Pith. sign in

REVIEW 2 major objections 4 minor 29 references

In exact conditioned diffusion models, the choice of heavy-tailed noise law cannot shape the generated distribution; only coupling the noise to the data restores control.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:05 UTC pith:GWVECYBO

load-bearing objection Correct math and an honest experiment program, but the abstract overstates the coupling claim and the invariance theorem's scope — still worth refereeing. the 2 major comments →

arxiv 2607.19218 v1 pith:GWVECYBO submitted 2026-07-21 q-fin.MF math.STstat.TH

Denoising Subordinated Probabilistic Models: Diffusion with a Tempered-Stable Volatility Clock, and What the Noise Mechanism Actually Controls

Junchi Shen , Helin Zhao This is my paper
classification q-fin.MF math.STstat.TH
keywords denoising diffusionheavy-tailed noisevolatility clusteringtempered-stable subordinatorvariance mixturestochastic volatilitynuisance parameterfinancial time series generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish two opposite things at once. Constructively, it defines a diffusion model whose per-coordinate noise variance follows a stationary AR(1) chain driven by tempered-stable jumps, so the noise itself carries volatility clustering; it derives closed-form formulas for kurtosis and squared-noise autocorrelation and an exactly identified, analytically invertible calibration, with Gaussian, i.i.d. stable, and Student-t noise as boundary cases. Delimitingly, it proves that if the denoiser is conditioned on the mixing chain and the chain is drawn independently of the data, the exact reverse process terminates at the data distribution regardless of the chain's law, so interventions on volatility do nothing. Controlled experiments confirm this invariance and show that a blind denoiser transmits the calibrated clustering exactly. The paper then breaks the independence with a variational encoder and recovers control over shocks, at the cost of unconditional kurtosis falling below the Gaussian baseline; the paper is explicit that fidelity and control are demonstrated separately, not jointly.

Core claim

The paper's central claim is a dichotomy. Constructively, the proposed Denoising Subordinated Probabilistic Model (DSPM) puts a stationary AR(1) tempered-stable chain inside the noise variance; conditioned on the chain, the denoising diffusion probabilistic model's posterior and objective are unchanged, and squared-noise autocorrelation is closed-form, giving exact calibration. Delimitingly, Proposition 5 proves that if the denoiser is conditioned on the chain and the chain is independent of the data, the exact reverse chain terminates at the data distribution — the chain's law is a nuisance and interventions do nothing. Experiments confirm both halves: conditioned models reproduce clusterin

What carries the argument

The central object is the variance chain A: a stationary AR(1) recursion A_i = φ A_{i-1} + η_i, with tempered-stable increments η_i normalized so E A_i = 1. Because every forward covariance is a multiple of the same diagonal matrix D_A, the conditional Gaussian posterior and denoising objective factorize, so the denoising diffusion machinery survives verbatim once the denoiser is fed √A and whitened inputs. The key identities are K = 3(1−α)/(θ(1+φ)) and ρ_r2(h) = v φ^h/(3v+2) with v = Var(A_i), giving an exactly identified calibration with analytic inverse and feasibility bound ρ_r2(1) < K/(3(K+2)). The load-bearing theorem is Proposition 5: under training-time independence of x0 and A, exac

Load-bearing premise

The invariance theorem presupposes that in training the data and the volatility chain are drawn independently, and that reverse kernels are exact; couple the two — as the paper's own coupled model does — and the theorem no longer applies.

