Multi-Asset Market Making Fundamentals

• Multi-asset market making is a liquidity provision strategy that manages quoting and risk across correlated financial instruments using dynamic stochastic control models.
• The methodology employs closed-form approximations, quadratic ansatz, and factor-model reductions to mitigate the curse of dimensionality and enhance computational efficiency.
• It is applied in both centralized order book and decentralized AMM settings, optimizing inventory management and capital allocation under adverse selection and volatility.

Multi-asset market making concerns liquidity provision, quoting, and risk management across a portfolio of correlated financial instruments. Its mathematical treatment generalizes classical single-asset limit order book market making to a vector-valued, high-dimensional dynamic stochastic control problem. Both centralized (order book) and decentralized (automated market maker, AMM) settings feature diverse frameworks for liquidity provision, driven by microstructure, risk preferences, inventory limits, and adverse selection. Multi-asset approaches highlight inventory-risk coupling via cross-asset correlations, capital allocation efficiency, and computational complexity arising from the "curse of dimensionality." The literature spans stochastic optimal control, convex-analytic AMM design, and more recent factor and dimensionality reduction techniques.

1. Model Formulations and Market Structure

Multi-asset market making models start with $d$ risky assets, whose reference prices $S_t = (S^1_t,\dots,S^d_t)^\top$ typically follow correlated Itô or Brownian processes:

$dS^i_t = \sigma^i dW^i_t,$

where $(W^1_t,\ldots,W^d_t)$ is a $d$-dimensional Brownian motion with covariance matrix $\Sigma = [\rho^{ij}\sigma^i\sigma^j]$ (Guéant, 2016). For OTC or exchange settings, the market maker's inventory $q_t = (q^1_t,\ldots,q^d_t)$ evolves according to executed trades, often modeled as Poisson arrivals with intensities $\lambda^{i,b}_t, \lambda^{i,a}_t$ determined by quoted bid/ask prices offset from reference mid-prices.

Automated market makers (AMMs) extend this concept to decentralized settings, where pools simultaneously offer trading across $n$ tokens with reserves $\alpha_i^t$ (Forgy et al., 2021). Each trade must be self-financing and maintain a pre-chosen rebalancing rule or portfolio weight vector.

Key state variables:

• Price vector $S_t = (S^1_t,\dots,S^d_t)^\top$0, inventory vector $S_t = (S^1_t,\dots,S^d_t)^\top$1, cash process $S_t = (S^1_t,\dots,S^d_t)^\top$2
• For AMMs: pool reserves $S_t = (S^1_t,\dots,S^d_t)^\top$3, implied prices $S_t = (S^1_t,\dots,S^d_t)^\top$4, portfolio weights $S_t = (S^1_t,\dots,S^d_t)^\top$5

2. Objective Functions and Dynamic Programming

Canonical optimization criteria arise:

• Model A (CARA utility):

$S_t = (S^1_t,\dots,S^d_t)^\top$6

• Model B (Risk-neutral P&L, quadratic penalty):

$S_t = (S^1_t,\dots,S^d_t)^\top$7

with terminal (and possibly running) quadratic penalties for risk.

The associated Hamilton-Jacobi-Bellman (HJB) PDE in the $S_t = (S^1_t,\dots,S^d_t)^\top$8-asset case admits reduction via an ansatz,

$S_t = (S^1_t,\dots,S^d_t)^\top$9

transforming the PDE in $dS^i_t = \sigma^i dW^i_t,$0 into a system of ODEs for $dS^i_t = \sigma^i dW^i_t,$1 (Guéant, 2016, Bergault et al., 2018). This reduction enables closed-form approximations and scalable computation. For AMMs, utility-optimization reduces to a convex-analytic mechanism design problem, with strong duality to the optimal transport formulation (Curry et al., 2024).

Table: Model Criteria

Model	Objective Functional	Typical Control Variable
Model A	CARA utility (exponential risk aversion)	Bid/ask quotes $dS^i_t = \sigma^i dW^i_t,$2
Model B	Linear P&L minus running inventory variance	Bid/ask quotes $dS^i_t = \sigma^i dW^i_t,$3
AMM Design	Convex utility/optimal menu	Trade schedule, price menu functions

3. Closed-form Approximations, Factorization, and Scalability

Exact solution of multi-asset HJBs is intractable for large $dS^i_t = \sigma^i dW^i_t,$4 due to the dimensionality. The literature develops systematic approximations:

• Taylor/PDE expansion: Expanding the ODEs in trade size and price offset, the optimal value function $dS^i_t = \sigma^i dW^i_t,$5 is approximated by a quadratic or series expansion (Guéant, 2016, Bergault et al., 2018).
• Quadratic-ansatz and Riccati methods: Seeking quadratic-in-inventory solutions, which reduce the backward ODEs to Riccati equations yielding explicit matrix solutions for hedging and quoting terms (Bergault et al., 2018, Bergault et al., 2019).
• Factor-model reduction: For $dS^i_t = \sigma^i dW^i_t,$6, factorizing the covariance matrix as $dS^i_t = \sigma^i dW^i_t,$7 with $dS^i_t = \sigma^i dW^i_t,$8 dominant factors allows reduction of the HJB to $dS^i_t = \sigma^i dW^i_t,$9 dimensions, approximating inventory risk and quoting in a lower-dimensional space (Bergault et al., 2019).
• Asymptotic/ergodic regimes: Long-horizon limits ($(W^1_t,\ldots,W^d_t)$0) yield stationary quadratic control rules with cross-inventory $(W^1_t,\ldots,W^d_t)$1 matrices quantifying correlation-induced risk coupling (Guéant, 2016, Bergault et al., 2018).

Table: Dimensionality Reduction

Technique	Dimension After Reduction	Core Mathematical Object
Quadratic-ansatz	ODE in $(W^1_t,\ldots,W^d_t)$2 (full)	Matrix Riccati equation for $(W^1_t,\ldots,W^d_t)$3
Factor-reduction	ODE in $(W^1_t,\ldots,W^d_t)$4 ($(W^1_t,\ldots,W^d_t)$5 factors)	Lower-dimensional HJB, $(W^1_t,\ldots,W^d_t)$6
Asymptotic	None (stationary)	Closed-form $(W^1_t,\ldots,W^d_t)$7 for cross-asset risk

4. Cross-Asset Coupling and Risk Management

In all models, the cross-asset correlation structure fundamentally alters quoting and risk. The off-diagonal elements of the “Gamma" matrix $(W^1_t,\ldots,W^d_t)$8 encode how inventory in one asset affects optimal quotes in another. For $(W^1_t,\ldots,W^d_t)$9, holding long asset $d$0 leads to increased risk, widening and/or shifting optimal quotes in $d$1.

Inventory is managed via quadratic penalties, defining ellipsoidal iso-risk surfaces in the inventory space, determined by covariance and sensitivity parameters. Factor-model approaches further clarify that well-diversified portfolios (orthogonal to principal risk factors) incur lower penalties (Guéant, 2016, Bergault et al., 2019).

Transaction costs and block/execution size further widen spreads; for each asset/size class, quotes incorporate skew and risk premium proportional to both inventory and cross-asset exposure (Bergault et al., 2019).

5. Automated Market Makers: Multi-Asset Pool Geometry

Decentralized AMMs extend these principles to on-chain market making. The family of multi-asset AMM invariants is derived from self-financing and fixed rebalancing:

$d$2

where $d$3 are growth factors, $d$4 the rebalancing parameter, and $d$5 portfolio weights (Forgy et al., 2021). Special cases include the constant-product rule (Uniswap/Balancer) and constant-sum (stablecoin pool).

AMM pricing functions are chosen to enforce internal no-arbitrage and can be optimized to align with a liquidity provider’s optimal static portfolio weights for efficiency (He et al., 2024). Sophisticated mechanisms use convex utility theory and optimal transport duality to design optimal multi-asset, menu-based pricing that incorporates mixed bundling and payment-in-kind (Curry et al., 2024).

6. Empirical and Numerical Insights

Case studies demonstrate the validity of closed-form approximations and reduction schemes. In two-asset settings (e.g., credit indices, SPX/VIX pairs), optimal quotes manifest as smooth, cross-inventory surfaces sensitive to correlation $d$6 and risk aversion $d$7, with strong concordance between numerical solutions and approximations for moderate inventory (Guéant, 2016, Rosenbaum et al., 2022).

Factor-model approaches show orders-of-magnitude speedup compared to full-grid PDEs, with up to 30-asset portfolios solvable via two-factor reduction (Bergault et al., 2019). Practical implementations confirm that cross-asset risk pooling and inventory diversification materially improve mean-variance trade-offs.

In advanced decentralized designs, protocols such as Dynamic Function Market Maker (DFMM) synchronize pool pricing to external order books, control inventory risk using synthetic accounting tokens, and utilize digital derivatives (swaptions) for robust inventory insurance, all while maintaining arbitrage-aligned, cross-asset execution (Abgaryan et al., 2023).

7. Theoretical Limits and Open Directions

Optimal multi-asset market making remains characterized by the tension between statistical optimality, computational tractability, and market microstructure constraints. Several robust findings:

• The cross-asset $d$8 term quantifies the full inventory-risk mutualization and should be regularly recomputed.
• Factor-model and quadratic-value-function proxies are generally valid for small/medium risk aversion and inventory size; the accuracy decreases in highly nonlinear or jump-driven markets.
• AMM designs with "menu-based" or "bundling" allocations strictly outperform single-asset bid-ask policies under adverse selection and provide quantifiable profit gains in simulation (Curry et al., 2024).
• The optimal AMM fee is an increasing function of the underlying exchange rate volatility, and pricing curvature should match the provider's mean-variance optimal allocation (He et al., 2024).

Active research extends these results to dynamic belief updates, high-dimensional option books, rough volatility regimes, and more complex adverse selection environments. Dimensionality reduction, mechanism design, and robust risk management remain central open avenues for multi-asset market making theory and implementation.