- Dev - dynare.org

Re: [DynareDev] [Trac] #28: Define a syntax for bytecode generation and the various Newton solvers (for steady and simul)
by Trac Dynare 28 Aug '09

28 Aug '09

#28: Define a syntax for bytecode generation and the various Newton solvers (for steady and simul) ---------------------------+------------------------------------------------ Reporter: sebastien | Owner: ferhat Type: enhancement | Status: accepted Priority: critical | Milestone: 4.1 Component: Preprocessor | Version: Resolution: | Keywords: ---------------------------+------------------------------------------------ Changes (by ferhat): * status: new => accepted -- Ticket URL: <https://www.dynare.org/trac/ticket/28#comment:1> Dynare <http://www.dynare.org> The Dynare project

1 0

Re: [DynareDev] [Trac] #21: Add partial information
by Trac Dynare 27 Aug '09

27 Aug '09

#21: Add partial information ---------------------------+------------------------------------------------ Reporter: sebastien | Owner: george Type: feature | Status: accepted Priority: major | Milestone: 4.2-alpha2 Component: Core M-files | Version: Resolution: | Keywords: ---------------------------+------------------------------------------------ Changes (by george): * owner: => george * status: new => accepted -- Ticket URL: <https://www.dynare.org/trac/ticket/21#comment:1> Dynare <http://www.dynare.org> The Dynare project

1 0

Trac for Dynare
by Sébastien Villemot 27 Aug '09

27 Aug '09

Hi, I have just installed a Trac for Dynare, available at the adresse: http://www.dynare.org/trac Trac is a project management software. It helps developers to deal with bugs, features to be added, deadlines, ... The basic item is a "ticket". A ticket is typically a bug report, or a feature to be added. The ticket is closed once the bug is fixed, or the feature is implemented. Users can spawn a discussion on a ticket. A ticket can be assigned: * a priority level * a type (bug, feature...) * a component (the part of Dynare to which it relates: the preprocessor, the core M-files, such DLL, the doc...) * a milestone (to be done for Dynare 4.1, for Dynare 4.2) * an owner (the person who is responsible for fixing the bug or adding the feature) * a version (4.0.1, 4.0.2, ...) (only useful for bug reports) I have already added some tickets, corresponding to the main release goals for Dynare 4.1, which is planned for the end of september. See "View tickets". The "Roadmap" lists the future milestones, currently 4.1 and 4.2. For each milestone, there is a deadline, and also the percentage of the tickets assigned to that milestone that have been closed. You can also browse the subversion repository through "Browse source". The "Timeline" lists all the recent changes (including SVN commits). You can syndicate to this timeline in your favorite RSS aggregator. Finally there is a wiki, which is a little bit redundant with the already existing DynareWiki. We may decide not to use it, or we may use it only for development specific stuff, and dedicate the other wiki to public documentation and other Dynare-related stuff. In order to add/modify tickets or wiki pages you need to be logged in. The usernames and password are exactly the same than for the SVN repository of Dynare. Please don't hesitate to add new tickets for the bugs you discover, or for the features you plan to add. Also please take ownership of existing tickets if you plan to deal with them. At a later date we may allow user to open tickets for bug reports, if we think we can avoid useless or wrong bug reports. Best, -- Sébastien Villemot

1 2

Re: [DynareDev] [Trac] #29: bug with varexo_det
by Trac Dynare 27 Aug '09

27 Aug '09

#29: bug with varexo_det ---------------------------+------------------------------------------------ Reporter: michel | Owner: michel Type: bug | Status: accepted Priority: major | Milestone: Component: Core M-files | Version: 4.0.4 Resolution: | Keywords: ---------------------------+------------------------------------------------ Changes (by sebastien): * version: => 4.0.4 -- Ticket URL: <https://www.dynare.org/trac/ticket/29#comment:3> Dynare <http://www.dynare.org> The Dynare project

