Skip to content

[approximation_errors, specification_tests, turbulence_inference, blackwell_kihlstrom] Add three lectures applying information projections - #1138

Merged
mmcky merged 3 commits into
mainfrom
nov_penn9
Nov 22, 2026
Merged

mmcky merged 3 commits into
mainfrom
nov_penn9

Conversation

@thomassargent30

@thomassargent30 thomassargent30 commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

This PR adds three lectures to the Statistics and Information section, right after information_projections. Each one applies m-projections, the main tool of information_projections.

Approximation Errors in Time Series Models (approximation_errors) studies what a misspecified time series model estimates. For Gaussian processes, the KL divergence per observation converges to an integral over frequencies. So the probability limit of a misspecified estimator is an m-projection that trades off errors at different frequencies. Hansen and Sargent's (1993) approximation criterion is twice this divergence rate, plus a constant. The lecture shows:

  • a misspecified autoregression, which is the m-projection onto the autoregressions, and what band deletion and filtering do to it;
  • least squares distributed lags as projections weighted by the regressor's spectral density, with replications of Sims (1971) on sampling continuous-time lags and Sims (1974) on seasonal noise;
  • seasonal adjustment in dynamic models. This covers Hansen and Sargent's experiment and Sims's (1993) permanent income examples, including two numbers in Sims's papers that we could not reproduce. The lecture flags them.

Specification Tests (specification_tests) describes specification tests as comparisons of two m-projections. It covers:

  • Wu's (1973) T2 as the ratio of the squared legs of a right triangle;
  • Hausman's (1978) lemma as a local Pythagorean theorem;
  • Hausman and McFadden's (1984) IIA test, whose LR, LM, and Hausman statistics are a KL divergence and quadratic approximations to it in three coordinates. The exact sizes and powers reproduce their Table III.3.
  • Hansen and Sargent's (1980) exogeneity test;
  • a frequency-band Hausman test, whose asymptotic power is compared with simulated power.

Turbulence, Earnings Instability, and Indirect Inference (turbulence_inference) reproduces the laissez-faire economy of Ljungqvist and Sargent (1998), matching their Table 5 and Figure 14. The lecture then covers:

  • Gottschalk and Moffitt's (1994) permanent/transitory decomposition, read as the m-projection onto a Gaussian random-effects model, with a periodogram and Fejér-kernel form;
  • Ljungqvist and Sargent's check against that evidence, as informal indirect inference;
  • a binding function in the turbulence parameter, and how far it falls short of the data;
  • a layoff-rate alternative that the two statistics barely distinguish;
  • Moffitt and Gottschalk's (2011) covariance-structure model as a second auxiliary model.

Garbling and the data processing inequality (second commit). A new section of approximation_errors uses the data processing inequality of blackwell_kihlstrom to organize what the lecture finds about filtering, seasonal adjustment, and sampling:

  • A filter applied to both the data and the model leaves the m-projection unchanged if its frequency response is zero only at isolated frequencies. This covers the Hodrick–Prescott filter and Sims's adjustment filter. A filter cannot delete a frequency band from the likelihood. Dropping the finite Fourier transforms in the band, which is how Hansen and Sargent define ideal adjustment, is a garbling at every sample size.
  • Hansen and Sargent's three kinds of data (unadjusted, seasonal means removed, ideally adjusted) form a Blackwell chain. Their misspecified econometrician gets the probability limits closest to the truth from the coarsest data, which reverses Blackwell's ranking.
  • Applied to two members of the model, the inequality gives the identification and efficiency costs of deleting frequencies. The lecture derives the covariance matrix of the frequency-domain estimator under misspecification (the sandwich form). It shows that, in the seasonal example, deleting the seasonal bands lowers both the bias and the variance.
  • Deleting low frequencies is also a garbling, but it moves the estimate away from the truth. The inequality orders divergences, not estimates.
  • Sampling a continuous-time process is a garbling, and aliasing (Hansen and Sargent 1983) is observational equivalence after it.
  • A new exercise computes the bias–variance trade-off in choosing a band width.

blackwell_kihlstrom gets a subsection, "Garbling a time series", and five rows in its summary table. specification_tests uses the sandwich covariance to reproduce the simulated standard deviations and correlations of the band test under the alternatives. divergence_measures gets a pointer to the new section.

