betaFeedback

WorksMathematics

Optimal lengths of commutators of idempotents and similarity of nilpotent operators to their negatives

Abstract

Let \HH be a complex separable infinite-dimensional Hilbert space. We prove that every operator in \BH is a sum of three commutators of idempotents and a product of two such commutators. Both numbers are best possible. The same is true for differences of idempotents, for operators similar to their negatives, and for operators anticommuting with an involution. Each of these four classes therefore has additive length three and multiplicative length two. We also construct nilpotent operators of order three that are similar to their negatives but cannot be conjugated to their negatives by any bounded involution. These results answer Question~7.22(a)–(c), the question in Remark~7.17, and the question about involutions in Section~2.9 of Marcoux, Radjavi and Zhang, J. Funct. Anal. 284 (2023), 109854.

Introduction and main results

Throughout this paper, H\mathcal{H} denotes a complex separable infinite-dimensional Hilbert space, and B(H)\mathcal{B}(\mathcal{H}) denotes the algebra of bounded linear operators on H\mathcal{H}. We write II for the identity operator and [A,B]:=AB−BA[A,B]:=AB-BA. Following [10], let

E={E∈B(H):E2=E},CE={[E,F]:E,F∈E},DE={E−F:E,F∈E}.\mathcal{E}=\{E\in\mathcal{B}(\mathcal{H}):E^2=E\},\qquad\mathcal{C}_{\mathcal{E}}=\{[E,F]:E,F\in\mathcal{E}\},\qquad\mathcal{D}_{\mathcal{E}}=\{E-F:E,F\in\mathcal{E}\}.

For Δ⊂B(H)\Delta\subset\mathcal{B}(\mathcal{H}) and a positive integer nn, let ∑nΔ\sum_n\Delta and ∏nΔ\prod_n\Delta denote the sets of sums and products of nn elements of Δ\Delta. Define

ℓ+(Δ)=min⁡{n≥1:∑nΔ=B(H)},ℓ×(Δ)=min⁡{n≥1:∏nΔ=B(H)}.\ell_{+}(\Delta)=\min\{n\geq1:\sum_n\Delta=\mathcal{B}(\mathcal{H})\},\qquad\ell_{\times}(\Delta)=\min\{n\geq1:\prod_n\Delta=\mathcal{B}(\mathcal{H})\}.

The minima are over positive integers. A length is ∞\infty if no such integer exists. The factors in a product need not be invertible. Each class in Theorem 1.1 below contains 00, so shorter sums can be extended by zero summands. An involution is an operator JJ with J2=IJ^2=I; it need not be self-adjoint. We write A∼BA\sim B for similarity and A≃BA\simeq B for unitary equivalence. The four classes of operators in Theorem 1.1 are invariant under unitary conjugation. We use fixed unitary identifications to regard operators on finite or countable orthogonal sums of copies of H\mathcal{H} as operators on H\mathcal{H}. Write

H(∞)=⨁n=1∞H.\mathcal{H}^{(\infty)}=\bigoplus_{n=1}^{\infty}\mathcal{H}.

For A∈B(H)A\in\mathcal{B}(\mathcal{H}), write

A(∞)=⨁n=1∞A.A^{(\infty)}=\bigoplus_{n=1}^{\infty}A.

For unrestricted bounded factors, Brown and Pearcy [3] proved that an operator in B(H)\mathcal{B}(\mathcal{H}) is a commutator if and only if it is not of the form λI+K\lambda I+K, where λ≠0\lambda\ne0 and KK is compact. See also Halmos [8 pp. 128–134] for a discussion of operator commutators. Wu [16] surveys factorizations into prescribed classes, including involutions and partial isometries. Here the two entries of each commutator are required to be idempotents, and we determine the least numbers of summands and factors needed to obtain every operator in B(H)\mathcal{B}(\mathcal{H}).

Marcoux, Radjavi and Zhang proved

B(H)=∑5CE=∏3CE=∑4DE,I∉∑2CE.\mathcal{B}(\mathcal{H})=\sum_{5}\mathcal{C}_{\mathcal{E}}=\prod_{3}\mathcal{C}_{\mathcal{E}}=\sum_{4}\mathcal{D}_{\mathcal{E}},\qquad I\notin\sum_{2}\mathcal{C}_{\mathcal{E}}.

see [10 Propositions 7.7 and 7.18, Theorem 7.16]. They also proved that ∏2E\prod_{2}\mathcal{E} is norm dense in B(H)\mathcal{B}(\mathcal{H}) and asked whether it equals B(H)\mathcal{B}(\mathcal{H}) [10 Remark 7.17]. Question 7.22 asks for the least number of differences of idempotents needed in a sum, whether every thin operator is a product of two such differences, and the least numbers of summands and factors for operators similar to their negatives. Here a thin operator is a scalar multiple of II plus a compact operator.

Theorem 1.1. Let Δ⊆B(H)\Delta\subseteq\mathcal{B}(\mathcal{H}) be one of the following four classes: commutators of idempotents, differences of idempotents, operators similar to their negatives, or operators anticommuting with an involution. Then,

ℓ+(Δ)=3,ℓ×(Δ)=2.\ell_{+}(\Delta)=3,\qquad\ell_{\times}(\Delta)=2.

This gives n=3n=3 in Question 7.22(a), an affirmative answer to Question 7.22(b), and m=2m=2, n=3n=3 in Question 7.22(c). The equality B(H)=∏2E\mathcal{B}(\mathcal{H})=\prod_{2}\mathcal{E} answers the question in Remark 7.17.

Note that the equality ℓ+(DE)=3\ell_{+}(\mathcal{D}_{\mathcal{E}})=3 also follows from earlier results. Pearcy and Topping proved that every operator in B(H)\mathcal{B}(\mathcal{H}) is a sum of five idempotents [12]; see also the proof of [13 Theorem 2]. Apply their theorem to T+3IT+3I, add a zero idempotent, and use E+F−I=E−(I−F)E+F-I=E-(I-F) to pair the six terms. The lower bound follows from [13 Theorem 1], which excludes −I-I from sums of four idempotents.

For every T∈B(H)T\in\mathcal{B}(\mathcal{H}) and every real c>∥T∥c>\lVert T\rVert, Theorem 3.3 gives A,B∈EEA,B\in\mathcal{E}_{\mathcal{E}} satisfying

T=AB,AA∗=B∗B=cI,∥A∥=∥B∥=c.T=AB,\qquad AA^{*}=B^{*}B=cI,\qquad\lVert A\rVert=\lVert B\rVert=\sqrt{c}.

Fong and Sourour [7] proved that, if a unitary is not a compact perturbation of a scalar operator, then every unitary is a finite product of operators unitarily equivalent to it. For the bilateral shift of multiplicity one, three factors suffice by [9 Theorem 5.3], and two suffice by [1 Theorem 1.1]. Proposition 3.4 uses the latter theorem to write every unitary TT as T=ABT=AB, where AA and BB are unitary commutators of idempotents. Both commutators are obtained by unitary conjugation of a fixed pair of idempotents. For this factorization, the identities above hold with c=1=∥T∥c=1=\lVert T\rVert. In Proposition 3.5, the right factor is independent of TT; both factors anticommute with involutions and are differences of idempotents.

Drnovšek, Radjavi and Rosenthal characterized commutators of idempotents in [6 Theorem 1]. Drnovšek gave an equivalent condition in terms of involutions in [5 Theorem 2.1]. We use this theorem to show that diag⁡(a,a,−a,−a)\operatorname{diag}(a,a,-a,-a) is a commutator of two idempotents over every unital associative algebra whose scalar field has characteristic different from two. Countable direct sums and [4 Lemma 2.6(ii)] give the additive result. For the multiplicative result, we write each strict contraction as V∗WV^{*}W, where V,WV,W are pure isometries with infinite-dimensional ker V∗V^{*} and ker W∗W^{*}. Proposition 2.1 then shows that the multiples of V∗V^{*} and WW used in the factorization belong to EE\mathcal{E}_{\mathcal{E}}.

