sign in sign up
<<
Home , Hermitian adjoint

arXiv:quant-ph/0105054v1 11 May 2001 h nepeaino unu-ehnclMdl ihNon- with Models Quantum-Mechanical of Interpretation The h hsclmaigo hs aefntosdeserves functions wave these study. of additional meaning physical the ota h oino o-emtct ntesneo the of sense the in non-Hermiticity of space notion the that so 36.w 36.d 73.20.Jc 03.65.Fd, 03.65.-w, Hatano of model the and Nelson. al. and et re Cannata model by th the introduced including Within cently examples, . various physical consider we of scheme algebra the of definition nnHriiiymathr ntesneo h space the of sense the in L here meant (non-Hermiticity n olbrtr nasre fppr 2 investigated [2] papers of series non-Hermitian, a Bender some in Next, collaborators de- Hamiltonian. a and non-Hermitian be that work a can suggested the by was superconductors scribed is it in where systems transition [1] such delocalization Nelson use and Hatano to of attempts earliest the em fHriinHmloin endi h sa Hilbert usual in the systems in such defined space of We Hamiltonians analysis Hermitian the of spectra. for terms framework real general a and up Hamiltonians set non-Hermitian with els in r o lmnso h ibr space Hilbert func- the wave of the elements that not means defini- values are This the tions complex [2,16]. extend to coordinate to functions the necessary wave of Bender example, by position-space for of treated is, tion models it the al. In et [12–15]. to addressed started be only have interpretation [3–11]. quantum-mechanical field this in triggered activity has of This lot eigenvalues. a real have they that guing † ∗ 2 -al [email protected] E-mail: -al [email protected] E-mail: eety oeswt o-emta Hamiltonians non-Hermitian with models Recently, ept hs oefnaetlise ocrigthe concerning issues fundamental some this, Despite ( esuyteqatmmcaia nepeaino mod- of interpretation quantum-mechanical the study We −∞ L L 2 2 , ( ∞ −∞ osntse ob sflhr.I u opinion, our In here. useful be to seem not does )hv trce o fitrs.Oeof One interest. of lot a attracted have )) , ∞ .Seilepai spto h correct the on put is emphasis Special ). .INTRODUCTION I. PT ymti aitnas ar- Hamiltonians, symmetric ntttfu hoeicePyi,Universit¨at Regensbu Physik, f¨ur Theoretische Institut otnIsiuefrNcerSuis asw Poland Warsaw, Studies, Nuclear for Institute Soltan aheec hsk nvri¨tSee,Germany Universit¨at Siegen, Physik, Fachbereich L 2 ( n elSpectra Real and −∞ .Szymanowski L. .Kretschmer R. , 1 a 2001) May (11 ∞ ), is - and 1 inI esatb osrcigaHletspace Hilbert a constructing by start we II tion lsdsse,adteeoehst eHriinin Hermitian be for to evolution has time therefore be space and the to the system, of is closed generator Hamiltonian a the the as that interpreted demanding by paths. product complex along defined g., are e. that consider, we products [2], scalar sug- collaborators approach and the Bender by by Motivated gested interpre- arbitrary. probabilistic quite a is to tation link the provides that product tutr of structure oini nyafnto ftoHriinoperators Hermitian Hamil- two of the which function theory. a in only formulation the is canonical in tonian a observables up physical set of con- We algebra we the Next sider interpretation. quantum-mechanical tent auso Hamiltonian a eigen- of real with values eigenfunctions of superpositions contains hoe yvnNuan[7 omptespace the map uniqueness to the [17] employ Neumann its von then of by can theorem generator One the and resp. variable translations, position-space interpreta- a an as have tion conditions some under may operators and space sbsdo h oe nrdcdb ant ta.[18] by al. given et is Cannata potential by example the introduced which second model in The the on does information. based formulation new is canonical any the contain that half show not real we the Here on Hermi- axis. problem well-known, Coulomb the one-dimensional simple on tian a is based example first model The non-Hermitian III. Section in examples able uniquely. eas osrc aoia formulation canonical a construct also we ihHermitian with ainde o lo h nepeainof interpretation the allow not does lation lie nApni .I pedxBw reycom- briefly we B Appendix In ex- are A. III Appendix Section in in used plained are that details of technical Some spirit the in [1] Nelson and approach. im- Hatano our phenomenologically this of the end model of We portant discussion unclear. a remains with in- the Section model that the so of resp., terpretation operators, momentum and position nti ae ewl drs hs usin.I Sec- In questions. these address will we paper this In ermv h rirrns o h hieo h scalar the of choice the for arbitrariness the remove We h hl osrcini ple otoeatysolv- exactly two to applied is construction whole The u ocuigrmrsaecnandi eto IV. Section in contained are remarks concluding Our p ∗ † c L ufiln aoia omtto eain.Such relations. commutation canonical fulfilling 2 nti a,tepyia nepeaini fixed is interpretation physical the way, this In . H hssillae oeabtaiesfrthe for arbitrariness some leaves still This . H u l loe hoishv consis- a have theories allowed all but , H g Germany, rg, u hnw hwta hsformu- this that show we then but , emta Hamiltonians Hermitian H eetecoc fascalar a of choice the Here . V ( x e = ) H x = 2i c x and H / H .Here 2. ( H x othe to p c that p , c x as c c ) ment on the connection between real spectra and from a unitary U that is defined to be an au- PT symmetric Hamiltonians. tomorphism, e. g. U : L2 L2, that respects the , † → −1 and therefore fulfills U = U .) Given an operator AH that is defined in , there is a corresponding operator II. GENERAL FRAMEWORK H −1 AL2 = T AHT (2.4) Suppose we have a non-Hermitian (in the usual sense of the space L ( , )) Hamiltonian H and some eigen- that is defined in L2 and fulfills 2 −∞ ∞ functions ψn(x) with real eigenvalues En, ′ ′ (ψ, AHψ )H = (Tψ,AL2 T ψ )L2 (2.5)

H(x, p)ψn = Enψn , En real . (2.1) ′ for all ψ and ψ in the domain of definition of AH. ∈ H Here the operators x and p act in the usual way as mul- If one denotes the of an operator AH ‡ tiplication by x and i times differentiation, resp., on with respect to the scalar product in by AH, − H ψn(x), and the eigenfunctions are calculated as solutions ′ ‡ ′ of the corresponding differential equation. They are not (ψ, AHψ )H = (AHψ, ψ )H , (2.6) required to be normalizable with respect to the norm one finds ψ L2 = (ψ, ψ)L2 of the L2( , ). k k ′ ∞ ∗ −∞′ ∞ (We use the notation (ψ, ψ ) = dx ψ (x)ψ (x) for † ‡ p L2 −∞ A = T A T −1 , (2.7) the scalar product of L .) L2 H 2 R Our aim is to set up a formulation in which we can so that, for example, A 2 is L Hermitian if AH is L 2 H interpret the superpositions of the eigenfunctions ψn Hermitian, and vice versa. quantum-mechanically. For simplicity, we will do this The equations (2.4) – (2.7) imply that theories defined only for the case of a discrete, infinite, non-degenerate in and L2 are physically indistinguishable. spectrum of H. H To start, we define the = span ψ ,n = V { n 0,... of the finite superpositions of the eigenfunctions. A. Hilbert spaces with Hermitian Hamiltonian Since} eigenfunctions corresponding to different eigenval- ues are always linearly independent, the dimension of We have, up to this point, not specified the details of is infinite. Next we define a scalar product (., .)V V the scalar product in ; the construction outlined above on this space. Besides requiring the usual properties works with every scalarH product that can be defined on ((ψ,c1ϕ1 + c2ϕ2)V = c1(ψ, ϕ1)V + c2(ψ, ϕ2)V , (ψ, ϕ)V = ∗ 2 the space . But obviously not only the value of ma- (ϕ, ψ) , ψ (ψ, ψ)V > 0 for ψ = 0 and ψ V = 0 V V k kV ≡ 6 k k trix elements, but the whole structure of the space for ψ = 0), we leave this scalar product arbitrary at this depends on this choice. And although any choice allowsH point. The scalar product turns into a separable Eu- V a consistent probabilistic interpretation, we will here con- clidean space, and we can use standard theorems [17] to sider only Hilbert spaces in which the Hamiltonian is complete this space, thereby defining a separable Hilbert Hermitian. The reason is that we consider the system space . As a result, we have and H V ⊂ H described by (2.1) as a closed system. (This appears to ′ ′ ′ be an implicit assumption in most of the models treated (ψ, ψ )H = (ψ, ψ )V for all ψ, ψ . (2.2) ∈V recently.) According to a theorem by Wigner [19] the (Recall that in addition to the finite superpositions of the of the time evolution for such systems is a con- ψ , n =0,..., that make up , the Hilbert space also sequence of fundamental properties of quantum theories n V H contains all limits ψ = limn→∞ fn, fn , of Cauchy as, for example, the linearity of the time-evolution oper- sequences of vectors of .) ∈ V ator and the fact that physical states are described by The Hilbert space Vis the natural choice as the space rays λψ; λ C, ψ rather than vectors ψ. If we { ∈ ∈ H} of states for the systemH described by the Hamiltonian H, want to keep these properties, and if the Hamiltonian H because it allows a consistent probabilistic interpretation is to be interpreted as the generator of time evolution, H of the model in terms of scalar products of states. has to be Hermitian in the underlying Hilbert space . H It is well-known [17] that all infinite-dimensional, sep- In our context, it is the reality of the spectrum that arable Hilbert spaces are unitarily equivalent to the makes the construction of Hilbert spaces with Hermitian Hilbert space L ( , ). This unitary equivalence H possible. Define, for example, 2 −∞ ∞ means that an isomorphism T : L2 exists that re- spects the scalar products in bothH spaces: → (ψn, ψm)H = δnm . (2.8) ′ ′ ′ It is easy to verify that (2.8) is a well-defined scalar prod- (ψ, ψ )H = (Tψ,Tψ )L2 for all ψ, ψ . (2.3) ∈ H uct in . (Here the linear independence of the eigenstates H The transformation T is called a unitary transformation ψn is important.) We have (from now on we denote H from onto L [17]. (This notion has to be distinguished by HH) H 2

2 (ψn,HHψm)H = Enδnm = (HHψn, ψm)H , (2.9) a formulation turns out to be impossible, we argue that the original problem (2.1) is inconsistent. On the other and since the ψn span the whole space, we can conclude hand, if a canonical formulation is possible, then the non- Hermiticity of (2.1) with respect to the space L is only ‡ † 2 HH = HH and HL2 = HL2 , (2.10) a superficial one, due to a choice of variables that is in- adequate for the given physical problem. where we have used (2.7). Thus, the choice (2.8) leads to A necessary condition for the self-adjointness of the op- Hermitian Hamiltonians HH and HL2 in their respective erators xc and pc is their Hermiticity. In physical consid- spaces. erations both notions are usually identified. We shall do Note that there is no problem to explicitly construct the same here, except in cases in which the difference be- a transformation T : L and the Hermitian oper- H → 2 comes essential for the physical conclusions. These cases ator HL2 . It is enough to take an arbitrary complete, will be clearly indicated. orthonormalized set of L2 functions, e. g. the eigenstates Of course, definition (2.8) is not the only scalar ϕn(x) of the harmonic oscillator, and define T in (2.3) as product giving a Hermitian H. The condition that a the linear transformation that fulfills T ψn = ϕn. Then scalar product (., .)′ in the space leads to a Her- −1 V V HL2 = THH(x, p)T has the required properties. The mitian Hamiltonian H follows from the requirement crucial point is that the transformed operators xL2 and ′ ′ (ψn, Hψm)V = (Hψn, ψm)V for all n,m, which gives p will have the same complicated properties concern- ′ L2 (En Em)(ψn, ψm) = 0. Thus all scalar products for ing their Hermitian adjoints as x and p and will therefore − V which the ψn are mutually orthogonal lead to a Hermi- in general be void of any physical significance. One has tian H. Therefore the most general ansatz for a scalar to find a transformation T that leads to a Hamiltonian product with Hermitian H is HL2 = H(xL2 ,pL2 ) that allows a clear physical interpre- ′ tation. (ψn, ψm)V = γnδnm , (2.14) For achieving this, one may employ the uniqueness theorem by von Neumann [17]. It states that all irre- where the γn are some positive constants. With respect c 2 ducible representations of two self-adjoint operators x , to this scalar product, vectors ψˆn = ψn/cn with cn = c | | p that are defined in a separable Hilbert space and fulfill γn are orthonormal. Let’s define the Euclidean space canonical commutation relations are unitarily equivalent. to be the space equipped with the scalar product This means that given two such operators in a separable (2.8),E and let ˆ be theV Euclidean space build from and E V Hilbert space , a unitary transformation T : L2 (2.14). Eventually, and ˆ would be completed to give H c c −1 c H → E E exists such that xL2 = T xHT and pL2 are the two the Hilbert spaces and ˆ, resp. Every linear operator canonical L2 operators H H

A : , Aψn = anmψm , (2.15) c c dϕ V→V (x ϕ)(x)= xϕ(x) , (p ϕ)(x)= i . (2.11) n,m L2 L2 − dx X has a meaning in both spaces , ˆ, but their properties, In this way the uniqueness theorem allows one to find E E a unique (modulo unitary equivalence) interpretation for for example concerning their Hermitian adjoints, will in quantum theories that are defined on separable Hilbert general be different in and ˆ. On the other hand, the Euclidean spaces andE ˆ areE connected by the unitary spaces other than L2. E E Thus our aim is to find a set of canonical operators transformation Tˆ : ˆ, c c E → E (xH,pH) that are self-adjoint in , H ˆ ˆ ψn ˆ ˆ ′ T ψn = ψn = , (ψn, ψm)V = (T ψn, T ψm)V , c ‡ c c ‡ c cn (xH) = xH , (pH) = pH , (2.12) (2.16) and fulfill and therefore A as an operator can be mapped to c c E→E [xH,pH] = i . (2.13) Aˆ = Tˆ ATˆ−1 : ˆ ˆ . (2.17) If we are able to express HH as a function of these two op- E → E c c erators, HH = H(xH,pH), we can immediately transform These operators, A : and Aˆ : ˆ ˆ, share the c c E →E ˆE → E the model into the space L2 with HL2 = H(xL2 ,pL2 ). same properties. The transformation T in (2.16) is only Here the physical meaning is clear, because one can in- a function of H, so that Hˆ = H. But in general, unless c c terpret xL2 as the position-space and pL2 as A commutes with Tˆ, one has Aˆ = A. c 6 the momentum (generator of xL2 translations). This means that to a certain extent one can adjust We emphasize that this canonical formulation HL2 = the properties of an operator A : by choos- c c V →V H(xL2 ,pL2 ) is not just an alternative formulation for the ing an appropriate scalar product. This will turn out original problem. In our opinion it is the only formula- to be helpful for the construction of sets of canonical tion which has a clear physical meaning. Thus, if such operators (Section IIIB). If, for example, an operator

3 −1 ′ F (H) AF (H): is Hermitian in , then A is Her- ′ ∗ dψ (ψ,pHψ )H = i ds [ψ(z(s))] E→Eˆ E ˆ − dz mitian in the space which is defined by ψn = F (H)ψn Z z=z(s) ˆ ˆ ′ E and (ψn, ψm)V = δnm. 1 = ds ϕ∗(s) pc ϕ′ ( s) dz(xc )/dxc L2 Z  L2 L2  B. Scalar products along complex paths 1 c ′ ′ = (ϕ, c pL2 ϕ )L2 (ϕ, pL2 ϕ )L2 , (2.27) dz/dxL2 ≡ In the work of Bender and collaborators [2] the eigen- and consequently functions ψn(z) are defined with z describing a curve in the complex plane which is chosen in such a way that the 1 c pH = c c pH . (2.28) eigenfunctions are decreasing asymptotically. Let a pa- dz(xH)/dxH rameterization for this curve be z = z(s),

4 ‡ Since (V (xH)) does not contain pH, a necessary con- A. The one-dimensional Coulomb problem on the dition for the Hermiticity of H is that the term linear real half axis in pH in (2.36) vanishes. This, together with relation ‡ (2.34) gives 3vw = w . Taking the Hermitian adjoint Let us start with a very simple example. We first dis- of this equation and using the unitarity of v then shows cuss it in a way inspired by Bender et al. [2,16] and then 2 ‡ c 2 that necessarily w = 0. Since w contains d xH/(dxH) , reinterpret it according to the framework of Section II. which is the operator version of d2z∗/ds2, one sees that Consider the one-dimensional Coulomb problem on the only straight lines, real half axis x> 0: The Hamiltonian reads

z(s)= a + bs , a,b C , (2.37) (0) c 2 a ∈ HL2 (a) = (pL2 ) c , (3.1) ∗ − xL2 are allowed. Then v is a constant phase, v(xH)= b/b = 2i arg b ‡ 2 4i arg b 2 2 e and (pH) =e pH. Hence pH is only Hermi- the eigenfunctions for real x> 0 and a> 0 that are finite tian if the parameter b satisfies at the origin, have a finite slope there, and are decreasing π for x are [20] arg b = n . (2.38) → ∞ 2 ax Φ (x; a)=e−ax/(2n) L(1) (ax/n) , (3.2) Under this condition, one has n n n−1 ‡ 2 ‡ H = pH + (V (xH)) , (2.39) (1) n = 1,..., with the Laguerre polynomials Ln−1. The and it may happen that a and b can be chosen in such a eigenvalues are way that also the potential becomes| | Hermitian. But the a 2 fact that only straight lines in the complex plane can lead E(0)(a)= . (3.3) to Hermitian Hamiltonians of the form (2.35) seems to n − 2n restrict the usefulness of scalar products along complex   It is clear that Φ (x; a) is normalizable in L (0, ) for paths severely. n 2 real and positive a; the spectrum in this case is a bound-∞ In the simple case of straight lines one can explicitly state spectrum. construct the transformation T : L (cf. (2.19)): it 2 We may now use the solutions (3.2) with a = iα, α real is given by a combination H → and positive, 1 T = Tt(a/b)Td(ln b) (2.40) iαx √b ψ (x)=Φ (x; iα)=e−iαx/(2n) L(1) (iαx/n) . n n n n−1 of a (complex) translation (3.4) Tt(a/b) = exp i(a/b)pH , (Tt(a/b)ψ)(s)= ψ(s + a/b) , The energies are then still real, (2.41) α 2 and a (complex) dilatation E = > 0 , (3.5) n 2n i   Td(ln b) = exp (xHpH + pHxH) ln b , (2.42) and the Hamiltonian is 2   iα (Td(ln b)ψ)(s)= √b ψ(bs) . (2.43) H = p2 , (3.6) − x

thus non-Hermitian with respect to L2. The asymptotic iϑ III. DISCUSSION OF VARIOUS MODELS behavior of ψn(x) for complex x = x e and large x is given by | | | | Unfortunately, most works dealing with non-Hermitian α x Hamiltonians are based on numerical methods which are ψ (x) exp | | sin ϑ , (3.7) | n |∼ 2n not very well adapted for studies of the structure of the   space . But note that an investigation in this direction H hence ψn converges most rapidly on the anti-Stokes line has been started in [12]. We will here consider exactly given by ϑ = 3π/2. The eigenfunctions (3.4) have no solvable models. Among those works that also treat ex- finite norm in L (0, ). One would therefore consider actly solvable models, we mention [13], where a similar 2 ∞ the ψn(x) on the anti-Stokes line construction of a Hilbert space as the one outlined in Sec- tion II has been recently applied to the study of certain x(s)=eiϑs , 0 s< . (3.8) coherent states. ≤ ∞ In the following we will throughout suppress the sub- In terms of our general discussion in Section II B, script for operators that act in the space ; for oper- we would define the scalar product in the space by H H H ators acting in L2 we keep the subscript L2. (cf. (2.18))

5 ∞ the lower-half x plane into vertical strips separated by ′ iϑ ∗ ′ iϑ (ψ, ψ )H = ds [ψ(e s)] ψ (e s) . (3.9) the lines Re x = nπ with integer n, and if one assumes Z0 that the curves along which the wave functions are de- fined have two asymptotes that are vertical lines going This corresponds to the special case (2.37) with a = 0 down, then the following configurations occur: If a curve iϑ and b =e . The canonical operators are given by (2.26) starts and ends in the same strip, no quantization con- and (2.28): dition appears, the spectrum is again unbounded from

c −iϑ c iϑ below. If the curve starts in one strip and ends in the x =e x , p =e p , (3.10) next strip, then no normalizable solutions exist. If, how- ever, the curve starts in one strip and ends in the next- therefore the canonical form of the Hamiltonian (3.6) is to-next strip (e. g. one asymptote of the curve is the anti- α α Stokes line at Re x = π/2 and the other asymptote is at H(xc,pc)=e−2iϑ(pc)2 ie−iϑ = (pc)2 + (3.11) − xc − xc Re x =5π/2), the quantization condition ν = n +1/2, n integer, appears. One then has the solution and is obviously Hermitian. Its L2 realization via T : iϑ 2 L2, ϕ(s) = (T ψ)(s)= ψ(e s), is (1) ix (n +1/2) H → ψn(x)= H (e ) , En = . (3.16) n+1/2 2 α H(xc ,pc )= (pc )2 + = H(0)(α) , (3.12) L2 L2 L2 c L2 In addition to this case, on which our subsequent discus- − xL2 − sion concentrates, there exist other curves with similar and the transformed eigenfunctions read spectra [18]. It follows from (3.16) that ϕn(s)=Φn(s; α) . (3.13) −n−1 ψ = i( 1) ψ− − . (3.17) n − n 1 In summary, the quantum-mechanical interpretation This means that the ψ with n 0 constitute a set of based on the canonical form (3.11) is completely equiv- n ≥ alent to the original, Hermitian problem; the change of linearly independent eigenstates, so that one can start from the space = span ψ , ψ ,... to construct . the sign of the spectrum (3.5) is reflected by the sign V { 0 1 } H The recursion relations zdZ /dz νZ = zZ ∓ for in (3.12). Hence the solutions (3.4) do not contain any ν ± ν ± ν 1 new information. Moreover, the kinetic energy in (3.11) Bessel functions can be used to find has—contrary to (3.6)—the wrong sign. This makes the 1 e−ixψ = (ψ + ψ ) , (3.18) model unphysical. n 2n +1 n−1 n+1 1 e−ixpψ = (ψ ψ ) . (3.19) n 2 n−1 n+1 B. The model of Cannata et al. − Thus the two operators e−ix and e−ixp are well-defined In the model introduced by Cannata, Junker and Trost operators . The two relations (3.18) and (3.19) H → H [18] the Hamiltonian is lead to the version −2ix 2 2 1 e (n +1/2) p ψn = ψn (3.20) H = (p2 +e2ix) (3.14) − 2 of the Schr¨odinger equation Hψ n = Enψn. (Note that (a similar model has been considered in [21]). The Schr¨o- these are algebraic relations that hold for all Bessel func- dinger equation for this Hamiltonian can be reduced to tions and all n C. It is (3.17) and the definition of ∈ V Bessel’s differential equation, so that its general solution that are specific to the solution (3.16).) is given by We first consider the question whether one can define a scalar product along a curve such that the Hamiltonian ν2 (3.14) becomes Hermitian. Since (3.14) is of the form Ψ (x)= c H(1)(eix)+ c H(2)(eix) , E = . ν 1 ν 2 ν ν 2 (2.35), it is clear that only straight lines x = a + bxc c (3.15) are possible. Then p = bp, and the Hamiltonian can be written as c For real x no normalizable solutions exist. But if x is (pc)2 e2i(a+bx ) chosen along curves in the complex plane, different pos- H(xc,pc)= + . (3.21) 2b2 2 sibilities occur. We briefly summarize the main cases; details can be found in [18]. If one wants this to be Hermitian, only lines with Re a = In the upper-half x plane, one can find normalizable nπ/2 and Re b = 0 are allowed. Although such lines may solutions for every real value of the energy, i. e. no quan- be the asymptotes of curves that lead to the solution tization condition appears and the spectrum has no lower (3.16), one cannot get a connected curve by joining these bound so that the system is unphysical. If one divides lines.

6 Therefore, we now return to the general discussion of These equations show Section II A and define the Hilbert space by imposing H ˆ −ix ˆ −ix ˆ ˆ the orthogonality condition (ψn, ψm)H = δnm for n,m (ψn, e ψm)Hˆ = (e ψn, ψm)Hˆ (3.28) 0, see (2.8). The equations (3.18) and (3.19) can then≥ −ix −ix for all n,m 0 except for n = m = 0, i. e. the restriction be used to calculate the adjoints of e and e p; the ≥ results are −ix e : ˆ1 ˆ (3.29) Vˆ1 V → H ix‡ 1 1 e ψn = ψn−1 + ψn+1 (3.22) 2(n 1)+1 2(n +1)+1 of e−ix to the space ˆ = span ψˆ ,... is Hermitian. If − 1 1 we define V { } and ψˆ ψˆ − ‡ − n−1 n+1 e ixp = e ixp . (3.23) ξn = + for n 1 , − n 1+1/2 n +1+1/2 ≥ − c  −ix It follows thatp ˜ = ie p is Hermitian. Furthermore, p p (3.30) [x, p] = i leads to [eix, p˜c]= i. Hence the two operators − −ix then the image of ˆ1 under e is given by ˆξ = 1 ‡ V V x˜c = eix +e−ix , p˜c = ie−ixp (3.24) span ξ1,... . Consider the series −2 { }   N c c are Hermitian and fulfill [˜x , p˜ ] = i so that they form a f (n) = ( 1)kξ (3.31) N − n+2k+1 canonical set. k=0 But it seems that one cannot write the Hamiltonian X c c ψˆ ψˆ (3.14) only as a function ofx ˜ andp ˜ . One can readily = n + ( 1)N n+2(N+1) . (3.32) express H as H = H(˜xc, p˜c, x˜c′), wherex ˜c′ = i(eix n +1/2 − [n + 2(N +1)]+1/2 ‡ − e−ix )/2 is another Hermitian operator. This operator It fulfillsp p fulfills [˜xc′, p˜c] = 0. This may mean thatx ˜c′ is only a c function ofp ˜ , but we were not able to show this. (n) ˆ fN ψn/ n +1/2 (3.33) One therefore has to find other canonical variables. We − Hˆ emphasize that for the following discussion we use the p1 = 0 for N term Hermitian in the precise sense of symmetry, i. e. an → → ∞ ′ ′ [n + 2(N +1)]+1/2 operator A is called Hermitian if (ψ, Aψ )H = (Aψ, ψ )H ′ for all ψ, ψ in the domain of definition of A. and thereforep converges in ˆ, i. e. We can use the method described at the end of Sec- H ∞ tion II A to try to find a new scalar product that makes H ψˆ − n k and e ix simultaneously Hermitian. As such a new scalar = ( 1) ξn+2k+1 for n 0 , (3.34) − ≥ ′ n +1/2 k=0 product is connected to the old one by (ψn, ψm)H = −ix X γn(ψn, ψm)H, γn > 0 (see (2.14)), a Hermitian e would p −ix which means that the linearly independent (although require real matrix elements (ψn, e ψn)H for all n. But −ix non-orthogonal) vectors ξn, n 1, lie dense in the space since ψ−1 = iψ0, one obtains (ψ0, e ψ0)H = i, so that ≥ −ix ˆ. We now define the operator one cannot construct a scalar product for which e is H Hermitian in the entire space . The best one can do −1 −ix H c −ix is to make e Hermitian in the subspace spanned by xˆ = e : ˆξ ˆ1 . (3.35) Vˆ1 V → V ψ , ψ ,... . This is achieved by defining { 1 2 }   This operator acts as 4 4 ψˆn = √H ψn = n +1/2 ψn/√2 for n 0 , (3.25) ≥ c xˆ ξn =2 n +1/2 ψˆn , (3.36) and introducing ap new Hilbert space ˆ with the new H p scalar product (ψˆ , ψˆ ) = δ for n,m 0. Then has a dense domain of definition in ˆ and is Hermitian: n m Hˆ nm ≥ H c c ˆ ˆ (ξn, xˆ ξm)Hˆ = 2(δn,m+1 + δn+1,m)=(ˆx ξn, ξm)Hˆ (3.37) −ix 1 ψn−1 ψn+1 e ψˆn = + 2 n +1/2 n 1+1/2 n +1+1/2! for all n,m 1. We usex ˆc as one of our canonical − variables. Its≥ conjugate momentum can be defined by p p p (3.26)

c i −ix −ix for n 1 and pˆ = (e p + p e ) . (3.38) ≥ −2 ˆ c −ix ˆ ˆ ψ1 The operatorx ˆ has to be distinguished from the op- e ψ0 = iψ0 + . (3.27) ix c √3 erator e . Furthermore, althoughx ˆ is well-defined in

7 ˆ the space ξ by (3.36), in acting on the eigenfunctions The general solution of the Schr¨odinger equation HL2 Φν V 2 ψˆn via (3.34) it leads to the divergent series = (ν +1/2) Φν /2 is given by [22] Φν (x) = c1ϕν (x)+ c2ϕ−ν−1(x) with xˆcψˆ =2 n +1/2 (3.39) n 1 ∞ ϕ (x)= J (x) , ν C . (3.46) p ν ν+1/2 ( 1)k n +2k +1+1/2 ψˆ √x ∈ × − n+2k+1 k=0 X p For the case corresponding to the spectrum (3.16), namely integer ν = n, these solutions are proportional to so thatx ˆcψˆ cannot be defined by this series. Because of n spherical Bessel functions c j (x)+ c j (x). In fact, this, the construction of (3.38) and the demonstration of 1 n 2 −n−1 the equations (3.26), (3.27) and (3.40) are closely related the commutation relation [ˆxc, pˆc] = i is a rather techni- to the recursion relations for spherical Bessel functions cal problem. The corresponding proofs are described in [23]. detail in Appendix A. Here we just note that The solutions (3.46) fulfill [22]

ˆ ˆ ∗ c i nψn−1 (n + 1)ψn+1 −iπ(ν −µ)/2 ∗ pˆ ψˆn = − + 2e sin π(ν µ) (ϕ , ϕ ) 2 = − (3.47) 2 n +1/2 n 1+1/2 n +1+1/2! ν µ L ν∗ + µ +1 π(ν∗ µ) − − p p p (3.40) ∗ for Re(ν + µ) > 1. Here the value of ϕν ( x), x > 0, has been chosen above− the cut, i. e. x =− eiπx. This for all n 0, and that indeed − ≥ result has quite interesting consequences: First of all, for c c integer n 0 only ϕn, but not ϕ−n−1 is L2 normalizable. (ξn, pˆ ξm) ˆ = (ˆp ξn, ξm) ˆ for all n,m 1 (3.41) H H ≥ Furthermore,≥ the functions and 2n +1 ϕˆ (x)= n +1/2 ϕ (x)= j (x) , n 0 , [ˆxc, pˆc]ξ = iξ for all n 1 . (3.42) n n π n ≥ n n ≥ r p (3.48) Now we are finally in the position to write down the canonical form of the Hamiltonian of the model of Can- are orthonormal in L2: nata et al.: From (A10) we have p = iˆxcpˆc +1/2, and ix c as explained in Appendix A, one can replace e byx ˆ in (ˆϕn, ϕˆm)L2 = δnm , (3.49) the Hamiltonian (3.14), so that H can be written as in other words, the linear transformation T : ˆ L2 1 1 ˆ H → H = H(ˆxc, pˆc)= (ˆxc)2(ˆpc)2 + 2iˆxcpˆc + +(ˆxc)2 . defined by T ψn =ϕ ˆn is an isometry. But it is not a 2 − 4 not   unitary transformation, because theϕ ˆn do span the (3.43) whole space L2. This can be seen most easily by consid- ering

One thing should be noted here: Although (3.14) mul- ∞ ∞ ˆ ˆ ∗ 2n +1 tiplied with ψn is of course well-defined in , the two D(x, y)= ϕˆn(x)ˆϕn(y)= jn(x)jn(y) . 2 2ix H π terms p ψˆ and e ψˆ are divergent; their divergencies n=0 n=0 n n X X cancel in the sum. With respect to (3.43) this means that (3.50) one can group the terms on the right-hand side in such a way that no divergencies appear in intermediate results. Completeness of theϕ ˆn would require D(x, y)= δ(x y), This is achieved by writing (see (A13)) but instead one obtains [23] − 1 1 1 H = xˆc xˆc 1 (ˆpc)2 + 2iˆpc + . (3.44) D(x, x)= . (3.51) 2 − 4 π      This form of the Hamiltonian is an operator that maps Even worse, the formally L2 Hermitian Hamiltonian ˆ = span ψˆ , ψˆ ,... to itself. In this sense, (3.43) is (3.45) is not even Hermitian: It follows from (3.47) that V { 0 1 } equivalent to (3.14) in the entire space ˆ. −π Im ν V 2 e sinh 2π Im ν 1 Let’s consider the realization of this Hamiltonian in ϕν L2 = for Re ν > , c c c k k Re ν +1/2 2π Im ν −2 L2( , ), i. e. HL2 = H(xL2 ,pL2 ), in which xL2 and c −∞ ∞ pL2 are identified with the canonical L2 operators: (3.52)

2 1 2 d d 1 2 thus HL2 has eigenfunctions with complex eigenvalues HL2 = x +2x + + x . (3.45) 2 C 2 dx2 dx 4 Eν = (ν +1/2) /2 in L2, so that (ϕν ,HL2 ϕν )L2 =   E ϕ 2 = E∗ ϕ ∈2 = (H ϕ , ϕ ) . ν k ν kL2 6 ν k ν kL2 L2 ν ν L2

8 Our initial aim was to apply the uniqueness theorem C. The model of Hatano and Nelson to the canonical formulation (3.43). But we have not c c checked whether the canonical operatorsx ˆ andp ˆ are In the context of studies of delocalization phenomena, self-adjoint (as required by the uniqueness theorem). Al- the model of Hatano and Nelson [1] has attracted a lot though the Hermiticity and dense domain of definition of of interest recently [24,25]. It is defined in one dimension these operators are a necessary condition for this [17], it by the non-Hermitian Hamiltonian is in general not an easy task to check the self-adjointness. In our case, a simple argument can be given that shows (p + ig)2 ˆ H(g)= + V (x) , (3.59) that no unitary transformation T : L2 that maps 2m xˆc andp ˆc to the canonical position andH → momentum op- c c −1 c c −1 where g is a real parameter connected to an externally erators of the space L2, xL2 = T xˆ T , pL2 = T pˆ T , exists: The operatorp ˆc is bounded, from (3.40) one finds applied magnetic field, and V (x) is a random potential. It has been demonstrated numerically that at a certain 3 critical value g = gc a localized wave function turns into pˆcψ ψ for all ψ ˆ . (3.53) k kHˆ ≤ 5 k kHˆ ∈ H a delocalized one, and it has been suggested that this r behaviour signals the occurrence of a delocalization phase Under a unitary transformation T the boundedness prop- transition. erties do not change, hence the bounded operatorp ˆc can- The numerical demonstration of a delocalization tran- c not be mapped to the pL2 . sition at the critical value gc is based on the use of a mod- This argument can even be generalized: The operator ified scalar product in the space of quantum-mechanical pˆc in (3.38) is not the only momentum operator that can states [1]. There has been some controversy about this be assigned tox ˆc, because if f(ˆxc) is a Hermitian function point [26], but at present a consensus seems to have been ofx ˆc, then reached that the delocalization transition is only visible if the density distribution of particles is calculated ac- πc =p ˆc + f(ˆxc) (3.54) cording to a scalar product based on the product of left- and right-eigenfunctions of the Hamiltonian (3.59) [27]. c c is also a Hermitian operator with [ˆx , π ] = i. Now sup- We want to point out here that a physical interpreta- c c pose that this set (ˆx , π ) can be mapped to the canoni- tion that is based on this modified scalar product is only cal position and momentum operators of L2 by a unitary consistent if the underlying Hilbert space is chosen ac- transformation T . Then for ϕ L , Eq. (3.54) gives H ∈ 2 cordingly. In fact, as we will show, this Hilbert space is just of the kind considered in Section II, i. e. below g dϕ c c c −1 the Hamiltonian (3.59) is Hermitian in . But it turns (ˆpL2 ϕ)(x) = (T pˆ T ϕ)(x)= i f(x)ϕ(x) (3.55) − dx − out that under reasonable conditions thisH Hilbert space d = ieiF (x) e−iF (x)ϕ(x) (3.56) explicitly depends on the parameter g, so that the g de- − dx pendence of the Hamiltonian (3.59) is not the complete c −1   = (UpL2 U ϕ)(x) , (3.57) dependence of the model of Hatano and Nelson on g. It is the aim of this Section to investigate this depen- where dF/dx = f(x), and where the U : dence. This is important for the phenomenological im- c L2 L2 is the gauge transformation U = exp iF (x 2 ). plications of the model of Hatano and Nelson, because → L Thus the canonical L2 momentum would be given by g is considered to be an external parameter that can be

c −1 c varied. pL2 = U pˆL2 U. (3.58) We want to emphasize that we are dealing here only with the quantum-mechanical model of [1], not with its Here all operators on the right-hand side are bounded, c applications to statistical-mechanical problems like vor- in contradistinction to the fact that pL2 is not bounded. c tex depinning in type-II superconductors or population This shows that ifp ˆ is bounded, then also all operators biology [24]. πc of the form (3.54) cannot be unitarily equivalent to Consider a set of eigenfunctions ψn(x; g) of (3.59), the canonical L2 momentum. Physically this means that one cannot interpretx ˆc and H(g)ψn(x; g)= En(g)ψn(x; g) , (3.60) pˆc or πc as position and momentum operators, resp. Thus our conclusion is that the physical interpretation of the determined as solutions of the corresponding differential model of Cannata et al. is still unclear. Of course, it may equation and being elements of a Hilbert space 0 (e. g. be possible that other, self-adjoint canonical operators the space L ( , ) or the space of periodic functions).H 2 −∞ ∞ can be found. As we have already pointed out, such In [1], examples are given in which the eigenvalues En(g) a formulation is, in our opinion, necessary for the very are real for g being smaller than some critical value gc, definition of the model. for g>gc they become complex. Let us first recall the construction of the modified scalar product as introduced by Hatano and Nelson. We

9 refer here to the detailed description given in [27], ac- This definition is always possible; it does not rely on cording to which the ψn can be chosen such that additional assumptions like (3.63). Condition (3.63) only has to do with the explicit form of the metric operator ∗ Mg relating the scalar products in and 0. dx ψn(x; g) ψm(x; g) = (ψn( g), ψm(g))H0 (3.61) − − We can now formulate the modelH of HatanoH and Nelson Z entirely in the space , and since the spectrum is real is always well-defined and given by H for g

(ψn(g), ψm(g))H = δnm . (3.67) Mg that is independent of g, since M−gpMg = p in this case. The second solution (3.74), −

10 M−gpMg = p + 2ig , (3.75) don’t know them, we have to rely on some assumptions, and here we will restrict our discussion to a physically requires a g dependent Mg. But the first solution is rather uninteresting case, namely the case of wave func- physically unacceptable. The reason is that (as H(g) tions that are defined with infinite, non-periodic bound- is Hermitian) we can assume that the time evolution is ary conditions. It is well-known [1] that under these con- governed by the Schr¨odinger equation, so that the expec- ditions there is no g dependence at all. The numerical tation value x (t) = (ψ(t), xψ(t))H evolves according to demonstrations of delocalization transitions have always h i been carried out for wave functions with periodic bound- d x (t) = i [H(g), x] (t)= v (t) (3.76) ary conditions. Still, the example of infinite, non-periodic dth i h i h i boundary conditions serves illustrative purposes. At the end, we will comment on the case of periodic boundary with v = (p + ig)/m. Hence p + ig pkin is the kinetic conditions. momentum. (This observable has also≡ been considered in The point is that for non-periodic boundary conditions [1].) The first solution (3.73) then corresponds to an anti- the unitary transformation T : L is known to be Hermitian velocity operator, v‡ = v, giving a purely g 2 only a function of x = xc [1]: H → imaginary expectation value for the− velocity. Only the second solution (3.74) leads to a real expectation value c Tg = Tg(x ) . (3.79) for the velocity. Therefore, we only consider (3.74) to c be physically meaningful, and consequently Mg has to Such a T acts on states ψ simply as (T (x )ψ)(x)= g ∈ H g depend on g. Tg(x)ψ(x), so that We now want to discuss how the quantum-mechanical predictions of the model vary with g. The situation in ′ ′ 2 ∗ ′ (ψ, ψ )H = (Tgψ,Tgψ )L2 = dx Tg(x) ψ (x)ψ (x) . is similar to the situation in classical | | Z Hamiltonian mechanics: There one has two canonically (3.80) conjugate variables xc and pc which are the independent variables, and the Hamiltonian has to be expressed in From this equation one can deduce how p‡ acts on ‡ ′ terms of these two variables: states of : A simple calculation based on (p ψ, ψ )H = ′ H (ψ,pψ )H and partial integration shows that H = H(g; xc,pc) . (3.77) d p‡ = p i ln(T T ‡) . (3.81) Varying any parameter g appearing in the Hamiltonian, − dxc g g the system reacts in precisely the way given by the g dependence of the formulation (3.77). In the quantum- Thus, from (3.78) we obtain mechanical case, these two canonically conjugate vari- i d ables xc and pc have to be two operators fulfilling canon- c ‡ p = p + c ln(TgTg ) . (3.82) ical commutation relations [xc,pc] = i and have to be 2 dx self-adjoint. By this last requirement, the structure of Substituting this into (3.59) leads to the expression the underlying Hilbert space comes into play. Having such a formulation, we can use the uniqueness theorem to H(g)= H(g; xc,pc) c c map H(g; x ,p ) uniquely (modulo unitary equivalence) 2 c c 1 c i d ‡ c to the L2 formulation HL2 = H(g; xL2 ,pL2 ). Thus, the = p + ln(TgTg ) + ig + V (x ) . (3.83) c c 2m 2 dxc g dependence of H(g; xL2 ,pL2 ) describes the complete g   dependence of the physical system. Equation (3.83) represents the canonical form of the Applying this reasoning to the model of Hatano and Hamiltonian of the model of Hatano and Nelson for in- Nelson, one has to find the two canonically conjugate finite, non-periodic boundary conditions. As before, in operators xc and pc. Again, we assume that x is Hermi- c this form the model can immediately be formulated in tian in , whereas p is non-Hermitian, see (3.74). But c c c c H the space L2 by just replacing x , p by xL2 , pL2 , resp. as [x, p] = i, one finds that one may choose As one can see, we have canonically coupled a nontrivial, 1 purely imaginary field xc = x and pc = (p + p‡) (3.78) 2 i d 2 A(x; g)= ln Tg(x) + ig (3.84) as the set of Hermitian operators with canonical commu- 2 dx | | tation relations. (For ggc. This means that H(g; x ,p ) is Hermitian through xc and pc. In order to do this, one has to know below gc and non-Hermitian above gc (see (3.68)). The how p‡ acts in for arbitrary g. Again, this would be H Hermiticity translates into a vanishing anti-commutator clear once we knew the states ψn explicitly. But since we pc, A(xc; g) = 0 for g

11 id/dx, A(x; g) ϕ(x) = 0 for all ϕ L , (3.85) spectrum of H(g) is real, the non-Hermiticity of H(g) {− } ∈ 2 with respect to the space 0 is a rather superficial one leads to the equation ϕ(dA/dx)= 2A(dϕ/dx) which— that can be avoided by consideringH the space . Thus − H as it has to hold for all ϕ L2—requires the time evolution can consistently be described by the ∈ Schr¨odinger equation, leading to a unitary time-evolution A(x; g)=0 (for gg the spectrum becomes complex, and A(x; g) = c 6 struct a Hilbert space in which the Hamiltonian is Her- 0 is possible. Hence one can write explicitly mitian and which at the same time contains eigenstates with complex eigenvalues. Although a completely con- c c 1 c c 2 c H(g; x ,p )= [p + Θ(g gc)A(x ; g)] + V (x ) . sistent probabilistic interpretation is still possible, the 2m − interpretation of the Hamiltonian as the generator of the (3.87) time evolution would then lead to a non-unitary time- evolution operator with all the problems mentioned in In fact, for infinite, non-periodic boundary conditions Section II A. it is well-known [1] that Tg is given by

T (xc) = exp( gxc) , (3.88) g − IV. CONCLUSIONS leading to In this work we have studied the quantum-mechanical A(x; g) = 0 for all g , (3.89) interpretation of models with non-Hermitian Hamiltoni- ans (in the usual sense of the space L2) and real spectra. i. e. gc = in (3.87), and the spectrum is real for all g. Assuming that the systems under consideration are ∞ The transformation Tg in (3.88) is often refered to as an closed, and that the Hamiltonians are the generators of “imaginary gauge transformation”. Our analysis shows the time evolution for these systems, we construct sepa- that this notion is quite misleading: Nothing has been rable Hilbert spaces in which the Hamiltonians are Her- gauged away by Tg; we just have two formulations of mitian. the same physics, realized in the two different, though Within this construction, we set up a canonical for- unitarily equivalent Hilbert spaces and L2. mulation in which the Hamiltonian is only a function of H The problem in the case of Hilbert spaces 0 and two Hermitian operators xc and pc that fulfill canonical H H that contain wave functions with periodic boundary con- commutation relations. Then a unique physical inter- ditions is that in these spaces only the differentiation pretation is obtained if these operators are self-adjoint, operator p is defined; the operator x has a vanishing do- because in this case the model can be formulated as a main of definition. This makes it impossible to construct Hermitian problem in the space L2. If, on the other a canonical formulation for which the uniqueness theo- hand, this equivalent L2 description cannot be achieved, rem may be used. Still, if one chooses a periodic poten- we consider the model to be quantum-mechanically in- tial in (3.59), our reasoning applies up to Eq. (3.75); only consistent. the argument given in (3.76) can—strictly speaking—no We apply the above construction to a number of mod- longer be used. This means that also in the case of peri- els recently discussed in the literature. odic boundary conditions it may happen that the space As a first example, we have analysed a simple non- is an unknown function of g, making it a priori impos- Hermitian model that is based on the Hermitian one- H sible to compare the predictions of the model for different dimensional Coulomb problem on the real half axis. Mo- values of g. tivated by the approach of Bender et al. [2], we extend To our knowledge, this aspect has not been discussed the eigenfunctions to complex values of the coordinate. in the literature up to now, but it is in our opinion of We then show that the canonical formulation of this ex- fundamental importance for the phenomenological inter- ample does not contain any new information compared pretation of the model of Hatano and Nelson. We see, for to the original, Hermitian version of the model. example, that the canonical formulation (3.87) describes The next example is the model introduced by Cannata a non-continuously differentiable change of the Hamil- et al. [18]. Although here we also find a canonical for- tonian at the transition point gc. Whether such a be- mulation, it turns out that no equivalent L2 formulation haviour should be called a phase transition is not clear exists. In our opinion this may indicate that the model to us; it seems to be more similar to the sudden switching- is intrinsically inconsistent. on of a magnetic field as the parameter g passes gc. It is Our last example is the phenomenologically important clear that the behaviour of the system will then change, model of Hatano and Nelson [1]. Here the appropriate localized states may become delocalized, but one would Hilbert space is already determined by the introduction hardly call this a phase transition. of the modified scalar product in [1]. Under these condi- Let us mention another aspect of the model of Hatano tions, the dependence of the Hamiltonian on the external and Nelson. We have seen in (3.68) that as long as the

12 parameter g that is responsible for the non-Hermiticity Quantum Mechanics of Atomic Spectra, Academic Press, of the Hamiltonian does not fully describe how the sys- New York, 1959. tem varies with g. For the special case of wave functions [20] R. Loudon, Am. J. Phys. 27, 649 (1959). with infinite, non-periodic boundary conditions, we give [21] C. M. Bender, S. Boettcher, H. F. Jones, V. M. Savage, the canonical L2 formulation which clearly exhibits this J. Phys. A 32, 6771 (1999). complete g dependence. It turns out that at the critical [22] I. S. Gradstein, I. M. Ryshik, Tables of Series, Products, and Integrals, Verlag Harri Deutsch, Thun, Frankfurt, value g = gc, where the states of the system undergo 1981. a delocalization transition, also the L2 version of the Hamiltonian changes in a non-continuously differentiable [23] M. Abramowitz, I. A. Stegun (eds.), Handbook of Mathe- way from Hermitian to non-Hermitian. In our opinion, matical Functions, Dover Publications, New York, 1965. 58 this makes the usual interpretation of this transition as [24] D. R. Nelson, N. M. Shnerb, Phys. Rev. E , 1383 (1998). a phase transition questionable. We argue that in the [25] J. Feinberg, A. Zee, Phys. Rev. E 59, 6433 (1999). phenomenologically relevant case with periodic bound- [26] P. G. Silvestrov, Phys. Rev. B 58, R10111 (1998). ary conditions similar effects may also occur. [27] N. Hatano, D. R. Nelson, Phys. Rev. B 58, 8384 (1998).

ACKNOWLEDGMENT APPENDIX A: PROPERTIES OF THE L. S. acknowledges support of Deutsche Forschungs- CANONICAL VARIABLES FOR THE MODEL OF gemeinschaft and the warm hospitality at Regensburg CANNATA ET AL. University. In (3.35) the operatorx ˆc is defined on the subset ξ1, ξ2,... which is dense in ˆ. Eq. (3.37) shows that {xˆc is Hermitian.} We can thereforeH use the fact that every densely defined, Hermitian operator has a closed exten- sion [17], to extend the domain of definition ofx ˆc to some larger space than ˆ . The problem is thatx ˆcψˆ is not [1] N. Hatano, D. R. Nelson, Phys. Rev. Lett. 77, 570 (1996); Vξ n N. Hatano, D. R. Nelson, Phys. Rev. B 56, 8651 (1997). defined by this procedure (cf. (3.39)). We will therefore ˆ 80 in the following define every operator on the space ξ, [2] C. M. Bender, S. Boettcher, Phys. Rev. Lett. , 5243 V (1998); check its algebraic properties there, and will then implic- C. M. Bender, S. Boettcher, P. Meisinger, J. Math. Phys. itly assume that we consider closed extensions where this 40, 2201 (1999). is possible. [3] M. Znojil, Phys. Lett. A 259, 220 (1999). To start, we rewrite (3.26) and (3.27) with the help of [4] B. Bagchi, R. Roychoudhury, J. Phys. A 33, L1 (2000). (3.30) and (3.34) in the form [5] A. A. Andrianov, F. Cannata, J.-P. Dedonder, M. V. 14 ξn−1 ξn+1 Ioffe, Int. J. Mod. Phys. A , 2675 (1999). e−ixξ = + (A1) [6] F. M. Fern´andez, R. Guardiola, J. Ros, M. Znojil, J. n 2(n 1+1/2) 2(n +1+1/2) Phys. A 32, 3105 (1999). − quant-ph/9912079 −ix ∞ k [7] M. Znojil, . for n 2, and e ξ1 = ξ2/5 k=1( i) ξk. It follows [8] P. Roy, R. Roychoudhury, quant-ph/0004034. that although≥ − − [9] C. M. Bender, S. Boettcher, H. F. Jones, P. N. Meisinger, P −ix c hep-th/0007248. e xˆ ξn = ξn for all n 1 , (A2) [10] B. Basu-Mallick, Bhabani Prasad Mandal, cond-mat/ ≥ 0101349. one obtains [11] P. Dorey, C. Dunning, R. Tateo, hep-th/0104119. c −ix [12] C. M. Bender, S. Boettcher, V. M. Savage, J. Math. Phys. xˆ e ξn = ξn only for n 2 , (A3) 41, 6381 (2000). ≥ c −ix [13] J. Beckers, J. F. Cari˜nena, N. Debergh, G. Marmo, Mod. whereasx ˆ e ξ1 is divergent. Phys. Lett. A 16, 91 (2001). In order to construct the canonically conjugate mo- [14] M. Znojil, math-ph/0104012. mentum operatorp ˆc, we first note that (3.19) leads to [15] G. S. Japaridze, quant-ph/0104077. [16] C. M. Bender, A. Turbiner, Phys. Lett. A 173, 442 ˆ ˆ −ix ˆ n +1/2 ψn−1 ψn+1 (1993). e pψn = 2 n 1+1/2 − n +1+1/2! [17] E. Prugoveˇcki, Quantum Mechanics in Hilbert Space, p − Academic Press, New York, 1981. p p (A4) [18] F. Cannata, G. Junker, J. Trost, Phys. Lett. A 246, 219 (1998). −ix for n 1 and e pψˆ0 = iψˆ0/2 ψˆ1/(2√3). Then we [19] E. P. Wigner, Group Theory and its Application to the use p e≥−ix e−ixp = e−ix together− with (A2) to get − −

