Optimal Progressive Capital Income Taxes in the Infinite Horizon Model

Why it matters, in detail. Panel B is inconsistent with Panel A under the stated model. The bound uses only c_t = c0e^{−στρt}, the bracket law of motion, x_T ≤ a*∞ and y ≤ 1 + ρE[a0], and is violated by factors of 1.8–7.7. The corrected times are long: P99 needs 47 (τ=1) to 191 (τ=0.25) years at σ = 0.25, and P99.99 needs 177–970 years. The WP's 'it would take many decades to reduce the very top fortunes' should read 'centuries'. Its 'after a century … virtually all individuals except the very top' remains roughly true for P99 when τ ≥ 0.5, but not at τ = 0.25 (P99 takes 191 years at σ = 0.25 and 106 at σ = 0.45), and P99.9 exceeds a century in 5 of 9 cells. No reading of the documented model reconciles the two panels: for P99 to exit in 9.1 years at σ = 0.25, τ = 1, the threshold would have to be ρa* ≥ 0.96, against 0.155 reported and 0.21–0.23 re-optimised.

Evidence. WP Table 1 (PDF p. 50) was transcribed and checked against the rendered page. Adjudicator bound (scripts/wp_table1_bound.py): a0 ≤ a*∞e^{−qT} + (ρa*∞+Y)e^{λT}J_{q+λ}(T) − Y·J_q(T), with q = ρ(1−τ) and λ = στρ. With Y = 1.25, 26/27 entries are infeasible and minimum/reported = 1.78–7.70 (median 2.53). Examples: σ=0.25, τ=1: P99 ≥ 42.9 yrs vs 9.1 reported, P99.9 ≥ 93.6 vs 47.3, P99.99 ≥ 170.8 vs 95.7. σ=0.25, τ=0.25: P99 ≥ 214.6 vs 64.2. σ=0.6, τ=0.25: P99 ≥ 62.3 vs 8.1. With Codex's looser Y = 2.4892, which does not assume non-negative wealth, the count is still 26/27. The one non-rejected cell (σ=0.6, τ=1, P99) has initial capital income 1.46 below that cell's threshold 3.216. Independent re-optimisation (scripts/wp_table1_reopt.py; L-BFGS-B, 4 starts; Claude used Powell+Nelder–Mead) gives years at the WP's own wealth levels of 47.0/99.5/176.8 (σ=0.25, τ=1; WP 9.1/47.3/95.7), 191.4/526.3/969.6 (σ=0.25, τ=0.25; WP 64.2/210.6/523.5), 24.8/59.2/108.1 (σ=0.45, τ=1; WP 5.3/17.1/55.4) and 20.0/47.3/70.7 (σ=0.6, τ=1; WP 4.1/6.3/20.7). These are within 0.1 year of Claude's replication. Codex's bound numbers (e.g. P99 ≥ 33.16 yrs at Y = 2.4892; 42.87 at Y = 1.25) match to the decimal.

Why this severity. The error is confined to the working paper's numerical illustration and leaves the published theorems untouched. But it is the only quantification of the policy's time scale the author ever published, and it understates the time needed to run down fortunes by a factor of 2–8.

The authors' response. None located. The published version (JPubE 2013) omits Section 5 and Table 1 without comment; the WP says the MATLAB programs are 'available upon request' (fn 31), and none are public.

Why it matters, in detail. With AW's preferred σ = 0.56–0.67, σα = 0.84–1.0 at α = 1.5 and 1.12–1.34 at α = 2. Truncation then requires the thickest observed tails and the lower AW estimate. Proposition 2 concerns the a0 → ∞ tail, so the relevant σ is that of the wealthiest. CRRA imposes a constant σ (fn 8 flags general utility), while micro evidence points to a higher IES among asset holders. The WP drew its strongest policy conclusion from this characterisation ('we would expect that the key condition … is empirically statisfied [sic], implying … progressive taxation should be used aggressively', WP p. 28). The published text hedges it, noting log-utility calibrations and near-zero uncompensated saving elasticities that imply σ ≈ 1.

Evidence. Saez p. 70 fn 34 (rendered page checked). AW, JPE 103(6), 1995: p. 1140 ('our favorite specification … 0.34 in the long sample and at 0.48 in the short one, but without much precision'; AW's text calls these columns 3 and 6, but in the printed Table 3 they are columns 2 and 5). Table 6 p. 1150: 0.559/0.668. p. 1149: 'This estimate of the EIS is higher and more precise than the ones obtained with just one price index'. p. 1150: total-consumption micro data 'yields even higher estimates' than Runkle's 0.45 (food). Other micro evidence available when the paper was written: Blundell, Browning & Meghir (1994, REStud 61(1), Table VI; UK data), with ISE at group means of 0.61–0.82 (model 1) and 1.08–2.21 (model 2); and Vissing-Jørgensen (2002, JPE 110(4), NBER WP 8896 abstract), with 0.3–0.4 for stockholders and 0.8–1 for bondholders, 'larger for households with larger asset holdings'.

Why this severity. This is the paper's only quantitative micro evidence for the key parameter of its main condition, and the error runs in the direction favourable to that condition. With the source's preferred values the condition is borderline at α = 1.5 and fails at α = 2. The published bottom line, that the condition holds 'for some but not all estimates', still stands, which limits the damage to a secondary claim.

Why it matters, in detail. Decumulation in the bracket proceeds at λ = στρ (1/λ = 80 yrs at σ=0.25, τ=1; 320 yrs at τ=0.25). Exit times grow like log(a0/ā0)/λ, so 'truncation' is reached only in the limit. For σα near 1 the optimal threshold is still rising well past t = 100: the sign of the threshold reform is governed by σα only for t̄ ≫ 1/λ, and the neglected welfare term is 2%–150% of the revenue term at t̄ = 100. Because the a*_t ≥ 0 constraint binds early, the optimum in the WP calibration is close to a uniform 100% (or τ) capital income tax for the first 5–20 years. Human-wealth effects make small fortunes dissave faster, so top shares rise first; this happens for σ = 0.25 under both calibrations and for σ = 0.45 under the continuous one. C4's superiority over linear taxation is therefore a long-run statement: it is confirmed here (long-run top-1% share 2.1–3.6% vs 21–23% under the optimal linear tax), but it emerges only after several decades. In a closed economy, aggregate decumulation of 40–66% would move r and w substantially over this long transition, which the paper's conjecture (Section 5.3) does not address.

Evidence. (i) Exit times under a constant threshold (scripts/time_scale.py; ρa* = 0.155–0.5, y = 1.05). σ=0.25, τ=1: 167–182 yrs for ρa0 = 31.4 (P99.99) and ~600 yrs for ρa0 = 10^4. σ=0.25, τ=0.25: 912–991 and ~2,740–2,820 yrs. σ=0.6, τ=1: 92–99 and ~280 yrs. The Proposition 1 analogue gives T(a0 = 10^4) = 360 yrs at σ = 0.25. (ii) Finite-horizon knife-edge. The adjudicator derived the exact per-person revenue effect of raising a* after t̄ for any τ: δTax/δa* = −(J_{ρ+λ}(T)/J_{q+λ}(T))[τρ∫_{t̄}^T e^{−qs}ds − e^{−qT}] − e^{−ρT}, checked against finite differences to 2.5e-7. Integrated over a Pareto tail, the critical σα is 0.62–0.65 (τ=1, σ=0.3), 0.89–0.90 (τ=1, σ=0.6), 0.54–0.67/0.81–0.88 (τ=0.5) and 0.45–0.80/0.74–0.98 (τ=0.25) at t̄ = 100. It reaches 1.000 only by t̄ ≈ 500–1,000 (τ = 1) and ≈ 3,000 (τ = 0.25), and at t̄ = 50 with τ = 0.25 it can exceed 1 (1.11–1.28). The asymptotic sign is that of τρ(ασ−1)/(q+αλ) for every τ, which confirms the τ-independence of the limit. (iii) Dynamics under the re-optimised WP policy (scripts/dynamics.py; α = 1.5, utilitarian, two calibrations of the undocumented tail mass). Thresholds are ρa* ≈ 0.000–0.04 for the first 5–20 years (t < 5 for σ ≥ 0.45; t < 10 for σ = 0.25, τ = 1; t < 20 for τ = 0.25), and the share of people taxed falls from ~100% to 11% (σ=0.25, τ=1) within 10 years as modest fortunes are run down. Top-1% wealth share, initial 17.5%: σ=0.25, τ=1: 22.3% (t=10), 17.0% (25), 9.5% (50), 4.2% (100), 2.1% (limit). σ=0.25, τ=0.25: 20.9% (25), 19.8% (50), 17.0% (100), 11.1% (200), 7.3% (300), 2.0% (1,000). Continuous-density calibration (initial 23.4%), σ=0.25, τ=0.25: 30.5% at t=50, 29.5% at 100, 22.6% at 200. For σ ≥ 0.45 at τ = 1 the share falls monotonically after t ≈ 10, by about half within 25–50 years. Aggregate private wealth ends at 34–60% of its initial level.

Why this severity. The limit theorem is correct and the paper calls it 'long-run'. The undisclosed time scale and transition dynamics matter a great deal for the paper's policy framing: in the model, truncation takes centuries, and at low IES concentration can be higher one or more decades after the start than at the start. They do not overturn any stated result.

The authors' response. No direct response. Saez's later work (Saez & Stantcheva 2018, JPubE 162: 120–142, Conclusion, p. 137) says the anticipated-reform approach, 'especially for very long-run tax rates, lacks realism both in terms of policy process and in terms of the infinite elasticities [it] generates in models without uncertainty'.

Why it matters, in detail. With idiosyncratic multiplicative return risk and a reflecting/reinjection mechanism below the threshold, a negative after-tax drift of log wealth in the bracket yields a stationary Pareto tail whose exponent ζ exceeds the no-tax exponent (a thinner tail), rather than a mass point at a*∞ (Kesten/Gabaix logic). The fraction taxed and the revenue then converge to positive constants. C1 ('fraction taxed vanishes') and the literal truncation in C2 would fail. Whether the optimal threshold path would still be bounded is open. Under τ = 1 with full loss offset, returns above the threshold are fully taxed, so regeneration would come only from below; the 'any τ > 0' claim is where the deterministic assumption bites hardest.

Evidence. Assumption 1 (p. 63); Section 2.3.1 (p. 64); fn 30 (p. 69); p. 62 ('ignores completely the issue of creation of new wealth'; 'not immediately clear in which direction would the introduction of entrepreneurs tilt the results'). Benhabib, Bisin & Zhu (2011, Econometrica 79(1): 123–157; NBER WP 14730 abstract) show that with capital income risk the stationary wealth distribution has a Pareto right tail driven by that risk, and that capital income and estate taxes 'can significantly reduce wealth inequality' (i.e. they act on the tail rather than truncating it). No stochastic optimal-tax version of Saez's problem was solved by the auditors or the adjudicator.

Why this severity. The truncation and vanishing-revenue results are exact properties of the deterministic model. Under a plausible extension, whose mechanism the paper itself invokes, they would likely be replaced by a thinner tail and a persistent tax base. The qualitative message that progressive taxes compress the top would probably survive. The impact on the optimal policy is not demonstrated, and the paper discloses that it ignores new wealth creation.

The authors' response. Disclosed as a scope limitation: the paper calls itself 'a theory of the taxation of rentiers' (p. 62). Saez & Stantcheva (2018, p. 137) later note that adding uncertainty is one way to obtain finite (rather than infinite) long-run elasticities.

Why it matters, in detail. The rate of decumulation is στρ, so time scales go as roughly 1/τ (about ×4 from τ = 1 to τ = 0.25). The finite-horizon critical σα also converges more slowly at low τ (F3). A politically capped τ therefore makes the redistributive effect much slower, even though the end point is qualitatively the same.

Evidence. Conclusion p. 72; p. 67 ('As long as τ > 0, none of our results stated in Propositions 2 and 3 are sensitive to the level of τ'); p. 70 ('apply for any τ > 0, no matter how small'). WP Table 1 Panel A: ρa* = 0.049/0.073/0.155 for τ = 0.25/0.5/1 (α = 1.5, σ = 0.25). Re-optimised P99 exit times: 191/93/47 years for τ = 0.25/0.5/1 (τ = 0.5 from Claude). Top-1% share after 100 years: 17.0% at τ = 0.25 vs 4.2% at τ = 1 (scripts/dynamics.py). The adjudicator's closed form confirms that the limiting knife-edge is τ-independent.

Why this severity. Read narrowly as a statement about the long-run classification the sentence is correct. The overstatement concerns the policy-relevance argument, which is interpretive.

Why it matters, in detail. Panel A cannot be certified without the code. Two independent re-implementations agree with each other but not with several WP cells, most strongly at τ = 0.25 (a factor of 3–5). Combined with F1, this suggests a coding or tabulation problem in the WP simulations that went beyond Panel B.

Evidence. WP Section 5, fns 29–31. Matching column (0) (ρP99 = 1.46) requires a Pareto tail mass of ≈ 8.5%. A density continuous at ā0 implies 40% and ρP99 = 1.95 (Claude and Codex agree; adjudicator confirmed analytically). Adjudicator re-optimisation (tail mass 8.5% / 40%), against the WP: ρa*∞ = 0.250/0.150 vs 0.049 (σ=0.25, τ=0.25); 0.228/0.213 vs 0.155 (σ=0.25, τ=1); 0.642/0.603 vs 0.697 (σ=0.45, τ=1); 2.000 (8.5%) vs 3.216 (σ=0.6, τ=1). These match Claude's replication to within ~1–2%. Capping all thresholds at the WP value lowers welfare: ΔW = −0.0132/−0.0046 (σ=0.25, τ=0.25) and −0.0075/−0.0012 (σ=0.25, τ=1). Claude reports 13 of 18 finite cells within a factor of 2 and 7 of 9 'infinite' cells as corners.

Why this severity. Only the WP's illustrative threshold levels are affected; the qualitative pattern (finite thresholds rising in σ and α when σα < 1) reproduces, and the table is absent from the published article.

Why it matters, in detail. The reversed inequality sits in the paragraph that motivates Proposition 2, but eq. (16), the proposition and the proof are all stated correctly. The ρδa* slip does not propagate into eqs. (13)–(14).

Evidence. Published pp. 68–70 rendered and checked; WP text (NBER PDF). The correct inequality appears in eq. (16), Proposition 2's first bullet and Appendix A (eq. 29).

Why this severity. These are typographical; the adjacent equations and the formal statements are correct, so no result is affected.

Why it matters, in detail. If G and p are scaled by the same constant, β is unchanged but the printed g changes, so the cutoff σα < 1 − ḡ would depend on an arbitrary normalisation. The appendix's g = β/(1 − dTax/da0) is the intended object.

Evidence. Published p. 70 (right column) and p. 71 (top left); Appendix A p. 73 (rendered pages checked).

Why this severity. This is a local printing error. Proposition 3 and its proof use the correct normalisation, which the adjudicator checked algebraically (ρa∞ = σy/(1−ḡ−σ)).

Why it matters, in detail. The misprint reverses the capital-stock ordering in the all-taxed steady state. The intended Piketty (2001a) steady-state results, and the conjecture that the optimal policy returns the economy to f′(k) = ρ, are unaffected; the conjecture remains unproven (see F3 on transitional decumulation).

Evidence. Published p. 71 (rendered page checked); fn 41's ordering a* ∈ (k′, k) is consistent only with the corrected formula. Codex's illustration with f(k) = k^0.3, ρ = 0.05, τ = 0.5: the correct k′ = 4.80 vs untaxed k = 12.93, while the misprint would give 34.8 (adjudicator re-computed).

Why this severity. The intended statements are clear from the adjacent equations, and the closed-economy extension is explicitly labelled a conjecture.

Why it matters, in detail. A lower α (a thicker tail) makes σα < 1 easier to satisfy, so the corrections would not hurt the paper. Fig. 6 measures capital income excluding realised gains, which proxies the wealth tail only under uniform returns. The 1996 PUF also excludes 113 extreme-value records (IRS 1996 PUF documentation, 'Disclosure avoidance procedures'), so the very top is not directly observed.

Evidence. Published p. 69 and fn 32 (rendered page); PS (2003) QJE fn 21; WP fn 26 cites PS (2001). Claude's computation from PS (2001) Table C1, spot-checked by the adjudicator (1927: α = 1.285 from top 1% to 0.1%, 1.561 from 0.1% to 0.01%): from top 0.1% to 0.01% the mean is 1.59 (range 1.40–1.70; 85% of years in [1.5, 1.7]); from top 1% to 0.1% it is 1.41 (1.24–1.56). Claude also finds α from Saez–Zucman (2016) shares of 1.31–2.36, ~75% of years in [1.5, 2] and ≈1.4 in 2012 (not re-run by the adjudicator).

Why this severity. The essential empirical point, a thick and stable Pareto tail with α near 1.5, is supported, and the imprecisions mostly run toward lower α, which favours the paper's condition.

Why it matters, in detail. The constancy of a*_t for σα < 1 is a corner of the imposed monotonicity constraint (a*′ ≥ 0), which is acknowledged in the proof. Without that constraint the government would want to lower thresholds, re-taxing exited wealth (see F13). The 'Pareto parameter α' in the σα > 1 case should be read as 'tail exponent α'.

Evidence. Appendix A pp. 72–73; Sections 3.3, 4.1, 5.1, 5.2. The adjudicator's numerical and closed-form checks support the main formulas and the τ < 1 knife-edge: the asymptotic revenue effect is ∝ τρ(ασ−1)/(q+αλ).

Why this severity. Heuristic asymptotic arguments of this kind are standard in applied optimal-tax theory, and every numerical check supports the stated results; no counterexample to the optimum is known.

Why it matters, in detail. The 'analogy to Chamley–Judd' in C1 is to a result that does not hold in those models for IES ≤ 1. Saez's vanishing of the taxed fraction is proven separately within his model and remains valid there.

Evidence. Published p. 62; Straub & Werning (2014) NBER WP 20441, Introduction and fn 1 (text checked); published as AER 110(1): 86–119 (2020). Saez's own Proposition 1 (σ < 1: tax the rich ever longer) is consistent with Straub & Werning's mechanism. In Saez's own uniform-linear model the optimal T is finite, because taxing forever would generate an unbounded human-wealth effect for low-wealth individuals (adjudicator's reasoning from eqs. 6–7; numerically T* = 8.9–15.5 yrs in the WP calibration, scripts/linear_vs_progressive.py).

Why this severity. This affects framing only: none of the paper's own results relies on the Chamley–Judd zero-tax result, and Straub & Werning postdate the paper.

Why it matters, in detail. Readers who relied on the widely circulated 2002 WP (cited, e.g., by Farhi, Sleet, Werning & Yeltekin 2012) received stronger empirical and policy conclusions than the analysis supports. See F1 and F2 on the numbers underlying them.

Evidence. NBER WP 9046 text (printed pages) pp. 3–4, 7, 28, 30–33; the published version hedges or removes each of these (pp. 62, 70–72).

Why this severity. These claims are confined to the WP; the version of record corrects them.

The authors' response. Implicitly corrected in the 2013 published version.

Why it matters, in detail. Re-optimisation at t > 0 reproduces the t = 0 incentive to tax inelastic wealth, so the announced path is not time-consistent. Credibility matters most precisely because the redistributive effect builds over centuries (F3).

Evidence. Sections 2.2 and 4; fn 23 p. 67; re-optimised paths (scripts/out/*.json). Farhi, Sleet, Werning & Yeltekin (2012, REStud 79(4)) analyse nonlinear capital taxation without commitment and describe this paper (2002 WP) as the commitment case.

Why this severity. Commitment is the standard benchmark assumption in the Ramsey literature that the paper extends. It is an undisclosed caveat for the policy reading, not an error.

Why it matters, in detail. Any additional base elasticity at the top would lower the revenue gain from taxing the super-rich longer, and so would tighten the truncation condition below σα < 1. The size of the effect is unknown.

Evidence. pp. 62–63, fn 9, Conclusion pp. 71–72. Corpus checklist (optimal-tax-behavior): 'One elasticity standing in for every margin'. Related evidence of large responses of top wealth to wealth taxes: Brülhart, Gruber, Krapf & Schmidheiny (2022, AEJ:EP; corpus optimal-tax-behavior-19).

Why this severity. This is a generic scope limitation of a deliberately stylised model, partly disclosed, with no demonstrated effect on the condition. The adjudicator downgrades it from Claude's 'moderate'.