Streamlined Krylov construction and classification of ergodic Floquet systems

Nikita Kolganov1 and Dmitrii A. Trunin2, ∗

1
Institute for Theoretical and Mathematical Physics,
Moscow State University, Leninskie Gory, GSP-1, Moscow, 119991, Russia
2
Department of Physics, Princeton University, Princeton, NJ 08544, USA
(Dated: December 30, 2024)
We generalize Krylov construction to periodically driven (Floquet) quantum systems using the
theory of orthogonal polynomials on the unit circle. Compared to other approaches, our method
works faster and maps any quantum dynamics to a one-dimensional tight-binding Krylov chain. We
also suggest a classification of chaotic and integrable Floquet systems based on the asymptotic be-
havior of Krylov chain hopping parameters (Verblunsky coefficients). We illustrate this classification
with random matrix ensembles, kicked top, and kicked Ising chain.
arXiv:2412.19797v1 [quant-ph] 27 Dec 2024

Introduction— The emergence of ergodic behavior in Lanczos algorithm— First of all, we briefly review
isolated quantum systems is an old but thriving subject. the standard Krylov construction [16–18]. Consider a
Although most local interacting systems were believed to quantum system with a constant Hamiltonian H and d-
be fully ergodic, i.e., obey strong Eigenstate Thermaliza- dimensional Hilbert space endowed pwith an inner product
tion Hypothesis [1–3], recent studies revealed several no- (A|B) = ⟨A|B⟩ and norm ∥A∥ = (A|A). The dynamics
table exceptions from this rule. In particular, such novel of a seeding state |ψ⟩ unfolds in the corresponding D-
ergodicity breaking scenarios encompass many-body lo- dimensional Krylov subspace (D ≤ d) [55]:
calization [4–6], quantum many-body scars [7–9], and
|ψ(t)⟩ ∈ span |ψ⟩, H|ψ⟩, H 2 |ψ⟩, · · · , H D−1 |ψ⟩ . (1)

Hilbert space fragmentation [9–12]. Besides, the emer-
gence of ergodic behavior is closely related to quantum A complete orthonormal basis {|Ψn ⟩} in this Krylov
chaos, which has recently become increasingly impor- subspace is recursively generated by the Lanczos algo-
tant for understanding quantum field theory and quan- rithm [16–18]:
tum gravity [13–17]. Therefore, it is crucial to develop
methods that capture the dynamics of ergodic and near- an = ⟨Ψn |H|Ψn ⟩, bn = ⟨Ψn |H|Ψn−1 ⟩,
ergodic quantum systems in the most efficient way. (2)
bn+1 |Ψn+1 ⟩ = H|Ψn ⟩ − an |Ψn ⟩ − bn |Ψn−1 ⟩,
Krylov construction is one of the most powerful tools
to study quantum dynamics [16–18]. This construc- with initial conditions |Ψ0 ⟩ = |ψ⟩/∥ψ∥ and

b0 = 0 until
tion maps an arbitrary continuous time evolution into bD = 0. This algorithm requires O Dd2 operations and
a one-dimensional local hopping problem, which allows generally works faster than direct diagonalization of the
one to identify the minimal subspace where time evo- Hamiltonian. Moreover, the Hamiltonian acquires a tri-
lution unfolds and follow this dynamics cost-effectively. diagonal form in the basis {|Ψn ⟩}:
The efficiency of Krylov construction explains its numer-
ous applications, including quantum computing and con- ⟨Ψm |H|Ψn ⟩ = am δm,n + bm δm,n+1 + bn δm,n−1 . (3)
trol [19–23], classification of topological phases [24–27], In other words, Lanczos algorithm maps the dynamics of
semiclassical chaos [28, 29], and holography [30–40]. a seeding state |ψ⟩ into a local one-dimensional hopping
In this Letter, we extend Krylov construction to dis- problem on the Krylov chain:
crete time evolution, i.e., to periodically driven (Floquet)
systems. Our algorithm exploits the theory of orthogo- ∂t ϕn (t) = ian ϕn (t) − bn+1 ϕn+1 (t) + bn ϕn−1 (t), (4)
nal polynomials on the unit circle [41–45], works as fast −n
as Lanczos algorithm, and maps any discrete time evo- where Krylov wave function is ϕn (t) = i Ψn Ψ0 (t) .
lution to a one-dimensional local model. Other existing So, dynamical properties of a quantum system, such
approaches either lack these advantages [46–52] or sub- as chaos and thermalization, are closely related to prop-
stantially restrict the set of initial states and evolution erties of the Krylov wave function. The most important
operators [53, 54]. Furthermore, we show that hopping properties are position of the wavepacket ϕn (t), which is
parameters of a Krylov chain have different asymptotic determined by the Krylov complexity (K-complexity):
behavior in integrable and chaotic ensembles. In the limit D−1
X
of infinitely large Hilbert space, this difference implies 2
K(t) = n |ϕn (t)| , (5)
Krylov localization for integrable ensembles and Krylov n=0
delocalization for chaotic ones. This observation resem-
bles the universal growth hypothesis suggested in [16]. and its localization length, which is captured by the ex-
ponent of the Krylov entropy (K-entropy):
D−1
X 2 2
∗
S(t) = − |ϕn (t)| log |ϕn (t)| . (6)
[email protected]
n=0
 2

In a chaotic system, K(t) grows exponentially and even-

tually saturates at K(t) ∼ d, whereas Krylov wave func-
tion quickly becomes delocalized. In an integrable system,
K(t) grows slower and saturates at a much smaller value,
K(t) ≪ d, since Krylov wave function is localized and
state |ψ(t)⟩ explores only a small fraction of the Hilbert
Figure 1. Quantum circuit implementing an arbitrary unitary
space. So, Krylov delocalization serves as a signature of U in the CMV basis (11). The seeding state |ψ⟩ = |00 · · · 0⟩.
quantum chaos in continuous time dynamics [56–61].
Szegő algorithm— Now, let us generalize construc-
tion (2) to Floquet case, i.e., to a periodic Hamiltonian The algorithm that builds the CMV basis from a scratch
H(t + T ) = H(t). As long as we restrict the evolution of has the same computational complexity as Szegő and
a Floquet system to discrete times t = nT , the Krylov Lanczos algorithms:
subspace of a seeding state |ψ⟩ has the following form:
 
ρ2n−1 ᾱ2n−1 ᾱ2n
|ψ(nT )⟩ ∈ span |ψ⟩, U |ψ⟩, U 2 |ψ⟩, · · · , U D−1 |ψ⟩ , (7)

|P2n+1 ⟩ = U |P2n−1 ⟩ − U+ |P2n ⟩,
ρ2n ρ2n ρ2n
where we introduce
h R the evolution
i operator in discrete ᾱ2n+1 = ⟨P2n |U |P2n+1 ⟩ /ρ2n ,
T
time U = Texp −i 0 dt H(t) . ρ2n †

α2n † α2n+1

To perform the generalization, we recall that Krylov |P2n+2 ⟩ = U |P2n ⟩ − U + |P2n+1 ⟩,
ρ2n+1 ρ2n+1 ρ2n+1
construction (2) is closely related to the theory of orthog-
onal polynomials [43–45]. Indeed, we can map |ψ⟩ → 1, ᾱ2n+2 = ⟨P2n+2 |U |P2n+1 ⟩ /ρ2n+1 ,
H → x, and |On ⟩ = Pn (H)|ψ⟩ → Pn (x), where {Pn (x)} (11)
is a sequence of orthonormal polynomials on the real line. with initial conditions |P0 ⟩ = |ψ⟩/∥ψ∥ and ᾱ−1 = −1
Such a sequence is fixed by recurrence relations (2). Sim- and ᾱ0 = ⟨P0 |U |P0 ⟩ until ρD−1 = 0. However, the choice
ilarly, we can map |ψ⟩ → 1, U → z, Qn (U )|ψ⟩ → Qn (z), of the CMV basis puts the evolution operator U to a
where {Qn (z)} is a sequence of orthonormal polynomi- computationally convenient five-diagonal form:
als on the unit circle |z| = 1. This sequence is fixed by ⟨Pm |U |Pn ⟩ = Lmk Mkn ,
Szegő recurrence relations [41–44]. Hence, we can apply (12)
the Szegő algorithm to construct a complete orthonormal L = 1 ⊕ Θ1 ⊕ Θ3 ⊕ · · · , M = Θ0 ⊕ Θ2 ⊕ · · · ,
basis {|Φn ⟩ = Qn (U )|ψ⟩} in the Krylov subspace (7):
where the emerging one-qubit unitary operators are
ᾱn = ⟨Φ̃n |U |Φn ⟩, ρ2n = 1 − |αn |2 ,    −iχ 
ᾱn ρn e n cos θn sin θn

|Φn+1 ⟩

1

U −ᾱn

|Φn ⟩

(8) Θn = = . (13)
= , −ρn αn − sin θn eiχn cos θn
|Φ̃n+1 ⟩ ρn −αn U 1 |Φ̃n ⟩
If we assume that the CMV basis coincides with the com-
where |Φ0 ⟩ = |Φ̃0 ⟩ = |ψ⟩/∥ψ∥ and ρD−1 = 0 [62]. Com- putational basis of an L-qubit quantum circuit (d = 2L ),
plex numbers αn , |αn | ≤ 1, are called the Verblunsky evolution operator (12) is implemented using two quan-
coefficients. We emphasize that computational complex- tum multiplexors and adding operations, see Fig. 1.
ity of this algorithm is equal to complexity of the Lanc- Hence, an arbitrary quantum dynamics in a Krylov basis
zos algorithm and significantly smaller than complexity is simulated using d = 2L CNOT gates. Of course, this
of Arnoldi iteration or direct diagonalization of U . result does not contradict the theoretical bound [64, 65]
The unitary evolution operator U in the basis {|Φn ⟩} because Krylov basis depends on the seeding state |ψ⟩.
has an upper Hessenberg form: Representation (12) also implies that we can define
 Qn−1 φn (t) = Pn P0 (t) and map the dynamics of a seeding
−ᾱn αm−1 k=m ρk , m < n + 1,

state |ψ⟩ into a one-dimensional tight-binding model:
⟨Φm |U |Φn ⟩ = ρn , m = n + 1, (9)
∂t2 φn (t) = (−αn ᾱn−1 − ᾱn αn−1 − 2) φn (t)

0 m > n + 1,
+ 2 (λn+1 ∂n ᾱn−1 + λn ∂n αn−1 ) ρn−1 φn−1 (t)
where we introduce α−1 = −1 for convenience. Note that
Arnoldi iteration [47–52] applied to unitary U reproduces + 2 (λn+1 ∂n ᾱn + λn ∂n αn ) ρn φn+1 (t)
matrix (9), but cannot discern its underlying structure. + ρn−1 ρn−2 φn−2 (t) + ρn ρn+1 φn+2 (t),
We also emphasize that unitary operator (9) describes a (14)
non-local evolution on the Krylov chain. where λn = [1 − (−1)n ] /2 and initial conditions are
CMV basis— Szegő algorithm can be optimized to
produce a five-diagonal matrix representation of U . To φn (0) = δn,0 and φn (T ) = ᾱ0 δn,0 + ρ0 δn,1 . (15)
show this, we define another complete and orthonormal
basis, the Cantero—Moral—Velazquez (CMV) basis [63]: For convenience, we introduce discrete derivatives over
the chain index, ∂n αn = (αn+1 − αn−1 )/2, and time,
|P2n ⟩ = U −n |Φ̃2n ⟩, |P2n+1 ⟩ = U −n |Φ2n+1 ⟩. (10) ∂t2 φn (t) = φn (t + T ) + φn (t − T ) − 2φn (t).
 3

In the high-frequency limit T → 0, coefficients αn are (a) (b)

expressed in terms of Lanczos coefficients, so bases (8)
and (11) reduce to the standard Krylov basis (2), see Ap-
pendix A. In this limit, model (14) reproduces model (4).
We also emphasize that the tight-binding model (14)
reproduces the dynamics of the boundary operator in
a non-interacting XY model, see Appendix B. In other
words, model (14) generalizes the approach of [53, 54]
from vectorized operators to arbitrary state dynamics.
Signatures of chaos and integrability— Inspired
by eigenvalue properties of matrix ensembles [66–70],
we conjecture that any quantum Floquet system with
Hilbert space dimension d generates one of the following
distributions of Verblunsky coefficients:
(i ) Degenerate case: |αn |2 < A/(d − n)1+ϵ , Figure 2. Krylov localization length averaged over time, ini-
tial position, and ensemble realization for different chaotic (a)
(ii ) Chaotic case: |αn |2 = A/ [β(d − n)/2 + 1], and integrable (b) ensembles (16). In all cases, K ∼ eS .
(iii ) Integrable case: |αn |2 > A/(d − n)1−ϵ ,
where ϵ > 0, β > 0, and constant A > 0 is chosen such So, we expect that Krylov wave function is delocalized
that all coefficients lie inside the unit circle. in ensembles (i ) and (ii ), eS ∼ d, but localized in inte-
To illustrate this classification, we consider random grable ensembles (iii ), eS ∼ d1−ϵ ≪ d. Such a localiza-
Verblunsky coefficients independently distributed on the tion was also proved rigorously for selected distributions,
unit disk with the following probability density functions: e.g., uniform distribution [78–80]. Numerical calculations
2 [(d−n)
(1+ϵ)
−1] confirm this qualitative analysis, see Fig. 2.
p(i)


n (αn ) ∼ 1 − |αn | , (16a) Another way to see localization-delocalization transi-
(ii) 2 [β(d−n)/2−1] tion is to notice that the tight-binding model (14) reduces


pn (αn ) ∼ 1 − |αn | , (16b)
(1−ϵ)
to a massless Klein-Gordon equation when |αn |2 ≪ 1:
2 [(d−n) −1]
p(iii)


n (αn ) ∼ 1 − |αn | . (16c)
∂t2 φn (t) ≈ 4∂n ∂n φn (t) + O(αn2 ), (18)
First of all, we discuss the eigenvalue distribution of
CMV matrices constructed using random Verblunsky co- where O(αn2 ) is a small noise term. Solutions to an un-
efficients. It was shown [66] that these distributions con- perturbed Eq. (18) with initially localized Krylov wave
verge to the clock ensemble with rigid level spacings (i ), function φn (0) = δn,0 are given by a sum of localized
circular β-ensemble (ii ), or Poisson ensemble (iii ), re- wave packets. In a degenerate ensemble (i ), wave pack-
spectively [71]. The mean level spacing ratio [72, 73] ets remain localized until they reach the far end of the
changes in the range 0.386 < ⟨r⟩ < 1 and grows with β Krylov chain. So, they explore the whole Hilbert space,
in ensemble (ii ), whereas ensembles (i ) and (iii ) saturate but remain localized up to times t ≫ d. In a beta ensem-
the upper and lower bounds of this inequality. In other ble (ii ), the noise term smoothes the propagating wave
words, energy levels are independently distributed in in- packet, so its localization length reaches eS ∼ d as soon
tegrable ensemble (iii ) and repel each other in chaotic as its center approaches the far half of the Krylov chain.
ensemble (ii ), which reproduces the conventional clas- In a Poisson ensemble (iii), the noise term is so strong
sification of ergodic systems [74–77]. In ensemble (i ), that the wave packet cannot propagate at all, see Fig. 3.
energy level repulsion is so strong that the system be-
comes integrable but preserves some features of chaotic
systems. (a)
Furthermore, classification (i )–(iii ) reflects the dy-
namical properties of chaotic and integrable ensembles,
i.e., implies Krylov localization in the integrable case and
delocalization in the chaotic and degenerate cases. Let us (b)
show this using the same example with random Verblun-
sky coefficients (16). To that end, we write down the
eigenvalue equation in the Krylov or CMV basis, average
it over the corresponding ensembles, and estimate the
probability distribution of an eigenvector φn (ω):

|φn (ω)|2
n−1
Y 1,
 (i), Figure 3. Ensemble-averaged localization length (a) and po-
2 −1/2β
∼ ρ k ∼ (d − n) , (ii), (17) sition (b) of an initially localized wave packet propagating on
|φ0 (ω)|2 exp −n/d1−ϵ
, (iii).
 Krylov chains with random Verblunsky coefficients (16).
k=0
 4

(a) (b) (a) (b)

Figure 4. (a) Distribution of Verblunsky coefficients in the Figure 5. (a) Distribution of Verblunsky coefficients in the
integrable and chaotic kicked top (J = 1000). (b) Localiza- integrable and chaotic kicked Ising chain (L = 13). (b) Local-
tion length of a seeding coherent state (circles) and average ization length of a seeding state averaged over random seeding
localization length of all Krylov basis elements (triangles). states from different symmetry sectors.

So, the behavior of Verblunsky coefficients serves as area A√ 0 ∼ ℏ [88]. A coherent state |Ω⟩ sweeps the area
a signature of quantum chaos and integrability similarly A ∼ L ℏ when Ω belongs to a classical trajectory
√ of fi-
to behavior of Lanczos coefficients [16]. In particular, in nite length L, so it spreads over D ∼ L/ ℏ eigenstates.
non-degenerate quantum systems with ⟨r⟩ < 1, the slow- Therefore, K-complexity and √ K-entropy of such a state
est possible growth of Verblunsky coefficients is achieved are proportional to K ∼ eS ∼ d ≪ d, while Verblunsky
for chaotic systems. In what follows, we will give several coefficients follow an integrable distribution (iii ). Nu-
examples where such kinds of behavior indeed emerge. merical simulations support this reasoning (Fig. 4).
Semiclassical Krylov localization— We start with In the chaotic regime, classical evolution brings any
the kicked top model, a paradigmatic example of a quan- point arbitrarily close to any other point. So, an ini-
tum Floquet system [74, 75, 81–84] with remarkable ex- tially localized classical distribution uniformly spreads
perimental realizations [85, 86]: over the phase space, while the corresponding quantum
 κ   κ  state uniformly covers the Hilbert space [88]. Conversely,
x z we can switch to a phase-space-local basis (e.g., basis of
Utop = exp −i Jx2 exp −i Jz2 exp[−i (b · J)] .
2J 2J Jz eigenstates) and think of Utop as a random unitary
(19) transformation of complex vectors written in this basis.
Here, J = (Jx , Jy , Jz ), [Jk , Jl ] = iϵklm Jm is the angular Applying Szegő algorithm to such an evolution, we obtain
momentum operator. The square J2 = J(J + 1) is con- that Verblunsky coefficients follow the chaotic distribu-
served, so we restrict our discussion to the d = 2J + 1 tion (ii ) with β = 2. This implies delocalization of the
dimensional Hilbert space. We also emphasize that Utop Krylov wave function, K ∼ eS ∼ d, see Fig. 4.
has reflection symmetry, so we further restrict its dynam- Spin chains— To show that similar distributions of
ics to one of the invariant sectors. Verblunsky coefficients emerge far from the semiclassical
In the semiclassical limit ℏ = 1/J → 0, Heisenberg evo- limit, we consider the kicked Ising chain [89–94]:
lution (19) reproduces the dynamics of a classical unit
spin j = J/J. with Poisson bracket {jk , jl } = ϵklm jm . L−1
X
! " L−1
X
#
z z
This dynamics can be regular, chaotic, or mixed depend- UKI = exp −iJ σk σk+1 exp −i (b · σ k ) ,
ing on the parameters of the model. We set κz = 0, k=0 k=0
b = (1.7, 0, 0) to study regular (integrable) regime and (20)
κz = 0.5, b = (0, 1.7, 0) to study chaotic regime [81, 82]. with σkx,y,z being the Pauli matrices on site k. We assume
In both cases, we generate an ensemble of evolution op- periodic boundary conditions, σL = σ0 , and restrict the
erators (19) by varying one of the angular momenta in dynamics to one of translation-invariant sectors.
the range 10 ≤ κx ≤ 12 with the step δκx = 0.05 [87]. To generate different ensembles of UKI , we fix the ex-
We take a spin coherent state |Ω⟩ as a natural seeding change constant J = 0.7 and vary the magnitude of the
state for the Krylov construction. magnetic field in the range 1.13 ≤ |b| ≤ 1.41 with the
In the regular regime, classical phase space breaks into step δb = 0.007. We fix the direction b/|b| = √ (1, 0, 0)√to
periodic trajectories. So, any initially localized distri- study integrable behavior and b/|b| = (1/ 2, 0, 1/ 2)
bution evolves in a region of area A sandwiched be- to study chaotic behavior [89, 90]. We choose a random
tween two adjacent trajectories. Semiclassical quanti- superposition of several computational basis states as a
zation maps periodic trajectories to eigenstates of Utop , natural low-entanglement seeding state.
each of which is concentrated in a phase space region of Similarly to the kicked top, kicked Ising chain produces
 5

0.0010 of the system and initial state. For example, local Pauli
Chaotic operators in a dual-unitary circuit on L sites generate
Dual-unitary zero Verblunsky coefficients up to n ∼ L, but reproduce
general chaotic evolution with |αn |2 ∼ 1/d ∼ 1/4L at
|αn |

0.0005
larger times L ≪ n ≪ d, see Fig. 6.
Conclusion— We suggested a new approach to
Krylov basis construction in quantum Floquet systems.
0.0000 Our algorithm works substantially faster than other ex-
0 20 40 60 80 100 isting approaches [49–54] and brings the unitary evolu-
tion operator to a five-diagonal form, which is extremely
n
convenient for simulations on both classical and quantum
computers. So, our algorithm streamlines numerical ex-
Figure 6. Verblunsky coefficients of a local operator in dual- periments and allows them to tackle larger Hilbert space
unitary and general chaotic Ising model (20) with L = 14.
dimensions for a given quantum system. Furthermore, it
provides a toolkit for engineering quantum Floquet sys-
tems with desired quantum dynamics, which is especially
Verblunsky coefficients with distribution (iii ) in the inte- useful for simulations on small quantum computers.
grable case and distribution (ii ) in the chaotic case. So, We also suggested a classification of ergodic Floquet
transition from chaos to integrability is again marked by systems based on the asymptotic behavior of Verblunsky
Krylov localization of low-entanglement states, see Fig. 5. coefficients. This classification reproduces the conven-
In other words, integrable evolution drags such states tional classification based on energy level repulsion, but
through small periodic cycles, whereas chaotic evolution provides additional insights into state dynamics. To illus-
explores the Hilbert space uniformly. trate our conjecture, we obtained the proposed distribu-
Convergence to maximally ergodic regime— It tions of Verblunsky coefficients for simple seeding states
was recently conjectured that Krylov basis of a generic in random matrix ensembles, kicked top, and kicked Ising
chaotic Floquet system converges to “maximally ergodic” chain. We emphasize that in general, distribution of
states with vanishing autocorrelation functions [51–54]. Verblunsky coefficients and Krylov localization length de-
Our observations contradict this conjecture. Indeed, such pends on the initial state [95]; the only exceptions are
a convergence implies that Verblunsky coefficients ap- chaotic circular ensembles with β = 1, 2, 4, which gener-
proach |αn |2 → 0 for n ≫ 1, i.e., belong to case (i ) ate distribution (ii ) for an arbitrary initial state. We will
from our classification. This, in turn, automatically im- study this dependence in the future work.
plies that the spectrum of evolution operator is rigid with There are many other directions for future research.
⟨r⟩ ≈ 1, which is not true for a generic chaotic system. First, we expect that the behavior of Veblunsky coef-
In fact, Verblunsky coefficients of a chaotic system are ficients might indicate the emergence of quantum scars
small but non-zero: |αn |2 ∼ 1/d for 1 ≪ n ≪ d, where d and Hilbert space fragmentation [7–12]. Second, our ap-
is the dimension of the Hilbert space. Although this dif- proach provides an alternative perspective for thermal-
ference seems innocuous in the limit d → ∞, it substan- ization and information spreading in random and dual-
tially affects state dynamics at large evolution times. In unitary quantum circuits [96–102]. Finally, our numerical
particular, it accelerates the spread of initially localized algorithm can be applied for simulations of such remark-
distribution in the Krylov space (Fig. 3). So, we suggest able driven systems as discrete time crystals [103–107]
to use asymptotic distribution (ii ) instead of maximally and Floquet topological insulators [108, 109].
ergodic bath approximation (i ) to improve the accuracy Acknowledgments— We thank David A. Huse,
of numerical simulations in chaotic Floquet systems. Alexander Gorsky, Anatoly Dymarsky, Dmitry Abanin,
However, we emphasize that the early-time behavior Philippe Suchsland, and Elizaveta Trunina for valuable
of Verblunsky coefficients is determined by the features discussions.

[1] J. M. Deutsch, Quantum statistical mechanics in a ics, Ann. Rev. Condensed Matter Phys. 6, 15 (2015),
closed system, Phys. Rev. A 43, 2046 (1991). arXiv:1404.0686.
[2] M. Srednicki, Chaos and quantum thermalization, Phys. [5] E. Altman and R. Vosk, Universal dynamics and
Rev. E 50, 888 (1994), arXiv:cond-mat/9403051. renormalization in many body localized systems,
[3] L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, Ann. Rev. Condensed Matter Phys. 6, 383 (2015),
From quantum chaos and eigenstate thermalization to arXiv:1408.2834.
statistical mechanics and thermodynamics, Adv. Phys. [6] D. A. Abanin, E. Altman, I. Bloch, and M. Serbyn,
65, 239 (2016), arXiv:1509.06411. Colloquium : Many-body localization, thermalization,
[4] R. Nandkishore and D. A. Huse, Many body localiza- and entanglement, Rev. Mod. Phys. 91, 021001 (2019),
tion and thermalization in quantum statistical mechan- arXiv:1804.11065.
 6

[7] M. Serbyn, D. A. Abanin, and Z. Papić, Quantum many- Rev. Lett. 124, 206803 (2020), arXiv:2002.00098.
body scars and weak breaking of ergodicity, Nature [25] P. Caputa and S. Liu, Quantum complexity and topo-
Phys. 17, 675 (2021), arXiv:2011.09486. logical phases of matter, Phys. Rev. B 106, 195125
[8] A. Chandran, T. Iadecola, V. Khemani, and R. Moess- (2022), arXiv:2205.05688.
ner, Quantum many-body scars: A quasiparticle per- [26] P. Caputa, N. Gupta, S. S. Haque, S. Liu, J. Murugan,
spective, Ann. Rev. Condensed Matter Phys. 14, 443 and H. J. R. Van Zyl, Spread complexity and topological
(2023), arXiv:2206.11528. transitions in the Kitaev chain, J. High Energy Phys.
[9] S. Moudgalya, B. A. Bernevig, and N. Regnault, Quan- 2023 (01), 120, arXiv:2208.06311.
tum many-body scars and Hilbert space fragmentation: [27] B. Bhattacharjee, S. Sur, and P. Nandy, Probing quan-
a review of exact results, Rept. Prog. Phys. 85, 086501 tum scars and weak ergodicity breaking through quan-
(2022), arXiv:2109.00548. tum complexity, Phys. Rev. B 106, 205150 (2022),
[10] P. Sala, T. Rakovszky, R. Verresen, M. Knap, and arXiv:2208.05503.
F. Pollmann, Ergodicity-breaking arising from Hilbert [28] B. Bhattacharjee, X. Cao, P. Nandy, and T. Pathak,
space fragmentation in dipole-conserving Hamiltonians, Krylov complexity in saddle-dominated scrambling, J.
Phys. Rev. X 10, 011047 (2020), arXiv:1904.04266. High Energy Phys. 2022 (05), 174, arXiv:2203.03534.
[11] V. Khemani, M. Hermele, and R. Nandkishore, Local- [29] K. Hashimoto, K. Murata, N. Tanahashi, and R. Watan-
ization from Hilbert space shattering: From theory to abe, Krylov complexity and chaos in quantum me-
physical realizations, Phys. Rev. B 101, 174204 (2020), chanics, J. High Energy Phys. 2023 (11), 040,
arXiv:1904.04815. arXiv:2305.16669.
[12] S. Moudgalya and O. I. Motrunich, Hilbert space frag- [30] A. Dymarsky and M. Smolkin, Krylov complexity in
mentation and commutant algebras, Phys. Rev. X 12, conformal field theory, Phys. Rev. D 104, L081702
011050 (2022), arXiv:2108.10324. (2021), arXiv:2104.09514.
[13] J. Maldacena, S. H. Shenker, and D. Stanford, A [31] A. Avdoshkin, A. Dymarsky, and M. Smolkin, Krylov
bound on chaos, J. High Energy Phys. 2016 (08), 106, complexity in quantum field theory, and beyond, J. High
arXiv:1503.01409. Energy Phys. 2024 (06), 066, arXiv:2212.14429.
[14] B. Swingle, Unscrambling the physics of out-of-time- [32] J. L. F. Barbón, E. Rabinovici, R. Shir, and R. Sinha,
order correlators, Nature Phys. 14, 988 (2018). On The evolution of operator complexity beyond
[15] S. Xu and B. Swingle, Scrambling dynamics and scrambling, J. High Energy Phys. 2019 (10), 264,
out-of-time-ordered correlators in quantum many- arXiv:1907.05393.
body systems, PRX Quantum 5, 010201 (2024), [33] E. Rabinovici, A. Sánchez-Garrido, R. Shir, and J. Son-
arXiv:2202.07060. ner, Operator complexity: a journey to the edge of
[16] D. E. Parker, X. Cao, A. Avdoshkin, T. Scaffidi, and Krylov space, J. High Energy Phys. 2021 (06), 062,
E. Altman, A universal operator growth hypothesis, arXiv:2009.01862.
Phys. Rev. X 9, 041017 (2019), arXiv:1812.08657. [34] E. Rabinovici, A. Sánchez-Garrido, R. Shir, and J. Son-
[17] P. Nandy, A. S. Matsoukas-Roubeas, P. Martínez- ner, Krylov complexity from integrability to chaos, J.
Azcona, A. Dymarsky, and A. del Campo, Quantum High Energy Phys. 2022 (07), 151, arXiv:2207.07701.
dynamics in Krylov space: methods and applications, [35] E. Rabinovici, A. Sánchez-Garrido, R. Shir, and J. Son-
arXiv:2405.09628. ner, A bulk manifestation of Krylov complexity, J. High
[18] V. S. Viswanath and G. Müller, The Recursion Method: Energy Phys. 2023 (08), 213, arXiv:2305.04355.
Applications to Many-Body Dynamics, Lecture Notes in [36] S.-K. Jian, B. Swingle, and Z.-Y. Xian, Complexity
Physics Monographs (Springer-Verlag, Berlin, 1994). growth of operators in the SYK model and in JT gravity,
[19] M. Motta, C. Sun, A. T. K. Tan, M. J. O. Rourke, E. Ye, J. High Energy Phys. 2021 (03), 014, arXiv:2008.12274.
A. J. Minnich, F. G. S. L. Brandão, and G. K.-L. Chan, [37] N. Iizuka and M. Nishida, Krylov complexity in the IP
Determining eigenstates and thermal states on a quan- matrix model, J. High Energy Phys. 2023 (11), 065,
tum computer using quantum imaginary time evolution, arXiv:2306.04805.
Nature Phys. 16, 205 (2019), arXiv:1901.07653. [38] A. Avdoshkin and A. Dymarsky, Euclidean operator
[20] N. H. Stair, R. Huang, and F. A. Evangelista, A mul- growth and quantum chaos, Phys. Rev. Res. 2, 043234
tireference quantum Krylov algorithm for strongly cor- (2020), arXiv:1911.09672.
related electrons, J. Chem. Theory Comput. 16, 2236 [39] P. Caputa, J. M. Magan, and D. Patramanis, Geometry
(2020), arXiv:1911.05163. of Krylov complexity, Phys. Rev. Res. 4, 013041 (2022),
[21] C. L. Cortes and S. K. Gray, Quantum Krylov sub- arXiv:2109.03824.
space algorithms for ground- and excited-state energy [40] A. Gorsky, S. Nechaev, and A. Valov, KPZ scaling from
estimation, Phys. Rev. A 105 105, 022417 (2022), the Krylov space, J. High Energy Phys. 2024 (09), 021,
arXiv:2109.06868. arXiv:2406.02782.
[22] M. Larocca and D. Wisniacki, Krylov-subspace ap- [41] G. Szegő, Orthogonal Polynomials, Colloquium Publi-
proach for the efficient control of quantum many- cations, Vol. 23 (American Mathematical Society, Prov-
body dynamics, Phys. Rev. A 105 103, 023107 (2021), idence, RI, 1939).
arXiv:2010.03598. [42] B. Simon, Orthogonal Polynomials on the Unit Circle,
[23] K. Takahashi and A. del Campo, Shortcuts to adiabatic- Colloquium Publications, Vol. 54 (American Mathemat-
ity in Krylov space, Phys. Rev. X 14, 011032 (2024), ical Society, Providence, RI, 2005).
arXiv:2302.05460. [43] M. E. H. Ismail, Classical and Quantum Orthogonal
[24] D. J. Yates, A. G. Abanov, and A. Mitra, Lifetime of al- Polynomials in One Variable, Encyclopedia of Math-
most strong edge-mode operators in one dimensional, in- ematics and its Applications (Cambridge University
teracting, symmetry protected topological phases, Phys. Press, Cambridge, 2005).
 7

[44] M. E. H. Ismail and W. Van Assche, Univariate Orthog- [63] M. J. Cantero, L. Moral, and L. Velazquez, Five-
onal Polynomials, Encyclopedia of Special Functions: diagonal matrices and zeros of orthogonal polynomials
The Askey-Bateman Project, Vol 1. (Cambridge Uni- on the unit circle, Linear Algebra Appl. 362, 29 (2003),
versity Press, Cambridge, 2020). arXiv:math/0204300.
[45] W. Mück and Y. Yang, Krylov complexity and orthog- [64] V. V. Shende, I. L. Markov, and S. S. Bullock,
onal polynomials, Nucl. Phys. B 984, 115948 (2022), Minimal universal two-qubit controlled-NOT-based cir-
arXiv:2205.12815. cuits, Phys. Rev. A 69, 062321 (2004), arXiv:quant-
[46] D. J. Yates, A. G. Abanov, and A. Mitra, Long-lived ph/0308033.
period-doubled edge modes of interacting and disorder- [65] V. V. Shende, S. S. Bullock, and I. L. Markov, Syn-
free Floquet spin chains, Commun. Phys. 5, 43 (2022), thesis of quantum-logic circuits, IEEE Trans. Comput.
arXiv:2105.13766. Aided Design Integr. Circuits Syst. 25, 1000 (2006),
[47] D. J. Yates and A. Mitra, Strong and almost strong arXiv:quant-ph/0406176.
modes of Floquet spin chains in Krylov subspaces, Phys. [66] R. Killip and M. Stoiciu, Eigenvalue statistics for CMV
Rev. B 104, 195121 (2021), arXiv:2105.13246. matrices: From Poisson to clock via random matrix en-
[48] H.-Y. Qi, Y. Wu, and W. Zheng, Topological origin sembles, Duke Math. J. 146, 361 (2009), arXiv:math-
of Floquet thermalization in periodically driven many- ph/0608002.
body systems, arXiv:2404.18052. [67] R. Killip and I. Nenciu, Matrix models for circular
[49] A. A. Nizami and A. W. Shrestha, Krylov construction ensembles, Int. Math. Res. Not. 2004, 2665 (2004),
and complexity for driven quantum systems, Phys. Rev. arXiv:math/0410034.
E 108, 054222 (2023), arXiv:2305.00256. [68] I. Dumitriu and A. Edelman, Matrix models for beta
[50] A. A. Nizami and A. W. Shrestha, Spread complexity ensembles, J. Math. Phys. 43, 5830 (2002), arXiv:math-
and quantum chaos for periodically driven spin chains, ph/0206043.
Phys. Rev. E 110, 034201 (2024), arXiv:2405.16182. [69] A. K. Das and A. Ghosh, Nonergodic extended states
[51] P. Suchsland, R. Moessner, and P. W. Claeys, Krylov in the β ensemble, Phys. Rev. E 105, 054121 (2022),
complexity and Trotter transitions in unitary circuit dy- arXiv:2112.11910.
namics, arXiv:2308.03851. [70] V. Balasubramanian, J. M. Magan, and Q. Wu, Tridiag-
[52] G. F. Scialchi, A. J. Roncaglia, C. Pineda, and D. A. onalizing random matrices, Phys. Rev. D 107, 126001
Wisniacki, Exploring quantum ergodicity of unitary evo- (2023), arXiv:2208.08452.
lution through the Krylov approach, arXiv:2407.06428. [71] Circular beta ensemble includes the orthogonal (β = 1),
[53] H.-C. Yeh and A. Mitra, Universal model of Floquet op- unitary (β = 2), and symplectic (β = 4) ensembles.
erator Krylov space, Phys. Rev. B 110, 155109 (2024), [72] V. Oganesyan and D. A. Huse, Localization of inter-
arXiv:2311.15116. acting fermions at high temperature, Phys. Rev. B 75,
[54] H.-C. Yeh and A. Mitra, Moment method and continued 155111 (2007), arXiv:cond-mat/0610854.
fraction expansion in Floquet Operator Krylov Space, [73] Y. Y. Atas, E. Bogomolny, O. Giraud, and G. Roux,
arXiv:2410.15223. Distribution of the ratio of consecutive level spacings in
[55] We can also map any operator O = di,j=1 Oij |ψi ⟩⟨ψj |
P
random matrix ensembles, Phys. Rev. Lett. 110, 084101
Pd
to a state |O⟩ = i,j=1 Oij |ψi ⟩⊗|ψj ⟩, i.e., to an element (2013), arXiv:1212.5611.
[74] F. Haake, Quantum Signatures of Chaos (Springer-
of d2 -dimensional vector
space endowed with an inner
Verlag, Berlin Heidelberg, 2010).
product (A|B) = tr A† B /d. Time evolution of such a
[75] H.-J. Stöckmann, Quantum Chaos: An Introduction
vectorized operator is generated by L = H ⊗ 1 − 1 ⊗ H.
(Cambridge University Press, Cambridge, England,
[56] A. Dymarsky and A. Gorsky, Quantum chaos as delo-
1999).
calization in Krylov space, Phys. Rev. B 102, 085137
[76] M. V. Berry and M. Tabor, Level clustering in the reg-
(2020), arXiv:1912.12227.
ular spectrum, Proc. R. Soc. Lond. A 356, 375 (1977).
[57] F. B. Trigueros and C.-J. Lin, Krylov complex-
[77] O. Bohigas, M. J. Giannoni, and C. Schmit, Character-
ity of many-body localization: Operator localiza-
ization of chaotic quantum spectra and universality of
tion in Krylov basis, SciPost Phys. 13, 037 (2022),
level fluctuation laws, Phys. Rev. Lett. 52, 1 (1984).
arXiv:2112.04722.
[78] A. V. Teplyaev, The pure point spectrum of random
[58] E. Rabinovici, A. Sánchez-Garrido, R. Shir, and J. Son-
orthogonal polynomials on the circle, Dokl. Akad. Nauk
ner, Krylov localization and suppression of complexity,
320, 49 (1991), [Dokl. Math. 44, 407 (1992)].
J. High Energy Phys. 2022 (03), 211, arXiv:2112.12128.
[79] C. Cedzich and A. H. Werner, Anderson localiza-
[59] H. G. Menzler and R. Jha, Krylov delocaliza-
tion for electric quantum walks and skew-shift CMV
tion/localization across ergodicity breaking, Phys. Rev.
matrices, Commun. Math. Phys. 387, 1257 (2021),
B 110, 125137 (2024), arXiv:2403.14384.
arXiv:1906.11931.
[60] Y. A. Alaoui and B. Laburthe-Tolra, Method to discrim-
[80] X. Zhu, Localization for random CMV matrices, J. Ap-
inate between localized and chaotic quantum systems,
prox. Theory 298, 106008 (2024), arXiv:2110.11386.
Phys. Rev. Res. 6, 043045 (2024), arXiv:2307.10706.
[81] F. Haake, M. Kuś, and R. Scharf, Classical and quantum
[61] M. Alishahiha and M. J. Vasli, Thermalization in Krylov
chaos for a kicked top, Z. Physik B 65, 381 (1987).
basis, arXiv:2403.06655.
[82] F. M. Izrailev, Simple models of quantum chaos: Spec-
[62] Note that the set of states |Φ̃n ⟩ is normalized but not
trum and eigenfunctions, Phys. Rept. 196, 299 (1990).
orthogonal: ⟨Φ̃n |Φ̃n ⟩ = 1 but ⟨Φ̃m |Φ̃n ⟩ ̸= 0 for m ̸= n. [83] E. B. Rozenbaum, S. Ganeshan, and V. Galitski, Lya-
However, these states are orthogonal to Krylov basis punov exponent and out-of-time-ordered correlator’s
states generated after step n: ⟨Φm |Φ̃n ⟩ = 0 for m > n. growth rate in a chaotic system, Phys. Rev. Lett. 118,
Besides, ⟨Φ̃m |U |Φ̃n ⟩ = 0 and ⟨Φ̃m |U |Φn ⟩ = 0 for m > n. 086801 (2017), arXiv:1609.01707.
 8

[84] E. Vallini and S. Pappalardi, Long-time freeness in the 109, 160401 (2012), arXiv:1202.2539.
kicked top, arXiv:2411.12050. [104] V. Khemani, A. Lazarides, R. Moessner, and S. L.
[85] S. Chaudhury, A. Smith, B. E. Anderson, S. Ghose, and Sondhi, Phase structure of driven quantum systems,
P. S. Jessen, Quantum signatures of chaos in a kicked Phys. Rev. Lett. 116, 250401 (2016), arXiv:1508.03344.
top, Nature 461, 768 (2009). [105] D. V. Else, B. Bauer, and C. Nayak, Floquet
[86] C. Neill et al., Ergodic dynamics and thermalization time crystals, Phys. Rev. Lett. 117, 090402 (2016),
in an isolated quantum system, Nature Phys. 12, 1037 arXiv:1603.08001.
(2016), arXiv:1601.00600. [106] S. Choi et al., Observation of discrete time-crystalline
[87] This approach allows us to get cleaner results for lower order in a disordered dipolar many-body system, Nature
dimensions of the Hilbert space. However, all discussed 543, 221 (2017), arXiv:1610.08057.
features are visible even without ensemble averaging for [107] J. Zhang et al., Observation of a discrete time crystal,
dimensions d ∼ 103 and larger. Nature 543, 217 (2017), arXiv:1609.08684.
[88] M. V. Berry, Semi-classical mechanics in phase space: [108] H. C. Po, L. Fidkowski, T. Morimoto, A. C. Potter,
A study of Wigner’s function, Philos. Trans. R. Soc. A and A. Vishwanath, Chiral Floquet phases of many-
287, 237 (1977). body localized bosons, Phys. Rev. X 6, 041070 (2016),
[89] T. Prosen, Exact time-correlation functions of quantum arXiv:1609.00006.
Ising chain in a kicking transversal magnetic field: spec- [109] F. Nathan, D. Abanin, E. Berg, N. H. Lindner, and
tral analysis of the adjoint propagator in Heisenberg pic- M. S. Rudner, Anomalous Floquet insulators, Phys.
ture, Prog. Theor. Phys. Suppl. 139, 191 (2000). Rev. B 99, 195133 (2019), arXiv:1712.02789.
[90] T. Prosen, On general relation between quantum ergod- [110] M. Derevyagin, L. Vinet, and A. Zhedanov, CMV ma-
icity and fidelity of quantum dynamics, Phys. Rev. E 65, trices and little and big −1 Jacobi polynomials, Constr.
036208 (2002), arXiv:quant-ph/0106149. Approx. 36, 513 (2012), arXiv:1108.3535.
[91] C. Pineda and T. Prosen, Non-universal level statis- [111] M. J. Cantero, F. Marcellán, L. Moral, and L. Ve-
tics in a chaotic quantum spin chain, Phys. Rev. E 76, lazquez, A CMV connection between orthogonal poly-
061127 (2007), arXiv:quant-ph/0702164. nomials on the unit circle and the real line, J. Approx.
[92] B. Bertini, P. Kos, and T. Prosen, Exact spectral form Theory 266, 105579 (2021), arXiv:2005.09772.
factor in a minimal model of many-body quantum chaos,
Phys. Rev. Lett. 121, 264101 (2018), arXiv:1805.00931.
[93] B. Bertini, P. Kos, and T. Prosen, Entanglement spread-
ing in a minimal model of maximal many-body quantum Appendix A: Reduction from Szegő to Lanczos
chaos, Phys. Rev. X 9, 021033 (2019), arXiv:1812.05090.
[94] M. Akila, D. Waltner, B. Gutkin, and T. Guhr, Particle- First of all, let us show that Krylov bases (8) and (11)
time duality in the kicked Ising spin chain, J. Phys. A: reduce to the standard Krylov basis (2) in the high-
Math. Theor. 49, 375101 (2016), arXiv:1602.07130. frequency limit T → 0. Exploiting the theory of or-
[95] The simplest example is quantized classical system with
thogonal polynomials, we rewrite the n-th element of the
a mixed phase space, where regular and chaotic regions
coexist. In such a system, coherent states from chaotic
basis (8) in the determinant form:
regions produce distributions (ii), whereas “regular” co-
c0 c1 ··· cn
 
herent states produce distributions (iii).
[96] M. P. A. Fisher, V. Khemani, A. Nahum, and S. Vi-  c−1 c0 ··· cn−1 
jay, Random quantum circuits, Ann. Rev. Condensed 1  .. 
|Φn ⟩ = p det 
 ··· ··· . ··· ,

Matter Phys. 14, 335 (2023), arXiv:2207.14280. ∆n ∆n−1
[97] V. Khemani, A. Vishwanath, and D. A. Huse, Opera-
c
1−n c 2−n ··· c 
1
tor spreading and the emergence of dissipation in uni- |ψ⟩ U |ψ⟩ · · · U n |ψ⟩
tary dynamics with conservation laws, Phys. Rev. X 8, (A1)
031057 (2018), arXiv:1710.09835. where we introduce the Toeplitz matrix:
[98] F. Arute et al., Quantum supremacy using a pro-
grammable superconducting processor, Nature 574, 505
 
c0 c1 · · · cn
(2019), arXiv:1910.11333.  c−1 c0 · · · cn−1 
[99] B. Bertini, P. Kos, and T. Prosen, Exact correla- ∆n = det 

,

(A2)
..
tion functions for dual-unitary lattice models in 1+1 ··· ··· . ··· 
dimensions, Phys. Rev. Lett. 123, 210601 (2019), c−n c1−n · · · c0
arXiv:1904.02140.
[100] F. Fritzsch and T. Prosen, Eigenstate thermaliza-
tion in dual-unitary quantum circuits: Asymptotics of
and coefficients cn = ⟨ψ|U n |ψ⟩. Elements of the standard
spectral functions, Phys. Rev. E 103, 062133 (2021), basis (2) are represented in a similar determinant form:
arXiv:2103.11694.
µ0 µ1 ··· µn
 
[101] S. Gopalakrishnan and A. Lamacraft, Unitary circuits of
finite depth and infinite width from quantum channels,  µ1 µ2 ··· µn+1 
Phys. Rev. B 100, 064309 (2019), arXiv:1903.11611. 1  .. 
|Ψn ⟩ = p det 
 ··· ··· . ··· ,

[102] A. Lerose, M. Sonner, and D. A. Abanin, Influence ma- Dn Dn−1
trix approach to many-body Floquet dynamics, Phys.
µ
n−1 µn−2 · · · µ2n−1 
Rev. X 11, 021040 (2021), arXiv:2009.10105. |ψ⟩ H|ψ⟩ · · · H n |ψ⟩
[103] F. Wilczek, Quantum time crystals, Phys. Rev. Lett. (A3)
 9

with the Hankel matrix: This construction is especially useful for a Hermitian
  seeding operator O† = O, which naturally produces real
µ0 µ1 · · · µn coefficients αn . In particular, correspondence (A12) al-
 µ1 µ2 · · · µn+1  lows us to use the most efficient algorithms, such as Par-
Dn = det  , (A4)
 
.. tial Re-Orthogonalization [33], and enlarge the accessi-
· · · · · · . ··· 
µn µn+1 · · · µ2n ble size of operator space by several orders of magnitude
compared to straightforward Arnoldi iteration. The ob-
and coefficients µn = ⟨ψ|H n |O⟩. In the limit T → 0, vious disadvantage is the need to calculate U 1/2 first. To
we can approximate U = exp(−iT H), where the instan- bypass this disadvantage, one may consider the evolution
taneous Hamiltonian is H = i(∂U/∂T )T =0 . Hence, the in doubled discrete time generated by U 2 . Since chaotic
leading approximation to coefficients cn is directly ex- systems remain chaotic and integrable systems remain in-
pressed through µn : tegrable no matter what time scale we choose, we expect
that the change U → U 2 does not qualitatively affect
∞
X (−inT )k operator dynamics and coefficients αn . In fact, we nu-
−inT H
cn ≈ ⟨ψ|e |ψ⟩ = µk . (A5) merically confirmed for several models that K-complexity
k!
k=0 and K-entropy in CMV bases generated by U and U 2 co-
incide up to insignificant O(1) factors.
In the same limit, Toeplitz matrices reduce to Hankel
matrices:
2
∆n ≈ T n +n
Dn , (A6)
Appendix B: Inhomogeneous XY model
so Krylov bases (8) and (11) coincide with Krylov ba-
sis (2) up to phase and higher-order corrections in T : An arbitrary unitary dynamics in Krylov space (11) of
n dimension d is mapped to an inhomogeneous XY model
|Pn ⟩ ≈ |Φn ⟩ ≈ (−i) |Ψn ⟩. (A7) on d sites:
Similarly, we prove that Verblunsky coefficients αn have
d/2−1 d/2−1
the following approximate expansion in terms of Lanczos Y Y
coefficients: UXY = U2k−1,2k U2k,2k+1 ,
k=1 k=0
T2 2
  iθk
iχk z z x x y y iχk z z
n 2

(σk −σk+1 ) (σk σk+1 +σk σk+1 ) (σk −σk+1 )
αn ≈ (−1) 1 + iT An − b − An + · · · , (A8) Uk,k+1 = e 4 e 2 , e 4
2 n+1 (B1)
Pn where χk ∈ [−π, π), θk ∈ [0, π/2] and d is even. This
where we denote An = k=0 ak for brevity. We also note
model is effectively non-interacting. In fact, unitary (B1)
that the leading approximation to ρn does not depend on
is bilinear in terms of Jordan-Wigner creation and anni-
the diagonal Lanczos coefficients: ρn ≈ T |bn+1 | + · · · .
hilation operators:
Surprisingly, Lanczos algorithm determines the se-
quence αn and matrix U exactly even away from the
k−1 k−1
limit T → 0 when all coefficients αn are real. In this case, 1 x 1 x
(σk − iσky ) c†k = (σk + iσky )
Y Y
we can employ the construction suggested in [110, 111]. ck = σjz , σjz .
2 j=0
2 j=0
Namely, we define the Hamiltonian H:
(B2)
So, operator evolution (B1) preserves the subspace of
 
H = i U 1/2 − U −1/2 , (A9)
one-particle fermion operators and reproduces the CMV
matrix (12) after identification αk = eiχk cos θk :
redefine the inner product and norm:
d d
!
(A|B) → (A|(1 − H/2)|B) , (A10) †
X X
UXY vk ck UXY = (LM v)k ck , (B3)
and generate sequences {an } and {bn } using the standard k=0 k=0
Lanczos algorithm (2). These sequences are unambigu-
ously related to coefficients αn of the original CMV basis: where vk are arbitrary complex numbers. In other words,
this mapping identifies one-particle operators ck with the
an = (−1)n (αn + αn−1 ) , bn = ρn−1 . (A11) CMV basis vectors. In particular, boundary operator c0
corresponds to the seeding state |P0 ⟩ = |ψ⟩/∥ψ∥.
Hence, we easily restore coefficients αn and matrix U :
Finally, this model can be rewritten as an inhomo-
n
X geneous transverse-field Ising model [53, 54] when all
αn = (−1)n ak + (−1)n . (A12) Verblunsky coefficients are real, i.e., χk = 0 for all k.
k=0