SciPost Submission Page
Flow Equations for Disordered Floquet Systems
by S. J. Thomson, D. Magano, M. Schiró
This Submission thread is now published as
Submission summary
Authors (as registered SciPost users):  Steven Thomson 
Submission information  

Preprint Link:  https://arxiv.org/abs/2009.03186v4 (pdf) 
Date accepted:  20210805 
Date submitted:  20210625 14:50 
Submitted by:  Thomson, Steven 
Submitted to:  SciPost Physics 
Ontological classification  

Academic field:  Physics 
Specialties: 

Approach:  Theoretical 
Abstract
In this work, we present a new approach to disordered, periodically driven (Floquet) quantum manybody systems based on flow equations. Specifically, we introduce a continuous unitary flow of Floquet operators in an extended Hilbert space, whose fixed point is both diagonal and timeindependent, allowing us to directly obtain the Floquet modes. We first apply this method to a periodically driven Anderson insulator, for which it is exact, and then extend it to driven manybody localized systems within a truncated flow equation ansatz. In particular we compute the emergent Floquet local integrals of motion that characterise a periodically driven manybody localized phase. We demonstrate that the method remains wellcontrolled in the weaklyinteracting regime, and allows us to access larger system sizes than accessible by numerically exact methods, paving the way for studies of twodimensional driven manybody systems.
Author comments upon resubmission
Dear Editor,
Please find attached the revised version of our manuscript "Flow Equations for Disordered Floquet Systems".
We are happy to see that both referees found our manuscript clear and well written, and we appreciate their constructive suggestions. The main feedback on the previous version from both referees was that this work would be significantly stronger with an improved error estimate that did not rely on comparison with exact numerics, such that when we consider larger systems than accessible with exact numerics we can still assess the accuracy of the method in a selfconsistent way. We have conducted a detailed and thorough investigation of this point, and the major change in the revised version of our manuscript is the addition of several new sections (Sections 4.3.1 and 6.4) which introduce the concept of flow invariants and demonstrate how they may be used to serve this purpose. We thank the referees for motivating us to address this.
We have additionally made many other smaller changes to the manuscript, detailed in our full response to the Referees below.
Yours sincerely, S. J. Thomson, D. Magano & M. Schiró
 Response to Second Report of First Referee 
We thank the Referee for their continued detailed reading of our manuscript and their constructive suggestions.
"1) My main criticism centers around the sentences:  Page 13: “There is, however, no divergence encountered due to resonant terms, as in typical perturbative treatments of disordered systems: the sole effect of closely separated onsite energies is the need for a large flow time lmax”  Page 20: “Note that this truncated form of K(l) is not the result of a perturbative analysis, but instead arises from a truncation in operator space which is valid in the localized phase. Flow equation methods based around Wegnertype generators therefore avoid problems with resonances [60] which can otherwise lead to divergent terms in perturbative treatments of interacting systems” in combination with “variational manifold” below Eq. (39).
"Indeed, for a quadratic Hamiltonian like the driven Anderson model (page 13) there are no problems with divergences because the unitary flow is exact in the sense that no higher order terms are being generated. So the only complication due to resonances are long flow times, exactly like the authors say. However, for interacting systems (page 20) the situation is different. The very slow decay of couplings in any given order of the flow equation calculation because of (near) resonances generically leads to large new interactions being generated in the next order. So for the interacting system the situation is worse than just needing long flow times, and the problem with resonances is not avoided (unfortunately). The way to expicitly see this is by going beyond the terms taken into account in the ansatz (39), which is of course a bit of work.
"However, on the plus side this does provide a check for the internal consistency of the method if the newly generated and neglected terms indeed remain small. It is not even necessary to take the whole flow of the newly generated terms into account, looking at the source term in their flow equations will be sufficient. One can speculate that normal ordering might be helpful, but there is no obvious reason why this should be the case for a localizationdelocalization scenario. For the above reasons the term “variational” is also somehow misleading."
