Introduction
Let 1 < p < ∞ 1<p<\infty 1 < p < ∞ , let S S S be the right shift on ℓ p ( N ) \ell^p(\mathbb{N}) ℓ p ( N ) , N = { 0 , 1 , 2 , … } \mathbb{N}=\{0,1,2,\ldots\} N = { 0 , 1 , 2 , … } , given by ( S x ) n = x n − 1 (Sx)_n=x_{n-1} ( S x ) n = x n − 1 with x − 1 = 0 x_{-1}=0 x − 1 = 0 , and let T T T be a contraction on an L p L^p L p space (∥ T ∥ ≤ 1 \|T\|\le1 ∥ T ∥ ≤ 1 ). The conjecture of V. I. Matsaev appears in Nikol’skii’s 1971 list of problems [9 Problem V] ; some authors date it to 1966 [13 ] . It states that
∥ P ( T ) ∥ L p → L p ≤ ∥ P ( S ) ∥ ℓ p → ℓ p for every polynomial P ( z ) = ∑ k = 0 deg P c k z k , c k ∈ C . (1) \|P(T)\|_{L^p\to L^p}\le\|P(S)\|_{\ell^p\to\ell^p}\qquad\text{for every polynomial }P(z)=\sum_{k=0}^{\deg P}c_kz^k,\quad c_k\in\mathbb{C}. \tag*{(1)} ∥ P ( T ) ∥ L p → L p ≤ ∥ P ( S ) ∥ ℓ p → ℓ p for every polynomial P ( z ) = k = 0 ∑ d e g P c k z k , c k ∈ C . (1)
We show that (1) fails for every p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) other than 2, with T T T a real 2 × 2 2\times2 2 × 2 matrix and P P P a real quadratic (Theorem A), and that it holds for every quadratic P P P with ∣ c 0 ∣ = ∣ c 2 ∣ |c_0|=|c_2| ∣ c 0 ∣ = ∣ c 2 ∣ (Theorem B).
Notation. We write ℓ 2 p \ell^p_2 ℓ 2 p for C 2 \mathbb{C}^2 C 2 (or R 2 \mathbb{R}^2 R 2 ) with the norm ∥ ξ ∥ p = ( ∣ ξ 1 ∣ p + ∣ ξ 2 ∣ p ) 1 / p \|\xi\|_p=(|\xi_1|^p+|\xi_2|^p)^{1/p} ∥ ξ ∥ p = ( ∣ ξ 1 ∣ p + ∣ ξ 2 ∣ p ) 1/ p (and ∥ ξ ∥ ∞ = max ( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) \|\xi\|_\infty=\max(|\xi_1|,|\xi_2|) ∥ ξ ∥ ∞ = max ( ∣ ξ 1 ∣ , ∣ ξ 2 ∣ ) ); it is the L p L^p L p space of a measure space with two atoms of mass one. For p = 1 p=1 p = 1 and p = ∞ p=\infty p = ∞ , S S S is the right shift on ℓ p ( N ) \ell^p(\mathbb{N}) ℓ p ( N ) given by the same formula as above. For an operator A A A on an L p L^p L p space (in particular on ℓ 2 p \ell^p_2 ℓ 2 p or ℓ p \ell^p ℓ p ), ∥ A ∥ p \|A\|_p ∥ A ∥ p denotes its operator norm (so (1) reads ∥ P ( T ) ∥ p ≤ ∥ P ( S ) ∥ p \|P(T)\|_p\le\|P(S)\|_p ∥ P ( T ) ∥ p ≤ ∥ P ( S ) ∥ p ), and I I I is the identity; ℓ p \ell^p ℓ p stands for ℓ p ( N ) \ell^p(\mathbb{N}) ℓ p ( N ) . We also write J ( ξ 1 , ξ 2 ) = ( ξ 2 , − ξ 1 ) J(\xi_1,\xi_2)=(\xi_2,-\xi_1) J ( ξ 1 , ξ 2 ) = ( ξ 2 , − ξ 1 ) , the clockwise rotation by a right angle, and R θ = exp ( θ J ) = cos θ I + sin θ J R_\theta=\exp(\theta J)=\cos\theta I+\sin\theta J R θ = exp ( θ J ) = cos θ I + sin θ J (as J 2 = − I J^2=-I J 2 = − I ), the clockwise rotation by the angle θ \theta θ . The table below compares our two main results with what was known before, and Figure 1 shows the conjecture for each p p p before and after this paper.
Figure 1. Matsaev’s conjecture for each p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) , placed by 1 / p 1/p 1/ p , so that p p p and its dual exponent p ′ = p / ( p − 1 ) p'=p/(p-1) p ′ = p / ( p − 1 ) lie symmetrically about p = 2 p=2 p = 2 . Top, before this paper: true for p = 2 p=2 p = 2 (von Neumann’s inequality), false for p = 4 p=4 p = 4 and p = 4 / 3 p=4/3 p = 4/3 (Drury [4 ] ), and not known otherwise (hatched). Bottom, this paper: false for every p ≠ 2 p\ne2 p = 2 (Theorem A). The endpoints p = ∞ p=\infty p = ∞ and p = 1 p=1 p = 1 (open circles) lie outside the conjecture; there (1) holds (see the Background).
Table 1.
Background. For p = 2 p=2 p = 2 , inequality (1) is von Neumann’s inequality. For every p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] and every contraction T T T , the triangle inequality and ∥ T k ∥ p ≤ 1 \lVert T^k\rVert_p\le1 ∥ T k ∥ p ≤ 1 give the trivial bound ∥ P ( T ) ∥ p ≤ ∑ k ∣ c k ∣ \lVert P(T)\rVert_p\le\sum_k\lvert c_k\rvert ∥ P ( T ) ∥ p ≤ ∑ k ∣ c k ∣ . For p = 1 p=1 p = 1 and p = ∞ p=\infty p = ∞ also ∥ P ( S ) ∥ p = ∑ k ∣ c k ∣ \lVert P(S)\rVert_p=\sum_k\lvert c_k\rvert ∥ P ( S ) ∥ p = ∑ k ∣ c k ∣ , so (1) holds trivially: for p = 1 p=1 p = 1 apply P ( S ) P(S) P ( S ) to the unit vector e 0 = ( 1 , 0 , 0 , … ) e_0=(1,0,0,\ldots) e 0 = ( 1 , 0 , 0 , … ) ; for p = ∞ p=\infty p = ∞ , with d = deg P d=\deg P d = deg P , take x n = c d − n / ∣ c d − n ∣ x_n=c_{d-n}/\lvert c_{d-n}\rvert x n = c d − n / ∣ c d − n ∣ for 0 ≤ n ≤ d 0\le n\le d 0 ≤ n ≤ d with c d − n ≠ 0 c_{d-n}\ne0 c d − n = 0 and x n = 0 x_n=0 x n = 0 otherwise, so that ( P ( S ) x ) d = ∑ k ∣ c k ∣ (P(S)x)_d=\sum_k\lvert c_k\rvert ( P ( S ) x ) d = ∑ k ∣ c k ∣ . The trivial bound proves (1), for a given p p p , for every polynomial with ∥ P ( S ) ∥ p = ∑ k ∣ c k ∣ \lVert P(S)\rVert_p=\sum_k\lvert c_k\rvert ∥ P ( S ) ∥ p = ∑ k ∣ c k ∣ . Every polynomial of degree at most one is of this kind. To see this, let P ( z ) = c 0 + c 1 z P(z)=c_0+c_1z P ( z ) = c 0 + c 1 z and choose ω , μ ∈ C \omega,\mu\in\mathbb{C} ω , μ ∈ C with ∣ ω ∣ = ∣ μ ∣ = 1 \lvert\omega\rvert=\lvert\mu\rvert=1 ∣ ω ∣ = ∣ μ ∣ = 1 , ω c 0 = ∣ c 0 ∣ \omega c_0=\lvert c_0\rvert ω c 0 = ∣ c 0 ∣ and ω μ c 1 = ∣ c 1 ∣ \omega\mu c_1=\lvert c_1\rvert ω μ c 1 = ∣ c 1 ∣ . The operator ω P ( μ S ) \omega P(\mu S) ω P ( μ S ) has the same norm as P ( S ) P(S) P ( S ) , because μ S = U S U − 1 \mu S=USU^{-1} μ S = U S U − 1 with the isometry ( U x ) n = μ n x n (Ux)_n=\mu^n x_n ( U x ) n = μ n x n . For x = ( 1 , … , 1 , 0 , 0 , … ) x=(1,\ldots,1,0,0,\ldots) x = ( 1 , … , 1 , 0 , 0 , … ) with N N N ones, ω P ( μ S ) x = ( ∣ c 0 ∣ , ∣ c 0 ∣ + ∣ c 1 ∣ , … , ∣ c 0 ∣ + ∣ c 1 ∣ , ∣ c 1 ∣ , 0 , 0 , … ) \omega P(\mu S)x=(\lvert c_0\rvert,\lvert c_0\rvert+\lvert c_1\rvert,\ldots,\lvert c_0\rvert+\lvert c_1\rvert,\lvert c_1\rvert,0,0,\ldots) ω P ( μ S ) x = (∣ c 0 ∣ , ∣ c 0 ∣ + ∣ c 1 ∣ , … , ∣ c 0 ∣ + ∣ c 1 ∣ , ∣ c 1 ∣ , 0 , 0 , … ) , with N − 1 N-1 N − 1 entries equal to ∣ c 0 ∣ + ∣ c 1 ∣ \lvert c_0\rvert+\lvert c_1\rvert ∣ c 0 ∣ + ∣ c 1 ∣ ; so ∥ ω P ( μ S ) x ∥ p / ∥ x ∥ p → ∣ c 0 ∣ + ∣ c 1 ∣ \lVert\omega P(\mu S)x\rVert_p/\lVert x\rVert_p\to\lvert c_0\rvert+\lvert c_1\rvert ∥ ω P ( μ S ) x ∥ p / ∥ x ∥ p → ∣ c 0 ∣ + ∣ c 1 ∣ as N → ∞ N\to\infty N → ∞ , and, with the trivial bound, ∥ P ( S ) ∥ p = ∣ c 0 ∣ + ∣ c 1 ∣ \lVert P(S)\rVert_p=\lvert c_0\rvert+\lvert c_1\rvert ∥ P ( S ) ∥ p = ∣ c 0 ∣ + ∣ c 1 ∣ .
For p ≠ 2 p\ne2 p = 2 , (1) is known to hold for positive contractions (T x ≥ 0 Tx\ge0 T x ≥ 0 whenever x ≥ 0 x\ge0 x ≥ 0 ), by the dilation theorem of Akcoglu and Sucheston [1 ] combined with the transference principle of Coifman, Rochberg and Weiss [3 ] , and, more generally, for contractions T T T with a positive contractive majorant (a positive contraction T + T_+ T + with ∣ T x ∣ ≤ T + ∣ x ∣ \lvert Tx\rvert\le T_+\lvert x\rvert ∣ T x ∣ ≤ T + ∣ x ∣ for every x x x ), because such contractions have isometric dilations (Peller [12 ] ; see [2 Sec. 1] ). This class contains the isometries, the Lamperti contractions (which map functions with disjoint supports to functions with disjoint supports) and the absolute contractions (contractions on both L 1 L^1 L 1 and L ∞ L^\infty L ∞ ); for the last two see also Peller [10 , 11 ] .
Drury [4 ] found a real 2 × 2 2\times2 2 × 2 contraction and P ( z ) = 1 + 2 z − 22 5 z 2 P(z)=1+2z-\frac{22}{5}z^2 P ( z ) = 1 + 2 z − 5 22 z 2 (of the form 1 + a z − B z 2 1+az-Bz^2 1 + a z − B z 2 used below, with a = 2 a=2 a = 2 , B = 22 5 B=\frac{22}{5} B = 5 22 ) that violate (1) for p = 4 p=4 p = 4 , hence, by duality (Lemma 2.3 ), for p = 4 / 3 p=4/3 p = 4/3 ; the norms ∥ P ( T ) ∥ 4 \lVert P(T)\rVert_4 ∥ P ( T ) ∥ 4 and ∥ P ( S ) ∥ 4 \lVert P(S)\rVert_4 ∥ P ( S ) ∥ 4 were bounded with computer assistance. Fackler [5 Sec. 2.3] noted that an analytic approach was missing, and, to our knowledge, no counterexample other than Drury’s was known (see [2 Sec. 1] and [5 Sec. 2.3] ). For polynomials in several commuting contractions, the analogue of (1) is known to fail for every 1 < p < ∞ 1<p<\infty 1 < p < ∞ once the number of contractions is large enough (Ray [13 ] ); this does not decide the one-variable conjecture studied here.
Theorem A. Let p ∈ ( 1 , ∞ ) p \in(1,\infty) p ∈ ( 1 , ∞ ) , p ≠ 2 p \ne2 p = 2 , B > 1 B>1 B > 1 , and P a ( z ) = 1 + a z − B z 2 P_a(z)=1+az-Bz^2 P a ( z ) = 1 + a z − B z 2 . For every sufficiently small a > 0 a>0 a > 0 (how small depends only on p p p and B B B ) there is a real 2 × 2 2\times2 2 × 2 matrix T T T with ∥ T ∥ p ≤ 1 \|T\|_p\le1 ∥ T ∥ p ≤ 1 on ℓ 2 p \ell_2^p ℓ 2 p such that
∥ P a ( T ) ∥ p > ∥ P a ( S ) ∥ p . \|P_a(T)\|_p>\|P_a(S)\|_p. ∥ P a ( T ) ∥ p > ∥ P a ( S ) ∥ p . For p > 2 p>2 p > 2 one can take T = e − ε D ∗ / 2 R π / 2 + ε / 2 T=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2} T = e − ε D ∗ /2 R π /2 + ε /2 , for every sufficiently small ε > 0 \varepsilon>0 ε > 0 (depending on p p p , B B B and a a a ), where D ∗ = max 0 ≤ t ≤ 1 ( t − t p − 1 ) / ( 1 + t p ) > 0 D_*=\max_{0\le t\le1}(t-t^{p-1})/(1+t^p)>0 D ∗ = max 0 ≤ t ≤ 1 ( t − t p − 1 ) / ( 1 + t p ) > 0 depends only on p p p (Lemma 3.4 ). For 1 < p < 2 1<p<2 1 < p < 2 one can take the transpose of the matrix above, built for the dual exponent p ′ = p / ( p − 1 ) > 2 p'=p/(p-1)>2 p ′ = p / ( p − 1 ) > 2 in place of p p p ; since the transpose of R θ R_\theta R θ is R − θ R_{-\theta} R − θ , this is T = e − ε D ∗ ′ / 2 R − π / 2 − ε / 2 T=e^{-\varepsilon D_*'/2}R_{-\pi/2-\varepsilon/2} T = e − ε D ∗ ′ /2 R − π /2 − ε /2 , where D ∗ ′ D_*' D ∗ ′ is D ∗ D_* D ∗ computed for p ′ p' p ′ : again a scaled rotation by slightly more than a right angle, now counterclockwise.
Theorem B. Let P ( z ) = c 0 + c 1 z + c 2 z 2 P(z)=c_0+c_1z+c_2z^2 P ( z ) = c 0 + c 1 z + c 2 z 2 with complex coefficients and ∣ c 0 ∣ = ∣ c 2 ∣ |c_0|=|c_2| ∣ c 0 ∣ = ∣ c 2 ∣ . Then ∥ P ( T ) ∥ p ≤ ∥ P ( S ) ∥ p \|P(T)\|_p\le\|P(S)\|_p ∥ P ( T ) ∥ p ≤ ∥ P ( S ) ∥ p for every 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ and every contraction T T T on every L p L^p L p space.
Idea of the proof of Theorem A. Let ( x n ) n ∈ Z (x_n)_{n\in\mathbb{Z}} ( x n ) n ∈ Z be a two-sided sequence with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n for all n n n , and shift it by the same formula, ( S x ) n = x n − 1 (Sx)_n=x_{n-1} ( S x ) n = x n − 1 . Then ( ( S x ) n , ( S x ) n − 1 ) = ( x n − 1 , x n − 2 ) = ( x n − 1 , − x n ) = J ( x n , x n − 1 ) ((Sx)_n,(Sx)_{n-1})=(x_{n-1},x_{n-2})=(x_{n-1},-x_n)=J(x_n,x_{n-1}) (( S x ) n , ( S x ) n − 1 ) = ( x n − 1 , x n − 2 ) = ( x n − 1 , − x n ) = J ( x n , x n − 1 ) : on pairs of consecutive terms of such sequences, S S S acts as J J J . Truncating long pieces of such sequences gives ∥ P ( J ) ∥ p ≤ ∥ P ( S ) ∥ p \|P(J)\|_p\le\|P(S)\|_p ∥ P ( J ) ∥ p ≤ ∥ P ( S ) ∥ p for every P P P (Lemma 2.2 ). For P 0 ( z ) = 1 − B z 2 P_0(z)=1-Bz^2 P 0 ( z ) = 1 − B z 2 we have P 0 ( J ) = I − B J 2 = ( 1 + B ) I P_0(J)=I-BJ^2=(1+B)I P 0 ( J ) = I − B J 2 = ( 1 + B ) I , so 1 + B = ∥ P 0 ( J ) ∥ p ≤ ∥ P 0 ( S ) ∥ p ≤ 1 + B 1+B=\|P_0(J)\|_p\le\|P_0(S)\|_p\le1+B 1 + B = ∥ P 0 ( J ) ∥ p ≤ ∥ P 0 ( S ) ∥ p ≤ 1 + B (Lemma 2.2 and the triangle inequality); thus ∥ P 0 ( S ) ∥ p = 1 + B \|P_0(S)\|_p=1+B ∥ P 0 ( S ) ∥ p = 1 + B , and P 0 ( S ) P_0(S) P 0 ( S ) has many almost-maximising vectors: long truncations x x x of every sequence with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n give ∥ P 0 ( S ) x ∥ p / ∥ x ∥ p \|P_0(S)x\|_p/\|x\|_p ∥ P 0 ( S ) x ∥ p /∥ x ∥ p close to 1 + B 1+B 1 + B .
For p ≠ 2 p\ne2 p = 2 and small a > 0 a>0 a > 0 we prove that P a ( S ) P_a(S) P a ( S ) still has exactly the norm C a = ∥ P a ( J ) ∥ p C_a=\|P_a(J)\|_p C a = ∥ P a ( J ) ∥ p (Theorem 3.6 ). For p > 2 p>2 p > 2 the upper bound comes from a Bellman function, a function G G G of two real variables with ∣ ( P a ( S ) x ) n ∣ p ≤ C a p ∣ x n ∣ p + G ( x n − 1 , x n − 2 ) − G ( x n , x n − 1 ) |(P_a(S)x)_n|^p\le C_a^p|x_n|^p+G(x_{n-1},x_{n-2})-G(x_n,x_{n-1}) ∣ ( P a ( S ) x ) n ∣ p ≤ C a p ∣ x n ∣ p + G ( x n − 1 , x n − 2 ) − G ( x n , x n − 1 ) for every real x ∈ ℓ p ( N ) x\in\ell^p(\mathbb{N}) x ∈ ℓ p ( N ) (with x − 1 = x − 2 = 0 x_{-1}=x_{-2}=0 x − 1 = x − 2 = 0 ) and every n ≥ 0 n\ge0 n ≥ 0 ; in the sum over n n n the G G G terms cancel, so ∥ P a ( S ) ∥ p ≤ C a \|P_a(S)\|_p\le C_a ∥ P a ( S ) ∥ p ≤ C a (Lemma 3.1 ). For 1 < p < 2 1<p<2 1 < p < 2 the upper bound follows by duality (Lemma 2.3 ).
