Lattice fermions and the doubling problem

Intuition Beginner

Suppose you want to simulate a single quantum particle — an electron, say — on a computer. You cannot store space as a smooth continuum, so you replace it with a grid of points, a lattice, and you approximate the rate of change of a field by comparing its values at neighbouring grid points. For most equations this works fine. For the equation that governs electrons, the Dirac equation, the naive grid version misbehaves badly: instead of one electron, the lattice describes a whole flock of them.

Where do the extra particles come from? The Dirac equation involves a first derivative, and the most symmetric way to approximate a first derivative on a grid is to subtract the value two steps to the left from the value two steps to the right. That subtraction is blind to a fast zig-zag pattern that flips sign at every site: such a pattern looks motionless to the grid even though it is wildly varying. Each independent zig-zag direction adds one phantom particle, so a $d$-dimensional grid hides $2^d$ of them. This is the doubling problem.

These phantoms — called doublers — are not a bug you can quietly delete. A deep result says that any honest, local recipe with the right symmetry will always grow them. So lattice physicists do something cleverer: they add a small correction that makes the phantoms very heavy, so heavy they drop out of the low-energy physics, leaving the one electron you wanted.

Visual Beginner

The picture shows a one-dimensional grid of sites. A smooth wave (the physical electron) varies slowly from site to site. Beneath it, a zig-zag wave flips sign at every site. The symmetric two-step difference used to approximate the derivative assigns both waves the same slope, so the grid cannot tell them apart — the zig-zag is the doubler.

The takeaway in one image: the symmetric difference has a blind spot at the fastest wiggle the grid supports, and that blind spot is exactly where each doubler lives.

Worked example Beginner

Work on a one-dimensional grid with spacing $a=1$, and approximate the derivative of a field $\psi$ at site $x$ by the symmetric difference $D\psi (x)=\frac{1}{2}(\psi (x+1)-\psi (x-1))$.

Feed it the slow wave $\psi (x)=1$ for all $x$ (a constant). Then $D\psi (x)=\frac{1}{2}(1-1)=0$. A constant has zero slope: correct.

Now feed it the zig-zag $\psi (x)=(-1{)}^x$, which reads $1,-1,1,-1,\dots$. At an even site, $\psi (x+1)=-1$ and $\psi (x-1)=-1$, so $D\psi (x)=\frac{1}{2}(-1-(-1))=0$. The fastest-varying field the grid can hold is assigned slope zero, the same answer as the constant.

So two completely different fields — one flat, one flipping at every step — both look like a zero-momentum, slowly-drifting particle to this difference rule. The grid sees two particles where the continuum has one.

What this tells us: the symmetric difference vanishes at both ends of the grid's momentum range, not just at zero. Every extra vanishing point is a doubler, and counting them in $d$ dimensions gives the factor $2^d$.

Check your understanding Beginner

Formal definition Intermediate+

Work on a $d$-dimensional cubic lattice $\Lambda =(a\mathbb{Z}{)}^d$ with spacing $a$, sites $x$, and unit lattice vectors $\hat{\mu }$. A Dirac field assigns to each site an anticommuting (Grassmann) spinor ${\psi }_x\in {\mathbb{C}}^{2^{\lfloor d/2\rfloor }}$ together with ${\bar{\psi }}_x$, living in the fermionic algebra of 08.14.02. Let ${\gamma }_{\mu }$ be the Euclidean gamma matrices, $\{{\gamma }_{\mu },{\gamma }_{\nu }\}=2{\delta }_{\mu \nu }$.

The naive lattice Dirac action replaces the continuum derivative by the symmetric difference: $$ S_{\text{naive}} = a^d \sum_x \left[ \sum_\mu \bar\psi_x, \gamma_\mu, \frac{\psi_{x+\hat\mu} - \psi_{x-\hat\mu}}{2a} + m, \bar\psi_x \psi_x \right]. $$ Passing to momentum space with ${\psi }_x={\int }_{\text{BZ}}\frac{d^{d}p}{(2\pi {)}^d}{e}^{ip\cdot x}\tilde{\psi }(p)$ over the Brillouin zone $p_{\mu }\in (-\pi /a,\pi /a]$, the action becomes diagonal with the lattice Dirac operator $$ D(p) = \frac{i}{a}\sum_\mu \gamma_\mu \sin(p_\mu a) + m . $$ The inverse $D(p{)}^{-1}$ is the lattice propagator. The doublers are the poles of the massless propagator: the points $p$ where $\sin (p_{\mu }a)=0$ for every $\mu$, namely $p_{\mu }\in \{0,\pi /a\}$. There are $2^d$ such corners of the Brillouin zone, and near each the operator linearises to a continuum Dirac operator, so the naive action describes $2^d$ degenerate Dirac species rather than one.

The Wilson action adds a second-difference term with parameter $r$ (usually $r=1$): $$ S_W = S_{\text{naive}} - \frac{r a^{d-1}}{2}\sum_{x,\mu} \bar\psi_x, (\psi_{x+\hat\mu} - 2\psi_x + \psi_{x-\hat\mu}), $$ whose momentum-space effect is to shift the mass to the momentum-dependent $$ M(p) = m + \frac{2r}{a}\sum_\mu \sin^2!\frac{p_\mu a}{2}. $$ A doubler at a corner with $k$ components equal to $\pi /a$ acquires mass $m+2rk/a$, so all $2^d-1$ doublers become heavy as $a\to 0$ while the physical mode at $p=0$ keeps mass $m$.

Counterexamples to common slips

• Setting the Wilson parameter $r=0$ returns the naive action and the doublers reappear; $r\ne 0$ is essential. The Wilson term is dimension-five (an irrelevant operator), so it does not alter the continuum limit of the physical mode, yet it is decisive for the doublers because it scales as $1/a$ at the zone corners.
• The Wilson term anticommutes with no ${\gamma }_{\mu }$ and so commutes with ${\gamma }_5$: it breaks chiral symmetry explicitly, $\{D_W,{\gamma }_5\}\ne 0$ at $m=0$. Forgetting this and expecting a chiral limit at fixed bare $m=0$ is the standard error; the chiral point must be tuned (the additive mass renormalisation).
• Staggered fermions do not remove doublers; they redistribute them. The $2^d$ doublers become $2^{d/2}$ "tastes," and treating those tastes as physically distinct flavours (rather than a discretisation artefact requiring a rooting prescription) is a frequent misreading.

Key theorem with proof Intermediate+

Theorem (Naive lattice fermions describe $2^d$ species). The massless naive lattice Dirac operator $D(p)=\frac{i}{a}{\sum }_{\mu }{\gamma }_{\mu }\sin (p_{\mu }a)$ on the Brillouin zone $p_{\mu }\in (-\pi /a,\pi /a]$ has exactly $2^d$ zeros, located at $p_{\mu }\in \{0,\pi /a\}$, and near each zero the operator reduces to a continuum massless Dirac operator. Hence the naive action propagates $2^d$ Dirac fermions.

Proof. The operator $D(p)$ is a sum of gamma matrices with real coefficients $\frac{1}{a}\sin (p_{\mu }a)$. Because $\{{\gamma }_{\mu },{\gamma }_{\nu }\}=2{\delta }_{\mu \nu }$, one has $D(p{)}^{†}D(p)=\frac{1}{a^2}{\sum }_{\mu }{\sin }^2(p_{\mu }a)\cdot 1$. Therefore $D(p)$ is singular (has a zero eigenvalue) precisely where ${\sin }^2(p_{\mu }a)=0$ for every $\mu$ simultaneously, i.e. $\sin (p_{\mu }a)=0$ for all $\mu$.

On the half-open interval $(-\pi /a,\pi /a]$ the equation $\sin (p_{\mu }a)=0$ has the two solutions $p_{\mu }=0$ and $p_{\mu }=\pi /a$. Independently across the $d$ components, the joint zero set is the product $\{0,\pi /a{\}}^d$, which contains $2^d$ points: the corners of the Brillouin zone.

Fix one such corner $p^{(\alpha )}$ and write $p=p^{(\alpha )}+k$ with $k$ small. For each component, $\sin ((p_{\mu }^{(\alpha )}+k_{\mu })a)=\cos (p_{\mu }^{(\alpha )}a)\sin (k_{\mu }a)=s_{\mu }^{(\alpha )}(k_{\mu }a)+O(k^3{a}^3)$, where $s_{\mu }^{(\alpha )}=\cos (p_{\mu }^{(\alpha )}a)=\pm 1$ according to whether $p_{\mu }^{(\alpha )}$ is $0$ or $\pi /a$. Substituting, $$ D(p^{(\alpha)} + k) = i\sum_\mu \big(s_\mu^{(\alpha)}\gamma_\mu\big), k_\mu + O(k^3). $$ Define ${\gamma }_{\mu }^{\prime }=s_{\mu }^{(\alpha )}{\gamma }_{\mu }$. Since each $s_{\mu }^{(\alpha )}=\pm 1$, the ${\gamma }_{\mu }^{\prime }$ satisfy the identical Clifford relations $\{{\gamma }_{\mu }^{\prime },{\gamma }_{\nu }^{\prime }\}=2{\delta }_{\mu \nu }$, so $i{\sum }_{\mu }{\gamma }_{\mu }^{\prime }{k}_{\mu }$ is a continuum massless Dirac operator in the variable $k$. Each of the $2^d$ corners thus supplies one continuum Dirac species, and the species are degenerate because the leading operator is the same up to the field redefinition ${\gamma }_{\mu }\mapsto {\gamma }_{\mu }^{\prime }$. $\square$

Bridge. This counting is the foundational reason a chiral lattice theory is so hard to build, and it builds toward the Nielsen-Ninomiya no-go theorem below, which promotes the $2^d$ accident into a topological obligation. The sign flips $s_{\mu }^{(\alpha )}=\pm 1$ are not cosmetic: half of the corners carry one handedness and half the other, so summed over the Brillouin zone the net chirality vanishes — this is exactly the cancellation the no-go theorem forces. The bridge is that the analytic statement "the symbol $D(p)$ has $2^d$ zeros" generalises to the topological statement "a smooth chiral charge on the Brillouin torus has vanishing total degree," and putting these together identifies the doubling count with an index. The same Clifford-algebra bookkeeping that fixes $2^{\lfloor d/2\rfloor }$ spinor components in 03.09.02 is dual to the bookkeeping that fixes the $2^d$ doublers here, and the central insight — that locality plus chirality plus the lattice forces extra fermions — appears again in the staggered and overlap constructions of the Master tier.

Exercises Intermediate+

Advanced results Master

The Nielsen-Ninomiya theorem elevates the doubling count to a topological obstruction. Consider a lattice Hamiltonian $H(p)$ for a single Weyl fermion, defined on the Brillouin torus $T^d=(\mathbb{R}/(2\pi /a)\mathbb{Z}{)}^d$. Locality of the action forces $H(p)$ to be a smooth (indeed analytic) periodic function of $p$; Hermiticity makes its eigenvalues real; and a Weyl fermion is a band touching where two bands cross with a definite chirality $\chi =\pm 1$, the sign of the Jacobian of the map from a small sphere around the crossing to the sphere of Hamiltonian directions. The theorem states that for any such lattice realisation the chiralities of all band-touching points sum to zero, ${\sum }_{\alpha }{\chi }_{\alpha }=0$ ^{[Nielsen-Ninomiya 1981]}. Consequently a single Weyl fermion cannot live on the lattice in isolation: it is always accompanied by a partner of opposite chirality.

The proof is a degree argument. Near a generic crossing in $d$ dimensions the Hamiltonian linearises to $H\approx {\sum }_{\mu }{v}_i^{\mu }(p-p_{\alpha }{)}_{\mu }{\sigma }^i$ (for the $d=3$ Weyl case, with Pauli matrices ${\sigma }^i$), and the chirality ${\chi }_{\alpha }=\mathrm{sgn}\det (v_i^{\mu })$ is the degree of the Gauss map $S^{d-1}\to S^{d-1}$, $\hat{n}\mapsto H(p_{\alpha }+ϵ\hat{n})/\Vert H\Vert$. The total chirality is the sum of these local degrees. Because the Brillouin zone is a torus — a closed manifold without boundary — the relevant differential form (a Berry curvature in $d=3$, a more general index density in higher $d$) integrates to zero by Stokes, exactly as a vector field on a torus has zero total index by Poincaré-Hopf. The zero net degree forces the crossings to pair up. The original 1981 papers gave first a homotopy-theoretic proof (part I) and then an intuitive topological proof in terms of the conservation of a fermionic charge under spectral flow (part II).

Wilson's resolution sacrifices chiral symmetry. The Wilson operator $D_W=D_{\text{naive}}+\frac{r}{2}a\square$, with $\square$ the lattice Laplacian, satisfies $\{D_W,{\gamma }_5\}=2m{\gamma }_5+ra{\gamma }_5\square \ne 0$ at $m=0$, so the no-go hypothesis of exact chiral symmetry fails and the doublers may be lifted ^{[Wilson 1977]}. The cost is real: an additive mass renormalisation (the bare mass must be tuned to a critical $m_c(g)$ that is no longer protected by symmetry), an explicit $O(a)$ discretisation error, and the loss of a clean lattice chirality. Kogut and Susskind instead diagonalised the spin structure, distributing the four Dirac components over the corners of an elementary hypercube; this keeps a continuous $U(1{)}_{ϵ}$ chiral symmetry that protects the mass, at the price of four degenerate "tastes" and the rooting controversy ^{[Kogut-Susskind 1975]}.

The modern resolution deforms chirality rather than abandoning it. Ginsparg and Wilson observed that a lattice Dirac operator obeying $D{\gamma }_5+{\gamma }_{5}D=aD{\gamma }_{5}D$ realises a lattice-deformed chiral symmetry under $\delta \psi ={\gamma }_5(1-\frac{a}{2}D)\psi$, $\delta \bar{\psi }=\bar{\psi }(1-\frac{a}{2}D){\gamma }_5$ ^{[Ginsparg-Wilson 1982]}. An operator $D$ satisfying this relation can have a single light fermion with an exact lattice chiral symmetry and a correct continuum index — the lattice index theorem $\mathrm{index}(D)=-\frac{a}{2}\mathrm{Tr}({\gamma }_{5}D)$ reproduces the topological charge. The no-go theorem is evaded because the symmetry algebra is the deformed one, not the undeformed $\{D,{\gamma }_5\}=0$. Neuberger's overlap operator and the equivalent domain-wall construction give explicit solutions.

Synthesis. The doubling problem is the foundational reason that putting chiral matter on a lattice is a topological, not merely numerical, problem, and the no-go theorem identifies the $2^d$ count with the vanishing of a net chiral charge on the Brillouin torus. This is exactly the same Poincaré-Hopf zero-total-index phenomenon that governs vector fields on closed manifolds, and it generalises the naive zero-counting of the Key theorem into a statement about degrees of Gauss maps. The central insight is that locality, Hermiticity, translation invariance, and exact chiral symmetry are jointly inconsistent on the lattice, so every practical scheme — Wilson, staggered, overlap — relaxes exactly one hypothesis: Wilson drops chiral symmetry outright, staggered keeps only a remnant $U(1)$ at the cost of tastes, and the Ginsparg-Wilson route deforms the symmetry algebra. Putting these together, the bridge to continuum physics is that all three schemes flow to the same chiral gauge theory as $a\to 0$, and the lattice index theorem identifies the deformed-chiral anomaly with the continuum axial anomaly, so the no-go obstruction is dual to the very anomaly it was forced by.

Full proof set Master

Proposition (Net lattice chirality vanishes for the naive operator). Let the naive massless lattice Dirac operator have its $2^d$ zeros at the corners $p^{(\alpha )}\in \{0,\pi /a{\}}^d$, and assign each the chirality ${\chi }_{\alpha }={\prod }_{\mu =1}^d\cos (p_{\mu }^{(\alpha )}a)$. Then ${\sum }_{\alpha }{\chi }_{\alpha }=0$ for every $d\ge 1$.

Proof. The corner set factorises as a $d$-fold Cartesian product, and the chirality is a product over components, so the total sum factorises: $$ \sum_{\alpha\in{0,\pi/a}^d} \prod_{\mu=1}^d \cos(p^{(\alpha)}\mu a) = \prod{\mu=1}^d \Big( \sum_{p_\mu \in {0,\pi/a}} \cos(p_\mu a) \Big) = \prod_{\mu=1}^d \big( \cos 0 + \cos\pi \big). $$ Each factor equals $1+(-1)=0$, so the whole product is $0$. The cancellation is exact and dimension-independent, and it pairs each chirality $+1$ corner with a $-1$ corner. $\square$

Proposition (Wilson term lifts all doublers and leaves one light mode). With Wilson parameter $r>0$ and bare mass $m$, the pole mass of the mode at a corner with $k$ components equal to $\pi /a$ is $M_k=m+2rk/a$. For $m$ of order unity and $a\to 0$, only the $k=0$ mode remains light; all $k\ge 1$ modes have mass diverging as $1/a$.

Proof. The Wilson momentum-space mass is $M(p)=m+\frac{2r}{a}{\sum }_{\mu }{\sin }^2(p_{\mu }a/2)$. At a corner with $p_{\mu }=\pi /a$ for $k$ values of $\mu$ and $p_{\mu }=0$ for the rest, each $\pi /a$ slot contributes ${\sin }^2(\pi /2)=1$ and each $0$ slot contributes ${\sin }^2(0)=0$. Hence ${\sum }_{\mu }{\sin }^2(p_{\mu }a/2)=k$ and $M_k=m+2rk/a$. The $k=0$ corner ($p=0$) keeps $M_0=m$, finite. For each $k\ge 1$, $M_k=m+2rk/a\to \infty$ as $a\to 0$, so those $2^d-1$ doublers decouple from the continuum spectrum. The physical mode near $p=0$ has propagator $D_W(k{)}^{-1}\approx (i{\sum }_{\mu }{\gamma }_{\mu }{k}_{\mu }+m{)}^{-1}+O(a)$, the continuum Dirac propagator. $\square$

Proposition (Ginsparg-Wilson operators carry a lattice chiral symmetry). If $D$ satisfies $D{\gamma }_5+{\gamma }_{5}D=aD{\gamma }_{5}D$, then the action $S=\bar{\psi }D\psi$ is invariant to first order under the deformed transformation $\delta \psi ={\gamma }_5(1-\frac{a}{2}D)\psi$, $\delta \bar{\psi }=\bar{\psi }(1-\frac{a}{2}D){\gamma }_5$.

Proof. Compute the variation $\delta S=\delta \bar{\psi }D\psi +\bar{\psi }D\delta \psi$: $$ \delta S = \bar\psi(1 - \tfrac{a}{2}D)\gamma_5 D\psi + \bar\psi D\gamma_5(1 - \tfrac{a}{2}D)\psi = \bar\psi\Big[ \gamma_5 D + D\gamma_5 - \tfrac{a}{2}\big(D\gamma_5 D + D\gamma_5 D\big)\Big]\psi. $$ The bracket is $({\gamma }_{5}D+D{\gamma }_5)-aD{\gamma }_{5}D$, which vanishes precisely by the Ginsparg-Wilson relation. Hence $\delta S=0$ and the deformed transformation is an exact symmetry of the lattice action. The undeformed chiral algebra is recovered as $a\to 0$, where the relation degenerates to $\{D,{\gamma }_5\}=0$. $\square$

Connections Master

• The link variables and gauge-invariant plaquette action of Wilson lattice gauge theory 08.08.01 supply the gauge field; coupling it to the lattice fermions defined here — by inserting the link $U_{\mu }(x)$ between ${\bar{\psi }}_x$ and ${\psi }_{x+\hat{\mu }}$ in the hopping term — assembles the full lattice QCD action. That unit owns the pure-gauge sector; this unit owns the matter sector that the no-go theorem complicates.

• The Grassmann integration of 08.14.02 is the measure in which these lattice fermions are defined: the fermionic partition function $\int {\prod }_{x}d{\bar{\psi }}_{x}d{\psi }_x{e}^{-\bar{\psi }D\psi }=\det D$ is exactly the determinant identity proved there, now with $D$ the lattice Dirac operator rather than the Ising transfer matrix. The doubler count is the multiplicity of zeros of this $\det D$ in the massless limit.

• The fermionic Fock space and anticommutation relations of 08.10.09 are the operator content on which the Hamiltonian (Kogut-Susskind) lattice formulation acts; the spin-diagonalisation that produces staggered "tastes" is a reorganisation of those creation and annihilation operators across a unit hypercube.

• The order-disorder and duality structure that organises the two-dimensional fermionic critical point appears again in 08.14.05, where the same lattice fermion bilinears carry the Kramers-Wannier and continuum-limit data; the chiral symmetry whose lattice fate is decided here is the symmetry whose breaking pattern that unit tracks.

Historical & philosophical context Master

Wilson introduced the lattice as a non-perturbative regulator of gauge theory in 1974, and in the same circle of work added the second-difference term — now the Wilson term — to remove the fermion doublers, presenting the lattice-fermion construction explicitly in the 1975 Erice lectures published in 1977 ^{[Wilson 1977]}. Kogut and Susskind gave the Hamiltonian lattice formulation in 1975 and the spin-diagonalised (staggered) treatment of fermions that trades doublers for tastes ^{[Kogut-Susskind 1975]}. The structural obstruction was made precise by Nielsen and Ninomiya in two 1981 papers in Nuclear Physics B: a homotopy-theoretic proof followed by an intuitive topological argument that the net chirality of any local, Hermitian, translation-invariant lattice fermion must vanish, so a single Weyl species — and with it a naive lattice formulation of the electroweak chiral gauge theory — is forbidden ^{[Nielsen-Ninomiya 1981]}. The deformation that ultimately reconciled chirality with the lattice came from Ginsparg and Wilson in 1982, whose relation lay largely dormant for fifteen years until Neuberger's overlap operator and the domain-wall construction realised it concretely ^{[Ginsparg-Wilson 1982]}.

Bibliography Master

@article{NielsenNinomiya1981a,
  author  = {Nielsen, H. B. and Ninomiya, M.},
  title   = {Absence of Neutrinos on a Lattice: (I). Proof by Homotopy Theory},
  journal = {Nuclear Physics B},
  volume  = {185},
  year    = {1981},
  pages   = {20--40}
}

@article{NielsenNinomiya1981b,
  author  = {Nielsen, H. B. and Ninomiya, M.},
  title   = {Absence of Neutrinos on a Lattice: (II). Intuitive Topological Proof},
  journal = {Nuclear Physics B},
  volume  = {193},
  year    = {1981},
  pages   = {173--194}
}

@article{Wilson1974Confinement,
  author  = {Wilson, Kenneth G.},
  title   = {Confinement of Quarks},
  journal = {Physical Review D},
  volume  = {10},
  year    = {1974},
  pages   = {2445--2459}
}

@incollection{Wilson1977Erice,
  author    = {Wilson, Kenneth G.},
  title     = {Quarks and Strings on a Lattice},
  booktitle = {New Phenomena in Subnuclear Physics (Erice 1975)},
  editor    = {Zichichi, A.},
  publisher = {Plenum Press},
  address   = {New York},
  year      = {1977},
  pages     = {69--142}
}

@article{KogutSusskind1975,
  author  = {Kogut, John and Susskind, Leonard},
  title   = {Hamiltonian Formulation of Wilson's Lattice Gauge Theories},
  journal = {Physical Review D},
  volume  = {11},
  year    = {1975},
  pages   = {395--408}
}

@article{GinspargWilson1982,
  author  = {Ginsparg, Paul H. and Wilson, Kenneth G.},
  title   = {A Remnant of Chiral Symmetry on the Lattice},
  journal = {Physical Review D},
  volume  = {25},
  year    = {1982},
  pages   = {2649--2657}
}

@book{ItzyksonDrouffe1989Vol1,
  author    = {Itzykson, Claude and Drouffe, Jean-Michel},
  title     = {Statistical Field Theory, Volume 1},
  publisher = {Cambridge University Press},
  year      = {1989}
}

@book{MontvayMunster1994,
  author    = {Montvay, Istvan and M\"unster, Gernot},
  title     = {Quantum Fields on a Lattice},
  publisher = {Cambridge University Press},
  year      = {1994}
}

@article{Neuberger1998Overlap,
  author  = {Neuberger, Herbert},
  title   = {Exactly Massless Quarks on the Lattice},
  journal = {Physics Letters B},
  volume  = {417},
  year    = {1998},
  pages   = {141--144}
}