,b^*)=B(\rho_s^*a^*,\rho_s^*b^*)\qquad(a^*,b^*\in V_s^*). B s ( a ∗ , b ∗ ) = B ( ρ s ∗ a ∗ , ρ s ∗ b ∗ ) ( a ∗ , b ∗ ∈ V s ∗ ) .
Since V s V_s V s is finite-dimensional, there is a unique tensor w s ∈ V s ⊗ V s w_s\in V_s\otimes V_s w s ∈ V s ⊗ V s with
( a ∗ ⊗ b ∗ ) ( w s ) = B ~ s ( a ∗ , b ∗ ) ( a ∗ , b ∗ ∈ V s ∗ ) . (a^*\otimes b^*)(w_s)=\widetilde{B}_s(a^*,b^*)\qquad(a^*,b^*\in V_s^*). ( a ∗ ⊗ b ∗ ) ( w s ) = B s ( a ∗ , b ∗ ) ( a ∗ , b ∗ ∈ V s ∗ ) .
Taking the supremum over the two unit balls gives exactly
∥ w s ∥ ε = ∥ B ∣ Y s × Y s ∥ ≤ 1. (28) \|w_s\|_\varepsilon=\|B|_{Y_s\times Y_s}\|\le1. \tag*{(28)} ∥ w s ∥ ε = ∥ B ∣ Y s × Y s ∥ ≤ 1. (28)
The analogous tensor for B 0 ∣ Y s × Y s B_0|_{Y_s\times Y_s} B 0 ∣ Y s × Y s is z 0 z_0 z 0 , since ( ρ s ∗ a ∗ ) ( e i ) = a ∗ ( e i ) (\rho_s^*a^*)(e_i)=a^*(e_i) ( ρ s ∗ a ∗ ) ( e i ) = a ∗ ( e i ) . Write
w s = ∑ j = 1 m x j ⊗ y j , x j , y j ∈ V s . w_s=\sum_{j=1}^{m}x_j\otimes y_j,\qquad x_j,y_j\in V_s. w s = j = 1 ∑ m x j ⊗ y j , x j , y j ∈ V s .
For the compressed trace, note first that ρ s ∗ P s e i ∗ = ρ 0 ∗ e i ∗ = u i \rho_s^*P_s e_i^*=\rho_0^*e_i^*=u_i ρ s ∗ P s e i ∗ = ρ 0 ∗ e i ∗ = u i , because P s ρ s = P P_s\rho_s=P P s ρ s = P . By the matrix identification (1) ,
Φ s ( w s ) = 1 2 tr ( ( P s ⊗ P s ) w s ) = 1 2 ∑ i = 1 2 e i ∗ ( P s w s P s ∗ e i ) = 1 2 ∑ i = 1 2 ( P s ∗ e i ∗ ) ( w ( P s ∗ e i ∗ ) ) = 1 2 ∑ i = 1 2 ( P s ∗ e i ∗ ) ( ∑ j = 1 m ( P s ∗ e i ∗ ) ( x j ) y j ) = 1 2 ∑ i = 1 2 ∑ j = 1 m ( P s ∗ e i ∗ ) ( x j ) ( P s ∗ e i ∗ ) ( y j ) = 1 2 ∑ i = 1 2 ( P s ∗ e i ∗ ⊗ P s ∗ e i ∗ ) ( w s ) = 1 2 ∑ i = 1 2 B ~ s ( P s ∗ e i ∗ , P s ∗ e i ∗ ) = 1 2 ∑ i = 1 2 B ( ρ s ∗ P s ∗ e i ∗ , ρ s ∗ P s ∗ e i ∗ ) = 1 2 ∑ i = 1 2 B ( u i , u i ) = B ( 1 2 ∑ i = 1 2 u i ⊗ u i ) = B ( F ) . \begin{aligned}
\Phi_s(w_s)&=\frac{1}{2}\operatorname{tr}((P_s\otimes P_s)w_s)=\frac{1}{2}\sum_{i=1}^{2}e_i^*(P_sw_sP_s^*e_i)=\frac{1}{2}\sum_{i=1}^{2}(P_s^*e_i^*)(w(P_s^*e_i^*))\\
&=\frac{1}{2}\sum_{i=1}^{2}(P_s^*e_i^*)\left(\sum_{j=1}^{m}(P_s^*e_i^*)(x_j)y_j\right)=\frac{1}{2}\sum_{i=1}^{2}\sum_{j=1}^{m}(P_s^*e_i^*)(x_j)(P_s^*e_i^*)(y_j)\\
&=\frac{1}{2}\sum_{i=1}^{2}(P_s^*e_i^*\otimes P_s^*e_i^*)(w_s)=\frac{1}{2}\sum_{i=1}^{2}\widetilde{B}_s(P_s^*e_i^*,P_s^*e_i^*)\\
&=\frac{1}{2}\sum_{i=1}^{2}B(\rho_s^*P_s^*e_i^*,\rho_s^*P_s^*e_i^*)=\frac{1}{2}\sum_{i=1}^{2}B(u_i,u_i)\\
&=B\left(\frac{1}{2}\sum_{i=1}^{2}u_i\otimes u_i\right)\\
&=B(F).
\end{aligned} Φ s ( w s ) = 2 1 tr (( P s ⊗ P s ) w s ) = 2 1 i = 1 ∑ 2 e i ∗ ( P s w s P s ∗ e i ) = 2 1 i = 1 ∑ 2 ( P s ∗ e i ∗ ) ( w ( P s ∗ e i ∗ )) = 2 1 i = 1 ∑ 2 ( P s ∗ e i ∗ ) ( j = 1 ∑ m ( P s ∗ e i ∗ ) ( x j ) y j ) = 2 1 i = 1 ∑ 2 j = 1 ∑ m ( P s ∗ e i ∗ ) ( x j ) ( P s ∗ e i ∗ ) ( y j ) = 2 1 i = 1 ∑ 2 ( P s ∗ e i ∗ ⊗ P s ∗ e i ∗ ) ( w s ) = 2 1 i = 1 ∑ 2 B s ( P s ∗ e i ∗ , P s ∗ e i ∗ ) = 2 1 i = 1 ∑ 2 B ( ρ s ∗ P s ∗ e i ∗ , ρ s ∗ P s ∗ e i ∗ ) = 2 1 i = 1 ∑ 2 B ( u i , u i ) = B ( 2 1 i = 1 ∑ 2 u i ⊗ u i ) = B ( F ) .
The trace defect is thus independent of s s s . Proposition 6.1 and (28) , applied also to the difference, yield
∥ ( B − B 0 ) ∣ Y s × Y s ∥ = ∥ w s − z 0 ∥ ε ≤ 2 1 − B ( F ) + 122 ( 1 − B ( F ) ) . (29) \lVert(B-B_0)|_{Y_s\times Y_s}\rVert=\lVert w_s-z_0\rVert_\varepsilon\leq2\sqrt{1-B(F)}+122(1-B(F)). \tag*{(29)} ∥( B − B 0 ) ∣ Y s × Y s ∥ = ∥ w s − z 0 ∥ ε ≤ 2 1 − B ( F ) + 122 ( 1 − B ( F )) . (29)
For completeness, if C C C is any bounded bilinear form on Y × Y Y\times Y Y × Y , then
∥ C ∥ = sup s ∥ C ∣ Y s × Y s ∥ . \lVert C\rVert=\sup_s\lVert C|_{Y_s\times Y_s}\rVert. ∥ C ∥ = s sup ∥ C ∣ Y s × Y s ∥ .
One inequality follows from restriction. For the other, fix f , g ∈ S Y f,g\in S_Y f , g ∈ S Y . By density and normalization choose f j , g j f_j,g_j f j , g j in the unit spheres of a common stage Y s j Y_{s_j} Y s j , with f j → f f_j\to f f j → f and g j → g g_j\to g g j → g . Bilinear continuity gives C ( f , g ) = lim j C ( f j , g j ) C(f,g)=\lim_j C(f_j,g_j) C ( f , g ) = lim j C ( f j , g j ) , whose absolute value is bounded by the displayed supremum. Now take the supremum over f , g ∈ S Y f,g\in S_Y f , g ∈ S Y . Applying this observation to (29) proves (27) . Since the exposing functional is evaluation at the element F ∈ Z π F\in Z_\pi F ∈ Z π , the strong exposure is weak-star strong exposure in the dual ball. □
Corollary 8.2. The operator T π z = − B 0 ( z ) F T_\pi z=-B_0(z)F T π z = − B 0 ( z ) F on Z π Z_\pi Z π has rank one, ∥ T π ∥ = 1 \lVert T_\pi\rVert=1 ∥ T π ∥ = 1 , and ∥ Id Z π + T π ∥ ≤ 199 / 100 \lVert\operatorname{Id}_{Z_\pi}+T_\pi\rVert\leq199/100 ∥ Id Z π + T π ∥ ≤ 199/100 .
Proof. The norm and rank assertions follow from (26) . For B ∈ Z π ∗ B\in Z_\pi^* B ∈ Z π ∗ one has
T π ∗ B = − B ( F ) B 0 . T_\pi^*B=-B(F)B_0. T π ∗ B = − B ( F ) B 0 . Let ∥ B ∥ ≤ 1 \lVert B\rVert\leq1 ∥ B ∥ ≤ 1 . If ∣ B ( F ) ∣ ≤ 99 / 100 |B(F)|\leq99/100 ∣ B ( F ) ∣ ≤ 99/100 , then
∥ B + T π ∗ B ∥ ≤ 1 + ∣ B ( F ) ∣ ≤ 199 / 100. \lVert B+T_\pi^*B\rVert\leq1+|B(F)|\leq199/100. ∥ B + T π ∗ B ∥ ≤ 1 + ∣ B ( F ) ∣ ≤ 199/100. Otherwise put C = sign ( B ( F ) ) B C=\operatorname{sign}(B(F))B C = sign ( B ( F )) B , so that C ( F ) = ∣ B ( F ) ∣ > 99 / 100 C(F)=|B(F)|>99/100 C ( F ) = ∣ B ( F ) ∣ > 99/100 . By Proposition 8.1 ,
∥ C − B 0 ∥ ≤ 2 1 / 100 + 122 / 100 = 71 / 50. \lVert C-B_0\rVert\leq2\sqrt{1/100}+122/100=71/50. ∥ C − B 0 ∥ ≤ 2 1/100 + 122/100 = 71/50. It follows that
∥ B + T π ∗ B ∥ = ∥ C − C ( F ) B 0 ∥ ≤ ∥ C − B 0 ∥ + 1 − C ( F ) ≤ 143 / 100. \lVert B+T_\pi^*B\rVert=\lVert C-C(F)B_0\rVert\leq\lVert C-B_0\rVert+1-C(F)\leq143/100. ∥ B + T π ∗ B ∥ = ∥ C − C ( F ) B 0 ∥ ≤ ∥ C − B 0 ∥ + 1 − C ( F ) ≤ 143/100. Consequently
∥ Id Z π + T π ∥ = ∥ Id Z π ∗ + T π ∗ ∥ ≤ 199 / 100. \lVert\operatorname{Id}_{Z_\pi}+T_\pi\rVert=\lVert\operatorname{Id}_{Z_\pi^*}+T_\pi^*\rVert\leq199/100. ∥ Id Z π + T π ∥ = ∥ Id Z π ∗ + T π ∗ ∥ ≤ 199/100. □
Propositions 3.2 and 7.2, together with Corollaries 6.3 and 8.2, prove Theorem 1.2 .
