Naive Fermions and Species Doubling

The symmetric nearest-neighbor discretization of the Dirac operator is local, translation invariant, and exactly chirally symmetric at zero mass, but its momentum symbol vanishes at every corner of the Brillouin zone. In even Euclidean dimension $d$, those $2^d$ zeros become $2^d$ relativistic continuum species with alternating chirality signs. Expanding only about $p=0$ therefore gives a locally correct dispersion relation and a globally wrong particle count.

Required background. Lattice Momentum, Propagators, and Cutoff Dispersion supplies the Brillouin zone and discrete Fourier transform used to locate poles. Grassmann Functional Integrals for Free Fermions supplies the quadratic fermion action and matrix-valued propagator.

Helpful background. Weyl Fields and Chirality supplies the continuum chirality operator whose sign is compared between lattice zeros.

The naive Dirac operator on the full Brillouin zone

Local regulator and convention card. We use a $d$-dimensional Euclidean hypercubic lattice $x=a n$, $n\in\mathbb Z^d$, with Hermitian matrices $\{\gamma_\mu,\gamma_\nu\}=2\delta_{\mu\nu}$. The free benchmark has periodic boundaries and momenta in $\mathcal B=(-\pi/a,\pi/a]^d$. For even $d$, $\gamma_*$ is Hermitian, squares to one, and anticommutes with every $\gamma_\mu$. Masses and momenta below are dimensionful. These choices extend the site’s global conventions; an interacting covariant difference is introduced only after the free zero count is fixed.

The centered difference

$$\nabla_\mu^{\mathrm s}\psi(x) =\frac{\psi(x+a\hat\mu)-\psi(x-a\hat\mu)}{2a}$$

gives the Euclidean action

$$S_{\mathrm n} =a^d\sum_x \bar\psi(x) \left(m+\sum_{\mu=1}^d\gamma_\mu\nabla_\mu^{\mathrm s}\right)\psi(x).$$

With $\psi(x)=V^{-1/2}\sum_{p\in\mathcal B}e^{ip\cdot x}\widetilde\psi(p)$, a translation by $\pm a\hat\mu$ contributes $e^{\pm iap_\mu}$. The exact finite-spacing symbol is therefore

$$D_{\mathrm n}(p) =m+\frac{i}{a}\sum_{\mu=1}^d\gamma_\mu\sin(ap_\mu),$$

and its inverse away from a zero is

$$S_{\mathrm n}(p) =\frac{m-\frac{i}{a}\sum_\mu\gamma_\mu\sin(ap_\mu)} {m^2+\frac{1}{a^2}\sum_\mu\sin^2(ap_\mu)}.$$

The denominator is a useful independent check: every term is nonnegative for real $m$ and $p$, so a massless pole occurs precisely when every $\sin(ap_\mu)$ vanishes. The symmetric derivative has the correct small-$p$ limit because $a^{-1}\sin(ap_\mu)=p_\mu+O(a^2p_\mu^3)$, but that local expansion says nothing about other points in the compact zone.

Every corner is a relativistic species

At $m=0$, label a Brillouin-zone corner by a bit vector $r\in\{0,1\}^d$,

$$p_\mu^{(r)}=\frac{\pi r_\mu}{a}.$$

There are $2^d$ such momenta. Put $p=p^{(r)}+q$ with $|a q_\mu|\ll1$. Since

$$\sin(\pi r_\mu+a q_\mu) =(-1)^{r_\mu}a q_\mu+O(a^3q_\mu^3),$$

the operator linearizes as

$$D_{\mathrm n}\bigl(p^{(r)}+q\bigr) =i\sum_\mu\gamma_\mu^{(r)}q_\mu+O(a^2q^3), \qquad \gamma_\mu^{(r)}=(-1)^{r_\mu}\gamma_\mu.$$

The matrices $\gamma_\mu^{(r)}$ obey the same Euclidean Clifford algebra, so each corner is not merely a zero: it has the linear Dirac dispersion of a continuum fermion. Its chirality matrix is

$$\gamma_*^{(r)} =(-1)^{r_1+\cdots+r_d}\gamma_*.$$

Consequently half the corners have the original chirality orientation and half have the opposite orientation. The total signed chirality is

$$\sum_{r\in\{0,1\}^d}(-1)^{\sum_\mu r_\mu} =(1-1)^d=0.$$

This cancellation is the concrete free-field precursor of the Nielsen–Ninomiya obstruction; the homotopy and topological-current arguments are developed in Nielsen and Ninomiya 1981, Nuclear Physics B 185, pp. 20–40 and Nielsen and Ninomiya 1981, Nuclear Physics B 193, pp. 173–194. The general theorem needs hypotheses beyond this particular nearest-neighbor action; the calculation here establishes the example, not the theorem.

An exactly checkable corner count

In two dimensions the entire calculation fits in four rows:

Corner $(ap_1,ap_2)$	Effective matrices	Chirality sign
$(0,0)$	$(\gamma_1,\gamma_2)$	$+1$
$(\pi,0)$	$(-\gamma_1,\gamma_2)$	$-1$
$(0,\pi)$	$(\gamma_1,-\gamma_2)$	$-1$
$(\pi,\pi)$	$(-\gamma_1,-\gamma_2)$	$+1$

Thus $d=2$ has four species, two of each chirality sign. In $d=4$ there are sixteen corners: the number with exactly $k$ components equal to $\pi/a$ is $\binom{4}{k}$, giving eight positive and eight negative signs. A periodic even-$N$ lattice contains both $0$ and $\pi/a$ on its discrete momentum grid, so the $N=8,12,16$ free fixtures all reproduce the same four zeros in two dimensions. This ladder tests the finite-volume momentum grid; because $a$ is held fixed, it is not a continuum extrapolation.

Antiperiodic temporal boundary conditions move the allowed temporal momenta away from both temporal corners. They can remove exact finite-volume zero eigenvalues without removing the doubled continuum branches. Counting only exact zeros after changing boundary conditions is therefore not a species test.

Chiral symmetry does not remove the extra modes

For $m=0$,

$$\{\gamma_*,D_{\mathrm n}\}=0,$$

so the action is invariant under the global axial rotation

$$\psi\longmapsto e^{i\alpha\gamma_*}\psi, \qquad \bar\psi\longmapsto\bar\psi e^{i\alpha\gamma_*}.$$

The same anticommutation holds at every corner. Exact naive chiral symmetry therefore protects all doublers together; it does not select the mode near the origin. This is why “exact chiral symmetry at nonzero $a$” and “one light fermion” cannot be treated as independent checkboxes for this operator.

In this page, species means the distinct relativistic poles of the full lattice propagator. Taste is reserved for the internal multiplicity obtained after the staggered reorganization. The two words can count related low-energy degrees of freedom, but they refer to different descriptions and should not be exchanged before the reconstruction is specified.

Gauge covariance replaces the two shifts by link transporters,

$$\nabla_\mu^{\mathrm s}\psi(x) \longmapsto \frac{U_\mu(x)\psi(x+a\hat\mu) -U_\mu^\dagger(x-a\hat\mu)\psi(x-a\hat\mu)}{2a}.$$

This preserves the local centered structure, but it removes momentum as a configuration-by-configuration quantum number. The free corner count remains the analytic benchmark; interacting claims require spectra, Ward identities, and continuum tests rather than a literal corner plot.

The obstruction map below turns the corner calculation into a topological statement. Inspect the alternating charges of the four two-dimensional zeros and the lower branches: every practical formulation changes a named finite-regulator premise rather than simply deleting unwanted modes.

The two-dimensional free symbol has four zeros with charges $+1,-1,-1,+1$, so the signed sum vanishes. Wilson, staggered, Ginsparg–Wilson, and domain-wall formulations alter different finite-regulator premises; a chiral-gauge target additionally requires anomaly cancellation and a globally integrable Weyl measure. The diagram is schematic and is not a cost or accuracy ranking.

Observable-level validation checklist

A calculation that claims to reproduce the naive free theory should pass all of the following observable-level checks:

• enumerate the full discrete Brillouin zone, not only $|p_\mu|\ll1$, and recover $2^d$ massless branches for periodic even extents;
• fit the dispersion about every corner and recover unit slope after the sign redefinition $\gamma_\mu\to\gamma_\mu^{(r)}$;
• verify the binomial multiplicities $\binom{d}{k}$ and the vanishing signed chirality sum;
• distinguish a boundary-condition shift of exact eigenvalues from removal of a low-energy branch;
• show that a nonzero mass opens the same mass gap at every naive corner; and
• for an interacting use, match the free limit and report which spectral or Ward-identity observable replaces momentum-space zero counting.

Adversarial failure. Suppose a code samples only $|ap_\mu|<\pi/2$ and finds one linear pole. Every value in that restricted window can agree with the continuum expansion while fifteen four-dimensional species remain outside it. The cure is not a finer fit at the origin; it is a zone-wide zero and degeneracy search.

You should now be able to (1) locate and linearize every zero of $D_{\mathrm n}(p)$ in arbitrary even dimension and (2) explain, using the compact Brillouin zone, why a correct small-momentum propagator does not establish the intended flavor content.

Exercises

Count corners with a fixed chirality sign

Show directly that a $d=2\ell$ dimensional naive fermion has $2^{d-1}$ corners of each chirality sign.

Solution

Choose the first $d-1$ bits of $r$ freely. There is exactly one choice of the last bit that makes $\sum_\mu r_\mu$ even and exactly one that makes it odd. Hence each sign occurs $2^{d-1}$ times. Equivalently, the even and odd binomial sums are both $2^{d-1}$ because $(1+1)^d=2^d$ and $(1-1)^d=0$.

Diagnose an odd lattice extent

For a periodic one-dimensional lattice with odd $N$, the allowed momenta are $p_n=2\pi n/(Na)$. Explain why $p=\pi/a$ is absent but a second continuum branch still appears as $N\to\infty$.

Solution

No integer $n$ solves $2\pi n/N=\pi$ when $N$ is odd, so there is no exact finite-$N$ zero at the zone edge. The nearest allowed momenta differ from $\pi/a$ by $\pi/(Na)$. Their eigenvalues scale as $a^{-1}\sin(\pi\mp\pi/N)\simeq\pi/(Na)$ and therefore approach zero as the physical extent $L=Na$ grows. The branch was shifted by the grid, not removed from the operator.

Where the calculation goes next

The Nielsen–Ninomiya theorem turns this example into an assumption-by-assumption obstruction. Wilson and Clover Fermions lift the corners by breaking the naive chiral anticommutation, while Staggered Fermions and Taste reorganize rather than eliminate the multiplicity.

References

• Nielsen, Holger Bech, and Masao Ninomiya. “Absence of Neutrinos on a Lattice. I. Proof by Homotopy Theory.” Nuclear Physics B 185 (1981): 20–40; erratum 195 (1982): 541. DOI.
• Nielsen, Holger Bech, and Masao Ninomiya. “Absence of Neutrinos on a Lattice. II. Intuitive Topological Proof.” Nuclear Physics B 193 (1981): 173–194. DOI.