Coded caching is a powerful technique designed to mitigate peak network traffic by proactively pushing content to user devices during off-peak hours. By carefully structuring what resides in local memory, a central server can broadcast coded transmissions (such as XORed file segments) that simultaneously satisfy the distinct requests of multiple users.
While early frameworks assumed uniform file distributions and independent content, real-world data is often highly correlated, layered by quality, or requested by users with diverse preferences. Recent research reveals a fundamental, unifying insight: the optimal strategy for allocating cache space is not static. Instead, it undergoes a structural regime shift dictated entirely by the available local cache capacity.
Understanding why these strategies shift requires looking at how local storage constraints interact with the mathematical efficiency of broadcast delivery mechanisms.
The Counter-Intuitive Case of Correlated Sources
When a library contains correlated files (such as video sequels or localized news updates), the data can be mathematically de-correlated into "common" descriptions shared across files and "private" descriptions unique to specific files.
Intuitively, one might assume that if user devices have severely limited cache capacity, the network should prioritize caching the common data. After all, storing shared information seems like the most efficient way to utilize scarce memory. However, information-theoretic analysis proves the exact opposite.
As demonstrated by Parisa Hassanzadeh et al. in Memory-rate trade-off for caching and delivery of correlated sources and Rate-Memory Trade-off for the Two-User Broadcast Caching Network with Correlated Sources, when cache capacity is low, systems must prioritize storing private data segments over common ones.
Low-Memory Regime:
[ User Cache ] ---> Prioritizes Private (Unique) Data
[ Broadcast ] ---> Delivers Common (Shared) Data via Simple Multicast
High-Memory Regime:
[ User Cache ] ---> Prioritizes Common (Shared) Data
[ Broadcast ] ---> Delivers Private (Unique) Data via Coded Multicast
The reason for this shift lies in the varying efficiency of the delivery phase. Coded multicasting is exceptionally effective at reducing the transmission rate of unique, private data because the server can blend different users' private requests into a single broadcast. Conversely, common data can be broadcast easily without complex coding, but it does not yield the same high-order multicast coding gains when mixed with other traffic.
By caching private components locally when memory is tight, the system eliminates the hardest-to-deliver unique bottlenecks. The remaining common data can then be cleared via simple, low-cost broadcasting. As cache capacity expands, the optimal strategy shifts: the authors show in Rate-Memory Trade-Off for Caching and Delivery of Correlated Sources that only when memory is abundant should the allocation transition to prioritizing common descriptions, balancing the global and local footprint of the data.
Hierarchical and Profile-Driven Prioritization
This capacity-dependent inversion also dictates how networks handle multi-layered content and diverse user preferences.
Successive Refinement and Quality Layers
Consider a network delivering video to heterogeneous devices where files are encoded using successive refinement—consisting of a critical "base layer" for standard resolution and subsequent "refinement layers" for high definition.
In Centralized Coded Caching for Heterogeneous Lossy Requests, Deniz Gunduz et al. establish that cache allocation must adapt strictly to total capacity thresholds. When memory is scarce, a strict hierarchical allocation strategy must prioritize base layers exclusively. This ensures that the absolute minimum requirements to satisfy the widest audience are met locally. Only when the cache budget crosses a critical threshold can the allocation strategy transition to a proportional model, distributing space between base and refinement layers to unlock higher-quality streams via coded delivery.
Global vs. Local Caching Gains
A parallel trade-off emerges when users are divided into distinct classes based on content preferences. In Coded Caching With Heterogeneous User Profiles, Ciyuan Zhang et al. highlight the tension between "global" caching gains (storing content shared across all classes) and "local" caching gains (storing content specific to an individual class).
The authors demonstrate that for small cache sizes, the system must aggressively favor global caching gains to preserve any meaningful opportunity for multicast delivery. As storage capacity grows, the optimal strategy shifts toward a balanced allocation that accommodates local class preferences, because the larger cache size provides enough headroom to exploit both shared and specialized multicast opportunities simultaneously.
Infrastructure and Physical Layer Adaptations
The size of the local cache constraint does not just change what content is stored; it alters the optimal architecture of the network itself, shifting how data is encoded and how antennas transmit signals.
Virtual Cache Extensions
In distributed storage networks where parity servers exist alongside standard content libraries, the local cache size changes the timing of data encoding. Tianqiong Luo et al. demonstrate in Coded Caching with Distributed Storage that systems with limited local memory benefit from treating parity servers as "virtual cache extensions."
In this low-memory regime, the network pre-calculates and stores linear combinations of data segments within the infrastructure. When user caches are large, this rigid pre-calculation becomes a bottleneck; instead, the optimal strategy shifts to storing plain segments and performing flexible, on-the-fly coding during the delivery phase to maximize transmission efficiency.
Coded Caching vs. Spatial Multiplexing
Ultimately, the available memory dictates whether the network should rely on storage-based coordination or physical-layer hardware coordination. In Scalable Content Delivery with Coded Caching in Multi-Antenna Fading Channels, Sheng Yang et al. map the trade-off between user memory and transmit power in multi-antenna (MIMO) systems.
[ Extreme Low Memory ] -------------------------> [ High Memory ]
Spatial Multiplexing Coded Caching
(Relies on Transmit Power/Antennas) (Relies on Local Storage)
When local device storage is extremely limited, coded caching algorithms fail to generate enough overlapping content demands to be useful. In this regime, the network must pivot to spatial multiplexing, using multi-antenna beamforming to separate user channels via raw transmit power. Coded caching only emerges as the dominant, scalable strategy when local cache sizes cross the threshold necessary to offload the burden from the physical antennas to local silicon.
Key Foundational Work
Rate-Memory Trade-off for the Two-User Broadcast Caching Network with Correlated Sources (2017) — contributed 3 concept elements to this direction.
Memory-rate trade-off for caching and delivery of correlated sources (2016) — contributed 3 concept elements to this direction.
Coded Caching with Distributed Storage (2016) — contributed 2 concept elements to this direction.