{"id":"20579706-c54b-403e-90c4-ffe600e93e54","arxiv_id":"1908.08800","paper_version":4,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Under an 'eventual discounting' condition, equivalent to a spectral radius below one, the standard optimality theorems of infinite-horizon dynamic programming hold with state-dependent discount factors.","lead":"This paper extends the mathematical theory of dynamic programming to settings where the discount factor changes over time and can even rise above one. It identifies a simple condition, based on the long-run average of the discount process, under which the usual guarantees (optimal policies exist, value iteration converges) still hold.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Necessity claim in A.6 is overstated: pointwise equivalence (51) fails for reducible processes, and the Krein–Rutman step needs unproved compactness of Lβ.","rationale":"Theorem 2.1 itself is well supported: Lemmas A.1 and A.2 correctly transfer eventual discounting of the exogenous process to eventual contractivity of Tσ and T, and the remaining steps are standard fixed-point and maximum-theorem arguments. No flaw was found in the sufficiency proof. The reader's weakest-assumption point about Assumption 2.1 is a limitation, not an error. The genuine issue is the paper's separate claim that eventual discounting 'cannot be significantly weakened,' which is asserted in the abstract and proven only in A.6. That proof contains a concrete false statement: (51) fails pointwise for reducible deterministic Markov chains, as the p/q/r example shows, and the Krein–Rutman step assumes compactness that general Feller kernels do not provide. This is an internal inconsistency rather than a disagreement with outside consensus. The main theorem can survive, but the paper should either prove a correctly formulated uniform necessity result under explicit hypotheses or temper the claim. Hence the verdict should be conditional rather than unconditional acceptance.","tokens_in":31342,"tokens_out":24643,"duration_ms":259963,"concrete_test":"Run the deterministic three-state example: Z={p,q,r}, transitions φ(p)=p, φ(q)=r, φ(r)=q, β(p)=0.5, β(q)=2, β(r)=1.2, and π_t≡1. Compute the discrete matrix of Lβ: on the {q,r} block it is [[0,2],[1.2,0]], so the spectral radius is √2.4>1; meanwhile E^p Σ_t ∏_{i<t}β_i = Σ_t 0.5^t = 2 <∞. This directly falsifies (51) as stated. Also verify whether any hypothesis silently assumed in A.6, such as irreducibility of Q or compactness of Lβ, is present anywhere in the paper; if not, the necessity claim must be reformulated or removed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Appendix A.6 claims (51): for Z compact, β∈bcZ and Q Markov, E^z Σ_{t>0} (∏_{i<t} β_i) π_t < ∞ iff r(Lβ)<1, with 0<a≤π_t≤b. This is the only support for the abstract's statement that eventual discounting 'cannot be significantly weakened.' Two problems arise. First, the proof of the r(Lβ)>1 direction invokes Krein–Rutman after 'By compactness of Lβ'; however, Lβh(z)=β(z)∫ h(z')Q(z,dz') is not compact for general Feller kernels (deterministic transition maps give composition operators), and no compactness or power-compactness hypothesis is imposed. Second, (51) is false pointwise as written. Take Z={p,q,r} with deterministic transitions p→p, q→r, r→q, and β(p)=0.5, β(q)=2, β(r)=1.2. Then the discount matrix on {q,r} has spectral radius √2.4>1, so r(Lβ)>1 and eventual discounting fails, yet E^p Σ_t (∏_{i<t} β_i) π_t = a/(1−0.5)<∞ for any bounded π with π_t≥a. The stated necessary condition therefore fails for initial state p. Theorem 2.1's sufficiency proof does not depend on this appendix, but the paper's strong 'cannot be significantly weakened' claim is not supported as stated; it needs a uniform or sup-over-z formulation, and any necessity result needs explicit conditions under which Lβ has the required spectral properties.