Relation with the operator Daugavet properties
The operator Daugavet property (ODP), introduced in [8 Definition 4.1] , means that, given x 1 , … , x n ∈ S E x_1,\ldots,x_n \in S_E x 1 , … , x n ∈ S E , η > 0 \eta> 0 η > 0 , and a slice S S S of B E B_E B E , there is h ∈ S h \in S h ∈ S such that for every q ∈ B E q \in B_E q ∈ B E there is an operator A : E → E A:E\to E A : E → E with ∥ A ∥ ≤ 1 + η \lVert A\rVert\leq1+\eta ∥ A ∥ ≤ 1 + η , ∥ A x i − x i ∥ < η \lVert Ax_i-x_i\rVert<\eta ∥ A x i − x i ∥ < η for every i i i , and A h = q Ah=q A h = q . The weak operator Daugavet property of [6 Definition 5.2] allows h h h to depend on q q q and requires only ∥ A h − q ∥ < η \lVert Ah-q\rVert<\eta ∥ A h − q ∥ < η .
Proposition 9.1. The space X X X has the ODP. The space Y Y Y does not have the weak operator Daugavet property.
Proof. Fix x 1 , … , x n ∈ S X x_1,\ldots,x_n\in S_X x 1 , … , x n ∈ S X , η > 0 \eta>0 η > 0 , and a slice S = Slice ( B X , x ∗ , α ) S=\operatorname{Slice}(B_X,x^*,\alpha) S = Slice ( B X , x ∗ , α ) . Choose v ∈ S X v\in S_X v ∈ S X with x ∗ ( v ) > 1 − α / 4 x^*(v)>1-\alpha/4 x ∗ ( v ) > 1 − α /4 . At a sufficiently late outer stage E N E_N E N , we can find y i , v 0 ∈ E N y_i,v_0\in E_N y i , v 0 ∈ E N with ∥ x i − y i ∥ < η / 3 \lVert x_i-y_i\rVert<\eta/3 ∥ x i − y i ∥ < η /3 , v 0 ∈ S E N v_0\in S_{E_N} v 0 ∈ S E N and x ∗ ( v 0 ) > 1 − α / 2 x^*(v_0)>1-\alpha/2 x ∗ ( v 0 ) > 1 − α /2 . Choose n 0 ≥ N n_0\geq N n 0 ≥ N with ε n 0 + 2 / m n 0 < α / 2 \varepsilon_{n_0}+2/m_{n_0}<\alpha/2 ε n 0 + 2/ m n 0 < α /2 , and choose p ∈ N n 0 p\in N_{n_0} p ∈ N n 0 with ∥ p − v 0 ∥ < ε n 0 \lVert p-v_0\rVert<\varepsilon_{n_0} ∥ p − v 0 ∥ < ε n 0 . At the extension V s + 1 = G ( V s , p , m n 0 ) V_{s+1}=G(V_s,p,m_{n_0}) V s + 1 = G ( V s , p , m n 0 ) corresponding to this point, the vectors h j = ( p , u j ) h_j=(p,u_j) h j = ( p , u j ) have an average at distance at most ε n 0 + 2 / m n 0 \varepsilon_{n_0}+2/m_{n_0} ε n 0 + 2/ m n 0 from v 0 v_0 v 0 by Lemma 2.1. Hence some h j h_j h j belongs to S S S . Fix this j j j .
For any q ∈ B X q\in B_X q ∈ B X , define
A q 0 : V s + 1 → X , A q 0 ( x , a ) = x − a j p + a j q . A_q^0:V_{s+1}\to X,\qquad A_q^0(x,a)=x-a_jp+a_jq. A q 0 : V s + 1 → X , A q 0 ( x , a ) = x − a j p + a j q . The extension norm gives
∥ A q 0 ( x , a ) ∥ ≤ ∥ x − a j p ∥ + ∣ a j ∣ ≤ ∥ ( x , a ) ∥ V s + 1 . \lVert A_q^0(x,a)\rVert\leq\lVert x-a_jp\rVert+|a_j|\leq\lVert(x,a)\rVert_{V_{s+1}}. ∥ A q 0 ( x , a )∥ ≤ ∥ x − a j p ∥ + ∣ a j ∣ ≤ ∥( x , a ) ∥ V s + 1 . Thus A q = A q 0 ρ s + 1 : X → X A_q=A_q^0\rho_{s+1}:X\to X A q = A q 0 ρ s + 1 : X → X is contractive. It fixes every vector of V s V_s V s , sends h j h_j h j to q q q , and satisfies
∥ A q x i − x i ∥ ≤ 2 ∥ x i − y i ∥ < η . \lVert A_qx_i-x_i\rVert\leq2\lVert x_i-y_i\rVert<\eta. ∥ A q x i − x i ∥ ≤ 2 ∥ x i − y i ∥ < η . The choice of h j h_j h j is independent of q q q , proving the ODP of X X X .