The Referee is quite correct to say that divergences may arise due to long flow times, as also seen in Ref. 61 for a timeindependent system of disordered, interacting fermions. In our previous response, we focused on issues due to resonances in the sense of energy denominators of the form $1/(E_iE_j)$ becoming illdefined when the two energies coincided, and leading to divergent terms. As no such energy denominators occur in this method, this precise problem is not encountered. However, as the Referee notes, at low drive frequencies (or small disorder strengths), the long flow times required can lead to eventual divergences of the interacting terms: the sharpeyed reader may note that this is visible in Fig. 9 for small distances at the lowest drive frequencies considered, and briefly discussed in the context of normalordering corrections. We thank the Referee for pointing out that this is indeed a consequence of (near) resonances, and in our revised manuscript we have given a more precise discussion of this point.
Such problems were encountered in previous work by two of us on localization in timeindependent systems (Refs 61 and 65). We found that the generation of new terms tended to remain wellcontrolled in localized phases, and could be quantified by looking at socalled `invariants of the flow', i.e. quantities which would be conserved by an exact unitary transform, but which are not conserved exactly by our approximate ansatz (Eq. 44). We have now extended the discussion of flow invariants to the Wegner Floquet flow (see the new Section 4.3.1) and have used this as a diagnostic to assess the validity of our ansatz (see also next point). We previously found that normalordering is a useful technical tool in taming these divergences as it typically leads to more stable numerical behaviour, as discussed in Refs 61 and 65, though its implementation in the driven case is beyond the scope of the present manuscript.
We have amended the term `variational', as also noted by the second referee.
"At a minimum the authors should change the sentences on page 13 and 20 to correctly describe the limitation of the method. Additionally, I would encourage them to do an error analysis by looking at the size of the newly generated terms in order to show the internal consistency of the approach. This would also address point 9) raised by the second referee. Once this is done I strongly support publication of this well written manuscript."
We have acted on the Referee's suggestion and have included a new Section 6.4 where an error analysis is presented based on the flow invariant, which would be exactly conserved in absence of any truncation of newly generated terms and therefore provides us with an estimate of the internal consistency of the method, along the lines of previous methods used in Refs. 61 and 65. The summary of this analysis is given in the new Figure 8 where the violation from exact invariance, $\delta I$, is plotted for the weakly interacting case as a function of the drive frequency, showing the frequency range in which the method remains controlled. Interestingly this range extends well within the disorder bandwidth and it is not therefore limited to the highfrequency regime.
"2) I appreciate the authors’ clarifying comments regarding the localizationdelocalization crossover vs. transition. On page 21 (Sect. 6.3) the authors still use the term transition, this should be changed."
We have removed the word transition' when referring to the delocalized system; we thank the Referee for spotting this. We now use the word
transition' only when discussing the interacting system, where we speculate that the breakdown of our ansatz Hamiltonian may signal the onset of a phase transition, as in the timeindependent case (Refs 61 and 65).
 Response to First Report of Second Referee 
We thank the Referee for their careful reading of the manuscript, and for their detailed feedback.
"1) Equation (12). I think the notation is not the best. LHS is a state, RHS is an operator. I do not think this equality is mathematically correct."
We thank the Referee for pointing this out, and we have corrected this notation in the revised version.
"2) I am not sure I agree with the discussion after Eqs. (26) and (27). The authors say that $\omega \to 0$ corresponds to the static limit. Is that so? I would say $\omega \to \infty$ does as the authors later explain. I would think the zero frequency limit is ill defined without cutoff in $N_h$, there is an extra infinite sum compared to the static case."
The Referee is correct that this is not, strictly speaking, the static limit and we have amended this sentence accordingly. In the limit of $\omega \to 0$, all terms with an explicit frequency dependence vanish and the flow equations take on the form of the timeindependent system, but indeed with an additional sum over the harmonic index n. In this manner, the equations more closely resemble the flow equations for a timeindependent twodimensional system.
"3) I wonder if in these equations there is any advantage in going to the comoving frame $h(n)i \to h(n)iexp[n^2 \omega^2l]$ and similarly for $J(n)_i$. Then there is an explicit exponential decay in all the nonzero terms in the sum and various approximations are more transparent."
This is a very interesting suggestion: we thank the Referee for drawing our attention to this possibility, and plan to return to it in a future work.
"4) It is very difficult to interpret different lines in Fig. 3 without legends/explanations. In the printed version in black and white the figure is simply unreadable."
Fig. 3 is intended to give the reader an overview of the behaviour of different classes of terms under the action of the continuous unitary transform, e.g. panels (c) and (c) show the flow of the zerofrequency onsite terms (solid) and hopping terms (dashed), while panels (e) and (f) show the flow of the higher frequency terms. The aim of these figures is to demonstrate that the flow is smooth and wellcontrolled, and also that the higher harmonics decay quickly to zero at high drive frequencies and more slowly in the case of low drive frequencies.
We do, however, agree with the Referee that Fig. 3 (and the similar Fig. 6) is difficult to read, and so we have streamlined this figure in the revised version, now showing the flow of only a small representative subset of the running couplings. We believe that the revised figure is clearer, while still conveying the same qualitative insight into the behaviour of the flow in different regimes.
"5) It looks to me that the bottom plot in Fig. 4 does not support very well the author's claim. If I only look at the figure then I would guess that the line does not stop at the asymptotes and keep going up (at small $\omega$) and keep decreasing (at large $\omega$). More points are needed to justify the conclusion. Also the choice of colors for the top panel is far from optimal with largest and smallest frequency being identical."
In fact, the central claim of Fig. 4b is that the flow equations and exact diagonalization results match extremely well, as this is a highly nontrivial quantity to compute using continuous unitary transforms. Additional results are shown in Appendix C from exact diagonalization demonstrate the behaviour over a greater range of frequencies and system sizes to illustrate the behaviour across a wider parameter. We believe that the data shown in Fig. 4b is sufficient to demonstrate our main point that flow equations give the same results as numerically exact methods in this regime.
We have nonetheless added an additional data point at lower frequencies to show that the level statistics reach a plateau. For this data point, the flow equations do not reproduce the exact diagonalization result, as a larger number of harmonics would be required to accurately capture the behaviour in this low frequency region. We have also updated the colour scheme of Fig. 4a.
"6) There is a mistake in the inline expression for the WignerDyson distribution on page (17). First of all there is a wrong sign and second this is not the WignerDyson distribution, this is its approximation, i.e. it is the Wigner surmise."
We thank the Referee for pointing this out, and have rectified this error in the revised manuscript.
"7) I am completely lost in the motivation behind Sec. 5.4, i.e. truncating the flow to a single harmonic and the top panel in Fig. 5. To me this approximation strongly devaluates the method. Why not use the same number of harmonics as everywhere. Also as I mention below keeping all the harmonics should allow, I believe, to have an independent way of estimating the error without need of any ED."
The Floquet eigenstates are timeindependent, and therefore should not have contributions from higher harmonics. In this section, we are approximating the eigenstates of the blockdiagonal timeindependent matrices sketched in Fig. 1c, not the full timedependent eigenstates of Fig. 1a. Strictly speaking, to do so exactly would require two unitary transforms: one `reverse' transform to give us the full timedependent eigenstates of the Floquet operator shown in Fig. 1a, followed by a second transform into the basis of Fig. 1c to give us the timeindependent Floquet eigenstates of the effective Floquet Hamiltonian. We included this section to show that at high frequencies, we can obtain the Floquet eigenstates to high accuracy without requiring this twostep procedure.
"8) What is variational in the ansatz (39)? Is it a "better" word for "truncated" or there is some variational optimization involved? What is minimized then? Also later the authors say that this ansatz is valid in the localized phase. Probably they mean "deeply localized" phase."
This point was also mentioned in the second report of the first referee, and we have removed the word `variational'. This ansatz has been shown to be valid in the localized phase of the corresponding timeindependent model, only breaking down close to the transition (see Ref. 61), however has also been used to study the delocalization transition in a timeindependent model with longrange couplings (Ref. 65), where it was able to capture the delocalized phase in the weaklyinteracting regime.
"9) Let me finish with what I think is the biggest issue with the paper. It does not provide any internal way of estimating the error within the method. Sometimes this is impossible and we have to do what the authors do, compare with ED in small systems and hope the error does not change with the system size. But this does not seem to be the case here. E.g. one can properly evolve the eigenstates backwards, as the authors explain, and then compute the variance of the photon number. This should be available for system sizes well beyond ED. One can probably do something different with the same idea. Right now the authors advocate the method as an efficient way to overcome difficulties of ED but never show any advantages as all the examples except the last one are confined to very small systems and the last example does not have any independent error estimate. To me the step of finding a clear and controlled example, where the method goes beyond other methods is a crucial part for advocating that it offers real advantages."
We thank the Referee for motivating us to further develop our error analysis. To this extent we have extended the concept of flow invariants, conserved quantities of the exact unitary flow used in static (timeindependent) applications of the flow equation method, to the Wegner Floquet generator used in this work (see new Sect.4.3.1). This allows us to assess the stability of our method in a case where truncation of the newly generated terms in the flow is necessary, such as in presence of interactions, which would result in a violation of the exact conservation law. In particular, as we now discuss in the new Sec.6.4, by monitoring the flow invariant as a function of the external parameters we can assess the regimes in which our truncated ansatz remains under control, as shown in Figure 8.
List of changes
New Section 4.3.1 on flow invariants and their uses.
Added brief discussion of the conservation of the flow invariant in Section 5.3 for the noninteracting system.
New Section 6.4 where we show the flow invariant for the interacting system, and added new Fig. 8.
Replaced Figs. 3 and 6 with cleaner figures showing fewer lines.
Corrected the notation in Eq. 12.
Minor changes to notation in Section 4.3 to be consistent with the new Sec. 4.3.1.
Replaced Fig. 4a with new colour scheme, and Fig. 4b now includes an extra point at low frequency.
Added sentence regarding the generation of higherorder terms by the flow, referencing work where this is discussed.
Added an additional citation in the introduction (Ref 73).
Published as SciPost Phys. 11, 028 (2021)
Reports on this Submission
Anonymous Report 2 on 2021723 (Invited Report)
 Cite as: Anonymous, Report on arXiv:2009.03186v4, delivered 20210723, doi: 10.21468/SciPost.Report.3287
Report
The authors addressed all the questions by myself, and I believe by the second referee. I have a few optional further suggestions, which authors might or might not want to take into account. Irrespective of their response I recommend the paper now.
1. The authors make some strong statements about MBL phase, which are not completely important for the rest of the paper. Nevertheless, recently many of the statements were challenged in the literature and many earlier claims were shown to be simply incorrect. E.g. on page 2 of https://arxiv.org/abs/2107.05642 by D. Huse and collaborators they openly challenge validity of a simple lbit picture of MBL and push the transition boundary to W>7 (instead of previously claimed W=3.5), a soon to be published work by D. Sels extended their method to larger system sizes and pushed the MBL transition point (if any) beyond $W_c>18$, which implies that basically all numerical results reporting MBL were actually finite size effects seen in the ergodic phase. Recently there emerged problems with the RG treatment of the transition as well. I believe the authors might soften a bit the language of their paper separating what was "shown" and what is "believed" to be true by some people.
2. On page (21) when the authors say that their method avoids the problem of small denominators, they might want to comment if it arises e.g. in Fig. 6. If the system sizes are small and $\Delta$ is also small (on the scale of level spacing) this problem might not even show up. I think the authors should be maximally fair about this point.
3. Page 22 "moreso" is written as a single word.
4. When the authors introduce their mistake in Eq. (50), I wonder what the absolute scale represents. I.e. is 0.01 small or large, does it change with the system size etc. To me it is more natural to normalize the mistake by dividing it by $I(\ell=0)$. Then the final value of the mistake tells us by how much offdiagonal elements were reduced by the flow.
5. For static systems an alternative to the flow method based on variational adiabatic transformations was introduced to introduce a nonperturbative SW transformation (https://arxiv.org/pdf/1910.11889.pdf). While there was no explicit discussion of the Floquet extension there it still might be worth commenting about that possibility as it also does not contain small denominators and allows for a systematic expansion.