GBB Semi-Feedback Fixed-Price Mechanisms

• The paper introduces a two-phase algorithm combining profit accumulation and Exp3 bandit learning to achieve near-optimal O(T^(2/3)) regret under strict GBB constraints.
• The mechanism leverages partial feedback by using unbiased surrogate estimators for gains-from-trade while maintaining nonnegative cumulative profit.
• The research delineates a regret landscape that clearly separates GBB protocols from stricter SBB/WBB models across independent and adversarial value settings.

Global–Budget–Balanced (GBB) semi-feedback fixed-price mechanisms are a class of online learning algorithms for repeated bilateral trade that optimize regret under strong budget constraints and limited information. In these protocols, a mechanism posts two prices for buyer and seller, learns only partial feedback about trade outcomes and seller value, and is required to maintain a nonnegative ex post profit over $T$ rounds. Recent research has rigorously characterized the achievable regret rates in this setting for both independent and adversarial value models, culminating in tight $\widetilde{O}(T^{2/3})$ upper and $\Omega(T^{2/3})$ lower bounds for adversarial values and $\widetilde{\Theta}(T^{2/3})$ for independent values (Jin, 23 Jan 2026, Chen et al., 6 Apr 2025). This establishes a sharp separation from settings with less restrictive budget constraints or more informative feedback.

1. Formal Model and Problem Setting

The $T$-round bilateral trade protocol considered in GBB semi-feedback fixed-price mechanisms involves a seller and buyer with private valuations $(v^S_t, v^B_t) \in [0,1]^2$ at each round $t$. The mechanism posts a pair of prices $(p_t, q_t)$, where $p_t$ is offered to the seller and $q_t$ to the buyer, both in $[0,1]$. Trade succeeds ($\mathbb{I}_t =1$) exactly when $v^S_t \le p_t$ and $v^B_t \ge q_t$. The realized gains-from-trade (GFT) per round are $G_t(p_t, q_t) = (v^B_t - v^S_t)\mathbb{I}_t$, and the profit (surplus) is $\Pi_t(p_t, q_t) = (q_t - p_t)\mathbb{I}_t$.

The mechanism satisfies the global-budget-balanced (GBB) constraint:

$\sum_{t=1}^T \Pi_t(p_t, q_t) \ge 0$

ensuring nonnegative cumulative profit across all rounds, unlike strong (per-round) budget balance (SBB), which is infeasible to achieve in this feedback regime. Semi-feedback means that in each round, the mechanism observes only $(v^S_t, \mathbb{I}_t)$: it knows the seller’s value and whether a trade occurred, but not the buyer’s value.

The regret is measured against the benchmark of the best single (SBB) price $p^*$:

$$\mathrm{Regret}(T) = \max_{p^* \in [0,1]} \sum_{t=1}^T \left[ \mathbf{1}\{v^S_t \le p^* \le v^B_t\} - \mathbf{1}\{v^S_t \le p_t \le v^B_t\} \right]$$

where $\mathbf{1}\{v^S_t \le p^* \le v^B_t\}$ is the indicator that trade would succeed at benchmark price $p^*$.

2. Main Algorithmic Paradigm and Upper Bound Construction

The state-of-the-art GBB semi-feedback fixed-price mechanism is a two-phase algorithm (“ALG”) consisting of profit accumulation followed by bandit-style learning on a nearly-diagonal discretization (Jin, 23 Jan 2026).

Phase I: Profit Accumulation

• Leverages a black-box subroutine (BCCF24) restricted to posting prices in the upper-left half-space (i.e., always $p \le q$), ensuring nonnegative per-round profit.
• Stops once cumulative profit exceeds $\beta = O(T^{2/3})$ or after $T$ rounds.
• Achieves $O(T^{2/3}\log^{5/3}T)$ regret with high probability and maintains GBB.

Phase II: Exp3-Type Bandit Learning

• Discretizes the SBB diagonal into $K \sim T^{1/3}/\mathrm{polylog}(T)$ grid points: $\{(k/K, (k-1)/K)\}_{k=1}^K$.
• In each subsequent round, forms exponential-weights (Exp3) over these grid points using importance-weighted unbiased estimators of a surrogate reward, based only on semi-feedback.
• Mixes “exploitation” (selecting according to weights) and “exploration” (random sampling pair of prices).
• Ensures GBB by allowing $p > q$ only when sufficient surplus buffer is accumulated.

