{"id":"460e7cc8-81b5-4fce-83a1-f8d6ec503531","arxiv_id":"2601.04438","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A power transformation W=V^(1-rho) makes the Epstein-Zin Euler equation invertible in closed form, yielding a root-finding-free endogenous grid method with large speed and accuracy gains.","lead":"A power transformation makes the Epstein-Zin consumption-savings problem compatible with the fast endogenous grid method, eliminating root-finding. In benchmarks the new EZ-EGM is 10–50x faster than value-function iteration or time iteration at equal grid size, and an order of magnitude more accurate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's convergence is asserted, not proved, and the paper's own ρ>1 robustness tests may lie outside the region where a well-defined infinite-horizon solution exists.","rationale":"The reader's weakest assumption points to the lack of a convergence proof for Algorithm 1. I agree that this is the central soft spot, but I would sharpen it: the paper's own Appendix F extends the method to ρ>1 without noticing that the power transformation turns the Bellman equation into a minimization, and does so in a parameter region where the paper's stated sufficient condition βR^θ<1 is violated. That makes the convergence issue more concrete than a generic missing proof. The benchmark at ρ=2/3 appears internally consistent and the core Euler-inversion idea is correct, so this does not overturn the main contribution; it does, however, reinforce the conditional nature of the acceptance. The proposed numerical test would settle whether the Appendix F claim is an artifact of the finite grid or a genuine solution of the stated infinite-horizon problem.","tokens_in":11864,"tokens_out":22506,"duration_ms":227161,"concrete_test":"Run Algorithm 1 for the Appendix F case ρ=2, γ=10, β=0.96, R=1.02 on a 300-point asset grid. Separately solve the same infinite-horizon model with globally convergent VFI on grids whose upper bounds are 20, 200, and 2000 times mean income. Check (i) whether the EGM policy matches the VFI policy to a max Euler error of 1e-4; (ii) whether the VFI value function stabilizes as the upper bound grows by more than 1e-6. If VFI fails to stabilize or EGM differs, the ρ>1 robustness claim and the unconditional convergence claim fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algebraic derivation of the inverted Euler equation (Proposition 2.1) is sound for the benchmark case ρ<1. The load-bearing risk is in the iteration used to find the infinite-horizon fixed point. Algorithm 1 alternately updates consumption through the Euler inversion and updates the value function with one Bellman step, stopping when ||c^n−c^(n−1)||<ε. No contraction, monotonicity, or global-convergence argument is given; the paper asserts 'standard conditions apply' without identifying them for this hybrid EGM-plus-Bellman operator. This is not a purely formal gap: for ρ>1 the transformation W=V^(1−ρ) reverses the order of the max, so Eq. (3) should be a minimization, not a maximization, and the first-order condition is necessary but not sufficient without a convexity/concavity result. Moreover, Appendix F claims convergence for ρ∈{1.1,1.5,2,3} with γ=10, β=0.96, R=1.02, but footnote 2's sufficient existence condition βR^θ<1 is violated in these cases (e.g., ρ=1.1 gives θ=90 and βR^θ≈5.7), suggesting the finite-grid runs may be solving a truncated or ill-posed problem. If the iteration is not globally convergent, the Section 4 speed/accuracy claims are evidence for one calibration, not a property of the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a power transformation W = V^{1-rho} that converts the Epstein-Zin Bellman equation into an additive form, derives a closed-form inverted Euler equation c(a,z) = (beta R mu^{1-theta} Xi)^{-1/rho} (Proposition 2.1), and presents Algorithm 1 (EZ-EGM), a root-finding-free endogenous grid method. The authors benchmark EZ-EGM against value function iteration and time iteration, reporting speedups of 10-50x at equal grid size, one order-of-magnitude accuracy improvements, and 150-630x speedups at equal accuracy. The paper also claims the method extends to rho > 1 (EIS < 1) with numerical verification in Appendix F.","tokens_in":12217,"tokens_out":11815,"duration_ms":106299,"significance":"The central insight is attractive and, if correct, practically important: it would make the endogenous grid method available for the widely used Epstein-Zin preferences, substantially accelerating structural estimation and model evaluation. The derivation of the inverted Euler equation is clean and the baseline calibration is standard. The paper also includes thoughtful benchmarking, an equal-accuracy comparison, and an independent welfare-cost check (below 0.1%). These are genuine strengths. However, the absence of a convergence proof for the hybrid