If Y Y Y had the weak operator Daugavet property, then [6 Theorem 5.4] would imply that Y ⊗ ^ π Y Y\widehat{\otimes}_{\pi}Y Y ⊗ π Y has that property, and hence the Daugavet property by [6 Remark 5.3] . This contradicts Corollary 8.2. □
There are also natural examples of spaces with the Daugavet property and without the weak operator Daugavet property. We use the following criterion:
Lemma 9.2 ([3 Lemma 2.2] ). Let E E E be a Banach space with the WODP, and let δ , ε > 0 \delta,\varepsilon>0 δ , ε > 0 . Suppose that x 1 , x 2 ∈ S E x_1,x_2\in S_E x 1 , x 2 ∈ S E satisfy
B E ( x 1 , 1 2 ) ∩ B E ( x 2 , 1 2 ) ≠ ∅ , B_E\left(x_1,\frac{1}{2}\right)\cap B_E\left(x_2,\frac{1}{2}\right)\ne\varnothing, B E ( x 1 , 2 1 ) ∩ B E ( x 2 , 2 1 ) = ∅ , and that every
z ∈ B E ( x 1 , 1 2 + 3 2 ε ) ∩ B E ( x 2 , 1 2 + 3 2 ε ) z\in B_E\left(x_1,\frac{1}{2}+\frac{3}{2}\varepsilon\right)\cap B_E\left(x_2,\frac{1}{2}+\frac{3}{2}\varepsilon\right) z ∈ B E ( x 1 , 2 1 + 2 3 ε ) ∩ B E ( x 2 , 2 1 + 2 3 ε ) satisfies ∥ z + x 1 ∥ > 2 − δ \lVert z+x_1\rVert>2-\delta ∥ z + x 1 ∥ > 2 − δ . Then, for every slice S S S of B E B_E B E , there exists s ∈ S s\in S s ∈ S such that
( 1 + ε ) ∥ s + u ∥ > 2 − δ − ε (1+\varepsilon)\lVert s+u\rVert>2-\delta-\varepsilon ( 1 + ε ) ∥ s + u ∥ > 2 − δ − ε for every u ∈ B E ( x 1 , 1 2 ) ∩ B E ( x 2 , 1 2 ) u\in B_E\left(x_1,\frac{1}{2}\right)\cap B_E\left(x_2,\frac{1}{2}\right) u ∈ B E ( x 1 , 2 1 ) ∩ B E ( x 2 , 2 1 ) .
Proposition 9.3. Let
M = R / ( 3 Z ) , d ( [ s ] , [ t ] ) = min k ∈ Z ∣ s − t + 3 k ∣ , M=\mathbb{R}/(3\mathbb{Z}),\qquad d([s],[t])=\min_{k\in\mathbb{Z}}|s-t+3k|, M = R / ( 3 Z ) , d ([ s ] , [ t ]) = k ∈ Z min ∣ s − t + 3 k ∣ ,
with distinguished point [ 0 ] [0] [ 0 ] . The Banach space
X = { f ∈ L ∞ ( 0 , 3 ; R ) : ∫ 0 3 f ( t ) d t = 0 } , ∥ f ∥ = ∥ f ∥ ∞ , X=\left\{f\in L_\infty(0,3;\mathbb{R}):\int_0^3 f(t)\,\mathrm{d}t=0\right\},\qquad\lVert f\rVert=\lVert f\rVert_\infty, X = { f ∈ L ∞ ( 0 , 3 ; R ) : ∫ 0 3 f ( t ) d t = 0 } , ∥ f ∥ = ∥ f ∥ ∞ ,
is linearly isometric to Lip 0 ( M ) \operatorname{Lip}_0(M) Lip 0 ( M ) and does not have the WODP while having the Daugavet property. Consequently, Lip 0 ( M ) \operatorname{Lip}_0(M) Lip 0 ( M ) does not have the WODP either. Here Lip 0 ( M ) \operatorname{Lip}_0(M) Lip 0 ( M ) is the space of real-valued Lipschitz functions on M M M vanishing at [ 0 ] [0] [ 0 ] , equipped with the Lipschitz norm.
The isometry is given by
D : Lip 0 ( M ) ⟶ X , D F = F ~ ′ ∣ ( 0 , 3 ) , D:\operatorname{Lip}_0(M)\longrightarrow X,\qquad DF=\widetilde{F}'\big|_{(0,3)}, D : Lip 0 ( M ) ⟶ X , D F = F ′ ( 0 , 3 ) ,
where F ~ ( t ) = F ( [ t ] ) \widetilde{F}(t)=F([t]) F ( t ) = F ([ t ]) is the 3 3 3 -periodic lift of F F F , and the derivative is understood almost everywhere.
