Re: Empirical CDF and InterpolatingFunction
- To: mathgroup at smc.vnet.net
- Subject: [mg36613] Re: Empirical CDF and InterpolatingFunction
- From: mark at markfisher.net (Mark Fisher)
- Date: Fri, 13 Sep 2002 23:33:13 -0400 (EDT)
- References: <alnv9e$7op$1@smc.vnet.net>
- Sender: owner-wri-mathgroup at wolfram.com
Daniel Lichtblau made two suggestions that allow one to use Interpolation the way I wanted. First, to make the resulting function right-continuous, change the sign twice. Second, to make the end point return the correct value, add an extra (phantom) observation (with an extra (irrelevant) value). (The phantom observation is made at the high end because of the sign reversals.) Here's the code I cooked up based on his suggestions: MakeEmpiricalCDF::usage = "MakeEmpiricalCDF[list] returns a function that evaluates the empirical CDF given the observations in the list. The function is defined on the entire real line." MakeEmpiricalCDF[list_?(VectorQ[#, NumericQ]&)] := Module[{n, s, a, r, idata}, n = Length[list]; s = Sort[list]; a = Append[s, s[[-1]] + 1]; (* phantom obs. *) r = Range[1/n, 1 + 1/n, 1/n]; (* phantom value 1 + 1/n *) idata = Last /@ Split[Transpose[{-a, r}], #1[[1]] == #2[[1]]&]; (* "-a" is the first sign change *) Block[{x}, Function @@ {x, Which @@ { x < s[[ 1]], 0., x > s[[-1]], 1., True, Interpolation[idata, InterpolationOrder -> 0][-x] (* "-x" is the second sign change *) }}] ] The construction "Last /@ Split[ ... ]" accounts for duplicate values. Here are two examples. Needs["Statistics`ContinuousDistributions`"] list1 = RandomArray[NormalDistribution[0, 1], 100]; f1 = MakeEmpiricalCDF[list1]; Plot[f1[x], {x, -4, 4}] list2 = Table[Random[Integer, {1, 10}], {10}]; f2 = MakeEmpiricalCDF[list2]; Plot[f2[x], {x, 0, 11}] --Mark mark at markfisher.net (Mark Fisher) wrote in message news:<alnv9e$7op$1 at smc.vnet.net>... > I'm trying to write a fast empirical cummulative distribution function > (CDF). Empirical CDFs are step functions that can be expressed in > terms of a Which statement. For example, given the list of > observations {1, 2, 3}, > > f = Which[# < 1, 0, # < 2, 1/3, # < 3, 2/3, True, 1]& > > is the empirical CDF. Note that f /@ {1, 2, 3} returns {1/3, 2/3, 1} > and f is continuous from the right. > > When the number of observations is large, the Which statement > evaluates fairly slowly (even if it has been Compiled). Since > InterpolationFunction evaluates so much faster in general, I've tried > to use Interpolation with InterpolationOrder -> 0. The problem is that > the resulting InterpolatingFunction doesn't behave the way (I think) > it ought to. For example, let > > g = Interpolation[{{1, 1/3}, {2, 2/3}, {3, 1}}, InterpolationOrder -> > 0] > > Then, g /@ {1, 2, 3} returns {2/3, 2/3, 1} instead of {1/3, 2/3, 1}. > In addition, g is continuous from the left rather than from the right. > > Obviously I am not aware of the considerations that went into > determining the behavior of InterpolationFunction when > InterpolationOrder -> 0. > > So I have two questions: > > (1) Does anyone have any opinions about how InterpolatingFunction > ought to behave with InterpolationOrder -> 0? > > (2) Does anyone have a faster way to evaluate an empirical CDF than a > compiled Which function? > > By the way, here's my current version: > > CompileEmpiricalCDF[list_?(VectorQ[#, NumericQ] &)] := > Block[{x}, Compile[{{x, _Real}}, Evaluate[ > Which @@ Flatten[ > Append[ > Transpose[{ > Thread[x < Sort[list]], > Range[0, 1 - 1/#, 1/#] & @ Length[list] > }], > {True, 1}]] > ]]] > > --Mark