Abstract
The purpose of this note is to review certain recent results concerning the pseudospectra and the eigenvalues asymptotics of non-selfadjoint semiclassical pseudo-differential operators subject to small random perturbations.
1 Introduction
The spectral theory of non-selfadjoint operators acting on a Hilbert space is an established and highly developed subject. Non-selfadjoint operators are prevalent naturally in a wide range of modern problems. For instance, in the field of quantum mechanics, the study of scattering systems naturally leads to the notion of quantum resonances. These can be described as the complex values of the meromorphic. The continuation of the scattering matrix or of the cut-off resolvent of the Hamiltonian to the non-physical sheet of the complex plane. Alternatively, through a complex deformation of the initial Hamiltonian, these resonances can be characterized as the genuine complex valued eigenvalues of a non-selfadjoint operator [, , ]. We recommend the reader to reference [] for an in-depth discussion of the mathematics of scattering poles. Another aspect of quantum mechanics is the examination of a small system that is linked to a larger environment. The effective dynamics of the small systems are governed by a non-selfadjoint operator: the Lindbladian [].
A major obstacle to the spectral analysis of non-selfadjoint operators is the possible strong spectral instability of their spectrum with respect to small perturbations. This phenomenon, sometimes referred to as the pseudospectral effect, was initially considered to be a drawback, as it could lead to the origin of immense numerical errors, see Embree and Trefethen [] and the references therein. However, a recent line of research has also demonstrated that the pseudospectral effect can provide novel insights into the spectral distribution of non-selfadjoint operators that are subjected to small generic perturbations.
2 Spectral instability of non-selfadjoint operators
We commence by recalling the definition of the pseudospectrum of a linear operator, a crucial concept that which quantifies its spectral instability. This notion appears to have originated in the second half of the 20th century in various contexts, see reference [] for a historic overview. It quickly became an important notion in numerical analysis, as it allows us to quantify how much eigenvalues can spread out under the influence of small perturbations, see references [, ] and the book []. We follow here the latter reference.
Let be a complex Hilbert space (assumed separable for simplicity) with norm ||·|| and scalar product (·|·). Let be a closed densely defined linear operator, with resolvent set ρ(P) and spectrum Spec(P) = ℂ\ρ(P).
Definition 1. For any ε > 0, we define the ε-pseudospectrum of P by
We note that some authors define the ε-pseudospectrum with a ⩾ rather than a >. We, however, follow here reference []. It is noteworthy that with this choice of non-strict inequality results in the Specε(P) being an open set in ℂ.
For P selfadjoint (or even normal), the spectral theorem implies that
where D(0, ε) ⊂ ℂ denotes the open disk with radius ε centered at 0. For P non-selfadjoint, the pseudospectrum of P can be much larger, as illustrated by the following example.
Example 1. For N ≫ 1, consider the Jordan block matrix
The spectrum of PN is given by {0}. Consider the vector , |z| ⩽ r < 1. Then,
So, Theorem 2 shows that for any ε > 0 and any r ∈ ]0, 1[ we have that for N > 1 sufficiently large
An immediate consequence of Equation 1 is the property that pseudospectra are nested. More precisely,
The set (Equation 1) describes a region of spectral instability of the operator P, since any point in the ε-pseudospectrum of P lies within the spectrum of a certain ε-perturbation of P [].
Theorem 1. Let ε > 0. Then
Proof. See reference [, p. 31].
A third, equivalent definition of the ε-pseudospectrum of P is provided by the existence of approximate solutions to the eigenvalue problem (P − z)u = 0.
Theorem 2. Let ε > 0 and z ∈ ℂ. Then the following statements are equivalent:
z ∈ Specε(P);
z ∈ Spec(P) or there exists a such that ||(P − z)uz|| < ε||uz||, where denotes the domain of P.
Proof. See reference [, p. 31].
Such a state uz is referred to as an ε-quasimode, or simply a quasimode of P − z.
Spectral instability of semiclassical pseudo-differential operators
Although the notion of ε-pseudospectrum defined in Definition 1 is valid in the context of semiclassical pseudo-differential operators, we present here a somewhat different, but still related notion, which is more suited to the semiclassical setting. Here, the term “semiclassical” implies that our operators are dependent on a parameter h ∈ ]0, 1] (often referred to as “Planck's parameter”), and that our focus we will be on the asymptotic (semiclassical) regime h ↘ 0. This small parameter will provide us with a natural threshold for defining the pseudospectrum, and thereby measuring the spectral instability. The following discussion is based on the studies of Davies [] and Dencker et al. [].
Let d ⩾ 1 and h ∈ ]0, 1]. An order function, is a function satisfying the following growth condition:
where denotes the “Japanese brackets.” We will also sometimes write (x, ξ) = ρ ∈ ℝ2d, so that ξ ∈ ℝd. To such an order function m, we may associate a semiclassical symbol class [, ]. We assert that a smooth function belongs to the symbol class S(m) if for any multiindex α ∈ ℕ2d when there exists a constant Cα > 0 such that
We recommend the reader for further reading on semiclassical analysis to [, , ].