EGM-plus-Bellman iteration and unresolved issues in the rho > 1 robustness analysis prevent the claims from being fully established as a general property of the method.","major_comments":[{"comment":"Convergence of Algorithm 1 to the infinite-horizon fixed point is asserted, not proved. The iteration is a hybrid: the policy is updated by the analytic Euler inversion (lines 7-11), then the value function is updated by exactly one Bellman evaluation (lines 12-16), and the loop stops when ||c^n - c^{n-1}|| < epsilon. No contraction mapping, monotonicity, or policy-iteration theorem is established for this operator. The stopping criterion does not by itself imply that the limit is a fixed point of the Bellman equation, and non-monotone endogenous grids are only addressed by a heuristic upper-envelope fix. Since all Section 4 speed/accuracy claims rest on this iteration, this is load-bearing. Please either supply a convergence theorem (for example by adapting policy-iteration theory for recursive preferences) or explicitly present the algorithm as heuristic and temper the benchmark claims","section":"§3, Algorithm 1, lines 11-18"},{"comment":"The robustness analysis for rho > 1 conflicts with the paper's own existence condition. Footnote 2 states that a sufficient condition for a well-defined infinite-horizon problem is beta R^theta < 1. For rho > 1 and gamma = 10, theta = (1-gamma)/(1-rho) is positive (e.g., theta = 90 for rho = 1.1), and beta R^theta is approximately 5.7, violating the condition. Yet Appendix F reports convergence with Euler errors near -5 for rho in {1.1, 1.5, 2, 3}. This suggests the finite-grid runs may be solving a truncated or ill-posed infinite-horizon problem. The paper must clarify whether a genuine infinite-horizon solution exists for these parameter values, or whether the reported numbers are for a modified (e.g., bounded-state or finite-horizon) problem.","section":"Appendix F vs Footnote 2"},{"comment":"For rho > 1, W = V^{1-rho} is decreasing in V, so the transformed Bellman equation should be a minimization, not a maximization. The paper acknowledges this only in a footnote and continues to write 'max' in Eq. (3). The first-order condition is necessary, but the paper provides no convexity/concavity argument to show it is sufficient for the optimum in the minimization problem. Since Proposition 2.1 is the analytical foundation of the method, the statement of the transformed Bellman equation for rho > 1 should be corrected and the sufficiency conditions stated, or the formal claims should be restricted to rho < 1 with rho > 1 presented only as a numerical extension.","section":"§2.1, Eq. (3) and Footnote 3"}],"minor_comments":[{"comment":"Appending the point (0,0) at the constraints is not sufficient to represent the borrowing-constrained policy. For m below the first endogenous grid point, the constrained policy is c(m,z) = m, not a linear interpolation through (0,0). Please specify that the constrained segment is imposed separately, as stated later in the text.","section":"Algorithm 1, line 10"},{"comment":"The header 'Euler Error' in Table 1 is ambiguous; it should be labeled 'Mean Euler Error (log10)' to match Table 2.","section":"Table 1 and Table 2"},{"comment":"The statement that V > 0 follows by induction from a terminal condition does not directly apply to the infinite-horizon problem; the limiting argument should be stated explicitly.","section":"Footnote 2"},{"comment":"The appendix reports results for rho in {0.5, 0.9, 1.1, 1.5, 2, 3} but does not provide a table of errors or iteration counts. Including the detailed numbers would help readers assess the robustness claim.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the baseline (rho < 1) derivation appears sound. The main risk is that the algorithm's convergence is not proved and the rho > 1 robustness tests seem to violate the paper's own existence condition. These issues are fixable in revision, but they are load-bearing, so I recommend major revision rather than acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution is real. Lujan shows that the power transform W=V^(1-rho) converts the Epstein-Zin Bellman equation into an additive form, and the Euler equation inverts in closed form (Eq. 9). Within the cited literature, that is new, and Proposition 2.1 is solid. The benchmark numbers are plausible and the equal-accuracy comparison is the right way to frame the speedups. The welfare-cost check below 0.1% gives an independent ground beyond the Euler-error metric. This is not a repackaging of known tricks; the application to EGM is a genuine step.