Referee reports, SISC October 2015 (decision: reject)

[labarba]

Decision : Reject (October 14, 2015)

Refers to v1 of arXiv:1506.05957.

SISC Associate Editor Comments

In essence, both referees state that the novelty in your current article is not sufficient to warrant a publication in SISC. The article might have been acceptable with more theoretical backing and/or a more systematic experimental evaluation.

Referee # 1:

This paper by Layton and Barba is basically centred around the following ideas:

1. The BEM for PDEs leads to dense linear systems, which however contains a lot of distance-based degradation of the importance of the coupling between two unknowns. This naturally leads to the idea of using the FMM to accelerate matrix-vector products.

2. Results from the previous decade indicate that Krylov solvers can be inexact, or better put, that matrix-vector products can be computed with some error without severely compromising the result. The results also indicate that the error can increase (within certain bounds) over iterations.

3. The contribution of this paper is the combination of these (existing) ideas, and to evaluate the approach with extensive numerical tests.

I think the idea is nice and deserves publication, despite not being earth-shatteringly novel. In its present form however, I cannot recommend acceptance one key reason: The paper is "too sloppy" for SISC, both in terms of language and more importantly, of maths.

• First, the not-so-important aspect of sloppyness: There is way too many colloquialism in language and grammar and wording. Please revise carefully.

Now, the sloppy math: Assuming that most of this can be addressed without severe impact on the results, the speedup results are nice, by the way.

• 1. The definition of a Krylov subspace involves residuals, not right hand sides, see for instance the classic textbook by Saad: With $r_0 := b-Ax_0$, we have $K_r = \span{r_0, Ar_0, A^2r_0, \ldots}$.
• 2. It is never mentioned which preconditioner is used. This is crucially important, because there's another inexact Krylov theory called inner-outer-methods or self-preconditioning which allow errors and inaccuracy in the preconditioner. To avoid confusion, please differentiate clearly from this avenue of reaearch. Actually, it would be nice to use FMM to accelerate the preconditioner as well, for instance by only using a small number of layers in the tree, which would be the inexact FMM analogue of a truncated "polynomial" preconditioner.
• 3. Similarly, why GMRES and not CG?
• 4. Sometimes, notation is weird. For instance, just below eq. 2.13: What is F_i, and since when is \Delta = \nabla^2? It's \Delta=\nabla\cdot\nabla!
• 5. In Sec.3, most parameter choices are not explained. Why are \Theta_{MAC} and p and TOL chosen the way they are? Because they give good grid convergence results and still good speedup? Also, why does the iterative tolerance change from test to test? Grid convergence is done with 1e-6, residual-over-p with 1e-5 etc. This does not lead to great confidence in the results.
• 6. Most severely, it is claimed that reducing p gives identical results (as long as the order is not reduced too aggressively). This may be so, but not one single numerical result backs up this claim! All BEM-FMM validation is done with comparisons of the (relative) algebraic error in some suitable norm (which one?), while all comparisons for reduced $p$ are done using norms of the residual. This is NOT the same! Residuals are a measure of how well a linear system is fulfilled, while the algebraic error measures how close you are to the true solution. Please be more rigorous here. Also, these numbers must be related to the condition number, using an appropriate argument with the perturbation theorem. Incorporating all these aspects would greatly increase the credibility of the results beyond the current plausibility based on circumstancial evidence.
• 7. The last sub-item is even made worse by the (nonsense) argument of just doing a couple additional iterations to simulate ill-conditioned problems.
• 8. In all tests, why is the numerically determined "lowest-possible "p" not related to the bound in eq. 2.19? If there is such a bound, I am crucially interested in how good/sharp it actually is, and if it is computable in the first place. This is not addressed.

Some minor comments:

• 9. The guy whom the delta was named after is Kronecker, not Krönecker.
• 10. Some cites on how the FMM works would be nice. Section 2.5 is only legible by experts on the subject, not by the BEM folks who seem to be the target audience.
• 11. Figure 3.10 claims to have a residual history, which I don't see. Uploaded the wrong figure?
• 12. Tables 3.10 and 3.11 have footnotes, which are not contained in the PDF.

---

Referee # 3:

A BEM solution strategy with relaxed application of the multipole method is proposed when solving the dense and large linear system. The idea of relaxing the accuracy of matrix-vector products was first introduced by Bouras & Fraysse, and then completely analyzed and developed in [17,18].
Since then, its applicability in multipole, multilevel and other complex strategies has been explored with success. Therefore, it is not surprising that also for this application the approach works well.
The idea is interesting, and the extensive experimental analysis also nice. However, it is not clear to me whether the contents of this manuscript is sufficiently novel to grant publication in SISC. The experimental discussion is not backed up by any theoretical discussion. Given the fact that the strategy has been used quite a bit, a deeper analysis would be welcome. Therefore, this manuscript
might be more suitable for a more application-oriented journal, or for a proceedings.

From a more technical point of view, I do not see any need to include Algorithm 1 and 2 for the given readership.

[labarba]

Referee # 1, comment # 2:

It is never mentioned which preconditioner is used.

See Issue #13 — "identity preconditioner" in all cases, i.e., no preconditioning applied. Clarified on revision, commit 737a13e

[labarba]

Referee # 1, comment # 1:

The definition of a Krylov subspace involves residuals, not right hand sides, see for instance the classic textbook by Saad: With $r_0 := b-Ax_0$, we have $K_r = \span{r_0, Ar_0, A^2r_0, \ldots}$.

The book by Saad (inventor of GMRES) does indeed present the Krylov subspace in terms of the residuals (p. 157). But other textbooks write it in terms of the right-hand vector b, e.g., Trefethen and Bau, p. 266. Of course, they are equivalent when the initial guess x_0 is zero—which is our case.

We have changed how it's written in the paper to match Saad's presentation. See commit b205d95.

[labarba]

Referee # 1, comment # 3:

Similarly, why GMRES and not CG?

Because the coefficient matrix A is not always symmetric. Element A_ij is the integral over panel j with the Green's function evaluated at the collocation point of panel i. Due to different shape and orientation of panels, A_ij ≠ A_ji.

[labarba]

Referee # 1, comment # 4:

Sometimes, notation is weird. For instance, just below eq. 2.13: What is F_i, and since when is \Delta = \nabla^2? It's \Delta=\nabla\cdot\nabla!

The referee may not appreciate a citation to Wikipedia, but here goes: "…the Laplacian … is usually denoted by the symbols ∇·∇, ∇², or Δ."

Plenty of textbooks adopt the ∇² notation, in fact. One of our favorites is "Vectors, tensors, and the basic equations of fluid mechanics" by Aris. Wonderful little book! Even the classic old "Foundations of Potential Theory" by Kellogg (1929) uses the ∇² notation.

The best answer we can find to "since when" ∇² is used to denote the Laplacian operator is: 1873, "A treatise on electricity and magnetism," Vol. 1, by James Clerk Maxwell, page 29. See: http://visualiseur.bnf.fr/CadresFenetre?O=NUMM-95175&I=3&M=tdm

[labarba]

Referee # 1, comment # 5:

In Sec.3, most parameter choices are not explained. Why are \Theta_{MAC} and p and TOL chosen the way they are? Because they give good grid convergence results and still good speedup? Also, why does the iterative tolerance change from test to test? Grid convergence is done with 1e-6, residual-over-p with 1e-5 etc. This does not lead to great confidence in the results.

In version 1 of the manuscript—arxiv:1506.05957v1—,we chose p (value of 10 in Laplace and 16 in Poisson) from experience, to give accuracy close to singe precision at the finest mesh. Then, p was kept the same at coarser meshes simply from a criterion of "changing one thing at a time." The iterative tolerance is tighter for the grid-convergence analysis to ensure that discretization error is what dominates. The tests of residual-over-p do not use the finest mesh, and thus the tolerance can be loosened. (In later versions of the manuscript, we have made extra effort to choose these parameters.) The value of ϴ_MAC=0.5 is standard in fast multipole method evaluations: it gives an acceptance criterion based on well-separated boxes.

In version 2 of the manuscript—1506.05957v2—, we matched the tolerance in the grid-convergence and the residual-over-p result, to avoid any misunderstanding. It does not affect the result in any substantial way.

In version 3 of the manuscript, we did two things. For the grid-convergence analysis, we show results with "tight" parameters, for high accuracy—high p, high k, small tolerance—, and "loose" parameters—reducing p, etc., until accuracy begins to deteriorate, but no more than 10%. For other tests, we made an effort to look for the optimal combination of parameters to match the discretization error according to grid coarseness. The whole process is very detailed, and we report it through additional supplementary materials (Jupyter notebooks).

[labarba]

Referee # 1, comment # 6:

All BEM-FMM validation is done with comparisons of the (relative) algebraic error in some suitable norm (which one?), while all comparisons for reduced $p$ are done using norms of the residual.

The referee is right to point out that we don't state the norm of the relative error used for the grid-convergence study. In version 2 of the manuscript, we added this information: we used the L2-norm of the error with respect to the analytical solution.

The referee is also right to point out that we should report the actual error, rather than only the residuals. Responding to comment # 11, we added the residual history in version 2 of the manuscript (it was, indeed, a mistake with the figure version used). But n version 3 of the manuscript, we removed the result showing residual history altogether, and instead report the relative error for each test case (in table form).

these numbers must be related to the condition number, using an appropriate argument with the perturbation theorem.

The tests show that only a few dozen iterations are needed to converge (without using a preconditioner), indicating that the system is not badly conditioned. Furthermore, we report the actual error, proving that the system is indeed converging to the right solution.

[labarba]

Referee # 1, comment # 7:

… made worse by the (nonsense) argument of just doing a couple additional iterations to simulate ill-conditioned problems.

We were not simulating ill-conditioned problems, but simply exploring how the speed-up might be in cases that required more iterations. Nevertheless, we removed this result in version 2 of the manuscript. It made no major claims: is only showed a trend to higher speed-ups.

[labarba]

Referee # 1, comment # 8:

why is the numerically determined "lowest-possible "p" not related to the bound in eq. 2.19? If there is such a bound, I am crucially interested in how good/sharp it actually is, and if it is computable in the first place. This is not addressed.

It is related to the bound. As explained in the text below the FMM error bound, we take r / a ≥ 2, which comes a geometrical consideration, and that gives the base-2 log in equation 2.19.

[labarba]

Referee # 1, comment # 12:

Tables 3.10 and 3.11 have footnotes, which are not contained in the PDF.

The footnotes appeared on the next pages — this is just a LaTeX quirk and we had no control over it (but would have been fixed in copy-editing).