At sigma=0.01 with gas of 2 USDC, the neural policy earns no more on average than a two-line rebalancing rule. ArrivalRebalance produces mean PnL of 49.72 USDC, beside 49.91 for PPO_narrow, p=0.81. The trained agent earns its keep in the 5% CVaR: 9.15 versus 6.86, p=0.0014. That pair of results carries the argument in Chionas, Kleitsikas, Leonardos, Sánchez-Betancourt and Ventre's paper. The authors acknowledge the mean tie and add that as volatility and gas rise "the mean-PnL advantage of the RL agents becomes less pronounced". The live question is what a policy network contributes beyond a two-parameter rule.
How the LP earns
An LP selects an interval of ticks, deposits both tokens, and receives a pro-rata share of the 0.3% fee on swaps crossing her range while the pool price remains inside it. A narrower range gives her more depth per tick and a larger fee share. Push the width too far and the price leaves the range. Fee income then stops, while she is left holding the wrong token. Impermanent loss and gas offset those fees, with each rebalance priced at 2, 4 or 6 USDC here. Repositioning also requires a swap to reconstruct the token mix demanded by the new range.
The paper formulates the problem as impulse control. At each chosen intervention time, the LP sets a lower and upper tick and pays a fixed cost g. Fees minus impermanent loss minus gas define the risk-neutral objective. In the risk-averse version, a running penalty phi times the squared deviation of risky-token holdings from a target is deducted as well. The resulting HJB quasi-variational inequality is intractable, which sends the authors to reinforcement learning. They train PPO in a simulator containing the full concentrated-liquidity mechanics. There are two parametrisations: PPO chooses both bounds, while PPO_narrow stays at the minimum two-tick width and learns where to centre it.
Each policy receives a training budget of 6x10^5 agent-decision transitions. Training spans 72 scenarios, made from six volatilities, three gas levels and four risk profiles, with ten independent seeds for each. The reported policies alter width, asymmetry and rebalancing frequency as mispricing, gas, volatility and inventory change. Risk aversion compresses the left tail. At sigma=0.03 with g=6, the agents generally refuse to deploy.
Two lemmas carry much of the economics. With one-tick jumps, the volume required to cross a tick scales with the total depth there. The LP receives her pro-rata fee share on that volume, causing total depth to cancel. Her fee revenue is linear in her own liquidity and independent of everyone else's. Crowding within a tick therefore leaves fees undiluted. Impermanent loss is also linear in the LP's own depth, making PnL affine in it, while gas enters as a constant.
Training and evaluation use no on-chain data. The paper's only real data is a normalised liquidity distribution for Optimism, Arbitrum and Ethereum, included in one qualitative figure. External prices follow a zero-drift geometric Brownian motion with S0=1000 and dt=0.001 over T=1. The volatility sweep is {0.005, 0.01, 0.015, 0.02, 0.025, 0.03}. Liquidity-taker arrivals have intensity 15 for baseline noise flow, plus 4000 times the mispricing between pool and external prices. The authors describe this order flow as mainly driven by arbitrage pressure. Gas costs 2, 4 or 6 USDC per rebalance, and starting capital is 1000 numéraire units. Actions stay within fifteen ticks on either side. Decisions arrive every 100 steps, giving ten intervention opportunities in each episode. Evaluation uses 1000 held-out trajectories, hierarchical paired t-tests for PnL, and a 10^4-replicate paired bootstrap for CVaR.
The easy regime flatters the heuristic
The heuristic ties in the friendliest cell, where volatility and gas are lowest. ArrivalRebalance posts at [-1,1], then rebalances every N_arr arrivals. Its N_arr is grid-searched from 10 to 800 separately for each scenario. The authors say they tuned it by "optimising both PnL and CVaR", the same objective used for evaluation. This treatment favours the baseline. PPO_narrow matching its mean while improving the tail is therefore the more useful reading. The passive benchmarks trail far behind: DeployNarrow records 15.85 PnL / -14.52 CVaR, while DeployWide returns 12.78 / -15.11.
The rest of the row complicates the victory. CDM, the closed-form frictionless allocation from Cartea, Drissi and Monga, records 35.18 PnL and 11.55 CVaR in the same cell. Its mean is lower, though its tail beats both risk-neutral and risk-averse PPO_narrow, at 9.15 and 9.32. At sigma=0.02 with g=4, CDM also exceeds both RL agents on mean PnL: 15.01 versus PPO_narrow's 6.51 and PPO's 5.33. Against the risk-neutral agents, it wins on CVaR as well, with -26.23 against -35.69. The risk-averse agents reverse that comparison, producing -8.86 against -26.23.
