Twisted Mass Fermions in Lattice QCD

• Twisted mass fermions are a Wilson-type lattice regularization that uses a chirally rotated quark mass term to achieve maximal twist and remove linear cutoff effects.
• Maximal twist is obtained by tuning the untwisted mass to its critical value, resulting in automatic O(a) improvement while necessitating clover terms to control O(a²) discretization effects.
• They enable precise spectroscopy, quark-mass determinations, and topological studies in lattice QCD, supported by advanced solvers like the DD-αAMG multigrid method.

Twisted mass fermions are a Wilson-type lattice-fermion regularization in which the quark mass term is chirally rotated. In the standard two-flavor formulation, the twisted mass enters as an $i\mu\,\gamma_5\tau_3$ term in the lattice Dirac operator; when the untwisted mass is tuned to its critical value, the theory is at maximal twist and parity-even physical observables are automatically free of linear cutoff effects in the lattice spacing. This formulation has been developed from $N_f=2$ simulations to $N_f=2+1+1$ simulations with dynamical strange and charm quarks, frequently supplemented by a clover term to reduce the characteristic charged–neutral pion mass splitting at finite $a$. It has been used for spectroscopy, hadron structure, quark-mass determinations, topology, quasi-distributions, and algorithmic developments in large-scale lattice QCD (0803.0224, 0810.3807, Finkenrath et al., 2017).

1. Wilson twisted-mass formulation

For two mass-degenerate flavors in the twisted basis $\chi=(u,d)^T$, the Wilson twisted-mass action is written as

$$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$

with Wilson operator

$$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$

and $\kappa=(8+2am_0)^{-1}$ for $r=1$ (0803.0224). The physical and twisted bases are related by an axial rotation, with maximal twist corresponding to $\omega=\pi/2$ (0803.0224).

In the ETMC $N_f=2$0 setup, the action is split into a light degenerate doublet and a heavy nondegenerate doublet. The light sector is

$N_f=2$1

while the heavy sector is

$N_f=2$2

Here the twist is along $N_f=2$3 and the strange–charm splitting along $N_f=2$4, a choice that keeps the heavy-sector determinant real (0810.3807). In clover-improved variants, the Wilson operator is augmented by the Sheikholeslami–Wohlert term; for example, physical-point $N_f=2$5 simulations use

$N_f=2$6

in both light and heavy sectors (Alexandrou et al., 2018).

This construction preserves the Wilson mechanism for doubler removal while shifting the leading lattice-artifact structure into a form that can be cancelled by tuning to maximal twist. The price is explicit flavor and parity breaking at finite lattice spacing, which reappears throughout the formalism as a central systematic effect.

2. Maximal twist, PCAC tuning, and automatic improvement

Maximal twist is reached by tuning the untwisted bare mass to its critical value so that the PCAC quark mass vanishes. In the two-flavor theory, the PCAC mass is defined as

$N_f=2$7

and the practical ETMC criterion is that the residual mistuning satisfy roughly $N_f=2$8 (0803.0224). The $N_f=2$9 program uses the analogous condition $N_f=2+1+1$0, with the same $N_f=2+1+1$1 benchmark (0810.3807). In clover-improved physical-point simulations, the operational criterion is

$N_f=2+1+1$2

or equivalently $N_f=2+1+1$3 up to small residual effects (Finkenrath et al., 2017, Alexandrou et al., 2018).

At maximal twist, parity-even observables are automatically $N_f=2+1+1$4-improved, so that leading discretization effects are pushed to $N_f=2+1+1$5 (0803.0224, Finkenrath et al., 2017). A standard practical consequence is the simple renormalized-mass relation

$N_f=2+1+1$6

used directly in quark-mass analyses (Finkenrath et al., 2017, Alexandrou et al., 2021).

A common misconception is that this statement is trivially preserved for any composite observable built from Wilson-type fermions. The topological-susceptibility analysis with spectral projectors shows why the issue is subtler. There, $N_f=2+1+1$7 is represented through integrated density chains that generate short-distance contact terms, and such terms could in principle spoil automatic improvement. Using the Symanzik effective theory and an OPE argument, the analysis shows that at maximal twist the relevant contact terms are odd under the $N_f=2+1+1$8 symmetry and therefore vanish; the spectral-projector/density-chain definition of $N_f=2+1+1$9 remains automatically $a$0-improved (Cichy et al., 2013). This result is stated to hold generally for density chains expressible in terms of $a$1, i.e. for chains with an even number of scalar and pseudoscalar densities (Cichy et al., 2013).

3. Heavy-quark sector, $a$2 implementations, and physical-point tuning

The nondegenerate heavy doublet is the defining extension from $a$3 to $a$4 twisted-mass QCD. Its renormalized strange and charm masses obey

$a$5

with the minus sign for strange and the plus sign for charm (0810.3807). Early ETMC studies tuned $a$6 and $a$7 by requiring the kaon mass to be physical and the charm quark to satisfy approximately $a$8 (0810.3807). First $a$9 production runs used the Iwasaki gauge action and PHMC, with separate $\chi=(u,d)^T$0 tuning for each light mass and about 5000 thermalized trajectories per ensemble in the initial study (0911.5244).

A technical complication of the heavy unitary action is the flavor non-diagonal and parity-odd Wilson structure, which makes neither flavor nor parity an exact quantum number at finite $\chi=(u,d)^T$1. In the first $\chi=(u,d)^T$2 study, this was already identified as a principal obstacle for $\chi=(u,d)^T$3- and especially $\chi=(u,d)^T$4-meson extraction, leading to the use of GEVP analyses, multi-exponential fits, and rotation back to the physical basis (0911.5244).

Later work established a physical-point strategy with maximally twisted mass fermions plus a clover term. A representative target ensemble has $\chi=(u,d)^T$5, lattice size $\chi=(u,d)^T$6, lattice spacing $\chi=(u,d)^T$7 fm, and parameters

$\chi=(u,d)^T$8

with stable molecular-dynamics evolution and an acceptance rate of about $\chi=(u,d)^T$9 (Finkenrath et al., 2017). A complementary physical-point analysis found $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$0, $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$1, and heavy-sector matching parameters

$$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$2

after using Osterwalder–Seiler valence quarks and the conditions $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$3 and $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$4 (Alexandrou et al., 2018).

The Osterwalder–Seiler construction became central beyond tuning. In the quark-mass determination, the sea strange and charm quarks are simulated with the nondegenerate twisted-mass action, while the valence strange and charm quarks are regularized as Osterwalder–Seiler fermions in order to avoid unwanted $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$5 strange–charm flavor mixing in physical observables (Alexandrou et al., 2021).

4. Flavor breaking, pion splitting, and phase structure

At finite lattice spacing, twisted mass fermions break isospin symmetry explicitly. The most visible manifestation is the charged–neutral pion splitting, estimated at leading order by

$$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$6

with $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$7 for twisted mass fermions, so that the neutral pion can become substantially lighter than the charged pion (Finkenrath et al., 2017). This effect is not removed by maximal twist; it is an $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$8 artifact. Physical-point studies therefore added a clover term specifically to suppress this splitting and to avoid the unphysical phase structure associated with too light a neutral pion (Finkenrath et al., 2017, Alexandrou et al., 2018).

A representative physical-point $$S_\mathrm{tm}^{\rm F}=a^4\sum_x \bar\chi_x\left[D_{\rm W}+m_0+i\gamma_5\tau_3\mu_q\right]\chi_x,$$9 ensemble with clover improvement measured

$$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$0

corresponding to a pion splitting of about $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$1, explicitly described as much smaller than in earlier twisted-mass ensembles without clover improvement (Alexandrou et al., 2018). In the same study, the $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$2-baryon relative splitting,

$$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$3

was consistent with zero, illustrating that visible isospin-breaking artifacts are highly channel dependent (Alexandrou et al., 2018).

The low-energy phase structure of twisted-mass Wilson fermions has been analyzed extensively with chiral perturbation theory. For nondegenerate $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$4 and $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$5 quarks, the continuum CP-violating phase and the lattice Aoki phase are continuously connected once the $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$6 term and $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$7 discretization effects are included (Horkel et al., 2014). With nonzero $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$8-equivalent phase $$D_{\rm W}=\frac{1}{2}\gamma_\mu(\nabla_\mu+\nabla_\mu^\ast)-\frac{ar}{2}\nabla_\mu\nabla_\mu^\ast,$$9, discretization effects, and twisted mass, the generic result is a first-order transition segment of finite length; only at extremal values $\kappa=(8+2am_0)^{-1}$0 and $\kappa=(8+2am_0)^{-1}$1 do second-order transitions occur (Horkel et al., 2015). At generic nonextremal parameters, pion masses are nonvanishing throughout the phase plane except at the endpoints of the first-order line (Horkel et al., 2015).

At finite isospin chemical potential, the phase diagram changes again. A specific result is that an infinitesimal isospin chemical potential destroys the Aoki phase at zero twist because the massless Goldstone bosons in that phase carry nonzero isospin charge (Janssen et al., 2015). For finite twist angle $\kappa=(8+2am_0)^{-1}$2, only two phases remain: one continuously connected to the Bose-condensed phase and another continuously connected to the normal phase (Janssen et al., 2015).

These analyses delimit the range of validity of physical-point simulations. Maximal twist suppresses $\kappa=(8+2am_0)^{-1}$3 artifacts, but it does not remove the nontrivial vacuum structure associated with Wilson-type flavor breaking.

5. Matrix elements, topology, spectroscopy, and quark masses

Twisted mass fermions have been especially effective in observables with disconnected contributions. In the scalar channel, the operator $\kappa=(8+2am_0)^{-1}$4 becomes in the twisted basis

$\kappa=(8+2am_0)^{-1}$5

and disconnected loops can be estimated through the difference of propagators with opposite Wilson terms. In the $\kappa=(8+2am_0)^{-1}$6 nucleon scalar-matrix-element calculation, this structure enabled a twisted-mass-specific one-end-trick estimator and purely multiplicative renormalization, yielding $\kappa=(8+2am_0)^{-1}$7 at fixed lattice spacing after a linear extrapolation in the pion mass (Dinter et al., 2011). A later dedicated excited-state study, also with maximally twisted mass fermions, found that the strange scalar matrix element is much more fragile: increasing the source–sink separation and statistics changed the plateau estimate by about $\kappa=(8+2am_0)^{-1}$8, and even $\kappa=(8+2am_0)^{-1}$9 fm was not shown to be fully asymptotic (Alexandrou et al., 2012).

In nucleon generalized form factors, twisted mass ensembles were used for $r=1$0, $r=1$1, and a clover-improved physical-pion-mass ensemble. The extraction employed plateau, summation, and two-state methods. On the B55.32 ensemble, disconnected contributions were computed from about 1,500,000 measurements on 4700 configurations; the disconnected part of $r=1$2 was nonzero and about $r=1$3 of the connected part, while the scalar charge required source–sink separations $r=1$4 fm for consistent plateau, summation, and two-state fits (Alexandrou et al., 2013). In the transversity sector, the tensor charge showed no sizable cutoff effects across the lattice spacings studied, with near-physical results $r=1$5 and $r=1$6 when disconnected contributions were neglected (Alexandrou et al., 2013).

The topological susceptibility has been computed with spectral projectors and dynamical twisted mass fermions. For $r=1$7, the mass dependence was compatible with the leading-order chiral prediction

$r=1$8

and the fit yielded

$r=1$9

in the $\omega=\pi/2$0 scheme at $\omega=\pi/2$1 GeV (Cichy et al., 2013). For $\omega=\pi/2$2, continuum extrapolations from three lattice spacings gave

$\omega=\pi/2$3

for fits with $\omega=\pi/2$4, and

$\omega=\pi/2$5

for fits over the full mass range (Cichy et al., 2013). The same study emphasized that typical errors on $\omega=\pi/2$6 were around $\omega=\pi/2$7–$\omega=\pi/2$8, with one high-statistics ensemble reaching about $\omega=\pi/2$9 precision (Cichy et al., 2013).

In spectroscopy, $N_f=2$00 maximally twisted mass fermions at three lattice spacings and three light masses were used for $N_f=2$01, $N_f=2$02, and charmonium spectroscopy. Joint TM and Osterwalder–Seiler continuum fits yielded good agreement with experiment for most ground states, including $N_f=2$03 MeV, $N_f=2$04 MeV, $N_f=2$05 MeV, and $N_f=2$06 MeV (Cichy et al., 2015).

The same formulation underlies precision quark-mass determinations. Using $N_f=2$07 ensembles at three lattice spacings and two physical-pion-mass ensembles, with separate meson- and baryon-sector analyses and their spread folded into the systematics, the final averaged renormalized masses were

$N_f=2$08

together with

$N_f=2$09

in the quoted $N_f=2$10 schemes (Alexandrou et al., 2021).

Twisted mass fermions with clover improvement have also been used in quasi-distribution calculations. Physical-point ensembles with $N_f=2$11 and $N_f=2$12 were employed for quasi-PDFs and exploratory quasi-GPDs, with the renormalized nonlocal matrix element

$N_f=2$13

The matching-scheme dependence was found to be a non-negligible source of uncertainty at one loop, while the integral of the exploratory $N_f=2$14 quasi-GPD gave $N_f=2$15, consistent with the value $N_f=2$16 from the local matrix element at $N_f=2$17 (Alexandrou et al., 2019).

6. Solvers, QCD+QED extensions, and Hamiltonian generalizations

The computational viability of twisted mass fermions at small quark mass depends heavily on solver technology. The DD-$N_f=2$18AMG multigrid method was extended from Wilson-clover fermions to two degenerate twisted-mass flavors by preserving the $N_f=2$19-compatibility of the coarse space and introducing a twisted-mass coarse operator

$N_f=2$20

A TM-specific modification,

$N_f=2$21

was shown to be crucial: on the physical ensemble $N_f=2$22, tuning to $N_f=2$23 reduced coarse-grid iterations by about an order of magnitude and improved the time to solution by a factor of about $N_f=2$24 (Bacchio et al., 2016). In a detailed performance study, the same framework yielded an inversion time of $N_f=2$25 core-hours versus $N_f=2$26 core-hours for CG on the physical ensemble, corresponding to a speed-up of about $N_f=2$27 for the solve alone (Alexandrou et al., 2016). The method was integrated into tmLQCD and used during HMC simulations at the physical point, including force-term, heat-bath, and acceptance-step inversions (Bacchio et al., 2016).

Twisted mass fermions have also been adapted to QCD+QED. A maximally twisted Wilson regularization for fully unquenched QCD+QED was constructed so that the continuum theory has vanishing effective $N_f=2$28-term, even though axial rotations are anomalous. In that framework, the QED-shifted critical masses are fixed by restoring continuum Ward identities, and a mixed-action setup allows RM123 computations of leading isospin-breaking effects on ETMC $N_f=2$29 pure-QCD ensembles with only $N_f=2$30 artifacts (Frezzotti et al., 2016).

Beyond Euclidean lattice QCD, twisted mass fermions were studied in the Hamiltonian formulation of the two-flavor Schwinger model with matrix product states. That analysis confirmed the expected $N_f=2$31 improvement in the free theory and found that it persists numerically in the interacting case. After tuning to maximal twist with an electric-field-based mass-renormalization method, the pion mass showed rapid continuum convergence; the charged–neutral pion splitting was significant, around $N_f=2$32 at the lattice spacing studied, and vanished toward the continuum (Schwägerl et al., 2 Sep 2025). The same work reported milder finite-volume dependence for twisted mass fermions than for staggered and Wilson discretizations in that benchmark model (Schwägerl et al., 2 Sep 2025).

Taken together, these developments define twisted mass fermions as a Wilson-based regularization whose central structural feature is the maximal-twist cancellation of $N_f=2$33 artifacts, but whose practical use depends equally on the control of $N_f=2$34 flavor breaking, phase-structure constraints, and solver technology.