An operator T∈B(H)T\in\mathcal{B}(\mathcal{H}) anticommutes with an involution if and only if it is a scalar multiple of a commutator of idempotents. Indeed, if JT=−TJJT=-TJ and J2=IJ^{2}=I, then [5 Corollary 2.4], applied to a sufficiently small positive multiple of TT relative to the eigenspace decomposition of JJ, gives idempotents P,QP,Q and δ>0\delta>0 such that [P,Q]=δT[P,Q]=\delta T. Conversely, 2P−I2P-I is an involution that anticommutes with every scalar multiple of [P,Q][P,Q]. This equivalence is also recorded in [2 Proposition 2.2].

In [10 Section 2.9], the authors also ask whether a nilpotent operator similar to its negative must be conjugated to its negative by an involution. We give a negative answer for nilpotent operators of order three.

Theorem 1.2. There are M∈B(H)M\in\mathcal{B}(\mathcal{H}) and an invertible S∈B(H)S\in\mathcal{B}(\mathcal{H}) such that

M3=0≠M2,S−1MS=−M,M^{3}=0\ne M^{2},\qquad S^{-1}MS=-M,

but no involution J∈B(H)J\in\mathcal{B}(\mathcal{H}) satisfies JMJ=−MJMJ=-M.

Section 4 allows the upper-right entry in the block form of [10 Section 2.6] to vary. Under the hypotheses of Theorem 4.2, a bounded involution exists exactly when the ratios of the block coefficients are bounded. The involution need not preserve the summands. Theorem 4.5 proves Theorem 1.2 with an operator whose singular values are pp-summable for every p>0p>0. For each η>0\eta>0, it gives such operators M,SM,S with max⁡{∥S∥,∥S−1∥}≤1+η\max\{\lVert S\rVert,\lVert S^{-1}\rVert\}\leq1+\eta.

The paper is organized as follows. § 2 determines the additive lengths. Section 3 proves the product representations, including the unitary case and the factorization with a fixed right factor. § 5 constructs the nilpotent examples.

Sums of three commutators

Proposition 2.1. Let AA be a unital associative algebra over a field of characteristic different from two. For every a∈Aa\in A, diag⁡(a,a,−a,−a)\operatorname{diag}(a,a,-a,-a) is a commutator of two idempotents in M4(A)M_4(A).

Proof. Let 11 denote the identity of AA, and set

a^=(a00a),r=(0a2+1410).\widehat{a}=\begin{pmatrix}a&0\\0&a\end{pmatrix},\qquad r=\begin{pmatrix}0&a^2+\frac{1}{4}\\1&0\end{pmatrix}.

Then, r2=a^ 2+14I2r^2=\widehat{a}^{\,2}+\frac{1}{4}I_2 and ra^=a^rr\widehat{a}=\widehat{a}r. In M4(A)M_4(A) put

t=(a^00−a^),s=(r00r),j=(0I2I20).t=\begin{pmatrix}\widehat{a}&0\\0&-\widehat{a}\end{pmatrix},\qquad s=\begin{pmatrix}r&0\\0&r\end{pmatrix},\qquad j=\begin{pmatrix}0&I_2\\I_2&0\end{pmatrix}.

These elements satisfy

j2=I4,jt=−tj,js=sj,st=ts,s2=t2+14I4.j^2=I_4,\qquad jt=-tj,\qquad js=sj,\qquad st=ts,\qquad s^2=t^2+\frac{1}{4}I_4.

These are the conditions in [5 Theorem 2.1(iii)]. Their equivalence with condition (i) proves that tt is a commutator of two idempotents. □

We use the alternating direct sums from the proof of [10 Proposition 7.18], together with Proposition 2.1, to obtain the following statements for commutators of idempotents.

Corollary 2.2. For every A∈B(H)A\in\mathcal{B}(\mathcal{H}),

A(∞)⊕(−A)(∞)∈Cε,A⊕0∈∑2Cε.A^{(\infty)}\oplus(-A)^{(\infty)}\in\mathcal{C}_{\varepsilon},\qquad A\oplus0\in\sum_{2}\mathcal{C}_{\varepsilon}.

where the zero summand acts on an infinite-dimensional space. There is also a function

f:B(H)→B(H)f:\mathcal{B}(\mathcal{H})\to\mathcal{B}(\mathcal{H})

such that A⊕f(A)∈CεA\oplus f(A)\in\mathcal{C}_{\varepsilon} for every A∈B(H)A\in\mathcal{B}(\mathcal{H}).

Proof. Apply Proposition 2.1 to A(∞)A^{(\infty)} and use A(∞)⊕A(∞)≃A(∞)A^{(\infty)}\oplus A^{(\infty)}\simeq A^{(\infty)}. Regrouping the copies of AA and −A-A gives the first assertion.

On H(∞)\mathcal{H}^{(\infty)}, define

YA=diag⁡(A,−A,A,−A,…),ZA=diag⁡(0,A,−A,A,−A,…).Y_A=\operatorname{diag}(A,-A,A,-A,\ldots),\qquad Z_A=\operatorname{diag}(0,A,-A,A,-A,\ldots).

The first assertion and a permutation of the summands give YA∈CεY_A\in\mathcal{C}_{\varepsilon}. Since ZA≃0⊕YAZ_A\simeq0\oplus Y_A and

0⊕[E,F]=[0⊕E,0⊕F],0\oplus[E,F]=[0\oplus E,0\oplus F],

we also have ZA∈CεZ_A\in\mathcal{C}_{\varepsilon}. Their sum is

YA+ZA=diag⁡(A,0,0,…)≃A⊕0.Y_A+Z_A=\operatorname{diag}(A,0,0,\ldots)\simeq A\oplus0.

Finally, fix a unitary W:H→H(∞)W:\mathcal{H}\to\mathcal{H}^{(\infty)} and put

f(A)=W∗diag⁡(−A,A,−A,A,…)W.f(A)=W^*\operatorname{diag}(-A,A,-A,A,\ldots)W.

Then, A⊕f(A)≃YA∈CεA\oplus f(A)\simeq Y_A\in\mathcal{C}_{\varepsilon}. □

We recall the following result of Davidson and Marcoux [4 Lemma 2.6(ii)].

Lemma 2.3. Let f:B(H)→B(H)f:\mathcal{B}(\mathcal{H})\to\mathcal{B}(\mathcal{H}) be a function, and let M1,M2\mathcal{M}_1,\mathcal{M}_2 be unitarily invariant subsets of B(H)\mathcal{B}(\mathcal{H}). Suppose that

A⊕f(A)∈M1,B⊕0∈M2(A,B∈B(H)),A\oplus f(A)\in\mathcal{M}_1,\qquad B\oplus0\in\mathcal{M}_2\qquad(A,B\in\mathcal{B}(\mathcal{H})),

where the zero summand is infinite-dimensional. Then, B(H)=M1+M2\mathcal{B}(\mathcal{H})=\mathcal{M}_1+\mathcal{M}_2.

Every commutator of idempotents anticommutes with an involution and is so similar to its negative; see [11 Proposition 2.1]. Every operator anticommuting with an involution is also a difference of idempotents: relative to the topological direct sum of the two eigenspaces, its diagonal blocks vanish, and the calculation in the proof of [11 Proposition 2.2] applies.

Theorem 2.4. One has

B(H)=∑3CE=∑3DE,−3I∉∑2DE.\mathcal{B}(\mathcal{H})=\sum_{3}\mathcal{C}_{\mathcal{E}}=\sum_{3}\mathcal{D}_{\mathcal{E}},\qquad-3I\notin\sum_{2}\mathcal{D}_{\mathcal{E}}.

Every operator is also a sum of three operators anticommuting with involutions. The identity is not a sum of two operators similar to their negatives. Hence ℓ+(Δ)=3\ell_{+}(\Delta)=3 for each of the four classes in Theorem 1.1.

Proof. The classes CE\mathcal{C}_{\mathcal{E}} and ∑2CE\sum_{2}\mathcal{C}_{\mathcal{E}} are unitarily invariant. Corollary 2.2 gives the hypotheses of Lemma 2.3 with M1=CE\mathcal{M}_1=\mathcal{C}_{\mathcal{E}} and M2=∑2CE\mathcal{M}_2=\sum_{2}\mathcal{C}_{\mathcal{E}}. Hence

B(H)=CE+∑2CE=∑3CE.\mathcal{B}(\mathcal{H})=\mathcal{C}_{\mathcal{E}}+\sum_{2}\mathcal{C}_{\mathcal{E}}=\sum_{3}\mathcal{C}_{\mathcal{E}}.

The inclusions stated before the theorem give the remaining three upper bounds.

The spectral argument in the proof of [10 Proposition 7.7] excludes I=A+BI=A+B whenever AA and BB are similar to their negatives. It therefore also excludes two commutators of idempotents or two operators anticommuting with involutions.

By [13 Theorem 1], −I-I is not a sum of four idempotents. If

−3I=(E1−F1)+(E2−F2),Ej,Fj∈E,-3I=(E_1-F_1)+(E_2-F_2),\qquad E_j,F_j\in\mathcal{E},

then

−I=E1+(I−F1)+E2+(I−F2)-I=E_1+(I-F_1)+E_2+(I-F_2)

would be such a sum. Thus, −3I∉∑2DE-3I\notin\sum_{2}\mathcal{D}_{\mathcal{E}}.

Products of two commutators

An isometry V∈B(H)V\in\mathcal{B}(\mathcal{H}) is called pure if ⋂k≥0VkH={0}\bigcap_{k\ge0}V^k\mathcal{H}=\{0\}. It has infinite multiplicity when ker⁡V∗\ker V^* is infinite-dimensional.

Theorem 3.1. For every C∈B(H)C\in\mathcal{B}(\mathcal{H}) with ∥C∥<1\lVert C\rVert<1, there are pure isometries V,W∈B(H)V,W\in\mathcal{B}(\mathcal{H}) such that C=V∗WC=V^*W and both ker⁡V∗\ker V^* and ker⁡W∗\ker W^* are infinite-dimensional.

Proof. Let D=(I−C∗C)1/2D=(I-C^*C)^{1/2}. Since

D≥1−∥C∥2I>0,D\ge\sqrt{1-\lVert C\rVert^2}I>0,

DD has a bounded inverse. On K=H⊕H⊕H\mathcal{K}=\mathcal{H}\oplus\mathcal{H}\oplus\mathcal{H}, define

E,F:H⟶K,Ex=(x,0,0),Fx=(Cx,Dx,0).E,F:\mathcal{H}\longrightarrow\mathcal{K},\qquad Ex=(x,0,0),\qquad Fx=(Cx,Dx,0).

The equality C∗C+D2=IC^*C+D^2=I shows that EE and FF are isometries, with E∗F=CE^*F=C. We choose a unitary U:K→HU:\mathcal{K}\to\mathcal{H} such that both UEUE and UFUF are pure.

Fix an orthonormal basis (ej)j≥1(e_j)_{j\ge1} of H\mathcal{H}, and put

Hn=span⁡‾{ej:j≥n},Mn=EHn+FHn(n≥1).H_n=\overline{\operatorname{span}}\{e_j:j\ge n\},\qquad M_n=EH_n+FH_n\qquad(n\ge1).

@page 4/13 starts_mid=0 ends_mid=1

Lemma 2.3. Let f:B(H)→B(H)f:\mathcal{B}(\mathcal{H})\to\mathcal{B}(\mathcal{H}) be a function, and let M1,M2\mathcal{M}_1,\mathcal{M}_2 be unitarily invariant subsets of B(H)\mathcal{B}(\mathcal{H}). Suppose that

A⊕f(A)∈M1,B⊕0∈M2(A,B∈B(H)),A\oplus f(A)\in\mathcal{M}_1,\qquad B\oplus0\in\mathcal{M}_2\qquad(A,B\in\mathcal{B}(\mathcal{H})),

where the zero summand is infinite-dimensional. Then, B(H)=M1+M2\mathcal{B}(\mathcal{H})=\mathcal{M}_1+\mathcal{M}_2.

Every commutator of idempotents anticommutes with an involution and is so similar to its negative; see [11 Proposition 2.1]. Every operator anticommuting with an involution is also a difference of idempotents: relative to the topological direct sum of the two eigenspaces, its diagonal blocks vanish, and the calculation in the proof of [11 Proposition 2.2] applies.

Theorem 2.4. One has

B(H)=∑3CE=∑3DE,−3I∉∑2DE.\mathcal{B}(\mathcal{H})=\sum_{3}\mathcal{C}_{\mathcal{E}}=\sum_{3}\mathcal{D}_{\mathcal{E}},\qquad-3I\notin\sum_{2}\mathcal{D}_{\mathcal{E}}.

Every operator is also a sum of three operators anticommuting with involutions. The identity is not a sum of two operators similar to their negatives. Hence ℓ+(Δ)=3\ell_{+}(\Delta)=3 for each of the four classes in Theorem 1.1.

Proof. The classes CE\mathcal{C}_{\mathcal{E}} and ∑2CE\sum_{2}\mathcal{C}_{\mathcal{E}} are unitarily invariant. Corollary 2.2 gives the hypotheses of Lemma 2.3 with M1=CE\mathcal{M}_1=\mathcal{C}_{\mathcal{E}} and M2=∑2CE\mathcal{M}_2=\sum_{2}\mathcal{C}_{\mathcal{E}}. Hence

B(H)=CE+∑2CE=∑3CE.\mathcal{B}(\mathcal{H})=\mathcal{C}_{\mathcal{E}}+\sum_{2}\mathcal{C}_{\mathcal{E}}=\sum_{3}\mathcal{C}_{\mathcal{E}}.

The inclusions stated before the theorem give the remaining three upper bounds.

The spectral argument in the proof of [10 Proposition 7.7] excludes I=A+BI=A+B whenever AA and BB are similar to their negatives. It therefore also excludes two commutators of idempotents or two operators anticommuting with involutions.

By [13 Theorem 1], −I-I is not a sum of four idempotents. If

−3I=(E1−F1)+(E2−F2),Ej,Fj∈E,-3I=(E_1-F_1)+(E_2-F_2),\qquad E_j,F_j\in\mathcal{E},

then

−I=E1+(I−F1)+E2+(I−F2)-I=E_1+(I-F_1)+E_2+(I-F_2)

would be such a sum. Thus, −3I∉∑2DE-3I\notin\sum_{2}\mathcal{D}_{\mathcal{E}}.

Products of two commutators

An isometry V∈B(H)V\in\mathcal{B}(\mathcal{H}) is called pure if ⋂k≥0VkH={0}\bigcap_{k\ge0}V^k\mathcal{H}=\{0\}. It has infinite multiplicity when ker⁡V∗\ker V^* is infinite-dimensional.

Theorem 3.1. For every C∈B(H)C\in\mathcal{B}(\mathcal{H}) with ∥C∥<1\lVert C\rVert<1, there are pure isometries V,W∈B(H)V,W\in\mathcal{B}(\mathcal{H}) such that C=V∗WC=V^*W and both ker⁡V∗\ker V^* and ker⁡W∗\ker W^* are infinite-dimensional.

Proof. Let D=(I−C∗C)1/2D=(I-C^*C)^{1/2}. Since

D≥1−∥C∥2I>0,D\ge\sqrt{1-\lVert C\rVert^2}I>0,

DD has a bounded inverse. On K=H⊕H⊕H\mathcal{K}=\mathcal{H}\oplus\mathcal{H}\oplus\mathcal{H}, define

E,F:H⟶K,Ex=(x,0,0),Fx=(Cx,Dx,0).E,F:\mathcal{H}\longrightarrow\mathcal{K},\qquad Ex=(x,0,0),\qquad Fx=(Cx,Dx,0).

The equality C∗C+D2=IC^*C+D^2=I shows that EE and FF are isometries, with E∗F=CE^*F=C. We choose a unitary U:K→HU:\mathcal{K}\to\mathcal{H} such that both UEUE and UFUF are pure.

Fix an orthonormal basis (ej)j≥1(e_j)_{j\ge1} of H\mathcal{H}, and put

Hn=span⁡‾{ej:j≥n},Mn=EHn+FHn(n≥1).H_n=\overline{\operatorname{span}}\{e_j:j\ge n\},\qquad M_n=EH_n+FH_n\qquad(n\ge1).

