Jiheng Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Jiheng Zhang has a total of 95 co-authors. Here are the top co-authors by number of shared publications:
- Hailun Zhang
CUHK, Shenzhen
9 shared publications - Liang Zhang
Hong Kong University of Science and Technology
6 shared publications - Rachel Q. Zhang
Tsingshan Institute for Advanced Business Studies and School of Management, Zhejiang University, Hangzhou, China
6 shared publications - Zhenghua Long
School of Business, Nanjing University, Nanjing 210093, China
6 shared publications - Haibin Kan
Shanghai Key Laboratory of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai 200433, China
4 shared publications - Jiahao He
Department of Industrial Engineering and Decision Analytics, The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong Special Administrative Region, China
3 shared publications
Latest Publications
- 12 November 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Manufacturing & Service Operations Management
Online Demand Fulfillment in Highly Asymmetric Markets
Problem definition: We study online demand fulfillment with limited flexibility in highly asymmetric markets, where some locations represented by the set [Formula: see text] have moderate to high market shares, whereas others have very low shares. Xu et al. was the first to establish that a positive generalized chaining gap (GCG) is a necessary and sufficient condition for bounded performance, which is defined as the expected lost sales as the total market size grows. Although a high GCG signals a better performance, the presence of high market asymmetry results in a low GCG and a larger performance bound, making the bound less useful for assessing actual system performance. Therefore, understanding system performance in highly asymmetric markets remains an important question. Methodology/results: We introduce a network condition called [Formula: see text]-connectivity, which ensures that after removing the markets with very low shares, the markets in [Formula: see text] are still connected through the available suppliers. We prove that [Formula: see text]-connectivity is both necessary and sufficient for maintaining bounded performance in highly asymmetric markets. This is achieved through a carefully designed online fulfillment policy that utilizes network partitioning and advanced analyses of mean-reverting processes using probability methods in an asymptotic context. Additionally, we offer strategies for network design to reduce the risk of [Formula: see text]-disconnectivity. Managerial implications: Our findings indicate that even the well-known long-chain structure may not ensure bounded performance in highly asymmetric markets, especially when low-demand markets are dispersed throughout the network. Instead, a “hub-and-spoke” network configuration with a shorter embedded long chain can help reduce the risk of [Formula: see text]-disconnectivity and enhance system performance. Funding: This research was supported in part by the National Natural Science Foundation of China [Grants 72671048 and 72192805] and the Hong Kong Research Grants Council [Grants 16200617, 16200821, 16200019, and 16500225]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2024.1002 .
- 12 November 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Manufacturing & Service Operations Management
Online Demand Fulfillment in Highly Asymmetric Markets
Problem definition: We study online demand fulfillment with limited flexibility in highly asymmetric markets, where some locations represented by the set [Formula: see text] have moderate to high market shares, whereas others have very low shares. Xu et al. was the first to establish that a positive generalized chaining gap (GCG) is a necessary and sufficient condition for bounded performance, which is defined as the expected lost sales as the total market size grows. Although a high GCG signals a better performance, the presence of high market asymmetry results in a low GCG and a larger performance bound, making the bound less useful for assessing actual system performance. Therefore, understanding system performance in highly asymmetric markets remains an important question. Methodology/results: We introduce a network condition called [Formula: see text]-connectivity, which ensures that after removing the markets with very low shares, the markets in [Formula: see text] are still connected through the available suppliers. We prove that [Formula: see text]-connectivity is both necessary and sufficient for maintaining bounded performance in highly asymmetric markets. This is achieved through a carefully designed online fulfillment policy that utilizes network partitioning and advanced analyses of mean-reverting processes using probability methods in an asymptotic context. Additionally, we offer strategies for network design to reduce the risk of [Formula: see text]-disconnectivity. Managerial implications: Our findings indicate that even the well-known long-chain structure may not ensure bounded performance in highly asymmetric markets, especially when low-demand markets are dispersed throughout the network. Instead, a “hub-and-spoke” network configuration with a shorter embedded long chain can help reduce the risk of [Formula: see text]-disconnectivity and enhance system performance. Funding: This research was supported in part by the National Natural Science Foundation of China [Grants 72671048 and 72192805] and the Hong Kong Research Grants Council [Grants 16200617, 16200821, 16200019, and 16500225]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2024.1002 .