Let the symbol p ∈ S(m), m ⩾ 1, be a “classical” symbol, which satisfies an asymptotic expansion in the limit h → 0:
where each pj ∈ S(m) is independent of h. We assume that there exists a z0 ∈ ℂ and a C0 > 0 such that
Here, T*ℝd ≃ ℝ2d denotes the cotangent space of ℝd. In this case, we call p0 the (semiclassical) principal symbol of p. We then define two subsets of ℂ associated with p0:
Here, the denotes the closure of the set , and we will use this notation in the sequel. The set Σ is the classical spectrum, and Σ∞ can be called the classical spectrum at infinity of the h-Weyl quantization of p was defined by
seen as an oscillatory integral in ξ. The operator Ph maps S → S, and by duality S′ → S′, continuously.
3.1 Semiclassical pseudospectrum
Similar to Dencker et al. [], we define for a symbol p ∈ S(m) as in Equation 8 the sets
where {·, ·} denotes the Poisson bracket. It should be noted that the condition is the classical analog of the . As in Dencker et al. [], we call the set
the semiclassical pseudospectrum.
Theorem 3 ([]). Suppose that n ⩾ 2, , and is compact for a dense set of values z ∈ ℂ. If , then
and for every z ∈ Λ+(p0) and every with
there exists such that
If, in addition, p has a bounded holomorphic continuation to to {ρ ∈ ℂ2d, |Imρ| ⩽ 1/C}, then Equation 14 holds with the h∞ replaced by exp(−1/(Ch)).
If n = 1, then the same conclusion holds, provided that in addition to the general assumptions, each component of ℂ\Σ∞ has a nonempty intersection with ∁Λ(p).2
This result can be extended to unbounded symbols p ∈ S(T*ℝd, m), as shown in Equation 8, and the corresponding operators Ph with principal symbol p0, by applying Theorem 3 to , with principal symbol and z0 as in Equation 9 and z0 ≠ z. Indeed, note that z ∈ Σ(p0) if and only if , and that with ±{Rep0, Imp0}(ρ) < 0 is equivalent to with . Furthermore, a quasimode u as in Theorem 3 for then provides, after a possible truncation, a quasimode for Ph − z in the same sense.
By replacing Ph with its formal adjoint, , and thus p with , Theorem 3 yields that for every z ∈ Λ−(p) and every with
there exists such that
The additional statements of Theorem 3 regarding symbols that permit a holomorphic extension to a complex neighborhood of ℝ2d, and the case where n = 1 hold as well.
Example 2. The case study to be considered is the case of the non-selfadjoint Harmonic oscillator
is seen as an unbounded operator L2(ℝ) → L2(ℝ). The principal symbol for Ph is given by p(x, ξ) = ξ2 + ix2 ∈ S(T*ℝ, m), with a weight function m(x, ξ) = 1 + ξ2 + x2. We equip Ph with the domain , where the operator on the right is the pseudo-differential inverse of Ph + 1. This choice of domain renders Ph a closed and densely defined operator. Using, for instance, the method of complex scaling, it can be observed that the spectrum of Ph is determined by
Furthermore, Σ is the closed first quadrant in the complex plane, whereas Σ∞ = ∅. For ρ = (x, ξ) ∈ T*ℝ, we find that
Thus, for every z ∈ 3 there exist points
such that
Using the WKB method, it is possible to construct quasimodes of the form with admitting an asymptotic expansion with and
see Davies [, ] for an explicit computation, and Dencker et al. [] for a more general construction.
In fact, the works of Davies [, ] provide an explicit WKB construction for a quasimode u for one-dimensional non-selfadjoint Schrödinger operators on L2(ℝ) with V ∈ C∞(ℝ) complex-valued and z = V(a)+η2, for some a ∈ ℝ, η > 0. Furthermore, one assumes that ImV′(a) ≠ 0. These studies served as the foundation for the quasimode construction of non-selfadjoint (pseudo-)differential operators. Zworski [] compared Davies' quasimode construction under the condition on the gradient of ImV to a quasimode construction under a non-vanishing condition of the Poisson bracket . Furthermore, Zworski [] established the link to the famous commutator condition of Hörmander [, ]. A full generalization of the quasimode construction under a non-vanishing condition of the poisson bracket, see Theorem 3, was then achieved by Dencker et al. []. Finally, Pravda-Starov [–] improved these results by modifying a quasimode construction by Moyer and Hörmander, see reference [, Lemma 26.4.14], for adjoints of operators that do not satisfy the Nirenberg-Tréves condition (Ψ) for local solvability.
For a quasimode construction for non-selfadjoint boundary value problems, we recommend the reader refer to the study of Galkowski [].
It is noteworthy, that Equation 14 (or Equation 17 in the aforementioned example) implies that if the resolvent exists then its norm is larger than any power of h when h → 0, or even larger than e1/Ch in the analytical case. Each family is an h∞-quasimode of Ph − z, or for short a quasimode of Ph − z.
From the quasimode Equation 14, it is easy to observe an operator Q of unity norm and a parameter , such that the perturbed operator Ph + δQ has an eigenvalue at z. For instance, if we call the error r+ = (Ph − z)e+, we may take the rank 1 operator . According to Theorem 3, it can be observed that the interior of the set Λ(p), situated away from the set Σ∞, is a zone of strong spectral instability for Ph. For this reason, we may refer to the semiclassical pseudospectrum Λ(p) also as the (h∞-) pseudospectrum of Ph. Finally, we recommend the reader also to the refer studies of Pravda-Starov [–] for further refinement of the notion of semiclassical pseudospectrum.
3.2 Outside the semiclassical pseudospectrum
When
then by condition (Equation 9), we have (p0(ρ) − z) ⩾ m(ρ)/C for some sufficiently large C > 0 and so we know that the inverse is a pseudo-differential operator with principal symbol . Hence, maps L2 → L2 and
uniformly in h > 0. Therefore, from the semiclassical point of view, we may consider ℂ\Σ as a zone of spectral stability.
3.3 At the boundary of the semiclassical pseudospectrum
At the boundary of the semiclassical pseudospectrum, a transition occurs between the zone of strong spectral instability and stability. Indeed, at the boundary we find an improvement over the resolvent bounds, assuming some additional non-degeneracy:
Splitting a symbol into real and imaginary part, p = p1 + ip2, we consider the iterated Poisson bracket
where I ∈ {1, 2}k, and |I| = k is called the order of the Poisson bracket. The order of p at ρ ∈ T*ℝd is given by
The order of z0 ∈ Σ\Σ∞ is the maximum of k(ρ) for .
Theorem 4. See Dencker et al. [, ] Assume that . Let and let z0 ∈ ∂Σ(p0)\Σ∞(p0). Assume that dp0 ≠ 0 at every point in , and that z0 has a finite order k ⩾ 1 for p. Then, k is equal and h > 0 is small enough for
In particular, there exists a c0 > 0, such that h > 0 is small enough for
This result was proven in dimension 1 by Zworski [], and in certain cases by Boulton []. Further refinements have been obtained from Sjöstrand []. Similar to the discussion after Theorem 3, we can extend Theorem 4 to unbounded symbols p ∈ S(T*ℝd, m) and their corresponding quantizations.
Example 3. Recall the non-selfadjoint Harmonic oscillator from Example 2. Here ∂Σ = ℝ+ ∪ iℝ+, so we see by Equation 16 that for 0 ≠ z0 ∈ Σ
However,
indicating that z0 is of order 2 for p = ξ2 + ix2, and Theorem 4 reveals that
In order for a the ε-pseudospectrum of Ph to reach the boundary of Σ, we require ε > h2/3/C.
3.4 Pseudospectra and random matrices
In this section, we present a brief discussion on pseudospectra for large N × N random matrices. One may interpret the 1/N, where N ≫ 1, as an analog to the semiclassical parameter. By recalling the example of the non-selfadjoint harmonic oscillator, as illustrated in Example 2, we see that pseudospectra can be very large in general. However, in a generic setting, they are typically much smaller.
Let M ∈ ℂN×N be a complex N × N matrix and let s1(M) ⩾ … ⩾ sN(M) ⩾ 0 denotes its singular values, which are the eigenvalues of ordered in a decreasing manner and counting multiplicities. It should be noted that if M − z is bijective for some z ∈ ℂ, then
In view of Equation 1, the ε-pseudospectrum of M is then characterized by the condition that z ∈ Specε(M)
A classical result from Sankar et al. [, Lemma 3.2] (stated there for real Gaussian random matrices) indicates that with a high probability, the smallest singular value of a deformed random matrix is not too small.
Theorem 5 ([]). There exists a constant C > 0 such that the following holds true. Let N ⩾ 2, let X0 be an arbitrary complex N × N matrix, and let Q be an N × N complex Gaussian random matrix, whose entries are all independent copies of a complex Gaussian random variable . Subsequently, for any δ > 0
Proof. For real matrices the proof can be found in Sankar et al. [, Lemma 3.2], see also reference [, Theorem 2.2]. For complex matrices a proof is presented for instance in Vogel [, Appendix A].
Theorem 5 states us that any fixed z ∈ ℂ is not included in the ε-pseudospectrum of X + δQ with a probability ⩾ 1 − CNε2δ−2. This result suggests that the pseudospectrum of random matrices is typically not too large. Theorem 5 has received many extensions. For instance Rudelson and Vershynin [] consider the case of random matrices with iid (independent and identically distributed) sub-Gaussian entries. Tao and Vu [] consider iid entries with a nonzero variance. Cook [] considers the case of random matrices whose of entries have an inhomogeneous variance profile under appropriate assumptions. We conclude this section by noting the following, quantitative outcome obtained by Tao and Vu.
Theorem 6 ([]). Let q be a random variable with a mean zero and a bounded second moment, and let γ ⩾ 1/2, A ⩾ 0 be constants. Then, there exists a constant C > 0, depending on q, γ, and A such that the following holds true. Let Q be the random matrix of size N, whose entries are independent and identically distributed copies of q, and let X0 be a deterministic matrix satisfying . Then,
Example 4. Consider the case where q is a random variable satisfying the moment conditions
Form [] reveals that Equation 20 implies that 𝔼[||Q||] ⩽ CN1/2, which, using Markov's inequality, yields that for any ε > 0
In this case (Equation 19) becomes
4 Eigenvalue asymptotics for non-selfadjoint (random) operators
Consider the operator depicted in Equations 8, 11, which is viewed as an unbounded operator L2(ℝd) → L2(ℝd). We equip Ph with the domain . It should be noted that exists for h > 0 that is sufficiently small by the elipticity condition (Equation 9). We will denote by ||u||m: = ||(Ph − z0)u|| the associated norm on H(m). Although this norm depends on the selection of the symbol p0 − z0, it is equivalent to the norm defined by any operator with an elliptic principal symbol q ∈ S(m), so that the space H(m) solely depends on the order function m. Since H(m) contains the Schwartz functions , it is dense in L2(ℝd).
Let us verify that Ph equipped with domain H(m) is closed. Let (Ph − z0)uj → v and uj → u in L2. Since is bijective, it follows that in H(m) and also in L2. So . In summary, Ph equipped with the domain H(m) is a densely defined closed linear operator.
Recall Equation 10, and let
be open, relatively compact, not entirely contained in Σ and so that . Using the ellipticity assumption (Equation 9), it was proven in reference [, Section 3] that
Spec(Ph) ∩ Ω is discrete for h > 0 small enough,
For all ε > 0 there exists an h(ε) > 0 such that
where D(0, ε) denotes the disc in ℂ of radius ε and centered at 0.
4.1 The selfadjoint setting
If Ph above is selfadjoint, which implies in particular that p is real-valued, we have the classical Weyl asymptotics. We follow here Dimassi and Sjöstrand [] for a brief review.
Theorem 7. Let Ω be as in Equation 23. For every h-independent interval I ⊂ Ω ∩ ℝ with ,
This result is, in increasing generality, attributed to Chazarin [], Helffer and Robert [, ], Petkov and Robert [] and Ivrii []. See also Dimassi and Sjöstrand [] for an overview. We highlight two special cases: when I = [a, b], a < b, and a, b are not critical points of p0, then the error term becomes , see Chazarin [], Helffer-Robert [], and Ivrii []. When additionally the unions of periodic Hp0 trajectories4 in the energy shell and are of the Liouville measure 0, then the error term is of the form
where Lλ denotes the Liouville measure on . See Petkov and Robert [] and Ivrii [] and Dimassi and Sjöstrand [] for details. Let us also highlight that similar results obtained from Theorem 7 are also valid for compact smooth manifolds, see, for instance, Grigis and Sjöstrand [, Chapter 12] and the references therein.
The corresponding results in the setting of self-adjoint partial differential operators in the high energy limit go back to the seminal study of Weyl [] and have a long and very rich history. These are, however, beyond the scope of this review.
Example 5. The guiding example to keep in mind is the self-adjoint Harmonic oscillator
seen as an unbounded operator. The principal symbol of Ph is represented by p(x, ξ) = ξ2 + x2 ∈ S(T*ℝ, m), and the weight function m(x, ξ) = 1 + ξ2 + x2. Ph is represented by the domain , where the operator on the right is the pseudo-differential inverse of Ph + 1. This choice of domain makes Ph a densely defined closed operator. It is widely acknowledged (see, for instance, reference [, Theorem 6.2]) that the spectrum of Ph is determined by
Counting the points (2n + 1)h contained in an interval [a, b], 0 ⩽ a < b < ∞, gives
Since , we confirm Theorem 7 for the Harmonic oscillator.
4.2 The non-self-adjoint setting
The natural counterpart of Theorem 7 for non-self-adjoint operators would be eigenvalue asymptotics in a complex domain Ω ⋐ ℂ as in Equation 23. Recall the non-self-adjoint Harmonic oscillator Ph from Example 2 with principal symbol p(x, ξ) = ξ2 + ix2. In this case, Σ = {z ∈ ℂ; Rez, Imz ⩾ 0} and Σ∞ = ∅. Any ∅ ≠ Ω ⋐ Σ away from the line , indicates the view of Equation 15 that
On the other hand,
This example suggests that a direct generalization of Theorem 7 to non-self-adjoint operators with a complex valued principal symbol cannot hold.
Let us comment on two settings where a form of Weyl asymptotics is known to hold: Upon assuming analyticity, one may recover a sort of Weyl asymptotics. More precisely, as shown in the studies of Melin and Sjöstrand [], Sjöstrand [], Hitrik and Sjöstrand [–], Hitrik et al. [], and Rouby [], the discrete spectrum of certain analytic non-self-adjoint pseudo-differential operators is confined to curves in Σ. Moreover, one can recover eigenvalue asymptotics using Bohr-Sommerfeld quantization conditions.
The second setting occurs when the non-self-adjointness of the operator Ph arises not from the principal symbol p0 (assumed to be real-valued), but from the subprincipal symbol p1. For instance, when studying the damped wave equation on a compact Riemannian manifold X, one is led to study the eigenvalues of the corresponding stationary operator
Here, Δ denotes the Laplace-Beltrami operator on X, and we call z ∈ ℂ an eigenvalue of Ph(z) if there exists a corresponding L2 function u is present in the kernel of Ph(z)−z. In fact, such a u is smooth by elliptic regularity. Using Fredholm theory, one can show that these eigenvalues form a discrete set in ℂ.
The principal part of Ph = Ph(z) is given by −h2Δ, and thus is self-adjoint. The principal symbol is (the norm here is with respect to the Riemannian metric on X). However, the subprincipal part is complex valued and non-self-adjoint.
Lebeau [] has established that there exists a± ∈ ℝ, wherein for every ε > 0 there exist a finite number of eigenvalues such that
Remark 1. In fact Lebeau provided precise expressions for a± in terms of the infimum and the supremum over the co-sphere bundle S*X of the long time average of the damping function a evolved via the geodesic flow. Further refinements have been obtained by Sjöstrand [], and when X is negatively curved by Anantharaman [] and Jin [].
Additionally, Markus and Matsaev [] and Sjöstrand [] have demonstrated the following analog of the Weyl law. For 0 < E1 < E2 < ∞ and for C > 0 sufficiently large
Finer results have been obtained by Anantharaman [] and Jin [] when X is negatively curved.