Proof. The isometry between X X X and Lip 0 ( M ) \operatorname{Lip}_0(M) Lip 0 ( M ) is clear by
f ⟼ F f , F f ( [ t ] ) = ∫ 0 t f ( r ) d r , 0 ≤ t ≤ 3 , f\longmapsto F_f,\qquad F_f([t])=\int_0^t f(r)\,dr,\qquad0\leq t\leq3, f ⟼ F f , F f ([ t ]) = ∫ 0 t f ( r ) d r , 0 ≤ t ≤ 3 ,
and Lip 0 ( M ) \operatorname{Lip}_0(M) Lip 0 ( M ) has the Daugavet property, since M M M is geodesic [2 ] . To prove that X X X does not have the WODP, set A = ( 0 , 1 ) A=(0,1) A = ( 0 , 1 ) , B = ( 1 , 2 ) B=(1,2) B = ( 1 , 2 ) , and C = ( 2 , 3 ) C=(2,3) C = ( 2 , 3 ) . We use ( a , b , c ) (a,b,c) ( a , b , c ) to denote the function taking the corresponding constant values on these intervals. In particular,
x 1 = ( 1 , − 1 , 0 ) , x 2 = ( 1 , 0 , − 1 ) x_1=(1,-1,0),\qquad x_2=(1,0,-1) x 1 = ( 1 , − 1 , 0 ) , x 2 = ( 1 , 0 , − 1 )
belong to S X S_X S X . The intersection B X ( x 1 , 1 2 ) ∩ B X ( x 2 , 1 2 ) B_X(x_1,\frac{1}{2})\cap B_X(x_2,\frac{1}{2}) B X ( x 1 , 2 1 ) ∩ B X ( x 2 , 2 1 ) is nonempty, since
x 1 + x 2 2 = ( 1 , − 1 2 , − 1 2 ) ∈ B X ( x 1 , 1 2 ) ∩ B X ( x 2 , 1 2 ) . \frac{x_1+x_2}{2}=\left(1,-\frac{1}{2},-\frac{1}{2}\right)\in B_X\left(x_1,\frac{1}{2}\right)\cap B_X\left(x_2,\frac{1}{2}\right). 2 x 1 + x 2 = ( 1 , − 2 1 , − 2 1 ) ∈ B X ( x 1 , 2 1 ) ∩ B X ( x 2 , 2 1 ) .
Choose
ε = 1 20 , δ = 1 5 , 1 2 + 3 2 ε = 23 40 , \varepsilon=\frac{1}{20},\qquad\delta=\frac{1}{5},\qquad\frac{1}{2}+\frac{3}{2}\varepsilon=\frac{23}{40}, ε = 20 1 , δ = 5 1 , 2 1 + 2 3 ε = 40 23 ,
and let z ∈ B X ( x 1 , 23 40 ) ∩ B X ( x 2 , 23 40 ) z\in B_X(x_1,\frac{23}{40})\cap B_X(x_2,\frac{23}{40}) z ∈ B X ( x 1 , 40 23 ) ∩ B X ( x 2 , 40 23 ) . On B B B , the two ball conditions give
− 63 40 ≤ z ≤ − 17 40 , − 23 40 ≤ z ≤ 23 40 -\frac{63}{40}\leq z\leq-\frac{17}{40},\qquad-\frac{23}{40}\leq z\leq\frac{23}{40} − 40 63 ≤ z ≤ − 40 17 , − 40 23 ≤ z ≤ 40 23
almost everywhere.
The same inequalities, with their roles interchanged, hold on C C C . Consequently,
− 23 40 ≤ z ≤ − 17 40 -\frac{23}{40}\leq z\leq-\frac{17}{40} − 40 23 ≤ z ≤ − 40 17
almost everywhere on B ∪ C B\cup C B ∪ C .
As z ∈ X z\in X z ∈ X and ∣ B ∣ = ∣ C ∣ = 1 |B|=|C|=1 ∣ B ∣ = ∣ C ∣ = 1 , it follows that
∫ A z = − ∫ B ∪ C z ≥ 17 20 . \int_A z=-\int_{B\cup C}z\geq\frac{17}{20}. ∫ A z = − ∫ B ∪ C z ≥ 20 17 .
Since ∣ A ∣ = 1 |A|=1 ∣ A ∣ = 1 and x 1 = 1 x_1=1 x 1 = 1 on A A A , we obtain
∥ z + x 1 ∥ ∞ ≥ ∫ A ( z + x 1 ) = 1 + ∫ A z ≥ 37 20 > 9 5 = 2 − δ . (30) \|z+x_1\|_\infty\geq\int_A(z+x_1)=1+\int_Az\geq\frac{37}{20}>\frac{9}{5}=2-\delta. \tag*{(30)} ∥ z + x 1 ∥ ∞ ≥ ∫ A ( z + x 1 ) = 1 + ∫ A z ≥ 20 37 > 5 9 = 2 − δ . (30)
Thus the geometric hypothesis of Lemma 9.2 is satisfied. Consider the functional φ ∈ X ∗ \varphi\in X^* φ ∈ X ∗ defined by φ ( f ) = − ∫ A f \varphi(f)=-\int_A f φ ( f ) = − ∫ A f . Clearly, ∥ φ ∥ ≤ 1 \|\varphi\|\leq1 ∥ φ ∥ ≤ 1 . Since ( − 1 , 1 2 , 1 2 ) ∈ S X (-1,\frac{1}{2},\frac{1}{2})\in S_X ( − 1 , 2 1 , 2 1 ) ∈ S X and φ ( − 1 , 1 2 , 1 2 ) = 1 \varphi(-1,\frac{1}{2},\frac{1}{2})=1 φ ( − 1 , 2 1 , 2 1 ) = 1 , we have ∥ φ ∥ = 1 \|\varphi\|=1 ∥ φ ∥ = 1 . Therefore
S = { s ∈ B X : φ ( s ) > 1 2 } = { s ∈ B X : ∫ A s < − 1 2 } S=\left\{s\in B_X:\varphi(s)>\frac{1}{2}\right\}
=\left\{s\in B_X:\int_A s<-\frac{1}{2}\right\} S = { s ∈ B X : φ ( s ) > 2 1 } = { s ∈ B X : ∫ A s < − 2 1 }
is a nonempty slice of B X B_X B X .