The PR also:

  • adds links to the new lectures from 16 existing lectures, including the Phillips suite, gallant_tauchen, ols, likelihood_var, sargent_surico, lq_permanent_income, and hansen_singleton_1983;
  • corrects phillips_misspecified's description of its criterion: it is twice the KL divergence rate of the true process from the perceived model;
  • makes the verbal KL convention ("divergence of p from q") consistent in five lectures;
  • labels the normal KL formula in exponential_families (ef_kl_normal);
  • adds 16 bibliography entries.

All three lectures execute without errors. A subset Jupyter Book build that includes the 16 modified lectures reports no warnings from the new lectures or from the new cross-references. Every remaining warning comes from a link to a lecture outside the subset.

After the second commit, a subset build that includes approximation_errors, specification_tests, and blackwell_kihlstrom reports no warnings from the new material. All three lectures execute without errors, and every number in the new prose matches the code output.

🤖 Generated with Claude Code

Add approximation_errors, specification_tests, and turbulence_inference
after information_projections in the Statistics and Information section.

- approximation_errors: misspecified time series estimation as
  m-projection with the KL divergence rate; Hansen and Sargent (1993),
  and Sims (1971, 1974, 1993) on sampling, seasonal noise, and seasonal
  adjustment.
- specification_tests: Wu, Hausman, Hausman-McFadden, and Hansen-Sargent
  (1980) specification tests as comparisons of m-projections, and a
  frequency-band Hausman test.
- turbulence_inference: the laissez-faire economy of Ljungqvist and
  Sargent (1998), Gottschalk and Moffitt's decomposition as an
  m-projection, a binding function in the turbulence parameter, a
  layoff-rate alternative, and Moffitt and Gottschalk's (2011)
  covariance-structure model.

Add links to the new lectures from 16 existing lectures, label the
normal KL formula in exponential_families, make the verbal KL convention
"divergence of p from q" consistent, correct the description of the
criterion in phillips_misspecified, and add 12 bibliography entries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copilot AI balanced review requested due to automatic review settings November 21, 2026 01:27

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 Changes recommended

The simulation drops partial batches and fails for small replication counts, while lecture formulas and spectral claims need corrected assumptions or qualifications.

4 open issues
What changed in this PR

Adds three lectures to the Statistics and Information section, extending information projections to misspecified time series models, specification tests, and indirect inference for earnings instability.

Changes:

  • Adds mathematical explanations, numerical replications, figures, and exercises for the three applications.
  • Connects the new material to existing lectures and adds 12 bibliography entries.
  • Corrects descriptions of KL divergence direction and labels the normal-distribution KL formula.
File Description
lectures/​turbulence_inference.md Adds search-model simulations and earnings-based indirect inference.
lectures/​specification_tests.md Explains estimator-comparison tests through information projections.
lectures/​sargent_surico.md Links filtering to frequency-weighted projections.
lectures/​phillips_self_confirming.md Connects beliefs and forecast errors to m-projections and divergence rates.
lectures/​phillips_policy_evaluation.md Adds interpretations of likelihood gaps and sampling effects.
lectures/​phillips_misspecified.md Corrects the criterion’s KL direction and links related applications.
lectures/​phillips_drifts_volatilities.md Connects structural-break test power to projection differences.
lectures/​ols.md Links the endogeneity test to Wu’s statistic.
lectures/​mccall_model_with_separation.md Links separation risk to skill losses and earnings evidence.
lectures/​lucas_prescott_search.md Links the turbulence model’s quantitative application.
lectures/​lq_permanent_income.md Links seasonal misspecification to inferred consumption responses.
lectures/​likelihood_var.md Adds the spectral interpretation and corrects KL wording.
lectures/​likelihood_ratio_process.md Corrects the verbal direction of KL divergence.
lectures/​likelihood_ratio_process_2.md Clarifies KL direction and belief comparisons.
lectures/​information_projections.md Introduces links to the three new applications.
lectures/​hansen_singleton_1983.md Connects likelihood-ratio tests to conditional KL divergences.
lectures/​gallant_tauchen.md Connects binding functions and specification tests to m-projections.
lectures/​exponential_families.md Labels the normal KL formula and links statistical applications.
lectures/​divergence_measures.md Links the new lectures and standardizes KL wording.
lectures/​approximation_errors.md Adds frequency-domain analysis of misspecification and seasonal adjustment.
lectures/​_toc.yml Places the three lectures after information projections.
lectures/​_static/​quant-econ.bib Adds 12 references supporting the new material.

Review effort: Balanced

Comment thread lectures/specification_tests.md Resolved
Comment thread lectures/approximation_errors.md Resolved
Comment thread lectures/turbulence_inference.md Resolved
Comment thread lectures/turbulence_inference.md Resolved
- approximation_errors: add the section "Garbling and the data processing
  inequality". It covers:
  - filters whose zeros are isolated, which leave the m-projection unchanged;
  - ideal adjustment as dropping Fourier transforms in the seasonal bands;
  - Hansen and Sargent's three kinds of data as a Blackwell chain whose
    ranking misspecification reverses;
  - Fisher-information contraction and the covariance of the estimator
    under misspecification;
  - low-frequency and Hodrick-Prescott filters;
  - sampling and aliasing.
  It also adds the exercise ae_ex_tradeoff on the bias-variance trade-off
  in choosing a band width.
- blackwell_kihlstrom: add the subsection "Garbling a time series" and
  rows to the DPI-Blackwell summary table.
- specification_tests: read the band Hausman test through the data
  processing inequality, and compute the variance of the estimator
  difference under the alternatives.
- divergence_measures: add a pointer to the new section.
- Add bib entries HansenSargent1983, Sargent1978Seasonal, Singleton1988,
  and HodrickPrescott1997.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thomassargent30 thomassargent30 changed the title [approximation_errors, specification_tests, turbulence_inference] Add three lectures applying information projections [approximation_errors, specification_tests, turbulence_inference, blackwell_kihlstrom] Add three lectures applying information projections Nov 21, 2026
…iew fixes and style pass

Address the four Copilot review comments:
- approximation_errors: an AR(1) can produce a narrow peak at pi with a
  coefficient near -1; qualify the sentence.
- turbulence_inference: the moment-matching characterization of the
  random-effects projection needs the same mean in every year.
- turbulence_inference: state the assumptions (positive definite S with a
  nonnegative average off-diagonal element) in ti_ex_projection, and note
  in the solution why sigma^2_mu >= 0 needs them.
- specification_tests: ols_iv now generates the final partial chunk, so R
  need not be a multiple of chunk. Output is unchanged for R = 20,000.

Style pass:
- Shorten the eight figure captions so each says what the figure shows.
- Change six bolded emphasis sentences to italic; definitions stay bold.

No other code changes; all printed output is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mmcky

mmcky commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

Thanks @thomassargent30, these are great additions. I reviewed all three lectures, checking the derivations and every printed number against the text, and found no blockers.

I pushed 7595f05 with:

  • the four Copilot fixes: three qualifying clauses, plus the ols_iv loop, which now includes the final partial batch (the output is unchanged)
  • shorter figure captions and six bold sentences changed to italic, per the style guide

The preview at 7595f05 builds cleanly in both HTML and PDF, with no execution failures and no new warnings.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants