The displayed range is independent of the set of data points which are being displayed.
In other words: even if you choose xmin
and xmax
in a way which hides all data points, pgfplots will still draw every data point.
If you want to select/discard subsets of your data points, you should study the filtering mechanisms, in particular: the restrict x to domain=1:9999
key with appropriate values. Filtering means to drop single coordinates. The unbounded coords=discard|jump
key controls how pgfplots should react (either as if it never saw the discarded points at all or by introducing jumps).
This happens because PGFPlots only uses one "stack" per axis: You're stacking the second confidence interval on top of the first. The easiest way to fix this is probably to use the approach described in "Is there an easy way of using line thickness as error indicator in a plot?": After plotting the first confidence interval, stack the upper bound on top again, using stack dir=minus
. That way, the stack will be reset to zero, and you can draw the second confidence interval in the same fashion as the first:
\documentclass{standalone}
\usepackage{pgfplots, tikz}
\usepackage{pgfplotstable}
\pgfplotstableread{
temps y_h y_h__inf y_h__sup y_f y_f__inf y_f__sup
1 0.237340 0.135170 0.339511 0.237653 0.135482 0.339823
2 0.561320 0.422007 0.700633 0.165871 0.026558 0.305184
3 0.694760 0.534205 0.855314 0.074856 -0.085698 0.235411
4 0.728306 0.560179 0.896432 0.003361 -0.164765 0.171487
5 0.711710 0.544944 0.878477 -0.044582 -0.211349 0.122184
6 0.671241 0.511191 0.831291 -0.073347 -0.233397 0.086703
7 0.621177 0.471219 0.771135 -0.088418 -0.238376 0.061540
8 0.569354 0.431826 0.706882 -0.094382 -0.231910 0.043146
9 0.519973 0.396571 0.643376 -0.094619 -0.218022 0.028783
10 0.475121 0.366990 0.583251 -0.091467 -0.199598 0.016664
}{\table}
\begin{document}
\begin{tikzpicture}
\begin{axis}
% y_h confidence interval
\addplot [stack plots=y, fill=none, draw=none, forget plot] table [x=temps, y=y_h__inf] {\table} \closedcycle;
\addplot [stack plots=y, fill=gray!50, opacity=0.4, draw opacity=0, area legend] table [x=temps, y expr=\thisrow{y_h__sup}-\thisrow{y_h__inf}] {\table} \closedcycle;
% subtract the upper bound so our stack is back at zero
\addplot [stack plots=y, stack dir=minus, forget plot, draw=none] table [x=temps, y=y_h__sup] {\table};
% y_f confidence interval
\addplot [stack plots=y, fill=none, draw=none, forget plot] table [x=temps, y=y_f__inf] {\table} \closedcycle;
\addplot [stack plots=y, fill=gray!50, opacity=0.4, draw opacity=0, area legend] table [x=temps, y expr=\thisrow{y_f__sup}-\thisrow{y_f__inf}] {\table} \closedcycle;
% the line plots (y_h and y_f)
\addplot [stack plots=false, very thick,smooth,blue] table [x=temps, y=y_h] {\table};
\addplot [stack plots=false, very thick,smooth,blue] table [x=temps, y=y_f] {\table};
\end{axis}
\end{tikzpicture}
\end{document}
Best Answer
Not quite an answer but a workaround is using an expression such as:
to subtract a large common number from the dataset. This still uses the math engine of pgf and works for your small test case.
For closely spaced data this still breaks down. The precision of the FPU engine can be shown with
Which prints zero as the change occurs at the 8th digit. So I would recommend using either gnuplot as suggested in the pgfmanual or preprocessing the dataset externally.
Alternatively the
fp
package is more suited to the problem as it operates with fixed point. I do not know how to integrate it with pgfplots.Update:
Whereas with LaTeX the precision is limited it is not a problem to use LuaTeX for the calculation here and the results are quite nice. If run with LuaTeX the following example produces the correct output: