WorksMathematics
Similarity cancellation and primitive operator roots
Abstract
Kaplansky's second test problem for similarity asks whether implies for bounded operators on a separable infinite-dimensional Hilbert space. We answer this question in the negative, even for compact operators whose singular values satisfy any prescribed positive upper bounds tending to zero. We also show that cancellation can fail under the stronger hypothesis considered by Marcoux, Radjavi, Troscheit, and Zhang \cite{MRTZ}. For these operators, finite amplifications and primitive roots are similar exactly at even multiplicities and even orders, respectively. In contrast, cancellation in fixed matrix size admits a uniform bound on the condition number. This extends the finite type~ results of Jiang, Ma, Shi, and Zhang \cite{JMSZ}. The bound also holds for simultaneous similarity of arbitrary families of matrices and can be chosen to tend to one as the amplified bound tends to one. We conclude that, in a finite type~ von Neumann algebra, cancellation of doubles and cancellation of primitive square roots are each equivalent to boundedness of the orders of its nonzero homogeneous summands.
Introduction
Let be a complex separable infinite-dimensional Hilbert space. We denote by the algebra of bounded operators on . Two operators and are similar, written , if for some bounded invertible operator . We use the same notation for operators on different Hilbert spaces, with acting between those spaces. All Hilbert space direct sums are orthogonal. For a positive integer , write
In his study of abelian groups, Kaplansky asked whether forces [16]. This is the second of his three test problems
2020 Mathematics Subject Classification. Primary 47A45, 47A65; Secondary 46L10, 47B10.
Key words and phrases. Similarity, finite amplification, primitive operator roots, compact operators, von Neumann algebras. for direct sums. Kadison and Singer [15] brought these questions to Hilbert space operators, and Azoff [1] studied them in several categories of operator algebras. Related versions of Kaplansky’s test problems for quasisimilarity were studied by Bercovici [2] and, for bitriangular operators, by Davidson and Herrero [5]. The answer depends on the equivalence relation being considered. For unitary equivalence, Kadison and Singer proved that
where denotes unitary equivalence. There is also a multiplicity theory for approximate unitary equivalence. Hadwin’s rank criterion for unital representations on a separable Hilbert space [8 Theorem 2.5] gives a formulation of Voiculescu’s theorem suited to this question; see also [4 §II.5]. In particular, finite amplifications cancel for approximate unitary equivalence in , as shown in [19 Proposition 4.4]. Related questions concerning distances between unitary orbits of matrix amplifications were studied by Marcoux, Sarkowicz, and Zhang [20]. Their results concern a different quantity from the condition number considered here. For similarity, the question is
Jiang [11 Theorem 4.6] proved cancellation of finite amplifications for strongly irreducible Cowen–Douglas operators. Jiang, Guo, and Ji [12] developed the similarity classification of Cowen–Douglas operators using the ordered -group of the commutant algebra. For weighted Hardy spaces of polynomial growth, Hou and Jiang [10 Theorem B] proved cancellation of doubles for the Hermitian holomorphic vector bundles induced by functions analytic in a neighbourhood of the closed unit disc.
Marcoux, Radjavi, Troscheit, and Zhang record this question in [19 §1.3], including the case in which and are compact.
The same authors consider a stronger hypothesis. Following their notation, for let
They call the primitive th root of . In particular,
Thus, similarity of and implies similarity of the doubles. The question in [19 Remark 5.3] is whether
Their Theorem 4.10 proves this implication when one of the operators is compact. A counterexample to cancellation in a unital ring, due to J. Bell, is recorded in [19 Example 5.4]. That algebraic example does not settle either question in .
We answer both questions in the negative. Our counterexample to (3) is noncompact, as required by the theorem just cited. The same pair distinguishes even from odd multiplicities, both for direct sums and for primitive roots.
Theorem 1.1. For every , there are noncompact operators with , such that
The operators have the same point spectrum. Every eigenspace is one-dimensional, and the eigenvectors of each operator have dense linear span.
The answer to (1) is negative even when both operators are compact. In fact, their singular values can lie below any prescribed positive sequence tending to zero. For a compact operator , let denote its singular values in nonincreasing order, counted with multiplicity.
Theorem 1.2. Let be a nonincreasing sequence of positive numbers tending to zero. There are compact operators such that
and
Both operators are injective and have dense range. Their common nonzero spectrum consists of positive algebraically simple eigenvalues, and their eigenvectors have dense linear span.
For an arbitrary positive null sequence , apply Theorem 1.2 to the nonincreasing sequence . Since and , the same upper bounds follow without a monotonicity assumption. In particular, and may belong to every Schatten class. Their positive eigenvalues should not be confused with positivity of the operators.
The counterexamples are orthogonal sums of diagonalizable matrices. In finite dimension, the Jordan form gives cancellation of finite amplifications; see [19 §1.3]. For an infinite orthogonal sum, however, the similarities chosen on the summands must have uniformly bounded norms and inverse norms. Pointwise similarity without such bounds is not enough, as the examples of Deckard and Pearcy already show [7 §2]. We therefore also consider the condition number of a similarity. For matrices , set
with when and are not similar. For and , put
For , this is the function introduced by Jiang, Ma, Shi, and Zhang [13 Definition 4.1].
Theorem 1.3. For every and ,
Jiang, Ma, Shi, and Zhang relate this bound to similarity cancellation in homogeneous type von Neumann algebras [13 Theorem 4.3]. They prove cancellation under property [13 Theorem 3.27] and, without this assumption, for [13 Proposition 4.7]. Theorem 1.3, together with the continuous selection theorem of Deckard and Pearcy [7], removes the extra assumption for every finite and every multiplicity .
Two further conclusions follow from the proof. The same uniform bound, with a possibly different constant, holds when one operator is required to intertwine an arbitrary family of matrices simultaneously. The constant does not depend on the number of matrices in the family. For a single pair, we also prove
Thus, a similarity with condition number close to one remains so after cancellation, provided the matrix size is fixed. Proposition 2.2 proves the same assertion for a common intertwiner of arbitrary families, uniformly in the index set.
The dependence on matrix size cannot be omitted. The finite-dimensional operators used in Theorems 1.1 and 1.2 have uniformly controlled similarities after doubling, while every similarity at an odd multiplicity has arbitrarily large condition number. Consequently, there is a fixed such that
These pairs also show that can occur, contrary to the equality suggested in [13 Remark 4.2]; see Corollary 5.1. The relation between a dimension-independent bound and cancellation in finite type I algebras was established in [13 Corollary 4.6]. We decide cancellation for each bounded product : both the double and primitive-square-root questions have affirmative answers for all pairs in the product if and only if . More generally, for a finite type I von Neumann algebra, the same two properties are equivalent to a common finite bound on the orders of its nonzero homogeneous summands. The centres of those summands may be diffuse.
The proofs of the positive and negative results both concern uniform bounds for finite-dimensional intertwiners. To prove Theorem 1.3, we take a limit of the commutants of inside a fixed matrix space. The limit is a finite-dimensional algebra, so Krull–Schmidt cancellation applies. The resulting intertwiner can be approximated by exact intertwiners for the original pairs.
For the counterexamples, the starting point is the even and odd decomposition of a Sobolev space on the sphere. Restriction to a finite set gives two Hilbert norms on the same vector space of functions. An intertwiner between multiplication operators with distinct eigenvalues must act pointwise. Applying it to constant functions and lifting to the odd Sobolev space produces a continuous odd matrix-valued function. At an odd multiplicity its determinant must vanish, by Borsuk–Ulam. Compactness of the inclusion into then forces the condition number to be large. At multiplicity two, a fixed invertible block operator induces similarities on all the quotients. Its entries are chosen to satisfy the additional identity for primitive square roots. Taking orthogonal sums with disjoint finite spectra gives the two pairs of operators.
The paper is organised as follows. § 2 proves the fixed-size bound, its simultaneous version, and (10), and then treats homogeneous type I algebras. § 3 constructs the Hilbert space quotients and their intertwiners. § 4 proves Theorems 1.1 and 1.2 by taking orthogonal sums. § 5 compares bounded and unbounded matrix sizes and characterises cancellation in finite type I von Neumann algebras.
Cancellation in fixed matrix size
We first prove the uniform estimate for matrices. The argument uses cancellation inside a finite-dimensional algebra and compactness of the set of orthogonal projections on a fixed matrix space.
By Krull–Schmidt cancellation, implies for finite-dimensional modules over a finite-dimensional algebra; see [17 Theorems 4.2 and 5.5].
We apply this result to right ideals in finite-dimensional operator algebras, without assuming self-adjointness. Let be a finite-dimensional unital algebra and let be idempotents. Suppose
Left multiplication by identifies with . Cancellation gives . An isomorphism between these right modules is left multiplication by an element , and its inverse is left multiplication by some . Thus,
We are now ready to prove Theorem 1.3.
Proof of Theorem 1.3. Fix , and , and suppose that . For each , choose with
By the definition of the infimum, there is an invertible such that
Multiplying by a positive scalar, we may assume
Consider the commutant
and the two fixed projections
Both projections belong to every . Equip with the Hilbert–Schmidt inner product, and let be the orthogonal projection onto the linear subspace . The maps act on a fixed finite-dimensional Hilbert space and have norm one. After passing to a subsequence, we may suppose that
in norm. The limit is an orthogonal projection. For every , (2.5) gives . The same lower bound holds for . Since is square and finite-dimensional, it is invertible, and
Put . We first check that is an algebra. If , then and . Since ,
Taking limits gives . Hence, . Moreover, , since each of these matrices is fixed by every . Thus, is a finite-dimensional unital algebra.
Write and in blocks of size . Equation (15) and its inverse give
If and , then implies . Therefore the corresponding corner matrices formed from the entries of and belong to and . Denote the resulting matrices over by and . The equations give (12).
By (13), there are and with and . Write
Then, , so is invertible.
For each , the matrix belongs to and tends to . Write its lower-left block as . We have
For all sufficiently large , is invertible and . Thus, stays bounded, contradicting (14). □
The proof does not require the sequences and to converge, or even to be bounded. It is the dimension of the ambient commutant space that is fixed. Also, the limiting algebra need not itself be a commutant.
Simultaneous similarity
The proof just given also applies to a joint commutant. This allows the same operator to intertwine any number of matrices, without a commutativity assumption on the family.
Theorem 2.1. Fix , and . There is a constant with the following property. Let and be families in , indexed by the same nonempty set. Suppose that one invertible matrix satisfies
Then, there is an invertible such that
The constant is independent of and of the two families.
We apply the preceding commutant argument to all the equations at once.
Proof. Suppose that no such constant exists. For each , choose two families and with a common amplified intertwiner of condition number at most , but no common intertwiner of condition number at most . After balancing , we have .
Apply the proof of Theorem 1.3 with the joint commutants
Each block of satisfies every intertwining equation at the th stage, and the blocks of satisfy the reverse equations. Thus the same limiting algebra and corner argument give an element with invertible lower-left block. The lower-left blocks of converge to that block and satisfy
For all sufficiently large , the matrices are invertible and their condition numbers are bounded, contrary to the choice of the families.
The index set may be infinite. The proof takes place in and uses only the common intertwining equations, not a bound on the number of those equations.
Condition numbers close to one
For an invertible matrix, condition number one means that a scalar multiple is unitary. The next proposition shows that this remains a uniform limiting statement after cancellation in a fixed size.
Proposition 2.2. For every fixed , ,
More generally, for every there is , depending only on , , , with the following property. Let and be families in indexed by the same nonempty set. If an invertible satisfies
then there is an invertible such that
In this case the limiting algebra contains the adjoints of the corner entries. This allows us to choose a partial isometry before returning to the original commutants.
Proof. We prove the assertion for families first. If it fails, there are and, for each , families and with a common amplified intertwiner of condition number at most , but no common intertwiner of condition number at most . After balancing,
Use the joint commutants from the proof of Theorem 2.1 and the projections and subsequences from the proof of Theorem 1.3. We obtain , , and the unital algebra . Here and have norm at most one, so is unitary.
Let . This is a finite-dimensional unital -algebra containing and . The lower-left corner matrices formed from the blocks of belong to . Their adjoints also belong to , since and the upper-right corner matrices formed from the blocks of have limits in . Hence, the matrices of (2.1) belong to the corresponding corners over , and .
Apply (2.2) in . We obtain and with and . Thus, is an invertible map from onto . By continuous functional calculus in the corner [4 §I.4], its polar part
belongs to and satisfies and . The inverse square root here is taken in , whose identity is . Write
then is unitary.
The lower-left blocks of satisfy
and converge to . For all sufficiently large , they are invertible and
contrary to the choice of the families.
Now fix and choose as in the assertion for families. For a single pair with , the definition of the infimum gives an amplified intertwiner with condition number less than . The assertion just proved yields . Hence , which proves the limit in (17). Finally, for , so .
The matrix size is fixed throughout (17). Section 5 shows that a bound valid in all sizes is a different question.
Homogeneous type I-n algebras
We next apply the fixed-size estimate to continuous matrix-valued functions. Deckard and Pearcy studied these algebras in [6, 7]; the uniform norm requirement in their similarity theorem is the point needed here.
A homogeneous type von Neumann algebra is isomorphic to for an abelian von Neumann algebra . We may write , where is compact and hyperstonean, hence Stonian; see [14 §6.6]. We use the following continuous selection result of Deckard and Pearcy [7 Theorem 1, p. 858]: if are pointwise similar on a dense subset of , with intertwiners whose norms and inverse norms are all less than one fixed number , then there is an invertible with and
Theorem 2.3. Let be a homogeneous type von Neumann algebra, with , and let . If in for some positive integer , then in . In particular, in implies for every .
