SciPost Submission Page
Continuous Matrix Product States for Inhomogeneous Quantum Field Theories: a BasisSpline Approach
by Martin Ganahl
Submission summary
Authors (as Contributors):  Martin Ganahl 
Submission information  

Preprint link:  scipost_201901_00046v1 
Date submitted:  20190118 01:00 
Submitted by:  Ganahl, Martin 
Submitted to:  SciPost Physics 
Ontological classification  

Academic field:  Physics 
Specialties: 

Approach:  Computational 
Abstract
Continuous Matrix Product States (cMPS) are powerful variational ansätze for ground states of continuous quantum field theories in (1+1) dimension. In this paper we introduce a novel parametrization of cMPS wave functions using basissplines, which we coin splinebased MPS (spMPS). We show how to variationally optimize spMPS by extending a recently developed ground state optimization algorithm for translational invariant cMPS [M. Ganahl, J. Rincón, G.Vidal. Phys.Rev.Lett. 118,220402 (2017)] to the case of spMPS, and use it to obtain the ground state of a gas of LiebLiniger bosons in a periodic potential. The resulting ground state wave function is virtually free of discretization artifacts, which we verify by comparing our results to lattice Density Matrix Renormalization Group (DMRG) calculations on extremely fine grids. We show that spMPS achieve a reduction of variational parameters by more than two orders of magnitude as compared to the lattice DMRG method, without any compromise in accuracy.
Current status:
Submission & Refereeing History
You are currently on this page
Reports on this Submission
Anonymous Report 2 on 2019426 (Invited Report)
Strengths
1detailed description of the methodology (i.e. the algorithm used, and specific routines used in Python)
2first demonstration of continuous matrix product states in an inhomogeneous setting
3fair discussion of limitations of the current approach
Weaknesses
1technique that this paper introduces is only illustrated on/applied to a single example
2specific quantitative details about the example/results are vague or missing: i.e what are the runtimes of the variational spMPS versus the reference DMRG, how many variational parameters has the former versus the latter, ...
3introduction to splines is not particularly enlightning
Report
In this manuscript, the author applies the continuous matrix product state ansatz of Verstraete and Cirac to inhomogeneous quantum field theories. So far, their use had been restricted to translationinvariant systems, in which case the variational parameters correspond to a single set of matrices. For an inhomogenous system, the variational parameters are instead matrixvalued functions of space, and a particular way needs to be chosen to represent those using a finite number of parameters stored on a computer. Here, the authors choose to combine spatial discretisation and spline interpolation.
The technical aspects of this approach and its implementation are described in a selfcontained way, without becoming too elaborate. The only subsection of Sections II to IV which I found to be of lesser quality is the introduction to splines. For example, the knot points are introduced, but their role or use is not really discussed (especially when they are degenerate). Furthermore, most of this section is about splines for open boundary conditions, whereas the author then uses splines with periodic boundary conditions.
The example (Section V) is somewhat weak. Only a single example is studied (and to know which potential this exactly represents, one has to browse back two pages, it would be better stated in Section V). Furthermore, this section provides very little quantitative details and leaves the reader with many questions, such as
* How much do the DMRG reference result (N=10^4) and the spMPS result differ, e.g. for the particle density, or for the total energy in one unit cell?
* What are the actual runtimes of the algorithms, the number of optimization steps?
* Does it work better or worse at larger bond dimensions? Is there a big improvement over GrossPitaevskii (D=1)? Are there specific effect the latter cannot capture?
* How much variational parameters do the spMPS and the DMRG/MPS ansatz contain?
* There is a big gap between N=40 and N=10^4, how would DMRG do with N = 400, 1000, 4000?
* Is there any limit in which comparison with an exact result or some other technique is possible?
Other comments and typos:
* Both in Section III and in the appendix, the central canonical form is introduced, with matrices (matrix functions) Gamma_Q(x) and Gamma_R(x). However, these do not seem to be used in the actual algorithm, only Q_l, Q_C, Q_r, R_l, R_C, R_r and C are used. What is the relevance of the Gamma_Q and Gamma_R or why are they also introduced?
* Abstract: "We show that spMPS achieve a reduction of variational parameters by two orders of magnitude...". Firstly, these numbers should be made more explicit in the Example section, as stated above. Secondly, the abstract should also point out the limitations, i.e. that optimization of this state is still hard, slow and unstable.
* "x1computational" in the first sentence
* "savely" > "safely" in the last paragraph of the appendix
Requested changes
* Improve the example section, addressing the questions posed above.
* Address other comments
Anonymous Report 1 on 201943 (Invited Report)
Strengths
1 Clear and concise
2 Proposes a technical implementation that may boost continusousspace MPS
3 Timely and potentially important for the field
Weaknesses
1 The comparison to algorithms based on naive discretization is very disappointing
2 Not clear whether the approach can be useful in practice or not
3 Benchmarking based on a single example
Report
The paper proposes a new implementation of matrix product state (MPS)  based tensor network techniques for continuousspace Hamiltonians. It builds upon an approach coauthored by the author of this contribution and introduces a spline parametrization of the MPS wave function. This is a standard approach, which usually allows for a better representation of continuousspace functions, as compared to naive, homogeneous discretization.
The paper is well written and easy to read. It contains all the pieces of information necessary to follow the approach, without excessive details. The limitations of the approach are discussed on the basis of calculations of the ground state of onedimensional bosons in a periodic lattice.
While the approach is a priori well justified and sound, it is far not clear that it provides any advantage in practice. This is a central point because the main contribution here is a rather technical implementation. The results presented in the paper are actually very diapointing and the conclusion is that it is better, in practice, to use a standard naive discretization. This is the main flaw of the paper. The author should provide convincing arguments that the method could be beneficial, maybe using different examples of physical systems to be solved or implementing the proposed improvement.
Alternatively, the abstract and the introduction should fairly state that the present implementation suffers strong limitations, but, possibly, that ways to overcome the latter are discussed. In the present form, I think that the sentence "We will demonstrate the power of the spMPS ansatz by obtaining the ground state of a gas of interacting LiebLiniger bosons in a periodic potential." is not a fair account of the results presented in the paper.
Requested changes
 The generic form in Eq. (1) should be justified
 The definition of the A^\sigma in terms of product matrices should be related to the wavefunction at the very beginning of Sec. III, where they are refered to
 The notations in the formulas for A^0 and A^\sigma are a bit misleading: Why is R(x) inside brackets, Q(x) is not?
 "For the cases considered in this manuscript, tensors A^\sigma with \sigma > 1 can be neglected." Why?
 Typo "x1" on the very first line
 The notion of "benchmarking" may be better stressed at the end of the introduction
 "an SVD" > "a SVD"
 Why is the form of the potential in Eq. (20) so complicated, and not just a sine function?
 "particle \langle n(x) \rangle" > "particle density \langle n(x) \rangle"