The operator

L:H⊕H⟶H⊕H,L(x,y)=(x+Cy,Dy)L:\mathcal{H}\oplus\mathcal{H}\longrightarrow\mathcal{H}\oplus\mathcal{H},\qquad L(x,y)=(x+Cy,Dy)

is boundedly invertible, with

L−1(a,b)=(a−CD−1b,D−1b).L^{-1}(a,b)=(a-CD^{-1}b,D^{-1}b).

Thus,

Mn=L(Hn⊕Hn)⊕{0}.M_n=L(\mathcal{H}_n\oplus\mathcal{H}_n)\oplus\{0\}.

In particular, the MnM_n are closed and decreasing, and

M1=H⊕H⊕{0},⋂n≥1Mn={0}.M_1=\mathcal{H}\oplus\mathcal{H}\oplus\{0\},\qquad\bigcap_{n\geq1}M_n=\{0\}.

The subspace Hn+1⊕Hn+1\mathcal{H}_{n+1}\oplus\mathcal{H}_{n+1} has codimension two in Hn⊕Hn\mathcal{H}_n\oplus\mathcal{H}_n. Since LL is invertible, Mn+1M_{n+1} has codimension two in MnM_n. Therefore,

Gn=Mn∩Mn+1⊥G_n=M_n\cap M_{n+1}^{\perp}

has dimension two. The orthogonal projections onto MnM_n decrease strongly to zero. Consequently,

Mn=⨁k≥nGk,K=M1⊥⊕⨁n≥1Gn.M_n=\bigoplus_{k\geq n}G_k,\qquad\mathcal{K}=M_1^\perp\oplus\bigoplus_{n\geq1}G_n.

Choose an orthonormal basis an,bna_n,b_n of each GnG_n, and an orthonormal basis (zn)n≥1(z_n)_{n\geq1} of M1⊥={0}⊕{0}⊕HM_1^\perp=\{0\}\oplus\{0\}\oplus\mathcal{H}. Define U:K→HU:\mathcal{K}\to\mathcal{H} on these basis vectors by

Uzn=e3n−2,Uan=e3n−1,Ubn=e3n.Uz_n=e_{3n-2},\qquad Ua_n=e_{3n-1},\qquad Ub_n=e_{3n}.

This defines a unitary. Moreover,

UMn=span⁡‾{e3k−1,e3k:k≥n}⊆Hn+1.UM_n=\overline{\operatorname{span}}\{e_{3k-1},e_{3k}:k\geq n\}\subseteq\mathcal{H}_{n+1}.

Now put V=UEV=UE and W=UFW=UF. These are isometries on H\mathcal{H}, and

V∗W=E∗U∗UF=E∗F=C.V^*W=E^*U^*UF=E^*F=C.

The inclusions EHn,FHn⊆MnE\mathcal{H}_n,F\mathcal{H}_n\subseteq M_n give

VHn⊆Hn+1,WHn⊆Hn+1.V\mathcal{H}_n\subseteq\mathcal{H}_{n+1},\qquad W\mathcal{H}_n\subseteq\mathcal{H}_{n+1}.

Since H1=H\mathcal{H}_1=\mathcal{H}, it follows by induction that

VkH⊆Hk+1,WkH⊆Hk+1(k≥0).V^k\mathcal{H}\subseteq\mathcal{H}_{k+1},\qquad W^k\mathcal{H}\subseteq\mathcal{H}_{k+1}\qquad(k\geq0).

The intersection of the Hn\mathcal{H}_n is zero, so both isometries are pure. Finally, ran⁡V,ran⁡W⊆UM1\operatorname{ran}V,\operatorname{ran}W\subseteq U M_1, and so

U(M1⊥)⊆ker⁡V∗∩ker⁡W∗.U(M_1^\perp)\subseteq\ker V^*\cap\ker W^*.

The subspace U(M1⊥)U(M_1^\perp) is infinite-dimensional, so both kernels are infinite-dimensional. □

Lemma 3.2. Let V∈B(H)V\in\mathcal{B}(\mathcal{H}) be a pure isometry with infinite-dimensional ker⁡V∗\ker V^*. Then, λV,λV∗∈CE\lambda V,\lambda V^*\in\mathcal{C}_{E} for every λ∈C\lambda\in\mathbb{C}. There is also a self-adjoint unitary GG such that GVG=−VGVG=-V and GV∗G=−V∗GV^*G=-V^*.

Proof. By the Wold decomposition [15 Chapter I, §1],

H=⨁k≥0VkD,D=ker⁡V∗.\mathcal{H}=\bigoplus_{k\geq0}V^k\mathcal{D},\qquad\mathcal{D}=\ker V^*.

On these summands define

G(Vkx)=(−1)kVkx(x∈D, k≥0).G(V^kx)=(-1)^kV^kx\qquad(x\in\mathcal{D},\ k\geq0).

Then, G=G∗=G−1G=G^*=G^{-1} and GVG=−VGVG=-V. Taking adjoints gives GV∗G=−V∗GV^*G=-V^*.

This decomposition identifies VV with the unilateral shift on ℓ2(N0,D)\ell^2(\mathbb{N}_0,\mathcal{D}), where D\mathcal{D} is separable and infinite-dimensional. Split D\mathcal{D} into four orthogonal infinite-dimensional subspaces. Each gives a reducing subspace on which the shift is unitarily equivalent to VV. Thus,

V≃V⊕V⊕V⊕V.V\simeq V\oplus V\oplus V\oplus V.

Together with V≃−VV \simeq-V, this yields

λV≃diag⁡(λV,λV,−λV,−λV).\lambda V \simeq\operatorname{diag}(\lambda V,\lambda V,-\lambda V,-\lambda V).

Proposition 2.1, applied to a=λVa=\lambda V, proves λV∈CE\lambda V \in\mathcal{C}_{\mathsf{E}}. Taking adjoints in the two unitary equivalences and applying the same proposition to a=λV∗a=\lambda V^{*} proves λV∗∈CE\lambda V^{*} \in\mathcal{C}_{\mathsf{E}}. □

Theorem 2.4 and the following theorem prove Theorem 1.1.

Theorem 3.3. For every T∈B(H)T \in\mathcal{B}(\mathcal{H}) and every real c>∥T∥c > \lVert T\rVert, there are A,B∈CEA,B \in\mathcal{C}_{\mathsf{E}} such that

T=AB,AA∗=B∗B=cI,∥A∥=∥B∥=c.T=AB,\qquad AA^{*}=B^{*}B=cI,\qquad\lVert A\rVert=\lVert B\rVert=\sqrt{c}.

Consequently,

B(H)=∏2CE=∏2DE.\mathcal{B}(\mathcal{H})=\prod_{2}\mathcal{C}_{\mathsf{E}}=\prod_{2}\mathcal{D}_{\mathsf{E}}.

Every operator is also a product of two operators anticommuting with involutions. Hence ℓ×(△)=2\ell_{\times}(\triangle)=2 for each of the four classes in Theorem 1.1.

Proof. Apply Theorem 3.1 to C=T/cC=T/c, and write T/c=V∗WT/c=V^{*}W. Set

A=cV∗,B=cW.A=\sqrt{c}V^{*},\qquad B=\sqrt{c}W.

Lemma 3.2 gives A,B∈CEA,B\in\mathcal{C}_{\mathsf{E}}, while

AB=T,AA∗=cV∗V=cI,B∗B=cW∗W=cI.AB=T,\qquad AA^{*}=cV^{*}V=cI,\qquad B^{*}B=cW^{*}W=cI.

Since VV and WW are isometries, ∥A∥=∥B∥=c\lVert A\rVert=\lVert B\rVert=\sqrt{c}. Each commutator anticommutes with an involution, is similar to its negative, and is a difference of idempotents. This gives the other three product representations.

By Theorem 2.4, each of the four classes is a proper subset of B(H)\mathcal{B}(\mathcal{H}). Hence one factor does not suffice. □

By [10 Corollary 7.12(a)], every unitary is a product of two commutators of idempotents; those factors need not be unitary. To obtain unitary factors, we use products of bilateral shifts. Marcoux, Omladič, Radjavi and Zhang proved that every unitary is a product of three bilateral shifts of multiplicity one [9 Theorem 5.3]. The author proved that two suffice [1 Theorem 1.1]. We first represent one such shift as a commutator of a fixed pair of idempotents. We use U(H)\mathcal{U}(\mathcal{H}) for the unitary group.

