Riemannian Geometry of Optimal Rebalancing in Dynamic Weight Automated Market Makers

This paper establishes that the optimal rebalancing trajectory for Dynamic Weight Automated Market Makers corresponds to a geodesic under the Fisher-Rao metric (SLERP in Hellinger coordinates), thereby proving that the recursive AM-GM bisection heuristic used in prior work lies exactly on this loss-minimizing path.

The Gist

Here is an explanation of the paper "Riemannian Geometry of Optimal Rebalancing in Dynamic Weight Automated Market Makers," translated into simple, everyday language with creative analogies.

The Big Picture: The "Smooth Move" Problem

Imagine you are managing a basket of fruits (tokens) in a smart vending machine (an Automated Market Maker or AMM). Every day, you decide to change the recipe: maybe you want more apples and fewer bananas.

In the old days, you would just swap the weights instantly. But if you do that, a "smart shopper" (an arbitrageur) sees the price is wrong and buys the cheap fruit and sells the expensive one until the prices match. This process costs the pool money. It's like trying to turn a heavy steering wheel all at once; the car jerks, and you lose control.

The Solution: Instead of turning the wheel instantly, you turn it slowly over several small steps. This spreads out the cost. But here is the tricky question: What is the perfect path to turn the wheel? Should you turn it in a straight line? A curve? A zig-zag?

This paper answers that question using a branch of math called Riemannian Geometry (which studies curved spaces), but we can understand it without the heavy math.

---

Analogy 1: The "Fruit Basket" and the "Tax"

When you change the weights of your fruit basket, you create a temporary "tax" or "loss" because the market has to adjust.

• The Paper's Discovery: The authors found that this "tax" isn't random. It behaves exactly like a concept in information theory called KL Divergence.
• The Metaphor: Think of the weight of your fruits as a map. If you move from Point A to Point B, the "cost" of the move depends on the shape of the terrain.
  • In a flat world (Euclidean geometry), the shortest path is a straight line.
  • In this "Fruit Basket World," the terrain is curved. The "straightest" line here is actually a curve when viewed from the outside.

Analogy 2: The "Hill" and the "Sphere"

The authors realized that the best way to move your weights is to imagine them not as numbers on a flat list, but as points on the surface of a sphere (like the Earth).

• The Transformation: They used a mathematical trick (called the Hellinger embedding) to turn your fruit weights into coordinates on a sphere.
• The Result: On this sphere, the "shortest path" between two points is a Great Circle (like the flight path of an airplane going from London to New York).
• The Name: The method to follow this path is called SLERP (Spherical Linear Interpolation).
  • Simple translation: Instead of moving in a straight line on a flat map, you are walking along the curve of a globe. This is the most efficient route to minimize the "tax" (arbitrage loss).

The "Magic Midpoint" Discovery

Before this paper, experts had a "rule of thumb" (a heuristic) for finding the perfect middle point between two weight settings. They took the Average (Arithmetic Mean) and the Geometric Mean of the weights, added them together, and normalized the result.

• The Surprise: The authors proved that this "rule of thumb" isn't just a lucky guess. It is exactly the same as the perfect middle point on the sphere (SLERP).
• The Metaphor: Imagine you are trying to find the exact middle of a curved road.
  • Method A (The Old Way): You guess by averaging the start and end points.
  • Method B (The New Way): You calculate the perfect curve.
  • The Paper's Finding: When you look at the exact middle of the road, Method A and Method B land on the exact same spot. The "rule of thumb" was actually a shortcut to the perfect solution all along!

The "No-Calculator" Trick

Calculating the perfect curve on a sphere usually requires complex trigonometry (sines, cosines, and arccos), which is hard and expensive for computers (especially on a blockchain).

• The Breakthrough: The authors found a way to build the perfect curve using only addition, multiplication, and square roots.
• The Analogy: Imagine you need to draw a perfect circle. Usually, you need a compass and complex math. But this paper says: "If you keep folding the paper in half and finding the middle of the fold, you can trace the perfect circle without ever using a compass."
• Why it matters: This allows the "smart vending machine" to calculate the perfect rebalancing path very quickly and cheaply, without needing heavy computing power.

The "Near the Edge" Problem

The math works perfectly when you have plenty of fruit. But what if you are almost out of one type of fruit (a weight near zero)?

• The paper shows that near the "edge" of the basket (where a weight is very small), the "tax" gets much higher.
• The "Perfect Curve" (SLERP) is still the best choice, but it gets slightly less perfect the closer you get to the edge. However, even in these difficult cases, it is still vastly superior to the old "straight line" method.

Summary: What Does This Mean for You?

1. The Problem: Changing the mix of assets in a crypto pool costs money if done clumsily.
2. The Geometry: The best way to change the mix is to follow a specific curved path (a geodesic) on a mathematical sphere, not a straight line.
3. The Shortcut: A previously popular "rule of thumb" (mixing averages) actually hits the perfect path exactly in the middle.
4. The Tool: We can now calculate this perfect path using simple math (no complex trigonometry needed), making it cheap and fast to run on blockchains.

In one sentence: This paper proves that the most efficient way to rebalance a crypto pool is to walk along a curved path on a sphere, and it gives us a simple, cheap recipe to do exactly that.

Technical Summary

Here is a detailed technical summary of the paper "Riemannian Geometry of Optimal Rebalancing in Dynamic Weight Automated Market Makers" by Matthew Willetts.

1. Problem Statement

In Temporal Function Market Making (TFMM), dynamic weight Automated Market Makers (AMMs) rebalance their asset holdings by adjusting the weights of a Geometric Mean Market Maker (G3M) pool over time. When weights change from a starting vector $w_{start}$ to a target $w_{end}$, the pool creates a temporary mispricing relative to the market, generating an arbitrage opportunity.

Arbitrageurs exploit this mispricing, extracting value from the pool (Loss Versus Rebalancing or LVR). The paper addresses the problem of minimizing this arbitrage loss when spreading a weight update over multiple intermediate steps ($f$ steps) rather than executing it in a single block. While prior work proposed an approximately optimal heuristic using the average of Arithmetic and Geometric means ((AM+GM)/normalise), the geometric justification for its effectiveness and the existence of a truly optimal trajectory were unexplained.

2. Methodology and Theoretical Framework

The author employs Information Geometry to model the weight space of the AMM.

• Arbitrage Loss as KL Divergence: The paper establishes that the log-loss incurred by the pool during a weight update is exactly the Kullback–Leibler (KL) divergence between the new and old weight vectors:
  $$-\log r = D_{KL}(w_{end} \parallel w_{start}) = \sum_i w_{end, i} \log \frac{w_{end, i}}{w_{start, i}}$$
  where $r$ is the retention ratio (final value / initial value).

• Riemannian Structure: The local second-order expansion of the KL divergence defines the Fisher–Rao metric on the probability simplex. The paper proves that the per-step arbitrage loss is proportional to the squared arc-length under this metric.
  $$\text{Loss} \approx \frac{1}{2} \sum_i \frac{(\Delta w_i)^2}{w_i}$$

• Hellinger Embedding: To simplify the geometry, the author maps the weight simplex $\Delta^{N-1}$ to the positive orthant of the unit sphere $S^{N-1}_+$ using the Hellinger coordinates $\eta_i = \sqrt{w_i}$. Under this isometry, the Fisher–Rao metric becomes a scaled version of the standard Euclidean metric on the sphere. Consequently, geodesics (shortest paths) on the weight simplex correspond to great circles on the sphere.

3. Key Contributions

A. Identification of the Optimal Trajectory (SLERP)

The paper proves that the trajectory minimizing the total quadratic arbitrage loss (the leading-order expansion of the KL cost) is the Spherical Linear Interpolation (SLERP) in Hellinger coordinates.

• Geometric Interpretation: The optimal path is a constant-speed geodesic on the unit sphere.
• Algorithm: The weights are interpolated by mapping $w \to \sqrt{w}$, performing linear interpolation on the sphere (great circle), and mapping back via squaring.