13 p = e−ixpxˆc 1. Combining this with (A4) gives for all Applying ξ to it, one should note that, for example, − n n 1 (ˆxc)2ξ =2 n +1/2ˆxcψˆ is not defined. But the terms ≥ n n on the right-hand side of (A12) can be arranged in the ˆ ψn+1 following way:p pξn = (n 1+1/2)ξn (2n + 1) . − − n +1+1/2 1 1 H = xˆc xˆc 1 (ˆpc)2 + 2iˆpc + . (A13) p (A5) 2 − 4   This equation shows that p is anti-Hermitian:    Then all intermediate results are well-defined in ˆ. One H (ξ , pξ ) = δ δ = (pξ , ξ ) (A6) obtains n m Hˆ n+2,m − n,m+2 − n m Hˆ c c for all n,m 1. Now define H(ˆx , pˆ )ξn (A14) ≥ (n 1+1/2)3/2 (n +1+1/2)3/2 i = − ψˆ − + ψˆ pˆc = (e−ixp + p e−ix)= ie−ix(p 1/2) . (A7) 2 n 1 2 n+1 −2 − − for n 2. We emphasize that this result cannot be used This operator fulfills ≥ ˆ (n) to check the eigenvalue equation for ψn = limN→∞ fN (cf. (3.31)): The Hamiltonian is an unbounded operator, c i n 1 n +2 pˆ ξ = − ξ − + ξ n 2 −n 1+1/2 n 1 n +1+1/2 n+1 hence one cannot expect that  −  (A8) H lim f (n) = lim Hf (n) . (A15) N→∞ N N→∞ N for all n 1. In the matrix element (ξ , pˆcξ ) = ≥ n m Hˆ (n) −ix In fact, limN→∞ HfN is divergent. But one can apply i(ξn, e (p 1/2)ξm)Hˆ (with n,m 1) the vector − − ≥ ψˆn directly to (A13). Again, all intermediate results are ψˆm+1 well-defined, for example (p 1/2)ξm = (m 1)ξm (2m + 1) − − − m +1+1/2 c c 2 c n(n + 1) xˆ 1 (ˆp ) + 2iˆp ψˆn = ξn for n 1 , p (A9) − 2 n +1/2 ≥    p (A16) never has a component proportional to ψˆ0 (see (3.30)), because for m = 1 the first term vanishes. Then, as a ˆ consequence of (3.28), the operator e−ix in (ξ , pˆcξ ) and one obtains the expected result Hψn = (n + n m Hˆ 2 ˆ can be shifted to the left side for all values n,m 1. 1/2) ψn/2. It even turns out that this result holds for This, together with the anti-Hermiticity of p shows≥ that n = 0, so that the representation (A13) is correct in the c ˆ ˆ ˆ pˆ is Hermitian in the entire space ˆ . entire space = span ψ0, ψ1,... . ξ V { } The commutator [ˆxc, pˆc] can be calculatedV in a similar c c c −ix way: One hasx ˆ pˆ ξn = iˆx e (p 1/2)ξn. Since (p − − − APPENDIX B: THE CONNECTION BETWEEN 1/2)ξn never has a ξ1 component, one can apply (A3), so that REAL SPECTRUM AND PT SYMMETRY

c c xˆ pˆ ξn = i(p 1/2)ξn for all n 1 . (A10) In [2], Bender et al. attribute the reality of the spec- − − ≥ c −ix trum of non-Hermitian Hamiltonians H to the in- Writingp ˆ = i(p +1/2)e , one can use (A2) to derive variance of H. In fact, many examples of non-Hermitian,PT pˆcxˆcξ = i(p−+1/2)ξ for all n 1. Combining these n − n ≥ symmetric Hamiltonians with real spectra have sub- relations leads to the desired result: sequentlyPT been found [3–6]. Still, also examples of non- Hermitian Hamiltonians with real spectra that are not [ˆxc, pˆc]ξ = iξ for all n 1 . (A11) n n ≥ invariant are known, see, e. g., [7,8,18]. A proof that PT c c invariance is related to the reality of the spectrum In summary, the operatorsx ˆ andp ˆ are Hermitian, PT densely defined operators in ˆ that fulfill canonical com- does not exist. Here we want to present some arguments mutation relations. H against such a relation. One can now express the Hamiltonian (3.14) as a func- Let us consider general properties of non-Hermitian, c c symmetric Hamiltonians. We put Θ in the tion ofx ˆ andp ˆ : If we restrict the discussion for the PT ≡ PT moment to the subspace span ξ2, ξ3,... , we can replace following and assume e2ix in (3.14) by (ˆxc)2. Then{ (A10) gives} [H, Θ]=0 . (B1) 2 1 c c 1 c 2 H = iˆx pˆ + +(ˆx ) . (A12) Since Θ is anti-unitary, we find for an H eigenstate ψEn , 2 2 "  # HψEn = EnψEn (En being complex in general),

14 ∗ HΘψEn =ΘHψEn =ΘEnψEn = EnΘψEn , (B2) showing firstly that ΘψEn is an eigenstate of H with ∗ eigenvalue En and secondly that the eigenvalues of a Θ symmetric, non-Hermitian Hamiltonian always occur in complex-conjugate pairs. Thus, if the spectrum of H is non-degenerate, one has

∗ ΘψEn = const ψEn . (B3)

We want to emphasize that the Θ symmetry of a non- Hermitian Hamiltonian H is not sufficient to ensure the reality of the spectrum of H [2]. Only if one makes the additional assumption that ψEn is a simultaneous eigen- state of Θ, i. e. ΘψEn = γψEn , one can conclude that the spectrum of H is real. But it is important to note that the last assumption is crucial for the reality of the spectrum, and that there is no reason to believe that this assumption holds in gen- eral. To see this, recall the usual argument that is used to show that a Hamiltonian H and a further linear oper- ator A can be simultaneously diagonalized if [H, A] = 0: If Hψn = Enψn, one has

HAψn = AHψn = AEnψn = EnAψn , (B4) showing that Aψn is an eigenstate with the same eigen- value as ψn. If now the spectrum of H is non-degenerate, one has Aψn = const ψn, and if the spectrum of H is de- generate, one may still diagonalize A in the eigenspace corresponding to En, so that the eigenstates of H with eigenvalue En may still be superimposed to give eigen- states of A. If, however, A is an anti-linear operator, ∗ then (B4) gives HAψn = EnAψn, so that Aψn is not an eigenstate with the eigenvalue En. Hence, in general one cannot simultaneously diagonalize H and an anti-linear operator A, even if their commutator vanishes.

15