1 0

MATLAB bug on Linux/64
by Sébastien Villemot 19 Aug '09

19 Aug '09

Hi everyone, I am hitting a MATLAB bug which is driving me crazy. I attach two files: plouf.m and plouf_static.m. They are derived from preprocessor-generated files with the EAGLE model. The script "plouf.m" calls the function "plouf_static.m". When I run "plouf.m" on a Linux/64 machine, MATLAB crashes with a bus error (tested on Karaba and Hal, with Matlab 7.5, 7.6 and 7.8). This does not happen on Linux/32 or Windows/32. I have not tested other platforms. As can be seen from the displayed marks which I have inserted, MATLAB crashes around line 131 of plouf_static.m. More precisely, it seems to crash just after having computed the expression there. I can't understand exactly what's going on. Can you reproduce this? Note that I have observed similar crashes with other big models on Karaba. Since so far I cannot precisely pinpoint what triggers the bug, I don't know how to make a workaround. So for the moment, for the development of EAGLE, I am going to use Matlab/32 in a virtual machine on Karaba. Probably we should send a report to Mathworks. Best, -- Sébastien Villemot

2 5

model derivatives with respect to parameters
by Michel Juillard 19 Aug '09

19 Aug '09

Dear Nikolay, we have started to work on producing the derivatives with respect to parameters. We are planing to create a new function, that will return a 3-dimensional array, by equation, variable and parameter. You could then select one parameter and squeeze the array in the parameter dimension to obtain a matrix of Jacobian derivatives with respect to a given parameter. Do you see any problem with this implementation? We should be able to provide in about two weeks. All the best, Michel

