Mobius Fermions: Improved Domain Wall Chiral Fermions

Worth your time, with one caveat about the headline number. The construction is genuinely new: they write the domain wall kernel as a general real Möbius transform of the Wilson operator, D = (b5+c5)Dw² + (b5-c5)Dw, which contains Shamir (c5=0) and Boriçi (b5=c5) as limits and adds an independent scale alpha = b5+c5. The observation that alpha drops out at Ls = infinity by scale invariance of the sign function but matters at finite Ls is the actual insight, and it is correct. They also show the 5-d matrix still reduces to an overlap operator with the standard polar-decomposition epsilon_Ls, and they handle the modified gamma5-Hermiticity with the D_- factor and fifth-dimension reflection properly. The 4-d even-odd preconditioning trick (even-odd on each slice, not alternating in the fifth direction, so the even-even and odd-odd blocks are gauge-independent and invertible analytically) is a nice practical point that later implementations leaned on. No extra Dirac applications per CG iteration, as advertised in abstract:s0 — that is the reason anyone cared. On the soft spots: the stress-test concern lands. Section 4 says M5 = -1.5 and alpha = 2.3 were the optimal choices at Ls = 8, while the Shamir baseline sits at -M5 = 1.8 with no scan reported. So the 'factor of 2 or more' in abstract:s2 is a tuned point against a default, not a like-for-like comparison. They also state that pion masses were adjusted so Ls = 8 matched Shamir at Ls = 16, which is honest but means the match is partly arranged. No error bars on mres despite 20 lattices, so even the size of the gap is unquantified. The LDU reduction to Eq. 4 is deferred to a forthcoming article — fine for a proceedings, not fine for a journal version. The more interesting claim, that mres falls fast at large Ls with large alpha and that there is a crossover near Ls = 16, is stated with even less support than the Ls = 8 result. None of this undermines the construction; it undermines the precision of the advertised gain. Read it as a method announcement, which is what it is. It is the right paper to cite for Möbius DWF, and the direction held up. A serious referee should see it, mainly to ask for the Shamir M5 scan and error bars.

The central claim (abstract:s2) is that Ls can be halved at fixed mres. Figure 1 compares three curves: Shamir at -M5=1.8, α=1.0; Möbius at -M5=1.5, α=2.3; and Möbius at -M5=1.5, α=4.0. Section 4 states the Möbius parameters were explicitly optimized: 'For Ls = 8 our tests found that M5 = -1.5 is the optimal choice' and 'the optimal scale factor was α = 2.3.' No corresponding scan of the Shamir M5 is reported; -1.8 reads as a conventional default. Because mres at fixed Ls depends strongly on M5 through the eigenvalue range of H = γ5 D(M5), an untuned Shamir baseline could sit well above its own minimum, inflating the apparent Möbius advantage. The structural argument is sound and suggests Möbius should win at finite Ls: α is a knob Shamir does not have, and it is irrelevant at Ls=∞ by scale invariance of the sign function. But the size of the win — 'a factor of 2 or more' — is exactly what a tuned-versus-default comparison cannot establish. Note that a5 is held at 1.0 for both operators, so the comparison is fair in a5; the asymmetry is confined to M5 and α. Compounding this, no error bars are shown on mres despite 20 lattices, so even the tuned-versus-default gap is not quantified statistically.

The reader's weakest_assumption identifies precisely the load-bearing soft spot: the Möbius parameters were tuned (M5=-1.5, α=2.3 at Ls=8) while the Shamir baseline at -M5=1.8 is not shown to be comparably optimized. This is the right concern because the headline number — a factor of two or more in Ls at fixed mres — is a quantitative comparison between two operators, and a tuned-versus-default comparison cannot establish its magnitude. The structural argument that α is a genuine new knob unavailable to Shamir, and irrelevant at Ls=∞, is correct and makes it plausible that Möbius wins at finite Ls; the open question is by how much. The other gaps the reader notes are real but secondary or already acknowledged. The LDU reduction to Eq. 4 is explicitly deferred by the authors ('The detailed algebraic steps will be published in a forthcoming article'), so it is an acknowledged limitation rather than a new finding, and Eq. 4 is stated explicitly enough to be checked. The absence of error bars on mres is a genuine reporting gap but is statistical rather than systematic, and it compounds rather than replaces the tuning concern. The paper's cost claim in abstract:s0 — no increase in Dirac applications per CG iteration — is supported by the 5-d structure, where D(i)± are linear in Dw, so that is not where the argument is weakest. I therefore agree with the reader's CONDITIONAL verdict and would leave it unchanged; the concern is a fair-comparison question that a targeted Shamir M5 scan would settle, not a refutation of the construction.

If the central claim holds, this is a practically useful result: a factor-of-two reduction in Ls at fixed chiral-symmetry violation translates directly into a factor-of-two reduction in the dominant cost of domain wall fermion simulations, obtained without adding Dirac applications per CG iteration. The construction is a genuine and economical extension of the Shamir operator — α is a knob that Shamir does not possess, and the argument that it is irrelevant at Ls = ∞ is correct — and the 4-d even-odd preconditioning trick, which keeps the even-even and odd-odd blocks gauge-field independent, is a sensible practical contribution in its own right. The paper also deserves credit for defining mres through the matrix elements of the GW breaking operator Δ_Ls rather than through an ad hoc ratio, which is the right measure and makes the comparison with Shamir meaningful in principle. The work is presented at proceedings length and the structural argument is sound; what is at issue is the strength of the numerical evidence for the size of the claimed gain.

Lattice QCD wants exact chiral symmetry at finite lattice spacing, and domain wall fermions approximate it by adding a fifth dimension of length $L_s$: the longer that dimension, the smaller the residual chiral breaking. Simulation cost grows with $L_s$, so the practical question is how short the fifth dimension can be while the breaking stays negligible. This paper argues that the standard Shamir kernel is a single point in a one-parameter family obtained by applying a real Möbius transformation to the Wilson operator, and that the new parameter $a = b_5 + c_5$ does real work at finite $L_s$ even though it cancels out of the sign function in the $L_s$ to infinity limit. On quenched lattices at $beta = 6.0$, the authors report that tuning $a$ lets $L_s = 8$ reach essentially the same residual mass as conventional Shamir at $L_s = 16$, which is where the factor-of-two saving comes from.

What carries the argument. The carrier of the argument is the five-dimensional domain wall matrix built from $D^{(i)}_+ = b_5 omega_i D_w(M_5) + 1$ and $D^{(i)}_- = c_5 omega_i D_w(M_5) - 1$, with the polar choice $omega_i = 1$ throughout. An LDU decomposition of that matrix reduces it to an effective four-dimensional overlap operator whose sign function is $epsilon_{L_s}[H] = [(1+H)^{L_s/2} - (1-H)^{L_s/2}]/[(1+H)^{L_s/2} + (1-H)^{L_s/2}]$ with Hamiltonian $H = gamma_5 D_{Mobius}(M_5)$; the chiral boundary fields $q_x = P_- Psi_{x,1} + P_+ Psi_{x,L_s}$ then reproduce the overlap propagator $langle q_x qbar_y rangle = [1/(1-m)][D_{ov}^{-1}(m) - 1]_{xy}$. The paper notes that $gamma_5$ Hermiticity now requires conjugating by $D_- = Diag[D^{(1)}_-, ..., D^{(L_s)}_-]$ together with a reflection in the fifth direction, which in turn changes the correct definition of the boundary states. Chirality violation is measured by the residual mass $m_{res} = Tr[Delta_{L_s}(H) D_{ov}^{-1} D_{ov}^{dagger -1}] / Tr[D_{ov}^{-1}D_{ov}^{dagger -1}]$, a specific matrix element of the Ginsparg-Wilson breaking term $Delta_{L_s} = (1 - epsilon_{L_s}^2)/4$, so that a vanishing residual mass implies unbroken Ginsparg-Wilson and axial Ward-Takahashi relations. Practical use also requires even-odd preconditioning, which the authors implement only on each four-dimensional slice (not alternating in the fifth direction) so that the even-even and odd-odd blocks are gauge-field independent and invertible analytically; they report roughly a factor of 3 improvement, comparable to what five-dimensional even-odd gave for Shamir. The paper defers the full algebraic steps of the decomposition to a forthcoming article and describes the numerical study as preliminary, with more detailed performance data across quenched and unquenched lattices still needed.

Core claim. The central claim is that replacing the Shamir overlap kernel $D_{Shamir}(M_5) = a_5 D_w(M_5)^2 + a_5 D_w(M_5)$ by the real Möbius kernel $D_{Mobius}(M_5) = (b_5+c_5)D_w(M_5)^2 + (b_5-c_5)D_w(M_5)$ gives a genuinely larger parameter space at no extra cost per conjugate-gradient iteration. Shamir's form is recovered by setting $b_5 = c_5 = a_5$; the new direction is the scale $a = b_5 + c_5$, which is invisible in the $L_s$ to infinity limit because the sign function satisfies $epsilon[a x] = epsilon[x]$, but which changes the finite-$L_s$ rational approximation. The authors' argument for why this helps is that a Möbius map can send three real values (for instance eigenvalues of the Wilson kernel) to arbitrary targets, so a single choice of $M_5$ can separate the physical chiral modes from doublers while simultaneously adjusting both the smallest and largest real eigenvalues entering the kernel. The evidence is a quenched study at $beta = 6.0$ on 20 configurations at bare quark mass $m = 0.06$: with $-M_5 = 1.5$ and $a = 2.3$, $L_s = 8$ gives essentially the same $m_{res}$ as Shamir at $L_s = 16$, and at larger $L_s$ the residual mass falls rapidly when large scale factors are used, with a crossover appearing at $L_s = 16$ for $a = 4$.

The premise it rests on. The factor-of-two saving rests on the assumption that the Shamir run at $-M_5 = 1.8$ is a fairly tuned reference, even though only the Möbius parameters were scanned for optimality.

If the result holds, a fixed chiral-accuracy target costs half as many fifth-dimension sites, which cuts memory and, at fixed iteration count per site, roughly halves the dominant cost of a domain wall simulation.

Because the scale factor is a free knob rather than a fixed constant, the same code can be retuned for different lattice spacings, volumes, and quark masses without changing the operator structure or the number of Dirac applications per iteration.

At large $L_s$, where valence-quark measurements live, the residual mass falls rapidly with large scale factors, so high-precision valence physics may become affordable at fifth-dimension lengths previously considered impractical.

The construction contains Shamir and Boriçi as limits, so existing domain wall codes can be extended by adding one parameter rather than by rewriting the inverter.

The four-dimensional-slice even-odd preconditioning scheme is a reusable algorithmic ingredient for any kernel of this Möbius type, independent of the chirality claim.

The comparison would be much stronger if the Shamir baseline were tuned as carefully as the Möbius point: the paper reports an optimal scale of 2.3 at $L_s = 8$ but no corresponding scan over Shamir's $M_5$, so part of the quoted factor of two could be a tuned-versus-default comparison rather than a like-for-like one.

The mechanism suggests a broader design principle: any kernel whose finite-$L_s$ rational approximation can be reshaped by a parameter that cancels in the infinite-$L_s$ limit is a candidate for the same treatment, including overlap formulations that are not literally domain wall.

Because the advantage is largest where the spectrum of the kernel is widest, one would expect the optimal scale factor to drift with lattice spacing and volume; a table of optimal scale versus $beta$ would be the natural next artifact and would make the method usable without per-ensemble tuning.

The reported crossover at $L_s = 16$ for a scale factor of 4 hints that the optimal scale is not monotonic in $L_s$, which would mean the tuning recipe is a two-dimensional problem in $(M_5, a)$ rather than a single recommended number.

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review