\n\nSoft spots, in roughly increasing order. First, no public code. Hand-coded JAX benchmarks, even if honest, are not independently checkable. Releasing code would settle it. Second, Algorithm 1's convergence is asserted, not proved. The hybrid EGM-plus-Bellman iteration may well be a contraction under standard conditions, but the paper doesn't say which conditions, and the stopping rule monitors only the policy. Monotonicity of the endogenous grid is assumed for 'typical calibrations' with an upper-envelope fix; if that fix fails in some regime, the speed claims are one-calibration evidence. Third, the rho>1 robustness tests look like they may sit outside the region where the infinite-horizon problem is well-defined. Footnote 2 gives a sufficient condition beta R^theta < 1; for rho=1.1, theta=90 and beta R^theta is about 5.7. The paper reports convergence anyway. Maybe the condition is only sufficient, but then the appendix should say so explicitly, and if the finite-grid runs are approximating a truncated problem, that needs to be disclosed. Also, 'sidesteps the speed-accuracy tradeoff entirely' is too strong: grid size remains a tradeoff axis. What EGM does is push the efficient frontier out, not remove the tradeoff.\n\nNet: this is a serious methods paper. A referee should ask for code, a convergence statement or a precise reference to existing conditions, and a careful treatment of the rho>1 region. The central derivation holds up, and the benchmark evidence, while not reproducible yet, is consistent and well presented. I would take it seriously, and I'd cite the transformation even before code appears.","headline":"A genuine extension of EGM to Epstein-Zin preferences with a clean closed-form Euler inversion and credible benchmark speedups; the main repair needed is a convergence proof for the hybrid iteration and a check of the rho>1 robustness claims.","tokens_in":766,"tokens_out":824,"would_cite":true,"duration_ms":30376,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B62","90C39"],"pacs":[],"model":"deepseek-v4-flash","headline":"A power transformation, W=V^(1-ρ), converts the Epstein-Zin Bellman equation into additive form, permitting a closed-form Euler inversion and a root-finding-free endogenous grid method that is 10-50x faster than value function iteration.","keywords":["endogenous grid method","Epstein-Zin preferences","recursive utility","consumption-savings","Euler equation inversion","power transformation","value function iteration","time iteration"],"falsifier":"Run EZ-EGM on a calibration with very patient agents (βR ≥ 1) or with a binding borrowing constraint and high curvature (e.g., ρ=3, γ=10) and check whether the iteration converges to the same fixed point as VFI/TI; specifically, check whether the endogenous grid remains monotone after the upper-envelope step and whether the mean Euler error stays near 1e-5. A failure to converge, or a grid that the upper-envelope cannot repair, would falsify the claim that EGM sidesteps the speed-accuracy tradeoff for Epstein-Zin preferences.","tokens_in":11755,"feed_emoji":"⚡","tokens_out":6775,"duration_ms":50242,"temperature":0.7,"pith_summary":"The paper claims that the apparent incompatibility between the endogenous grid method and Epstein-Zin recursive utility is resolved by a power transformation of the value function. Raising V to the power 1-ρ turns the CES Bellman recursion into an additive one, and the resulting Euler equation can be inverted analytically for consumption as a function of end-of-period assets. This yields an algorithm with no root-finding that tracks consumption and transformed value simultaneously. Benchmarks in a standard consumption-savings model show speedups of one to two orders of magnitude over value function iteration and time iteration at equal grid size, with Euler errors near 1e-5 and up to three orders of magnitude speedup holding accuracy constant. If correct, this makes recursive-utility models dramatically cheaper to estimate and solve.","feed_headline":"One power transform makes Epstein-Zin models 50x faster to solve","feed_subtitle":"W=V^(1-ρ) makes the Bellman equation additive, giving closed-form Euler inversion with no root-finding.","key_machinery":"The load-bearing object is the power transformation W=V^(1-ρ) and the additive Bellman equation it induces. The certainty equivalent μ(a,z)=(E[W(m',z')^θ])^(1/θ) with θ=(1-γ)/(1-ρ) and the expectation Ξ(a,z)=E[W^(θ-1) c^(-ρ)] together make the Euler equation c^(-ρ)=βR μ^(1-θ) Ξ invertible in closed form (Proposition 2.1). This inversion is what lets EGM compute consumption directly from the end-of-period asset grid, sidestepping root-finding. The algorithm is a hybrid of EGM and Howard policy iteration, updating c and V/W together until both stabilize; the borrowing constraint is anchored by appending (0,0), and non-monotonic grids are handled with the upper envelope.","core_discovery":"The central claim is that the transformation W(m,z)=V(m,z)^(1-ρ) (with θ=(1-γ)/(1-ρ)) converts the Epstein-Zin Bellman equation into the additive form W(m,z)=max_c[(1-β)c^(1-ρ)+β μ(m-c,z)], where μ is the power-mean certainty equivalent of next-period transformed value. Differentiation then yields an Euler equation c^(-ρ)=βR μ^(1-θ) Ξ that is invertible in closed form: c(a,z)=(βR μ^(1-θ) Ξ)^(-1/ρ). Given an exogenous grid over end-of-period assets, this makes consumption computable directly without numerical root-finding, and the endogenous grid is recovered as m=c+a. The algorithm iterates on both the policy and the value function until convergence, and the paper reports large speed and acc","pith_inferences":["The same inversion strategy may apply to other recursive utility forms whose Bellman equations become additive under a suitable transform — for instance, models with multiple assets or portfolio choice — since the essential requirement is only that the Euler equation be invertible in closed form after the transformation.","Because the endogenous grid evaluates the Euler equation exactly at the grid points, combining EZ-EGM with higher-order or adaptive interpolation could push accuracy beyond the reported 1e-5 at negligible additional cost; the paper does not explore this.","The reported speedups are hardware- and implementation-dependent (JAX on CPU), but the structural advantage of eliminating root-finding is implementation-independent; a careful comparison on GPU or with alternative interpolation methods could change the constants but not the qualitative ordering.","The convergence of the hybrid EGM-plus-Bellman iteration is asserted rather than proven; a formal contraction argument or a counterexample for a pathological calibration would determine whether the approach applies beyond the parameterizations tested."],"forward_implications":["An analyst solving an infinite-horizon consumption-savings model with Epstein-Zin preferences can replace root-finding-based time iteration or VFI with EZ-EGM and obtain mean Euler errors around 1e-5 at roughly 1/10 to 1/50 the runtime at equal grid size.","Holding solution accuracy constant, EZ-EGM is 150 to 630 times faster than VFI in the reported benchmarks, turning previously intractable estimation or uncertainty-quantification tasks into routine computations.","The closed-form Euler inversion eliminates the speed-accuracy tradeoff that characterizes VFI and time iteration: the endogenous grid places evaluation points exactly where the Euler equation holds, so accurate-mode precision is achieved at fast-mode cost.","The method extends to the logarithmic limit ρ=1 (unit EIS) and to the risk-sensitive preferences of Hansen-Sargent, and is verified numerically for ρ>1 (EIS<1), covering the empirically relevant low-EIS range.","Non-monotonic endogenous grids, when they occur, can be repaired with the upper-envelope procedure, so the approach does not require a monotonicity assumption for the baseline implementation."],"fun_headline_variants":["Power transform eliminates root-finding in Epstein-Zin EGM","One power transform makes Epstein-Zin EGM root-free","Closed-form Euler inversion for Epstein-Zin via power transform","Epstein-Zin models solved 100x faster with one transform","EGM for Epstein-Zin: closed-form inversion, orders of magnitude faster"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Algorithm 1's convergence to the infinite-horizon fixed point is assumed to hold under 'standard conditions' without a proof that the EGM-plus-Bellman iteration is a contraction; if the iteration fails to converge or the upper-envelope repair cannot fix non-monotonic endogenous grids in some economically relevant region, the reported speed and accuracy gains would not generalize beyond the calibrated case.","fun_headline_variants_meta":{"raw":{"variants":["Power transform eliminates root-finding in Epstein-Zin EGM","One power transform makes Epstein-Zin EGM root-free","Closed-form Euler inversion for Epstein-Zin via power transform","Epstein-Zin models solved 100x faster with one transform","EGM for Epstein-Zin: closed-form inversion, orders of magnitude faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1740,"prompt_tokens":663,"completion_tokens":1077,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":1003}},"tokens_in":407,"tokens_out":1077,"duration_ms":8249,"temperature":1.0,"reasoning_tokens":1003,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T12:02:39.158674+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EZ-EGM on a calibration with very patient agents (βR ≥ 1) or with a binding borrowing constraint and high curvature (e.g., ρ=3, γ=10) and check whether the iteration converges to the same fixed point as VFI/TI; specifically, check whether the endogenous grid remains monotone after the upper-envelope step and whether the mean Euler error stays near 1e-5. A failure to converge, or a grid that the upper-envelope cannot repair, would falsify the claim that EGM sidesteps the speed-accuracy tradeoff for Epstein-Zin preferences.","supporting_citations":[],"review_version":1}