4 6

Re: [DynareDev] lyapunov kalman filter options
by Michel Juillard 30 Jul '09

30 Jul '09

G. Perendia wrote: > Dear Michel > > Should we include Matlab disclyap_fast.m into (next major release of) Dynare > Matlab suite as an alternative or replacement for lyapunov_symm.m and shall > we/I then also: > > we should just replace lyapunov_symm by (renamed) disclyap_fast.m > a) change Matlab disclyap_fast.m to accept tolerance as parameter and > include it into Dynare Matlab suite, > yes > b) change its name, > yes > d) prepare and include C++ disclyap_fast_dll.dll as separate module, again > as alternative to lyapunov_symm.m, > no, let's not multiply the DLLs called from Matlab. We will use the C code as part of a big DLL that performs all the estimation code > d) change its name too, > e) change DsgeLikelihood.m to use either Matlab and/or C++ disclyap_fast.m > > This is a bit complicated. Dsgelikelihood (and dsgelikelihood_hh) calls lyapunov_symm for stationary models. kalman_algo = 1 or 2 Equivalent code is present in dsgelikelihood.m for nonstationary models kalman_algo = 3 or 4 Best Michel > Best regards > > George > ----- Original Message ----- > From: "Michel Juillard" <michel.juillard(a)ens.fr> > To: "List for Dynare developers" <dev(a)dynare.org> > Sent: Wednesday, July 29, 2009 7:36 PM > Subject: Re: [DynareDev] lyapunov kalman filter options > > > Sounds good, thanks > > Michel > > G. Perendia wrote: > >> Dear Michel (et al) >> >> On my slower laptop I run few tests and got timings roughly on lines of >> > the > >> following timings for an 8x8 T matrix and 10,000 loops: >> >> Matlab: >> >> lyapunov_symm.m time = 19.2270 s >> Pearlman Matlab disclyap_fast.m time = 13.0280 s >> >> New C++ disclyap_fast DLL: >> >> disclyap_fast_dll_time = 7.8010 sec for tolerance= 1e-16 >> disclyap_fast_dll_time = 6.5600 sec for tolerance= 1e-12 >> >> Best regards >> >> George >> >> ----- Original Message ----- >> From: "Michel Juillard" <michel.juillard(a)ens.fr> >> To: "G. Perendia" <george(a)perendia.orangehome.co.uk>; "List for Dynare >> developers" <dev(a)dynare.org> >> Sent: Wednesday, July 29, 2009 10:33 AM >> Subject: Re: [DynareDev] lyapunov kalman filter options >> >> >> Tanks for the update. >> The false value when A contains unit root is a fault of design. The >> information on which columns are useable is in the list of stationary >> variables, >> >> Best >> >> Michel >> >> G. Perendia wrote: >> >> >>> Dear Michel and Stephane >>> >>> Re the below warning: the results from lyapunov_symm.m seem now correct - >>> >>> >> It >> >> >>> seems that my Matlab variables memory was corrupted and after a reboot >>> > and > >>> restart it seems ok now - sorry for the alarm. >>> >>> It however reports (an incorrect ) X=[0's] instead of indeterminate or >>> >>> >> NaN >> >> >>> if I pass a=I. >>> >>> PS: and I still need to test and debug new C++lyapunov because it may >>> > have > >>> been causing the variable corruption. >>> >>> >>> Best regards >>> >>> George >>> >>> ----- Original Message ----- >>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>> Cc: "List for Dynare developers" <dev(a)dynare.org> >>> Sent: Wednesday, July 29, 2009 9:27 AM >>> Subject: lyapunov kalman filter options >>> >>> >>> >>> >>> >>>> I would use lyapunov_tol = 1e-12 >>>> (in the example disclyap_fat.m, the criterium is 1e-16) >>>> and >>>> kalman_tol = 1e-12 >>>> >>>> Both options should be accessible to users and obtained from >>>> options_. The initialization should take place in >>>> > global_initialization.m > >>>> Best >>>> >>>> Michel >>>> >>>> >>>> >>>> >>>> >>>>> Errata: >>>>> 1) it is options_.lyapunov_complex_threshold=1e-15 that is passed in >>>>> >>>>> >>>>> >>> calls >>> >>> >>> >>>>> to the >>>>> existing lyapunov_symm.m and I can use that as tolerance instead the >>>>> original 1e-6 if needed. >>>>> >>>>> Warning: >>>>> >>>>> If I pass a=T=[0] and b=RQR'~0 to lyapunov_symm.m, I get result Pstar >>>>> >>>>> >>>>> >>> very >>> >>> >>> >>>>> different from b although the solution to Lyapunov x-a*x*a' = b is >>>>> >>>>> >>>>> >>> x=b=RQR' >>> >>> >>> >>>>> and both, the new C++ and old matlab disklyap_fast return X=b=RQR as >>>>> solution!! >>>>> >>>>> Best regards >>>>> >>>>> George >>>>> >>>>> ----- Original Message ----- >>>>> From: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>>>> To: "Michel Juillard" <michel.juillard(a)ens.fr>; "Stéphane Adjemian" >>>>> <stephane.adjemian(a)gmail.com> >>>>> Sent: Wednesday, July 29, 2009 7:46 AM >>>>> Subject: Re: measurement errors + steady KF >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>>> Hi Michel, Stephane >>>>>> >>>>>> What tolerances should we use to pass: >>>>>> >>>>>> 1)for the new fast iterative C++ Lyapunov solution convergence check >>>>>> >>>>>> if max( max( abs( P1 - P0 ) ) )>tol >>>>>> >>>>>> The original Joe Pearlman's matlab code used 1e-6. which is same as >>>>>> >>>>>> >>>>>> >>> Dynare >>> >>> >>> >>>>>> default ricatti_tol (also used for a similar purpose in calls to the >>>>>> existing lyapunov_symm.m.). So, shall we use the riccati_tol or set a >>>>>> separate, lyapunov_tol one at the same default level? >>>>>> >>>>>> 2) and, as per below, for the Old style DiffuseLikelihood.m like >>>>>> > Pstar > >>>>>> convergence/differences, that has been built into the C++KF i.e.: >>>>>> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >>>>>> where crit is options_.kalman_tol = 1.0000e-012 - I guess we can >>>>>> >>>>>> >>>>>> >>>>>> >>>>> continue >>>>> >>>>> >>>>> >>>>> >>>>>> to use the same >>>>>> though the new likelihood/kalman_filter*.m suite uses riccati_tol >>>>>> >>>>>> >>>>>> >>> instead >>> >>> >>> >>>>>> but for F convergence check. >>>>>> >>>>>> Best regards >>>>>> >>>>>> George >>>>>> Mob. +44(0)7951415480 >>>>>> artilogica(a)btconnect.com >>>>>> >>>>>> ----- Original Message ----- >>>>>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>>>>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk>; "Stéphane >>>>>> >>>>>> >>>>>> >>> Adjemian" >>> >>> >>> >>>>>> <stephane.adjemian(a)gmail.com> >>>>>> Sent: Tuesday, July 14, 2009 11:47 PM >>>>>> Subject: Re: measurement errors + steady KF >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>> I don't know. This is something that we need to discuss with >>>>>>> > Stephane. > >>>>>>> All the best, >>>>>>> >>>>>>> Michel >>>>>>> >>>>>>> G. Perendia wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>>> Thanks Michel >>>>>>>> >>>>>>>> I am finishing algorithm for checking if KF P (and/or K) are steady >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>> for >>>>> >>>>> >>>>> >>>>> >>>>>> the >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>>> C++KF. >>>>>>>> >>>>>>>> Old DiffuseLikelihood.m uses Pstar differences, i.e.: >>>>>>>> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >>>>>>>> where crit is options_.kalman_tol = 1.0000e-012 >>>>>>>> >>>>>>>> whilst the new kalman_filter.m uses K to check: >>>>>>>> notsteady = max(max(abs(K-oldK))) > riccati_tol; >>>>>>>> where >>>>>>>> K = Pstar*Z'*invF; >>>>>>>> and options_.riccati_tol= 1.0000e-006 >>>>>>>> >>>>>>>> Which tolerance (and check) should I use for C++KF? >>>>>>>> I have already built a system using Pstar -oldPstar and that works >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>> well >>>>> >>>>> >>>>> >>>>> >>>>>> but >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>>> question is which tolerance to use since riccati seems sufficient >>>>>>>> > for > >>>>>>>> >>>>>>>> >>>>> P >>>>> >>>>> >>>>> >>>>> >>>>>>>> difference too. >>>>>>>> >>>>>>>> Best regards >>>>>>>> >>>>>>>> George >>>>>>>> Mob. +44(0)7951415480 >>>>>>>> artilogica(a)btconnect.com >>>>>>>> >>>>>>>> ----- Original Message ----- >>>>>>>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>>>>>>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>>>>>>> Sent: Tuesday, July 14, 2009 6:57 AM >>>>>>>> Subject: Re: measurement errors >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>>> Hi George, >>>>>>>>> >>>>>>>>> it is like for exogenous variables. If GDP is an endogenous >>>>>>>>> > variable > >>>>>>>>> estimated_params; >>>>>>>>> stderr GDP, uniform_pdf,,,0,1; >>>>>>>>> .... >>>>>>>>> end; >>>>>>>>> >>>>>>>>> will introduce a measurement error on GDP. >>>>>>>>> >>>>>>>>> Best >>>>>>>>> >>>>>>>>> Michel >>>>>>>>> G. Perendia wrote: >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>>> Dear Michel >>>>>>>>>> >>>>>>>>>> For testing Part Info (and the other) Kalman filters, I would need >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>> to >>>>> >>>>> >>>>> >>>>> >>>>>>>>>> set situation with measurement errors so that H~=0. >>>>>>>>>> >>>>>>>>>> Can you please let me know how one specifies measurement errors >>>>>>>>>> in Dynare for estimation so that H ~ =0? >>>>>>>>>> Best regards >>>>>>>>>> >>>>>>>>>> George >>>>>>>>>> artilogica(a)btconnect.com <mailto:artilogica@btconnect.com> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>> a >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>> >>>>> >>> >>> >> _______________________________________________ >> Dev mailing list >> Dev(a)dynare.org >> http://www.dynare.org/cgi-bin/mailman/listinfo/dev >> >> >> >> _______________________________________________ >> Dev mailing list >> Dev(a)dynare.org >> http://www.dynare.org/cgi-bin/mailman/listinfo/dev >> >> >> > > > _______________________________________________ > Dev mailing list > Dev(a)dynare.org > http://www.dynare.org/cgi-bin/mailman/listinfo/dev > > > >

2 2

Re: [DynareDev] lyapunov kalman filter options
by Michel Juillard 29 Jul '09

29 Jul '09

Tanks for the update. The false value when A contains unit root is a fault of design. The information on which columns are useable is in the list of stationary variables, Best Michel G. Perendia wrote: > Dear Michel and Stephane > > Re the below warning: the results from lyapunov_symm.m seem now correct - It > seems that my Matlab variables memory was corrupted and after a reboot and > restart it seems ok now - sorry for the alarm. > > It however reports (an incorrect ) X=[0's] instead of indeterminate or NaN > if I pass a=I. > > PS: and I still need to test and debug new C++lyapunov because it may have > been causing the variable corruption. > > > Best regards > > George > Mob. +44(0)7951415480 > artilogica(a)btconnect.com > > ----- Original Message ----- > From: "Michel Juillard" <michel.juillard(a)ens.fr> > To: "G. Perendia" <george(a)perendia.orangehome.co.uk> > Cc: "List for Dynare developers" <dev(a)dynare.org> > Sent: Wednesday, July 29, 2009 9:27 AM > Subject: lyapunov kalman filter options > > > >> I would use lyapunov_tol = 1e-12 >> (in the example disclyap_fat.m, the criterium is 1e-16) >> and >> kalman_tol = 1e-12 >> >> Both options should be accessible to users and obtained from >> options_. The initialization should take place in global_initialization.m >> >> Best >> >> Michel >> >> >> >>> Errata: >>> 1) it is options_.lyapunov_complex_threshold=1e-15 that is passed in >>> > calls > >>> to the >>> existing lyapunov_symm.m and I can use that as tolerance instead the >>> original 1e-6 if needed. >>> >>> Warning: >>> >>> If I pass a=T=[0] and b=RQR'~0 to lyapunov_symm.m, I get result Pstar >>> > very > >>> different from b although the solution to Lyapunov x-a*x*a' = b is >>> > x=b=RQR' > >>> and both, the new C++ and old matlab disklyap_fast return X=b=RQR as >>> solution!! >>> >>> Best regards >>> >>> George >>> Mob. +44(0)7951415480 >>> artilogica(a)btconnect.com >>> >>> ----- Original Message ----- >>> From: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>> To: "Michel Juillard" <michel.juillard(a)ens.fr>; "Stéphane Adjemian" >>> <stephane.adjemian(a)gmail.com> >>> Sent: Wednesday, July 29, 2009 7:46 AM >>> Subject: Re: measurement errors + steady KF >>> >>> >>> >>> >>>> Hi Michel, Stephane >>>> >>>> What tolerances should we use to pass: >>>> >>>> 1)for the new fast iterative C++ Lyapunov solution convergence check >>>> >>>> if max( max( abs( P1 - P0 ) ) )>tol >>>> >>>> The original Joe Pearlman's matlab code used 1e-6. which is same as >>>> > Dynare > >>>> default ricatti_tol (also used for a similar purpose in calls to the >>>> existing lyapunov_symm.m.). So, shall we use the riccati_tol or set a >>>> separate, lyapunov_tol one at the same default level? >>>> >>>> 2) and, as per below, for the Old style DiffuseLikelihood.m like Pstar >>>> convergence/differences, that has been built into the C++KF i.e.: >>>> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >>>> where crit is options_.kalman_tol = 1.0000e-012 - I guess we can >>>> >>>> >>> continue >>> >>> >>>> to use the same >>>> though the new likelihood/kalman_filter*.m suite uses riccati_tol >>>> > instead > >>>> but for F convergence check. >>>> >>>> Best regards >>>> >>>> George >>>> Mob. +44(0)7951415480 >>>> artilogica(a)btconnect.com >>>> >>>> ----- Original Message ----- >>>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk>; "Stéphane >>>> > Adjemian" > >>>> <stephane.adjemian(a)gmail.com> >>>> Sent: Tuesday, July 14, 2009 11:47 PM >>>> Subject: Re: measurement errors + steady KF >>>> >>>> >>>> >>>> >>>>> I don't know. This is something that we need to discuss with Stephane. >>>>> >>>>> All the best, >>>>> >>>>> Michel >>>>> >>>>> G. Perendia wrote: >>>>> >>>>> >>>>>> Thanks Michel >>>>>> >>>>>> I am finishing algorithm for checking if KF P (and/or K) are steady >>>>>> >>>>>> >>> for >>> >>> >>>> the >>>> >>>> >>>>>> C++KF. >>>>>> >>>>>> Old DiffuseLikelihood.m uses Pstar differences, i.e.: >>>>>> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >>>>>> where crit is options_.kalman_tol = 1.0000e-012 >>>>>> >>>>>> whilst the new kalman_filter.m uses K to check: >>>>>> notsteady = max(max(abs(K-oldK))) > riccati_tol; >>>>>> where >>>>>> K = Pstar*Z'*invF; >>>>>> and options_.riccati_tol= 1.0000e-006 >>>>>> >>>>>> Which tolerance (and check) should I use for C++KF? >>>>>> I have already built a system using Pstar -oldPstar and that works >>>>>> >>>>>> >>> well >>> >>> >>>> but >>>> >>>> >>>>>> question is which tolerance to use since riccati seems sufficient for >>>>>> >>>>>> >>> P >>> >>> >>>>>> difference too. >>>>>> >>>>>> Best regards >>>>>> >>>>>> George >>>>>> Mob. +44(0)7951415480 >>>>>> artilogica(a)btconnect.com >>>>>> >>>>>> ----- Original Message ----- >>>>>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>>>>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>>>>> Sent: Tuesday, July 14, 2009 6:57 AM >>>>>> Subject: Re: measurement errors >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>>>> Hi George, >>>>>>> >>>>>>> it is like for exogenous variables. If GDP is an endogenous variable >>>>>>> >>>>>>> estimated_params; >>>>>>> stderr GDP, uniform_pdf,,,0,1; >>>>>>> .... >>>>>>> end; >>>>>>> >>>>>>> will introduce a measurement error on GDP. >>>>>>> >>>>>>> Best >>>>>>> >>>>>>> Michel >>>>>>> G. Perendia wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>>> Dear Michel >>>>>>>> >>>>>>>> For testing Part Info (and the other) Kalman filters, I would need >>>>>>>> >>>>>>>> >>> to >>> >>> >>>>>>>> set situation with measurement errors so that H~=0. >>>>>>>> >>>>>>>> Can you please let me know how one specifies measurement errors >>>>>>>> in Dynare for estimation so that H ~ =0? >>>>>>>> Best regards >>>>>>>> >>>>>>>> George >>>>>>>> Mob. +44(0)7951415480 >>>>>>>> artilogica(a)btconnect.com <mailto:artilogica@btconnect.com> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>> >>>>>> >>>>> a >>>>> >>>>> >>>>> >>> >>> >>> >>> >> > > > >

