Compute feasibility of the matrix constraints (A) and matrix variables (Y)
Note that only these in pen/bar (non-strict treatment) are considered
because these in the strict barrier are feasible by design.
(This routine is called only during computation so any point must be
feasibile with respect to constraints in log-barrier. Similarly, it cannot
be more infeasible than the matrix penalty parameter.)
TODO
Suggested but not implemented...
nFast = 0 ... no speed up; if infeasible, go back with perturbation
to compute the feasibility up to PRECISION
nFast = 1 ... if infeasible, stop on first infeasible matrix constr.
and don't go back (==> the (in)feasibility is overestimated
up to SCALEUP factor times for this constraint and the other
are not checked --> could be even worse)
nFast = 2 ... if infeasible, don't go back with the perturbation
but at least check all the constraints that there is nothing
worse
how to make pstart&ptol dynamic to save some time...
+ allow how to refine it-if it looks big, tolerance 1e-10 is time wasting
add changing order? start with the worst one from the last step??0001 % Compute feasibility of the matrix constraints (A) and matrix variables (Y) 0002 % Note that only these in pen/bar (non-strict treatment) are considered 0003 % because these in the strict barrier are feasible by design. 0004 % (This routine is called only during computation so any point must be 0005 % feasibile with respect to constraints in log-barrier. Similarly, it cannot 0006 % be more infeasible than the matrix penalty parameter.) 0007 % TODO 0008 % Suggested but not implemented... 0009 % nFast = 0 ... no speed up; if infeasible, go back with perturbation 0010 % to compute the feasibility up to PRECISION 0011 % nFast = 1 ... if infeasible, stop on first infeasible matrix constr. 0012 % and don't go back (==> the (in)feasibility is overestimated 0013 % up to SCALEUP factor times for this constraint and the other 0014 % are not checked --> could be even worse) 0015 % nFast = 2 ... if infeasible, don't go back with the perturbation 0016 % but at least check all the constraints that there is nothing 0017 % worse 0018 % 0019 % how to make pstart&ptol dynamic to save some time... 0020 % + allow how to refine it-if it looks big, tolerance 1e-10 is time wasting 0021 % add changing order? start with the worst one from the last step?? 0022 % 0023 function [mfeas] = feas_ay(obj) 0024 0025 mfeas=0; 0026 pstart=1e-10; 0027 pstop=1e-7; 0028 0029 % check only pen/bar; strict barrier must be feasible by design 0030 for k=obj.Yboxindphi 0031 0032 pkx=obj.PYbox(k); %2*p(sdpdata.Ng+k); 0033 Ykx = obj.Y{obj.Yboxmap(k)}; 0034 M=obj.Yboxshift(k)*speye(size(Ykx)) + obj.Yboxmlt(k)*Ykx; 0035 % so far negative definite ... feasm is posdef checker 0036 %mfeas = feasm(-M, mfeas, Inf, pstart, pstop); 0037 mfeas = feasm(-M, mfeas, pkx, pstart, pstop); 0038 end 0039 0040 for k=obj.Aindphi 0041 0042 pkx=obj.PA(k); % I used to use 2* !!!!!!!! 0043 % TODO need to map the matrix first! - is it correct??? 0044 kuser=obj.Amap(k); 0045 [Akuserx, obj.userdata] = obj.mconfun(obj.x, obj.Y, kuser, obj.userdata); 0046 M = obj.Ashift(k)*speye(size(Akuserx)) + obj.Amlt(k) .* Akuserx; 0047 %mfeas = feasm(-M, mfeas, Inf, pstart, pstop); 0048 mfeas = feasm(-M, mfeas, pkx, pstart, pstop); 0049 end 0050 0051 0052 end 0053 0054