Soft Photons and Infrared-Finite QED

An exclusive amplitude between charged Fock states is not by itself an infrared-finite prediction: an arbitrarily soft photon can be emitted at arbitrarily small energy cost, and virtual photons probe the same region. QED nevertheless predicts finite quantities when the observable treats experimentally indistinguishable states consistently. The calculation below keeps charged-particle masses nonzero and derives the leading soft correction to an inclusive cross section with a stated resolution. Coherent dressed states give a different, amplitude-level construction; taking charged masses to zero introduces a separate collinear question.

Required background. Tree-level QED processes supplies the charged amplitudes and external-state conventions. Bloch–Nordsieck and KLN cancellation supplies the general degeneracy criterion for infrared-safe probabilities.

Helpful background. Dressed states and infrared-finite scattering develops coherent asymptotic states beyond the inclusive construction used first below.

Soft factorization in QED

Let a hard Born amplitude $\mathcal M_n$ have external massive charged particles with future-directed on-shell momenta $p_i$, signed charges $q_i=eQ_i$ with $e>0$ and $Q_e=-1$, and signs

$$\eta_i= \begin{cases} +1,&\text{outgoing},\\ -1,&\text{incoming}. \end{cases}$$

Keep the hard kinematics fixed away from thresholds and additional nearly on-shell internal propagators. For one additional outgoing photon of momentum $k$, polarization $\varepsilon_\mu(k)$, and energy much smaller than every resolved hard scale, attachment to the external charged legs gives

$$\mathcal M_{n+1}(k,\varepsilon) =-e\,\varepsilon_\mu^*(k)J_{\mathrm s}^{\mu}(k)\,\mathcal M_n +\mathcal O(k^0), \qquad J_{\mathrm s}^{\mu}(k) =\sum_i\eta_iQ_i\frac{p_i^\mu}{p_i\cdot k}.$$

The overall minus sign follows from the site’s $+iq_i\gamma^\mu$ vertex and the factor $i$ in the adjacent fermion propagator. For an outgoing fermion, the leading attachment is

$$\bar u(p)(iq_i\gamma^\mu) \frac{i(p\!\!\!/+k\!\!\!/+m)}{(p+k)^2-m^2} =-q_i\frac{p^\mu}{p\cdot k}\bar u(p)+O(k^0).$$

For an incoming fermion the denominator is $-2p\cdot k$, giving $+q_i p^\mu/(p\cdot k)$. Crossing includes antiparticles with their opposite signed charge. Thus the leading factor is $-\sum_i\eta_iq_i p_i\cdot\varepsilon^*/(p_i\cdot k)$. Reversing the potential to $A^{\rm em}=-A$ also reverses its photon polarization and gives the frequently used positive-sign soft factor in that convention. The probability is unchanged.

The $1/(p_i\cdot k)$ poles are universal: the leading soft factor depends only on the external charges and momenta. The hard amplitude already contains spin and short-distance dynamics; additional dependence in the soft factor enters beyond leading order. Under the gauge replacement $\varepsilon^\mu\to\varepsilon^\mu+c\,k^\mu$,

$$k_\mu J_{\mathrm s}^{\mu}=\sum_i\eta_iQ_i=0,$$

where the last equality is charge conservation. This is a fast sign check on incoming particles, antiparticles, and crossed amplitudes. The external-line derivation and charge-conservation test appear in Weinberg 1995, § 13.1, pp. 535–537, Eqs. (13.1.1)–(13.1.4). The explicit vertex calculation above fixes the translation from his metric and amplitude conventions to the site’s sign.

After summing physical photon polarizations, the unresolved one-photon probability contains the positive eikonal integral

$$e^2\int_{\lambda<\omega<\Delta E} \frac{d^3\mathbf k}{(2\pi)^3\,2\omega} \sum_{r=1,2} \left|\varepsilon^{(r)*}(k)\cdot J_{\mathrm s}(k)\right|^2.$$

Here $k^2=0$, $\omega=|\mathbf k|$, and $\lambda$ is an auxiliary lower cutoff on the soft photon energy in a specified frame, not a photon mass. The upper bound $\Delta E$ is a physical resolution in that frame, with $\lambda\ll\Delta E$ well below the hard scales. A photon-mass prescription would instead change the dispersion relation and phase space; dimensional regularization is another option. Whichever prescription is used, its real and virtual terms must match before the regulator is removed.

To extract the coefficient, write $k^\mu=\omega n^\mu$ with $n^\mu=(1,\widehat{\mathbf k})$, and put

$$j^\mu(\widehat{\mathbf k}) =\sum_i\eta_iQ_i\frac{p_i^\mu}{p_i\cdot n}, \qquad J_{\mathrm s}^\mu=\frac{j^\mu}{\omega}.$$

Charge conservation gives $n\cdot j=0$, so the two physical polarization sum is $-j^2=|\mathbf j_\perp|^2\geq0$. The massless phase-space measure then separates into an energy logarithm and an angular coefficient:

$$\begin{aligned} B&=\frac{1}{4\pi}\int\!\mathrm d\Omega\,[-j^2],\\ \delta_{\mathrm{real}}^{\mathrm{soft}} &=\frac{\alpha}{\pi}B \int_\lambda^{\Delta E}\frac{\mathrm d\omega}{\omega} =\frac{\alpha}{\pi}B\ln\frac{\Delta E}{\lambda}. \end{aligned}$$

Here $\alpha=e^2/(4\pi)$ and $\delta_{\mathrm{real}}^{\mathrm{soft}}$ is the correction relative to a nonzero Born rate at the fixed hard kinematics. The angular integral includes every photon direction unresolved as its energy tends to zero. Positive charged masses keep the angular denominators nonzero. For a wider hard bin, integrate the Born rate times this kinematics-dependent coefficient rather than assigning one constant $B$ to every event.

The soft coefficient for a Compton event

Consider the hard Compton process, with an additional unresolved photon. In the incoming electron’s rest frame, rotate the recoil electron onto the positive $z$ axis and write

$$p=m(1,0,0,0),\qquad p'=m\gamma(1,0,0,\beta),\qquad \gamma=(1-\beta^2)^{-1/2},\quad 0\leq\beta<1.$$

The two hard photons carry no electric charge, so only the incoming and outgoing electron enter $j$. For $x=\cos\theta$ between the soft photon and recoil electron, their current gives

$$-j^2=\frac{\beta^2(1-x^2)}{(1-\beta x)^2}, \qquad B(\beta)=\frac{\beta^2}{2} \int_{-1}^1\!\mathrm dx\, \frac{1-x^2}{(1-\beta x)^2}.$$

For $\beta>0$, substitute $y=1-\beta x$; the numerator divided by $y^2$ becomes $(\beta^2-1)/y^2+2/y-1$. Its elementary integral yields

$$\begin{aligned} B(\beta) &=\frac{1}{\beta}\ln\frac{1+\beta}{1-\beta}-2\\ &=2\left(\frac{\operatorname{artanh}\beta}{\beta}-1\right) =\frac{2\beta^2}{3}+O(\beta^4). \end{aligned}$$

The continuous limit is $B(0)=0$: with no change in the electron velocity, the two leading soft currents cancel. For nonzero recoil, $B>0$. As $\beta\to1$, the growing logarithm also warns that collinear enhancement becomes important. This is the single-charge velocity-change coefficient in Weinberg 1995, § 13.2, pp. 542–543, where the rate exponent is $A=(\alpha/\pi)B$. It determines the leading soft logarithm, not the complete finite one-loop Compton correction.

Build the observable before cancelling the regulator

A necessary soft-safety test is that the measurement function becomes insensitive to an unresolved photon,

$$F_\Delta(X+\gamma(k))\longrightarrow F_\Delta(X) \qquad (k\to0).$$

In a detector this statement requires more than an energy cut. One must say which frame defines the photon-energy threshold, which angular region is accepted, whether a photon close to a charged track is recombined with it, and which charged-particle masses are retained. These choices are part of the observable, not properties of the regulator. The limiting equality is not by itself a convergence proof: after real–virtual subtraction, the difference of measurement weights must be integrable against the soft measure $\mathrm d\omega/\omega$. A positive-power approach to the limit is sufficient away from bin boundaries; bins are handled after phase-space integration. The canonical measurement-function test states the wider soft and collinear requirements.

For the same fixed hard kinematics and a global unresolved-energy cut, use the coefficient $B\geq0$ derived above. To first relative order in $\alpha$, the virtual correction and unresolved-real contribution have the schematic but regulator-explicit form

$$\begin{aligned} \sigma_{\mathrm{virt}}(\lambda) &=\sigma_0\frac{\alpha}{\pi}\left[-B \ln\frac{Q}{\lambda}+c_{\mathrm{virt}}\right],\\ \sigma_{\mathrm{real}}(\lambda,\Delta E) &=\sigma_0\frac{\alpha}{\pi}\left[+B \ln\frac{\Delta E}{\lambda}+c_{\mathrm{real}}(\Delta)\right]. \end{aligned}$$

$Q$ denotes a positive hard reference scale, while $c_{\mathrm{virt}}$ and $c_{\mathrm{real}}$ are finite as $\lambda\to0$ at fixed masses and resolution. They need not be numerically order one: mass ratios or angular cuts can produce large finite logarithms. The latter coefficient retains dependence on the complete prescription $\Delta$. Hard loop regions and terms beyond the leading soft approximation belong in these coefficients; the logarithm does not extend the eikonal approximation to arbitrary photon energies. Adding the Born term gives

$$\sigma_{\mathrm{incl}}(\Delta) =\sigma_0\left[ 1+\frac{\alpha}{\pi}\left( -B\ln\frac{Q}{\Delta E} +c_{\mathrm{virt}}+c_{\mathrm{real}}(\Delta) \right) \right]+\mathcal O(\alpha^2),$$

so that

$$\frac{\partial\sigma_{\mathrm{incl}}}{\partial\ln\lambda}=0 \qquad\text{at this order.}$$

The virtual coefficient must include the complete amplitude and its external-state normalization, as discussed on the Ward-identity page. Omitting the field-residue contribution loses part of the soft coefficient. The equal real and virtual coefficients, including these external-line terms, are derived in Weinberg 1995, §§ 13.2–13.3, pp. 542–547.

The cancellation occurs with the same measurement weight in each unresolved limit. A photon outside the detector’s angular acceptance is normally unobserved and must be summed over, not vetoed down to zero energy. A finite-energy angular veto may be soft safe, but it changes the finite logarithms. Energy and angular resolution are worked out explicitly in Schwartz 2014, § 20.2, pp. 364–365, Eqs. (20.52)–(20.57).

The finite logarithm $\ln(Q/\Delta E)$ is physical resolution dependence. If $(\alpha/\pi)B\ln(Q/\Delta E)$ is large, fixed-order perturbation theory is poorly organized even though the regulator has cancelled; multiple-emission exponentiation or another resummation is then appropriate. At leading soft accuracy an independent per-photon energy cut produces the factor $(\Delta E/Q)^{\alpha B/\pi}$ multiplying the hard rate. A cut on the total emitted energy requires the corresponding multi-photon convolution and changes finite terms Weinberg 1995, § 13.3, pp. 545–547, Eqs. (13.3.6)–(13.3.11). The original massive-electron soft approximation already exhibits the vanishing fixed-photon-number probability and finite inclusive sum Bloch and Nordsieck 1937, § III, pp. 58–59, Eqs. (26)–(31).

Soft and collinear limits are distinct

For a charged particle of energy $E\gg m$, emission parallel to its momentum produces logarithms such as $\ln(E^2/m^2)$. The physical mass regulates the strict collinear divergence, but it need not make the logarithm small. In a massless approximation, final-state collinear safety requires an appropriate measurement, such as a specified photon–lepton recombination rule. If the observable resolves that splitting, a factorized description may instead require a fragmentation function and its scheme/scale dependence. Factorization does not make the original measurement collinear safe. Prepared incoming beams have their own mass-logarithm or distribution-function treatment; final-state recombination does not repair an initial-state collinear singularity Schwartz 2014, § 20.3.2, pp. 370–371.

The relevant limits therefore answer different questions:

Region	Momentum geometry	What makes a prediction usable?
Soft	$k^0\to0$ at generic angle	Sum over states unresolved by the energy/time resolution
Collinear	$k\parallel p_i$ at finite energy fraction	Keep the mass, or define recombination/factorization consistently
Soft-collinear	Both limits together	Avoid double counting when factorizing or resumming regions

The broader KLN statement concerns perturbative probability sums over complete degenerate initial and final sets, under unitarity, regularity and renormalization hypotheses. It is not a statement about arbitrary exclusive amplitudes or convergence of the perturbation series Lee and Nauenberg 1964, § II, pp. B1550–B1551, and § III, points 3 and 6–7, pp. B1552–B1554. The canonical discussion states these conditions. In practice, an initial-state beam is prepared rather than averaged over arbitrarily; surviving mass logarithms or factorization dependence must then be retained rather than attributed to KLN cancellation.

Three infrared-finite constructions

Inclusive probabilities. Sum ordinary Fock-space probabilities over every state that the stated detector resolution cannot distinguish. This is the usual collider and decay-rate construction. It predicts explicit, measurable dependence on the resolution.

Coherently dressed states. Incorporate the long-range field associated with the massive charged configuration into its asymptotic state. For a nonzero soft radiative kernel and a fixed soft matching scale $\Lambda$, the photon-number integral is proportional to $\int_\lambda^\Lambda\mathrm d\omega/\omega$, while its energy is proportional to $\int_\lambda^\Lambda\mathrm d\omega$. Thus a cloud can contain infinitely many soft photons with finite energy as $\lambda\to0$, and need not lie in the original finite-particle Fock space. Compatible dressings can cancel soft singularities at amplitude level, but the charges, velocities, Coulomb phases and Gauss-law constraints must match; a transverse photon cloud alone is not a complete gauge-invariant state prescription. Kulish and Faddeev construct the modified asymptotic dynamics and its phase operators, without proving general existence of the limiting scattering operator Kulish and Faddeev 1970, §§ 3–5, Russian original pp. 160–161 and 166–167, PDF. A dressed amplitude is not obtained merely by deleting the soft logarithm from an inclusive rate: the state space and observable have changed. The dressed-state construction and its limits are treated separately.

Finite-time or detector observables. A finite observation interval introduces a frequency-resolution scale of order its inverse, while a finite detector has energy and angular thresholds. Such descriptions can make the physical origin of the resolution transparent, but finite duration alone is not a proof of infrared finiteness: the charged-state preparation, measurement and limiting procedure must still be specified. Sending the observation time to infinity while reverting to bare charged Fock states restores the original infrared problem.

Virtual soft exchange can also produce a regulator-dependent Coulomb phase Weinberg 1995, § 13.2, p. 542, Eq. (13.2.6). A common phase cancels from a probability, including interference between amplitudes with the same external asymptotic currents. Comparing different charged configurations instead requires a consistent state and measurement prescription; the bare regulator-dependent phase is not itself an observable. Dressed asymptotic evolution must account for these phases as well as the real soft factor.

A calculation checklist

1. Define the measured bin. State the energy frame, angular acceptance, recombination rule, particle masses, and whether extra charged pairs are degenerate.
2. Identify every singular region. Test soft, collinear, and overlapping limits of both loop and phase-space integrands.
3. Use one regulator consistently. Keep the same photon mass, dimensional regulator, or auxiliary prescription in the real and virtual pieces until after addition.
4. Verify coefficients before finite terms. The coefficient multiplying $\ln\lambda$ or the infrared pole must agree with opposite sign between degenerate real and virtual contributions.
5. Remove only the regulator. The final answer may and generally should retain $\Delta E$, angular cuts, masses, and factorization scales.
6. Assess large logarithms. Check the actual enhanced combinations, including $(\alpha/\pi)B\ln(Q/\Delta E)$ and any additional soft-collinear logarithms, before trusting fixed order.

Three especially useful diagnostics are the gauge test $k\cdot J_{\mathrm s}=0$, the regulator derivative of the inclusive result, and monotonicity: increasing the unresolved phase space must not decrease the real-emission contribution.

Apply them alongside the hard-process normalization and Ward checks: infrared cancellation cannot correct a missing spin average or flux factor.

Common pitfalls

Calling a regulator an experimental resolution. A fictitious photon mass or an infrared dimensional pole must disappear. The energy threshold, angular acceptance, and recombination radius are physical definitions and remain in the answer.

Adding rates for nondegenerate events. Real–virtual cancellation occurs only after the measurement maps the two configurations to the same outcome. A vetoed resolved photon cannot cancel a virtual correction in the accepted bin.

Using “KLN” as a blanket eraser. The theorem requires a specified sum over degenerate states. Prepared initial states, non-collinear-safe lepton definitions, or an incomplete sum can leave mass logarithms and factorization dependence without contradicting it.

Equating inclusive and dressed predictions. Both can be infrared finite, but they answer questions about different observables and asymptotic states. Their finite terms need not agree without a matching prescription.

Exercises

Resolution dependence and an angular veto. Keep the Compton kinematics fixed at recoil speed $\beta=3/5$, with the unresolved energy defined in the incoming electron’s rest frame. At leading soft accuracy, how much does the relative inclusive rate change when the threshold increases from $\Delta E$ to $2\Delta E$, with both well below every hard scale? Then suppose photons in an angular patch $\mathcal A$ are vetoed at every energy, however small. Define $B_{\mathcal A}=(4\pi)^{-1}\int_{\mathcal A}(-j^2)\,\mathrm d\Omega>0$ and find the uncancelled logarithm.

Solution

At $\beta=3/5$, the angular coefficient is $B=(10/3)\ln2-2>0$. Increasing the threshold adds the positive real-emission integral over the energy interval from $\Delta E$ to $2\Delta E$; the virtual term is unchanged. Thus, within leading soft accuracy,

$$\frac{\sigma_{\mathrm{incl}}(2\Delta E)-\sigma_{\mathrm{incl}}(\Delta E)}{\sigma_0} =\frac{\alpha}{\pi} \left(\frac{10}{3}\ln2-2\right)\ln2.$$

This does not determine power corrections from recoil, finite-energy emission or higher orders. For the angular veto, the virtual logarithm still contains the full $B$, but the real coefficient becomes $B-B_{\mathcal A}$. Relative to the soft-inclusive result, the missing contribution is

$$\delta_{\mathrm{veto}}-\delta_{\mathrm{incl}} =-\frac{\alpha}{\pi}B_{\mathcal A} \ln\frac{\Delta E}{\lambda}.$$

It diverges negatively at this fixed order as $\lambda\to0$. The veto changes the measurement weight even for an arbitrarily soft photon in $\mathcal A$, so it fails the soft-safety test. A veto only above a fixed positive energy is different: the truly soft states are still summed, leaving a finite resolution logarithm. An uninstrumented angular patch is also different, since its unobserved photons are included rather than forbidden.

References

• Bloch, Felix, and Arnold Nordsieck. “Note on the Radiation Field of the Electron.” Physical Review 52 (1937): 54–59. DOI.
• Kulish, P. P., and L. D. Faddeev. “Asymptotic Conditions and Infrared Divergences in Quantum Electrodynamics.” Theoretical and Mathematical Physics 4 (1970): 745–757. DOI. Russian original: Teoreticheskaya i Matematicheskaya Fizika 4, no. 2 (1970): 153–170. Open PDF, Russian original.
• Lee, T. D., and M. Nauenberg. “Degenerate Systems and Mass Singularities.” Physical Review 133 (1964): B1549–B1562. DOI.
• Schwartz, Matthew D. Quantum Field Theory and the Standard Model. Cambridge: Cambridge University Press, 2014. DOI.
• Weinberg, Steven. The Quantum Theory of Fields. Volume I: Foundations. Cambridge: Cambridge University Press, 1995. DOI.

Further reading

• Kinoshita, Toichiro. “Mass Singularities of Feynman Amplitudes.” Journal of Mathematical Physics 3 (1962): 650–677. DOI.