What would settle it

Train a conditioned denoiser with independent A, let it converge, and sample under a designed ×8 volatility path; Proposition 5 predicts the amplitude envelope ratio stays near 1 (measured 0.88±0.03), against the naive √A prediction of 2.83 — a measured ratio far from 1 would falsify the invariance.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Conditioned models reproduce the data's volatility clustering regardless of whether the noise is Gaussian, i.i.d. stable, shared Student-t scale, or a persistent tempered-stable chain; all measured lag-one squared-return autocorrelations land within 0.017 of the data's 0.159.
  • In the exact-denoiser limit, scenario design through the mixing chain is inert: a ×8 volatility shock changes the generated amplitude envelope by a factor 0.88±0.03, where √A-scaling would predict 2.83.
  • Blind denoisers transmit the mechanism as calibrated: blind DSPM samples show lag-one squared-return autocorrelation 0.151 against the calibrated 0.161, with the predicted geometric decay, while the i.i.d. variant's clustering collapses to 0.008.
  • Coupling the chain to data through a variational encoder restores quantitative control: the same ×8 shock scales the envelope by 3.00±0.14 and the calm regime by 0.52±0.01, bracketing naive √A predictions.
  • The simplified denoising loss cannot train the volatility encoder on its own because the whitened residual is nearly scale-free in A; the exact stochastic-volatility likelihood's log-determinant term is required.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Testable extension: run the four-way conditioned comparison at larger scale with a strict measure of the exact-denoiser limit; because Proposition 5 is scale-free, any systematic between-noise difference in dependence structure would expose a finite-T or receptive-field effect the paper leaves open.
  • The feasibility bound ρ_r2(1) < K/(3(K+2)) (always below 1/3) is a cheap pre-training screen: a dataset with lag-one squared autocorrelation above the bound at matched kurtosis cannot be described by a single AR(1) mixing chain, pushing toward multi-factor (supOU) volatility before any neural training.
  • For financial scenario generation, the paper's dichotomy means conditioned heavy-tailed diffusion is inert for volatility stress-testing unless the latent is coupled to the path; the decisive test would be path-dependent pricing under generated measures, which the paper leaves for future work.
  • The paper states plainly that v2 runs on a log-normal volatility clock, not the tempered-stable chain of the constructive section; a model carrying the tempered-stable prior via Monte-Carlo KL, or with a supOU multi-factor chain, is the natural step toward joint fidelity and control.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper introduces the Denoising Subordinated Probabilistic Model (DSPM), a DDPM whose per-coordinate noise variance is a stationary AR(1) chain driven by tempered-stable increments. Conditionally on the chain, the standard DDPM posterior, ELBO, and ancestral sampling hold unchanged (Propositions 1--2). The authors prove closed-form identities for excess kurtosis and squared-noise autocorrelation (Proposition 3) and give an analytic calibration inverse with a feasibility bound (Corollary 1); DDPM, i.i.d.-stable, and shared-scale Student-t noise are recovered as boundary cases (Proposition 4). Proposition 5 states that, with exact conditional reverse kernels, the generated distribution is invariant to the law of the mixing chain if the denoiser is conditioned on it. Experiments with GARCH-t data confirm that conditioned models do not transmit the mixing law, blind models do, and a variational encoder (v2) that couples A to x0 restores control. The paper ships reproducible scripts for all numbers.

Significance. The DSPM construction is elegant and the closed-form calibration is a genuine practical contribution. The paper is unusually honest: code is provided, every headline number is seed-averaged where it matters, and the limitations of the blind ablation and the v2 fidelity/control trade-off are stated explicitly. The nuisance-invariance theorem is a useful delimiting result for the heavy-tailed diffusion literature. Its scope, however, is narrower than the abstract says: it applies to a latent A that is independent of the data at training time, not to any conditioning on A. The v2 experiment is a well-designed controlled demonstration of exactly this boundary.