Fix s ∈ S s\in S s ∈ S , put
a = ∫ A s ∈ [ − 1 , − 1 2 ) , c = 1 2 ( 1 − a ) ∈ [ 1 4 , 1 3 ) , a=\int_A s\in\left[-1,-\frac{1}{2}\right),\qquad c=\frac{1}{2(1-a)}\in\left[\frac{1}{4},\frac{1}{3}\right), a = ∫ A s ∈ [ − 1 , − 2 1 ) , c = 2 ( 1 − a ) 1 ∈ [ 4 1 , 3 1 ) ,
and define
u ( t ) = { 1 2 + c ( 1 − s ( t ) ) , t ∈ A , − 1 2 , t ∈ B ∪ C . u(t)=
\begin{cases}
\frac{1}{2}+c(1-s(t)), & t\in A,\\
-\frac{1}{2}, & t\in B\cup C.
\end{cases} u ( t ) = { 2 1 + c ( 1 − s ( t )) , − 2 1 , t ∈ A , t ∈ B ∪ C .
We have
∫ A u = 1 2 + c ( 1 − a ) = 1 , ∫ B ∪ C u = − 1. \int_A u=\frac{1}{2}+c(1-a)=1,\qquad\int_{B\cup C}u=-1. ∫ A u = 2 1 + c ( 1 − a ) = 1 , ∫ B ∪ C u = − 1.
Hence u ∈ X u\in X u ∈ X . Moreover, since − 1 ≤ s ≤ 1 -1\le s\le1 − 1 ≤ s ≤ 1 almost everywhere,
1 2 ≤ u ≤ 1 2 + 2 c < 7 6 < 3 2 \frac{1}{2}\le u\le\frac{1}{2}+2c<\frac{7}{6}<\frac{3}{2} 2 1 ≤ u ≤ 2 1 + 2 c < 6 7 < 2 3
almost everywhere on A A A .
On A A A both x 1 x_1 x 1 and x 2 x_2 x 2 equal 1 1 1 , while on B ∪ C B\cup C B ∪ C the function u u u equals − 1 2 -\frac{1}{2} − 2 1 and each of x 1 , x 2 x_1,x_2 x 1 , x 2 takes values in { − 1 , 0 } \{-1,0\} { − 1 , 0 } . Thus
∥ u − x 1 ∥ ∞ ≤ 1 2 , ∥ u − x 2 ∥ ∞ ≤ 1 2 , \lVert u-x_1\rVert_\infty\le\frac{1}{2},\qquad\lVert u-x_2\rVert_\infty\le\frac{1}{2}, ∥ u − x 1 ∥ ∞ ≤ 2 1 , ∥ u − x 2 ∥ ∞ ≤ 2 1 ,
so u ∈ B X ( x 1 , 1 2 ) ∩ B X ( x 2 , 1 2 ) u\in B_X(x_1,\frac{1}{2})\cap B_X(x_2,\frac{1}{2}) u ∈ B X ( x 1 , 2 1 ) ∩ B X ( x 2 , 2 1 ) .
On A A A we have
s + u = 1 2 + c + ( 1 − c ) s . s+u=\frac{1}{2}+c+(1-c)s. s + u = 2 1 + c + ( 1 − c ) s .
Since 1 − c > 0 1-c>0 1 − c > 0 and − 1 ≤ s ≤ 1 -1\le s\le1 − 1 ≤ s ≤ 1 almost everywhere, this gives
− 1 2 + 2 c ≤ s + u ≤ 3 2 . -\frac{1}{2}+2c\le s+u\le\frac{3}{2}. − 2 1 + 2 c ≤ s + u ≤ 2 3 .
The lower bound is nonnegative because c ≥ 1 4 c\ge\frac{1}{4} c ≥ 4 1 . On B ∪ C B\cup C B ∪ C , we have ∣ s + u ∣ = ∣ s − 1 2 ∣ ≤ 3 2 |s+u|=|s-\frac{1}{2}|\le\frac{3}{2} ∣ s + u ∣ = ∣ s − 2 1 ∣ ≤ 2 3 . Consequently,
∥ s + u ∥ ∞ ≤ 3 2 . (31) \lVert s+u\rVert_\infty\le\frac{3}{2}. \tag*{(31)} ∥ s + u ∥ ∞ ≤ 2 3 . (31)
Suppose now that X X X has the WODP. By (30) , Lemma 9.2 provides an s ∈ S s\in S s ∈ S such that
( 1 + ε ) ∥ s + u ∥ ∞ > 2 − δ − ε = 7 4 (1+\varepsilon)\lVert s+u\rVert_\infty>2-\delta-\varepsilon=\frac{7}{4} ( 1 + ε ) ∥ s + u ∥ ∞ > 2 − δ − ε = 4 7
for every u ∈ B X ( x 1 , 1 2 ) ∩ B X ( x 2 , 1 2 ) u\in B_X(x_1,\frac{1}{2})\cap B_X(x_2,\frac{1}{2}) u ∈ B X ( x 1 , 2 1 ) ∩ B X ( x 2 , 2 1 ) . For this s s s , however, the preceding construction gives an element u u u in the same intersection satisfying (9.2). Therefore
( 1 + ε ) ∥ s + u ∥ ∞ ≤ 21 20 ⋅ 3 2 = 63 40 < 7 4 , (1+\varepsilon)\lVert s+u\rVert_\infty\le\frac{21}{20}\cdot\frac{3}{2}=\frac{63}{40}<\frac{7}{4}, ( 1 + ε ) ∥ s + u ∥ ∞ ≤ 20 21 ⋅ 2 3 = 40 63 < 4 7 ,
a contradiction. □
Isomorphic result
For the main result of this section, we need the following theorem from [1 ] .
Theorem 10.1. Suppose that E E E has the Daugavet property and F F F is a closed subspace of E E E . If the quotient E / F E/F E / F has the Radon–Nikodým property, then F F F also has the Daugavet property with its inherited norm.
Theorem 10.2. Let X X X be a nonzero Banach space and put
κ = max { dens ( X ) , ℵ 0 } , \kappa=\max\{\operatorname{dens}(X),\aleph_0\}, κ = max { dens ( X ) , ℵ 0 } , where dens ( X ) \operatorname{dens}(X) dens ( X ) denotes the density character of X X X . Let Γ \Gamma Γ be a set such that ∣ Γ ∣ > κ |\Gamma|>\kappa ∣Γ∣ > κ . Then the Banach space Y = X ⊕ 2 ℓ 2 ( Γ ) Y=X\oplus_2\ell_2(\Gamma) Y = X ⊕ 2 ℓ 2 ( Γ ) has the following properties:
X X X embeds linearly and isometrically into Y Y Y .
No equivalent norm on Y Y Y has the Daugavet property.
Consequently, Y Y Y is not isomorphic to any Banach space with the Daugavet property.
Proof. Equip Y = X ⊕ 2 ℓ 2 ( Γ ) Y = X \oplus_{2} \ell_{2}(\Gamma) Y = X ⊕ 2 ℓ 2 ( Γ ) with the norm
∥ ( x , h ) ∥ 0 = ( ∥ x ∥ 2 2 + ∥ h ∥ 2 2 ) 1 / 2 . \lVert(x,h)\rVert_{0} = \left(\lVert x\rVert_{2}^{2}+\lVert h\rVert_{2}^{2}\right)^{1/2}. ∥( x , h ) ∥ 0 = ( ∥ x ∥ 2 2 + ∥ h ∥ 2 2 ) 1/2 .
The canonical map
J : X ⟶ Y , J ( x ) = ( x , 0 ) , J:X\longrightarrow Y,\qquad J(x)=(x,0), J : X ⟶ Y , J ( x ) = ( x , 0 ) ,
is a linear isometry. We identify X X X with J ( X ) J(X) J ( X ) throughout the proof.
Suppose, towards a contradiction, that p p p is an equivalent norm on Y Y Y for which ( Y , p ) (Y,p) ( Y , p ) has the Daugavet property. Since equivalent norms induce the same topology, the unit sphere of ( X , p ∣ X ) (X,p|_{X}) ( X , p ∣ X ) has a dense subset D D D with ∣ D ∣ ≤ κ |D|\leq\kappa ∣ D ∣ ≤ κ . For each d ∈ D d\in D d ∈ D , the Hahn–Banach theorem gives F d ∈ ( Y , p ) ∗ F_{d}\in(Y,p)^{*} F d ∈ ( Y , p ) ∗ such that
∥ F d ∥ p , ∗ = 1 , F d ( d ) = 1. \lVert F_{d}\rVert_{p,*}=1,\qquad F_{d}(d)=1. ∥ F d ∥ p , ∗ = 1 , F d ( d ) = 1.
This family is norming for ( X , p ∣ X ) (X,p|_{X}) ( X , p ∣ X ) ; that is,
p ( x ) = sup d ∈ D ∣ F d ( x ) ∣ ( x ∈ X ) . (32) p(x)=\sup_{d\in D}|F_{d}(x)|\qquad(x\in X). \tag*{(32)} p ( x ) = d ∈ D sup ∣ F d ( x ) ∣ ( x ∈ X ) . (32)
Indeed, ∣ F d ( x ) ∣ ≤ p ( x ) |F_{d}(x)|\leq p(x) ∣ F d ( x ) ∣ ≤ p ( x ) for every d ∈ D d\in D d ∈ D and x ∈ X x\in X x ∈ X . Conversely, if p ( x ) = 1 p(x)=1 p ( x ) = 1 and ε > 0 \varepsilon>0 ε > 0 , choose d ∈ D d\in D d ∈ D with p ( x − d ) < ε p(x-d)<\varepsilon p ( x − d ) < ε . Then
∣ F d ( x ) ∣ ≥ ∣ F d ( d ) ∣ − ∣ F d ( x − d ) ∣ ≥ 1 − p ( x − d ) > 1 − ε . |F_{d}(x)|\geq|F_{d}(d)|-|F_{d}(x-d)|\geq1-p(x-d)>1-\varepsilon. ∣ F d ( x ) ∣ ≥ ∣ F d ( d ) ∣ − ∣ F d ( x − d ) ∣ ≥ 1 − p ( x − d ) > 1 − ε .
Letting ε → 0 \varepsilon\to0 ε → 0 gives (10.1) for unit vectors, and the general case follows by homogeneity.
Identify ℓ 2 ( Γ ) \ell_{2}(\Gamma) ℓ 2 ( Γ ) with the closed subspace H = { 0 } ⊕ ℓ 2 ( Γ ) H=\{0\}\oplus\ell_{2}(\Gamma) H = { 0 } ⊕ ℓ 2 ( Γ ) of Y Y Y . Since p p p is equivalent to ∥ ⋅ ∥ 0 \lVert\cdot\rVert_{0} ∥ ⋅ ∥ 0 , each restriction F d ∣ H F_{d}|_{H} F d ∣ H is continuous for the original Hilbert space norm. By the Riesz representation theorem, it is represented by a vector a d ∈ ℓ 2 ( Γ ) a_{d}\in\ell_{2}(\Gamma) a d ∈ ℓ 2 ( Γ ) . Every such vector has at most countable support, so
S = ⋃ d ∈ D supp ( a d ) satisfies ∣ S ∣ ≤ ∣ D ∣ ⋅ ℵ 0 ≤ κ . S=\bigcup_{d\in D}\operatorname{supp}(a_{d})\qquad\text{satisfies}\qquad|S|\leq|D|\cdot\aleph_{0}\leq\kappa. S = d ∈ D ⋃ supp ( a d ) satisfies ∣ S ∣ ≤ ∣ D ∣ ⋅ ℵ 0 ≤ κ .
As ∣ Γ ∣ > κ |\Gamma|>\kappa ∣Γ∣ > κ , we can choose γ 0 ∈ Γ ∖ S \gamma_{0}\in\Gamma\setminus S γ 0 ∈ Γ ∖ S . Set e = ( 0 , e γ 0 ) e=(0,e_{\gamma_{0}}) e = ( 0 , e γ 0 ) , where e γ 0 e_{\gamma_{0}} e γ 0 is the corresponding canonical unit vector of ℓ 2 ( Γ ) \ell_{2}(\Gamma) ℓ 2 ( Γ ) . By construction,
F d ( e ) = 0 ( d ∈ D ) . F_{d}(e)=0\qquad(d\in D). F d ( e ) = 0 ( d ∈ D ) .
For every x ∈ X x\in X x ∈ X and t ∈ R t\in\mathbb{R} t ∈ R , it follows from (10.1) that
p ( x ) = sup d ∈ D ∣ F d ( x ) ∣ = sup d ∈ D ∣ F d ( x + t e ) ∣ ≤ p ( x + t e ) . \begin{aligned}
p(x)&=\sup_{d\in D}|F_{d}(x)|\\
&=\sup_{d\in D}|F_{d}(x+te)|\\
&\leq p(x+te).
\end{aligned} p ( x ) = d ∈ D sup ∣ F d ( x ) ∣ = d ∈ D sup ∣ F d ( x + t e ) ∣ ≤ p ( x + t e ) .
Thus, on the closed subspace Z = X ⊕ R e Z=X\oplus\mathbb{R}e Z = X ⊕ R e , the projection
Q : Z ⟶ Z , Q ( x + t e ) = x , Q:Z\longrightarrow Z,\qquad Q(x+te)=x, Q : Z ⟶ Z , Q ( x + t e ) = x ,
is contractive. Since Q Q Q is the identity on the nonzero subspace X X X , we have ∥ Q ∥ p = 1 \lVert Q\rVert_{p}=1 ∥ Q ∥ p = 1 .
The quotient Y / Z Y/Z Y / Z , equipped with the quotient norm induced by ∥ ⋅ ∥ 0 \lVert\cdot\rVert_{0} ∥ ⋅ ∥ 0 , is isometric to ℓ 2 ( Γ ∖ { γ 0 } ) \ell_{2}(\Gamma\setminus\{\gamma_{0}\}) ℓ 2 ( Γ ∖ { γ 0 }) . Equivalent norms on Y Y Y induce equivalent quotient norms, so ( Y , p ) / Z (Y,p)/Z ( Y , p ) / Z is isomorphic to this Hilbert space and is therefore reflexive. In particular, it has the Radon–Nikodým property. The hereditary theorem stated above now implies that ( Z , p ∣ Z ) (Z,p|_{Z}) ( Z , p ∣ Z ) has the Daugavet property.
Consider P = I Z − Q P=I_{Z}-Q P = I Z − Q . Since
P ( x + t e ) = t e ( x ∈ X , t ∈ R ) , P(x+te)=te\qquad(x\in X,\ t\in\mathbb{R}), P ( x + t e ) = t e ( x ∈ X , t ∈ R ) ,
the operator P P P is a nonzero rank-one projection. Moreover, ∥ P ∥ p ≥ 1 \lVert P\rVert_p \ge1 ∥ P ∥ p ≥ 1 , because P ( e ) = e P(e)=e P ( e ) = e . Applying the Daugavet equation in ( Z , p ∣ Z ) (Z,p|_Z) ( Z , p ∣ Z ) to the rank-one operator − P -P − P gives
1 = ∥ Q ∥ p = ∥ I Z − P ∥ p = 1 + ∥ P ∥ p ≥ 2 , 1=\lVert Q\rVert_p=\lVert I_Z-P\rVert_p=1+\lVert P\rVert_p\ge2, 1 = ∥ Q ∥ p = ∥ I Z − P ∥ p = 1 + ∥ P ∥ p ≥ 2 ,
a contradiction. Hence no equivalent norm on Y Y Y has the Daugavet property. □ \square □
Acknowledgments
The author used OpenAI’s GPT-6 Astra to assist with mathematical exploration, the development and checking of proofs, and the presentation of arguments, particularly those concerning injective and projective tensor products. The author takes full responsibility for the mathematical content and any remaining errors.
References
Mathematisches Seminar, Christian-Albrechts-Universität zu Kiel, Kiel, Germany
Email address: hamad@math.uni-kiel.de