Factorization and coefficient functions
The theorem, MS̄ scheme, higher twist — why PDFs are universal and where that fails.
Asymptotic freedom (§2.1) makes the photon–quark collision computable; confinement makes the proton itself incomputable. The factorization theorem is the licence to have both at once: it states that a DIS structure function splits — up to controlled power corrections — into a perturbative coefficient function convolved with a non-perturbative but universal PDF. This chapter states the theorem precisely, shows where the split actually happens in a calculation (mass factorization of collinear logarithms), derives why nothing physical depends on the artificial scale $\mu_F$ the split introduces, and quantifies the $\mathcal{O}(\Lambda^2/Q^2)$ corrections our runcard cuts are designed to keep out.
The theorem, stated precisely
For DIS structure functions, collinear factorization reads:
— compactly, $F_2 = x\sum_i C_i \otimes f_i$, where $\otimes$ denotes exactly that scale-invariant convolution. Read every factor: $f_i(\xi,\mu_F^2)$ is the PDF — non-perturbative, unknowable analytically, universal (the same function in every process); $C_i$ is the coefficient function — computed order by order in $\alpha_s$, process-dependent (different for $F_2$, $F_L$, Drell–Yan) but target-independent; $\mu_F$ is the factorization scale, the bookkeeping boundary between "short-distance" (into $C$) and "long-distance" (into $f$); and the error term is power-suppressed, not $\alpha_s$-suppressed — a different expansion entirely (last section). Universality is the entire business model of a global fit: extract $f_i$ from DIS, predict Drell–Yan, LHC, neutrino scattering. It is a falsifiable claim, and forty years of global fits are its ongoing test.
Why it works — timescales in the Breit frame
Sit in the Breit frame (§1.3), where the proton carries momentum $P_z \to \infty$ against the photon. Two clocks tick at utterly different rates. The hard clock: the virtual photon is absorbed over a time $\sim 1/Q$ — at $Q = 10$ GeV, about $0.02$ fm/$c$. The soft clock: the partons rearrange, exchange soft gluons, and maintain the bound state over the hadronic timescale $\sim 1/\Lambda$ in their rest frame, which time dilation stretches by $\gamma \sim Q/(x\Lambda)$ in the Breit frame. The photon therefore photographs a frozen configuration: during the absorption, the struck quark cannot communicate with the spectators. No interference between emission (long-time) and absorption (short-time) survives the mismatch — amplitudes decohere, and probabilities multiply: flux of quarks at momentum fraction $\xi$, times cross-section on a free quark. That product-of-probabilities structure is the convolution above. The formal proof (Collins, Soper, Sterman) upgrades this picture to all orders — showing soft-gluon exchanges between the struck quark and the remnant cancel in inclusive DIS — and that proof is precisely what fails at $\mathcal{O}(\Lambda^2/Q^2)$, where multi-parton correlations (higher twist) enter.
Where the split happens: collinear singularities and mass factorization
Factorization becomes concrete — and subtle — the moment you compute $C_i$ beyond leading order. Compute DIS on a free quark target at $\mathcal{O}(\alpha_s)$: the quark can emit a gluon before absorbing the photon. When the gluon is emitted exactly parallel to the quark, the intermediate propagator $1/(p-k)^2 \to \infty$: a collinear singularity. Regulate it with a small quark mass $m$ and the partonic structure function comes out as
The logarithm diverges as $m \to 0$ — perturbation theory on a parton target is infrared sick. But look at what multiplies it: $P_{qq}(z)$, a universal function (the splitting kernel we derive in §2.3), independent of the photon probe. The divergence comes from long-distance propagation of a nearly-on-shell quark — physics that in a real proton belongs to the bound state, not the probe. So absorb it where it belongs:
Split the offending logarithm at an arbitrary intermediate scale $\mu_F$:
$$ \ln\frac{Q^2}{m^2} \;=\; \underbrace{\ln\frac{Q^2}{\mu_F^2}}_{\text{short-distance}} \;+\; \underbrace{\ln\frac{\mu_F^2}{m^2}}_{\text{long-distance}} $$
Define the (renormalized) quark-in-quark PDF to contain the long-distance half:
$$ f_{q/q}(\xi, \mu_F^2) \;=\; \delta(1-\xi) \;+\; \frac{\alpha_s}{2\pi}\, P_{qq}(\xi)\, \ln\frac{\mu_F^2}{m^2} \;+\; \mathcal{O}(\alpha_s^2) $$
and demand $\hat F_2^{\,q} = z\sum_i C_i \otimes f_{i/q}$ order by order. At $\mathcal{O}(\alpha_s^0)$: $C_q^{(0)}(z) = e_q^2\,\delta(1-z)$. At $\mathcal{O}(\alpha_s)$, the convolution of $C^{(0)}$ with the $\alpha_s$ piece of $f_{q/q}$ supplies $e_q^2\,(\alpha_s/2\pi) P_{qq} \ln(\mu_F^2/m^2)$; subtracting it from $\hat F_2^{\,q}$ leaves
$$ C_q^{(1)}(z) \;=\; e_q^2\, \frac{\alpha_s}{2\pi} \left( P_{qq}(z) \ln\frac{Q^2}{\mu_F^2} + c_q(z) \right) $$
— finite as $m \to 0$. All sensitivity to the quark mass (in reality: to confinement-scale physics) now lives inside $f$, which for a real proton we were never going to compute anyway — we fit it. The divergence hasn't been solved; it has been filed correctly. Crucially, this works only because the collinear log is universal: the same $P_{qq}\ln(\mu_F^2/m^2)$ arises in Drell–Yan, in jet production, in everything — so one and the same redefinition of $f$ cures them all. That is why factorized PDFs are transferable between processes.
$\mu_F$ was invented in the previous step; $F_2$ is measured and cannot know about it. Demand:
$$ 0 \;=\; \frac{d\,F_2}{d\ln\mu_F^2} \;=\; x\sum_i\left[ \frac{d C_i}{d\ln\mu_F^2} \otimes f_i \;+\; C_i \otimes \frac{d f_i}{d\ln\mu_F^2} \right] $$
From Derivation 1, $dC_q^{(1)}/d\ln\mu_F^2 = -e_q^2 (\alpha_s/2\pi) P_{qq}$. Substituting into the first term and requiring cancellation against the second forces, at this order,
$$ \frac{d f_q}{d\ln\mu_F^2} \;=\; \frac{\alpha_s}{2\pi}\, P_{qq} \otimes f_q \;+\; \ldots $$
— which is DGLAP. The evolution equation of §2.3 is not an extra postulate: it is the integrability condition of factorization, the statement that shuffling the boundary between $C$ and $f$ changes nothing physical. The cancellation is exact only order by order: truncate at $\mathcal{O}(\alpha_s^n)$ and a residual $\mu_F$-dependence of $\mathcal{O}(\alpha_s^{n+1})$ survives. Varying $\mu_F$ (and $\mu_R$) is therefore the standard estimator of missing higher orders — the MHOU covariance matrices of §2.4 are built from exactly this residual.
Scheme dependence — what "$\overline{\rm MS}$ PDF" actually means
Derivation 1 hid a choice. The split of $\ln(Q^2/m^2)$ was forced, but finite terms are up for grabs: one can move any $z$-dependent constant from $c_q(z)$ into the definition of $f$. Each convention is a factorization scheme. The universal choice is $\overline{\rm MS}$: regulate dimensionally instead of with a mass, absorb the pole $1/\epsilon$ plus the bookkeeping constants $\ln 4\pi - \gamma_E$ into $f$, and nothing more. The consequences are worth stating bluntly: a PDF is a scheme-dependent object — "the gluon at $x = 10^{-3}$" has no meaning without "in $\overline{\rm MS}$"; coefficient functions are scheme-dependent in the compensating way; only the product $C \otimes f$ is physical. Mixing an $\overline{\rm MS}$ coefficient with a differently-schemed PDF is a silent, order-$\alpha_s$ error. Our entire pipeline — MSHT20 parametrization, EKO evolution, FK-table coefficients under theory id 40000000 — is $\overline{\rm MS}$ end to end, which is what makes its pieces interchangeable with the global-fit world's.
LO recovers the parton model; NLO changes its character
Insert $C_q^{(0)} = e_q^2\,\delta(1-z)$ into the master formula: the delta collapses the convolution,
— Chapter 1's parton model, now derived as the leading term of a systematic expansion, with its $x$-scaling violated only by the slow $\mu_F^2$-dependence of the PDFs. At NLO, two qualitative changes. First, the quark coefficient acquires the terms $\big[\tfrac{\ln(1-z)}{1-z}\big]_+$ and $\ln(1-z)$ — remnants of soft-gluon emission (the plus-distribution is defined carefully in §2.3) — which grow large as $z \to 1$ and make large-$x$ predictions delicate; their all-order pattern is what threshold resummation organizes. Second, a gluon channel opens: $\gamma^* g \to q\bar q$ contributes $C_g \otimes g$, so $F_2$ becomes directly sensitive to the gluon at $\mathcal{O}(\alpha_s)$ — the main handle on $g(x)$ in DIS-only fits like ours, and the reason our gluon uncertainties balloon where DIS data run out (the wide gluon bands in our results).
Hadron colliders: the double convolution
The same theorem, with two hadrons in the initial state:
Everything doubles: two PDFs, a sum over parton pairs, and — after discretization — predictions that are quadratic rather than linear in the PDF vector. This is exactly why our FK tables come in two kinds (linear for DIS, quadratic for hadronic data, §3.3), and why adding collider data to a fit is a structural upgrade, not just more rows.
Factorization is a theorem for inclusive DIS and Drell–Yan — proven to all orders in $\alpha_s$, up to $\mathcal{O}(\Lambda^2/Q^2)$. For more exclusive observables (transverse-momentum spectra, diffraction, low-$p_T$ hadroproduction) collinear factorization can be modified or genuinely broken. When our fits quote a PDF uncertainty, they are quoting it inside the factorization framework; the framework's own accuracy — power corrections, the perturbative order of $C$, evolution — is a separate, theory-side error budget. Keeping those two budgets distinct is a core discipline of this lab.
Power corrections: higher twist and target mass
The $\mathcal{O}(\Lambda^2/Q^2)$ term is where the frozen-spectator argument leaks. Two distinct sources. Higher twist: the photon resolves two partons at once — the struck quark reinteracts with the remnant during absorption; these multi-parton correlations scale as $\mu^2/Q^2$ with $\mu^2$ a hadronic matrix element, empirically up to $\sim\!0.5$ GeV$^2$ at large $x$. They are not part of any PDF and, if present in fitted data, masquerade as distorted partons. Target-mass corrections: purely kinematic $M^2 x^2/Q^2$ effects from the proton's mass (Nachtmann variable $\xi_N = 2x/(1 + \sqrt{1 + 4M^2x^2/Q^2})$ replacing $x$), calculable, but signalling proximity to the resonance region. The defence is the runcard: cuts of order $Q^2 \ge 3.5$ GeV$^2$ and $W^2 \ge 12.5$ GeV$^2$ (the NNPDF-style defaults our pipeline inherits) fence the dataset into the leading-twist region — worked example 2 puts numbers on how much they exclude.
Worked example 1 — a toy convolution, end to end
Given: a caricature NLO coefficient $C(z) = \delta(1-z) + \frac{\alpha_s}{2\pi}\, 2C_F \ln(1-z)$ (the genuine NLO $c_q$ contains exactly such logs, organized into plus-distributions), a valence-like PDF $f(\xi) = \xi^{-1/2}(1-\xi)^3$, and $\alpha_s = 0.30$ (low-$Q^2$, per §2.1). Evaluate $(C \otimes f)(x) = \int_x^1 \frac{d\xi}{\xi}\, C(x/\xi)\, f(\xi)$ at $x = 0.1$.
Leading term (exact): the $\delta(1 - x/\xi)$ fires at $\xi = x$ and the $d\xi/\xi$ Jacobian cancels: contribution $= f(0.1) = (0.1)^{-1/2}(0.9)^3 = 3.1623 \times 0.729 = 2.305$.
Correction integral: $I = \int_{0.1}^{1} d\xi\; \xi^{-3/2} (1-\xi)^3 \ln\!\big(1 - 0.1/\xi\big)$. The integrand has an integrable log singularity at $\xi = 0.1$: near threshold, $\int_{0.1}^{0.11} \approx 21 \int_0^{0.01} \ln(u/0.105)\, du = 21 \times 0.01 \times (\ln 0.095 - 1) \approx -0.70$. Simpson's rule on the rest (sampled at $\xi = 0.11, 0.13, 0.15, 0.175, 0.2, 0.25, 0.3, \ldots, 1$): segments contribute $-0.94$, $-0.35$, $-0.20$, $-0.05$, $-0.02$, $-0.002$. Total $I \approx -2.3$.
Assemble: correction $= \frac{0.30}{2\pi} \times \frac{8}{3} \times (-2.3) = -0.29$, so $(C\otimes f)(0.1) \approx 2.305 - 0.29 = 2.01$ — a $-13\%$ NLO shift.
Three lessons. (1) A convolution samples $f$ at all $\xi \ge x$: the prediction at $x = 0.1$ knows about the PDF up to $\xi = 1$ — which is why data at one $x$ constrain parameters everywhere, and why PDF uncertainties are so correlated across $x$. (2) $\ln(1-z)$ made the correction largest from the $\xi \approx x$ region (soft emission), foreshadowing the plus-distributions of §2.3. (3) Corrections of tens of percent at $\alpha_s \approx 0.3$ are normal: NNLO is not a luxury at our $Q_0$ — it is the working order of theory id 40000000.
Worked example 2 — how big is the fine print?
Generic size: with $\Lambda \approx 0.3$ GeV, $\Lambda^2/Q^2$ at the $Q^2$ cut ($3.5$ GeV$^2$) is $0.09/3.5 = 2.6\%$ — comparable to fixed-target statistical errors, so the cut sits exactly where it must. At the top of BCDMS coverage ($Q^2 \approx 230$ GeV$^2$): $0.04\%$, negligible. At HERA's $Q^2 \sim 10^4$ GeV$^2$: $10^{-5}$ — power corrections are a fixed-target problem.
Large $x$ is worse than generic: empirical higher-twist fits find $F_2 \to F_2^{\rm LT} \big(1 + C_{\rm HT}(x)/Q^2\big)$ with $C_{\rm HT}(0.75) \sim 0.5$–$1$ GeV$^2$: at $Q^2 = 5$ GeV$^2$ that is a $10$–$20\%$ distortion. Check what the $W^2$ cut does there: $W^2 = M^2 + Q^2(1-x)/x \ge 12.5$ GeV$^2$ at $x = 0.75$ requires $Q^2 \ge (12.5 - 0.88) \times 0.75/0.25 = 34.9$ GeV$^2$ — pushing the correction down to $1.4$–$2.9\%$, and the target-mass ratio to $M^2x^2/Q^2 = 0.88 \times 0.5625/34.9 = 1.4\%$. The $W^2$ cut is doing precision work: it removes precisely the high-$x$, low-$Q^2$ corner where "PDF" would silently mean "PDF plus quark–gluon correlations".