4.3 Probabilistic Weyl asymptotics
In a series of studies by Hager [–] and Sjöstrand [, ], the authors proved a Weyl law, with overwhelming probability, for the eigenvalues in a compact set Ω ⋐ ℂ as in Equation 23 for randomly perturbed operators
where Ph is as per in Section 3, and the random perturbation Qω is one of the following two types.
4.3.1 Random matrix
Let N(h) → ∞ sufficiently fast as h → 0. Let qj,k, 0 ⩽ j, k < N(h) be independent copies of a complex Gaussian random variable . We consider the random matrix
where is an orthonormal basis and for u ∈ L2(ℝ). The condition on N(h) is determined by the requirement that the microsupport of the vectors in the orthonormal system {ej}j < N(h), “covers” the compact set , where p0 is the principal symbol of Ph. For instance, we could consider the first N(h) eigenfunctions (ordered according to increasing eigenvalues) of the Harmonic oscillator on ℝd. The number N(h) is then determined by the condition that the semiclassical wavefront sets of ej, j ⩾ N(h), are disjoint from . Alternatively, as in Hager and Sjöstrand [], one may also take N(h) = ∞; however, then one must conjugate Qω by suitable elliptic Hilbert–Schmidt operators. We recommend the reader to Hager and Sjöstrand [] for further information.
4.3.2 Random potential
We take N(h) and an orthonormal family (ek)k ∈ ℕ as above. Let v be real or complex random vector in ℝN(h) or ℂN(h), respectively, with joint probability law
where Zh > 0 is a normalization constant, B(0, R) is either the real ball ⋐ℝN(h) or the complex ball ⋐ℂN(h) of radius R = R(h) ≫ 1, and centered at 0, L(dv) denotes the Lebesgue measure on either ℝN(h) or ℂN(h) and ϕ ∈ C1 with
uniformly, for an arbitrary but fixed value of κ4 ⩾ 0. In Hager [] the case of non-compactly supported probability law was considered. More precisely, the entries of the random vector v were supposed to be independent and identically distributed (iid) complex Gaussian random variables . In Sjöstrand [, ], the law Equation 29 was considered. For the sake of simplicity, we will not elaborate here the precise conditions on the ek, R(h), and N(h), in this case, but refer the reader to Sjöstrand [, ]. However, one example of a random vector v with law (Equation 30) is a truncated complex or real Gaussian random variables with expectation 0, and uniformly bounded covariances. In fact, the methods in Sjöstrand [, ] can be extended to non-compactly supported probability distributions, provided sufficient decay conditions at infinity are assumed. For instance, iid complex Gaussian random variables, as in the one dimensional case [], are permissable. Finally, we conclude that the methods in Sjöstrand [, ] can probably also be modified to allow for the case of more general independent and identically distributed random variables. We define the random function as
We call this perturbation a “random potential,” even though Vω is complex valued. When we consider this type of perturbation, we will make the additional symmetry assumption:
Let Ω ⋐ ℂ be an open simply connected set as in Equation 23. For z ∈ Ω and 0 ⩽ t ≪ 1 we set
Let Γ ⋐ Ω be open with boundary and make the following non-flatness assumption
The above mentioned works have yielded the following result.
Theorem 8 (Probabilistic Weyl's law). Let Ω be as in Equation 23. Let Γ ⋐ Ω be open with boundary. Let be a randomly perturbed operators as in Equation 27 with e−1/Ch ≪ δ ⩽ hθ with θ > 0 sufficiently large. Then, in the limit h → 0,
for some fixed η > 0.
The studies [–, , ] also provide an explicit control over θ, the error term in Weyl's law, and the error term in the probability estimate. Theorem 8 is remarkable because such Weyl laws are typically a feature of self-adjoint operator, whereas in the non-selfadjoint case they generally fail. Indeed, as laid out in Section 4.2, the discrete spectrum of the (unperturbed) non-selfadjoint operator Ph is usually localized to curves in the pseudospectrum Σ, see Melin and Sjöstrand [], Hitrik and Sjöstrand [–], and Rouby []. In contrast, Theorem 8 shows that a “generic” perturbation of size is sufficient for the spectrum to “fill out” Σ.
To illustrate this phenomenon, recall the non-selfadjoint harmonic oscillator on ℝ from Example 2. Its spectrum is given by {eiπ/4(2n + 1)h; n ∈ ℕ} [] on the line . The Theorem 8 shows that a “generic” perturbation of arbitrarily small size is sufficient to produce spectrum roughly equidistributed in any fixed compact set in its classical spectrum Σ, which is in this case the upper right quadrant of ℂ.
As observed in Christiansen and Zworski [], the real analytic p condition (Equation 34) consistently holds for some κ > 0. Similarly, when p is truly analytical and such that Σ ⊂ ℂ has non-empty interior, then
For smooth p, we have that when for every z ∈ ∂Ω
Observe that dp and are linearly independent at ρ when , where {a, b} = ∂ξa·∂xb − ∂xa·∂ξb denotes the Poisson bracket. Moreover, in dimension d = 1, the condition on p−1(z) is equivalent to dp, with being linearly independent at every point of p−1(z). However, in dimensions d > 1, this cannot in hold general, as the integral of with respect to the Liouville measure on p−1(z) vanishes on every compact connected component of p−1(z), see reference [, Lemma 8.1]. Furthermore, condition (Equation 37) cannot hold when z ∈ ∂Σ. However, some iterated Poisson brackets may not have zero there. For example, it has been observed in [, Example 12.1] that if
4.3.3 Related results
Theorem 8 has also been extended to the case of elliptic semiclassical differential operators on compact manifolds by Sjöstrand [], to the Toeplitz quantization of the torus by Christiansen and Zworski [] and Vogel [], and to general Berezin-Toeplitz quantizations on compact Kähler manifolds by Oltman [] in the context of complex Gaussian noise. A further extension of Theorem 8 has been achieved by Becker, Oltman and the author in Becker et al. []. There we prove a probabilistic Weyl law for the non-selfadjoint off-diagonal operators of the Bistritzer-MacDonald Hamiltonian [] for twisted bilayer graphene, see also Cancés et al. [] and Watson et al. [], subject to random tunneling potentials. This probabilistic Weyl has an interesting physical consequence as it demonstrates the instability of the so-called magic angels for this model of twisted bilayer graphene. Similar results have been achieved in random matrix theory. The case of Toeplitz matrices is represented by symbols on T2 of the form , (x, ξ) ∈ T2, has been conducted in a series of recent studies by Śniady [], Davies and Hager [], Guionnet et al. [], Basak et al. [, ], Sjöstrand and the author of this text [–]. Such symbols amount to the case of symbols which are constant in the x variable. In these studies the non-selfadjointness of the problem, however, does not come from the symbol itself, but from the boundary conditions destroying it. The periodicity of the symbol in x is achieved by allowing for a discontinuity. Nevertheless, these studies demonstrate that by adding a small random matrix, the limit of the empirical eigenvalues counting measure μN of the perturbed operator converges in probability (or even almost surely in some cases) to p*(dρ).
Statements
Author contributions
MV: Writing – original draft, Writing – review & editing.
Funding
The author(s) declare financial support was received for the research, authorship, and/or publication of this article. I am partly supported by the ANR Grant ADYCT ANR-20-CE40-0017.
Conflict of interest
The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Footnotes
1.^This implies that the semiclassical wavefront set of e+ is defined by ρ0. In other words, the state e+ is concentrated in position and frequency near the point ρ0. See, for instance, Zworski [] for a definition. For u = (u(h))h ∈ (0, 1) a bounded family in L(ℝ), its semiclassical wavefront set WFh(u) denotes the phase space region where u is h-microlocalized:where a denotes the Weyl quantization of a, and ∁U denotes the complement of a given set U.
2.^∁Λ(p) denotes the complement of the set Λ(p).
3.^ denotes the interior of the set Σ.
4.^Hp0 denotes the Hamilton vector field induced by p0.
References
1.
AguilarJCombesJ-M. A class of analytic perturbations for one-body schrödinger Hamiltonians. Commun Math Phys. (1971) 22:269–79. 10.1007/BF01877510
2.
AnantharamanN. Spectral deviations for the damped wave equation. Geom Funct Anal. (2010) 20:3. 10.1007/s00039-010-0071-x
3.
BalslevECombesJ-M. Spectral properties of many-body Schrödinger operators with dilation analytic interactions. Comm Math Phys. (1971) 22:280–94. 10.1007/BF01877511
4.
BasakAPaquetteEZeitouniO. Regularization of non-normal matrices by Gaussian noise - the banded Toeplitz and twisted Toeplitz cases. Forum of Mathematics, Sigma. (2019) 7:e3. 10.1017/fms.2018.29
5.
BasakAPaquetteEZeitouniO. Spectrum of random perturbations of Toeplitz matrices with finite symbols. Trans Amer Math Soc. (2020) 373:4999–5023. 10.1090/tran/8040
6.
BeckerSOltmanIVogelM. Absence of small magic angles for disordered tunneling potentials in twisted bilayer graphene. arXiv preprint arXiv:2402.12799. (2024).
7.
BistritzerRMacDonaldA. Moiré bands in twisted double-layer graphene. PNAS. (2011) 108:12233–7. 10.1073/pnas.1108174108
8.
BoultonL. Non-self-adjoint harmonic oscillator, compact semigroups and pseudospectra. J Operator Theory. (2002) 47:413–29.
9.
CancèsEGarrigueLGontierD. A simple derivation of moiré-scale continuous models for twisted bilayer graphene. Phys Rev B. (2023) 107:155403. 10.1103/PhysRevB.107.155403
10.
ChazarainJ. Spectre d'un hamiltonien quantique et mécanique classique. Comm PDE. (1980) 5:595–644. 10.1080/0360530800882148
11.
ChristiansenTJZworskiM. Probabilistic Weyl laws for quantized tori. Comm Math Phys. (2010) 299:305–34. 10.1007/s00220-010-1047-2
12.
CookN. Lower bounds for the smallest singular value of structured random matrices. Ann Probab. (2018) 46:3442–500. 10.1214/17-AOP1251
13.
DaviesEB. Pseudo–spectra, the harmonic oscillator and complex resonances. Proc Royal Soc London A. (1999) 455:585–99. 10.1098/rspa.1999.0325
14.
DaviesEB. Semi-classical states for non-self-adjoint Schrödinger operators. Comm Math Phys. (1999) 200:35–41. 10.1007/s002200050521
15.
DaviesEBHagerM. Perturbations of Jordan matrices. J Approx Theory. (2009) 156:82–94. 10.1016/j.jat.2008.04.021
16.
DenckerNSjöstrandJZworskiM. Pseudospectra of semiclassical (pseudo-) differential operators. Commun Pure Appl Mathem. (2004) 57:384–415. 10.1002/cpa.20004
17.
DimassiMSjöstrandJ. Spectral Asymptotics in the Semi-Classical Limit. London Mathematical Society Lecture Note Series 268, Cambridge: Cambridge University Press (1999).
18.
DyatlovSZworskiM. Mathematical Theory of Scattering Resonances. London: American Mathematical Society (2019).
19.
EmbreeMTrefethenLN. Spectra and Pseudospectra: The Behavior of Nonnormal Matrices and Operators. Princeton: Princeton University Press (2005). 10.1515/9780691213101
20.
GalkowskiJ. Pseudospectra of semiclassical boundary value problems. J Inst Math Jussieu. (2014) 2:405–49. 10.1017/S1474748014000061
21.
GrigisASjöstrandJ. Microlocal Analysis for Differential Operators, London Mathematical Society Lecture Note Series 196. Cambridge: Cambridge University Press (1994).
22.
GuionnetAWoodPMZeitouniO. Convergence of the spectral measure of non-normal matrices. Proc AMS. (2014) 142:667–79. 10.1090/S0002-9939-2013-11761-2
23.
HagerM. Instabilité Spectrale Semiclassique d'Opérateurs Non-Autoadjoints II. Ann Henri Poincare. (2006) 7:1035–64. 10.1007/s00023-006-0275-7
24.
HagerM. Instabilité spectrale semiclassique pour des opérateurs non-autoadjoints I: un modèle. Ann faculté des Sci Toulouse Sé. (2006) 6:243–80. 10.5802/afst.1121
25.
HagerMSjöstrandJ. Eigenvalue asymptotics for randomly perturbed non-selfadjoint operators. Mathem Ann. (2008) 342:177–243. 10.1007/s00208-008-0230-7
26.
HelfferBRobertD. Comportement semi-classique du spectre des hamiltoniens quantiques elliptiques. Ann Inst Fourier, Grenoble. (1981) 31:169–223. 10.5802/aif.844
27.
HelfferBRobertD. Calcul fonctionnel par la transformation de mellin et opérateurs admissibles. J Funct Anal. (1983) 53:246–68. 10.1016/0022-1236(83)90034-4
28.
HitrikMSjöstrandJ. Non-selfadjoint perturbations of selfadjoint operators in 2 dimensions I. Ann Henri Poincaré. (2004) 5:1–73. 10.1007/s00023-004-0160-1
29.
HitrikMSjöstrandJ. Non-selfadjoint perturbations of selfadjoint operators in 2 dimensions II. Vanishing averages. Comm Partial Differ Equat. (2005) 30:1065–1106. 10.1081/PDE-200064447
30.
HitrikMSjöstrandJ. Non-selfadjoint perturbations of selfadjoint operators in 2 dimensions III a. One branching point. Canadian J Math. (2008) 60:572–657. 10.4153/CJM-2008-028-3
31.
HitrikMSjöstrandJVũ NgọcS. Diophantine tori and spectral asymptotics for non-selfadjoint operators. Amer J Math. (2007) 129:105–82. 10.1353/ajm.2007.0001
32.
HörmanderL. Differential Equations without Solutions. Math Annalen. (1960) 140:169–173. 10.1007/BF01361142
33.
HörmanderL. Differential operators of principal type. Math Annalen. (1960) 140:124–146. 10.1007/BF01360085
34.
HörmanderL. The Analysis of Linear Partial Differential Operators IV, Grundlehren der mathematischen Wissenschaften, vol. 275. Cham: Springer-Verlag (1985).
35.
IvriiV. Microlocal Analysis and Precise Spectral Asymptotics. New York: Springer (1998). 10.1007/978-3-662-12496-3
36.
JinL. Damped wave equations on compact hyperbolic surfaces. Comm Math Phys. (2020) 373:771–94. 10.1007/s00220-019-03650-x
37.
R. Latala. Some estimates of norms of random matrices. Proc Amer Math Soc. (2005) 133:1273–82. 10.1090/S0002-9939-04-07800-1
38.
LebeauG. Équation des ondes amorties, Algebraic and Geometric Methods in Mathematical Physics (Kaciveli, 1993). Dordrecht: Kluwer Academy Publication (1996). 10.1007/978-94-017-0693-3_4
39.
LindbladG. On the generators of quantum dynamical semigroups. Commun Math Phys. (1976) 48:119–30. 10.1007/BF01608499
40.
MarkusASMatsaevVI. Comparison theorems for spectra of linear operators and spectral asymptotics Trudy Moskov. Mat Obshch. (1982) 45:133–81.
41.
MartinezA. An Introduction to Semiclassical and Microlocal Analysis. Cham: Springer (2002). 10.1007/978-1-4757-4495-8
42.
MelinASjöstrandJ. Determinants of pseudodifferential operators and complex deformations of phase space. Methods Appl Anal. (2002) 9:177–237. 10.4310/MAA.2002.v9.n2.a1
43.
MelinASjöstrandJ. Bohr-Sommerfeld quantization condition for non-selfadjoint operators in dimension 2. Astérisque. (2003) 284:181–244.
44.
OltmanI. A probabilistic Weyl-law for Berezin-Toeplitz operators. J Spectral Theory. (2023) 13:727–54. 10.4171/jst/459
45.
PetkovVRobertD. Asymptotique semi-classique du spectre d'hamiltoniens quantiques et trajectoires classiques périodiques. Comm in PDE. (1985) 10:365–90. 10.1080/03605308508820382
46.