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops infinite-horizon dynamic programming with state-dependent discount factors of the form β(Z_t), where {Z_t} is an exogenous Markov process. The central condition is 'eventual discounting': sup_z E^z ∏_{t=0}^{n-1} β_t < 1 for some n. Under a Lipschitz condition on the aggregator (Assumption 2.1), Theorem 2.1 shows that the policy operators T_σ and the Bellman operator T are eventually contracting, that the value function v* is the unique fixed point of T in bcS, that an optimal policy exists, that Bellman's principle of optimality holds, and that both value function iteration and Howard policy iteration converge. The paper connects eventual discounting to the spectral radius of the operator L_β (Proposition 4.1), gives finite-state and AR(1) tests, and extends the theory to unbounded rewards via homogeneous functions and local contractions (Section 5) and to Epstein–Zin preferences (Section 6). The paper also claims a necessity result in Appendix A.6, stated as equivalence between finiteness of expected discounted rewards and r(L_β) < 1.","tokens_in":31700,"tokens_out":10832,"duration_ms":114565,"significance":"If the main results hold, the paper is a useful and nontrivial contribution to dynamic programming theory. Theorem 2.1 is a clean generalization of Blackwell's contraction argument, and the eventual-contraction modulus r_β^n is the right object: the proofs of Lemmas A.1 and A.2 are self-contained and sound. Proposition 4.1, which identifies r(L_β) with the asymptotic geometric mean of the discount process via a local spectral radius argument, is also carefully presented. The paper has clear practical value because it gives a checkable spectral condition for models in which β_t exceeds one with positive probability, and it applies the condition to parameterizations from Christiano et al. (2011), Hills et al. (2019), Hubmer et al. (2020), and Albuquerque et al. (2016). The extensions to homogeneous growth, local contractions, and Epstein–Zin preferences broaden the scope considerably. The main substantive gap is the overstated necessity claim in Appendix A.6 and the corresponding abstract statement that the condition 'cannot be significantly weakened'; this claim needs repair, but it does not affect the soundness of the sufficiency theory.","major_comments":[{"comment":"The claimed equivalence E^z ∑_{t>0} (∏_{i<t} β_i) π_t < ∞ ⇔ r(L_β) < 1 is false pointwise as stated. Consider Z = {p,q,r} with deterministic transitions p→p, q→r, r→q, discount factors β(p) = 1/2, β(q) = 2, β(r) = 6/5, and any reward bounded below by a > 0. Then L_β restricted to {q,r} has spectral radius √(12/5) > 1, so r(L_β) > 1 and eventual discounting fails, yet from initial state p the expected discounted reward is a/(1−1/2) < ∞. A correct necessity statement therefore requires an irreducibility or positive-recurrence assumption, or a uniform condition over initial states, together with additional spectral assumptions that rule out boundary cases. As written, this appendix does not support the abstract's claim that eventual discounting 'cannot be significantly weakened.'","section":"Appendix A.6, Eq. (51)"},{"comment":"The proof invokes Krein–Rutman after the sentence 'By compactness of L_β', but L_β is not compact for a general Feller kernel: deterministic transition maps give composition operators, which need not be compact. No compactness or power-compactness hypothesis is imposed in the paper. The existence of a positive eigenfunction e with L_β e = r(L_β) e therefore requires additional structural conditions (for example, finite Z, or a kernel with a density and an appropriate compact embedding). Since Theorem 2.1's sufficiency proof does not use this step, the main theorem is unaffected, but the advertised necessity result lacks proof as stated.","section":"Appendix A.6, proof of the r(L_β) > 1 direction"}],"minor_comments":[{"comment":"The phrase 'in many cases, not just sufficient but also necessary' is too vague; once Appendix A.6 is corrected, the paper should specify the exact class of processes for which necessity holds (for example, finite irreducible chains).","section":"Section 2.4, last paragraph"},{"comment":"The augmented state \\tilde Z_{t+1} = (Z_{t+1}, Z_t) and the induced kernel \\tilde Q on \\tilde Z = Z^2 should be defined explicitly in the main text, since the aggregator formula uses \\tilde Q(z, dz') with z already augmented.","section":"Section 6.1, footnote 21"},{"comment":"The notation r_β^n in the inequality ‖T^n v − T^n w‖_j ≤ r_β^n ‖v − w‖_j is easy to confuse with the pointwise function β; renaming the eventual-discounting modulus (for example, γ_n) would improve readability.","section":"Section 5.2, proof of Proposition 5.2"},{"comment":"Please specify the initialization of the simulated paths used to compute s\\u005cn, since reproducibility requires knowing whether the paths are started from a fixed state or drawn from a stationary distribution.","section":"Section 6.2.4, Table 2"},{"comment":"The