Files
lmsr-amm/src/LMSRStabilizedBalancedPair.sol
2025-10-03 13:50:41 -04:00

193 lines
8.6 KiB
Solidity

// SPDX-License-Identifier: UNLICENSED
pragma solidity ^0.8.30;
import {ABDKMath64x64} from "../lib/abdk-libraries-solidity/ABDKMath64x64.sol";
import {LMSRStabilized} from "./LMSRStabilized.sol";
/// @notice Specialized functions for the 2-asset stablecoin case
library LMSRStabilizedBalancedPair {
using ABDKMath64x64 for int128;
// Precomputed Q64.64 representation of 1.0 (1 << 64).
int128 private constant ONE = 0x10000000000000000;
/// @notice Specialized 2-asset balanced approximation of swapAmountsForExactInput.
/// - Assumes exactly two assets and that the two assets' internal balances are within ~1% of parity.
/// - Implements a gas-optimized two-tier Taylor approximation to avoid most exp()/ln() calls:
/// * Tier 1 (quadratic, cheapest): for small u = a/b (u <= 0.1) we compute
/// X = u*(1 + δ) - u^2/2
/// ln(1+X) ≈ X - X^2/2
/// and return amountOut ≈ b * lnApprox. This Horner-style form minimizes multiplies/divides
/// and temporaries compared to the earlier a^2/a^3 expansion.
/// * Tier 2 (cubic correction): for moderate u (0.1 < u <= 0.5) we add the X^3/3 term:
/// ln(1+X) ≈ X - X^2/2 + X^3/3
/// which improves accuracy while still being significantly cheaper than full exp/ln.
/// - For cases where |δ| (the per-asset imbalance scaled by b) or u are outside the safe ranges,
/// or when limitPrice handling cannot be reliably approximated, the function falls back to the
/// numerically-exact swapAmountsForExactInput(...) implementation to preserve correctness.
/// - The goal is to keep relative error well below 0.001% in the intended small-u, near-parity regime,
/// while substantially reducing gas in the common fast path.
function swapAmountsForExactInput(
LMSRStabilized.State storage s,
uint256 i,
uint256 j,
int128 a,
int128 limitPrice
) internal view returns (int128 amountIn, int128 amountOut) {
// Quick index check
require(i < s.nAssets && j < s.nAssets, "LMSR: idx");
// If not exactly a two-asset pool, fall back to the general routine.
if (s.nAssets != 2) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Compute b and inverse early (needed to evaluate delta and limit-price)
int128 b = LMSRStabilized._computeB(s);
// Guard: if b not positive, fallback to exact implementation (will revert there if necessary)
if (!(b > int128(0))) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
int128 invB = ABDKMath64x64.div(ONE, b);
// Small-signal delta = (q_i - q_j) / b (used to approximate r0 = exp(delta))
int128 delta = s.qInternal[i].sub(s.qInternal[j]).mul(invB);
// If a positive limitPrice is given, attempt a 2-asset near-parity polynomial solution
if (limitPrice > int128(0)) {
// Approximate r0 = exp(delta) using Taylor: 1 + δ + δ^2/2 + δ^3/6
int128 delta_sq = delta.mul(delta);
int128 delta_cu = delta_sq.mul(delta);
int128 r0_approx = ONE
.add(delta)
.add(delta_sq.div(ABDKMath64x64.fromUInt(2)))
.add(delta_cu.div(ABDKMath64x64.fromUInt(6)));
// If limitPrice <= r0 (current price) we must revert (same semantic as original)
if (limitPrice <= r0_approx) {
revert("LMSR: limitPrice <= current price");
}
// Ratio = limitPrice / r0_approx
int128 ratio = limitPrice.div(r0_approx);
// x = ratio - 1; use Taylor for ln(1+x) when |x| is small
int128 x = ratio.sub(ONE);
int128 absX = x >= int128(0) ? x : x.neg();
// Acceptable range for ln Taylor approx: |x| <= 0.1 (conservative)
int128 X_MAX = ABDKMath64x64.divu(1, 10); // 0.1
if (absX > X_MAX) {
// Too large to safely approximate; fall back to exact computation
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// ln(1+x) ≈ x - x^2/2 + x^3/3
int128 x_sq = x.mul(x);
int128 x_cu = x_sq.mul(x);
int128 lnRatioApprox = x
.sub(x_sq.div(ABDKMath64x64.fromUInt(2)))
.add(x_cu.div(ABDKMath64x64.fromUInt(3)));
// aLimitOverB = ln(limitPrice / r0) approximated
int128 aLimitOverB = lnRatioApprox;
// Must be > 0; otherwise fall back
if (!(aLimitOverB > int128(0))) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// aLimit = b * aLimitOverB (in Q64.64)
int128 aLimit64 = b.mul(aLimitOverB);
// If computed aLimit is less than requested a, use the truncated value.
