rpsmooth.m

这个帖子中我想讨论的是移动窗口多项式最小二乘拟和平滑方法

function f=rpsmooth(x,nambda)

%   Roughness Penalty Smoothing.
%    
%   f=rpsmooth(x,nambda);
%
%   x ------- Input data for smoothing. It must be a vector.
%
%   nambda -- The smoothing parameter. It measures the rate of exchange
%             between fit to the data. As nambda become larger and larger,
%             functions which are not linear get more penalty,vice versa.
%             You can input one number for nambda, if you have some
%             knowladge about your data and if not, you can input a vector
%             contain two number, which input a scale to choose the best
%             parameter for your data smoothing by leave-one-out
%             Cross-Validation.
%
%   DNP. 2007.12.25
%   Merry Christmas!

tic
f=[];
if nargin<1
    error('Not enough inputs!');
%     return
elseif nargin>2
    error('Too many inputs!');
%     return
elseif nargin==1     % If you have not input the parameter, the default changable ranger of nambda is
    nambda_low=1;    % from 1:100, with the change step is 10.
    step_wise=10;
    nambda_up=100;
elseif nargin==2
    if (length(nambda)==1 & nambda > 0) | (length(nambda)==2 & isequal(nambda(1),nambda(2)))
        nambda_low=nambda(1);
        step_wise=[];
        nambda_up=nambda(1);
    elseif nambda==0
        f=x;
        return
    elseif any(nambda<0)
        error('The bandwidth must be nonnegative.')
%         return
    elseif length(nambda)>2
        error('The element in parameter vector must not more then 2.')
%         return
    else
        nambda_low=min(nambda);                
        step_wise=(max(nambda)-min(nambda))/20; % The parameter change step is (max(nambda)-min(nambda))/20
        nambda_up=max(nambda);
    end
end

[r c]=size(x);
if ~any([r c]==1)
    error('Input data must be a vector!');
%     return
elseif r==1
    x=x';
end


% *****************************************************************************
% ... Creat Matrix Q and R for calculating K. ................

Q=[];
R=[];
n=max(r,c);
h=ones(n,1);         % The width of intervals.
for i=1:n-2
    Q(i:i+2,i)=[1;-2;1];
%     Q(i,i)=1/h(i);
%     Q(i+1,i)=-1/h(i)-1/h(i+1);
%     Q(i+2,i)=1/h(i+1);
    R(i,i)=(h(i)+h(i+1))/3;                      
    if i==n-2
        break
    else
        R(i,i+1)=h(i+1)/6;
        R(i+1,i)=R(i,i+1);
    end
end

K=Q*inv(R)*Q';
% toc

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% ....... Leave-one-out Cross Validation. .........

rescv=[]; % Store the CV value.
para=[];  % Store the paramters.
j=0;
for alfa=nambda_low:step_wise:nambda_up
    j=j+1;
    para(j)=alfa;
    A=inv(eye(n)+alfa*K);
%     rescv(j)=0;
    f=A*x;
    rescv(j)=sum((x-f)./(1-diag(A)).^2);
%     for i=1:n
%         rescv(j)=rescv(j)+(x(i)-f(i))/(1-A(i,i))^2;
%     end
end

if length(rescv)>1
    subplot(2,1,1);
    plot(rescv);
    [rescvmin index]=min(rescv); 
    A=inv(eye(n)+para(index)*K);
    f=A*x;
    subplot(2,1,2);
    plot(f)  
    t=toc
else
    A=inv(eye(n)+nambda*K);
    f=A*x;
    plot(f)  
    t=toc
end