Possible Bug With `gridinterplayer = 1` and Semi-Exogenous States

1

I found what looks like a bug in the finite-horizon stationary-distribution code for models with semi-exogenous states when gridinterplayer = 1.

Summary

  • The model runs through with simoptions.gridinterplayer = 0.

  • The same model fails when simoptions.gridinterplayer = 1.

  • The failure happens in the stationary-distribution step, not in VFI itself.

  • The error appears to come from a length mismatch inside a sparse(...) call in the semi-exogenous nProbs distribution code.

Reproduction

In my application:

  • finite horizon

  • one endogenous asset state

  • one decision variable d

  • semi-exogenous states semiz = (l, g, z3)

  • no exogenous Markov state z

  • permanent types handled through *_PType wrappers

The toolkit is called through:

  • ValueFnIter_Case1_FHorz_PType

  • StationaryDist_Case1_FHorz_PType

When I run with:


vfoptions.gridinterplayer = 0;

the toolkit solve and stationary distribution complete successfully.

When I run with:


vfoptions.gridinterplayer = 1;

vfoptions.ngridinterp = 15;

VFI completes, but the code fails during the stationary-distribution step.

Error

The stack trace is:


Error using sparse

Vectors must be the same length.

Error in StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_raw (line 47)

Gammatranspose=sparse(Policy_aprimesemiz(:,:,jj),II2,PolicyProbs(:,:,jj),N_a*N_semiz,N_a*N_semiz);

Error in StationaryDist_FHorz_SemiExo (line 99)

StationaryDist=StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_raw(...)

Error in StationaryDist_FHorz_Case1 (line 201)

StationaryDist=StationaryDist_FHorz_SemiExo(...)

Error in StationaryDist_Case1_FHorz_PType (line 234)

StationaryDist_ii=StationaryDist_FHorz_Case1(...)

Relevant toolkit functions:

  • StationaryDist_FHorz_SemiExo

  • StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_raw

  • also likely the sibling nProbs semi-exogenous routines

Why this seems to fail

Inside:

  • StationaryDist/FHorz/nProbs/StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_raw.m

the code first compresses the semi-exogenous transition matrix to keep only the nonzero successors:


N_semizshort = max(max(max(sum((pi_semiz_J>0),2))));

Then Policy_aprimesemiz and PolicyProbs are built using N_probs * N_semizshort entries per current state.

But II2 is built using N_probs * N_semiz:


II2 = repelem((1:1:N_a*N_semiz)',1,N_semiz*N_probs);

So if the semi-exogenous transition matrix is sparse and N_semizshort < N_semiz, then:

  • Policy_aprimesemiz(:,:,jj) has width N_probs * N_semizshort

  • PolicyProbs(:,:,jj) has width N_probs * N_semizshort

  • II2 has width N_probs * N_semiz

and the sparse(...) inputs no longer have matching lengths.

That seems consistent with the observed error.

Proposed fix

In:

  • StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_raw.m

replace:


II2 = repelem((1:1:N_a*N_semiz)',1,N_semiz*N_probs);

with:


II2 = repelem((1:1:N_a*N_semiz)',1,N_semizshort*N_probs);

Likely related files

The same issue may also be present in the sibling semi-exogenous nProbs routines:

  • StationaryDist_FHorz_Iteration_SemiExo_nProbs_noz_e_raw.m

  • StationaryDist_FHorz_Iteration_SemiExo_nProbs_e_raw.m

Those files also appear to construct II2 using N_semiz * N_probs even though the corresponding policy objects are expanded with N_semizshort.

For reference, the z version:

  • StationaryDist_FHorz_Iteration_SemiExo_nProbs_raw.m

already uses N_semizshort * N_probs when building II2, which looks like the consistent implementation.

Bottom line

My current understanding is:

  • gridinterplayer = 0 works

  • gridinterplayer = 1 fails only in the semi-exogenous stationary-distribution nProbs path

  • the likely cause is an index-length mismatch caused by using N_semiz instead of N_semizshort when constructing II2

minimal reproducible example based on my toy model: GitHub - aledinola/life-cycle-search · GitHub

2

I have also sent a PR with a fix.

1 Like
3

Merged the PR. Thanks!

1 Like