The main theorem asserts that for absolute constants $C,\alpha > 0$:

$\mathrm{Regret}(T) \le C T^{2/3} \log^\alpha T$

with GBB holding ex post (indeed, $\alpha=5/3$ and $C \approx 310$) (Jin, 23 Jan 2026).

3. Tight Regret Lower Bound: Adversarial and Independent Values

Matching lower bounds have been established for GBB semi-feedback mechanisms. In particular, [CJLZ25, see (Jin, 23 Jan 2026)] proves that no GBB mechanism in this setting can obtain regret $o(T^{2/3})$, even for independent seller and buyer values.

The construction partitions the rounds into $K \approx T^{1/3}$ contiguous blocks. In each block $k$, value pairs $(v^S_t, v^B_t)$ are concentrated near two points such that the optimal SBB price is near $k/K$. Any exploration outside that diagonal incurs large local regret within the block, while information-theoretic constraints and the structure of the feedback signal preclude circumventing exploration cost. A counting argument yields overall regret $\Omega(T^{2/3})$.

A plausible implication is that the $\Theta(T^{2/3})$ scaling is intrinsic to this feedback-budget regime. For correlated or adversarial values under GBB and semi-feedback, prior work showed higher $\Theta(T^{3/4})$ regret (Chen et al., 6 Apr 2025).

4. Regret Landscape Across Value, Feedback, and Balance Models

The latest research provides a unified minimax regret landscape for fixed-price bilateral trade mechanisms, covering all combinations of:

• Value Models: Independent, correlated, adversarial,
• Feedback: Full, two-bit/one-bit (partial), semi (semi-transparent).
• Budget Balance: Strong (SBB), weak (WBB), global (GBB).

The following table (from (Jin, 23 Jan 2026, Chen et al., 6 Apr 2025, Cesa-Bianchi et al., 2023)) summarizes tight minimax regret rates (ignoring polylogarithms):

Feedback & BB	Independent Values	Correlated/Adversarial Values
Full + any BB	$\Theta(T^{1/2})$	$\Theta(T^{1/2})$
Partial+SBB/WBB	$\Theta(T)$	$\Theta(T)$
Partial+GBB	$\widetilde{\Theta}(T^{2/3})$	$\widetilde{\Theta}(T^{3/4})$
Semi+GBB	$\widetilde{\Theta}(T^{2/3})$	$\widetilde{\Theta}(T^{2/3})$

This suggests GBB is the critical relaxation enabling sublinear regret in minimal-feedback mechanisms, distinguishing it from SBB/WBB, which suffer linear regret under partial or semi-feedback.

5. Technical Insights: Surrogate Estimation and Semi-Feedback

Semi-feedback presents a fundamental obstacle: the buyer’s value is unobserved, only trade success/failure and seller’s value are revealed. Modern algorithms circumvent this by constructing unbiased surrogate estimators for gains-from-trade at candidate price pairs, using only available signals and importance weighting. In Phase II, surrogate reward $g^t_k$ at grid index $k$ combines the observed $v^S_t$ and $\mathbb{I}_t$:

$$g^t_k = [v^S_t - \frac{k-1}{K}]_+ \, \mathbf{1}\{v^S_t \le \frac{k}{K}\} + [\frac{k}{K} - v^S_t]_+ \, \mathbf{1}\{\frac{k-1}{K} \le v^S_t\}$$

Algorithmic analysis leverages the Exp3 framework for contextual bandits, controlling discretization, exploration cost, and surplus buffer to guarantee both GBB and near-optimal regret.

6. Broader Implications and Open Directions

The resolution of the $\widetilde{\Theta}(T^{2/3})$ regret rate for GBB semi-feedback mechanisms completes the theory of regret minimization in fixed-price bilateral trade under budget constraints and partial information (Jin, 23 Jan 2026, Chen et al., 6 Apr 2025). Extensions of interest include: sharpening constants, incorporating richer feedback (for example, glimpses of buyer’s value), and generalizing to settings with multi-unit or multi-dimensional trade and adversarial/budget constraints.

A plausible implication is that methodologies for surrogate reward estimation and profit-buffered two-phase algorithms may generalize to other settings where minimal feedback and tight budget constraints interact—such as dynamic markets, mechanism design for multi-agent scenarios, and combinatorial auctions.