- Open Access
- 7 November 2026
- Published by Elsevier in Blockchain: Research and Applications
Publicly Accountable Data Trading for Web 3.0 with Fine-Grained Access Control and Diversified Payment Mechanisms
- 7 November 2026
- Published by Elsevier in Blockchain: Research and Applications
Open AccessPublicly Accountable Data Trading for Web 3.0 with Fine-Grained Access Control and Diversified Payment Mechanisms
- Open Access
- 16 October 2026
- Published by Elsevier in SSRN Electronic Journal
Minimax Learning for Contextual Pricing and Inventory Control under Censored Demand
We study online joint pricing and inventory control for a perishable product over T periods when the seller observes only censored sales. Demand combines an unknown linear function of d-dimensional covariates and price with i.i.d. additive noise from an unknown common distribution. Covariates may depend on past observations, while current noise remains independent of the past and current covariates. Profit optimized over inventory need not be concave in price. Stockouts hide unmet demand and shortage costs, preventing direct profit feedback. We develop a policy combining online quantile calibration, shared residual-probability prediction, and joint priceinventory exploration. An inventory floor based on the quantile forecast preserves observable demand comparisons, while integrated residual predictions rank actions without observing realized profit. Under a feasible quantile prior, bounded noise, and density regularity, the policy achieves expected regret O » T d log(1 + √ T /d). A multiscale lower bound matches this rate for every d ≥ 2 and T ≥ 1, with constants independent of dimension and horizon. The lower bound holds even when the demand intercept, price sensitivity, and noise distribution are known and full demand is observed. Thus, sales censoring and learning the common noise law need not increase the minimax regret order within this model class for d ≥ 2. Numerical experiments illustrate the horizon and dimension dependence and show lower mean regret than four benchmark implementations at every tested horizon on a nonconcave instance.
- 16 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMinimax Learning for Contextual Pricing and Inventory Control under Censored Demand
We study online joint pricing and inventory control for a perishable product over T periods when the seller observes only censored sales. Demand combines an unknown linear function of d-dimensional covariates and price with i.i.d. additive noise from an unknown common distribution. Covariates may depend on past observations, while current noise remains independent of the past and current covariates. Profit optimized over inventory need not be concave in price. Stockouts hide unmet demand and shortage costs, preventing direct profit feedback. We develop a policy combining online quantile calibration, shared residual-probability prediction, and joint priceinventory exploration. An inventory floor based on the quantile forecast preserves observable demand comparisons, while integrated residual predictions rank actions without observing realized profit. Under a feasible quantile prior, bounded noise, and density regularity, the policy achieves expected regret O » T d log(1 + √ T /d). A multiscale lower bound matches this rate for every d ≥ 2 and T ≥ 1, with constants independent of dimension and horizon. The lower bound holds even when the demand intercept, price sensitivity, and noise distribution are known and full demand is observed. Thus, sales censoring and learning the common noise law need not increase the minimax regret order within this model class for d ≥ 2. Numerical experiments illustrate the horizon and dimension dependence and show lower mean regret than four benchmark implementations at every tested horizon on a nonconcave instance.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (10) , 16412-16425
MASH: Mask Shuffling-Based Federated Learning for Both Cross-Silo and Cross-Device Environments
Federated Learning (FL) has been extensively studied over the past few years. Cross-silo FL typically involves tens of reliable organizations with larger-scale, high-quality data sources, while cross-device FL may include thousands of devices, each with limited computation and communication capabilities. However, existing research has yet to simultaneously address cross-silo and cross-device scenarios. We propose a novel mask-shuffling scheme that protects clients' gradient privacy and ensures all masks added to gradients are ultimately canceled out. While gradients are treated as exponentiation values, we eliminate the need to solve the discrete logarithm problem by leveraging a precomputed lookup table. Additionally, we partition all clients into small groups, each performing the mask-shuffling scheme to accelerate mask generation. This partitioning strategy enhances the framework's resilience to client dropouts. To resist forgery attacks on aggregated gradients, each client publishes a commitment to enable verification of its contribution's correctness. We further propose a batch verification method for d-dimensional gradients, reducing the complexity from O(d) to O(1). Besides, the blockchain smart contract serves as a verifier, ensuring public trustworthiness. We formally analyze all the required security properties and conduct a comprehensive comparison with related works. Furthermore, we perform systematic experiments to evaluate overall performance, considering both cross-silo and cross-device environments. The results demonstrate reasonable efficiency, even for models with 105-dimensional gradients across 104 devices and 1010-dimensional gradients across 10 organizations.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (10) , 16412-16425
MASH: Mask Shuffling-Based Federated Learning for Both Cross-Silo and Cross-Device Environments
Federated Learning (FL) has been extensively studied over the past few years. Cross-silo FL typically involves tens of reliable organizations with larger-scale, high-quality data sources, while cross-device FL may include thousands of devices, each with limited computation and communication capabilities. However, existing research has yet to simultaneously address cross-silo and cross-device scenarios. We propose a novel mask-shuffling scheme that protects clients' gradient privacy and ensures all masks added to gradients are ultimately canceled out. While gradients are treated as exponentiation values, we eliminate the need to solve the discrete logarithm problem by leveraging a precomputed lookup table. Additionally, we partition all clients into small groups, each performing the mask-shuffling scheme to accelerate mask generation. This partitioning strategy enhances the framework's resilience to client dropouts. To resist forgery attacks on aggregated gradients, each client publishes a commitment to enable verification of its contribution's correctness. We further propose a batch verification method for d-dimensional gradients, reducing the complexity from O(d) to O(1). Besides, the blockchain smart contract serves as a verifier, ensuring public trustworthiness. We formally analyze all the required security properties and conduct a comprehensive comparison with related works. Furthermore, we perform systematic experiments to evaluate overall performance, considering both cross-silo and cross-device environments. The results demonstrate reasonable efficiency, even for models with 105-dimensional gradients across 104 devices and 1010-dimensional gradients across 10 organizations.
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (5) , 11658-11670
Verifiable Encrypted Timed Signature Scheme and Its Application in Blockchains
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 23 (5) , 11658-11670
Verifiable Encrypted Timed Signature Scheme and Its Application in Blockchains
- Open Access
- 21 August 2026
- Published by Elsevier in SSRN Electronic Journal
MEELF: Morphology Evidence and Expert Learning Framework for Sperm Classification
Background and Objectives: Sperm morphology assessment is a routine component of semen analysis, yet manual evaluationremains labor-intensive and prone to variability. Existing learning-based methods can improve classification, but they often lackexplicit and reusable morphology representations and do not fully exploit complementary local abnormalities. This study aimsto develop a morphology-aware framework that improves sperm classification while preserving interpretability, efficiency, andtransferability.Methods: We propose the Morphology Evidence and Expert Learning Framework (MEELF) for sperm microscopy analysis.The framework combines a frozen self-supervised visual backbone, a deterministic morphology evidence pipeline, head- andtail-focused local experts, and a split-conditional fusion module. The method was evaluated on a three-class sperm microscopybenchmark using stratified five-fold cross-validation, significance testing against a reproduced baseline, external validation on ahead-focused reference dataset, and cross-backbone transfer experiments.Results: MEELF achieved a mean accuracy of 94.47% and reached 100% accuracy on the Non-Sperm category. Compared withthe reproduced baseline, it produced statistically significant improvements in fold-wise accuracy, and paired sample-level testingfurther supported this advantage. Ablation studies, qualitative analyses, and cross-backbone experiments consistently showed thatexplicit morphology evidence and complementary local experts contributed to the observed gains.Conclusions: The proposed framework improves sperm classification while maintaining interpretability, parameter efficiency, andportability across encoders. It also provides a reusable morphology-aware interface for semen image analysis. The project sourcecode is publicly available at https://github.com/wgwgr/MEELF.git.