Proposition 3.4. There exist an orthogonal projection PP and an idempotent QQ on H\mathcal{H} such that [P,Q][P,Q] is a bilateral shift of multiplicity one and every U∈U(H)U\in\mathcal{U}(\mathcal{H}) has a representation

U=[R1PR1∗,R1QR1∗][R2PR2∗,R2QR2∗]U=[R_{1}PR_{1}^{*},R_{1}QR_{1}^{*}][R_{2}PR_{2}^{*},R_{2}QR_{2}^{*}]

for some R1,R2∈U(H)R_{1},R_{2}\in\mathcal{U}(\mathcal{H}). In particular,

U(H)=∏2(U(H)∩CE),\mathcal{U}(\mathcal{H})=\prod_{2}\bigl(\mathcal{U}(\mathcal{H})\cap\mathcal{C}_{\mathsf{E}}\bigr),

and of course one factor does not suffice.

Proof. Identify H\mathcal{H} with L⊕L\mathcal{L}\oplus\mathcal{L}, where L=ℓ2(Z)\mathcal{L}=\ell^{2}(\mathbb{Z}), and let Wek=ek+1We_{k}=e_{k+1} on its standard orthonormal basis. We need an operator RR commuting with WW and satisfying

R2=W+14I.R^{2}=W+\frac{1}{4}I.

The equation Wx=−xWx=-x has no nonzero square-summable solution, so I+WI+W and its adjoint are injective. In the polar decomposition

I+W=ZF,F=∣I+W∣,I+W=ZF,\qquad F=\lvert I+W\rvert,

ZZ is unitary and commutes with FF and WW, since I+WI+W is normal. Moreover,

Z2F2=(I+W)2=W(2I+W+W∗)=WF2.Z^{2}F^{2}=(I+W)^{2}=W(2I+W+W^{*})=WF^{2}.

The range of F2F^2 is dense, hence Z2=WZ^2=W. The norm-convergent binomial series gives

R=Z∑j=0∞(1/2j)(W∗4)j,RW=WR,R2=W+14I.R=Z\sum_{j=0}^{\infty}\binom{1/2}{j}\left(\frac{W^*}{4}\right)^j,\qquad RW=WR,\qquad R^2=W+\frac{1}{4}I.

Set

P=(I000),B=(0WI0).P=\begin{pmatrix} I & 0 \\ 0 & 0 \end{pmatrix},\qquad B=\begin{pmatrix} 0 & W \\ I & 0 \end{pmatrix}.

Apply [5 Theorem 2.2] with both square roots equal to RR. The identities R2=W+14IR^2=W+\frac{1}{4}I and RW=WRRW=WR verify its hypotheses and give an idempotent QQ with B=[P,Q]B=[P,Q]. We keep this pair P,QP,Q fixed. The vectors

f2k=(ek,0),f2k+1=(0,ek)(k∈Z)f_{2k}=(e_k,0),\qquad f_{2k+1}=(0,e_k)\qquad(k\in\mathbb{Z})

form an orthonormal basis, and Bfj=fj+1Bf_j=f_{j+1} for every integer jj. Thus, BB is a bilateral shift of multiplicity one.

By [1 Theorem 1.1], every unitary UU can be written as

U=(R1BR1∗)(R2BR2∗)U=(R_1BR_1^*)(R_2BR_2^*)

for unitary R1,R2R_1,R_2. Substituting B=[P,Q]B=[P,Q] proves the stated representation. Finally, a commutator of idempotents is similar to its negative, whereas II is not. Thus, one factor does not suffice. □

Thus, for a unitary UU, the identities AA∗=B∗B=cIAA^*=B^*B=cI hold with c=1=∥U∥c=1=\lVert U\rVert. In this case the factors are unitary, so A∗A=BB∗=IA^*A=BB^*=I as well.

For differences of idempotents, one can fix the right factor in advance.

Proposition 3.5. There are a fixed isometry V∈EeV\in\mathcal{E}_e and fixed self-adjoint unitaries J,G∈B(H)J,G\in\mathcal{B}(\mathcal{H}) such that, for every T∈B(H)T\in\mathcal{B}(\mathcal{H}), the operator

AT=TV∗−JTV∗JA_T=TV^*-JTV^*J

satisfies

T=ATV,JATJ=−AT,GVG=−V.T=A_TV,\qquad JA_TJ=-A_T,\qquad GVG=-V.

The map T↦ATT\mapsto A_T is bounded and linear. Both ATA_T and VV are differences of idempotents.

Proof. Identify H\mathcal{H} with ⨁n≥0K\bigoplus_{n\geq0}\mathcal{K}, where K\mathcal{K} is a separable infinite-dimensional Hilbert space. Let

V(x0,x1,x2,…)=(0,x0,x1,…),V(x_0,x_1,x_2,\ldots)=(0,x_0,x_1,\ldots),
G(x0,x1,x2,…)=(x0,−x1,x2,…).G(x_0,x_1,x_2,\ldots)=(x_0,-x_1,x_2,\ldots).

Then, V∗V=IV^*V=I, G=G∗=G−1G=G^*=G^{-1} and GVG=−VGVG=-V. The isometry VV is pure of infinite multiplicity, so V∈EeV\in\mathcal{E}_e by Lemma 3.2.

Put M=ran⁡VM=\operatorname{ran}V and N=ker⁡V∗N=\ker V^*. Both spaces are infinite-dimensional and H=M⊕N\mathcal{H}=M\oplus N. Choose a unitary R:M→NR:M\to N and define

J(m+n)=Rm+R∗n(m∈M, n∈N).J(m+n)=Rm+R^*n\qquad(m\in M,\ n\in N).

This is a self-adjoint unitary interchanging MM and NN. Therefore,

V∗JV=0.V^*JV=0.

For AT=TV∗−JTV∗JA_T=TV^*-JTV^*J, we obtain

JATJ=JTV∗J−TV∗=−ATJA_TJ=JTV^*J-TV^*=-A_T

and

ATV=TV∗V−JTV∗JV=T.A_TV=TV^*V-JTV^*JV=T.

The formula defines a bounded linear map T↦ATT\mapsto A_T. Since ATA_T and VV anticommute with involutions, both are differences of idempotents. □

Nilpotent operators similar to their negatives

In [10 Section 2.9], Marcoux, Radjavi and Zhang ask whether a nilpotent operator similar to its negative is conjugate to its negative by an involution. We replace the identity in the upper-right corner of their block form [10 Section 2.6] by an operator CC. For D,C∈B(K)D,C \in\mathcal{B}(\mathcal{K}), put

MD,C=(0DC00D000)on H=K⊕K⊕K.(1)M_{D,C}=\begin{pmatrix}0&D&C\\0&0&D\\0&0&0\end{pmatrix}\quad\text{on }\mathcal{H}=\mathcal{K}\oplus\mathcal{K}\oplus\mathcal{K}. \tag*{(1)}

Throughout this section, DD is injective. Then, MD,C3=0≠MD,C2M_{D,C}^{3}=0\ne M_{D,C}^{2}, with D2D^{2} in position (1,3)(1,3) of MD,C2M_{D,C}^{2}. We take DD and CC block diagonal on the same orthogonal sum. Theorem 4.2 below determines exactly when an operator in this family is conjugate to its negative by a bounded involution. Choosing CC compact as well as DD will give compact counterexamples.

Lemma 4.1. Let L\mathcal{L} be a finite-dimensional complex Hilbert space of odd dimension. For every A∈B(L)A\in\mathcal{B}(\mathcal{L}) and every unitary W∈B(L)W\in\mathcal{B}(\mathcal{L}),

∥A2−I∥+∥A+WAW∗2∥2≥1.(2)\lVert A^{2}-I\rVert+\left\lVert\frac{A+WAW^{*}}{2}\right\rVert^{2}\ge1. \tag*{(2)}

Proof. Put δ=∥A2−I∥\delta=\lVert A^{2}-I\rVert. If δ≥1\delta\ge1, the inequality is immediate. Suppose δ<1\delta<1 and write A=X+iYA=X+iY, where XX and YY are self-adjoint. Taking real parts gives

X2−Y2=A2+(A∗)22≥(1−δ)I,X^{2}-Y^{2}=\frac{A^{2}+(A^{*})^{2}}{2}\ge(1-\delta)I,

and hence X2≥(1−δ)IX^{2}\ge(1-\delta)I. This implies that every eigenvalue of XX has absolute value at least a=1−δa=\sqrt{1-\delta}.

