arXiv:1908.07029v2 [math.FA] 21 Aug 2019 identify ieroeaosatn on acting operators linear pcrmof spectrum omloeaos(..toeoperators those (i.e. operators normal rils[,1]poien ee hn8 uhcniin ntematrix the in conditions norm such a 89 than fewer no provide 14] [6, articles ra elo okdn opoueatraecniin nagive a on conditions alternate produce to done work of deal great a ti ayt e htif that see to easy is it to equivalent either o all for n prtr(..seta)nrsaebt ntrl-nain nor unitarily-invariant both are norms spectral) (i.e. operator and hn N. 1014 1717 1011 n h hn Scholar China Prov the Jilin and of 11201171) Scientists 11671167, Young 11601104, for (No.: Foundation China Science Natural the by spectrum. Lavrentieff values, singular Let mns h otsuidadbs nesodcaso operator of class understood best and studied most the Amongst h rsn ril dsytaohrcniint hsseigyend seemingly this to condition another yet adds article present The ..Mroxsrsac sspotdi atb SR Cnd) Y (Canada); NSERC by part in supported is research Marcoux’s L.W. e od n phrases. and words Key 2010 Date H uut2,2019. 22, August : X ahmtc ujc Classification. Subject Mathematics N epc ote( the to respect htif that then omloeao hs pcrmnihrdsoncstepaen plane the disconnects neither spectrum whose operator normal nain omon norm invariant oml hl ftesetu of spectrum the if while normal, operator mle that operator implies an extent what to examine smti eaieto relative isometric n w ieroperators linear two and l one ieroeaosatn on acting operators linear bounded all Abstract. k · k B eacmlx eaal ibr pc n eoeby denote and space Hilbert separable complex, a be PRTR HC R OYOILYISOMETRIC POLYNOMIALLY ARE WHICH OPERATORS r opc prtr with operators compact are B ∈ ( H A u A ihtealgebra the with ) ,N A, snra.W hnetn hsrsl oteifiiedmninlsett infinite-dimensional the to result this extend then We normal. is on ( H B B ∈ n o l ntr operators unitary all for and ) B uhthat such A Let B ∈ ( ( H H T sisl oml oeepiil,w rtso htif that show first we explicitly, More normal. itself is ,p c, ARN .MROXADYAHN ZHANG YUANHANG AND MARCOUX W. LAURENT ssi obe to said is ) ( H ). H en oml ra es ml that imply least at or normal, being -omsuidb hn iadT,then Tu, and Li Chan, by studied )-norm M T oyoilyioerc omloeaos ntrl-nain norm, unitarily-invariant operators, normal isometric, polynomially r oyoilyioercrltv oteoeao omand norm operator the to relative isometric polynomially are ) eacmlx eaal ibr pc and space Hilbert separable complex, a be k · k n B B ∈ ( C OANRA OPERATOR NORMAL A TO H and ,if ), u A When . ( if H and N M k ,N A, and ) k p N X n ( r oyoilyioerc ial,w hwta if that show we Finally, isometric. polynomially are N A ( B k unitarily-invariant C oml n if and normal, 1. ) snto hsfr,te hr laseit non-normal a exists always there then form, this of not is u k ∈ A in of ) 71,1A0 15A21. 15A60, 47B15, u = k N H Introduction M en oyoilyioerct omloperator normal a to isometric polynomially being x T = B H n k ( B ∈ ie ntrl-nain norm unitarily-invariant a Given . ( n sfiiedmninlwt dim with finite-dimensional is k H UX k C p ,w hl a that say shall we ), × ( r oyoilyioercand isometric polynomially are ) = B 1 ( k H n ) k k u u T o which for ) ,V U, ope arcs ednt by denote We matrices. complex A = o l polynomials all for ∗ x and k k XV B ∈ o all for N k hpCouncil. ship r oyoilyioercwith isometric polynomially are if u ( T H A ne(o:210008H,NS of NNSF 20190103028JH), (No.: ince NN .Freape h Frobenius the example, For ). saannormal. again is x snra 2 1.Frexample, For 11]. [2, normal is B A B ( H hn’ eerhi supported is research Zhang’s .H. rhsitro,then interior, has or H ∈ H ( and ∗ eoeteagbaof algebra the denote ) H k · k = son ms h ler fbounded of algebra the ) p then , nti ae,we paper, this In . B N n in s u are ∗ setting. T sayunitarily- any is esls.Rcl that Recall list. less N n yshowing by ing k · k M B B ∈ H N .Teehsbeen has There ). polynomially T ( n u H snormal, is ( = snra.The normal. is ( on C stestof set the is ) H N ). A < n B hc are which ) A and ( sa is ( H ,p c, σ N is ) ( ∞ A )-norm, the ) we , Given an operator T ( ), and given a unitarily-invariant norm on ( ), we are ∈ B H k·ku B H interested in determining to what extent the values of p(T ) u, p a polynomial determine the normality of T . To make this notion precise, we introducek k the following definition.

1.1. Definition. Let A, B ( ), and let u be a unitarily-invariant norm on ( ). We say that A and B are polynomially∈ B H isometrick·k relative to if B H k·ku p(A) = p(B) for all polynomials p. k ku k ku

We remark that when is finite-dimensional with dim = n < , it follows from the Cayley-Hamilton TheoremH that one need only consider polynomialsH of∞ degree at most n 1. We should mention that variants of this notion have at times been referred to by saying that− “A and B have the same norm behaviour”. We feel that the current terminology is more precise. 1.2. The underlying theme of our investigation is that if A and B are linear operators which are polynomially isometric with respect to a given norm, then A and B share many common properties. For example, in the setting of matrices, it is easily seen that if A, B M (C) are ∈ n polynomially isometric relative to a unitarily-invariant norm u, then A and B share a common minimal polynomial, and in particular, σ(A)= σ(B).k·k In the case of the Frobenius and operator (i.e. spectral) norms, Greenbaum and Trefethen [13] studied the relationship between polynomial and resolvent growth for a pair A, B Mn(C), showing in particular that if A and B are polynomially isometric relative to any∈ matrix norm, then A and B share common . Later, Viswanath and Trefethen [22] showed that if A Mn(C) is unitarily equivalent to A1 A2 where σ(A1) σ(A2) = ∅, and if B M (C∈) is polynomially isometric to A, then B⊕also admits a direct∩ sum decomposition ∈ n B = B1 B2. As yet another example, in [9, Theorem 2.1], [10, Theorem 2.1], it was shown that if A⊕ M (C) is an upper triangular Toeplitz matrix with nonzero superdiagonal and ∈ n B Mn(C) is polynomially isometric to A with respect to either the Frobenius or operator norm,∈ then A and B are unitarily equivalent. The main question we study is the following.

Question: Let A, N ( ) be two operators with N normal, and suppose that ( ) is ∈ B H B H equipped with a unitarily-invariant norm u. If A and N are polynomially isometric with respect to , must A be normal? k·k k·ku

In the matrix setting, some positive answers have already been obtained. For Mn(C) equipped with the Frobenius norm, a positive answer to this question was given in [11], while the case of Mn(C) equipped with the usual operator norm was subsequently handled in [2]. In Section 2 below, we show that for any unitarily-invariant norm on M (C), if k·ku n A Mn(C) is polynomially isometric to a normal matrix N, then A is normal. Furthermore, if the∈ norm is able to distinguish between projections of different rank, then A and N must in fact be unitarily equivalent. From Section 3 on, we assume that is an infinite-dimensional, complex, separable . The most important unitarily-invariantH norm in the infinite-dimensional setting is the usual operator norm on ( ). In Section 3 we consider the case where A, N ( ) are k·k B H 2 ∈ B H polynomially isometric relative to this norm, and N is normal. A compact subset K C is said to be Lavrientiev if C K is connected, and if the interior Ko of K is empty.⊆ We prove that if σ(N) is Lavrentieff,\ then A is normal, while if σ(N) is not Lavrentieff, then there always exists a non- B such that B and N are polynomially isometric relative to the operator norm. In Section 4, we extend above result to compact operators A and N (with N normal) which are polynomially isometric with respect to the (c,p)-norms introduced by Chan, Li and Tu in [3]. Since the operator norm on ( ) is an example of a (c,p)-norm, we see from above that the result does not extend to allB H pairs of operators A and N in ( ) with N normal. B H We conclude with a few remarks concerning polynomial isometry when restricted to poly- nomials without constant term.

2. Unitarly-invariant norms on Mn(C) We first deal with the case where is finite-dimensional. The reader is referred to [7, 18, 17] for more information about theH relationship between unitarily-invariant norms and symmetric gauge functions.

2.1. Definition. Let x = (x , , x ),y = (y , ,y ) Rn. We shall write x y to 1 ··· n 1 ··· n ∈ ≤ mean that xi yi(i = 1, , n). Let us also denote by x˜1 , x˜n the numbers x1, , xn arranged in non-ascending≤ ···order of magnitude, and let y˜ ···, y˜ be defined analogously.··· If 1 ··· n x˜ + +˜x y˜ + +˜y (k =1, ..., n), 1 ··· k ≤ 1 ··· k we shall write x 4 y. Next, we need to recall the following fact which will be used later.

2.2. Lemma. Let p,q,k N, and 0

Proof. Set q p r r r yˆ =(0, , 0; 1+ 1 , , 1+ 1 ;1+ 1 ,, 1+ r , , 1+ r ). ··· 2q ··· 2q 2 2 ··· k z }| { Then z }| { yˆ 4 y, and q k p r r r r yˆ z := (0, , 0; 1+ 1 , , 1+ 1 ; 1+ 1 , , 1+ 1 ). ≥ ··· 2q ··· 2q 2q ··· 2q z }| { z }|3 { z }| { Therefore, by [7, Theorem 4] (also see [18, Theorem 1]), for a fixed symmetric gauge function Φ in p + q + k variables, r Φ(y) Φ(ˆy) Φ(z)=(1+ 1 )Φ(x) > Φ(x). ≥ ≥ 2q  For convenience, we formulate a variant of [22, Proposition 1.3] as the next lemma, which will prove very useful.