We can now apply the selection theorem to the pointwise similarities.
Proof. Write as above. Suppose that for an invertible , and set . Evaluation at gives
By Theorem 1.3,
Choose . For each , the definition of the infimum gives a pointwise intertwiner with condition number less than . A scalar multiple makes both its norm and its inverse norm less than . The theorem of Deckard and Pearcy applies on all of and gives a similarity in . For primitive roots, take th powers and apply the assertion just proved.
Hilbert quotients and finite-dimensional intertwiners
We now construct finite-dimensional pairs with uniformly bounded similarities of their doubles, but without such a bound before amplification. We work with two fixed Hilbert spaces and take their quotients by the functions that vanish on a finite set.
Even and odd quotient spaces
We choose the Hilbert spaces so that evaluation is continuous and smooth multiplication is bounded. Let be the complex Sobolev space on the unit sphere in , equipped with an orthogonally invariant Sobolev norm, normalised so that . Multiplication by every smooth function is bounded on , and the natural inclusion
is compact. These are consequences of the smooth-multiplier, Sobolev embedding, and Rellich theorems; see [23 Chapter 4, §§2–3]. In particular, compactness follows by factoring through for , using Propositions 3.3–3.4 of that chapter. We use the continuous representative of each element of .
The operator
is a unitary involution. Its orthogonal eigenspaces
consist of the even and odd functions. A smooth even multiplier preserves both spaces, while a smooth odd multiplier maps each into the other.
Let and denote evaluation at on and . Orthogonal invariance gives constants such that
The first functional is nonzero because it evaluates the constant function. The second is nonzero because a smooth odd function can be chosen not to vanish at any prescribed point. By (3.1) and the Arzelà–Ascoli theorem [3 Theorem VI.3.8], the unit ball of is uniformly equicontinuous. Hence,
The norm in this formula is the dual norm on .
Let be a finite nonempty subset of containing no antipodal pair. Let and be the restrictions to of and , with the quotient norms
Both spaces contain every function on . Indeed, prescribe the values and at and , respectively. Smooth interpolation on , followed by the corresponding symmetrisation, gives an even or odd function with restriction ; see the bump-function construction in [18 Chapter 2]. Thus, and are the same vector space of functions, equipped with different Hilbert norms.
The restriction maps have closed kernels. By the Hilbert space projection theorem [3 Theorems I.2.5–I.2.7], each quotient has a linear isometric lift obtained by choosing the representative of least norm. In particular, there is an isometry
The constant function on has -norm at most one.
For , evaluation on either quotient has the same norm as on the corresponding full space, by the isometric identification of the dual of a quotient with the annihilator of its kernel [3 Theorem III.10.2]. On the -fold sums, write
for the componentwise evaluations. By (19) and the Riesz representation theorem,
For any scalar function , let and be multiplication by on and :
If is injective, both operators are diagonalizable with simple eigenvalues . The eigenspace at consists of the functions supported at . Pointwise bounds on alone need not bound these operator norms uniformly as varies.
Odd multiplicities
We show that the norms of a similarity and its inverse cannot both stay bounded as the finite sets and their antipodes fill the sphere.
We call an -net if each point of the sphere is at distance less than from this set. We shall use the fact that every continuous odd function has a zero. This is the Borsuk–Ulam theorem applied to the real and imaginary parts; see [9 Theorem 1.10, p. 32].
Proposition 3.1. For every , there is with the following property. Let contain no antipodal pair, suppose that is an -net, and let be injective. For every odd positive integer , each invertible operator
satisfies
The choice of is independent of , of , and of the values of .
The constant function and the evaluation operators give the required estimate.
Proof. Put . By (20), choose such that
Suppose that such a has . After a positive scalar rescaling, we may assume
Since the eigenvalues in (23) are distinct, the intertwining equation makes act pointwise:
In particular, . Equation (22) gives
Consequently,
Let send to the constant function with value . Since , the operator maps into and has norm at most . Define the matrix-valued function by
It is continuous and odd, and for . Under the Hilbert space identification , componentwise evaluation is . Hence
It follows that
Since is odd, is a continuous odd complex-valued function. Borsuk–Ulam gives a point where is singular. Choose a unit vector and a point with . At this point, is or for some . Equations (28), (25), and (30) give the contradiction
An invertible block intertwiner
The intertwining equation for primitive square roots requires equal diagonal blocks; see [19 Lemma 4.9]. We choose those blocks, and the off-diagonal entries, so that the inverse can be written explicitly.
Lemma 3.2. Let , , be commuting bounded operators on a Hilbert space such that
and suppose that is invertible. Set
Then,
If a bounded involution anticommutes with , , , then commutes with , while and interchange the two eigenspaces of .
Proof. All the operators in the block entries commute. The relation gives
Multiplying the two block matrices in either order gives the identity, with lower-left inverse entry . Comparing the products with proves the intertwining identity. If anticommutes with , , , then , , commute with , and each entry of and anticommutes with . This proves the last assertion. □
We realise these operators on . For a smooth scalar function , let denote multiplication by , and put
They are bounded, commute, and satisfy by the equation of the sphere. All three anticommute with . Here is multiplication by the conjugate coordinate, not the Hilbert space adjoint of .
The operator is multiplication by . Its reciprocal is smooth, so is a bounded multiplier. This argument does not require to be self-adjoint for the Sobolev inner product. The operator is multiplication by the smooth even function
Lemma 3.2 therefore gives a bounded invertible operator
Both and its inverse have smooth multiplication operators as entries. They are fixed independently of the finite set .
We now pass this operator and its inverse to the finite-dimensional quotients.
Proposition 3.3. There is such that, for each finite nonempty containing no antipodal pair, an invertible operator
satisfies and
for every scalar function on . Moreover,
and
Proof. Multiplication preserves vanishing on . Thus, maps the restriction kernel in into that in , and maps the latter kernel back into the former. The two inclusions show that maps the first kernel onto the second. Hence, the induced quotient maps are inverses. By the quotient norms,
For example, is at most for every lift of ; taking the infimum over lifts gives the first bound. The inverse is treated in the same way.
At each , the matrix defining commutes with . This proves (3.18), even when is specified only on . The identity with in (31) passes to the quotients and gives (36). Finally, preserves both parity spaces and both restriction kernels. Its quotient operators have norm at most , proving (37).
The distinction between the two intertwining equations will be used below. Equation (35) allows any finite choice of eigenvalues. Equation (36) requires the particular multiplier .
Counterexamples on Hilbert space
We first choose the finite sets and recall how disjoint spectra control the intertwiners of an orthogonal sum. Both constructions will then use the operators from Section 3.
Finite sets and spectral separation
The finite sets must become dense after adjoining their antipodes. We also choose them so that no eigenvalue occurs in two different summands.
Let
Choose finite nonempty sets such that
and all third coordinates of all selected points are distinct. Such sets can be chosen inductively. For a prescribed radius, compactness of the sphere and density of give a finite antipodal net with smaller radius. Perturb its points within to avoid the third coordinates chosen earlier and to make the new coordinates distinct. Only finitely many values are excluded, and every nonempty open subset of contains an interval of possible third coordinates. The perturbations can be made small enough to retain the prescribed covering radius.
Suppose that and act on finite-dimensional Hilbert spaces, their norms are uniformly bounded, and
If , then its blocks satisfy
For , the Sylvester–Rosenblum theorem [21] gives . Thus, is block diagonal. If is invertible, its diagonal blocks are invertible and
The same conclusion holds after a common finite amplification. This is the orthogonal-sum version of the spectral separation used in [19, Lemma 2.16]. Only disjointness of the finite spectra is needed; their union need not have isolated points.
Similar primitive square roots
We are now ready to prove Theorem 1.1. The multiplier gives uniformly bounded finite-dimensional operators, and the maps give a similarity of their primitive square roots.
Proof of Theorem 1.1. On and , respectively, set
and define
Both sums are bounded by (37). Proposition 3.3 also shows that
are bounded. Reordering the orthogonal sums and applying (36) gives
Taking squares gives similarity of the doubles. Repetition gives similarity at every even multiplicity.
For , the function in (33) satisfies
This is strictly decreasing as increases from zero to one. Hence, the values at all selected points are nonzero and pairwise distinct, including values from different sets. The finite blocks in (41) therefore satisfy (39).
Suppose that an invertible intertwines an odd amplification of and . Spectral separation gives invertible restrictions satisfying (40). Choose with . By (38), every sufficiently large satisfies the net condition in Proposition 3.1. Hence,
contrary to (40). We have proved
The point spectrum of either operator is exactly . Indeed, the components of an eigenvector must solve the corresponding finite-block eigenvalue equations. Since all the eigenvalues are distinct, every eigenspace is one-dimensional. The eigenvectors span each finite block, so their linear span is dense in each direct sum.
Neither operator is compact. Fix a point on the equator and choose points in converging to it. Their third coordinates tend to zero, so the corresponding eigenvalue moduli tend to by (43). For either operator, choose unit eigenvectors in these distinct finite blocks. They form an orthonormal sequence whose images do not tend to zero. The compact-operator criterion in [3 Chapter II, §4] excludes compactness.
To obtain the assertion about all primitive roots, we use two identities from [19]: similarity of and implies similarity of and by Lemma 2.4(c), and
by Proposition 2.5(a). Applying these facts to (42) proves similarity of the primitive roots at every even order. At an odd order , such a similarity would give after taking th powers, contrary to (44).
For , let . The scalar-rescaling identity recorded in [19 Remark 5.3] is
Choose with , and put , . The intertwining equations for finite amplifications are unchanged by a common nonzero scalar. Equation (45) preserves the similarity of the primitive square roots. The preceding even-order argument and the odd-power argument then give (1.5) for this scaled pair. The assertions about point spectra, eigenspaces, and noncompactness follow from those proved above. Finally, identify the two separable infinite-dimensional Hilbert spaces unitarily with . □
Compact operators with prescribed singular value bounds
For the compact pair, we keep the quotient spaces and the similarities of the doubles, but choose positive eigenvalues separately on each finite set. Small scalar multiples will give the singular-value bounds in Theorem 1.2.
Proof of Theorem 1.2. Choose any injective function
on each finite set, and now put
Their norms are finite, although they may grow with . Proposition 3.3 gives the same uniformly bounded similarities of their doubles, while Proposition 3.1 still applies at every odd multiplicity.
Write , , and . Choose positive numbers inductively so that
and the finite sets are pairwise disjoint. At each step, (46) leaves a nonempty interval of choices, whereas spectral disjointness excludes only finitely many ratios of previous eigenvalues to the values of . Hence, both requirements can be satisfied.
Define
By (46), the tails tend to zero in operator norm, so and are compact. For , the first blocks have total rank less than , and the remaining tail has norm at most
The singular-value approximation principle [22 Chapter 1],
therefore gives (1.6) for both operators.
The maps and their inverses are uniformly bounded. Their direct sums are bounded inverses of one another, and (3.18), applied to , gives . Repetition gives all even multiplicities. If an invertible intertwined an odd amplification, the disjoint finite spectra would make it block diagonal. Its restrictions would satisfy (4.3), whereas Proposition 3.1 contradicts those bounds for sufficiently large . This proves (1.7).
Every finite block is invertible. Thus, are injective, and their ranges contain all finitely supported vectors, so both ranges are dense. Their eigenvectors span each finite block and have dense linear span. By the Riesz–Schauder theorem [3 Chapter VII, §7], every nonzero spectral value of a compact operator is an eigenvalue. The block equations identify these values as
They are positive and distinct. Each belongs to a single diagonalizable finite block. The same block equations for powers of and show that the generalized eigenspace at each nonzero eigenvalue is one-dimensional. Hence, each such eigenvalue is algebraically simple. Unitary identifications put both operators on .
For , a compact operator belongs to when ; see [22 Chapter 1]. Choosing a sequence whose positive powers are all summable gives the following consequence.
Corollary 4.1. There are injective compact operators with dense range, belonging to every Schatten class, such that but .
Proof. This follows from Theorem 1.2 by taking .
For the compact pair, by [19 Theorem 4.10]. The two constructions use the same quotient spaces and the same doubled intertwiners. Changing the eigenvalues and choosing small weights preserves (3.18), but not the additional identity (3.19) required for primitive square roots.
Matrix size and bounded products
We first compare the similarities of the finite-dimensional pairs before and after amplification. We then use those pairs to decide cancellation in bounded products of matrices and in finite type I von Neumann algebras.
The following corollary records the finite-dimensional pairs obtained from Propositions 3.1 and 3.3.
Corollary 5.1. There is with the following property. For every , there are invertible diagonalizable matrices of the same size, with distinct eigenvalues and , and an invertible such that
while every invertible satisfying for an odd positive integer has
The matrix size may depend on .
Proof. Take the spaces and multiplication by . For sufficiently large , Proposition 3.1 gives (5.2) simultaneously for all odd . The eigenvalues are distinct and nonzero by (4.6) and the choice of the sets. Proposition 3.3 gives operator norms at most and an intertwiner of the primitive square roots whose norm and inverse norm are uniformly bounded. Choose one fixed scalar such that , and multiply both operators by . This leaves the intertwining equations for finite amplifications unchanged. By (45), the primitive-root intertwiner is conjugated by . Its norm and inverse norm are multiplied by at most , a fixed number. Thus, (48) holds with a constant independent of and . Represent the two finite-dimensional Hilbert spaces in orthonormal bases to obtain the stated matrices.
Taking squares in (48) gives . Equation (49) with , together with Theorem 1.3, proves (1.11) with . Thus, the failure of a bound independent of matrix size occurs even for invertible diagonalizable contractions with simple spectra, and even when their primitive square roots have uniformly bounded similarities. Taking in Corollary 5.1 also gives , so the equality suggested in [13 Remark 4.2] does not hold in general.
Let be a sequence of positive integers. In the theorem below,
denotes the von Neumann algebra of all sequences with , equipped with the supremum norm. This is the bounded product, not the norm-closed direct sum. An element is invertible in precisely when every is invertible and both and are finite.
The reduction for the full sequence of matrix sizes at multiplicity two is given in [13 Corollary 4.6]. We now decide cancellation for each sequence of sizes and include the primitive-square-root question.
Theorem 5.2. For the algebra in (50), the following are equivalent.
(i) .
(ii) For all , similarity of and in implies similarity of and in .
(iii) For all , similarity of and in implies similarity of and in .
When these conditions hold, cancellation is valid for every finite amplification and every primitive root.
The fixed-size theorem proves one direction. For the converse, the matrices in Corollary 5.1 fit into a subsequence of factors of increasing size.
Proof. Assume (i), and choose with for every . Suppose that for an invertible , and put . Evaluation in the th factor gives
By Theorem 1.3,
For each , choose with condition number less than , and balance by a positive scalar. Then,
Thus, is invertible in and satisfies . This proves (ii) and cancellation at every finite multiplicity. Taking powers gives cancellation for every primitive root. In particular, (ii) implies (iii) directly by squaring.
Suppose now that is unbounded. For each , choose matrices and an intertwiner as in Corollary 5.1, with . Choose distinct indices inductively so that . In the th factor set
and put in the other factors. These sequences define contractions .
A permutation of the direct summands identifies with
and gives the same identification for . The map intertwines these two operators. Its norm and inverse norm are at most , after increasing the fixed constant to at least one. In the other factors use the identity. The resulting sequence is an invertible element of , so
If and were similar in , their coordinate similarities would have uniformly bounded condition numbers. In the th factor, invertibility of and separates their spectra from the additional zero block. The Sylvester–Rosenblum theorem therefore forces any such similarity to restrict to a similarity between and , with no increase in its norm or inverse norm. By (49), the condition number of this restriction is greater than . This contradicts the uniform bound. Thus, (5.4) holds without similarity of and , which disproves (iii). We have shown that (iii) implies (i), completing the proof.
Finite type I von Neumann algebras
The centres of the homogeneous summands do not change the conclusion. We use the central decomposition of a finite type I von Neumann algebra [14 §6.6], in the form
where and each is a nonzero abelian von Neumann algebra. The product in (52) is again the bounded product. We exclude the zero algebra, so is nonempty.
Theorem 5.3. Let be a nonzero finite type I von Neumann algebra, with decomposition (52). The following conditions are equivalent.
(i) The set is bounded.
(ii) For all , similarity of and in implies similarity of and in .
(iii) For all , similarity of and in implies similarity of and in .
When these conditions hold, cancellation is valid for every finite amplification and every primitive root.
To prove the converse, it is important not to infer nonsimilarity in merely from nonsimilarity in a subalgebra. The abelian centres provide homomorphisms back to the scalar matrix factors.
Proof. If is bounded, it is finite. Apply Theorem 2.3 in each of the finitely many homogeneous summands. The resulting similarities and their inverses form bounded families, so they give a similarity in . This proves cancellation for every finite amplification. Taking powers gives the assertion for primitive roots. As before, (ii) implies (iii) by squaring.
Suppose that is unbounded, and put
By Theorem 5.2, there are whose primitive square roots are similar in but which are not similar in . Let
be the isometric unital inclusion. It carries the given similarity to one between and . For each , choose a character on . Such a character exists because is a nonzero unital abelian -algebra; see [4 §I.3]. The characters are not assumed to be normal. Applying entrywise gives a unital -homomorphism
Each coordinate map is contractive, so is well defined on the bounded product, and . If an invertible intertwined and , then would intertwine and and would have inverse . This contradicts their nonsimilarity in . Hence, (iii) fails when is unbounded, proving the theorem. □
References
Department of Mathematics, Faculty of Sciences of Monastir, University of Monastir, Tunisia
Email address: [email protected]
Email address: [email protected]
References
- [1]E. A. Azoff, Test problems for operator algebras, Trans. Amer. Math. Soc. 347 (1995), no. 8, 2989–3001.DOI
- [2]H. Bercovici, Three test problems for quasisimilarity, Canad. J. Math. 39 (1987), no. 4, 880–892. doi:10.4153/CJM-1987-043-x.DOI
- [3]J. B. Conway, A Course in Functional Analysis, 2nd ed., Graduate Texts in Mathematics, vol. 96, Springer-Verlag, New York, 1990.
- [4]K. R. Davidson, C*-Algebras by Example, Fields Institute Monographs, vol. 6, American Mathematical Society, Providence, RI, 1996.DOI
- [5]K. R. Davidson and D. A. Herrero, The Jordan form of a bitriangular operator, J. Funct. Anal. 94 (1990), no. 1, 27–73. doi:10.1016/0022-1236(90)90027-I.DOI
- [6]D. Deckard and C. Pearcy, On matrices over the ring of continuous complex-valued functions on a Stonian space, Proc. Amer. Math. Soc. 14 (1963), no. 2, 322–328.DOI
- [7]D. Deckard and C. Pearcy, On continuous matrix-valued functions on a Stonian space, Pacific J. Math. 14 (1964), no. 3, 857–869. doi:10.2140/pjm.1964.14.857.DOI
- [8]D. W. Hadwin, Nonseparable approximate equivalence, Trans. Amer. Math. Soc. 266 (1981), no. 1, 203–231. doi:10.1090/S0002-9947-1981-0613792-6.
- [9]A. Hatcher, Algebraic Topology, Cambridge University Press, Cambridge, 2002.
- [10]B. Hou and C. Jiang, The Jordan decomposition and Kaplansky’s second test problem for Hermitian holomorphic vector bundles, preprint, 2025, arXiv:2503.03415v1.arxiv.org/abs/2503.03415
- [11]C. Jiang, Similarity classification of Cowen–Douglas operators, Canad. J. Math. 56 (2004), no. 4, 742–775. doi:10.4153/CJM-2004-034-8.DOI
- [12]C. Jiang, X. Guo, and K. Ji, K-group and similarity classification of operators, J. Funct. Anal. 225 (2005), no. 1, 167–192. doi:10.1016/j.jfa.2004.12.008.DOI
- [13]C. Jiang, M. Ma, R. Shi, and Y. Zhang, Kaplansky’s second test problem in operator algebras, preprint, 2026, arXiv:2607.20593v1.
- [14]R. V. Kadison and J. R. Ringrose, Fundamentals of the Theory of Operator Algebras. Vol. II: Advanced Theory, Graduate Studies in Mathematics, vol. 16, American Mathematical Society, Providence, RI, 1997.
- [15]R. V. Kadison and I. M. Singer, Three test problems in operator theory, Pacific J. Math. 7 (1957), no. 2, 1101–1106. doi:10.2140/pjm.1957.7.1101.DOI
- [16]I. Kaplansky, Infinite Abelian Groups, University of Michigan Press, Ann Arbor, MI, 1954.
- [17]H. Krause, Krull–Schmidt categories and projective covers, Expo. Math. 33 (2015), no. 4, 535–549. doi:10.1016/j.exmath.2015.10.001.DOI
- [18]J. M. Lee, Introduction to Smooth Manifolds, 2nd ed., Graduate Texts in Mathematics, vol. 218, Springer, New York, 2013.
- [19]L. W. Marcoux, H. Radjavi, S. Troitscheit, and Y. Zhang, Stability relations for Hilbert space operators and a problem of Kaplansky, Math. Z. 308 (2024), no. 3, Paper No. 39, 46 pp. doi:10.1007/s00209-024-03590-w; arXiv:2306.11202v1.DOI
- [20]L. W. Marcoux, P. Sarkowicz, and Y. Zhang, Kaplansky’s problem and unitary orbits in matrix amplifications, J. Operator Theory, to appear, arXiv:2508.13834.
- [21]M. Rosenblum, On the operator equation BX − XA = Q, Duke Math. J. 23 (1956), no. 2, 263–269. doi:10.1215/S0012-7094-56-02324-9.DOI
- [22]B. Simon, Trace Ideals and Their Applications, 2nd ed., Mathematical Surveys and Monographs, vol. 120, American Mathematical Society, Providence, RI, 2005.
- [23]M. E. Taylor, Partial Differential Equations I: Basic Theory, 2nd ed., Applied Mathematical Sciences, vol. 115, Springer, New York, 2011.