Since dim⁡L\dim\mathcal{L} is odd, either the positive or the negative spectral subspace M\mathcal{M} of XX has dimension greater than dim⁡L/2\dim\mathcal{L}/2. Choose the corresponding sign η∈{1,−1}\eta\in\{1,-1\}, so that

η⟨Xx,x⟩≥a∥x∥2(x∈M).\eta\langle Xx,x\rangle\ge a\lVert x\rVert^{2}\qquad(x\in\mathcal{M}).

The spaces M\mathcal{M} and WMW\mathcal{M} have a nonzero intersection. For a unit vector xx in that intersection, both xx and W∗xW^{*}x belong to M\mathcal{M}. Thus,

ηRe⁡⟨(A+WAW∗)x,x⟩=η⟨(X+WXW∗)x,x⟩≥2a.\eta\operatorname{Re}\langle(A+WAW^{*})x,x\rangle=\eta\langle(X+WXW^{*})x,x\rangle\ge2a.

It follows that ∥A+WAW∗∥≥21−δ\lVert A+WAW^{*}\rVert\ge2\sqrt{1-\delta}, which proves (2).

The unitary blocks in the next theorem are arbitrary. The involution need not preserve the summands Kn\mathcal{K}_{n}.

Theorem 4.2. Let K=⨁n≥1Kn\mathcal{K}=\bigoplus_{n\ge1}\mathcal{K}_{n}, where each Kn\mathcal{K}_{n} has finite odd dimension, and let UnU_{n} be unitary on Kn\mathcal{K}_{n}. Let (an)(a_{n}) and (cn)(c_{n}) be bounded sequences of positive numbers. Suppose that, for some 0<r<10<r<1,

an+1≤ran,an+12cn+1≤ran2cnfor all sufficiently large n.(3)a_{n+1}\le ra_{n},\qquad\frac{a_{n+1}^{2}}{c_{n+1}}\le r\frac{a_{n}^{2}}{c_{n}}\quad\text{for all sufficiently large }n. \tag*{(3)}

Set

D=⨁n≥1anUn,C=⨁n≥1cnIKn.D=\bigoplus_{n\ge1}a_{n}U_{n},\qquad C=\bigoplus_{n\ge1}c_{n}I_{\mathcal{K}_{n}}.

A bounded involution JJ satisfying JMD,CJ=−MD,CJM_{D,C}J=-M_{D,C} exists if and only if

sup⁡n≥1cnan<∞.(4)\sup_{n\ge1}\frac{c_{n}}{a_{n}}<\infty. \tag*{(4)}

Moreover, every such involution satisfies, for all sufficiently large nn,

1≤∥J∥2(2r(1+r)1−rancn+an2cn2).(5)1\le\lVert J\rVert^{2}\left(\frac{2r(1+r)}{1-r}\frac{a_{n}}{c_{n}}+\frac{a_{n}^{2}}{c_{n}^{2}}\right). \tag*{(5)}

Proof. Suppose first that (4.4) holds. The operator

R=⨁n≥1cnanUn∗R=\bigoplus_{n\ge1}\frac{c_n}{a_n}U_n^*

is bounded and satisfies DR=CDR=C. Direct multiplication gives

J=(I000−I−2R00I),J2=I,JMD,C=−MD,CJ.(6)J=\begin{pmatrix} I&0&0\\0&-I&-2R\\0&0&I \end{pmatrix},\qquad J^2=I,\qquad JM_{D,C}=-M_{D,C}J. \tag*{(6)}

Conversely, suppose that JJ is a bounded involution anticommuting with MD,CM_{D,C}, and put b=∥J∥b=\lVert J\rVert. Since DD is injective,

ker⁡MD,C=K⊕0⊕0,ker⁡MD,C2=K⊕K⊕0.\ker M_{D,C}=\mathcal K\oplus0\oplus0,\qquad\ker M_{D,C}^2=\mathcal K\oplus\mathcal K\oplus0.

Both spaces are invariant under JJ. Hence

J=(J1J12J130J2J2300J3),J12=J22=J32=I.J=\begin{pmatrix}J_1&J_{12}&J_{13}\\0&J_2&J_{23}\\0&0&J_3\end{pmatrix},\qquad J_1^2=J_2^2=J_3^2=I.

Comparing entries gives

J1D=−DJ2,J2D=−DJ3,J1C+CJ3=−DJ23−J12D.J_1D=-DJ_2,\qquad J_2D=-DJ_3,\qquad J_1C+CJ_3=-DJ_{23}-J_{12}D.

In particular,

J1D2=D2J3,J1C+CJ3=−DJ23−J12D.(7)J_1D^2=D^2J_3,\qquad J_1C+CJ_3=-DJ_{23}-J_{12}D. \tag*{(7)}

For C=IC=I, these are the relations in the proof of [10 Proposition 2.8].

Fix a sufficiently large index nn, and put t=ant=a_n, s=cns=c_n and U=UnU=U_n. Use the orthogonal decomposition

K=K−⊕Kn⊕K+,K−=⨁j<nKj,K+=⨁j>nKj.\mathcal K=\mathcal K_-\oplus\mathcal K_n\oplus\mathcal K_+,\qquad\mathcal K_-=\bigoplus_{j<n}\mathcal K_j,\qquad\mathcal K_+=\bigoplus_{j>n}\mathcal K_j.

Write D=D−⊕tU⊕D+D=D_-\oplus tU\oplus D_+ and C=C−⊕sI⊕C+C=C_-\oplus sI\oplus C_+. Define the bounded operator

H+=⨁j>naj2cjUj2,D+2=C+H+.H_+=\bigoplus_{j>n}\frac{a_j^2}{c_j}U_j^2,\qquad D_+^2=C_+H_+.

Both sequences aja_j and aj2/cja_j^2/c_j eventually decrease geometrically by (4.3). Choose nn so large that these inequalities hold from n−1n-1 onward and, for each sequence, the term at n−1n-1 is no greater than any preceding term. Then,

∥D−−1∥≤rt,∥D−−2C−∥≤rst2,∥D+∥≤rt,∥H+∥≤rt2s.(8)\lVert D_-^{-1}\rVert\le\frac{r}{t},\qquad\lVert D_-^{-2}C_-\rVert\le\frac{rs}{t^2},\qquad\lVert D_+\rVert\le rt,\qquad\lVert H_+\rVert\le\frac{rt^2}{s}. \tag*{(8)}

The operator D−D_- is invertible on the finite-dimensional space K−\mathcal K_-. The second condition in (4.3) makes H+H_+ bounded.

Relative to this decomposition, write

J3=(∗X∗ZAT∗Y∗),J_3=\begin{pmatrix} *&X&*\\ Z&A&T\\ *&Y&* \end{pmatrix},

and let BB be the middle diagonal entry of J1J_1. Every entry of J1,J3,J12,J23J_1,J_3,J_{12},J_{23} has norm at most bb.

Let L−,R−L_-,R_- be the entries of J23,J12J_{23},J_{12} from Kn\mathcal K_n to K−\mathcal K_-. The corresponding entries in (4.7), with the entry of J1J_1 eliminated, give

sD−2X+C−(tU)2=−(D−L−+R−tU)(tU)2.sD_-^2X+C_-(tU)^2=-(D_-L_-+R_-tU)(tU)^2.

Multiplication on the left by s−1D−−2s^{-1}D_-^{-2} and (4.8) yield

(1−r)∥X∥≤bts(r+r2),∥X∥≤btsr(1+r)1−r.(1-r)\lVert X\rVert\le b\frac{t}{s}(r+r^2),\qquad\lVert X\rVert\le b\frac{t}{s}\frac{r(1+r)}{1-r}.

Next consider T:K+→KnT:\mathcal{K}_{+}\to\mathcal{K}_{n}. Let B+B_{+} be the corresponding entry of J1J_{1}, and let L+,R+L_{+},R_{+} be those of J23,J12J_{23},J_{12}. Equations (7) give

B+D+2=(tU)2T,B+C++sT=−tUL+−R+D+.B_{+}D_{+}^{2}=(tU)^{2}T,\qquad B_{+}C_{+}+sT=-tUL_{+}-R_{+}D_{+}.

