A Simpler Theory of Optimal Capital Taxation

Why it matters, in detail. In the paper's own concave wealth-in-utility model, positive steady-state elasticity and saddle-path stability are the same condition (correspondence principle). Under it the unstable root mu_u = delta + tau r - mu_s exceeds delta, so lim_{T->inf} e_K^a is finite: 0.17-2.43 in the cases computed, 50-100% of e_K, and exactly e_K at tau = 0. The T -> infinity anticipated criterion therefore gives tau_K = (1 - g_K)/(1 - g_K + e_K^a(inf)) > 0 whenever g_K < 1. In the cases computed this is at least as high as the steady-state-approach rate, not zero. Divergence occurs only in the limit a'' -> 0, where e_K itself is infinite (Chamley-Judd). The paper's contrast between a zero-tax anticipated approach and its own steady-state approach therefore does not hold in its own model.

Evidence. Adjudicator code (scripts/anticipation_check.py; output/anticipation_check.txt), exact linear saddle-path solution plus nonlinear two-point BVP. Auditor B's log-log case (r = .04, delta = .08, b = .5, tau = .4): roots -0.0316 and 0.1276 (> delta); steady-state e_K = 0.667; e_K^a(T) = 0.189, 0.409, 0.477, 0.503, 0.506 at T = 0, 25, 50, 100, 200; limit 0.506024 (reproduces B to six decimals); nonlinear BVP with dtau = +/-1e-4 gives 0.40949/0.40952 vs linear 0.40950 at T = 25 and 0.50329/0.50333 vs 0.50331 at T = 100. Continuous-time analogue of Auditor A's base case (log-log, delta = 5%, r = 4%, tau = 20%, k/y = 5): e_K = 2.222, limit 1.951 (A's discrete-time version: 1.758 vs 2.162). Three further saddle-stable CRRA cases: limits 2.431, 0.510, 0.169 vs e_K 4.861, 0.521, 0.176. At tau = 0 the limit equals e_K exactly (2.000 vs 2.000). Two-type economy (scripts/anticipated_positive_tax.py): at tau = 0, dSWF/dtau is proportional to +0.464, and the stationary tax under the T -> infinity anticipated criterion is 47.38% (g_K = 0.634, e_K^a = 0.406), reproducing Auditor B exactly. Extension (exploratory, likely): Table 1's Judd (1985) endogenous-discount row rests on Piketty-Saez (2013b, Online App. A.3), which uses the same flat-shift step ('a future reform shifts the entire (pre-reform) consumption path'); in a saddle-stable Uzawa specification (discount rate increasing in c) the adjudicator also finds a finite e_K^a (equal to e_K at tau = 0, tending to 0 for tau > 0; linear 7.5017 vs nonlinear 7.5012-7.5022 at T = 100) (scripts/uzawa_anticipation.py, uzawa_nonlinear_check.py).

Why this severity. It reverses a stated proposition and a Table 1 entry and removes one of the paper's two main reasons for rejecting the anticipated-reform approach in favour of its new steady-state approach. The sufficient-statistics formulas (Props. 8-10), the simple model and the U.S. calculations are untouched, and the paper's other objections to the anticipated approach remain: weight concentrated on the most patient agents, and the unrealism of reforms announced infinitely far ahead.

Why it matters, in detail. Equal asymptotic top rates require e_K/e_L = a_L/a_K: 1.16 with the stated parameters, 1.16 (P99.99) to 1.31 ($10M) in the reproduction depending on the threshold (1.21 at P99.9), 1.10 excluding realized gains in 2007, and 0.99 in 2000. The sign of a_K < a_L holds in 17/17 years and is reinforced by reclassifying business income. The ranking of rates is not robust to plausible elasticity differences. JJKZ's long-run e_K of about 1.15 for the very wealthy, with a top ETI of 0.25, implies about 39% vs 71%, and cross-base responses push the capital rate lower still. The result is also specific to revenue-maximizing top rates (G -> 0 at the top of both distributions).

Evidence. Adjudicator synthetic-file reproduction (scripts/synthetic_facts.py; output/synthetic_facts.txt), 2007: - Top-bracket Pareto a for labor/capital is 1.619/1.336 at P99.9 (ratio 1.21), 1.598/1.315 at $1M (1.22) and 1.94/1.48 at $10M (1.31). - Excluding realized gains, a_K(P99.9) = 1.474 and the ratio falls to 1.10. - Over 1993-2012 the ratio is 1.14-1.38 with gains and 0.99 (2000) to 1.30 without. - The 2007 capital share, 26.3%, is the 1993-2012 maximum. Digitized Fig. 5 rates at $1M imply alpha/(1-G) = 1.32 (capital) vs 1.78 (labor) (scripts/digitize_fig4_fig5.py). External evidence (verified in source PDFs): - Jakobsen, Jakobsen, Kleven & Zucman (2020 QJE 135(1)) estimate a long-run elasticity of taxable wealth to the net-of-tax return of 0.58-0.77 (moderately wealthy) and 1.15 (very wealthy). This gives 1/(1+1.38*1.15) = 38.7% for capital vs 71.4% for labor at e_L = 0.25. - Lefebvre, Lehmann & Sicsic (2024, Scand. J. Econ.) find a French capital-income Laffer rate of about 57% without the cross-elasticity and about 43% with it. - Agersnap & Zidar (NBER w27705) find revenue-maximizing capital-gains rates of 38-47%.

Why this severity. The sign a_K < a_L is robust, so the conditional statement survives. But its policy content (a higher top rate on capital) holds only if e_K exceeds e_L by less than about 16% (about 10% without realized gains). The plotted gap is about twice the asymptotic one, and the paper's wording implies a robustness it does not show. This is an important undisclosed caveat on the paper's main numerical result.

The authors' response. No direct response. The paper frames the result as conditional on equal elasticities, reports e = 0.25/0.5/1, notes 'a paucity of empirical estimates on the capital elasticity' (p.130) and calls for estimating cross-elasticities (pp.121, 137). JJKZ (2020), Lefebvre et al. (2024) and Agersnap-Zidar (2020/21) post-date the paper.

Why it matters, in detail. Taxing asset j has a fiscal externality on asset s's base equal to tau^s r^s dk^{m,s}, so cross terms must be weighted by relative tax bases. Eq. (20) mis-scales each cross term by the factor r^j/r^s: it understates the term when r^s > r^j and overstates it when r^s < r^j. The qualitative implications are unchanged: a higher own elasticity, a higher g_K^j, or complementarity with taxed assets each lowers tau^j. Quantitative use of eq. (20) for assets with different returns (e.g., housing vs equities) would be biased.

Evidence. Checked visually on the rendered p.128. Adjudicator test (scripts/formula_checks.py): returns 3% and 7%, tau_2 = 35%, complementary quadratic utility of wealth. Direct revenue maximization gives tau_1 = 67.50%, equal to the corrected formula; the fixed point of printed eq. (20) is 79.17%; with equal returns both give 57.50%. Auditor B's independent example: 56.67% (true) vs 63.33% (printed).

Why this severity. A stated proposition, offered as a formula 'easily taken to the data', is wrong in exactly the case it is designed for (unequal returns with cross-effects); in simple examples it moves the rate by 7-12 pp. It does not affect the single-asset formulas, the U.S. calculations or the qualitative lessons drawn from Prop. 7.

Why it matters, in detail. Readers comparing Fig. 5 with the text see a 6.0-7.2 pp capital-labor gap at $1M, while the asymptotic rates implied by the stated parameters differ by 2.9-3.6 pp. The $1M values reflect local alpha below the tail. Because a_K and a_L both drift upward above $1M, the stated 'constant' parameters approximate only roughly the $0.5-2M range.

Evidence. Digitization by scripts/digitize_fig4_fig5.py, which matches Auditor A's to three decimals: - Fig. 4 at $1M: labor 1.583 (top-bracket) / 1.695 (local), capital 1.309 / 1.270, total about 1.385 / 1.354. - Fig. 5 at $1M: capital 75.2/60.2/43.1%, labor 69.2/52.9/36.0%. Synthetic file (scripts/synthetic_facts.py): - Top-bracket a at $1M is 1.598/1.315/1.393 with the extreme-record correction. - Without the correction capital is 1.377, but labor (1.64) and total (1.45) then no longer match the text, so the text's three values do not come from one consistent calculation in the reproduction.

Why this severity. This is a presentation inconsistency. It changes top rates by about 1-2.5 pp and makes the plotted capital-labor gap (6-7 pp) about twice the stated-parameter gap (2.9-3.6 pp); the direction of the ranking is unaffected.

