1: \subsection{Stability of PDF fits \protect
2: \footnote{Contributing authors: J.~Huston, J.~Pumplin.}
3: \label{sec:cteq}}
4:
5: One of the issues raised at the workshop is the reliability of determinations of
6: parton distribution functions (PDFs), which might be compromised for example by
7: the neglect of NNLO effects or non-DGLAP evolution in the standard analysis, or
8: hidden assumptions made in parameterizing the PDFs at nonperturbative scales.
9: We summarize the results of the CTEQ PDF group on this issue.
10: For the full story see \cite{Huston:2005jm}.
11:
12: \subsubsection{Stability of PDF determinations}
13: The stability of NLO global analysis was seriously challenged
14: by an analysis \cite{epj:c35:325} which found a 20\% variation
15: in the cross section predicted for $W$ production at the LHC -- a critical
16: ``standard candle'' process for hadron colliders -- when certain cuts on input
17: data are varied. If this instability were confirmed, it would significantly
18: impact the phenomenology of a wide range of physical processes for the
19: Tevatron Run II and the LHC. The CTEQ PDF group therefore performed an
20: independent study of this issue within their global analysis framework.
21: In addition, to explore the dependence of the results on assumptions about the
22: parameterization of PDFs at the starting scale $Q_{0} \! = \! 1.3\, \mathrm{GeV}$,
23: we also studied the effect of allowing a negative gluon distribution at small $x$ --
24: a possibility that is favored by the MRST NLO analysis, and that is closely tied
25: to the W cross section controversy.
26:
27: The stability of the global analysis was investigated by varying the inherent
28: choices that must be made to perform the analysis. These choices include the
29: selection of experimental data points based on kinematic cuts, the functional
30: forms used to parameterize the initial nonperturbative parton distribution functions,
31: and the treatment of $\alpha_s$.
32:
33: The stability of the results is most
34: conveniently measured by differences in the global $\chi^2$ for the relevant fits.
35: To quantitatively define a change of $\chi^{2}$ that characterizes a significant
36: change in the quality of the PDF fit is a difficult issue in global QCD analysis.
37: In the context of the current analysis, we have argued that an increase by
38: $\Delta\chi^{2}\sim 100$ (for $\sim$ \! 2000 data points) represents roughly
39: a 90\% confidence level uncertainty on PDFs due to the uncertainties of the
40: current input experimental data~\cite{%
41: Pumplin:2000vx,Stump:2001gu,Pumplin:2001ct,cteq}.
42: In other words, PDFs with $\chi^2 - \chi^2_{\mathrm{Best Fit}} > 100$ are
43: regarded as not tolerated by current data.
44:
45: The CTEQ6 and previous CTEQ global fits imposed ``standard'' cuts
46: $Q > 2 \,\mathrm{GeV}$ and $W > 3.5 \, \mathrm{GeV}$ on the input data set,
47: in order to suppress higher-order terms in the perturbative expansion and the
48: effects of resummation and power-law (``higher twist'') corrections.
49: We examined the effect of stronger cuts on $Q$ to see if the fits are stable.
50: We also examined the effect of imposing cuts on $x$, which should serve
51: to suppress any errors due to deviations from DGLAP evolution, such as those
52: predicted by BFKL. The idea is that any inconsistency in the global fit due to
53: data points near the boundary of the accepted region will be revealed by an
54: improvement in the fit to the data that remain after those near-boundary
55: points have been removed.
56: In other words, the decrease in $\chi^2$ for the subset of data that is retained,
57: when the PDF shape parameters are refitted to that subset alone, measures the
58: degree to which the fit to that subset was distorted in the original fit by
59: compromises imposed by the data at low $x$ and/or low $Q$.
60:
61: The main results of this study are presented in Table~\ref{tab:tableI}.
62: Three fits are shown, from three choices of the cuts on input data as specified
63: in the table. They are labeled `standard', `intermediate' and `strong'.
64: $N_{\rm pts}$ is the number of data points that pass the cuts in each case,
65: and $\chi^2_{N_{\rm pts}}$ is the $\chi^2$ value for that subset of data.
66: The fact that the changes in $\chi^2$ in each column are insignificant compared
67: to the uncertainty tolerance is strong evidence that our NLO global fit results
68: are very stable with respect to choices of kinematic cuts.
69: %=========== Table I ================
70: \begin{table}[t]
71: \begin{center}
72: \begin{tabular}{||r|r|r|r||c|c|c|c||c||}
73: \hline \hline
74: Cuts & $Q_{\mathrm{min}}$ & $x_{\mathrm{min}}$ & $N_{\mathrm{pts}}$ &
75: $\chi^2_{1926}$ & $\chi^2_{1770}$ & $\chi^2_{1588}$ &
76: $\sigma_W^{\mathrm{LHC}} \! \times \! B_{\ell \nu} \, [\mathrm{nb}]$
77: \\
78: \hline
79: standard & $2 \, \mathrm{GeV}$ & $0$ & $1926$ & $2023$ &
80: $1850$ & $1583$ & $20.02$
81: \\
82: intermediate & $2.5 \, \mathrm{GeV}$ & $0.001$ & $1770$ & -- &
83: $1849$ & $1579$ & $20.10$
84: \\ strong & $3.162 \, \mathrm{GeV}$ & $0.005$ & $1588$ & -- &
85: -- & $1573$ & $20.34$
86: \\
87: \hline \hline
88: \end{tabular}
89: \caption{Comparisons of three fits with different choices of the cuts on input
90: data at the $Q$ and $x$ values indicated. In these fits, a conventional
91: positive-definite gluon parameterization was used.
92: \label{tab:tableI}}
93: \end{center}
94: \end{table}
95: %%=========== Table II ================
96: \begin{table}[ht]
97: \begin{center}
98: \begin{tabular}{||r|r|r|r||c|c|c|c||c||}
99: \hline \hline
100: Cuts & $Q_{\mathrm{min}}$ & $x_{\mathrm{min}}$ & $N_{\mathrm{pts}}$ &
101: $\chi^2_{1926}$ & $\chi^2_{1770}$ & $\chi^2_{1588}$ &
102: $\sigma_W^{\mathrm{LHC}} \! \times \! B_{\ell \nu} \, [\mathrm{nb}]$
103: \\
104: \hline
105: standard & $2 \, \mathrm{GeV}$ & $0$ & $1926$ & $2011$ & $1845$ &
106: $1579$ & $19.94$
107: \\
108: intermediate & $2.5 \, \mathrm{GeV}$ & $0.001$ & $1770$ & -- & $1838$ &
109: $1574$ & $19.80$
110: \\
111: strong & $3.162 \, \mathrm{GeV}$ & $0.005$ & $1588$ & -- & -- &
112: $1570$ & $19.15$
113: \\
114: \hline \hline
115: \end{tabular}
116: \caption{Same as Table~\ref{tab:tableI} except that the gluon parameterization
117: is extended to allow negative values. \label{tab:tableII} }
118: \end{center}
119: \end{table}
120:
121: We extended the analysis to a series of fits in which the gluon distribution
122: $g(x)$ is allowed to be negative at small $x$, at the scale
123: $Q_0 \! = \! 1.3 \, \mathrm{GeV}$ where we begin the DGLAP evolution.
124: The purpose of this additional study is to determine whether the feature
125: of a negative gluon PDF is a key element in the stability puzzle,
126: as suggested by the findings of~\cite{epj:c35:325}.
127: The results are presented in Table~\ref{tab:tableII}.
128: Even in this extended case, we find no evidence of instability.
129: For example, $\chi^2$ for the subset of 1588 points that pass
130: the \textit{strong} cuts increases only from 1570 to 1579 when the fit is
131: extended to include the full standard data set.
132:
133: Comparing the elements of Table~\ref{tab:tableI} and Table~\ref{tab:tableII}
134: shows that our fits with $g(x) < 0$ have slightly smaller values of$\chi^2$:
135: e.g., $2011$ versus $2023$ for the standard cuts.
136: However, the difference $\Delta \chi^2 \! = \! 12$ between these values is again
137: not significant according to our tolerance criterion.
138:
139: \subsubsection{W cross sections at the LHC}
140: The last columns of Tables~\ref{tab:tableI} and \ref{tab:tableII} show the
141: predicted cross section for $W^+ + W^-$ production at the LHC. This prediction
142: is also very stable: it changes by only $1.6 \%$ for the positive-definite
143: gluon parameterization, which is substantially less than the overall PDF uncertainty
144: of $\sigma_W$ estimated previously with the standard cuts.
145: For the negative gluon parameterization, the change is $4 \%$--larger,
146: but still less than the overall PDF uncertainty.
147: These results are explicitly displayed, and compared to the MRST results
148: in Fig.~\ref{fig:WtotXs}.
149: %
150: \begin{figure}[ht]\centering
151: \resizebox*{0.6\textwidth}{!} {\includegraphics[clip=true]{WLHC4ZZ.eps} }
152: \caption{ Predicted total cross section of $W^+ + W^-$ production at the LHC
153: for the fits obtained in our stability study, compared to the NLO results of
154: Ref.~\protect\cite{epj:c35:325}.
155: The $Q$-cut values associated with the CTEQ points are given in the two tables.
156: The overall PDF uncertainty of the prediction is $\sim 5\%$. }
157: \label{fig:WtotXs}
158: \end{figure}
159: %
160: We see that this physical prediction is indeed insensitive to the kinematic cuts
161: used for the fits, and to the assumption on the positive definiteness of the gluon
162: distribution.
163:
164: We also studied the stability of the prediction for $\sigma_{W}$
165: using the Lagrange Multiplier (LM) method of Refs.~\cite{%
166: Pumplin:2000vx,Stump:2001gu,Pumplin:2001ct}.
167: Specifically, we performed a series of fits to the global data set that are
168: constrained to specific values of $\sigma_W$ close to the best-fit prediction.
169: The resulting variation of $\chi^{2}$ versus $\sigma_{W}$ measures the uncertainty
170: of the prediction. We repeated the constrained fits for each case of fitting
171: choices (parameterization and kinematic cuts).
172: In this way we gain an understanding of the stability of the uncertainty,
173: in addition to the stability of the central prediction.
174:
175: Figure~\ref{fig:ChiVsSigCutsPG} shows the results of the LM study for the three
176: sets of kinematic cuts described in Table~\ref{tab:tableI}, all of which have
177: a positive-definite gluon distribution.
178: %
179: \begin{figure}[htb]\centering
180: \resizebox*{0.4\textwidth}{!}{\includegraphics[clip=true]{figLM5newx.eps}}
181: \caption{ Lagrange multiplier results for the $W$ cross section (in $\mathrm{nb}$)
182: at the LHC using a positive-definite gluon. The three curves, in order of
183: decreasing steepness, correspond to the three sets of kinematic cuts labeled
184: standard/intermediate/strong in Table~\ref{tab:tableI}.
185: \label{fig:ChiVsSigCutsPG}}
186: \end{figure}
187: %
188: The $\chi^2$ shown along the vertical axis is normalized to its value for the best
189: fit in each series. In all three series, $\chi^2$ depends almost quadratically on
190: $\sigma_W$. We observe several features:
191: \begin{itemize}
192: \item
193: The location of the minimum of each curve represents the best-fit prediction
194: for $\sigma_W^{\mathrm{LHC}}$ for the corresponding choice of cuts. The fact
195: that the three minima are close together displays the stability of the predicted
196: cross section already seen in Table~\ref{tab:tableI}.
197: \item
198: Although more restrictive cuts make the global fit less sensitive to possible
199: contributions from resummation, power-law and other nonperturbative effects,
200: the loss of constraints caused by the removal of precision HERA data points at
201: small $x$ and low $Q$ results directly in increased uncertainties on the PDF
202: parameters and their physical predictions. This is shown in
203: Fig.~\ref{fig:ChiVsSigCutsPG} by the increase of the width of the curves with
204: stronger cuts. The uncertainty of the predicted $\sigma_W$ increases by more
205: than a factor of 2 in going from the standard cuts to the strong cuts.
206: \end{itemize}
207:
208: Figure~\ref{fig:ChiVsSigCutsNG} shows the results of the LM study for the three
209: sets of kinematic cuts described in Table~\ref{tab:tableII}, all of which have
210: a gluon distribution which is allowed to go negative.
211:
212: \begin{figure}[htb]\centering
213: \resizebox*{0.4\textwidth}{!}{\includegraphics[clip=true]{figLM6new.eps}}
214: \caption{ Lagrange multiplier results for the $W$ cross section (in $\mathrm{nb}$)
215: at the LHC using a functional form where the gluon is not required to be
216: positive-definite. The three curves, in order of decreasing steepness,
217: correspond to the three sets of kinematic cuts labeled standard/intermediate/strong
218: in Table~\ref{tab:tableII}.
219: \label{fig:ChiVsSigCutsNG}}
220: \end{figure}
221:
222: We observe:
223: \begin{itemize}
224: \item
225: Removing the positive definiteness condition necessarily lowers the value of
226: $\chi^2$, because more possibilities are opened up in the $\chi^2$ minimization
227: procedure. But the decrease is insignificant compared to other sources of
228: uncertainty. Thus, a negative gluon PDF is allowed, but not required.
229: \item
230: The minima of the two curves occur at approximately the same $\sigma_{W}$.
231: Allowing a negative gluon makes no significant change in the central prediction
232: -- merely a decrease of about $1\,\%$, which is small compared to the overall
233: PDF uncertainty.
234: \item
235: For the standard set of cuts, allowing a negative gluon PDF would expand the
236: uncertainty range only slightly. For the intermediate and strong cuts, allowing
237: a negative gluon PDF would significantly expand the uncertainty range.
238: \end{itemize}
239:
240: \begin{figure}[hbt]
241: \mbox{
242: \resizebox{0.48\textwidth}{!}{
243: \includegraphics[clip=true,height=.25\textheight]{GluUncQ3s3.ps}}
244: \hfill
245: \resizebox{0.48\textwidth}{!}{
246: \includegraphics[clip=true,height=.25\textheight]{GluUncQ3s5.ps}}
247: }
248: \caption{Left: mrst2002 NLO (solid) and NNLO (dotted);
249: Right: mrst2004 NLO (solid) and NNLO (dotted);
250: Shaded region is uncertainty according to the 40 eigenvector sets of CTEQ6.1.
251: \label{fig:GluUncQ3s35}}
252: \end{figure}
253: We examined a number of aspects of our analysis that might account for the
254: difference in conclusions between our stability study and that
255: of~\cite{epj:c35:325}.
256: A likely candidate seems to be that in order to obtain stability, it is necessary
257: to allow a rather free parametrization of the input gluon distribution.
258: This suspicion is seconded by recent work by MRST \cite{Martin:2004ir}, in which
259: a different gluon parametrization appears to lead to a best-fit gluon distribution
260: that is close to that of CTEQ6. In summary, we found that the NLO PDFs and their
261: physical predictions at the Tevatron and LHC are quite stable with respect to
262: variations of the kinematic cuts and the PDF parametrization after all.
263:
264: \subsubsection{NLO and NNLO}
265: In recent years, some preliminary next-to-next-leading-order (NNLO) analyses
266: for PDFs have been carried out either for DIS alone \cite{Alekhin:2003yh},
267: or in a global analysis context \cite{Martin:2002dr} -- even if all the necessary
268: hard cross sections, such as inclusive jet production, are not yet available
269: at this order.
270: Determining the parton distributions at NNLO is obviously desirable on theoretical
271: grounds, and it is reasonable to plan for having a full set of tools for a true
272: NNLO global analysis in place by the time LHC data taking begins.
273: At the moment, however, NNLO fitting is not a matter of pressing necessity,
274: since the difference between NLO and NNLO appears to be very small compared
275: to the other uncertainties in the PDF analysis.
276: This is demonstrated in Fig.~\ref{fig:GluUncQ3s35}, which shows the NLO and NNLO
277: gluon distributions extracted by the MRST group.
278: The difference between the two curves is much smaller than the other uncertainties
279: measured by the 40 eigenvector uncertainty sets of CTEQ6.1, which is shown by
280: the shaded region. The difference is also much smaller than the difference
281: between CTEQ and MRST best fits.
282: Similar conclusions \cite{Pumplin:2005yf} can be found using the NLO and NNLO
283: fits by Alekhin.
284:
285: