A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics

We present a formal geometric framework for the study of adiabatic quantum mechanics for arbitrary finite-dimensional non-degenerate Hamiltonians. This framework generalizes earlier holonomy interpretations of the geometric phase to non-cyclic states appearing for non-Hermitian Hamiltonians. We star...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Symmetry, Integrability and Geometry: Methods and Applications
Datum:2022
Hauptverfasser: Pap, Eric J., Boer, Daniël, Waalkens, Holger
Format: Artikel
Sprache:Englisch
Veröffentlicht: Інститут математики НАН України 2022
Online Zugang:https://nasplib.isofts.kiev.ua/handle/123456789/211542
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Назва журналу:Digital Library of Periodicals of National Academy of Sciences of Ukraine
Zitieren:A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics. Eric J. Pap, Daniël Boer and Holger Waalkens. SIGMA 18 (2022), 003, 42 pages

Institution

Digital Library of Periodicals of National Academy of Sciences of Ukraine
_version_ 1859609232317349888
author Pap, Eric J.
Boer, Daniël
Waalkens, Holger
author_facet Pap, Eric J.
Boer, Daniël
Waalkens, Holger
citation_txt A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics. Eric J. Pap, Daniël Boer and Holger Waalkens. SIGMA 18 (2022), 003, 42 pages
collection DSpace DC
container_title Symmetry, Integrability and Geometry: Methods and Applications
description We present a formal geometric framework for the study of adiabatic quantum mechanics for arbitrary finite-dimensional non-degenerate Hamiltonians. This framework generalizes earlier holonomy interpretations of the geometric phase to non-cyclic states appearing for non-Hermitian Hamiltonians. We start with an investigation of the space of non-degenerate operators on a finite-dimensional state space. We then show how the energy bands of a Hamiltonian family form a covering space. Likewise, we demonstrate that the eigenrays form a bundle, a generalization of a principal bundle, which admits a natural connection that yields the (generalized) geometric phase. This bundle also provides a natural generalization of the quantum geometric tensor and derived tensors, and we demonstrate how it can incorporate the non-geometric dynamical phase as well. We conclude by demonstrating how the bundle can be recast as a principal bundle, allowing both the geometric phases and the permutations of eigenstates to be expressed simultaneously using standard holonomy theory.
first_indexed 2026-03-14T04:12:43Z
format Article
fulltext Symmetry, Integrability and Geometry: Methods and Applications SIGMA 18 (2022), 003, 42 pages A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics Eric J. PAP ab, Daniël BOER b and Holger WAALKENS a a) Bernoulli Institute, University of Groningen, P.O. Box 407, 9700 AK Groningen, The Netherlands E-mail: e.j.pap@rug.nl, h.waalkens@rug.nl URL: http://www.rug.nl/staff/e.j.pap/, http://www.rug.nl/staff/h.waalkens/ b) Van Swinderen Institute, University of Groningen, 9747 AG Groningen, The Netherlands E-mail: d.boer@rug.nl URL: http://www.rug.nl/staff/d.boer/ Received July 23, 2021, in final form December 28, 2021; Published online January 13, 2022 https://doi.org/10.3842/SIGMA.2022.003 Abstract. We present a formal geometric framework for the study of adiabatic quantum mechanics for arbitrary finite-dimensional non-degenerate Hamiltonians. This framework generalizes earlier holonomy interpretations of the geometric phase to non-cyclic states ap- pearing for non-Hermitian Hamiltonians. We start with an investigation of the space of non-degenerate operators on a finite-dimensional state space. We then show how the energy bands of a Hamiltonian family form a covering space. Likewise, we show that the eigenrays form a bundle, a generalization of a principal bundle, which admits a natural connection yielding the (generalized) geometric phase. This bundle provides in addition a natural gen- eralization of the quantum geometric tensor and derived tensors, and we show how it can incorporate the non-geometric dynamical phase as well. We finish by demonstrating how the bundle can be recast as a principal bundle, so that both the geometric phases and the permutations of eigenstates can be expressed simultaneously by means of standard holonomy theory. Key words: adiabatic quantum mechanics; geometric phase; exceptional point; quantum geometric tensor 2020 Mathematics Subject Classification: 81Q70; 81Q12; 55R99 1 Introduction The eigenvalues of a matrix or operator are of great interest in many fields of mathematics and physics. In quantum mechanics, the eigenvalues of an observable are the possible measurement outcomes for this observable. The dynamics is governed by the Hamilton operator, whose eigen- values define the energy levels. This relation between quantum mechanics and linear algebra, or more general functional analysis, is even tighter in the subfield of adiabatic quantum mechanics. In general, if the Hamiltonian changes in time, then initial eigenstates need not evolve into instantaneous eigenstates of the Hamiltonian at a later time. However, this could be realized for sufficiently slow change of the Hamiltonian [7, 16], known as the adiabatic approximation. The interest in adiabatic quantum mechanics gained momentum with the discovery of the geometric phase by Berry [3]. He showed that the phase picked up by an eigenstate upon varying parameters along a closed loop has in addition to the usual dynamical contribution a purely geometric contribution. This geometric phase depends only on the traversed loop, i.e., it is invariant under reparametrization of the path. Geometric phases have been found in numerous physical systems, see the overviews in [8, 9, 31], where in fact the earliest observation mailto:e.j.pap@rug.nl mailto:h.waalkens@rug.nl http://www.rug.nl/staff/e.j.pap/ http://www.rug.nl/staff/h.waalkens/ mailto:d.boer@rug.nl http://www.rug.nl/staff/d.boer/ https://doi.org/10.3842/SIGMA.2022.003 2 E.J. Pap, D. Boer and H. Waalkens of such a phase had been reported by Pancharatnam [26] in polarized light. On the theoretical side, the geometric phase was readily identified to be the holonomy of a bundle of eigenstates over system parameters [32]. In addition, generalizations have been studied, e.g., for degenerate Hamiltonians [37] or for any so called cyclic state returning to its original ray in a Hilbert space [1]. Geometric phase is also studied for non-Hermitian Hamiltonians, which allow for gain and loss of energy. Using left-eigenstates instead of bra’s, a generalization of Berry’s phase for cyclic eigenstates of a non-degenerate non-Hermitian Hamiltonian was presented in [12], and this was related to a geometric model in [21]. Another prominent feature of a non-Hermitian Hamiltonian family is that, when one follows a loop in the space of system parameters, the energies may swap places, which renders the evolution non-cyclic. This effect indicates a non-trivial topology of the energy bands; they need not be separated as in the Hermitian case. The interchange of energies can be related to degeneracies in the space of system parameters, which are known as exceptional points (EPs). These were first mentioned, with slightly different meaning, in [17], and have been found in many experimental setups, see the overviews, e.g., in [14, 23]. Many of these experiments revolve around parity-time or PT symmetric systems. Such systems are typically described by non-Hermitian Hamiltonians, and the breaking of this symmetry is often associated with an EP. Although PT -symmetry is not necessary for EP theory, it does provide experimentally accessible realizations. We remark that this phenomenology is not adiabatic in the standard way, as non-adiabatic effects cannot be avoided. The adiabatic theorem in the non-Hermitian case only holds for the state with highest relative gain [25]. This leads to significant limitations on how well the state exchange around an EP can be measured [5, 36]. A solution is to consider quasi-static set-ups; the path in parameter space is discretized into points, and one measures per configuration, see also [15, 22]. In this paper we introduce a geometric formalism to properly describe the adiabatic evolution of eigenstates of any non-degenerate finite-dimensional operator playing the role of a Hamilto- nian. In particular, the eigenstates need not be cyclic and the Hamiltonian need not be Her- mitian. To this end we introduce a bundle that directly follows from the eigenvalue problem and consists of triples “matrix-eigenvalue-eigenvector”. This bundle will not be principal, rather it has the structure of a semi-principal bundle as rigorously defined and studied in [28]. It is naturally equipped with a connection, whose parallel transport corresponds to the generalized geometric phase. As a result, both the geometric phase and the swaps of eigenstates arising from EPs can be described in a single holonomy description. In fact, the same bundle also allows for the incorporation of the non-geometric dynamical phase. We also treat its associated “frame” bundle, which is principal and provides a rigorous geometric argument behind the matrices used to describing state evolution. The condition of non-degeneracy of the operator will be crucial, and hence we study the space this defines in Section 2. The geometry of the eigenvalues of these operators we study in Section 3. This will yield a natural way to treat the interchanges of energies around EPs. In Section 4, this is extended to include eigenvectors, which yields a natural parallel transport theory for geometric phases, also in the presence of EPs. Following up on this, we show in Section 5 how these physical phenomena can be interpreted via a single holonomy description. We finish with a discussion in Section 6. 2 The non-degeneracy space We start with the mathematical objects that will provide the basis for all our arguments con- cerning eigenvalues. At this moment, we do not consider any specific operator family, instead only a complex vector space V of finite dimension n is given. An important remark is that we A Unified View on Geometric Phases and Exceptional Points 3 do not endow V with an inner product; we only use the topological vector space and manifold properties of V . Let us establish some notation by reviewing the following definitions and facts. We recall that endomorphisms or operators, i.e., linear maps A : V → V , form the space End(V ), which is a complex manifold of complex dimension n2. The set of all eigenvalues of an operator A is called the spectrum of A, and we denote it as Spec(A). If (v1, . . . , vn) is a frame, or basis, of V , we call it an eigenframe of A if each vi is an eigenvector of A. The set of all bases of V we denote as Fr(V ), and the set of all eigenframes of A by EigFr(A). Any basis (v1, . . . , vn) of V has a dual basis ( θ1, . . . , θn ) of the dual space V ∨ defined by the condition θi(vj) = δij = { 1 if i = j, 0 if i ̸= j. An eigencovector of A is a non-zero covector θ such that θA = λθ, where λ ∈ C is necessarily an eigenvalue of A. An eigencoframe of A is a frame of V ∨ consisting of eigencovectors of A. One can verify the following facts about eigenframes. Lemma 2.1. Let A ∈ End(V ), then (1) A has an eigenframe if and only if A is diagonalizable, (2) (v1, . . . , vn) is an eigenframe of A if and only if the dual basis ( θ1, . . . , θn ) is an eigen- coframe. We will now focus on the operators A which are non-degenerate, i.e., which have n distinct eigenvalues. The subspace in End(V ) of non-degenerate operators we denote by N(V ). We will go over different ways to formally define this subspace. First, we inspect an algebraic argument, which allows for a straightforward result on the manifold properties of N(V ). Afterwards, we consider other formulations of non-degeneracy, which will naturally guide us to symmetries and bundle properties of N(V ). 2.1 Discriminant definition The first characterisation of N(V ) we will inspect is based on the discriminant, and will form our algebraic definition of N(V ). Naturally, the eigenvalues of A ∈ End(V ) are the zeros of the characteristic polynomial p(A, z) of A, and whether or not the zeros of a polynomial are distinct can be inferred from the discriminant. That is, one first has the map p : End(V ) → C[z], A 7→ p(A, z) := det(zI −A) and by evaluating the discriminant one obtains a composite function d : End(V ) → C, A 7→ discrim(p(A, z), z) known as the discriminant of operators. Its zero set ∆(V ) = { A ∈ End(V ) | d(A) = 0 } is called the discriminant set and consists of all matrices that are degenerate. Clearly N(V ) is the complement of the discriminant set in End(V ), which yields our formal definition of N(V ). 4 E.J. Pap, D. Boer and H. Waalkens Definition 2.2 (non-degeneracy space). Given a finite-dimensional complex vector space V , its space of non-degenerate operators, or non-degeneracy space, is N(V ) = { A ∈ End(V ) | d(A) ̸= 0 } . It is well-known that d becomes a polynomial function in the matrix elements of the operator. Hence one may readily conclude the following. Lemma 2.3. The space N(V ) is an (algebraic) open and dense subset of End(V ). In particular, N(V ) is a submanifold of complex dimension n2 and real codimension 2. We next consider alternative characterizations of the non-degeneracy space. 2.2 Parametrizing the non-degeneracy space Another way to describeN(V ) is by explicitly parametrizing its elements. One such parametriza- tion readily follows from the fact that any A ∈ N(V ) is similar to a diagonal matrix where the diagonal entries are distinct. In other words, there is a frame f̃ = (f1, . . . , fn) of V so that the matrix of A w.r.t. f̃ takes the form diag(λ1, . . . , λn), where all the λi are distinct. We will denote the space of tuples λ̃ = (λ1, . . . , λn) of n distinct complex numbers as Cn; more background on this space can be found in Appendix A. We abbreviate diag(λ1, . . . , λn) as diag ( λ̃ ) . The above decomposition can be recast as the following parametrization. For convenience, let us identify a frame f̃ ∈ Fr(V ) with the map Sf̃ : C n → V given by (z1, . . . , zn) 7→ ∑n i=1 zifi, which is a linear isomorphism by definition of a frame. The parametrization of N(V ) is then formally described using the map Ξ: Fr(V )× Cn → N(V ),( f̃ , λ̃ ) 7→ Sf̃ diag ( λ̃ ) S−1 f̃ , which is a smooth surjection. Clearly, this map is not injective for two reasons; firstly, Ξ is indifferent concerning a non-zero scaling of the eigenvectors, and secondly, if we permute both the vectors of f̃ and the values in λ̃ we obtain the same operator. This is a symmetry of Ξ, which we phrase using group actions. First, writing the group of non-zero complex numbers as C×, the scaling symmetry is given by the (C×)n-action z̃ · ( f̃ , λ̃ ) = ( z̃f̃ , λ̃ ) , (2.1) where z̃f̃ is the entry-wise product (z1f1, . . . , znfn) ∈ Fr(V ). To describe the permutations, let us agree that a permutation acts on a tuple by permuting its entries, e.g., for σ ∈ Sn and a tuple λ̃ ∈ Cn, σ · λ̃ = ( λσ−1(1), . . . , λσ−1(n) ) . Then the permutation action on Fr(V )× Cn is simply σ · ( f̃ , λ̃ ) = ( σf̃ , σλ̃ ) . (2.2) The scaling and permutation actions merge into a single action of the wreath product C× ≀ In, where In := {1, . . . , n} is an index set with n elements. This group is the semi-direct product (C×)n ⋊ Sn, whose defining action of Sn on (C×)n is exactly the tuple permutation as we used above. The group multiplication of C× ≀ In reads (z̃1, σ1) · (z̃2, σ2) = ( z̃1(σ1z̃2), σ1σ2 ) . A Unified View on Geometric Phases and Exceptional Points 5 The action of the wreath product on Fr(V )×Cn is obtained by performing first the permutation and then the scaling, i.e., C× ≀ In × (Fr(V )× Cn) → Fr(V )× Cn, (z̃, σ) · ( f̃ , λ̃ ) = ( z̃ ( σf̃ ) , σλ̃ ) . (2.3) Let us make some remarks at this point. First, C× ≀ In has a faithful representation by complex generalized permutation matrices. These are matrices with a single non-zero complex number in every row and column, or, equivalently, products DP with D a diagonal matrix without zeros on the diagonal and P a standard permutation matrix. Given the conjugating nature of Ξ, the appearance of C× ≀ In is natural, as the generalized permutation matrices form the stabilizer of the diagonal matrices. Let us also refer to some earlier reports on the wreath product in combination with this parametrization. For example, in [38] there is firstly a quotient by scaling and secondly the quotient by permutations. Although not reported as such, we recognize the wreath product there. In [19, Lemma 1.1], it is shown that a restriction of Ξ defines a principal bundle for a wreath product group, where the number field need not be C. As we will argue now, the parametrization Ξ also defines a principal C× ≀ In-bundle. Let us use that a projection is a principal bundle if it is given by the quotient of a free and proper action. The latter is readily verified, hence it remains to be shown that Ξ coincides with this projection. Lemma 2.4. There is a principal bundle C× ≀ In Fr(V )× Cn N(V ).Ξ Proof. As stated, it suffices to show that Ξ and the action quotient map are isomorphic as bundle maps. As we already saw that the action preserves the fibers of Ξ, it only remains to show that each group orbit exhausts the fiber in which it lies. Therefore, assume Ξ ( f̃ , λ̃ ) = Ξ ( f̃ ′, λ̃′ ) =: A. Then both λ̃ and λ̃′ constitute Spec(A), hence they are related by a permutation σ ∈ Sn as λ̃ = σλ̃′. Then fi needs to be parallel to f ′σ−1(i) for each i, so f̃ and σf̃ ′ are equal up to an element of (C×)n. However, then ( f̃ , λ̃ ) and ( f̃ ′, λ̃′ ) differ only up to an element of C× ≀ In. ■ This result has various consequences. First, it shows that N(V ) can be realized as a quotient space. Another important observation concerns the existence of local sections of Ξ. These sections provide local moving eigenframes and corresponding smooth eigenvalues, as we will use in Section 4. We state the details in the following corollaries. Corollary 2.5. Identifying Fr(V ) with GL(n,C), the space of non-degenerate operators is rea- lized as a quotient space as N(V ) ∼= GL(n,C)× Cn C× ≀ In . Corollary 2.6. Let A0 ∈ N(V ). There is a neighborhood U of A0 on which one has smooth local eigenvalue functions λ1, . . . , λn : U → C, exhausting the spectrum at each point, and smooth local eigenvector functions f1, . . . , fn : U → V . That is, for all A ∈ U and i ∈ In one has Afi(A) = λi(A)fi(A) such that Spec(A) = {λ1(A), . . . , λn(A)} and (f1(A), . . . , fn(A)) is a basis of V . Moreover, the tuple (λ1(A), . . . , λn(A)) may be taken to be any given ordering of Spec(A), and the tuple (f1(A), . . . , fn(A)) any eigenframe of A following the same ordering. Proof. The functions fi(A) and λi(A) are the components of a local section s : U → Fr(V )×Cn, which can be taken through an arbitrary point above A. ■ 6 E.J. Pap, D. Boer and H. Waalkens 2.3 Spectrum map on non-degeneracy space Non-degenerate operators can also be characterized based on their spectrum. Indeed, an operator A ∈ End(V ) is non-degenerate if and only if Spec(A) consists of n distinct elements. In other words, Spec(A) should not be any subset of C, but belong to the set (C n ) of all subsets of C consisting of n distinct elements. Taking the spectrum can thus be written as a map Spec: N(V ) → ( C n ) , A 7→ Spec(A). We will now continue by showing that this map Spec defines a fiber bundle. In this way, we find N(V ) realized as a total space instead of a base space. In addition, the map Spec will reappear when we discuss EPs; formally it is this map that associates the change in spectrum to a change in Hamiltonian. To start, let us verify that Spec is smooth. For the manifold structure on (C n ) , we follow the idea that the space (C n ) can be obtained from Cn by reducing an (ordered) tuple to the (unordered) set of its elements. We write q : Cn → (C n ) for the quotient map, and use the manifold structure on (C n ) for which q defines a principal Sn-bundle with the standard permutation action of Sn on Cn. For more details, we refer to Appendix A. We can see that Spec is smooth using a geometric argument, based on the projection q and the map Ξ from the previous part. The key observation is the equality Spec ( Ξ ( f̃ , λ̃ )) = q ( λ̃ ) for any ( f̃ , λ̃ ) ∈ Fr(V ) × Cn; by construction λ̃ lists the eigenvalues of the operator Ξ ( f̃ , λ̃ ) . In other words, we have the following commutative diagram: Fr(V )× Cn Cn N(V ) (C n ) . prCn Ξ q Spec (2.4) Smoothness of Spec is now clear; as the upper route in the diagram is smooth and Ξ is a surjective submersion the claim follows. Let us continue by discussing the model fiber of Spec. This is facilitated by viewing operators according to their spectral decomposition. That is, any A ∈ N(V ) can be written as a sum∑n i=1 λiPi, where λi ∈ Spec(A) and each Pi : V → V is a projection. Geometrically, this expresses that A ∈ N(V ) is completely specified by the pairs of eigenvalue and the corresponding eigenrays. Hence, if we fix the spectrum, then only the choice of eigenrays remains. This means that any fiber of Spec is diffeomorphic to this space of possible choices of eigenrays, or, equivalently, to the space of suitable tuples of projectors. The latter space we can describe in more detail. Clearly, each Pi projects to a one-dimensional subspace in V , and together these projectors satisfy PiPj = δijPi and ∑n i=1 Pi = idV . That is, the tuple (P1, . . . , Pn) must lie in the space PFr(V ) := { (P1, . . . , Pn) ∈ End(V ) | PiPj = δijPi, dimC(im(Pi)) = 1 } , which is thus also the model fiber for Spec. One can think of PFr(V ) as the space of all resolutions of the identity which are compatible with non-degenerate operators, i.e., those corresponding to n rays in V . Clearly, PFr(V ) relates to Fr(V ) by sending a basis (f1, . . . , fn), with dual basis( θ1, . . . , θn ) , to the tuple ( f1θ 1, . . . , fnθ n ) ∈ PFr(V ). This map is surjective, but not injective as individual scaling of the basis vectors will yield the same projectors. Hence we find that Fr(V )/(C×)n is canonically isomorphic to PFr(V ), and it is the former form that we will see in the upcoming proofs. A Unified View on Geometric Phases and Exceptional Points 7 At this point we have discussed the relevant spaces, but did not yet discuss their symmetry. This symmetry can be found from the observation that operators with the same spectrum differ by a similarity transformation. Let us implement these transformations in the language of the action of GL(V ) on N(V ) given by GL(V )×N(V ) → N(V ), S ·A = SAS−1. (2.5) The map Spec is invariant w.r.t. this action by the familiar rule Spec ( SAS−1 ) = Spec(A). Hence each fiber of Spec is a GL(V )-manifold as well, and so we wish to view PFr(V ) as a GL(V )- manifold, also equipped with the conjugation action. Observe that this action on PFr(V ) is naturally inherited from Fr(V ), on which it reads S · (f1, . . . , fn) = (Sf1, . . . , Sfn). We thus arrive at the following result. Lemma 2.7. The GL(V )-action on N(V ) is transitive on the fibers of Spec. Moreover, any fiber of Spec is isomorphic to PFr(V ) as GL(V )-manifolds. Proof. As any two non-degenerate operators with the same spectrum differ by a similarity transformation, the action is transitive on the fibers of Spec. Hence, every fiber of Spec is a homo- geneous GL(V )-space. The stabilizer subgroup at A ∈ N(V ) consists of the maps that preserve all eigenrays of A individually, hence is isomorphic to (C×)n. In order to parametrize this fiber, let λ̃ be an ordering of Spec(A), and consider Ξ restricted to the subset Fr(V ) × {λ̃}. Clearly, this surjects on the fiber of Spec containing A. Moreover, Ξ is equivariant w.r.t. the canonical GL(V )-action on Fr(V ) × Cn. Hence, the fiber of Spec containing A is isomorphic, as GL(V )- manifold, to the quotient of Fr(V ) by the stabilizer. The stabilizer (C×)n is now straightforward; it appears via the (C×)n-action given in equation (2.1). Hence we found a GL(V )-equivariant isomorphism to Fr(V )/(C×)n, and so to PFr(V ). ■ The model fiber of Spec is thus PFr(V ), which we view as a GL(V )-manifold, emphasizing its close relation to similarity transformations. We thus wish to prove that Spec defines a fiber bundle that respects the GL(V )-action. That is, Spec defines a GL(V )-manifold bundle in the language of [28]; both fiber and total space are endowed with a GL(V )-action, and local trivializations can be taken GL(V )-equivariant. We then arrive at the following statement, which summarizes the results of this section. Proposition 2.8. The spectrum map induces the GL(V )-manifold bundle PFr(V ) N(V ) (C n ) . Spec Proof. Pick a point {λ1, . . . , λn} ∈ (C n ) , and let U ⊂ (C n ) be a neighborhood of {λ1, . . . , λn} on which a local section s : U → Cn of q is defined. Consider the map U × Fr(V ) → N(V ),( u, f̃ ) 7→ Ξ ( f̃ , s(u) ) , which for each u ∈ U surjects on the fiber of Spec above u. By applying Lemma 2.7 fiber-wise we find that the reduced map U × Fr(V )/(C×)n → N(V ) is well-defined. In fact, if we look at the spectral decomposition, this map pairs the tuple (P1, . . . , Pn) ∈ PFr(V ) determined by f̃ with the values in the tuple s(u). We thus obtained a GL(V )-equivariant local trivialization of Spec around {λ1, . . . , λn}, hence the claim follows. ■ 8 E.J. Pap, D. Boer and H. Waalkens 2.4 Summarizing diagram If one takes the bundles defined by Ξ and Spec plus the diagram in (2.4), then one readily obtains the following diagram of bundle sequences: (C×)n C× ≀ In Sn Fr(V ) Fr(V )× Cn Cn PFr(V ) N(V ) (C n ) . prCn Ξ q Spec (2.6) The direct product Fr(V ) × Cn we view as a bundle over Cn. The remaining bundles are straightforward; on top is the defining decomposition of the wreath product C× ≀In, on the right the quotient map q, and on the left the quotient realization of PFr(V ). Observe that all rows and columns are group-space bundles, i.e., each one is related to a group action. With the exception of Spec all bundles are principal; Spec itself defines a bundle of homogeneous GL(V )-spaces. 3 Eigenvalue bundle and exceptional points Let us study the geometry that describes how eigenvalues and eigenvectors depend on the op- erator. We will find that eigenvalues and eigenvectors form bundles over the non-degenerate operators. This space N(V ) of non-degenerate operators will reappear as the region free of singularities. In this way, N(V ) shows us where we can use results from geometry, in partic- ular concerning parallel transport, which in turn provides a framework for adiabatic quantum mechanics. We will deal with the eigenvalues in this section, and provide an extended similar argument in the next section concerning the eigenvectors. 3.1 The spectrum bundle We start by describing a natural abstract model for the energy bands. Namely, when studying EPs, we want to follow eigenvalues as a function of the operator. We wish to view this in a geometric way. For example, we wish to view an eigenvalue function λ = λ(A), with λ(A) an eigenvalue of the operator A, as a local section. Note that such functions λ(A) are necessarily local; otherwise EPs could not exist. We will call such functions local eigenvalues. We quickly come to the conclusion that we should restrict the operators. Namely, degenerate operators pose a problem as they will form singularities. Indeed, around a degenerate energy, the energy bands do not resemble a smooth manifold. On the other hand, for A ∈ N(V ) such issues do not occur. As the following lemma shows, the implicit function theorem yields that simple eigenvalues always admit an extension to a local eigenvalue. An immediate consequence is that the restriction from End(V ) to N(V ) is minimal in order to obtain a smooth structure. Lemma 3.1. Given A ∈ End(V ), then A ∈ N(V ) if and only if ∂p ∂z (A, λi) ̸= 0, ∀λi ∈ Spec(A). We can now formalize the bundle which has the local eigenvalues as its local sections. That is, its local sections are of the form A 7→ (A, λ(A)) with λ(A) a local eigenvalue. This could be used to define the bundle bottom-up, but we prefer to use the following more explicit top-down method. Clearly, the total space of the bundle consists of all pairs (A, λ) such that A ∈ N(V ) A Unified View on Geometric Phases and Exceptional Points 9 and λ ∈ Spec(A). We observe that this set is the zero set of the characteristic polynomial map p, restricted to non-degenerate operators. This viewpoint will form our primary definition1 of the bundle, because of its algebraic convenience. We use the name spectrum bundle: the fiber above A ∈ N(V ) is simply Spec(A), and the term bundle we will justify in Theorem 3.4. Definition 3.2 (spectrum bundle). Given the vector space V , define its spectrum bundle to be the space Spec(V ) = { (A, λ) ∈ N(V )× C | p(A, λ) = 0 } . Furthermore, we write πλ : Spec(V ) → N(V ) for the projection (A, λ) 7→ A. Our first step in proving the bundle property of πλ is showing that Spec(V ) is a smooth manifold. This readily follows from the derivative characterization of N(V ) in Lemma 3.1. Proposition 3.3. The space Spec(V ) is a closed submanifold of N(V )× C of complex dimen- sion n2. Proof. Consider the restricted characteristic polynomial p : N(V ) × C → C. As Spec(V ) is the zero set of this map, by the Submersion Theorem it suffices to show that 0 is a regular value. Hence we consider the differential dp(A, z), which contains the term ∂p ∂z (A, z)dz. This is a surjection whenever ∂p ∂z (A, z) ̸= 0, which holds on all of Spec(V ) by Lemma 3.1. Hence Spec(V ) is a closed submanifold, of the same dimension as N(V ). ■ We are now in place to deduce the bundle structure of πλ. A fiber is of the form Spec(A) with A ∈ N(V ), hence by definition a set of n distinct points. We thus take the model fiber to be In. Inspection of πλ yields the following argument. Theorem 3.4. The map πλ : Spec(V ) → N(V ) defines a fiber bundle with model fiber In. Proof. First, πλ is a surjective submersion; the implicit function theorem provides local eigen- values, and so local sections, through any point of Spec(V ). By dimension count, πλ is a local diffeomorphism. As πλ is also proper, it is a covering map [18]. Then, as each fiber has exactly n elements, πλ is an In-bundle. Indeed, a local trivialization is a map of the form ϕ : U × In → Spec(V )|U , (A, i) 7→ (A, λi(A)) (3.1) with λ1, . . . , λn distinct local eigenvalues. ■ This result formalizes the idea that Spec(V ) is locally the union of the graphs of n local eigenvalues. We remark that the inverse of the local trivialization ϕ in equation (3.1) above can be written as (A, λ) 7→ (A,#(A, λ)), where #: Spec(V )|U → In yields the label of the graph in which a point lies. This map # one can interpret as the “local labeling” induced by the chosen local eigenvalues. We will see this map again when discussing the eigenvectors in Section 4. 3.2 Geometry behind swaps of energies We will now treat how the abstract theory discussed above facilitates the study of instantaneous energies in adiabatic quantum mechanics, including the swaps of energies related to exceptional points (EPs). We assume that instantaneous eigenstates will remain instantaneous eigenstates. 1We remark the similarity with the space M in [21]. However, as we explicitly list the eigenvalue as a coor- dinate, it is not a multi-valued function here. 10 E.J. Pap, D. Boer and H. Waalkens Hence instantaneous energies are well-defined. The change of energy in time we will treat formally using covering theory. To start, we assume that an experimental set-up is captured by a Hamiltonian operator. Typically, this Hamiltonian depends on the available system parameters, e.g., cavity size and field strength in optics and photonics (see, e.g., the review in [23]). This induces a manifold of system parameter values, which we denote by M . We thus obtain a family of experimental set-ups, and so a family of Hamiltonians, described by a map H : M → End(V ), which sends a configuration of system parameters x ∈M to the HamiltonianH(x) corresponding to that configuration. For simplicity, we will assume that M and H are smooth. We will refer to H as the Hamiltonian family, where each H(x) is a Hamiltonian operator on V . We do however not require the Hamiltonian operators to be Hermitian. We refer to the eigenvalues as energies and to the eigenvectors as eigenstates. The idea is now to vary the system parameters. In practice this means we follow a path γ in M , whose initial point x0 serves as a reference. This results in the time-dependent Hamilto- nian H(γ(t)). A state ψ(t) is called an instantaneous eigenstate at time t if it satisfies the eigenvalue problem H(γ(t))ψ(t) = E(t)ψ(t), with E(t) an energy of H(γ(t)). The idea of the adiabatic approximation is that this relation is preserved in time, at least approximately. However, this means one first has to make sure that the function E(t) is well-defined. This fundamental fact can now easily be deduced from the covering properties of Spec(V ). First, we remark that in order to unambiguously follow a specific energy level of H(γ(t)), the energy level may not become degenerate at any time. Hence we require H(γ(t)) to be non- degenerate for all t, i.e., t 7→ H(γ(t)) is a path in N(V ). Let E0 be the energy level of the initial Hamiltonian H(x0) that we wish to follow in time. Clearly, this defines the point (H(x0), E0) in the fiber of Spec(V ) above H(x0). Hence, it specifies a unique lift of H ◦ γ to Spec(V ), which is of the form (H(γ(t)), E(t)) for some function E = E(t). In this way, we obtain the instantaneous energy for all relevant times in a formal way. The intuition of this lifting argument is to follow the initial energy along the energy bands of the family H. If H is fixed, it is convenient to consider these energy bands directly. This can be done by taking the pull-back of the bundle Spec(V ) → N(V ) along H. To obtain this pull-back, we must restrict the system parameters to those were H is non-degenerate, i.e., we must restrict M to the subspace N(H) := H−1(N(V )) = { x ∈M | H(x) has non-degenerate energy levels } . The energy bands of H, with the degenerate points omitted, then form the space Spec(H) := { (x,E) ∈ N(H)× C | E ∈ Spec(H(x)) } , i.e., the space of all pairs of a specific configuration x of the system parameters and an energy E of the system for this configuration. We call Spec(H) the spectrum bundle of H; it has a natural projection πHλ : (x,E) 7→ x to N(H), so that the fiber above x is simply Spec(H(x)). The bundle property of πHλ is guaranteed by the pull-back construction, and relates to πλ as given by the pull-back diagram below: Spec(H) Spec(V ) (x,E) (H(x), E) N(H) N(V ), x H(x). πH λ πλ H A Unified View on Geometric Phases and Exceptional Points 11 We see that the path γ can be lifted directly to Spec(H), which is then both an intuitive and formally correct way to obtain the function E(t) directly. Indeed, by assumption γ lies in N(H), the initial point is now (x0, E0), and the lift is of the form (γ(t), E(t)). We see that the study of Spec(H) is a straightforward generalization of Kato’s original set- ting [17]. There, H was assumed to depend analytically on a single complex variable, so that the energies were locally given by analytic functions. By analytically continuing them along a path around a degeneracy, the eigenvalues could return in a different order, and such a degen- eracy was called an EP. Here, we see that Spec(H) allows us to drop the assumption of analytic dependence; instead of analytic continuation we can use lifting along a covering map. Hence we broadened the perspective from complex analytic to general continuous dependence. Note that this is necessary to include Hermitian or PT -symmetric Hamiltonians, as these are defined by a conjugate-linear operation and hence do not fit in an analytic setting. The study of the permutations of energies in the family H is in this construction tantamount to the study of the covering properties of Spec(H). In fact, the permutations of energies ofH(x0) are naturally described by the monodromy action at x0, as we will argue now. First, as we wish to compare energies, we should restore the original system configuration at the end (see also [27] for a practical argument for this). That is, the path γ in N(H) should return to our reference x0 ∈ N(H). In other words, γ should be a loop based at x0. Let us write Loop(N(H), x0) for the set of such loops. We are then interested in how the energies of H(x0) return upon following them along γ. For an energy E ∈ Spec(H(x0)), this is given by the lift of γ to Spec(H); the final energy is the endpoint of the lift, which we write as pγ(E). Doing this for every energy of H(x0) yields the map pγ : Spec(H(x0)) → Spec(H(x0)). This map is a bijection/permutation of Spec(H(x0)); its inverse is given by traversing γ in the opposite direction. Going over all loops in Loop(N(H), x0) then yields the group of all possible permutations of Spec(H(x0)), namely the group Hol Spec(H) N(H) (x0) = { pγ ∈ Aut(Spec(H(x0))) | γ ∈ Loop(N(H), x0) } , where the automorphisms are simply bijections. This is indeed a holonomy group as lifting along a covering can be regarded as parallel transport. We remark that this group was called Λ(x0) in [27], which is a straightforward generalization of the permutation group studied by Kato [17] from complex analytic to continuous Hamiltonian families. Example 3.5. Let us consider a standard case, namely a family with EP2, where an EP2 is an EP permuting 2 energies. Explicitly, consider H : C → End ( C2 ) given by H(x) = ( 1 x x −1 ) . The eigenvalues are given by the multi-valued function √ 1 + x2, so N(H) = C \ {±i} and Spec(H) is the graph of this multi-valued function, which in this case is a Riemann surface. Let us take a branch cut, and write E±(x) for the two energy branches. Clearly, Spec(H(x0)) = {E+(x0), E−(x0)}, and if γ is a loop based at x0 encircling an EP once, it induces the map pγ : E±(x0) 7→ E∓(x0). Of course, if γ does not encircle any EP, we obtain pγ = idSpec(H(x0)). The association γ 7→ pγ brings us to a monodromy action as follows. The map pγ is invariant under continuous deformation of γ, i.e., it only depends on its homotopy class [γ]. The set of such homotopy classes of loops based at x0 is the (based) fundamental group π1(N(H), x0). The ex- change of energies can then be expressed by the homomorphism π1(N(H), x0) → Hol Spec(H) N(H) (x0), 12 E.J. Pap, D. Boer and H. Waalkens [γ] 7→ pγ , or equivalently as an action of π1(N(H), x0) on Spec(H(x0)). This is known as the monodromy action, which we write as π1(N(H), x0)× Spec(H(x0)) → Spec(H(x0)), [γ] · (x0, E) = (x0, pγ(E)). A picture to have in mind is Figure 1; following the energy level along γ defines a path in Spec(H), which may end at a different energy. An explicit example we treat in Example 3.6 below. We note that the group π1(N(H), x0) is typically non-trivial. Indeed, by Lemma 2.3 the space N(V ) has real codimension 2 in End(V ) and so a non-trivial fundamental group, which then holds for a generic N(H) as well. We remark that the relevance of the monodromy action for EPs was reported in [33, 34]. The space used there is either Spec(H) or a related space we obtain in Section 5. Figure 1. Exchange of energies around an EP is given by the monodromy of the chosen path. Plotted is the real part of the energy levels of the Hamiltonian family in Example 3.5. Upon changing system parameters (black circle) from the reference value (red dot below), the energy moves (blue line) along the sheets and does not return to itself. Such behavior only happens around special degeneracies, namely the EPs. Here the black + marks the EP of the system related to the drawn exchange. Example 3.6. Continuing Example 3.5, it follows that π1(N(H), x0) is isomorphic to the fun- damental group of the “figure 8”, which is a free group on two generators, for any reference x0. Picking x0 = 0, the monodromy action maps both generators of π1(N(H), 0) to the interchange of +1 and −1 in Spec(H(0)). The realization of Spec(H) as a pull-back of Spec(V ) allows us to distinguish between inherent geometric properties coming from Spec(V ) versus artifacts of the family H. Such artifacts can result in different, i.e., non-isomorphic, spaces Spec(H) for the different H. For example, we can explicitly add an artifact to a particular H : M → End(V ) by introducing a fictitious system parameter, i.e., an artificial variable that does not correspond to any change in the experimental set-up. This increases the dimensions of both M and Spec(H), but there is no additional physical information. Another example is to include multiple degenerate levels in H which do not depend on the system parameters. In this case N(H) becomes empty. Of course, these examples are highly artificial, yet we see it is advisable to be careful not to include any non-physical information in H. A Unified View on Geometric Phases and Exceptional Points 13 3.2.1 Merging path method A common way to demonstrate the existence of an EP in an experimental set-up is to follow a loop in parameter space and trace the eigenvalues accordingly. If the eigenvalues do not return to themselves, then there must be (at least one) EP structure inside the loop, provided that any path in the parameter space M can be contracted to a point. Such an argument can already be found in [13]. We observe that this method does not attribute a swap to a particular EP – these are not even present in N(H) – but rather to a homotopy class of loops, which in turn signals the presence of a degeneracy structure. In practice, the traversing of a loop is done in a stroboscopic way by measuring the energy for sufficiently close discrete parameter values (e.g., in [39]). The formal mathematics behind this argument is straightforward. Let us go through this merging path method step by step. The change of system parameters is given by a loop γ in N(H). The spectrum then changes according to the loop Spec(H(γ(t))) in (C n ) , with n the dimension of the state space. To extract the permutation of the energies, note that we cannot work in (C n ) alone as the (unordered) spectrum will return to itself regardless of the presence of EPs. We must thus lift the loop Spec(H(γ(t))) to Cn in order to observe and extract a permutation. This argument translates directly into the diagram Cn N(H) N(V ) (C n ) ,H Spec (3.2) which is the bottom right corner of diagram (2.6) plus the adaptation to the specific system defined by H. 3.2.2 Real eigenvalue case The space Spec(H) can have additional properties if the family H satisfies additional assump- tions. One interesting assumption onH is thatH(x) has real energies for all x ∈M . This occurs, e.g., when H(x) is always Hermitian w.r.t. some given inner product on V , or when all H(x) have exact PT symmetry (see [24] on the relation between these). In this case, Spec(H) is trivial over N(H), which means that all energy bands are globally disconnected. In particular, no exchanges can occur and EPs cannot be present. We first prove this on the abstract level, and then use the pull-back by H to go to Spec(H). We remark that the argument also holds if the eigenvalues are taken in another totally ordered subset of C. Proposition 3.7. Write R(V ) for the subspace of N(V ) of matrices with real eigenvalues. The bundle Spec(V ) restricted to R(V ) is trivial. Proof. If A ∈ R(V ), then there is a unique ordering of Spec(A) by indexing the eigenvalues from lowest to highest λ1(A) < · · · < λn(A). This establishes a global labelling map Spec(V )|R(V ) → R(V )× In defined as (A, λi) 7→ (A, i), which is a homeomorphism. ■ Corollary 3.8. If H(x) has real eigenvalues for all x ∈ N(H), then Spec(H) is a trivial bundle. In particular, there are n distinct energy bands E1(x), . . . , En(x) defined continuously on all of N(H), and the family H has no EPs. 4 Eigenvector bundle and geometric phases In the previous section we showed how the energies of an adiabatic quantum system, in particular their exchanges around EPs, can be treated using the covering space Spec(V ). We will now extend this formalism to eigenstates, where again non-Hermitian Hamiltonians are allowed. 14 E.J. Pap, D. Boer and H. Waalkens 4.1 The eigenvector bundle Let us start by studying how eigenvectors vary with the operator. We will use the following notation. The set of all eigenvectors of an operator A ∈ End(V ) corresponding to an eigenvalue λ ∈ Spec(A) we denote as Eigλ(A) = { v ∈ V \ {0} | Av = λv } = ker(A− λI) \ {0}. We call this the space of eigenvectors, or eigenvector space, of A corresponding to λ. This terminology emphasizes that the zero vector is excluded. In this way, each Eigλ(A) is a free C×-manifold; any non-zero multiple of an eigenvector is again an eigenvector. Similarly, we define the eigenvector space Eig(A) of A to be the set of all eigenvectors of A. Naturally, Eig(A) = ⊔ λ∈Spec(A) Eigλ(A), which is immediately the partition of Eig(A) into its connected components. We see that Eig(A) is also endowed with C×-scaling, but need not be a manifold as the eigenvector spaces of the eigenvalues need not have the same dimension. However, in case A ∈ N(V ) clearly Eig(A) is a C×-manifold of complex dimension 1, and the above decomposition shows Eig(A) as a union of C×-torsors. Here we already found a hint that also concerning eigenvectors we need to restrict to non- degenerate operators. Indeed, if we wish to view eigenvectors varying with the operator as local sections of a bundle, i.e., locally defined functions v(A) so that v(A) ∈ Eig(A), then degenerate operators will again give rise to singularities. We encounter the same situation as before, be it with more facets. The problem with eigenvalues was that the number of distinct ones suddenly drops at a degenerate operator. Here, this implies that Eig(A) suddenly consists of less than n connected components, which renders a bundle structure impossible. In addition, we see for degenerate yet diagonalizable operators that the eigenvector space can consist of parts with different dimension, hence Eig(A) is not even a manifold for such operators. Hence we again restrict to non-degenerate operators. The space that we obtain this way we state in a definition because of its importance. Definition 4.1 (eigenvector bundle). Given the vector space V , define its eigenvector bundle to be the space Eig(V ) = { (A, λ, v) ∈ Spec(V )× V \ {0} | v ∈ Eigλ(A) } . Again, we have used the term bundle straightaway, and in the remainder of this section we will justify this term. In fact, we will find that Eig(V ) is a bundle with respect to two different projections. The idea is that we can view Eig(V ) as a union of total eigenvector spaces or as a disjoint union of individual eigenrays; Eig(V ) = ⊔ A∈N(V ) Eig(A) = ⊔ (A,λ)∈Spec(V ) Eigλ(A). Both of these viewpoints come with their own projection. For the first, one projects (A, λ, v) 7→A, which we write as the map πλv : Eig(V ) → N(V ). For the second, one only omits the eigenvector and projects to (A, λ), which defines a map πv : Eig(V ) → Spec(V ). These two projections are related by the projection πλ, as summarized in the diagram below: Eig(V ) Spec(V ) (A, λ, v) (A, λ) N(V ), A. πv πλv πλ (4.1) A Unified View on Geometric Phases and Exceptional Points 15 Our first step in proving the bundle claims is to show that Eig(V ) is a smooth manifold. Clearly Eig(V ) is the zero set of P : Spec(V )× V \ {0} → V, (A, λ, v) 7→ (λI −A)v. In this way we obtain the following structure on Eig(V ). Proposition 4.2. The space Eig(V ) is a closed algebraic C×-submanifold of Spec(V )×V \ {0} of complex dimension n2 + 1. Moreover, the projection maps πλv and πv are smooth. Proof. Pick a point (A0, λ0, v0) ∈ Eig(V ) and pick local coordinates (A, λ) on Spec(V ) around (A0, λ0); this implicitly defines λ = λ(A) with λ(A0) = λ0. Consider now the differential dP = (λ(A)I −A)dv + d(λ(A)I −A)v. As P is constant on the C×-orbits, dP has rank at most n − 1. On the other hand, the term (λ(A)I − A)dv has the same rank as λ(A)I − A, which is n − 1 by non-degeneracy. Hence dP has constant rank n − 1, and so by the constant rank theorem [18] the fibers of P are closed submanifolds of (complex) dimension ( n2 + n ) − (n − 1) = n2 + 1. The maps πλv and πv are then restrictions of smooth projections to a submanifold, hence smooth. ■ Both πλv and πv are invariant w.r.t. the C×-action on Eig(V ), which hints at a principal bundle structure. For πv, this holds; its fibers are eigenrays Eigλ(A), which are C×-torsors, and πv coincides with the quotient of the action on Eig(V ). However, πλv is clearly not principal. The fiber of πλv above A ∈ N(V ) is the total eigenvector space Eig(A) of A, which is a union of C×-torsors, hence not a C×-torsor itself for n > 1. Still, with the eye on adiabatic quantum mechanics, we are motivated to describe πλv in more detail. We thus turn to a wider class of bundles, which extends the principal bundles. In short, principal G-bundles are bundles endowed with a G-action so that the fibers are G-torsors (hence diffeomorphic to G) and the projection admits G-equivariant local trivializations. Let us only change the model fiber; we allow it to be a semi-torsor, i.e., a disjoint union of torsors (hence diffeomorphic to G× I for some index set I). Bundles satisfying this more general condition we call semi-principal G-bundles [28]. Clearly, as any torsor is a semi-torsor, this is an extension of the principal bundles. In case the semi-torsor consists of n torsors, where n is finite, we write the model fiber as G× In. More details on semi-principal bundles can be found in [28]. Let us show that πλv : Eig(V ) → N(V ) is a semi-principal C×-bundle with model fiber C×× In. The main property left to show is the equivariant local triviality, for which we prepare ourselves with the following extension property. We use the language of smooth maps, but they may be taken to be algebraic. Lemma 4.3. For every point A ∈ N(V ), there is an open neighborhood U ⊂ N(V ) of A and n eigenvector functions v1, . . . , vn : U → V which constitute an eigenframe at each point of U . Moreover, this moving eigenframe can be chosen to extend any eigenframe of A. Proof. The local eigenvector functions exist by Corollary 2.6. The eigenframe they form at A can be changed to any given eigenframe of A by acting with the appropriate group element in C× ≀ In, as we did in the proof of Lemma 2.4. ■ Theorem 4.4. The map πλv : Eig(V ) → N(V ) defines a semi-principal C×-bundle with model fiber C× × In. 16 E.J. Pap, D. Boer and H. Waalkens Proof. Pick A0 ∈ N(V ), let U ⊂ N(V ) be a neighborhood of A0 and for i = 1, . . . , n let vi : U → V be a local eigenvector corresponding to eigenvalue λi(A0), according to Lemma 4.3. One has a map Φ: U × In × C× → Eig(V )|U , (A, i, z) 7→ (A, λi(A), zvi(A)), which is smooth and clearly intertwines the C×-actions. Its inverse is (A, λ, v) 7→ ( A,#(A, λ), [ v/v#(A,λ)(A) ]) , which satisfies the same properties. Here # is the local labelling, and the last entry denotes the scalar z such that v = zv#(A,λ)(A). It follows that Φ is a C×-equivariant local trivialization, hence the bundle property follows. ■ We observe that πλ and πv are closely tied to πλv. Namely, it holds that every semi-principal G-bundle decomposes naturally into a principal G-bundle and a covering space [28]. These maps form exactly such a decomposition. Proposition 4.5. The rule πλv = πλ ◦ πv is the decomposition of πλv in a principal bundle projection followed by a covering map. 4.2 Connection on the eigenvector bundle Our next step is to show that Eig(V ) supports a canonical connection. This connection is compatible with both the semi-principal projection πλv and the principal projection πv. Hence, we prefer to avoid the term “principal connection”, and make reference only to the action instead. We will thus use the concept of a G-connection defined as follows. Let G be a Lie group with algebra g; a 1-form ω on a G-manifold M is a G-connection if ωm ◦ am = idg for all m ∈ M , with am the infinitesimal action at m, and L∗ gω = Adg(ω) for any g ∈ G. See also [28] on this. To describe the connection 1-form on Eig(V ), it is helpful to extend our notation. The set of eigencovectors of A ∈ End(V ) we denote as Eig∨(A), and those corresponding to a specific eigenvalue λ by Eig∨λ(A). Note that the zero covector is excluded. As we saw in Lemma 2.1, for diagonalizable A, there is a bijection between eigenframes and eigencoframes. For non- degenerate A a stronger statement holds, namely, the bijection already appears on the level of individual eigenvectors and eigencovectors. Lemma 4.6. For a non-degenerate matrix A, there is a canonical bijection Eig(A) → Eig∨(A) sending v ∈ Eigλ(A) to the unique covector θ ∈ Eig∨λ(A) defined by θ(v) = 1. Proof. Following the decomposition of V by the eigenspaces of A, define the covector θ to vanish on every eigenspace except the one of λ, where it is fixed by setting θ(v) = 1. ■ This correspondence shows that the scaling action on Eig(A) corresponds to inverse scaling on Eig∨(A), i.e., we obtain the C×-action on Eig∨(A) given by z · θ = θz−1. The bijection Eig(A) → Eig∨(A) is then an isomorphism of C×-manifolds. In addition, as any triple (A, λ, v) ∈ Eig(V ) defines a unique θ, we are free to augment the triple to (A, λ, v, θ). This will be a convenient notation when defining maps on Eig(V ), such as the connection 1-form we seek at the moment. Similarly, one is free to omit λ from the triple. We will use these alternative notations frequently. A Unified View on Geometric Phases and Exceptional Points 17 We are now set to describe the connection 1-form on Eig(V ). By definition, this 1-form is a right-inverse of the infinitesimal action, which is specified by the vector field ∂ ∂v ∣∣∣∣ (A,v) := d dt ∣∣∣∣ t=0 exp(t) · (A, v) = d dt ∣∣∣∣ t=0 ( A, etv ) , which can be thought of as a unit vector field pointing along the eigenray of A through v. Hence, given a path Γ(t) = (A(t), λ(t), v(t), θ(t)) in Eig(V ), the connection should measure the component of v̇ along v. This cannot be done with the vector space structure of V alone, but given the operator this is possible. Indeed, the eigenrays of A(t) provide a natural decomposition of V at time t. According to this choice, the coefficient of v̇ along v is simply θ(v̇). Let us describe this argument formally. To obtain the quantity v̇ from Γ̇, we can use the differential of the projection prV : Eig(V ) → V , (A, λ, v) 7→ v. Technically, this will take values in the tangent space T(A,v)V , but this can be identified with V in a canonical way. Let us write dv for (prV )∗ with the image viewed in V . This map can be followed up with θ, which yields the connection 1-form. Proposition 4.7. The 1-form ω ∈ Ω1(Eig(V ),C) given by ω = θdv is a C×-connection. Proof. As C× is commutative, the equivariance of ω reduces to invariance. This holds by the opposite scaling of θ and v: for z ∈ C×, (Lz) ∗ω = ( θz−1 ) (d(zv)) = θdv = ω. The left-inverse property of ω follows as ω(∂v) = θdv(∂v) = θ(v) = 1. ■ For future reference, we also remark that this induces the global curvature form K on Eig(V ) in the standard way, i.e., K := dω+ 1 2 [ω, ω] = dω. In coordinates it reads K = dθ∧dv, where dθ is defined similar to dv. As C× is commutative, the curvature K is invariant under the group action. Hence it admits push-forward along the quotient map πv to Spec(V ), resulting in the following. Lemma 4.8. The curvature form K on Eig(V ) reduces to a unique 2-form k ∈ Ω2(Spec(V ),C), which satisfies K = π∗v(k). 4.3 Geometry behind the geometric phase We will now show that the connection ω yields the geometric phases in adiabatic quantum mechanics in a natural way. That is, Eig(V ) provides a geometric model for the geometric phase. This is also applicable to non-cyclic states, which appear in the presence of EPs of non-Hermitian Hamiltonian families. An argument similar to the one that led us from Spec(V ) to Spec(H) in Section 3.2 holds here. Namely, the space Eig(V ) can be seen as a general abstract model, which can be adjusted for a specific Hamiltonian family using pull-back by H. This pull-back of Eig(V ) along the Hamiltonian family H yields the eigenstate bundle of H, given by Eig(H) = { (x,E, ψ) ∈ Spec(H)× V \ {0} | ψ ∈ EigE(H(x)) } . We observe that the pull-back construction of this space immediately guarantees various prop- erties of Eig(H). Clearly, Eig(H) is a smooth manifold. One can also write an element as (x,E, ψ, χ), where χ is the unique left-eigenstate of H(x) corresponding to E so that χ(ψ) = 1. Similarly, one may omit E from the tuple. The projection πHλv : Eig(H) → N(H) is a semi- principal C×-bundle, and the projection πHv : Eig(H) → Spec(H) is a principal C×-bundle. 18 E.J. Pap, D. Boer and H. Waalkens The latter bundle confirms that also Eig(H) can have non-trivial topology arising from the permutations around the EPs of H. The local sections of πHλv are in correspondence with local eigenstates. Here, a local eigenstate is a smooth function ψ : U → Eig(H), with U ⊂ N(H) open, such that ψ(x) is an eigenstate ofH(x). Clearly, this fixes a local eigenvalue E(x). This data is summarized in the corresponding local section ψ̂ as ψ̂ : x 7→ (x,E(x), ψ(x)). Observe that for each x ∈ U , knowing E(x) and ψ(x) fixes a unique left-eigenstate χ(x) of H(x), which depends smoothly on x. Hence we can include χ(x) as the fourth component of the local section ψ̂, following our earlier remark on notation of elements of Eig(H). The space Eig(H) also inherits a connection, which we write as ωH . Again, for any family H, the smoothness and other properties of Eig(H) are immediate from the pull-back construction. It is now easy to verify that parallel transport w.r.t. this connection is equivalent to studying the geometric phase. Namely, using the local section ψ̂, we can express ωH locally as ψ̂∗ωH = χdψ. This is indeed the expression reported by Garrison and Wright [12] as the generalization of the Berry connection. That is, the parallel transport on Eig(H) defined via the horizontal lifts w.r.t. ωH is equivalent to the calculation of geometric phase, for both Hermitian and non- Hermitian Hamiltonians. We may thus study geometric phases by studying the space Eig(H). For example, the connection ωH on Eig(H) can be flat, in which case the geometric phase is actually of a topological nature. By this we mean that the geometric phase due to a loop γ is invariant under continuous deformation of γ, which is equivalent to vanishing of the curva- ture KH of Eig(H). Hence, only non-contractible loops can yield a non-trivial geometric phase. Flatness is also of practical relevance, as the following result shows. Proposition 4.9. The connection ωH on Eig(H) is flat in case � the Hamiltonian family H is an analytic function in a single complex variable z. � H(x) is a symmetric matrix w.r.t. a fixed basis of V for all x ∈ N(H). Explicitly, if a local eigenstate ψ is such that ψTψ = 1, then ψ̂∗ωH = 0. Proof. If H = H(z) is analytic, we can find an analytic local eigenstate ψ(z). Consequently, ψ̂∗ωH = f(z)dz for some analytic function f . Using coordinates z and z̄ for the parameter space, ψ̂∗KH is proportional to ∂f(z) ∂z̄ . As this derivative vanishes, it follows that KH = 0. In the symmetric case, ψT is a non-zero multiple of the left-eigenstate χ corresponding to ψ. Hence the function ψTψ is non-vanishing, and one may always (locally) scale ψ so that ψTψ = 1, implying χ = ψT. By partial integration, ψ̂∗ωH = 1 2(χdψ − dχψ) = 1 2 ( ψTdψ − dψTψ ) = 0. ■ 4.3.1 Calculating the lift via an ansatz An explicit calculation of the lift of a path to Eig(H) can be done using an ansatz technique. For this technique, it does not matter if the path lies in N(H) or Spec(H). Indeed, if a path γ in N(H) and an initial eigenstate ψ0 of H(x0) is chosen, then ψ0 has an energy E0, and covering theory yields the corresponding path γ̄(t) = (γ(t), E(t)) in Spec(H). Conversely, any path γ̄ in Spec(H) yields a path in N(H), which obviously lifts to γ̄. If we restrict to loops the story changes, as we will see after discussing the ansatz technique. A Unified View on Geometric Phases and Exceptional Points 19 The ansatz technique here is simply the application of the usual one to the parallel transport equation. Let us continue with the path γ in N(V ) and the initial state ψ0. This data fixes a unique lift to Eig(H), which we denote by Γ and which is given by Γ(t) = (γ(t), E(t),Ψ(t)), where Ψ(t) is the adiabatically evolved eigenstate at time t, up to dynamical phase. Naturally, this lift is also given as (γ̄(t),Ψ(t)) in the perspective of lifting a path γ̄ from Spec(V ) to Eig(V ). In practice, one would not calculate Ψ(t) directly, but instead approach this problem using reference instantaneous eigenstates ψ(t). That is, for each t, one finds an eigenstate ψ(t) of H(t) with energy E(t), or ψ(t) ∈ EigE(t)(H(t)) for short. For example, ψ(t) can be obtained by explicitly calculating an eigenstate of H(γ(t)), and then letting t vary. We may assume that t 7→ ψ(t) is differentiable and satisfies the initial condition ψ(0) = ψ0. These instantaneous eigenstates define a path Γ0(t) = (γ(t), E(t), ψ(t), χ(t)) in Eig(H), where χ(t) is the path of accompanying left-eigenstates. Of course, Γ0 does not need to be the lift Γ, as equivalently ψ(t) does not need to be equal the actual state Ψ(t). However, Γ0 can function as an ansatz to calculate Γ explicitly. Indeed, both ψ(t) and Ψ(t) must lie in EigE(t)(H(t)), and so differ only by a scale factor ef(t). That is, one has Γ(t) = ef(t) · Γ0(t) = ( γ(t), E(t), ef(t)ψ(t), χ(t)e−f(t) ) with f(t) some differentiable complex-valued function satisfying f(0) = 0. This f depends solely on the ansatz ψ by imposing the lift condition 0 = (ωH)Γ(t) ( Γ̇(t) ) = χ(t)e−f(t) d dt ( ef(t)ψ(t) ) = ḟ(t) + χ(t)ψ̇(t). Solving for f(t), we thus find the actual state from the ansatz by application of the scale factor ef(t); Ψ(t) = exp ( − ∫ t 0 χ(t′)ψ̇(t′)dt′ ) ψ(t). (4.2) This is the general expression for a state ψ0 that undergoes parallel transport along a path γ or γ̄, also in the non-Hermitian case, expressed using an ansatz. Although the integral∫ t 0 χ(t ′)ψ̇(t′)dt′ is famous for its relation to the geometric phase, without additional assumptions it does not have physical significance. Indeed, the integral is not ansatz independent. That is, given another ansatz ψ′(t), then ψ′(t) = exp(a(t))ψ(t) for some complex-valued function a = a(t), and ∫ t 0 χ(t ′)ψ̇(t′)dt′ becomes ∫ t 0 χ ′(t′)ψ̇′(t′)dt′ + [a(t)− a(0)], which can in principle be any complex-valued continuous function. The actual state Ψ(t) is of course ansatz independent; assuming the same initial condition, i.e., ψ′(0) = ψ0, then a(0) = 0 and as desired we obtain Ψ′(t) = exp ( − ∫ t 0 χ′(t′)ψ̇′(t′)dt′ ) ψ′(t) = exp ( − ∫ t 0 χ(t′)ψ̇(t′)dt′ − a(t) ) ea(t)ψ(t) = Ψ(t). This confirms that the integral is, in general, only a correction factor to the specific ansatz ψ(t); it merely quantifies how much our ansatz ψ(t) was away from being the lift Ψ(t). In fact, just from physical arguments one should not expect the integral to yield an observable; we did not assume the state to be cyclic, so that there is no particular phase the integral can be equal to. Of course, if the state is cyclic, say the state returns at time T , then a geometric phase is well-defined. With an additional assumption on the ansatz ψ, namely ψ(T ) = ψ0, this geometric phase is indeed given by the integral at t = T . However, even in this cyclic 20 E.J. Pap, D. Boer and H. Waalkens case, if we consider intermediate times t, i.e., 0 < t < T , the integral does not yield a physical quantity; we fixed a(T ) = 0, but for intermediate times a(t) can still be anything. In other words, concerning a state acquiring a geometric phase, there is no well-defined rate at which this happens. This reflects that a geometric phase depends only on the locus of a (closed) path, not its parametrization. Still, the integral ∫ t 0 χ ′(t′)ψ̇′(t′)dt′ can be related to path “lengths”, as we consider in Section 4.6.3. We will consider the state evolution in more detail in the following. We will distinguish between the cyclic and the non-cyclic case. The approach is summarized in diagram (4.3) below, which is the pullback of diagram (4.1) along H. Namely, we see that the projection πHλv : Eig(H) → N(H) can be considered as the main bundle to model the evolution of states. Indeed, given any loop in N(H), it will certainly induce a holonomy operation on Eig(H), regardless of states being cyclic or non-cyclic. In contrast, one can also consider the bundle πHv : Eig(H) → Spec(H). Clearly, loops in Spec(H) correspond to cyclic evolution only.2 How- ever, this bundle is principal and so has the advantage that its holonomies are easier to describe. We thus remark that for cyclic states one uses πHv , and one uses πHλv primarily for non-cyclic states. We also emphasize that the calculation of Ψ(t) above can be used in both cases: Eig(H) Spec(H) (x,E, ψ) (x,E) N(H), x. cyclic states non-cyclic states energy swaps πH v πH λv πH λ (4.3) 4.3.2 The cyclic case Let us start with the cyclic case. As said, this concerns the bundle πHv : Eig(H) → Spec(H). We remark that the only difference between a path in N(H) and a path in Spec(H) is that the latter not only specifies the change in system parameters, but also which energy level is of interest. Moreover, as we have seen, fixing an initial energy E0 uniquely specifies a path in Spec(H) given a path in N(H). It is thus natural to consider a loop γ̄ in Spec(H) as the input data for cyclic evolution. Let us now consider the evolution of the eigenstates. Denote the basepoint of γ̄ again by (x0, E0), and assume γ̄ returns to this point at time t = T . The state evolution is then described by the parallel transport map Pγ̄ , e.g., Pγ̄(ψ0) = Ψ(T ) following the previous calculation. As πHv defines a principal bundle and ωH is a principal connection, Pγ̄ follows from standard holonomy theory. That is, Pγ̄ is an automorphism of the C×-torsor EigE0 (H(x0)), and thus amounts to scaling by a unique element in C×. This element is clearly the geometric phase factor that any state in EigE0 (H(x0)) acquires due to following γ̄. As we allow for non-Hermitian Hamiltonians, this phase factor need not be unitary. We hence obtain a definition of (generalized) geometric phase from holonomy as follows. Definition 4.10 (geometric phase). Let γ̄ be a loop in Spec(H) based at (x0, E0). The geometric phase γgeo due to γ̄ is defined, modulo 2π, via Pγ̄ = eiγgeo · idEigE0 (H(x0)). It now remains to calculate γgeo explicitly. For this, we can use the earlier result from the ansatz technique. As usual, one only has to compare the phase difference between the final state Ψ(T ) and the initial state ψ0, i.e., Ψ(T ) = Pγ̄(ψ0) = exp(iγgeo)ψ0. We thus con- sider the unique non-zero complex scalar [Ψ(T )/ψ0] such that Ψ(T ) = [Ψ(T )/ψ0]ψ0. Equating 2In an adiabatic setting, with cyclic we also assume that the system parameters are restored. A Unified View on Geometric Phases and Exceptional Points 21 exp(iγgeo) = [Ψ(T )/Ψ(0)] and substituting our expressing for Ψ(t) in terms of our ansatz ψ(t), we find the expression γgeo = i ∫ T 0 χ(t)ψ̇(t)dt− i ln([ψ(T )/ψ0]), (4.4) where the logarithm term yields the usual modulo 2π of a phase. By construction this is invariant under both replacing ψ0 with another state in the same ray and picking another ansatz ψ. We see two terms of different nature in equation (4.4). The integral term is clearly the usual integral for the geometric phase, and corrects for changes of ψ along its own direction. The logarithm term is a correction for ψ not closing on itself. Observe that only the two terms together are invariant under changing the choice of the ansatz ψ. The logarithm term vanishes whenever ψ(1) = ψ(0), which happens, e.g., if ψ is built using a local eigenstate. On the other hand, the integral term vanishes, e.g., when ψ is chosen to be the lift Ψ, in which case the integrand is identically zero. In this case γgeo can be computed using the end points only, see Example 4.11 below for an explicit example. Example 4.11. Let us consider a typical system with a diabolic point (DP), named after the diabolo shape of the energy bands [6]. Let us pick V = C2 with standard basis (e1, e2), in which the Hamiltonian family reads H(a, b) = ( a b b −a ) , where a, b are real numbers, i.e., M = R2. The energy bands are given as λ±(a, b) = ± √ a2 + b2, which have a single degeneracy at the origin. This is located at the apex of the diabolo, and is the DP of this system. Let us show how the geometric phase integral can be seen as a correction factor. Therefore, let us traverse the unit circle using the path γ(t) = (cos(t), sin(t)) with time interval [0, 2π], and consider the λ+ level (λ− is similar). By looking at H(γ(t))− λ+(γ(t))I, one readily finds an nsatz ψ(t) for the evolution of the eigenstate, together with a left-eigenstate path χ(t), as ψ(t) = 1 2 ( 1 + cos(t) sin(t) ) , χ(t) = 1 1 + cos(t) ( 1 + cos(t) sin(t) )T . The factor 1/2 is introduced to have ψ(0) = e1. Note that ψ and χ are defined only for t ∈ [0, π); being a (left-)eigenstate, they are not allowed to vanish. Still, for these times we can calculate the lift of γ(t) starting at e1. We can correct the ansatz ψ(t) following equation (4.2). We thus evaluate the geometric phase integral, whose integrand is χ(t)ψ̇(t) = − sin(t) 2(1 + cos(t)) = 1 2 d dt ln(1 + cos(t)). Hence we find the lift of γ starting at e1 to be Ψ(t) = √ 1 2(1 + cos(t)) ( 1 + cos(t) sin(t) ) = ( cos(t/2) sin(t/2) ) . As the scale factor is real, we can interpret it as a length correction. This length interpretation clearly shows that our original ψ(t) was varying along itself. 22 E.J. Pap, D. Boer and H. Waalkens Note that our final expression of Ψ(t) can be extended to arbitrary times, hence yields the full lift of γ. Consequently, it is now convenient to obtain the geometric phase via Ψ(t). For lifts, only the logarithm term in equation (4.4) contributes, and we find the phase (modulo 2π) γgeo = −i ln([Ψ(2π)/ψ0]) = −i ln([−e1/e1]) = −i ln(−1) = π. In the perspective of loops in Spec(H), the above concerns the loop γ̄+(t) := (γ(t), λ+(γ(t))) = (γ(t), 1), revealing that Pγ̄+ = −id. A similar argument shows that the loop γ̄−(t) = (γ(t),−1), i.e., considering the other energy band, yields Pγ̄− = −id as well. In some situations, one can express the geometric phase alternatively as a curvature integral. A standard argument is as follows. Let U ⊂ N(H) be a neighborhood on which a local eigenstate ψ = ψ(x) is defined, which has as partner the local left-eigenstate χ. If γ is a loop in U that forms the boundary of a surface S also contained in U , then the geometric phase is given by γgeo = i ∫ γ χdψ = i ∫ ∂S ψ̂∗ωH = i ∫ S ψ̂∗(dωH) = i ∫ ψ̂(S) KH . We see that the assumption of γ being a boundary is essential. If γ is a loop but not a boundary, then the geometric phase is not given by a curvature integral. This motivates us to consider homology theory. In order to do this, it is convenient to work on Spec(H) instead. Indeed, the integral above can be rewritten as an integral over kH , which is the reduced curvature on Spec(H). Clearly, the local eigenstate ψ(x) defines an energy function E(x). The map Ê : U → Spec(H), x 7→ (x,E(x)) is then a local section of Spec(H). Moreover, as Ê = πHv ◦ ψ̂, where πHv is the projection Eig(H) → Spec(H), we have γgeo = i ∫ S ψ̂∗(KH) = i ∫ S ψ̂∗((πHv )∗ (kH) ) = i ∫ S Ê∗(kH) = i ∫ Ê(S) kH . Let us interpret this result. First, we remark that the integral over kH is manifestly inde- pendent of the chosen local eigenstate ψ. More precisely, only the energy bands matter, not the exact eigenstates. In addition, we now consider surfaces in Spec(H), hence the above argument can be used whenever our path γ̄ is a boundary. We thus find that the homology of Spec(H), i.e., the homology of the energy bands, plays a key role in the relation of geometric phase to cur- vature. Moreover, in case Eig(H) is flat, the homology theory allows one to see the topological nature of the geometric phase in an explicit way, as shown in the following. Proposition 4.12. Let γ̄ be a loop in Spec(H). If γ̄ is the boundary of a surface Σ in Spec(H), then the geometric phase acquired by traversing γ̄ equals i ∫ Σ kH . Corollary 4.13. If Eig(H) is flat, then the geometric phase due to a loop in Spec(H) only depends on the class of the loop in the first homology group H1(Spec(H)). Example 4.14. Let us continue Example 4.11. As H is symmetric, by Proposition 4.9 the con- nection is flat. Hence the geometric phases only depend on the homology of Spec(H), and thus are of topological nature. Clearly, Spec(H) is the diabolo minus the DP, which is homeomorphic to two punctured planes. Hence H1(Spec(H)) ∼= Z2, generated by (the classes of) the images of the unit circle under λ+ and λ−. It thus suffices to know the phase due to each generator, which is π following our earlier calculation in Example 4.11. It is now straightforward to recognize the bundles introduced by Simon [32] in the bundle Eig(H) → Spec(H). According to Corollary 3.8, for Hermitian H the bundle Spec(H) → N(H) is trivial, i.e., Spec(H) separates into n distinct energy bands. Let us label these energy bands A Unified View on Geometric Phases and Exceptional Points 23 by 1, . . . , n and write Speck(H) for energy band k, and similarly Eigk(H) for the subbundle over Speck(H). Then Eigk(H) → Speck(H) is a principal C×-bundle, which is the bundle used by Simon for energy level k adapted to our language. It is thus geometrically clear why this approach does not apply to non-Hermitian Hamiltonians; in that case the energy bands in Spec(H) need not be separated but connected via “spiral staircase” like structures. If this is the case, i.e., if non-cyclic states appear, then the bundle Eig(H) → Spec(H) can still model these states by parallel transport, but as the path γ̄ is then not a loop this does not fit in the framework of holonomy. 4.3.3 The non-cyclic case Let us now demonstrate how non-cyclic states do have a holonomy description when using the other projection, i.e., the bundle πHλv : Eig(H) → N(H). Here, holonomy is to be understood in the context of a semi-principal bundle. Similarly to the case above, parallel transport along a loop γ in N(H) based at x0 induces the parallel transport map Pγ , which is an automorphism of the C×-semi-torsor Eig(H(x0)). The set of all Pγ for a fixed base point x0 then defines the holonomy group Hol Eig(H) N(H) (x0) = { Pγ ∈ Aut(Eig(H(x0))) | γ ∈ Loop(N(H), x0) } . However, these automorphisms are harder to describe. Whereas we could previously identify a map Pγ̄ with an element in C×, this need not be for Pγ . In particular, by definition a non-cyclic state does not return to the same group orbit, i.e., the initial eigenray, hence no element in C× can relate the initial and final states. Instead, one must find how the eigenrays are transported individually, as illustrated in the following example. Example 4.15. Let us consider a standard EP2 example, using the Hamiltonian family of Example 3.5. We take x0 = 0 again as our reference point, and from there we encircle the EP at x = +i in positive direction by following a path γ. One can take γ to be a circular path, but as the connection is flat by Proposition 4.9, the exact shape of γ does not play a role. The evolution of the states due to encircling γ is captured by the map Pγ : Eig(H(0)) → Eig(H(0)). Clearly, Eig(H(0)) is the disjoint union of the two eigenrays, i.e., Eig(H(0)) = C×e1 ⊔ C×e2, where e1 = (1, 0)T and e2 = (0, 1)T are the standard basis vectors. Hence, an element of Eig(H(0)) is of the form ze1 or ze2 with z ∈ C×. We emphasize that Eig(H(0)) is a semi- torsor and not a vector space; linear combinations of e1 and e2 are not present, as these are not eigenstates of H(0). Accordingly, Pγ is equivariant and not linear. We can now calculate Pγ as follows. First, it is sufficient to know Pγ(e1) and Pγ(e2) due to equivariance; Pγ(zek) = zPγ(ek) for k = 1, 2. That is, it is sufficient to follow e1 and e2 around the EP. This can be done using an explicit parametrization of Spec(H). One can easily find eigenstates depending on x, and following Proposition 4.9 we “normalize” them so that no geometric phase will appear.3 Hence we arrive at the expressions ψ±(x) = ∓ 1√ 2(1 + x2 ∓ √ 1 + x2) ( −x 1∓ √ 1 + x2 ) , where we remark that ψ+ has a removable singularity at x = 0; the limit reads limx→0 ψ+(x) = e1. Hence at x0 = 0, ψ+(0) = e1 and ψ−(0) = e2 as desired. The lifts of these states to 3Note that this uses flatness of the connection. In the non-flat case the geometric phase depends on the exact shape of γ and so cannot be expressed using local eigenstates. One should then use, e.g., the ansatz method instead. 24 E.J. Pap, D. Boer and H. Waalkens Eig(H) can thus be found by inspection of ψ±(x). Encircling the EP in positive direction swaps the ± signs, and in addition the overall root in the denominator obtains a factor of i. Thus, after following ψ±(x) around the EP, we return with i−1ψ∓(x). At our reference x0 = 0, this rule becomes e1 7→ −ie2, e2 7→ −ie1. This information is enough to specify Pγ , which we thus find to be given by Pγ : ze1 7→ −ize2, ze2 7→ −ize1. (4.5) In order to study Pγ , let us consider the following statement showing how a general auto- morphism of a semi-torsor can be studied by its invariant subspaces, similar to linear algebra. Proposition 4.16. Let G be a Lie group, F a G-semi-torsor and ϕ : F → F an automorphism. Then F = ⊔i∈I′Fi, for some index set I ′, decomposes F into minimal ϕ-invariant subspaces, i.e., ϕ(Fi) = Fi for all i. Moreover, if G is Abelian and a particular Fi consists of k <∞ orbits, then (ϕ|Fi) k equals translation by an element of G. Proof. Clearly, the index set I ′ is the original orbit space modulo the relation that orbits mapped into one another by ϕ are identified. If Fi consists of k orbits, then by minimality (ϕ|Fi) k preserves the orbits in Fi. Picking f ∈ Fi, we find ϕk(f) = g(f)f for some g(f) ∈ G. If G is Abelian, then g(f) is constant on the orbit through f . In addition, as equivariance yields ϕk(ϕ(f)) = g(f)ϕ(f), we see ϕ(f) is scaled by the same element. It follows that g(f) is constant on the subspace Fi. ■ For the map Pγ , it is clear that a minimal subspace is any minimal union of eigenrays whose energies are permuted upon traversing γ. The element of C× associated to such a minimal union is a phase factor. Indeed, if there are k rays in the union, then a ray first returns to itself by following γ exactly k times, after which is has obtained precisely this phase factor. Again, if k > 1 the state is non-cyclic and there is no definite phase between an initial state ψ0 in the union and its transport Pγ(ψ0). We come back to this in Section 5, where we treat how Pγ can be expressed via a holonomy matrix. Of course, if k = 1, i.e., ψ0 is cyclic, then we do recover the geometric phase. We summarize these findings in the following. Corollary 4.17. Given a loop γ, the invariant subspaces of Pγ are the minimal unions of eigenrays. If a union consists of k eigenrays, then the characteristic phase of the union equals the geometric phase obtained by traversing the loop γ̄k, where γ̄ is the lift of γ to Spec(H) with an energy corresponding to any of the eigenrays in the union. It is also clear that Pγ contains the information of the underlying permutation pγ of the energies. This is formally expressed by πHv being a bundle map that is equivariant w.r.t. the collapsing quotient homomorphism C× → 1. As πHv maps Eig(H) to Spec(H), this quotient reduces Pγ to pγ . Proposition 4.18. For any γ ∈ Loop(N(H), x0), the map Pγ induces the following commutative diagram: Eig(H(x0)) Eig(H(x0)) Spec(H(x0)) Spec(H(x0)). Pγ πH v πH v pγ A Unified View on Geometric Phases and Exceptional Points 25 4.4 Including the dynamical phase The geometric properties of adiabatic dynamics for non-degenerate operators, Hermitian or non- Hermitian, can be described on Eig(V ) using the connection ω. This leaves out an important non-geometrical property, namely the dynamical phase. Nevertheless, Eig(V ) does support a calculation of the dynamical phase. This builds on a particular complex-valued function, which simply extracts the eigenvalue from the elements in Eig(V ), i.e., E : Eig(V ) → C, (A, λ, v) 7→ λ. Given a lift Γ in Eig(V ), the corresponding dynamical phase is then the integral γdyn = − i ℏ ∫ E(Γ(t))dt. It is even possible to put the dynamical phase explicitly in the lift. Clearly, the lift Γ without dynamical phase vanishes under the covariant derivative D = d + ω. If one modifies this to Dtot = d + ω + i ℏ E and calculates the lift Γtot of γ given by DtotΓtot = 0, then Γtot is the path of the eigenstate, including the dynamical phase, assuming γ is parametrized by physical time. 4.5 Relation with the work of Aharonov and Anandan We found that Eig(V ) provides a framework suitable for any non-degenerate finite-dimensional Hamiltonian, including non-Hermitian Hamiltonians in particular. However, this brings us to the question how the theory of Hermitian systems relates to Eig(V ). The geometric framework for such Hermitian cases was pioneered by Aharonov and Anandan in [1] (see also the elaboration in [2]). The geometric spaces found there differ significantly from the spaces obtained here. Nevertheless, we will show that they can be obtained from Eig(V ). Let us summarize the theory of [1], rephrasing it in line with our approach to Eig(V ). First, the state space V is now assumed to be a Hilbert space, i.e., V should be equipped with a Hermi- tian inner product ⟨ | ⟩. This allows one to define the unit sphere S1(V ) inside V by restricting to norm 1 states. As norm 1 fixes a state up to a U(1)-phase, S1(V ) is naturally a U(1)-manifold, and the quotient is the projective space P (V ) of V . This quotient defines the principal bundle U(1) S1(V ) P (V ), (4.6) which is the central object in the formalism. The base P (V ) can be viewed as the space of rays in V , but also as the set of projectors projecting to a line. That is, the ray through |ψ⟩ ∈ S1(V ) can be identified with the projection operator |ψ⟩ ⟨ψ|. The metric on V obtained from the inner product ⟨ | ⟩ induces a metric tensor on S1(V ), and hence a connection 1-form. On a path |ψ(t)⟩ in S1(V ), this connection yields ⟨ψ(t)|ψ̇(t)⟩. If one lifts a loop from P (V ) to S1(V ), it follows that the final state must lie in the same ray as the initial state. In this case, one says that the state is cyclic. If the path |ψ(t)⟩ satisfies |ψ(1)⟩ = |ψ(0)⟩, then the obtained Aharonov–Anan- dan (AA) phase is given by the integral i ∫ 1 0 ⟨ψ(t)|ψ̇(t)⟩dt. This coincides with the adiabatic Berry phase if |ψ(t)⟩ evolves adiabatically. 26 E.J. Pap, D. Boer and H. Waalkens The AA phase is thus a generalization of the Berry phase from adiabatic state evolution to any path of states. Hence the AA phase is viewed as a non-adiabatic generalization. Still, we argue that it can be obtained from Eig(V ), equipped with the “adiabatic” connection ω. That is, we will show that the bundle in equation (4.6), including connection, can be obtained by collapsing the principal bundle πv : Eig(V ) → Spec(V ). This can be done in two steps. First, we use the inner product to restrict the operators to the Hermitian ones and the vectors to normalized ones. After this, listing the operator will be redundant, and the second step is to discard it. Let us describe the first step. The only additional ingredient we use is the chosen inner product ⟨ | ⟩ on V . It allows us to talk about Hermitian operators, and we restrict N(V ) accordingly to the closed subset N⟨ | ⟩(V ) := { A ∈ N(V ) | A† = A with respect to ⟨ | ⟩ } of all non-degenerate Hermitian operators on V . This is a non-canonical subset of N(V ); a dif- ferent inner product may yield a different subset. Consequently, there are the subbundles over N⟨ | ⟩(V ) given by Spec⟨ | ⟩(V ) := { (A, λ) ∈ Spec(V ) | A† = A with respect to ⟨ | ⟩ } , Eig⟨ | ⟩(V ) := { (A, λ, v) ∈ Eig(V ) | A† = A, ∥v∥ = 1 with respect to ⟨ | ⟩ } . The C×-action on Eig(V ) reduces to U(1)-phase rotation on Eig⟨ | ⟩(V ), which has quotient space Spec⟨ | ⟩(V ). The key observation now is that, after this restriction, we no longer need to know A and λ to compute the eigencovector and eigenprojectors. Indeed, given the normalized vector |v⟩, the covector θ is ⟨v| and the eigenprojector is |v⟩ ⟨v|, regardless of the exact A and λ. Hence, we are motivated to discard A and λ. To do so, it is convenient to describe Spec(V ) using pro- jectors. Clearly, any pair (A, λ) ∈ Spec(V ) defines an eigenprojector P(A,λ) projecting on the eigenspace of A corresponding to λ. Conversely, given an eigenprojector P , the eigenvalue can be retrieved from the identity AP = PA = λP . Hence we may write (A, λ) ∈ Spec(V ) equiva- lently as (A,P(A,λ)). For Spec⟨ | ⟩(V ), we may even write an element as (A, |v⟩ ⟨v|), where |v⟩ is a normalized eigenvector of A. We can now perform the second step, i.e., the reduction. Reducing Spec⟨ | ⟩(V ) to the space P (V ) is straightforward using the projector description where the element (A, |v⟩ ⟨v|) goes to |v⟩ ⟨v|. Reducing Eig⟨ | ⟩(V ) to S1(V ) is similar; we only keep the vector. Together, these maps define a morphism of bundles as follows: Eig⟨ | ⟩(V ) S1(V ) (A, λ, |v⟩) |v⟩ Spec⟨ | ⟩(V ) P (V ), (A, |v⟩ ⟨v|) |v⟩ ⟨v| . πv Our final claim is that the connection ω on Eig(V ) also carries over to S1(V ). This again follows the same two steps. First, clearly ω restricts to a U(1)-connection on Eig⟨ | ⟩(V ). Second, this restricted form admits push-forward to S1(V ). Concerning explicit formulas, this push- forward is given by substituting θ = ⟨v|, which yields exactly the connection used to define the AA phase. The following then summarizes these findings. Proposition 4.19. The restricted projection Eig⟨ | ⟩(V ) → Spec⟨ | ⟩(V ) has a canonical reduction to the bundle S1(V ) → P (V ). Moreover, the restricted connection on Eig⟨ | ⟩(V ) admits push- forward to S1(V ), yielding the standard U(1)-connection as used for the AA phase. A Unified View on Geometric Phases and Exceptional Points 27 We conclude that the above projection allows us to translate the general theory of Eig(V ) to more specific results, enabled by a Hermitian inner product. We will use the phrase “in the Hermitian case” to indicate such a passage has happened. We already saw that notationally this amounts to replacing θ by ⟨v|, but the broad picture contains more. For example, Eig(V ) is locally partitioned according to eigenvalues, while this is completely absent in S1(V ). 4.6 Quantum geometric tensor We will show that the covariant derivatives naturally define a tensor on Eig(V ), which is a straightforward generalization of the quantum geometric tensor. We also consider its reduced version on Spec(V ). Before we discuss the tensor itself, we first introduce convenient bases of tangent spaces of Eig(V ) and Spec(V ). Afterwards we comment on a relation between geometric phase and distance. 4.6.1 Bases for tangent space of spectrum and eigenvector bundles Let us formulate (complex) bases for the tangent spaces T(A,λ,v) Eig(V ) and T(A,λ) Spec(V ). The idea is that we pass the v, λ and A components one-by-one. In this way, both tangent spaces can be described in a similar way. Let us start with the eigenvector part. Clearly, the main difference between the two tan- gent spaces is that T(A,λ,v) Eig(V ) has a tangent along the eigenray Eigλ(A). This direction is naturally spanned by the fundamental tangent vector ∂v originating from the scaling action. An advantage of picking ∂v is that we may pick the remaining tangent vectors in T(A,λ,v) Eig(V ) to be horizontal. These tangent are then the horizontal lifts of unique tangents in T(A,λ) Spec(V ), which means we cover both tangent spaces simultaneously. We continue with the eigenvalue part. Obviously, a change of eigenvalue must be accompanied by a change of operator. Hence, let us change the operator only by what is absolutely necessary. Writing P = vθ for the eigenprojector of A corresponding to λ, a shift in λ is then given by the tangent ∂λ = d dt ∣∣∣∣ t=0 (A+ tP, λ+ t, v, θ). We must then find n − 1 other tangent vectors for the other eigenvalues of A. These can be obtained via paths of the form (A + tP ′, λ, v, θ), with P ′ the corresponding eigenprojector. Note that λ is constant as we vary another eigenvalue, but still consider tangents at (A, λ, v). The combination of these n tangent vectors defines the tuple ∂λ̃, where we pick an ordering of Spec(A). Writing the coefficients of these vectors as ∆λ̃, again cf the ordering, we obtain the linear combination ∆λ̃∂λ̃. The corresponding tangents in T(A,λ) Spec(V ) are similar. It thus remains to describe all changes in operator and eigenvector, where the eigenvector should not change along itself. As we should now avoid to change the spectrum of A, let us use similarity transformations. That is, we conjugate by the operator eitX , where iX ∈ End(V ) can be viewed as a generator; the extra i is for later convenience. We then obtain the infinitesimal conjugation action (extending equation (2.5)), which we view as the linear map End(V ) → T(A,λ,v) Eig(V ) given by iX 7→ d dt ∣∣∣∣ t=0 ( eitXAe−itX , λ, eitXv, θe−itX ) . As is, this parametrization of T(A,λ,v) Eig(V ) is not compatible with our earlier choices. For instance, if iX = P we retrieve ∂v, and for other eigenprojectors the obtained tangent vanishes. Hence we require X to be free of eigenprojectors of A, which means that the matrix of X w.r.t. 28 E.J. Pap, D. Boer and H. Waalkens any eigenframe of A has zero diagonal. In this case, we say X is A-free. Note that the obtained tangent is horizontal if and only if θXv = 0, which is automatically satisfied for A-free X. In summary, we may express a general element v ∈ T(A,λ,v,θ) Eig(V ) as v = iX +∆λ̃∂λ̃ + z∂v, where for T(A,λ) Spec(V ) the first two terms suffice. This is a basis if we impose X to be A-free, in which case the terms span a subspace of dimension n2 − n, n and 1, respectively. The values of dv and dθ then read dv(v) = d dt ∣∣∣∣ t=0 eitXetzv = (iX + z)v, dθ(v) = −θ(iX + z). We finally remark that it is possible to not impose X to be A-free; the map to T(A,λ,v) Eig(V ) is then still surjective, but no longer injective. This can be convenient in practice; X can play the role of a Schrödinger Hamiltonian, which need not be A-free. We will keep this in mind when describing the tensors in the following. 4.6.2 Generalized quantum geometric tensor On the space Eig(V ) there is a canonical tensor which in the Hermitian case reduces to the quantum geometric tensor (QGT). This QGT was first reported in [30], where it was found by looking at infinitesimal distance between states. Its anti-symmetric part was later recognized to essentially be the Berry curvature, demonstrating its relevance to the geometric phase, while its symmetric part yields a metric on parameter space [4]. We now show that Eig(V ) supports a more general tensor, which relates directly to covariant derivatives. Moreover, this generalized tensor makes no reference to an inner product. It is thus also incorporates a generalization of the QGT based on PT -symmetry as reported in [40]. Let us start from the standard expression. Fixing local coordinates on N(H) and a local normalized eigenstate |ψ⟩ = |ψ(x)⟩, the QGT is given by Tij = ⟨∂iψ| (1− |ψ⟩ ⟨ψ|) |∂jψ⟩. We regard this to be the pull-back of a more abstract tensor G on Eig(V ). Clearly, this G is simply given by G(A,v,θ)(v1, v2) = dθ(v1)(1− vθ)dv(v2), v1, v2 ∈ T(A,v,θ) Eig(V ). (4.7) Because of this straightforward generalization from the Hermitian case to Eig(V ), we refer to G as the (generalized) quantum geometric tensor. In addition, G is a natural tensor on Eig(V ) in the following sense. Observe that the projector 1 − vθ is naturally obtained by taking the covariant derivative, as ∇v(v) = dv(v)− ω(v)v = dv(v)− θ(dv(v))v = (1− vθ)dv(v), v ∈ T(A,v,θ) Eig(V ), and similarly ∇θ = dθ(1− vθ). We thus observe that G is the natural combination G(A,v,θ)(v1, v2) = ∇θ(v1)(∇v(v2)), v1, v2 ∈ T(A,v,θ) Eig(V ), which directly displays its scale invariance. We can write G more explicitly using our description of the tangent space T(A,λ,v,θ) Eig(V ). Introducing labels 1 and 2 according to the two tangent vectors, this yields G(A,v,θ)(v1, v2) = θ(X1X2)v − (θX1v)(θX2v). A Unified View on Geometric Phases and Exceptional Points 29 One can also write this using scale invariant quantities only. Writing points of Eig(V ) as (A,P, v, θ), we find G(A,P,v,θ)(v1, v2) = tr(PX1X2)− tr(PX1) tr(PX2). This also provides an explicit form of the reduced QGT defined on Spec(V ). We observe that in these expressions we need not impose the Xi to be A-free. This is similar to correcting for a non-zero mean in probability theory; the second term provides a correction that vanishes for A-free Xi. Hence we choose not to impose the A-free condition, and instead keep the second term. In the Hermitian case, the QGT is the sum of a symmetric tensor, known as the quantum metric, plus an anti-symmetric part proportional to the Berry curvature [4]. We find a similar decomposition here. The anti-symmetric part of G is readily seen to be K/2, as Galt (A,v,θ)(v1, v2) = θ [X1, X2] 2 v = 1 2 dθ ∧ dv(v1, v2) = 1 2 K(A,v,θ)(v1, v2). We may also write this as tr ( P [X1,X2] 2 ) , which in addition yields an explicit form of the reduced curvature k. The symmetric part M := Gsym of G is then a tensor on Eig(V ) generalizing the standard quantum metric tensor (QMT). Hence we will refer to M using the same name. A scale invariant explicit form is M(A,P,v,θ)(v1, v2) = covP (X1, X2) := tr ( P {X1, X2} 2 ) − tr(PX1) tr(PX2), where covP is the covariance of non-commutative operators w.r.t. the density P . We hence obtain a decomposition similar to [4, equation (30)], but now for the generalized tensors on Eig(V ). Proposition 4.20. The QGT on Eig(V ) decomposes as a linear combination of the QMT and the curvature as G = M+ 1 2 K, which is the decomposition of G into its symmetric and anti-symmetric part, respectively. Notable properties that do not generalize from the Hermitian case are the following. Clearly M and K are complex rather than real-valued tensors, and no longer the real resp. imaginary part of G. In addition, M is a degenerate form, hence does not follow a standard metric interpretation. We also wish to comment on reducing the QMT to a “metric” on pa- rameter space. In the Hermitian case this can be done for each energy band separately. That is, for a fixed energy band, one can define a global eigenstate and so obtain a metric on N(H) by pull-back of M. In the non-Hermitian case, or better, whenever Spec(H) is non-trivial, this need not be possible as a global eigenstate could be unavailable. 4.6.3 Relation between geometric phase and distance Let us comment on a relation between geometric phase and distance. This was pointed out in [29] for the Hermitian case, but was also considered for the non-Hermitian case in [11]. The idea is to compare two distance functions defined on normalized states. Using a given inner product on V and the corresponding bundle S1(V ) → P (V ) from Section 4.5, these distances are given by L(ψ,ψ′) = ∥∥ψ − ψ′∥∥, D(ψ,ψ′) = √ 2− 2|⟨ψ|ψ′⟩|, ψ, ψ′ ∈ S1(V ). 30 E.J. Pap, D. Boer and H. Waalkens Here, L is the inner product metric on V restricted to S1(V ). The ray space P (V ) also inherits a metric, whose pull-back to S1(V ) is D. In this way, L measures total length, whereas D only measures the underlying change of rays. Their difference is then related to movement along the ray, which is where we find geometric phase. To make this concrete, one works infinitesimally. Setting ψ′ = ψ+ ψ̇dt+ 1 2 ψ̈dt 2 + · · · and keeping terms up to dt2 only, one finds dL2 = ⟨ψ̇|ψ̇⟩dt2 and dD2 = [ ⟨ψ̇|ψ̇⟩+ ⟨ψ|ψ̇⟩2 ] dt2. One thus finds dL2−dD2 = (i⟨ψ|ψ̇⟩dt)2, so that the geometric phase integrand equals √ dL2 − dD2, which is one of the main results in [29]. It is tempting to write “dγgeo” for i⟨ψ|ψ̇⟩dt, but we refrain from doing that. Following the discussion after equation (4.2), there is no definite rate at which a geometric phase is acquired, although the contrary is suggested by such notation. There is also no need to introduce more notation; the tensor field corresponding to this quantity is simply iω. We now generalize to the non-Hermitian setting of Eig(V ). First, the distances L and D do not carry over. Concerning L, we see L2 = 2 − ⟨ψ|ψ′⟩ − ⟨ψ′|ψ⟩ generalizes to the function 2 − θ1(v2)− θ2(v1) on Eig(V ), but its square root would be multi-valued on Eig(V ), meaning that L itself does not generalize. Similar issues appear for D2 = 2 − 2 √ ⟨ψ|ψ′⟩⟨ψ′|ψ⟩. Nevertheless, the tensor fields dL2 and dD2 do have a generalization to Eig(V ). One can check that dD2 is actually M for the Hermitian case, see also [10, 11]. Finally, dL2 generalizes to the tensor field L := (dθdv)sym, which is the non-covariant counterpart of M. The relation between L, M and ω is readily found from equation (4.7); first rewriting this to G = dθdv + (−dθv)⊗ (θdv) = dθdv + ω ⊗ ω and then taking the symmetric part yields M = L+ ω ⊗ ω. This is thus an equality of tensor fields on Eig(V ), generalizing the relation “dD2 = dL2−dγ2geo”. 5 Explicit description of the holonomy In Sections 3 and 4 we saw how the bundles Spec(V ) and Eig(V ) provide a holonomy interpreta- tion for the physics of geometric phases and exceptional points. However, in the non-cyclic case, we did not provide an easier description of the maps pγ and Pγ in terms of explicit permutations like (12) resp. holonomy matrices. We will now treat how these can be obtained, and afterwards show that this naturally follows holonomy theory as well. 5.1 Explicit permutations and holonomy matrices Let us start with the permutations of energies. Our goal is to find a formalism so that a permu- tation pγ of the eigenvalues of some A ∈ N(V ) can be expressed using an standard permutation σ ∈ Sn. A first attempt would be to label the eigenvalues, i.e., call them λ1, . . . , λn, and define σ by the relation pγ(λi) = λσ(i). However, this approach has a serious disadvantage. Although this works fine if we consider the eigenvalues of only the operator A, as the authors did in [27], this method does not extend to all of N(V ). The reason is that knowing the action of Sn on a spectrum is equivalent to having a labelling of the eigenvalues [28]. For example, if there are 3 eigenvalues, then from the action of (12) one finds which eigenvalue is labeled 3. Hence, if the above extends to a global Sn-action, then Spec(H) admits a global labelling and is thus trivial. We thus wish to obtain a method that also applies to non-trivial Spec(H). Instead of con- sidering eigenvalues separately, let us consider specific tuples of eigenvalues. Namely, for an operator A, we consider the n-tuples that list all eigenvalues exactly once, which brings us to the set Spec(A)! = { λ̃ = (λ1, . . . , λn) ∈ Spec(A)n | all eigenvalues of A appear once in λ̃ } . A Unified View on Geometric Phases and Exceptional Points 31 Observe that this set has n! elements, hence the factorial notation. In addition, this space is naturally endowed with the Sn-action on n-tuples, cf. our earlier notation in equation (2.2). It is this Sn-action that will help us with describing the map pγ . Of course, one can act with pγ on a tuple λ̃ entry-wise, which leads to the map p̃γ : (λ1, . . . , λn) 7→ (pγ(λ1), . . . , pγ(λn)). As p̃γ ( λ̃ ) is a reordering of λ̃, we obtain a unique permutation σ, depending on λ̃, by stating that p̃γ ( λ̃ ) = σ · λ̃. Let us say a few more words on how the tuple method is subtly yet significantly different from the labelling method we started with. First, although we write the eigenvalues with indices (just to distinguish them), in principle they are never labelled. Indeed, like in Example 5.1 below, we are interested in the values of the energies, not the names we use for them. Of course, any tuple induces a labelling by assigning label k to the energy appearing in position k, like we also used above. However, σ is not based on these labels, but instead on the tuple positions. That is, σ will send the energy at place k to place σ(k), regardless of what energy is located at place k or which value we designated as λk. Example 5.1. Let us continue with the EP2 example of Example 3.5, where E+(x0) and E−(x0) are exchanged when one encircles an EP. Picking the ordering of Spec(H(x0)) to be (E+(x0), E−(x0)), the above reads p̃γ(E+(x0), E−(x0)) = (E−(x0), E+(x0)) = (12) · (E+(x0), E−(x0)) so that σ = (12). In this case, as n = 2, one also obtains (12) for the alternative ordering (E−(x0), E+(x0)). Consider now the case where we add a separate constant energy level E0(x) = 0, i.e., con- sider the Hamiltonian H ′(x) = H(x) ⊕ 0 which brings us to n = 3. Picking the ordering (E+(x0), E0(x0), E−(x0)) of Spec(H ′(x0)) one finds p̃γ(E+(x0), E0(x0), E−(x0))= (E−(x0), E0(x0), E+(x0))= (13) · (E+(x0), E0(x0), E−(x0)), which thus yields (13). Of course, in the ordering (E+(x0), E−(x0), E0(x0)) we would again obtain (12). Finally, let us demonstrate how the tuple method differs from labelling the energies. Consider we label E+(x0) = E1, E0(x0) = E2 and E−(x0) = E3. In the ordering (E1, E2, E3), clearly σ will replace the labels as expected (up to a convention; here the label Ek will become Eσ−1(k) instead of Eσ(k)). However, in another ordering this need not be so, for example (12) · (E3, E1, E2) = (E1, E3, E2) instead of the tuple (E3, E2, E1) one would obtain by checking the labels. The method of using a reference tuple thus circumvents the need for explicit labels. We can extend this approach to describe the evolution of states, as given by a map Pγ , by a holonomy matrix. Similarly to the previous approach with eigenvalues, we will consider a tuple formulation. In this case, for an operator A ∈ N(V ), we consider tuples f̃ = (f1, . . . , fn) of eigenvectors of A such that each eigenray is represented exactly once. That is, f̃ should be an eigenframe of A, and we write EigFr(A) for the space of all eigenframes of A. This space has a natural action by the wreath product C× ≀ In, isomorphic to the group of complex generalized permutation matrices, which we already encountered in Section 2.2. In this case, the action reads C× ≀ In × EigFr(A) → EigFr(A), ((z1, . . . , zn), σ) · (f1, . . . , fn) = (z1fσ−1(1), . . . , znfσ−1(n)), 32 E.J. Pap, D. Boer and H. Waalkens i.e., the action is similar to the earlier C× ≀ In-action in equation (2.3). Using this action, we can express a map Pγ by an element (z̃, σ) ∈ C× ≀ In once we have chosen a reference eigenframe f̃ . Writing P̃γ for the map on EigFr(A) obtained by applying Pγ entry-wise, we obtain the group element by stating P̃γ(f̃) = (z̃, σ) · f̃ . If one represents (z̃, σ) by a permutation matrix, which we will denote by Mf̃ , then Mf̃ is the matrix of the linear extension of Pγ to the entire state space. The dependence ofMf̃ on f̃ (again keeping the loop γ fixed) follows the familiar conjugation rule. Example 5.2. In case of the EP2, we found Pγ as given by equation (4.5). With respect to the eigenframe (e1, e2) of H(0), we find P̃γ(e1, e2) = (Pγ(e1), Pγ(e2)) = (−ie2,−ie1) = ((−i,−i), (12)) · (e1, e2), i.e., we find ((−i,−i), (12)) ∈ C× ≀ I2. The holonomy matrix M(e1,e2) is thus M(e1,e2) = ( −i 0 0 −i )( 0 1 1 0 ) = ( 0 −i −i 0 ) . Or, alternatively,( Pγ(e1) Pγ(e2) ) = ( −ie2 −ie1 ) = ( 0 −i −i 0 )( e1 e2 ) . The more familiar pattern ψ1 7→ ψ2 7→ −ψ1 is obtained by picking, e.g., (ie1, e2) as the reference frame. Indeed, as P̃γ(ie1, e2) = (e2,−ie1) = ((1,−1), (12)) · (ie1, e2) one finds the group element ((1,−1), (12)) and the holonomy matrix reads M(ie1,e2) = ( 1 0 0 −1 )( 0 1 1 0 ) = ( 0 1 −1 0 ) . The matrices are different only because we picked a different reference frame; both matrices still describe the map Pγ . 5.2 Holonomy interpretation on the frame bundle We now discuss how the approach of Section 5.1 complies with a more general theory on semi- torsors. As shown in [28], one can define the notion of a frame of a semi-torsor. Let us briefly review this here for completeness. If G is a Lie group and F a G-semi-torsor consisting of n orbits, then G×In ∼= F as (left) G-spaces. Such an isomorphism is always of the form (g, i) 7→ gfi for a unique tuple f̃ = (f1, . . . , fn) of elements in F . This tuple f̃ we call a basis of F , similar to linear algebra. Observe that n elements of F form a basis if and only if each orbit in F is represented exactly once. The set of all frames of F we denote by Fr(F ). Clearly, one can translate or permute the elements of a basis. This is captured by a natural action of the wreath product G ≀ In, i.e., G ≀ In × Fr(F ) → Fr(F ), ((g1, . . . , gn), σ) · (f1, . . . , fn) = ( g1fσ−1(1), . . . , gnfσ−1(n) ) . In fact, this action is free and transitive, so that, in other words, Fr(F ) is a G ≀ In-torsor. A Unified View on Geometric Phases and Exceptional Points 33 This is exactly what we did in the previous section. For the eigenvalues, we see that Spec(A) is a set containing n points, hence trivially a semi-torsor with G the trivial group. Each eigenvalue is thus an orbit by itself, and a frame is thus any tuple λ̃ = (λ1, . . . , λn) listing all elements of Spec(A) once. These are exactly the tuples we considered before, i.e., Fr(Spec(A)) = Spec(A)!. This space is naturally endowed with the Sn-action on n-tuples, making it an Sn-torsor. For the eigenstates, we consider the space Eig(A), which we already found to be a C×-semi-torsor. The orbits in this case are the eigenrays, and hence a frame of Eig(A) is a tuple f̃ of eigenvectors such that each eigenray is represented exactly once. Indeed, we again find that f̃ should be an eigenframe of A, i.e., Fr(Eig(A)) = EigFr(A). And this space is clearly a C× ≀ In-torsor. The next step is to lift this procedure of taking frames to the level of semi-principal bundles. We remark that this theory closely resembles that of the frame bundle of a vector bundle. As shown in [28], the operation of taking the frame space induces a functor from semi-principal bundles to principal bundles. Given a semi-principal G-bundle π : B →M , we obtain its frame bundle π! : Fr(B) →M by applying the above fiber-wise; Fr(B) = ⊔ m∈M Fr(Bm). It holds that π! : Fr(B) → M is a principal G ≀ In-bundle. Moreover, any G-connection on B becomes a G ≀ In-connection on Fr(B) by stating that a path in Fr(B) is horizontal if and only if all frame elements traverse horizontal paths. The map B 7→ Fr(B) induces a retracting functor from the semi-principal bundles to the principal bundles. The holonomy interpretation of the explicit permutations and holonomy matrices we found above can now be obtained by applying this frame bundle functor to Spec(V ) resp. Eig(V ). 5.2.1 Frame bundle of the spectrum bundle Let us start with Spec(V ). We found in Theorem 3.4 that πv : Spec(V ) → N(V ) is an In-bundle, hence it is a semi-principal bundle with trivial structure group. The frame bundle functor then yields the following. Proposition 5.3. The frame bundle of Spec(V ) is the principal Sn-bundle E(V ) := Fr(Spec(V )) = {( A, λ̃ ) ∈ N(V )× Cn | λ̃ ∈ Fr(Spec(A)) } , with action σ · (λ1, . . . , λn) = ( λσ−1(1), . . . , λσ−1(n) ) and projection πλ! : E(V ) → N(V ),( A, λ̃ ) 7→ A. Remarkably, we observe that πλ! is also the pull-back of q : Cn → (C n ) along the map Spec: N(V ) → (C n ) ; the fiber Fr(Spec(A)) of E(V ) equals q−1(Spec(A)). In addition, when 34 E.J. Pap, D. Boer and H. Waalkens a Hamiltonian operator family H : M → End(V ) is given, it is again natural to consider the pull-back along H. This yields the bundle πHλ ! : E(H) → N(H), (x, Ẽ) 7→ x, where Ẽ is then a tuple listing the energies of H(x), and is the frame bundle of πHλ . Observe that both pull- backs together result in the following commutative diagram, where all vertical maps are principal Sn-bundles: E(H) E(V ) Cn (x, Ẽ) (H(x), Ẽ) Ẽ N(H) N(V ) (C n ) , x H(x) Spec(H(x)). πH λ ! πλ! q H Spec We observe that E(H) and E(V ) are also the natural spaces behind the merging path method for EP detection, which we reviewed in Section 3.2.1. Namely, the above diagram can alternatively be obtained by completing diagram (3.2) using pull-backs. The holonomy interpretation behind the permutations can be explained as follows. Fix a loop γ in N(H) based at x0, and let Ẽ be the ordering of Spec(H(x0)) w.r.t. which we find the permutation σ expressing pγ . Observe that the picking of the ordering Ẽ = (E1, . . . , En) of Spec(H(x0)) is equivalent to picking the point ( x0, Ẽ ) in the fiber of E(H) above x0. This point ( x0, Ẽ ) fixes a unique lift of γ to E(H). Physically, this lift traces all of the energies simultaneously and in a particular order. The permutation σ now naturally appears via the holonomy theory of principal bundles; the end point of the lift is σ ( x0, Ẽ ) , which fixes σ uniquely. One can alternatively argue by means of holonomy groups. We already found that pγ lies in the holonomy group Hol Spec(H) N(H) (x0). Clearly, pγ extends to a map p̃γ on frames by applying pγ on each frame element separately, and so p̃γ is an element of Hol E(H) N(H)(x0). Both of these groups are based on a point in the base space, and so consist of “abstract” fiber automorphisms. An explicit group element is obtained by considering the holonomy group at a point in the total space. In this case, we move to the holonomy group at ( x0, Ẽ ) defined as HolE(H) ( x0, Ẽ ) = { σ ∈ Sn | ( x0, Ẽ ) ∼ σ ( x0, Ẽ )} , where the equivalence relation ∼ holds if and only if the two points can be connected by a path. The groups are related by the isomorphism Hol E(H) N(H)(x0) → HolE(H) ( x0, Ẽ ) , p̃γ 7→ [ p̃γ ( Ẽ ) /Ẽ ] , where [−/−] denotes the unique group element translating the right entry to the left one. We summarize these findings in the following. Lemma 5.4. Given a loop γ in N(H) based at x0, the permutation σ expressing pγ w.r.t. the or- dering Ẽ of Spec(H(x0)) is the holonomy element at the point ( x0, Ẽ ) in the frame bundle E(H). It follows that we should consider σ to lie in the group HolE(H) ( x0, Ẽ ) . This automatically accounts for the change in σ if we choose another ordering Ẽ′. Geometrically, we see this amounts to picking a different reference point in E(H), namely (x0, Ẽ ′) instead of ( x0, Ẽ ) . Hence, the permutation σ′ obtained w.r.t. Ẽ′ lies in HolE(H) ( x0, Ẽ ′) rather than HolE(H) ( x0, Ẽ ) . We observe that the relation between σ and σ′ is given by the usual conjugation relation between holonomy groups. Namely, there is a unique τ ∈ Sn such that Ẽ′ = τẼ, and consequently σ′ = τστ−1. Note that the frame dependence of σ can also be regarded as a gauge dependence. The gauge freedom is then in the choice of frame of Spec(A), and different gauges are related by an Sn-symmetry. Of course, the gauge invariant behind the permutations σ and σ′ is the map of eigenvalues pγ , or equivalently its frame version p̃γ . Using the framework of holonomy groups, we can be more explicit on how permutations arising from different loops γ are related, as the authors also studied in [27]. Like Spec(H), also A Unified View on Geometric Phases and Exceptional Points 35 E(H) is a covering space of N(H), and so the permutations are prescribed by the monodromy action. With respect to the ordering Ẽ of Spec(H(x0)), this is captured by the homomorphism π1(N(H), x0) → HolE(H) ( x0, Ẽ ) , [γ] 7→ [ p̃γ ( Ẽ ) /Ẽ ] . Clearly this homomorphism is surjective, and its kernel consists of all classes whose paths induce a cyclic change of eigenvectors. In addition, it shows how the holonomy matrix of a complicated loop can be decomposed in holonomy matrices of more elementary loops. This present argument hence improves on a previous proof of this fact given in [27]. Let us briefly compare the holonomy groups of E(V ) and E(H). For E(V ), every holonomy group exhausts all of Sn as all possible paths are present. Indeed, the bundle πλ! : E(V ) → N(V ) does not admit a reduction of the structure group. However, E(H) can have smaller holonomy groups, and admit a reduction of the structure group. In an extreme case, e.g., if the Hamiltonian family is always Hermitian w.r.t. a given inner product on V , then E(H) is trivial as Spec(H) is,4 and the structure group can be reduced to the trivial group 1. Intermediate cases are also possible, and express in what way the energy bands are connected. For example, if k energy bands are disconnected from the remaining n−k, clearly the structure group Sn can be reduced to Sk×Sn−k. Practically, this can be achieved by restricting to tuples Ẽ in which these k energy bands appear only in the first k entries. Example 5.5. Let us continue the study of the standard EP2 system as we described in Ex- ample 3.5. We found N(H) = C \ {±i}, which means π1(N(H), 0) has two generators. Define generators a± to wind once around ±i in positive direction. We found in Example 3.6 that both generators exchange the eigenvalues. Let us label the spectrum Spec(H(0)) = {±1} as (+1,−1). Then the permutations arising from the EPs are expressed by the homomorphism π1(N(H), 0) → HolE(H)(0, (+1,−1)), a± 7→ (12). Clearly, this confirms that a permutation occurs if and only if a loop decomposes into an odd number of a±. As seen in Example 5.1, if an extra level is present, the exact form of the permutation does indeed depend on the labelling. In addition, we observe that for the extended system H ′ presented there, the principal S3-bundle E(H ′) reduces to a S2-bundle, which is isomorphic to E(H). We note that the space E(V ) has appeared in previous papers. We already remarked that [34] mentioned the relevance of the monodromy action in the theory of EPs. The covering used there is of the form E(H), using the one-to-one correspondence between eigenvalues and eigenpro- jectors for non-degenerate operators. We also recognize E(V ) as the space M̃ defined in [21]. However, where [21] views E(V ) primarily as an extended parameter space, we consider it rather as the bundle over N(V ) which yields the permutations due to EPs by its holonomy. 5.2.2 Frame bundle of the eigenvector bundle We now proceed by considering the frame bundle of πλv : Eig(V ) → N(V ). This is a semi- principal C×-bundle by Theorem 4.4 and endowed with the C×-connection ω as in Proposi- tion 4.7. Applying the frame bundle functor then yields the following. 4An alternative argument: if the energies of H are real, then E(H) is a pull-back of Rn → (R n ) , hence trivial by Proposition A.3. 36 E.J. Pap, D. Boer and H. Waalkens Proposition 5.6. The frame bundle of Eig(V ) is the principal C× ≀ In-bundle EigFr(V ) = {( A, f̃ ) ∈ N(V )× Fr(V ) | f̃ is an eigenframe of A } , where the element ( A, f̃ ) can also be written as ( A, λ̃, f̃ ) with λ̃ listing the eigenvalues of A in the order of f̃ . The C× ≀ In-action on EigFr(V ) reads (z̃, σ) · ( A, λ̃, f̃ ) = ( A, σλ̃, z̃ ( σf̃ )) and the projection is πλv! : EigFr(V ) → N(V ),( A, λ̃, f̃ ) 7→ A. Furthermore, EigFr(V ) is naturally endowed with the C× ≀ In-connection ω!(A,λ̃,f̃) = ω(A,λ1,f1) ⊕ · · · ⊕ ω(A,λn,fn). Recall that we already found the principal C× ≀ In-bundle Ξ: Fr(V ) × Cn → N(V ) in Section 2.2. We observe that Ξ and πλv! are reformulations of each other, i.e., they are canonically isomorphic bundles. Lemma 5.7. The map EigFr(V ) → Fr(V )× Cn,( A, λ̃, f̃ ) 7→ ( f̃ , λ̃ ) is a canonical isomorphism of bundles over N(V ). Hence, in particular, EigFr(V ) is canonically isomorphic to Fr(V ) × Cn as a C× ≀ In-manifold, and consequently E(V ) is canonically diffeo- morphic to PFr(V )× Cn. Again, given a Hamiltonian family H, it is more natural to study the pull-back bundle. This yields the space EigFr(H), whose elements we write as ( x, Ẽ, ψ̃ ) . The connection form ω! also carries over to the connection form ωH !. Note that EigFr(H) may admit a reduction of the structure group, e.g., to U(1)n when H is Hermitian. We can summarize the pull-back and relation with Ξ in diagram (5.1). EigFr(H) EigFr(V ) Fr(V )× Cn ( x, Ẽ, ψ̃ ) ( H(x), Ẽ, ψ̃ ) ( ψ̃, Ẽ ) N(H) N(V ) N(V ), x H(x) H(x). πH λv ! ∼ πλv ! Ξ H id (5.1) We can now view the holonomy interpretation of the map on eigenstates as follows. Again, we fix a family H and a loop γ in N(H) based at some reference x0. We then pick an eigenframe ψ̃ of H(x0), which fixes the tuple Ẽ of corresponding energies. That is, we pick a point ( x0, Ẽ, ψ̃ ) in the fiber of EigFr(V ) above x0. This point fixes a unique lift of γ to EigFr(V ), which traces all states in the tuple ψ simultaneously. The end point of this lift then equals (z̃, σ) ( x0, Ẽ, ψ̃ ) , where (z̃, σ) ∈ C× ≀ In we obtained as before. Hence it lies in the holonomy group HolEigFr(H) ( x0, Ẽ, ψ̃ ) = { (z̃, σ) ∈ C× ≀ In | ( x0, Ẽ, ψ̃ ) ∼ (z̃, σ) ( x0, Ẽ, ψ̃ )} , where the equivalence relation ∼ means that the points can be connected by a horizontal path in EigFr(H), or equivalently, that there is a path in N(H) whose lift to EigFr(H) through one A Unified View on Geometric Phases and Exceptional Points 37 point ends at the other. Again, we consider the holonomy group at a point in the total space rather than in the base space. Clearly, σ is the same permutation as we would find when looking only at the energies. This we already found in Proposition 4.18, which is based on the map πv : Eig(V ) → Spec(V ). However, we can also phrase this on the frame bundles as follows. Namely, the frame bun- dle functor turns πv into the bundle map πv! : EigFr(V ) → E(V ), ( A, λ̃, f̃ ) 7→ ( A, λ̃ ) . This map is equivariant w.r.t. the canonical quotient C× ≀ In → Sn, and is compatible with the connections. Fixing a point ( A, λ̃, f̃ ) , this quotient reappears as a map on holonomy groups HolEigFr(V ) ( A, λ̃, f̃ ) → HolE(V ) ( A, λ̃ ) . This map simply extracts σ, which confirms that σ is the holonomy from E(V ). We summarize the above findings as follows. Lemma 5.8. Given a loop γ in N(H) based at x0, the element (z̃, σ) expressing Pγ w.r.t. the eigenframe ψ̃ of H(x0) is the holonomy element at the point (x0, ψ̃) in the frame bundle EigFr(H). Writing Ẽ for the corresponding ordering of Spec(H(x0)), σ is the permutation representing pγ w.r.t. Ẽ. The element (z̃, σ) thus lies in the group HolEigFr(H)(x0, ψ̃). We can view the picking of ψ̃ as a gauge freedom, just like the picking of Ẽ earlier. A change of gauge thus corresponds to switching holonomy group, with the usual similarity transformation of the holonomy matrix. The gauge invariant in this case is the map Pγ . Indeed, Pγ is defined without using any gauge, hence is trivially independent of gauge. Let us then take a closer look at the phase factors z̃ appearing in the holonomy matrix. Given the parallel transport formalism, each of these may be called a “geometric” phase factor. However, we also see that they are gauge dependent. Namely, the order gets permuted by changing the order of the eigenframe ψ̃, and scaling of the individual states in ψ̃ affects the phase factors in z̃ corresponding to non-cyclic evolution. On the other hand, z̃ is subject to constraints coming from the characteristic phases of Pγ . Parametrizing zj = exp(iαj), if we follow a state until it returns to its original ray, its acquired phase is ∑ cycle αi, where we sum over the indices belonging to the corresponding cycle. Hence ∑ cycle αi must be equal to the characteristic phase of the cycle, independent of the chosen gauge. Note however that this does not yield a canonical phase for a non-cyclic state after following γ once; in one gauge all αi in the cycle could be equal, whereas in another all except one could vanish. In terms of the holonomy matrix, we can express this as follows. By permuting ψ̃ the phase factors can be moved within the matrix, which is subject to the underlying cycle structure of the permutation. The matrix is block-diagonal if and only if states from the same minimal union are listed adjacently. The size of a block is k × k, with k the length of the corresponding cycle. The k individual phase factors within this block can be gauged to arbitrary values, with the only condition that their product is fixed. Let us treat in more detail how a seemingly Abelian connection ωH ! can still induce non- Abelian holonomy. That is, given the technique of path-ordered exponential and the diagonal form of ωH !, one may wonder how the end result can be non-diagonal. In short, the reason is that if the loop γ lifts to non-cyclic states, then any instantaneous eigenframe returns with a twist, and this twist yields an additional permutation matrix making the theory non-Abelian. That is, in order to write ωH ! as a diagonal gauge potential, in practice one would need to extend the initial eigenframe ψ̃ to a path of instantaneous eigenframes over γ. Upon returning to x0, one cannot use the initial eigenframe ψ̃ again due to the presence of non-cyclic states. Consequently, one needs to correct for the difference between final and initial eigenframe. This basis transformation can be expressed by an element in C× ≀ In, and thus permutation matrices enter the end result. Naturally, this issue can only occur when E(H) is non-trivial and γ is a non-contractible loop, and so is completely absent in case π1(N(H), x0) is trivial. 38 E.J. Pap, D. Boer and H. Waalkens Let us also compare EigFr(V ) to earlier work in [21], which is likewise aimed to provide a holonomy description for geometric phases in the presence of EPs. There, one considers cyclic states under the adiabatic evolution due to a change of a Hamiltonian, which may be non-Hermitian. The general geometric model consists of n complex line bundles L1, . . . , Ln over E(V ), there called M̃, where each line bundle corresponds to a certain self-chosen part of the energy bands. A main disadvantage of this approach is that the connection information is spread over the Lk. Hence, calculation of a lift can in general not be done in a single Lk, i.e., multiple line bundles need to be used. That is, formally the Lk do not support a holonomy interpretation for geometric phases in the presence of EPs. In contrast, we found that EigFr(V ) is a single bundle allowing to lift a path directly from N(V ), which can moreover be used to treat non-cyclic states. It is also the case that the theory of EigFr(V ) provides a geometric model for the off-diagonal phases reported by Manini and Pistolesi [20], and extends the generalized permutation matrices found by Tanaka, Cheon and Kim [35]. The generalization of their work to non-Hermitian Hamiltonians can be obtained from EigFr(H) in the following way. First, let us clarify that there are 3 main classes of paths that are considered in this theory. In the above, we considered loops γ only, i.e., the final point equals the initial point x0. In con- trast, [20] starts by considering general paths, so that the final point x1 need not equal x0. They then consider a more special case, which is also treated in [35] and there called an “adiabatic loop”. This term refers to a path γ for which x1 need not equal x0, but still Eig(H(x0)) = Eig(H(x1)) (as subsets of V , strictly speaking not as fibers of Eig(H)). We thus observe that an “adiabatic loop” need not be a “loop” in the sense we use, and forms an inter- mediate subset:{ loops; x0 = x1 } ⊂ { “adiabatic loops”; Eig(H(x0)) = Eig(H(x1)) } ⊂ { general paths; no relation x0, x1 } . Let us now demonstrate that EigFr(V ) also describes the theory of the general paths, so that we obtain the “adiabatic loops” as a special case. Let γ be any path in N(H), so belonging to the most right set in the above comparison. Lifting to EigFr(V ), it induces the parallel transport map P̃γ : EigFr(H(x0)) → EigFr(H(x1)). More explicitly, given an initial eigenframe ψ̃ of H(x0), one obtains a lift Γ̃ of γ starting at ψ̃, and the final frame is ψ̃′ = P̃γ(ψ̃). However, as ψ̃′ and ψ̃ are (generally) not in the same fiber, the two frames cannot be compared using the bundle structure of EigFr(H). The solution is to consider the frames by themselves; one neglects the fact that they are associated to different values of the system parameters. Clearly, any two frames of V are related by a unique matrix; in this case there is a unique U ∈ GL(n,C) such that ψ̃′ = Uψ̃. Mathematically, this argument is captured by employing the canonical projection EigFr(V ) → Fr(V ), and using that Fr(V ) is a GL(n,C)-torsor. We also remark that if both ψ̃ and ψ̃′ are orthonormal w.r.t. some inner product, then U is indeed unitary, and can be calculated by taking inner products between the states, as used in [20]. Of course, the matrix U is not unique; if we change ψ̃ to Gψ̃, with G ∈ C× ≀ In a generalized permutation matrix, then ψ̃′ = P̃γ ( ψ̃ ) becomes P̃γ ( Gψ̃ ) = GP̃γ ( ψ̃ ) = Gψ̃′, and so U becomes GUG−1. That is, U is unique up to conjugation by C× ≀ In. When looking for invariants, if we write G = (g̃, τ), we may set τ = id; non-trivial τ simply rearrange the elements of U . It follows that Uij becomes giUijg −1 j . Hence, one obtains an invariant quantity for each cycle in Sn A Unified View on Geometric Phases and Exceptional Points 39 by taking the corresponding product, e.g., (12) yields the invariant U12U21 and (123) yields the invariant U12U23U31, in addition to the diagonal elements Ukk which are also invariant. Whenever such a product is non-zero, its argument is a gauge-invariant phase, as reported in [20]. The “adiabatic loops” are then the special case where U is also a generalized permutation matrix, which was studied further in [35]. The similarity of this case with our study of the holo- nomy on EigFr(V ) is straightforward to explain; in both cases EigFr(H(x0)) = EigFr(H(x1)), so that we are studying an automorphism of a semi-torsor. The method above, based on the matrix U , then reduces to the argument of Section 5.1, and invariant phases follow from Propo- sition 4.16. 6 Discussion In this paper we introduced a framework that facilitates the mathematical description of the adiabatic quantum mechanics of finite-dimensional, non-degenerate but otherwise arbitrary and hence not necessarily Hermitian Hamiltonians. We started with a preparatory study of the space of non-degenerate operators, which was summarized in diagram (2.6). The following step was to find the geometry behind the adiabatic change of energies of a Hamiltonian fam- ily H. This is captured by the covering space formed by the energy bands, where we re- strict to system parameter values where H is non-degenerate. Moreover, the specific covering needed for H we found as the pull-back of an abstract yet explicit model covering, which makes the covering property and hence bundle structure immediate for any smooth Hamiltonian fa- mily H. We continued by introducing a bundle for the eigenstates. A remarkable property of this bundle is that it is not a principal bundle, but has the structure of a semi-principal bundle, which is described in [28]. That is, each of its fibers consists of a collection of eigenrays rather than a single eigenray. We found a natural connection on this bundle, which provides a par- allel transport description of the adiabatic geometric phase as found by Garrison and Wright [12]. Because of the semi-principal structure, this incorporates non-cyclic states as well. This formalism hence extends previous holonomy formulations, which are based on principal bun- dles. Among these are the first holonomy interpretation reported by Simon [32], which treats the Hermitian case with separate energy bands, and the formalism in [21], which considered the non-Hermitian case restricted to cyclic states. We furthermore found that the Aharonov– Anandan formalism [1] for general Hermitian systems can be obtained by a reduction from the present formalism using a given inner product. Moreover, we found that the dynami- cal phase can be included as a non-geometric addition, which means that the full adiabatic evolution of a state can be calculated naturally on this bundle. We then obtained a gener- alized quantum geometric tensor, which we found to be directly related to covariant deriva- tives. We highlight that the presented formalism, as it allows one to also describe non-cyclic states by means of parallel transport, is a natural formalism for the study of exceptional points. The description including geometric phases is then given on the full eigenstate bundle, whereas the permutations of energies can already be accurately described with the monodromy theory of the energy band covering. Both of these spaces have a semi-principal bundle structure, which is the key to incorporate non-cyclic behavior. However, by switching to a multi-state approach, one can describe the permutations of energies and the accompanying geometric phase factors simultaneously in a more explicit way, namely via holonomy matrices. This is given by a straightforward application of the frame bundle technique as described in [28]. This provides a more explicit holonomy description of state evolution, also in the non-cyclic case enabled by exceptional points. 40 E.J. Pap, D. Boer and H. Waalkens A Bundle structure of tuples of distinct complex numbers If one considers n distinct complex numbers, one can do so with or without an ordering of these numbers. Let us start with the ordered one. This brings us to the space Cn = { (z1, . . . , zn) ∈ Cn | i ̸= j =⇒ zi ̸= zj } , where we borrow the notation of falling factorials; if we would replace C by a finite set F , then |Fn| = |F |(|F | − 1) · · · (|F | − n), which in combinatorics is known as the falling factorial |F |n. The space Cn is a submanifold of Cn, and more is true. Lemma A.1. The space Cn is an (algebraic) open and dense submanifold of Cn. Proof. For i ̸= j, define the function mij : Cn → C, (z1, . . . , zn) 7→ zi− zj , and write D(mij) ={ (z1, . . . , zn) ∈ Cn | mij(z1, . . . , zn) ̸= 0 } . Then Cn = ∩i ̸=jD(mij) expresses the space as a finite intersection of (algebraic) open subsets, hence is open in Cn and so a submanifold. It is also dense; given a tuple with repeated numbers, these become unequal by an arbitrarily small perturbation. ■ We can now proceed to sets of n distinct complex numbers, which is the unordered variant of the ordered tuple case as just discussed. The idea is to view the unordered sets as a quotient of the tuple space. This quotient is specified by a permutation action; upon identifying rearranged tuples, clearly the ordering vanishes. The action is the usual permutation of slots, i.e., for σ ∈ Sn and (z1, . . . , zn) ∈ Cn we have σ · (z1, . . . , zn) = (zσ−1(1), . . . , zσ−1(n)). This action is smooth as it is the restriction of the smooth permutation action on Cn to an invariant submanifold. Taking the quotient results in the space of subsets of C with n distinct complex numbers, which we write as( C n ) := Cn/Sn = { S ⊂ C | |S| = n } . Again, we borrow our notation from combinatorics. We write q for the quotient map Cn → (C n ) . Clearly, the fiber of q above a certain set in (C n ) consists of all the possible orderings of its elements. Via q one obtains a unique manifold structure on (C n ) , and in this way q is principal bundle. Proposition A.2. There is a unique manifold structure on (C n ) such that the quotient map q : Cn → (C n ) is a surjective submersion. The permutation action on Cn thus induces the principal bundle Sn Cn (C n ) . q Moreover, this bundle is non-trivial for n > 1. Proof. Clearly the Sn-action on Cn is free and proper, hence the quotient map defines a prin- cipal bundle. Assume it is trivial for n > 1 for contradiction. This would imply that Cn and (C n ) ×Sn are homeomorphic. However, as Cn and (C n ) are path-connected but Sn is not, this cannot be. ■ A Unified View on Geometric Phases and Exceptional Points 41 The non-triviality of the bundle is related to the topology of C not being compatible with a total ordering. To expand on this, observe that for any manifold M one can define the space of distinct tuples Mn. This similarly has a Sn-action defined on it, and one obtains a principal bundle over the space ( M n ) of unordered subsets of n elements. IfM has an order topology given by a total ordering, such as R, this bundle is trivial. Indeed, given any element S ∈ ( M n ) , then there is a standard way to list the elements of S, e.g., from lowest to highest. This induces triviality as follows. Proposition A.3. If a manifold M is orderable, then the principal Sn-bundle M n → ( M n ) is trivial for all n. Proof. Let ≤ be a total ordering on M compatible with the topology. Define an ordering map o : ( M n ) →Mn by the condition that o({m1, . . . ,mn}) = (m1, . . . ,mn), where m1 < · · · < mn. As all the mi are distinct this is well-defined, and by assumption on the topology of M also continuous. It follows that o is a global section of the principal bundle, which is thus trivial. ■ Acknowledgements The authors thank the anonymous referees whose careful remarks contributed to the quality of the paper. References [1] Aharonov Y., Anandan J., Phase change during a cyclic quantum evolution, Phys. Rev. Lett. 58 (1987), 1593–1596. [2] Anandan J., Aharonov Y., Geometry of quantum evolution, Phys. Rev. Lett. 65 (1990), 1697–1700. [3] Berry M.V., Quantal phase factors accompanying adiabatic changes, Proc. Roy. Soc. London Ser. A 392 (1984), 45–57. [4] Berry M.V., The quantum phase, five years after, in Geometric Phases in Physics, Editors A. Shapere, F. Wilczeck, Adv. Ser. Math. Phys., Vol. 5, World Sci. Publ., Teaneck, NJ, 1989, 7–28. [5] Berry M.V., Uzdin R., Slow non-Hermitian cycling: exact solutions and the Stokes phenomenon, J. Phys. A: Math. Theor. 44 (2011), 435303, 26 pages. [6] Berry M.V., Wilkinson M., Diabolical points in the spectra of triangles, Proc. Roy. Soc. London Ser. A 392 (1984), 15–43. [7] Born M., Fock V., Beweis des Adiabatensatzes, Z. Phys. 51 (1928), 165–180. [8] Chruściński D., Jamio lkowski A., Geometric phases in classical and quantum mechanics, Progress in Mathe- matical Physics, Vol. 36, Birkhäuser Boston, Inc., Boston, MA, 2004. [9] Cohen E., Larocque H., Bouchard F., Nejadsattari F., Gefen Y., Karimi E., Geometric phase from Aharonov– Bohm to Pancharatnam–Berry and beyond, Nature Rev. Phys. 1 (2019), 437–449, arXiv:1912.12596. [10] Cui X.-D., Zheng Y., Geometric phases in non-Hermitian quantum mechanics, Phys. Rev. A 86 (2012), 064104, 4 pages. [11] Cui X.-D., Zheng Y., Unification of the family of Garrison–Wright’s phases, Sci. Rep. 4 (2014), 5813, 8 pages. [12] Garrison J.C., Wright E.M., Complex geometrical phases for dissipative systems, Phys. Lett. A 128 (1988), 177–181. [13] Heiss W.D., Phases of wave functions and level repulsion, Eur. Phys. J. D 7 (1999), 1–4, arXiv:quant- ph/9901023. [14] Heiss W.D., The physics of exceptional points, J. Phys. A: Math. Theor. 45 (2012), 444016, 11 pages, arXiv:1210.7536. https://doi.org/10.1103/PhysRevLett.58.1593 https://doi.org/10.1103/PhysRevLett.65.1697 https://doi.org/10.1098/rspa.1984.0023 https://doi.org/10.1088/1751-8113/44/43/435303 https://doi.org/10.1088/1751-8113/44/43/435303 https://doi.org/10.1098/rspa.1984.0022 https://doi.org/10.1007/BF01343193 https://doi.org/10.1007/978-0-8176-8176-0 https://doi.org/10.1007/978-0-8176-8176-0 https://doi.org/10.1038/s42254-019-0071-1 https://arxiv.org/abs/1912.12596 https://doi.org/10.1103/PhysRevA.86.064104 https://doi.org/10.1038/srep05813 https://doi.org/10.1016/0375-9601(88)90905-X https://doi.org/10.1007/s100530050339 https://arxiv.org/abs/quant-ph/9901023 https://arxiv.org/abs/quant-ph/9901023 https://doi.org/10.1088/1751-8113/45/44/444016 https://arxiv.org/abs/1210.7536 42 E.J. Pap, D. Boer and H. Waalkens [15] Höller J., Read N., Harris J.G.E., Non-Hermitian adiabatic transport in spaces of exceptional points, Phys. Rev. A 102 (2020), 032216, 9 pages, arXiv:1809.07175. [16] Kato T., On the adiabatic theorem of quantum mechanics, J. Phys. Soc. Japan 5 (1950), 435–439. [17] Kato T., Perturbation theory for linear operators, Grundleheren der Mathematischen Wissenschaften, Vol. 132, Springer, Berlin – Heidelberg, 1966. [18] Lee J.M., Introduction to smooth manifolds, Graduate Texts in Mathematics, Vol. 218, Springer, New York, 2012. [19] Mandel A., Gonçalves J.Z., Construction of open sets of free k-tuples of matrices, Proc. Edinburgh Math. Soc. 30 (1987), 121–131. [20] Manini N., Pistolesi F., Off-diagonal geometric phases, Phys. Rev. Lett. 85 (2000), 3067–3071, arXiv:quant- ph/9911083. [21] Mehri-Dehnavi H., Mostafazadeh A., Geometric phase for non-Hermitian Hamiltonians and its holonomy interpretation, J. Math. Phys. 49 (2008), 082105, 17 pages, arXiv:0807.3405. [22] Milburn T.J., Doppler J., Holmes C.A., Portolan S., Rotter S., Rabl P., General description of quasiadiabatic dynamical phenomena near exceptional points, Phys. Rev. A 92 (2015), 052124, 12 pages, arXiv:1410.1882. [23] Miri M.-A., Alù A., Exceptional points in optics and photonics, Science 363 (2019), 42 pages. [24] Mostafazadeh A., Exact PT -symmetry is equivalent to Hermiticity, J. Phys. A: Math. Gen. 36 (2003), 7081–7091, arXiv:quant-ph/0304080. [25] Nenciu G., Rasche G., On the adiabatic theorem for nonselfadjoint Hamiltonians, J. Phys. A: Math. Gen. 25 (1992), 5741–5751. [26] Pancharatnam S., Generalized theory of interference, and its applications. I. Coherent pencils, Proc. Indian Acad. Sci. Sect. A 44 (1956), 247–262. [27] Pap E.J., Boer D., Waalkens H., Non-Abelian nature of systems with multiple exceptional points, Phys. Rev. A 98 (2018), 023818, 9 pages, arXiv:1805.04291. [28] Pap E.J., Waalkens H., Frames of group-sets and their application in bundle theory, arXiv:2010.03913. [29] Pati A.K., Relation between “phases” and “distance” in quantum evolution, Phys. Lett. A 159 (1991), 105–112. [30] Provost J.P., Vallee G., Riemannian structure on manifolds of quantum states, Comm. Math. Phys. 76 (1980), 289–301. [31] Shapere A., Wilczek F. (Editors), Geometric phases in physics, Advanced Series in Mathematical Physics, Vol. 5, World Sci. Publ. Co., Inc., Teaneck, NJ, 1989. [32] Simon B., Holonomy, the quantum adiabatic theorem, and Berry’s phase, Phys. Rev. Lett. 51 (1983), 2167– 2170. [33] Tanaka A., Cheon T., Bloch vector, disclination and exotic quantum holonomy, Phys. Lett. A 379 (2015), 1693–1698, arXiv:1409.5211. [34] Tanaka A., Cheon T., Path topology dependence of adiabatic time evolution, in Functional Analysis and Operator Theory for Quantum Physics, Editors J. Dittrich, H. Kovař́ık, A. Laptev, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich, 2017, 531–542, arXiv:1512.06983. [35] Tanaka A., Cheon T., Kim S.W., Gauge invariants of eigenspace and eigenvalue anholonomies: examples in hierarchical quantum circuits, J. Phys. A: Math. Theor. 45 (2012), 335305, 20 pages, arXiv:1203.5412. [36] Uzdin R., Mailybaev A., Moiseyev N., On the observability and asymmetry of adiabatic state flips generated by exceptional points, J. Phys. A: Math. Theor. 44 (2011), 435302, 8 pages. [37] Wilczek F., Zee A., Appearance of gauge structure in simple dynamical systems, Phys. Rev. Lett. 52 (1984), 2111–2114. [38] Wojcik C.C., Sun X.-Q., Bzdušek T., Fan S., Homotopy characterization of non-Hermitian Hamiltonians, Phys. Rev. B 101 (2020), 205417, 17 pages, arXiv:1911.12748. [39] Xu H., Mason D., Jiang L., Harris J.G.E., Topological energy transfer in an optomechanical system with exceptional points, Nature 537 (2016), 80–83, arXiv:1602.06881. [40] Zhang D.-J., Wang Q.-H., Gong J., Quantum geometric tensor in PT-symmetric quantum mechanics, Phys. Rev. A 99 (2018), 042104, 13 pages, arXiv:1811.04638. https://doi.org/10.1103/physreva.102.032216 https://doi.org/10.1103/physreva.102.032216 https://arxiv.org/abs/1809.07175 https://doi.org/10.1143/JPSJ.5.435 https://doi.org/10.1007/978-3-662-12678-3 https://doi.org/10.1007/978-1-4419-9982-5 https://doi.org/10.1017/S0013091500018046 https://doi.org/10.1017/S0013091500018046 https://doi.org/10.1103/PhysRevLett.85.3067 https://arxiv.org/abs/quant-ph/9911083 https://arxiv.org/abs/quant-ph/9911083 https://doi.org/10.1063/1.2968344 https://arxiv.org/abs/0807.3405 https://doi.org/10.1103/physreva.92.052124 https://arxiv.org/abs/1410.1882 https://doi.org/10.1126/science.aar7709 https://doi.org/10.1088/0305-4470/36/25/312 https://arxiv.org/abs/quant-ph/0304080 https://doi.org/10.1088/0305-4470/25/21/027 https://doi.org/10.1007/BF03046050 https://doi.org/10.1007/BF03046050 https://doi.org/10.1103/physreva.98.023818 https://doi.org/10.1103/physreva.98.023818 https://arxiv.org/abs/1805.04291 https://arxiv.org/abs/2010.03913 https://doi.org/10.1016/0375-9601(91)90255-7 https://doi.org/10.1007/BF02193559 https://doi.org/10.1142/0613 https://doi.org/10.1103/PhysRevLett.51.2167 https://doi.org/10.1016/j.physleta.2015.05.009 https://arxiv.org/abs/1409.5211 https://doi.org/10.4171/175-1/26 https://arxiv.org/abs/1512.06983 https://doi.org/10.1088/1751-8113/45/33/335305 https://arxiv.org/abs/1203.5412 https://doi.org/10.1088/1751-8113/44/43/435302 https://doi.org/10.1103/PhysRevLett.52.2111 https://doi.org/10.1103/physrevb.101.205417 https://arxiv.org/abs/1911.12748 https://doi.org/10.1038/nature18604 https://arxiv.org/abs/1602.06881 https://doi.org/10.1103/PhysRevA.99.042104 https://doi.org/10.1103/PhysRevA.99.042104 https://arxiv.org/abs/1811.04638 1 Introduction 2 The non-degeneracy space 2.1 Discriminant definition 2.2 Parametrizing the non-degeneracy space 2.3 Spectrum map on non-degeneracy space 2.4 Summarizing diagram 3 Eigenvalue bundle and exceptional points 3.1 The spectrum bundle 3.2 Geometry behind swaps of energies 3.2.1 Merging path method 3.2.2 Real eigenvalue case 4 Eigenvector bundle and geometric phases 4.1 The eigenvector bundle 4.2 Connection on the eigenvector bundle 4.3 Geometry behind the geometric phase 4.3.1 Calculating the lift via an ansatz 4.3.2 The cyclic case 4.3.3 The non-cyclic case 4.4 Including the dynamical phase 4.5 Relation with the work of Aharonov and Anandan 4.6 Quantum geometric tensor 4.6.1 Bases for tangent space of spectrum and eigenvector bundles 4.6.2 Generalized quantum geometric tensor 4.6.3 Relation between geometric phase and distance 5 Explicit description of the holonomy 5.1 Explicit permutations and holonomy matrices 5.2 Holonomy interpretation on the frame bundle 5.2.1 Frame bundle of the spectrum bundle 5.2.2 Frame bundle of the eigenvector bundle 6 Discussion A Bundle structure of tuples of distinct complex numbers References
id nasplib_isofts_kiev_ua-123456789-211542
institution Digital Library of Periodicals of National Academy of Sciences of Ukraine
issn 1815-0659
language English
last_indexed 2026-03-14T04:12:43Z
publishDate 2022
publisher Інститут математики НАН України
record_format dspace
spelling Pap, Eric J.
Boer, Daniël
Waalkens, Holger
2026-01-05T12:30:38Z
2022
A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics. Eric J. Pap, Daniël Boer and Holger Waalkens. SIGMA 18 (2022), 003, 42 pages
1815-0659
2020 Mathematics Subject Classification: 81Q70; 81Q12; 55R99
arXiv:2107.02497
https://nasplib.isofts.kiev.ua/handle/123456789/211542
https://doi.org/10.3842/SIGMA.2022.003
We present a formal geometric framework for the study of adiabatic quantum mechanics for arbitrary finite-dimensional non-degenerate Hamiltonians. This framework generalizes earlier holonomy interpretations of the geometric phase to non-cyclic states appearing for non-Hermitian Hamiltonians. We start with an investigation of the space of non-degenerate operators on a finite-dimensional state space. We then show how the energy bands of a Hamiltonian family form a covering space. Likewise, we demonstrate that the eigenrays form a bundle, a generalization of a principal bundle, which admits a natural connection that yields the (generalized) geometric phase. This bundle also provides a natural generalization of the quantum geometric tensor and derived tensors, and we demonstrate how it can incorporate the non-geometric dynamical phase as well. We conclude by demonstrating how the bundle can be recast as a principal bundle, allowing both the geometric phases and the permutations of eigenstates to be expressed simultaneously using standard holonomy theory.
The authors thank the anonymous referees whose careful remarks contributed to the quality of the paper.
en
Інститут математики НАН України
Symmetry, Integrability and Geometry: Methods and Applications
A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
Article
published earlier
spellingShingle A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
Pap, Eric J.
Boer, Daniël
Waalkens, Holger
title A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
title_full A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
title_fullStr A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
title_full_unstemmed A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
title_short A Unified View on Geometric Phases and Exceptional Points in Adiabatic Quantum Mechanics
title_sort unified view on geometric phases and exceptional points in adiabatic quantum mechanics
url https://nasplib.isofts.kiev.ua/handle/123456789/211542
work_keys_str_mv AT papericj aunifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics
AT boerdaniel aunifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics
AT waalkensholger aunifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics
AT papericj unifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics
AT boerdaniel unifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics
AT waalkensholger unifiedviewongeometricphasesandexceptionalpointsinadiabaticquantummechanics