B. Resolution of the (AM+GM)/Normalise Heuristic

A major contribution is explaining why the heuristic from prior work [1], defined as $\tilde{w} \propto \text{AM}(w_{start}, w_{end}) + \text{GM}(w_{start}, w_{end})$, is effective.

• Theorem 2: The paper proves that the midpoint of the SLERP geodesic is exactly equal to the (AM+GM)/normalise midpoint for any number of tokens and any magnitude of weight change.
• Mechanism: This identity arises from the binomial expansion of the squared sum of square roots: $(\sqrt{a} + \sqrt{b})^2 = (a+b) + 2\sqrt{ab}$, which corresponds to $2(\text{AM} + \text{GM})$.
• Implication: The heuristic lies exactly on the optimal geodesic at the midpoint, explaining its high performance (~95% of full optimization).

C. Trigonometric-Free Implementation (Corollary 3)

While standard SLERP requires trigonometric functions (sin, cos, arccos), the paper derives a recursive bisection algorithm for step counts $f = 2^d$.

• By recursively inserting the (AM+GM)/normalise midpoint between existing points, one can generate the exact SLERP trajectory at power-of-two intervals.
• Significance: This allows for an on-chain implementation using only elementary arithmetic (addition, multiplication, square root, normalization) without expensive transcendental functions.

D. Sub-optimality Bounds

The paper provides a rigorous bound on the sub-optimality of SLERP relative to the exact (non-quadratic) KL cost minimization.

• Theorem 3: The relative sub-optimality is bounded by $O(\Omega^2 / f^2)$, where $\Omega$ is the total angular distance and $f$ is the number of steps.
• This confirms that for reasonable step counts, the difference between the SLERP geodesic and the true global optimum is negligible.

4. Results and Validation

The author validates these theoretical findings through numerical experiments with a 3-token pool:

• Loss Uniformity: SLERP achieves near-perfect uniformity in per-step arbitrage loss (standard deviation/mean $\approx 0.0002$), whereas linear interpolation and the (AM+GM) heuristic show significantly higher variance.
• Comparison with Brute Force: Using L-BFGS-B optimization to find the true global optimum, SLERP was found to be virtually indistinguishable from the optimal path. At $f=1000$, the relative cost improvement of brute-force over SLERP was $\sim 1.2 \times 10^{-5}$.
• Boundary Behavior: In near-boundary scenarios (where weights approach 0), the Fisher–Rao metric diverges. While SLERP remains robust, the paper notes that for single-step ($f=2$) updates near the boundary, the Lambert W function (which optimizes the exact retention ratio) can outperform SLERP by up to 10%. However, for multi-step rebalancing, SLERP remains superior.
• Visual Confirmation: Plots of weight trajectories show that while (AM+GM)/normalise and SLERP are visually indistinguishable at high step counts, their block-to-block weight changes differ slightly, with SLERP maintaining constant metric speed.

5. Significance and Conclusion

This paper provides a unifying geometric perspective on AMM rebalancing:

1. Theoretical Unification: It identifies the Fisher–Rao metric as the natural geometry governing arbitrage costs in geometric mean markets, linking information geometry directly to DeFi economics.
2. Practical Optimization: It validates the (AM+GM)/normalise heuristic as a mathematically sound approximation of the geodesic midpoint, justifying its use in production systems.
3. Implementation Efficiency: The recursive bisection method offers a practical, gas-efficient way to implement optimal rebalancing on-chain without complex math libraries.
4. Generalizability: The framework suggests that any AMM parameterized by a probability simplex vector will have its rebalancing costs governed by information-geometric divergences, opening avenues for optimizing other AMM designs (e.g., concentrated liquidity).

In summary, the paper transforms the problem of dynamic AMM rebalancing from a numerical optimization challenge into a geometric one, proving that SLERP in Hellinger coordinates is the natural, near-optimal solution, and that existing heuristics are exact realizations of this geometry at key points.