if (aLimit64 < a) {
a = aLimit64;
} else {
// console2.log("balanced2: limitPrice does not truncate input");
}
// Note: after potential truncation we continue with the polynomial approximation below
}
// Small-signal delta already computed above; reuse it
int128 absDelta = delta >= int128(0) ? delta : delta.neg();
// Allow balanced pools only: require |delta| <= 1% (approx ln(1.01) ~ 0.00995; we use conservative 0.01)
int128 DELTA_MAX = ABDKMath64x64.divu(1, 100); // 0.01
if (absDelta > DELTA_MAX) {
// Not balanced within 1% -> use exact routine
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Scaled input u = a / b (Q64.64). For polynomial approximation we require moderate u.
int128 u = a.mul(invB);
if (u <= int128(0)) {
// Non-positive input -> behave like exact implementation (will revert if invalid)
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Restrict to a conservative polynomial radius for accuracy; fallback otherwise.
// We choose u <= 0.5 (0.5 in Q64.64) as safe for cubic approximation in typical parameters.
int128 U_MAX = ABDKMath64x64.divu(1, 2); // 0.5
if (u > U_MAX) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Now compute a two-tier approximation using Horner-style evaluation to reduce mul/divs.
// Primary tier (cheap quadratic): accurate for small u = a/b.
// Secondary tier (cubic correction): used when u is moderate but still within U_MAX.
// Precomputed thresholds
int128 U_TIER1 = ABDKMath64x64.divu(1, 10); // 0.1 -> cheap quadratic tier
int128 U_MAX_LOCAL = ABDKMath64x64.divu(1, 2); // 0.5 -> still allowed cubic tier
// u is already computed above
// Compute X = u*(1 + delta) - u^2/2
int128 u2 = u.mul(u);
int128 X = u.mul(ONE.add(delta)).sub(u2.div(ABDKMath64x64.fromUInt(2)));
// Compute X^2 once
int128 X2 = X.mul(X);
int128 lnApprox;
if (u <= U_TIER1) {
// Cheap quadratic ln(1+X) ≈ X - X^2/2
lnApprox = X.sub(X2.div(ABDKMath64x64.fromUInt(2)));
} else if (u <= U_MAX_LOCAL) {
// Secondary cubic correction: ln(1+X) ≈ X - X^2/2 + X^3/3
int128 X3 = X2.mul(X);
lnApprox = X.sub(X2.div(ABDKMath64x64.fromUInt(2))).add(X3.div(ABDKMath64x64.fromUInt(3)));
} else {
// u beyond allowed range - fallback
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
int128 approxOut = b.mul(lnApprox);
// Safety sanity: approximation must be > 0
if (approxOut <= int128(0)) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Cap to available j balance: if approximated output exceeds q_j, it's likely approximation break;
// fall back to the exact solver to handle capping/edge cases.
int128 qj64 = s.qInternal[j];
if (approxOut >= qj64) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
// Everything looks fine; return approximated amountOut and used amountIn (a)
amountIn = a;
amountOut = approxOut;
// Final guard: ensure output is sensible and not NaN-like (rely on positivity checks above)
if (amountOut < int128(0)) {
return LMSRStabilized.swapAmountsForExactInput(s, i, j, a, limitPrice);
}
return (amountIn, amountOut);
}
}