I would like to be able to resample a histograms bins without having access tot he raw data. And just to be clear, by resample, I mean to change the number of bins and still provide a good estimate of the original probabilities of those bins.

I can think of many ways to do this, but having trouble figuring out which is the best method which maintains the same probability in the resulting histogram. The easy one would be if the input histogram X had x bins and the desired result histogram Y has y bins where x = y. This is a simple 1 - 1 sampling of the original bins. The problem forms to me as I decrease y lower than x or increase it above x.

For example: If x = 10 bins and y = 20 bins, it seems like you could simply double each of x's bins so you have Y = { x1, x1, x2, x2, x3, x3, ..., x10, x10 } but this seems like a naive approach as it seems like the 2nd copy of the previous bin should be influenced also by the next bin value y2 = (x1 + x2)/2 for example.

If x = 20 bins and y = 7 bins I can see it isn't fair to simply sample a value based upon a linear interpolation between data point as there might be 3 or 4 points on either side of the sample that should be a part of the probability for the resampled data.

I would also like to consider the possibility that the histogram is contained on the ends, so in the case of measuring water temperature below freezing isn't a likely temp for water nor above boiling for the standard cases. I would like to be able to consider the probability beyond one or both extreme bins to be 0.

Is there a standard algorithm which can be coded in C++/C# or something in pseudo code that I can convert to code for the above re-sampling / re-sizing?

  • $\begingroup$ Is it that you only have access to the histogram or do you have the data but do not want reprocess it? Asking as there is a better starting point than a histogram if you have access to the underlying data. $\endgroup$
    – Edmund
    Dec 28 '19 at 1:58
  • $\begingroup$ The goal here is to only use the bins to resample to a larger or smaller size histogram. The raw data is generally unavailable (or massive in size to reprocess, doesn't matter). My thoughts are around preserving the "probability density" of the original as best I can through resampling. $\endgroup$ Dec 30 '19 at 17:38

You may use Interpolation on the the bins and heights of the HistogramList to produce a smooth PDF for use in ProbabilityDistribution. With a smooth distribution then you need not worry about increasing or decreasing bin sizes.

For example, with data

data = RandomVariate[WeibullDistribution[2, 2], 10^6];

has Histogram

hist = Histogram[data, Automatic, "PDF", PlotRange -> Full]

Mathematica graphics

You have the histogram bins and heights so I'll just use HistogramList to collect them and display with ListStepPlot.

hlData = HistogramList[data, Automatic, "PDF"];
lstep = ListStepPlot[Transpose@{First@hlData, Append[Last@hlData, 0]}, Mesh -> Full]

Mathematica graphics

A smooth InterpolatingFunction can be constructed from the bins and heights with Interpolation.

ifData = Interpolation[Transpose@{First@hlData, Append[Last@hlData, 0]}];
ifPlot = Plot[ifData[x], {x, Sequence @@ First@ifData["Domain"]}, PlotStyle -> Purple]

Mathematica graphics

It can be seen that it matches the bins and heights.

Show[hist, ifPlot, lstep]

Mathematica graphics

We want to use this function as the PDF of a ProbabilityDistribution so that NProbability can be calculated and RandomVariates can be generated.

The function needs to be non-negative over its domain.

NMinimize[ifData[x], {x} ∈ Interval @@ ifData["Domain"]]
{0., {x -> 7.4}}

Minimum value is zero at x equal to 7.4.

Must Integrate to 1 over its domain.

NIntegrate[ifData[x], {x} ∈ Interval @@ ifData["Domain"]] // N

The integral is just shy of 1 but we can ask ProbabilityDistribution to "Normalize" the PDF.

dist = ProbabilityDistribution[ifData[x], {x, Sequence @@ First@ifData["Domain"]}, 
   Method -> "Normalize"];

This distribution can be used in RandomVariate, Probability and other functions of the Random Variables guide.

Probabilities can be calculated

NProbability[0.9 < x < 2.2, x \[Distributed] dist]

Pseudo-random numbers generated

RandomVariate[dist, 5]
{0.494373, 1.16545, 2.94366, 4.06116, 1.72519}

Properties like the SurvivalFunction, CDF, and others can be calculated.

Plot[SurvivalFunction[dist, x], {x, Sequence @@ First@ifData["Domain"]}, 
 PlotStyle -> Purple]

Mathematica graphics


Widening the bins will result in a loss of information so I would not recommend it. However, if you must then Subdivide the InterpolatingFunction "Domain" into the number of bins you required.

downsamples =
    With[{bins = Subdivide[##, numbins] & @@ First@ifData["Domain"]},
      NProbability[#1 < x < #2, x \[Distributed] dist] & @@@ 
       Partition[bins, 2, 1]
    ] /@ Range[36, 12, -6];

Addressing your question in the comment; The total probability in each of the 5 cases is still one.

Total[downsamples[[All, 2]], {2}]
{1., 1., 1., 1., 1.}

However, notice the differences in the PDF histograms.

 Transpose[{First@#, Append[Last@#, 0]}] & /@ downsamples,
 Mesh -> Full,
 PlotRange -> All,
 PlotLegends -> StringTemplate["`` bins"] /@ Range[36, 12, -6]]

Mathematica graphics

The initial histogram had 37 bins. The plot above shows how the information is lost as the bin width is widened (number of bins decreased). I would recommend working with dist instead.

Hope this helps.

  • $\begingroup$ Thanks Edmund, this looks like a good approach, one I was considering but couldn't articulate. The example you show, would this work for significant down sampling. Say to just 3 bins? It seems like the total probability will be way off and loss of the hump near the beginning. Maybe I missed how to handle this as it seems like I need to integrated the PDF for the new bins? $\endgroup$ Jan 1 '20 at 23:41
  • 1
    $\begingroup$ @mazecreator Yes, wider bins will distort shape of the PDF. However, in the example above the ProbabilityDistribution (dist) is a smooth approximation of the population's distribution based on the histogram. Why would you want to discard that level of detail for wider bins? With dist you can generate pseudo-random numbers, calculate probabilities on any interval, construct transformed distributions, and so on. In any case see the update for widening the bins. $\endgroup$
    – Edmund
    Jan 2 '20 at 0:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.