For an ETH-USDC liquidity provider, the implied fee rate in Duarte, Ferreira, Gobet and Saporito is a useful hurdle to clear. The paper calls it "the compensation required to provide liquidity". Treating that rate as a forecast of pool income goes too far. Yet the paper invites that reading: its backtest says model-implied fees "track realized fee accrual closely", and the abstract describes tighter ranges as "earning larger implied returns". The distinction remains unresolved.
The rate expresses the option market's price for an LP's impermanent loss as an annualized yield. Swap traders supply the LP's income through fees, collected when she withdraws. In the paper's identity, expected fees equal today's pool value less the option price of the pool's terminal payoff. The authors calibrate that price using Deribit ETH options from April 1, 2019 to May 17, 2023.
What does the option price buy?
At horizon T, an LP holds the tokens remaining in the pool and the fees earned along the way. The remaining tokens have a value g_R that depends on price. For Uniswap v2, g_R is 2√P. A v3 range [p_a, p_b) has a concave payoff inside the band and a flat payoff above p_b, giving it a covered-call shape. The authors equate today's position value with the discounted risk-neutral expectation of its value at T. Expected discounted fees therefore equal today's pool value less the option price of the terminal pool payoff. Divide by today's pool value to get the implied total fee return R. Concavity, with non-negative rates, bounds that return between 0 and 1.
R is the fee income needed to make pool entry zero-NPV against an option hedge of the concave payoff. To price it, the authors use a Bates stochastic-volatility jump-diffusion model. They fit the model to daily Deribit ETH option quotes from April 1, 2019 to May 17, 2023, setting both interest rates to zero. A Lewis-style Fourier integral prices the v3 payoff. The v2 price needs no integral: it is 2e^{-rθ}G(1/2), the log-price transform evaluated at one half. The charts report y = (1/θ)log(1+R) for tenors from 0.10 to 0.50 years.
Every date and range shown has positive R, and most curves slope downward. Across dates, the curves for [0.7S0, 1.4S0] and [0.8S0, 1.2S0] span roughly 20% to 120%, read from the figure axis. A PCA attributes 81.0% of the variation to a short-end factor, 15.5% to slope and 3.1% to curvature. Changes in the curve thus have few dimensions.
The contribution is narrower than it first appears. Earlier work already treats impermanent loss as a contingent claim; the authors cite papers on static replication and hedging that take this approach. Their addition is to interpret its price as a fee hurdle and lay out that hurdle by maturity, using the full option surface, including jumps. The authors say nobody has previously calibrated a fee term structure to option prices. My reading of the diffusion component is that g(S_t) minus E[g(S_T)] equals the expected integral of ½σ²S²|g''|. This is the loss-versus-rebalancing rate of Milionis et al., cited in the paper. On that reading, y is close to option-implied LVR plus a jump term.
Tight Uniswap v3 ranges sell more convexity per dollar
The range comparisons follow naturally. In the top panels, the authors compare [0.8S0, 1.2S0] with [0.7S0, 1.4S0]. For April 1, 2023 (S0 = 1,821.52), they compare six ranges, running from [0.8S0, 1.5S0] to [0.1S0, 1.3S0]. They report higher rates for tighter ranges centered near spot than for wider or more asymmetric ranges. A narrow band places more payoff curvature near spot and commits less capital. The concavity costs more per dollar, raising the fee hurdle. On the March 1, 2022 surface, the authors move a fixed minimal two-tick range across prices; at short tenors, rates peak with p_b near spot.
Concentrated liquidity, in the authors' words, "requires greater risk-neutral compensation". That explanation is right: it describes the earnings a range needs. Clearing the higher hurdle depends on the volume traded inside the tight range. The paper leaves that question untested for v3, since its realized-fee comparison uses only the v2 pool.
The downward slope and its arithmetic limit
Within the plotted grid of 0.10 to 0.50 years, arithmetic explains none of the slope. R remains below 1, so y cannot rise above ln2/θ. The ceiling falls from 693% at 0.10 years to 139% at 0.50. It never binds within the plotted grid, where the option surface determines the downward slope.
Beyond the grid, the cap begins to matter. It falls to 69% at one year and 35% at two, below the 100% reached by the stressed short-tenor series. In stressed periods, the 0.25-year series at M = 1.20 reaches about 100%. Such a rate cannot persist at any horizon beyond 0.69 years.
Even within the grid, the slope varies. At extreme moneyness on the 2022 surface, parts of the curve flatten or turn upward.
The calibration also warrants scrutiny. The April 1, 2023 fit gives σ_v = 12.664 against b_v = 21.335. This puts 2b_v at 42.7 and σ_v² at 160.4, far outside the Feller condition. The jump parameters look odd as well: κ = 1.96 with intensity λ = 0.004. I did not find reported calibration pricing errors. These parameters suggest weak identification of the jump and vol-of-vol terms. Because y is a functional of the fitted surface, uncertainty in the jump parameters affects it at every tenor.
Two windows of v2 fees
The validation follows a constant-liquidity v2 position of 100,000 USDC, leaving fees un-reinvested. Its two windows run from December 13, 2022 to March 13, 2023, and from March 14 to May 17, 2023. On an axis from 0 to 2.5%, implied and realized accumulated fees appear to move together. I found no tracking-error statistic. The exercise covers one pool and no v3 ranges. It measures fee accrual alone, leaving out the inventory leg.
The authors describe the result as a "useful approximation to realized fee income" under a constant-liquidity strategy. The figure supports that claim. In the introduction, the wording is more guarded: implied fees are "an important component of realized LP profitability". The conclusion expands this to "realized LP profitability", which the exercise does not measure.
Accept the fit at face value and the implication is sharper. If collected fees roughly matched the option cost of the concave payoff, v2 LPs over those five months received about breakeven compensation for the convexity they sold. There is no visible premium left, before gas and under the assumption that the on-chain (DEX) pool price equals the exchange (CEX) price.
We could not rerun this.
Recalibrating the model on each date requires historical Deribit ETH option chains. The exercise also needs on-chain fee accrual per range. ETH spot prices provide neither.
Use y as the yield a range must beat against an option hedge. A v3 test across many windows would change my view: realized range fees minus hedged inventory P&L, compared with this hurdle and accompanied by an error metric.