For p > 2 p>2 p > 2 we compose J J J with the rotation R ε / 2 R_{\varepsilon/2} R ε /2 and multiply by e − ε D ∗ / 2 e^{-\varepsilon D_*/2} e − ε D ∗ /2 ; this gives the matrix T ε = e − ε D ∗ / 2 R ε / 2 J = e − ε D ∗ / 2 R π / 2 + ε / 2 T_\varepsilon=e^{-\varepsilon D_*/2}R_{\varepsilon/2}J=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2} T ε = e − ε D ∗ /2 R ε /2 J = e − ε D ∗ /2 R π /2 + ε /2 of Theorem A. The factor e − ε D ∗ / 2 e^{-\varepsilon D_*/2} e − ε D ∗ /2 compensates the largest possible rate of growth of the ℓ p \ell^p ℓ p norm under the rotations R ε / 2 R_{\varepsilon/2} R ε /2 (at most D ∗ / 2 D_*/2 D ∗ /2 per unit of ε \varepsilon ε ), so that T ε T_\varepsilon T ε is still a contraction (Figure 3 ). To see this, put ϕ ( y ) = sgn y ∣ y ∣ p − 1 \phi(y)=\operatorname{sgn}y|y|^{p-1} ϕ ( y ) = sgn y ∣ y ∣ p − 1 (so that d d y ∣ y ∣ p = p ϕ ( y ) \frac{d}{dy}|y|^p=p\phi(y) d y d ∣ y ∣ p = pϕ ( y ) ) and, for real ξ ≠ 0 \xi\ne0 ξ = 0 ,
Figure 3. The real unit circle of ℓ 2 4 \ell_2^4 ℓ 2 4 (black); its image under the rotation R π / 2 + ε / 2 R_{\pi/2+\varepsilon/2} R π /2 + ε /2 (red, dashed), which leaves the unit ball; and its image under T ε = e − ε D ∗ / 2 R π / 2 + ε / 2 T_\varepsilon=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2} T ε = e − ε D ∗ /2 R π /2 + ε /2 (blue), which stays inside it (Lemma 4.1). Here ε = 0.8 \varepsilon=0.8 ε = 0.8 , larger than in Theorem A, so that the curves are easy to tell apart. Computed numerically; an illustration, not used in the proofs.
Q ( ξ ) = ϕ ( ξ 1 ) ξ 2 − ϕ ( ξ 2 ) ξ 1 ∥ ξ ∥ p p , so that Q ( 1 , t ) = t − t p − 1 1 + t p ( 0 ≤ t ≤ 1 ) . Q(\xi)=\frac{\phi(\xi_1)\xi_2-\phi(\xi_2)\xi_1}{\|\xi\|_p^p},
\qquad\text{so that}\qquad
Q(1,t)=\frac{t-t^{p-1}}{1+t^p}\quad(0\le t\le1). Q ( ξ ) = ∥ ξ ∥ p p ϕ ( ξ 1 ) ξ 2 − ϕ ( ξ 2 ) ξ 1 , so that Q ( 1 , t ) = 1 + t p t − t p − 1 ( 0 ≤ t ≤ 1 ) .
Since d d ε R ε / 2 ξ ∣ ε = 0 = 1 2 J ξ = 1 2 ( ξ 2 , − ξ 1 ) \left.\frac{d}{d\varepsilon}R_{\varepsilon/2}\xi\right|_{\varepsilon=0}=\frac{1}{2}J\xi=\frac{1}{2}(\xi_2,-\xi_1) d ε d R ε /2 ξ ε = 0 = 2 1 J ξ = 2 1 ( ξ 2 , − ξ 1 ) , the chain rule in each coordinate gives
d d ε ∥ R ε / 2 ξ ∥ p p ∣ ε = 0 = p 2 ( ϕ ( ξ 1 ) ξ 2 − ϕ ( ξ 2 ) ξ 1 ) = p 2 Q ( ξ ) ∥ ξ ∥ p p . \left.\frac{d}{d\varepsilon}\|R_{\varepsilon/2}\xi\|_p^p\right|_{\varepsilon=0}
=\frac{p}{2}\bigl(\phi(\xi_1)\xi_2-\phi(\xi_2)\xi_1\bigr)
=\frac{p}{2}Q(\xi)\|\xi\|_p^p. d ε d ∥ R ε /2 ξ ∥ p p ε = 0 = 2 p ( ϕ ( ξ 1 ) ξ 2 − ϕ ( ξ 2 ) ξ 1 ) = 2 p Q ( ξ ) ∥ ξ ∥ p p .
hence, dividing by p ∥ ξ ∥ p p − 1 p\|\xi\|_p^{p-1} p ∥ ξ ∥ p p − 1 , d d ε ∥ R ε / 2 ξ ∥ p ∣ ε = 0 = 1 2 Q ( ξ ) ∥ ξ ∥ p \left.\frac{d}{d\varepsilon}\|R_{\varepsilon/2}\xi\|_p\right|_{\varepsilon=0}=\frac{1}{2}Q(\xi)\|\xi\|_p d ε d ∥ R ε /2 ξ ∥ p ε = 0 = 2 1 Q ( ξ ) ∥ ξ ∥ p . Lemma 3.4 shows that the maximum of Q Q Q over all ξ ≠ 0 \xi\ne0 ξ = 0 is attained at ξ = ( 1 , t ∗ ) \xi=(1,t_*) ξ = ( 1 , t ∗ ) for some t ∗ ∈ ( 0 , 1 ) t_*\in(0,1) t ∗ ∈ ( 0 , 1 ) ; so max Q = D ∗ \max Q=D_* max Q = D ∗ , and the derivative above is at most D ∗ 2 ∥ ξ ∥ p \frac{D_*}{2}\|\xi\|_p 2 D ∗ ∥ ξ ∥ p , with equality at ξ = ( 1 , t ∗ ) \xi=(1,t_*) ξ = ( 1 , t ∗ ) . As R ( ε + h ) / 2 = R h / 2 R ε / 2 R_{(\varepsilon+h)/2}=R_{h/2}R_{\varepsilon/2} R ( ε + h ) /2 = R h /2 R ε /2 , the same bound holds at every ε \varepsilon ε , with ξ \xi ξ replaced by R ε / 2 ξ R_{\varepsilon/2}\xi R ε /2 ξ . Thus d d ε log ∥ R ε / 2 ∥ p ≤ D ∗ 2 \frac{d}{d\varepsilon}\log\|R_{\varepsilon/2}\|_p\le\frac{D_*}{2} d ε d log ∥ R ε /2 ∥ p ≤ 2 D ∗ , and integration from 0 0 0 to ε \varepsilon ε gives ∥ R ε / 2 ξ ∥ p ≤ e ε D ∗ / 2 ∥ ξ ∥ p \|R_{\varepsilon/2}\xi\|_p\le e^{\varepsilon D_*/2}\|\xi\|_p ∥ R ε /2 ξ ∥ p ≤ e ε D ∗ /2 ∥ ξ ∥ p for ε ≥ 0 \varepsilon\ge0 ε ≥ 0 . As J J J is an isometry of ℓ 2 p \ell_2^p ℓ 2 p (it permutes the coordinates up to sign), ∥ T ε ξ ∥ p = e − ε D ∗ / 2 ∥ R ε / 2 J ξ ∥ p ≤ ∥ J ξ ∥ p = ∥ ξ ∥ p \|T_\varepsilon\xi\|_p=e^{-\varepsilon D_*/2}\|R_{\varepsilon/2}J\xi\|_p\le\|J\xi\|_p=\|\xi\|_p ∥ T ε ξ ∥ p = e − ε D ∗ /2 ∥ R ε /2 J ξ ∥ p ≤ ∥ J ξ ∥ p = ∥ ξ ∥ p , so T ε T_\varepsilon T ε is a contraction on real vectors, and hence, by Lemma 2.1 , on ℓ 2 p \ell_2^p ℓ 2 p (see also Lemma 4.1 ).
Let ξ a \xi_a ξ a be a real vector at which ∥ P a ( J ) ξ ∥ p / ∥ ξ ∥ p \|P_a(J)\xi\|_p/\|\xi\|_p ∥ P a ( J ) ξ ∥ p /∥ ξ ∥ p is maximal, normalised by ( ξ a ) 1 = 1 (\xi_a)_1=1 ( ξ a ) 1 = 1 and close to ( 1 , t ∗ ) (1,t_*) ( 1 , t ∗ ) (such a vector exists by Lemma 3.5). The derivative in ε \varepsilon ε at ε = 0 \varepsilon=0 ε = 0 of ∥ P a ( T ε ) ξ a ∥ p \|P_a(T_\varepsilon)\xi_a\|_p ∥ P a ( T ε ) ξ a ∥ p is
1 2 p ( p − 1 ) ( 1 + B ) p − 2 E ∗ a ( B − 1 ) + o ( a ) ( a → 0 ) , E ∗ = t ∗ p − 2 ( 1 + t ∗ 2 ) 1 + t ∗ p > 0 \frac{1}{2}p(p-1)(1+B)^{p-2}E_*a(B-1)+o(a)\quad(a\to0),\qquad E_*=\frac{t_*^{p-2}(1+t_*^2)}{1+t_*^p}>0 2 1 p ( p − 1 ) ( 1 + B ) p − 2 E ∗ a ( B − 1 ) + o ( a ) ( a → 0 ) , E ∗ = 1 + t ∗ p t ∗ p − 2 ( 1 + t ∗ 2 ) > 0
(Lemma 4.1). At ε = 0 \varepsilon=0 ε = 0 we have T 0 = J T_0=J T 0 = J , so ∥ P a ( T 0 ) ξ a ∥ p = ∥ P a ( J ) ∥ p ∥ ξ a ∥ p \|P_a(T_0)\xi_a\|_p=\|P_a(J)\|_p\|\xi_a\|_p ∥ P a ( T 0 ) ξ a ∥ p = ∥ P a ( J ) ∥ p ∥ ξ a ∥ p by the choice of ξ a \xi_a ξ a . Hence, for B > 1 B>1 B > 1 and every sufficiently small a > 0 a>0 a > 0 , we get ∥ P a ( T ε ) ξ a ∥ p > ∥ P a ( J ) ∥ p ∥ ξ a ∥ p \|P_a(T_\varepsilon)\xi_a\|_p>\|P_a(J)\|_p\|\xi_a\|_p ∥ P a ( T ε ) ξ a ∥ p > ∥ P a ( J ) ∥ p ∥ ξ a ∥ p , and so ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p>\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p , for all sufficiently small ε > 0 \varepsilon>0 ε > 0 ; how small ε \varepsilon ε must be depends on a a a . The gain ∥ P a ( T ε ) ∥ p − ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p-\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p − ∥ P a ( J ) ∥ p is positive, but it is small: it tends to 0 0 0 as ε → 0 \varepsilon\to0 ε → 0 , and the argument gives no explicit lower bound for it. Therefore an upper bound ∥ P a ( S ) ∥ p ≤ ∥ P a ( J ) ∥ p + r \|P_a(S)\|_p\le\|P_a(J)\|_p+r ∥ P a ( S ) ∥ p ≤ ∥ P a ( J ) ∥ p + r with an error term r > 0 r>0 r > 0 would prove nothing unless r r r were shown to be smaller than this gain. This is why we prove the exact equality ∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p \|P_a(S)\|_p=\|P_a(J)\|_p ∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p . For B = 1 B=1 B = 1 the gain vanishes to first order, and indeed for B = 1 B=1 B = 1 inequality (1) holds (Theorem B).
Further results and relation to earlier work.
For 1 < p < ∞ 1<p<\infty 1 < p < ∞ , p ≠ 2 p\ne2 p = 2 , we show in two cases that ∥ P ( S ) ∥ p \|P(S)\|_p ∥ P ( S ) ∥ p equals the norm of an explicit 2 × 2 2\times2 2 × 2 matrix: for every quadratic with ∣ c 0 ∣ = ∣ c 2 ∣ |c_0|=|c_2| ∣ c 0 ∣ = ∣ c 2 ∣ (Corollary 5.1), and for P a ( z ) = 1 + a z − B z 2 P_a(z)=1+az-Bz^2 P a ( z ) = 1 + a z − B z 2 , B > 0 B>0 B > 0 , with 0 < a < a 0 0<a<a_0 0 < a < a 0 (Theorem 3.6; a 0 > 0 a_0>0 a 0 > 0 depends on p p p and B B B and is not explicit). This norm can be strictly smaller than ∑ ∣ c k ∣ \sum|c_k| ∑ ∣ c k ∣ (Remark 5.2). To our knowledge, these are the first exact formulas for ∥ P ( S ) ∥ p \|P(S)\|_p ∥ P ( S ) ∥ p , p ≠ 2 p\ne2 p = 2 , other than the trivial formula ∥ P ( S ) ∥ p = ∑ ∣ c k ∣ \|P(S)\|_p=\sum|c_k| ∥ P ( S ) ∥ p = ∑ ∣ c k ∣ ; for real P P P with nonzero coefficients, the polynomials with ∥ P ( S ) ∥ p = ∑ ∣ c k ∣ \|P(S)\|_p=\sum|c_k| ∥ P ( S ) ∥ p = ∑ ∣ c k ∣ are characterised in [2 Prop. 6.1] .
An expanded version of this paper, with many other results, is posted on Zenodo [6 ] . It gives explicit counterexamples for every p ∈ [ 40 / 39 , 40 ] p\in[40/39,40] p ∈ [ 40/39 , 40 ] , p ≠ 2 p\ne2 p = 2 (previously, to our knowledge, the only explicit counterexample was Drury’s, for p = 4 p=4 p = 4 and p = 4 / 3 p=4/3 p = 4/3 [4 ] ). It also gives, with a proof by hand, an explicit counterexample in which the polynomial and the rotation angle do not depend on p p p : P ( z ) = 1 + 2 z − 6 5 z 2 P(z)=1+2z-\frac{6}{5}z^2 P ( z ) = 1 + 2 z − 5 6 z 2 and T = R θ / ∥ R θ ∥ p T=R_\theta/\|R_\theta\|_p T = R θ /∥ R θ ∥ p , where θ ∈ ( π / 2 , π ) \theta\in(\pi/2,\pi) θ ∈ ( π /2 , π ) and cos θ = − 1 12 \cos\theta=-\frac{1}{12} cos θ = − 12 1 , violate (1) for every p ≠ 2 p\ne2 p = 2 close enough to 2. These results are not used here.
Basic facts
As in the Introduction, for a real number y y y put ϕ ( y ) = sgn y ∣ y ∣ p − 1 \phi(y)=\operatorname{sgn}y|y|^{p-1} ϕ ( y ) = sgn y ∣ y ∣ p − 1 , so that d d y ∣ y ∣ p = p ϕ ( y ) \frac{d}{dy}|y|^p=p\phi(y) d y d ∣ y ∣ p = pϕ ( y ) for p > 1 p>1 p > 1 ; on real ξ ∈ ℓ 2 p \xi\in\ell_2^p ξ ∈ ℓ 2 p , ϕ \phi ϕ and ∣ ⋅ ∣ p − 2 |\cdot|^{p-2} ∣ ⋅ ∣ p − 2 act coordinatewise, and a product of two vectors, such as ∣ ξ ∣ p − 2 ψ |\xi|^{p-2}\psi ∣ ξ ∣ p − 2 ψ , is taken coordinatewise. With ⟨ ξ , ψ ⟩ = ξ 1 ψ 1 + ξ 2 ψ 2 \langle\xi,\psi\rangle=\xi_1\psi_1+\xi_2\psi_2 ⟨ ξ , ψ ⟩ = ξ 1 ψ 1 + ξ 2 ψ 2 , d d h ∥ ξ + h ψ ∥ p p ∣ h = 0 = p ⟨ ϕ ( ξ ) , ψ ⟩ \frac{d}{dh}\|\xi+h\psi\|_p^p\big|_{h=0}=p\langle\phi(\xi),\psi\rangle d h d ∥ ξ + h ψ ∥ p p h = 0 = p ⟨ ϕ ( ξ ) , ψ ⟩ for real ξ , ψ \xi,\psi ξ , ψ and p > 1 p>1 p > 1 .
Lemma 2.1 (real and complex norms). Let 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ . A real 2 × 2 2\times2 2 × 2 matrix has the same norm on the real and on the complex ℓ 2 p \ell_2^p ℓ 2 p , and for a real polynomial P P P the operator P ( S ) P(S) P ( S ) has the same norm on the real and on the complex ℓ p \ell^p ℓ p .
Proof. This is the Marcinkiewicz–Zygmund theorem [8 ] (see also [7 ] ). For a complex number z = y 1 + i y 2 z=y_1+iy_2 z = y 1 + i y 2 we have y 1 cos θ + y 2 sin θ = ∣ z ∣ cos ( θ − arg z ) y_1\cos\theta+y_2\sin\theta=|z|\cos(\theta-\arg z) y 1 cos θ + y 2 sin θ = ∣ z ∣ cos ( θ − arg z ) , so ∫ 0 2 π ∣ y 1 cos θ + y 2 sin θ ∣ p d θ = β p ∣ z ∣ p \int_0^{2\pi}|y_1\cos\theta+y_2\sin\theta|^p\,d\theta=\beta_p|z|^p ∫ 0 2 π ∣ y 1 cos θ + y 2 sin θ ∣ p d θ = β p ∣ z ∣ p with β p = ∫ 0 2 π ∣ cos θ ∣ p d θ \beta_p=\int_0^{2\pi}|\cos\theta|^p\,d\theta β p = ∫ 0 2 π ∣ cos θ ∣ p d θ . Let A A A be a real matrix (a 2 × 2 2\times2 2 × 2 matrix, or the infinite matrix of P ( S ) P(S) P ( S ) ), and let x = u + i v x=u+iv x = u + i v with real vectors u , v u,v u , v . Applying this identity to each coordinate of A x = A u + i A v Ax=Au+iAv A x = A u + i A v gives ∥ A x ∥ p p = β p − 1 ∫ 0 2 π ∥ A ( u cos θ + v sin θ ) ∥ p p d θ \|Ax\|_p^p=\beta_p^{-1}\int_0^{2\pi}\|A(u\cos\theta+v\sin\theta)\|_p^p\,d\theta ∥ A x ∥ p p = β p − 1 ∫ 0 2 π ∥ A ( u cos θ + v sin θ ) ∥ p p d θ , and the same identity for A = I A=I A = I gives ∥ x ∥ p p = β p − 1 ∫ 0 2 π ∥ u cos θ + v sin θ ∥ p p d θ \|x\|_p^p=\beta_p^{-1}\int_0^{2\pi}\|u\cos\theta+v\sin\theta\|_p^p\,d\theta ∥ x ∥ p p = β p − 1 ∫ 0 2 π ∥ u cos θ + v sin θ ∥ p p d θ . Write ∥ A ∥ r e a l \|A\|_{\mathrm{real}} ∥ A ∥ real for the norm of A A A on real vectors. Since ∥ A ( u cos θ + v sin θ ) ∥ p ≤ ∥ A ∥ r e a l ∥ u cos θ + v sin θ ∥ p \|A(u\cos\theta+v\sin\theta)\|_p\le\|A\|_{\mathrm{real}}\|u\cos\theta+v\sin\theta\|_p ∥ A ( u cos θ + v sin θ ) ∥ p ≤ ∥ A ∥ real ∥ u cos θ + v sin θ ∥ p for every θ \theta θ , we get ∥ A x ∥ p ≤ ∥ A ∥ r e a l ∥ x ∥ p \|Ax\|_p\le\|A\|_{\mathrm{real}}\|x\|_p ∥ A x ∥ p ≤ ∥ A ∥ real ∥ x ∥ p : the complex norm of A A A is at most its real norm. The reverse inequality is clear. □ \square □
Lemma 2.2 (a lower bound for ∥ P ( S ) ∥ p \|P(S)\|_p ∥ P ( S ) ∥ p ). For every polynomial P P P and 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ , ∥ P ( J ) ∥ p ≤ ∥ P ( S ) ∥ p \|P(J)\|_p\le\|P(S)\|_p ∥ P ( J ) ∥ p ≤ ∥ P ( S ) ∥ p .
Proof. Let ξ ∈ C 2 \xi\in\mathbb{C}^{2} ξ ∈ C 2 and let x x x be the sequence with ( x n , x n − 1 ) = J − n ξ (x_n,x_{n-1})=J^{-n}\xi ( x n , x n − 1 ) = J − n ξ for n ∈ Z n\in\mathbb{Z} n ∈ Z . This is consistent: if J − n − 1 ξ = ( b 1 , b 2 ) J^{-n-1}\xi=(b_1,b_2) J − n − 1 ξ = ( b 1 , b 2 ) , then J − n ξ = J ( b 1 , b 2 ) = ( b 2 , − b 1 ) J^{-n}\xi=J(b_1,b_2)=(b_2,-b_1) J − n ξ = J ( b 1 , b 2 ) = ( b 2 , − b 1 ) , so both rules give x n = b 2 x_n=b_2 x n = b 2 . Then x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n (as J − 2 = − I J^{-2}=-I J − 2 = − I ), and with X n = ( x n , x n − 1 ) X_n=(x_n,x_{n-1}) X n = ( x n , x n − 1 ) we have X n − k = J k X n X_{n-k}=J^kX_n X n − k = J k X n , so y n : = ∑ k c k x n − k = ( P ( J ) X n ) 1 y_n:=\sum_k c_kx_{n-k}=(P(J)X_n)_1 y n := ∑ k c k x n − k = ( P ( J ) X n ) 1 . Since X n + 1 = J − 1 X n = − J X n X_{n+1}=J^{-1}X_n=-JX_n X n + 1 = J − 1 X n = − J X n and P ( J ) P(J) P ( J ) commutes with J J J , ( P ( J ) X n + 1 ) 1 = − ( P ( J ) X n ) 2 (P(J)X_{n+1})_1=-(P(J)X_n)_2 ( P ( J ) X n + 1 ) 1 = − ( P ( J ) X n ) 2 . Hence, with Y n = P ( J ) X n Y_n=P(J)X_n Y n = P ( J ) X n (so Y n + 1 = − J Y n Y_{n+1}=-JY_n Y n + 1 = − J Y n and Y n + 2 = − Y n Y_{n+2}=-Y_n Y n + 2 = − Y n ), the numbers ∣ y n ∣ , ∣ y n + 1 ∣ , ∣ y n + 2 ∣ , ∣ y n + 3 ∣ |y_n|,|y_{n+1}|,|y_{n+2}|,|y_{n+3}| ∣ y n ∣ , ∣ y n + 1 ∣ , ∣ y n + 2 ∣ , ∣ y n + 3 ∣ are ∣ ( Y n ) 1 ∣ , ∣ ( Y n ) 2 ∣ , ∣ ( Y n ) 1 ∣ , ∣ ( Y n ) 2 ∣ |(Y_n)_1|,|(Y_n)_2|,|(Y_n)_1|,|(Y_n)_2| ∣ ( Y n ) 1 ∣ , ∣ ( Y n ) 2 ∣ , ∣ ( Y n ) 1 ∣ , ∣ ( Y n ) 2 ∣ ; since ∥ Y n ∥ p = ∥ J − n P ( J ) ξ ∥ p = ∥ P ( J ) ξ ∥ p \|Y_n\|_p=\|J^{-n}P(J)\xi\|_p=\|P(J)\xi\|_p ∥ Y n ∥ p = ∥ J − n P ( J ) ξ ∥ p = ∥ P ( J ) ξ ∥ p (J J J is an isometry of ℓ 2 p \ell_2^p ℓ 2 p ), over any four consecutive indices ∑ ∣ y n ∣ p = 2 ∥ P ( J ) ξ ∥ p p \sum|y_n|^p=2\|P(J)\xi\|_p^p ∑ ∣ y n ∣ p = 2∥ P ( J ) ξ ∥ p p , and likewise ∑ ∣ x n ∣ p = 2 ∥ ξ ∥ p p \sum|x_n|^p=2\|\xi\|_p^p ∑ ∣ x n ∣ p = 2∥ ξ ∥ p p . For j ≥ 1 j\ge1 j ≥ 1 let x ( j ) ∈ ℓ p ( N ) x^{(j)}\in\ell^p(\mathbb{N}) x ( j ) ∈ ℓ p ( N ) agree with x x x for 0 ≤ n < 4 j 0\le n<4j 0 ≤ n < 4 j and vanish elsewhere. Then ∥ x ( j ) ∥ p = 2 j ∥ ξ ∥ p \|x^{(j)}\|_p=2j\|\xi\|_p ∥ x ( j ) ∥ p = 2 j ∥ ξ ∥ p , and ( P ( S ) x ( j ) ) n = y n (P(S)x^{(j)})_n=y_n ( P ( S ) x ( j ) ) n = y n for deg P ≤ n < 4 j \deg P\le n<4j deg P ≤ n < 4 j , while ( P ( S ) x ( j ) ) n = 0 (P(S)x^{(j)})_n=0 ( P ( S ) x ( j ) ) n = 0 for n ≥ 4 j + deg P n\ge4j+\deg P n ≥ 4 j + deg P ; so ∥ P ( S ) x ( j ) ∥ p p = 2 j ∥ P ( J ) ξ ∥ p p + O ( 1 ) \|P(S)x^{(j)}\|_p^p=2j\|P(J)\xi\|_p^p+O(1) ∥ P ( S ) x ( j ) ∥ p p = 2 j ∥ P ( J ) ξ ∥ p p + O ( 1 ) , and the ratio ∥ P ( S ) x ( j ) ∥ p / ∥ x ( j ) ∥ p \|P(S)x^{(j)}\|_p/\|x^{(j)}\|_p ∥ P ( S ) x ( j ) ∥ p /∥ x ( j ) ∥ p tends to ∥ P ( J ) ξ ∥ p / ∥ ξ ∥ p \|P(J)\xi\|_p/\|\xi\|_p ∥ P ( J ) ξ ∥ p /∥ ξ ∥ p as j → ∞ j\to\infty j → ∞ (the sums ∑ n ≥ 0 ∣ ( P ( S ) x ( j ) ) n ∣ p \sum_{n\ge0}|(P(S)x^{(j)})_n|^p ∑ n ≥ 0 ∣ ( P ( S ) x ( j ) ) n ∣ p and ∑ 0 ≤ n < 4 j ∣ y n ∣ p = 2 j ∥ P ( J ) ξ ∥ p p \sum_{0\le n<4j}|y_n|^p=2j\|P(J)\xi\|_p^p ∑ 0 ≤ n < 4 j ∣ y n ∣ p = 2 j ∥ P ( J ) ξ ∥ p p differ only in the at most 2 deg P 2\deg P 2 deg P terms with n < deg P n<\deg P n < deg P or 4 j ≤ n < 4 j + deg P 4j\le n<4j+\deg P 4 j ≤ n < 4 j + deg P , and in these terms ∣ ( P ( S ) x ( j ) ) n ∣ |(P(S)x^{(j)})_n| ∣ ( P ( S ) x ( j ) ) n ∣ and ∣ y n ∣ |y_n| ∣ y n ∣ are at most ∑ k ∣ c k ∣ max m ∣ x m ∣ \sum_k|c_k|\max_m|x_m| ∑ k ∣ c k ∣ max m ∣ x m ∣ ). Hence ∥ P ( S ) ∥ p ≥ ∥ P ( J ) ξ ∥ p / ∥ ξ ∥ p \|P(S)\|_p\ge\|P(J)\xi\|_p/\|\xi\|_p ∥ P ( S ) ∥ p ≥ ∥ P ( J ) ξ ∥ p /∥ ξ ∥ p for every ξ ≠ 0 \xi\ne0 ξ = 0 , which is the claim. For p = ∞ p=\infty p = ∞ the inequality is trivial: ∥ P ( J ) ∥ ∞ ≤ ∑ ∣ c k ∣ = ∥ P ( S ) ∥ ∞ \|P(J)\|_\infty\le\sum|c_k|=\|P(S)\|_\infty ∥ P ( J ) ∥ ∞ ≤ ∑ ∣ c k ∣ = ∥ P ( S ) ∥ ∞ . □
For 1 < p < ∞ 1<p<\infty 1 < p < ∞ , Lemma 2.2 also follows from Peller’s results [12 ] , as J J J is an isometry of ℓ 2 p \ell_2^p ℓ 2 p ; we give the direct proof because its sequences with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n reappear in Section 3 (Lemma 3.2).
Lemma 2.3 (duality). For a polynomial P P P and 1 < p < ∞ 1<p<\infty 1 < p < ∞ , ∥ P ( S ) ∥ p = ∥ P ( S ) ∥ p ′ \|P(S)\|_p=\|P(S)\|_{p'} ∥ P ( S ) ∥ p = ∥ P ( S ) ∥ p ′ , p ′ = p / ( p − 1 ) p'=p/(p-1) p ′ = p / ( p − 1 ) . If T T T is a real 2 × 2 2\times2 2 × 2 matrix with transpose T t T^{\mathrm{t}} T t , then ∥ P ( T t ) ∥ p = ∥ P ( T ) ∥ p ′ \|P(T^{\mathrm{t}})\|_p=\|P(T)\|_{p'} ∥ P ( T t ) ∥ p = ∥ P ( T ) ∥ p ′ and ∥ T t ∥ p = ∥ T ∥ p ′ \|T^{\mathrm{t}}\|_p=\|T\|_{p'} ∥ T t ∥ p = ∥ T ∥ p ′ (norms on ℓ 2 p \ell_2^p ℓ 2 p and on ℓ 2 p ′ \ell_2^{p'} ℓ 2 p ′ ).
Proof. Let C C C be convolution by ( c k ) (c_k) ( c k ) on ℓ p ( Z ) \ell^p(\mathbb{Z}) ℓ p ( Z ) , ( C x ) n = ∑ k c k x n − k (Cx)_n=\sum_k c_kx_{n-k} ( C x ) n = ∑ k c k x n − k . Then ∥ P ( S ) ∥ p = ∥ C ∥ p \|P(S)\|_p=\|C\|_p ∥ P ( S ) ∥ p = ∥ C ∥ p : P ( S ) P(S) P ( S ) is the restriction of C C C to the sequences that vanish at negative indices, so ∥ P ( S ) ∥ p ≤ ∥ C ∥ p \|P(S)\|_p\le\|C\|_p ∥ P ( S ) ∥ p ≤ ∥ C ∥ p ; conversely, a finitely supported x ∈ ℓ p ( Z ) x\in\ell^p(\mathbb{Z}) x ∈ ℓ p ( Z ) can be translated into ℓ p ( N ) \ell^p(\mathbb{N}) ℓ p ( N ) , translations are isometries that commute with C C C , and finitely supported vectors are dense, so ∥ C ∥ p ≤ ∥ P ( S ) ∥ p \|C\|_p\le\|P(S)\|_p ∥ C ∥ p ≤ ∥ P ( S ) ∥ p . The adjoint of C C C , acting on ℓ p ′ ( Z ) \ell^{p'}(\mathbb{Z}) ℓ p ′ ( Z ) , is convolution by ( c − k ) (c_{-k}) ( c − k ) , which the reflection n ↦ − n n\mapsto-n n ↦ − n (an isometry) turns back into convolution by ( c k ) (c_k) ( c k ) ; hence ∥ C ∥ p = ∥ C ∥ p ′ \|C\|_p=\|C\|_{p'} ∥ C ∥ p = ∥ C ∥ p ′ , and the same holds for P ( S ) P(S) P ( S ) . The statements about matrices follow from ∥ A t ∥ p = ∥ A ∥ p ′ \|A^{\mathrm{t}}\|_p=\|A\|_{p'} ∥ A t ∥ p = ∥ A ∥ p ′ (duality) and P ( T ) t = P ( T t ) P(T)^{\mathrm{t}}=P(T^{\mathrm{t}}) P ( T ) t = P ( T t ) . □
The norm of P a ( S ) P_a(S) P a ( S ) for small a a a
Fix p > 2 p>2 p > 2 , B > 0 B>0 B > 0 , and P a ( z ) = 1 + a z − B z 2 P_a(z)=1+az-Bz^2 P a ( z ) = 1 + a z − B z 2 ; put C a : = ∥ P a ( J ) ∥ p C_a:=\|P_a(J)\|_p C a := ∥ P a ( J ) ∥ p (by Lemma 2.1, a maximum over real vectors). For a real sequence x x x , the window at n n n is ( u , v , w ) = ( x n , x n − 1 , x n − 2 ) (u,v,w)=(x_n,x_{n-1},x_{n-2}) ( u , v , w ) = ( x n , x n − 1 , x n − 2 ) ; then ( P a ( S ) x ) n = u + a v − B w (P_a(S)x)_n=u+av-Bw ( P a ( S ) x ) n = u + a v − B w .
Lemma 3.1 (a Bellman-function bound for ∥ P a ( S ) ∥ p \|P_a(S)\|_p ∥ P a ( S ) ∥ p ). Let C > 0 C>0 C > 0 , and let G : R 2 → R G:\mathbb{R}^2\to\mathbb{R} G : R 2 → R be continuous with G ( 0 , 0 ) = 0 G(0,0)=0 G ( 0 , 0 ) = 0 . If
F ( u , v , w ) : = C p ∣ u ∣ p + G ( v , w ) − G ( u , v ) − ∣ u + a v − B w ∣ p ≥ 0 for all ( u , v , w ) ∈ R 3 , (2) F(u,v,w):=C^p|u|^p+G(v,w)-G(u,v)-|u+av-Bw|^p\ge0\qquad\text{for all }(u,v,w)\in\mathbb{R}^3, \tag*{(2)} F ( u , v , w ) := C p ∣ u ∣ p + G ( v , w ) − G ( u , v ) − ∣ u + a v − B w ∣ p ≥ 0 for all ( u , v , w ) ∈ R 3 , (2) then ∥ P a ( S ) ∥ p ≤ C \|P_a(S)\|_p\le C ∥ P a ( S ) ∥ p ≤ C . We call such a G G G a Bellman function for P a P_a P a and C C C .
Proof. For a finitely supported real x ∈ ℓ p ( N ) x\in\ell^p(\mathbb{N}) x ∈ ℓ p ( N ) (with x − 1 = x − 2 = 0 x_{-1}=x_{-2}=0 x − 1 = x − 2 = 0 , and x n = 0 x_n=0 x n = 0 for n > n 0 n>n_0 n > n 0 ) sum (2) over the windows at $0\le n\le n_0+2: the G G G terms telescope to G ( 0 , 0 ) − G ( 0 , 0 ) = 0 G(0,0)-G(0,0)=0 G ( 0 , 0 ) − G ( 0 , 0 ) = 0 , so ∥ P a ( S ) x ∥ p p ≤ C p ∥ x ∥ p p \|P_a(S)x\|_p^p\le C^p\|x\|_p^p ∥ P a ( S ) x ∥ p p ≤ C p ∥ x ∥ p p . By density this holds for all real x ∈ ℓ p ( N ) x\in\ell^p(\mathbb{N}) x ∈ ℓ p ( N ) , and by Lemma 2.1 also for complex x x x . □
Lemma 3.2 (the case a = 0 a=0 a = 0 ). With G 0 ( v , w ) = B ( 1 + B ) p − 1 ( ∣ v ∣ p + ∣ w ∣ p ) G_0(v,w)=B(1+B)^{p-1}(|v|^p+|w|^p) G 0 ( v , w ) = B ( 1 + B ) p − 1 ( ∣ v ∣ p + ∣ w ∣ p ) and C = 1 + B C=1+B C = 1 + B , (2) holds for a = 0 a=0 a = 0 : F 0 ( u , v , w ) = ( 1 + B ) p − 1 ( ∣ u ∣ p + B ∣ w ∣ p ) − ∣ u − B w ∣ p ≥ 0 F_0(u,v,w)=(1+B)^{p-1}(|u|^p+B|w|^p)-|u-Bw|^p\ge0 F 0 ( u , v , w ) = ( 1 + B ) p − 1 ( ∣ u ∣ p + B ∣ w ∣ p ) − ∣ u − B w ∣ p ≥ 0 , with equality exactly on the plane Π = { w = − u } \Pi=\{w=-u\} Π = { w = − u } , the set of windows of the sequences with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n .
Proof. The identities G 0 ( v , w ) − G 0 ( u , v ) = B ( 1 + B ) p − 1 ( ∣ w ∣ p − ∣ u ∣ p ) G_0(v,w)-G_0(u,v)=B(1+B)^{p-1}(|w|^p-|u|^p) G 0 ( v , w ) − G 0 ( u , v ) = B ( 1 + B ) p − 1 ( ∣ w ∣ p − ∣ u ∣ p ) and ( 1 + B ) p − B ( 1 + B ) p − 1 = ( 1 + B ) p − 1 (1+B)^p-B(1+B)^{p-1}=(1+B)^{p-1} ( 1 + B ) p − B ( 1 + B ) p − 1 = ( 1 + B ) p − 1 give the formula for F 0 F_0 F 0 . Write u − B w = ( 1 + B ) ( 1 1 + B u + B 1 + B ( − w ) ) u-Bw=(1+B)\left(\frac{1}{1+B}u+\frac{B}{1+B}(-w)\right) u − B w = ( 1 + B ) ( 1 + B 1 u + 1 + B B ( − w ) ) ; strict convexity of ∣ ⋅ ∣ p |\cdot|^p ∣ ⋅ ∣ p (p > 1 p>1 p > 1 ) gives ∣ u − B w ∣ p ≤ ( 1 + B ) p − 1 ( ∣ u ∣ p + B ∣ w ∣ p ) |u-Bw|^p\le(1+B)^{p-1}(|u|^p+B|w|^p) ∣ u − B w ∣ p ≤ ( 1 + B ) p − 1 ( ∣ u ∣ p + B ∣ w ∣ p ) , with equality if and only if u = − w u=-w u = − w . □
Lemma 3.3 (growth of F 0 F_0 F 0 away from Π \Pi Π ). With ν = u + w \nu= u+w ν = u + w , there is γ 1 = γ 1 ( p , B ) > 0 \gamma_1=\gamma_1(p,B)>0 γ 1 = γ 1 ( p , B ) > 0 with F 0 ( u , v , w ) ≥ γ 1 ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 F_0(u,v,w)\geq\gamma_1(|u|+|\nu|)^{p-2}\nu^2 F 0 ( u , v , w ) ≥ γ 1 ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 for all real u , v , w u,v,w u , v , w .
Proof. The function F 0 F_0 F 0 does not depend on v v v , and both sides are homogeneous of degree p p p in ( u , ν ) (u,\nu) ( u , ν ) , so it suffices to consider K = { ∣ u ∣ + ∣ ν ∣ = 1 } K=\{|u|+|\nu|=1\} K = { ∣ u ∣ + ∣ ν ∣ = 1 } . On K K K with ν ≠ 0 \nu\ne0 ν = 0 , F 0 > 0 F_0>0 F 0 > 0 by Lemma 3.2. Near ν = 0 \nu=0 ν = 0 on K K K , ∣ u ∣ |u| ∣ u ∣ is close to 1 1 1 , and the bases of the three p p p -th powers in F 0 F_0 F 0 (the numbers y y y in its terms ∣ y ∣ p |y|^p ∣ y ∣ p ), namely u u u , w = ν − u w=\nu-u w = ν − u and u − B w = ( 1 + B ) u − B ν u-Bw=(1+B)u-B\nu u − B w = ( 1 + B ) u − B ν , are close to ± 1 , ∓ 1 , ± ( 1 + B ) \pm1,\mp1,\pm(1+B) ± 1 , ∓ 1 , ± ( 1 + B ) ; so F 0 F_0 F 0 is real-analytic there, F 0 = 0 F_0=0 F 0 = 0 at ν = 0 \nu=0 ν = 0 , ∂ ν F 0 = p B ( 1 + B ) p − 1 φ ( ν − u ) + p B φ ( ( 1 + B ) u − B ν ) \partial_\nu F_0=pB(1+B)^{p-1}\varphi(\nu-u)+pB\varphi((1+B)u-B\nu) ∂ ν F 0 = pB ( 1 + B ) p − 1 φ ( ν − u ) + pB φ (( 1 + B ) u − B ν ) , which at ν = 0 \nu=0 ν = 0 is − p B ( 1 + B ) p − 1 φ ( u ) + p B ( 1 + B ) p − 1 φ ( u ) = 0 -pB(1+B)^{p-1}\varphi(u)+pB(1+B)^{p-1}\varphi(u)=0 − pB ( 1 + B ) p − 1 φ ( u ) + pB ( 1 + B ) p − 1 φ ( u ) = 0 , and ∂ ν 2 F 0 = p ( p − 1 ) B ( 1 + B ) p − 2 ∣ u ∣ p − 2 > 0 \partial_\nu^2F_0=p(p-1)B(1+B)^{p-2}|u|^{p-2}>0 ∂ ν 2 F 0 = p ( p − 1 ) B ( 1 + B ) p − 2 ∣ u ∣ p − 2 > 0 at ν = 0 \nu=0 ν = 0 . Hence F 0 / ν 2 F_0/\nu^2 F 0 / ν 2 extends to a positive continuous function on K K K , and γ 1 \gamma_1 γ 1 is its minimum. □
The function Q Q Q . As in the Introduction, for ξ ∈ R 2 ∖ { 0 } \xi\in\mathbb{R}^2\setminus\{0\} ξ ∈ R 2 ∖ { 0 } put Q ( ξ ) = ⟨ φ ( ξ ) , J ξ ⟩ / ∥ ξ ∥ p p = ( φ ( ξ 1 ) ξ 2 − φ ( ξ 2 ) ξ 1 ) / ∥ ξ ∥ p p Q(\xi)=\langle\varphi(\xi),J\xi\rangle/\|\xi\|_p^p=(\varphi(\xi_1)\xi_2-\varphi(\xi_2)\xi_1)/\|\xi\|_p^p Q ( ξ ) = ⟨ φ ( ξ ) , J ξ ⟩ /∥ ξ ∥ p p = ( φ ( ξ 1 ) ξ 2 − φ ( ξ 2 ) ξ 1 ) /∥ ξ ∥ p p . Since J 2 = − I J^2=-I J 2 = − I , P a ( J ) = ( 1 + B ) I + a J P_a(J)=(1+B)I+aJ P a ( J ) = ( 1 + B ) I + a J , so ∥ P a ( J ) ξ ∥ p p = ( 1 + B ) p ∥ ξ + h J ξ ∥ p p \|P_a(J)\xi\|_p^p=(1+B)^p\|\xi+hJ\xi\|_p^p ∥ P a ( J ) ξ ∥ p p = ( 1 + B ) p ∥ ξ + h J ξ ∥ p p with h = a / ( 1 + B ) h=a/(1+B) h = a / ( 1 + B ) . The derivative formula of § 2 gives the term of first order in h h h , namely h p ⟨ φ ( ξ ) , J ξ ⟩ = h p Q ( ξ ) ∥ ξ ∥ p p hp\langle\varphi(\xi),J\xi\rangle=hpQ(\xi)\|\xi\|_p^p h p ⟨ φ ( ξ ) , J ξ ⟩ = h pQ ( ξ ) ∥ ξ ∥ p p ; the second derivative in h h h , namely p ( p − 1 ) ∑ i = 1 2 ∣ ξ i + h ( J ξ ) i ∣ p − 2 ( J ξ ) i 2 p(p-1)\sum_{i=1}^2|\xi_i+h(J\xi)_i|^{p-2}(J\xi)_i^2 p ( p − 1 ) ∑ i = 1 2 ∣ ξ i + h ( J ξ ) i ∣ p − 2 ( J ξ ) i 2 , is bounded for ξ \xi ξ on the Euclidean unit circle and ∣ h ∣ ≤ 1 |h|\leq1 ∣ h ∣ ≤ 1 because p > 2 p>2 p > 2 . Taylor’s formula in h h h , and homogeneity in ξ \xi ξ , therefore give
∥ P a ( J ) ξ ∥ p p = ( 1 + B ) p ∥ ξ ∥ p p ( 1 + p a 1 + B Q ( ξ ) + O ( a 2 ) ) . (3) \|P_a(J)\xi\|_p^p=(1+B)^p\|\xi\|_p^p\left(1+\frac{pa}{1+B}Q(\xi)+O(a^2)\right). \tag*{(3)} ∥ P a ( J ) ξ ∥ p p = ( 1 + B ) p ∥ ξ ∥ p p ( 1 + 1 + B p a Q ( ξ ) + O ( a 2 ) ) . (3)
so, to first order in a a a , the vectors ξ \xi ξ that maximise ∥ P a ( J ) ξ ∥ p / ∥ ξ ∥ p \|P_a(J)\xi\|_p/\|\xi\|_p ∥ P a ( J ) ξ ∥ p /∥ ξ ∥ p are the maximisers of Q Q Q . For real σ \sigma σ put f ( σ ) = Q ( 1 , σ ) = σ ( 1 − ∣ σ ∣ p − 2 ) / ( 1 + ∣ σ ∣ p ) f(\sigma)=Q(1,\sigma)=\sigma(1-|\sigma|^{p-2})/(1+|\sigma|^p) f ( σ ) = Q ( 1 , σ ) = σ ( 1 − ∣ σ ∣ p − 2 ) / ( 1 + ∣ σ ∣ p ) .
Lemma 3.4 (the maximum of Q Q Q ). Let p > 2 p>2 p > 2 . The function f f f has exactly one critical point t ∗ t_* t ∗ in ( 0 , 1 ) (0,1) ( 0 , 1 ) ; it is a nondegenerate maximum, and D ∗ : = max ξ ≠ 0 Q ( ξ ) = f ( t ∗ ) > 0 D_*:=\max_{\xi\ne0}Q(\xi)=f(t_*)>0 D ∗ := max ξ = 0 Q ( ξ ) = f ( t ∗ ) > 0 is attained exactly on the nonzero multiples of ( 1 , t ∗ ) (1,t_*) ( 1 , t ∗ ) and of J ( 1 , t ∗ ) J(1,t_*) J ( 1 , t ∗ ) . Moreover
1 + t ∗ 2 p − 2 = ( p − 1 ) t ∗ p − 2 ( 1 + t ∗ 2 ) . (4) 1+t_*^{2p-2}=(p-1)t_*^{p-2}(1+t_*^2). \tag*{(4)} 1 + t ∗ 2 p − 2 = ( p − 1 ) t ∗ p − 2 ( 1 + t ∗ 2 ) . (4)
Proof. For 0 < t < 1 0<t<1 0 < t < 1 , f ( t ) = ( t − t p − 1 ) / ( 1 + t p ) f(t)=(t-t^{p-1})/(1+t^p) f ( t ) = ( t − t p − 1 ) / ( 1 + t p ) , and by the quotient rule f ′ ( t ) = g ( t ) / ( 1 + t p ) 2 f'(t)=g(t)/(1+t^p)^2 f ′ ( t ) = g ( t ) / ( 1 + t p ) 2 with g ( t ) = 1 + t 2 p − 2 − ( p − 1 ) t p − 2 ( 1 + t 2 ) g(t)=1+t^{2p-2}-(p-1)t^{p-2}(1+t^2) g ( t ) = 1 + t 2 p − 2 − ( p − 1 ) t p − 2 ( 1 + t 2 ) . Next g ′ ( t ) = ( p − 1 ) t p − 3 χ ( t ) g'(t)=(p-1)t^{p-3}\chi(t) g ′ ( t ) = ( p − 1 ) t p − 3 χ ( t ) with χ ( t ) = 2 t p − p t 2 − ( p − 2 ) \chi(t)=2t^p-pt^2-(p-2) χ ( t ) = 2 t p − p t 2 − ( p − 2 ) ; since χ ( 0 ) = − ( p − 2 ) < 0 \chi(0)=-(p-2)<0 χ ( 0 ) = − ( p − 2 ) < 0 and χ ′ ( t ) = 2 p t ( t p − 2 − 1 ) < 0 \chi'(t)=2pt(t^{p-2}-1)<0 χ ′ ( t ) = 2 pt ( t p − 2 − 1 ) < 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) , χ < 0 \chi<0 χ < 0 there, so g g g is strictly decreasing on ( 0 , 1 ) (0,1) ( 0 , 1 ) , from g ( 0 + ) = 1 g(0+)=1 g ( 0 + ) = 1 to g ( 1 ) = 4 − 2 p < 0 g(1)=4-2p<0 g ( 1 ) = 4 − 2 p < 0 . Hence g g g has exactly one zero t ∗ t_* t ∗ , which is (4), and f ′ ′ ( t ∗ ) = g ′ ( t ∗ ) / ( 1 + t ∗ p ) 2 < 0 f''(t_*)=g'(t_*)/(1+t_*^p)^2<0 f ′′ ( t ∗ ) = g ′ ( t ∗ ) / ( 1 + t ∗ p ) 2 < 0 . As f > 0 f>0 f > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) and f ( 0 ) = f ( 1 ) = 0 f(0)=f(1)=0 f ( 0 ) = f ( 1 ) = 0 , t ∗ t_* t ∗ is the only maximiser of f f f on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Finally, Q Q Q is invariant under ξ ↦ J ξ \xi\mapsto J\xi ξ ↦ J ξ , ξ ↦ − ξ \xi\mapsto-\xi ξ ↦ − ξ and ξ ↦ c ξ \xi\mapsto c\xi ξ ↦ c ξ ( c > 0 ) (c>0) ( c > 0 ) . Replacing ξ \xi ξ by J ξ J\xi J ξ if necessary, we may assume ∣ ξ 2 ∣ ≤ ∣ ξ 1 ∣ |\xi_2|\leq|\xi_1| ∣ ξ 2 ∣ ≤ ∣ ξ 1 ∣ ; then ξ \xi ξ is a multiple of ( 1 , σ ) (1,\sigma) ( 1 , σ ) with ∣ σ ∣ ≤ 1 |\sigma|\leq1 ∣ σ ∣ ≤ 1 , and Q ( ξ ) = f ( σ ) Q(\xi)=f(\sigma) Q ( ξ ) = f ( σ ) . Since f ( σ ) ≤ 0 f(\sigma)\leq0 f ( σ ) ≤ 0 for σ ∈ [ − 1 , 0 ] \sigma\in[-1,0] σ ∈ [ − 1 , 0 ] , the maximum of Q Q Q is f ( t ∗ ) f(t_*) f ( t ∗ ) , attained exactly at the nonzero multiples of ( 1 , t ∗ ) (1,t_*) ( 1 , t ∗ ) and of J ( 1 , t ∗ ) J(1,t_*) J ( 1 , t ∗ ) . □
For example, t ∗ = 2 sin ( π / 12 ) t_*=2\sin(\pi/12) t ∗ = 2 sin ( π /12 ) , D ∗ = 2 − 3 / 2 D_*=2^{-3/2} D ∗ = 2 − 3/2 for p = 4 p=4 p = 4 , and t ∗ = ( 5 − 1 ) / 2 t_*=(\sqrt{5}-1)/2 t ∗ = ( 5 − 1 ) /2 , D ∗ = 1 2 D_*=\frac{1}{2} D ∗ = 2 1 for p = 6 p=6 p = 6 (Figure 2 ).
Figure 2. The function f ( t ) = ( t − t p − 1 ) / ( 1 + t p ) f(t)=(t-t^{p-1})/(1+t^p) f ( t ) = ( t − t p − 1 ) / ( 1 + t p ) of Lemma 3.4 on [ 0 , 1 ] [0,1] [ 0 , 1 ] , for p = 3 , 4 , 6 p=3,4,6 p = 3 , 4 , 6 . The dot marks its maximum D ∗ = f ( t ∗ ) D_* = f(t_*) D ∗ = f ( t ∗ ) (for p = 4 p=4 p = 4 and p = 6 p=6 p = 6 , the values given above). Plotted numerically; an illustration, not used in the proofs.
Lemma 3.5 (the maximising vectors of P a ( J ) P_a(J) P a ( J ) ). Let p > 2 p>2 p > 2 and B > 0 B>0 B > 0 . We measure distances on the Euclidean unit circle of R 2 \mathbb{R}^2 R 2 by angle. Let e ∗ = ( 1 , t ∗ ) / ∥ ( 1 , t ∗ ) ∥ 2 e_*=(1,t_*)/\|(1,t_*)\|_2 e ∗ = ( 1 , t ∗ ) /∥ ( 1 , t ∗ ) ∥ 2 , let O 0 = { ± e ∗ , ± J e ∗ } \mathcal{O}_0=\{\pm e_*,\pm Je_*\} O 0 = { ± e ∗ , ± J e ∗ } , and for s ∈ R 2 s\in\mathbb{R}^2 s ∈ R 2 put Σ a ( s ) = 2 ( C a ∥ s ∥ p p − ∥ P a ( J ) s ∥ p p ) \Sigma_a(s)=2(C_a\|s\|_p^p-\|P_a(J)s\|_p^p) Σ a ( s ) = 2 ( C a ∥ s ∥ p p − ∥ P a ( J ) s ∥ p p ) ; thus Σ a ≥ 0 \Sigma_a\geq0 Σ a ≥ 0 by the definition of C a C_a C a . There are ρ ∈ ( 0 , 1 2 arctan t ∗ ) \rho\in(0,\frac{1}{2}\arctan t_*) ρ ∈ ( 0 , 2 1 arctan t ∗ ) , a 1 > 0 a_1>0 a 1 > 0 and η > 0 \eta>0 η > 0 such that for 0 < a < a 1 0<a<a_1 0 < a < a 1 :
( i ) (i) ( i ) the function Σ a \Sigma_a Σ a vanishes on the unit circle at exactly four points: if s a s_a s a denotes the one nearest to e ∗ e_* e ∗ , these points are O a = { ± s a , ± J s a } \mathcal{O}_a=\{\pm s_a,\pm Js_a\} O a = { ± s a , ± J s a } ; they are the maximisers of ∥ P a ( J ) s ∥ p / ∥ s ∥ p \|P_a(J)s\|_p/\|s\|_p ∥ P a ( J ) s ∥ p /∥ s ∥ p on the unit circle, and O a \mathcal{O}_a O a lies within angular distance ρ / 2 \rho/2 ρ /2 of O 0 \mathcal{O}_0 O 0 ;
( i i ) (ii) ( ii ) if s s s is on the unit circle at angular distance at least ρ \rho ρ from O a \mathcal{O}_a O a , then Σ a ( s ) ≥ a δ \Sigma_a(s)\geq a\delta Σ a ( s ) ≥ a δ ;
( i i i ) (iii) ( iii ) if s s s is on the unit circle at angular distance r ( s ) ≤ ρ r(s)\leq\rho r ( s ) ≤ ρ from O a \mathcal{O}_a O a , then Σ a ( s ) ≥ a η r ( s ) 2 \Sigma_a(s)\geq a\eta r(s)^2 Σ a ( s ) ≥ a η r ( s ) 2 ;
( i v ) (iv) ( i v ) with s 0 : = e ∗ s_0:=e_* s 0 := e ∗ , the maps a ↦ s a a\mapsto s_a a ↦ s a and a ↦ C a p a\mapsto C_a^p a ↦ C a p are real-analytic on [ 0 , a 1 ) [0,a_1) [ 0 , a 1 ) .
:::
Proof. With θ \theta θ the angle on the unit circle and θ ∗ = arctan t ∗ \theta_*=\arctan t_* θ ∗ = arctan t ∗ , we have Q ( cos θ , sin θ ) = f ( tan θ ) Q(\cos\theta,\sin\theta)=f(\tan\theta) Q ( cos θ , sin θ ) = f ( tan θ ) near θ ∗ \theta_* θ ∗ , and f ′ ( t ∗ ) = 0 f'(t_*)=0 f ′ ( t ∗ ) = 0 gives ∂ θ 2 Q ( θ ∗ ) = f ′ ′ ( t ∗ ) sec 4 θ ∗ < 0 \partial_\theta^2 Q(\theta_*)=f''(t_*)\sec^4\theta_*<0 ∂ θ 2 Q ( θ ∗ ) = f ′′ ( t ∗ ) sec 4 θ ∗ < 0 ; by the invariance of Q Q Q under ξ ↦ J ξ \xi\mapsto J\xi ξ ↦ J ξ and ξ ↦ − ξ \xi\mapsto-\xi ξ ↦ − ξ , the same holds at the other three points of O 0 \mathcal{O}_0 O 0 . The points ( 0 , ± 1 ) (0,\pm1) ( 0 , ± 1 ) , ( ± 1 , 0 ) (\pm1,0) ( ± 1 , 0 ) are at angular distance at least arctan t ∗ \arctan t_* arctan t ∗ from O 0 \mathcal{O}_0 O 0 . Hence we can fix ρ ∈ ( 0 , 1 2 arctan t ∗ ) \rho\in(0,\frac{1}{2}\arctan t_*) ρ ∈ ( 0 , 2 1 arctan t ∗ ) and m > 0 m>0 m > 0 such that ∂ θ 2 Q ≤ − m \partial_\theta^2Q\leq-m ∂ θ 2 Q ≤ − m on the four closed arcs of angular radius 2 ρ 2\rho 2 ρ centred at the points of O 0 \mathcal{O}_0 O 0 (the 2 ρ 2\rho 2 ρ -arcs); these arcs are disjoint and avoid the coordinate axes. Write
Σ a ( s ) a = 2 ∥ s ∥ p p ( c ( a ) − H ( a , s ) ) , where \frac{\Sigma_a(s)}{a}=2\lVert s\rVert_p^p\bigl(c(a)-H(a,s)\bigr),\qquad\text{where} a Σ a ( s ) = 2 ∥ s ∥ p p ( c ( a ) − H ( a , s ) ) , where
H ( a , s ) = 1 a ( ∥ P a ( J ) s ∥ p p ∥ s ∥ p p − ( 1 + B ) p ) , c ( a ) = max s ≠ 0 H ( a , s ) = C a p − ( 1 + B ) p a . H(a,s)=\frac{1}{a}\left(\frac{\lVert P_a(J)s\rVert_p^p}{\lVert s\rVert_p^p}-(1+B)^p\right),\qquad c(a)=\max_{s\ne0}H(a,s)=\frac{C_a^p-(1+B)^p}{a}. H ( a , s ) = a 1 ( ∥ s ∥ p p ∥ P a ( J ) s ∥ p p − ( 1 + B ) p ) , c ( a ) = s = 0 max H ( a , s ) = a C a p − ( 1 + B ) p .
By (3), H ( a , s ) ⟶ H ( 0 , s ) : = p ( 1 + B ) p − 1 Q ( s ) H(a,s)\longrightarrow H(0,s):=p(1+B)^{p-1}Q(s) H ( a , s ) ⟶ H ( 0 , s ) := p ( 1 + B ) p − 1 Q ( s ) uniformly on the unit circle, hence c ( a ) ⟶ p ( 1 + B ) p − 1 D ∗ c(a)\longrightarrow p(1+B)^{p-1}D_* c ( a ) ⟶ p ( 1 + B ) p − 1 D ∗ and, uniformly on the unit circle,
Σ a ( s ) a ⟶ Σ ′ ( s ) : = 2 p ( 1 + B ) p − 1 ∥ s ∥ p p ( D ∗ − Q ( s ) ) ( a → 0 + ) . \frac{\Sigma_a(s)}{a}\longrightarrow\Sigma'(s):=2p(1+B)^{p-1}\lVert s\rVert_p^p(D_*-Q(s))\qquad(a\to0^+). a Σ a ( s ) ⟶ Σ ′ ( s ) := 2 p ( 1 + B ) p − 1 ∥ s ∥ p p ( D ∗ − Q ( s )) ( a → 0 + ) .
(As Σ 0 = 0 \Sigma_0=0 Σ 0 = 0 , Σ ′ ( s ) \Sigma'(s) Σ ′ ( s ) is the right derivative of Σ a ( s ) \Sigma_a(s) Σ a ( s ) at a = 0 a=0 a = 0 .) On the 2 ρ 2\rho 2 ρ -arcs both coordinates of s s s have absolute value at least sin ( arctan t ∗ − 2 ρ ) \sin(\arctan t_* - 2\rho) sin ( arctan t ∗ − 2 ρ ) , so for ∣ a ∣ < 1 2 ( 1 + B ) sin ( arctan t ∗ − 2 ρ ) |a| < \frac{1}{2}(1+B)\sin(\arctan t_* - 2\rho) ∣ a ∣ < 2 1 ( 1 + B ) sin ( arctan t ∗ − 2 ρ ) no coordinate of P a ( J ) s = ( 1 + B ) s + a J s P_a(J)s=(1+B)s+aJs P a ( J ) s = ( 1 + B ) s + a J s vanishes there. For such a a a , ∥ P a ( J ) s ∥ p p / ∥ s ∥ p p \lVert P_a(J)s\rVert_p^p/\lVert s\rVert_p^p ∥ P a ( J ) s ∥ p p / ∥ s ∥ p p is real-analytic in ( a , s ) (a,s) ( a , s ) and equals ( 1 + B ) p (1+B)^p ( 1 + B ) p at a = 0 a=0 a = 0 , so H H H (with the value H ( 0 , s ) H(0,s) H ( 0 , s ) above) is real-analytic in ( a , s ) (a,s) ( a , s ) , also at a = 0 a=0 a = 0 . In particular H ( a , ⋅ ) → H ( 0 , ⋅ ) H(a,\cdot)\to H(0,\cdot) H ( a , ⋅ ) → H ( 0 , ⋅ ) in C 2 C^2 C 2 on these arcs. By Lemma 3.4, Σ ′ ≥ 0 \Sigma'\geq0 Σ ′ ≥ 0 vanishes on the unit circle exactly on O 0 \mathcal O_0 O 0 . Put
δ = 1 2 min { Σ ′ ( s ) : ∥ s ∥ 2 = 1 , s at angular distance ≥ ρ / 2 from O 0 } > 0 , m 1 = 1 2 p ( 1 + B ) p − 1 m . \delta=\frac{1}{2}\min\{\Sigma'(s):\lVert s\rVert_2=1,\ s\text{ at angular distance }\geq\rho/2\text{ from }\mathcal O_0\}>0,\qquad m_1=\frac{1}{2}p(1+B)^{p-1}m. δ = 2 1 min { Σ ′ ( s ) : ∥ s ∥ 2 = 1 , s at angular distance ≥ ρ /2 from O 0 } > 0 , m 1 = 2 1 p ( 1 + B ) p − 1 m .
By the uniform convergence, and the C 2 C^2 C 2 convergence with ∂ θ 2 H ( 0 , ⋅ ) = p ( 1 + B ) p − 1 ∂ θ 2 Q ≤ − 2 m 1 \partial_\theta^2H(0,\cdot)=p(1+B)^{p-1}\partial_\theta^2Q\leq-2m_1 ∂ θ 2 H ( 0 , ⋅ ) = p ( 1 + B ) p − 1 ∂ θ 2 Q ≤ − 2 m 1 , there is a 1 > 0 a_1>0 a 1 > 0 such that for 0 < a < a 1 0<a<a_1 0 < a < a 1 : Σ a / a ≥ δ \Sigma_a/a\geq\delta Σ a / a ≥ δ at angular distance ≥ ρ / 2 \geq\rho/2 ≥ ρ /2 from O 0 \mathcal O_0 O 0 , and ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 \partial_\theta^2H(a,\cdot)\leq-m_1 ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 on the closed 2 ρ 2\rho 2 ρ -arcs. Decreasing a 1 a_1 a 1 if necessary, we may also assume that a 1 ≤ 1 2 ( 1 + B ) sin ( arctan t ∗ − 2 ρ ) a_1\leq\frac{1}{2}(1+B)\sin(\arctan t_* - 2\rho) a 1 ≤ 2 1 ( 1 + B ) sin ( arctan t ∗ − 2 ρ ) , so that H H H is real-analytic on [ 0 , a 1 ) [0,a_1) [ 0 , a 1 ) times the 2 ρ 2\rho 2 ρ -arcs.
Proof of (i) and (iv). The maximisers of H ( a , ⋅ ) H(a,\cdot) H ( a , ⋅ ) on the unit circle are the zeros of Σ a \Sigma_a Σ a there, so they lie within angular distance ρ / 2 \rho/2 ρ /2 of O 0 \mathcal O_0 O 0 ; by the invariance under s ↦ J s s\mapsto Js s ↦ J s and s ↦ − s s\mapsto-s s ↦ − s there is a maximiser near each point of O 0 \mathcal O_0 O 0 . As H ( a , ⋅ ) H(a,\cdot) H ( a , ⋅ ) is strictly concave on each 2 ρ 2\rho 2 ρ -arc, the arc around a point o ∈ O 0 o\in\mathcal O_0 o ∈ O 0 contains exactly one maximiser, namely the only zero of ∂ θ H ( a , ⋅ ) \partial_\theta H(a,\cdot) ∂ θ H ( a , ⋅ ) on that arc. For a = 0 a=0 a = 0 this maximiser is o o o , as H ( 0 , ⋅ ) = p ( 1 + B ) p − 1 Q H(0,\cdot)=p(1+B)^{p-1}Q H ( 0 , ⋅ ) = p ( 1 + B ) p − 1 Q (Lemma 3.4). Since ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 < 0 \partial_\theta^2H(a,\cdot)\leq-m_1<0 ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 < 0 on the arc for every a ∈ [ 0 , a 1 ) a\in[0,a_1) a ∈ [ 0 , a 1 ) , the implicit function theorem for ∂ θ H = 0 \partial_\theta H=0 ∂ θ H = 0 , applied at each a ∈ [ 0 , a 1 ) a\in[0,a_1) a ∈ [ 0 , a 1 ) , shows that this maximiser is a real-analytic function of a ∈ [ 0 , a 1 ) a\in[0,a_1) a ∈ [ 0 , a 1 ) with value o o o at a = 0 a=0 a = 0 ; for o = e ∗ o=e_* o = e ∗ it is s a s_a s a . Hence c ( a ) = H ( a , s a ) c(a)=H(a,s_a) c ( a ) = H ( a , s a ) is real-analytic in a a a , and so is C a p = ( 1 + B ) p + a c ( a ) C_a^p=(1+B)^p+ac(a) C a p = ( 1 + B ) p + a c ( a ) .
Proof of (ii). If s s s is at angular distance ≥ ρ \geq\rho ≥ ρ from O a \mathcal O_a O a , it is at angular distance ≥ ρ / 2 \geq\rho/2 ≥ ρ /2 from O 0 \mathcal O_0 O 0 , so Σ a ( s ) ≥ a δ \Sigma_a(s)\geq a\delta Σ a ( s ) ≥ a δ .
Proof of (iii). Let o a o_a o a be the point of O a \mathcal O_a O a nearest to s s s , so that r ( s ) r(s) r ( s ) is the angular distance from s s s to o a o_a o a , and let o o o be the point of O 0 \mathcal O_0 O 0 within ρ / 2 \rho/2 ρ /2 of o a o_a o a . The arc from o a o_a o a to s s s lies within 3 ρ / 2 < 2 ρ 3\rho/2<2\rho 3 ρ /2 < 2 ρ of o o o ; on this arc ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 \partial_\theta^2H(a,\cdot)\leq-m_1 ∂ θ 2 H ( a , ⋅ ) ≤ − m 1 , and ∂ θ H ( a , o a ) = 0 \partial_\theta H(a,o_a)=0 ∂ θ H ( a , o a ) = 0 , so c ( a ) − H ( a , s ) ≥ 1 2 m 1 r ( s ) 2 c(a)-H(a,s)\geq\frac{1}{2}m_1r(s)^2 c ( a ) − H ( a , s ) ≥ 2 1 m 1 r ( s ) 2 . Hence Σ a ( s ) ≥ a η r ( s ) 2 \Sigma_a(s)\geq a\eta r(s)^2 Σ a ( s ) ≥ a η r ( s ) 2 with η = m 1 min ∥ s ∥ 2 = 1 ∥ s ∥ p p \eta=m_1\min_{\lVert s\rVert_2=1}\lVert s\rVert_p^p η = m 1 min ∥ s ∥ 2 = 1 ∥ s ∥ p p . □
Theorem 3.6 (∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p \lVert P_a(S)\rVert_p=\lVert P_a(J)\rVert_p ∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p for small a a a ). For every p > 2 p>2 p > 2 and B > 0 B>0 B > 0 there is a 0 = a 0 ( p , B ) > 0 a_0=a_0(p,B)>0 a 0 = a 0 ( p , B ) > 0 such that ∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p \lVert P_a(S)\rVert_p=\lVert P_a(J)\rVert_p ∥ P a ( S ) ∥ p = ∥ P a ( J ) ∥ p for 0 ≤ a < a 0 0\leq a<a_0 0 ≤ a < a 0 ; for 1 < p < 2 1<p<2 1 < p < 2 the same holds for 0 ≤ a < a 0 ( p ′ , B ) 0\leq a<a_0(p',B) 0 ≤ a < a 0 ( p ′ , B ) .
Proof. By Lemma 2.2 only ∥ P a ( S ) ∥ p ≤ C a \lVert P_a(S)\rVert_p\leq C_a ∥ P a ( S ) ∥ p ≤ C a needs proof (and only this inequality is used for Theorem A). For a = 0 a=0 a = 0 it is Lemma 3.2. For small a > 0 a>0 a > 0 we construct G = G a G=G_a G = G a with C = C a C=C_a C = C a in Lemma 3.1.
Outline. Write a window as s ^ + ν e 3 \hat{s}+\nu e_3 s ^ + ν e 3 with the state s = ( u , v ) s=(u,v) s = ( u , v ) , s ^ = ( u , v , − u ) ∈ Π \hat{s}=(u,v,-u)\in\Pi s ^ = ( u , v , − u ) ∈ Π , ν = v + w \nu=v+w ν = v + w and e 3 = ( 0 , 0 , 1 ) e_3=(0,0,1) e 3 = ( 0 , 0 , 1 ) . At a = 0 a=0 a = 0 , F 0 ≥ 0 F_0\geq0 F 0 ≥ 0 vanishes exactly on Π \Pi Π and grows like ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 (|u|+|\nu|)^{p-2}\nu^2 ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 off it (Lemmas 3.2 and 3.3). First, along a sequence with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n the state runs through the cycle s , − J s , − s , J s s,-Js,-s,Js s , − J s , − s , J s , and the sum of F F F (with C = C a C=C_a C = C a ) over the four windows does not depend on G G G : it is Σ a ( s ) \Sigma_a(s) Σ a ( s ) . Second, we take G a = G 0 + 1 2 Δ a + k a G_a=G_0+\frac{1}{2}\Delta_a+k_a G a = G 0 + 2 1 Δ a + k a , with Δ a \Delta_a Δ a and k a k_a k a defined in Steps 1 and 2, and F a F_a F a denotes F F F with C = C a C=C_a C = C a and G = G a G=G_a G = G a . Then F a F_a F a takes the same value 1 4 Σ a ( s ) \frac{1}{4}\Sigma_a(s) 4 1 Σ a ( s ) at each of the four windows of the cycle; in particular F a ( s ^ ) = 1 4 Σ a ( s ) ≥ 0 F_a(\hat{s})=\frac{1}{4}\Sigma_a(s)\geq0 F a ( s ^ ) = 4 1 Σ a ( s ) ≥ 0 , which is zero exactly when s = 0 s=0 s = 0 or s s s lies on a ray through a point of O a \mathcal O_a O a , and grows quadratically off these rays (Lemma 3.5). Third, since we need F a ≥ 0 F_a\geq0 F a ≥ 0 , the gradient of F a F_a F a must vanish wherever F a F_a F a vanishes. Along Π \Pi Π this holds automatically. In the direction e 3 e_3 e 3 , across Π \Pi Π , it is achieved by the term k a k_a k a , which is of size O ( a ) O(a) O ( a ) and does not change F a F_a F a on Π \Pi Π . Fourth, F a − F 0 F_a-F_0 F a − F 0 and its gradient are of size O ( a ) O(a) O ( a ) on bounded sets (Step 5). Away from O a \mathcal O_a O a , Σ a ≥ a δ \Sigma_a\geq a\delta Σ a ≥ a δ absorbs this error for small ν \nu ν , and F 0 F_0 F 0 absorbs it for large ν \nu ν ; near O a \mathcal O_a O a , F a F_a F a is bounded below by a positive quadratic form in the distance to O a \mathcal O_a O a and ν \nu ν . Throughout, J ( u , v ) = ( v , − u ) J(u,v)=(v,-u) J ( u , v ) = ( v , − u ) , and ρ \rho ρ , a 1 a_1 a 1 , δ \delta δ , η \eta η , O a \mathcal O_a O a , s a s_a s a are those of Lemma 3.5; the constant a 1 a_1 a 1 will be decreased finitely many times. Step 1: the sum over a cycle of windows. For a sequence with x n + 2 = − x n x_{n+2}=-x_n x n + 2 = − x n , if the window at n n n is s ^ \hat{s} s ^ , then the window at n + 1 n+1 n + 1 is − J s -Js − J s , and ( v , w ) = ( v , − u ) = J s (v,w)=(v,-u)=Js ( v , w ) = ( v , − u ) = J s on Π \Pi Π . Thus, with
Δ a ( s ) = C a p ∣ u ∣ p − ∣ ( 1 + B ) u + a v ∣ p , \Delta_a(s)=C_a^p|u|^p-\lvert(1+B)u+av\rvert^p, Δ a ( s ) = C a p ∣ u ∣ p − ∣( 1 + B ) u + a v ∣ p ,
F ( s ^ ) = Δ a ( s ) + G ( J s ) − G ( s ) F(\hat{s})=\Delta_a(s)+G(Js)-G(s) F ( s ^ ) = Δ a ( s ) + G ( J s ) − G ( s ) on Π \Pi Π , for C = C a C=C_a C = C a . Since Δ a \Delta_a Δ a is even and − J s , − s , J s -Js,-s,Js − J s , − s , J s complete the cycle of s s s , the G G G terms telescope around the cycle and
∑ s ′ ∈ { s , − J s , − s , J s } F ( s ^ ′ ) = 2 ( Δ a ( s ) + Δ a ( J s ) ) = Σ a ( s ) . (5) \sum_{s'\in\{s,-Js,-s,Js\}}F(\hat{s}')=2\bigl(\Delta_a(s)+\Delta_a(Js)\bigr)=\Sigma_a(s). \tag*{(5)} s ′ ∈ { s , − J s , − s , J s } ∑ F ( s ^ ′ ) = 2 ( Δ a ( s ) + Δ a ( J s ) ) = Σ a ( s ) . (5)
because Δ a ( s ) + Δ a ( J s ) = C a p ( ∣ u ∣ p + ∣ v ∣ p ) − ∣ ( 1 + B ) u + a v ∣ p − ∣ ( 1 + B ) v − a u ∣ p \Delta_a(s)+\Delta_a(Js)=C_a^p\bigl(|u|^p+|v|^p\bigr)-\lvert(1+B)u+av\rvert^p-\lvert(1+B)v-au\rvert^p Δ a ( s ) + Δ a ( J s ) = C a p ( ∣ u ∣ p + ∣ v ∣ p ) − ∣( 1 + B ) u + a v ∣ p − ∣( 1 + B ) v − a u ∣ p and P a ( J ) = ( 1 + B ) I + a J P_a(J)=(1+B)I+aJ P a ( J ) = ( 1 + B ) I + a J .
Step 2: the Bellman function. By Step 1, F ( s ^ ) = 1 4 Σ a ( s ) F(\hat{s})=\frac{1}{4}\Sigma_a(s) F ( s ^ ) = 4 1 Σ a ( s ) for every s s s if G ( J s ) − G ( s ) = 1 2 ( Δ a ( J s ) − Δ a ( s ) ) G(Js)-G(s)=\frac{1}{2}\bigl(\Delta_a(Js)-\Delta_a(s)\bigr) G ( J s ) − G ( s ) = 2 1 ( Δ a ( J s ) − Δ a ( s ) ) ; the choice G = 1 2 Δ a G=\frac{1}{2}\Delta_a G = 2 1 Δ a satisfies this. Adding to G G G a function invariant under J J J does not change F F F on Π \Pi Π . Accordingly, with G 0 G_0 G 0 as in Lemma 3.2, we put
G a = G 0 + 1 2 Δ a + k a , k a = λ 1 ( a ) k ( 1 ) + λ 2 ( a ) k ( 2 ) , k ( 1 ) ( s ) = ( u 2 + v 2 ) p / 2 , k ( 2 ) ( s ) = ∣ u ∣ p + ∣ v ∣ p . G_a=G_0+\frac{1}{2}\Delta_a+k_a,\qquad k_a=\lambda_1(a)k^{(1)}+\lambda_2(a)k^{(2)},\qquad k^{(1)}(s)=(u^2+v^2)^{p/2},\quad k^{(2)}(s)=|u|^p+|v|^p. G a = G 0 + 2 1 Δ a + k a , k a = λ 1 ( a ) k ( 1 ) + λ 2 ( a ) k ( 2 ) , k ( 1 ) ( s ) = ( u 2 + v 2 ) p /2 , k ( 2 ) ( s ) = ∣ u ∣ p + ∣ v ∣ p .
with λ ( a ) = ( λ 1 ( a ) , λ 2 ( a ) ) \lambda(a)=(\lambda_1(a),\lambda_2(a)) λ ( a ) = ( λ 1 ( a ) , λ 2 ( a )) to be fixed in Step 4. The functions G 0 , k ( 1 ) , k ( 2 ) G_0,k^{(1)},k^{(2)} G 0 , k ( 1 ) , k ( 2 ) are even, J J J -invariant and homogeneous of degree p p p ; the functions k ( 1 ) k^{(1)} k ( 1 ) and k ( 2 ) k^{(2)} k ( 2 ) are C 2 C^2 C 2 (as p > 2 p>2 p > 2 ); and G a ( 0 , 0 ) = 0 G_a(0,0)=0 G a ( 0 , 0 ) = 0 . For a = 0 a=0 a = 0 , F a F_a F a is F 0 F_0 F 0 of Lemma 3.2, as Δ 0 = 0 \Delta_0=0 Δ 0 = 0 and λ ( 0 ) = 0 \lambda(0)=0 λ ( 0 ) = 0 (see Step 4). On Π \Pi Π the J J J -invariant parts cancel, so
F a ( s ^ ) = Δ a ( s ) + 1 2 ( Δ a ( J s ) − Δ a ( s ) ) = 1 4 Σ a ( s ) ≥ 0. (6) F_a(\hat{s})=\Delta_a(s)+\frac{1}{2}\bigl(\Delta_a(Js)-\Delta_a(s)\bigr)=\frac{1}{4}\Sigma_a(s)\geq0. \tag*{(6)} F a ( s ^ ) = Δ a ( s ) + 2 1 ( Δ a ( J s ) − Δ a ( s ) ) = 4 1 Σ a ( s ) ≥ 0. (6)
Step 3: the bases of the p p p -th powers near the zeros of F a F_a F a . The function ∣ y ∣ p |y|^p ∣ y ∣ p is real-analytic in y y y for y ≠ 0 y\ne0 y = 0 , and we use its analyticity only there. We therefore show that the bases of all the p p p -th powers in F a F_a F a stay away from 0 0 0 near the windows s ^ \hat{s} s ^ , s ∈ O a s\in\mathcal{O}_a s ∈ O a . For a > 0 a>0 a > 0 these bases are u , v , w u,v,w u , v , w (from ∣ u ∣ p , G 0 , k ( 2 ) |u|^p,G_0,k^{(2)} ∣ u ∣ p , G 0 , k ( 2 ) ), ( 1 + B ) u + a v (1+B)u+av ( 1 + B ) u + a v , ( 1 + B ) v + a w (1+B)v+aw ( 1 + B ) v + a w (from Δ a ( u , v ) \Delta_a(u,v) Δ a ( u , v ) and Δ a ( v , w ) \Delta_a(v,w) Δ a ( v , w ) ) and u + a v − B w u+av-Bw u + a v − B w . The function k ( 1 ) k^{(1)} k ( 1 ) is smooth where u 2 + v 2 u^2+v^2 u 2 + v 2 and v 2 + w 2 v^2+w^2 v 2 + w 2 are positive. For ν 0 > 0 \nu_0>0 ν 0 > 0 let
U ( ν 0 ) = { ( s , ν ) : ∥ s ∥ 2 = 1 , s within angular distance 2 ρ of O 0 , ∣ ν ∣ ≤ ν 0 } , \mathcal{U}(\nu_0)=\{(s,\nu):\|s\|_2=1,\ s\text{ within angular distance }2\rho\text{ of }\mathcal{O}_0,\ |\nu|\leq\nu_0\}, U ( ν 0 ) = {( s , ν ) : ∥ s ∥ 2 = 1 , s within angular distance 2 ρ of O 0 , ∣ ν ∣ ≤ ν 0 } ,
with s = ( u , v ) s=(u,v) s = ( u , v ) and w = ν − u w=\nu-u w = ν − u . At a = 0 a=0 a = 0 and ν = 0 \nu=0 ν = 0 the six bases are u , v , − u , ( 1 + B ) u , ( 1 + B ) v , ( 1 + B ) u u,v,-u,(1+B)u,(1+B)v,(1+B)u u , v , − u , ( 1 + B ) u , ( 1 + B ) v , ( 1 + B ) u , nonzero as s s s avoids the axes. Hence we can fix ν 0 ∈ ( 0 , 1 ) \nu_0\in(0,1) ν 0 ∈ ( 0 , 1 ) and κ , q ∗ > 0 \kappa,q_*>0 κ , q ∗ > 0 such that on U = U ( ν 0 ) \mathcal{U}=\mathcal{U}(\nu_0) U = U ( ν 0 ) the six numbers u , v , w , ( 1 + B ) u , ( 1 + B ) v , u − B w u,v,w,(1+B)u,(1+B)v,u-Bw u , v , w , ( 1 + B ) u , ( 1 + B ) v , u − B w have absolute value at least κ \kappa κ , and ∂ ν 2 F 0 ≥ 4 q ∗ \partial_\nu^2F_0\geq4q_* ∂ ν 2 F 0 ≥ 4 q ∗ (at ν = 0 \nu=0 ν = 0 , ∂ ν 2 F 0 = p ( p − 1 ) B ( 1 + B ) p − 2 ∣ u ∣ p − 2 \partial_\nu^2F_0=p(p-1)B(1+B)^{p-2}|u|^{p-2} ∂ ν 2 F 0 = p ( p − 1 ) B ( 1 + B ) p − 2 ∣ u ∣ p − 2 , by the proof of Lemma 3.3). For a > 0 a>0 a > 0 the bases ( 1 + B ) u + a v (1+B)u+av ( 1 + B ) u + a v , ( 1 + B ) v + a w (1+B)v+aw ( 1 + B ) v + a w and u + a v − B w u+av-Bw u + a v − B w differ from ( 1 + B ) u , ( 1 + B ) v (1+B)u,(1+B)v ( 1 + B ) u , ( 1 + B ) v and u − B w u-Bw u − B w by a v , a w av,aw a v , a w and a v av a v ; as ∣ v ∣ ≤ 1 |v|\leq1 ∣ v ∣ ≤ 1 and ∣ w ∣ ≤ 2 |w|\leq2 ∣ w ∣ ≤ 2 on U \mathcal{U} U , these differences are at most 2 a 2a 2 a in absolute value. Therefore all six bases have absolute value at least κ / 2 \kappa/2 κ /2 on U \mathcal{U} U once a ≤ κ / 4 a\leq\kappa/4 a ≤ κ /4 , and we shrink a 1 a_1 a 1 to at most κ / 4 \kappa/4 κ /4 . The terms k ( 1 ) ( u , v ) k^{(1)}(u,v) k ( 1 ) ( u , v ) and k ( 1 ) ( v , w ) k^{(1)}(v,w) k ( 1 ) ( v , w ) are also real-analytic on U \mathcal{U} U , as u 2 + v 2 = 1 u^2+v^2=1 u 2 + v 2 = 1 and v 2 + w 2 ≥ w 2 ≥ κ 2 v^2+w^2\geq w^2\geq\kappa^2 v 2 + w 2 ≥ w 2 ≥ κ 2 there.
