betaFeedback

WorksPhysics and Astronomy

Impact of errors-on-errors on the PDF determination from combined HERA data

Abstract

We study how the determination of parton distribution functions (PDFs) from the combined HERA inclusive deep-inelastic-scattering data is affected when the quoted sizes of the systematic errors are themselves treated as uncertain. We refer to these uncertainties informally as errors-on-errors and account for them using the Gamma Variance Model, which treats the squared quoted systematic errors as Gamma-distributed estimates of unknown variances. As a result, the usual quadratic nuisance parameter constraints in the χ2χ^2 are replaced with logarithmic ones. We present two applications of the framework. In the first, a robustness test estimates the largest errors-on-errors effects that could be present in our reproduction of the HERAPDF2.0 NNLO fit, reducing χ2/ndofχ^2/n_{\rm dof} from 1363/1131 to 1306/1131, with the main impact concentrated at low Bjorken xx and low squared momentum transfer, Q2Q^2. Comparable reductions are observed with a more flexible small-xx parametrisation, both at fixed NNLO and with NNLO+NLLx_x resummation. In the second, the error-on-error parameters are instead assigned a priori, as they would be in a PDF determination intended for publication. Taken together, the results indicate that errors-on-errors can account for a substantial part of the apparent data-theory tension in the combined HERA dataset, improving the inferred compatibility and widening the fitted PDF uncertainties without appreciably displacing the PDFs or cross-section predictions themselves.

Introduction

Modern global parton distribution function (PDF) determinations obtain their constraining power by combining precision measurements from different experiments, processes and kinematic regions. This variety of inputs is essential for separating parton flavours and determining their dependence on xx, the proton momentum fraction carried by the parton, but can also expose tensions when datasets that are individually well described prefer different PDFs in a simultaneous fit. Such tensions are often reflected in poor goodness of fit, with values of χ2/ndof\chi^2/n_{\rm dof} (the reduced χ2\chi^2) significantly above unity, where χ2\chi^2 is evaluated at its minimum and ndofn_{\rm dof} denotes the number of degrees of freedom.

The combined HERA I+II inclusive deep-inelastic-scattering (DIS) dataset [1], released with its reference QCD analysis HERAPDF2.0, is a cornerstone of modern global PDF determinations. It brings together the H1 and ZEUS neutral- and charged-current e±pe^\pm p cross sections over a broad range in Bjorken xx and Q2Q^2, the negative squared four-momentum transfer. The lower-Q2Q^2 neutral-current measurements constrain the low-xx sea and, through their scale dependence, the gluon, while the high-Q2Q^2 neutral- and charged-current data constrain the valence distributions. At next-to-next-to-leading order (NNLO), the HERAPDF2.0 fit to these data yields χ2/ndof=1363/1131\chi^2/n_{\rm dof}=1363/1131, well above the range expected from nominal statistical fluctuations.

Several effects may contribute to an increase in the observed χ2\chi^2, including limitations of the perturbative calculation or heavy-flavour treatment, insufficient flexibility of the input PDF parametrisation and assumptions in the experimental uncertainty and correlation model. For example, Bonvini and Giuli [2] showed that a more flexible low-xx PDF parametrisation reduced the minimum χ2\chi^2 by 62 units compared with the fit using the HERAPDF2.0 parametrisation. They also found that including next-to-leading logarithmic small-xx (NLLx_x) resummation matched to NNLO reduced the minimum χ2\chi^2 by 28–29 units relative to the corresponding fixed-order NNLO fit. This comparison used the separate APFEL+HELL/FONLL-C setup introduced in Ref. [3].

We focus here on imperfect knowledge of the quoted systematic uncertainties of the input measurements. Their magnitudes are often estimated from calibration or simulated samples of finite size, or by varying modelling assumptions and analysis choices. Limited sample sizes and assumptions about the choice and range of these variations introduce uncertainty into the estimated systematic-error scales themselves. We refer to this additional uncertainty as errors-on-errors. Treating the quoted systematic uncertainties as exact can make discrepancies between data and theory appear more significant than they are, particularly when those uncertainties are underestimated. This can bias the extracted PDFs and lead to underestimated uncertainties. We relax this approximation by explicitly accounting for uncertainty in the systematic-error scales within the fit using the Gamma Variance Model (GVM), introduced in [4] and further developed in [5, 6].

Previously, the broader issue of imperfect experimental-uncertainty modelling has been investigated from several perspectives: PDF fits to differential ttˉt\bar{t} data have shown a marked sensitivity to the assumed correlation model of the dominant systematic uncertainties, with a non-negligible effect on the fitted gluon [7]; the effect of deliberately introduced data inconsistencies on the fitted PDFs has been studied with idealised pseudodata [8]; and closure tests have been used to assess whether PDF uncertainties remain reliable when correlated experimental uncertainties are underestimated [9]. In these approaches, the quoted uncertainties enter the fit as exact inputs; here instead the uncertainty in the systematic-error scales is introduced explicitly as part of the statistical model through the GVM.

We apply this framework to the combined HERA inclusive dataset by implementing the GVM in xFITTER, an open-source framework for QCD analyses and PDF determinations [10, 11, 12]. Sections 2 and 3 present the model, implementation and fit setup. We then consider two complementary applications. First, Sec. 4 presents a robustness test designed to estimate the largest plausible impact of errors-on-errors on the minimum χ2\chi^2, fitted PDFs and cross-section predictions. To distinguish flexibility in the uncertainty model from limitations of the PDF shape and small-xx theory, Sec. 5 repeats the study with the more flexible Bonvini–Giuli parametrisation and with NNLO+NLLx_x resummed predictions. Second, Sec. 6 presents an illustrative informed application, in which errors-on-errors are assigned a priori using information external to the PDF fit rather than tuned to reduce χ2\chi^2. Conclusions are given in Sec. 7.

Errors-on-errors in PDF fits

The GVM provides a statistical treatment of uncertainties in the assigned sizes of systematic errors, referred to here as errors-on-errors [4, 5, 6]. We denote the conditional density of the measurements by P(y∣a,b)P(\boldsymbol{y}\mid\boldsymbol{a},\boldsymbol{b}), where y\boldsymbol{y} is the vector of measured values, a\boldsymbol{a} denotes the parameters of interest (POIs) entering the theoretical predictions, and b\boldsymbol{b} denotes the nuisance parameters (NPs) introduced to account for systematic uncertainties. For PDF fits, a\boldsymbol{a} denotes the parameters defining the PDFs at the input scale, while systematic effects are modelled as biases in the measured values y\boldsymbol{y} that depend linearly on b\boldsymbol{b}. To follow the notation commonly used in the PDF-fitting community, these parameters are denoted by a\boldsymbol{a} and b\boldsymbol{b}, rather than by μ\boldsymbol{\mu} and θ\boldsymbol{\theta}, as in Refs. [4, 5, 6].

For each systematic source ss, the NP bsb_s is constrained by a Gaussian auxiliary estimate usu_s, which represents the best estimate of the bias induced in the data by the source. The uncertainty assigned to this estimate, σus\sigma_{u_s}, is commonly referred to as the systematic uncertainty for source ss. The GVM treats the square of this uncertainty, i.e. the variance σus2\sigma_{u_s}^2, as an adjustable parameter and models its estimate vsv_s as Gamma-distributed. This choice is motivated by the sample variance of independent observations from a Gaussian distribution, which follows a scaled χ2\chi^2 distribution and hence a Gamma distribution; see Ref. [4] for further details. Up to terms independent of the fitted parameters, the full χ2\chi^2 function of the model is

χ2(a,b,σu2)=−2log⁡P(y∣a,b)+∑s[(us−bs)2σus2+(1+12ϵs2)log⁡σus2+vs2ϵs2σus2].(1)\chi^2(\boldsymbol{a},\boldsymbol{b},\sigma_{u}^2)=-2\log P(\boldsymbol{y}\mid\boldsymbol{a},\boldsymbol{b})+\sum_s\left[\frac{(u_s-b_s)^2}{\sigma_{u_s}^2}+\left(1+\frac{1}{2\epsilon_s^2}\right)\log\sigma_{u_s}^2+\frac{v_s}{2\epsilon_s^2\sigma_{u_s}^2}\right]. \tag*{(1)}

The parameter ϵs\epsilon_s quantifies how well the systematic uncertainty σus\sigma_{u_s} is known: it represents the relative uncertainty on its estimate vs\sqrt{v_s}, and is therefore referred to as the error-on-error parameter.1 For example, ϵs=0.3\epsilon_s=0.3 corresponds approximately to a 30% relative uncertainty on the quoted size of the systematic error. We adopt a frequentist approach and profile over the systematic variances at fixed a\boldsymbol{a} and b\boldsymbol{b}. 2 This gives

σ^us2(bs)=vs+2ϵs2(us−bs)21+2ϵs2.(2)\widehat{\sigma}_{u_s}^2(b_s)=\frac{v_s+2\epsilon_s^2(u_s-b_s)^2}{1+2\epsilon_s^2}. \tag*{(2)}

Substituting into Eq. (1) and subtracting additive terms independent of the fitted parameters gives

χ2(a,b)=−2log⁡P(y∣a,b)+∑s(1+12ϵs2)log⁡[1+2ϵs2(us−bs)2vs].(3)\chi^2(\boldsymbol{a},\boldsymbol{b})=-2\log P(\boldsymbol{y}\mid\boldsymbol{a},\boldsymbol{b})+\sum_s\left(1+\frac{1}{2\epsilon_s^2}\right)\log\left[1+2\epsilon_s^2\frac{(u_s-b_s)^2}{v_s}\right]. \tag*{(3)}

This expression shows how uncertainty in the size of a systematic error affects the constraint on the corresponding NP. A larger ϵs\epsilon_s means that the nominal estimate of σus\sigma_{u_s} is less precisely known. The logarithmic constraint accounts for this by allowing greater freedom for large shifts of the NP away from its auxiliary estimate usu_s: the penalty for such shifts grows more slowly than in the usual quadratic treatment. The fit can therefore accommodate experimental tensions associated with that source through a larger systematic adjustment. For small ϵs\epsilon_s, expanding the logarithmic terms in Eq. (3) gives

(1+12ϵs2)log⁡[1+2ϵs2(us−bs)2vs]=(us−bs)2vs+O(ϵs2),(4)\left(1+\frac{1}{2\epsilon_s^2}\right)\log\left[1+2\epsilon_s^2\frac{(u_s-b_s)^2}{v_s}\right]=\frac{(u_s-b_s)^2}{v_s}+\mathcal{O}(\epsilon_s^2), \tag*{(4)}

so the usual quadratic constraint is recovered as ϵs→0\epsilon_s\to0.

The best estimates of the model parameters, denoted by a^\widehat{\boldsymbol{a}} and b^\widehat{\boldsymbol{b}}, are defined by

(a^,b^)=arg min⁡a,b χ2(a,b).(5)(\widehat{\boldsymbol{a}},\widehat{\boldsymbol{b}})=\underset{\boldsymbol{a},\boldsymbol{b}}{\operatorname{arg\,min}}\ \chi^2(\boldsymbol{a},\boldsymbol{b}). \tag*{(5)}

These are the maximum-likelihood estimators. The minimum χmin⁡2=χ2(a^,b^)\chi^2_{\min}=\chi^2(\widehat{\boldsymbol{a}},\widehat{\boldsymbol{b}}) is used as the goodness-of-fit statistic. It measures how well the fitted model describes the data relative to their assigned uncertainties. In the Gaussian limit and for a correctly specified model it is approximately χ2\chi^2-distributed with ndof=Ndata−nan_{\mathrm{dof}}=N_{\mathrm{data}}-n_a degrees of freedom, where NdataN_{\mathrm{data}} is the number of cross-section measurements and nan_a the number of fitted POIs. Hence, for a satisfactory fit, the reduced χ2\chi^2 should be close to unity, with relative fluctuations of order 2/ndof\sqrt{2/n_{\mathrm{dof}}}. Significantly larger values indicate a poor description of the data or underestimated uncertainties, whereas significantly smaller values can indicate overestimated uncertainties or overfitting.

Confidence intervals in parameter space can be constructed from the χ2\chi^2 function in two closely related ways. In the Δχ2\Delta\chi^2 approach, a value of a given parameter is retained as long as the increase from the global minimum, Δχ2=χ2−χmin⁡2\Delta\chi^2 = \chi^2 - \chi^2_{\min}, stays below a prescribed threshold. Alternatively, near the minimum the χ2\chi^2 surface can be approximated quadratically, with its local curvatures and parameter correlations encoded in the Hessian

H≡12∇(a,b)2χ2∣(a^,b^).(6)H \equiv\frac{1}{2}\nabla^2_{(a,b)}\chi^2\bigg|_{(\hat{a},\hat{b})}. \tag*{(6)}

In this quadratic limit H−1H^{-1} is the estimated covariance matrix of the fitted parameters and the two constructions coincide: for a single parameter of interest, the standard choice Δχ2≤1\Delta\chi^2 \le1 gives an approximate 68.3% confidence interval, a^m±(H−1)amam\hat{a}_m \pm\sqrt{(H^{-1})_{a_m a_m}}. Away from the quadratic regime the two differ, the Δχ2\Delta\chi^2 intervals being in general asymmetric. For further details on the Hessian method for computing uncertainties in PDF fits, see, e.g., Ref. [13]. For nonzero errors-on-errors, corrections are needed when using the minimum χ2\chi^2 to assess goodness of fit and when constructing confidence intervals, as discussed in § 2.2.

Specialisation to xFitter

To apply the GVM to PDF fits, we implement it in xFitter [10, 11, 12], an open-source framework for QCD analyses and PDF determinations used by both the theory and experimental communities. The HERAPDF2.0 analysis was performed within this framework, then known as HERAFitter, and xFitter has since been widely used in LHC measurements to assess their impact on PDFs; see, for example, Refs. [14, 15, 16, 17, 18].

For the present study, we use the HERAPDF2.0 χ2\chi^2 definition implemented in xFitter. At fixed values of the NPs, the likelihood is a product of independent Gaussian terms for the individual measurements. The statistical and uncorrelated systematic uncertainties determine the point-by-point variance, while correlated systematic effects enter through NP shifts.3 The resulting data contribution to the χ2\chi^2 is

−2log⁡P(y∣a,b)=∑i[yi−Ti(a)(1−∑sγisbs)]2δi,stat2yiTi(a)+δi,uncor2Ti2(a)+∑ilog⁡(δi,uncor2Ti2(a)+δi,stat2yiTi(a)(δi,uncor2+δi,stat2)yi2).(7)-2\log P(\boldsymbol{y}\mid\boldsymbol{a},\boldsymbol{b}) = \sum_i \frac{\left[y_i-T_i(\boldsymbol{a})\left(1-\sum_s\gamma_{is}b_s\right)\right]^2}{\delta_{i,\mathrm{stat}}^2 y_i T_i(\boldsymbol{a})+\delta_{i,\mathrm{uncor}}^2 T_i^2(\boldsymbol{a})} + \sum_i \log\left(\frac{\delta_{i,\mathrm{uncor}}^2T_i^2(\boldsymbol{a})+\delta_{i,\mathrm{stat}}^2y_iT_i(\boldsymbol{a})}{\left(\delta_{i,\mathrm{uncor}}^2+\delta_{i,\mathrm{stat}}^2\right)y_i^2}\right). \tag*{(7)}

Here, yiy_i denotes the measured cross section and Ti(a)T_i(\boldsymbol{a}) the corresponding theoretical prediction. The coefficients γis\gamma_{is} describe the relative response of measurement ii to systematic source ss. Since the same NP controls the effect of that source on all affected measurements, it links their shifts and introduces correlations. The quantities δi,stat\delta_{i,\mathrm{stat}} and δi,uncor\delta_{i,\mathrm{uncor}} denote the relative statistical and uncorrelated systematic uncertainties. The uncorrelated systematic variance scales with Ti2T_i^2, corresponding to a multiplicative uncertainty. The statistical term instead uses yiTiy_iT_i, reflecting Poisson counting statistics. The second sum is the logarithmic normalisation required because this variance depends on the PDF parameters a\boldsymbol{a} through Ti(a)T_i(\boldsymbol{a}); its parameter-independent denominator makes the logarithmic term vanish when Ti=yiT_i=y_i, without changing the fitted parameter values. Further details on the construction of this χ2\chi^2 can be found in Ref. [1].

To incorporate errors-on-errors, we substitute the data contribution in Eq. (7) into the profiled GVM of Eq. (3). We assume that any known systematic bias has already been removed from the data, so that νs=0\nu_s=0. We express bsb_s in units of the nominal systematic uncertainty, absorbing this scale into the corresponding column of the γis\gamma_{is} matrix, so that νs=1\nu_s=1. With these conventions, the resulting χ2\chi^2 is

χ2(a,b)=−2log⁡P(y∣a,b)+∑pbp2+∑q(1+12ϵq2)log⁡(1+2ϵq2bq2).(8)\chi^2(\mathbf{a},\mathbf{b})=-2\log P(\mathbf{y}\mid\mathbf{a},\mathbf{b})+\sum_p b_p^2 +\sum_q\left(1+\frac{1}{2\epsilon_q^2}\right)\log\left(1+2\epsilon_q^2b_q^2\right). \tag*{(8)}

The index pp runs over correlated systematic sources whose uncertainties are treated as exactly known, whereas qq runs over correlated systematic sources assigned a non-zero error-on-error ϵq\epsilon_q. The use of relative systematic errors does not alter the GVM construction: for each source, bsb_s rescales the complete absolute response vector with components TiγisT_i\gamma_{is}.

In this application, errors-on-errors affect the correlated systematic sources, while the statistical and aggregated uncorrelated uncertainties in Eq. (7) are unchanged. The contributions of individual sources to the uncorrelated systematic uncertainty could also be treated through NPs, as discussed in Ref. [6]. However, the combined HERA dataset provides only the total uncorrelated systematic uncertainty for each measurement, without a breakdown by source.

Departures from the Gaussian limit

For nonzero error-on-error parameters, the logarithmic NP constraints in Eq. (8) depart from the quadratic form recovered in the Gaussian limit of Eq. (4). This departure has two practical consequences. The NP profiling equations, used to minimise χ2\chi^2 over the NPs at fixed PDF parameters, are linear for quadratic constraints but become nonlinear with errors-on-errors. Moreover, the minimum χ2\chi^2 can depart from its usual approximate distribution, and the confidence intervals described above may be too narrow, requiring statistical corrections. We summarise the treatment of both effects here, while their derivations and implementation details are given in Apps. § 9 and § 10.

The PDF parameters a\mathbf{a} are determined numerically. The NPs can either be included in this numerical minimisation or profiled for each hypothesised value of a\mathbf{a}. The former, referred to here as external minimisation, provides the reference solution of Eq. (8) but increases the dimension of the numerical problem by the number of correlated systematic sources. For quadratic constraints, the NPs can be profiled analytically in closed form, avoiding this increase in the dimension of the numerical minimisation; see, e.g., Ref. [19]. Its extension to the GVM is, however, non-trivial, since the logarithmic constraints lead to coupled nonlinear equations for the NPs. The iterative procedure summarised below provides an approximate analytic profiling of this system. This method was developed in Ref. [6], where the perturbative results were checked against numerical minimisation.

To write the profiling equations in a compact form, we define

ri(a)=yi−Ti(a),r_i(\mathbf{a})=y_i-T_i(\mathbf{a}),
σi2(a)=δi,stat2Ti2(a)+δi,uncor2Ti2(a),\sigma_i^2(\mathbf{a})=\delta_{i,\mathrm{stat}}^2T_i^2(\mathbf{a})+\delta_{i,\mathrm{uncor}}^2T_i^2(\mathbf{a}),
Γis(a)=−Ti(a)γis.(9)\Gamma_{is}(\mathbf{a})=-T_i(\mathbf{a})\gamma_{is}. \tag*{(9)}

For fixed a\mathbf{a}, the NPs are updated at iteration kk according to

bs(k+1)=∑p(A(k))sp−1Bp.(10)b_s^{(k+1)}=\sum_p\left(A^{(k)}\right)_{sp}^{-1}B_p. \tag*{(10)}

Here, the matrix AA and vector BB are defined as

Asp(k)=δspσ^us2(k)+∑iΓisΓipσi2,A_{sp}^{(k)}=\frac{\delta_{sp}}{\widehat{\sigma}_{us}^{2(k)}}+\sum_i\frac{\Gamma_{is}\Gamma_{ip}}{\sigma_i^2},
Bs=∑iΓisriσi2.(11)B_s=\sum_i\frac{\Gamma_{is}r_i}{\sigma_i^2}. \tag*{(11)}

where the systematic variance is re-evaluated at each iteration from Eq. (2),

σ^us2(k)=1+2ϵs2(bs(k))21+2ϵs2.(12)\widehat{\sigma}_{us}^{2(k)}=\frac{1+2\epsilon_s^2\left(b_s^{(k)}\right)^2}{1+2\epsilon_s^2}. \tag*{(12)}

The procedure is initialised with σ^us2(0)=1\widehat{\sigma}_{us}^{2(0)}=1, so the first iteration reproduces the analytic Gaussian profiling solution for quadratic NP constraints. Subsequent iterations update the systematic variances and thus account for the non-quadratic GVM terms. At convergence, the solution minimises the χ2\chi^2 with respect to the NPs at fixed a\mathbf{a}. Sources with large NP shifts or poor convergence are instead minimised numerically. The derivation, convergence condition and numerical validation for the HERA application are given in App. § 9.

With the profiling treatment in place, we now turn to the consequences of the logarithmic constraints for the goodness-of-fit statistic and confidence intervals. The goodness-of-fit statistic χmin⁡2\chi^2_{\min}, obtained by minimising Eq. (8), is not exactly χ2\chi^2-distributed, because the logarithmic NP constraints depart from the quadratic Gaussian form underlying the usual χ2\chi^2 approximation. Furthermore, confidence intervals based on a quadratic approximation near the minimum may be too narrow. Ignoring these effects slightly underestimates dataset compatibility and yields narrower PDF uncertainties. To correct for these effects, we introduce two correction factors, denoted by Cχ2C_{\chi^2} and CCIC_{\mathrm{CI}}, following the rescaling method introduced by Bartlett [20] and subsequently developed in Refs. [21, 22]. Its application to the GVM was studied in Refs. [5, 6], including validation with Monte Carlo simulations in the models considered.

Cχ2C_{\chi^2} rescales the goodness-of-fit statistic, while CCIC_{\mathrm{CI}} rescales the squared confidence-interval (CI) widths. These factors are defined as

Cχ2=1+bχ2ndof,CCI=1+bCIna.(13)\begin{aligned} C_{\chi^2} &= 1 + \frac{b_{\chi^2}}{n_{\mathrm{dof}}},\\ C_{\mathrm{CI}} &= 1 + \frac{b_{\mathrm{CI}}}{n_a}. \tag*{(13)} \end{aligned}

where ndof=Ndata−nan_{\mathrm{dof}} = N_{\mathrm{data}} - n_a is the number of degrees of freedom, NdataN_{\mathrm{data}} is the number of fitted measurements, and nan_a is the number of fitted PDF parameters (POIs). The raw minimum is divided by Cχ2C_{\chi^2} and each interval half-width is multiplied by CCI\sqrt{C_{\mathrm{CI}}}, without changing the best-fit parameter values:

χB2=χmin⁡2Cχ2,Δam,±B=CCI Δam,±.(14)\begin{aligned} \chi^2_{\mathrm{B}} &= \frac{\chi^2_{\min}}{C_{\chi^2}},\\ \Delta a^{\mathrm{B}}_{m,\pm} &= \sqrt{C_{\mathrm{CI}}}\,\Delta a_{m,\pm}. \tag*{(14)} \end{aligned}

Here Δam,−\Delta a_{m,-} and Δam,+\Delta a_{m,+} denote the lower and upper half-widths of the uncorrected confidence interval for parameter ama_m, respectively. The leading correction terms bχ2b_{\chi^2} and bCIb_{\mathrm{CI}} are obtained from the analytical expressions developed by Lawley [23] and Cordeiro [21] and specialised to the GVM in Refs. [5, 6]. Here we express the Bartlett correction in matrix form for direct use in xFitter; details are given in App. § 10.

Relation to tolerance prescriptions

A common method used in global PDF fits to address experimental tensions and obtain more realistic confidence intervals is to introduce a tolerance. In its simplest form, this prescription enlarges the nominal one-parameter confidence region from Δχ2≤1\Delta\chi^2 \le1 to Δχ2≤T2\Delta\chi^2 \le T^2, using a fixed tolerance T>1T > 1. More refined prescriptions allow TT to depend on the direction in PDF-parameter space [8, 24]. Ideally one would prefer to avoid introducing an ad hoc tolerance, as it complicates the statistical interpretation of the resulting parameter region.

Tolerance prescriptions enlarge the confidence region without modifying the objective function (the χ2\chi^2 minimised in the fit) or attributing the enlargement to individual experimental uncertainty sources. The GVM instead modifies the objective function by allowing for uncertainty in the sizes of systematic errors whose sources are assigned nonzero εs\varepsilon_s. Within the assumed GVM model, improvements in data–theory compatibility can therefore be traced to specific sources, helping to identify uncertainty estimates that warrant experimental reassessment.

The two approaches can nevertheless be complementary. The study by M. Redder [25] of ATLAS WW, ZZ and inclusive-jet data within the MSHT framework suggests that errors-on-errors can account for part of an effective tolerance, with T2∼1.5T^2 \sim1.5–2 in the examples considered. The tolerance required in PDF fits may also reflect other effects, such as uncertainties in the assumed bin-to-bin correlations of individual systematic sources. In a combined approach, the tolerance would be applied when constructing confidence intervals from the GVM objective function. Since this function already accounts for the assigned errors-on-errors, the tolerance would address the remaining tensions. One would therefore expect a smaller tolerance than in an otherwise equivalent fit with Gaussian constraints.

HERAPDF2.0, NNLO, RTOPT fit setup

The fits are performed using the combined HERA I+II inclusive reduced cross sections measured by the H1 and ZEUS experiments [1]. The combined dataset comprises seven samples: four neutral-current (NC) e+pe^{+}p samples at proton-beam energies of 920, 820, 575 and 460 GeV; one NC e−pe^{-}p sample; and two charged-current (CC) samples for e+pe^{+}p and e−pe^{-}p scattering. The latter three are quoted at a reference proton-beam energy of 920 GeV. The DIS kinematics are described by Q2Q^{2}, the negative squared four-momentum transfer, and Bjorken xx, the dimensionless scaling variable. Following the HERAPDF2.0 analysis, we retain measurements with Q2≥3.5 GeV2Q^{2} \ge3.5\ \mathrm{GeV}^{2}. This leaves 1145 measurements, covering xx values from about 4×10−54 \times10^{-5} to 0.65 and extending up to Q2=5×104 GeV2Q^{2} = 5 \times10^{4}\ \mathrm{GeV}^{2}. In broad terms, the low-Q2Q^{2} NC data constrain the low-xx sea and, through their scale dependence, the gluon, whereas the difference between the high-Q2Q^{2} NC e+pe^{+}p and e−pe^{-}p cross sections, together with the CC data, constrains the valence distributions. The complementary flavour sensitivities of the NC and CC measurements help separate the up- and down-type quark combinations.

The predictions are calculated at NNLO in QCD in the MS scheme. The PDFs are evolved with QCDNUM [26], using renormalisation and factorisation scales μR=μF=Q\mu_{R} = \mu_{F} = Q. Following the HERAPDF2.0 setup, the heavy-quark coefficient functions entering the NC structure functions are evaluated in the optimised Thorne–Roberts (RTOPT) general-mass variable-flavour-number scheme, whereas the zero-mass approximation is used for the CC structure functions, since all fitted HERA CC data satisfy Q2≫Mb2Q^{2} \gg M_{b}^{2} [27, 28, 29]. The evolution uses a variable-flavour-number scheme with nf=5n_{f} = 5, heavy-quark mass parameters Mc=1.43 GeVM_{c} = 1.43\ \mathrm{GeV} and Mb=4.50 GeVM_{b} = 4.50\ \mathrm{GeV}, and fixed αs(MZ)=0.118\alpha_{s}(M_{Z}) = 0.118. Electroweak effects are included at leading order, using the boson masses MZ=91.1876 GeVM_{Z} = 91.1876\ \mathrm{GeV} and MW=80.385 GeVM_{W} = 80.385\ \mathrm{GeV}, the weak mixing angle θW\theta_{W} specified by sin⁡2θW=0.23127\sin^{2}\theta_{W} = 0.23127, and the Fermi constant GF=1.16638×10−5 GeV−2G_{F} = 1.16638 \times10^{-5}\ \mathrm{GeV}^{-2}.

At the starting scale Q02=1.9 GeV2Q_{0}^{2} = 1.9\ \mathrm{GeV}^{2}, below the charm threshold, five distributions are parametrised: gg, uvu_{v}, dvd_{v}, U‾=u‾\overline{U} = \overline{u} and D‾=d‾+s‾\overline{D} = \overline{d} + \overline{s}. Here gg is the gluon distribution, uv=u−u‾u_{v} = u - \overline{u} and dv=d−d‾d_{v} = d - \overline{d} are the up- and down-valence distributions, and U‾\overline{U} and D‾\overline{D} are the up- and down-type sea-antiquark combinations, respectively. Apart from the additional gluon term specified below,

their generic form is

xf(x,Q02)=AfxBf(1−x)CfPf(x),(15)x f(x,Q_{0}^{2}) = A_{f}x^{B_{f}}(1-x)^{C_{f}}P_{f}(x), \tag*{(15)}

where AfA_{f} is an overall normalisation coefficient and BfB_{f} and CfC_{f} control the low- and high-xx behaviour. The factor Pf(x)P_{f}(x) is a polynomial that allows additional shape flexibility where required. For uvu_{v} and U‾\overline{U}, it is chosen as Puv(x)=1+Euvx2P_{u_{v}}(x) = 1 + E_{u_{v}}x^{2} and PU‾(x)=1+DU‾xP_{\overline{U}}(x) = 1 + D_{\overline{U}}x, respectively, while Pf(x)=1P_{f}(x) = 1 for dvd_{v}, D‾\overline{D} and the leading gluon term. The full gluon parametrisation contains an additional negative term, −Ag′xBg′(1−x)Cg′-A'_{g}x^{B'_{g}}(1-x)^{C'_{g}}, with a separate fitted low-xx exponent Bg′B'_{g} and Cg′=25C'_{g} = 25. The coefficients AgA_{g}, AuvA_{u_{v}} and AdvA_{d_{v}} are fixed by the momentum and quark-number sum rules. At Q02Q_{0}^{2}, the remaining constraints are BU‾=BD‾B_{\overline{U}} = B_{\overline{D}}, s‾=s=fsD‾\overline{s} = s = f_{s}\overline{D} with the strange suppression factor fs=0.4f_{s} = 0.4, and AU‾=(1−fs)AD‾A_{\overline{U}} = (1-f_{s})A_{\overline{D}}. Together, these give the standard 14-parameter HERAPDF2.0 fit and hence ndof=1145−14=1131n_{\mathrm{dof}} = 1145 - 14 = 1131.

For each combined measurement, the HERA combination [1] provides a statistical uncertainty, one aggregate uncorrelated systematic uncertainty and 169 correlated systematic errors. The uncorrelated systematic components were added in quadrature before averaging, so their source-by-source breakdown is not available. The correlated sources, by contrast, are resolved individually. Of the 169, 162 originate from the H1 and ZEUS measurements, including normalisations; they were propagated through the joint fit of the combined cross sections and published in an orthogonal basis, and are denoted sysHZComb. The remaining seven are procedural, introduced by the combination itself: they cover the choice between additive and multiplicative errors, possible correlations of the photoproduction background and hadronic energy scale across datasets, and four modes associated with groups of sources having unusually large pulls. Setting ϵs=0\epsilon_{s} = 0 for all 169 correlated sources, this setup gives χmin⁡2=1363\chi_{\min}^{2} = 1363 for 1131 degrees of freedom, in agreement with the published HERAPDF2.0 NNLO result, χmin⁡2/ndof=1363/1131\chi_{\min}^{2}/n_{\mathrm{dof}} = 1363/1131. To distinguish this setup from those introduced later, we label HERAPDF2.0, NNLO, RTOPT and refer to the fit with Gaussian constraints as the corresponding Gaussian baseline.

Robustness study on HERA data

The first application we present uses the errors-on-errors framework as a robustness test of the HERAPDF2.0, NNLO, RTOPT fit. Rather than assigning the uncertainty on each systematic source, we ask what the largest realistic effect of errors-on-errors on the analysis is. In particular, we assess how much of the observed tension they could account for, how they would affect the fitted PDFs and cross-section predictions, and in which kinematic regions these effects would be concentrated. The values of ϵs\epsilon_s are therefore chosen to probe the largest plausible effect, rather than being assigned a priori from source-specific knowledge. This type of study is especially useful when such information is unavailable, as in retrospective analyses where the original expert knowledge is no longer accessible. It can also assess the potential impact of errors-on-errors and identify sources that merit experimental follow-up, without committing to precise ϵs\epsilon_s values for individual sources. An a priori assignment of ϵs\epsilon_s is considered in § 6.

Forward selection

To identify the correlated systematic sources to which the fit is most sensitive, we build the set of selected sources iteratively. Starting from the HERAPDF2.0, NNLO, RTOPT baseline, in which all ϵs=0\epsilon_s = 0, we first test each of the 169 correlated sources in turn by assigning that source ϵs=0.6\epsilon_s = 0.6 and refitting the PDF parameters and NPs. The source producing the largest reduction in the Bartlett-corrected minimum χB2\chi^2_{\mathrm{B}}, defined in Eq. (14), is retained. We use χB2\chi^2_{\mathrm{B}} because it provides a more accurate measure of data–theory compatibility when errors-on-errors are nonzero. At the next step, this source remains assigned ϵs=0.6\epsilon_s = 0.6, and each of the remaining sources is tested one at a time in addition to it. The source giving the largest further reduction is again retained. The same procedure is then repeated: after each step, all previously selected sources remain active, while each remaining source is tested in turn on top of the updated fit. This iterative procedure is referred to as forward selection. At each step, the improvement from adding a candidate source ss is quantified by

δχB2(s)=χB2(current fit)−χB2(current fit+s).(16)\delta\chi^2_{\mathrm{B}}(s) = \chi^2_{\mathrm{B}}(\text{current fit}) - \chi^2_{\mathrm{B}}(\text{current fit} + s). \tag*{(16)}

The iteration stops when the best remaining source gives δχB2≤1\delta\chi^2_{\mathrm{B}} \le1. A simpler scan in which each source is tested only against the initial baseline can provide a useful first picture at lower computational cost, but its reductions cannot be combined directly because adding one source changes the fit minimum and therefore the impact of the others. We set ϵs=0.6\epsilon_s = 0.6 to probe a large but plausible effect. As a reference, ϵ=1/2≃0.7\epsilon= 1/\sqrt{2} \simeq0.7 corresponds to the value one would assign to a “two-point systematic”, for which the uncertainty size is estimated from the spread of two independent determinations.4 Our reference value lies slightly below this benchmark; however, the choice of ϵs\epsilon_s is indicative: one could also consider values such as 0.5 or 0.7, or explore the dependence through a scan, as done below. The emphasis is on understanding the impact for large but plausible values of ϵs\epsilon_s, rather than on fine-tuning a parameter whose value can itself only be estimated approximately.

The selection reported in Table 1 identifies seven correlated systematic sources and lowers χB2\chi^2_{\mathrm{B}} by 57 units. The leading source, sysHZComb1053, accounts for 37 units, while the remaining six contribute about 21 units. In an ongoing analysis or an active experiment, the large effect associated with the leading source would motivate experimental follow-up to reassess its uncertainty estimate and its impact on the measurement.

RankSystematicχB2\chi^{2}_{B}δχB2\delta\chi^{2}_{B}
1sysHZComb1053132736.78
2proc_tb2113197.36
3sysHZComb106413153.80
4sysHZComb111413122.91
5sysHZComb114413102.53
6sysHZComb111113082.19
7sysHZComb111313061.88
8 (stop)proc_nrl13050.81

Table 1. Forward selection for HERAPDF2.0, NNLO, RTOPT, with ϵs=0.6\epsilon_s = 0.6. Rows 1–7 list the selected sources in order. χB2\chi^{2}_{B} is the corrected minimum after each addition, and δχB2\delta\chi^{2}_{B} its reduction from the previous step.

A preliminary indication of sources worth testing for errors-on-errors can be obtained from sizeable NP shifts in the Gaussian baseline fit, expressed in units of the nominal systematic uncertainty. The magnitude of a shift alone, however, does not determine the effect, since the improvement also depends on the data constraints and correlations with the other fitted parameters. Further details on NP shifts and their uncertainties are given in App. § 11.

In the final seven-source fit, the raw χ2\chi^2 minimum falls from 1363 to 1313, and the Bartlett correction gives χB2=1306\chi^{2}_{B} = 1306. With 1131 degrees of freedom, the corrected reduced χ2\chi^2 is 1306/1131, compared with 1363/1131 for the Gaussian baseline. The compatibility is substantially improved, although the reduced χ2\chi^2 remains above unity (the expectation for an ideal description).

To identify the kinematic regions that contribute most to the improvement in fit quality, we examine the change in the data contribution χdata2\chi^2_{\mathrm{data}} between the baseline and errors-on-errors fits. Here χdata2\chi^2_{\mathrm{data}} denotes the data term defined by the first sum in Eq. (7), evaluated at the respective best fits before the Bartlett correction, with the logarithmic normalisation and NP constraint terms excluded. The largest contribution to its 58-unit decrease comes from the NC e+pe^{+}p sample at 920 GeV (377 fitted points), whose contribution falls from 443 to 418. The reductions for the 820, 575 and 460 GeV samples are 7, 8 and 12, respectively; the NC e−pe^{-}p and CC samples also improve, though more moderately.

The distribution of this improvement in Q2Q^2 and xx is shown in Table 2, which gives Δχdata2=χdata2(selected)−χdata2(baseline)\Delta\chi^2_{\mathrm{data}} = \chi^2_{\mathrm{data}}(\mathrm{selected}) - \chi^2_{\mathrm{data}}(\mathrm{baseline}). Here, selected denotes the final seven-source configuration. Each populated cell gives the change followed, in parentheses, by the number of measurements in the bin; negative values indicate a better description. The largest contribution comes from the low-Q2Q^2, low-xx region. The bins with Q2<10 GeV2Q^2 < 10\ \mathrm{GeV}^2 account for −25-25, of which −21-21 arises at x<10−3x < 10^{-3}. As

Q2Q^{2} [GeV2^{2}]x<10−3x < 10^{-3}10−3≤x<10−210^{-3} \le x < 10^{-2}10−2≤x<10−110^{-2} \le x < 10^{-1}x≥10−1x \ge10^{-1}total
< 10−21.0-21.0 (108)−2.5-2.5 (17)−1.6-1.6 (4)–−25.1-25.1
10≤Q2<3010 \le Q^{2} < 30−0.8-0.8 (109)−6.9-6.9 (56)+1.4+1.4 (9)–−6.2-6.2
30≤Q2<10030 \le Q^{2} < 100−7.0-7.0 (27)−5.2-5.2 (128)−2.2-2.2 (25)+0.0+0.0 (2)−14.3-14.3
100≤Q2<1000100 \le Q^{2} < 1000–+1.5+1.5 (128)−2.6-2.6 (195)−5.3-5.3 (118)−6.4-6.4
≥1000\ge1000––−5.7-5.7 (79)+0.0+0.0 (140)−5.7-5.7
total−28.7-28.7−13.1-13.1−10.7-10.7−5.2-5.2−57.7-57.7

Table 2. Change Δχdata2\Delta\chi_{\mathrm{data}}^{2} in the pointwise data contribution between the final seven-source errors-on-errors configuration and the HERAPDF2.0, NNLO, RTOPT baseline fit, binned in Q2Q^{2} and xx. Each populated cell gives this change, followed in parentheses by the number of fitted measurements in the bin.

a complementary check of the low-Q2Q^2 sensitivity seen in Table 2, we varied the minimum Q2Q^2 in the HERAPDF2.0, NNLO, RTOPT configuration. Raising the cut from 3.5 to 10 GeV2\mathrm{GeV}^2 reduces the number of selected systematic sources, as it excludes many data points sensitive to errors-on-errors. Conversely, lowering it from 3.5 to 2 GeV2\mathrm{GeV}^2 leaves the relative reduction in the overall χ2\chi^2 largely unchanged and does not alter the conclusions of the robustness study.

Finally, we assess how the reduction in χB2\chi^{2}_{B} depends on the assumed size of errors-on-errors. We keep the seven sources selected in Table 1 fixed and vary their common ϵ\epsilon from 0 to 0.7, refitting the PDF parameters and NPs at each point, while all remaining sources retain ϵs=0\epsilon_s = 0. Figure 1 shows the corrected minimum χB2\chi^{2}_{B} as a function of ϵ\epsilon, recovering the Gaussian baseline at ϵ=0\epsilon= 0. The reduction is appreciable already at ϵ≳0.1\epsilon\gtrsim0.1 and slows beyond ϵ≈0.3\epsilon\approx0.3. However, χB2/ndof\chi^{2}_{B}/n_{\mathrm{dof}} remains above unity throughout the range, indicating that errors-on-errors alone cannot account for the observed HERA tension. Furthermore, repeating the forward selection at different ϵ\epsilon values would have little impact on the conclusions. The leading sources remain unchanged, while the number selected increases with ϵ\epsilon and the ordering of lower-ranked sources may vary slightly.

Figure 1. Bartlett-corrected minimum χB2\chi^{2}_{B} versus the common errors-on-errors parameter ϵ\epsilon for the seven sources selected in Table 1, using the HERAPDF2.0, NNLO, RTOPT setup.

Impact on PDFs and cross sections

We now turn to the impact of these systematic adjustments on the fitted PDFs and cross-section predictions. Figure 2 compares the final seven-source errors-on-errors fit with the HERAPDF2.0, NNLO, RTOPT baseline for xuvx u_{\mathrm{v}}, xdvx d_{\mathrm{v}} and xgx g at Q2=4 GeV2Q^{2} = 4\ \mathrm{GeV}^{2}, after evolution from the input scale. The central values shift mainly at low xx but generally remain within the baseline uncertainty bands; only the uu-valence PDF lies marginally outside in some regions. The confidence intervals broaden by up to about 25% for uvu_{\mathrm{v}} and 15% for dvd_{\mathrm{v}} at low xx, and by about 10–13% for the gluon over most of the high-xx region. Despite the substantial reduction in χB2/ndof\chi_{\mathrm{B}}^{2}/n_{\mathrm{dof}}, the PDFs remain broadly compatible with the baseline. The improvement in fit quality is accompanied by a broadening of the PDF confidence intervals, while the central values remain stable and show no large displacement.

Absolute PDFs and ratios for the HERAPDF2.0, NNLO, RTOPT Gaussian baseline and the forward-selection errors-on-errors fit

Figure 2. Absolute PDFs and ratios for the HERAPDF2.0, NNLO, RTOPT Gaussian baseline and the forward-selection errors-on-errors (EoE) fit at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2. From left to right, the three columns show xuvx u_v, xdvx d_v, and xgx g, with the absolute distributions in the upper row and their ratios to the baseline central PDF in the lower row.

The corresponding cross sections show the same stability. Across all 1145 fitted measurements, the median point-by-point absolute change in the unshifted theory prediction, i.e. the cross section obtained from the fitted PDFs without accounting for NP shifts, is only 0.20%. This stability is also present in the region giving the largest reduction in the data term in Table 2, Q2<10 GeV2Q^{2} < 10\ \mathrm{GeV}^{2} and x<10−3x < 10^{-3}. Among its 108 measurements, the median change is 0.19%. These observations suggest that errors-on-errors improve the fit through the cumulative effect of small adjustments across many measurements, rather than large shifts in a few cross sections. Figure 3 illustrates this stability in the two lowest-Q2Q^{2} bins of the NC e+pe^{+}p sample, at Q2=3.5Q^{2} = 3.5 and 4.5 GeV24.5\ \mathrm{GeV}^{2}. At both scales, the two fits remain close in their central predictions and in the widths of their confidence intervals. The confidence-interval broadening is more pronounced in the CC e−pe^{-}p sample, which is more directly sensitive to the uu-type valence distributions: across that sample, the median increase in width is about 9%.

Figure 3. Representative low-Q2Q^2 reduced NC e+pe^+p cross sections at a proton-beam energy of 920 GeV, at Q2=3.5 GeV2Q^2=3.5\ \mathrm{GeV}^2 (left) and Q2=4.5 GeV2Q^2=4.5\ \mathrm{GeV}^2 (right). The HERAPDF2.0, NNLO, RTOPT Gaussian baseline and forward-selection EoE fit are compared. The lower panels show the predictions divided by the data; the dashed curves include the fitted NP shifts, and the hatched bands are the 68.3% Pumplin confidence intervals.

Taken together, the errors-on-errors robustness test shows that the tension in the HERAPDF2.0, NNLO, RTOPT baseline is substantially reduced once the systematic-uncertainty model is given additional flexibility. The GVM allows the fit to accommodate small coherent fluctuations along the response patterns of the correlated sources, as expected when some systematic uncertainties are underestimated. This markedly improves data–theory compatibility without requiring comparably large shifts in the fitted PDFs or cross-section predictions. The stability of the central values reflects the nature of the tension in this fit: the discrepancies are spread over many measurements, with no clear outlier pulling the PDFs in a particular direction. The situation may be different in global analyses combining different experiments and processes, for instance when LHC data are added, where tensions between datasets or an outlying dataset in a given kinematic region could lead errors-on-errors to shift the central values as well, in the direction preferred by the mutually consistent and more precise data, down-weighting outlying or less precise datasets.

The conclusion of this section is limited to the 169 correlated sources, which account for about one quarter of the total uncertainty budget. The uncorrelated systematic components were aggregated before averaging and are not available source by source. The published data therefore do not allow us to assess the additional effect of applying errors-on-errors to all systematic uncertainties before the H1 and ZEUS measurements are combined.

Interplay with the small-xx description

The largest errors-on-errors effects occur in the same low-xx, low-Q2Q^2 region in which the fit is particularly sensitive to the assumed PDF shape and perturbative description. We therefore consider separately the additional flexibility of the input parametrisation and the perturbative small-xx description, tested with matched fixed-order and NNLO+NLLx_x FONLL-C fits. This allows us to test whether these changes replace, overlap with, or remain complementary to the errors-on-errors treatment.

We first examine whether the errors-on-errors effect depends on the assumed PDF functional form by considering the parametrisation proposed by Bonvini and Giuli [2], which we denote HERAPDF-BG. At the same starting scale Q02=1.9 GeV2Q_0^2 = 1.9\ \mathrm{GeV}^2, its generic form is

xf(x,Q02)=AfxBf(1−x)Cf×[1+Dfx+Efx2+Ffln⁡x+Gf(ln⁡x)2].(17)x f(x,Q_0^2) = A_f x^{B_f}(1-x)^{C_f} \times\left[1 + D_f x + E_f x^2 + F_f \ln x + G_f(\ln x)^2\right]. \tag*{(17)}

We choose these coefficients following Ref. [2]. For the gluon, the terms proportional to ln⁡x\ln x and (ln⁡x)2(\ln x)^2, with coefficients FgF_g and GgG_g, replace the separate negative term used in HERAPDF2.0. Both logarithmic terms are also included for uvu_v, with coefficients FuvF_{u_v} and GuvG_{u_v}. For the sea distributions, we retain only the linear logarithmic terms, with a shared coefficient FU‾=FD‾F_{\overline{U}} = F_{\overline{D}}, and include an additional term DD‾xD_{\overline{D}}x for D‾\overline{D}. All remaining coefficient choices and constraints follow the prescription of § 3. The advantage of this parametrisation lies in the additional flexibility provided by the logarithmic terms, which allow a wider range of PDF shapes at small xx while vanishing as x→1x \to1. The resulting parametrisation has 18 fitted parameters, four more than the HERAPDF2.0 form.

A distinct possibility is that part of the low-xx behaviour reflects limitations of the fixed-order perturbative description itself. At small xx, contributions enhanced by powers of log⁡(1/x)\log(1/x) can become increasingly important and reduce the reliability of the fixed-order expansion [30]. The underlying leading-logarithmic small-xx (LLx_x) results were developed in Refs. [31, 32, 33, 34], and the NLLx_x results in Refs. [35, 36, 37, 38]. We follow Ref. [2] and resum these contributions using the default NLL variant of the High-Energy Large Logarithms (HELL) code, version 3.0 [39, 40, 41, 42], including their effects in the Dokshitzer–Gribov–Lipatov–Altarelli–Parisi (DGLAP) splitting functions, the heavy-quark matching conditions and the DIS coefficient functions. The calculation is matched to NNLO and interfaced to xFitter through A PDF Evolution Library (APFEL) [3, 43]. Since this implementation uses the Fixed Order plus Next-to-Leading Logarithms (FONLL) general-mass variable-flavour-number scheme [44], we adopt its FONLL-C variant, appropriate at NNLO, and use an otherwise identical fixed-order FONLL-C fit to isolate the effect of resummation.

With the HERAPDF-BG parametrisation and the original RTOPT setup of § 3, we obtain χmin⁡2/ndof=1301/1127\chi^2_{\min}/n_{\mathrm{dof}}=1301/1127 under standard Gaussian constraints, about 63 units below the HERAPDF2.0, NNLO, RTOPT baseline. This reproduces the result of Ref. [2], where the clearest improvement occurs at low xx and low Q2Q^2. We refer to this setup as HERAPDF-BG, NNLO, RTOPT. For the matched FONLL-C comparison, both fits retain the same parametrisation, data and cuts, and αs(MZ)=0.118\alpha_s(M_Z)=0.118, but use Q0=1.6 GeVQ_0=1.6\ \mathrm{GeV}, mc=1.46 GeVm_c=1.46\ \mathrm{GeV} and μc=1.12mc\mu_c=1.12m_c, as in that reference. We refer to this fixed-order reference as HERAPDF-BG, NNLO, FONLL-C. With these settings, it gives χmin⁡2=1312\chi^2_{\min}=1312. Including NLLx_x resummation lowers this value to 1283. We refer to this resummed setup as HERAPDF-BG, NNLO+NLLx_x, FONLL-C. The HERAPDF-BG, NNLO, RTOPT and HERAPDF-BG, NNLO+NLLx_x, FONLL-C Gaussian baselines are used for the errors-on-errors study in the next subsection.

Errors-on-errors analysis

We repeat the forward selection of § 4 for the HERAPDF-BG, NNLO, RTOPT and HERAPDF-BG, NNLO+NLLx_x, FONLL-C baselines, using the same ϵs=0.6\epsilon_s = 0.6 assignment and retaining sources with δχB2>1\delta\chi^2_{\mathrm{B}} > 1. Table 3 and Table 4 give the resulting sequences. Eight sources are retained with the HERAPDF-BG, NNLO, RTOPT baseline. All seven sources selected in the HERAPDF2.0, NNLO, RTOPT fit are selected again, although in a different order, and the selection additionally retains sysHZComb1062. In the HERAPDF-BG, NNLO+NLLx_x, FONLL-C fit, the five retained sources are precisely the first five selected with the HERAPDF-BG, NNLO, RTOPT baseline, although their marginal effects are changed by the different theoretical description. This overlap nevertheless suggests that the fits remain sensitive to many of the same systematic sources when the low-xx parametrisation or theoretical description is changed. At the stopping point, errors-on-errors reduce χB2/ndof\chi^2_{\mathrm{B}}/n_{\mathrm{dof}} from 1301/1127 to 1254/1127 for the HERAPDF-BG, NNLO, RTOPT fit, and from 1283/1127 to 1247/1127 for the HERAPDF-BG, NNLO+NLLx_x, FONLL-C fit.

RankSystematicχB2\chi^2_{\mathrm{B}}δχB2\delta\chi^2_{\mathrm{B}}
1sysHZComb1053127030.69
2sysHZComb106212663.67
3sysHZComb111112633.02
4sysHZComb111412612.61
5sysHZComb114412582.62
6proc_tb2112571.18
7sysHZComb106412551.69
8sysHZComb111312541.41
9 (stop)proc_nrl12530.78

Table 3. Forward selection from the HERAPDF-BG, NNLO, RTOPT Gaussian baseline, with ϵs=0.6\epsilon_s = 0.6. Entries give Bartlett-corrected minima and marginal reductions.

RankSystematicχB2\chi^2_{\mathrm{B}}δχB2\delta\chi^2_{\mathrm{B}}
1sysHZComb1053126221.74
2sysHZComb106212557.02
3sysHZComb111112522.85
4sysHZComb111412502.29
5sysHZComb114412472.68
6 (stop)proc_tb2412460.99

Table 4. Forward selection from the HERAPDF-BG, NNLO+NLLx_x, FONLL-C Gaussian baseline, with ϵs=0.6\epsilon_s = 0.6. Entries give Bartlett-corrected minima and marginal reductions.

Figure 4 and Figure 5 compare the Gaussian baselines with the corresponding final errors-on-errors fits for xuvx u_v, xdvx d_v and xgx g at Q2=4Q^2 = 4 GeV2^2. In the first configuration, the central-value shifts are concentrated at the smallest xx and are most visible in uvu_v and the gluon, while dvd_v remains comparatively stable. The characteristic shapes generated by the additional parametrisation freedom are preserved. The valence confidence intervals broaden in parts of the xx range, most visibly for uvu_v at small xx, while the bands of the two fits overlap throughout the displayed range. In the second configuration, the central-value shifts are modest and remain within the corresponding Gaussian confidence bands throughout the displayed range. The valence confidence intervals broaden moderately in parts of the xx range, while the low-xx gluon uncertainty is nearly unchanged. In both configurations, the improvement in fit quality is therefore accompanied by local central-value adjustments and a selective expansion of the PDF confidence intervals.

PDFs and ratios for HERAPDF-BG, NNLO, RTOPT

Figure 4. PDFs and ratios for HERAPDF-BG, NNLO, RTOPT at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2, comparing the Gaussian baseline and forward-selection EoE fit. Columns show xuvx u_v, xdvx d_v and xgx g; ratios use the baseline central PDF.

Absolute PDFs and ratios for the HERAPDF-BG, NNLO+NLL$_x$, FONLL-C Gaussian baseline and forward-selection EoE fit

Figure 5. Absolute PDFs and ratios for the HERAPDF-BG, NNLO+NLLx_x, FONLL-C Gaussian baseline and forward-selection EoE fit at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2. From left to right, the columns show xuvx u_v, xdvx d_v and xgx g, with the absolute distributions in the upper row and their ratios to the corresponding Gaussian baseline central PDF in the lower row.

For both configurations, the effect of errors-on-errors on the cross sections is similar to that found in the HERAPDF2.0, NNLO, RTOPT study. Relative to the corresponding Gaussian fit, the central predictions remain broadly stable, with moderate changes in their confidence intervals. Table 5 summarises the forward-selection results for all three fit configurations. The additional parametrisation freedom and the change in theoretical treatment mainly affect the fitted low-xx PDF shapes, whereas errors-on-errors provide additional freedom predominantly through correlated NP shifts. The ordering of the fit minima is preserved after introducing errors-on-errors. The substantial errors-on-errors improvements that remain in all three configurations indicate that their effect is not absorbed by the additional PDF flexibility or the change in theoretical description.

Fit setupSelected sourcesGaussian χmin⁡2\chi^2_{\min}Final χB2\chi^2_{\mathrm{B}}
HERAPDF2.0, NNLO, RTOPT713631306
HERAPDF-BG, NNLO, RTOPT813011254
HERAPDF-BG, NNLO+NLLx_x, FONLL-C512831247

Table 5. Summary of the forward-selection robustness tests. Each final fit assigns ϵs=0.6\epsilon_s = 0.6 to the selected sources and zero to all remaining sources.

Illustrative informed application

The tests described above help assess the robustness of the PDF fits and identify systematic sources that merit closer investigation. For a PDF determination intended for physics analyses, each ϵs\epsilon_s should be assigned a priori from information on how the corresponding systematic uncertainty was constructed. Such assignments are necessarily approximate: ϵs\epsilon_s is more naturally regarded as a scale reflecting the reliability of the quoted uncertainty than as a precisely known parameter. In practice, a coarse classification with representative values such as ϵs=0\epsilon_s = 0, 0.30.3 and 0.60.6 can distinguish between systematic uncertainties whose quoted sizes are regarded as well determined, moderately uncertain or poorly determined, respectively. Although approximate, such an assignment can represent an improvement relative to the Gaussian treatment (ϵs=0\epsilon_s = 0 for all sources), provided the categories reasonably reflect the reliability of the quoted uncertainties. The Gaussian treatment is itself an approximation, imposing the more restrictive and often unrealistic assumption that all systematic uncertainty sizes are known exactly. The task of assigning errors-on-errors is also simplified when most of the effect is concentrated in a few sources. Reasonable estimates for these dominant sources may then capture the main effect, without requiring equally detailed information for every systematic uncertainty. Where the estimates remain uncertain, their impact can be made explicit by presenting the relevant results as functions of the assumed ϵs\epsilon_s, as illustrated for the fit minimum in Figure 1.

For the HERA application, we adopt the following coarse classification. We assign ϵs=0.6\epsilon_s = 0.6 to the seven procedural uncertainties introduced in the combination, which quantify its dependence on the treatment of systematic uncertainties as additive or multiplicative, assumptions on correlations between data sets, and four variations derived from groups of sources with large pulls; see Sec. 4.5 of Ref. [1]. Their estimation from discrete alternative procedures limits the precision of their assigned sizes, motivating a value close to the two-point-systematic benchmark ϵs=1/2≃0.7\epsilon_s = 1/\sqrt{2} \simeq0.7.

Then, we assign ϵs=0.3\epsilon_s = 0.3 to the 40 sysHZComb sources identified by the HERA combination pull study, discussed in Sec. 4.5.3 and Figure 3 of Ref. [1] and explicitly listed in Sec. 3.4.3 (p. 57; see also Figs. 3.5–3.6) of Ref. [45]. Their normalised combination pulls, defined in Eq. (3.12) of that reference, exceed 1.3 in absolute value. As noted in Ref. [1], such large pulls might indicate that these uncertainties were underestimated or do not fulfil the assumptions of the linear averaging procedure. Their quoted sizes are therefore less reliable than those of the remaining sources, motivating the intermediate value ϵs=0.3\epsilon_s = 0.3. These pulls are evaluated using all inclusive NC and CC measurements simultaneously over the full published Q2Q^2 and xx coverage, before imposing the Q2Q^2 cut of our PDF fits. We therefore use the sources identified in the published HERA combination study, without adjusting the selection to our PDF-fit results. The selected sources span very-low-Q2Q^2 HERA I data, reduced-energy HERA II data, medium- and high-Q2Q^2 data, and normalisation effects. The remaining 122 sources retain ϵs=0\epsilon_s = 0.5

We apply the ϵs\epsilon_s assignment described above to the three fit setups in Table 6, refitting all PDF parameters and NPs in each case. The informed configurations give χB2/ndof\chi^2_{\mathrm{B}}/n_{\mathrm{dof}} equal to 1328/1131, 1276/1127 and 1267/1127, compared with 1363/1131, 1301/1127 and 1283/1127 for the respective Gaussian baselines. The informed treatment therefore has a significant impact on data–theory compatibility in all three setups, as expected: several of the procedural and pull-study sources, in particular sysHZComb1053, were also identified by the forward selections of § 4 and § 5.1. Figure 6 compares the PDF sets obtained with the informed errors-on-errors assignment at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2. At small xx, the two fits with the HERAPDF-BG parametrisation yield uvu_v distributions whose confidence bands remain clearly separated from the HERAPDF2.0, NNLO, RTOPT band. In the same region, the three gluon distributions have non-overlapping confidence bands.

Fit setupndofn_{\mathrm{dof}}Gaussian χmin⁡2\chi^2_{\min}Informed χB2\chi^2_{\mathrm{B}}
HERAPDF2.0, NNLO, RTOPT113113631328
HERAPDF-BG, NNLO, RTOPT112713011276
HERAPDF-BG, NNLO+NLLx_x, FONLL-C112712831267

Table 6. Goodness of fit for the three Gaussian baselines and the corresponding informed errors-on-errors configurations.

Comparison of the informed errors-on-errors PDF sets

Figure 6. Comparison of the informed errors-on-errors PDF sets at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2 for HERAPDF2.0, NNLO, RTOPT; HERAPDF-BG, NNLO, RTOPT; and HERAPDF-BG, NNLO+NLLx_x, FONLL-C. All three fits use the same ϵs\epsilon_s assignment described in the text. From left to right, the columns show xuvx u_v, xdvx d_v and xgx g, with absolute distributions and their confidence bands in the upper row and ratios to the informed HERAPDF2.0, NNLO, RTOPT central PDFs in the lower row.

This exercise illustrates how experimental information can guide the assignment of errors-on-errors in a PDF fit. Here, the chosen values of ϵs\epsilon_s are indicative but entirely realistic; future analyses of new datasets could assign ϵs\epsilon_s more precisely

using detailed information on how each systematic uncertainty was estimated. The configuration needed to reproduce the informed HERAPDF2.0, NNLO, RTOPT fit is made available in the errors-on-errors folder of the xFITTER examples directory.

Conclusions

PDF fits often yield values of χ2/ndof\chi^2/n_{\mathrm{dof}} significantly larger than unity, indicating residual tensions that may arise not only from the theory or the PDF parametrisation, but also from the experimental uncertainty model. In this work, we have investigated, using the combined HERA inclusive dataset, to what extent this discrepancy can be attributed to uncertainty in the assigned sizes of systematic errors. We have done so by modelling uncertainties in the assignment of systematic errors, informally referred to as errors-on-errors, with the Gamma Variance Model implemented in xFITTER.

To quantify the largest errors-on-errors effect accessible within the combined HERA dataset, we first performed a robustness study of the fit. The study uses forward selection of correlated systematic sources at a common stress value ϵs=0.6\epsilon_s = 0.6 to probe the sensitivity of the fit to errors-on-errors. By identifying the systematic sources whose errors-on-errors most strongly affect the fit, the study can also help prioritise experimental follow-up to reassess their uncertainty estimates. The forward selection identifies seven systematic sources with appreciable errors-on-errors effects in the HERAPDF2.0, NNLO, RTOPT baseline, which follows [1], reducing χB2/ndof\chi^2_{\mathrm{B}}/n_{\mathrm{dof}} from 1363/1131 to 1306/1131, with the improvement driven particularly by the data at low Q2Q^2 and low xx. Despite this substantial improvement, the PDF central values remain broadly stable, while the confidence intervals expand moderately. The Gamma Variance Model thus improves the inferred data–theory compatibility by allowing greater freedom in the coherent shifts associated with correlated systematic sources, without requiring large changes in the fitted PDFs or individual cross-section predictions. This behaviour is consistent with some of the quoted systematic uncertainties being underestimated.

Since both the more flexible HERAPDF-BG parametrisation and small-xx resummation improve the description of the same low-xx, low-Q2Q^2 region where the errors-on-errors effect is largest, it is natural to ask how much of this effect remains once these improvements are included. Repeating the study with the HERAPDF-BG, NNLO, RTOPT and HERAPDF-BG, NNLO+NLLx_x, FONLL-C configurations still yields substantial reductions in χB2\chi_B^2. The effect is therefore not absorbed by either the additional PDF-shape flexibility or the improved perturbative description.

While the forward selection estimates the largest errors-on-errors effect accessible in the present setup, a PDF determination including errors-on-errors should assign the ϵs\epsilon_s values a priori, using information on how the various sources of uncertainty were quantified. As an illustration, we use the published HERA studies to classify the systematic sources: the seven procedural uncertainties are assigned ϵs=0.6\epsilon_s = 0.6, the 40 correlated sources highlighted by the HERA pull study are assigned ϵs=0.3\epsilon_s = 0.3, and all remaining sources retain ϵs=0\epsilon_s = 0. The resulting fit minima and PDF distributions are presented in Table 6 and Figure 6, respectively. Although illustrative, these results show that a transparent, externally motivated errors-on-errors treatment improves the goodness of fit across all three fit setups while preserving the characteristic low-xx differences in the uvu_v and gluon distributions.

Taken together, these studies show that, within a fixed PDF parametrisation and theoretical setup, errors-on-errors improve the inferred data–theory compatibility and increase the uncertainties on fitted parameters, without substantially changing the central PDF and cross-section predictions relative to the corresponding Gaussian fits. This stability may not persist in global fits combining different experiments and processes, where datasets may be in tension: errors-on-errors could then shift the central values as well, by down-weighting outlying or less precise datasets in favour of the mutually consistent and more precise ones.

Acknowledgments

The authors are grateful to Robert Thorne, Matthew Reader, Amanda Cooper-Sarkar, Davide Malito and Gitanjali Poddar for helpful discussions. E. C. and G. C. thank the UK Science and Technology Facilities Council for its support. The work of O. Z. has been supported by the MSCA4Ukraine Programme of the European Commission through the Alexander von Humboldt Foundation.

Iterative profiling of NPs

As described in § 2.2, minimisation of χ2(a,b)\chi^2(\boldsymbol{a},\boldsymbol{b}) can be made easier by first replacing the nuisance parameters b\boldsymbol{b} with their profiled values at fixed a\boldsymbol{a}. Here we describe an iterative procedure for finding these values. For fixed a\boldsymbol{a}, differentiation of Eq. (8) with respect to bsb_s gives

∑iΓisσi2(ri−∑pΓipbp)−(1+2ϵs2)bs1+2ϵs2bs2=0(18)\sum_i \frac{\Gamma_{is}}{\sigma_i^2}\left(r_i-\sum_p \Gamma_{ip}b_p\right)-\frac{(1+2\epsilon_s^2)b_s}{1+2\epsilon_s^2b_s^2}=0 \tag*{(18)}

where rir_i, σi2\sigma_i^2 and Γis\Gamma_{is} are defined in Eq. (9). For ϵs≠0\epsilon_s \ne0, these coupled equations are nonlinear. Here we implement the iterative form of the perturbative procedure derived in [6] in matrix form for use in xFITTER. At iteration kk, we define an effective constraint variance

σ^us2(k)=1+2ϵs2(bs(k))21+2ϵs2,(19)\hat{\sigma}_{u_s}^{2(k)}=\frac{1+2\epsilon_s^2\left(b_s^{(k)}\right)^2}{1+2\epsilon_s^2}, \tag*{(19)}

and construct the matrices

Asp(k)=δspσ^us2(k)+∑iΓisΓipσi2,Bs=∑iΓisriσi2.(20)A_{sp}^{(k)}=\frac{\delta_{sp}}{\hat{\sigma}_{u_s}^{2(k)}}+\sum_i\frac{\Gamma_{is}\Gamma_{ip}}{\sigma_i^2}, \qquad B_s=\sum_i\frac{\Gamma_{is}r_i}{\sigma_i^2}. \tag*{(20)}

The updated NPs are obtained from

bs(k+1)=∑p(A(k))sp−1Bp.(21)b_s^{(k+1)}=\sum_p\left(A^{(k)}\right)^{-1}_{sp}B_p. \tag*{(21)}

At convergence, Eqs. (19)–(21) solve Eq. (18) and therefore minimise the χ2\chi^2 with respect to the NPs at fixed a\boldsymbol{a}. The procedure is initialised with σ^us2(0)=1\hat{\sigma}_{u_s}^{2(0)}=1. The first iteration therefore uses

Asp(0)=δsp+∑iΓisΓipσi2,bs(1)=∑p(A(0))sp−1Bp.(22)A_{sp}^{(0)}=\delta_{sp}+\sum_i\frac{\Gamma_{is}\Gamma_{ip}}{\sigma_i^2}, \qquad b_s^{(1)}=\sum_p\left(A^{(0)}\right)^{-1}_{sp}B_p. \tag*{(22)}

which is precisely the analytic Gaussian profiling used by xFITTER for quadratic NPs. Subsequent iterations update the effective constraint widths and incorporate the non-quadratic GVM terms.

Local convergence of the expansion underlying this procedure is guaranteed provided

2ϵs2bs2<1.(23)2\epsilon_s^2b_s^2<1 . \tag*{(23)}

In practice, the iterative solution has been observed to converge even when this condition is not satisfied. In such cases, however, numerical convergence should be checked by increasing the number of iterations and verifying that the fitted results stabilise, particularly in the presence of large NP shifts, large errors-on-errors, or competing local minima. If convergence is not achieved, the NPs are minimised numerically.

Convergence and numerical validation

To demonstrate how the iterative profiling converges in practice, we consider the seven-source configuration obtained by forward selection from the HERAPDF2.0, NNLO, RTOPT baseline, as reported in Table 1. At convergence, further iterations leave the NP values unchanged, and the χ2\chi^2 minimisation equations at fixed a\boldsymbol{a} are satisfied. We therefore compare the profiled fit with a reference fit in which all seven NPs are minimised numerically. In the profiled fit, sysHZComb1053 and proc_tb21, which have large fitted shifts, are still minimised numerically, while the other five NPs are profiled iteratively. We compare the two fits as the number of profiling iterations is increased.

Figure 7 shows this comparison for the uvu_v, dvd_v, and gluon distributions. The columns correspond to the three PDFs and the rows to 2, 4, and 8 profiling iterations. Each panel shows the ratio of the profiled fit to the numerically minimized one for the indicated number of iterations. Two iterations already give a good approximation to the reference fit; after four, only a small residual difference remains in the dvd_v confidence interval, while after eight the profiled and externally minimised results are indistinguishable at the resolution of the figure. We therefore use eight iterations whenever errors-on-errors sources are profiled; the rapid convergence confirms the numerical accuracy of the iterative procedure.

Figure 7. Convergence of the iteratively profiled HERAPDF2.0, NNLO, RTOPT fit towards the fit in which all seven sources are minimised numerically, at Q2=4 GeV2Q^2 = 4\ \mathrm{GeV}^2. From left to right, the columns show the ratios for uvu_v, dvd_v, and the gluon; from top to bottom, the rows correspond to 2, 4, and 8 profiling iterations.

Evaluation of the Bartlett correction

The correction terms bχ2b_{\chi^2} and bCIb_{\mathrm{CI}}, entering the factors defined in Eq. (13), follow from the expansion developed by Lawley [23] and Cordeiro [21] and specialised to the GVM in Refs. [5, 6]. The confidence-interval correction is evaluated for all PDF parameters together, profiling over the NPs. The resulting factor CCIC_{\mathrm{CI}} is the appropriate one for the joint confidence region of the PDF parameters, and we use it as a reasonable approximation for the individual confidence intervals, multiplying their half-widths by CCI\sqrt{C_{\mathrm{CI}}}.

The two correction terms are

bCI=∑s(βsfree−βsfix),b_{\mathrm{CI}}=\sum_s\left(\beta_s^{\mathrm{free}}-\beta_s^{\mathrm{fix}}\right),
bχ2=∑s(3ϵs2−βsfree).(24)b_{\chi^2}=\sum_s\left(3\epsilon_s^2-\beta_s^{\mathrm{free}}\right). \tag*{(24)}

where the sums run over the systematic sources with ϵs>0\epsilon_s>0. Each source enters through a single quantity βs\beta_s: βsfix\beta_s^{\mathrm{fix}} is evaluated with the POIs held fixed at their best-fit values, whereas βsfree\beta_s^{\mathrm{free}} is evaluated with them retained as free fit parameters, so that their correlations with the NPs are included. The quantities βs\beta_s are proportional to ϵs2\epsilon_s^2 and are given by

βsfix=ϵs2[4Vsfixσ^us2−(Vsfixσ^us2)2],\beta_s^{\mathrm{fix}}=\epsilon_s^2\left[4\frac{V_s^{\mathrm{fix}}}{\hat{\sigma}_{u_s}^2}-\left(\frac{V_s^{\mathrm{fix}}}{\hat{\sigma}_{u_s}^2}\right)^2\right],
βsfree=ϵs2[4Vsfreeσ^us2−(Vsfreeσ^us2)2].(25)\beta_s^{\mathrm{free}}=\epsilon_s^2\left[4\frac{V_s^{\mathrm{free}}}{\hat{\sigma}_{u_s}^2}-\left(\frac{V_s^{\mathrm{free}}}{\hat{\sigma}_{u_s}^2}\right)^2\right]. \tag*{(25)}

Here VsfixV_s^{\mathrm{fix}} and VsfreeV_s^{\mathrm{free}} are the local variances of b^s\hat{b}_s, with the PDF parameters fixed and free respectively. The common denominator σ^us2\hat{\sigma}_{u_s}^2 is the profiled systematic variance in Eq. (2), evaluated at the fitted shift bs=b^sb_s=\hat{b}_s. Each ratio Vs/σ^us2V_s/\hat{\sigma}_{u_s}^2 is close to unity when the source is constrained mainly by its auxiliary measurement and decreases as the data determine its shift more precisely. Correspondingly, βs\beta_s ranges from zero to 3ϵs23\epsilon_s^2, which is the maximum contribution from one source. A source constrained similarly with the PDFs fixed or free contributes mainly to the goodness-of-fit correction, whereas an increase in variance when the PDFs are released transfers the correction to the PDF confidence intervals. A source unconstrained by the data in either case contributes to neither. Since only the variances enter, a contribution is possible even without a visible shift of the fitted central values.

The variances VsfixV_s^{\mathrm{fix}} and VsfreeV_s^{\mathrm{free}} entering Eq. (25) are obtained from the covariance matrix of the fitted parameters, namely the inverse Hessian of the χ2\chi^2 at its minimum, evaluated in the quadratic limit with all ϵs\epsilon_s set to zero. In this limit, the Hessian HH defined in Eq. (6) has the block form

H=(KGTGA)(26)H=\begin{pmatrix}K&G^{\mathsf{T}}\\G&A\end{pmatrix} \tag*{(26)}

Here AA is the nb×nbn_b\times n_b NP Hessian at fixed PDFs, where nbn_b is the number of fitted NPs. The na×nan_a\times n_a block KK is the PDF-parameter Hessian at fixed NPs, and the nb×nan_b\times n_a block GG couples the two parameter sets. At the fit minimum, the three blocks are given by

A=ΓTWΓ+Λ,A=\Gamma^{\mathsf{T}}W\Gamma+\Lambda,
G=ΓTWD,G=\Gamma^{\mathsf{T}}WD,
K=DTWD.(27)K=D^{\mathsf{T}}WD. \tag*{(27)}

Here WW is the inverse covariance matrix of the measurements at the fitted prediction. In the absence of statistical covariance blocks it is diagonal, with entries 1/σi2(a^)1/\sigma_i^2(\hat{a}), where σi\sigma_i is defined in Eq. (9); otherwise, WW is the inverse of the full covariance matrix. The correlated-shift matrix Γ\Gamma, also defined in Eq. (9), is evaluated at a^\hat{a}. The elements Dim=∂Ti/∂am∣a^D_{im}=\left.\partial T_i/\partial a_m\right|_{\hat{a}} are evaluated numerically using finite differences. Finally, Λ\Lambda is the diagonal NP constraint matrix, with Λss=1/σ^us2\Lambda_{ss}=1/\hat{\sigma}_{u_s}^2. These expressions neglect the dependence of Γ\Gamma and WW on the PDF parameters, so that they hold in general, independently of the multiplicative treatment of the uncertainties. We have verified that the neglected terms affect both the χ2\chi^2 and confidence-interval corrections at the sub-percent level in the examples considered. Inverting HH block by block, the variances entering Eq. (25) are

Vsfix=(A−1)ss,V_s^{\mathrm{fix}}=(A^{-1})_{ss},
Vsfree=[A−1+A−1G(K−GTA−1G)−1GTA−1]ss.(28)V_s^{\mathrm{free}}=\left[A^{-1}+A^{-1}G\left(K-G^{\mathsf{T}}A^{-1}G\right)^{-1}G^{\mathsf{T}}A^{-1}\right]_{ss}. \tag*{(28)}

The second term in VsfreeV_s^{\mathrm{free}} is generated by the mixed block GG and is positive semidefinite, so releasing the PDF parameters can only increase the NP variance relative to VsfixV_s^{\mathrm{fix}}. This occurs because a non-zero GG allows a change in the NP to be partly compensated by a correlated change in the PDFs, leaving the description of the data approximately unchanged. The two variances therefore coincide when G=0G=0, while their difference quantifies this degeneracy.

NP shifts and uncertainties

Here we examine the effect of errors-on-errors on NP shifts and their uncertainties for the seven-source configuration obtained by forward selection from the HERAPDF2.0, NNLO, RTOPT baseline fit in § 4 (Table 1), with ϵs=0.6\epsilon_s = 0.6 for the selected sources.

The response of these NPs helps explain how the fit accommodates the data–theory tension. Table 7 compares their fitted shifts and conditional uncertainties before and after introducing errors-on-errors. The quantities bsb_s reported here are fitted NP shifts expressed in units of the nominal systematic uncertainty vs\sqrt{v_s} in Eq. (1), with us=0u_s = 0 and vs=1v_s = 1 in our convention. Thus, ∣bs∣=1|b_s| = 1 denotes a displacement equal to one nominal systematic uncertainty; it does not by itself specify the statistical significance of that displacement. The conditional uncertainties δbs\delta b_s describe how precisely the fit determines these shifts. They are expressed in the same units and evaluated with the PDF parameters held fixed at each fit minimum.

SystematicGaussian baseline bs±δbsb_s \pm\delta b_sErrors-on-errors bs±δbsb_s \pm\delta b_s
sysHZComb1053−2.67±0.92-2.67 \pm0.92−16.41±2.36-16.41 \pm2.36
proc_tb21−3.10±0.67-3.10 \pm0.67−6.31±1.05-6.31 \pm1.05
sysHZComb1064+0.71±0.93+0.71 \pm0.93+9.94±2.76+9.94 \pm2.76
sysHZComb1114+1.77±0.81+1.77 \pm0.81+4.24±1.39+4.24 \pm1.39
sysHZComb1144+1.48±0.84+1.48 \pm0.84+4.08±1.57+4.08 \pm1.57
sysHZComb1111−1.84±0.55-1.84 \pm0.55−2.53±0.67-2.53 \pm0.67
sysHZComb1113+1.38±0.84+1.38 \pm0.84+3.81±1.56+3.81 \pm1.56

Table 7. Fitted NP shifts bsb_s and conditional uncertainties δbs\delta b_s for the seven NPs selected in the HERAPDF2.0, NNLO, RTOPT forward-selection fit.

Sizeable shifts can indicate sensitivity to errors-on-errors, since the logarithmic penalty in Eq. (3) grows more slowly than the Gaussian quadratic penalty at large shifts. One should keep in mind that, during forward selection, assigning a nonzero ϵs\epsilon_s to an additional systematic source and refitting can change the fit minimum and modify the shifts of the other NPs. Shifts that are small in the Gaussian baseline can therefore become sizeable at later steps. For example, sysHZComb1064 is selected even though its Gaussian-baseline shift is only 0.71, with a conditional uncertainty of 0.93. Conversely, a large shift does not necessarily imply a large errors-on-errors effect, which also depends on the data constraints and parameter correlations.

The shift of sysHZComb1053 may seem large at about 16 times the nominal systematic uncertainty. Its size should, however, be interpreted differently in the GVM than in the Gaussian case. For ϵs=0.6\epsilon_s = 0.6, a deviation of this size under the GVM constraint is about as rare as a Gaussian fluctuation of 2–3 standard deviations [4].

The conditional uncertainties on the fitted shifts, δbs\delta b_s, increase for all seven selected NPs. In the Gaussian treatment with the PDFs held fixed, the unit-width auxiliary constraints already limit these uncertainties to one, and the data can only reduce them further. This bound does not hold in general in the GVM: the profiled systematic variance in Eq. (2) increases with bsb_s, so the fitted uncertainties can also grow with the shifts and exceed one.

Another related diagnostic is provided by normalised pulls, which combine information from the NP shifts and their uncertainties. In the Gaussian baseline, a source that is weakly constrained by the data is expected to shift only slightly. A sizeable shift accompanied by little reduction in uncertainty therefore gives a large pull. Following the HERA combination definition [1], with us=0u_s = 0 and vs=1v_s = 1, one can define

ps=bs1−(δbs)2.(29)p_s = \frac{b_s}{\sqrt{1-(\delta b_s)^2}}. \tag*{(29)}

These pulls can therefore help identify sources worth testing for errors-on-errors. However, this definition cannot be applied generally once errors-on-errors are introduced: the uncertainties δbs\delta b_s can exceed unity, making the argument of the square root negative. We therefore use the NP shifts rather than the pulls to compare the Gaussian and errors-on-errors fits.

References

  1. [1]H1 and ZEUS Collaborations: Combination of measurements of inclusive deep inelastic e±p scattering cross sections and QCD analysis of HERA data. Eur. Phys. J. C 75(12), 580 (2015) https://doi.org/10.1140/epjc/s10052-015-3710-4 arXiv:1506.06042 [hep-ex]
  2. [2]Bonvini, M., Giuli, F.: A new simple PDF parametrization: improved description of the HERA data. Eur. Phys. J. Plus 134(10), 531 (2019) https://doi.org/10.1140/epjp/i2019-12872-x arXiv:1902.11125 [hep-ph]
  3. [3]Abdolmaleki, H., et al.: Impact of low-x resummation on QCD analysis of HERA data. Eur. Phys. J. C 78(8), 621 (2018) https://doi.org/10.1140/epjc/s10052-018-6090-8 arXiv:1802.00064 [hep-ph]
  4. [4]Cowan, G.: Statistical models with uncertain error parameters. Eur. Phys. J. C 79(2), 133 (2019) https://doi.org/10.1140/epjc/s10052-019-6644-4 arXiv:1809.05778 [physics.data-an]
  5. [5]Canonero, E., Brazzale, A.R., Cowan, G.: Higher-order asymptotic corrections and their application to the Gamma Variance Model. Eur. Phys. J. C 83, 1100 (2023) https://doi.org/10.1140/epjc/s10052-023-12263-7 arXiv:2304.10574 [physics.data-an]
  6. [6]Canonero, E., Cowan, G.: Correlated systematic uncertainties and errors-on-errors in measurement combinations with an application to the 7–8 TeV ATLAS–CMS top quark mass combination. Eur. Phys. J. C 85(2), 156 (2025) https://doi.org/10.1140/epjc/s10052-025-13884-w arXiv:2407.05322 [physics.data-an]
  7. [7]Bailey, S., Harland-Lang, L.A.: Differential top quark pair production at the LHC: Challenges for PDF fits. Eur. Phys. J. C 80, 60 (2020) https://doi.org/10.1140/epjc/s10052-020-7633-3 arXiv:1909.10541 [hep-ph]
  8. [8]Watt, G., Thorne, R.S.: Study of Monte Carlo approach to experimental uncertainty propagation with MSTW 2008 PDFs. JHEP 08, 052 (2012) https://doi.org/10.1007/JHEP08(2012)052 arXiv:1205.4024 [hep-ph]
  9. [9]Barontini, A., Costantini, M.N., De Crescenzo, G., Forte, S., Ubiali, M.: Evaluating the faithfulness of PDF uncertainties in the presence of inconsistent data. arXiv:2503.17447 [hep-ph] (2025)arxiv.org/abs/2503.17447
  10. [10]Alekhin, S., et al.: HERAFitter, open source QCD fit project. Eur. Phys. J. C 75(7), 304 (2015) https://doi.org/10.1140/epjc/s10052-015-3480-z arXiv:1410.4412 [hep-ph]
  11. [11]Bertone, V., et al.: xFitter 2.0.0: An open source QCD fit framework. PoS DIS2017, 203 (2018) https://doi.org/10.22323/1.297.0203 arXiv:1709.01151 [hep-ph]
  12. [12]Abdolmaleki, H., et al.: xFitter: An Open Source QCD Analysis Framework. A resource and reference document for the Snowmass study. arXiv:2206.12465 [hep-ph] (2022)arxiv.org/abs/2206.12465
  13. [13]Pumplin, J., Stump, D., Brock, R., Casey, D., Huston, J., Kalk, J., Lai, H.L., Tung, W.K.: Uncertainties of predictions from parton distribution functions. II. The Hessian method. Phys. Rev. D 65, 014013 (2001) https://doi.org/10.1103/PhysRevD.65.014013 arXiv:hep-ph/0101032
  14. [14]ATLAS Collaboration: Precision measurement and interpretation of inclusive \(W^+\), \(W^-\) and \(Z/\gamma^*\) production cross sections with the ATLAS detector. Eur. Phys. J. C 77(6), 367 (2017) https://doi.org/10.1140/epjc/s10052-017-4911-9 arXiv:1612.03016 [hep-ex]
  15. [15]ATLAS Collaboration: Determination of the parton distribution functions of the proton from ATLAS measurements of differential \(W^\pm\) and \(Z\) boson production in association with jets. JHEP 07, 223 (2021) https://doi.org/10.1007/JHEP07(2021)223 arXiv:2101.05095 [hep-ex]
  16. [16]ATLAS Collaboration: Determination of the parton distribution functions of the proton using diverse ATLAS data from \(pp\) collisions at \(\sqrt{s}=7\), 8 and 13 TeV. Eur. Phys. J. C 83, 438 (2022) https://doi.org/10.1140/epjc/s10052-022-10217-z arXiv:2112.11266 [hep-ex]
  17. [17]CMS Collaboration: Measurement of associated production of a \(W\) boson and a charm quark in proton-proton collisions at \(\sqrt{s}=13\) TeV. Eur. Phys. J. C 79, 269 (2019) https://doi.org/10.1140/epjc/s10052-019-6752-1 arXiv:1811.10021 [hep-ex]
  18. [18]CMS Collaboration: Measurement and QCD analysis of double-differential inclusive jet cross sections in proton-proton collisions at \(\sqrt{s}=13\) TeV. JHEP 02, 142 (2022) https://doi.org/10.1007/JHEP02(2022)142 arXiv:2111.10431 [hep-ex]. Addendum: JHEP 12 (2022) 035, doi:10.1007/JHEP12(2022)035
  19. [19]Stump, D., Pumplin, J., Brock, R., Casey, D., Huston, J., Kalk, J., Lai, H.L., Tung, W.K.: Uncertainties of predictions from parton distribution functions. I. The Lagrange multiplier method. Phys. Rev. D 65, 014012 (2001) https://doi.org/10.1103/PhysRevD.65.014012 arXiv:hep-ph/0101051
  20. [20]Bartlett, M.S.: Properties of sufficiency and statistical tests. Proc. Roy. Soc. Lond. A 160, 268–282 (1937) https://doi.org/10.1098/rspa.1937.0109
  21. [21]Cordeiro, G.M.: General matrix formulae for computing Bartlett corrections. Stat. Probab. Lett. 16(1), 11–18 (1993) https://doi.org/10.1016/0167-7152(93)90115-Y
  22. [22]Cordeiro, G.M., Cribari-Neto, F.: An Introduction to Bartlett Correction and Bias Reduction. Springer, Berlin, Heidelberg (2014). https://doi.org/10.1007/978-3-642-55255-7
  23. [23]Lawley, D.N.: A general method for approximating the distribution of likelihood ratio criteria. Biometrika 43(3/4), 295–303 (1956) https://doi.org/10.2307/2332908
  24. [24]Bailey, S., et al.: Parton distributions from LHC, HERA, Tevatron and fixed target data: MSHT20 PDFs. Eur. Phys. J. C 81, 341 (2021) https://doi.org/10.1140/epjc/s10052-021-09057-0 arXiv:2012.04684 [hep-ph]
  25. [25]Reader, M.: A study of systematic uncertainties within the MSHT PDF framework. PoS DIS2024, 059 (2025) https://doi.org/10.22323/1.469.0059 arXiv:2408.12922 [hep-ph]
  26. [26]Botje, M.: QCDNUM: Fast QCD evolution and convolution. Comput. Phys. Commun. 182, 490–532 (2011) https://doi.org/10.1016/j.cpc.2010.10.020 arXiv:1005.1481 [hep-ph]
  27. [27]Thorne, R.S., Roberts, R.G.: An ordered analysis of heavy flavor production in deep inelastic scattering. Phys. Rev. D 57, 6871–6898 (1998) https://doi.org/10.1103/PhysRevD.57.6871 arXiv:hep-ph/9709442 [hep-ph]
  28. [28]Thorne, R.S.: A variable-flavor number scheme for NNLO. Phys. Rev. D 73, 054019 (2006) https://doi.org/10.1103/PhysRevD.73.054019 arXiv:hep-ph/0601245 [hep-ph]
  29. [29]Thorne, R.S.: Effect of changes of variable flavour number scheme on parton distribution functions and predicted cross sections. Phys. Rev. D 86, 074017 (2012) https://doi.org/10.1103/PhysRevD.86.074017 arXiv:1201.6180 [hep-ph]
  30. [30]Ball, R.D., Bertone, V., Bonvini, M., Marzani, S., Rojo, J., Rottoli, L.: Parton distributions with small-x resummation: evidence for BFKL dynamics in HERA data. Eur. Phys. J. C 78(4), 321 (2018) https://doi.org/10.1140/epjc/s10052-018-5774-4 arXiv:1710.05935 [hep-ph]
  31. [31]Kuraev, E.A., Lipatov, L.N., Fadin, V.S.: Multi-Reggeon processes in the Yang-Mills theory. Sov. Phys. JETP 44, 443–450 (1976)
  32. [32]Kuraev, E.A., Lipatov, L.N., Fadin, V.S.: The Pomeranchuk singularity in nonabelian gauge theories. Sov. Phys. JETP 45, 199–204 (1977)
  33. [33]Balitsky, I.I., Lipatov, L.N.: The Pomeranchuk singularity in quantum chromodynamics. Sov. J. Nucl. Phys. 28, 822–829 (1978)
  34. [34]Jaroszewicz, T.: Gluonic Regge singularities and anomalous dimensions in QCD. Phys. Lett. B 116, 291–294 (1982) https://doi.org/10.1016/0370-2693(82)90345-8
  35. [35]Catani, S., Hautmann, F.: High-energy factorization and small x deep inelastic scattering beyond leading order. Nucl. Phys. B 427, 475–524 (1994) https://doi.org/10.1016/0550-3213(94)90636-X arXiv:hep-ph/9405388
  36. [36]Catani, S., Hautmann, F.: Quark anomalous dimensions at small x. Phys. Lett. B 315, 157–163 (1993) https://doi.org/10.1016/0370-2693(93)90174-G
  37. [37]Fadin, V.S., Lipatov, L.N.: BFKL pomeron in the next-to-leading approximation. Phys. Lett. B 429, 127–134 (1998) https://doi.org/10.1016/S0370-2693(98)00473-0 arXiv:hep-ph/9802290
  38. [38]Ciafaloni, M., Camici, G.: Energy scale(s) and next-to-leading BFKL equation. Phys. Lett. B 430, 349–354 (1998) https://doi.org/10.1016/S0370-2693(98)00551-6 arXiv:hep-ph/9803389
  39. [39]Bonvini, M., Marzani, S., Peraro, T.: Small-x resummation from HELL. Eur. Phys. J. C 76(11), 597 (2016) https://doi.org/10.1140/epjc/s10052-016-4445-6 arXiv:1607.02153 [hep-ph]
  40. [40]Bonvini, M., Marzani, S., Muselli, C.: Towards parton distribution functions with small-x resummation: HELL 2.0. JHEP 12, 117 (2017) https://doi.org/10.1007/JHEP12(2017)117 arXiv:1708.07510 [hep-ph]
  41. [41]Bonvini, M., Marzani, S.: Four-loop splitting functions at small x. JHEP 06, 145 (2018) https://doi.org/10.1007/JHEP06(2018)145 arXiv:1805.06460 [hep-ph]
  42. [42]Bonvini, M.: Small-x phenomenology at the LHC and beyond: HELL 3.0 and the case of the Higgs cross section. Eur. Phys. J. C 78(10), 834 (2018) https://doi.org/10.1140/epjc/s10052-018-6315-x arXiv:1805.08785 [hep-ph]
  43. [43]Bertone, V., Carrazza, S., Rojo, J.: APFEL: A PDF evolution library with QED corrections. Comput. Phys. Commun. 185, 1647–1668 (2014) https://doi.org/10.1016/j.cpc.2014.03.007 arXiv:1310.1394 [hep-ph]
  44. [44]Forte, S., Laenen, E., Nason, P., Rojo, J.: Heavy quarks in deep-inelastic scattering. Nucl. Phys. B 834, 116–162 (2010) https://doi.org/10.1016/j.nuclphysb.2010.03.014 arXiv:1001.2312 [hep-ph]
  45. [45]Turkot, O.: Search for contact interactions and finite quark radius using combined e⁺p data at HERA. PhD thesis, Universität Hamburg (2017). https://doi.org/10.3204/PUBDB-2017-09190 · DESY-THESIS-2017-035
  46. [46]Sinervo, P.: Definition and Treatment of Systematic Uncertainties in High Energy Physics and Astrophysics. eConf C030908, 122–129 (2003)

Paper details

vlnc:W_9mw2samw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents