{"id":"1d0399fd-d816-4569-a650-724d9120bab3","arxiv_id":"2505.02959","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new prediction-market pricing rule, the Smooth Quadratic Prediction Market, replaces the Bregman fee with a quadratic fee, preserves the main DCFMM axioms, and makes traders collectively perform steepest descent toward their beliefs.","lead":"This paper proposes a new way to price trades in prediction markets, called the Smooth Quadratic Prediction Market, where traders effectively take small gradient-descent steps toward their beliefs. The design charges a simple quadratic fee, keeps the classic no-arbitrage and information guarantees, and gives the market maker a lower worst-case loss than the standard Duality-based Cost Function Market Maker.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1/t price-convergence rate claimed in Theorems 6 and 8 is not implied by the cited optimization bounds, and the proof assumes a minimizer that need not exist for boundary beliefs.","rationale":"The reader's weakest_assumption emphasizes the homogeneous-belief and boundary issues; my stress-test identifies an additional, more technical gap: the convergence-rate claim in Theorems 6/8 is not a consequence of the cited optimization theorems. The market update being a steepest-descent step is correct, and the qualitative convergence may be salvageable with a proper Lyapunov analysis or by adding a coercivity/attainment assumption. The proof gaps are correctable and do not invalidate the market design, so the reader's CONDITIONAL verdict stands. I would not escalate to REJECT because the central mechanism-design insight (separating fee from price update, traders implementing steepest descent) is independently supported by the first-order conditions and by the axiom proofs for Axioms 1-4, which are sound. However, the paper should either correct the rate to O(1/√t) or provide a valid proof of 1/t, and explicitly state the minimizer-existence condition.","tokens_in":14431,"tokens_out":24664,"duration_ms":264568,"concrete_test":"Independently re-derive the gradient-norm bound from the proof of Thm 6: set F=C-⟨μ,q⟩, γ=1/L, apply Thm 5 to get F(q_t)-F* ≤ ||q_0-q*||^2/(2γt), then apply Lem 8 to get ||∇F(q_t)||_2 ≤ L||q_0-q*||/√t (up to constants). This shows the best rate implied by the cited theorems is O(1/√t), not O(1/t). If the authors cannot supply a Lyapunov/descent argument achieving 1/t, the rate claim in Thm 6 and Thm 8 should be weakened or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core convergence theorems (Thm 6, §3.2; Thm 8, §3.3) assert lim_t ∇C(q_t)=μ at rate 1/t. The proofs show the trader's optimal bundle makes q_{t+1} a gradient/steepest-descent step on F(q)=C(q)-⟨μ,q⟩, and then cite Thm 5/Thm 7 plus Lem 8. But Thm 5 (resp. Thm 7) bounds the function-value gap by O(1/t); Lem 8 only converts that into ||∇F(q_t)||_* ≤ √(2L(F(q_t)-F*)) = O(1/√t). The claimed 1/t rate for the market price does not follow from these results. Moreover, both cited theorems require a fixed minimizer q* with F(q*)=inf F; the CIIP condition (closure of gradient range equals Δ_d) does not guarantee such q* for boundary μ (e.g., softmax attains only interior distributions). The proof therefore has an unstated minimizer-existence assumption, and the boundary case included in the theorem statement is not covered. These gaps leave the central 'incremental incentive compatibility' claim only partially established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new automated market maker for Arrow-Debreu securities, the Smooth Quadratic Prediction Market, whose payment rule is Pay_L(q_t, r_t) = ⟨∇C(q_t), r_t⟩ + (L/2)‖r_t‖² for a CIIP cost function C. The authors show that this mechanism satisfies the DCFMM axioms of instantaneous pricing, information incorporation, no arbitrage, and expressiveness, and that its worst-case loss is no worse than that of the DCFMM. They then characterize traders' expected-profit-maximizing behavior: for the ℓ₂-based market this is exactly gradient descent on F(q)=C(q)-⟨μ,q⟩, and for ℓ_p-based markets it is general steepest descent; they claim this implies that prices converge to a common trader belief μ at rate 1/t. The paper also analyzes budget-bounded traders, a buy-only market, and an adaptive-liquidity extension.","tokens_in":14696,"tokens_out":4273,"duration_ms":48119,"significance":"The construction is elegant and, if the convergence claims are properly established, would be a useful contribution to prediction-market mechanism design: it cleanly separates the instantaneous price term from a quadratic fee, preserves the main DCFMM axioms, and gives a concrete optimization-theoretic interpretation of trader behavior. The elementary axiom proofs (Lemmas 4–7) and the worst-case-loss comparison (Theorem 4) are direct and convincing, and the paper is honest about the exploratory status of the adaptive-liquidity section. However, the central incremental-incentive-compatibility claim currently rests on convergence arguments that are not fully justified: the stated 1/t price rate does not follow from the cited function-value bounds, the minimizer-existence assumption is not guaranteed for boundary beliefs, and the buy-only market claim is asserted without a proof of convergence. These issues are fixable but substantive, so the paper needs a major revision rather than minor polishing.","major_comments":[{"comment":"The claimed rate lim_{t→∞} ∇C(q_t)=μ at rate 1/t is not established by the proofs. The proofs identify the trader update with gradient/steepest descent on F(q)=C(q)-⟨μ,q⟩ and invoke Theorem 5 or Theorem 7, which give F(q_t)-inf F = O(1/t). Lemma 8 then converts this only into ‖∇F(q_t)‖_* = O(1/√t). Since ∇F(q_t) = ∇C(q_t)-μ, the price error is O(1/√t), not O(1/t), unless an additional strong-convexity or duality argument is supplied. Please either correct the rate to O(1/√t) or add assumptions that genuinely yield the 1/t rate.","section":"§3.2–3.3, Theorems 6 and 8"},{"comment":"Both theorems assume the existence of a minimizer q* of F(q)=C(q)-⟨μ,q⟩ with ∇C(q*)=μ, but the CIIP condition only gives closure({∇C(q) | q∈R^d}) = Δ_d. For natural choices such as softmax, ∇C(q) lies in the relative interior of Δ_d, so no q* exists for boundary beliefs μ. Since the theorem statements allow arbitrary μ∈Δ_d, the boundary case is not covered by the cited convergence results. Please restrict the statements to beliefs for which a minimizer exists, or handle boundary beliefs by an explicit approximation or limiting argument.","section":"§3.2–3.3, Theorems 6 and 8"},{"comment":"The buy-only subsection asserts that the KKT-derived update qt+1 = qt + (1/L)((∇C(qt)-μ)_+ - (∇C(qt)-μ)) implies incremental incentive compatibility, but no theorem or proof is provided. The KKT conditions characterize only the one-step optimal bundle; they do not by themselves imply lim_{t→∞} ∇C(qt)=μ. This is a load-bearing claim in the paper's motivation, so it needs a precise statement and a proof, including treatment of coordinates where the constraint is active and of boundary beliefs.","section":"§4, Buy-Only Market"}],"minor_comments":[{"comment":"There is a typo in the abstract: 'the incentivizes agents' should read 'that incentivizes agents'. Similar proofreading issues appear elsewhere, e.g., 'a a curvature fee' in Section 2.4 and 'instsantaneous' in Definition 2.","section":"Abstract and Section 1"},{"comment":"The proof uses the same symbol C for both the market cost function and the auxiliary function F(q)=C(q)-⟨μ,q⟩; this makes the argument harder to follow and invites the incorrect impression that the market cost itself is being minimized. Please introduce a separate notation such as F(q).","section":"Theorem 6 proof"},{"comment":"The text says the result holds 'not just experimentally but analytically', but the convergence claim is not actually proved in the main text or Appendix F. Please either supply the proof or soften the claim to 'we conjecture' / 'we leave the proof to future work'.","section":"§4, Buy-Only Market"},{"comment":"The captions state that softmax is not ℓ₁-smooth but is nevertheless used with ℓ₁ norms 'for the sake of comparison'; this is fine, but the captions should also state how convergence is measured (e.g., which quantity is plotted on the axes and over how many iterations).","section":"Figures 1–3"}],"recommendation":"major_revision","confidential_remarks":"The core mechanism-design idea is sound and the axiomatic part is clean, but the paper's central convergence claims need substantive repair: the price-rate statement is too strong for the cited tools, the boundary-belief case is unhandled, and the buy-only market's convergence is asserted without proof. These are fixable within the scope of the paper, so I recommend major revision rather than rejection. I do not see a circularity problem: the steepest-descent interpretation is built into the payment by design, which is legitimate mechanism design rather than an independent empirical prediction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a genuinely useful piece of mechanism design: replacing the Bregman fee with a quadratic one gives a weakly better worst-case loss for the market maker and preserves the DCFMM axioms, and the trader-as-steepest-descent reading is a real insight. Second, the main convergence-rate claim is over-claimed. The stress-test note is right: the proof of Theorems 6 and 8 bounds the function value by O(1/t), but Lemma 8 only converts that into an O(1/√t) bound on the gradient norm, so the stated 1/t rate for the market price simply does not follow. The minimizer-existence issue for boundary beliefs is also real; the CIIP property only gives closure of the gradient range, not attainment, so the theorem statements as written cover cases the proof does not handle.\n\nWhat is new and good: the SQPM payment rule itself, the steepest-descent incentive analysis, and the appended formal connection between CFTRL and DCFMM (which, as the authors say, was previously only hinted at). Lemmas 4–7 and Theorem 4 are clean and correct. The paper is refreshingly honest about its own limits: it does not claim full incentive compatibility, explicitly defers varying beliefs, and labels the adaptive-liquidity section introductory. No fabricated evidence here, and the code and experiments are straightforward.\n\nWhere it falls short: the rate claims in Section 3.2 and 3.3 need to be corrected or downgraded to O(1/√t), and the buy-only convergence result is asserted rather than proved in the main text (the appendix gives a KKT derivation but no rate). These are fixable. The central design idea survives; the paper just needs to be more careful about what it promises.\n\nWho is this for? Someone working on prediction-market design or on the interface between online learning and market mechanisms. It will not reorganize the field, but it is a solid, citable contribution after revision. I would give it a serious referee, and if I were the editor, I would send it out with a request to fix the rate claims.\n\nRecommendation: engage with it, treat the convergence-rate statements with suspicion, and suggest a revision rather than rejection.","headline":"A clean, honest mechanism-design paper that adds a simpler fee schedule to prediction markets and ties trader updates to steepest descent, but the claimed 1/t price-convergence rate does not follow from the cited theorems.","tokens_in":15186,"tokens_out":1655,"would_cite":false,"duration_ms":19582,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A prediction market that swaps the DCFMM's Bregman fee for a quadratic fee keeps the core guarantees, lowers the worst-case loss, and makes traders run steepest descent toward their shared belief.","keywords":["prediction markets","cost function market makers","Arrow-Debreu securities","incremental incentive compatibility","steepest descent","convex optimization","market liquidity","online learning"],"falsifier":"Simulate the market: choose a smooth cost function such as softmax with $L=1$, fix a belief $\\mu$ in the interior of the simplex, and run a long sequence of traders each solving their one-step expected-profit maximization; Theorem 6 predicts $\\|\\nabla C(q_t)-\\mu\\|$ decays as $O(1/t)$, so a price that stalls, cycles, or converges at a slower rate would contradict the incremental incentive-compatibility claim.","tokens_in":14200,"feed_emoji":"📊","tokens_out":18450,"duration_ms":172626,"temperature":0.7,"pith_summary":"The paper proposes a new pricing rule for prediction markets over mutually exclusive outcomes. The standard Duality-based Cost Function Market Maker (DCFMM) charges a trade as the change in a convex cost function, which amounts to a linear price term plus a Bregman divergence fee. The Smooth Quadratic Prediction Market keeps the linear term and replaces that fee with a simple quadratic fee, $\\frac{L}{2}\\|r_t\\|^2$. The paper argues this preserves the DCFMM's guarantees of instantaneous prices, information incorporation, expressiveness, and no arbitrage, while lowering the market maker's worst-case loss. It also shows that a sequence of traders who share a belief $\\mu$ and maximize only current profit will move prices to $\\mu$ at rate $1/t$, because each trade is exactly a steepest-descent step on $C(q)-\\langle\\mu, q\\rangle$.","feed_headline":"Quadratic fee gives prediction markets lower worst-case loss","feed_subtitle":"Traders sharing a belief push prices toward it by steepest descent; the market maker loses no more than the classic one.","key_machinery":"The load-bearing object is the payment rule $Pay_L(q_t, r_t) = \\langle\\nabla C(q_t), r_t\\rangle + \\frac{L}{2}\\|r_t\\|^2$, a 'Price-Plus-Fee' market with a quadratic fee. The linear term sets the instantaneous price; the quadratic fee replaces the Bregman divergence that the DCFMM charges implicitly through the cost difference $C(q_t+r_t)-C(q_t)$. Because $L$-smoothness bounds the Bregman divergence by the quadratic fee, the market maker collects at least as much revenue as the DCFMM on every trade. And because each trader's objective is, up to constants, $\\langle\\nabla C(q_t), r_t\\rangle + \\frac{L}{2}\\|r_t\\|^2 - \\langle\\mu, r_t\\rangle$, minimizing it over the next market state is exactly the definition of general steepest descent on $C(q)-\\langle\\mu, q\\rangle$; standard smoothness-based convergence theorems then give the $1/t$ rate.","core_discovery":"The central claim is that the payment rule $Pay_L(q_t, r_t) = \\langle\\nabla C(q_t), r_t\\rangle + \\frac{L}{2}\\|r_t\\|^2$, defined for any smooth convex cost function $C$ whose gradient maps to the probability simplex, defines a prediction market that satisfies the standard axioms of instantaneous price, information incorporation, no arbitrage, and expressiveness, and has, for every fixed trade history, a worst-case monetary loss no larger than the DCFMM's. It does not satisfy the standard one-shot incentive compatibility axiom; instead, when traders appear sequentially with a common belief $\\mu$ and each maximizes expected payout minus payment, the market price $\\nabla C(q_t)$ converges to $\\mu$ at rate $1/t$. The mechanism is that each trader's best response is exactly one step of general steepest descent on $C(q) - \\langle\\mu, q\\rangle$. The paper also derives a closed-form update for buy-only markets and gives an adaptive-liquidity extension.","pith_inferences":["A natural extension not explored in the paper: the price/fee split suggests a generic recipe, whereby any learning algorithm with gradient-descent-like updates could seed a market whose fee encodes the algorithm's one-step cost; the Smooth Quadratic Market is one instantiation.","If trader beliefs drift slowly over time, the same descent dynamics should track a moving consensus; a stochastic-gradient treatment with randomly selected traders, which the authors mention as future work, would likely turn this into a quantitative tracking bound.","By choosing the fee's norm or scaling, a designer can shape the path prices take to the consensus without changing the worst-case loss or no-arbitrage guarantees; the paper's simulations with different $\\ell^p$ fees illustrate this freedom."],"forward_implications":["For any fixed trade history, the market maker's collected revenue is at least that of the DCFMM, so the worst-case loss is no larger; with any nonzero trade it is strictly smaller because the quadratic fee dominates the Bregman fee pointwise.","A market run by a sequence of traders with identical belief $\\mu$ reaches $\\mu$ in price: with an $\\ell^2$ fee the price error decays as $O(1/t)$, and with an $\\ell^p$ fee the traders implement general steepest descent with the same convergence guarantee.","The buy-only constraint does not break convergence: the optimal trade has the closed form $\\frac{1}{L}((\\nabla C(q_t)-\\mu)_+ - (\\nabla C(q_t)-\\mu))$, which is a coordinate gradient-descent step, so prices still approach the shared belief.","The payment rule can be extended to adaptive liquidity by letting the smoothness parameter shrink as trade volume grows; the volume-parameterized extension preserves no arbitrage, while the remaining axioms are left to future work."],"supporting_citations":[{"why":"Defines the DCFMM payment rule and its cost-function axioms; supplies the no-arbitrage and worst-case-loss baseline the new market is compared against.","marker":"[Abernethy et al., 2013]"},{"why":"Provides the $O(1/t)$ gradient-descent convergence theorem used to prove incremental incentive compatibility in the $\\ell^2$ case.","marker":"[Garrigos and Gower, 2023]"},{"why":"Supplies the smoothness lemmas connecting gradient norm to function suboptimality, used in the convergence proofs.","marker":"[Sidford, 2024]"},{"why":"Gives the general steepest-descent convergence rate that extends the result to $\\ell^p$ norms.","marker":"[Kelner et al., 2014]"},{"why":"Supplies the volume-parameterized market-making framework that the adaptive-liquidity extension adapts.","marker":"[Abernethy et al., 2014]"},{"why":"Introduces the buy-only market constraint whose KKT analysis yields the closed-form coordinate-descent update.","marker":"[Li and Vaughan, 2013]"},{"why":"States the strong-convexity/smoothness duality used to relate smooth CIIP cost functions to their Fenchel conjugates.","marker":"[Kakade et al., 2009]"}],"fun_headline_variants":["Quadratic fee steepest-descent market lowers worst-case loss","Steepest descent market maker with quadratic fees cuts loss","Quadratic fee prediction market trims worst-case loss","Smooth quadratic market improves worst-case loss bound","New prediction market uses quadratic fees to cut worst-case loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence guarantee requires an infinite sequence of traders who all share exactly the same belief and each maximizes only their immediate one-step profit, so the paper does not cover changing beliefs, traders who plan ahead, or beliefs on the boundary of the probability simplex.","fun_headline_variants_meta":{"raw":{"variants":["Quadratic fee steepest-descent market lowers worst-case loss","Steepest descent market maker with quadratic fees cuts loss","Quadratic fee prediction market trims worst-case loss","Smooth quadratic market improves worst-case loss bound","New prediction market uses quadratic fees to cut worst-case loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001038,"raw_usage":{"total_tokens":4364,"prompt_tokens":936,"completion_tokens":3428,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":3349}},"tokens_in":552,"tokens_out":3428,"duration_ms":30667,"temperature":1.0,"reasoning_tokens":3349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:39:25.784753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the market: choose a smooth cost function such as softmax with $L=1$, fix a belief $\\mu$ in the interior of the simplex, and run a long sequence of traders each solving their one-step expected-profit maximization; Theorem 6 predicts $\\|\\nabla C(q_t)-\\mu\\|$ decays as $O(1/t)$, so a price that stalls, cycles, or converges at a slower rate would contradict the incremental incentive-compatibility claim.","supporting_citations":[{"cited_title":"Optimization algorithms","cited_arxiv_id":null,"evidence_quote":"Supplies the smoothness lemmas connecting gradient norm to function suboptimality, used in the convergence proofs."},{"cited_title":"An almost-linear-time algorithm for approximate max flow in undirected graphs, and its multicommodity generalizations","cited_arxiv_id":null,"evidence_quote":"Gives the general steepest-descent convergence rate that extends the result to $\\ell^p$ norms."},{"cited_title":"A general volume-parameterized market making framework","cited_arxiv_id":null,"evidence_quote":"Supplies the volume-parameterized market-making framework that the adaptive-liquidity extension adapts."},{"cited_title":"On the duality of strong convexity and strong smoothness: Learning applications and matrix regularization","cited_arxiv_id":null,"evidence_quote":"States the strong-convexity/smoothness duality used to relate smooth CIIP cost functions to their Fenchel conjugates."}],"review_version":1}