Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Optimal Dynamic Fees in Automated Market Makers

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper derives closed-form optimal dynamic fees for a constant function AMM and identifies two fee regimes: one that penalizes arbitrageurs and one that attracts noise traders.

desk verdict A useful first pass at optimal dynamic AMM fees, with a clean exact result in the constant-price case and a heuristic quadratic approximation that needs an error bound before the near-optimality claim is taken literally. read the letter →

arxiv 2506.02869 v3 pith:EWZZWY3A submitted 2025-06-03 q-fin.TR q-fin.MF

classification q-fin.TRq-fin.MF MSC 93E2091B70
keywords decentralizedfinanceautomatedmarketmakersoptimalfeesarbitrageursnoisetradingdynamicstochasticcontrolloss-versus-rebalancing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks how an automated market maker should set transaction fees over time to maximize the fees it collects. Working on a finite inventory grid and assuming order arrivals react exponentially to the gap between the pool's fee-inclusive exchange rate and an external oracle price, the authors solve the stochastic control problem in two tractable cases: a constant oracle price and a quadratic approximation with a stochastic price. In both settings the optimal fees are not constant: the pool charges high fees when its quoted price is favorable to arbitrageurs and low or negative fees in the other direction to stimulate noise trading. The paper argues that a fee rule that is linear in inventory and sensitive to the oracle price is a near-optimal, implementation-ready approximation.

What carries the argument

The central object is a controlled point-process model of order flow, with buy and sell intensities proportional to $\exp\left(k\left(Z_+^p(y)-(s+\zeta)\right)\Delta_+(y)\right)$ and its symmetric counterpart for sells. The exponential form, adapted from market-making intensity models, converts the Hamilton-Jacobi-Bellman equation into a linear system through the substitution $w=e^{kg}$; the tridiagonal matrix $A$ in Theorem 4.1 encodes arrival rates, exchange rates, and inventory grid, and optimal fees are read from log-ratios of $w$ across neighboring inventory states. In the stochastic-price case the same HJB equation is replaced by a quadratic expansion, yielding a Riccati ODE for the ansatz $g(t,y,s)=y^2A(t)+yB(t,s)+C(t,s)$, from which the fees are affine in $y$ and $s$.

What would settle it

Take a pool with a known fee rule and record, for each one-step inventory level, the empirical log arrival rate of buys and sells as a function of the fee-inclusive exchange rate minus the oracle price. If the log rate is not well approximated by a line with slope $k$ over the range of typical gaps, the exponential-intensity premise is violated and the optimality result does not carry over. A second check is whether the predicted revenue gap between optimal and constant fees appears in a live or simulated pool under a different order-flow model.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is an explicit solution to the AMM fee-control problem. When the oracle price is constant, the value function takes the form $v(t,y,c)=c+\frac1k\log w(t,y)$ with $w(t,y)=\exp(A(T-t))\mathbf{1}$, and the optimal fees are $p^*(t,y_i)=\frac{1}{kZ_+(y_i)\Delta_+(y_i)}(1+\log(w(t,y_i)/w(t,y_{i+1})))$ and the symmetric formula for $m^*$; in the quadratic approximation these become fees that are affine in inventory $y$ and oracle price $s$. The fee schedule splits into two regimes, penalizing arbitrageurs on one side of the oracle price and subsidizing noise traders on the other. Simulations show the optimal policy outperforms constant fees by a clear margin, and the linearized version gives nearly identical revenue.

Load-bearing premise

All results rest on the assumption that buy and sell order arrivals are controlled point processes whose intensity is exponential in the gap between the pool's fee-inclusive exchange rate and the oracle price, with trades of fixed size and constant pool depth; if real order flow does not react this way, the derived fee formulas are model outputs rather than robust prescriptions.

Editorial extensions

If this is right

  • The fee formulas in Theorem 4.1 can be precomputed for a given inventory grid and applied as a deterministic function of time and inventory, enabling real-time fee setting without per-trade optimization.
  • Under the quadratic approximation, optimal fees are affine in inventory and oracle price and do not depend on price volatility, giving a simple implementation rule that only needs the current pool inventory and an oracle price feed.
  • The two-regime structure implies that a uniform constant fee is revenue-dominated; the simulated gap between optimal and constant fees grows with arrival intensity and shrinks as traders become less price-sensitive.
  • Because the linearized policy reproduces the optimal revenue in the paper's simulations, AMM designers can restrict attention to a small family of linear fee schedules rather than solving the full control problem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, a testable prediction is that pools switching from constant to inventory-linear dynamic fees should see arbitrage trades concentrated in the high-fee direction and noise-trading volume rise in the low-fee direction, something the paper's simulations exhibit but do not validate on market data.
  • The affine fee rule in the stochastic-price case could be implemented on-chain with periodic recalibration of its coefficients, since the oracle price enters linearly and no per-trade optimization is needed.
  • The analysis fixes pool depth and excludes liquidity-provider entry and competition among pools; if those are endogenized, the slope of the optimal linear fee may shift, but the two-regime shape is likely to persist.
  • The quadratic approximation comes without formal error bounds, so the near-optimality of linear fees is a numerical finding rather than a proven theorem; a rigorous error estimate would be the direct next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies optimal dynamic fee setting in a constant-function market maker. Order flow is modelled by controlled doubly stochastic point processes with exponential intensities depending on the gap between the fee-inclusive AMM exchange rate and an external oracle price, on a finite inventory grid. The authors derive an HJB equation, solve it exactly in the constant-oracle-price case (Theorem 4.1) via the transformation e^{kg}=w, and in the stochastic-price case replace exponentials by quadratic polynomials and posit a quadratic value function to obtain closed-form fee formulas (Theorem 5.2). Simulations compare these formulas with constant fees and with a linearized rule and report that the two-regime structure (penalize arbitrageurs versus attract noise traders) and the linear rule are revenue-close to the optimal policies.

Significance. If valid, the paper offers a tractable framework for dynamic fee design in AMMs and a concrete linear fee rule implementable in venues such as Uniswap v4. Theorem 4.1 is an elegant exact solution of the reduced HJB system, and the public code is a strength. However, the central near-optimality claim rests on an approximate HJB with no error control and on simulations that compare only against constant and first-order policies; the significance is therefore conditional on fixing the verification and approximation issues below.

major comments (4)
  1. [Section 3, footnote 4] The control problem is never given a rigorous weak formulation, and no verification theorem is stated or proved showing that the solution of HJB (3.4) equals the value function of the controlled point-process system. Consequently Theorem 4.1 shows that v solves an HJB equation and gives candidate optimizers, but it does not establish that these policies are optimal for the original model. Please add a verification argument (or a formal statement with the required regularity) and a precise probability-space construction.
  2. [Section 5, Eq. (5.1)] The replacement of the exponentials by second-order Taylor polynomials, the constant trade sizes δ±, and the quadratic ansatz g(t,y,s)=y^2 A(t)+y B(t,s)+C(t,s) are made without any error estimate. No argument shows that the solution of (5.1) is close to the value function of (3.4), so the abstract's claim that linear-in-inventory, price-sensitive fees are a 'good approximation of the optimal fee structure' is not quantified. The authors' own observation after Figure 8 that the second-order approximation degrades as k→0 indicates the error is not uniformly small; an explicit bound or a numerical error study against the unapproximated HJB is needed.
  3. [Theorem 5.2] The statement asserts existence of 26 constants {ψ_i} whose full definition is relegated to a GitHub repository. As written, the theorem is not self-contained: the ODE system (ψ0+ψ1A+ψ2A²+A'=0, etc.) cannot be verified or implemented from the paper alone, and the claim of uniqueness of solutions to the ODE system is not justified. Please include the constants and the proof, or move them to an appendix.
  4. [Section 5.1, simulation table] The numerical comparison measures revenue under the same order-flow model used to derive the policies. The 'Linear' rule in Section 4.1 is a Taylor linearization of the optimal fee rule in that same model, so the near-identical revenue is partly by construction; no comparison to the true optimal policy of the original, unapproximated model is provided, and no confidence intervals are reported. To support the near-optimality claim, please benchmark against a numerical solution of the original HJB (or at least report the approximation error in the value function) and report standard errors.
minor comments (5)
  1. [Throughout] There are typographical errors such as 'assety' and 'the the grid'; please proofread the manuscript carefully.
  2. [Figures 1 and 7] The captions for Figures 1a/1b and 7a/7b are identical or nearly identical; please make the captions distinguish the two approximations explicitly.
  3. [Eq. (3.1)] In Eq. (3.1), the notation in the jump terms is inconsistent, writing both v(t, y, c, s) and v(t, c, y, s) for the same object; please use a uniform argument order.
  4. [Corollaries 4.2 and 5.3] Both corollaries assert limits as k→0, but the paper does not state whether these limits are uniform in t and y; please clarify the mode of convergence.
  5. [Section 2] The sentence 'the depth of the pool p² is fixed throughout [0,T]' is an important modelling limitation and should be highlighted in the introduction as an assumption, not only in the model section.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the fee formulas are derived forward from explicit intensity-model assumptions, and the linear-fee claim is an in-model approximation check rather than a fitted prediction; the only self-citation is a minor, non-load-bearing technical footnote.

full rationale

The derivation is forward. Order-flow intensities are specified exogenously; the dynamic programming principle leads to the HJB equation (3.1)-(3.4); under constant oracle price the substitution e^{kg}=w turns (4.1) into the linear ODE system (4.2), whose exact solution yields the fee formulas (4.3)-(4.4). The second approximation is explicit about its inputs: second-order Taylor expansion of the exponentials, constant trade sizes, a linearized exchange rate, and the quadratic ansatz g(t,y,s)=y^2 A(t)+y B(t,s)+C(t,s). Formula (5.2) then follows algebraically from (3.3) and that ansatz, so the linear-in-inventory and linear-in-price form is an assumption-consequence, not an independently discovered empirical fact; the paper does not claim to have derived the ansatz from the original control problem. The 'good approximation' claim rests on simulations within the same model, comparing the linearized policy to the exact first-approximation optimum and to constant fees. Because the linear rule is a Taylor linearization of the optimal policy itself, the near-identical revenue is partly by construction, but this is a self-consistency check, not a fitted parameter relabeled as prediction. No parameters are calibrated to external data, and no resulting quantity is renamed as an out-of-sample prediction. The only self-citation is footnote 4, which points to Barucci et al. (2025) for a rigorous weak formulation; that technical point is not load-bearing for the fee formulas or the numerical comparisons. The absence of an error bound between the solution of (5.1) and the true value function is a correctness and robustness gap, not a circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

All numerical results are within the model; no external data are used. The free parameters are hand-picked and the sensitivity analysis covers a few values of k, lambda and phi. The central theorems hold for given parameters, but the economic claims are demonstrated only in the chosen parameter range.

free parameters (5)
  • k (intensity sensitivity) = 2, 1, 0.5, 0.25, 0.1 (used in figures and tables)
    Controls how strongly order flow reacts to the gap between fee-inclusive AMM price and oracle price. Not fitted; ranges are scanned in Section 4.1. The fee magnitudes and the shape of the regimes depend on it.
  • lambda +/- (baseline order intensities) = 50, 100, 150 per side
    Set by hand for simulations; revenues scale with lambda but relative comparisons are insensitive; sensitivity is reported in the tables.
  • sigma (oracle volatility) = 0 (Section 4), 0.2 (Section 5)
    In the quadratic approximation the optimal fees turn out independent of sigma; chosen for the simulation only.
  • delta +/- (trade size / grid step) = delta = 0.5 and exchange rate step 0.1
    Trades move inventory by exactly one grid step; this fixes the size of shocks that drive the volatility-attracting regime. Results may depend on this step.
  • phi (penalty weight) = 0 for main results; scanned up to 100
    Penalty P = phi (Z(y)-s)^2 is set to zero for headline results; the paper argues concavity does not require it.
assumptions (6)
  • domain assumption Order arrivals are controlled Cox processes with exponential intensities of Avellaneda-Stoikov form, symmetric in buy/sell with coefficient k.
    Section 2 defines lambda +/-; the entire optimality analysis hangs on this functional form.
  • domain assumption Trade sizes are fixed to one grid step and the pool depth p^2 is constant over [0,T]; no LP entry or exit.
    Section 2 states T is short enough that liquidity is constant; grid increments define the state jumps.
  • domain assumption The oracle price is an arithmetic Brownian motion S_t = S_0 + sigma W_t.
    Section 2; used to derive the sigma^2/2 second derivative term in the HJB equation.
  • ad hoc to paper Fees p and m are allowed to be arbitrary real numbers, unbounded, negative, and above 100 percent.
    Footnote 3 relaxes p,m in [0,1] because negative fees arise as optimal; the magnitude of negative rebates is not constrained.
  • domain assumption The dynamic programming principle and HJB verification hold for the controlled point-process problem.
    Equation (3.1) is asserted from the DPP; footnote 4 admits the weak formulation is left out, so admissibility and verification are not proved.
  • ad hoc to paper Section 5 approximations: exponential terms are replaced by second-order Taylor expansions; g is a quadratic polynomial in y; delta +/- are constant; exchange rates are linearized.
    Remark 5.1 and surrounding text; no error bounds are given for these approximations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal Dynamic Fees in Automated Market Makers." pith.science (2026). https://pith.science/paper/EWZZWY3A

@misc{pith2026250602869,
  author       = {Pith},
  title        = {Pith review of: Optimal Dynamic Fees in Automated Market Makers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWZZWY3A}},
  note         = {Machine review of arXiv:2506.02869}
}
read the original abstract

Automated Market Makers (AMMs) are emerging as a popular decentralised trading platform. In this work, we determine the optimal dynamic fees in a constant function market maker. We find approximate closed-form solutions to the control problem and study the optimal fee structure. We find that there are two distinct fee regimes: one in which the AMM imposes higher fees to deter arbitrageurs, and another where fees are lowered to increase volatility and attract noise traders. Our results also show that dynamic fees that are linear in inventory and are sensitive to changes in the external price are a good approximation of the optimal fee structure and thus constitute suitable candidates when designing fees for AMMs.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Optimal Fees for Liquidity Provision in Automated Market Makers

    q-fin.TR 2025-08 conditional novelty 7.0 of 10

    Optimal AMM fees sit just below all-in CEX trading costs in normal markets, rise with volatility, and become effectively infinite (halt trading) in extreme volatility.

  2. Reinforcement Learning for Execution under Dynamic Fees in a Closed-Loop DEX Simulator

    cs.LG 2026-07 accept novelty 6.0 of 10

    In a reserved 1,000-seed closed-loop DEX simulation with forced completion, a small DQN cuts implementation shortfall 13.3 bps versus tuned one-step routing under dynamic fees, with no detectable edge under constant fees.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages · cited by 2 Pith papers

  1. [1]

    Moallemi, Sara Reynolds, and Dan Robinson

    Austin Adams, Ciamac C. Moallemi, Sara Reynolds, and Dan Robinson. am-amm: An auction-managed automated market maker.arXiv preprint arXiv:2403.03367,

  2. [5]

    The adoption of blockchain-based decentralized exchanges.arXiv preprint arXiv:2103.08842,

    Agostino Capponi and Ruizhe Jia. The adoption of blockchain-based decentralized exchanges.arXiv preprint arXiv:2103.08842,

  3. [7]

    Optimal fees for geometric mean market makers

    Alex Evans, Guillermo Angeris, and Tarun Chitra. Optimal fees for geometric mean market makers. In Financial Cryptography and Data Security. FC 2021 International Workshops, Lecture Notes in Computer Science,

  4. [8]

    A note on optimal fees for constant function market makers

    Robin Fritsch. A note on optimal fees for constant function market makers. InProceedings of the 2021 ACM CCS Workshop on Decentralized Finance and Security. Association for Computing Machinery,

  5. [11]

    Optimal Design of Automated Market Makers on Decentralized Exchanges

    SSRN preprint. Xue Dong He, Chen Yang, and Yutian Zhou. Optimal design of automated market makers on decentralized exchanges.arXiv preprint arXiv:2404.13291,

  6. [12]

    Automated market making and loss-versus-rebalancing.arXiv preprint arXiv:2208.06046,

    22 Jason Milionis, Ciamac C Moallemi, Tim Roughgarden, and Anthony Lee Zhang. Automated market making and loss-versus-rebalancing.arXiv preprint arXiv:2208.06046,

  7. [13]

    ZeroSwap: Data-driven Optimal Market Making in DeFi

    Viraj Nadkarni, Jiachen Hu, Ranvir Rana, Chi Jin, Sanjeev Kulkarni, and Pramod Viswanath. Zeroswap: Data-driven optimal market making in defi.arXiv preprint arXiv:2310.09413,

  8. [14]

    Adaptive Curves for Optimally Efficient Market Making

    Viraj Nadkarni, Sanjeev Kulkarni, and Pramod Viswanath. Adaptive curves for optimally efficient market making.arXiv preprint arXiv:2406.13794,

Show all 15 references
  1. [15]

    The case for variable fees in constant product markets: An agent-based simulation

    Marc Sabate-Vidales and David ˇSiˇ ska. The case for variable fees in constant product markets: An agent-based simulation. InFinancial Cryptography and Data Security. FC 2022 International Workshops, Lecture Notes in Computer Science,

  2. [2016]

    Ivan Guo, Shijia Jin, and Kihun Nam

    ISBN 978-1-4987-2547-7; 978-1-4987-2548-4. Ivan Guo, Shijia Jin, and Kihun Nam. Macroscopic market making games.arXiv preprint arXiv:2406.05662,

  3. [2021]

    Market making with fads, informed, and uninformed traders.arXiv preprint arXiv:2501.03658,

    Emilio Barucci, Adrien Mathieu, and Leandro S´ anchez-Betancourt. Market making with fads, informed, and uninformed traders.arXiv preprint arXiv:2501.03658,

  4. [2022]

    doi: 10.2139/ssrn.4144743

    ISSN 1556-5068. doi: 10.2139/ssrn.4144743. ´Alvaro Cartea, Fay¸ cal Drissi, and Marcello Monga. Predictable losses of liquidity provision in constant function markets and concentrated liquidity markets.Applied Mathematical Finance, 30(2):69–93,

  5. [2023]

    Equilibrium reward for liquidity providers in automated market makers.arXiv preprint arXiv:2503.22502,

    Alif Aqsha, Philippe Bergault, and Leandro S´ anchez-Betancourt. Equilibrium reward for liquidity providers in automated market makers.arXiv preprint arXiv:2503.22502,

  6. [2024]

    Liquidity provision of utility indifference type in decentralized exchanges.arXiv preprint arXiv:2502.01931,

    Masaaki Fukasawa, Basile Maire, and Marcus Wunsch. Liquidity provision of utility indifference type in decentralized exchanges.arXiv preprint arXiv:2502.01931,

  7. [2025]

    Price-aware automated market makers: Models beyond Brownian prices and static liquidity.arXiv preprint arXiv:2405.03496, 2024a

    Philippe Bergault, Louis Bertucci, David Bouba, Olivier Gu´ eant, and Julien Guilbert. Price-aware automated market makers: Models beyond Brownian prices and static liquidity.arXiv preprint arXiv:2405.03496, 2024a. Philippe Bergault, Louis Bertucci, David Bouba, and Olivier Gu...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.