2.3. Lemma. Let P M (C) be a projection, Q M (C) be an idempotent, and be ∈ n ∈ n k·ku a unitarily-invariant norm of Mn(C). If P u = Q u and In P u = In Q u, then Q is a projection. k k k k k − k k − k Proof. Suppose that rank(P ) = r, rank(Q) = s. If s = 0, then Q = 0, while if s = n, then rank(In Q) = 0, that is, Q = In. We therefore− assume that 1 s n 1. Relative to the decomposition Cn = ran Q (ran Q)⊥, we may write ≤ ≤ − ⊕ I Y Q = s , 0 0  n−s Note that either s r or n s > n r. ≥ − − (1) When s r, we note that since the singular values of Q are same (counting mul- ≥ ∗ ∗ tiplicity) as the singular values of Q , it follows that Q u = Q u. If Y = 0, then – by Lemma 2.2 and the one-to-one correspondencek betweenk unitarily-invariantk k 6 norms and symmetric gauge functions due to Von Neumann (see [17, Theorem 1.1]) – Q u > P u. (2) Whenk k n ks >k n r, we consider − − 0 Y In Q = − . − 0 In−s   In an argument similar to the one above, In Q u > In P u by Lemma 2.2 unless Y = 0. k − k k − k Therefore, Y = 0, and so Q is a projection.  Suppose that P and Q M (C) are two projections of equal rank. Clearly P and Q ∈ n are unitarily similar, and thus if u is any unitarily-invariant norm on Mn(C), then P = Q . In general, however,k·k it often happens that P = Q even if the ranks of k ku k ku k ku k ku P and Q differ; one need only consider the operator norm on Mn(C). Let us agree to say that a unitarily-invariant norm separates projections by rank if, whenever P and k·ku Q are projections and rank P = rank Q, then P u = Q u. This happens for a large family of unitarily-invariant norms, including,6 for example,k k 6 thek kp-norms 1 n p p T p := sk(T ) , T Mn(C), k k ! ∈ Xk=1 where s1(T ) s2(T ) sn(T ) 0 are the singular numbers of T . (See Section 4 for the definition of the≥ singular≥··· numbers≥ of T ( ).) ∈ B H4 The following theorem generalises the equivalence of conditions (i) and (vii) in the main Theorem of [11] as well as [2, Corollary 1], and it adds yet one more equivalence to the 89 other characterisations of normal matrices that appear in [14] and [6].

2.4. Theorem. Let n N and be a unitarily-invariant norm on M (C). Let N be ∈ k·ku n an n n normal complex matrix and A Mn(C). Suppose that A and N are polynomially isometric× relative to . Then A is normal.∈ k·ku If, furthermore, we suppose that u separates projections by rank, then A and N are unitarily similar. k·k Proof. Let µ be the minimal polynomial of N. Then σ(N) = λ : µ (λ)=0 . Since N { N } µN (N) = 0, it follows that µN (A) = 0, and thus µ (σ(A)) = σ(µ (A)) = 0 . N N { } In particular, σ(A) σ(N). By symmetry, ⊂ σ(N)= σ(A) := λ , ,λ . { 1 ··· m} If m = 1, then N = λI, and since A λI = N λI =0, k − k k − k we have A = λI. Next, suppose that m 2. Define ≥ z λj pi(z)= − , 1 i m. λi λj ≤ ≤ 1≤jY≤m,j6=i − Set

Pi := pi(N), Qi := pi(A), it follows that P = Q , and I P = I Q . k iku k iku k n − iku k n − iku From the basic properties of the continuous for normal matrices, we find P is a projection; that is P 2 P = 0 and P = P ∗, 1 i m; indeed, P is the • i i − i i i ≤ ≤ i spectral projection for N corresponding to the eigenvalue λi; PiPj =0, 1 i = j m; • m P I≤= 0;6 ≤ • i=1 i − Therefore, P 2 2 Qi Qi u = Pi Pi u = 0, so that Qi is an idempotent, 1 i m; in fact, the • kcondition− k abovek that− Pk = Q and I P = I Q≤ ≤, combined with k iku k iku k n − iki k n − iku Lemma 2.3, yields that each Qi is a projection. QiQj u = PiPj u = 0, and thus QiQj =01 i = j m; • k m kQ kI =k m P I = 0, and therefore≤ 6 ≤ m Q = I. • k i=1 i − ku k i=1 i − ku i=1 i Set P P m P z λj p(z) := z λi − . − λi λj i=1 1≤j≤m,j6=i − ! X 5 Y Then m m A λ Q = p(A) = p(N) = N λ P =0, k − i iku k ku k ku k − i iku i=1 i=1 X X and thus m

A = λiQi is normal. i=1 X The last statement follows trivially from the fact that P = Q implies that rank P = k iku k iku i rank Qi for all 1 i m. ≤ ≤ 

2.5. Remark. If we replace the hypothesis that u separates projections by rank by the hypothesis that A and N share a common characteristick·k polynomial, then the fact that A must be unitarily similar to N still follows, as is easy to check. In the absence of both hypotheses, multiplicity of eigenvalues becomes an issue. For example, consider the unitarily- invariant norm T := s (T )+ s (T ), T M (C), k ku 1 2 ∈ n where as always, s1(T ) s2(T ) sn(T ) 0 are the singular numbers of T . Let P = diag (1, 1, 1, 0, 0), Q≥= diag (1≥, 1, ···0, 0, 0) be≥ diagonal projections of ranks 3 and 2 re- spectively. For any polynomial p we have that p(P ) = diag (p(1),p(1),p(1),p(0),p(0)), while p(Q) = diag (p(1),p(1),p(0),p(0),p(0)), and so p(P ) u = p(Q) u. Clearly P is not unitar- ily equivalent to Q. k k k k

3. The infinite-dimensional setting: the operator norm 3.1. In [16], Lavrentieff established the following result: 3.2. Theorem. Let K C be a compact set, and denote by (K) the set of all polynomials with complex coefficients.⊆ The following conditions are equivalent:P (a) The family (K) is uniformly dense in (K) := f : K C : f is continuous . (b) The set C PK is connected and the interiorC Ko of{ K is→ empty. } \ In view of this result, we refer to a compact set K satisfying condition (b) as a Lavrentieff set. The relevance of this result for us is that as a consequence of the for normal operators (see, e.g. [8, Theorem 2.7 (iii) (iv)] ), if N ( ) is a normal operator, then N ∗ belongs to the norm-closed unital algebra⇐⇒ Alg(N) generated∈ B H by N if and only if σ(N) is a Lavrentieff set. In other words, Alg(N) = C∗(N) if and only if σ(N) is a Lavrentieff set. Our first result is general enough to hold in the context of C∗-algebras. 3.3. Theorem. Let A, B be unital C∗-algebras, N A be a with Lavrentieff spectrum and A B. Suppose that N and A are polynomially∈ isometric. Then A is normal. ∈ Proof. Let Alg(N) (resp. Alg(A)) denote the unital commutative algebra generated by N (resp. by A). Since σ(N) is a Lavrentieff set, Alg(N)= C∗(N), as noted above. Define Φ : p(N) p(A), p a polynomial. 0 7→ Then Φ0 is a unital isometric isomorphism, and can thus be uniquely extended to an isometric isomorphism Φ : C∗(N) B whose range is Alg(A). It is known, however (see, e.g. [1, → 6 Proposition A.5.8]) that a contractive homomorphism between C∗-algebras is necessarily a ∗-homomorphism, and so Alg(A) must also be a C∗-algebra, being isometrically isomorphic to C∗(N). In particular, since Alg(A) is abelian and A∗ Alg(A), A must be normal. ∈  3.4. Remark. The result of Brooks and Condori [2, Corollary 3] is a finite-dimensional version of Theorem 3.3, as a finite subset of C is obviously a Lavrentieff set.

3.5. Definition. Let K be a compact subset of C, define the polynomially convex hull of K to be the set K given by

K := z C : p(z) p K for every polynomial p , b { ∈ | |≤k k } where b p := max p(z) : z K . k kK {| | ∈ } In fact, C K is exactly the unbounded component of C K (see [4, p206, Proposition 5.3]). \ \ Recall thatb given an element a of a , the of a is spr(a) := max λ : λ σ(a) . A {| | ∈ } 3.6. Theorem. Let N B( ) be normal. The following are equivalent: ∈ H (a) σ(N) is a Lavrentieff set. (b) If A B( ) is polynomially isometric to N, then A is normal. ∈ H Proof. (a) implies (b): This is an immediate consequence of Theorem 3.3. [ (b) implies (a): Suppose that σ(N) is not a Lavrentieff set. Let L = σ(N). Then Lo = . By translating both A and N by the same scalar operator, we may assume without6 ∅ loss of generality that 0 Lo. ∈ Choose ε > 0, such that K := B(0,ε) Lo. Denote by S the unilateral shift operator on l2. Set = ℓ2. Define ⊂ K H ⊕ T := N εS B( ). ⊕ ∈ K As S is subnormal, given any polynomial p, we have that p(εS) is subnormal, hence by [15, Proposition 6.10], p(εS) = spr(p(εS)). k k By the Maximal Modulus Principle and the Spectral Mapping Theorem, p(εS) = spr(p(εS)) k k = max p(λ) : λ σ(εS) {| | ∈ } max p(µ) : µ σ(N) ≤ {| | ∈ } = spr(p(N)) = p(N) . k k It follows that N and T are polynomially isometric. Clearly, however, T is not normal. Since is isomorphic to ℓ2, we see that T is unitarily equivalent to an A B( ) whichH must also be non-normalH⊕ and polynomially isometric to N. ∈ H  7 4. On the (c,p) norms of Chan, Li and Tu. 4.1. We now wish to extend the results of Section 3 to a family of unitarily-invariant norms on ( ) first introduced by Chan, Li and Tu in [3]. Before doing so, we first remind the readerB H of the definition of singular values for not-necessarily compact operators. We denote by ( ) the ideal of compact operators in ( ). For K in ( ), the sin- K H ∞ B H K H gular numbers (sn(K))n=1 of K are defined to be the eigenvalues (repeated according to 1 multiplicity) of K =(K∗K) 2 . This notion was extended to non-compact operators in [12] | | 1 as follows. Let T ( ) be an arbitrary operator. Then T := (T ∗T ) 2 is a positive ∈ B H | | operator. Let α := max λ : λ σe( T ) , where σe(X) is the of X ( ), that is, the spectrum{ of∈ the| image| } π(X) of X ( ) under the canonical map π :∈ B( H) ( )/ ( ), the Calkin algebra. It is a consequence∈ B H of the work of Wolf [23] thatB ifHQ →:= BE H((α,K HT ]) denotes the spectral projection of T corresponding to the set |T | k k | | (α, T ] R, then the spectrum of the compression (Q T Q)QH of T to the range of Q is thek closurek ⊆ of a sequence (perhaps finite) of isolated eigenvalues| | of finite| | multiplicity of T tending to α. Thus, | |