Multiply the second equality on the right by H+H_{+}. Since C+H+=D+2C_{+}H_{+}=D_{+}^{2}, this gives

(tU)2T+sTH+=−(tUL++R+D+)H+.(tU)^{2}T+sTH_{+}=-(tUL_{+}+R_{+}D_{+})H_{+}.

Multiplication on the left by (tU)−2(tU)^{-2} now gives

(1−r)∥T∥≤bts(1+r),∥T∥≤btsr(1+r)1−r.(1-r)\lVert T\rVert\le b\frac{t}{s}(1+r),\qquad\lVert T\rVert\le b\frac{t}{s}\frac{r(1+r)}{1-r}.

The middle diagonal entry of J32=IJ_{3}^{2}=I is A2−I=−ZX−TYA^{2}-I=-ZX-TY. Thus,

∥A2−I∥≤b(∥X∥+∥T∥)≤b2ts2r(1+r)1−r.\lVert A^{2}-I\rVert\le b(\lVert X\rVert+\lVert T\rVert)\le b^{2}\frac{t}{s}\frac{2r(1+r)}{1-r}.

Taking middle diagonal entries in (7) also gives

B=U2A(U∗)2,∥A+B∥≤2bts.B=U^{2}A(U^{*})^{2},\qquad\lVert A+B\rVert\le2b\frac{t}{s}.

Apply Lemma 4.1 on Kn\mathcal{K}_{n} with W=U2W=U^{2}. It follows that

1≤∥A2−I∥+∥A+B2∥2≤b2(2r(1+r)1−rts+t2s2).1\le\lVert A^{2}-I\rVert+\left\lVert\frac{A+B}{2}\right\rVert^{2}\le b^{2}\left(\frac{2r(1+r)}{1-r}\frac{t}{s}+\frac{t^{2}}{s^{2}}\right).

This proves (5). If (cn/an)(c_{n}/a_{n}) were unbounded, the last expression would be less than one for a suitable large index. Hence (4) is necessary. □

The restriction to odd dimensions is needed. On C2\mathbb{C}^{2}, put

U=(100i),K=(0110).U=\begin{pmatrix}1&0\\0&i\end{pmatrix},\qquad K=\begin{pmatrix}0&1\\1&0\end{pmatrix}.

Then, KU2=−U2KKU^{2}=-U^{2}K, and the self-adjoint unitary diag⁡(−K,−UKU∗,K)\operatorname{diag}(-K,-UKU^{*},K) anticommutes with MaU,cIM_{aU,cI} for every a,c>0a,c>0. Taking orthogonal sums gives a bounded involution regardless of the sizes of cn/anc_{n}/a_{n} in this even-dimensional example.

For classical similarity, an unbounded ratio cn/anc_{n}/a_{n} can be compensated by the factor 1+qn21+q_{n}^{2}. This is the condition used in the following construction.

Proposition 4.3. Let (an)(a_{n}), (cn)(c_{n}) be bounded sequences of positive numbers. Let dn≥3d_{n}\ge3 be odd integers and choose qn∈Cq_{n}\in\mathbb{C} such that qndn=1q_{n}^{d_{n}}=1 and

sup⁡n≥1cnan∣1+qn2∣<∞.(9)\sup_{n\ge1}\frac{c_{n}}{a_{n}}\lvert1+q_{n}^{2}\rvert<\infty. \tag*{(9)}

There are unitaries UnU_{n} on Kn=Cdn\mathcal{K}_{n}=\mathbb{C}^{d_{n}} for which

D=⨁n≥1anUn,C=⨁n≥1cnIKnD=\bigoplus_{n\ge1}a_{n}U_{n},\qquad C=\bigoplus_{n\ge1}c_{n}I_{\mathcal{K}_{n}}

satisfy MD,C∼−MD,CM_{D,C}\sim-M_{D,C}. If the supremum in (9) is at most η>0\eta>0, there is an invertible SS with S−1MD,CS=−MD,CS^{-1}M_{D,C}S=-M_{D,C} and

max⁡{∥S∥,∥S−1∥}≤1+η.(10)\max\{\lVert S\rVert,\lVert S^{-1}\rVert\}\le1+\eta. \tag*{(10)}

If (3) holds and (cn/an)(c_{n}/a_{n}) is unbounded, no bounded involution conjugates MD,CM_{D,C} to −MD,C-M_{D,C}. If an→0a_{n}\to0 and cn→0c_{n}\to0, then MD,CM_{D,C} is compact.

Proof. On the standard orthonormal basis en,0,…,en,dn−1e_{n,0},\ldots,e_{n,d_n-1}, set

Unen,j=qnjen,j,Pnen,j=en,j+1mod⁡dn.U_n e_{n,j}=q_n^j e_{n,j},\qquad P_n e_{n,j}=e_{n,j+1\mathbin{\operatorname{mod}}d_n}.

Since qndn=1q_n^{d_n}=1, both operators are unitary and UnPn=qnPnUnU_nP_n=q_nP_nU_n. Put

P=⨁nPn,Q=⨁nqnIKn.P=\bigoplus_n P_n,\qquad Q=\bigoplus_n q_n I_{K_n}.

Then, P,QP,Q are unitary, DP=QPDDP=QPD, and CC commutes with D,P,QD,P,Q. Define

Xn=−cnan(1+qn2)Un∗Pn,X=⨁nXn.X_n=-\frac{c_n}{a_n}(1+q_n^2)U_n^*P_n,\qquad X=\bigoplus_n X_n.

By (9), XX is bounded and

DX=−C(I+Q2)P.DX=-C(I+Q^2)P.

Set

S=(Q2P000−QPX00P)(11)S= \begin{pmatrix} Q^2P & 0 & 0\\ 0 & -QP & X\\ 0 & 0 & P \end{pmatrix} \tag*{(11)}

The three diagonal blocks are unitary and XX is bounded, so the upper triangular operator SS is bounded and invertible. The only possibly nonzero entries of SMD,C+MD,CSSM_{D,C}+M_{D,C}S are

(SMD,C+MD,CS)12=Q2PD−DQP=0,(SM_{D,C}+M_{D,C}S)_{12}=Q^2PD-DQP=0,
(SMD,C+MD,CS)23=−QPD+DP=0,(SM_{D,C}+M_{D,C}S)_{23}=-QPD+DP=0,
(SMD,C+MD,CS)13=Q2PC+DX+CP=0.(SM_{D,C}+M_{D,C}S)_{13}=Q^2PC+DX+CP=0.

Thus, S−1MD,CS=−MD,CS^{-1}M_{D,C}S=-M_{D,C}.

For the norm estimate, let S0=diag⁡(Q2P,−QP,P)S_0=\operatorname{diag}(Q^2P,-QP,P) and E=S−S0E=S-S_0. The operator S0S_0 is unitary, ∥E∥=∥X∥≤η\lVert E\rVert=\lVert X\rVert\leq\eta, and (S0∗E)2=0(S_0^*E)^2=0. Hence

S−1=(I−S0∗E)S0∗,S^{-1}=(I-S_0^*E)S_0^*,

which proves (10). Theorem 4.2 gives the assertion about involutions. If an,cn→0a_n,c_n\to0, the finite-rank truncations of DD and CC converge in norm, so every entry of MD,CM_{D,C} is compact. □

We next take C=∣D∣αC=|D|^\alpha. In this case an involution exists exactly for α≥1\alpha\geq1.

Corollary 4.4. For each 0<r<10<r<1, there is a compact normal injective operator DD such that, for every α≥0\alpha\geq0,

Mα=(0D∣D∣α00D000)M_\alpha= \begin{pmatrix} 0 & D & |D|^\alpha\\ 0 & 0 & D\\ 0 & 0 & 0 \end{pmatrix}

satisfies Mα∼−MαM_\alpha\sim-M_\alpha.

Here ∣D∣0=I|D|^0=I. A bounded involution conjugates MαM_\alpha to −Mα-M_\alpha if and only if α≥1\alpha\geq1. Moreover, MαM_\alpha is compact if and only if α>0\alpha>0.

Proof. Choose odd integers dn≥3d_n\geq3 and dnd_n-th roots of unity qnq_n with ∣1+qn2∣≤rn|1+q_n^2|\leq r^n. Such choices are possible because the roots of unity of odd order are dense in the unit circle. Let UnU_n be as in Proposition 4.3, and set D=⨁nrnUnD=\bigoplus_n r^nU_n. Then, ∣D∣α=⨁nrαnI|D|^\alpha=\bigoplus_n r^{\alpha n}I. For an=rna_n=r^n and cn=rαnc_n=r^{\alpha n},

cnan∣1+qn2∣≤rαn≤1.\frac{c_n}{a_n}|1+q_n^2|\leq r^{\alpha n}\leq1.

Proposition 4.3 gives Mα∼−MαM_\alpha\sim-M_\alpha for every α≥0\alpha\geq0.

If 0≤α<10\leq\alpha<1, then an2/cn=r(2−α)na_n^2/c_n=r^{(2-\alpha)n} satisfies (4.3) with the same rr, whereas cn/an=r(α−1)nc_n/a_n=r^{(\alpha-1)n} is unbounded. Theorem 4.2 excludes an involution. If α≥1\alpha\geq1, the ratio cn/anc_n/a_n is bounded and (4.6) supplies one. Finally, for α>0\alpha>0 both DD and ∣D∣α|D|^\alpha are compact. For α=0\alpha=0, the upper-right entry is the identity on an infinite-dimensional space, so M0M_0 is not compact. □

For 0<α<10<\alpha<1, this gives a compact counterexample. We now impose summability of every positive power of the singular values. For p>0p>0, let Sp(H)S_p(\mathcal{H}) denote the Schatten pp-class of compact operators TT on H\mathcal{H} whose singular values (sj(T))j≥1(s_j(T))_{j\geq1} satisfy

∑j=1∞sj(T)p<∞\sum_{j=1}^{\infty}s_j(T)^p<\infty

For background on singular values and operator ideals, see [14].

Theorem 4.5. For every η>0\eta>0, there are M∈⋂p>0Sp(H)M\in\bigcap_{p>0}S_p(\mathcal{H}) and an invertible S∈B(H)S\in\mathcal{B}(\mathcal{H}) such that

M3=0≠M2,S−1MS=−M,max⁡{∥S∥,∥S−1∥}≤1+η,M^3=0\ne M^2,\qquad S^{-1}MS=-M,\qquad\max\{\lVert S\rVert,\lVert S^{-1}\rVert\}\leq1+\eta,

but no bounded involution JJ satisfies JMJ=−MJMJ=-M.

Proof. Fix 0<r<10<r<1 and set

an=rn2+n,cn=rn2(n≥1).(12)a_n=r^{n^2+n},\qquad c_n=r^{n^2}\qquad(n\geq1). \tag*{(12)}

Then,

an+1an=r2n+2,an+12/cn+1an2/cn=r2n+3,cnan=r−n.\frac{a_{n+1}}{a_n}=r^{2n+2},\qquad\frac{a_{n+1}^2/c_{n+1}}{a_n^2/c_n}=r^{2n+3},\qquad\frac{c_n}{a_n}=r^{-n}.

Thus, (4.3) holds and (cn/an)(c_n/a_n) is unbounded.

We choose odd dn≥3d_n\geq3 and dnd_n-th roots qnq_n so that

∣1+qn2∣≤ηrn,dn≤Lηr−n,\lvert1+q_n^2\rvert\leq\eta r^n,\qquad d_n\leq L_\eta r^{-n},

where LηL_\eta is independent of nn.

Now, construct DD, CC and M=MD,CM=M_{D,C} by Proposition 4.3. The coefficients tend to zero, so MM is compact. Since

cnan∣1+qn2∣≤η,\frac{c_n}{a_n}\lvert1+q_n^2\rvert\leq\eta,

that proposition gives SS with the asserted norm bounds. Theorem 4.2 excludes every bounded involution, and M3=0≠M2M^3=0\ne M^2 follows from the injectivity of DD.

Regrouping the orthogonal summands writes MM as a direct sum of matrices MnM_n on Kn3\mathcal{K}_n^3. The range of each MnM_n is Kn⊕Kn⊕{0}\mathcal{K}_n\oplus\mathcal{K}_n\oplus\{0\}, so its rank is 2dn2d_n. Also,

∥Mn∥≤an+cn≤2cn.\lVert M_n\rVert\leq a_n+c_n\leq2c_n.

Consequently, for every p>0p>0,

∑jsj(M)p=∑n∑jsj(Mn)p≤2p+1Lη∑n≥1rpn2−n<∞.\sum_j s_j(M)^p=\sum_n\sum_j s_j(M_n)^p\leq2^{p+1}L_\eta\sum_{n\geq1}r^{pn^2-n}<\infty.

The series converges for every p>0p>0. This completes the proof. □

Declarations

Competing interests. The author declares that he has no competing interests.

Funding. The author received no funding.

Data availability. No datasets were generated during the current study.

Department of Mathematics, Faculty of Sciences of Monastir, University of Monastir, 5019 Monastir, Tunisia

Email address: [email protected]

Email address: [email protected]

References

  1. [1]M. A. Aouichaoui, All unitary operators are bicycles, Proc. Amer. Math. Soc. 154 (2026), no. 12, 5325–5330.DOI
  2. [2]M. A. Aouichaoui, The set of commutators of quadratic operators and its norm closure, Proc. Amer. Math. Soc., doi:10.1090/proc/17996.DOI
  3. [3]A. Brown and C. Pearcy, Structure of commutators of operators, Ann. of Math. (2) 82 (1965), no. 1, 112–127.DOI
  4. [4]K. R. Davidson and L. W. Marcoux, Linear spans of unitary and similarity orbits of a Hilbert space operator, J. Operator Theory 52 (2004), no. 1, 113–132.
  5. [5]R. Drnovšek, On commutators of idempotents, Linear Multilinear Algebra 73 (2025), no. 4, 649–654.DOI
  6. [6]R. Drnovšek, H. Radjavi and P. Rosenthal, A characterization of commutators of idempotents, Linear Algebra Appl. 347 (2002), nos. 1–3, 91–99.DOI
  7. [7]C. K. Fong and A. R. Sourour, The semigroup generated by a similarity orbit or a unitary orbit of an operator, Proc. Amer. Math. Soc. 131 (2003), no. 10, 3203–3210.DOI
  8. [8]P. R. Halmos, A Hilbert Space Problem Book, 2nd ed., Grad. Texts in Math., vol. 19, Springer-Verlag, New York, 1982.
  9. [9]L. W. Marcoux, M. Omladič, H. Radjavi and Y. Zhang, All unitary operators are tricycles, and most are bicycles, Bull. Lond. Math. Soc. 58 (2026), no. 5, e70378.DOI
  10. [10]L. W. Marcoux, H. Radjavi and Y. Zhang, Around the closure of the set of commutators of idempotents in B(H): Biquasitrangularity and factorisation, J. Funct. Anal. 284 (2023), no. 8, Paper No. 109854, 45 pp.DOI
  11. [11]L. W. Marcoux, H. Radjavi and Y. Zhang, Around the closures of the set of commutators and the set of differences of idempotent elements of B(H), J. Operator Theory 91 (2024), no. 1, 97–124.
  12. [12]C. Pearcy and D. Topping, Sums of small numbers of idempotents, Michigan Math. J. 14 (1967), no. 4, 453–465.DOI
  13. [13]V. I. Rabanovich and Yu. S. Samoilenko, When a sum of idempotents or projections is a multiple of the identity, Funct. Anal. Appl. 34 (2000), no. 4, 311–313.DOI
  14. [14]B. Simon, Trace Ideals and Their Applications, 2nd ed., Math. Surveys Monogr., vol. 120, American Mathematical Society, Providence, RI, 2005.
  15. [15]B. Sz.-Nagy, C. Foias, H. Bercovici and L. Kérchy, Harmonic Analysis of Operators on Hilbert Space, 2nd ed., Universitext, Springer, New York, 2010.
  16. [16]P. Y. Wu, The operator factorization problems, Linear Algebra Appl. 117 (1989), 35–63.DOI

Paper details

vlnc:W_un6vnts8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents