1 ## Author: Paul Kienzle <pkienzle@gmail.com>
2 ## This program is granted to the public domain.
4 ## [y,dy] = polyconf(p,x,s)
6 ## Produce prediction intervals for the fitted y. The vector p
7 ## and structure s are returned from polyfit or wpolyfit. The
8 ## x values are where you want to compute the prediction interval.
10 ## polyconf(...,['ci'|'pi'])
12 ## Produce a confidence interval (range of likely values for the
13 ## mean at x) or a prediction interval (range of likely values
14 ## seen when measuring at x). The prediction interval tells
15 ## you the width of the distribution at x. This should be the same
16 ## regardless of the number of measurements you have for the value
17 ## at x. The confidence interval tells you how well you know the
18 ## mean at x. It should get smaller as you increase the number of
19 ## measurements. Error bars in the physical sciences usually show
20 ## a 1-alpha confidence value of erfc(1/sqrt(2)), representing
21 ## one standandard deviation of uncertainty in the mean.
23 ## polyconf(...,1-alpha)
25 ## Control the width of the interval. If asking for the prediction
26 ## interval 'pi', the default is .05 for the 95% prediction interval.
27 ## If asking for the confidence interval 'ci', the default is
28 ## erfc(1/sqrt(2)) for a one standard deviation confidence interval.
31 ## [p,s] = polyfit(x,y,1);
32 ## xf = linspace(x(1),x(end),150);
33 ## [yf,dyf] = polyconf(p,xf,s,'ci');
34 ## plot(xf,yf,'g-;fit;',xf,yf+dyf,'g.;;',xf,yf-dyf,'g.;;',x,y,'xr;data;');
35 ## plot(x,y-polyval(p,x),';residuals;',xf,dyf,'g-;;',xf,-dyf,'g-;;');
37 function [y,dy] = polyconf(p,x,varargin)
40 for i=1:length(varargin)
42 if isstruct(v), s = v;
43 elseif ischar(v), typestr = v;
44 elseif isscalar(v), alpha = v;
48 if (nargout>1 && (isempty(s)||nargin<3)) || nargin < 2
56 ## For a polynomial fit, x is the set of powers ( x^n ; ... ; 1 ).
59 if columns(s.R) == n, ## fit through origin
60 A = (x(:) * ones (1, n)) .^ (ones (k, 1) * (n:-1:1));
63 A = (x(:) * ones (1, n+1)) .^ (ones (k, 1) * (n:-1:0));
66 [y(:),dy(:)] = confidence(A,p,s,alpha,typestr);
72 %! # data from Hocking, RR, "Methods and Applications of Linear Models"
73 %! temperature=[40;40;40;45;45;45;50;50;50;55;55;55;60;60;60;65;65;65];
74 %! strength=[66.3;64.84;64.36;69.70;66.26;72.06;73.23;71.4;68.85;75.78;72.57;76.64;78.87;77.37;75.94;78.82;77.13;77.09];
75 %! [p,s] = polyfit(temperature,strength,1);
76 %! [y,dy] = polyconf(p,40,s,0.05,'ci');
77 %! assert([y,dy],[66.15396825396826,1.71702862681486],200*eps);
78 %! [y,dy] = polyconf(p,40,s,0.05,'pi');
79 %! assert(dy,4.45345484470743,200*eps);
81 ## [y,dy] = confidence(A,p,s)
83 ## Produce prediction intervals for the fitted y. The vector p
84 ## and structure s are returned from wsolve. The matrix A is
85 ## the set of observation values at which to evaluate the
86 ## confidence interval.
88 ## confidence(...,['ci'|'pi'])
90 ## Produce a confidence interval (range of likely values for the
91 ## mean at x) or a prediction interval (range of likely values
92 ## seen when measuring at x). The prediction interval tells
93 ## you the width of the distribution at x. This should be the same
94 ## regardless of the number of measurements you have for the value
95 ## at x. The confidence interval tells you how well you know the
96 ## mean at x. It should get smaller as you increase the number of
97 ## measurements. Error bars in the physical sciences usually show
98 ## a 1-alpha confidence value of erfc(1/sqrt(2)), representing
99 ## one standandard deviation of uncertainty in the mean.
101 ## confidence(...,1-alpha)
103 ## Control the width of the interval. If asking for the prediction
104 ## interval 'pi', the default is .05 for the 95% prediction interval.
105 ## If asking for the confidence interval 'ci', the default is
106 ## erfc(1/sqrt(2)) for a one standard deviation confidence interval.
108 ## Confidence intervals for linear system are given by:
109 ## x' p +/- sqrt( Finv(1-a,1,df) var(x' p) )
110 ## where for confidence intervals,
111 ## var(x' p) = sigma^2 (x' inv(A'A) x)
112 ## and for prediction intervals,
113 ## var(x' p) = sigma^2 (1 + x' inv(A'A) x)
115 ## Rather than A'A we have R from the QR decomposition of A, but
116 ## R'R equals A'A. Note that R is not upper triangular since we
117 ## have already multiplied it by the permutation matrix, but it
118 ## is invertible. Rather than forming the product R'R which is
119 ## ill-conditioned, we can rewrite x' inv(A'A) x as the equivalent
120 ## x' inv(R) inv(R') x = t t', for t = x' inv(R)
121 ## Since x is a vector, t t' is the inner product sumsq(t).
122 ## Note that LAPACK allows us to do this simultaneously for many
123 ## different x using sqrt(sumsq(X/R,2)), with each x on a different row.
125 ## Note: sqrt(F(1-a;1,df)) = T(1-a/2;df)
127 ## For non-linear systems, use x = dy/dp and ignore the y output.
128 function [y,dy] = confidence(A,p,S,alpha,typestr)
129 if nargin < 4, alpha = []; end
130 if nargin < 5, typestr = 'ci'; end
133 case 'ci', pred = 0; default_alpha=erfc(1/sqrt(2));
134 case 'pi', pred = 1; default_alpha=0.05;
135 otherwise, error("use 'ci' or 'pi' for interval type");
137 if isempty(alpha), alpha = default_alpha; end
138 s = tinv(1-alpha/2,S.df)*S.normr/sqrt(S.df);
139 dy = s*sqrt(pred+sumsq(A/S.R,2));