Introduction
In the 70s, Brown, Douglas, and Fillmore [2 , 3 ] beautifully classified essentially normal operators by developing the operator theoretic realization of K K K -homology theory. This development inspired a fundamental interaction between functional analysis and topology, e.g. [5 ] .
Let T T T be an essentially normal operator. The image [ T ] [T] [ T ] of T T T in the Calkin algebra is normal. The Brown-Douglas-Fillmore theory shows that the essential spectrum of T T T defined by [ T ] [T] [ T ] and the Fredholm index of T − z I T-zI T − z I on the complement of the essential spectrum determine its unitary equivalence class modulo the compact operators. Since then, researchers have been exploring an analogous problem for almost-normal operators and Hilbert–Schmidt perturbations.
Let S 1 S_1 S 1 and S 2 S_2 S 2 denote the trace class and Hilbert–Schmidt class, equipped with the norms
∥ K ∥ 1 = Tr ∣ K ∣ , ∥ K ∥ 2 2 = Tr ( K ∗ K ) . \lVert K\rVert_1 = \operatorname{Tr}|K|,\qquad\lVert K\rVert_2^2 = \operatorname{Tr}(K^*K). ∥ K ∥ 1 = Tr ∣ K ∣ , ∥ K ∥ 2 2 = Tr ( K ∗ K ) .
An operator T T T is called almost normal if
[ T ∗ , T ] ∈ S 1 [T^*,T] \in S_1 [ T ∗ , T ] ∈ S 1
Write T = A + i B T=A+iB T = A + i B , where A A A and B B B are bounded and self-adjoint. Observe that [ T ∗ , T ] = 2 i [ A , B ] [T^*,T]=2i[A,B] [ T ∗ , T ] = 2 i [ A , B ] . So T T T is almost normal if and only if [ A , B ] ∈ S 1 [A,B]\in S_1 [ A , B ] ∈ S 1 .
Helton and Howe [9 , 10 ] associated to a pair of bounded self-adjoint operators A A A , B B B with [ A , B ] ∈ S 1 [A,B]\in S_1 [ A , B ] ∈ S 1 a compactly supported finite signed Borel measure P T P_T P T on R 2 \mathbb{R}^2 R 2 , where T = A + i B T=A+iB T = A + i B , that records traces of polynomial commutators. With the normalization of [26 §2.3] , P T P_T P T satisfies
Tr [ P ( A , B ) , Q ( A , B ) ] = 1 2 π i ∫ R 2 ( ∂ x P ∂ y Q − ∂ y P ∂ x Q ) d P T (1) \operatorname{Tr}[P(A,B),Q(A,B)] = \frac{1}{2\pi i}\int_{\mathbb{R}^2}(\partial_xP\,\partial_yQ-\partial_yP\,\partial_xQ)\,dP_T \tag*{(1)} Tr [ P ( A , B ) , Q ( A , B )] = 2 π i 1 ∫ R 2 ( ∂ x P ∂ y Q − ∂ y P ∂ x Q ) d P T (1)
2020 Mathematics Subject Classification . Primary 47B20. Secondary 47A20, 47B10.
Key words and phrases . Almost normal operators, normal dilation, Hilbert–Schmidt perturbations, magnetic holonomy, Helton–Howe measure.
for P , Q ∈ C [ x , y ] P,Q \in\mathbb{C}[x,y] P , Q ∈ C [ x , y ] . The measure P T P_T P T is absolutely continuous, and its density is called the Pincus principal function. Carey and Pincus [4 ] related the principal function to the Fredholm index. The trace formula (1.1) gives invariance of P T P_T P T under trace-class perturbations, which change each polynomial evaluation by a trace-class operator.
In 1981, Voiculescu [23 ] proved that two almost-normal operators differing by a Hilbert–Schmidt operator have the same Helton–Howe measure when one has finite multicyclicity. He asked for analogues of the Brown–Douglas–Fillmore theory for almost normal operators with Hilbert–Schmidt perturbations.
Voiculescu [23 p. 317] , [25 ] , and [26 §2.5] proposed the following conjectures about the classification of almost normal operators up to Hilbert–Schmidt perturbations on a separable infinite-dimensional Hilbert space H \mathcal{H} H .
Conjecture 1.1. ([26 Conjecture 3] ) If T 1 T_1 T 1 and T 2 T_2 T 2 are almost normal on a Hilbert space H \mathcal{H} H and P T 1 = P T 2 P_{T_1}=P_{T_2} P T 1 = P T 2 , then there are a bounded normal operator N N N on H \mathcal{H} H and a unitary W W W on H ⊕ H \mathcal{H}\oplus\mathcal{H} H ⊕ H such that
T 1 ⊕ N − W ( T 2 ⊕ N ) W ∗ ∈ S 2 . (2) T_1\oplus N-W(T_2\oplus N)W^* \in\mathcal{S}_2. \tag*{(2)} T 1 ⊕ N − W ( T 2 ⊕ N ) W ∗ ∈ S 2 . (2)
Conjecture 1.2. ([26 Conjecture 4] ) For every almost normal T T T on H \mathcal{H} H , there are an almost-normal S S S on H \mathcal{H} H and a bounded normal M M M on H ⊕ H \mathcal{H}\oplus\mathcal{H} H ⊕ H such that
T ⊕ S − M ∈ S 2 . (3) T\oplus S-M\in\mathcal{S}_2. \tag*{(3)} T ⊕ S − M ∈ S 2 . (3)
Several classes of operators are known to satisfy the normal dilation conclusion. In 1983, Pasnicu proved [19 ] that a unilateral weighted shift with nonnegative weights a n ↑ a a_n \uparrow a a n ↑ a satisfies Conjecture 1.2 if ∑ n ( a − a n ) p < ∞ \sum_n(a-a_n)^p<\infty ∑ n ( a − a n ) p < ∞ for some p > 0 p>0 p > 0 . Lauric [14 , 15 , 16 ] identified further affirmative cases, including subnormal operators with trace-class self-commutator, hyponormal contractions T T T with I − T T ∗ ∈ S 1 I-TT^*\in\mathcal{S}_1 I − T T ∗ ∈ S 1 , and almost-normal operators whose square is normal.
In this note, we prove the following theorem.
Theorem 1.3. There is a compact operator T T T on a separable infinite-dimensional Hilbert space H \mathcal{H} H with [ T ∗ , T ] ∈ S 1 [T^*,T]\in\mathcal{S}_1 [ T ∗ , T ] ∈ S 1 and P T = 0 P_T=0 P T = 0 such that
T ⊕ S − M ∉ S 2 T\oplus S-M\notin\mathcal{S}_2 T ⊕ S − M ∈ / S 2 for every bounded operator S S S on a separable Hilbert space K \mathcal{K} K and every bounded normal operator M M M on H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K .
The proof of Theorem 1.3 uses a finite-dimensional measure of normal comparison. For a self-adjoint pair ( A , B ) (A,B) ( A , B ) , its frame error, Definition 2.1 , is the least total squared residual over finite labelled Parseval frames. This quantity is bounded above by the squared Hilbert–Schmidt cost of intertwining ( A , B ) (A,B) ( A , B ) with a commuting self-adjoint pair on a larger space. For a bounded direct sum of finite-dimensional blocks, we show in Proposition 2.5 that normal dilation modulo S 2 \mathcal{S}_2 S 2 is equivalent to summability of the block frame errors.
We therefore seek finite-dimensional pairs with uniformly bounded commutator trace norms and unbounded frame errors. We introduce in § 3 such pairs from signed shifts on finite square tori. We consider signs that are supported on a finite string with endpoints giving two flux defects, and compute that disjoint loops surrounding one defect yield a positive operator with logarithmically growing trace.
The key estimate in our study is Proposition 4.1 which shows that for the pair ( A L , B L ) (A_L,B_L) ( A L , B L ) constructed in § 3 , their frame error satisfies
∂ ( A L , B L ) ≥ c log L − C 0 , \partial(A_L,B_L)\ge c\log L-C_0, ∂ ( A L , B L ) ≥ c log L − C 0 ,
where d d d is defined by (4) and c , C 0 c,C_0 c , C 0 are independent of L L L . In § 5 , we form
T = ⨁ j λ j ( A L j + i B L j ) , λ j = 2 − j . T=\bigoplus_j \lambda_j(A_{L_j}+iB_{L_j}),\qquad\lambda_j=2^{-j}. T = j ⨁ λ j ( A L j + i B L j ) , λ j = 2 − j .
The uniform bound on the block commutator trace norms, together with ∑ j λ j 2 < ∞ \sum_j\lambda_j^2<\infty ∑ j λ j 2 < ∞ , ensures that T T T is almost normal. We choose L j L_j L j so that λ j 2 log L j → ∞ \lambda_j^2\log L_j\to\infty λ j 2 log L j → ∞ . Proposition 4.1 and the criterion in Proposition 2.5 then exclude every bounded companion S S S and bounded normal comparator M M M . This study of frame error connects Hilbert–Schmidt intertwining methods [22 , 6 ] with quadratic frame approximation [8 , 1 ] and spectral variance in joint diagonalization [18 §7] .
Theorem 1.3 shows that Conjectures 1.1 and 1.2 do not hold in their current forms. New invariants are needed to completely classify almost normal operators up to stable unitary equivalence modulo the Hilbert–Schmidt class. In Definition 6.4 , we introduce a quantitative invariant R ( T ) \mathcal{R}(T) R ( T ) , the rank profile, for an almost normal operator T T T . In Proposition 6.5 , we show that the rank profile invariant must agree for two almost normal operators that are stably unitarily equivalent modulo the Hilbert–Schmidt class. In particular, the rank profile distinguishes the operator T T T constructed in Theorem 1.3 from zero. In a forthcoming paper, we will investigate refinements of this invariant and possible connections with quantitative K K K -theory, with the aim of developing a complete invariant for stable unitary equivalence of almost normal operators modulo Hilbert–Schmidt perturbations.
The magnetic shifts used in our construction define discrete magnetic connections on square tori with two localized flux defects. Their normalized Hilbert–Schmidt defect tends to zero as the lattice size grows. The construction also suggests analogous models on surfaces of higher genus, which we plan to investigate in future work.
Acknowledgements: We would like to thank Professor Ronald G. Douglas for introducing the problem of classification of almost normal operators and the inspiring discussion. Our research was supported in part by DMS-2350181 and Simons Foundation Collaboration Grant MPS-TSM-00007714.
Frame errors and normal comparisons
Frame error measures the least squared cost of comparing a finite-dimensional self-adjoint pair with a commuting pair on a larger Hilbert space. We first define the quantity in terms of Parseval frames.
Let A , B A,B A , B be self-adjoint operators on a finite-dimensional Hilbert space H \mathcal{H} H . Recall that a finite Parseval frame is a family ( ψ ν ) ν (\psi_\nu)_\nu ( ψ ν ) ν in H \mathcal{H} H satisfying ∑ ν ψ ν ψ ν ∗ = I H \sum_\nu\psi_\nu\psi_\nu^*=I_{\mathcal{H}} ∑ ν ψ ν ψ ν ∗ = I H , with ψ ν ψ ν ∗ \psi_\nu\psi_\nu^* ψ ν ψ ν ∗ the rank-one operator associated to ψ ν \psi_\nu ψ ν . For such a frame, assign a real label x ν x_\nu x ν to the first coordinate and y ν y_\nu y ν to the second.
Definition 2.1. Define the frame error of the pair ( A , B ) (A,B) ( A , B ) by
∂ ( A , B ) = inf ∑ ν ( ∥ ( A − x ν I ) ψ ν ∥ 2 + ∥ ( B − y ν I ) ψ ν ∥ 2 ) . (4) \partial(A,B)=\inf\sum_\nu\left(\lVert(A-x_\nu I)\psi_\nu\rVert^2+\lVert(B-y_\nu I)\psi_\nu\rVert^2\right). \tag*{(4)} ∂ ( A , B ) = inf ν ∑ ( ∥( A − x ν I ) ψ ν ∥ 2 + ∥( B − y ν I ) ψ ν ∥ 2 ) . (4) where the infimum ranges over all finite Parseval frames and all real labels.
We remark that a similar definition has appeared in [21 Eq. (114)] . In particular, for n = dim H n=\dim\mathcal{H} n = dim H , ∂ ( A , B ) \partial(A,B) ∂ ( A , B ) can be identified with n I ( { A , B } , I H / n ) nI(\{A,B\},I_{\mathcal{H}}/n) n I ({ A , B } , I H / n ) for the convex-roof functional I I I in [21 ] .
We omit any zero vectors for a fixed Parseval frame. Given ψ ν \psi_\nu ψ ν , we notice that the functions x ↦ ∥ ( A − x I ) ψ ν ∥ 2 x \mapsto\lVert(A-xI)\psi_\nu\rVert^2 x ↦ ∥( A − x I ) ψ ν ∥ 2 and y ↦ ∥ ( B − y I ) ψ ν ∥ 2 y \mapsto\lVert(B-yI)\psi_\nu\rVert^2 y ↦ ∥( B − y I ) ψ ν ∥ 2 reach their minimum at
x ν = ⟨ A ψ ν , ψ ν ⟩ ∥ ψ ν ∥ 2 , y ν = ⟨ B ψ ν , ψ ν ⟩ ∥ ψ ν ∥ 2 . x_\nu=\frac{\langle A\psi_\nu,\psi_\nu\rangle}{\lVert\psi_\nu\rVert^2},\qquad y_\nu=\frac{\langle B\psi_\nu,\psi_\nu\rangle}{\lVert\psi_\nu\rVert^2}. x ν = ∥ ψ ν ∥ 2 ⟨ A ψ ν , ψ ν ⟩ , y ν = ∥ ψ ν ∥ 2 ⟨ B ψ ν , ψ ν ⟩ .
These labels satisfy
∣ x ν ∣ ≤ ∥ A ∥ , ∣ y ν ∣ ≤ ∥ B ∥ . (5) \lvert x_\nu\rvert\leq\lVert A\rVert,\qquad\lvert y_\nu\rvert\leq\lVert B\rVert. \tag*{(5)} ∣ x ν ∣ ≤ ∥ A ∥ , ∣ y ν ∣ ≤ ∥ B ∥ . (5)
Thus the infimum defining ∂ ( A , B ) \partial(A,B) ∂ ( A , B ) is unchanged if we restrict to { x ν , y ν } \{x_\nu,y_\nu\} { x ν , y ν } satisfying these bounds, i.e.
∂ ( A , B ) = inf ∣ x ν ∣ ≤ ∥ A ∥ , ∣ y ν ∣ ≤ ∥ B ∥ , ( ψ ν ) ν ∑ ν ( ∥ ( A − x ν I ) ψ ν ∥ 2 + ∥ ( B − y ν I ) ψ ν ∥ 2 ) . (6) \partial(A,B)=\inf_{\lvert x_\nu\rvert\leq\lVert A\rVert,\lvert y_\nu\rvert\leq\lVert B\rVert,(\psi_\nu)_\nu}\sum_\nu\left(\lVert(A-x_\nu I)\psi_\nu\rVert^2+\lVert(B-y_\nu I)\psi_\nu\rVert^2\right). \tag*{(6)} ∂ ( A , B ) = ∣ x ν ∣ ≤ ∥ A ∥ , ∣ y ν ∣ ≤ ∥ B ∥ , ( ψ ν ) ν inf ν ∑ ( ∥( A − x ν I ) ψ ν ∥ 2 + ∥( B − y ν I ) ψ ν ∥ 2 ) . (6)
Furthermore, using (4) , we observe that for λ ∈ R \lambda\in\mathbb{R} λ ∈ R rescaling A , B A,B A , B to λ A , λ B \lambda A,\lambda B λ A , λ B gives λ 2 \lambda^2 λ 2 multiple of ∂ ( A , B ) \partial(A,B) ∂ ( A , B ) ,
∂ ( λ A , λ B ) = λ 2 ∂ ( A , B ) , ∀ λ ∈ R . (7) \partial(\lambda A,\lambda B)=\lambda^2\partial(A,B),\qquad\forall\lambda\in\mathbb{R}. \tag*{(7)} ∂ ( λ A , λ B ) = λ 2 ∂ ( A , B ) , ∀ λ ∈ R . (7)
Lemma 2.2. Let A , B A,B A , B be self-adjoint operators on a finite-dimensional Hilbert space H \mathcal{H} H . Then ∂ ( A , B ) = 0 \partial(A,B)=0 ∂ ( A , B ) = 0 if and only if A A A and B B B commute.
Proof. We start the proof with establishing the following inequality,
∥ [ A , B ] ∥ 2 2 ≤ 4 ( ∥ A ∥ 2 + ∥ B ∥ 2 ) ∂ ( A , B ) . (8) \lVert[A,B]\rVert_2^2\leq4(\lVert A\rVert^2+\lVert B\rVert^2)\partial(A,B). \tag*{(8)} ∥[ A , B ] ∥ 2 2 ≤ 4 (∥ A ∥ 2 + ∥ B ∥ 2 ) ∂ ( A , B ) . (8) Fix a finite Parseval frame ( ψ ν ) ν (\psi_\nu)_\nu ( ψ ν ) ν and labels satisfying (5) , as permitted by (6) . Set
a ν = ( A − x ν I ) ψ ν , b ν = ( B − y ν I ) ψ ν . a_\nu=(A-x_\nu I)\psi_\nu,\qquad b_\nu=(B-y_\nu I)\psi_\nu. a ν = ( A − x ν I ) ψ ν , b ν = ( B − y ν I ) ψ ν . Since scalar operators commute with A A A and B B B ,
[ A , B ] ψ ν = ( A − x ν I ) b ν − ( B − y ν I ) a ν . [A,B]\psi_\nu=(A-x_\nu I)b_\nu-(B-y_\nu I)a_\nu. [ A , B ] ψ ν = ( A − x ν I ) b ν − ( B − y ν I ) a ν . So we have the following bound,
∥ [ A , B ] ψ ν ∥ ≤ 2 ∥ A ∥ ∥ b ν ∥ + 2 ∥ B ∥ ∥ a ν ∥ ≤ 2 ( ∥ A ∥ 2 + ∥ B ∥ 2 ) 1 / 2 ( ∥ a ν ∥ 2 + ∥ b ν ∥ 2 ) 1 / 2 . (9) \begin{aligned}
\lVert[A,B]\psi_\nu\rVert&\leq2\lVert A\rVert\lVert b_\nu\rVert+2\lVert B\rVert\lVert a_\nu\rVert\\
&\leq2(\lVert A\rVert^2+\lVert B\rVert^2)^{1/2}(\lVert a_\nu\rVert^2+\lVert b_\nu\rVert^2)^{1/2}. \tag*{(9)}
\end{aligned} ∥[ A , B ] ψ ν ∥ ≤ 2 ∥ A ∥ ∥ b ν ∥ + 2 ∥ B ∥ ∥ a ν ∥ ≤ 2 (∥ A ∥ 2 + ∥ B ∥ 2 ) 1/2 (∥ a ν ∥ 2 + ∥ b ν ∥ 2 ) 1/2 . (9) The second inequality is the scalar Cauchy–Schwarz inequality. Moreover, the Parseval identity gives
∑ ν ∥ [ A , B ] ψ ν ∥ 2 = Tr ( [ A , B ] ∗ [ A , B ] ∑ ν ψ ν ψ ν ∗ ) \sum_\nu\lVert[A,B]\psi_\nu\rVert^2=\operatorname{Tr}\left([A,B]^*[A,B]\sum_\nu\psi_\nu\psi_\nu^*\right) ν ∑ ∥[ A , B ] ψ ν ∥ 2 = Tr ( [ A , B ] ∗ [ A , B ] ν ∑ ψ ν ψ ν ∗ ) = ∥ [ A , B ] ∥ 2 2 . =\lVert[A,B]\rVert_2^2. = ∥[ A , B ] ∥ 2 2 . Squaring the preceding estimate (9) and summing over ν \nu ν , we have
∥ [ A , B ] ∥ 2 2 ≤ 4 ( ∥ A ∥ 2 + ∥ B ∥ 2 ) ∑ ν ( ∥ ( A − x ν I ) ψ ν ∥ 2 + ∥ ( B − y ν I ) ψ ν ∥ 2 ) . \lVert[A,B]\rVert_2^2\leq4(\lVert A\rVert^2+\lVert B\rVert^2)\sum_\nu\left(\lVert(A-x_\nu I)\psi_\nu\rVert^2+\lVert(B-y_\nu I)\psi_\nu\rVert^2\right). ∥[ A , B ] ∥ 2 2 ≤ 4 (∥ A ∥ 2 + ∥ B ∥ 2 ) ν ∑ ( ∥( A − x ν I ) ψ ν ∥ 2 + ∥( B − y ν I ) ψ ν ∥ 2 ) . Taking the infimum, we have proved (8) . Accordingly, ∂ ( A , B ) = 0 \partial(A,B)=0 ∂ ( A , B ) = 0 implies [ A , B ] = 0 [A,B]=0 [ A , B ] = 0 .
In the converse direction, when A , B A,B A , B commute, since A , B A,B A , B are self-adjoint and the space is finite dimensional, they have a common orthonormal eigenbasis. Such a basis, with labels being the corresponding eigenvalues, has zero frame cost. □
We say that a self-adjoint operator Y Y Y on H \mathcal{H} H gives a lower bound to ∂ ( A , B ) \partial(A,B) ∂ ( A , B ) whenever
Y ≤ ( A − x I ) 2 + ( B − y I ) 2 ( x , y ∈ R ) . (10) Y \le(A-xI)^2+(B-yI)^2 \qquad(x,y\in\mathbb{R}). \tag*{(10)} Y ≤ ( A − x I ) 2 + ( B − y I ) 2 ( x , y ∈ R ) . (10)
Apply (10) to ψ ν \psi_\nu ψ ν with its own labels and sum. We get
∑ ν ⟨ Y ψ ν , ψ ν ⟩ = Tr Y , \sum_\nu\langle Y\psi_\nu,\psi_\nu\rangle=\operatorname{Tr}Y, ν ∑ ⟨ Y ψ ν , ψ ν ⟩ = Tr Y ,
and
∂ ( A , B ) ≥ Tr Y . (11) \partial(A,B)\ge\operatorname{Tr}Y. \tag*{(11)} ∂ ( A , B ) ≥ Tr Y . (11)
The next lemma identifies frame error with the least squared intertwining cost against commuting self-adjoint pairs.
Lemma 2.3. Assume that H \mathcal{H} H is finite dimensional. Let J : H → K J:\mathcal{H}\to\mathcal{K} J : H → K be an isometry and let C , D C,D C , D be bounded commuting self-adjoint operators on K \mathcal{K} K . Then
∂ ( A , B ) ≤ ∥ C J − J A ∥ 2 2 + ∥ D J − J B ∥ 2 2 . (12) \partial(A,B)\le\lVert CJ-JA\rVert_2^2+\lVert DJ-JB\rVert_2^2. \tag*{(12)} ∂ ( A , B ) ≤ ∥ C J − J A ∥ 2 2 + ∥ D J − J B ∥ 2 2 . (12) The infimum of the right-hand side over J , C , D J,C,D J , C , D and K \mathcal{K} K equals ∂ ( A , B ) \partial(A,B) ∂ ( A , B ) .
Proof. We start with considering the special case that C = ∑ k = 1 s x k E k C=\sum_{k=1}^s x_kE_k C = ∑ k = 1 s x k E k and D = ∑ k = 1 s y k E k D=\sum_{k=1}^s y_kE_k D = ∑ k = 1 s y k E k for mutually orthogonal projections ( E k ) (E_k) ( E k ) satisfying
∑ k E k = I K . \sum_k E_k=I_{\mathcal{K}}. k ∑ E k = I K .
For each k k k , take an orthonormal basis ( v k , l ) l (v_{k,l})_l ( v k , l ) l of E k J H E_kJ\mathcal{H} E k J H . Define ψ k , l = J ∗ v k , l \psi_{k,l}=J^*v_{k,l} ψ k , l = J ∗ v k , l . We compute
∀ u ∈ H , ∑ k , l ∣ ⟨ ψ k , l , u ⟩ ∣ 2 = ∑ k ∥ E k J u ∥ 2 = ∥ u ∥ 2 , \forall u\in\mathcal{H},\qquad\sum_{k,l}|\langle\psi_{k,l},u\rangle|^2=\sum_k\lVert E_kJu\rVert^2=\lVert u\rVert^2, ∀ u ∈ H , k , l ∑ ∣ ⟨ ψ k , l , u ⟩ ∣ 2 = k ∑ ∥ E k J u ∥ 2 = ∥ u ∥ 2 ,
which shows that the vectors ψ k , l = J ∗ v k , l \psi_{k,l}=J^*v_{k,l} ψ k , l = J ∗ v k , l form a Parseval frame.
The E k E_k E k -component of C J − J A CJ-JA C J − J A is
E k ( ∑ j x j E j J − ∑ j E j J A ) = E k J ( x k I − A ) . E_k\left(\sum_jx_jE_jJ-\sum_jE_jJA\right)=E_kJ(x_kI-A). E k ( j ∑ x j E j J − j ∑ E j J A ) = E k J ( x k I − A ) .
The range of E k ( C J − J A ) E_k(CJ-JA) E k ( C J − J A ) is in E k J H E_kJ\mathcal{H} E k J H . We compute the Hilbert–Schmidt norm of its adjoint on the basis ( v k , l ) l (v_{k,l})_l ( v k , l ) l by
∥ E k ( C J − J A ) ∥ 2 2 = ∥ ( E k J ( x k I − A ) ) ∗ ∥ 2 2 = ∑ l ∥ ( E k J ( x k I − A ) ) ∗ v k , l ∥ 2 = ∑ l ∥ ( A − x k I ) J ∗ E k v k , l ∥ 2 = ∑ l ∥ ( A − x k I ) J ∗ v k , l ∥ 2 = ∑ l ∥ ( A − x k I ) ψ k , l ∥ 2 . \begin{aligned}
\lVert E_k(CJ-JA)\rVert_2^2
&=\lVert(E_kJ(x_kI-A))^*\rVert_2^2\\
&=\sum_l\lVert(E_kJ(x_kI-A))^*v_{k,l}\rVert^2\\
&=\sum_l\lVert(A-x_kI)J^*E_kv_{k,l}\rVert^2\\
&=\sum_l\lVert(A-x_kI)J^*v_{k,l}\rVert^2\\
&=\sum_l\lVert(A-x_kI)\psi_{k,l}\rVert^2.
\end{aligned} ∥ E k ( C J − J A ) ∥ 2 2 = ∥( E k J ( x k I − A ) ) ∗ ∥ 2 2 = l ∑ ∥( E k J ( x k I − A ) ) ∗ v k , l ∥ 2 = l ∑ ∥( A − x k I ) J ∗ E k v k , l ∥ 2 = l ∑ ∥( A − x k I ) J ∗ v k , l ∥ 2 = l ∑ ∥( A − x k I ) ψ k , l ∥ 2 .
The same calculation for D J − J B DJ-JB D J − J B proves (12) in this case, with labels ( x k , y k ) (x_k,y_k) ( x k , y k ) .
For general C , D C,D C , D , using the joint spectral theorem, we write
C = ∫ R x d E ( x , y ) , D = ∫ R y d E ( x , y ) , C=\int_{\mathbb{R}}x\,dE(x,y),\qquad D=\int_{\mathbb{R}}y\,dE(x,y), C = ∫ R x d E ( x , y ) , D = ∫ R y d E ( x , y ) ,
where R ⊂ R 2 R \subset\mathbb{R}^{2} R ⊂ R 2 is a bounded rectangle containing their joint spectrum, and E E E is a projection valued measure supported on R R R .
We subdivide R R R into finitely many pairwise disjoint rectangles Q 1 , … , Q m Q_{1},\ldots,Q_{m} Q 1 , … , Q m , each of diameter at most δ \delta δ . Choose ( x k , y k ) ∈ Q k (x_{k},y_{k}) \in Q_{k} ( x k , y k ) ∈ Q k and define E k = E ( Q k ) E_{k}=E(Q_{k}) E k = E ( Q k ) for k = 1 , … , m k=1,\ldots,m k = 1 , … , m . Define
C δ = ∑ k x k E k , D δ = ∑ k y k E k . C_{\delta}=\sum_{k}x_{k}E_{k}, \qquad D_{\delta}=\sum_{k}y_{k}E_{k}. C δ = k ∑ x k E k , D δ = k ∑ y k E k .
On E k K E_{k}\mathcal{K} E k K , the diameter bound of Q k Q_{k} Q k implies that
∥ C δ − C ∥ ≤ δ , ∥ D δ − D ∥ ≤ δ . \|C_{\delta}-C\|\leq\delta, \qquad\|D_{\delta}-D\|\leq\delta. ∥ C δ − C ∥ ≤ δ , ∥ D δ − D ∥ ≤ δ .
Choose an orthonormal basis e 1 , … , e n e_{1},\ldots,e_{n} e 1 , … , e n of H \mathcal{H} H . Since J J J is an isometry, we compute
∥ ( C δ − C ) J ∥ 2 2 = ∑ j = 1 n ∥ ( C δ − C ) J e j ∥ 2 ≤ ∥ C δ − C ∥ 2 ∑ j = 1 n ∥ J e j ∥ 2 = dim H ∥ C δ − C ∥ 2 , \begin{aligned}
\|(C_{\delta}-C)J\|_{2}^{2}=\sum_{j=1}^{n}\|(C_{\delta}-C)Je_{j}\|^{2} \\
&\leq\|C_{\delta}-C\|^{2}\sum_{j=1}^{n}\|Je_{j}\|^{2} \\
&=\dim\mathcal{H}\|C_{\delta}-C\|^{2},
\end{aligned} ∥ ( C δ − C ) J ∥ 2 2 = j = 1 ∑ n ∥ ( C δ − C ) J e j ∥ 2 ≤ ∥ C δ − C ∥ 2 j = 1 ∑ n ∥ J e j ∥ 2 = dim H ∥ C δ − C ∥ 2 ,
which gives
∥ ( C δ J − J A ) − ( C J − J A ) ∥ 2 = ∥ ( C δ − C ) J ∥ 2 ≤ dim H δ ⟶ 0. \|(C_{\delta}J-JA)-(CJ-JA)\|_{2}=\|(C_{\delta}-C)J\|_{2}\leq\sqrt{\dim\mathcal{H}}\,\delta\longrightarrow0. ∥ ( C δ J − J A ) − ( C J − J A ) ∥ 2 = ∥ ( C δ − C ) J ∥ 2 ≤ dim H δ ⟶ 0.
Similarly,
∥ ( D δ J − J B ) − ( D J − J B ) ∥ 2 = ∥ ( D δ − D ) J ∥ 2 ≤ dim H δ ⟶ 0. \|(D_{\delta}J-JB)-(DJ-JB)\|_{2}=\|(D_{\delta}-D)J\|_{2}\leq\sqrt{\dim\mathcal{H}}\,\delta\longrightarrow0. ∥ ( D δ J − J B ) − ( D J − J B ) ∥ 2 = ∥ ( D δ − D ) J ∥ 2 ≤ dim H δ ⟶ 0.
Applying the special case to C δ , D δ C_{\delta},D_{\delta} C δ , D δ , we obtain
∂ ( A , B ) ≤ ∥ C δ J − J A ∥ 2 2 + ∥ D δ J − J B ∥ 2 2 . \partial(A,B)\leq\|C_{\delta}J-JA\|_{2}^{2}+\|D_{\delta}J-JB\|_{2}^{2}. ∂ ( A , B ) ≤ ∥ C δ J − J A ∥ 2 2 + ∥ D δ J − J B ∥ 2 2 .
Passing to the limit δ → 0 \delta\to0 δ → 0 , we get
∂ ( A , B ) ≤ ∥ C J − J A ∥ 2 2 + ∥ D J − J B ∥ 2 2 . \partial(A,B)\leq\|CJ-JA\|_{2}^{2}+\|DJ-JB\|_{2}^{2}. ∂ ( A , B ) ≤ ∥ C J − J A ∥ 2 2 + ∥ D J − J B ∥ 2 2 .
Conversely, for any finite labelled Parseval frame ( ψ ν ) ν , ν = 1 , … , N (\psi_{\nu})_{\nu},\ \nu=1,\ldots,N ( ψ ν ) ν , ν = 1 , … , N , we define the isometry J : H → C N J:\mathcal{H}\to\mathbb{C}^{N} J : H → C N by
J u = ( ⟨ u , ψ 1 ⟩ , … , ⟨ u , ψ N ⟩ ) . Ju=(\langle u,\psi_{1}\rangle,\ldots,\langle u,\psi_{N}\rangle). J u = (⟨ u , ψ 1 ⟩ , … , ⟨ u , ψ N ⟩) .
We observe that J J J is linear under the convention as the inner product is linear in its first variable. The isometry property is implied by the Parseval frame property. Take C , D C,D C , D diagonal with entries x ν , y ν x_{\nu},y_{\nu} x ν , y ν . We directly check that rows of C J − J A CJ-JA C J − J A and D J − J B DJ-JB D J − J B are the functionals corresponding to ( x ν I − A ) ψ ν (x_{\nu}I-A)\psi_{\nu} ( x ν I − A ) ψ ν and ( y ν I − B ) ψ ν (y_{\nu}I-B)\psi_{\nu} ( y ν I − B ) ψ ν . Their norm squares sum to the frame cost. Taking infima gives equality. □ \square □
Lemma 2.4. Given a bounded operator T T T on H \mathcal{H} H , there exists a bounded operator S S S and a normal operator M M M such that
T ⊕ S − M ∈ S 2 T\oplus S-M\in\mathcal{S}_{2} T ⊕ S − M ∈ S 2 if and only if there is an almost-normal operator S ~ \widetilde{S} S such that
T ⊕ S ~ − M ∈ S 2 . T\oplus\widetilde{S}-M\in\mathcal{S}_{2}. T ⊕ S − M ∈ S 2 .
Proof. Let J : H → K J:\mathcal{H}\to\mathcal{K} J : H → K be an isometry, P = J J ∗ P=JJ^* P = J J ∗ , and M M M bounded normal on K \mathcal{K} K . Suppose T − J ∗ M J ∈ S 2 T-J^*MJ\in\mathcal{S}_2 T − J ∗ M J ∈ S 2 and [ P , M ] ∈ S 2 [P,M]\in\mathcal{S}_2 [ P , M ] ∈ S 2 . Then the two off-diagonal corners P M ( I − P ) PM(I-P) P M ( I − P ) and ( I − P ) M P (I-P)MP ( I − P ) M P are Hilbert–Schmidt. Set Q = I − P Q=I-P Q = I − P and S = Q M Q ∣ Q K S=QMQ|_{Q\mathcal{K}} S = QM Q ∣ Q K . Identifying P K P\mathcal{K} P K with H \mathcal{H} H by J J J gives T ⊕ S − M ∈ S 2 T\oplus S-M\in\mathcal{S}_2 T ⊕ S − M ∈ S 2 . Moreover, normality of M M M gives
[ S ∗ , S ] = Q M P M ∗ Q − Q M ∗ P M Q on Q K . (13) [S^*,S]=QMPM^*Q-QM^*PMQ\quad\text{on }Q\mathcal{K}. \tag*{(13)} [ S ∗ , S ] = QM P M ∗ Q − Q M ∗ P M Q on Q K . (13)
Since Q M P QMP QM P and P M Q PMQ P M Q are Hilbert–Schmidt, both Q M P M ∗ Q QMPM^*Q QM P M ∗ Q and Q M ∗ P M Q QM^*PMQ Q M ∗ P M Q are trace class operators.
Conversely, if T ⊕ S − M ∈ S 2 T\oplus S-M\in\mathcal{S}_2 T ⊕ S − M ∈ S 2 , take J J J to be the inclusion of the first summand. The diagonal corner of the difference gives T − J ∗ M J ∈ S 2 T-J^*MJ\in\mathcal{S}_2 T − J ∗ M J ∈ S 2 , and [ P , M ] = [ P , M − ( T ⊕ S ) ] ∈ S 2 [P,M]=[P,M-(T\oplus S)]\in\mathcal{S}_2 [ P , M ] = [ P , M − ( T ⊕ S )] ∈ S 2 . □
The following criterion for infinite sums provides the key tool in our later proof of Theorem 1.3 .
Proposition 2.5. Let T = ⨁ j ≥ 1 ( A j + i B j ) T=\bigoplus_{j\ge1}(A_j+iB_j) T = ⨁ j ≥ 1 ( A j + i B j ) be bounded, with A j , B j A_j,B_j A j , B j self-adjoint on finite-dimensional spaces H j \mathcal{H}_j H j . Let H = ⨁ j H j \mathcal{H}=\bigoplus_j\mathcal{H}_j H = ⨁ j H j . There is a bounded operator S S S on K \mathcal{K} K and a bounded normal M M M on H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K with T ⊕ S − M ∈ S 2 T\oplus S-M\in\mathcal{S}_2 T ⊕ S − M ∈ S 2 if and only if
∑ j ≥ 1 ∂ ( A j , B j ) < ∞ . (14) \sum_{j\ge1}\partial(A_j,B_j)<\infty. \tag*{(14)} j ≥ 1 ∑ ∂ ( A j , B j ) < ∞. (14) If (2.11) holds, S S S may be chosen almost normal. If, in addition, T T T is compact, both S S S and M M M may be chosen compact.
Proof. We separate the proof into three parts.
I: Necessity. Suppose E = T ⊕ S − M ∈ S 2 E=T\oplus S-M\in\mathcal{S}_2 E = T ⊕ S − M ∈ S 2 , and write M = C + i D M=C+iD M = C + i D with C , D C,D C , D commuting and self-adjoint. For the inclusion J j J_j J j of the j j j -th block into H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K ,
C J j − J j A j = − ( Re E ) J j , D J j − J j B j = − ( Im E ) J j . CJ_j-J_jA_j=-(\operatorname{Re}E)J_j,\qquad DJ_j-J_jB_j=-(\operatorname{Im}E)J_j. C J j − J j A j = − ( Re E ) J j , D J j − J j B j = − ( Im E ) J j .
The block domains are mutually orthogonal. Apply Lemma 2.3 and sum squared column norms to obtain
∑ j ≥ 1 ∂ ( A j , B j ) ≤ ∥ Re E ∥ 2 2 + ∥ Im E ∥ 2 2 = ∥ E ∥ 2 2 . (15) \sum_{j\ge1}\partial(A_j,B_j)\leq\|\operatorname{Re}E\|_2^2+\|\operatorname{Im}E\|_2^2=\|E\|_2^2. \tag*{(15)} j ≥ 1 ∑ ∂ ( A j , B j ) ≤ ∥ Re E ∥ 2 2 + ∥ Im E ∥ 2 2 = ∥ E ∥ 2 2 . (15)
The last identity follows from the parallelogram identity applied to E E E and E ∗ E^* E ∗ in the Hilbert space S 2 \mathcal{S}_2 S 2 , i.e.
∥ E + E ∗ ∥ 2 2 + ∥ E − E ∗ ∥ 2 2 = 2 ( ∥ E ∥ 2 2 + ∥ E ∗ ∥ 2 2 ) = 4 ∥ E ∥ 2 2 . \|E+E^*\|_2^2+\|E-E^*\|_2^2=2(\|E\|_2^2+\|E^*\|_2^2)=4\|E\|_2^2. ∥ E + E ∗ ∥ 2 2 + ∥ E − E ∗ ∥ 2 2 = 2 ( ∥ E ∥ 2 2 + ∥ E ∗ ∥ 2 2 ) = 4∥ E ∥ 2 2 .
II. Sufficiency. Choose a finite labelled frame on each block with cost at most ∂ ( A j , B j ) + 2 − j \partial(A_j,B_j)+2^{-j} ∂ ( A j , B j ) + 2 − j . Write ( ψ j , ν ) ν = 1 N j (\psi_{j,\nu})_{\nu=1}^{N_j} ( ψ j , ν ) ν = 1 N j for the chosen Parseval frame on H j \mathcal{H}_j H j , with real labels x j , ν x_{j,\nu} x j , ν and y j , ν y_{j,\nu} y j , ν . We first arrange that
∣ x j , ν ∣ ≤ ∥ A j ∥ , ∣ y j , ν ∣ ≤ ∥ B j ∥ . |x_{j,\nu}|\leq\|A_j\|,\qquad|y_{j,\nu}|\leq\|B_j\|. ∣ x j , ν ∣ ≤ ∥ A j ∥ , ∣ y j , ν ∣ ≤ ∥ B j ∥.
To do this, replace every x j , ν > ∥ A j ∥ x_{j,\nu}>\|A_j\| x j , ν > ∥ A j ∥ by ∥ A j ∥ \|A_j\| ∥ A j ∥ and every x j , ν < − ∥ A j ∥ x_{j,\nu}<-\|A_j\| x j , ν < − ∥ A j ∥ by − ∥ A j ∥ -\|A_j\| − ∥ A j ∥ , leaving the other labels unchanged. Make the corresponding replacements for y j , ν y_{j,\nu} y j , ν using the endpoints ± ∥ B j ∥ \pm\|B_j\| ± ∥ B j ∥ . These replacements do not increase the frame error. Indeed, if x ^ \widehat{x} x is the replacement of x x x , then
∣ t − x ^ ∣ ≤ ∣ t − x ∣ for every t ∈ [ − ∥ A j ∥ , ∥ A j ∥ ] . |t-\widehat{x}|\leq|t-x|\qquad\text{for every }t\in[-\|A_j\|,\|A_j\|]. ∣ t − x ∣ ≤ ∣ t − x ∣ for every t ∈ [ − ∥ A j ∥ , ∥ A j ∥ ] .
Since the spectrum of A j A_j A j lies in this interval, the spectral theorem gives
∥ ( A j − x ~ I ) v ∥ 2 ≤ ∥ ( A j − x I ) v ∥ 2 ( v ∈ H j ) . \lVert(A_j-\tilde{x}I)v\rVert^2\leq\lVert(A_j-xI)v\rVert^2\qquad(v\in\mathcal{H}_j). ∥( A j − x ~ I ) v ∥ 2 ≤ ∥( A j − x I ) v ∥ 2 ( v ∈ H j ) .
The same argument applies to B j B_j B j . We continue to denote the modified labels by x j , ν x_{j,\nu} x j , ν and y j , ν y_{j,\nu} y j , ν . Their frame costs
ε j = ∑ ν = 1 N j ( ∥ ( A j − x j , ν I ) ψ j , ν ∥ 2 + ∥ ( B j − y j , ν I ) ψ j , ν ∥ 2 ) \varepsilon_j=\sum_{\nu=1}^{N_j}\left(\lVert(A_j-x_{j,\nu}I)\psi_{j,\nu}\rVert^2+\lVert(B_j-y_{j,\nu}I)\psi_{j,\nu}\rVert^2\right) ε j = ν = 1 ∑ N j ( ∥( A j − x j , ν I ) ψ j , ν ∥ 2 + ∥( B j − y j , ν I ) ψ j , ν ∥ 2 )
therefore still satisfy ε j ≤ d ( A j , B j ) + 2 − j \varepsilon_j\leq d(A_j,B_j)+2^{-j} ε j ≤ d ( A j , B j ) + 2 − j , and hence ∑ j ε j < ∞ \sum_j\varepsilon_j<\infty ∑ j ε j < ∞ .
Set K j = C N j \mathcal{K}_j=\mathbb{C}^{N_j} K j = C N j and define
J j u = ( ⟨ u , ψ j , ν ⟩ ) ν , C j = diag ( x j , ν ) , D j = diag ( y j , ν ) . J_ju=\left(\langle u,\psi_{j,\nu}\rangle\right)_\nu,\qquad C_j=\operatorname{diag}(x_{j,\nu}),\qquad D_j=\operatorname{diag}(y_{j,\nu}). J j u = ( ⟨ u , ψ j , ν ⟩ ) ν , C j = diag ( x j , ν ) , D j = diag ( y j , ν ) .
The Parseval identity makes J j J_j J j an isometry. By the calculation at the end of the proof of Lemma 2.3,
∥ C j J j − J j A j ∥ 2 2 + ∥ D j J j − J j B j ∥ 2 2 = ε j . \lVert C_jJ_j-J_jA_j\rVert_2^2+\lVert D_jJ_j-J_jB_j\rVert_2^2=\varepsilon_j. ∥ C j J j − J j A j ∥ 2 2 + ∥ D j J j − J j B j ∥ 2 2 = ε j .
Moreover,
∥ C j ∥ ≤ ∥ A j ∥ , ∥ D j ∥ ≤ ∥ B j ∥ . \lVert C_j\rVert\leq\lVert A_j\rVert,\qquad\lVert D_j\rVert\leq\lVert B_j\rVert. ∥ C j ∥ ≤ ∥ A j ∥ , ∥ D j ∥ ≤ ∥ B j ∥ .
Now put
H = ⨁ j H j , K = ⨁ j K j , J = ⨁ j J j , C = ⨁ j C j , D = ⨁ j D j . \mathcal{H}=\bigoplus_j\mathcal{H}_j,\qquad\mathcal{K}=\bigoplus_j\mathcal{K}_j,\qquad J=\bigoplus_jJ_j,\qquad C=\bigoplus_jC_j,\qquad D=\bigoplus_jD_j. H = j ⨁ H j , K = j ⨁ K j , J = j ⨁ J j , C = j ⨁ C j , D = j ⨁ D j .
Then J : H → K J:\mathcal{H}\to\mathcal{K} J : H → K is an isometry. Since T = ⨁ j ( A j + i B j ) T=\bigoplus_j(A_j+iB_j) T = ⨁ j ( A j + i B j ) is bounded,
sup j ∥ A j ∥ ≤ ∥ T ∥ , sup j ∥ B j ∥ ≤ ∥ T ∥ . \sup_j\lVert A_j\rVert\leq\lVert T\rVert,\qquad\sup_j\lVert B_j\rVert\leq\lVert T\rVert. j sup ∥ A j ∥ ≤ ∥ T ∥ , j sup ∥ B j ∥ ≤ ∥ T ∥ .
Consequently, C C C and D D D are bounded commuting self-adjoint operators. Set M : = C + i D M:=C+iD M := C + i D . It follows that M M M is bounded and normal.
Write A = ⨁ j A j A=\bigoplus_jA_j A = ⨁ j A j , B = ⨁ j B j B=\bigoplus_jB_j B = ⨁ j B j , and set
E C = C J − J A , E D = D J − J B . E_C=CJ-JA,\qquad E_D=DJ-JB. E C = C J − J A , E D = D J − J B .
The direct-sum formula for the Hilbert–Schmidt norm gives
∥ E C ∥ 2 2 + ∥ E D ∥ 2 2 = ∑ j ε j < ∞ . \lVert E_C\rVert_2^2+\lVert E_D\rVert_2^2=\sum_j\varepsilon_j<\infty. ∥ E C ∥ 2 2 + ∥ E D ∥ 2 2 = j ∑ ε j < ∞.
In particular, E C , E D ∈ S 2 E_C,E_D\in\mathcal{S}_2 E C , E D ∈ S 2 .
Let P = J J ∗ P=JJ^* P = J J ∗ and Q = I − P Q=I-P Q = I − P . Using the identity J ∗ J = I J^*J=I J ∗ J = I , we obtain
J ∗ M J − T = J ∗ E C + i J ∗ E D ∈ S 2 . J^*MJ-T=J^*E_C+iJ^*E_D\in\mathcal{S}_2. J ∗ M J − T = J ∗ E C + i J ∗ E D ∈ S 2 .
Furthermore, we observe that Q J = 0 QJ=0 Q J = 0 . Hence,
Q C P = Q ( C J − J A ) J ∗ = Q E C J ∗ ∈ S 2 , QCP=Q(CJ-JA)J^*=QE_CJ^*\in\mathcal{S}_2, QC P = Q ( C J − J A ) J ∗ = Q E C J ∗ ∈ S 2 ,
and similarly Q D P ∈ S 2 QDP\in\mathcal{S}_2 Q D P ∈ S 2 . Since C C C and D D D are self-adjoint, we know
P C Q = ( Q C P ) ∗ ∈ S 2 , P D Q = ( Q D P ) ∗ ∈ S 2 . PCQ=(QCP)^*\in\mathcal{S}_2,\qquad PDQ=(QDP)^*\in\mathcal{S}_2. P C Q = ( QC P ) ∗ ∈ S 2 , P D Q = ( Q D P ) ∗ ∈ S 2 .
It follows that both Q M P QMP QM P and P M Q PMQ P M Q are Hilbert–Schmidt, and therefore
[ P , M ] = P M Q − Q M P ∈ S 2 . [P,M]=PMQ-QMP\in\mathcal{S}_2. [ P , M ] = P M Q − QM P ∈ S 2 .
Define the operator S S S on Q K Q\mathcal{K} Q K by
S = Q M Q ∣ Q K . S=QMQ|_{Q\mathcal{K}}. S = QM Q ∣ Q K .
We directly check that
U : H ⊕ Q K ⟶ K , U ( u , v ) = J u + v , U:\mathcal{H}\oplus Q\mathcal{K}\longrightarrow\mathcal{K},\qquad U(u,v)=Ju+v, U : H ⊕ Q K ⟶ K , U ( u , v ) = J u + v ,
is unitary. In these coordinates,
T ⊕ S − U ∗ M U = ( T − J ∗ M J − J ∗ M ∣ Q K − Q M J 0 ) ∈ S 2 , T\oplus S-U^{*}MU=
\begin{pmatrix}
T-J^{*}MJ & -J^{*}M|_{Q\mathcal{K}}\\
-QMJ & 0
\end{pmatrix}\in\mathcal{S}_{2}, T ⊕ S − U ∗ M U = ( T − J ∗ M J − QM J − J ∗ M ∣ Q K 0 ) ∈ S 2 ,
because all three nonzero entries are Hilbert–Schmidt. Since U ∗ M U U^{*}MU U ∗ M U is bounded and normal, this is the required normal operator. Finally, normality of M M M gives, as in (2.10),
[ S ∗ , S ] = Q M P M ∗ Q − Q M ∗ P M Q on Q K . [S^{*},S]=QMPM^{*}Q-QM^{*}PMQ\quad\text{on }Q\mathcal{K}. [ S ∗ , S ] = QM P M ∗ Q − Q M ∗ P M Q on Q K .
Both terms are products of Hilbert–Schmidt operators, so [ S ∗ , S ] ∈ S 1 [S^{*},S]\in\mathcal{S}_{1} [ S ∗ , S ] ∈ S 1 . Thus the operator S S S is also almost normal.
III. Compactness. If T T T is compact, then ∥ A j ∥ , ∥ B j ∥ → 0 \lVert A_j\rVert,\lVert B_j\rVert\to0 ∥ A j ∥ , ∥ B j ∥ → 0 . For instance, failure of the block norms of T T T to tend to zero would give an orthonormal sequence with images bounded away from zero. The label bounds now give ∥ C j + i D j ∥ → 0 \lVert C_j+iD_j\rVert\to0 ∥ C j + i D j ∥ → 0 . Thus M M M , a sum of finite matrices with vanishing block norms, is compact, and its complementary compression S S S is compact as well. □
Finite lattice model on torus
On the square torus, we assign minus signs to a finite string of horizontal bonds. The resulting commutator is supported near the two endpoints. Loops surrounding one endpoint have holonomy − 1 -1 − 1 , and their contributions define a positive operator W L W_L W L whose trace grows logarithmically with L L L . This provides the lower bound needed later to obstruct normal comparison.
Our construction in this section is inspired by the magnetic flux model in mathematical physics, e.g. [12 §1] , [11 §11.1] . De Nittis and Schulz-Baldes [7 §2.2] represent a half-flux tube on the infinite square lattice by assigning the sign − 1 -1 − 1 to each bond crossing a half-line. We take two such tubes with opposite fluxes and put the resulting finite string of bond signs on a square torus.
Let m ≥ 1 m\ge1 m ≥ 1 be an integer and set L = 8 m L=8m L = 8 m . Write G L = ( Z / L Z ) 2 G_L=(\mathbb{Z}/L\mathbb{Z})^2 G L = ( Z / L Z ) 2 , and let e p , q e_{p,q} e p , q denote the standard basis vectors of H L = ℓ 2 ( G L ) \mathcal{H}_L=\ell^2(G_L) H L = ℓ 2 ( G L ) . In Z / L Z \mathbb{Z}/L\mathbb{Z} Z / L Z , we take the representative q q q in { 0 , … , L − 1 } \{0,\ldots,L-1\} { 0 , … , L − 1 } . Define
s ( p , q ) = { − 1 , p = 0 , 1 ≤ q ≤ L / 2 , 1 , otherwise . s(p,q)=
\begin{cases}
-1, & p=0,\quad1\le q\le L/2,\\
1, & \text{otherwise}.
\end{cases} s ( p , q ) = { − 1 , 1 , p = 0 , 1 ≤ q ≤ L /2 , otherwise .
We consider two unitary operators U , V U,V U , V on H L \mathcal{H}_L H L ,
U e p , q = s ( p , q ) e p + 1 , q , V e p , q = e p , q + 1 . (16) Ue_{p,q}=s(p,q)e_{p+1,q},\qquad Ve_{p,q}=e_{p,q+1}. \tag*{(16)} U e p , q = s ( p , q ) e p + 1 , q , V e p , q = e p , q + 1 . (16)
The subscripts on the right are read modulo L L L . Thus U U U is the signed horizontal shift associated with s s s and V V V is the ordinary vertical shift. The pair ( U , V ) (U,V) ( U , V ) defines a parallel transport on G L G_L G L . At vertex ( p , q ) ∈ G L (p,q)\in G_L ( p , q ) ∈ G L , the parallel transport along the paths from ( p , q ) (p,q) ( p , q ) to ( p + 1 , q + 1 ) (p+1,q+1) ( p + 1 , q + 1 ) are different by a sign s ( p , q ) s ( p , q + 1 ) s(p,q)s(p,q+1) s ( p , q ) s ( p , q + 1 ) . It is direct to check that at vertices ( 0 , 0 ) (0,0) ( 0 , 0 ) and ( 0 , L / 2 ) (0,L/2) ( 0 , L /2 ) , the signs are − 1 -1 − 1 .
Lemma 3.1. The commutator of U U U and V V V has two nonzero singular values, both equal to 2. For A L = Re U A_L=\operatorname{Re}U A L = Re U and B L = Re V B_L=\operatorname{Re}V B L = Re V ,
∥ [ U , V ] ∥ 1 = 4 , ∥ [ U , V ] ∥ 2 2 = 8 , ∥ [ A L , B L ] ∥ 1 = 4. (17) \lVert[U,V]\rVert_{1}=4,\qquad\lVert[U,V]\rVert_{2}^{2}=8,\qquad\lVert[A_L,B_L]\rVert_{1}=4. \tag*{(17)} ∥[ U , V ] ∥ 1 = 4 , ∥[ U , V ] ∥ 2 2 = 8 , ∥[ A L , B L ] ∥ 1 = 4. (17)
Proof. By (16) , the operator D = [ U , V ] D=[U,V] D = [ U , V ] acts by
D e p , q = ( s ( p , q + 1 ) − s ( p , q ) ) e p + 1 , q + 1 . D e_{p,q}=(s(p,q+1)-s(p,q))e_{p+1,q+1}. D e p , q = ( s ( p , q + 1 ) − s ( p , q )) e p + 1 , q + 1 .
The value s ( p , q + 1 ) − s ( p , q ) s(p,q+1)-s(p,q) s ( p , q + 1 ) − s ( p , q ) is nonzero only at ( 0 , 0 ) (0,0) ( 0 , 0 ) and ( 0 , L / 2 ) (0,L/2) ( 0 , L /2 ) . This gives the property about singular values of D D D .
For the real parts, we expand the commutator and use unitarity to obtain
4 [ A L , B L ] = [ U + U ∗ , V + V ∗ ] = D − V ∗ D V ∗ − U ∗ D U ∗ − D ∗ . 4[A_L,B_L]=[U+U^*,V+V^*]=D-V^*DV^*-U^*DU^*-D^*. 4 [ A L , B L ] = [ U + U ∗ , V + V ∗ ] = D − V ∗ D V ∗ − U ∗ D U ∗ − D ∗ .
For b = 0 , L / 2 b=0,L/2 b = 0 , L /2 , set δ b = s ( 0 , b + 1 ) − s ( 0 , b ) \delta_b=s(0,b+1)-s(0,b) δ b = s ( 0 , b + 1 ) − s ( 0 , b ) . The terms D − D ∗ D-D^* D − D ∗ give a skew-symmetric two-by-two block on e 0 , b , e 1 , b + 1 e_{0,b},e_{1,b+1} e 0 , b , e 1 , b + 1 with off-diagonal entries of absolute value 2. The term − V ∗ D V ∗ -V^*DV^* − V ∗ D V ∗ sends e 0 , b + 1 e_{0,b+1} e 0 , b + 1 to − δ b e 1 , b -\delta_b e_{1,b} − δ b e 1 , b . The term − U ∗ D U ∗ -U^*DU^* − U ∗ D U ∗ sends e 1 , b e_{1,b} e 1 , b to δ b e 0 , b + 1 \delta_b e_{0,b+1} δ b e 0 , b + 1 . This is a second skew-symmetric two-by-two block. The four resulting blocks for b = 0 , L / 2 b=0,L/2 b = 0 , L /2 act on disjoint sets of basis vectors and vanish on the complementary space. Thus [ A L , B L ] [A_L,B_L] [ A L , B L ] has eight singular values equal to 1 / 2 1/2 1/2 .
For 0 ≤ r < m 0\le r<m 0 ≤ r < m , let γ r \gamma_r γ r be the image in G L G_L G L of the lattice boundary of the square [ − r , r + 1 ] 2 [-r,r+1]^2 [ − r , r + 1 ] 2 . Its number of edges and its number of vertices are both ℓ r = 4 ( 2 r + 1 ) \ell_r=4(2r+1) ℓ r = 4 ( 2 r + 1 ) . The union of the lifted boundaries lies in [ − m + 1 , m ] 2 [-m+1,m]^2 [ − m + 1 , m ] 2 , which projects injectively to G L G_L G L on lattice vertices. The boundaries are thus vertex-disjoint. On γ r \gamma_r γ r , the top side meets one negative bond1 at height q = r + 1 q=r+1 q = r + 1 . The bottom side has height q = 0 q=0 q = 0 for r = 0 r=0 r = 0 and residue L − r L-r L − r for r > 0 r>0 r > 0 , so on all its bonds s s s has positive sign. Thus the product of signs of bonds along every γ r \gamma_r γ r is − 1 -1 − 1 . Figure 1 shows the first three loops.
Figure 1. The first three loops for m ≥ 3 m\ge3 m ≥ 3 , in the lifted lattice. Each loop crosses one negative edge. The other flux plaquette lies beyond the part of the string shown.
Let P γ r P_{\gamma_r} P γ r denote the orthogonal projection onto the subspace generated by the vertices of γ r \gamma_r γ r . Define the positive operator
W L = ∑ r = 0 m − 1 4 ℓ r 2 P γ r . (18) W_L=\sum_{r=0}^{m-1}\frac{4}{\ell_r^2}P_{\gamma_r}. \tag*{(18)} W L = r = 0 ∑ m − 1 ℓ r 2 4 P γ r . (18)
The trace contribution of γ r \gamma_r γ r is 4 / ℓ r 4/\ell_r 4/ ℓ r . In particular,
Tr W L = ∑ r = 0 m − 1 1 2 r + 1 ≥ 1 2 ∑ k = 1 m 1 k ≥ 1 2 log ( m + 1 ) . (19) \operatorname{Tr}W_L=\sum_{r=0}^{m-1}\frac{1}{2r+1}\ge\frac{1}{2}\sum_{k=1}^{m}\frac{1}{k}\ge\frac{1}{2}\log(m+1). \tag*{(19)} Tr W L = r = 0 ∑ m − 1 2 r + 1 1 ≥ 2 1 k = 1 ∑ m k 1 ≥ 2 1 log ( m + 1 ) . (19)
Lemma 3.2. For every ζ , η ∈ T = { z ∈ C ∣ ∣ z ∣ = 1 } \zeta,\eta\in\mathbb{T}=\{z\in\mathbb{C}\mid|z|=1\} ζ , η ∈ T = { z ∈ C ∣ ∣ z ∣ = 1 } ,
( U − ζ I ) ∗ ( U − ζ I ) + ( V − η I ) ∗ ( V − η I ) ≥ W L . (20) (U-\zeta I)^*(U-\zeta I)+(V-\eta I)^*(V-\eta I)\ge W_L. \tag*{(20)} ( U − ζ I ) ∗ ( U − ζ I ) + ( V − η I ) ∗ ( V − η I ) ≥ W L . (20)
Proof. For u ∈ ℓ 2 ( G L ) u \in\ell^{2}(G_L) u ∈ ℓ 2 ( G L ) , we consider the square error
E ( u ) : = ∥ ( U − ζ I ) u ∥ 2 + ∥ ( V − η I ) u ∥ 2 . \mathcal{E}(u) := \lVert(U-\zeta I)u\rVert^{2}+\lVert(V-\eta I)u\rVert^{2}. E ( u ) := ∥( U − ζ I ) u ∥ 2 + ∥( V − η I ) u ∥ 2 .
With the shift convention (3.1), we compute E ( u ) \mathcal{E}(u) E ( u ) as
∑ p , q ( ∣ u ( p + 1 , q ) − ζ ‾ s ( p , q ) u ( p , q ) ∣ 2 + ∣ u ( p , q + 1 ) − η ‾ u ( p , q ) ∣ 2 ) . \sum_{p,q}\left(|u(p+1,q)-\overline{\zeta}s(p,q)u(p,q)|^{2}+|u(p,q+1)-\overline{\eta}u(p,q)|^{2}\right). p , q ∑ ( ∣ u ( p + 1 , q ) − ζ s ( p , q ) u ( p , q ) ∣ 2 + ∣ u ( p , q + 1 ) − η u ( p , q ) ∣ 2 ) .
We assign the phases ζ ‾ s ( p , q ) \overline{\zeta}s(p,q) ζ s ( p , q ) and η ‾ \overline{\eta} η to forward bonds and their inverses to reversed bonds.
Fix a loop γ r \gamma_r γ r of length ℓ = ℓ r \ell=\ell_r ℓ = ℓ r , and let R r R_r R r be the unitary cyclic shift on ℓ 2 ( γ r ) \ell^{2}(\gamma_r) ℓ 2 ( γ r ) carrying the above bond phases. Note that the ζ , η \zeta,\eta ζ , η factors along γ r \gamma_r γ r cancel, and its bond-sign product is − 1 -1 − 1 . Hence
R r ℓ = − I . R_r^{\ell}=-I. R r ℓ = − I .
We consider the factorization,
2 I = I − R r ℓ = ( I + R r + ⋯ + R r ℓ − 1 ) ( I − R r ) . 2I=I-R_r^{\ell}=(I+R_r+\cdots+R_r^{\ell-1})(I-R_r). 2 I = I − R r ℓ = ( I + R r + ⋯ + R r ℓ − 1 ) ( I − R r ) .
Since R r R_r R r is unitary, for every v ∈ ℓ 2 ( γ r ) v\in\ell^{2}(\gamma_r) v ∈ ℓ 2 ( γ r ) , we have the following estimates,
2 ∥ v ∥ ≤ ℓ ∥ ( I − R r ) v ∥ , ∥ ( I − R r ) v ∥ 2 ≥ 4 ℓ 2 ∥ v ∥ 2 . 2\lVert v\rVert\leq\ell\lVert(I-R_r)v\rVert,\qquad\lVert(I-R_r)v\rVert^{2}\geq\frac{4}{\ell^{2}}\lVert v\rVert^{2}. 2 ∥ v ∥ ≤ ℓ ∥( I − R r ) v ∥ , ∥( I − R r ) v ∥ 2 ≥ ℓ 2 4 ∥ v ∥ 2 .
Consider the restriction u r = u ∣ γ r u_r=u|_{\gamma_r} u r = u ∣ γ r . We observe that the quantity ∥ ( I − R r ) u r ∥ 2 \lVert(I-R_r)u_r\rVert^{2} ∥( I − R r ) u r ∥ 2 is exactly the bond energy along γ r \gamma_r γ r . Since the loops have disjoint edge sets, we have
E ( u ) ≥ ∑ r = 0 m − 1 ∥ ( I − R r ) u r ∥ 2 ≥ ∑ r = 0 m − 1 4 ℓ r 2 ∥ P γ r u ∥ 2 = ⟨ W L u , u ⟩ , \mathcal{E}(u)\geq\sum_{r=0}^{m-1}\lVert(I-R_r)u_r\rVert^{2}\geq\sum_{r=0}^{m-1}\frac{4}{\ell_r^{2}}\lVert P_{\gamma_r}u\rVert^{2}=\langle W_Lu,u\rangle, E ( u ) ≥ r = 0 ∑ m − 1 ∥( I − R r ) u r ∥ 2 ≥ r = 0 ∑ m − 1 ℓ r 2 4 ∥ P γ r u ∥ 2 = ⟨ W L u , u ⟩ ,
which proves Lemma 3.2. □
The loop inequality concerns the unitary coordinates U , V U,V U , V . The frame criterion requires an inequality for their real parts, uniform in the real labels x , y x,y x , y . We obtain this by applying a fixed smooth cutoff supported on an upper arc of the circle.
The real parts of the shifts
We convert the loop inequality, Lemma 3.2, into a logarithmic lower bound for the frame error of A L = Re U A_L=\operatorname{Re}U A L = Re U and B L = Re V B_L=\operatorname{Re}V B L = Re V . The resulting self-adjoint test operator retains the logarithmic trace growth of W L W_L W L , giving d ( A L , B L ) ≥ c log L − C 0 \mathfrak{d}(A_L,B_L)\geq c\log L-C_0 d ( A L , B L ) ≥ c log L − C 0 . This is the finite-dimensional obstruction used in the counterexample construction.
Recall A L = Re U A_L=\operatorname{Re}U A L = Re U and B L = Re V B_L=\operatorname{Re}V B L = Re V from Lemma 3.1. Both are self-adjoint contractions.
Proposition 4.1. There are constants c > 0 c>0 c > 0 and C 0 < ∞ C_0<\infty C 0 < ∞ , independent of L = 8 m L=8m L = 8 m , such that for every such L L L there is a self-adjoint operator Y L Y_L Y L on H L H_L H L satisfying
Y L ≤ ( A L − x I ) 2 + ( B L − y I ) 2 ( x , y ∈ R ) Y_L\leq(A_L-xI)^{2}+(B_L-yI)^{2}\qquad(x,y\in\mathbb{R}) Y L ≤ ( A L − x I ) 2 + ( B L − y I ) 2 ( x , y ∈ R ) and
Tr Y L ≥ c log L − C 0 . \operatorname{Tr}Y_L\geq c\log L-C_0. Tr Y L ≥ c log L − C 0 . Consequently,
d ( A L , B L ) ≥ c log L − C 0 . (21) \mathfrak{d}(A_L,B_L)\geq c\log L-C_0. \tag*{(21)} d ( A L , B L ) ≥ c log L − C 0 . (21)
Proof. We split the proof into three steps.
I. Construction of Y L Y_L Y L . Fix once and for all a nonzero f ∈ C ∞ ( T ) f \in C^\infty(\mathbb{T}) f ∈ C ∞ ( T ) with 0 ≤ f ≤ 1 0 \le f \le1 0 ≤ f ≤ 1 and support in { z ∈ T ∣ Im z ≥ 1 / 2 } \{z \in\mathbb{T} \mid\operatorname{Im} z \ge1/2\} { z ∈ T ∣ Im z ≥ 1/2 } . We can choose f f f to be supported in a small arc about i i i . Set
F = f ( U ) f ( V ) , K 1 = [ A L , f ( V ) ] , K 2 = [ V , f ( U ) ] f ( V ) . F=f(U)f(V), \qquad K_1=[A_L,f(V)], \qquad K_2=[V,f(U)]f(V). F = f ( U ) f ( V ) , K 1 = [ A L , f ( V )] , K 2 = [ V , f ( U )] f ( V ) .
We first transfer the inequality (20) to the real coordinates.
For x ∈ [ − 1 , 1 ] x \in[-1,1] x ∈ [ − 1 , 1 ] , set ζ = x + i 1 − x 2 \zeta=x+i\sqrt{1-x^2} ζ = x + i 1 − x 2 . For z ∈ T z \in\mathbb{T} z ∈ T ,
2 z ( Re z − x ) = ( z − ζ ) ( z − ζ ‾ ) . 2z(\operatorname{Re}z-x)=(z-\zeta)(z-\overline{\zeta}). 2 z ( Re z − x ) = ( z − ζ ) ( z − ζ ) .
On supp f \operatorname{supp} f supp f , the distance ∣ z − ζ ‾ ∣ |z-\overline{\zeta}| ∣ z − ζ ∣ is at least 1 / 2 1/2 1/2 . So ∣ ( z − ζ ) f ( z ) ∣ ≤ 4 ∣ Re z − x ∣ |(z-\zeta)f(z)| \le4|\operatorname{Re}z-x| ∣ ( z − ζ ) f ( z ) ∣ ≤ 4∣ Re z − x ∣ on all z ∈ T z \in\mathbb{T} z ∈ T . By the spectral theorem for U U U , we get
∥ ( U − ζ I ) f ( U ) v ∥ ≤ 4 ∥ ( A L − x I ) v ∥ . (22) \|(U-\zeta I)f(U)v\| \le4\|(A_L-xI)v\|. \tag*{(22)} ∥ ( U − ζ I ) f ( U ) v ∥ ≤ 4∥ ( A L − x I ) v ∥. (22)
Similarly, we have the corresponding estimate for V , B L V,B_L V , B L
∥ ( V − η I ) f ( V ) u ∥ ≤ 4 ∥ ( B L − y I ) u ∥ , (23) \|(V-\eta I)f(V)u\| \le4\|(B_L-yI)u\|, \tag*{(23)} ∥ ( V − η I ) f ( V ) u ∥ ≤ 4∥ ( B L − y I ) u ∥ , (23)
where y ∈ [ − 1 , 1 ] y \in[-1,1] y ∈ [ − 1 , 1 ] and η = y + i 1 − y 2 \eta=y+i\sqrt{1-y^2} η = y + i 1 − y 2 . Apply inequality (22) to v = f ( V ) u v=f(V)u v = f ( V ) u and commute A L − x I A_L-xI A L − x I past f ( V ) f(V) f ( V ) . For the second inequality use
( V − η I ) F = f ( U ) ( V − η I ) f ( V ) + [ V , f ( U ) ] f ( V ) . (V-\eta I)F=f(U)(V-\eta I)f(V)+[V,f(U)]f(V). ( V − η I ) F = f ( U ) ( V − η I ) f ( V ) + [ V , f ( U )] f ( V ) .
Since f ( U ) f(U) f ( U ) and f ( V ) f(V) f ( V ) are contractions, we obtain
∥ ( U − ζ I ) F u ∥ ≤ 4 ∥ ( A L − x I ) u ∥ + 4 ∥ K 1 u ∥ , (24) \|(U-\zeta I)Fu\| \le4\|(A_L-xI)u\|+4\|K_1u\|, \tag*{(24)} ∥ ( U − ζ I ) F u ∥ ≤ 4∥ ( A L − x I ) u ∥ + 4∥ K 1 u ∥ , (24)
∥ ( V − η I ) F u ∥ ≤ 4 ∥ ( B L − y I ) u ∥ + ∥ K 2 u ∥ . (25) \|(V-\eta I)Fu\| \le4\|(B_L-yI)u\|+\|K_2u\|. \tag*{(25)} ∥ ( V − η I ) F u ∥ ≤ 4∥ ( B L − y I ) u ∥ + ∥ K 2 u ∥. (25)
Fix x , y ∈ [ − 1 , 1 ] x,y \in[-1,1] x , y ∈ [ − 1 , 1 ] , and let ζ , η ∈ T \zeta,\eta\in\mathbb{T} ζ , η ∈ T be the corresponding phases used in (24) and (25) . Applying Lemma 3.2 to F u Fu F u gives
⟨ W L F u , F u ⟩ ≤ ∥ ( U − ζ I ) F u ∥ 2 + ∥ ( V − η I ) F u ∥ 2 . (26) \langle W_LFu,Fu\rangle\le\|(U-\zeta I)Fu\|^2+\|(V-\eta I)Fu\|^2. \tag*{(26)} ⟨ W L F u , F u ⟩ ≤ ∥ ( U − ζ I ) F u ∥ 2 + ∥ ( V − η I ) F u ∥ 2 . (26)
By (24) and the inequality ( a + b ) 2 ≤ 2 a 2 + 2 b 2 (a+b)^2 \le2a^2+2b^2 ( a + b ) 2 ≤ 2 a 2 + 2 b 2 , we obtain
∥ ( U − ζ I ) F u ∥ 2 ≤ ( 4 ∥ ( A L − x I ) u ∥ + 4 ∥ K 1 u ∥ ) 2 ≤ 32 ∥ ( A L − x I ) u ∥ 2 + 32 ∥ K 1 u ∥ 2 . (27) \begin{aligned}
\|(U-\zeta I)Fu\|^2
&\le\bigl(4\|(A_L-xI)u\|+4\|K_1u\|\bigr)^2\\
&\le32\|(A_L-xI)u\|^2+32\|K_1u\|^2. \tag*{(27)}
\end{aligned} ∥ ( U − ζ I ) F u ∥ 2 ≤ ( 4∥ ( A L − x I ) u ∥ + 4∥ K 1 u ∥ ) 2 ≤ 32∥ ( A L − x I ) u ∥ 2 + 32∥ K 1 u ∥ 2 . (27)
Similarly, (25) gives
∥ ( V − η I ) F u ∥ 2 ≤ ( 4 ∥ ( B L − y I ) u ∥ + ∥ K 2 u ∥ ) 2 ≤ 32 ∥ ( B L − y I ) u ∥ 2 + 2 ∥ K 2 u ∥ 2 . (28) \begin{aligned}
\|(V-\eta I)Fu\|^2
&\le\bigl(4\|(B_L-yI)u\|+\|K_2u\|\bigr)^2\\
&\le32\|(B_L-yI)u\|^2+2\|K_2u\|^2. \tag*{(28)}
\end{aligned} ∥ ( V − η I ) F u ∥ 2 ≤ ( 4∥ ( B L − y I ) u ∥ + ∥ K 2 u ∥ ) 2 ≤ 32∥ ( B L − y I ) u ∥ 2 + 2∥ K 2 u ∥ 2 . (28)
Combining estimates (26) , (27) , (28) and dividing by 32 32 32 , we obtain
1 32 ⟨ F ∗ W L F u , u ⟩ ≤ ∥ ( A L − x I ) u ∥ 2 + ∥ ( B L − y I ) u ∥ 2 + ∥ K 1 u ∥ 2 + 1 16 ∥ K 2 u ∥ 2 . (29) \frac{1}{32}\langle F^*W_LF u,u\rangle\le\|(A_L-xI)u\|^2+\|(B_L-yI)u\|^2+\|K_1u\|^2+\frac{1}{16}\|K_2u\|^2. \tag*{(29)} 32 1 ⟨ F ∗ W L F u , u ⟩ ≤ ∥ ( A L − x I ) u ∥ 2 + ∥ ( B L − y I ) u ∥ 2 + ∥ K 1 u ∥ 2 + 16 1 ∥ K 2 u ∥ 2 . (29)
Consequently, the self-adjoint operator
Y L = 1 32 F ∗ W L F − K 1 ∗ K 1 − 1 16 K 2 ∗ K 2 (30) Y_L=\frac{1}{32}F^*W_LF-K_1^*K_1-\frac{1}{16}K_2^*K_2 \tag*{(30)} Y L = 32 1 F ∗ W L F − K 1 ∗ K 1 − 16 1 K 2 ∗ K 2 (30)
satisfies
⟨ Y L u , u ⟩ ≤ ∥ ( A L − x I ) u ∥ 2 + ∥ ( B L − y I ) u ∥ 2 . (31) \langle Y_Lu,u\rangle\le\|(A_L-xI)u\|^2+\|(B_L-yI)u\|^2. \tag*{(31)} ⟨ Y L u , u ⟩ ≤ ∥ ( A L − x I ) u ∥ 2 + ∥ ( B L − y I ) u ∥ 2 . (31)
Since the inequality (31) holds for every u ∈ H L u \in\mathcal{H}_L u ∈ H L , it is equivalent to the operator inequality
Y L ≤ ( A L − x I ) 2 + ( B L − y I ) 2 Y_L \le(A_L-xI)^2+(B_L-yI)^2 Y L ≤ ( A L − x I ) 2 + ( B L − y I ) 2
which is (10) . The operator Y L Y_L Y L is independent of x , y x,y x , y , so the same inequality holds throughout [ − 1 , 1 ] 2 [-1,1]^2 [ − 1 , 1 ] 2 .
To extend the inequality to arbitrary real x , y x,y x , y , define
x ^ = max { − 1 , min { x , 1 } } , y ^ = max { − 1 , min { y , 1 } } . \widehat{x}=\max\{-1,\min\{x,1\}\}, \qquad\widehat{y}=\max\{-1,\min\{y,1\}\}. x = max { − 1 , min { x , 1 }} , y = max { − 1 , min { y , 1 }} .
For every t ∈ [ − 1 , 1 ] t \in[-1,1] t ∈ [ − 1 , 1 ] ,
∣ t − x ^ ∣ ≤ ∣ t − x ∣ , ∣ t − y ^ ∣ ≤ ∣ t − y ∣ . |t-\widehat{x}| \le|t-x|, \qquad|t-\widehat{y}| \le|t-y|. ∣ t − x ∣ ≤ ∣ t − x ∣ , ∣ t − y ∣ ≤ ∣ t − y ∣.
Because A L A_L A L and B L B_L B L are self-adjoint contractions, the spectral theorem therefore gives
( A L − x ^ I ) 2 ≤ ( A L − x I ) 2 , ( B L − y ^ I ) 2 ≤ ( B L − y I ) 2 . (32) (A_L-\widehat{x}I)^2 \le(A_L-xI)^2, \qquad(B_L-\widehat{y}I)^2 \le(B_L-yI)^2. \tag*{(32)} ( A L − x I ) 2 ≤ ( A L − x I ) 2 , ( B L − y I ) 2 ≤ ( B L − y I ) 2 . (32)
Applying (31) at ( x ^ , y ^ ) (\widehat{x},\widehat{y}) ( x , y ) and then using (32) , we obtain
Y L ≤ ( A L − x ^ I ) 2 + ( B L − y ^ I ) 2 ≤ ( A L − x I ) 2 + ( B L − y I ) 2 . (33) Y_L \le(A_L-\widehat{x}I)^2+(B_L-\widehat{y}I)^2 \le(A_L-xI)^2+(B_L-yI)^2. \tag*{(33)} Y L ≤ ( A L − x I ) 2 + ( B L − y I ) 2 ≤ ( A L − x I ) 2 + ( B L − y I ) 2 . (33)
Thus (10) holds for all real labels.
Finally, by (11) and (33) , we conclude
∂ ( A L , B L ) ≥ Tr Y L . (34) \partial(A_L,B_L) \ge\operatorname{Tr}Y_L. \tag*{(34)} ∂ ( A L , B L ) ≥ Tr Y L . (34)
Estimates of the correction terms
We prove that the two correction terms have trace bounded independently of L L L . Write the Fourier series f ( z ) = ∑ k ∈ Z c k z k f(z)=\sum_{k\in\mathbb{Z}}c_kz^k f ( z ) = ∑ k ∈ Z c k z k and set M f = ∑ k ∈ Z ∣ k ∣ ∣ c k ∣ < ∞ M_f=\sum_{k\in\mathbb{Z}}|k||c_k|<\infty M f = ∑ k ∈ Z ∣ k ∣∣ c k ∣ < ∞ . For a unitary R R R , the identity [ X , R k ] = ∑ a = 0 k − 1 R a [ X , R ] R k − 1 − a [X,R^k]=\sum_{a=0}^{k-1}R^a[X,R]R^{k-1-a} [ X , R k ] = ∑ a = 0 k − 1 R a [ X , R ] R k − 1 − a for k ≥ 1 k\ge1 k ≥ 1 gives ∥ [ X , R k ] ∥ 2 ≤ k ∥ [ X , R ] ∥ 2 \|[X,R^k]\|_2\le k\|[X,R]\|_2 ∥ [ X , R k ] ∥ 2 ≤ k ∥ [ X , R ] ∥ 2 . The bound for negative powers follows from [ X , R − 1 ] = − R − 1 [ X , R ] R − 1 [X,R^{-1}]=-R^{-1}[X,R]R^{-1} [ X , R − 1 ] = − R − 1 [ X , R ] R − 1 . Thus ∥ [ X , f ( R ) ] ∥ 2 ≤ M f ∥ [ X , R ] ∥ 2 \|[X,f(R)]\|_2\le M_f\|[X,R]\|_2 ∥ [ X , f ( R )] ∥ 2 ≤ M f ∥ [ X , R ] ∥ 2 . Since [ U ∗ , V ] = − [ U , V ] U ∗ [U^*,V]=-[U,V]U^* [ U ∗ , V ] = − [ U , V ] U ∗ , we have ∥ [ A L , V ] ∥ 2 ≤ ∥ [ U , V ] ∥ 2 \|[A_L,V]\|_2\le\|[U,V]\|_2 ∥ [ A L , V ] ∥ 2 ≤ ∥ [ U , V ] ∥ 2 . Together with (17) and ∥ f ( V ) ∥ ≤ 1 \|f(V)\|\le1 ∥ f ( V ) ∥ ≤ 1 , these estimates give
∥ K 1 ∥ 2 2 + 1 16 ∥ K 2 ∥ 2 2 ≤ 17 2 M f 2 . (35) \|K_1\|_2^2+\frac{1}{16}\|K_2\|_2^2\le\frac{17}{2}M_f^2. \tag*{(35)} ∥ K 1 ∥ 2 2 + 16 1 ∥ K 2 ∥ 2 2 ≤ 2 17 M f 2 . (35)
Estimate of Tr ( F ∗ W L F ) \operatorname{Tr}(F^*W_LF) Tr ( F ∗ W L F )
It remains to obtain a lower bound for the positive contribution Tr ( F ∗ W L F ) \operatorname{Tr}(F^*W_LF) Tr ( F ∗ W L F ) in (30) . Since W L W_L W L is diagonal in the vertex basis, a lower bound for Tr ( F ∗ W L F ) \operatorname{Tr}(F^*W_LF) Tr ( F ∗ W L F ) follows from a uniform lower bound for ∥ F ∗ e p , q ∥ 2 \|F^*e_{p,q}\|^2 ∥ F ∗ e p , q ∥ 2 . The row and column decompositions allow us to compute these norms.
For ω ∈ T \omega\in\mathbb{T} ω ∈ T , define
α L ( ω ) : = 1 L ∑ k = 0 L − 1 ∣ f ( ω e 2 π i k / L ) ∣ 2 , a f : = ∫ 0 2 π ∣ f ( e i θ ) ∣ 2 d θ 2 π . \alpha_L(\omega):=\frac{1}{L}\sum_{k=0}^{L-1}|f(\omega e^{2\pi ik/L})|^2,\qquad a_f:=\int_0^{2\pi}|f(e^{i\theta})|^2\frac{d\theta}{2\pi}. α L ( ω ) := L 1 k = 0 ∑ L − 1 ∣ f ( ω e 2 π ik / L ) ∣ 2 , a f := ∫ 0 2 π ∣ f ( e i θ ) ∣ 2 2 π d θ .
Since f f f is continuous and not identically zero, a f > 0 a_f>0 a f > 0 . We first observe that
α L ( ω ) → a f (36) \alpha_L(\omega)\to a_f \tag*{(36)} α L ( ω ) → a f (36)
uniformly in ω \omega ω as L → ∞ L\to\infty L → ∞ .
We next compute the norm of f ( U ) ∗ e p , q f(U)^*e_{p,q} f ( U ) ∗ e p , q . Both U U U and U ∗ U^* U ∗ preserve the row subspace
R q = span { e p , q : 0 ≤ p < L } , \mathcal{R}_q=\operatorname{span}\{e_{p,q}:0\le p<L\}, R q = span { e p , q : 0 ≤ p < L } ,
so f ( U ) ∗ e p , q f(U)^*e_{p,q} f ( U ) ∗ e p , q also belongs to R q \mathcal{R}_q R q . Define σ q : = ∏ p = 0 L − 1 s ( p , q ) \sigma_q:=\prod_{p=0}^{L-1}s(p,q) σ q := ∏ p = 0 L − 1 s ( p , q ) , and choose ω q ∈ T \omega_q\in\mathbb{T} ω q ∈ T satisfying ω q L = σ q \omega_q^L=\sigma_q ω q L = σ q . Let S q S_q S q be the ordinary cyclic shift on this row:
S q e p , q = e p + 1 , q . S_qe_{p,q}=e_{p+1,q}. S q e p , q = e p + 1 , q .
Set d 0 = 1 d_0=1 d 0 = 1 and define successively
d p + 1 : = ω q s ( p , q ) − 1 d p . d_{p+1}:=\omega_qs(p,q)^{-1}d_p. d p + 1 := ω q s ( p , q ) − 1 d p .
We observe that d p d_p d p has modulus one, and after traversing the row,
d L = ω q L ( ∏ p = 0 L − 1 s ( p , q ) ) − 1 d 0 = σ q σ q − 1 = d 0 . d_L=\omega_q^L\left(\prod_{p=0}^{L-1}s(p,q)\right)^{-1}d_0=\sigma_q\sigma_q^{-1}=d_0. d L = ω q L ( p = 0 ∏ L − 1 s ( p , q ) ) − 1 d 0 = σ q σ q − 1 = d 0 .
Thus these entries are consistent with the periodic indexing. Define G q e p , q : = d p e p , q G_qe_{p,q}:=d_pe_{p,q} G q e p , q := d p e p , q . We directly check that
G q U G q ∗ e p , q = s ( p , q ) d p + 1 d p ‾ e p + 1 , q = ω q e p + 1 , q . G_qUG_q^*e_{p,q}=s(p,q)d_{p+1}\overline{d_p}e_{p+1,q}=\omega_qe_{p+1,q}. G q U G q ∗ e p , q = s ( p , q ) d p + 1 d p e p + 1 , q = ω q e p + 1 , q .
We have constructed a diagonal unitary G q G_q G q on R q \mathcal{R}_q R q such that
G q ( U ∣ R q ) G q ∗ = ω q S q . (37) G_q(U|_{\mathcal{R}_q})G_q^*=\omega_qS_q. \tag*{(37)} G q ( U ∣ R q ) G q ∗ = ω q S q . (37)
Write ρ k = e 2 π i k / L \rho_k=e^{2\pi ik/L} ρ k = e 2 π ik / L . Define v q , k ∈ R q v_{q,k}\in\mathcal{R}_q v q , k ∈ R q by
v q , k : = 1 L ∑ r = 0 L − 1 ρ k − r e r , q , 0 ≤ k < L . v_{q,k}:=\frac{1}{\sqrt{L}}\sum_{r=0}^{L-1}\rho_k^{-r}e_{r,q},\qquad0\leq k<L. v q , k := L 1 r = 0 ∑ L − 1 ρ k − r e r , q , 0 ≤ k < L .
We check directly that ( v q , k ) k (v_{q,k})_k ( v q , k ) k form an orthonormal basis of R q \mathcal{R}_q R q and satisfy
S q v q , k = ρ k v q , k , ∣ ⟨ e p , q , v q , k ⟩ ∣ 2 = 1 L . S_qv_{q,k}=\rho_kv_{q,k},\qquad|\langle e_{p,q},v_{q,k}\rangle|^2=\frac{1}{L}. S q v q , k = ρ k v q , k , ∣ ⟨ e p , q , v q , k ⟩ ∣ 2 = L 1 .
Applying (37) , we compute the norm of f ( U ) ∗ e p , q f(U)^*e_{p,q} f ( U ) ∗ e p , q by
∥ f ( U ) ∗ e p , q ∥ 2 = ∥ f ( ω q S q ) ∗ e p , q ∥ 2 = ∑ k = 0 L − 1 ∣ f ( ω q ρ k ) ∣ 2 ∣ ⟨ e p , q , v q , k ⟩ ∣ 2 = 1 L ∑ k = 0 L − 1 ∣ f ( ω q ρ k ) ∣ 2 = α L ( ω q ) . \begin{aligned}
\|f(U)^*e_{p,q}\|^2&=\|f(\omega_qS_q)^*e_{p,q}\|^2\\
&=\sum_{k=0}^{L-1}|f(\omega_q\rho_k)|^2|\langle e_{p,q},v_{q,k}\rangle|^2\\
&=\frac{1}{L}\sum_{k=0}^{L-1}|f(\omega_q\rho_k)|^2\\
&=\alpha_L(\omega_q).
\end{aligned} ∥ f ( U ) ∗ e p , q ∥ 2 = ∥ f ( ω q S q ) ∗ e p , q ∥ 2 = k = 0 ∑ L − 1 ∣ f ( ω q ρ k ) ∣ 2 ∣ ⟨ e p , q , v q , k ⟩ ∣ 2 = L 1 k = 0 ∑ L − 1 ∣ f ( ω q ρ k ) ∣ 2 = α L ( ω q ) .
In particular, we observe that this norm is independent of the horizontal coordinate p p p . So we can write
f ( U ) ∗ e p , q = ∑ r = 0 L − 1 c r e r , q , ∑ r = 0 L − 1 ∣ c r ∣ 2 = α L ( ω q ) . (38) f(U)^*e_{p,q}=\sum_{r=0}^{L-1}c_re_{r,q},\qquad\sum_{r=0}^{L-1}|c_r|^2=\alpha_L(\omega_q). \tag*{(38)} f ( U ) ∗ e p , q = r = 0 ∑ L − 1 c r e r , q , r = 0 ∑ L − 1 ∣ c r ∣ 2 = α L ( ω q ) . (38)
Similarly, for each r r r , the column subspace
C r = span { e r , t : 0 ≤ t < L } \mathcal{C}_r=\operatorname{span}\{e_{r,t}:0\leq t<L\} C r = span { e r , t : 0 ≤ t < L }
is preserved by V V V , V ∗ V^* V ∗ , and hence f ( V ) ∗ f(V)^* f ( V ) ∗ . The spaces C r \mathcal{C}_r C r are mutually orthogonal. Consequently, for different r r r , the vectors f ( V ) ∗ e r , q f(V)^*e_{r,q} f ( V ) ∗ e r , q belonging to different columns are orthogonal. On each column, V V V is the ordinary cyclic shift, so the same Fourier calculation gives
∥ f ( V ) ∗ e r , q ∥ 2 = 1 L ∑ k = 0 L − 1 ∣ f ( ρ k ) ∣ 2 = α L ( 1 ) . \|f(V)^*e_{r,q}\|^2=\frac{1}{L}\sum_{k=0}^{L-1}|f(\rho_k)|^2=\alpha_L(1). ∥ f ( V ) ∗ e r , q ∥ 2 = L 1 k = 0 ∑ L − 1 ∣ f ( ρ k ) ∣ 2 = α L ( 1 ) .
We compute the norm of F ∗ e p , q F^*e_{p,q} F ∗ e p , q as follows:
∥ F ∗ e p , q ∥ 2 = ∥ f ( V ) ∗ f ( U ) ∗ e p , q ∥ 2 = ∥ ∑ r = 0 L − 1 c r f ( V ) ∗ e r , q ∥ 2 = ∑ r = 0 L − 1 ∣ c r ∣ 2 ∥ f ( V ) ∗ e r , q ∥ 2 = α L ( 1 ) ∑ r = 0 L − 1 ∣ c r ∣ 2 = α L ( 1 ) α L ( ω q ) . \begin{aligned}
\lVert F^*e_{p,q}\rVert^2 &= \lVert f(V)^*f(U)^*e_{p,q}\rVert^2 \\
&= \left\lVert\sum_{r=0}^{L-1} c_r f(V)^*e_{r,q}\right\rVert^2 \\
&= \sum_{r=0}^{L-1} \lvert c_r\rvert^2\lVert f(V)^*e_{r,q}\rVert^2 \\
&= \alpha_L(1)\sum_{r=0}^{L-1}\lvert c_r\rvert^2 \\
&= \alpha_L(1)\alpha_L(\omega_q).
\end{aligned} ∥ F ∗ e p , q ∥ 2 = ∥ f ( V ) ∗ f ( U ) ∗ e p , q ∥ 2 = r = 0 ∑ L − 1 c r f ( V ) ∗ e r , q 2 = r = 0 ∑ L − 1 ∣ c r ∣ 2 ∥ f ( V ) ∗ e r , q ∥ 2 = α L ( 1 ) r = 0 ∑ L − 1 ∣ c r ∣ 2 = α L ( 1 ) α L ( ω q ) .
By the uniform convergence (36) , we conclude that there is L 0 > 0 L_0>0 L 0 > 0 such that
∥ F ∗ e p , q ∥ 2 = α L ( 1 ) α L ( ω q ) ≥ a f 2 4 (39) \lVert F^*e_{p,q}\rVert^2=\alpha_L(1)\alpha_L(\omega_q)\geq\frac{a_f^2}{4} \tag*{(39)} ∥ F ∗ e p , q ∥ 2 = α L ( 1 ) α L ( ω q ) ≥ 4 a f 2 (39)
for every vertex whenever L ≥ L 0 L\geq L_0 L ≥ L 0 .
Since W L W_L W L is positive and diagonal in the vertex basis, write
W L e p , q = w p , q e p , q , w p , q ≥ 0. W_Le_{p,q}=w_{p,q}e_{p,q},\qquad w_{p,q}\geq0. W L e p , q = w p , q e p , q , w p , q ≥ 0.
Using the cyclicity of trace and (39) , we obtain the following trace computation,
Tr ( F ∗ W L F ) = Tr ( W L F F ∗ ) = ∑ ( p , q ) ∈ G L w p , q ⟨ F F ∗ e p , q , e p , q ⟩ = ∑ ( p , q ) ∈ G L w p , q ∥ F ∗ e p , q ∥ 2 ≥ a f 2 4 ∑ ( p , q ) ∈ G L w p , q = a f 2 4 Tr W L . (40) \begin{aligned}
\operatorname{Tr}(F^*W_LF)&=\operatorname{Tr}(W_LFF^*)\\
&=\sum_{(p,q)\in G_L}w_{p,q}\langle FF^*e_{p,q},e_{p,q}\rangle\\
&=\sum_{(p,q)\in G_L}w_{p,q}\lVert F^*e_{p,q}\rVert^2\\
&\geq\frac{a_f^2}{4}\sum_{(p,q)\in G_L}w_{p,q}\\
&=\frac{a_f^2}{4}\operatorname{Tr}W_L. \tag*{(40)}
\end{aligned} Tr ( F ∗ W L F ) = Tr ( W L F F ∗ ) = ( p , q ) ∈ G L ∑ w p , q ⟨ F F ∗ e p , q , e p , q ⟩ = ( p , q ) ∈ G L ∑ w p , q ∥ F ∗ e p , q ∥ 2 ≥ 4 a f 2 ( p , q ) ∈ G L ∑ w p , q = 4 a f 2 Tr W L . (40)
Therefore, by (30) , (35) , and (40) , we conclude
Tr Y L = 1 32 Tr ( F ∗ W L F ) − ∥ K 1 ∥ 2 − 1 16 ∥ K 2 ∥ 2 ≥ a f 2 128 Tr W L − 17 2 M f 2 . (41) \begin{aligned}
\operatorname{Tr}Y_L&=\frac{1}{32}\operatorname{Tr}(F^*W_LF)-\lVert K_1\rVert^2-\frac{1}{16}\lVert K_2\rVert^2\\
&\geq\frac{a_f^2}{128}\operatorname{Tr}W_L-\frac{17}{2}M_f^2. \tag*{(41)}
\end{aligned} Tr Y L = 32 1 Tr ( F ∗ W L F ) − ∥ K 1 ∥ 2 − 16 1 ∥ K 2 ∥ 2 ≥ 128 a f 2 Tr W L − 2 17 M f 2 . (41)
Using (19) and (34) , we obtain
∂ ( A L , B L ) ≥ a f 2 256 log ( 1 + L / 8 ) − 17 2 M f 2 (42) \partial(A_L,B_L)\geq\frac{a_f^2}{256}\log(1+L/8)-\frac{17}{2}M_f^2 \tag*{(42)} ∂ ( A L , B L ) ≥ 256 a f 2 log ( 1 + L /8 ) − 2 17 M f 2 (42)
for every admissible L ≥ L 0 L\geq L_0 L ≥ L 0 .
Finally, enlarge L 0 L_0 L 0 , if necessary, so that L 0 ≥ 8 L_0\geq8 L 0 ≥ 8 , and set
c = a f 2 256 , D f = 17 2 M f 2 , C 0 = D f + c log L 0 . c=\frac{a_f^2}{256},\qquad D_f=\frac{17}{2}M_f^2,\qquad C_0=D_f+c\log L_0. c = 256 a f 2 , D f = 2 17 M f 2 , C 0 = D f + c log L 0 .
For L ≥ L 0 L \ge L_0 L ≥ L 0 , we derive from Equations (41) and (19) that
Tr Y L ≥ c log ( 1 + L / 8 ) − D f ≥ c log L − C 0 . \operatorname{Tr} Y_L \ge c\log(1+L/8)-D_f \ge c\log L-C_0. Tr Y L ≥ c log ( 1 + L /8 ) − D f ≥ c log L − C 0 .
For L < L 0 L<L_0 L < L 0 , by the positivity of F ∗ W L F F^*W_LF F ∗ W L F , together with (30) and (35) , we obtain
Tr Y L ≥ − D f ≥ c log L − C 0 . \operatorname{Tr} Y_L \ge-D_f \ge c\log L-C_0. Tr Y L ≥ − D f ≥ c log L − C 0 .
Thus the asserted trace bound holds for every admissible L L L . Applying (11) yields
d ( A L , B L ) ≥ Tr Y L ≥ c log L − C 0 , d(A_L,B_L)\ge\operatorname{Tr}Y_L\ge c\log L-C_0, d ( A L , B L ) ≥ Tr Y L ≥ c log L − C 0 ,
which completes the proof. □ \square □
The compact sum and the two conjectures
We prove Theorem 1.3 by taking the direct sum of a sequence of carefully chosen finite dimensional models.
Proof of Theorem 1.3 . We choose positive numbers λ j \lambda_j λ j and integers L j ∈ 8 N L_j\in8\mathbb{N} L j ∈ 8 N such that
∑ j λ j 2 < ∞ , λ j 2 log L j ⟶ ∞ . \sum_j\lambda_j^2<\infty,\qquad\lambda_j^2\log L_j\longrightarrow\infty. j ∑ λ j 2 < ∞ , λ j 2 log L j ⟶ ∞.
For j ≥ 1 j\ge1 j ≥ 1 , choose λ j = 2 − j \lambda_j=2^{-j} λ j = 2 − j and L j = 8 ⋅ 2 8 j L_j=8\cdot2^{8^j} L j = 8 ⋅ 2 8 j . On the Hilbert space H = ⨁ j ≥ 1 H L j \mathcal{H}=\bigoplus_{j\ge1}\mathcal{H}_{L_j} H = ⨁ j ≥ 1 H L j , we consider the operator
T = ⨁ j ≥ 1 λ j ( A L j + i B L j ) . (43) T=\bigoplus_{j\ge1}\lambda_j(A_{L_j}+iB_{L_j}). \tag*{(43)} T = j ≥ 1 ⨁ λ j ( A L j + i B L j ) . (43)
The norm of the j j j -th block is at most 2 λ j 2\lambda_j 2 λ j , so the finite partial sums converge in operator norm to T T T . The self-commutator is
[ T ∗ , T ] = ⨁ j ≥ 1 2 i λ j 2 [ A L j , B L j ] . [T^*,T]=\bigoplus_{j\ge1}2i\lambda_j^2[A_{L_j},B_{L_j}]. [ T ∗ , T ] = j ≥ 1 ⨁ 2 i λ j 2 [ A L j , B L j ] .
By Lemma 3.1 it has trace norm 8 ∑ j ≥ 1 4 − j = 8 / 3 8\sum_{j\ge1}4^{-j}=8/3 8 ∑ j ≥ 1 4 − j = 8/3 .
By (7) and Proposition 4.1 , we have
d ( λ j A L j , λ j B L j ) ≥ λ j 2 ( c log L j − C 0 ) . d(\lambda_jA_{L_j},\lambda_jB_{L_j})\ge\lambda_j^2(c\log L_j-C_0). d ( λ j A L j , λ j B L j ) ≥ λ j 2 ( c log L j − C 0 ) .
Since λ j 2 log L j = 4 − j log 8 + 2 j log 2 \lambda_j^2\log L_j=4^{-j}\log8+2^j\log2 λ j 2 log L j = 4 − j log 8 + 2 j log 2 , the lower bound tends to infinity. It follows from Lemma 2.4 and Proposition 2.5 (in particular the estimate (15) holds for every bounded S S S ) that T ⊕ S − M T\oplus S-M T ⊕ S − M cannot be Hilbert–Schmidt for any bounded S S S and bounded normal M M M .
We next compute the Helton–Howe measure P T P_T P T . Write A = Re T A=\operatorname{Re}T A = Re T and B = Im T B=\operatorname{Im}T B = Im T . We evaluate polynomials in A , B A,B A , B by placing all powers of A A A before those of B B B .
The trace class S 1 ( H ) \mathcal{S}_1(\mathcal{H}) S 1 ( H ) is a two-sided ideal in B ( H ) B(\mathcal{H}) B ( H ) . Since [ A , B ] ∈ S 1 ( H ) [A,B]\in\mathcal{S}_1(\mathcal{H}) [ A , B ] ∈ S 1 ( H ) , the images of A A A and B B B commute in the algebraic quotient B ( H ) / S 1 ( H ) B(\mathcal{H})/\mathcal{S}_1(\mathcal{H}) B ( H ) / S 1 ( H ) . Consequently, for every P , Q ∈ C [ x , y ] P,Q\in\mathbb{C}[x,y] P , Q ∈ C [ x , y ] , the corresponding commutator is in trace class, i.e.
C : = [ P ( A , B ) , Q ( A , B ) ] ∈ S 1 ( H ) . C:=[P(A,B),Q(A,B)]\in\mathcal{S}_1(\mathcal{H}). C := [ P ( A , B ) , Q ( A , B )] ∈ S 1 ( H ) .
Each finite-dimensional summand H L j \mathcal{H}_{L_j} H L j reduces both A A A and B B B , and hence
C = ⨁ j ≥ 1 C j , C=\bigoplus_{j\ge1}C_j, C = j ≥ 1 ⨁ C j ,
where
C j = [ P ( λ j A L j , λ j B L j ) , Q ( λ j A L j , λ j B L j ) ] . C_j=[P(\lambda_jA_{L_j},\lambda_jB_{L_j}),Q(\lambda_jA_{L_j},\lambda_jB_{L_j})]. C j = [ P ( λ j A L j , λ j B L j ) , Q ( λ j A L j , λ j B L j )] .
Each C j C_j C j is a commutator of finite matrices, so Tr C j = 0 \operatorname{Tr} C_j=0 Tr C j = 0 . Moreover, note that
∑ j ≥ 1 ∣ Tr C j ∣ ≤ ∑ j ≥ 1 ∥ C j ∥ 1 = ∥ C ∥ 1 < ∞ . \sum_{j\geq1}|\operatorname{Tr} C_j|\leq\sum_{j\geq1}\|C_j\|_1=\|C\|_1<\infty. j ≥ 1 ∑ ∣ Tr C j ∣ ≤ j ≥ 1 ∑ ∥ C j ∥ 1 = ∥ C ∥ 1 < ∞.
Therefore, we have
Tr [ P ( A , B ) , Q ( A , B ) ] = ∑ j ≥ 1 Tr C j = 0. \operatorname{Tr}[P(A,B),Q(A,B)]=\sum_{j\geq1}\operatorname{Tr} C_j=0. Tr [ P ( A , B ) , Q ( A , B )] = j ≥ 1 ∑ Tr C j = 0.
Given any polynomial p ( x , y ) p(x,y) p ( x , y ) , we take P ( x , y ) = x P(x,y)=x P ( x , y ) = x and Q ( x , y ) = ∫ 0 y p ( x , t ) d t Q(x,y)=\int_0^y p(x,t)\,\mathrm{d}t Q ( x , y ) = ∫ 0 y p ( x , t ) d t . Using (1.1), we verify directly that the Jacobian of P , Q P,Q P , Q equals p p p . Thus the integral of every polynomial against P T P_T P T vanishes. Since polynomials are uniformly dense in the continuous functions on the compact support of P T P_T P T , the measure P T P_T P T vanishes. □
Corollary 5.1. The operator in (43) is a counterexample to Voiculescu’s Conjecture 1.2. Together with 0 H 0_{\mathcal H} 0 H , it is a counterexample to Conjecture 1.1.
Proof. The conclusion of Conjecture 1.2 is precisely (1.3) with an almost-normal S S S . Theorem 1.3 excludes the existence of any such bounded operator S S S . For Conjecture 1.1, use T 1 = T T_1=T T 1 = T and T 2 = 0 H T_2=0_{\mathcal H} T 2 = 0 H , which have the same zero Helton–Howe measure. If (1.2) held, then T ⊕ N T\oplus N T ⊕ N would be Hilbert–Schmidt close to the normal operator W ( 0 H ⊕ N ) W ∗ W(0_{\mathcal H}\oplus N)W^* W ( 0 H ⊕ N ) W ∗ , again contradicting Theorem 1.3 . □
Remark 5.2. Theorem 1.3 also answers A1 in [23 p. 317] , the zero-measure case of Guess A.
The rank profile of the joint residual
Proposition 4.1 bounds frame error by constructing self-adjoint tests dominated by every labelled joint residual. Suppose a finite-dimensional pair is a reducing block of X = A + i B X=A+iB X = A + i B , and P P P projects onto that block. The zero extension of a test on the block satisfies
Y ~ ≤ ( A − x I ) P ( A − x I ) + ( B − y I ) P ( B − y I ) ( x , y ∈ R ) . \widetilde{Y}\leq(A-xI)P(A-xI)+(B-yI)P(B-yI)\qquad(x,y\in\mathbb{R}). Y ≤ ( A − x I ) P ( A − x I ) + ( B − y I ) P ( B − y I ) ( x , y ∈ R ) .
We now allow P P P to be any finite-rank orthogonal projection and maximize the trace of such tests subject to a rank bound. This defines quantities that do not depend on a chosen block decomposition.
Let X = A + i B X=A+iB X = A + i B be bounded on a Hilbert space H \mathcal{H} H , with A A A and B B B self-adjoint. For a finite-rank orthogonal projection P P P on H \mathcal{H} H , define
Q X , P ( x , y ) = ( A − x I ) P ( A − x I ) + ( B − y I ) P ( B − y I ) . (44) Q_{X,P}(x,y)=(A-xI)P(A-xI)+(B-yI)P(B-yI). \tag*{(44)} Q X , P ( x , y ) = ( A − x I ) P ( A − x I ) + ( B − y I ) P ( B − y I ) . (44)
Its quadratic form on v ∈ H v\in\mathcal{H} v ∈ H is the sum of the squared residuals ∥ P ( A − x I ) v ∥ 2 + ∥ P ( B − y I ) v ∥ 2 \|P(A-xI)v\|^2+\|P(B-yI)v\|^2 ∥ P ( A − x I ) v ∥ 2 + ∥ P ( B − y I ) v ∥ 2 .
Definition 6.1. A finite-rank self-adjoint operator Y Y Y is admissible for ( X , P ) (X,P) ( X , P ) if
Y ≤ Q X , P ( x , y ) , ∀ x , y ∈ R . (45) Y\leq Q_{X,P}(x,y),\qquad\forall x,y\in\mathbb{R}. \tag*{(45)} Y ≤ Q X , P ( x , y ) , ∀ x , y ∈ R . (45)
For n ≥ 1 n\geq1 n ≥ 1 , define r n ( X ) ≥ 0 r_n(X)\geq0 r n ( X ) ≥ 0 by
r n ( X ) 2 = sup P orthogonal , rank P ≤ n Y admissible for ( X , P ) Tr Y . (46) r_n(X)^2=\sup_{\substack{P\ \text{orthogonal},\ \operatorname{rank}P\leq n\\Y\ \text{admissible for }(X,P)}}\operatorname{Tr}Y. \tag*{(46)} r n ( X ) 2 = P orthogonal , rank P ≤ n Y admissible for ( X , P ) sup Tr Y . (46)
We point out that Y = 0 Y=0 Y = 0 is admissible, so the above supremum is nonnegative.
At x = y = 0 x=y=0 x = y = 0 , (45) implies Tr Y ≤ Tr ( A P A + B P B ) \operatorname{Tr}Y\leq\operatorname{Tr}(APA+BPB) Tr Y ≤ Tr ( A P A + B P B ) . Therefore, Tr Y \operatorname{Tr}Y Tr Y is at most rk ( P ) ( ∥ A ∥ 2 2 + ∥ B ∥ 2 2 ) \operatorname{rk}(P)(\lVert A\rVert_{2}^{2}+\lVert B\rVert_{2}^{2}) rk ( P ) (∥ A ∥ 2 2 + ∥ B ∥ 2 2 ) . The sequence r n ( X ) r_n(X) r n ( X ) is therefore finite at each rank and nondecreasing. For a fixed P P P , the operators Q X , P ( x , y ) Q_{X,P}(x,y) Q X , P ( x , y ) are supported on E = ran P + A ( ran P ) + B ( ran P ) E=\operatorname{ran}P+A(\operatorname{ran}P)+B(\operatorname{ran}P) E = ran P + A ( ran P ) + B ( ran P ) . We observe that if we compress Y Y Y to the subspace E E E by the orthogonal projection P E P_E P E , the compressed operator P E Y P E P_EYP_E P E Y P E is still admissible. Furthermore, Tr ( Y ) \operatorname{Tr}(Y) Tr ( Y ) is less than or equal to Tr ( P E Y P E ) \operatorname{Tr}(P_EYP_E) Tr ( P E Y P E ) since the compression of Y Y Y to E ⊥ E^\perp E ⊥ is nonpositive. Thus the supremum may be restricted to tests supported on a subspace of dimension at most 3 n 3n 3 n . We also observe that on a finite dimension block with P = I P=I P = I , (45) is precisely (10) .
Proposition 6.2. The quantities r n r_n r n satisfy the following properties.
∀ n \forall n ∀ n , r n ( X ) r_n(X) r n ( X ) is invariant under unitary conjugacy.
For bounded operators X , X ′ X,X' X , X ′ on the same Hilbert space with X − X ′ ∈ S 2 X-X'\in\mathcal{S}_2 X − X ′ ∈ S 2 ,
∣ r n ( X ) − r n ( X ′ ) ∣ ≤ ∥ X − X ′ ∥ 2 ( n ≥ 1 ) . (47) \lvert r_n(X)-r_n(X')\rvert\leq\lVert X-X'\rVert_2\qquad(n\geq1). \tag*{(47)} ∣ r n ( X ) − r n ( X ′ )∣ ≤ ∥ X − X ′ ∥ 2 ( n ≥ 1 ) . (47) For every bounded normal N N N , r n ( X ⊕ N ) = r n ( X ) r_n(X\oplus N)=r_n(X) r n ( X ⊕ N ) = r n ( X ) .
Proof. (1) It is straightforward to check that r n ( X ) r_n(X) r n ( X ) is equal to r n ( U ∗ X U ) r_n(U^*XU) r n ( U ∗ X U ) for a unitary operator U U U as unitary conjugacy carries both P P P and Y Y Y to tests with the same rank and trace.
(2) Let X X X and X ′ X' X ′ be bounded operators with X − X ′ ∈ S 2 X-X'\in\mathcal{S}_2 X − X ′ ∈ S 2 . Write X = A + i B X=A+iB X = A + i B and X ′ = A ′ + i B ′ X'=A'+iB' X ′ = A ′ + i B ′ , and set K = X − X ′ K=X-X' K = X − X ′ , Δ A = A − A ′ \Delta_A=A-A' Δ A = A − A ′ , and Δ B = B − B ′ \Delta_B=B-B' Δ B = B − B ′ . For a fixed finite-rank orthogonal projection P P P , denote R P = Δ A P Δ A + Δ B P Δ B R_P=\Delta_AP\Delta_A+\Delta_BP\Delta_B R P = Δ A P Δ A + Δ B P Δ B . Then
Tr R P = ∥ P Δ A ∥ 2 2 + ∥ P Δ B ∥ 2 2 ≤ ∥ Δ A ∥ 2 2 + ∥ Δ B ∥ 2 2 = ∥ K ∥ 2 2 . \operatorname{Tr}R_P=\lVert P\Delta_A\rVert_2^2+\lVert P\Delta_B\rVert_2^2\leq\lVert\Delta_A\rVert_2^2+\lVert\Delta_B\rVert_2^2=\lVert K\rVert_2^2. Tr R P = ∥ P Δ A ∥ 2 2 + ∥ P Δ B ∥ 2 2 ≤ ∥ Δ A ∥ 2 2 + ∥ Δ B ∥ 2 2 = ∥ K ∥ 2 2 .
For any v ∈ H v\in\mathcal{H} v ∈ H , we compute
P ( A − x I ) v = P ( A ′ − x I ) v + P Δ A v . P(A-xI)v=P(A'-xI)v+P\Delta_Av. P ( A − x I ) v = P ( A ′ − x I ) v + P Δ A v .
For t > 0 t>0 t > 0 , applying the squared-norm inequality ∥ u + v ∥ 2 ≤ ( 1 + t ) ∥ u ∥ 2 + ( 1 + t − 1 ) ∥ v ∥ 2 \lVert u+v\rVert^2\leq(1+t)\lVert u\rVert^2+(1+t^{-1})\lVert v\rVert^2 ∥ u + v ∥ 2 ≤ ( 1 + t ) ∥ u ∥ 2 + ( 1 + t − 1 ) ∥ v ∥ 2 to the above ∥ P ( A − x I ) v ∥ 2 \lVert P(A-xI)v\rVert^2 ∥ P ( A − x I ) v ∥ 2 , we get
∥ P ( A − x I ) v ∥ 2 ≤ ( 1 + t ) ∥ P ( A ′ − x I ) v ∥ 2 + ( 1 + t − 1 ) ∥ P Δ A v ∥ 2 . \lVert P(A-xI)v\rVert^2\leq(1+t)\lVert P(A'-xI)v\rVert^2+(1+t^{-1})\lVert P\Delta_Av\rVert^2. ∥ P ( A − x I ) v ∥ 2 ≤ ( 1 + t ) ∥ P ( A ′ − x I ) v ∥ 2 + ( 1 + t − 1 ) ∥ P Δ A v ∥ 2 .
This gives
⟨ ( A − x I ) P ( A − x I ) v , v ⟩ ≤ ( 1 + t ) ⟨ ( A ′ − x I ) P ( A ′ − x I ) v , v ⟩ + ( 1 + t − 1 ) ⟨ Δ A P Δ A v , v ⟩ . \langle(A-xI)P(A-xI)v,v\rangle\leq(1+t)\langle(A'-xI)P(A'-xI)v,v\rangle+(1+t^{-1})\langle\Delta_AP\Delta_Av,v\rangle. ⟨( A − x I ) P ( A − x I ) v , v ⟩ ≤ ( 1 + t ) ⟨( A ′ − x I ) P ( A ′ − x I ) v , v ⟩ + ( 1 + t − 1 ) ⟨ Δ A P Δ A v , v ⟩ .
Similarly, we have
⟨ ( B − y I ) P ( B − y I ) v , v ⟩ ≤ ( 1 + t ) ⟨ ( B ′ − y I ) P ( B ′ − y I ) v , v ⟩ + ( 1 + t − 1 ) ⟨ Δ B P Δ B v , v ⟩ . \langle(B-yI)P(B-yI)v,v\rangle\leq(1+t)\langle(B'-yI)P(B'-yI)v,v\rangle+(1+t^{-1})\langle\Delta_BP\Delta_Bv,v\rangle. ⟨( B − y I ) P ( B − y I ) v , v ⟩ ≤ ( 1 + t ) ⟨( B ′ − y I ) P ( B ′ − y I ) v , v ⟩ + ( 1 + t − 1 ) ⟨ Δ B P Δ B v , v ⟩ .
Summarizing the above estimates, we obtain
Q X , P ( x , y ) ≤ ( 1 + t ) Q X ′ , P ( x , y ) + ( 1 + t − 1 ) R P . (48) Q_{X,P}(x,y)\leq(1+t)Q_{X',P}(x,y)+(1+t^{-1})R_P. \tag*{(48)} Q X , P ( x , y ) ≤ ( 1 + t ) Q X ′ , P ( x , y ) + ( 1 + t − 1 ) R P . (48)
Inequality (48) shows that an admissible Y Y Y for ( X , P ) (X,P) ( X , P ) gives the admissible test ( Y − ( 1 + t − 1 ) R P ) / ( 1 + t ) (Y-(1+t^{-1})R_P)/(1+t) ( Y − ( 1 + t − 1 ) R P ) / ( 1 + t ) for ( X ′ , P ) (X',P) ( X ′ , P ) . Taking traces and then the supremum, we obtain
r n ( X ) 2 ≤ ( 1 + t ) r n ( X ′ ) 2 + ( 1 + t − 1 ) ∥ K ∥ 2 2 . r_n(X)^2\leq(1+t)r_n(X')^2+(1+t^{-1})\lVert K\rVert_2^2. r n ( X ) 2 ≤ ( 1 + t ) r n ( X ′ ) 2 + ( 1 + t − 1 ) ∥ K ∥ 2 2 .
The infimum over t > 0 t>0 t > 0 on the right is ( r n ( X ′ ) + ∥ K ∥ 2 ) 2 (r_n(X')+\lVert K\rVert_2)^2 ( r n ( X ′ ) + ∥ K ∥ 2 ) 2 . So we have
r n ( X ) − r n ( X ′ ) ≤ ∥ K ∥ 2 . r_n(X)-r_n(X')\leq\lVert K\rVert_2. r n ( X ) − r n ( X ′ ) ≤ ∥ K ∥ 2 .
Exchanging X X X and X ′ X' X ′ , we obtain
r n ( X ′ ) − r n ( X ) ≤ ∥ K ∥ 2 , r_n(X')-r_n(X)\leq\lVert K\rVert_2, r n ( X ′ ) − r n ( X ) ≤ ∥ K ∥ 2 ,
and conclude (47) .
:::
(3) For the stabilization assertion, first take N = C + i D N=C+iD N = C + i D normal. Partition the joint spectrum of C , D C,D C , D into finitely many pairwise disjoint Borel sets, called cells, of diameter at most ε \varepsilon ε . Let E k E_k E k be their joint spectral projections and ( x k , y k ) (x_k,y_k) ( x k , y k ) the corresponding labels, chosen in the cells. As Y Y Y is admissible, Y Y Y satisfies
Y ≤ ( C − x I ) P ( C − x I ) + ( D − y I ) P ( D − y I ) , ∀ x , y ∈ R . Y \le(C-xI)P(C-xI)+(D-yI)P(D-yI),\quad\forall x,y\in\mathbb{R}. Y ≤ ( C − x I ) P ( C − x I ) + ( D − y I ) P ( D − y I ) , ∀ x , y ∈ R .
Taking x = x k , y = y k x=x_k,y=y_k x = x k , y = y k and multiplying E k E_k E k on both sides, we obtain
E k Y E k ≤ E k ( C − x k I ) P ( C − x k I ) E k + E k ( D − y k I ) P ( D − y k I ) E k . E_kYE_k\le E_k(C-x_kI)P(C-x_kI)E_k+E_k(D-y_kI)P(D-y_kI)E_k. E k Y E k ≤ E k ( C − x k I ) P ( C − x k I ) E k + E k ( D − y k I ) P ( D − y k I ) E k .
Hence,
Tr ( E k Y E k ) ≤ ∥ P ( C − x k I ) E k ∥ 2 2 + ∥ P ( D − y k I ) E k ∥ 2 2 . \operatorname{Tr}(E_kYE_k)\le\lVert P(C-x_kI)E_k\rVert_2^2+\lVert P(D-y_kI)E_k\rVert_2^2. Tr ( E k Y E k ) ≤ ∥ P ( C − x k I ) E k ∥ 2 2 + ∥ P ( D − y k I ) E k ∥ 2 2 .
We choose the cell defining E k E_k E k to be sufficiently small so that on E k E_k E k , we have
∥ ( C − x k I ) E k ∥ ≤ ε , ∥ ( D − y k I ) E k ∥ ≤ ε . \lVert(C-x_kI)E_k\rVert\le\varepsilon,\qquad\lVert(D-y_kI)E_k\rVert\le\varepsilon. ∥( C − x k I ) E k ∥ ≤ ε , ∥( D − y k I ) E k ∥ ≤ ε .
Then we have
Tr ( E k Y E k ) ≤ ∥ P ( C − x k I ) E k ∥ 2 2 + ∥ P ( D − y k I ) E k ∥ 2 2 ≤ ∥ P E k ∥ 2 2 ∥ ( C − x k I ) E k ∥ 2 + ∥ P E k ∥ 2 2 ∥ ( D − y k I ) E k ∥ 2 ≤ 2 ε 2 ∥ P E k ∥ 2 2 . \begin{aligned}
\operatorname{Tr}(E_kYE_k)&\le\lVert P(C-x_kI)E_k\rVert_2^2+\lVert P(D-y_kI)E_k\rVert_2^2\\
&\le\lVert PE_k\rVert_2^2\lVert(C-x_kI)E_k\rVert^2+\lVert PE_k\rVert_2^2\lVert(D-y_kI)E_k\rVert^2\\
&\le2\varepsilon^2\lVert PE_k\rVert_2^2.
\end{aligned} Tr ( E k Y E k ) ≤ ∥ P ( C − x k I ) E k ∥ 2 2 + ∥ P ( D − y k I ) E k ∥ 2 2 ≤ ∥ P E k ∥ 2 2 ∥( C − x k I ) E k ∥ 2 + ∥ P E k ∥ 2 2 ∥( D − y k I ) E k ∥ 2 ≤ 2 ε 2 ∥ P E k ∥ 2 2 .
Summing over k k k , we have
Tr ( Y ) ≤ ∑ k 2 ε 2 ∥ P E k ∥ 2 2 = 2 ε 2 ∑ k Tr ( E k P E k ) = 2 ε 2 Tr P . \operatorname{Tr}(Y)\le\sum_k2\varepsilon^2\lVert PE_k\rVert_2^2=2\varepsilon^2\sum_k\operatorname{Tr}(E_kPE_k)=2\varepsilon^2\operatorname{Tr}P. Tr ( Y ) ≤ k ∑ 2 ε 2 ∥ P E k ∥ 2 2 = 2 ε 2 k ∑ Tr ( E k P E k ) = 2 ε 2 Tr P .
And taking ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , we have shown that
Tr Y ≤ 0. (49) \operatorname{Tr}Y\le0. \tag*{(49)} Tr Y ≤ 0. (49)
Let X = A + i B X=A+iB X = A + i B be a bounded operator on H \mathcal{H} H , and N = C + i D N=C+iD N = C + i D be normal on K \mathcal{K} K . Fix n ≥ 1 n\ge1 n ≥ 1 , and suppose that P P P is an orthogonal projection on H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K with rank P ≤ n P\le n P ≤ n . Let Y Y Y be a finite-rank self-adjoint operator admissible for ( X ⊕ N , P ) (X\oplus N,P) ( X ⊕ N , P ) .
Denote the canonical inclusions of the two summands by
ι 1 : H ⟶ H ⊕ K , ι 2 : K ⟶ H ⊕ K . \iota_1:\mathcal{H}\longrightarrow\mathcal{H}\oplus\mathcal{K},\qquad\iota_2:\mathcal{K}\longrightarrow\mathcal{H}\oplus\mathcal{K}. ι 1 : H ⟶ H ⊕ K , ι 2 : K ⟶ H ⊕ K .
Define the compressions
P i = ι i ∗ P ι i , Y i = ι i ∗ Y ι i , i = 1 , 2. P_i=\iota_i^*P\iota_i,\qquad Y_i=\iota_i^*Y\iota_i,\qquad i=1,2. P i = ι i ∗ P ι i , Y i = ι i ∗ Y ι i , i = 1 , 2.
Each Y i Y_i Y i is finite rank and self-adjoint. Each P i P_i P i is a positive contraction, since
⟨ P i h , h ⟩ = ∥ P ι i h ∥ 2 ≤ ∥ h ∥ 2 . \langle P_ih,h\rangle=\lVert P\iota_ih\rVert^2\le\lVert h\rVert^2. ⟨ P i h , h ⟩ = ∥ P ι i h ∥ 2 ≤ ∥ h ∥ 2 .
Although P i P_i P i need not be an orthogonal projection, we can replace it by a larger orthogonal projection without exceeding the rank bound.
Let L i = ι i ∗ ( ran P ) L_i=\iota_i^*(\operatorname{ran}P) L i = ι i ∗ ( ran P ) , and Π i \Pi_i Π i be the orthogonal projection onto L i L_i L i . This subspace is finite dimensional, with
rank Π i = dim L i ≤ rank P ≤ n . \operatorname{rank}\Pi_i=\dim L_i\le\operatorname{rank}P\le n. rank Π i = dim L i ≤ rank P ≤ n .
If h ⊥ L i h\perp L_i h ⊥ L i , then ι i h ⊥ ran P \iota_ih\perp\operatorname{ran}P ι i h ⊥ ran P , so P ι i h = 0 P\iota_ih=0 P ι i h = 0 and hence P i h = 0 P_ih=0 P i h = 0 . Thus P i = Π i P i Π i P_i=\Pi_iP_i\Pi_i P i = Π i P i Π i . Since P i P_i P i is a positive contraction,
⟨ P i h , h ⟩ = ⟨ P i Π i h , Π i h ⟩ ≤ ∥ Π i h ∥ 2 = ⟨ Π i h , h ⟩ . \langle P_ih,h\rangle=\langle P_i\Pi_ih,\Pi_ih\rangle\le\lVert\Pi_ih\rVert^2=\langle\Pi_ih,h\rangle. ⟨ P i h , h ⟩ = ⟨ P i Π i h , Π i h ⟩ ≤ ∥ Π i h ∥ 2 = ⟨ Π i h , h ⟩ .
Consequently, we have
0 ≤ P i ≤ Π i . 0\le P_i\le\Pi_i. 0 ≤ P i ≤ Π i .
The real and imaginary parts of X ⊕ N X \oplus N X ⊕ N are A ⊕ C A \oplus C A ⊕ C and B ⊕ D B \oplus D B ⊕ D . Compressing (45) to the first summand therefore gives, for every x , y ∈ R x,y \in\mathbb{R} x , y ∈ R ,
Y 1 ≤ ( A − x I ) P 1 ( A − x I ) + ( B − y I ) P 1 ( B − y I ) ≤ ( A − x I ) Π 1 ( A − x I ) + ( B − y I ) Π 1 ( B − y I ) = Q X , Π 1 ( x , y ) . Y_1 \le(A-xI)P_1(A-xI)+(B-yI)P_1(B-yI)
\le(A-xI)\Pi_1(A-xI)+(B-yI)\Pi_1(B-yI)
=Q_{X,\Pi_1}(x,y). Y 1 ≤ ( A − x I ) P 1 ( A − x I ) + ( B − y I ) P 1 ( B − y I ) ≤ ( A − x I ) Π 1 ( A − x I ) + ( B − y I ) Π 1 ( B − y I ) = Q X , Π 1 ( x , y ) .
The second inequality follows from P 1 ≤ Π 1 P_1 \le\Pi_1 P 1 ≤ Π 1 : the difference between the corresponding terms is, for example,
( A − x I ) ( Π 1 − P 1 ) ( A − x I ) ≥ 0. (A-xI)(\Pi_1-P_1)(A-xI)\ge0. ( A − x I ) ( Π 1 − P 1 ) ( A − x I ) ≥ 0.
Thus Y 1 Y_1 Y 1 is admissible for ( X , Π 1 ) (X,\Pi_1) ( X , Π 1 ) . The same argument in the second summand gives
Y 2 ≤ ( C − x I ) P 2 ( C − x I ) + ( D − y I ) P 2 ( D − y I ) ≤ ( C − x I ) Π 2 ( C − x I ) + ( D − y I ) Π 2 ( D − y I ) = Q N , Π 2 ( x , y ) . Y_2 \le(C-xI)P_2(C-xI)+(D-yI)P_2(D-yI)
\le(C-xI)\Pi_2(C-xI)+(D-yI)\Pi_2(D-yI)
=Q_{N,\Pi_2}(x,y). Y 2 ≤ ( C − x I ) P 2 ( C − x I ) + ( D − y I ) P 2 ( D − y I ) ≤ ( C − x I ) Π 2 ( C − x I ) + ( D − y I ) Π 2 ( D − y I ) = Q N , Π 2 ( x , y ) .
Hence Y 2 Y_2 Y 2 is admissible for ( N , Π 2 ) (N,\Pi_2) ( N , Π 2 ) .
By (46) and (49) , we have that,
Tr Y 1 ≤ r n ( X ) 2 , Tr Y 2 ≤ 0. \operatorname{Tr}Y_1\le r_n(X)^2,\qquad\operatorname{Tr}Y_2\le0. Tr Y 1 ≤ r n ( X ) 2 , Tr Y 2 ≤ 0.
The union of orthonormal bases of H \mathcal{H} H and K \mathcal{K} K is an orthonormal basis of their direct sum. Computing the trace of the finite-rank operator Y Y Y in this basis gives
Tr Y = Tr Y 1 + Tr Y 2 ≤ r n ( X ) 2 . \operatorname{Tr}Y=\operatorname{Tr}Y_1+\operatorname{Tr}Y_2\le r_n(X)^2. Tr Y = Tr Y 1 + Tr Y 2 ≤ r n ( X ) 2 .
Taking the supremum over all admissible Y Y Y and all orthogonal projections P P P of rank at most n n n yields
r n ( X ⊕ N ) 2 ≤ r n ( X ) 2 . (50) r_n(X\oplus N)^2\le r_n(X)^2. \tag*{(50)} r n ( X ⊕ N ) 2 ≤ r n ( X ) 2 . (50)
For the reverse, let P 0 P_0 P 0 be any orthogonal projection on H \mathcal{H} H with rank P 0 ≤ n P_0\le n P 0 ≤ n , and let Y 0 Y_0 Y 0 be admissible for ( X , P 0 ) (X,P_0) ( X , P 0 ) . Extend both operators by zero to H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K :
P ^ = P 0 ⊕ 0 , Y ^ = Y 0 ⊕ 0. \widehat{P}=P_0\oplus0,\qquad\widehat{Y}=Y_0\oplus0. P = P 0 ⊕ 0 , Y = Y 0 ⊕ 0.
Then rank P ^ = rank P 0 \operatorname{rank}\widehat{P}=\operatorname{rank}P_0 rank P = rank P 0 , and
Q X ⊕ N , P ^ ( x , y ) = Q X , P 0 ( x , y ) ⊕ 0. Q_{X\oplus N,\widehat{P}}(x,y)=Q_{X,P_0}(x,y)\oplus0. Q X ⊕ N , P ( x , y ) = Q X , P 0 ( x , y ) ⊕ 0.
Therefore Y ^ \widehat{Y} Y is admissible for ( X ⊕ N , P ^ ) (X\oplus N,\widehat{P}) ( X ⊕ N , P ) and has trace Tr Y 0 \operatorname{Tr}Y_0 Tr Y 0 . Taking the supremum over ( P 0 , Y 0 ) (P_0,Y_0) ( P 0 , Y 0 ) proves
r n ( X ⊕ N ) 2 ≥ r n ( X ) 2 . (51) r_n(X\oplus N)^2\ge r_n(X)^2. \tag*{(51)} r n ( X ⊕ N ) 2 ≥ r n ( X ) 2 . (51)
Since these quantities are nonnegative, combining (50) and (51) , we conclude that
r n ( X ⊕ N ) = r n ( X ) . r_n(X\oplus N)=r_n(X). r n ( X ⊕ N ) = r n ( X ) .
Corollary 6.3. If there exist a bounded operator S S S and a normal operator M M M such that X ⊕ S − M ∈ S 2 X\oplus S-M\in\mathcal{S}_2 X ⊕ S − M ∈ S 2 , r n ( X ) ≤ ∥ X ⊕ S − M ∥ 2 r_n(X)\le\lVert X\oplus S-M\rVert_2 r n ( X ) ≤ ∥ X ⊕ S − M ∥ 2 for every n ≥ 1 n\ge1 n ≥ 1 .
Proof. By extending any admissible Y Y Y for ( X , P ) (X,P) ( X , P ) to H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K by zero on K \mathcal{K} K , we obtain r n ( X ) ≤ r n ( X ⊕ S ) r_n(X)\le r_n(X\oplus S) r n ( X ) ≤ r n ( X ⊕ S ) for any bounded S S S . Since M M M is normal, Proposition 6.2, Part (2) and (3) gives
r n ( X ) ≤ r n ( X ⊕ S ) = r n ( X ⊕ S ) − r n ( M ) ≤ ∥ X ⊕ S − M ∥ 2 . r_n(X)\le r_n(X\oplus S)=r_n(X\oplus S)-r_n(M)\le\lVert X\oplus S-M\rVert_2. r n ( X ) ≤ r n ( X ⊕ S ) = r n ( X ⊕ S ) − r n ( M ) ≤ ∥ X ⊕ S − M ∥ 2 .
Definition 6.4. R ( X ) : = [ ( r n ( X ) ) n ≥ 1 ] \mathcal{R}(X):=[(r_n(X))_{n\geq1}] R ( X ) := [( r n ( X ) ) n ≥ 1 ] in R N / ℓ ∞ \mathbb{R}^{\mathbb{N}}/\ell^\infty R N / ℓ ∞ is called the rank profile of X X X . We use the algebraic quotient by the bounded real sequences for R N / ℓ ∞ \mathbb{R}^{\mathbb{N}}/\ell^\infty R N / ℓ ∞ .
The following proposition directly follows from Proposition 6.2 .
Proposition 6.5. The rank profile R ( X ) \mathcal{R}(X) R ( X ) satisfies the following properties.
(1) R ( X ) \mathcal{R}(X) R ( X ) is invariant under unitary conjugacy.
(2) If X − X ′ X-X' X − X ′ is a Hilbert-Schmidt operator, R ( X ) = R ( X ′ ) \mathcal{R}(X)=\mathcal{R}(X') R ( X ) = R ( X ′ ) .
(3) For every bounded normal N N N , R ( X ⊕ N ) = R ( X ) \mathcal{R}(X\oplus N)=\mathcal{R}(X) R ( X ⊕ N ) = R ( X ) .
(4) Let X , X ′ X,X' X , X ′ be bounded operators on H , H ′ \mathcal{H},\mathcal{H}' H , H ′ , and N , N ′ N,N' N , N ′ be normal operators on K , K ′ \mathcal{K},\mathcal{K}' K , K ′ , and W W W a unitary operator from H ⊕ K \mathcal{H}\oplus\mathcal{K} H ⊕ K to H ′ ⊕ K ′ \mathcal{H}'\oplus\mathcal{K}' H ′ ⊕ K ′ . If
X ⊕ N − W ∗ ( X ′ ⊕ N ′ ) W ∈ S 2 , X\oplus N-W^*(X'\oplus N')W\in\mathcal{S}_2, X ⊕ N − W ∗ ( X ′ ⊕ N ′ ) W ∈ S 2 , then R ( X ) = R ( X ′ ) \mathcal{R}(X)=\mathcal{R}(X') R ( X ) = R ( X ′ ) .
(5) If there exist a bounded operator S S S and a normal operator M M M such that X ⊕ S − M ∈ S 2 X\oplus S-M\in\mathcal{S}_2 X ⊕ S − M ∈ S 2 , R ( X ) = 0 \mathcal{R}(X)=0 R ( X ) = 0 .
Proof. Part (1), (2), (3) are corollaries of the corresponding part in Proposition 6.2 . Part (4) follows from
R ( X ) = ( 3 ) R ( X ⊕ N ) = ( 2 ) R ( W ∗ ( X ′ ⊕ N ′ ) W ) = ( 1 ) R ( X ′ ⊕ N ′ ) = ( 3 ) R ( X ′ ) . \mathcal{R}(X)\mathrel{\overset{(3)}{=}}\mathcal{R}(X\oplus N)\mathrel{\overset{(2)}{=}}\mathcal{R}(W^*(X'\oplus N')W)\mathrel{\overset{(1)}{=}}\mathcal{R}(X'\oplus N')\mathrel{\overset{(3)}{=}}\mathcal{R}(X'). R ( X ) = ( 3 ) R ( X ⊕ N ) = ( 2 ) R ( W ∗ ( X ′ ⊕ N ′ ) W ) = ( 1 ) R ( X ′ ⊕ N ′ ) = ( 3 ) R ( X ′ ) .
Part (5) follows from Corollary 6.3 . □
Look at example T T T on H \mathcal{H} H constructed in the proof of Theorem 1.3 . For L j = 8 ⋅ 2 8 j L_j=8\cdot2^{8j} L j = 8 ⋅ 2 8 j , on the Hilbert space H = ⨁ j ≥ 1 H L j \mathcal{H}=\bigoplus_{j\geq1}\mathcal{H}_{L_j} H = ⨁ j ≥ 1 H L j , consider
T = ⨁ j ≥ 1 2 − j ( A L j + i B L j ) . T=\bigoplus_{j\geq1}2^{-j}(A_{L_j}+iB_{L_j}). T = j ≥ 1 ⨁ 2 − j ( A L j + i B L j ) .
Let P j P_j P j be the orthogonal projection onto the j j j -th component H L j \mathcal{H}_{L_j} H L j with dimension d j = L j 2 = 64 ⋅ 2 2 ⋅ 8 j d_j=L_j^2=64\cdot2^{2\cdot8j} d j = L j 2 = 64 ⋅ 2 2 ⋅ 8 j . We construct in Equation (30) a self-adjoint operator Y L j Y_{L_j} Y L j satisfying Equation (33) , i.e.
Y L j ≤ ( A L j − x I ) 2 + ( B L j − y I ) 2 , x , y ∈ R . Y_{L_j}\leq(A_{L_j}-xI)^2+(B_{L_j}-yI)^2,\qquad x,y\in\mathbb{R}. Y L j ≤ ( A L j − x I ) 2 + ( B L j − y I ) 2 , x , y ∈ R .
Let ι j : H L j ⟶ H \iota_j:\mathcal{H}_{L_j}\longrightarrow\mathcal{H} ι j : H L j ⟶ H be the canonical inclusion, so that P j = ι j ι j ∗ P_j=\iota_j\iota_j^* P j = ι j ι j ∗ , and let I j I_j I j denote the identity on H L j \mathcal{H}_{L_j} H L j . Extend the scaled block operator by zero to H \mathcal{H} H , defining
Y ~ j = 4 − j ι j Y L j ι j ∗ . \widetilde{Y}_j=4^{-j}\iota_jY_{L_j}\iota_j^*. Y j = 4 − j ι j Y L j ι j ∗ .
This is a finite-rank self-adjoint operator. Since H L j \mathcal{H}_{L_j} H L j reduces both the real and imaginary parts of T T T , for every x , y ∈ R x,y\in\mathbb{R} x , y ∈ R we have
Q T , P j ( x , y ) = ι j ( ( 2 − j A L j − x I j ) 2 + ( 2 − j B L j − y I j ) 2 ) ι j ∗ = 4 − j ι j ( ( A L j − 2 j x I j ) 2 + ( B L j − 2 j y I j ) 2 ) ι j ∗ \begin{aligned}
Q_{T,P_j}(x,y)=\iota_j\left((2^{-j}A_{L_j}-xI_j)^2+(2^{-j}B_{L_j}-yI_j)^2\right)\iota_j^* \\
&=4^{-j}\iota_j\left((A_{L_j}-2^jxI_j)^2+(B_{L_j}-2^jyI_j)^2\right)\iota_j^*
\end{aligned} Q T , P j ( x , y ) = ι j ( ( 2 − j A L j − x I j ) 2 + ( 2 − j B L j − y I j ) 2 ) ι j ∗ = 4 − j ι j ( ( A L j − 2 j x I j ) 2 + ( B L j − 2 j y I j ) 2 ) ι j ∗
≥ 4 − j ι j Y L j ι j ∗ = Y ~ j . \geq4^{-j}\iota_jY_{L_j}\iota_j^*=\widetilde{Y}_j. ≥ 4 − j ι j Y L j ι j ∗ = Y j .
Here the inequality follows from (4.13), applied with labels 2 j x 2^jx 2 j x and 2 j y 2^jy 2 j y . Thus Y ~ j \widetilde{Y}_j Y j is admissible for ( T , P j ) (T,P_j) ( T , P j ) . Furthermore, we have
rank P j = d j , Tr H Y ~ j = 4 − j Tr H L j Y L j , \operatorname{rank}P_j=d_j,\qquad\operatorname{Tr}_{\mathcal{H}}\widetilde{Y}_j=4^{-j}\operatorname{Tr}_{\mathcal{H}_{L_j}}Y_{L_j}, rank P j = d j , Tr H Y j = 4 − j Tr H L j Y L j ,
and therefore
r d j ( T ) 2 ≥ 4 − j Tr Y L j . r_{d_j}(T)^2\geq4^{-j}\operatorname{Tr}Y_{L_j}. r d j ( T ) 2 ≥ 4 − j Tr Y L j .
Recall we proved in Proposition 4.1 that
Tr ( Y L j ) ≥ c log L j − C 0 , \operatorname{Tr}(Y_{L_j}) \ge c\log L_j-C_0, Tr ( Y L j ) ≥ c log L j − C 0 ,
with the constants c c c , C 0 C_0 C 0 fixed there. Therefore,
r d j ( T ) 2 ≥ 4 − j ( c log L j − C 0 ) = 4 − j ( c log 8 − C 0 ) + c ( log 2 ) 2 j . r_{d_j}(T)^2 \ge4^{-j}(c\log L_j-C_0)=4^{-j}(c\log8-C_0)+c(\log2)2^j. r d j ( T ) 2 ≥ 4 − j ( c log L j − C 0 ) = 4 − j ( c log 8 − C 0 ) + c ( log 2 ) 2 j .
We can find some a > 0 a>0 a > 0 for all sufficiently large j j j such that
r d j ( T ) ≥ a 2 j / 2 . r_{d_j}(T) \ge a2^{j/2}. r d j ( T ) ≥ a 2 j /2 .
Furthermore, for all sufficiently large n n n , we find j j j with d j ≤ n < d j + 1 d_j\le n<d_{j+1} d j ≤ n < d j + 1 . We have
r n ( T ) ≥ r d j ( T ) ≥ a 2 j / 2 , log n ≤ log d j + 1 ≤ C ~ 8 j , r_n(T)\ge r_{d_j}(T)\ge a2^{j/2},\qquad\log n\le\log d_{j+1}\le\widetilde C8^j, r n ( T ) ≥ r d j ( T ) ≥ a 2 j /2 , log n ≤ log d j + 1 ≤ C 8 j ,
for some positive C ~ \widetilde C C . Combining the above two estimates, we conclude
r n ( T ) ≥ a C ~ − 1 / 6 ( log n ) 1 / 6 . r_n(T)\ge a\widetilde C^{-1/6}(\log n)^{1/6}. r n ( T ) ≥ a C − 1/6 ( log n ) 1/6 .
Hence r n ( T ) → ∞ r_n(T)\to\infty r n ( T ) → ∞ as n → ∞ n\to\infty n → ∞ . We conclude that
R ( T ) ≠ 0. (52) \mathcal{R}(T)\ne0. \tag*{(52)} R ( T ) = 0. (52)
By Proposition 6.5 (5), this recovers the normal-dilation obstruction in Theorem 1.3 .
AI Disclosure: The author used ChatGPT 6.1 Sol-Ultra (OpenAI) to assist with the discovery and development of the main constructions, the exploration and checking of arguments, and the preparation and revision of the exposition. The author reviewed and rewrote the resulting material and takes full responsibility for the mathematical claims, references, and final manuscript.
References
Department of Mathematics, Washington University, St. Louis, MO 63130
Email address: xtang@math.wustl.edu