SciPost Submission Page
Tangentspace methods for truncating uniform MPS
by Bram Vanhecke, Maarten Van Damme, Jutho Haegeman, Laurens Vanderstraeten, Frank Verstraete
Submission summary
As Contributors:  Jutho Haegeman · Laurens Vanderstraeten · Bram Vanhecke 
Arxiv Link:  https://arxiv.org/abs/2001.11882v2 (pdf) 
Date submitted:  20210113 10:20 
Submitted by:  Vanhecke, Bram 
Submitted to:  SciPost Physics Core 
Academic field:  Physics 
Specialties: 

Approach:  Theoretical 
Abstract
A central primitive in quantum tensor network simulations is the problem of approximating a matrix product state with one of a lower bond dimension. This problem forms the central bottleneck in algorithms for time evolution and for contracting projected entangled pair states. We formulate a tangentspace based variational algorithm to achieve this for uniform (infinite) matrix product states. The algorithm exhibits a favourable scaling of the computational cost, and we demonstrate its usefulness by several examples involving the multiplication of a matrix product state with a matrix product operator.
Current status:
Submission & Refereeing History
You are currently on this page
Reports on this Submission
Anonymous Report 2 on 2021122 Invited Report
Report
The authors have addressed enough of the points brought up in my previous report: I thus recommend the publication of the manuscript in SciPost. I invite the authors to nevertheless take a look at the following comments:
1. I appreciate the reformulation of the definition of the cost function. I still find the wording around it a bit confusing though. Among others, the sentence just before Eq. 6 seems to refer to $\langle \Psi(A^*)\Psi(M)\rangle$ as a "norm", which it is not in the mathematical sense of the term, and this can be confusing. It would be more accurate to speak of "overlaps" or "inner products".
2. I find the sentence right after Eq. 9 rather confusing. In the previous manuscript, the rhs of Eq. 7 was the part proportional to $\Psi(A)\rangle$, which vanished under the projector $\mathcal P_A$, and it prompted requested change #2 in my previous report. Now, that equation is no longer there. What I understand as a reader is that Eq. 9 comes from taking the vector whose overlap with the tangent space we need to vanish, i.e. $\Psi(M)\rangle\dfrac{\langle\Psi(A^*)\Psi(M)\rangle}{\langle\Psi(A^*)\Psi(A)\rangle}\Psi(A)\rangle$ from Eq. 8, applying the projector (which annihilates the part proportional to $\Psi(A)\rangle$) and equating that to 0. The rhs of Eq. 9 arises from this last equating to zero part, i.e., it was never anything other than zero. In what sense can we "keep track of it"?
3. I still believe the "if and only if" right before Eq. 11 (e.g. point 3 of my previous report) is not entirely obvious. In their reply, the authors mention an added comment right after Eq.17, which says that said equation vanishes only if the gradient does. Again, it is selfevident that (17) vanishing implies (9) is satisfied, while the converse, though it seems intuitively correct, could use more justification at the mathematical level (since the reader is likely unfamiliar with the mathematical standing of infinite sums and subtractions of infinite MPS). It is however a valid choice by the authors to omit said justificacion if it is technical enough that it would derail the proper flow of the paper.
4. What does $\sim$ mean in the sentence after Eq. 14: "Eq. (11) can only be satisfied if $C'\sim C$ and $A'_C\sim A_C$"?
5. I guess the reader might still be a bit lost regarding how significant the fidelity improvement between the two methods in Fig. 1 is in a practical sense, though I understand it may be difficult to assess it generally.
6. In the power method section, I understand it would be cumbersome to prove that the antiferromagnetic fixed points can be found as the two sublattice rotated ferromagnetic fixed points, thus allowing the latter to be used as a benchmark for the antiferromagnet obtained via power method (which, it is to be guessed, could not be obtained via vumps as there is no single fixed point). Just to reiterate my opinion, stating this rather than implying it could ease the reader's task.
7. Regarding #14 in my previous report, the authors's understanding probably exceeds mine in this, but I keep wondering if the difference between the improvements in computational cost for the cases with and without substructure justifies singling out the former. In the paper it is argued in detail that in the case with substructure, the cost of their scheme (time,memory) is $$O(\chi^3 D d + \chi^2 D^2 d), O(\chi^2Dd)$$ versus $$O(\chi^3 D^2 d + \chi^2 D^3 d + \chi^3 D^3), O(\chi^2D^2d)$$ for the local truncation scheme. If I have not made mistakes, in the nosubstructure case, where we have to approximate an MPS of bond dimension $\chi_2$ by one of dimension $\chi$, their scheme would give $$O(\chi^2 \chi_2 d + \chi \chi_2^2 d), O(\chi_2^2d)$$ versus $$O(\chi_2^3 d), O(\chi_2^2d)$$ for the local truncation. Indeed, the savings in memory cost disappear, but in computational time they could be comparable to the other case, since $\chi_2$ could be though of being $\sim\chi D$. Once more, though, it is up to the authors to emphasize what they find more relevant in their work.
Report 1 by Johannes Hauschild on 2021121 Invited Report
Report
The authors have properly addressed the requested changes from the previous review, so I think this paper ready for publication.
Requested changes
None