The cost basis differs. The authors disclose it directly: "CDM does not directly account gas costs." They deliberately allow CDM to rebalance for free at every decision point, treating it as an optimistic analytical benchmark. The medium-vol mean-PnL loss therefore comes against an agent that pays nothing to move.
Risk aversion supplies the paper's cleanest trade. At sigma=0.02 and g=4, changing PPO_narrow from risk-neutral to phi=50 lowers mean PnL from 6.51 to 2.27 and improves CVaR from -35.69 to -8.86. Roughly 4 USDC of carry buys roughly 27 of tail. The mechanism differs across the agents. Flexible PPO widens positions under the inventory penalty, which the authors identify as the main force moving the LP toward wider positions. PPO_narrow has no scope to widen, so it shifts its centre and intervenes less often. The evaluation uses 5% CVaR, while training penalises squared inventory. Those functionals are related, and part of the table reflects PPO optimising the objective it received.
At sigma=0.03 and gas 6, staying out wins
At sigma=0.03 and g=6, the risk-averse PPO reports 0.00 +/- 0.04 PnL with -0.68 CVaR. PPO_narrow reports -0.01 with -0.16. Both mostly stay out. Every alternative suffers much heavier tail losses in that cell: DeployNarrow reaches -56.17, CDM -56.93 and ArrivalRebalance -62.64. DeployNarrow's 0.10 still edges the RL agents on mean. Gas by itself cannot explain the withdrawal. At g=6 with sigma=0.01, the agents continue to deploy and simply retain their positions through large mispricing (Figure 6, right panel).
The strongest downside result is a no-trade decision.
The arithmetic supports it. Ten intervention opportunities at 6 USDC apiece allow gas costs of up to 60 USDC against 1000 units of capital. Meanwhile, the highest mean PnL anywhere in the tables is about 50. The authors state the implication directly: "Under sufficiently adverse conditions, the economically meaningful learned behaviour is endogenous market exit rather than superior active liquidity management." A desk can reach the same conclusion without a policy network. Its practical value lies in placing a numerical marker where liquidity provision ceases to pay.
A policy laboratory, under fixed conditions
Every policy is trained and tested within the same stationary regime, using the same sigma, g and phi. The authors describe their findings as "conditional on correctly identified, approximately stationary transition dynamics". They say empirical validation "would require on-chain data analysis, which is beyond the scope of this work". Their broader claim is equally restrained: "We do not claim universal superiority of RL; rather, we show that it can produce interpretable, state-dependent policies where closed-form solutions are generally unavailable." Those three qualifications set the standard for reading the work. Since interpretability is the claimed output, the state-dependent maps in Figures 5 and 6 deserve as much attention as the PnL table. Mispricing determines placement direction in those maps, while gas governs whether the agent moves.
Two features of the environment matter more than the training design. Arbitrage intensity is 4000 times mispricing, compared with noise intensity 15, and the authors characterise the resulting flow as mainly driven by arbitrage pressure. Friction enters as one fixed g for each intervention. Yet the introduction also identifies the extra trades needed to obtain the token composition of a new position, and slippage on that swap remains unpriced. Price parameters come from the Cartea-Drissi-Monga estimates. T=1 has no mapping to calendar time, leaving the wall-clock severity of the "adversarial" cell undefined.
We could not test these claims on our own data. Concentrated-liquidity LP PnL depends on tick-level depth, swap-by-swap fee growth, tick-crossing events and gas for each transaction. Our data consists of crypto OHLCV bars and contains none of that pool state. A spot-trading substitute would remove the mechanism being studied.
The paper reports mean PnL, 5% CVaR, cross-seed standard deviations and p-values. It gives no risk-adjusted metric at all: no Sharpe, no capacity, no turnover cost. Evidence that would change my view is the same policy family trained in one regime and evaluated in another, with the recomposition swap charged. For now, Lemma 3.1 is the result I would bring to a desk. Under one-tick order flow, and for an LP whose liquidity is small relative to the pool, other LPs' depth within a tick does not dilute her fees. Live swaps cross multiple ticks. A live test would probe precisely that one-tick assumption.