Step 4: the derivative across Π \Pi Π . At the windows s ^ \hat{s} s ^ , s ∈ O a s\in\mathcal{O}_a s ∈ O a , F a = 1 4 Σ a = 0 F_a=\frac{1}{4}\Sigma_a=0 F a = 4 1 Σ a = 0 by (6) and Lemma 3.5(i); since we need F a ≥ 0 F_a\geq0 F a ≥ 0 , its gradient must vanish there. Along Π \Pi Π it does, by (6); across Π \Pi Π we impose it by the choice of λ ( a ) \lambda(a) λ ( a ) . Let N a ( s ) = ∂ ν F a ( s ^ + ν e 3 ) ∣ ν = 0 N_a(s)=\partial_\nu F_a(\hat{s}+\nu e_3)|_{\nu=0} N a ( s ) = ∂ ν F a ( s ^ + ν e 3 ) ∣ ν = 0 . The term k a ( u , v ) − k a ( u , v ) k_a(u,v)-k_a(u,v) k a ( u , v ) − k a ( u , v ) contributes ∂ 2 k a ( J s ) \partial_2k_a(Js) ∂ 2 k a ( J s ) to N a ( s ) N_a(s) N a ( s ) , where ∂ 1 , ∂ 2 \partial_1,\partial_2 ∂ 1 , ∂ 2 are the partial derivatives in the first and second argument. For a J J J -invariant C 1 C^1 C 1 function k k k (such as k a k_a k a ), differentiating k ( J ξ ) = k ( ξ ) k(J\xi)=k(\xi) k ( J ξ ) = k ( ξ ) gives ∇ k ( J ξ ) = J ∇ k ( ξ ) \nabla k(J\xi)=J\nabla k(\xi) ∇ k ( J ξ ) = J ∇ k ( ξ ) (as J t = J − 1 J^t=J^{-1} J t = J − 1 ), so this contribution is − ∂ 1 k a ( s ) -\partial_1k_a(s) − ∂ 1 k a ( s ) , and its contribution to N a ( J s ) N_a(Js) N a ( J s ) is − ∂ 1 k a ( J s ) = − ∂ 2 k a ( s ) -\partial_1k_a(Js)=-\partial_2k_a(s) − ∂ 1 k a ( J s ) = − ∂ 2 k a ( s ) . Write N a 0 N_a^0 N a 0 for N a N_a N a with k a = 0 k_a=0 k a = 0 . We choose λ ( a ) \lambda(a) λ ( a ) so that
∇ k a ( s a ) = ( N a 0 ( s a ) , N a 0 ( J s a ) ) , i.e. N a ( s a ) = N a ( J s a ) = 0. \nabla k_a(s_a)=\bigl(N_a^0(s_a),N_a^0(Js_a)\bigr),\qquad\text{i.e. }N_a(s_a)=N_a(Js_a)=0. ∇ k a ( s a ) = ( N a 0 ( s a ) , N a 0 ( J s a ) ) , i.e. N a ( s a ) = N a ( J s a ) = 0.
This is a 2 × 2 2\times2 2 × 2 linear system. Here s 0 = e ∗ s_0=e_* s 0 = e ∗ is a positive multiple of ( 1 , t ∗ ) (1,t_*) ( 1 , t ∗ ) ; at ( 1 , t ∗ ) (1,t_*) ( 1 , t ∗ ) the gradients of k ( 1 ) , k ( 2 ) k^{(1)},k^{(2)} k ( 1 ) , k ( 2 ) are p ( 1 + t ∗ 2 ) ( p − 2 ) / 2 ( 1 , t ∗ ) p(1+t_*^2)^{(p-2)/2}(1,t_*) p ( 1 + t ∗ 2 ) ( p − 2 ) /2 ( 1 , t ∗ ) and p ( 1 , t ∗ p − 1 ) p(1,t_*^{p-1}) p ( 1 , t ∗ p − 1 ) , and at s 0 s_0 s 0 they are positive multiples of these, so they are independent because 0 < t ∗ < 1 0<t_*<1 0 < t ∗ < 1 and p > 2 p>2 p > 2 . Moreover N a 0 = ∂ ν F 0 = 0 N_a^0=\partial_\nu F_0=0 N a 0 = ∂ ν F 0 = 0 on all of Π \Pi Π (F 0 ≥ 0 F_0\geq0 F 0 ≥ 0 vanishes on Π \Pi Π ), and N a 0 N_a^0 N a 0 and s a s_a s a are real-analytic near a = 0 a=0 a = 0 (Lemma 3.5(iv)); so, after strictly decreasing a 1 a_1 a 1 , the system is uniquely solvable for 0 ≤ a ≤ a 1 0\leq a\leq a_1 0 ≤ a ≤ a 1 , λ ( a ) \lambda(a) λ ( a ) is real-analytic there, λ ( 0 ) = 0 \lambda(0)=0 λ ( 0 ) = 0 , and ∣ λ ( a ) ∣ ≤ Λ a |\lambda(a)|\leq\Lambda a ∣ λ ( a ) ∣ ≤ Λ a there, for some constant Λ \Lambda Λ . As F a F_a F a is even, N a ( − s ) = − N a ( s ) N_a(-s)=-N_a(s) N a ( − s ) = − N a ( s ) , so N a = 0 N_a=0 N a = 0 on all of O a O_a O a .
Step 5: a uniform C 1 C^1 C 1 bound. Put E a = F a − F 0 E_a=F_a-F_0 E a = F a − F 0 , i.e.
E a = ( C a p − ( 1 + B ) p ) ∣ u ∣ p + 1 2 ( Δ a ( v , w ) − Δ a ( u , v ) ) + k a ( v , w ) − k a ( u , v ) − ( ∣ u + a v − B w ∣ p − ∣ u − B w ∣ p ) . E_a=\left(C_a^p-(1+B)^p\right)|u|^p+\frac{1}{2}\left(\Delta_a(v,w)-\Delta_a(u,v)\right)+k_a(v,w)-k_a(u,v)-\left(|u+av-Bw|^p-|u-Bw|^p\right). E a = ( C a p − ( 1 + B ) p ) ∣ u ∣ p + 2 1 ( Δ a ( v , w ) − Δ a ( u , v ) ) + k a ( v , w ) − k a ( u , v ) − ( ∣ u + a v − B w ∣ p − ∣ u − B w ∣ p ) .
With ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ the maximum norm on R 3 \mathbb{R}^3 R 3 there is L L L , depending on p p p , B B B and Λ \Lambda Λ , with
∣ E a ( x ) ∣ ≤ L a ∥ x ∥ ∞ p , ∣ ∇ E a ( x ) ∣ ≤ L a ∥ x ∥ ∞ p − 1 (windows x = ( u , v , w ) ∈ R 3 , 0 ≤ a < a 1 ). (7) |E_a(x)|\leq La\|x\|_\infty^p,\qquad|\nabla E_a(x)|\leq La\|x\|_\infty^{p-1}\qquad\text{(windows }\mathbf{x}=(u,v,w)\in\mathbb{R}^3,\ 0\leq a<a_1\text{).} \tag*{(7)} ∣ E a ( x ) ∣ ≤ L a ∥ x ∥ ∞ p , ∣∇ E a ( x ) ∣ ≤ L a ∥ x ∥ ∞ p − 1 (windows x = ( u , v , w ) ∈ R 3 , 0 ≤ a < a 1 ). (7)
Indeed, ∣ C a − ( 1 + B ) ∣ ≤ a ∥ J ∥ p = a |C_a-(1+B)|\leq a\|J\|_p=a ∣ C a − ( 1 + B ) ∣ ≤ a ∥ J ∥ p = a , so ∣ C a p − ( 1 + B ) p ∣ ≤ p ( 1 + B + a ) p − 1 a |C_a^p-(1+B)^p|\leq p(1+B+a)^{p-1}a ∣ C a p − ( 1 + B ) p ∣ ≤ p ( 1 + B + a ) p − 1 a by the mean value theorem. By the mean value theorem, ∣ ∣ y + h ∣ p − ∣ y ∣ p ∣ ≤ p ( ∣ y ∣ + ∣ h ∣ ) p − 1 ∣ h ∣ ||y+h|^p-|y|^p|\leq p(|y|+|h|)^{p-1}|h| ∣∣ y + h ∣ p − ∣ y ∣ p ∣ ≤ p ( ∣ y ∣ + ∣ h ∣ ) p − 1 ∣ h ∣ and (as p > 2 p>2 p > 2 ) ∣ ϕ ( y + h ) − ϕ ( y ) ∣ ≤ ( p − 1 ) ( ∣ y ∣ + ∣ h ∣ ) p − 2 ∣ h ∣ |\phi(y+h)-\phi(y)|\leq(p-1)(|y|+|h|)^{p-2}|h| ∣ ϕ ( y + h ) − ϕ ( y ) ∣ ≤ ( p − 1 ) ( ∣ y ∣ + ∣ h ∣ ) p − 2 ∣ h ∣ ; with y = u − B w y=u-Bw y = u − B w , h = a v h=av h = a v , ∣ y ∣ ≤ ( 1 + B ) ∥ x ∥ ∞ |y|\leq(1+B)\|x\|_\infty ∣ y ∣ ≤ ( 1 + B ) ∥ x ∥ ∞ and ∣ h ∣ ≤ a ∥ x ∥ ∞ |h|\leq a\|x\|_\infty ∣ h ∣ ≤ a ∥ x ∥ ∞ , these bound ∣ u + a v − B w ∣ p − ∣ u − B w ∣ p |u+av-Bw|^p-|u-Bw|^p ∣ u + a v − B w ∣ p − ∣ u − B w ∣ p and its gradient (whose remaining part, a p ϕ ( u + a v − B w ) ap\phi(u+av-Bw) a pϕ ( u + a v − B w ) in the v v v -direction, is at most a p ( 1 + B + a ) p − 1 ∥ x ∥ ∞ p − 1 ap(1+B+a)^{p-1}\|x\|_\infty^{p-1} a p ( 1 + B + a ) p − 1 ∥ x ∥ ∞ p − 1 ); after we write Δ a ( s ) = ( C a p − ( 1 + B ) p ) ∣ u ∣ p − ( ∣ ( 1 + B ) u + a v ∣ p − ∣ ( 1 + B ) u ∣ p ) \Delta_a(s)=\left(C_a^p-(1+B)^p\right)|u|^p-\left(|(1+B)u+av|^p-|(1+B)u|^p\right) Δ a ( s ) = ( C a p − ( 1 + B ) p ) ∣ u ∣ p − ( ∣ ( 1 + B ) u + a v ∣ p − ∣ ( 1 + B ) u ∣ p ) , they also bound the terms with Δ a \Delta_a Δ a . The terms with k a k_a k a are bounded because ∣ λ ( a ) ∣ ≤ Λ a |\lambda(a)|\leq\Lambda a ∣ λ ( a ) ∣ ≤ Λ a .
Step 6: F a ≥ 0 F_a\geq0 F a ≥ 0 . Since F a F_a F a is homogeneous of degree p p p , it suffices to consider windows s ^ + ν e 3 \hat{s}+\nu e_3 s ^ + ν e 3 with ∥ s ∥ 2 = 1 \|s\|_2=1 ∥ s ∥ 2 = 1 or s = 0 s=0 s = 0 . If s = 0 s=0 s = 0 , F a ( 0 , 0 , w ) = F 0 ( 0 , 0 , w ) + E a ( 0 , 0 , w ) ≥ ( B ( 1 + B ) p − 1 − B p − L a ) ∣ w ∣ p ≥ 0 F_a(0,0,w)=F_0(0,0,w)+E_a(0,0,w)\geq\left(B(1+B)^{p-1}-B^p-La\right)|w|^p\geq0 F a ( 0 , 0 , w ) = F 0 ( 0 , 0 , w ) + E a ( 0 , 0 , w ) ≥ ( B ( 1 + B ) p − 1 − B p − L a ) ∣ w ∣ p ≥ 0 once L a ≤ B ( 1 + B ) p − 1 − B p La\leq B(1+B)^{p-1}-B^p L a ≤ B ( 1 + B ) p − 1 − B p , a positive number as 1 + B > B 1+B>B 1 + B > B . If ∥ s ∥ 2 = 1 \|s\|_2=1 ∥ s ∥ 2 = 1 and x = s ^ + ν e 3 x=\hat{s}+\nu e_3 x = s ^ + ν e 3 , then F a ( x ) = F 0 ( x ) + E a ( s ^ ) + ( E a ( x ) − E a ( s ^ ) ) F_a(x)=F_0(x)+E_a(\hat{s})+(E_a(x)-E_a(\hat{s})) F a ( x ) = F 0 ( x ) + E a ( s ^ ) + ( E a ( x ) − E a ( s ^ )) with E a ( s ^ ) = F a ( s ^ ) = 1 4 Σ a ( s ) E_a(\hat{s})=F_a(\hat{s})=\frac{1}{4}\Sigma_a(s) E a ( s ^ ) = F a ( s ^ ) = 4 1 Σ a ( s ) ; along the segment from s ^ \hat{s} s ^ to x x x the maximum norm is at most 1 + ∣ ν ∣ 1+|\nu| 1 + ∣ ν ∣ , so (7) and Lemma 3.3 give
F a ( x ) ≥ 1 4 Σ a ( s ) + γ 1 ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 − L a ( 1 + ∣ ν ∣ ) p − 1 ∣ ν ∣ . (8) F_a(x)\geq\frac{1}{4}\Sigma_a(s)+\gamma_1(|u|+|\nu|)^{p-2}\nu^2-La(1+|\nu|)^{p-1}|\nu|. \tag*{(8)} F a ( x ) ≥ 4 1 Σ a ( s ) + γ 1 ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 − L a ( 1 + ∣ ν ∣ ) p − 1 ∣ ν ∣. (8)
Case 1: s s s at angular distance ≥ ρ \geq\rho ≥ ρ from O a O_a O a . (This case contains s = ( 0 , ± 1 ) s=(0,\pm1) s = ( 0 , ± 1 ) and s = ( ± 1 , 0 ) s=(\pm1,0) s = ( ± 1 , 0 ) , where u = 0 u=0 u = 0 or v = 0 v=0 v = 0 and F a F_a F a need not be real-analytic; the argument uses only (8).) Let ν 1 = min { 1 , δ / ( 2 p + 2 L ) } \nu_1=\min\{1,\delta/(2^{p+2}L)\} ν 1 = min { 1 , δ / ( 2 p + 2 L )} . For ∣ ν ∣ ≤ ν 1 |\nu|\leq\nu_1 ∣ ν ∣ ≤ ν 1 , (8) and Lemma 3.5(ii) give F a ≥ a δ / 4 − 2 p − 1 L a ν 1 ≥ a δ / 8 F_a\geq a\delta/4-2^{p-1}La\nu_1\geq a\delta/8 F a ≥ a δ /4 − 2 p − 1 L a ν 1 ≥ a δ /8 . For ∣ ν ∣ ≥ ν 1 |\nu|\geq\nu_1 ∣ ν ∣ ≥ ν 1 , ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 ≥ ∣ ν ∣ p (|u|+|\nu|)^{p-2}\nu^2\geq|\nu|^p ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 ≥ ∣ ν ∣ p and ( 1 + ∣ ν ∣ ) p − 1 ≤ ( 1 + 1 / ν 1 ) p − 1 ∣ ν ∣ p − 1 (1+|\nu|)^{p-1}\leq(1+1/\nu_1)^{p-1}|\nu|^{p-1} ( 1 + ∣ ν ∣ ) p − 1 ≤ ( 1 + 1/ ν 1 ) p − 1 ∣ ν ∣ p − 1 , so F a ≥ ∣ ν ∣ p ( γ 1 − L a ( 1 + 1 / ν 1 ) p − 1 ) ≥ 0 F_a\geq|\nu|^p\left(\gamma_1-La(1+1/\nu_1)^{p-1}\right)\geq0 F a ≥ ∣ ν ∣ p ( γ 1 − L a ( 1 + 1/ ν 1 ) p − 1 ) ≥ 0 for small a a a .
Case 2: s s s within angular distance ρ \rho ρ of O a O_a O a . Write r = r ( s ) r=r(s) r = r ( s ) . As O a O_a O a lies within ρ / 2 \rho/2 ρ /2 of O 0 O_0 O 0 , s s s lies within 3 ρ / 2 < 2 ρ 3\rho/2<2\rho 3 ρ /2 < 2 ρ of O 0 O_0 O 0 , so for ∣ ν ∣ ≤ ν 0 |\nu|\leq\nu_0 ∣ ν ∣ ≤ ν 0 the point ( s , ν ) (s,\nu) ( s , ν ) is in U U U , where all bases of all p p p -th powers in F a F_a F a are at least κ / 2 \kappa/2 κ /2 in absolute value (Step 3). Then F a F_a F a is real-analytic in ( a , s , ν ) (a,s,\nu) ( a , s , ν ) there, and by Taylor’s formula in ν \nu ν ,
F a ( s ^ + ν e 3 ) = 1 4 Σ a ( s ) + ν N a ( s ) + ν 2 q a ( s , ν ) , q a ( s , ν ) = ∫ 0 1 ( 1 − τ ) ∂ ν 2 F a ( s ^ + τ ν e 3 ) d τ . F_a(\hat{s}+\nu e_3)=\frac{1}{4}\Sigma_a(s)+\nu N_a(s)+\nu^2q_a(s,\nu),\qquad q_a(s,\nu)=\int_0^1(1-\tau)\partial_\nu^2F_a(\hat{s}+\tau\nu e_3)\,d\tau. F a ( s ^ + ν e 3 ) = 4 1 Σ a ( s ) + ν N a ( s ) + ν 2 q a ( s , ν ) , q a ( s , ν ) = ∫ 0 1 ( 1 − τ ) ∂ ν 2 F a ( s ^ + τ ν e 3 ) d τ .
By Step 3, ∂ ν 2 F 0 ≥ 4 q ∗ \partial_\nu^2F_0\geq4q_* ∂ ν 2 F 0 ≥ 4 q ∗ on U U U ; as F a F_a F a is real-analytic in ( a , s , ν ) (a,s,\nu) ( a , s , ν ) on the compact set [ 0 , a 1 ] × U [0,a_1]\times U [ 0 , a 1 ] × U , there is L 2 L_2 L 2 with ∣ ∂ ν 2 ( F a − F 0 ) ∣ ≤ L 2 a |\partial_\nu^2(F_a-F_0)|\leq L_2a ∣ ∂ ν 2 ( F a − F 0 ) ∣ ≤ L 2 a there, so q a ≥ q ∗ q_a\geq q_* q a ≥ q ∗ on U U U once L 2 a ≤ 2 q ∗ L_2a\leq2q_* L 2 a ≤ 2 q ∗ . Since N 0 = N 0 ′ = 0 N_0=N_0'=0 N 0 = N 0 ′ = 0 on all of Π \Pi Π (Step 4, as λ ( 0 ) = 0 \lambda(0)=0 λ ( 0 ) = 0 ) and N a N_a N a is real-analytic in a a a , the derivative of N a N_a N a along the unit circle is O ( a ) O(a) O ( a ) on the 2 ρ 2\rho 2 ρ -arcs. As N a N_a N a vanishes on O a O_a O a , there is a constant L 1 L_1 L 1 with ∣ N a ( s ) ∣ ≤ L 1 a r |N_a(s)|\leq L_1ar ∣ N a ( s ) ∣ ≤ L 1 a r . With Lemma 3.5(iii),
F a ≥ 1 4 a n r 2 − L 1 a r ∣ ν ∣ + q ∗ ν 2 ≥ 0 once ( L 1 a ) 2 ≤ a n q ∗ , F_a\geq\frac{1}{4}anr^2-L_1ar|\nu|+q_*\nu^2\geq0\qquad\text{once }(L_1a)^2\leq anq_*, F a ≥ 4 1 an r 2 − L 1 a r ∣ ν ∣ + q ∗ ν 2 ≥ 0 once ( L 1 a ) 2 ≤ an q ∗ ,
since then the discriminant ( L 1 a ) 2 − 4 ⋅ 1 4 a n ⋅ q ∗ (L_1a)^2-4\cdot\frac{1}{4}an\cdot q_* ( L 1 a ) 2 − 4 ⋅ 4 1 an ⋅ q ∗ of this quadratic form in ( r , ∣ ν ∣ ) (r,|\nu|) ( r , ∣ ν ∣ ) is ≤ 0 \leq0 ≤ 0 . For ∣ ν ∣ > ν 0 |\nu|>\nu_0 ∣ ν ∣ > ν 0 , (8) with Σ a ≥ 0 \Sigma_a\geq0 Σ a ≥ 0 , ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 ≥ ∣ ν ∣ p (|u|+|\nu|)^{p-2}\nu^2\geq|\nu|^p ( ∣ u ∣ + ∣ ν ∣ ) p − 2 ν 2 ≥ ∣ ν ∣ p and ( 1 + ∣ ν ∣ ) p − 1 ≤ ( 1 + 1 / ν 0 ) p − 1 ∣ ν ∣ p − 1 (1+|\nu|)^{p-1}\leq(1+1/\nu_0)^{p-1}|\nu|^{p-1} ( 1 + ∣ ν ∣ ) p − 1 ≤ ( 1 + 1/ ν 0 ) p − 1 ∣ ν ∣ p − 1 gives F a ≥ ∣ ν ∣ p ( γ 1 − L a ( 1 + 1 / ν 0 ) p − 1 ) ≥ 0 F_a\geq|\nu|^p(\gamma_1-La(1+1/\nu_0)^{p-1})\geq0 F a ≥ ∣ ν ∣ p ( γ 1 − L a ( 1 + 1/ ν 0 ) p − 1 ) ≥ 0 for small a a a .
Order of the constants. The constants are chosen in this order: γ 1 \gamma_1 γ 1 (Lemma 3.3); ρ \rho ρ , m m m , δ \delta δ , η \eta η and a 1 a_1 a 1 (Lemma 3.5); ν 0 \nu_0 ν 0 , κ \kappa κ , q ∗ q_* q ∗ (Step 3); Λ \Lambda Λ (Step 4, after the last shrinking of a 1 a_1 a 1 ); L L L (Step 5); L 1 L_1 L 1 , L 2 L_2 L 2 , ν 1 \nu_1 ν 1 (Step 6); finally, a 0 a_0 a 0 is the smallest of a 1 a_1 a 1 and the finitely many upper bounds on a a a imposed in Step 6. For 0 < a < a 0 0<a<a_0 0 < a < a 0 , (2) holds with C = C a C=C_a C = C a , and Lemma 3.1 gives ∥ P a ( S ) ∥ p ≤ C a = ∥ P a ( J ) ∥ p \|P_a(S)\|_p\leq C_a=\|P_a(J)\|_p ∥ P a ( S ) ∥ p ≤ C a = ∥ P a ( J ) ∥ p .
The case 1 < p < 2 1<p<2 1 < p < 2 . By Lemma 2.3 and the case p ′ > 2 p'>2 p ′ > 2 , ∥ P a ( S ) ∥ p = ∥ P a ( S ) ∥ p ′ = ∥ P a ( J ) ∥ p ′ = ∥ P a ( J t ) ∥ p \|P_a(S)\|_p=\|P_a(S)\|_{p'}=\|P_a(J)\|_{p'}=\|P_a(J^t)\|_p ∥ P a ( S ) ∥ p = ∥ P a ( S ) ∥ p ′ = ∥ P a ( J ) ∥ p ′ = ∥ P a ( J t ) ∥ p for 0 ≤ a < a 0 ( p ′ , B ) 0\leq a<a_0(p',B) 0 ≤ a < a 0 ( p ′ , B ) . Since J t = − J = V J V J^t=-J=VJV J t = − J = V J V with the isometry V = diag ( 1 , − 1 ) V=\operatorname{diag}(1,-1) V = diag ( 1 , − 1 ) of ℓ 2 p \ell_2^p ℓ 2 p , P a ( J t ) = V P a ( J ) V P_a(J^t)=VP_a(J)V P a ( J t ) = V P a ( J ) V has the norm of P a ( J ) P_a(J) P a ( J ) . □
A scaled rotation by more than a right angle
Lemma 4.1 (a contraction T ε T_\varepsilon T ε with ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p>\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p ). Let p > 2 p>2 p > 2 , B > 1 B>1 B > 1 , K ∗ = 1 2 ( J − D ∗ I ) K_*=\frac{1}{2}(J-D_*I) K ∗ = 2 1 ( J − D ∗ I ) and T ε = exp ( ε K ∗ ) J T_\varepsilon=\exp(\varepsilon K_*)J T ε = exp ( ε K ∗ ) J . Then ∥ T ε ∥ p ≤ 1 \|T_\varepsilon\|_p\leq1 ∥ T ε ∥ p ≤ 1 for all ε ≥ 0 \varepsilon\geq0 ε ≥ 0 , on both the real and the complex ℓ 2 p \ell_2^p ℓ 2 p , and there is a 2 > 0 a_2>0 a 2 > 0 such that for each 0 < a < a 2 0<a<a_2 0 < a < a 2 , ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p>\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p for all sufficiently small ε > 0 \varepsilon>0 ε > 0 (depending on a a a ).
Proof. The matrix T ε T_\varepsilon T ε is a contraction. For real ξ \xi ξ and ξ ( ε ) = exp ( ε K ∗ ) ξ \xi(\varepsilon)=\exp(\varepsilon K_*)\xi ξ ( ε ) = exp ( ε K ∗ ) ξ ,
d d ε ∥ ξ ( ε ) ∥ p p = p ⟨ ϕ ( ξ ( ε ) ) , K ∗ ξ ( ε ) ⟩ = p 2 ∥ ξ ( ε ) ∥ p p ( Q ( ξ ( ε ) ) − D ∗ ) ≤ 0. \frac{d}{d\varepsilon}\|\xi(\varepsilon)\|_p^p=p\langle\phi(\xi(\varepsilon)),K_*\xi(\varepsilon)\rangle=\frac{p}{2}\|\xi(\varepsilon)\|_p^p\bigl(Q(\xi(\varepsilon))-D_*\bigr)\leq0. d ε d ∥ ξ ( ε ) ∥ p p = p ⟨ ϕ ( ξ ( ε )) , K ∗ ξ ( ε )⟩ = 2 p ∥ ξ ( ε ) ∥ p p ( Q ( ξ ( ε )) − D ∗ ) ≤ 0.
Hence exp ( ε K ∗ ) \exp(\varepsilon K_*) exp ( ε K ∗ ) , and so T ε = exp ( ε K ∗ ) J T_\varepsilon=\exp(\varepsilon K_*)J T ε = exp ( ε K ∗ ) J (J J J is an isometry), is a contraction of the real ℓ 2 p \ell_2^p ℓ 2 p , and by Lemma 2.1 also of the complex ℓ 2 p \ell_2^p ℓ 2 p . (As exp ( ε K ∗ ) = e − ε D ∗ / 2 R ε / 2 \exp(\varepsilon K_*)=e^{-\varepsilon D_*/2}R_{\varepsilon/2} exp ( ε K ∗ ) = e − ε D ∗ /2 R ε /2 , T ε = e − ε D ∗ / 2 R π / 2 + ε / 2 T_\varepsilon=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2} T ε = e − ε D ∗ /2 R π /2 + ε /2 .)
Choice of test vector. By Lemma 3.5(i), (iv), for 0 < a < a 1 0<a<a_1 0 < a < a 1 the real maximisers of ∥ P a ( J ) ξ ∥ p / ∥ ξ ∥ p \|P_a(J)\xi\|_p/\|\xi\|_p ∥ P a ( J ) ξ ∥ p /∥ ξ ∥ p are exactly the nonzero real multiples of s a s_a s a and J s a Js_a J s a , with s a s_a s a real-analytic in a a a ; we take ξ a = s a / ( s a ) 1 \xi_a=s_a/(s_a)_1 ξ a = s a / ( s a ) 1 , so ξ a = ξ 0 + a ξ ˙ + o ( a ) \xi_a=\xi_0+a\dot{\xi}+o(a) ξ a = ξ 0 + a ξ ˙ + o ( a ) with ξ 0 = ( 1 , t ∗ ) \xi_0=(1,t_*) ξ 0 = ( 1 , t ∗ ) and ξ ˙ = d d a ξ a ∣ a = 0 \dot{\xi}=\left.\frac{d}{da}\xi_a\right|_{a=0} ξ ˙ = d a d ξ a a = 0 . Put y a = P a ( J ) ξ a y_a=P_a(J)\xi_a y a = P a ( J ) ξ a .
Derivative in ε \varepsilon ε . Recall P a ( T ε ) = I + a T ε − B T ε 2 P_a(T_\varepsilon)=I+aT_\varepsilon-BT_\varepsilon^2 P a ( T ε ) = I + a T ε − B T ε 2 and T 0 = J T_0=J T 0 = J . Since d d ε T ε ∣ 0 = K ∗ J \left.\frac{d}{d\varepsilon}T_\varepsilon\right|_0=K_*J d ε d T ε 0 = K ∗ J and d d ε T ε 2 ∣ 0 = K ∗ J 2 + J K ∗ J = − K ∗ + J K ∗ J \left.\frac{d}{d\varepsilon}T_\varepsilon^2\right|_0=K_*J^2+JK_*J=-K_*+JK_*J d ε d T ε 2 0 = K ∗ J 2 + J K ∗ J = − K ∗ + J K ∗ J ,
d d ε ∥ P a ( T ε ) ξ a ∥ p p ∣ ε = 0 = p Γ ( a ) , Γ ( a ) = ⟨ ϕ ( y a ) , M a ξ a ⟩ , M a = a K ∗ J + B ( K ∗ − J K ∗ J ) . \left.\frac{d}{d\varepsilon}\|P_a(T_\varepsilon)\xi_a\|_p^p\right|_{\varepsilon=0}=p\Gamma(a),\qquad\Gamma(a)=\langle\phi(y_a),M_a\xi_a\rangle,\qquad M_a=aK_*J+B(K_*-JK_*J). d ε d ∥ P a ( T ε ) ξ a ∥ p p ε = 0 = p Γ ( a ) , Γ ( a ) = ⟨ ϕ ( y a ) , M a ξ a ⟩ , M a = a K ∗ J + B ( K ∗ − J K ∗ J ) .
Since J 2 = − I J^2=-I J 2 = − I , K ∗ J = − 1 2 ( I + D ∗ J ) K_*J=-\frac{1}{2}(I+D_*J) K ∗ J = − 2 1 ( I + D ∗ J ) and K ∗ − J K ∗ J = J − D ∗ I K_*-JK_*J=J-D_*I K ∗ − J K ∗ J = J − D ∗ I . At a = 0 a=0 a = 0 , y 0 = ( 1 + B ) ξ 0 y_0=(1+B)\xi_0 y 0 = ( 1 + B ) ξ 0 and Γ ( 0 ) = ( 1 + B ) p − 1 B ∥ ξ 0 ∥ p p ( Q ( ξ 0 ) − D ∗ ) = 0 \Gamma(0)=(1+B)^{p-1}B\|\xi_0\|_p^p(Q(\xi_0)-D_*)=0 Γ ( 0 ) = ( 1 + B ) p − 1 B ∥ ξ 0 ∥ p p ( Q ( ξ 0 ) − D ∗ ) = 0 . Differentiating Γ ( a ) = ⟨ ϕ ( y a ) , M a ξ a ⟩ \Gamma(a)=\langle\phi(y_a),M_a\xi_a\rangle Γ ( a ) = ⟨ ϕ ( y a ) , M a ξ a ⟩ at a = 0 a=0 a = 0 , with d d a y a ∣ 0 = ( 1 + B ) ξ ˙ + J ξ 0 \left.\frac{d}{da}y_a\right|_0=(1+B)\dot{\xi}+J\xi_0 d a d y a 0 = ( 1 + B ) ξ ˙ + J ξ 0 , d ϕ ( y ) = ( p − 1 ) ∣ y ∣ p − 2 d y d\phi(y)=(p-1)|y|^{p-2}dy d ϕ ( y ) = ( p − 1 ) ∣ y ∣ p − 2 d y , d d a M a = K ∗ J \frac{d}{da}M_a=K_*J d a d M a = K ∗ J and M 0 = B ( J − D ∗ I ) M_0=B(J-D_*I) M 0 = B ( J − D ∗ I ) , gives
Γ ′ ( 0 ) = ( p − 1 ) ( 1 + B ) p − 2 B ⟨ ∣ ξ 0 ∣ p − 2 ( ( 1 + B ) ξ ˙ + J ξ 0 ) , ( J − D ∗ I ) ξ 0 ⟩ + ( 1 + B ) p − 1 ⟨ ϕ ( ξ 0 ) , K ∗ J ξ 0 ⟩ + ( 1 + B ) p − 1 B ⟨ ϕ ( ξ 0 ) , ( J − D ∗ I ) ξ ˙ ⟩ . \Gamma'(0)=(p-1)(1+B)^{p-2}B\left\langle|\xi_0|^{p-2}\bigl((1+B)\dot{\xi}+J\xi_0\bigr),(J-D_*I)\xi_0\right\rangle
+(1+B)^{p-1}\langle\phi(\xi_0),K_*J\xi_0\rangle+(1+B)^{p-1}B\langle\phi(\xi_0),(J-D_*I)\dot{\xi}\rangle. Γ ′ ( 0 ) = ( p − 1 ) ( 1 + B ) p − 2 B ⟨ ∣ ξ 0 ∣ p − 2 ( ( 1 + B ) ξ ˙ + J ξ 0 ) , ( J − D ∗ I ) ξ 0 ⟩ + ( 1 + B ) p − 1 ⟨ ϕ ( ξ 0 ) , K ∗ J ξ 0 ⟩ + ( 1 + B ) p − 1 B ⟨ ϕ ( ξ 0 ) , ( J − D ∗ I ) ξ ˙ ⟩ .
The two terms containing ξ ˙ \dot{\xi} ξ ˙ add up to ( 1 + B ) p − 1 B (1+B)^{p-1}B ( 1 + B ) p − 1 B times the derivative at ξ 0 \xi_0 ξ 0 , in the direction ξ ˙ \dot{\xi} ξ ˙ , of Ψ ( ξ ) = ⟨ ϕ ( ξ ) , ( J − D ∗ I ) ξ ⟩ = ∥ ξ ∥ p p ( Q ( ξ ) − D ∗ ) \Psi(\xi)=\langle\phi(\xi),(J-D_*I)\xi\rangle=\|\xi\|_p^p(Q(\xi)-D_*) Ψ ( ξ ) = ⟨ ϕ ( ξ ) , ( J − D ∗ I ) ξ ⟩ = ∥ ξ ∥ p p ( Q ( ξ ) − D ∗ ) . As Ψ ≤ 0 \Psi\leq0 Ψ ≤ 0 and Ψ ( ξ 0 ) = 0 \Psi(\xi_0)=0 Ψ ( ξ 0 ) = 0 , ξ 0 \xi_0 ξ 0 is a maximum of Ψ \Psi Ψ , so this derivative is 0. To evaluate the remaining terms, put t = t ∗ t=t_* t = t ∗ . Then ∥ ξ 0 ∥ p p = 1 + t p \|\xi_0\|_p^p=1+t^p ∥ ξ 0 ∥ p p = 1 + t p , ⟨ ϕ ( ξ 0 ) , J ξ 0 ⟩ = D ∗ ( 1 + t p ) \langle\phi(\xi_0),J\xi_0\rangle=D_*(1+t^p) ⟨ ϕ ( ξ 0 ) , J ξ 0 ⟩ = D ∗ ( 1 + t p ) , and
E ∗ : = ⟨ ∣ ξ 0 ∣ p − 2 J ξ 0 , ( J − D ∗ I ) ξ 0 ⟩ = t 2 + t p − 2 − D ∗ ( t − t p − 1 ) = t p − 2 ( 1 + t 2 ) 2 1 + t p ; E_*:=\langle|\xi_0|^{p-2}J\xi_0,(J-D_*I)\xi_0\rangle=t^2+t^{p-2}-D_*(t-t^{p-1})=\frac{t^{p-2}(1+t^2)^2}{1+t^p}; E ∗ := ⟨ ∣ ξ 0 ∣ p − 2 J ξ 0 , ( J − D ∗ I ) ξ 0 ⟩ = t 2 + t p − 2 − D ∗ ( t − t p − 1 ) = 1 + t p t p − 2 ( 1 + t 2 ) 2 ;
the last equality follows from D ∗ = ( t − t p − 1 ) / ( 1 + t p ) D_*=(t-t^{p-1})/(1+t^p) D ∗ = ( t − t p − 1 ) / ( 1 + t p ) and the identity ( t 2 + t p − 2 ) ( 1 + t p ) − ( t − t p − 1 ) 2 = t p − 2 ( 1 + t 2 ) 2 (t^2+t^{p-2})(1+t^p)-(t-t^{p-1})^2=t^{p-2}(1+t^2)^2 ( t 2 + t p − 2 ) ( 1 + t p ) − ( t − t p − 1 ) 2 = t p − 2 ( 1 + t 2 ) 2 , valid for every t > 0 t>0 t > 0 . Using K ∗ J = − 1 2 ( I + D ∗ J ) K_*J=-\frac{1}{2}(I+D_*J) K ∗ J = − 2 1 ( I + D ∗ J ) , the remaining terms are ( p − 1 ) ( 1 + B ) p − 2 B E ∗ − 1 2 ( 1 + B ) p − 1 ( ∥ ξ 0 ∥ p p + D ∗ ⟨ ϕ ( ξ 0 ) , J ξ 0 ⟩ ) (p-1)(1+B)^{p-2}BE_*-\frac{1}{2}(1+B)^{p-1}\bigl(\|\xi_0\|_p^p+D_*\langle\phi(\xi_0),J\xi_0\rangle\bigr) ( p − 1 ) ( 1 + B ) p − 2 B E ∗ − 2 1 ( 1 + B ) p − 1 ( ∥ ξ 0 ∥ p p + D ∗ ⟨ ϕ ( ξ 0 ) , J ξ 0 ⟩ ) , that is
Γ ′ ( 0 ) = ( 1 + B ) p − 2 [ − 1 2 ( 1 + B ) ( 1 + t p ) ( 1 + D ∗ 2 ) + ( p − 1 ) B E ∗ ] . \Gamma'(0)=(1+B)^{p-2}\left[-\frac{1}{2}(1+B)(1+t^p)(1+D_*^2)+(p-1)BE_*\right]. Γ ′ ( 0 ) = ( 1 + B ) p − 2 [ − 2 1 ( 1 + B ) ( 1 + t p ) ( 1 + D ∗ 2 ) + ( p − 1 ) B E ∗ ] .
By (4), ( 1 + t p ) 2 ( 1 + D ∗ 2 ) = ( 1 + t p ) 2 + ( t − t p − 1 ) 2 = ( 1 + t 2 ) ( 1 + t 2 p − 2 ) = ( p − 1 ) t p − 2 ( 1 + t 2 ) 2 (1+t^p)^2(1+D_*^2)=(1+t^p)^2+(t-t^{p-1})^2=(1+t^2)(1+t^{2p-2})=(p-1)t^{p-2}(1+t^2)^2 ( 1 + t p ) 2 ( 1 + D ∗ 2 ) = ( 1 + t p ) 2 + ( t − t p − 1 ) 2 = ( 1 + t 2 ) ( 1 + t 2 p − 2 ) = ( p − 1 ) t p − 2 ( 1 + t 2 ) 2 , that is ( 1 + t p ) ( 1 + D ∗ 2 ) = ( p − 1 ) E ∗ (1+t^p)(1+D_*^2)=(p-1)E_* ( 1 + t p ) ( 1 + D ∗ 2 ) = ( p − 1 ) E ∗ . Therefore
Γ ( a ) = a ( 1 + B ) p − 2 ( p − 1 ) E ∗ B − 1 2 + o ( a ) , \Gamma(a)=a(1+B)^{p-2}(p-1)E_*\frac{B-1}{2}+o(a), Γ ( a ) = a ( 1 + B ) p − 2 ( p − 1 ) E ∗ 2 B − 1 + o ( a ) ,
which is positive for 0 < a < a 2 0<a<a_2 0 < a < a 2 , for some a 2 ∈ ( 0 , a 1 ] a_2\in(0,a_1] a 2 ∈ ( 0 , a 1 ] , when B > 1 B>1 B > 1 . For such a a a , p Γ ( a ) p\Gamma(a) p Γ ( a ) is positive, so for all small ε > 0 \varepsilon>0 ε > 0 , ∥ P a ( T ε ) ∥ p ≥ ∥ P a ( T ε ) ξ a ∥ p / ∥ ξ a ∥ p > ∥ P a ( J ) ξ a ∥ p / ∥ ξ a ∥ p = ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p\geq\|P_a(T_\varepsilon)\xi_a\|_p/\|\xi_a\|_p>\|P_a(J)\xi_a\|_p/\|\xi_a\|_p=\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p ≥ ∥ P a ( T ε ) ξ a ∥ p /∥ ξ a ∥ p > ∥ P a ( J ) ξ a ∥ p /∥ ξ a ∥ p = ∥ P a ( J ) ∥ p . □
Proofs of Theorems A and B
Proof of Theorem A. Let p > 2 p>2 p > 2 and B > 1 B>1 B > 1 . For 0 < a < min ( a 0 , a 2 ) 0<a<\min(a_0,a_2) 0 < a < min ( a 0 , a 2 ) , Lemma 4.1 gives a contraction T ε T_\varepsilon T ε of ℓ 2 \ell_2 ℓ 2 with ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p \|P_a(T_\varepsilon)\|_p>\|P_a(J)\|_p ∥ P a ( T ε ) ∥ p > ∥ P a ( J ) ∥ p , and Theorem 3.6 gives ∥ P a ( J ) ∥ p ≥ ∥ P a ( S ) ∥ p \|P_a(J)\|_p\ge\|P_a(S)\|_p ∥ P a ( J ) ∥ p ≥ ∥ P a ( S ) ∥ p . Since ℓ 2 p \ell_2^p ℓ 2 p is an L p L^p L p space, this contradicts (1) . This holds for real and for complex scalars alike: T ε T_\varepsilon T ε is a contraction of both spaces (Lemma 4.1), and the real operators P a ( T ε ) P_a(T_\varepsilon) P a ( T ε ) , P a ( J ) P_a(J) P a ( J ) and P a ( S ) P_a(S) P a ( S ) have the same norms on the real and on the complex spaces (Lemma 2.1). For 1 < p < 2 1<p<2 1 < p < 2 , apply this at p ′ > 2 p'>2 p ′ > 2 and use Lemma 2.3: T ε 2 T_\varepsilon^2 T ε 2 is a contraction of ℓ 2 p \ell_2^p ℓ 2 p and ∥ P a ( T ε 2 ) ∥ p = ∥ P a ( T ε ) ∥ p ′ > ∥ P a ( S ) ∥ p ′ = ∥ P a ( S ) ∥ p \|P_a(T_\varepsilon^2)\|_p=\|P_a(T_\varepsilon)\|_{p'}>\|P_a(S)\|_{p'}=\|P_a(S)\|_p ∥ P a ( T ε 2 ) ∥ p = ∥ P a ( T ε ) ∥ p ′ > ∥ P a ( S ) ∥ p ′ = ∥ P a ( S ) ∥ p . Finally, if T T T is any of these contractions, T ⊕ 0 T\oplus0 T ⊕ 0 is a contraction of ℓ 2 p ⊕ p ℓ p p = ℓ p p \ell_2^p\oplus_p\ell_p^p=\ell_p^p ℓ 2 p ⊕ p ℓ p p = ℓ p p with P a ( T ⊕ 0 ) = P a ( T ) ⊕ I P_a(T\oplus0)=P_a(T)\oplus I P a ( T ⊕ 0 ) = P a ( T ) ⊕ I , hence ∥ P a ( T ⊕ 0 ) ∥ p = max ( ∥ P a ( T ) ∥ p , 1 ) ≥ ∥ P a ( T ) ∥ p \|P_a(T\oplus0)\|_p=\max(\|P_a(T)\|_p,1)\ge\|P_a(T)\|_p ∥ P a ( T ⊕ 0 ) ∥ p = max ( ∥ P a ( T ) ∥ p , 1 ) ≥ ∥ P a ( T ) ∥ p , so (1) also fails on ℓ p p \ell_p^p ℓ p p itself. □
Proof of Theorem B. If c 0 = c 2 = 0 c_0=c_2=0 c 0 = c 2 = 0 , then P ( z ) = c 1 z P(z)=c_1z P ( z ) = c 1 z , and ∥ P ( T ) ∥ p ≤ ∣ c 1 ∣ = ∥ P ( S ) ∥ p \|P(T)\|_p\le|c_1|=\|P(S)\|_p ∥ P ( T ) ∥ p ≤ ∣ c 1 ∣ = ∥ P ( S ) ∥ p because S S S is an isometry. Otherwise choose ζ ∈ C \zeta\in\mathbb{C} ζ ∈ C with ∣ ζ ∣ = 1 |\zeta|=1 ∣ ζ ∣ = 1 and ζ 2 = − c 2 / c 0 \zeta^2=-c_2/c_0 ζ 2 = − c 2 / c 0 ; then P ( z ) = c 0 P ~ ( ζ z ) P(z)=c_0\widetilde P(\zeta z) P ( z ) = c 0 P ( ζ z ) with P ~ ( z ) = 1 + α z − z 2 \widetilde P(z)=1+\alpha z-z^2 P ( z ) = 1 + α z − z 2 , α = c 1 / ( c 0 ζ ) ∈ C \alpha=c_1/(c_0\zeta)\in\mathbb{C} α = c 1 / ( c 0 ζ ) ∈ C . Since ζ T \zeta T ζ T is a contraction and ζ S = U S U − 1 \zeta S=USU^{-1} ζ S = U S U − 1 with the isometry ( U x ) n = ζ n x n (Ux)_n=\zeta^n x_n ( U x ) n = ζ n x n (as in the Background), it suffices to treat P ~ \widetilde P P : if ∥ P ~ ( T ′ ) ∥ p ≤ ∥ P ~ ( S ) ∥ p \|\widetilde P(T')\|_p\le\|\widetilde P(S)\|_p ∥ P ( T ′ ) ∥ p ≤ ∥ P ( S ) ∥ p for every contraction T ′ T' T ′ , then ∥ P ( T ) ∥ p = ∣ c 0 ∣ ∥ P ~ ( ζ T ) ∥ p ≤ ∣ c 0 ∣ ∥ P ~ ( S ) ∥ p = ∣ c 0 ∣ ∥ U P ~ ( S ) U − 1 ∥ p = ∣ c 0 ∣ ∥ P ~ ( ζ S ) ∥ p = ∥ P ( S ) ∥ p \|P(T)\|_p=|c_0|\|\widetilde P(\zeta T)\|_p\le|c_0|\|\widetilde P(S)\|_p=|c_0|\|U\widetilde P(S)U^{-1}\|_p=|c_0|\|\widetilde P(\zeta S)\|_p=\|P(S)\|_p ∥ P ( T ) ∥ p = ∣ c 0 ∣∥ P ( ζ T ) ∥ p ≤ ∣ c 0 ∣∥ P ( S ) ∥ p = ∣ c 0 ∣∥ U P ( S ) U − 1 ∥ p = ∣ c 0 ∣∥ P ( ζ S ) ∥ p = ∥ P ( S ) ∥ p . For a contraction T T T on L p = L p ( Ω , m ) L^p=L^p(\Omega,m) L p = L p ( Ω , m ) and x ∈ L p x\in L^p x ∈ L p ,
P ~ ( T ) x = 1 2 [ ( 2 x + α T x ) + T ( α x − 2 T x ) ] . \widetilde P(T)x=\frac{1}{2}\bigl[(2x+\alpha Tx)+T(\alpha x-2Tx)\bigr]. P ( T ) x = 2 1 [ ( 2 x + α T x ) + T ( α x − 2 T x ) ] .
By convexity of ∣ ⋅ ∣ p |\cdot|^p ∣ ⋅ ∣ p (pointwise) and ∥ T ∥ p ≤ 1 \|T\|_p\le1 ∥ T ∥ p ≤ 1 ,
∥ P ~ ( T ) x ∥ p p ≤ 1 2 ∥ 2 x + α T x ∥ p p + 1 2 ∥ α x − 2 T x ∥ p p = 1 2 ∫ Ω ∥ M ( x ( ω ) , T x ( ω ) ) ∥ p p d m ( ω ) ≤ 1 2 ∥ M ∥ p p ( ∥ x ∥ p p + ∥ T x ∥ p p ) ≤ ∥ M ∥ p p ∥ x ∥ p p , \begin{aligned}
\|\widetilde P(T)x\|_p^p\le\frac{1}{2}\|2x+\alpha Tx\|_p^p+\frac{1}{2}\|\alpha x-2Tx\|_p^p=\frac{1}{2}\int_\Omega\|M(x(\omega),Tx(\omega))\|_p^p\,\mathrm{d}m(\omega) \\
&\le\frac{1}{2}\|M\|_p^p(\|x\|_p^p+\|Tx\|_p^p)\le\|M\|_p^p\|x\|_p^p,
\end{aligned} ∥ P ( T ) x ∥ p p ≤ 2 1 ∥2 x + α T x ∥ p p + 2 1 ∥ α x − 2 T x ∥ p p = 2 1 ∫ Ω ∥ M ( x ( ω ) , T x ( ω )) ∥ p p d m ( ω ) ≤ 2 1 ∥ M ∥ p p ( ∥ x ∥ p p + ∥ T x ∥ p p ) ≤ ∥ M ∥ p p ∥ x ∥ p p ,
where M = ( 2 α α − 2 ) M=\begin{pmatrix}2&\alpha\\\alpha&-2\end{pmatrix} M = ( 2 α α − 2 ) acts pointwise on ( x , T x ) (x,Tx) ( x , T x ) . As M = V P ~ ( J ) M=V\widetilde P(J) M = V P ( J ) with V = diag ( 1 , − 1 ) V=\operatorname{diag}(1,-1) V = diag ( 1 , − 1 ) , ∥ M ∥ p = ∥ P ~ ( J ) ∥ p ≤ ∥ P ~ ( S ) ∥ p \|M\|_p=\|\widetilde P(J)\|_p\le\|\widetilde P(S)\|_p ∥ M ∥ p = ∥ P ( J ) ∥ p ≤ ∥ P ( S ) ∥ p by Lemma 2.2. (For p = ∞ p=\infty p = ∞ , replace each average of p p p -th powers by a maximum and the integral by an essential supremum.) □
Corollary 5.1 (exact norms for ∣ c 0 ∣ = ∣ c 2 ∣ |c_0|=|c_2| ∣ c 0 ∣ = ∣ c 2 ∣ ). Let P ( z ) = c 0 + c 1 z + c 2 z 2 P(z)=c_0+c_1z+c_2z^2 P ( z ) = c 0 + c 1 z + c 2 z 2 with ∣ c 0 ∣ = ∣ c 2 ∣ > 0 |c_0|=|c_2|>0 ∣ c 0 ∣ = ∣ c 2 ∣ > 0 , and let ζ ∈ C \zeta\in\mathbb{C} ζ ∈ C satisfy ∣ ζ ∣ = 1 |\zeta|=1 ∣ ζ ∣ = 1 and ζ 2 = − c 2 / c 0 \zeta^2=-c_2/c_0 ζ 2 = − c 2 / c 0 . For 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ , ∥ P ( S ) ∥ p = ∥ P ( ζ ˉ J ) ∥ p = max ∣ μ ∣ = 1 ∥ P ( μ J ) ∥ p \|P(S)\|_p=\|P(\bar\zeta J)\|_p=\max_{|\mu|=1}\|P(\mu J)\|_p ∥ P ( S ) ∥ p = ∥ P ( ζ ˉ J ) ∥ p = max ∣ μ ∣ = 1 ∥ P ( μ J ) ∥ p .
Proof. With P ( z ) = c 0 P ~ ( ζ z ) P(z)=c_0\widetilde P(\zeta z) P ( z ) = c 0 P ( ζ z ) as in the proof of Theorem B, ∥ P ( S ) ∥ p = ∣ c 0 ∣ ∥ P ~ ( S ) ∥ p \|P(S)\|_p=|c_0|\|\widetilde P(S)\|_p ∥ P ( S ) ∥ p = ∣ c 0 ∣∥ P ( S ) ∥ p . The proof of Theorem B with T = S T=S T = S gives ∥ P ~ ( S ) ∥ p ≤ ∥ P ~ ( J ) ∥ p \|\widetilde P(S)\|_p\le\|\widetilde P(J)\|_p ∥ P ( S ) ∥ p ≤ ∥ P ( J ) ∥ p , and Lemma 2.2 gives equality; here c 0 P ~ ( J ) = P ( ζ ˉ J ) c_0\widetilde P(J)=P(\bar\zeta J) c 0 P ( J ) = P ( ζ ˉ J ) . For every ∣ μ ∣ = 1 |\mu|=1 ∣ μ ∣ = 1 , Lemma 2.2 applied to z ↦ P ( μ z ) z\mapsto P(\mu z) z ↦ P ( μ z ) gives ∥ P ( μ J ) ∥ p ≤ ∥ P ( μ S ) ∥ p = ∥ P ( S ) ∥ p \|P(\mu J)\|_p\le\|P(\mu S)\|_p=\|P(S)\|_p ∥ P ( μ J ) ∥ p ≤ ∥ P ( μ S ) ∥ p = ∥ P ( S ) ∥ p . (The factor ζ ˉ \bar\zeta ζ ˉ cannot be omitted: for P ( z ) = 1 + z 2 P(z)=1+z^2 P ( z ) = 1 + z 2 , P ( J ) = 0 P(J)=0 P ( J ) = 0 , while, with ζ = i \zeta=i ζ = i , ∥ P ( S ) ∥ p = ∥ P ( − i J ) ∥ p = ∥ 2 I ∥ p = 2. ) \|P(S)\|_p=\|P(-iJ)\|_p=\|2I\|_p=2.) ∥ P ( S ) ∥ p = ∥ P ( − i J ) ∥ p = ∥2 I ∥ p = 2. ) □
Remark 5.2 (Theorem B is not the trivial bound). Theorem B does not reduce to the trivial bound ∥ P ( T ) ∥ p ≤ ∑ ∣ c k ∣ \|P(T)\|_p\le\sum|c_k| ∥ P ( T ) ∥ p ≤ ∑ ∣ c k ∣ : for P ~ ( z ) = 1 + α z − z 2 \widetilde P(z)=1+\alpha z-z^2 P ( z ) = 1 + α z − z 2 with real α ≠ 0 \alpha\ne0 α = 0 (ζ = 1 \zeta=1 ζ = 1 in Corollary 5.1) and 1 < p < ∞ 1<p<\infty 1 < p < ∞ , Corollary 5.1 gives ∥ P ~ ( S ) ∥ p = ∥ 2 I + α J ∥ p < 2 + ∣ α ∣ = ∑ ∣ c k ∣ \|\widetilde P(S)\|_p=\|2I+\alpha J\|_p<2+|\alpha|=\sum|c_k| ∥ P ( S ) ∥ p = ∥2 I + α J ∥ p < 2 + ∣ α ∣ = ∑ ∣ c k ∣ (consistent with [2 Prop. 6.1] ). Indeed, by strict convexity of ℓ 2 p \ell_2^p ℓ 2 p , ∥ 2 ξ + α J ξ ∥ p = 2 ∥ ξ ∥ p + ∣ α ∣ ∥ J ξ ∥ p \|2\xi+\alpha J\xi\|_p=2\|\xi\|_p+|\alpha|\|J\xi\|_p ∥2 ξ + α J ξ ∥ p = 2∥ ξ ∥ p + ∣ α ∣∥ J ξ ∥ p for a real ξ ≠ 0 \xi\ne0 ξ = 0 would make J ξ J\xi J ξ a real multiple of ξ \xi ξ , impossible as J 2 = − I J^2=-I J 2 = − I ; so the maximum of ∥ ( 2 I + α J ) ξ ∥ p \|(2I+\alpha J)\xi\|_p ∥ ( 2 I + α J ) ξ ∥ p over the compact real unit sphere is < 2 + ∣ α ∣ <2+|\alpha| < 2 + ∣ α ∣ , and Lemma 2.1 gives the same on the complex ℓ 2 p \ell_2^p ℓ 2 p .
Acknowledgements
The author thanks S. K. Ray for correspondence about this work.3 This work was done overwhelmingly by AI models and supervised by the author. The models were Anthropic’s Claude and OpenAI’s ChatGPT. The author takes full responsibility for the paper.
References
Email address: david44hello@gmail.com