Research

I use stochastic geometry and information theory to answer engineering questions about wireless networks: how much coverage a deployment delivers, how quickly a short packet can be decoded reliably, and which investment pays off.

PhD research · 2021–2024 · Inria Paris / ENS, with Nokia Networks France

Stochastic geometry for RIS-assisted cellular networks

Reconfigurable intelligent surfaces (RIS) are thin panels that redirect radio waves toward users. Deployed across a city, their benefit depends on where base stations and surfaces happen to be. I modeled both as random point processes and derived coverage and rate for the whole network, then linked these results to deployment cost.

Explore in RIS Lab (opens in a new tab)

RIS-assisted cellular network drawn as a random spatial model Base stations at random positions partition the plane into cells. Reconfigurable intelligent surfaces and users are scattered at random. A typical user at the center is linked directly to its nearest base station and indirectly through two nearby surfaces.
One random realization of the model: base stations with their cells, randomly placed surfaces and users. The typical user at the center receives a direct signal from its serving base station plus copies reflected by nearby surfaces.
  1. Stochastic-geometry framework for RIS-assisted OFDM cellular networks with random base-station and RIS deployments; coverage probability and ergodic rate. IEEE Transactions on Wireless Communications, 2026 · arXiv
  2. Closed-form analysis of signals reflected by multiple surfaces in RIS-aided networks. IEEE WiOpt 2025 · arXiv
  3. Techno-economic analysis with Nokia researchers: when deploying RIS is more cost-effective than densifying the network. IEEE WiOpt 2025 · arXiv

Funded by the ERC Advanced Grant NEMO (Network Mobility, PI François Baccelli) and a CIFRE contract with Nokia Networks France.

Postdoctoral research · 2025–2026 · Inria Centre at Université Côte d’Azur · PEPR Future Networks, project FOUND

Variable-length stop-feedback codes for short packets

With variable-length stop-feedback (VLSF) codes, the transmitter keeps sending until the receiver is confident enough to decode and returns a one-bit stop signal. Feedback cuts the average delay, but a practical receiver can only attempt decoding at a few instants. My work quantifies the resulting trade-offs for short packets.

Explore in VLSF Lab (opens in a new tab)

Sparse variable-length stop-feedback decoding The accumulated information density grows over channel uses like a random walk with positive drift. The receiver checks it against a threshold gamma only at four decoding instants. It is below the threshold at the first two instants and above it at the third, where the receiver decodes and sends the stop signal. channel uses information density n₁ n₂ n₃ n₄ threshold γ decode, send stop continue continue
Sparse VLSF decoding (illustration). The receiver tests the accumulated information density only at the decoding instants n₁, …, n₄. The first test above the threshold γ triggers decoding and the stop feedback; later instants are not used.
  1. Optimization of a limited set of decoding instants for sparse VLSF codes using saddlepoint approximations. IEEE ICC 2026 · arXiv
  2. Computable bounds on the information density over correlated noncoherent fading channels, enabling VLSF decoding with reliability guarantees when exact evaluation is intractable. IEEE ISIT 2026 · arXiv
  3. Extension to multiple access: performance guarantees of SIC-based decoding in a hardcore-regulated MAC, from a research visit I initiated at INSA Lyon. IEEE GLOBECOM 2026 · arXiv

Collaborators: Samir M. Perlaza (Inria), Philippe Mary (INSA Rennes, IETR), Jean-Marie Gorce and Kevin Zagalo (INSA Lyon, CITI).

Earlier and related work

Nokia Bell Labs · 2020

Retransmissions for URLLC

Iterative resolution and optimal scheduling of blind retransmissions for multi-user ultra-reliable low-latency communication.

IEEE ICC Workshops 2021

RWTH Aachen · 2020–2021

Polar-coded modulation

Multi-level polar-coded modulation designed by density evolution, for short packets in the finite-blocklength regime.

IEEE ISWCS 2021

RWTH Aachen · 2018–2021

Resource allocation and UAV networks

Finite-blocklength relaying and NOMA; resource allocation for ground networks serving UAVs; UAV trajectories for wireless power transfer.

IEEE TVT 2020 · IEEE ICC Workshops 2022

Inria · 2026

Online caching

Optimistic online caching with switching cost and dynamic batched requests.

MAIN 2026 · Best paper runner-up

Software

2021–2026

risgeo: open-source C++20 library for RIS-assisted network analysis

Header-only · CMake, CTest · OpenMP · MIT license · github.com/Sun8Guden/RISassistedCellularFramework

  • Numerical solver for the multidimensional integrals of the stochastic-geometry framework
  • Heap instead of per-step sorting: 33× less overhead · OpenMP: 9.4 s → 2.2 s
  • Identical serial and parallel results · validated against 4 million Monte Carlo samples
2026

erc_phy: a 5G modem physical layer in C

Personal project · C11, about 8,300 lines · 12 test programs

  • Floating-point DSP modules behind small C APIs: FFT, OFDM, channel estimation, LDPC decoding
  • Slot-by-slot sample processing · 5G FAPI message interface · fault injection and message traces in tests
2026

5G NR radio link in Python

Personal project · NumPy, 23 modules, 20 unit tests

  • From raw IQ samples to decoded bits, driven by command-line scripts

Research simulations: Python, C/C++, MATLAB and Julia; Monte Carlo and system-level network simulation.

Industry collaboration

CIFRE doctorate with Nokia Networks France (2021–2024) and an internship in the 3GPP RAN2 standardization team of Nokia Bell Labs (2020). These collaborations produced three papers co-authored with Nokia researchers (IEEE TWC 2026, WiOpt 2025, ICC Workshops 2021).

They taught me to translate industrial requirements into mathematical models and to evaluate deployment trade-offs, for example comparing RIS deployment with network densification in terms of investment.