2 2

lyapunov kalman filter options
by Michel Juillard 29 Jul '09

29 Jul '09

I would use lyapunov_tol = 1e-12 (in the example disclyap_fat.m, the criterium is 1e-16) and kalman_tol = 1e-12 Both options should be accessible to users and obtained from options_. The initialization should take place in global_initialization.m Best Michel > Errata: > 1) it is options_.lyapunov_complex_threshold=1e-15 that is passed in calls > to the > existing lyapunov_symm.m and I can use that as tolerance instead the > original 1e-6 if needed. > > Warning: > > If I pass a=T=[0] and b=RQR'~0 to lyapunov_symm.m, I get result Pstar very > different from b although the solution to Lyapunov x-a*x*a' = b is x=b=RQR' > and both, the new C++ and old matlab disklyap_fast return X=b=RQR as > solution!! > > Best regards > > George > Mob. +44(0)7951415480 > artilogica(a)btconnect.com > > ----- Original Message ----- > From: "G. Perendia" <george(a)perendia.orangehome.co.uk> > To: "Michel Juillard" <michel.juillard(a)ens.fr>; "Stéphane Adjemian" > <stephane.adjemian(a)gmail.com> > Sent: Wednesday, July 29, 2009 7:46 AM > Subject: Re: measurement errors + steady KF > > > >> Hi Michel, Stephane >> >> What tolerances should we use to pass: >> >> 1)for the new fast iterative C++ Lyapunov solution convergence check >> >> if max( max( abs( P1 - P0 ) ) )>tol >> >> The original Joe Pearlman's matlab code used 1e-6. which is same as Dynare >> default ricatti_tol (also used for a similar purpose in calls to the >> existing lyapunov_symm.m.). So, shall we use the riccati_tol or set a >> separate, lyapunov_tol one at the same default level? >> >> 2) and, as per below, for the Old style DiffuseLikelihood.m like Pstar >> convergence/differences, that has been built into the C++KF i.e.: >> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >> where crit is options_.kalman_tol = 1.0000e-012 - I guess we can >> > continue > >> to use the same >> though the new likelihood/kalman_filter*.m suite uses riccati_tol instead >> but for F convergence check. >> >> Best regards >> >> George >> Mob. +44(0)7951415480 >> artilogica(a)btconnect.com >> >> ----- Original Message ----- >> From: "Michel Juillard" <michel.juillard(a)ens.fr> >> To: "G. Perendia" <george(a)perendia.orangehome.co.uk>; "Stéphane Adjemian" >> <stephane.adjemian(a)gmail.com> >> Sent: Tuesday, July 14, 2009 11:47 PM >> Subject: Re: measurement errors + steady KF >> >> >> >>> I don't know. This is something that we need to discuss with Stephane. >>> >>> All the best, >>> >>> Michel >>> >>> G. Perendia wrote: >>> >>>> Thanks Michel >>>> >>>> I am finishing algorithm for checking if KF P (and/or K) are steady >>>> > for > >> the >> >>>> C++KF. >>>> >>>> Old DiffuseLikelihood.m uses Pstar differences, i.e.: >>>> notsteady = ~(max(max(abs(Pstar-oldPstar)))<crit); >>>> where crit is options_.kalman_tol = 1.0000e-012 >>>> >>>> whilst the new kalman_filter.m uses K to check: >>>> notsteady = max(max(abs(K-oldK))) > riccati_tol; >>>> where >>>> K = Pstar*Z'*invF; >>>> and options_.riccati_tol= 1.0000e-006 >>>> >>>> Which tolerance (and check) should I use for C++KF? >>>> I have already built a system using Pstar -oldPstar and that works >>>> > well > >> but >> >>>> question is which tolerance to use since riccati seems sufficient for >>>> > P > >>>> difference too. >>>> >>>> Best regards >>>> >>>> George >>>> Mob. +44(0)7951415480 >>>> artilogica(a)btconnect.com >>>> >>>> ----- Original Message ----- >>>> From: "Michel Juillard" <michel.juillard(a)ens.fr> >>>> To: "G. Perendia" <george(a)perendia.orangehome.co.uk> >>>> Sent: Tuesday, July 14, 2009 6:57 AM >>>> Subject: Re: measurement errors >>>> >>>> >>>> >>>> >>>>> Hi George, >>>>> >>>>> it is like for exogenous variables. If GDP is an endogenous variable >>>>> >>>>> estimated_params; >>>>> stderr GDP, uniform_pdf,,,0,1; >>>>> .... >>>>> end; >>>>> >>>>> will introduce a measurement error on GDP. >>>>> >>>>> Best >>>>> >>>>> Michel >>>>> G. Perendia wrote: >>>>> >>>>> >>>>>> Dear Michel >>>>>> >>>>>> For testing Part Info (and the other) Kalman filters, I would need >>>>>> > to > >>>>>> set situation with measurement errors so that H~=0. >>>>>> >>>>>> Can you please let me know how one specifies measurement errors >>>>>> in Dynare for estimation so that H ~ =0? >>>>>> Best regards >>>>>> >>>>>> George >>>>>> Mob. +44(0)7951415480 >>>>>> artilogica(a)btconnect.com <mailto:artilogica@btconnect.com> >>>>>> >>>>>> >>>> >>>> >>>> >>> a >>> >>> > > > > >

1 0

next step
by Michel Juillard 23 Jul '09

23 Jul '09

Dear George, the next step in the C implementation of the estimation routine is to have a compiled version of the Lyapunov equation solver. The corresponding Matlab code is in lyapunov_symm.m or around lines 185-242 of DsgeLikelihood.m. I probably got the idea of the algorithm in "Numerical Recepees in C++" or "Matrix Computations", this last one by Golub and van Loan. lyapunov_symm.m returns the covariance matrix (and the list of stationary variables) while in DsgeLikelihood, we keep the Schur decomposition of the state space transition matrix in order to have a quasi-triangular transition matrix in the Kalman filter (matrix T) and the Schur vectors to build matrix Z. Currently we only use it for the diffuse filter, but there is no reason not to use for the regular filter if it is more efficient. I see two ways to implement it. Write it from scratch in C++ using direct calls to the necessary Lapack functions and/or use RECSY http://www.cs.umu.se/~isak/recsy/ (we need of course to evaluate it first). The problem/advantage of RECSY is that it uses parallel computing via OpenMP, but currently (as of 2008 versions) MKL distributed with Matlab is incompatible with OpenMP (see the discussion lead by Stephane Adjemian a few months ago). My feeling is that building the estimation DLL will be easier with a Lyapunov solver built from scratch and that we can keep dealing with RESY for further improvments of efficiency. Note that there must be a Lyapunov solver in dynare++, but I'm afraid that it will be plagued with the same problems as the Kalman filter and my feeling is that it takes longer to modify it than to write clean code from the start. Please tell me your thoughts, Best Michel

1 0