Scale versus Conformal Invariance: Hypotheses and Counterexamples

Scale invariance does not imply conformal invariance by algebra alone: dilatations form a proper subgroup of the conformal group. Enhancement is a dynamical statement about the trace of a local stress tensor. If a scale-invariant theory has $T^\mu{}_{\mu}=\partial_\mu V^\mu$, conformal invariance follows when the virial current can be removed by a local improvement. Whether this must happen depends on dimension, unitarity or reflection positivity, locality, the spectrum of scaling dimensions, the existence and transformation of the stress tensor and scale current, boundary conditions, and possible anomalies.

Required background. Conserved Currents and the Stress Tensor supplies Ward identities and stress-tensor improvement. Local RG and Trace Identities supplies renormalized trace relations and flavor-rotation ambiguities. Helpful background. Fixed Points and Linearized RG Flow supplies fixed-point and limit-cycle language. What Is an Anomaly? separates genuine anomalies from removable local terms.

The virial-current criterion

Work in Lorentzian signature with the mostly-minus metric and assume a local symmetric conserved stress tensor,

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

A local dilatation current can be written

$$D^\mu=x_\nu T^{\mu\nu}-V^\mu.$$

Its divergence is

$$\partial_\mu D^\mu =T^\mu{}_{\mu}-\partial_\mu V^\mu.$$

Thus unbroken scale invariance with a local scale current requires, modulo equations of motion and contact terms,

$$T^\mu{}_{\mu}=\partial_\mu V^\mu.$$

The virial current is not unique: it may be shifted by a conserved current or by a superpotential without changing its divergence. The question is whether its nonconserved part can be written as the divergence of a local tensor in precisely the form needed for a stress-tensor improvement. A common sufficient condition is

$$V_\mu=\partial_\mu L, \qquad T^\mu{}_{\mu}=\partial^2L.$$

Then

$$T'_{\mu\nu} =T_{\mu\nu} +\frac1{d-1} \left(\partial_\mu\partial_\nu -\eta_{\mu\nu}\partial^2\right)L$$

is symmetric, conserved, and traceless. More generally, $T^\mu{}_{\mu}=\partial_\mu\partial_\nu L^{\mu\nu}$ can admit a tensor improvement; its exact form depends on the irreducible decomposition of $L^{\mu\nu}$. The existence, locality, gauge invariance, and scaling properties of the required $L$ are hypotheses, not consequences of manipulating the trace formally. The virial criterion and its qualifications are reviewed in Nakayama 2015, §§ 2.3–2.4.

Once a traceless stress tensor exists, the special-conformal currents

$$K_{(\alpha)}^\mu =\left(2x_\alpha x^\nu-x^2\delta_\alpha{}^\nu\right) T^\mu{}_{\nu}$$

are conserved. Direct differentiation using symmetry and conservation gives

$$\partial_\mu K_{(\alpha)}^\mu=2x_\alpha T^\mu{}_{\mu}.$$

This calculation proves the enhancement after improvement. It does not prove that the improvement operator exists.

Benchmark: the free massless scalar

For $d>2$, consider

$$S=\frac12\int d^dx\,\partial_\mu\phi\,\partial^\mu\phi, \qquad \partial^2\phi=0.$$

The canonical symmetric tensor is

$$T^{\mathrm{can}}_{\mu\nu} =\partial_\mu\phi\partial_\nu\phi -\frac12\eta_{\mu\nu}(\partial\phi)^2.$$

On the equation of motion,

$$T^{\mathrm{can}\,\mu}{}_{\mu} =-\frac{d-2}{2}(\partial\phi)^2 =-\frac{d-2}{4}\partial^2\phi^2.$$

The virial current is a gradient, $V_\mu=-(d-2)\partial_\mu\phi^2/4$. Add

$$\Delta T_{\mu\nu} =\xi\left(\eta_{\mu\nu}\partial^2 -\partial_\mu\partial_\nu\right)\phi^2, \qquad \xi=\frac{d-2}{4(d-1)}.$$

Its trace is $\xi(d-1)\partial^2\phi^2$, which cancels the canonical trace. The result is the conformal stress tensor. This example checks the sign and normalization of the improvement and illustrates why a trace that is nonzero before improvement does not disprove conformal invariance. The construction is the basic improved tensor of Callan, Coleman, and Jackiw 1970.

There is also a boundary to the example. If $\phi$ is compact or only derivative operators are admitted, $\phi^2$ may fail to be a globally well-defined local operator. An improvement allowed in the noncompact local algebra may then be forbidden in the specified theory. Local operator content and global identifications must be checked, not inferred from the equation of motion.

What representative results actually establish

The following table states the claim, hypotheses, and failure diagnosis together. It is intentionally narrower than the slogan “scale implies conformal.”

Setting	Enhancement statement	Hypotheses that do substantive work	Concrete test	What invalidates the inference	Source and claim strength
Relativistic QFT in $d=2$	Scale invariance implies a traceless stress tensor and hence conformal invariance	Unitarity or reflection positivity, Poincaré invariance and causality, discrete scaling spectrum, a local scale current, unbroken scale invariance, and the stress-tensor regularity used in the $c$-function argument	Construct stress-tensor two-point functions; scale invariance makes the $c$-function stationary, positivity forces the trace two-point function to vanish, and locality promotes this to an operator statement	Riva–Cardy elasticity violates positivity; noncompact or incomplete sigma-model targets can violate spectral or boundary assumptions	Theorem under the stated assumptions: Polchinski 1988; hypothesis summary in Nakayama 2015, § 5.1
Perturbative renormalizable fixed point	A genuine fixed point is conformal when the scheme-covariant trace coefficients vanish and every virial candidate is conserved, redundant, or improvable	Local renormalization, unitarity in the physical sector, a controlled perturbative expansion, complete operator mixing, and quotienting flavor rotations and equations of motion	Compute the local trace identity, enumerate dimension-$(d-1)$ singlet vectors, and test whether $V_\mu$ is a gradient, conserved current, or flavor rotation	An omitted vector operator, a non-diagonalizable mixing problem, gauge-noninvariant improvement, or mistaking a flavor orbit for physical running	Perturbative result within its assumptions: Nakayama 2015, §§ 7 and 8.5
Unitary relativistic theory in $d=4$	Dilaton and anomaly arguments can force $T^\mu{}_{\mu}=\partial^2L$ and hence conformality	A local symmetric conserved stress tensor, a local scale current, unitarity, appropriate current algebra, an admissible infrared regulator or flow, regular forward amplitudes, and assumptions connecting trivial dilaton scattering to a free field redefinition	Couple the trace to a background dilaton, constrain amplitudes with anomalies and unitarity, then reconstruct a local $L$	Infrared singularities, compact-field obstructions, generalized free sectors, or failure of the S-matrix-to-local-operator step	Conditional physical argument, explicitly not presented as a general mathematical theorem: Dymarsky et al. 2015, §§ 1 and 3–4
Two-dimensional sigma model with metric and $B$ field	Scale invariance implies conformal invariance to all orders in sigma-model perturbation theory for a compact target	Sigma-model perturbation theory and compactness, or a sufficient completeness condition, of the target	Express beta functions as a generalized Ricci soliton and use the compact-target integral argument	Noncompact or incomplete targets admit the geometric obstructions displayed by the authors	Specialized perturbative theorem: Papadopoulos and Witten 2024, §§ 2.4–2.5
General dimension or nonstandard theory	No enhancement is assumed on this page without a named result or an explicit local improvement	Must be established for the particular dimension, locality class, spectrum, stress tensor, positivity notion, and boundary conditions	Exhibit a traceless improved $T$, conserved special-conformal charges, and their Ward identities	Nonlocality, absent stress tensor, nonunitarity, continuous spectrum, boundaries, defects, anomalies, or spontaneous breaking can defeat intermediate steps	Case-by-case conclusion; the diagnostic framework is reviewed in Nakayama 2015

The table does not turn evidence in one row into a theorem in another. In particular, a perturbative fixed-point calculation is not a nonperturbative proof for a distant strongly coupled theory, and a vanishing separated-point trace correlator implies a vanishing operator only with the required positivity and locality theorem.

Diagnosing a perturbative fixed point

For renormalized couplings $g^I$ multiplying local operators $[\mathcal O_I]$, a flat-space trace identity has the schematic form

$$T^\mu{}_{\mu} =B^I[\mathcal O_I] +\partial_\mu J^\mu +\text{equations of motion} +\text{contact terms}.$$

Here $B^I$ denotes a scheme- and flavor-covariant beta function. If ordinary beta functions contain a flavor rotation, $\beta^I=(Qg)^I$, combining the dilatation generator with that internal symmetry can give $B^I=0$; a closed trajectory in coupling coordinates is then not automatically a physical scale-without-conformal cycle.

Use the following finite workflow at a proposed perturbative fixed point:

1. Choose a renormalized operator basis including equations of motion, total derivatives, and all internal-singlet vectors of dimension $d-1$ to the retained order.
2. Compute the local trace identity and the operator-mixing matrix, not only constant-coupling beta functions.
3. Quotient basis changes and flavor rotations to obtain the covariant coefficients $B^I$.
4. If $B^I=0$, test whether the remaining $J^\mu$ is conserved or removable by improvement.
5. Verify the special-conformal Ward identity on separated correlators and track contact or anomaly terms independently.

For the one-component $\phi^4$ Wilson–Fisher fixed point near four dimensions, the obvious $\mathbb Z_2$-even singlet vector $\phi\partial_\mu\phi$ is $(1/2)\partial_\mu\phi^2$, hence a gradient rather than an obstruction. After the mass is tuned and the fixed-point beta coefficient vanishes, no independent perturbative virial candidate remains in that operator sector. This establishes conformality order by order within the epsilon expansion; it does not by itself prove a nonperturbative statement at a finite value of $\epsilon$ Nakayama 2015, § 4.3.

A counterexample that identifies a missing hypothesis

Two-dimensional elasticity provides a local Euclidean scale-invariant theory with vector displacement fields $u_\mu$ and action

$$S=\frac12\int d^2x\, \left[2g\,u_{\mu\nu}u^{\mu\nu} +k\,(u^\sigma{}_{\sigma})^2\right], \qquad u_{\mu\nu}=\frac12(\partial_\mu u_\nu+\partial_\nu u_\mu).$$

For generic elastic moduli, its symmetric stress tensor has a trace that is a virial divergence, but the virial current is not a gradient and cannot be improved away. The theory is scale invariant but not conformally invariant. This does not contradict the two-dimensional theorem: after Lorentzian continuation the field transforms as a spacetime vector, and the Hamiltonian has negative directions; equivalently, the Euclidean theory fails reflection positivity. Riva and Cardy compute both the non-gradient virial and the positivity failure in Riva and Cardy 2005, pp. 339–342.

The example teaches more than the existence of an exception. It shows exactly why the implication

$$\langle T^\mu{}_{\mu}(x)T^\nu{}_{\nu}(0)\rangle=0 \quad\Longrightarrow\quad T^\mu{}_{\mu}=0$$

requires positivity. In an indefinite theory, an operator can have zero self-two-point function yet nonzero mixed correlators and need not vanish.

Other ways the hypotheses can fail

• No local stress tensor. Generalized free fields can be conformally covariant at the level of correlators without containing a local stress tensor. A stress-tensor enhancement theorem cannot be applied to them as though they were ordinary local QFTs.
• No local scale current. A scale charge may exist while every gauge-invariant local representative of the current is obstructed. The virial equation must then be reformulated; it cannot simply be assumed.
• Continuous or non-diagonalizable spectrum. Noncompact fields, logarithmic theories, and Jordan blocks can invalidate the discrete spectral decomposition used in positivity arguments.
• Forbidden improvement. Compactness, gauge invariance, or operator identifications may exclude the apparent local $L$.
• Boundaries and defects. Bulk improvement can leave boundary or defect virial terms, so the preserved conformal subgroup needs a separate analysis.
• Anomalies and contact terms. Flat separated-point tracelessness does not imply Weyl invariance on curved space. Allowed local counterterms and anomaly cohomology must be tracked.
• Spontaneous breaking. A scale-invariant action may have a vacuum that is not scale invariant. The theorem hypotheses normally concern an invariant vacuum and Ward identities without an uncompensated dilaton state.

For a proposed model, the strongest conclusion supported here is either an explicit traceless improvement with checked Ward identities, a named theorem whose hypotheses have all been verified, or a diagnosis of the missing hypothesis. Theorem-first reconstruction and exceptional nonperturbative domains hand off to Mathematical QFT; beta functions and local trace coefficients hand back to Renormalization and Effective Field Theory.

Common pitfalls

A zero beta function is the whole proof. It removes the explicit $B^I[\mathcal O_I]$ term only after scheme and flavor rotations are handled. Virial, improvement, anomaly, and contact terms remain to be tested.

A trace that is a divergence is already zero. That is the condition for a conserved scale current. Conformal symmetry requires the divergence to have the stronger improvable form.

The two-dimensional theorem has no hypotheses. Positivity, locality, a discrete scaling spectrum, a local scale current, and an invariant vacuum are essential. Elasticity evades the theorem by failing positivity.

A classical improvement survives quantization automatically. The improvement operator can mix, cease to be well defined, or be forbidden by gauge or global identifications. Verify the renormalized trace identity.

Exercises

Derive the improvement coefficient for the free scalar.

Solution

On shell, $T^{\mathrm{can}\,\mu}{}_{\mu}=-(d-2)\partial^2\phi^2/4$. The trace of $\xi(\eta_{\mu\nu}\partial^2-\partial_\mu\partial_\nu)\phi^2$ is $\xi(d-1)\partial^2\phi^2$. Setting their sum to zero gives

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

Suppose $T^\mu{}_{\mu}=\partial_\mu V^\mu$ and $V_\mu=\partial_\mu L+J_\mu$ with $\partial_\mu J^\mu=0$. Show that the conserved part does not obstruct improvement.

Solution

The trace is $\partial^2L$ because the divergence of $J_\mu$ vanishes. Adding $(\partial_\mu\partial_\nu-\eta_{\mu\nu}\partial^2)L/(d-1)$ makes the trace zero. The conserved current changes the representative of the dilatation current but not its divergence.

References

• Callan, Curtis G., Sidney Coleman, and Roman Jackiw. “A New Improved Energy–Momentum Tensor.” Annals of Physics 59 (1970): 42–73. DOI
• Dymarsky, Anatoly, Zohar Komargodski, Adam Schwimmer, and Stefan Theisen. “On Scale and Conformal Invariance in Four Dimensions.” Journal of High Energy Physics 2015, no. 10 (2015): 171. DOI; Open PDF
• Nakayama, Yu. “Scale Invariance versus Conformal Invariance.” Physics Reports 569 (2015): 1–93. DOI; Open PDF
• Papadopoulos, Georgios, and Edward Witten. “Scale and Conformal Invariance in 2d Sigma Models, with an Application to $\mathcal N=4$ Supersymmetry.” arXiv:2404.19526 (2024). Open PDF
• Polchinski, Joseph. “Scale and Conformal Invariance in Quantum Field Theory.” Nuclear Physics B 303 (1988): 226–236. DOI
• Riva, Valentina, and John Cardy. “Scale and Conformal Invariance in Field Theory: A Physical Counterexample.” Physics Letters B 622 (2005): 339–342. DOI; Open PDF