if rank Q = 0, we set sn(T )= α for all n 1. • If 0 < rank Q = m < , we denote by ≥s (T ) m the eigenvalues of (Q T Q) , • ∞ { n }n=1 | | QH repeated according to multiplicity, and for n m + 1, we set sn(T )= α. ∞ ≥ If rank Q = , we denote by sn(T ) n=1 the eigenvalues of (Q T Q)QH, written in • non-increasing∞ order repeated according{ } to their multiplicity. | | We refer the interested reader to [12] and to [19] for more details regarding the singular numbers of operators. 4.2. We are now in a position to describe the norms of Chan, Li and Tu. Let 1 p< , ≤ ∞ and let n 1 be an integer. Suppose that we are given real numbers c1 c2 cn > 0. For T ≥( ), we define the (c,p)-norm of T to be ≥ ≥···≥ ∈ B H n 1 T =( c s (T )p) p . k kc,p j j j=1 X In [3], they proved (amongst other things) that these are indeed unitarily-invariant norms on ( ), and that for all T ( ), B H ∈ B H n 1/p c1/p T T c T . 1 k k≤k kc,p ≤ j k k j=1 ! X This family of norms includes

the operator norm T = s1(T ); • k k n the Ky-Fan n-norm T [n] := j=1 sj(T ); • k k n the weighted Ky-Fan n-norm T := c s (T ), where c = 1 and c c P [c,n] j=1 j j 1 j j+1 • for all 1 j n 1, as well as k k ≥ ≤ ≤ − P 1/p the [p, n]-singular norm T := n sp(T ) . • k k[p,n] j=1 j We also mention the useful fact that if k  n, then the (c,p)-norm defines a unitarily- ≥P invariant norm on Mk(C). Our main result in this section is the following. 8 4.3. Theorem. Let 1 p < , n N, c1 c2 cn > 0. Suppose that A and N are compact operators, and≤ that∞N is∈ normal.≥ If A≥···≥and N are polynomially isometric with respect to , then A is normal. k·kc,p ∞ Proof. Let en n=1 be an orthonormal basis for . Set K := σ(N) σ(A). Since N, A are compact operators,{ } K is countable, and if it is infinite,H it consists of∪ a sequence converging to 0 (along with 0 itself). Case One. K is infinite. Let us write K = λ : i N 0 , where λ λ > 0, and λ = λ whenever { i ∈ }∪{ } | 1|≥| 2|≥··· i 6 j i = j, and let us choose 1 n1 < n2 < < nk < such that λnk > λnk+1 . 6 For each k N, choose 0≤<ε such that··· ··· | | | | ∈ k (i) for i = j 1, , nk , B(λi, 2εk) B(λj, 2εk)= ∅; and (ii) λ : 6 j >∈{ n ···B(0, }λ 3ε ). ∩ { j k} ⊂ | nk | − k Define E =( B(λ ,ε )) B(0, λ 2ε ). k i k | nk | − k 1≤i≤n [ k [ Then Ek is a compact set of C and C Ek is connected. Define \ ∆k = B(λi,εk), 1≤i≤n [ k and denote by χ∆k be the characteristic function of ∆k. Observe that χ∆k is holomorphic, hence continuous on Ek. Let

Pk = EN (∆k), (resp. Qk = EA(∆k))

be the Riesz idempotent for N (resp. for A) corresponding to ∆k. Since N is normal, Pk is in fact an orthogonal projection. As N and A are compact, rank P < and rank Q < . k ∞ k ∞ For k N, consider the following function ∈ 1 , z B(λ , 2ε ); z 1≤i≤nk i k fk(z)= ∈ 0, z B(0, λ 3ε ).  ∈ S | nk | − k Then fk is also holomorphic on a neighbourhood of K. (k) By Mergelyan’s theorem (see [21, Theorem 20.5]), we can choose polynomials tj , j 1, such that ≥ (k) lim tj fk Ek =0. j→∞k − k Thus, by the Riesz Functional Calculus [4, p201. 4.7(e)], (k) (k) lim tj (N) fk(N) =0, and lim tj (A) fk(A) =0, j→∞k − k j→∞k − k where denotes the operator norm in ( ). By [3, Proposition 2.4.(a)], k·k B H (k) (k) lim tj (N)N fk(N)N c,p =0, and lim tj (A)A fk(A)A c,p =0, j→∞k − k j→∞k − k Therefore, (k) (k) Qk c,p = fk(A)A c,p = lim tj (A)A c,p = lim tj (N)N c,p = fk(N)N c,p = Pk c,p. k k k k j→∞k k j→∞k k k k k k 9 Similarly, I Qk c,p = I Pk c,p. k ⊥ − k k − k Denote Hk = Ran Pk (ker Qk) Ran Qk span e1, e2,...,eαk , where αk := rank Pk + dim(ker Q )⊥ + rank Q ∨+3k. Set R∨ be the∨ (finite{ rank) orthogonal} projection of onto k k k H k. Clearly k is reducing for both Pk and Qk, and thus Pk = RkPkRk, Qk = RkQkRk. HMoreover, it isH readily verified that R P = I P = I Q = R Q . k k − kkc,p k − kkc,p k − kkc,p k k − kkc,p Using the fact that the (c,p)-norm on ( ) is unitarily-invariant, it now follows Lemma 2.3 B Hk that Qk is a projection. Define

Nk := PkN = PkNPk, Ak := QkA = QkAQk.

We now show that Ak and Nk are in fact polynomially isometric. Let q be an arbitrary polynomial, and write q = q(0) + q1, where q1(0) = 0. Observe that using [3, Proposition 2.4.(a)] and the fact that A and N are polynomially isometric, we obtain

q(N ) = (q(0))I + P q (N) k k kc,p k k 1 kc,p (k) = lim q(0)I + tj (N)Nq1(N) c,p j→∞k k (k) = lim q(0)I + tj (A)Aq1(A) c,p j→∞k k = q(0)I + Q q (A) k k 1 kc,p = q(A ) . k k kc,p

Notice also that q(N ) = q(0)R + q (N ) , q(A ) = q(0)R + q (A ) . k k kc,p k k 1 k kc,p k k kc,p k k 1 k kc,p We may view each Nk as an operator on k, and similarly view Ak as an operator on k. Once again, the fact that the (c,p)-normH on ( ) is unitarily-invariant, combined withH B Hk Theorem 2.4 implies that Ak is normal, and σ(N )= σ(A )= λ , ,λ 0 . k k { 1 ··· nk }∪{ } ∞ Note that Qk k=1 is an increasing sequence of projections (of finite rank), and thus if we define { } Q = sup Qk : k N = SOT lim Qk, { ∈ } − k→∞ then Q is also a projection. Since QkA = AQk, for all k N, and since Ak = QkQQk is normal for all k, we see that AQ = QA and that QAQ is normal.∈ Similarly, P ∞ is an increasing sequence of projections (of finite rank), and so { k}k=1 P = sup Qk : k N = SOT lim Pk, { ∈ } − k→∞ also defines a projection. As in the previous case, the fact that PkN = NPk for all k N implies that P N = NP . ∈ Write PNP 0 Ran P QAQ 0 Ran Q N = , A = . 0 0 Ran P ⊥ 0 Z Ran Q⊥   10   If Z = 0, we can choose k∗ N sufficiently large so that 6 ∈ n p 1 p ( c ) λ ∗ < ( Z ) . j | nk +1| 3k k j=1 X it follows that

⊥ ⊥ Z = Q (A A ∗ )Q k kc,p k − k kc,p A Q ∗ A ≤k − k kc,p (k∗) = lim A tj (A)AA c,p j→∞k − k (k∗) = lim N tj (N)NN c,p j→∞k − k = N P ∗ N k − k kc,p = N N ∗ k − k k(c,p) n 1 p p (( c ) λ ∗ ) ≤ j | nk +1| j=1 X 1 < Z . 3k kc,p This obvious contradiction implies, therefore, that Z = 0. Since QAQ is normal and Z = 0, A is normal. Case Two. K is finite. The argument here is similar to that of Case One. The main difference is that instead of considering an infinite sequence of projections Pk and Qk, it suffices to consider a single pair of projections P and Q, defined as follows.

After writing K = λ1,λ2,...,λk0 0 , with λ1 λ2 λk0 > 0 and λi = λj if i = j, we only choose{ one value of }∪{ε> 0} such that| |≥| |≥···≥| | 6 6 (i) for i = j 1, 2,...,k0 , B(λi, 2ε) B(λj, 2ε)= ∅; and (ii) B(0, 26 ε) ∈{B(λ , 2ε)= ∅}, 1 i k∩. ∩ i ≤ ≤ 0 We then consider ∆ := 1≤i≤k0 B(λi,ε). Arguing as in Case One, we see that if P and Q are the Riesz projections for∪ N and A corresponding to the set ∆, then P is a projection (as N is normal) and we argue as in Case One that Q is also selfadjoint. The remainder of the proof follows as above. 

4.4. Remark. It is clear from our work in Section 3 that this result does not extend directly to general operators on ( ), since the result fails for the operator norm, which we have seen to be an example of aB (c,pH )-norm.

5. Polynomials without constant term 5.1. It is tempting to try to extend the results above to the classes of compact operators lying in the various Schatten p-ideals. Recall that with the singular numbers of a compact 11 operator defined as in Section 4, for 1 p< , the Schatten p-class ( ) is defined as ≤ ∞ Cp H ∞ ( ) := K ( ): K is compact and (s (K))p < , Cp H { ∈ B H n ∞} n=1 X and it is equipped with the unitarily-invariant norm

∞ 1/p K := (s (K))p . k kp n n=1 ! X Each p( ) is well-known to be an ideal of ( ). Because of this, the identity operator I ( C) forH any 1 p< , and so at best, whenB H considering polynomial isometry of two 6∈ Cp H ≤ ∞ elements A and B of p( ), we must restrict our attention to those polynomials q which satisfy q(0) = 0. C H We now produce an example of two operators A and N in 2( ) (the ideal of Hilbert- Schmidt operators on ) such that N is normal, A and N areC polynomiallyH isometric (for polynomials which vanishH at 0), and yet A is not normal.

∞ 5.2. Example. Let ek k=1 be an orthonormal basis for , and for x, y , denote by ∗ { } ∗ H ∈ H∗ ∗ x y the rank-one operator x y (z) = z,y x, z . Define N = e1 e1 + e2 e2, ⊗ ∗ ∗ ⊗ h i ∈ H ⊗ ⊗ A = e1 e1 + e1 e2. Then N is a projection and A is an idempotent but it is not a projection.⊗ In particular,⊗ A is not normal. Nevertheless, we find that

1 1 N = tr(N ∗N) 2 = √2, A = tr(A∗A) 2 = √2. k k2 k k2 Hence for every polynomial f(z)= a zn + + a z, n ··· 1 n n n n f(N) = ( a )N = √2 a , f(A) = ( a )A = √2 a . k k2 k i k2 | i| k k2 k i k2 | i| i=1 i=1 i=1 i=1 X X X X That is, A and N are polynomially isometric relative to 2, considering only polynomials which vanish at 0. k·k

5.3. Remark. It is clear that we did not need to consider the infinite-dimensional Hilbert 2 2 1/2 space setting for the above example. Indeed, if T 2 = ((s1(T )) +(s2(T )) ) denotes the k k 2 Frobenius norm on M2(C), and if e1, e2 is an orthonormal basis for C , then the operators N and A defined as in the above example{ } again show that Theorem 2.4 fails if we consider only polynomials without constant term. The problem lies in the fact that in the absence of the identity operator, the Frobenius norm on M (C) (and more generally the norms 2 k·kp on Mn(C) and on p( )) are not able to distinguish between projections and idempotents. This shows that bothC H of the norm conditions of Lemma 2.3 are necessary. Despite this, there is still something we can say, at least when both A are N are invertible and acting on finite-dimensional spaces.

12 5.4. Corollary. Let A Mm(C), N Mn(C) be invertible matrices, and N be normal. ∈ ∈ [m] [n] Let κ := max(m, n), 1 p< , and let p (resp. p ) denote the p-norm on Mm(C) ≤ ∞ k·k [m]k·k [n] (resp. the p-norm on Mn(C)). Suppose that r(A) p = r(N) p for all polynomials r of degree at most κ for which r(0) = 0. Then k k k k (1) n = m; (2) A is also normal; and (3) A and N are unitarily similar.

Proof. Let µN ,µA denote the minimal polynomials of N and of A respectively. Then NµN (N) =0, and so 0= Nµ (N) [n] = Aµ (A) [m], k N kp k N kp whence AµN (A) = 0. Since A is invertible, it follows that µN (A)= 0, and so µA divides µN . This part of the argument clearly does not rely upon the fact that N is normal, and so by symmetry, we find that µN divides µA as well. Without loss of generality, we may therefore assume that µA = µN . We denote this common polynomial by µ. The fact that both A and N are invertible implies that µ(0) = 0. As such, and keeping in mind that µ(A)= µ(N) = 0, we set 6 ν(z)= µ(z) µ(0), z C. − ∈ Clearly ν is a polynomial of degree at most κ and ν(0) = 0. By hypothesis, µ(0)I [m] = ν(A) [m] = ν(N) [n] = µ(0)I [n]. k mkp k kp k kp k nkp [n] [m] From this it follows that n = m, and we simply write to denote p = p . k·kp k·k k·k Let r(z)= r + r z + r z2 + + r zn be a polynomial of degree at most n. Set 0 1 2 ··· n q(z)= r µ(0)−1ν(z)+(r z + r z2 + + r zn), 0 1 2 ··· n so that q is a polynomial of degree at most n and q(0) = 0. By our hypothesis, r(A) = r I +(r A + r A2 + r An) k kp k 0 n 1 2 ··· n kp = r ( µ(0))−1ν(A)+(r A + r A2 + r An) k 0 − 1 2 ··· n kp = q(A) k kp = q(N) k kp = r ( µ(0))−1ν(N)+(r N + r N 2 + r N n) k 0 − 1 2 ··· n kp = r I +(r N + r N 2 + r N n) k 0 n 1 2 ··· n kp = r(N) . k kp Since µA = µN is a polynomial of degree at most n, it follows that if t is any polynomial, then t(A)= s(A) and t(N) = s(N) for some polynomial s of degree at most n. Combining this with the above calculation yields t(A) = t(N) for all polynomials t. k kp k kp Since the p-norm separates projections by rank, it follows from Theorem 2.4 that A and N are unitarily similar.  13 Acknowledgments. This paper was written when the second author was spending his sabbatical leave at the University of Waterloo. He would like to thank Department of Pure Mathematics of the University of Waterloo, and in particular Professor L.W. Marcoux, for their kind support.

References [1] D. P. Blecher and C. Le Merdy, Operator algebras and their modules – an operator space approach, London Math. Soc. Monographs 30, Clarendon Press, Oxford, 2004. [2] C. D. Brooks and A. A. Condori, A resolvent criterion for normality, Amer. Math. Monthly 125 (2018), no. 2, 149–156. [3] Jor-Ting Chan, Chi-Kwong Li, and Charlies C. N. Tu, A class of unitarily invariant norms on B(H), Proc. Amer. Math. Soc. 129 (2001), no. 4, 1065–1076. [4] J. B. Conway, A course in , second ed., Graduate Texts in Mathematics, vol. 96, Springer-Verlag, New York, 1990. [5] K. R. Davidson, Nest algebras, Pitman Research Notes in Mathematics Series, vol. 191, Longman Scientific & Technical, Harlow, 1988, Triangular forms for operator algebras on Hilbert space. [6] L. Elsner and Kh. D. Ikramov, Normal matrices: an update, Linear Algebra Appl. 285 (1998), no. 1-3, 291–303. [7] K. Fan, Maximum properties and inequalities for the eigenvalues of completely continuous operators, Proc. Nat. Acad. Sci., U. S. A. 37 (1951), 760–766. [8] D. R. Farenick, B. E. Forrest, and L. W. Marcoux, Amenable operators on Hilbert spaces, J. Reine Angew. Math. 582 (2005), 201–228. [9] D.R. Farenick, V. Futorny, T. G. Gerasimova, V. V. Sergeichuk, and N. Shvai, A criterion for unitary similarity of upper triangular matrices in general position, Linear Algebra Appl. 435 (2011), no. 6, 1356–1369. [10] D.R. Farenick, T.G. Gerasimova, and N. Shvai, A complete unitary similarity invariant for unicellular matrices, Linear Algebra Appl. 435 (2011), no. 2, 409–419. [11] T.G. Gerasimova, Unitary similarity to a normal matrix, Linear Algebra Appl. 436 (2012), no. 9, 3777–3783. [12] I.C. Gohberg and M.G. Krein, Introduction to the theory of linear nonselfadjoint operators, Amer. Math. Soc., Providence, R.I., (1969). [13] A. Greenbaum and L.N. Trefethen, Do the pseudospectra of a matrix determine its behavior?, Technical Report TR 93-1371, Cornell University, 1993. ecommons.cornell.edu/bitstream/handle/1813/6145/93- 1371.pdf. [14] R. Grone, C. R. Johnson, E. M. Sa, and H. Wolkowicz, Normal matrices, Linear Algebra Appl. 87 (1987), 213–225. [15] C. S. Kubrusly, The elements of operator theory, second ed., Birkh¨auser/Springer, New York, 2011. [16] M.A. Lavrentieff, Sur les fonctions d’une variable complexe repr´esentables par les series de polynˆomes, Acta. Sci. Ind. 441 (1936). [17] C.K. Li and N.K Tsing, On the unitarily-invariant norms and some related results, Linear and Multi- linear Algebra 20 (1987), no. 2, 107–119. [18] L. Mirsky, Symmetric gauge functions and unitarily-invariant norms, Quart. J. Math. Oxford Ser. (2) 11 (1960), 50–59. [19] A. Pietsch, Eigenvalues and S-numbers, Cambridge Studies in Advanced Mathematics 13, Cambridge Univ. Press, Cambridge, 1987. [20] J. R. Ringrose, Super-diagonal forms for compact linear operators, Proc. London Math. Soc. (3) 12 (1962), 367–384. [21] W. Rudin, Real and complex analysis, third ed., McGraw-Hill Book Co., New York, 1987. [22] D. Viswanath and L.N. Trefethen, Matrix behaviour, unitary reducibility, and Hadamard products, Tech- nical Report TR96-1596, Cornell University, 1996. ecommons.cornell.edu/handle/1813/7251. [23] F. Wolf, On the essential spectrum of partial differential boundary problems, Comm. Pure Appl. Math. 12 (1959), 211-228. 14 Laurent W. Marcoux, Department of Pure Mathematics, University of Waterloo, Wa- terloo, Ontario N2L 3G1 CANADA E-mail address: [email protected] Yuanhang Zhang, School of Mathematics, Jilin University, Changchun 130012, P.R. CHINA E-mail address: [email protected]

15