Symmetry-Protected Operators, Currents, and Improvement

A conserved current is protected only after three statements have been separated: the quantum Ward identity is exact, its charge is normalized to generate a specified symmetry, and the local current has been defined modulo improvements and redundant operators. Conservation alone does not make an arbitrary representative unique, while a zero diagonal entry in one anomalous-dimension matrix does not establish protection.

This page turns Ward identities into a test on the operator-mixing matrix. The clean example is the global $O(N)$ current of scalar theory, whose normalized current class has zero anomalous dimension. The stress tensor then shows why exact charge normalization coexists with finite improvement freedom. A QED example supplies the essential caution: gauge fields provide an identically conserved operator with which a conserved current can mix.

Required background. Operator Mixing and Renormalization Matrices supplies closed sectors and redundant directions; Quantum Currents, Improvements, and Conservation supplies the current and charge construction; and Operator Anomalous-Dimension Matrices fixes the convention $O_0=ZO$ and $\mathcal D O=-\gamma O$.

Helpful background. Localized Transformations and Ward–Takahashi Identities derives the contact terms generated by a local symmetry parameter. What Is an Anomaly? explains when no choice of local counterterms can preserve all desired identities.

Protection as a Ward-identity statement

Let $j_A^\mu$ be the current for a continuous internal generator $A$, and let

$$X=\Phi_1(x_1)\cdots\Phi_n(x_n)$$

be a product of renormalized fields or operators. With a fixed convention for the infinitesimal variation $\delta_A$, the renormalized local identity has the form

$$\begin{aligned} \partial_\mu \left\langle T\,j_A^\mu(x)X\right\rangle_{\rm R} ={}& -i\sum_{r=1}^n \delta^{(d)}(x-x_r) \left\langle T\,\Phi_1\cdots(\delta_A\Phi_r)\cdots\Phi_n \right\rangle_{\rm R} \\ &+ \left\langle T\,\mathcal A_A(x)X\right\rangle_{\rm R}. \end{aligned}$$

The delta functions are not violations of conservation. They state how the charge acts on every insertion. The final term $\mathcal A_A$ is an anomalous or explicit breaking term; it vanishes for the exact symmetry considered first. Products on the coincidence diagonals must use the renormalized contact terms developed on Contact Terms and Renormalized Operator Products. Omitting them can make an exact identity appear to fail.

Integrating over a thin time slab around one insertion gives the charge action. With boundary conditions that remove spatial flux,

$$Q_A(t)=\int d^{d-1}\mathbf x\,j_A^0(t,\mathbf x)$$

generates $\delta_A$. This fixes the normalization of the charge-carrying class of the current. Multiplying $j_A^\mu$ by an arbitrary finite constant would multiply the contact terms and would therefore describe a differently normalized generator.

The statement needed for RG protection is stronger than $\partial_\mu j_A^\mu=0$:

1. the renormalized Ward identity is exact, including its contact terms;
2. the charge is fixed to act as the declared generator;
3. the operator sector contains every current with the same dimension and quantum numbers, including improvements, equations of motion, and BRST-exact directions when applicable;
4. regulator breaking has been removed by allowed local counterterms; and
5. no anomaly remains in the identity.

Under these conditions the scale derivative of the normalized current obeys the homogeneous Ward identity: it has zero charge action. It can therefore be only a current with vanishing charge, such as an improvement or another redundant representative. Denoting the physical quotient by square brackets,

$$\boxed{ \mathcal D[\,j_A^\mu\,]_{\rm phys}=0. }$$

If a constant row vector $v_A^{\mathsf T}$ selects the current class from a closed operator column $O$,

$$[\,j_A^\mu\,]_{\rm phys} = v_A^{\mathsf T}[\,O^\mu\,]_{\rm phys},$$

then the chapter convention gives

$$\mathcal D[\,j_A^\mu\,]_{\rm phys} = -v_A^{\mathsf T}\gamma_{\rm phys} [\,O^\mu\,]_{\rm phys}.$$

Protection is consequently the left-null-vector condition

$$\boxed{ v_A^{\mathsf T}\gamma_{\rm phys}=0. }$$

In a basis whose first element is the protected current, the entire first row of the induced physical anomalous-dimension matrix vanishes. A zero $\gamma_{11}$ is insufficient if $\gamma_{1j}\ne0$ for some $j$. In an enlarged off-shell sector, that row may contain coefficients multiplying directions that vanish in the physical quotient; those entries describe representative dependence, not running of the charge.

Collins proves the renormalized Ward identity and the nonrenormalization result for exact linear internal symmetries, while also isolating the gauge-field and stress-tensor qualifications Collins 1984/2023, § 6.6, pp. 160–163.

Global O(N) current: a protected scalar-theory example

Consider $N$ real scalars with

$$\mathcal L = \frac12\partial_\mu\phi^a\partial^\mu\phi^a -\frac{m^2}{2}\phi^a\phi^a -\frac{\lambda}{4!} \left(\phi^a\phi^a\right)^2 .$$

The metric is the site convention $(+---)$. For an infinitesimal rotation in the $a$-$b$ plane, a current representative is

$$j_{ab}^\mu = \phi_a\partial^\mu\phi_b -\phi_b\partial^\mu\phi_a, \qquad j_{ab}^\mu=-j_{ba}^\mu .$$

At the bare level, the current is constructed from the complete bare action. Rewriting it in renormalized fields automatically brings along the field counterterm:

$$j_{0,ab}^\mu = Z_\phi \left( \phi_a\partial^\mu\phi_b -\phi_b\partial^\mu\phi_a \right).$$

It would be wrong to drop $Z_\phi$ and then infer an extra current renormalization from the remaining diagrams. The counterterm content of the complete Noether current is part of the operator.

The exact $O(N)$ Ward identity fixes the charge to act as

$$[Q_{ab},\phi_c] = i\left( \delta_{ac}\phi_b-\delta_{bc}\phi_a \right)$$

with the displayed sign defining the charge convention. In a nongauge renormalizable scalar theory there is no other dimension-three, antisymmetric-$O(N)$ vector whose divergence vanishes identically and whose charge is zero. Equation-of-motion terms can affect contact insertions but not the normalized physical current class. The Ward identity therefore fixes

$$\mathcal D[\,j_{ab}^\mu\,]_{\rm phys}=0, \qquad \gamma_{j,{\rm phys}}=0.$$

This conclusion is all-orders and scheme-independent as a statement about the normalized class, provided the finite scheme change preserves the Ward normalization. At an RG fixed point the current consequently has scaling dimension $d-1$, as required for a conserved spin-one operator. The elementary fields can have nonzero anomalous dimensions at the same fixed point; their contributions inside the composite current are cancelled by its operator counterterms.

The example also distinguishes a theorem from a diagrammatic accident. One-loop vertex and wave-function poles may cancel, but the Ward identity is what says how that cancellation extends to every order and which finite normalization is intended.

Why a conserved QED current can still mix

The argument changes when the closed sector contains another identically conserved vector. In QED,

$$R^\mu=\partial_\nu F^{\nu\mu}$$

has the right dimension and quantum numbers, and

$$\partial_\mu R^\mu = \partial_\mu\partial_\nu F^{\nu\mu} =0$$

identically by antisymmetry of $F^{\nu\mu}$. Adding $\kappa R^\mu$ to a current leaves its local divergence and ordinary Ward contact terms unchanged. The Ward identity alone therefore cannot fix $\kappa$.

In dimensional regularization with minimal subtraction, the electron-number current has precisely this mixing. In the conventions of Collins, Manohar, and Wise,

$$j_{{\rm MS}}^\mu = Z_2\bar\psi\gamma^\mu\psi + \frac{Z_3-1}{e\mu^\epsilon} \partial_\nu F^{\nu\mu}.$$

The minimally subtracted local current has a nonzero anomalous dimension. A definite finite shift is needed to obtain the electron-number operator with the intended normalization; for a massless photon, the integrated improvement is related to electric flux at spatial infinity and need not vanish. Thus “conserved” does not imply “unrenormalized local representative” in a gauge theory Collins, Manohar, and Wise 2006, pp. 1–4.

This example does not contradict charge protection. It shows that three objects must not be conflated:

Object	What fixes it	What may still vary
Ward identity	Quantum symmetry and its contact terms	Addition of an identically conserved current
Integrated charge	Generator normalization plus boundary conditions	Surface terms when long-range gauge fields are present
Local current representative	A renormalization prescription and finite conditions	Scheme-dependent mixing inside the zero-charge sector

The full mixing convention record is therefore required even for a protected charge. The relevant matrix must contain $R^\mu$, and the boundary conditions used to turn a local identity into a charge statement must be declared.

Translation protection and stress-tensor improvement

The translation Ward identity fixes the normalization of the stress tensor through

$$\begin{aligned} \partial_\mu \left\langle T\,T^{\mu\nu}(x)X\right\rangle_{\rm R} = -i\sum_r\delta^{(d)}(x-x_r) \frac{\partial}{\partial x_{r\nu}} \left\langle T\,X\right\rangle_{\rm R}, \end{aligned}$$

for scalar insertions, with the corresponding local transformation terms for tensor or spinor insertions. Its charges

$$P^\nu = \int d^{d-1}\mathbf x\,T^{0\nu}$$

must generate translations. This protects the translation-current class, but it does not select a unique local tensor.

For the scalar theory, a symmetric flat-space family is

$$\begin{aligned} T_{\mu\nu}^{(\xi)} ={}& \partial_\mu\phi^a\partial_\nu\phi^a -\eta_{\mu\nu} \left[ \frac12\partial_\rho\phi^a\partial^\rho\phi^a -V(\phi) \right] \\ &+ \xi \left( \eta_{\mu\nu}\Box-\partial_\mu\partial_\nu \right) [\,\phi^a\phi^a\,]. \end{aligned}$$

The last line is the improvement

$$\Delta T_{\mu\nu} = \xi \left( \eta_{\mu\nu}\Box-\partial_\mu\partial_\nu \right) [\,\phi^2\,].$$

It is symmetric and identically conserved:

$$\partial^\mu\Delta T_{\mu\nu}=0.$$

Under boundary conditions for which the spatial surface terms vanish, it leaves $P^\nu$ unchanged. Nevertheless it changes local stress-tensor matrix elements, contact terms, the trace, and the response to a background metric. Charge normalization and local representative are again different layers.

For the free massless theory, the trace is

$$\begin{aligned} T^{(\xi)\mu}{}_\mu = \left(1-\frac d2\right) (\partial_\rho\phi^a)^2 + \xi(d-1)\Box(\phi^a\phi^a). \end{aligned}$$

Using $\Box\phi^a=0$ gives

$$\Box(\phi^a\phi^a) = 2(\partial_\rho\phi^a)^2.$$

The choice

$$\boxed{ \xi_c=\frac{d-2}{4(d-1)} }$$

makes the on-shell trace vanish; in four dimensions $\xi_c=1/6$. Callan, Coleman, and Jackiw show how this improved tensor preserves the Poincaré generators while giving finite renormalized matrix elements Callan, Coleman, and Jackiw 1970, pp. 42–73.

In an interacting theory, $[\,\phi^2\,]$ itself is a renormalized composite operator, and the improvement coefficient is part of the stress-tensor renormalization conditions. The exact translation Ward identity still fixes the physical $T_{\mu\nu}$ class. It does not force a particular $\xi$.

Conservation is not tracelessness

Translation symmetry and scale symmetry have different Ward identities. In a massless theory with renormalized dimensionless couplings $g_i$, a useful operator form of the trace identity is

$$[\,T^\mu{}_\mu\,] = \sum_i\beta_i(g) \left[ \frac{\partial\mathcal L}{\partial g_i} \right] + \text{EOM terms} + \partial_\mu V^\mu .$$

Mass terms and other relevant parameters add their own dimension and anomalous-dimension contributions. The equation-of-motion terms matter in contact Green functions even though they vanish in appropriate on-shell matrix elements. The total derivative contains improvement or virial-current freedom.

For the scalar interaction used above,

$$\frac{\partial\mathcal L}{\partial\lambda} = -\frac1{4!} \left(\phi^a\phi^a\right)^2,$$

so the beta-function part of the trace is

$$-\frac{\beta_\lambda}{4!} \left[ \left(\phi^a\phi^a\right)^2 \right].$$

The sign is tied to the displayed Lagrangian convention. This term can remain nonzero even though

$$\partial_\mu T^{\mu\nu}=0.$$

Accordingly:

• zero anomalous dimension of the normalized stress-tensor class follows from translations;
• vanishing trace requires a separate scale or conformal statement;
• a nonzero beta function produces a trace anomaly in the massless theory;
• at a fixed point, $\beta_i=0$ removes the beta-function terms, but a nonremovable virial current would require separate analysis.

Improving the stress tensor cannot remove a genuine beta-function anomaly. It can only rearrange total derivatives and terms that belong to the allowed improvement sector.

Regulator restoration and anomaly checks

A regulator need not preserve the symmetry used to define the current. Protection is asserted only after the following finite-renormalization problem has been solved.

Question	Required evidence
What identity is imposed?	State the local renormalized Ward identity, including contact terms and any breaking insertion.
What fixes normalization?	Specify the action of the integrated charge on a complete set of fields or physical states.
What is the closed sector?	Include same-quantum-number currents, improvements, EOM operators, BRST-exact operators, and lower-dimensional mixing allowed by the regulator.
What did the regulator break?	List the local breaking terms and the counterterms used to cancel them.
Is restoration possible?	Show that the breaking is the variation of allowed local counterterms and that all required identities can be satisfied together.
Is there an anomaly?	Demonstrate whether a residual cohomologically nontrivial breaking remains.
What is protected?	Identify the RG-invariant class or left null vector, not merely one bare-looking formula.
What freedom remains?	Record finite improvement coefficients, boundary conditions, and scheme choices.

Dimensional regularization preserves the linear $O(N)$ symmetry of the scalar example, so its current identity can be maintained by minimal subtraction. A cutoff that breaks a chiral symmetry, or a lattice regulator that breaks continuous translations, generally requires a larger mixing problem and finite restoration conditions. If a residual breaking cannot be removed without violating another required identity, it is an anomaly rather than an unfortunate scheme choice.

This distinction also controls interpretation of numerical results. A small fitted anomalous dimension is evidence for protection only after the Ward residual, continuum limit, sector closure, and normalization condition have been checked. Symmetry labels alone do not supply those checks.

Common pitfalls

Inferring protection from conservation at separated points. A separated-point divergence can vanish while contact terms carry the charge action or while an identically conserved operator mixes with the current. Test the full distributional Ward identity.

Checking only one matrix element. The condition $\gamma_{jj}=0$ does not prevent the current from evolving into other physical operators. Identify the complete left null vector of the physical mixing matrix.

Calling every improvement physically irrelevant. An improvement may leave an integrated charge unchanged under chosen boundary conditions while changing local correlators, trace identities, boundary observables, and background-field couplings.

Treating a trace anomaly as failure of translation conservation. The stress tensor can remain exactly conserved and correctly normalized while its trace contains beta functions and relevant couplings.

Exercises

1. The O(N) charge fixes a left null vector

Let $O^\mu$ be a closed column of antisymmetric-$O(N)$ vector operators and

$$j_{ab}^\mu=v_{ab}^{\mathsf T}O^\mu .$$

Assume $v_{ab}$ is scale independent and the exact Ward identity fixes the normalized current class. Derive the condition on $\gamma_{\rm phys}$.

Solution

The operator equation is

$$\mathcal D[O^\mu]_{\rm phys} = -\gamma_{\rm phys}[O^\mu]_{\rm phys}.$$

Therefore

$$\mathcal D[j_{ab}^\mu]_{\rm phys} = -v_{ab}^{\mathsf T} \gamma_{\rm phys} [O^\mu]_{\rm phys}.$$

The Ward identity fixes the current class and hence makes the left-hand side zero. Linear independence in the physical quotient gives

$$v_{ab}^{\mathsf T}\gamma_{\rm phys}=0.$$

Thus $v_{ab}^{\mathsf T}$ is a left zero-eigenvector. A vanishing diagonal entry in a basis that does not align with $v_{ab}$ is not the required statement.

2. Verify the scalar improvement

For

$$\Delta T_{\mu\nu} = \xi(\eta_{\mu\nu}\Box-\partial_\mu\partial_\nu)\phi^2,$$

show that it is conserved identically, that its contribution to $P^\nu$ is a spatial surface term, and that $\xi_c=(d-2)/[4(d-1)]$ makes the free massless trace vanish on shell.

Solution

Commutativity of derivatives gives

$$\partial^\mu\Delta T_{\mu\nu} = \xi(\partial_\nu\Box-\Box\partial_\nu)\phi^2 =0.$$

For $\nu=0$, the time-derivative pieces in $\eta^{00}\Box-\partial^0\partial^0$ cancel, leaving a spatial Laplacian. For spatial $\nu$, the integral is likewise a spatial divergence. Hence

$$\int d^{d-1}\mathbf x\,\Delta T^{0\nu}$$

is a boundary term and vanishes for the declared falloff.

The trace of the improvement is

$$\Delta T^\mu{}_\mu = \xi(d-1)\Box\phi^2.$$

For $\Box\phi=0$,

$$\Box\phi^2=2(\partial\phi)^2.$$

Adding this to the canonical trace gives the coefficient

$$1-\frac d2+2\xi(d-1).$$

Setting it to zero yields

$$\xi_c=\frac{d-2}{4(d-1)}.$$

3. Diagnose the QED loophole

Suppose a Ward identity fixes $\partial_\mu j^\mu$, and the operator sector also contains $R^\mu=\partial_\nu F^{\nu\mu}$. Explain why the identity does not fix the coefficient of $R^\mu$ in $j^\mu$, and name the additional information needed to define the charge.

Solution

Antisymmetry of $F^{\nu\mu}$ gives

$$\partial_\mu R^\mu = \partial_\mu\partial_\nu F^{\nu\mu} =0$$

without using equations of motion. Therefore $j^\mu+\kappa R^\mu$ obeys the same local divergence identity for every $\kappa$. One must add a finite renormalization condition that fixes the intended local representative and specify spatial boundary conditions. With a massless photon, the integral of $R^0$ is related by Gauss’s law to electric flux at infinity, so it cannot be discarded automatically when defining the charge.

Continue

• Continue to Nonperturbative Renormalization Schemes and Step Scaling to impose operator-normalization conditions without weak coupling at the starting scale.
• Return to Dual Evolution of Operators and Wilson Coefficients to evolve an unprotected closed sector while preserving the coefficient–operator pairing.
• Continue to Renormalization-Group Equations and Running for beta functions and trace identities along complete RG trajectories.

References

• Callan, Curtis G., Jr., Sidney Coleman, and Roman Jackiw. 1970. “A New Improved Energy-Momentum Tensor.” Annals of Physics 59 (1): 42–73. DOI.
• Collins, John C. 1984; open-access reissue 2023. Renormalization: An Introduction to Renormalization, the Renormalization Group and the Operator-Product Expansion. Cambridge Monographs on Mathematical Physics. Cambridge University Press. DOI and Open PDF.
• Collins, John C., Aneesh V. Manohar, and Mark B. Wise. 2006. “Renormalization of the Vector Current in QED.” Physical Review D 73: 105019. DOI. Open PDF.