- 21 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMEELF: Morphology Evidence and Expert Learning Framework for Sperm Classification
Background and Objectives: Sperm morphology assessment is a routine component of semen analysis, yet manual evaluationremains labor-intensive and prone to variability. Existing learning-based methods can improve classification, but they often lackexplicit and reusable morphology representations and do not fully exploit complementary local abnormalities. This study aimsto develop a morphology-aware framework that improves sperm classification while preserving interpretability, efficiency, andtransferability.Methods: We propose the Morphology Evidence and Expert Learning Framework (MEELF) for sperm microscopy analysis.The framework combines a frozen self-supervised visual backbone, a deterministic morphology evidence pipeline, head- andtail-focused local experts, and a split-conditional fusion module. The method was evaluated on a three-class sperm microscopybenchmark using stratified five-fold cross-validation, significance testing against a reproduced baseline, external validation on ahead-focused reference dataset, and cross-backbone transfer experiments.Results: MEELF achieved a mean accuracy of 94.47% and reached 100% accuracy on the Non-Sperm category. Compared withthe reproduced baseline, it produced statistically significant improvements in fold-wise accuracy, and paired sample-level testingfurther supported this advantage. Ablation studies, qualitative analyses, and cross-backbone experiments consistently showed thatexplicit morphology evidence and complementary local experts contributed to the observed gains.Conclusions: The proposed framework improves sperm classification while maintaining interpretability, parameter efficiency, andportability across encoders. It also provides a reusable morphology-aware interface for semen image analysis. The project sourcecode is publicly available at https://github.com/wgwgr/MEELF.git.
- 2 June 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
Estimation Errors as Regret Lower Bounds for Linear Contextual Bandits
Linear contextual bandits and their variants represent a fundamental class of models with wide real-world applications, usually solved using algorithms guided by parameter estimation. The Cauchy-Schwarz inequality established analytically that estimation errors dominate algorithm regrets. Therefore, accurate parameter estimation suffices to guarantee algorithms with low regrets. In this paper, we establish the necessity of accurate estimations in effective algorithms for linear contextual bandit problems by first constructing an estimator for any given algorithm. We then show that algorithm regrets dominate the estimation errors of their induced estimators under mild conditions. In other words, low-regret algorithms must imply accurate estimators, and developing low-regret algorithms is equivalent to finding efficient estimators, either implicitly or explicitly. Thus, our analysis reduces regret lower bounds to estimation errors, bridging lower bound analysis in bandit problems and regression analysis. This provides a framework for finding practical and informative regret lower bounds by leveraging the extensive estimation literature in Statistics. It leads to insightful lower bounds for a variety of contextual bandit problems in the literature, which are either new or tighter than existing ones. This paper was accepted by J. George Shanthikumar, data science. Funding: Financial support from the Hong Kong Research Grants Council [Grants 16200821, 16500023, 16500225, and T32-615/24-R] is gratefully acknowledged. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2023.02827 .
- 2 June 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
Estimation Errors as Regret Lower Bounds for Linear Contextual Bandits
Linear contextual bandits and their variants represent a fundamental class of models with wide real-world applications, usually solved using algorithms guided by parameter estimation. The Cauchy-Schwarz inequality established analytically that estimation errors dominate algorithm regrets. Therefore, accurate parameter estimation suffices to guarantee algorithms with low regrets. In this paper, we establish the necessity of accurate estimations in effective algorithms for linear contextual bandit problems by first constructing an estimator for any given algorithm. We then show that algorithm regrets dominate the estimation errors of their induced estimators under mild conditions. In other words, low-regret algorithms must imply accurate estimators, and developing low-regret algorithms is equivalent to finding efficient estimators, either implicitly or explicitly. Thus, our analysis reduces regret lower bounds to estimation errors, bridging lower bound analysis in bandit problems and regression analysis. This provides a framework for finding practical and informative regret lower bounds by leveraging the extensive estimation literature in Statistics. It leads to insightful lower bounds for a variety of contextual bandit problems in the literature, which are either new or tighter than existing ones. This paper was accepted by J. George Shanthikumar, data science. Funding: Financial support from the Hong Kong Research Grants Council [Grants 16200821, 16500023, 16500225, and T32-615/24-R] is gratefully acknowledged. Supplemental Material: The online appendix is available at https://doi.org/10.1287/mnsc.2023.02827 .
- Open Access
- 30 April 2026
- Published by Elsevier in SSRN Electronic Journal
Staffing under Taylor's Law: A Unifying Framework for Bridging Square-root and Linear Safety Rules
Staffing rules are an essential management tool in service industries for meeting target service levels. The square-root safety rule, based on the Poisson arrival assumption, has been commonly used. However, empirical findings suggest that arrivals often exhibit "over-dispersion", meaning that the variance exceeds the mean. In this paper, we develop a new doubly stochastic Poisson process model that captures two key features of over-dispersed arrivals: (i) Taylor's law, which links the variance to the mean through a power-law relationship, and (ii) temporal correlation decay, where the correlation between arrival counts in disjoint time intervals decreases as the time gap grows. Using this model, we study how over-dispersion affects staffing and derive a closed-form staffing formula to ensure a desired service level. Our formula shows that the safety level grows as a power of the nominal load. The exponent lies between 1/2 (the square-root safety rule) and 1 (the linear safety rule). It depends on the degree of over-dispersion, and it implies that Taylor's law is the dominant factor in determining staffing levels in heavy traffic. Extensive numerical experiments with both simulated and real arrival data show that our model and staffing rules significantly outperform various alternatives.
- 30 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessStaffing under Taylor's Law: A Unifying Framework for Bridging Square-root and Linear Safety Rules
Staffing rules are an essential management tool in service industries for meeting target service levels. The square-root safety rule, based on the Poisson arrival assumption, has been commonly used. However, empirical findings suggest that arrivals often exhibit "over-dispersion", meaning that the variance exceeds the mean. In this paper, we develop a new doubly stochastic Poisson process model that captures two key features of over-dispersed arrivals: (i) Taylor's law, which links the variance to the mean through a power-law relationship, and (ii) temporal correlation decay, where the correlation between arrival counts in disjoint time intervals decreases as the time gap grows. Using this model, we study how over-dispersion affects staffing and derive a closed-form staffing formula to ensure a desired service level. Our formula shows that the safety level grows as a power of the nominal load. The exponent lies between 1/2 (the square-root safety rule) and 1 (the linear safety rule). It depends on the degree of over-dispersion, and it implies that Taylor's law is the dominant factor in determining staffing levels in heavy traffic. Extensive numerical experiments with both simulated and real arrival data show that our model and staffing rules significantly outperform various alternatives.
- Open Access
- 10 March 2026
- Published by Elsevier in SSRN Electronic Journal
Dynamic Pricing under Revenue-Based Financing: Global Stability versus Local Inefficiency
Revenue-Based Financing (RBF) is an emerging funding mechanism in which firms repay investors through a fixed fraction of future revenues until cumulative repayments reach a predetermined cap. Despite rapid market growth, how RBF's revenue-contingent repayment structure affects firms' operational decisions remains an open question. We study this question by developing a dynamic programming model in which a retailer sets prices over a finite horizon under demand uncertainty, managing the trade-off between revenue maximization and debt repayment. Our analysis reveals that RBF exhibits a dual nature: local inefficiency, where pricing distortions concentrate at intermediate debt levels in a non-monotonic pattern, and global stability, where worst-case efficiency losses vanish as the planning horizon grows. This contrasts sharply with traditional zero-liability debt, where fixed repayment obligations can generate persistent distortions even over long horizons. Among contracts delivering equivalent expected repayments to the financier, we show that lower revenue shares paired with higher caps generally preserve operational efficiency more effectively than higher shares with lower caps. Incorporating inventory constraints, we find that scarcity concentrates distortions into a narrower debt region but intensifies their magnitude within that region. Time discounting introduces a double-edged effect: it can improve efficiency by devaluing distant distortions, but can also intensify current-period pricing pressure. These findings provide actionable guidance for structuring RBF contracts to balance repayment security with operational efficiency.
- 10 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDynamic Pricing under Revenue-Based Financing: Global Stability versus Local Inefficiency
Revenue-Based Financing (RBF) is an emerging funding mechanism in which firms repay investors through a fixed fraction of future revenues until cumulative repayments reach a predetermined cap. Despite rapid market growth, how RBF's revenue-contingent repayment structure affects firms' operational decisions remains an open question. We study this question by developing a dynamic programming model in which a retailer sets prices over a finite horizon under demand uncertainty, managing the trade-off between revenue maximization and debt repayment. Our analysis reveals that RBF exhibits a dual nature: local inefficiency, where pricing distortions concentrate at intermediate debt levels in a non-monotonic pattern, and global stability, where worst-case efficiency losses vanish as the planning horizon grows. This contrasts sharply with traditional zero-liability debt, where fixed repayment obligations can generate persistent distortions even over long horizons. Among contracts delivering equivalent expected repayments to the financier, we show that lower revenue shares paired with higher caps generally preserve operational efficiency more effectively than higher shares with lower caps. Incorporating inventory constraints, we find that scarcity concentrates distortions into a narrower debt region but intensifies their magnitude within that region. Time discounting introduces a double-edged effect: it can improve efficiency by devaluing distant distortions, but can also intensify current-period pricing pressure. These findings provide actionable guidance for structuring RBF contracts to balance repayment security with operational efficiency.
- 1 March 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Operations Research
- Vol. 74 (2) , 879-897
Minimax Optimality in Contextual Dynamic Pricing with General Valuation Models
We study contextual dynamic pricing, where a decision maker posts personalized prices based on observable contexts and receives binary purchase feedback indicating whether the customer’s valuation exceeds the price. Each valuation is modeled as an unknown latent function of the context, corrupted by independent and identically distributed market noise from an unknown distribution. Relying only on Lipschitz continuity of the noise distribution and bounded valuations, we propose a minimax-optimal algorithm. To accommodate the unknown distribution, our method discretizes the relevant noise range to form a finite set of candidate prices, then applies layered data partitioning to obtain confidence bounds substantially tighter than those derived via the elliptical potential lemma. A key advantage is that estimation bias in the valuation function cancels when comparing upper confidence bounds, eliminating the need to know the Lipschitz constant. The framework extends beyond linear models to general function classes through offline regression oracles. Our regret analysis depends solely on the oracle’s estimation error, typically governed by the statistical complexity of the class. These techniques yield a regret upper bound matching the minimax lower bound up to logarithmic factors. Furthermore, we refine these guarantees under additional structures—for example, linear valuation models, second-order smoothness, sparsity, and known noise distribution or observable valuations—and compare our bounds and assumptions with prior dynamic-pricing methods. Finally, numerical experiments corroborate the theory and show clear improvements over benchmark methods. Funding: X. Gong’s research is generously supported by the National Natural Science Foundation of China [Grant 72501238]. W. You’s research is generously supported by the Hong Kong Research Grants Council [Grant GRF 16212823]. J. Zhang’s research is generously supported by the Hong Kong Research Grants Council [Theme-based Research Project T32-615/24-R]. Supplemental Material: All supplemental materials, including the code, data, and files required to reproduce the results are available at https://doi.org/10.1287/opre.2025.1779 .
- 1 March 2026
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Operations Research
- Vol. 74 (2) , 879-897
Minimax Optimality in Contextual Dynamic Pricing with General Valuation Models
We study contextual dynamic pricing, where a decision maker posts personalized prices based on observable contexts and receives binary purchase feedback indicating whether the customer’s valuation exceeds the price. Each valuation is modeled as an unknown latent function of the context, corrupted by independent and identically distributed market noise from an unknown distribution. Relying only on Lipschitz continuity of the noise distribution and bounded valuations, we propose a minimax-optimal algorithm. To accommodate the unknown distribution, our method discretizes the relevant noise range to form a finite set of candidate prices, then applies layered data partitioning to obtain confidence bounds substantially tighter than those derived via the elliptical potential lemma. A key advantage is that estimation bias in the valuation function cancels when comparing upper confidence bounds, eliminating the need to know the Lipschitz constant. The framework extends beyond linear models to general function classes through offline regression oracles. Our regret analysis depends solely on the oracle’s estimation error, typically governed by the statistical complexity of the class. These techniques yield a regret upper bound matching the minimax lower bound up to logarithmic factors. Furthermore, we refine these guarantees under additional structures—for example, linear valuation models, second-order smoothness, sparsity, and known noise distribution or observable valuations—and compare our bounds and assumptions with prior dynamic-pricing methods. Finally, numerical experiments corroborate the theory and show clear improvements over benchmark methods. Funding: X. Gong’s research is generously supported by the National Natural Science Foundation of China [Grant 72501238]. W. You’s research is generously supported by the Hong Kong Research Grants Council [Grant GRF 16212823]. J. Zhang’s research is generously supported by the Hong Kong Research Grants Council [Theme-based Research Project T32-615/24-R]. Supplemental Material: All supplemental materials, including the code, data, and files required to reproduce the results are available at https://doi.org/10.1287/opre.2025.1779 .