major comments (2)
  1. [Abstract; §3.4, Eq. (9)] Proposition 5's conclusion p(x0|A)=q(x0|A)=q(x0) uses q(x0|A)=q(x0), which is valid only because training draws x0 and A independently (stated in §3.1). The abstract's sentence 'when the denoiser is conditioned on the mixing variables, their law is a nuisance' omits this premise and is false as written: v2 (§3.5, Table 4) conditions the denoiser on A, yet interventions move the envelope by 3.00±0.14 exactly because the encoder breaks x0⊥A. Please add the independence qualifier to the abstract and to the formal statement of Proposition 5, and rephrase the headline claim as 'an independently drawn mixing chain is a nuisance in the exact-denoiser limit.'
  2. [§4.5, Table 3] The blind-model quantitative claim ('blind models transmit the mechanism exactly as calibrated') rests on a single training run and on a self-described 'crude' pairing of a marginal denoiser with A-scaled updates. This is acceptable as an ablation but not as the 'exactly as calibrated' headline. Please either supply multiple seeds for the blind rows or weaken the wording in the abstract and Section 4.5.
minor comments (4)
  1. [§3.4] The independence hypothesis x0⊥A should be part of the formal statement of Proposition 5, not only a sentence in the preceding paragraph. As written, readers may quote the proposition without the premise.
  2. [Remark 2, Eq. (10)] The claim that restoring the log-determinant term is 'necessary and sufficient' for the encoder to train is stronger than the evidence: the paper shows two failure modes and one success. Suggest phrasing it as 'necessary in our experiments' or providing a formal argument.
  3. [§4.5] The single-run blind ablation should either report seed variability or be clearly marked in the abstract, since the abstract presents the 0.151 vs 0.161 agreement as a headline result.
  4. [References] References to 'Anonymous, 2026' and 'Baule, 2026' are working papers without full author lists. Unless the submission is double-blind, these should be updated to complete archival citations before publication.

Circularity Check

0 steps flagged

No significant circularity: all load-bearing results are derived from stated premises or tested by out-of-sample Monte Carlo checks.

full rationale

The derivation chain is self-contained. Definition 1 fixes the model via (2)-(3). Propositions 1-2 are coordinate-wise algebra showing the conditional posterior and KL reduce to DDPM form. Proposition 3 and Corollary 1 are exact moment computations and their algebraic inversion; the feasibility bound is just v>0 and phi<1 applied to (7), not an empirical fit. Proposition 4 is a limit computation. These are falsifiable: Section 4.2's Monte Carlo recovery matches to three digits, and the blind model's rho_r2(5)=0.019 is a cross-moment value not used in calibration, so it is a genuine prediction rather than a renamed input. Proposition 5 is explicitly conditional on the stated training protocol x0 ⊥ A (Section 3.1) and the exact-denoiser limit; its proof invokes q(x0|A)=q(x0) exactly as that premise. The paper does not conceal this: it states the premise, names it as load-bearing, and v2 (Section 3.5) is presented precisely as breaking it. That is a scope limitation, not circularity. The anonymous reference (Anonymous, 2026) is contextual and not load-bearing; no uniqueness claim or ansatz is imported from a self-citation. Empirical comparisons are out-of-sample against synthetic GARCH ground truth with fixed seeds, and the paper openly states its limitations (small scale, crude blind sampler, v2 fidelity/control trade-off). No circular step can be quoted or exhibited.

Axiom & Free-Parameter Ledger

6 free parameters · 9 axioms · 2 invented entities

The paper's derivations are mostly parameter-free given the model definition (tempered-stable moment formulas, time-reversal identity), which keeps the circularity burden moderate. The load-bearing free parameters are the calibrated pair (φ,θ)=(0.638,0.0876) fit by (8) to training-set moments, the hand-fixed α=0.7, and v2's empirical-Bayes-learned (a,s²). The v1 independence x₀⊥A is a domain assumption the authors themselves identify and then break in v2; the v2 log-normal substitution is ad hoc and explicitly flagged. The exact-denoiser limit and the GARCH-t synthetic ground truth are the two domain assumptions that all empirical sharpness rests on.

free parameters (6)
  • α (tempered-stable activity index) = 0.7 (fixed by hand)
    Section 3.3: 'we fix α=0.7'. Not fitted or learned; shapes the noise marginal beyond the first two moments.
  • φ (chain persistence) = 0.638 in main experiments
    From inverse map (8) applied to training-set targets (K̂,ρ̂₁)=(6.27,0.161); 0.750 in the recovery experiment. Controls the geometric decay of squared-noise autocorrelation.
  • θ (tempering rate) = 0.0876 in main experiments
    From the same calibration map (8); sets the tail regime (small θ → apparent power-law behavior).
  • v2 log-normal prior memory a and variance s² = a: 0.638 → 0.803 (learned); s²=log(1+K/3)
    Section 3.5/4.6(c): initialized from the moment calibration, moved by empirical Bayes toward the GARCH persistence a+b=0.95 (directional claim across different parametrizations).
  • Noise schedule β_t = T=100, linear 10⁻³→0.2
    Hand-chosen DDPM schedule used in all experiments; inherited rather than fitted.
  • Coarse bottleneck d_c = 16 (path length d=64)
    Section 3.5: chosen to structurally block the latent from copying per-coordinate detail; directly affects v2's kurtosis collapse.
axioms (9)
  • standard math Tempered-stable Lévy Laplace transform and cumulants: Ee^{-uη}=exp(-δ[(θ+u)^α-θ^α]), κ_n=δ α Γ(n-α)/Γ(1-α) θ^{α-n}
    Section 2, Eq. (1); the basis of Proposition 3's moment identities and the sampler.
  • standard math Kanter's one-sided stable representation with exponential-tilting rejection samples TS(α,θ,δ) correctly
    Section 4.1; sampler validated against the Laplace transform to ≤2.1×10⁻⁴. Load-bearing for all Monte Carlo, hence Tables 1–4.
  • standard math Exact-reverse-chain time-reversal identity: composing exact posterior kernels from the exact terminal marginal reproduces the forward joint law
    Proposition 5 proof (Section 3.4). The whole delimiting result is this identity applied conditionally on A.
  • standard math Conditionally on A, the Gaussian DDPM posterior/ELBO identities hold verbatim (Propositions 1–2)
    Section 3.2; relies on the noise covariance being a scalar multiple of one diagonal D_A.
  • domain assumption x₀ and A are drawn independently in v1 training
    Section 3.1 ('a fact that Proposition 5 will make consequential'); used at Eq. (9) as q(x₀|A)=q(x₀). This is the premise v2 breaks to regain control.
  • domain assumption The exact-denoiser limit is reached by the trained network (converged training, receptive field ≥ d)
    Section 3.4 + experiments; empirical sharpness of the invariance depends on the conditioned denoiser being near-exact (RF 69≥64, 6000 steps). The small-RF ablation (RF=13) is admitted to be suggestive only.
  • domain assumption GARCH(1,1)-t (ω=0.05, a=0.10, b=0.85, ν=6) is an adequate synthetic stand-in for the target stylized facts
    Section 4; limitation (4) defers real-index tests. All fidelity conclusions are measured against this generator.
  • ad hoc to paper v2 substitutes a log-normal 'Clark' clock for the tempered-stable chain of Definition 1
    Section 3.5 explicit flag: chosen for closed-form Gaussian KL; the coupling contribution therefore does not use the TS mechanism, only its calibration logic and initialization.
  • ad hoc to paper Restoring the log-determinant term is necessary and sufficient for the encoder to train (Remark 2)
    Section 3.5 Remark 2: predicted and verified via two documented failure modes (posterior collapse at β=1; copy-cheat at β=0.1); the necessity claim is empirical, not a theorem.
invented entities (2)
  • DSPM volatility chain A (AR(1) with tempered-stable increments along the data axis) independent evidence
    purpose: Persistent, calibrated heavy-tailed noise variance inside a DDPM — the discrete BNS stochastic-volatility clock
    Provides falsifiable external handles: the feasibility bound ρ₁<K/(3(K+2)) (always <1/3) constrains real data, the closed-form geometric ACF decay (7) is an identifiable fingerprint, and the object is anchored in prior BNS/time-changed-Lévy literature rather than invented ad hoc.
  • Variational volatility encoder q_ψ(A|x₀) with coarse log-normal AR(1) prior (v2) no independent evidence
    purpose: Breaks x₀⊥A so designed variance paths become control inputs; filters latent volatility from observed paths
    Validated only in-paper against synthetic GARCH(1,1)-t ground truth (per-path correlation 0.567); no external falsifiable handle yet — the authors defer real-index validation to future work (limitation 4).

pith-pipeline@v1.3.0-alltime-deepseek · 13269 in / 30264 out tokens · 266196 ms · 2026-08-01T13:05:12.632063+00:00 · methodology

0 comments
read the original abstract

Heavy-tailed diffusion models replace Gaussian noise by a Gaussian variance mixture: denoising Levy probabilistic models (DLPM) take the mixing variables i.i.d. across coordinates, while Student-t EDM shares one mixing variable per sample. Neither has dynamics, yet temporal dependence of the noise amplitude - volatility clustering - is the defining stylized fact of financial returns. We introduce the Denoising Subordinated Probabilistic Model (DSPM), whose mixing vector is a stationary AR(1) chain driven by tempered-stable subordinator increments (the discrete Barndorff-Nielsen-Shephard volatility process) along the data axis. Conditionally on the chain the DDPM machinery survives verbatim; kurtosis and squared-noise autocorrelation are closed-form in the chain parameters, giving an exactly identified, analytically invertible calibration; DDPM, DLPM and Student-t noise are boundary cases of one memory parameter. We then prove a delimiting result: when the denoiser is conditioned on the mixing variables, their law is a nuisance - in the exact-denoiser limit the generated distribution is invariant to it and interventions on the chain do nothing. Experiments confirm both halves: conditioned models match the data's clustering whatever the mixing law, a designed x8 volatility shock moves the envelope by under 13%, while blind models transmit the mechanism exactly as calibrated. Finally, coupling the chain to the data by a variational volatility encoder - trained with the stochastic-volatility likelihood whose log-determinant the simplified denoising loss provably drops - restores control (shock response 3.07 vs. naive 2.83), recovers latent volatility (correlation 0.76), and learns the prior memory toward the true persistence.

Figures

Figures reproduced from arXiv: 2607.19218 by Helin Zhao, Junchi Shen.

Figure 1
Figure 1. Figure 1: Sampler validation. Left: empirical vs. exact Laplace transform, TS(0 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: exact ACF (7) vs. Monte Carlo. Middle: recovered ( ˆφ, ˆθ) over 200 replications (star = truth). Right: noise path ri = √ Aizi with latent √ Ai—clustering before any learning [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Conditioned comparison. Left: pooled marginals (log scale). Middle: pooled ACF of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Invariance under interventions. Left: amplitude envelopes under the three designed [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: v2. Left: under identical designed interventions, the v2 envelope responds along the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 4 linked inside Pith

  1. [1]

    Baeumer and M

    B. Baeumer and M. M. Meerschaert. Tempered stable L\'evy motion and transient super-diffusion. J. Comput. Appl. Math., 233:2438--2448, 2010

    2010

  2. [2]

    O. E. Barndorff-Nielsen and N. Shephard. Non-Gaussian Ornstein--Uhlenbeck-based models and some of their uses in financial economics. J. R. Stat. Soc. B, 63(2):167--241, 2001

    2001

  3. [3]

    R. Baule. Score-based generative models with jump-diffusion forward processes. Working paper, 2026

    2026

  4. [4]

    Beck and E

    C. Beck and E. G. D. Cohen. Superstatistics. Physica A, 322:267--275, 2003

    2003

  5. [5]

    Bollerslev

    T. Bollerslev. Generalized autoregressive conditional heteroskedasticity. J. Econometrics, 31:307--327, 1986

    1986

  6. [6]

    P. Carr, H. Geman, D. B. Madan, and M. Yor. The fine structure of asset returns: an empirical investigation. J. Business, 75(2):305--332, 2002

    2002

  7. [7]

    P. Carr, H. Geman, D. B. Madan, and M. Yor. Stochastic volatility for L\'evy processes. Math. Finance, 13(3):345--382, 2003

    2003

  8. [8]

    A. V. Chechkin, F. Seno, R. Metzler, and I. M. Sokolov. Brownian yet non-Gaussian diffusion: from superstatistics to subordination of diffusing diffusivities. Phys. Rev. X, 7:021002, 2017

    2017

  9. [9]

    M. V. Chubynsky and G. W. Slater. Diffusing diffusivity: a model for anomalous, yet Brownian, diffusion. Phys. Rev. Lett., 113:098302, 2014

    2014

  10. [10]

    P. K. Clark. A subordinated stochastic process model with finite variance for speculative prices. Econometrica, 41(1):135--155, 1973

    1973

  11. [11]

    R. Cont. Empirical properties of asset returns: stylized facts and statistical issues. Quant. Finance, 1:223--236, 2001

    2001

  12. [12]

    L. Devroye. Non-Uniform Random Variate Generation. Springer, 1986

    1986

  13. [13]

    R. F. Engle. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica, 50(4):987--1007, 1982

    1982

  14. [14]

    Gopikrishnan, V

    P. Gopikrishnan, V. Plerou, L. A. N. Amaral, M. Meyer, and H. E. Stanley. Scaling of the distribution of fluctuations of financial market indices. Phys. Rev. E, 60:5305--5316, 1999

    1999

  15. [15]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840--6851, 2020

    2020

  16. [16]

    M. Kanter. Stable densities under change of scale and total variation inequalities. Ann. Probab., 3(4):697--707, 1975

    1975

  17. [17]

    Diffusion-based generative models for financial time series via geometric Brownian motion

    Kim et al. Diffusion-based generative models for financial time series via geometric Brownian motion. arXiv:2507.19003, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    I. Koponen. Analytic approach to the problem of convergence of truncated L\'evy flights towards the Gaussian stochastic process. Phys. Rev. E, 52:1197--1199, 1995

    1995

  19. [19]

    R. N. Mantegna and H. E. Stanley. Stochastic process with ultraslow convergence to a Gaussian: the truncated L\'evy flight. Phys. Rev. Lett., 73:2946--2949, 1994

    1994

  20. [20]

    Mikosch and C

    T. Mikosch and C. St a ric a . Limit theory for the sample autocorrelations and extremes of a GARCH(1,1) process. Ann. Statist., 28(5):1427--1451, 2000

    2000

  21. [21]

    E. W. Montroll and G. H. Weiss. Random walks on lattices. II. J. Math. Phys., 6:167--181, 1965

    1965

  22. [22]

    Nobis et al

    G. Nobis et al. Generative fractional diffusion models. arXiv:2310.17638, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Pandey et al

    K. Pandey et al. Heavy-tailed diffusion models. ICLR, 2025. arXiv:2410.14171

    Pith/arXiv arXiv 2025

  24. [24]

    Sabino and N

    P. Sabino and N. Cufaro Petroni. Fast simulation of tempered stable Ornstein--Uhlenbeck processes. Comput. Statist., 37:2517--2551, 2022

    2022

  25. [25]

    Shariatian, U

    D. Shariatian, U. Simsekli, and A. Durmus. Denoising L\'evy probabilistic models. arXiv:2407.18609, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. ICML, 2015

    2015

  27. [27]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. ICLR, 2021

    2021

  28. [28]

    E. B. Yoon, K. Park, S. Kim, and S. Lim. Score-based generative models with L\'evy processes. NeurIPS, 36, 2023

    2023

  29. [29]

    Working paper, 2026

    Do heavy tails help diffusion? A statistical-estimation perspective. Working paper, 2026

    2026