Why it matters, in detail. Since dV/dc(t) = exp(-int delta)[u_c - delta'(c) V_t], the printed weight mis-states relative g_i, and hence g_K, whenever u_i and delta_i' vary across individuals. The stability requirement delta'(c) > 0 (increasing marginal impatience) is the opposite of the paper's stated sign.

Evidence. Rendered pp.137 and App. p.4. Finite differences (scripts/formula_checks.py): - With delta' < 0: 30.19 (matches the minus-sign formula) vs printed 2.48. - With delta' > 0: 2.20 vs printed 26.78. - A consumption bump at t = 10: numerical 0.878, correct 0.888, printed 0.073. Auditor B independently: 12 vs printed 8. Linearization (scripts/uzawa_anticipation.py) with delta' < 0: roots 0.016, 0.016 and 0.04, none stable.

Why this severity. This is a derivation slip in a section that connects the paper to earlier literature. Once the sign is corrected, the formula keeps its sufficient-statistics structure (eq. 25/27 with a redefined g_i), and nothing else in the paper uses it.

Why it matters, in detail. The paper attributes the zero limit to infinite anticipation elasticities around a steady state. Straub & Werning's counter-examples involve non-convergence to an interior steady state or binding bounds, which lie outside the paper's local analysis. As a general characterization of what Chamley-Judd imply, the summary is incorrect.

Evidence. Text search of the version files (rev0 line 218, rev1, rev2: present; rev3, the Oct 2017 version and JPubE: absent). Straub & Werning WP 20441: Props. 2-3 (pp.10-11) and Prop. 7 (pp.22-23). Crossref: AER 110(1):86-119, doi 10.1257/aer.20150210.

Why this severity. This concerns how the paper describes prior literature in an extension. The paper's local steady-state statements are not directly contradicted, and the omitted results, if anything, support the paper's own view that zero capital taxes are fragile.

The authors' response. Earlier drafts (NBER rev0-rev2) acknowledged Straub & Werning in a footnote; the citation disappeared when the related-work section was removed in the Sept 2017 revision.

Why it matters, in detail. With SYZZ's human-capital share of pass-through profits, capital is a majority of income only above P99.9 (0.53-0.54) and about 73% in the top 0.01%; between P99 and P99.9 it is 0.29-0.42. a_K falls under reclassification, so the a_K < a_L ranking and the optimal-rate results are robust. The 'top income is mostly capital' framing is not robust below the top 0.1%.

Evidence. Synthetic 2007 file (scripts/synthetic_facts.py), capital fraction of total income for P99-99.5 / P99.5-99.9 / P99.9-99.99 / P99.99-100: - baseline: 0.43/0.57/0.69/0.83; - 75% of partnership plus S-corporation profit treated as labor (Smith, Yagan, Zidar & Zwick 2019): 0.33/0.42/0.54/0.73; - additionally 70% of Schedule C/F treated as labor: 0.29/0.39/0.53/0.73; - realized gains excluded: 0.33/0.43/0.52/0.59. Auditor B (IRS Table 1.4, AGI of $10M or more) finds 83.7% falling to 72.6% under the SYZZ split. Reclassification lowers a_K at $1M to 1.28, which widens rather than narrows the tail gap.

Why this severity. The descriptive fact holds for the top 0.1% under every classification tried. Only its breadth (P99-P99.9) and the related normative remark depend on the fiscal classification, which the paper discloses. The optimal-tax ranking is unaffected.

The authors' response. Fn. 26-27 acknowledge that business profits mix labor and capital income and that tax-based components differ from economic income; the paper notes that any separate capital tax must use tax-based components.

Why it matters, in detail. With a progressive consumption tax, working more raises the consumption tax on the income stream from wealth and so lowers real wealth. The correct labor wedge is therefore larger than 1 - T_C', not smaller, so the appendix misstates the direction in which the labor-tax equivalence breaks down under elastic labor.

Evidence. Rendered appendix p.14. Adjudicator finite difference (scripts/formula_checks.py; q = .04, delta = .09, k = 10, z = 1, T_C = .05y^2) gives 0.810, equal to the corrected expression, against 0.910 for the printed one. Auditor B obtained the same numbers independently.

Why this severity. The error is confined to the appendix. Prop. 5 concerns linear consumption taxes and is unaffected, and the appendix already says the equivalence is not exact with elastic labor.

Why it matters, in detail. With c >= 0, upward adjustments become most-rapid-approach paths (c = 0 until k reaches its target). Anticipated/unanticipated equivalence then holds only for reforms that lower desired wealth, and fn. 11's levy avoidance fails because wealth cannot be rebuilt instantly after the levy.

Evidence. pp.122 and 126-127; fn. 11. The static equivalent U_i = c_i + a_i(k_i) - h_i(z_i) + delta_i (k_i^init - k_i) embeds a signed consumption impulse of size k_i^init - k_i (both auditors' derivations).

Why this severity. For small reforms around an interior steady state with c > 0, upward adjustment under c >= 0 takes a time of order dk, a second-order effect on welfare and revenue, so Props. 1-7 are unaffected. The 'no transition from any initial wealth' and levy-avoidance statements hold fully only for reductions in wealth.

Why it matters, in detail. Without R_US, the inversion schedule or the smoothing, Fig. 5 below about the top 1% cannot be replicated exactly; Auditor A's approximate inversion matched the $1M rates within 2.4 pp. The top rates depend only on tail parameters and elasticities, which are reported.

Evidence. pp.129-130. Both auditors searched the author, NBER and publisher materials. Adjudicator reproduction from the IRS-hosted Saez-Zucman synthetic files is close: capital share 26.3% vs 26%; top-1% share 62.5-64.1% vs 63%; P99 thresholds within 0.1-2.5%; Fig. 4 parameters at $1M within 0.015. The undocumented choices still matter: the clipping convention and whether the extreme-record correction is included move the top-1% share by about 2 pp and a_K at $1M between 1.31 and 1.38.

Why this severity. The asymptotic results follow from the reported parameters, and the descriptive facts and top rates reproduce closely from public data. The gaps mainly prevent exact replication of the schedules below the top 1%.

Why it matters, in detail. The correct normalizations appear in eq. (19) and Sec. 3.3.1 (cross-elasticities), in the revenue-loss equation of the Prop. 2 proof, and in Online App. A.1 and A.3 (heterogeneous-delta and top-bracket weights). Readers should use those versions.

Evidence. Rendered pp.133 and 135; text of pp.123, 130, 138-139 and fn. 28; Online App. A.1 and A.3.

Why this severity. Each slip is either corrected elsewhere in the paper or appendix, or harmless; none changes a result.

Why it matters, in detail. The cited works remain identifiable; the errors have no analytical consequence.

Evidence. Crossref records for DOIs 10.2307/1911310, 10.1093/qje/qjr020, 10.3982/ECTA10712 and 10.1086/508032, checked by the adjudicator; the errors were found by both auditors.

Why this severity. These are bibliographic only.

Why it matters, in detail. Alternative mechanisms for fat-tailed wealth, such as idiosyncratic returns, would change transition dynamics and possibly the relevant policy-elasticity concept, but not the structure of the sufficient-statistics formulas.

Evidence. pp.120 and 124; fn. 16. de Vries & Toda (2022 Rev. Income Wealth 68(4)) reproduce fat capital-income tails with capital-income risk. Jakobsen et al. (2020) report that wealthy Danes keep accumulating into old age, which is consistent with utility of residual wealth.

Why this severity. The formulas are expressed in elasticities, so the microfoundation matters mainly for delivering a finite e_K and the absence of transitions. The assumption is plausible and has some later support.

Why it matters, in detail. The example does not illustrate the model's mechanism (taste-driven re-accumulation). If anything it describes a case where the consumption tax's initial-levy component would be effective.

Evidence. SS p.121. NPZ (2018), abstract and the privatization discussion (text lines 607-741 and 1146 of the cached copy).

Why this severity. The example is purely rhetorical and bears on no proposition or calculation.

Why it matters, in detail. Prop. A2 is a fixed point of a local reform criterion with allocation- and direction-dependent weights. Sher's characterization implies such criteria can rank revenue-equivalent policies intransitively, so the 'optimum' need not maximize any coherent objective. The Sec. 4 weights, 1/disposable income fixed at the status quo, are equivalent to Pareto weights under quasi-linear utility and are consistent.

Evidence. Online App. A.4 text, including 'Hence, the weights also need to depend on deltaT(z, rk), the direction of the tax reform'. Sher (2024) Theorems 1-3 as summarized in corpus entry optimal-tax-behavior-13. Auditor B examined this point and did not report it as confirmed.

Why this severity. This affects one of two arguments for comprehensive taxation. The efficiency argument, infinitely elastic income shifting (Prop. 4), and the fixed-weight U.S. calibration are unaffected.