2025​‌Activity reportProject-TeamNEO​​

7.1.1 Marmote

• Name:​​
  MARkovian MOdeling: The Environment​​​‌
• Keyword:
  Markov model
• Functional​ Description:

  Marmote is a​‌ library for modeling with​​ Markov chains. It is​​​‌ written in C++ with​ a Python interface. It​‌ consists in a reduced​​ set of high-level abstractions​​​‌ for constructing state spaces,​ transition structures and Markov​‌ chains (discrete-time and continuous-time).​​ It provides the ability​​​‌ of constructing hierarchies of​ Markov models, from the​‌ most general to the​​ particular, and equip each​​​‌ level with specifically optimized​ solution methods. The current​‌ release features the library​​ MarmoteMDP for modeling Markov​​​‌ Decision Processes and solving​ them.

  This software was​‌ started within the ANR​​ MARMOTE project: ANR-12-MONU-00019 under​​​‌ the name marmoteCore. Within​ the Marmote project, the​‌ code conforms the latest​​ C++ standards and the​​​‌ library is available on​ multiple platforms via a​‌ conda distribution.

• Release Contributions:​​

  Version 1.3.1 was released​​​‌ at the end of​ 2025. Together with previous​‌ version 1.3.0, it brought​​ many improvements to the​​​‌ 'user experience'. To quote​ a few: generalized use​‌ of the logging facility,​​ generalized use of the​​​‌ 'policy' feature, including in​ numerical algorithms, improved formatting​‌ of large objects (sets,​​ matrices, ...). Technical improvements​​​‌ concern primarily the computation​ of hitting times in​‌ Markov chains, and the​​ related phase-type distributions. The​​​‌ RLGL algorithm for computing​ stationary distributions, which is​‌ an exclusivity of Marmote,​​ was enriched with an​​​‌ "adaptive relaxation" strategy.

  A​ complete set of examples​‌ is available on the​​ project's website, both for​​​‌ the C++ and the​ Python languages, and both​‌ for Marmote and MarmoteMDP,​​ the library for manipulating​​​‌ Markov Decision Processes. For​ Python, a notebook-based tutorial​‌ demonstrates the principal functionalities​​ of the library.

• News​​ of the Year:
  The​​​‌ software was awarded an‌ Inria ADT (Action of‌​‌ Technological Development). Thanks to​​ the help of software​​​‌ engineers of Inria’s SED‌ (Service Experimentation Development), versions‌​‌ 1.3.0 and 1.3.1 were​​ developed during the year.​​​‌ See above for the‌ improvements brought by these‌​‌ versions.
• URL:
  https://­marmote.­gitlabpages.­inria.­fr/­marmote/
• Publications:​​
  hal-03770430v1, hal-03781620v1,​​​‌ hal-02395100v1, hal-04176076v1
• Contact:‌
  Alain Jean-Marie
• Participants:
  Alain‌​‌ Jean-Marie, Patrick Brown, Emmanuel​​ Hyon
• Partner:
  Université Paris​​​‌ Nanterre

8.1.1 Markov Processes​​

Markov processes are a​​​‌ cornerstone of Neo’s‌ stochastic modeling work, providing‌​‌ a unifying framework for​​ analyzing complex random dynamics​​​‌ in networked systems. Over‌ the last year, we‌​‌ pursued complementary contributions spanning​​ (i) fundamental properties of​​​‌ continuous-time Markov processes, (ii)‌ decision-making under partial observability‌​‌ in Markovian environments, and​​ (iii) software support for​​​‌ Markov modeling and numerical‌ analysis through continued development‌​‌ of the Marmote library.​​

In 14, Konstantin​​​‌ Avrachenkov together with Flora‌ Spieksma (Leiden University, Netherlands)‌​‌ studies the deviation matrix​​ of denumerable state space,​​​‌ multi-chain Markov processes in‌ continuous time. The deviation‌​‌ matrix is a measure​​ of the cumulative deviation​​​‌ from the limiting probabilities‌ and it plays an‌​‌ important role in many​​ application domains. First, the​​​‌ authors provide several equivalent‌ necessary and sufficient conditions‌​‌ for the existence of​​ the deviation matrix. Next,​​​‌ they study a relation‌ between the deviation matrix‌​‌ and rank-one perturbations of​​ Markov processes. Based on​​​‌ a rank-one perturbation, they‌ derive a versatile formula‌​‌ for the deviation matrix​​ and apply this formula​​​‌ to Markov processes with‌ restart. Along the way,‌​‌ they establish several new​​ properties of rank-one perturbed​​​‌ Markov processes. They feel‌ that those properties can‌​‌ be useful in their​​ own right.

In 28​​​‌, Konstantin Avrachenkov ,‌ together with Madhu Dhiman‌​‌ and Kavitha Veerarun (IIT​​ Bombay, India), studies Markov​​​‌ Decision Processes (MDPs) with‌ intermittent state information, with‌​‌ periods of missing observations.​​ Linear programming (LP) methods​​​‌ can play a crucial‌ role in solving MDPs,‌​‌ in particular, with constraints.​​ However, the resultant belief​​​‌ MDPs lead to infinite‌ dimensional LPs, even when‌​‌ the original MDP is​​ with finite state and​​​‌ action spaces. The verification‌ of strong duality becomes‌​‌ non-trivial. This work investigates​​ the conditions for no​​​‌ duality gap in average-reward‌ finite Markov decision process‌​‌ with intermittent state observations.​​ The authors first establish​​​‌ that in such MDPs,‌ the belief MDP is‌​‌ unichain if the original​​ Markov chain is recurrent.​​​‌ Furthermore, they establish strong‌ duality of the problem,‌​‌ under the same assumption.​​​‌ Finally, the authors provide​ a wireless channel example,​‌ where the belief state​​ depends on the last​​​‌ channel state received and​ the age of the​‌ channel state. The numerical​​ results indicate interesting properties​​​‌ of the solution.

The​ development of the Marmote​‌ library has continued, led​​ by Alain Jean-Marie and​​​‌ Emmanuel Hyon of Université​ Paris-Nanterre. The software was​‌ awarded an Inria ADT​​ (Action of Technological Development),​​​‌ during which several software​ engineers of Inria's SED​‌ (Service Experimentation Development) helped​​ improve the robustness and​​​‌ portability of the library.​ This year's improvements have​‌ been primarily targeted at​​ a better `user experience'.​​​‌ Through an adequate configuration,​ Marmote programmers have access​‌ to a fine control​​ on the behavior of​​​‌ the program: its algorithms​ and its output. On​‌ the side of Markov​​ modeling and numerical analysis,​​​‌ the principal improvement is​ about the computation of​‌ hitting times and the​​ related phase-type distributions. Also,​​​‌ the RLGL algorithm, co-developed​ by Konstantin Avrachenkov and​‌ Patrick Brown , was​​ enriched with a "adaptive​​​‌ relaxation" strategy for vector​ updates. Version 1.3.1 of​‌ Marmote has been released​​ at the end of​​​‌ 2025.

8.1.2 Queueing Systems​

Queueing systems with multiple​‌ queues can be controlled​​ in different ways depending​​​‌ on who makes the​ decision. In one setting,​‌ each arriving customer chooses​​ which queue to join,​​​‌ leading to load-balancing policies​ that aim to reduce​‌ congestion and delay. In​​ another setting, a single​​​‌ server chooses which queue​ to serve (and when​‌ to switch), yielding polling-type​​ control problems where stability​​​‌ and delay trade-offs depend​ on the switching rule​‌ and its overhead. Over​​ the last year, we​​​‌ contributed to both perspectives.​

Join-the-shortest-queue and its variants​‌ have often been used​​ in solving load balancing​​​‌ problems. The aim of​ such policies is to​‌ minimize the average system​​ occupation, e.g., the customer's​​​‌ system time. In 36​, Eitan Altman ,​‌ together with Andrea Fox​​ and Francesco De Pellegrini​​​‌ (Avignon Université, France), Arnob​ Ghosh (New Jersey Institute​‌ of Technology, USA), and​​ Ness Shroff (The Ohio​​​‌ State University, USA), extends​ the load balancing setting​‌ to include constraints that​​ may be imposed, e.g.,​​​‌ due to the communication​ network. First, the authors​‌ cast the problem in​​ the framework of constrained​​​‌ MDPs: this permits them​ to address both action-dependent​‌ constraints, such as bandwidth​​ limitation, and state-dependent constraints,​​​‌ such as minimum queue​ utilization. Hence, unlike the​‌ state-of-the-art approaches in load​​ balancing, they derive new​​​‌ policies that satisfy the​ constraints while minimizing system​‌ occupancy. Extensive numerical simulations​​ evaluate the policies' performance​​​‌ under various system settings.​

In 27, Konstantin​‌ Avrachenkov , together with​​ Kousik Das, Kavitha Veeraruna​​​‌ (IIT Bombay, India) and​ Vartika Singh (University of​‌ Colorado, US), considers a​​ polling system with two​​​‌ queues, where a single​ server is attending the​‌ queues in a cyclic​​ order and requires non-zero​​​‌ switching times to switch​ between the queues. The​‌ aim is to identify​​ a fairly general and​​​‌ comprehensive class of Markovian​ switching policies that renders​‌ the system stable, potentially​​ a class of policies​​ that can cover the​​​‌ Pareto frontier related to‌ individual-queue-centric performance measures like‌​‌ the stationary expected number​​ of waiting customers in​​​‌ each queue. For instance,‌ such a class of‌​‌ policies has been identified​​ recently for a polling​​​‌ system near the fluid‌ regime (with large arrival‌​‌ and departure rates), and​​ the authors aim to​​​‌ include that class. They‌ also aim to include‌​‌ a second class that​​ facilitates switching between the​​​‌ queues at the instance‌ the occupancy in the‌​‌ opposite queue crosses a​​ threshold and when that​​​‌ in the visiting queue‌ is below a threshold‌​‌ (this inclusion facilitates design​​ of `robust' polling systems).​​​‌ Towards this, the authors‌ consider a class of‌​‌ two-phase switching policies, which​​ includes the above mentioned​​​‌ classes. In the maximum‌ generality, the policies can‌​‌ be represented by eight​​ parameters, while two parameters​​​‌ are sufficient to represent‌ the aforementioned classes. The‌​‌ authors provide simple conditions​​ to identify the sub-class​​​‌ of switching policies that‌ ensure system stability. By‌​‌ numerically tuning the parameters​​ of the proposed class,​​​‌ they illustrate that the‌ proposed class can cover‌​‌ the Pareto frontier for​​ the stationary expected number​​​‌ of customers in the‌ two queues.

8.1.3 Approximate‌​‌ Counting

Count-Min Sketch with​​ Conservative Updates (CMS-CU) is​​​‌ a memory-efficient hash-based algorithm‌ used to estimate the‌​‌ occurrences of items within​​ a data stream. CMS-CU​​​‌ employs $d$ hash functions‌ and a total of‌​‌ $m$ counters arranged in​​ $d$ rows, with each​​​‌ hash function mapping an‌ item to one counter‌​‌ per row. In 29​​, Younes Ben Mazziane​​​‌ (Avignon Université, France) and‌ Othmane Marfoq (Meta, USA)‌​‌1 study a similar​​ algorithm, which they refer​​​‌ to as CU-S, where‌ $d$ hash functions map‌​‌ each item to $\mathrm{d}$ distinct counters in a​​​‌ single array of size‌ $m$. Specifically, they‌​‌ present an analytical method​​ to quantify the trade-off​​​‌ between memory usage and‌ the Counting Error (CE)‌​‌ of an item in​​ CU-S, defined as the​​​‌ discrepancy between the estimated‌ and actual number of‌​‌ its occurrences. The first​​ result of this paper​​​‌ shows that items absent‌ from the stream experience‌​‌ the highest CE. They​​ refer to their error​​​‌ as the Unseen items‌ Error (UE). The second‌​‌ result is an upper​​ bound on UE of​​​‌ any stream, expressed in‌ terms of the Reference‌​‌ Error (RE), which is​​ the UE of a​​​‌ reference stream where each‌ item appears at most‌​‌ once. They also identify​​ streams for which this​​​‌ bound is provably tight.‌ The direct computation of‌​‌ RE involves dealing with​​ an infinite state space​​​‌ Markov process similar to‌ the Balls and Bins‌​‌ model with the Power​​ of Choice, but more​​​‌ difficult to handle analytically.‌ Instead, they construct two‌​‌ finite state space Markov​​ processes, parametrized by a​​​‌ positive integer $g$,‌ that bound the original‌​‌ chain and yield efficiently​​ computable lower and upper​​​‌ bounds on RE. Increasing‌ $g$ narrows the gap‌​‌ between these bounds and​​ enlarges the state space​​​‌ to $\left(\genfrac{}{}{0pt}{}{m+\mathrm{g‌}-d}{g}\right)$,‌​‌ thus increasing the computation​​​‌ time. For $d=m-1$,​‌ $g=1$,​​ and as $m\to ‌\infty$, they prove​ that the lower and​‌ upper bounds coincide. Their​​ bounds are accurate for​​​‌ small values of $\mathrm{g}$ and for more general​‌ values of $m$ and​​ $d$, as shown​​​‌ by numerical computation. Finally,​ simulations on various streams​‌ confirm that bounding RE​​ in this manner enables​​​‌ efficient and accurate computation​ of UE.

8.1.4 Characterization​‌ of Integrated Autoregressive Time​​ Series

Integrated autoregressive time​​​‌ series are widely used​ to model non-stationary dynamics​‌ in economics, finance, and​​ other applications, where observed​​​‌ variables exhibit persistent trends​ and become approximately stationary​‌ only after differencing. In​​ 21, Konstantin Avrachenkov​​​‌ , together with with​ Phil Howlett (University of​‌ South Australia, Australia), Brendan​​ Beare (University of Sydney,​​​‌ Australia), Massimo Franchi (Sapienza​ Universita di Roma, Italy),​‌ and John Boland (University​​ of South Australia, Australia),​​​‌ proves an extended Granger-Johansen​ Representation Theorem (GJRT) for​‌ finite—or infinite—order integrated autoregressive​​ time series on Banach​​​‌ space. The authors assume​ only that the resolvent​‌ of the autoregressive polynomial​​ for the series is​​​‌ analytic on and inside​ the unit circle except​‌ for an isolated singularity​​ at unity. If the​​​‌ singularity is a pole​ of finite order, the​‌ time series is integrated​​ of the same order.​​​‌ If the singularity is​ an essential singularity, the​‌ time series is integrated​​ of order infinity. When​​​‌ there is no deterministic​ forcing, the value of​‌ the series at each​​ time is the sum​​​‌ of an almost surely​ convergent stochastic trend, a​‌ deterministic term depending on​​ the initial conditions and​​​‌ a finite sum of​ embedded white noise terms​‌ in the prior observations.​​ This is the extended​​​‌ GJRT. In each case​ the original series is​‌ the sum of two​​ separate autoregressive time series​​​‌ on complementary subspaces—a singular​ component which is integrated​‌ of the same order​​ as the original series​​​‌ and a regular component​ which is not integrated.​‌ The extended GJRT applies​​ to all integrated autoregressive​​​‌ processes irrespective of the​ spatial dimension, the number​‌ of stochastic trends and​​ cointegrating relations in the​​​‌ system and the order​ of integration.

8.1.5 Coding​‌ Schemes for Distributed Storage​​

Distributed storage over heterogeneous​​​‌ networks offers a cost-effective​ alternative to cloud storage.​‌ Most existing systems rely​​ on erasure coding to​​​‌ ensure data durability despite​ significant peer unreliability. However,​‌ these solutions face the​​ fundamental trade-off between data​​​‌ safety and storage overhead,​ both of which are​‌ of great importance to​​ users. In addition, resource​​​‌ consumption, particularly CPU and​ bandwidth usage, remains a​‌ major concern. In an​​ ongoing research, Sara Alouf​​​‌ and Kyrylo Tymchenko ,​ together with Frederic Giroire​‌ and Stephane Perennes (​​Coati Inria team), investigate​​​‌ novel coding schemes which​ combine reliability and fast​‌ coding operations. They use​​ Markovian modeling to evaluate​​​‌ the performance of the​ coding schemes designed and​‌ compare them with classical​​ erasure coding and locally​​​‌ repairable codes. This work​ has been submitted for​‌ publication and is under​​ review.

8.2.1 Information-theoretic Foundations​​​‌ for Statistical Learning

Over‌ the last year, we‌​‌ developed measure-theoretic and information-theoretic​​ tools that clarify how​​​‌ learning objectives and regularization‌ relate to probability measures.‌​‌

In 52, 47​​, Yaiza Bermudez and​​​‌ Samir Medina Perlaza ,‌ together with Gaetan Bisson‌​‌ (Université de la Polynésie​​ française, France) and Iñaki​​​‌ Esnaola (University of Sheffield,‌ United Kingdom), present rigorous‌​‌ statements and formal proofs​​ for both foundational and​​​‌ advanced folklore theorems on‌ the Radon-Nikodym derivative. The‌​‌ cases of conditional and​​ marginal probability measures are​​​‌ carefully considered, which leads‌ to an identity involving‌​‌ the sum of mutual​​ and lautum information suggesting​​​‌ a new interpretation for‌ such a sum.

In‌​‌ 23, 56,​​ Samir Medina Perlaza ,​​​‌ together with Gaetan Bisson‌ (Université de la Polynésie‌​‌ française, France), presents closed-form​​ expressions for the variation​​​‌ of the expectation of‌ a given function due‌​‌ to changes in the​​ probability measure (probability distribution​​​‌ drifts). These expressions unveil‌ interesting connections with Gibbs‌​‌ probability measures, information projections,​​ Pythagorean identities for relative​​​‌ entropy, mutual information, and‌ lautum information.

In 54‌​‌, 31, José​​ Francisco Daunas Torres and​​​‌ Samir Medina Perlaza ,‌ together with Iñaki Esnaola‌​‌ (University of Sheffield, United​​ Kingdom), introduce the dual​​​‌ formulation of empirical risk‌ minimization (ERM) with f-divergence‌​‌ regularization (ERM-fDR). The authors​​ connect the solution of​​​‌ the dual optimization problem‌ to the ERM-f DR‌​‌ to the notion of​​ normalization function introduced as​​​‌ an implicit function. This‌ dual approach leverages the‌​‌ Legendre-Fenchel transform and the​​ implicit function theorem to​​​‌ provide a nonlinear ODE‌ expression to the normalization‌​‌ function. Furthermore, the nonlinear​​ ODE expression and its​​​‌ properties provide a computationally‌ efficient method to calculate‌​‌ the normalization function of​​ the ERM-fDR solution under​​​‌ a mild condition.

In‌ 18, José Francisco‌​‌ Daunas Torres and Samir​​ Medina Perlaza , together​​​‌ with Iñaki Esnaola (University‌ of Sheffield, United Kingdom)‌​‌ and H. Vincent Poor​​ (Princeton University, USA), analyze​​​‌ the effect of relative‌ entropy asymmetry in the‌​‌ context of empirical risk​​ minimization with relative entropy​​​‌ regularization (ERM-RER). The authors‌ consider two regularizations: (a)‌​‌ the relative entropy of​​ the measure to be​​​‌ optimized with respect to‌ a reference measure (Type-I‌​‌ ERM-RER); and (b) the​​ relative entropy of the​​​‌ reference measure with respect‌ to the measure to‌​‌ be optimized (Type-II ERM-RER).​​ The main result is​​​‌ the characterization of the‌ solution to the Type-II‌​‌ ERM-RER problem and its​​ key properties. By comparing​​​‌ the well-understood Type-I ERM-RER‌ with Type-II ERM-RER, the‌​‌ effects of entropy asymmetry​​ are highlighted. The analysis​​​‌ shows that in both‌ cases, regularization by relative‌​‌ entropy forces the support​​ of the solution to​​​‌ collapse into the support‌ of the reference measure,‌​‌ introducing a strong inductive​​ bias that negates the​​​‌ evidence provided by the‌ training data. The authors‌​‌ show finally that Type-II​​​‌ regularization is equivalent to​ Type-I regularization with an​‌ appropriate transformation of the​​ empirical risk function.

In​​​‌ 58, Xinying Zou​ and Samir Medina Perlaza​‌ , together with Iñaki​​ Esnaola (University of Sheffield,​​​‌ United Kingdom), address the​ challenge of learning models​‌ that stay reliable under​​ distribution shifts when only​​​‌ finite training data is​ available. The authors propose​‌ a novel training-dependent minimax​​ problem to design learning​​​‌ algorithms that are robust​ to the worst-case datagenerating​‌ distribution. For the ambiguity​​ sets, they construct Kullback-Leibler​​​‌ (KL) divergence neighborhoods on​ both model and data​‌ distributions. They obtain an​​ analytical solution to this​​​‌ minimax problem, referred to​ as the robust learner.​‌ They show that the​​ robust learner follows a​​​‌ Gibbs distribution, in which​ the prior can be​‌ chosen as any baseline​​ learning algorithm to be​​​‌ robustified. Such a robust​ learner minimizes the worst-case​‌ generalization gap within a​​ KL-divergence neighborhood of unseen​​​‌ new data and it​ provides a smaller generalization​‌ error compared with the​​ baseline learning algorithm Q.​​​‌ Under certain conditions, the​ robust learner also guarantees​‌ a smaller expected testing​​ error than Q. They​​​‌ also provide a training-dependent​ PAC-Bayes bound on the​‌ robust learner's performance on​​ unseen data. Closed-form expressions​​​‌ for generalization error and​ expected loss of the​‌ robust learner are given​​ in terms of mutual​​​‌ information, lautum information, and​ KL-divergence. As a by-product,​‌ they show that the​​ proposed minimax problem admits​​​‌ a two-player zero-sum game​ formulation, for which a​‌ unique Nash equilibrium exists.​​ This enhances the understanding​​​‌ of learning algorithm robustification.​ Numerical experiments validate the​‌ applicability of the results​​ and the benefits of​​​‌ robustification.

8.2.2 Unlearning

Machine​ unlearning aims at removing​‌ the influence of specific​​ data points from a​​​‌ trained model, striving to​ achieve this at a​‌ fraction of the cost​​ of full model retraining.​​​‌ In 45, Giovanni​ Neglia , together with​‌ Martin van Waerebeke, Kevin​​ Scaman (Argo Inria​​​‌ team), and Marco Lorenzi​ (Epione Inria team),​‌ analyzes the efficiency of​​ unlearning methods. The authors​​​‌ establish the first upper​ and lower bounds on​‌ minimax computation times for​​ this problem, characterizing the​​​‌ performance of the most​ efficient algorithm against the​‌ most difficult objective function.​​ Specifically, for strongly convex​​​‌ objective functions and under​ the assumption that the​‌ forget data is inaccessible​​ to the unlearning method,​​​‌ they provide a phase​ diagram for the unlearning​‌ complexity ratio—a novel metric​​ that compares the computational​​​‌ cost of the best​ unlearning method to full​‌ model retraining. The phase​​ diagram reveals three distinct​​​‌ regimes: one where unlearning​ at a reduced cost​‌ is infeasible, another where​​ unlearning is trivial because​​​‌ adding noise suffices, and​ a third where unlearning​‌ achieves significant computational advantages​​ over retraining. These findings​​​‌ highlight the critical role​ of factors such as​‌ data dimensionality, the number​​ of samples to forget,​​​‌ and privacy constraints in​ determining the practical feasibility​‌ of unlearning.

8.2.3 Learning​​ Strategies for Email Marketing​​​‌

Email marketing is increasingly​ criticized due to ethical​‌ concerns, as bulk email​​ campaigns often result in​​ spam, reduced engagement, and​​​‌ negative user experiences. In‌ addition, there is increasing‌​‌ awareness of the environmental​​ impact, as these large-scale​​​‌ campaigns contribute to carbon‌ emissions. To address these‌​‌ issues, in 35,​​ Ibtihal El Mimouni and​​​‌ Konstantin Avrachenkov introduce QWIC-Fair‌ (Q-learning Whittle Index with‌​‌ Context and Fairness), an​​ algorithm that operates within​​​‌ a Contextual Restless Multi-Armed‌ Bandit framework. QWIC-Fair leverages‌​‌ implicit feedback to learn​​ the dynamics of user​​​‌ interactions and thus target‌ users with relevant content.‌​‌ In this model, each​​ user represents an arm​​​‌ of the bandit, evolving‌ as a Markov Decision‌​‌ Process that captures state​​ transitions reflecting their interactions​​​‌ with email contents, while‌ accounting for contextual information.‌​‌ The algorithm also incorporates​​ a fairness constraint to​​​‌ ensure balanced selection and‌ to avoid repetitive targeting‌​‌ of the same users.​​ The experiments conducted, using​​​‌ synthetic and real-world data,‌ show that QWIC-Fair outperforms‌​‌ existing email marketing approaches.​​

In 34, Ibtihal​​​‌ El Mimouni and Konstantin‌ Avrachenkov introduce DQWIC, a‌​‌ novel algorithm that combines​​ Deep Reinforcement Learning and​​​‌ Whittle index theory within‌ the Contextual Restless Multi-Armed‌​‌ Bandit framework for the​​ discounted criterion. The authors​​​‌ design DQWIC to learn‌ in evolving environments typical‌​‌ of real-world applications, such​​ as recommender systems, where​​​‌ user preferences and environmental‌ dynamics evolve over time.‌​‌ In particular, they apply​​ DQWIC to the problem​​​‌ of optimizing email recommendations,‌ where it tackles the‌​‌ dual challenges of enhancing​​ content relevance and reducing​​​‌ spam messages, thereby addressing‌ ethical concerns related to‌​‌ intrusive emailing. The algorithm​​ leverages two neural networks:​​​‌ a Q-network for approximating‌ action-value functions and a‌​‌ Whittle-network for estimating Whittle​​ indices, both of which​​​‌ integrate contextual features to‌ inform decision-making. In addition,‌​‌ the inclusion of context​​ allows them to handle​​​‌ many heterogeneous users in‌ a scalable way. The‌​‌ learning process occurs through​​ a two time scale​​​‌ stochastic approximation, with the‌ Q-network updated frequently to‌​‌ minimize the loss between​​ predicted and target Q-values,​​​‌ and the Whittle-network updated‌ on a slower time‌​‌ scale. To evaluate its​​ effectiveness, they conducted experiments​​​‌ in partnership with Smartprofile,‌ a company specializing in‌​‌ digital marketing. Their results,​​ derived from both synthetic​​​‌ and real-world data, show‌ that DQWIC outperforms existing‌​‌ email marketing baselines.

8.2.4​​ Graph Clustering

Graph clustering​​​‌ (aka community detection) is‌ one of the fundamental‌​‌ problems in data science​​ which consists of partitioning​​​‌ graph nodes into disjoint‌ communities. In 19,‌​‌ 40, Konstantin Avrachenkov​​ , together with Lucas​​​‌ Lopes Felipe and Daniel‌ Sadoc Menasché (UFRJ, Brazil),‌​‌ presents a game-theoretic perspective​​ on the Constant Potts​​​‌ Model (CPM) for partitioning‌ graphs into disjoint communities,‌​‌ emphasizing its efficiency, robustness,​​ and accuracy. Efficiency: CPM​​​‌ is reinterpreted as a‌ potential hedonic game by‌​‌ decomposing its global Hamiltonian​​ into local utility functions,​​​‌ where the local utility‌ gain of each agent‌​‌ matches the corresponding increase​​ in global utility. Leveraging​​​‌ this equivalence, the authors‌ prove that local optimization‌​‌ of the CPM objective​​ via better-response dynamics converges​​​‌ in pseudo-polynomial time to‌ an equilibrium partition. Robustness:‌​‌ The authors introduce and​​​‌ relate two stability criteria,​ namely a strict criterion​‌ based on a novel​​ notion of robustness—requiring nodes​​​‌ to simultaneously maximize neighbors​ and minimize non-neighbors within​‌ communities—and a relaxed utility​​ function based on a​​​‌ weighted sum of these​ objectives, controlled by a​‌ resolution parameter. Accuracy: In​​ community tracking scenarios, where​​​‌ initial partitions are used​ to bootstrap the Leiden​‌ algorithm with partial ground-truth​​ information, the experiments reveal​​​‌ that robust partitions yield​ higher accuracy in recovering​‌ ground-truth communities.

In 38​​, Konstantin Avrachenkov and​​​‌ Louis Hauseux , together​ with Nahuel Soprano-Loto (​‌Mathnet Inria team), study​​ Gibbs distributions with competing​​​‌ interactions and propose a​ higher-order extension of the​‌ Swendsen-Wang dynamics that incorporates​​ triangular bonds. The new​​​‌ dynamics preserves the same​ stationary distribution, alleviates frustration,​‌ and yields markedly better​​ sampling. When applied to​​​‌ a synthetic Euclidean-graph community-detection​ benchmark, the proposed algorithm​‌ outperforms existing methods.

8.2.5​​ Density-based Clustering

Many clustering​​​‌ algorithms are based on​ density estimates in ${\mathrm{ℝ‌}}^d$. Building geometric​​ graphs on a dataset​​​‌ $X\subset {ℝ}^{\mathrm{d}}$ is an elegant way​‌ to achieve this. In​​ fact, the connected components​​​‌ of a geometric graph​ match exactly with the​‌ high-density clusters of the​​ 1-Nearest Neighbor density estimator​​​‌ or Single-Linkage algorithm. In​ 20, Louis Hauseux​‌ and Konstantin Avrachenkov ,​​ together with Josiane Zerubia​​​‌ (Ayana Inria team),​ analyze and generalize the​‌ classical Single-Linkage clustering algorithm,​​ which performs hierarchical clustering​​​‌ by iteratively merging the​ two closest clusters. Single-Linkage​‌ and its robust version​​ are still widely used​​​‌ in modern clustering techniques​ like the state-of-the-art HDBSCAN.​‌ Single-Linkage can be understood​​ from three perspectives: (i)​​​‌ it conducts persistent analysis​ on geometric graphs; (ii)​‌ it identifies high-density clusters​​ using the 1-Nearest Neighbor​​​‌ density estimator; and (iii)​ it is implemented via​‌ the minimum spanning tree​​ of the data. The​​​‌ authors extend Single-Linkage to​ higher-order interactions by replacing​‌ geometric graphs with hypergraphs​​ and introducing a stricter​​​‌ notion of connected components,​ named $K$-polyhedra. Specifically,​‌ for $K=\mathrm{2}$, their method employs​​​‌ “triangle connectivity”. They prove​ that $K$-polyhedra correspond​‌ to high-density clusters of​​ the $K$-Nearest Neighbors​​​‌ density estimator. In practice,​ this approach is implemented​‌ by identifying a minimum​​ $K$-tree. The authors​​​‌ also introduce original geometric​ optimizations for efficiently computing​‌ the 2-generalization of Single-Linkage​​ in low-dimensional Euclidean spaces.​​​‌ Experimental results demonstrate that​ even when $K=‌2$ is used, the​​ proposed method already surpasses​​​‌ the state-of-the-art clustering methods​ on synthetic and real-world​‌ datasets.

8.2.6 Conjectural Learning​​

In the context of​​​‌ dynamic games in which​ players have limited information,​‌ Alain Jean-Marie , together​​ with Mabel Tidball (​​​‌INRAE, Montpellier, France)​ and Tania Jiménez (Avignon​‌ Université, France), introduces in​​ 60 the family of​​​‌ Conjectural Learning procedures. With​ Conjectural Learning, agents form​‌ conjectures about what the​​ opponent will play, as​​​‌ a function of their​ action or some state​‌ variable and may revise​​ these conjectures at each​​​‌ interaction. The authors prove​ general properties of Conjectural​‌ Learning procedures, comparing their​​ steady-states to the cooperative​​ solution of a corresponding​​​‌ static game with complete‌ information. They then specify‌​‌ simple functional forms of​​ conjectures and analyze the​​​‌ resulting dynamic systems, in‌ terms of steady states‌​‌ and their relation with​​ Pareto solutions of the​​​‌ complete information framework. They‌ illustrate the transient and‌​‌ stationary behavior of these​​ Conjectural Learning procedures in​​​‌ the Fish War model‌ of Levhari and Mirman.‌​‌

In a related forthcoming​​ publication, the same authors​​​‌ endow the otherwise selfish‌ and myopic agents with‌​‌ a certain degree of​​ forward-looking behavior, in the​​​‌ form of a subjective‌ valuation of future stocks,‌​‌ incorporated in their utility.​​ They analyze the interaction​​​‌ of three specific conjectural‌ learning procedures and this‌​‌ short-range forward-looking behavior, in​​ the case of a​​​‌ management problem of groundwater‌ resource with two symmetric‌​‌ players. The performance is​​ measured with both the​​​‌ asymptotic stock and the‌ total discounted gain. They‌​‌ conduct numerical experiments using​​ data from the La​​​‌ Mancha aquifer in Spain.‌ They conclude that learning‌​‌ processes can be Pareto​​ improving at each period.​​​‌ However, this result depends‌ on the valuation that‌​‌ agents place on future​​ stocks, the inertia of​​​‌ the learning process and‌ the initial value of‌​‌ conjectures, which must be​​ appropriately chosen. In particular,​​​‌ if the valuation of‌ the future stock is‌​‌ too small, the performance​​ of these schemes is​​​‌ worse than that of‌ the farsighted Nash solution,‌​‌ in terms of both​​ profits and resource levels.​​​‌

8.3.1 Theoretical‌ Insights on Federated Learning‌​‌ Algorithms

A sound theoretical​​ understanding is essential to​​​‌ characterize when FL works,‌ how fast it converges,‌​‌ and how well it​​ generalizes under the statistical​​​‌ dependencies and heterogeneity that‌ arise in practice. Over‌​‌ the last year, we​​​‌ contributed to this agenda​ with complementary results on​‌ optimization under dependent client​​ data streams and on​​​‌ generalization error in FL​ through an information-theoretic lens.​‌

Most theoretical and empirical​​ FL studies rely on​​​‌ the assumption that clients​ have access to pre-collected​‌ data sets, with limited​​ investigation into scenarios where​​​‌ clients continuously collect data.​ In many real-world applications,​‌ particularly when data is​​ generated by physical or​​​‌ biological processes, client data​ streams are often modeled​‌ by non-stationary Markov processes.​​ Unlike standard i.i.d. sampling,​​​‌ the performance of FL​ with Markovian data streams​‌ remains poorly understood due​​ to the statistical dependencies​​​‌ between client samples over​ time. In 39,​‌ Giovanni Neglia , together​​ with Tan-Khiem Huynh, Malcolm​​​‌ Egan, and Jean-Marie Gorce​ (Maracas Inria team),​‌ investigates whether FL can​​ still support collaborative learning​​​‌ with Markovian data streams.​ Specifically, the authors analyze​‌ the performance of Minibatch​​ SGD (Stochastic Gradient Descent),​​​‌ Local SGD, and a​ variant of Local SGD​‌ with momentum. They answer​​ affirmatively under standard assumptions​​​‌ and smooth non-convex client​ objectives: the sample complexity​‌ is proportional to the​​ inverse of the number​​​‌ of clients, with a​ communication complexity comparable to​‌ the i.i.d. scenario. However,​​ the sample complexity for​​​‌ Markovian data streams remains​ higher than for i.i.d.​‌ sampling. Their analysis is​​ validated via experiments with​​​‌ real pollution monitoring time​ series data.

In 53​‌, Yaiza Bermudez and​​ Samir Medina Perlaza ,​​​‌ together with Iñaki Esnaola​ (University of Sheffield, United​‌ Kingdom) and H. Vincent​​ Poor (Princeton University, USA),​​​‌ characterize the generalization error​ of FL systems through​‌ a novel statistical framework.​​ Central to this framework​​​‌ is the concept of​ a meta-federated learning algorithm,​‌ defined as a probability​​ measure over a client's​​​‌ local models conditioned on​ the datasets of all​‌ participating clients. By means​​ of this abstraction, the​​​‌ authors state several fundamental​ properties of FL systems​‌ and derive closed-form expressions​​ for the generalization error.​​​‌ More specifically, they extend​ to FL the method​‌ of gaps, originally introduced​​ for non-federated settings, and​​​‌ obtain closed-form expressions for​ the generalization error in​‌ terms of classical information​​ measures, including relative entropy,​​​‌ mutual information, and lautum​ information. A central role​‌ in these new expressions​​ is played by specific​​​‌ Gibbs probability measures (Gibbs​ algorithms). More importantly, they​‌ reveal that the challenge​​ of evaluating the generalization​​​‌ error in FL is​ reduced to two distinct​‌ tasks: (a) measuring the​​ dependence of client model​​​‌ choices on the datasets​ of all clients; and​‌ (b) distinguishing the meta-federated​​ learning algorithm from a​​​‌ Gibbs algorithm trained solely​ on local data. Through​‌ these findings, they establish​​ new links between generalization​​​‌ in FL, mismatched hypothesis​ testing, Shannon's information measures,​‌ and Pythagorean identities for​​ the generalization error.

8.3.2​​​‌ Scalable Algorithms

Scalability is​ a central requirement for​‌ federated and distributed learning​​ systems: algorithms must remain​​​‌ effective when the number​ of clients grows, communication​‌ is constrained, and the​​ amount of data per​​​‌ client varies widely. Over​ the last year, we​‌ investigated scalability both by​​ reducing the number of​​ communication rounds in FL​​​‌ and by enabling large‌ numbers of clients to‌​‌ collaborate with limited per-client​​ resources.

One-Shot FL enables​​​‌ multiple clients to cooperatively‌ learn a global model‌​‌ in a single round​​ of communication with a​​​‌ central server. In 44‌, Giovanni Neglia ,‌​‌ together with Jacopo Talpini​​ and Marco Savi (University​​​‌ of Milano-Bicocca, Italy), analyzes‌ the One-Shot FL problem‌​‌ through the lens of​​ Bayesian inference. The authors​​​‌ propose FedBEns, an algorithm‌ that leverages the inherent‌​‌ multimodality of local loss​​ functions to find better​​​‌ global models. Their algorithm‌ leverages a mixture of‌​‌ Laplace approximations for the​​ clients' local posteriors, which​​​‌ the server then aggregates‌ to infer the global‌​‌ model. They conduct extensive​​ experiments on various datasets,​​​‌ demonstrating that the proposed‌ method outperforms competing baselines‌​‌ that typically rely on​​ unimodal approximations of the​​​‌ local losses.

While FL‌ clients may want to‌​‌ collaborate because they do​​ not have enough data​​​‌ to learn an accurate‌ model on their own,‌​‌ collaboration also introduces a​​ bias–variance trade-off when local​​​‌ data distributions differ. A‌ key challenge is for‌​‌ each client to identify​​ clients with similar distributions​​​‌ while learning the model,‌ a problem that remains‌​‌ largely unresolved. In 37​​, Giovanni Neglia ,​​​‌ together with Franco Galante‌ and Emilio Leonardi (Politechnic‌​‌ University of Turin, Italy),​​ focuses on a particular​​​‌ instance of this challenge,‌ where each client collects‌​‌ samples from a real-valued​​ distribution over time to​​​‌ estimate its mean. Existing‌ algorithms face impractical per-client‌​‌ space and time complexities​​ (linear in the number​​​‌ of clients $\left|\mathrm{A‌}\right|$). To address‌​‌ scalability challenges, the authors​​ propose a framework where​​​‌ clients self-organize into a‌ graph, allowing each client‌​‌ to communicate with only​​ a selected number of​​​‌ peers $r$. They‌ propose two collaborative mean‌​‌ estimation algorithms: one employs​​ a consensus-based approach, while​​​‌ the other uses a‌ message-passing scheme, with complexity‌​‌ $𝒪(r)$ and $𝒪(\mathrm{r‌}log|A|‌)$, respectively. They‌​‌ establish conditions for both​​ algorithms to yield asymptotically​​​‌ optimal estimates and they‌ provide a theoretical characterization‌​‌ of their performance.

8.3.3​​ Carbon-aware Distributed Learning

Training​​​‌ large-scale machine learning models‌ incurs substantial carbon emissions.‌​‌ Since FL distributes computation​​ across geographically dispersed clients,​​​‌ it offers a natural‌ framework to exploit regional‌​‌ and temporal variations in​​ Carbon Intensity (CI) through​​​‌ carbon-aware participation and scheduling.‌

In 42, Daniel‌​‌ Richards Ravi Arputharaj and​​ Giovanni Neglia , together​​​‌ with Charlotte Rodriguez (Accenture‌ Labs, France) and Angelo‌​‌ Rodio (Linköping University, Sweden),​​ investigate how to reduce​​​‌ emissions in FL through‌ carbon-aware client selection and‌​‌ training scheduling. They first​​ quantify the emission savings​​​‌ of a carbon-aware scheduling‌ policy that leverages slack‌​‌ time—permitting a modest extension​​ of the training duration​​​‌ so that clients can‌ defer local training rounds‌​‌ to lower-carbon periods. They​​ then examine the performance​​​‌ trade-offs of such scheduling,‌ which stem from statistical‌​‌ heterogeneity among clients, selection​​ bias in participation, and​​​‌ temporal correlation in model‌ updates. To leverage these‌​‌ trade-offs, they construct a​​​‌ carbon-aware scheduler that integrates​ slack time, $\alpha$-fair​‌ carbon allocation, and a​​ global fine-tuning phase. Experiments​​​‌ on real-world CI data​ show that their scheduler​‌ outperforms slack-agnostic baselines, achieving​​ higher model accuracy across​​​‌ a wide range of​ carbon budgets, with especially​‌ strong gains under tight​​ carbon constraints.

8.3.4 Privacy​​​‌ Attacks to Federated Learning​

Although FL keeps raw​‌ data local, the training​​ phase can leak sensitive​​​‌ information through exchanged updates,​ enabling reconstruction attacks that​‌ recover either attributes of​​ the training data or​​​‌ the data themselves. Over​ the last year, we​‌ strengthened the empirical toolbox​​ for privacy evaluation in​​​‌ FL with two attacks​ that substantially broaden the​‌ scope and effectiveness of​​ reconstruction in realistic settings.​​​‌

First, while attribute inference​ attacks (AIA) have been​‌ widely studied for classification,​​ their impact on federated​​​‌ regression had remained largely​ unexplored. In 32,​‌ Giovanni Neglia , together​​ with Francesco Diana, Chuan​​​‌ Xu, and Frederic Giroire​ (Coati Inria team),​‌ Othmane Marfoq (Meta, USA)​​ and Eoin Thomas (Amadeus,​​​‌ France), addresses this gap​ by proposing novel model-based​‌ AIAs specifically designed for​​ regression tasks in FL​​​‌ environments. The authors' approach​ considers scenarios where adversaries​‌ can either eavesdrop on​​ exchanged messages or directly​​​‌ interfere with the training​ process. They benchmark their​‌ proposed attacks against state-of-the-art​​ methods using real-world datasets.​​​‌ The results demonstrate a​ significant increase in reconstruction​‌ accuracy, particularly in heterogeneous​​ client datasets, a common​​​‌ scenario in FL. The​ efficacy of their model-based​‌ AIAs makes them better​​ candidates for empirically quantifying​​​‌ privacy leakage for federated​ regression tasks.

Second, existing​‌ data reconstruction attacks often​​ rely on assumptions about​​​‌ clients' data distributions and​ their effectiveness degrades sharply​‌ when batch sizes exceed​​ a few tens of​​​‌ samples. In 33,​ Giovanni Neglia , together​‌ with Francesco Diana, André​​ Nusser, and Chuan Xu​​​‌ (Coati Inria team),​ introduces a novel data​‌ reconstruction attack that overcomes​​ these limitations. The method​​​‌ leverages a new geometric​ perspective on fully connected​‌ layers to craft malicious​​ model parameters, enabling the​​​‌ perfect recovery of arbitrarily​ large data batches in​‌ classification tasks without any​​ prior knowledge of clients'​​​‌ data. Through extensive experiments​ on both image and​‌ tabular datasets, the authors​​ demonstrate that their attack​​​‌ outperforms existing methods and​ achieves perfect reconstruction of​‌ data batches two orders​​ of magnitude larger than​​​‌ the state of the​ art.

8.4.1 Games with Irrational​‌ Players

The classical game​​ theory considers rational players​​​‌ and proposes Nash equilibrium​ (NE) as the solution.​‌ However, real-world scenarios rarely​​ feature rational players; instead,​​​‌ players make inconsistent and​ irrational decisions. Often, irrational​‌ players exhibit herding behavior​​ by simply following the​​​‌ majority. In 26,​ Khushboo Agarwal and Konstantin​‌ Avrachenkov , together with​​ Veeraruna Kavitha and Raghupati​​​‌ Vyas (IIT Bombay, India),​ consider a mean-field game​‌ with $\alpha$-fraction of​​ rational players and the​​​‌ rest being herding-irrational players.​ For such a game,​‌ the authors introduce a​​ novel concept of equilibrium​​ named $\alpha$-Rational NE​​​‌ (in short, $\alpha$-RNE).‌ They extensively analyze the‌​‌ $\alpha$-RNEs and their​​ implications in games with​​​‌ two actions. Due to‌ herding-irrational players, new equilibria‌​‌ may arise, and some​​ classical NEs may be​​​‌ deleted. They establish that‌ the rational players are‌​‌ not harmed but benefit​​ from the presence of​​​‌ irrational players. More interestingly,‌ in some examples, the‌​‌ rational players attain higher​​ utility (under $\alpha$-RNE)​​​‌ than even the social‌ optimal utility (in the‌​‌ classical setting), by leveraging​​ upon the herding behavior​​​‌ of irrational players. Surprisingly,‌ the irrational players may‌​‌ also benefit by not​​ being rational. They observe​​​‌ that irrational players do‌ not lose compared to‌​‌ some classical NEs for​​ participation and bandwidth-sharing games.​​​‌ Importantly, in bandwidth-sharing game,‌ the irrational players also‌​‌ receive utility near social​​ optimal utility. Such examples​​​‌ indicate that it may‌ sometimes be `rational' to‌​‌ be irrational.

In 13​​, Khushboo Agarwal and​​​‌ Konstantin Avrachenkov , together‌ with Veeraruna Kavitha and‌​‌ Raghupati Vyas (IIT Bombay,​​ India), consider one more​​​‌ realistic behavioral game dynamics‌ where the players choose‌​‌ actions in a turn-by-turn​​ manner and exhibit two​​​‌ prominent behavioral traits—$\mathrm{\alpha ‌}$-fraction of them are‌​‌ myopic players who strategically​​ choose optimal actions against​​​‌ the empirical distribution of‌ the previous plays, while‌​‌ the rest exhibit herding​​ behavior by choosing the​​​‌ most popular action till‌ then. The utilities are‌​‌ realized for all, at​​ the end of the​​​‌ game, and each player‌ gets to play only‌​‌ once. The analysis focuses​​ on scenarios when players​​​‌ encounter two possible choices,‌ common in applications like‌​‌ participation games (e.g., crowd-sourcing)​​ or minority games. To​​​‌ begin with, the authors‌ derive the almost sure‌​‌ mean-field limits of such​​ dynamics. The proof is​​​‌ constructive and progressively narrows‌ down the potential limit‌​‌ set and finally establishes​​ the existence of a​​​‌ unique limit for almost‌ all sample paths. The‌​‌ authors argue that the​​ dynamics at the limit​​​‌ is captured by a‌ differential inclusion (and not‌​‌ the usual ordinary differential​​ equation) due to the​​​‌ discontinuities arising from the‌ switching behavioral choices. It‌​‌ is noteworthy that the​​ presented methodology can be​​​‌ easily modified to analyze‌ the avoid-the-crowd behavior, in‌​‌ place of herding behavior.​​ The work is concluded​​​‌ with two interesting examples,‌ named participation game and‌​‌ routing game, which encapsulate​​ several real-life scenarios.

8.4.2​​​‌ Strategic Queueing with Information‌ Cost

Consider an M/M/1-type‌​‌ queue where joining attains​​ a known reward, but​​​‌ a known waiting cost‌ is paid per time‌​‌ unit spent queueing. In​​ the 1960s, Naor showed​​​‌ that any arrival optimally‌ joins the queue if‌​‌ its length is less​​ than a known threshold.​​​‌ Yet acquiring knowledge of‌ the queue length often‌​‌ brings an additional cost,​​ e.g., website loading time​​​‌ or data roaming charge.‌ Therefore, their model presents‌​‌ any arrival with three​​ options: join blindly, balk​​​‌ blindly, or pay a‌ known inspection cost to‌​‌ make the optimal joining​​ decision by comparing the​​​‌ queue length to Naor's‌ threshold. In a recent‌​‌ paper, Hassin and Roet-Green​​​‌ prove that a unique​ Nash equilibrium always exists​‌ and classify regions where​​ the equilibrium probabilities are​​​‌ non-zero. In 17,​ Konstantin Avrachenkov and Eitan​‌ Altman , together with​​ Jake Clarkson (National Highways,​​​‌ United Kingdom), complement these​ findings with new closed-form​‌ expressions for the equilibrium​​ probabilities in the majority​​​‌ of cases. Further, Hassin​ and Roet-Green show that​‌ minimizing inspection cost maximizes​​ social welfare. Envisaging a​​​‌ queue operator choosing where​ to invest, the authors​‌ of 17 compare the​​ effects of lowering inspection​​​‌ cost and increasing the​ queue-joining reward on social​‌ welfare. They prove that​​ the former dominates and​​​‌ that the latter can​ even have a detrimental​‌ effect on social welfare.​​

8.4.3 Kelly Mechanism

The​​​‌ Kelly mechanism is a​ proportional allocation auction widely​‌ adopted in decentralized resource​​ allocation systems to share​​​‌ an infinitely divisible resource​ among competing agents.

In​‌ 41, Eitan Altman​​ , together with Cleque​​​‌ Marlain Mboulou-Moutoubi, Younes Ben​ Mazziane, and Francesco De​‌ Pellegrini (Avignon Université, France),​​ analyze the sequential game​​​‌ the Kelly allocation induces​ when agents have $\mathrm{\alpha ‌}$-fair utilities and behave​​ strategically. The authors' main​​​‌ result proves that synchronous​ best-response updates drive bids​‌ to the unique Nash​​ equilibrium at a linear​​​‌ rate for $\alpha \in \{0,\mathrm{1‌},2\}$.​​ Extensive simulations reveal that​​​‌ best-response dynamics reach equilibrium​ significantly faster than previously​‌ proposed no-regret learning algorithms.​​

When agents are aware​​​‌ of the allocation mechanism,​ their interactions form a​‌ game. The properties of​​ its Nash equilibria are​​​‌ well understood under the​ simplifying assumption of unbounded​‌ budgets. In 30,​​ Eitan Altman , together​​​‌ with Younes Ben Mazziane,​ Cleque-Marlain Mboulou-Moutoubi, and Francesco​‌ De Pellegrini (Avignon Université,​​ France), analyzes the game​​​‌ in a more realistic​ budget-constrained setting, motivated by​‌ its optimality in terms​​ of the liquid price​​​‌ of anarchy. Specifically, the​ authors establish a sufficient​‌ condition for the uniqueness​​ of the Nash equilibrium​​​‌ and design a distributed​ sequential learning procedure that​‌ provably converges to the​​ equilibrium. In particular, their​​​‌ sufficient condition holds when​ the payoff functions of​‌ the agents are of​​ the proportional fair type​​​‌ in the allocated fraction.​ Finally, extensive numerical experiments​‌ shed light on the​​ interplay between the heterogeneity​​​‌ of the payoff functions​ and the agents' budgets.​‌

8.4.4 Applications to Energy​​ Markets

In 22,​​​‌ 59, Eitan Altman​ , together with Hélène​‌ Le Cadre, Mathis Guckert​​ (Inocs Inria team)​​​‌ and Mandar Datar (CEA-Leti,​ France), consider a peer-to-peer​‌ electricity market modeled as​​ a private network game,​​​‌ where end users minimize​ their cost by computing​‌ their demand and controllable​​ generation. Their nominal demand​​​‌ constitutes sensitive information that​ they might want to​‌ keep private. The authors​​ prove that the private​​​‌ network game admits a​ unique variational equilibrium, which​‌ depends on the private​​ information of all end​​​‌ users. Thus, to update​ their strategy, end users​‌ rely on randomized readings.​​ They introduce a data​​​‌ aggregator, which aims to​ learn the end users’​‌ private information, while remunerating​​ them depending on the​​ quality of their readings.​​​‌ Using performative prediction, they‌ define a decision-dependent game‌​‌ explicitly taking into account​​ the distribution shift caused​​​‌ by the endusers’ hidden‌ ability. The decision-dependent game‌​‌ coincides with a Stackelberg​​ game when the end​​​‌ users’ hidden abilities are‌ best responses. Further, the‌​‌ market robustness can be​​ quantified by evaluating the​​​‌ efficiency loss as the‌ difference between the social‌​‌ cost in the performatively​​ stable equilibrium and the​​​‌ optimum. The authors show‌ that under mild assumptions,‌​‌ the performatively stable equilibrium​​ can be found by​​​‌ distributed and sequential variants‌ of the repeated stochastic‌​‌ gradient method while they​​ propose a two-timescale stochastic​​​‌ approximation method to learn‌ Stackelberg equilibrium. Finally, they‌​‌ formulate the data aggregator’s​​ optimal contract design as​​​‌ a bilevel optimization problem‌ that they cast as‌​‌ a more tractable non-linear​​ non-convex optimization problem which​​​‌ can be solved using‌ simulated annealing. Simulations on‌​‌ small and large scale​​ problem instances illustrate the​​​‌ results.

8.5.1​​​‌ Scalable Online Learning for‌ Caching

Online learning algorithms‌​‌ address sequential decision-making problems​​ where, at each round,​​​‌ a learner selects an‌ action, observes feedback (e.g.,‌​‌ a loss or reward),​​ and updates its decision​​​‌ rule on the fly.‌ A key appeal of‌​‌ these methods is their​​ worst-case, distribution-free guarantee: they​​​‌ ensure sublinear regret with‌ respect to the best‌​‌ fixed decision in hindsight,​​ without requiring statistical assumptions​​​‌ on how requests are‌ generated. This makes online‌​‌ learning particularly attractive for​​ caching, where traffic can​​​‌ be non-stationary and hard‌ to predict. Over the‌​‌ last year, we studied​​ how to make such​​​‌ no-regret caching policies practical‌ at scale, by addressing‌​‌ constraints that arise in​​ real deployments—limited observability of​​​‌ requests, and reduced computational‌ and memory budgets of‌​‌ online algorithms operating over​​ very large catalogs.

Most​​​‌ existing algorithms involve computationally‌ expensive operations and require‌​‌ knowledge of all past​​ requests, which may not​​​‌ be feasible in practical‌ scenarios such as femtocaching,‌​‌ where a base station​​ (BS) jointly decides the​​​‌ content of many edge‌ caches and visibility of‌​‌ all requests at the​​ BS requires constant communication​​​‌ between these caches and‌ the BS. To capture‌​‌ this constraint, in 15​​, Sara Alouf and​​​‌ Giovanni Neglia , together‌ with Younes Ben Mazziane‌​‌ (Université Avignon, France) and​​ Francescomaria Faticanti (ENS, Lyon,​​​‌ France), study a single‌ cache problem under a‌​‌ more restrictive setting, that​​ they refer to as​​​‌ the Bernoulli Partial Observability‌ (BPO) model, in which‌​‌ the caching policy only​​ observes a request with​​​‌ probability $p$, reflecting‌ the fraction of requests‌​‌ forwarded from the edge​​ caches to the BS​​​‌ in the femtocaching example.‌ They propose a policy,‌​‌ based on the classic​​​‌ online learning algorithm Follow-the-Perturbed-Leader​ (FPL), that achieves an​‌ asymptotically optimal regret bound​​ of $𝒪(\sqrt{\mathrm{C‌}T/p})$ under BPO in $\mathrm{𝒪‌}\left(1\right)$ amortized​​ time complexity as $\mathrm{T‌}$ goes to infinity, where​ $C$ is the cache​‌ size and $T$ is​​ the number of requests.​​​‌ Moreover, they show that​ their policy extends to​‌ bipartite caching albeit with​​ a sublinear $\alpha$-regret​​​‌ for $\alpha =\mathrm{1}-1/\mathrm{e‌}$ and a higher computational​​ cost. The experimental evaluation​​​‌ compares the proposed solution​ with classic caching policies​‌ and validates the proposed​​ approach using both synthetic​​​‌ and real-world request traces.​

Caching policies based on​‌ online learning algorithms often​​ suffers high computation complexity,​​​‌ which hinders their practical​ adoption. In 16,​‌ Giovanni Neglia and Xufeng​​ Zhang , together with​​​‌ Damiano Carra (University of​ Verona, Italy), introduce a​‌ new variant of the​​ gradient-based online caching policy​​​‌ that achieves groundbreaking logarithmic​ computational complexity relative to​‌ catalog size, while also​​ providing regret guarantees. This​​​‌ advancement allows the authors​ to test the policy​‌ on large-scale, real-world traces​​ featuring millions of requests​​​‌ and items—a significant achievement,​ as such scales have​‌ been beyond the reach​​ of existing policies with​​​‌ regret guarantees. The regret​ guarantees and the low​‌ complexity are also maintained​​ in cases where items​​​‌ have non-uniform sizes. To​ the best of their​‌ knowledge, the proposed solution​​ is the only low-complexity​​​‌ no-regret policy for such​ a case, and their​‌ experimental results demonstrate for​​ the first time that​​​‌ the regret guarantees of​ gradient-based caching policies offer​‌ substantial benefits in practical​​ scenarios.

Online learning algorithms​​​‌ provide robust performance in​ caching problems but require​‌ substantial memory to store​​ per-file historical data, limiting​​​‌ their scalability to large-catalog​ systems. To overcome this​‌ challenge, in 46,​​ Xufeng Zhang , Sara​​​‌ Alouf , and Giovanni​ Neglia propose a dimensionality​‌ reduction algorithm based on​​ the Follow-the-Perturbed-Leader framework and​​​‌ the Johnson–Lindenstrauss lemma. Their​ method significantly reduces memory​‌ consumption while preserving sublinear​​ regret, making it well-suited​​​‌ for caching under resource​ constraints. Experiments on both​‌ synthetic and real-world traces​​ demonstrate its advantages over​​​‌ other memory-efficient approaches.

8.5.2​ Simultaenous Information and Energy​‌ Transmission in Wireless Networks​​

Simultaneous wireless information and​​​‌ power transfer aims to​ use the same radio-frequency​‌ signal to convey information​​ and deliver energy that​​​‌ can be harvested by​ the receiver. A central​‌ challenge is the inherent​​ trade-off between reliable communication​​​‌ and sufficient energy transfer,​ especially in practical regimes​‌ with finite block-length codes​​ and finite input constellations.​​​‌

In 25, Samir​ Medina Perlaza , together​‌ with Sadaf Ul Zuhra​​ and H. Vincent Poor​​​‌ (Princeton University, USA) and​ Mikael Skoglund (KTH Royal​‌ Institute of Technology, Sweden),​​ characterize the trade-offs between​​​‌ information and energy transmission​ over an additive white​‌ Gaussian noise channel in​​ the finite block-length regime​​​‌ with finite channel input​ symbols. The authors characterize​‌ these trade-offs in the​​ form of inequalities involving​​​‌ the information transmission rate,​ energy transmission rate, decoding​‌ error probability (DEP) and​​ energy outage probability (EOP)​​ for a given finite​​​‌ block-length code. The first‌ set of results identify‌​‌ a set of necessary​​ conditions that a given​​​‌ code must satisfy for‌ simultaneous information and energy‌​‌ transmission. They propose a​​ novel method for constructing​​​‌ a family of codes‌ that can satisfy a‌​‌ target information rate, energy​​ rate, DEP and EOP.​​​‌ Finally, achievability results identify‌ the set of tuples‌​‌ of information rate, energy​​ rate, DEP and EOP​​​‌ that can be simultaneously‌ achieved by the constructed‌​‌ family of codes.

8.5.3​​ Covert Communications

Covert communications​​​‌ study how a transmitter‌ (or adversary) can operate‌​‌ while keeping its very​​ presence hard to detect,​​​‌ typically by ensuring that‌ any statistical test at‌​‌ a warden cannot reliably​​ distinguish "signal-present" from "signal-absent."​​​‌ This viewpoint naturally connects‌ to sequential detection: if‌​‌ an observer monitors a​​ process for a change,​​​‌ a covert adversary may‌ adapt its behavior so‌​‌ that the change remains​​ barely detectable and detection​​​‌ is delayed as much‌ as possible under a‌​‌ prescribed false-alarm constraint.

In​​ 57, 24,​​​‌ Philippe Nain , together‌ with Amir Reza Ramtin‌​‌ and Don Towsley (University​​ of Massachusetts Amherst, USA),​​​‌ investigate the problem of‌ covert quickest change detection‌​‌ in a continuous-time setting,​​ where a Brownian motion​​​‌ experiences a drift change‌ at an unknown time.‌​‌ Unlike classical formulations, the​​ authors consider a covert​​​‌ adversary who adjusts the‌ post-change drift $\mu =‌‌\mu (\gamma )$ as a function of​​​‌ the false alarm constraint‌ parameter $\gamma$, with‌​‌ the goal of remaining​​ undetected for as long​​​‌ as possible. Leveraging the‌ exact expressions for the‌​‌ average detection delay (ADD)​​ and average time to​​​‌ false alarm (AT2FA) known‌ for the continuous-time CuSum‌​‌ procedure, the authors rigorously​​ analyze how the asymptotic​​​‌ behavior of ADD evolves‌ as $\mu \left(\mathrm{\gamma ‌‌}\right)\to 0$ with​​ increasing $\gamma$. Their​​​‌ results reveal that classical‌ detection delay characterizations no‌​‌ longer hold in this​​ regime. They derive sharp​​​‌ asymptotic expressions for the‌ ADD under various convergence‌​‌ rates of $\mu (\gamma )$, identify​​​‌ precise conditions for maintaining‌ covertness, and characterize the‌​‌ total damage inflicted by​​ the adversary. They show​​​‌ that the adversary achieves‌ maximal damage when the‌​‌ drift scales as $\mathrm{\mu }\left(\gamma \right)=‌\Theta (1/‌\sqrt{\gamma })$, marking‌​‌ a fundamental trade-off between​​ stealth and impact in​​​‌ continuous time detection systems.‌

9.1.1 Cifre contract with‌ EDF “Automated and responsible‌​‌ recommendation systems for digital​​ marketing” (February 2025 –​​​‌ January 2028)

Participants: Gaspard‌ Gerard Philippe Berthelier,‌​‌ Samir Medina Perlaza,​​ Giovanni Neglia.

• Contractor​​​‌: EDF
• Collaborators:‌ Etienne Le Naour, Tahar‌​‌ Nabil, Richard Niamke

In​​ collaboration with EDF, in​​​‌ the framework of Gaspard‌ Gerard Philippe Berthelier 's‌​‌ PhD thesis, we develop​​​‌ federated learning methods tailored​ to time series forecasting​‌ in the energy sector.​​ Our project explores personalized​​​‌ FL to allow each​ client to learn specialized​‌ models while benefiting from​​ collaborative training. We also​​​‌ investigate the use of​ public or non-sensitive data​‌ to improve representation learning​​ and model initialization. This​​​‌ work aims to deliver​ privacy-preserving, adaptive FL methods​‌ suitable for real-world industrial​​ time series applications.

9.1.2​​​‌ Inria challenge with Hivenet​ (September 2025–August 2029)

Participant:​‌ Giovanni Neglia.

• Project​​ Acronym:
  CUPSELI
• Project Title:​​​‌
  Collaborative Unified Platform for​ a Scalable and Efficient​‌ Learning Infrastructure
• Duration:
  September​​ 2025–August 2029
• Abstract:
  The​​​‌ CUPSELI challenge ambitions to​ push the boundaries of​‌ distributed computing and artificial​​ intelligence to offer a​​​‌ sovereign, secure, and sustainable​ alternative to centralized cloud​‌ solutions. It mobilizes a​​ large scientific community, bringing​​​‌ together 11 Inria research​ teams from six research​‌ centers: Rennes, Bordeaux, Lorraine,​​ Côte d’Azur, Lyon, and​​​‌ Paris. The CUPSELI challenge​ addresses three major technological​‌ challenges: sustainable and distributed​​ AI across diverse computing​​​‌ hardware, distributed and secure​ computing, and large-scale distributed​‌ computing. Two PhD students​​ will be recruited during​​​‌ 2026 in the framework​ of this project.

9.1.3​‌ Contracts with Nokia

9.1.4 Cifre contract​​ with NSP-SmartProfile “Automated and​​​‌ responsible recommendation systems for‌ digital marketing” (August 2022‌​‌ – May 2026)

Participants:​​ Konstantin Avrachenkov, Ibtihal​​​‌ El Mimouni.

• Contractor‌: NSP-SmartProfile
• Collaborators:‌​‌ Hervé Baile, Julien Musso​​

SmartProfile is a marketing​​​‌ platform that allows to‌ collect, to enhance and‌​‌ to analyze marketing data.​​ Digital marketing campaigns continue​​​‌ to expand across all‌ digital channels and media.‌​‌ The 'mass marketing' strategies​​ implemented by most companies​​​‌ show limits in terms‌ of performance and acceptance‌​‌ by clients, as well​​ as in terms of​​​‌ their impact on the‌ environment. In opposite to‌​‌ these practices, we believe​​ that current technologies, particularly​​​‌ in terms of Artificial‌ Intelligence (AI), should make‌​‌ marketing interactions more efficient​​ and virtuous. Through this​​​‌ research project, we want‌ to create an alternative‌​‌ solution to mass marketing​​ by switching to an​​​‌ intelligent, automated and eco-responsible‌ system, which will support‌​‌ the heterogeneity of data​​ and the diversity of​​​‌ sectors, and whose purpose‌ is to recommend the‌​‌ best content by determining​​ the most relevant target​​​‌ and taking into account‌ the communication constraints. This‌​‌ contract complements the Cifre​​ thesis of Ibtihal El​​​‌ Mimouni. Relevant publications: 34‌, 35.

9.1.5‌​‌ Cifre contract with Orange​​ Labs “Analytical modeling of​​​‌ large-scale wireless networks integrating‌ RIS” (September 2023 –‌​‌ September 2026)

Participants: Eitan​​ Altman, Konstantin Avrachenkov​​​‌, Julian Alfonso Santos‌ Bustos.

• Contractor:‌​‌ Orange Labs
• Collaborators:​​ Jean-Marc Kelif

A Reconfigurable​​​‌ Intelligent Surface (RIS) is‌ a programmable surface structure‌​‌ that allows one to​​ control the reflection of​​​‌ electromagnetic (EM) waves by‌ changing the electric and‌​‌ magnetic properties of the​​ surface. In the absence​​​‌ of RIS, short wavelentghs‌ signals as in 5G,‌​‌ are subject to a​​ huge attenuation when there​​​‌ is no direct line‌ of sight channel. Within‌​‌ our collaboration, we shall​​ evaluate and optimize the​​​‌ position of RIS.

This‌ contract complements the Cifre‌​‌ thesis of J. Santos.​​

10.1.1‌ Inria associate team not‌​‌ involved in an IIL​​ or an international program​​​‌

10.1.2 STIC/MATH/CLIMAT​‌ AmSud projects

10.1.3 Participation in​​ other International Programs

10.2.1 Visits​​​‌ of international scientists

10.2.2 Visits to​​​‌ international teams

10.3.1​​ Horizon Europe

NF-FOUNDS PC9​‌ PEPR 5G

Participants: Khushboo​​ Agarwal, Eitan Altman​​​‌, Samir Medina Perlaza​, Guodong Sun.​‌

• Project Acronym:
  NF-FOUNDS
• Project​​ Title:
  Networks of the​​​‌ Future - Foundations of​ Future Communications Networks
• Program:​‌
  ANR-22-PEFT-0010
• Coordinator:
  CEA (Dmitri​​ Kténas), CNRS (Serge Verdeyme),​​​‌ IMT (Daniel Koffman)
• Duration:​
  2023 - 2030
• Other​‌ Partners:
  EURECOM
• Summary:
  The​​ 5G network and the​​​‌ networks of the future​ represent a key issue​‌ for French and European​​ industry, society and digital​​​‌ sovereignty. This is why​ the French government has​‌ decided to launch a​​ dedicated national strategy. One​​​‌ of this strategy's priority​ ambitions is to produce​‌ significant public research efforts​​ so the national scientific​​​‌ community contributes fully to​ making progress that clearly​‌ responds to the challenges​​ of 5G and the​​​‌ networks of the future.​ In this context, the​‌ CNRS, the CEA and​​ the Institut Mines-Télécom (IMT)​​​‌ are co-leading the '5G'​ acceleration PEPR to support​‌ upstream research into the​​ development of advanced technologies​​​‌ for 5G and the​ networks of the future.​‌ Neo  is involved in​​ the theme "Networks and​​​‌ Telecommunications" and more specifically​ in the targeted projet​‌ 9 (PC9) Foundations of​​ Future Communications Networks (FOUNDS).​​​‌
• Publications in 2025:
  13​, 18, 23​‌, 25, 26​​, 31, 47​​​‌, 52, 53​, 54, 56​‌.

ANR PARFAIT

Participants:​​ Eitan Altman, Samir​​​‌ Medina Perlaza, Xinying​ Zou.

• Project Acronym:​‌
  PARFAIT
• Project Title:
  Planning​​ And leaRning For AI-Edge​​ compuTing
• Coordinator:
  Avignon Univ.​​​‌
• Duration:
  October 2021 -‌ September 2025
• Other Partners:‌​‌
  Conservatoire National des Arts​​ et Métiers (CNAM), Univ.​​​‌ Savoie Mont Blanc (USMB)‌
• Summary:
  The PARFAIT project‌​‌ develops theoretical foundations for​​ distributed and scalable resource​​​‌ allocation schemes on edge‌ computing infrastructures tailored for‌​‌ AI-based processing tasks. Algorithmic​​ solutions will be developed​​​‌ based on the theory‌ of constrained, delayed, and‌​‌ distributed Markov decision processes​​ to account for edge​​​‌ service orchestration actions and‌ quantify the effect of‌​‌ orchestration policies. Furthermore, using​​ both game and team​​​‌ formulations, the project will‌ pave the way for‌​‌ a theory of decentralized​​ orchestration, a missing building​​​‌ block necessary to match‌ the application quest for‌​‌ data proximity and the​​ synchronization problems that arise​​​‌ when multiple edge orchestrators‌ cooperate under local or‌​‌ partial system view. Finally,​​ to achieve efficient online​​​‌ edge service orchestration, such‌ solutions will be empowered‌​‌ with reinforcement learning techniques​​ to define a suit​​​‌ of orchestration algorithms able‌ to at once adapt‌​‌ over time to the​​ applications' load and cope​​​‌ with the uncertain information‌ available from AI-based applications'‌​‌ footprints.
• Publications in 2025:​​
  18, 23,​​​‌ 31, 36,‌ 47, 52,‌​‌ 53, 54,​​ 56, 58.​​​‌

Inria Challenge FedMalin

Participant:‌ Giovanni Neglia.

• Project‌​‌ Acronym:
  FedMalin
• Project Title:​​
  FEDerated MAchine Learning over​​​‌ the INternet
• Coordinator:
  Giovanni‌ Neglia and Aurélien Bellet‌​‌ (Premedical Inria team)​​
• Duration:
  November 2022 -​​​‌ November 2026
• Summary:

  In‌ many use-cases of Machine‌​‌ Learning (ML), data is​​ naturally decentralized: medical data​​​‌ is collected and stored‌ by different hospitals, crowdsensed‌​‌ data is generated by​​ personal devices, etc. Federated​​​‌ Learning (FL) has recently‌ emerged as a novel‌​‌ paradigm where a set​​ of entities with local​​​‌ datasets collaboratively train ML‌ models while keeping their‌​‌ data decentralized.

  FedMalin is​​ a research project that​​​‌ spans 10 Inria research‌ teams and aims to‌​‌ push FL research and​​ concrete use-cases through a​​​‌ multidisciplinary consortium involving expertise‌ in ML, distributed systems,‌​‌ privacy and security, networks,​​ and medicine. We propose​​​‌ to address a number‌ of challenges that arise‌​‌ when FL is deployed​​ over the Internet, including​​​‌ privacy and fairness, energy‌ consumption, personalization, and location/time‌​‌ dependencies. FedMalin will also​​ contribute to the development​​​‌ of open-source tools for‌ FL experimentation and real-world‌​‌ deployments, and use them​​ for concrete applications in​​​‌ medicine and crowdsensing. The‌ FedMalin Inria Challenge is‌​‌ supported by Groupe La​​ Poste, sponsor of the​​​‌ Inria Foundation.

• Publications in‌ 2025:
  33, 37‌​‌, 39, 42​​, 44.

DIAMOND​​​‌

Participants: Yaiza Bermudez,‌ Samir Medina Perlaza.‌​‌

• Project Acronym:
  DIAMOND
• Project​​ Title:
  Data-Injection Attacks in​​​‌ Supervised Machine Learning Systems‌
• Funding Agency:
  Agence de‌​‌ l'Innovation de Défense (AID)​​
• Principal Investigators:
  Samir Medina​​​‌ Perlaza and Iñaki Esnaola‌ (University of Sheffield)
• Duration:‌​‌
  November 2024 - October​​ 2027
• Summary:
  This project​​​‌ aims at studying the‌ impact on the generalization‌​‌ capabilities of federated learning​​ systems of data-injection attacks​​​‌ (DIA) in the context‌ of military applications. A‌​‌ DIA refers to any​​​‌ modification on the local​ training datasets aiming to​‌ tamper with the global​​ performance. The focus is​​​‌ on the special class​ of Stealth DIAs (S-DIA),​‌ which exhibit the lowest-probability​​ of detection. Before and​​​‌ after such attacks, training​ datasets exhibit empirical probabilities​‌ that are sufficiently close​​ in relative entropy, which​​​‌ makes the probability of​ attack detection arbitrarily close​‌ to zero. The project​​ funds the PhD scholarship​​​‌ of Yaiza Bermudez .​
• Publications in 2025:
  47​‌, 52, 53​​.

11.1.1​ Scientific events: organization

11.1.2 Journal​​​‌

11.1.3 Invited​​​‌ talks

• Konstantin Avrachenkov delivered​
  • a plenary talk “Accuracy​‌ and Efficiency of Semi-Supervised​​ Graph Clustering Methods” at​​​‌ the 20th Workshop on​ Modelling and Mining Networks​‌ (WAW 2025), June 30​​ - July 3, 2025,​​​‌ Vilnius, Lithuania;
  • an invited​ talk “Introduction to Graph​‌ Clustering” at Universidade Federal​​ do Rio de Janeiro,​​​‌ Rio de Janeiro, Brazil,​ December 8, 2025.
• Louis​‌ Hauseux delivered
  • a seminar​​ "(Hyper-)graphs, percolation and clustering​​​‌ performances" at the Mathnet​/Dyogene Inria team​‌ at Inria-Paris center, March​​ 2025;
  • a seminar "Méthodes​​​‌ de clustering avec graphes.​ Ou « Comment hacker​‌ (H)DBSCAN ? » Théorie,​​ algorithmes & applications" at​​​‌ Valeo AI, Paris, November​ 2025.
• Giovanni Neglia delivered​‌
  • an invited talk "Breaking​​ Privacy in Federated Learning:​​​‌ Advances in Attribute Inference​ and Data Reconstruction Attacks"​‌ at University of Palermo,​​ Italy, February 13, 2025;​​​‌
  • an invited talk "Federated​ Learning" at the Colloquium​‌ on Networks and Learning,​​ Universidade Federal do Rio​​​‌ de Janeiro, Rio de​ Janeiro, Brazil, October 23,​‌ 2025.
• Samir Medina Perlaza​​ delivered
  • a seminar "Variations​​​‌ of the Expectation due​ to Changes in the​‌ Measure: Applications to Generalization​​ and Game Theory" in​​​‌ the series Cambridge Information​ Theory Seminar, Centre for​‌ Mathematical Sciences, University of​​ Cambridge, February 26, 2025;​​​‌
  • an invited talk "Generalization​ Error of Machine Learning​‌ Algorithms" at the Information​​ Theory and Tapas Workshop,​​​‌ Universidad Carlos III de​ Madrid, April 2-5, 2025.​‌

11.1.4 Leadership within the​​ scientific community

• Eitan Altman​​​‌
  • is Fellow Member of​ IEEE;
  • is Member of​‌ WG 7.3 of IFIP​​ on Computer System Modeling;​​
  • is the elected Vice​​​‌ Chairman of WG 6.3‌ of IFIP on Performance‌​‌ of Communications Systems.
• Konstantin​​ Avrachenkov is a member​​​‌ of Conseil Scientifique &‌ Pédagogique EUR DS4H Univ.‌​‌ Côte d'Azur.
• Samir Medina​​ Perlaza
  • is a member​​​‌ of the Digital Presence‌ Committee of the IEEE‌​‌ Information Theory Society;
  • is​​ a workpackage leader of​​​‌ the PEPR - Réseaux‌ du Futur – A‌​‌ project funded by the​​ French National Agency for​​​‌ Research (ANR) via the‌ project n°ANR-22-PEFT-0010 of the‌​‌ France 2030 program;
  • is​​ the organizer of the​​​‌ PC9 Seminar on Wireless‌ Communications, a national online‌​‌ seminar part of the​​ PEPR – Réseaux du​​​‌ Futur.
• Giovanni Neglia
  • holds‌ a Chair by the‌​‌ Interdisciplinary Institute for Artificial​​ Intelligence 3IA Côte d'Azur,​​​‌ in the theme "Core‌ Elements of AI."

11.1.5‌​‌ Research administration

• Sara Alouf​​
  • is a member of​​​‌ the Colloquium Jacques Morgenstern‌ Committee of Inria center‌​‌ at Université Côte d'Azur,​​ since March 2023;
  • is​​​‌ a member of NICE,‌ the Invited Researchers Committee‌​‌ of Inria center at​​ Université Côte d'Azur, since​​​‌ June 2020;
  • is vice-head‌ of project-team Neo since‌​‌ January 2017;
  • is a​​ member of the Selection​​​‌ Committee of the dAIEGDE‌ European project.
• Konstantin Avrachenkov‌​‌ is an alternate representative​​ (suppléant, collège A) on​​​‌ the Center Committee of‌ Inria center at Université‌​‌ Côte d'Azur.
• Louis Hauseux​​ is
  • a representative (titulaire,​​​‌ collège C) on the‌ Center Committee of Inria‌​‌ center at Université Côte​​ d'Azur;
  • a representantive on​​​‌ the catering commission of‌ Inria center at Université‌​‌ Côte d'Azur.
• Alain Jean-Marie​​
  • has been leader of​​​‌ project-team Neo from January‌ 2017 until March 2025.‌​‌
• Samir Medina Perlaza
  • is​​ an alternate representative (suppléant,​​​‌ collège A) on the‌ Center Committee of Inria‌​‌ center at Université Côte​​ d'Azur;
  • is a Member​​​‌ of the Bureau of‌ the Réseaux, Information et‌​‌ Société Numérique Excellence Academy​​ of Université Côte d'Azur;​​​‌
  • represents Inria at the‌ Conseil du Département Disciplinaire‌​‌ Informatique of Université Côte​​ d'Azur.
• Giovanni Neglia
  • was​​​‌ a member of the‌ competitive exam jury for‌​‌ a researcher position (CRCN/ISFP)​​ at Inria Lille in​​​‌ 2025 (but did not‌ participate to the interviews‌​‌ because of a conflict​​ of interest);
  • is an​​​‌ elected member of Inria‌ evaluation committee, since September‌​‌ 2023. In this role,​​ he also participated in​​​‌ a working group that‌ proposed a set of‌​‌ reflections, guidelines, and recommendations​​ on the responsible use​​​‌ of generative AI in‌ research professions. To ensure‌​‌ wider dissemination, the corresponding​​ document is available both​​​‌ in English 49 and‌ in French 50.‌​‌
  • is a member of​​ the steering commitee of​​​‌ Université Côte d’Azur Graduate‌ School of Digital Systems‌​‌ for Humans (DS4H) since​​ September 2022.
  • has been​​​‌ leader of project-team Neo‌ since April 2025.

Master

• Sara‌ Alouf , Alain Jean-Marie‌​‌ , "Performance Evaluation of​​ Networks," 24H, M2 Ubinet,​​​‌ Université Côte d’Azur, France.‌
• Louis Hauseux , Konstantin‌​‌ Avrachenkov , "Statistical Analysis​​ of Networks," 21H, M2​​​‌ Data Science and Artificial‌ Intelligence, Université Côte d’Azur,‌​‌ France.
• Louis Hauseux ,​​​‌ "Statistical Inference," 30H, M1​ Data Science and Artificial​‌ Intelligence, Université Côte d’Azur,​​ France.
• Giovanni Neglia ,​​​‌ "Machine Learning: Theory and​ Algorithms," 24H, M2 Ubinet,​‌ Université Côte d’Azur, France.​​
• Giovanni Neglia together with​​​‌ Francesco Diana and Chuan​ Xu (Coati Inria​‌ team) "Federated Learning &​​ Data Privacy," 24H, M2​​​‌ Data Science and Artificial​ Intelligence, Université Côte d’Azur,​‌ France.
• Adrien Sardi "Optimisation​​ Différentiable: Théorie et Algorithmes,"​​​‌ 11H ETD, M1, ENSTA​ - IPP Paris, France.​‌
• Xufeng Zhang , "Mathématiques​​ pour l'IA," 24H, Bac+5​​​‌ Formations expertes, CentraleDigitalLab @LaPlateforme_,​ Centrale Méditerranée, France.
• Xufeng​‌ Zhang , "Machine Learning,"​​ 22H, Bac+5 Formations expertes,​​​‌ CentraleDigitalLab @LaPlateforme_, Centrale Méditerranée,​ France.

11.2.1 Supervision

11.2.2​​ Juries

International​​ journals

• 13 articleK.​​​‌Khushboo Agarwal, K.​Konstantin Avrachenkov, R.​‌Raghupati Vyas and V.​​Veeraruna Kavitha. Two​​​‌ Choice Behavioral Game Dynamics​ with Myopic-Rational and Herding​‌ Players.Proceedings of​​ the ACM on Measurement​​​‌ and Analysis of Computing​ Systems 912025​‌, 1-26HALDOI​​back to textback​​​‌ to textback to​ text
• 14 articleK.​‌Konstantin Avrachenkov and F.​​Flora Spieksma. Deviation​​​‌ matrix in denumerable Markov​ processes: Characterisation, rank-one perturbation​‌ and restart.The​​ Annals of Applied Probability​​​‌356December 2025​, 4354-4380HALDOI​‌back to text
• 15​​ articleY.Younes Ben​​​‌ Mazziane, F.Francescomaria​ Faticanti, S.Sara​‌ Alouf and G.Giovanni​​ Neglia. Efficient and​​​‌ Optimal No-Regret Caching Under​ Partial Observation.IEEE​‌ Transactions on NetworkingDecember​​ 2025, 1-11HAL​​​‌DOIback to text​
• 16 articleD.Damiano​‌ Carra, G.Giovanni​​ Neglia and X.Xufeng​​​‌ Zhang. Low-Complexity online​ learning for caching.​‌Computer Networks273December​​ 2025, 111743HAL​​​‌DOIback to text​back to textback​‌ to text
• 17 article​​ J.Jake Clarkson,​​​‌ K.Konstantin Avrachenkov and​ E.Eitan Altman.​‌ Queues with inspection cost:​​ To see or not​​​‌ to see? Queueing Systems​ 109 13 March 2025​‌ HAL DOI back to​​ text back to text​​​‌
• 18 articleF.Francisco​ Daunas, I.Iñaki​‌ Esnaola, S. M.​​Samir M. Perlaza and​​​‌ H. V.H. Vincent​ Poor. Asymmetry of​‌ the Relative Entropy in​​ the Regularization of Empirical​​​‌ Risk Minimization.IEEE​ Transactions on Information Theory​‌7182025,​​ 6198 – 6226HAL​​​‌DOIback to text​back to textback​‌ to text
• 19 article​​L. L.Lucas Lopes​​ Felipe, K.Konstantin​​​‌ Avrachenkov and D. S.‌Daniel Sadoc Menasché.‌​‌ From Leiden to Pleasure​​ Island: The Constant Potts​​​‌ Model for Community Detection‌ as a Hedonic Game‌​‌.Physica A: Statistical​​ Mechanics and its Applications​​​‌680December 2025,‌ 130989HALDOIback‌​‌ to text
• 20 article​​L.Louis Hauseux,​​​‌ K.Konstantin Avrachenkov and‌ J.Josiane Zerubia.‌​‌ Generalization of Single-Linkage with​​ Higher-Order Interactions.Applied​​​‌ Network ScienceNovember 2025‌. In press. HAL‌​‌DOIback to text​​
• 21 articleP.Phil​​​‌ Howlett, B. K.‌Brendan K Beare,‌​‌ M.Massimo Franchi,​​ J.John Boland and​​​‌ K.Konstantin Avrachenkov.‌ The Granger–Johansen representation theorem‌​‌ for integrated time series​​ on Banach space.​​​‌Journal of Time Series‌ Analysis462025,‌​‌ 432-457HALDOIback​​ to text
• 22 article​​​‌H.Hélène Le Cadre‌, M.Mandar Datar‌​‌, M.Mathis Guckert​​ and E.Eitan Altman​​​‌. Learning Market Equilibria‌ Preserving Statistical Privacy Using‌​‌ Performative Prediction.IEEE​​ Transactions on Automatic Control​​​‌7011November 2025‌, 7125-7140HALDOI‌​‌back to text
• 23​​ articleS. M.Samir​​​‌ M. Perlaza and G.‌Gaetan Bisson. Variations‌​‌ on the Expectation Due​​ to Changes in the​​​‌ Probability Measure.Entropy‌278August 2025‌​‌, 865HALDOI​​back to textback​​​‌ to textback to‌ text
• 24 articleA.‌​‌ R.Amir Reza Ramtin​​, P.Philippe Nain​​​‌ and D.Don Towsley‌. Quickest Change Detection‌​‌ in Continuous-Time in Presence​​ of a Covert Adversary​​​‌.IEEE Signal Processing‌ Letters32November 2025‌​‌, pp. 4299-4303HAL​​DOIback to text​​​‌
• 25 articleS.Sadaf‌ Ul Zuhra, S.‌​‌ M.Samir M Perlaza​​, H. V.H.​​​‌ Vincent Poor and M.‌Mikael Skoglund. Simultaneous‌​‌ Information and Energy Transmission​​ with Short Packets and​​​‌ Finite Constellations.IEEE‌ Transactions on Communications73‌​‌9March 2025,​​ 7505-7516HALDOIback​​​‌ to textback to‌ text

International peer-reviewed conferences‌​‌

• 26 inproceedingsK.Khushboo​​ Agarwal, K.Konstantin​​​‌ Avrachenkov, V.Veeraruna‌ Kavitha and R.Raghupati‌​‌ Vyas. Balancing rationality​​ and social influence: Alpha-rational​​​‌ Nash equilibrium in games‌ with herding.GameNets‌​‌ 2025 - 13th EAI​​ International Conference on Game​​​‌ Theory for NetworksCambridge,‌ United KingdomMarch 2025‌​‌HALDOIback to​​ textback to text​​​‌back to text
• 27‌ inproceedingsK.Konstantin Avrachenkov‌​‌, K.Kousik Das​​, V.Veeraruna Kavitha​​​‌ and V.Vartika Singh‌. Stability of Polling‌​‌ Systems for a Large​​ Class of Markovian Switching​​​‌ Policies.CDC 2025‌ - IEEE Conference on‌​‌ Decision and ControlRio​​ de Janeiro, BrazilDecember​​​‌ 2025HALback to‌ textback to text‌​‌
• 28 inproceedingsK.Konstantin​​ Avrachenkov, M.Madhu​​​‌ Dhiman and V.Veeraruna‌ Kavitha. Constrained Average-Reward‌​‌ Intermittently Observable MDPs.​​CDC 2025 - IEEE​​​‌ Conference on Decision and‌ ControlRio de Janeiro,‌​‌ BrazilDecember 2025HAL​​back to textback​​​‌ to text
• 29 inproceedings‌Y.Younes Ben Mazziane‌​‌ and O.Othmane Marfoq​​​‌. Universal and Tight​ Bounds on Counting Errors​‌ of Count-Min Sketch with​​ Conservative Updates.International​​​‌ Conference on Measurement and​ Modeling of Computer Systems​‌ (SIGMETRICS)92Stony​​ brook, NY, United States​​​‌June 2025, 1-32​HALDOIback to​‌ text
• 30 inproceedingsY.​​Younes Ben Mazziane,​​​‌ C.-M.Cleque-Marlain Mboulou-Moutoubi,​ F.Francesco De Pellegrini​‌ and E.Eitan Altman​​. Learning to Bid​​​‌ in Proportional Allocation Auctions​ with Budget Constraints.​‌23rd International Symposium on​​ Modeling and Optimization in​​​‌ Mobile, Ad Hoc, and​ Wireless Networks (WiOpt)Linkoping,​‌ SwedenMay 2025,​​ 1-8HALDOIback​​​‌ to text
• 31 inproceedings​F.Francisco Daunas,​‌ I.Iñaki Esnaola and​​ S. M.Samir M.​​​‌ Perlaza. A Dual​ Optimization View to Empirical​‌ Risk Minimization with f​​ -Divergence Regularization.IEEE-ITW​​​‌ 2025 - Information Theory​ WorkshopSidney, AustraliaSeptember​‌ 2025HALDOIback​​ to textback to​​​‌ textback to text​
• 32 inproceedingsF.Francesco​‌ Diana, O.Othmane​​ Marfoq, C.Chuan​​​‌ Xu, G.Giovanni​ Neglia, F.Frédéric​‌ Giroire and E.Eoin​​ Thomas. Attribute Inference​​​‌ Attacks for Federated Regression​ Tasks.39th Annual​‌ AAAI Conference on Artificial​​ Intelligence - AAAI 2025​​​‌3915Philadelphia (Pennsylvania),​ United StatesFebruary 2025​‌, 16271-16279HALDOI​​back to textback​​​‌ to text
• 33 inproceedings​F.Francesco Diana,​‌ A.André Nusser,​​ C.Chuan Xu and​​​‌ G.Giovanni Neglia.​ Cutting Through Privacy: A​‌ Hyperplane-Based Data Reconstruction Attack​​ in Federated Learning.​​​‌41st Conference on Uncertainty​ in Artificial Intelligence (UAI​‌ '25)Rio de Janeiro,​​ BrazilJuly 2025,​​​‌ 959–980HALDOIback​ to textback to​‌ textback to text​​back to text
• 34​​​‌ inproceedingsI.Ibtihal El​ Mimouni and K.Konstantin​‌ Avrachenkov. Deep Q-Learning​​ with Whittle Index for​​​‌ Contextual Restless Bandits: Application​ to Email Recommender Systems​‌.NLDL 2025 -​​ 6th Northern Lights Deep​​​‌ Learning ConferencePMLR-265Tromso,​ NorwayJanuary 2025,​‌ 176-183HALback to​​ textback to text​​​‌
• 35 inproceedingsI.Ibtihal​ El Mimouni and K.​‌Konstantin Avrachenkov. Fostering​​ Responsibility in Email Marketing:​​​‌ A Contextual Restless Bandit​ Framework.ECML PKDD​‌ 2025 - European Conference​​ on Machine Learning and​​​‌ Principles and Practice of​ Knowledge Discovery in Databases​‌LNAI-16020Porto, PortugalSeptember​​ 2025, 366-383HAL​​​‌DOIback to text​back to text
• 36​‌ inproceedingsA.Andrea Fox​​, F.Francesco De​​​‌ Pellegrini, E.Eitan​ Altman, A.Arnob​‌ Ghosh and N.Ness​​ Shroff. Performing Load​​​‌ Balancing under Constraints.​WiOpt 2025 - 23rd​‌ International Symposium on Modeling​​ and Optimization in Mobile,​​​‌ Ad Hoc, and Wireless​ NetworksLinkoping, SwedenMay​‌ 2025, 103-110HAL​​DOIback to text​​​‌back to text
• 37​ inproceedingsF.Franco Galante​‌, G.Giovanni Neglia​​ and E.Emilio Leonardi​​​‌. Scalable Decentralized Algorithms​ for Online Personalized Mean​‌ Estimation.39th Annual​​ AAAI Conference on Artificial​​​‌ Intelligence - AAAI 2025​Philadelphia (Pennsylvania), United States​‌February 2025HALDOI​​back to textback​​ to textback to​​​‌ text
• 38 inproceedingsL.‌Louis Hauseux, N.‌​‌Nahuel Soprano-Loto and K.​​Konstantin Avrachenkov. Higher-order​​​‌ Monte Carlo cluster dynamics‌ for community detection in‌​‌ Euclidean graphs.Asilomar​​ Conference on Signals, Systems,​​​‌ and Computers 2025Pacific‌ Grove (CA), United States‌​‌October 2025HALback​​ to text
• 39 inproceedings​​​‌T.-K.Tan-Khiem Huynh,‌ M.Malcolm Egan,‌​‌ G.Giovanni Neglia and​​ J.-M.Jean-Marie Gorce.​​​‌ Streaming Federated Learning with‌ Markovian Data.NeurIPS‌​‌ 2025 - Thirty-Ninth Annual​​ Conference on Neural Information​​​‌ Processing SystemsSan Diego,‌ United StatesDecember 2025‌​‌HALback to text​​back to textback​​​‌ to textback to‌ textback to text‌​‌
• 40 inproceedingsL.Lucas​​ Lopes Felipe, K.​​​‌Konstantin Avrachenkov and D.‌Daniel Sadoc Menasché.‌​‌ Robustness against Frustration in​​ Community Detection.SIGMETRICS​​​‌ 2025 - ACM Sigmetrics‌ 2025 (Student Competition)53‌​‌2Stony Brook, United​​ StatesAugust 2025,​​​‌ 128-130HALDOIback‌ to text
• 41 inproceedings‌​‌C.Cleque Marlain Mboulou-Moutoubi​​, Y.Younes Ben​​​‌ Mazziane, F.Francesco‌ De Pellegrini and E.‌​‌Eitan Altman. Best-Response​​ Learning in Budgeted α-Fair​​​‌ Kelly Mechanisms.NETGCOOP‌ 2025 - 12th International‌​‌ Conference of Networks, Games,​​ Control and OptimizationBilbao,​​​‌ SpainOctober 2025HAL‌DOIback to text‌​‌
• 42 inproceedingsD.Daniel​​ Richards Arputharaj, C.​​​‌Charlotte Rodriguez, A.‌Angelo Rodio and G.‌​‌Giovanni Neglia. Green​​ Federated Learning via Carbon-Aware​​​‌ Client and Time Slot‌ Scheduling.MASCOTS 2025‌​‌ - 33rd International Symposium​​ on the Modeling, Analysis,​​​‌ and Simulation of Computer‌ and Telecommunication SystemParis,‌​‌ FranceOctober 2025HAL​​DOIback to text​​​‌back to textback‌ to textback to‌​‌ textback to text​​
• 43 inproceedingsA.Angelo​​​‌ Rodio, G.Giovanni‌ Neglia, Z.Zheng‌​‌ Chen and E. G.​​Erik G Larsson.​​​‌ A Unified Convergence Analysis‌ for Semi-Decentralized Learning: Sampled-to-Sampled‌​‌ vs. Sampled-to-All Communication.​​AAAI-26 - 40th Annual​​​‌ AAAI Conference on Artificial‌ IntelligenceSingapore, SingaporeJanuary‌​‌ 2026HAL
• 44 inproceedings​​J.Jacopo Talpini,​​​‌ M.Marco Savi and‌ G.Giovanni Neglia.‌​‌ FedBEns: One-Shot Federated Learning​​ based on Bayesian Ensemble​​​‌.ICML 2025 -‌ 42nd International Conference on‌​‌ Machine LearningPMLR 267​​Vancouver, CanadaJuly 2025​​​‌HALback to text‌back to textback‌​‌ to textback to​​ text
• 45 inproceedingsM.​​​‌Martin van Waerebeke,‌ M.Marco Lorenzi,‌​‌ G.Giovanni Neglia and​​ K.Kevin Scaman.​​​‌ When to Forget? Complexity‌ Trade-offs in Machine Unlearning‌​‌.International Conference in​​ Machine Learning (ICML 2025)​​​‌Vancouver (BC), CanadaJuly‌ 2025HALback to‌​‌ text
• 46 inproceedingsX.​​Xufeng Zhang, S.​​​‌Sara Alouf and G.‌Giovanni Neglia. Memory-efficient‌​‌ Online Caching Policies with​​ Regret Guarantees.IFIP​​​‌ NETWORKING 2025Limassol, Cyprus‌May 2025HALback‌​‌ to textback to​​ textback to text​​​‌

Conferences without proceedings

• 47‌ inproceedingsY.Yaiza Bermudez‌​‌, G.Gaetan Bisson​​, I.Iñaki Esnaola​​​‌ and S.Samir Perlaza‌. Démonstrations des théorèmes‌​‌ folkloriques sur la dérivée​​​‌ de Radon-Nikodym.GRETSI’25​ - XXXe Colloque Francophone​‌ de Traitement du Signal​​ et des ImagesStrasbourg,​​​‌ FranceAugust 2025HAL​back to textback​‌ to textback to​​ textback to text​​​‌

Edition (books, proceedings, special​ issue of a journal)​‌

• 48 periodicalS.Sara​​ Alouf, O.Oliver​​​‌ Hohlfeld and Z.Zhiyuan​ Jiang, eds. Preface:​‌ Special issue on ITC​​ 2023.Performance Evaluation​​​‌167March 2025,​ 102462HALDOIback​‌ to text

Reports &​​ preprints

• 49 reportS.​​​‌Soraya Arias, M.​Michel Bergmann, F.​‌Fabien Campillo, M.-A.​​Marie-Agnès Enard, C.​​​‌Christian Fabre, F.​Frédérick Garcia, B.​‌Benjamin Guedj, E.​​Emmanuel Jeannot, G.​​​‌Giovanni Neglia, D.​Diane Peurichard, D.​‌Daniel Racoceanu, B.​​Benoît Sagot and G.​​​‌Gaëlle Tworkowski. Reflections​ on the Use of​‌ Generative AI for Research​​ Professions.InriaJuly​​​‌ 2025HALback to​ text
• 50 reportS.​‌Soraya Arias, M.​​Michel Bergmann, F.​​​‌Fabien Campillo, M.-A.​Marie-Agnès Enard, C.​‌Christian Fabre, F.​​Frédérick Garcia, B.​​​‌Benjamin Guedj, E.​Emmanuel Jeannot, G.​‌Giovanni Neglia, D.​​Diane Peurichard, D.​​​‌Daniel Racoceanu, B.​Benoît Sagot and G.​‌Gaëlle Tworkowski. Réflexions​​ sur l'usage de l'IA​​​‌ générative pour les métiers​ de la recherche.​‌Inria2025, 1-10​​HALback to text​​​‌
• 51 miscK.Konstantin​ Avrachenkov, L.Lucas​‌ Gamertsfelder and B.Bernard​​ Mourrain. Weighted Moment-SoS​​​‌ approach to POMDPs with​ polynomial data.November​‌ 2025HAL
• 52 report​​Y.Yaiza Bermudez,​​​‌ G.Gaetan Bisson,​ I.Iñaki Esnaola and​‌ S. M.Samir M.​​ Perlaza. Proofs for​​​‌ Folklore Theorems on the​ Radon-Nikodym Derivative.RR-9591​‌InriaJuly 2025HAL​​back to textback​​​‌ to textback to​ textback to text​‌
• 53 reportY.Yaiza​​ Bermudez, S. M.​​​‌Samir M. Perlaza,​ I.Iñaki Esnaola and​‌ H. V.H. Vincent​​ Poor. Statistical Federated​​​‌ Learning and Generalization.​RR-9599InriaOctober 2025​‌HALback to text​​back to textback​​​‌ to textback to​ text
• 54 reportF.​‌Francisco Daunas, I.​​Iñaki Esnaola and S.​​​‌ M.Samir M. Perlaza​. A Dual Optimization​‌ View to Empirical Risk​​ Minimisation with f-Divergence Regularization​​​‌.RR-9594InriaJuly​ 2025HALback to​‌ textback to text​​back to text
• 55​​​‌ reportM.Maxime Nicaise​, Y.Yaiza Bermudez​‌ and S. M.Samir​​ M. Perlaza. On​​​‌ optimal transport with f-divergence​ regularization.RR-9607Inria​‌January 2026HAL
• 56​​ reportS. M.Samir​​​‌ M. Perlaza and G.​Gaetan Bisson. Variations​‌ on the Expectation due​​ to Changes in the​​​‌ Probability Measure.RR-9592​InriaJuly 2025,​‌ 28HALback to​​ textback to text​​​‌back to text
• 57​ miscA. R.Amir​‌ Reza Ramtin, P.​​Philippe Nain and D.​​​‌Don Towsley. Quick​ Change Detection in Continuous-Time​‌ in Presence of a​​ Covert Adversary.September​​ 2025HALback to​​​‌ text
• 58 reportX.‌Xinying Zou, S.‌​‌ M.Samir M. Perlaza​​ and I.Iñaki Esnaola​​​‌. A Minimax Approach‌ for Robustifying Statistical Learning‌​‌ Algorithms.9600Inria​​December 2025HALback​​​‌ to textback to‌ text

Other scientific publications‌​‌

• 59 miscH.Hélène​​ Le Cadre, M.​​​‌Mandar Datar, M.‌Mathis Guckert and E.‌​‌Eitan Altman. Supplementary​​ Material to "Learning Market​​​‌ Equilibria Preserving Statistical Privacy‌ Using Performative Prediction".‌​‌April 2025HALback​​ to text