abstract states that the condition 'cannot be significantly weakened,' while Section 7 explicitly leaves open how close to necessary the condition is for recursive-preference models; the abstract should be qualified to match the scope of the necessity result actually proved.","section":"Abstract and Section 7"}],"recommendation":"major_revision","confidential_remarks":"The main sufficiency theorem is sound and the paper is publishable after revision. The only substantive technical gap is the necessity claim in Appendix A.6, which is peripheral to Theorem 2.1 but central to the abstract's 'cannot be significantly weakened' statement. A revision that restricts the necessity result to finite irreducible processes (or otherwise supplies the missing spectral assumptions) and adjusts the abstract accordingly would resolve my concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on Stachurski and Zhang. The core result is worth knowing: under a spectral-radius condition on the state-dependent discount operator, the standard DP machinery—Bellman equation, optimal policy, convergence of VFI and policy iteration—survives even when beta_t exceeds one with positive probability. The eventual discounting condition and its link to r(L_beta) (Prop 4.1) are genuinely new relative to the earlier literature, which either imposed beta <= 1 or was confined to optimal savings. The paper also handles unbounded rewards via homogeneous and local-contraction methods, and extends to Epstein–Zin preferences, with an honest discussion of where necessity is still open.\n\nThe sufficiency proofs are careful and complete. I checked the Blackwell-type argument and the iterated bound in Lemmas A.1–A.2; the eventual contracting property with modulus r_beta^n is clean. The numerical spectral radius calculations for Christiano et al., Hills et al., and others are a useful practical contribution.\n\nNow the soft spot. Appendix A.6 claims that, for compact Z and Feller Q, finiteness of the expected discounted sum is equivalent to r(L_beta) < 1. That is not true pointwise. The proof of the r>1 direction invokes compactness of L_beta without justification; for deterministic Feller transitions, L_beta is a composition operator and generally not compact. More importantly, the claim fails for reducible processes. A simple three-state example with an absorbing state where beta < 1 and a two-state cycle with beta > 1 gives r(L_beta) > 1 but finite expected discounted rewards when starting in the absorbing state. So the abstract's claim that the condition 'cannot be significantly weakened' is not supported as written. This does not damage Theorem 2.1, which is a sufficiency result, but the necessity appendix needs a uniform or sup-over-z formulation and explicit hypotheses under which L_beta has the required spectral properties.\n\nMy bottom line: the paper is a solid contribution for macro/finance theorists and applied DP users. It deserves a serious referee and, with a revision fixing A.6, should be accepted. I'd take it to reading group and would cite it for the sufficiency condition.","headline":"Useful sufficiency theory for state-dependent discounting; the main theorems hold up, but the appendixed necessity claim that the condition 'cannot be significantly weakened' is false as stated and needs repair.","tokens_in":32220,"tokens_out":3229,"would_cite":true,"duration_ms":29322,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C39","90C40","91B62"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single 'eventual discounting' condition restores Bellman optimality and convergence when discount factors vary by state.","keywords":["dynamic programming","state-dependent discounting","eventual discounting","Bellman equation","spectral radius","recursive utility","unbounded rewards","policy iteration"],"falsifier":"Use the parameterization reported in Section 6.2.4, which the paper computes to have stationary geometric mean s = 1.0168 > 1 so eventual discounting fails, and run value iteration on the Epstein–Zin Bellman equation (28) on a fine grid; if iteration converges to a finite fixed point and an optimal policy exists, then in that recursive-preference setting eventual discounting is not necessary for the standard optimality results, undermining the paper's claim that the condition cannot be significantly weakened.","tokens_in":62,"feed_emoji":"📈","tokens_out":6830,"duration_ms":129356,"temperature":0.7,"pith_summary":"This paper asks when the standard toolkit of infinite-horizon dynamic programming survives state-dependent discounting, where the discount factor is a random process that may even exceed one. The authors show that a single condition, called eventual discounting — the expected product of discount factors over some finite horizon is uniformly below one — is enough to recover the Bellman equation, existence of an optimal policy, Bellman's principle of optimality, and convergence of value iteration and Howard policy iteration. The condition is checkable in practice through the spectral radius of a discount operator, and the paper shows it is nearly necessary in many settings. This matters because modern macro-finance models routinely let discount rates vary with the state, and until now the theory did not cover them.","feed_headline":"Discount factors above one are fine under one spectral condition","feed_subtitle":"Bellman optimality, existence and convergence all survive state-dependent discounting when a spectral radius stays below one.","key_machinery":"The central object is the discount operator (L_β h)(z) = β(z)∫ h(z′)Q(z,dz′), whose spectral radius r(L_β) characterizes eventual discounting: the condition holds exactly when r(L_β) < 1 (Proposition 4.1). In finite irreducible settings this reduces to a Perron–Frobenius eigenvalue or the stationary geometric mean s_β = lim (E ∏ β_t)^{1/n}. This spectral condition does the work by making the Bellman operator and policy operators eventually contracting in the supremum norm, so the standard Banach fixed point and policy iteration arguments apply even though individual β_t may exceed one.","core_discovery":"The paper's central theorem (Theorem 2.1) states that for a regular dynamic program whose continuation aggregator H satisfies the Lipschitz bound |H(x,z,x',v)−H(x,z,x',w)| ≤ β(z)∫|v−w| Q(z,dz'), the eventual discounting condition sup_z E^z ∏_{t=0}^{n−1} β_t < 1 for some n implies that the Bellman operator and every policy operator are eventually contracting, the value function is finite, continuous, and the unique fixed point, an optimal policy exists, Bellman's principle of optimality holds, and both value iteration and Howard policy iteration converge. The discount process may take values above one with positive probability, as long as its long-run growth in expectation is below one.","pith_inferences":["The spectral-radius test could serve as a pre-estimation diagnostic in calibrated macro models: before solving a model with state-dependent discounting, compute r(L_β) from the estimated discount process; values at or above one warn that the optimization problem may be ill-posed rather than economically meaningful.","The framework's reliance on eventual contractivity suggests that a similar condition might be derivable for continuous-time recursive utility models, where the discount operator becomes an infinitesimal generator and the spectral radius is replaced by a growth rate; the paper lists continuous time as an open question.","The Epstein–Zin analysis implies a testable extension: in models where preference-shock volatility is high and ψ is close to one, the 'explosive responses' reported in some applied studies might be artifacts of a failed eventual discounting condition rather than genuine economic mechanisms.","Because the proof uses only the order-preserving and Lipschitz structure of the aggregator, the main theorem might extend to non-Markov discount processes by replacing the kernel Q with conditional distributions and keeping the same spectral-radius formulation."],"forward_implications":["For additively separable problems with bounded rewards, eventual discounting makes value iteration and Howard policy iteration convergent, so optimal policies can be computed numerically in models where the discount factor is state-dependent.","Models that allow discount factors above one, such as New Keynesian zero-lower-bound settings, are admissible whenever the spectral radius of the discount operator stays below one.","Unbounded rewards can be handled by adapting the eventual discounting condition to include growth bounds: for homogeneous problems the condition becomes (24) with an extra α^θ term, and for general unbounded rewards a weighted-norm local contraction setup works.","For Epstein–Zin preferences, the modified condition (29) shows that the elasticity of intertemporal substitution matters: discount-factor volatility becomes increasingly destabilizing as ψ approaches one.","In many settings with bounded rewards, eventual discounting is also necessary for finite lifetime values, so the condition cannot be substantially weakened without breaking the theory."],"supporting_citations":[{"why":"Supplies the abstract dynamic programming framework and the policy-iteration convergence argument that the paper adapts to state-dependent discounting.","marker":"Bertsekas (2013)"},{"why":"Provides the standard infinite-horizon theory, regularity conditions, and the monotonicity/concavity results that the paper extends.","marker":"Stokey et al. (1989)"},{"why":"Supplies the homogeneous-function treatment of unbounded rewards that Section 5.1 extends to state-dependent discounting.","marker":"Alvarez and Stokey (1998)"},{"why":"Provides the local contraction method for unbounded rewards that Section 5.2 builds on.","marker":"Rincón-Zapatero and Rodríguez-Palmero (2003)"},{"why":"Supplies the weighted-norm Banach space and 0-local contraction fixed point result used in Section 5.2.","marker":"Matkowski and Nowak (2011)"},{"why":"Provides the monotone fixed point theorem that underpins the Epstein–Zin extension in Section 6.2.","marker":"Marinacci and Montrucchio (2010)"},{"why":"Supplies the local spectral radius formula for positive operators used to equate eventual discounting with r(L_β) < 1.","marker":"Krasnosel'skii et al. (1972)"},{"why":"Defines the policy iteration algorithm whose convergence is established in Theorem 2.1(g).","marker":"Howard (1960)"}],"fun_headline_variants":["One spectral condition makes discount rates above one safe for optimality","State-dependent discounting: one spectral bound keeps optimality intact","Above-one discounting works if a spectral radius is below one","Eventual discounting condition lets discount rates exceed one safely"],"cache_read_input_tokens":34304,"weakest_assumption_plain":"The load-bearing premise is that the continuation aggregator's Lipschitz modulus is exactly the exogenous discount process β(z) in inequality (8); if the modulus were a different process — say action-dependent or endogenous — eventual discounting of (β,Q) would not control the Bellman operator, and the proofs of Lemmas A.1–A.2 would fail.","fun_headline_variants_meta":{"raw":{"variants":["One spectral condition makes discount rates above one safe for optimality","State-dependent discounting: one spectral bound keeps optimality intact","Above-one discounting works if a spectral radius is below one","Eventual discounting condition lets discount rates exceed one safely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2516,"prompt_tokens":749,"completion_tokens":1767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":365,"completion_tokens_details":{"reasoning_tokens":1697}},"tokens_in":365,"tokens_out":1767,"duration_ms":11532,"temperature":1.0,"reasoning_tokens":1697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:29:06.838306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the parameterization reported in Section 6.2.4, which the paper computes to have stationary geometric mean s = 1.0168 > 1 so eventual discounting fails, and run value iteration on the Epstein–Zin Bellman equation (28) on a fine grid; if iteration converges to a finite fixed point and an optimal policy exists, then in that recursive-preference setting eventual discounting is not necessary for the standard optimality results, undermining the paper's claim that the condition cannot be significantly weakened.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the abstract dynamic programming framework and the policy-iteration convergence argument that the paper adapts to state-dependent discounting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard infinite-horizon theory, regularity conditions, and the monotonicity/concavity results that the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the homogeneous-function treatment of unbounded rewards that Section 5.1 extends to state-dependent discounting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the local contraction method for unbounded rewards that Section 5.2 builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the weighted-norm Banach space and 0-local contraction fixed point result used in Section 5.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the local spectral radius formula for positive operators used to equate eventual discounting with r(L_β) < 1."}],"review_version":1}