DeFi and AMM mechanics
Pool arithmetic can validate a quote. Realized execution also needs transaction ordering, actual reserves, fees and routing. Do not apply constant-product formulas to a different invariant merely because the venue calls itself an AMM.
✅ Source-verified 2026-09-14 through extracted PDF text: the Uniswap v2 whitepaper describes the reserve invariant, input fee and optional protocol fee. The Uniswap v3 whitepaper describes concentrated ranges, separate fee balances and inactive liquidity outside a range. Those are historical protocol specifications, not a claim that today's deployed pools use only their initial fee tiers.
All measurements below were reproduced on 2026-09-14 by scripts/amm.py, seed 20260910.
The simulations use continuous quantities and ignore on-chain integer rounding.
Constant-product impact: account for the fee retained in reserves
For reserves (x, y), selling dx of X, fee f, g = 1-f, and u = dx/x:
dy = g * dx * y / (x + g * dx)
x_after = x + dx
y_after = y - dy
execution_price / initial_spot = g / (1 + g*u)
post_trade_spot / initial_spot = 1 / ((1 + u) * (1 + g*u))
🚨 1/(1+u)^2 is the pool-price ratio only at zero fee. The full input enters reserves,
while the fee-reduced input determines output. cp_impact() is tested against the reserve
changes in cp_swap() at nonzero fees as well as at zero.
✅ Measured: a zero-fee trade equal to 5% of the input reserve executes 4.7619% below the initial spot and moves the reserve price down 9.2971%. Its average execution price is the geometric mean of the initial and final reserve prices. Walking the curve in 200,000 slices returns 47,619.0476190473 units versus the closed form 47,619.0476190476: relative difference 6.7e-15.
That slicing identity assumes zero fees and no other flow. Retained fees, arbitrage between slices, block ordering and routing change the result. It is not a live execution-scheduling rule.
Impermanent loss compares with holding the same starting basket
For a full-range, equal-value constant-product position without fees, rebalanced to a new
relative price r = P1/P0:
IL = LP_value / hold_value - 1 = 2*sqrt(r)/(1+r) - 1
✅ Measured independently from the ending reserves: a halving and a doubling both give
−5.7191%; a fourfold move gives −20.0000%; a tenfold move gives −42.5040%. The largest absolute
closed-form/reserve difference in the printed table is 1.11e-16. The formula is symmetric in
r and 1/r and is zero at the starting price.
This is underperformance against holding, not necessarily a negative absolute return. A negative convexity description can be useful; claiming the position is exactly a sold straddle is not an identity. Exiting realizes the current difference, while subsequent price reversals would have changed an unclosed position's difference.
Fees and a conditional break-even calculation
✅ Source-verified 2026-09-14, v2 whitepaper §2.4: fee-related growth can be expressed through
sqrt(k). V3 §3.2.1 keeps accrued fees separately, so they are not automatically reinvested in
active liquidity. An auto-compounding wrapper is a separate strategy and needs separate costs.
✅ Measured: with the illustrative fee 0.30%, 4,000 synthetic two-sided trades increase k by
2.7829%. 1 - sqrt(k0/k1) is 1.3630%, and sqrt(k1/k0)-1 is 1.3819%; converting between those
denominators agrees to 1.21e-16. This demo excludes protocol-fee LP-token dilution.
il_vs_fees() draws 20,000 terminal prices over thirty days, then solves the approximation
fee * daily_volume / TVL * days = -mean_IL:
| Assumed annual volatility | Mean IL | Median IL | 10th-percentile IL | Break-even daily volume / TVL |
|---|---|---|---|---|
| 30% | −0.093% | −0.043% | −0.250% | 1.03% |
| 60% | −0.370% | −0.170% | −1.002% | 4.11% |
| 100% | −1.024% | −0.471% | −2.775% | 11.38% |
| 150% | −2.287% | −1.081% | −6.221% | 25.41% |
🚨 These are scenario break-evens with fixed TVL, no gas, no protocol cut and no joint model of price and fee-generating flow. They are not empirical thresholds for profitable pools. Expected IL grows approximately with variance for small moves; the exact expression is not globally quadratic. Use consistent value denominators and actual fee ownership for a live LP analysis.
Concentrated liquidity: narrow ranges amplify exposure
For a symmetric range [p0/m, p0*m], the capital multiple at entry relative to the full curve
at the same liquidity L is 1/(1 - 1/sqrt(m)). cl_amounts() clips the reserve calculation at
range boundaries. Outside the range one token balance is zero and the position supplies no
active liquidity until price returns.
✅ Measured: range [1/1.10, 1.10] has a capital multiple of 21.5. For an in-range +2% price
move its divergence loss relative to its own starting basket is amplified by that same factor.
At +20%, which is outside this range, its loss is −6.87% versus −0.41% for full range; the
in-range multiplier cannot be extrapolated past the boundary.
✅ Measured: across 4,000 thirty-day paths with assumed 60% annual volatility, the same range is active at 62% of observations. The printed fee-multiple illustration is capital multiple times active-time share, 13.2. This assumes constant fee flow per unit of active liquidity; actual fees depend on competing liquidity and where volume occurs. It is neither a guaranteed return nor a rigorous upper bound. A range order can reverse if price crosses back before liquidity is removed.
Quote versus realized execution
✅ Measured: in the script's pool with reserves 1,000,000 of each token, victim input 20,000, fee 0.30% and slippage tolerance 1%, a simulated sandwich can push victim output to the limit. The resulting attacker profit is 0.862% of victim input. The function solves the limit-saturating size, not globally optimal profit. The fraction captured changes with reserves, trade size, fees, ordering and costs; it is not a universal 86% rule.
The script cannot infer live gas or sandwich frequency from synthetic prices. Historical receipts, transaction traces, mempool observations and aggregator routes can support a richer replay; it would need those actual inputs. Gas is paid even when some transactions revert, ordering determines whether an attack executes, and a multi-pool route changes the reserve arithmetic. Report unmeasured costs separately rather than inventing realized performance.
Scripts and routing
Run python plugins/fin-crypto/skills/defi-and-amm-mechanics/scripts/amm.py.
Public functions cover swaps, fee-aware impact, sliced execution, fee growth, full-range IL,
conditional break-even volume, concentrated reserves and a hypothetical sandwich. Everything
runs offline without credentials or order submission.
../crypto-data-and-execution/SKILL.md— venue clients and data sourcing.../perpetuals-and-funding/SKILL.md— perpetual funding; check whether the actual on-chain venue uses an AMM or an order book before borrowing this model.../crypto-token-events/SKILL.md— wrappers, rebases and token-unit changes inside pools.../crypto-market-structure/SKILL.md— stablecoin quote conversion and counterparty events.../../../fin-strategies/skills/market-making-models/SKILL.md— order-book quoting.../../../fin-microstructure/skills/limit-order-book-models/SKILL.md— passive queue fills.../../../fin-strategies/skills/execution-algorithms/SKILL.md— order scheduling.../../../fin-core/skills/execution-cost-analysis/SKILL.md— realized execution costs.