Pravda-StarovK. A general result about the pseudo-spectrum of Schrödinger operators. Proc R Soc London Ser A Math Phys Eng Sci. (2004) 460:471–7. 10.1098/rspa.2003.1194
47.
Pravda-StarovK. Étude du pseudo-spectre d'opérateurs non auto-adjoints. Ph.D. thesis (2006).
48.
Pravda-StarovK. Pseudo-spectrum for a class of semi-classical operators. Bull Soc Math France. (2008) 136:329–372. 10.24033/bsmf.2559
49.
RoubyO. Bohr-Sommerfeld quantization conditions for non-selfadjoint perturbations of selfadjoint operators in dimension one. Int Math Res Not IMRN. (2018) 7:2156–207. 10.1093/imrn/rnw309
50.
RudelsonMVershyninR. The least singular value of a randomsquare matrix is O(n−1/2). C R Math Acad Sci Paris. (2008) 346:893–6. 10.1016/j.crma.2008.07.009
51.
SankarASpielmannDATengSH. Smoothed analysis of the condition numbers and growth factors of matrices. SIAM J, Matrix Anal Appl. (2006) 28:446–76. 10.1137/S0895479803436202
52.
SjöstrandJ. Asymptotic Distribution of Eigenfrequencies for Damped Wave Equations. Publ RIMS, Kyoto Univ. (2000) 36:573–611. 10.2977/prims/1195142811
53.
SjöstrandJ. Perturbations of selfadjoint operators with periodic classical flow. In: Wave Phenomena and Asymptotic Analysis. RIMS Kokyuroku (2003). p. 1315.
54.
SjöstrandJ. Eigenvalue distribution for non-self-adjoint operators with small multiplicative random perturbations. Ann Fac Sci Toulouse. (2009) 18:739–795. 10.5802/afst.1223
55.
SjöstrandJ. Eigenvalue distribution for non-self-adjoint operators on compact manifolds with small multiplicative random perturbations. Ann Fac Toulouse. (2010) 19:277–301. 10.5802/afst.1244
56.
SjöstrandJ. Resolvent estimates for non-selfadjoint operators via semigroups. In: Around the Research of Vladimir Maz'ya III, International Mathematical Series 13. New York: Springer (2010). p. 359–384. 10.1007/978-1-4419-1345-6_13
57.
SjöstrandJVogelM. Large bi-diagonal matrices and random perturbations. J Spectral Theory. (2016) 6:977–1020. 10.4171/jst/150
58.
SjöstrandJVogelM. General Toeplitz matrices subject to Gaussian perturbations. Ann Henri Poincaré. (2021) 22:49–81. 10.1007/s00023-020-00970-w
59.
SjöstrandJVogelM. Toeplitz band matrices with small random perturbations. Indagationes Mathem. (2021) 32:275–322. 10.1016/j.indag.2020.09.001
60.
SjöstrandJZworskiM. Complex scaling and the distribution of scattering poles. J Am Mathem Soc. (1991) 43:4. 10.2307/2939287
61.
ŚniadyP. Random regularization of brown spectral measure. J Functional Anal. (2002) 193:291–313. 10.1006/jfan.2001.3935
62.
TaoTVuV. Random matrices: the circular law. Commun Contemp Math. (2008) 10:261–307. 10.1142/S0219199708002788
63.
TaoTVuV. Smooth analysis of the condition number and the least singular value. Math Comp. (2010) 79:2333–52. 10.1090/S0025-5718-2010-02396-8
64.
TrefethenLN. Pseudospectra of matrices. Numer Anal. (1992) 91:234–66.
65.
TrefethenLN. Pseudospectra of linear operators. SIAM Rev. (1997) 39:383–406. 10.1137/S0036144595295284
66.
VogelM. Almost sure Weyl law for quantized tori. Comm Math Phys 378. (2020) 2:1539–85. 10.1007/s00220-020-03797-y
67.
WatsonABKongTMacDonaldAHLuskinM. Bistritzer-macdonald dynamics in twisted bilayer graphene. J Math Phys. (2023) 64:031502. 10.1063/5.0115771
68.
WeylH. Das asymptotische verteilungsgesetz der eigenwerte linearer partieller differentialgleichungen (mit einer anwendung auf die theorie der hohlraumstrahlung). Math Ann. (1912) 71:441–79. 10.1007/BF01456804
69.
ZworskiM. A remark on a paper of E.B. Davies. Proc AMS. (2001) 129:2955–2957. 10.1090/S0002-9939-01-05909-3
70.
ZworskiM. Pseudospectra of semi-classical operators. Unpublished Leture, King's College (2001).
71.
ZworskiM. Semiclassical Analysis, Graduate Studies in Mathematics 138. New York: American Mathematical Society (2012). 10.1090/gsm/138
Summary
Keywords
semiclassical analysis, non-selfadjoint operators, random matrix, spectral theory, partial differential equation (PDE)
Citation
Vogel M (2024) Pseudospectra and eigenvalue asymptotics for disordered non-selfadjoint operators in the semiclassical limit. Front. Appl. Math. Stat. 10:1508973. doi: 10.3389/fams.2024.1508973
Received
10 October 2024
Accepted
11 November 2024
Published
02 December 2024
Volume
10 - 2024
Edited by
Jose Luis Jaramillo, Université de Bourgogne, France
Reviewed by
Wenfeng Chen, SUNY Polytechnic Institute, United States
Updates
Copyright
© 2024 Vogel.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Martin Vogel vogel@math.unistra.fr
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.