📊 statistics

Revenue Maximization Under Sequential Price Competition Via The Estimation Of s-Concave Demand Functions

This paper proposes a dynamic pricing policy for multiple sellers in a sequential price competition setting with unknown nonlinear demand, proving that the policy converges to the Nash equilibrium at a rate of $O(T^{-1/7})$ and achieves $O(T^{5/7})$ regret by leveraging semi-parametric least-squares estimation under $s$ -concave demand constraints.

Original authors: Daniele Bracale, Moulinath Banerjee, Cong Shi, Yuekai Sun

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniele Bracale, Moulinath Banerjee, Cong Shi, Yuekai Sun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling marketplace where dozens of sellers are trying to sell the same kind of product, like lemonade or video games. In this world, every seller is constantly watching their neighbors. If one seller drops their price, they hope to steal customers from the others. But here's the tricky part: they can't see how many cups of lemonade their neighbors actually sold, only what price the neighbors are charging. This creates a high-stakes guessing game. The field of science that studies these strategic moves is called game theory, and when it involves learning over time, it's known as online learning. The big question researchers ask is: Can sellers figure out the perfect price to charge just by watching the market, without knowing the secret recipe for how customers react? If they can't, they might lose money or get stuck in a cycle of price wars that hurts everyone.

This paper tackles that exact puzzle. The authors, a team of statisticians and economists, propose a new way for sellers to learn and compete. They suggest that instead of guessing a simple, straight-line relationship between price and sales (which is often too simple for real life), sellers should use a flexible, shape-shifting model. They call this model "s-concave," which is a fancy way of saying the demand curve has a specific, predictable bend to it, but doesn't have to be a straight line. The researchers designed an algorithm where sellers first spend some time experimenting with random prices to gather data, and then use that data to estimate the hidden shape of the demand curve. Once they have a good guess, they switch to setting prices that maximize their profit.

The paper finds that if all sellers use this specific learning strategy, they eventually settle down into a stable state called a "Nash Equilibrium." In this state, no seller can make more money by changing their price alone, even though they started with zero knowledge of how customers behave. The authors proved mathematically that the sellers' prices will get closer and closer to this perfect balance as time goes on. They also calculated exactly how fast this happens and how much money a seller might lose while learning (a concept called "regret"). Their results show that the learning process is efficient, with the sellers' mistakes shrinking at a predictable rate. Through computer simulations, they confirmed that their method works well even when the market is noisy or when the sellers have different levels of sensitivity to price changes. The study essentially provides a roadmap for how smart, data-driven sellers can navigate a chaotic market and find a stable, profitable rhythm without needing to know the future.

Technical Summary: Revenue Maximization Under Sequential Price Competition via the Estimation of s-Concave Demand Functions

Problem Formulation

The paper addresses the problem of dynamic pricing in a competitive market consisting of $N$ sellers over a finite horizon of $T$ periods. In each period, sellers simultaneously set prices, which are made public, while their realized demands remain private. The demand for seller $i$ depends nonlinearly on the vector of all sellers' prices. Specifically, the mean demand follows a monotone single-index model:
$\mathbb{E}[y_i | p] = \psi_i(\langle \theta_i, p \rangle) = \psi_i(-\beta_i p_i + \langle \gamma_i, p_{-i} \rangle)$
where $\theta_i$ is an unknown parameter vector and $\psi_i$ is an unknown, monotone, and $s$ -concave link function. The sellers do not know the demand parameters or the functional form of the link functions. The objective is to design a pricing policy that maximizes cumulative revenue (minimizes regret) relative to a dynamic benchmark, while ensuring that the sequence of prices converges to the Nash Equilibrium (NE) that would exist under full information.

Methodology

The authors propose a two-phase algorithm, SPE-BR (Semi-Parametric Estimation then Best-Response):

Exploration Phase: Sellers sample prices from a common, known distribution (assumed to be elliptically symmetric) for a duration of $\tau \propto T^\xi$ $τ \propto T^{ξ}$ .
- Parametric Estimation: Using a subset of the data, sellers estimate the linear index parameters $\theta_i$ (specifically the direction of the vector) using a least-squares estimator adapted for single-index models under elliptical symmetry.
- Nonparametric Estimation: Using the remaining data and the estimated index, sellers estimate the link function $\psi_i$ . Crucially, this is done via shape-constrained least squares estimation under the assumption that $\psi_i$ is $s$ -concave. The authors establish that $s$ -concavity is equivalent to the condition that the seller's virtual valuation function is increasing, a necessary condition for the existence of a unique NE.
Exploitation Phase: For the remaining periods, sellers set prices by solving the estimated best-response problem based on the learned demand model.

Key Contributions

1. Theoretical Framework: s-Concavity and Equilibrium

The paper generalizes existing linear and log-concave demand models by introducing $s$ -concavity as the shape constraint for the demand link function.

The authors prove that the condition required for the existence and uniqueness of a Nash Equilibrium (specifically, that the virtual valuation $\phi_i(u) = u + \psi_i(u)/\psi'_i(u)$ is increasing with a bounded derivative) is mathematically equivalent to the link function $\psi_i$ being $(c_i - 1)$ -concave.
This connection allows the use of fully data-driven, tuning-free shape-constrained estimators to guarantee equilibrium convergence.

2. Statistical Contributions: Concentration Inequalities

A significant technical contribution is the derivation of new concentration results for the least-squares estimator under shape constraints.

The authors establish a uniform (supremum-norm) convergence rate for the nonparametric estimator of an $s$ -concave regression function. Previous literature often provided only $L_2$ convergence rates, which are insufficient for establishing regret bounds in dynamic pricing where the price distribution in the exploitation phase may differ from the exploration phase.
They also provide a concentration inequality for the parametric component $\theta_i$ under elliptically symmetric designs, improving upon prior results that only established convergence in probability or distribution.

3. Regret Bounds and Convergence Rates

The paper derives the performance guarantees of the proposed policy:

Regret: The individual regret for each seller is bounded by $\tilde{O}(N^{3/2} T^{5/7})$ . This is achieved by optimizing the exploration length $\tau$ and the split of data between parametric and nonparametric estimation.
Convergence to NE: The joint prices converge to the Nash Equilibrium at a rate of $\tilde{O}(N^{3/4} T^{-1/7})$ .
The analysis demonstrates that the optimal exploration exponent is $\xi = 5/7$ , balancing the trade-off between learning accuracy and revenue loss during exploration.

Results and Validation

The theoretical findings are supported by numerical experiments involving markets with $N \in \{2, 4, 6\}$ sellers.

Simulations confirm that the algorithm converges to the Nash Equilibrium and that the empirical regret rates match or exceed the theoretical upper bounds.
The study investigates the robustness of the algorithm to the misspecification of the concavity parameter $s$ . Results indicate that choosing a conservative $s' \leq s$ preserves the convergence rates, provided the contraction condition for the best-response map is maintained.
The experiments also show that the algorithm remains robust even when sellers have slightly misaligned exploration phases, though the theoretical guarantees rely on a common exploration phase.

Significance and Claims

The paper claims to offer significant insights into dynamic competition-aware pricing by moving beyond restrictive linear or fixed-parametric demand assumptions. By leveraging $s$ -concavity, the authors provide a flexible, non-parametric framework that:

Generalizes existing models: It strictly generalizes linear demand (where $s=0$ corresponds to log-concavity) and allows for more complex, realistic demand behaviors.
Ensures Equilibrium: It provides a rigorous theoretical proof that learning-based pricing policies can converge to a Nash Equilibrium without requiring sellers to share private demand information.
Advances Nonparametric Learning: The derivation of uniform convergence rates for $s$ -concave estimators contributes to the broader study of non-parametric learning in strategic decision-making, offering tools that are applicable beyond just pricing.

The authors emphasize that their approach is "fully data-driven" and "tuning-parameter-free" regarding the shape constraints, relying only on the structural assumption of $s$ -concavity which is economically motivated by the properties of virtual valuations. They do not claim minimax optimality of the regret rate itself, but rather that the derived rates are optimal within the context of their derived upper bounds.