Ferkans — Interactive Telecom Tutor

Why GDoF?

DoF analysis (Chapter 5) assumes all users have the same high-SNR scaling. On a degraded BC with user SNRs $\rho_k = \rho^{\alpha_k}$ for different $\alpha_k$ , the classical DoF framework is insufficient — it cannot distinguish user 1 (SNR $\rho$ ) from user 2 (SNR $\rho^{1/2}$ ). Both have DoF = 1 individually.

Generalized Degrees of Freedom (GDoF) resolve this by tracking the SNR exponents. User $k$ 's GDoF is $\mathrm{GDoF}_k \;=\; \lim_{\rho \to \infty} \frac{R_k}{\log_2 \rho},$ which can exceed 1 in the sense that $\mathrm{GDoF}_k \leq \alpha_k$ (up to the user's own SNR ceiling).

The JLEC 2019 characterization is naturally in GDoF terms. We flesh out the notation and verify the optimal separation result in this richer framework.

Definition:
Generalized Degrees of Freedom

Let the SNRs of the $K$ users scale as $\rho_k = \rho^{\alpha_k}$ for a common scaling parameter $\rho \to \infty$ and fixed exponents $\alpha_1 \geq \ldots \geq \alpha_K \geq 0$ . The per-user GDoF is $\mathrm{GDoF}_k \;=\; \lim_{\rho \to \infty} \frac{R_k}{\log_2 \rho}.$ The GDoF region is the set of achievable $(\mathrm{GDoF}_1, \ldots, \mathrm{GDoF}_K)$ tuples.

GDoF generalizes DoF: when all $\alpha_k = 1$ , GDoF = DoF. The framework was introduced by Etkin, Tse, and Wang (2008) for the interference channel and has since become standard for heterogeneous channel analyses.

Theorem: GDoF Region for Mixed Cache-Aided BC (Detailed)

For the cache-aided $L$ -antenna degraded BC with user SNR exponents $\alpha_1 \geq \ldots \geq \alpha_K$ and memory ratio $\mu$ , the GDoF region of mixed cacheable $R_c^{(k)}$ and uncacheable $R_u^{(k)}$ rates satisfies, for every subset $\mathcal{S} \subseteq [K]$ : $\sum_{k \in \mathcal{S}} \mathrm{GDoF}_c^{(k)}/\alpha_k + \sum_{k \in \mathcal{S}} \mathrm{GDoF}_u^{(k)}/(\alpha_k \cdot L/(t+L)) \leq |\mathcal{S}|.$ The region is achieved by a separation-plus-weighting scheme.

Each user's cacheable and uncacheable "costs" are normalized by its SNR exponent (stronger channel = cheaper delivery). The full region is the intersection of $2^K$ linear constraints, one per subset $\mathcal{S}$ . In practice only a few constraints are active at the optimum; the active ones define the Pareto boundary.

Proof

Per-subset constraint

For any subset $\mathcal{S}$ , a cut-set argument bounds the aggregate GDoF delivered to $\mathcal{S}$ by the BC capacity scaled by caches and antennas.

Intersection

The achievable GDoF region is the intersection of all $2^K$ subset constraints. Typically polytope with $O(K)$ effective constraints.

Achievability via weighted separation

Allocate fractions $\theta_k$ of channel uses to user $k$ 's cacheable mode and uncacheable mode. Optimize $\theta$ to touch the boundary. Closed form depends on $\alpha_k, t, L$ .

Special case: uniform $\alpha$

If all $\alpha_k = 1$ , the region reduces to the DoF region of Chapter 5 for the cacheable part and MU-MIMO for the uncacheable part. The JLEC pentagon of §6.3 emerges. $\blacksquare$

Example: Two-Class Heterogeneous Users

Consider $K = 10$ users: 5 "strong" with $\alpha_s = 1$ , 5 "weak" with $\alpha_w = 0.5$ . $L = 2$ antennas, $\mu = 0.2$ (so $t = 2$ ). Compute the GDoF achievable for each class.

Solution

DoF pools

Cacheable DoF: $t + L = 4$ . Uncacheable DoF: $L = 2$ .

Per-user GDoF

For user $k$ : cacheable GDoF contribution $\propto \alpha_k$ ; uncacheable similarly. Symmetric allocation to within each class: Strong user ( $\alpha_s = 1$ ): $\mathrm{GDoF}_c + \mathrm{GDoF}_u \leq 1$ (individual cap). Weak user ( $\alpha_w = 0.5$ ): $\mathrm{GDoF}_c + \mathrm{GDoF}_u \leq 0.5$ .

Pool constraints

Cacheable pool: $\sum_k \mathrm{GDoF}_c^{(k)} \leq (t+L) = 4$ , with weighted per-user caps. Uncacheable pool: $\sum_k \mathrm{GDoF}_u^{(k)} \leq L = 2$ .

Pareto point

At symmetric operation: all strong users get 0.5 cacheable + 0.25 uncacheable = 0.75 GDoF total. Sum from strong users: $5 \cdot 0.5 = 2.5$ cacheable + $5 \cdot 0.25 = 1.25$ uncacheable. Similarly 5 weak users at 0.375 total => cacheable + uncacheable sums of 1.5 + 0.75. Combined cacheable: 4 (at cap); uncacheable: 2 (at cap). Both constraints active — Pareto-optimal.

Degraded BC Rate Region: Superposition vs Time-Sharing

For a two-user degraded BC with strong user SNR $\rho_s$ and weak user SNR $\rho_w$ , visualize the superposition-coding capacity region (Cover '72) and the time-sharing lower bound. Superposition strictly dominates time-sharing except at the corners. This is the underlying mechanism that the JLEC separation scheme exploits.

Parameters

Strong user SNR (dB)20

Weak user SNR (dB)5

Key Takeaway

The JLEC 2019 separation scheme is GDoF-optimal across all heterogeneity profiles. Cacheable traffic uses Lampiris-Caire (DoF $t+L$ ); uncacheable traffic uses MU-MIMO (DoF $L$ ); time-sharing between them is optimal in GDoF. The converse argument uses the fundamental fact that caches cannot help uncacheable content.

⚠️Engineering Note

From GDoF Theory to Deployed Schemes

The GDoF-optimal JLEC separation has a clean deployment path:

Time-domain multiplexing. Allocate dedicated time slots to cacheable vs. uncacheable delivery. Operationally simple — matches 5G NR's subcarrier/slot structure.
Rate adaptation. Each mode uses channel-aware MCS. Cacheable mode benefits from MAN + multicast; uncacheable mode from ZF unicast.
Cache warm-up. Off-peak pre-placement, refreshed at library turnover intervals (Ch 1 engineering note).
Admission control. Uncacheable latency-critical traffic (live calls) can preempt cacheable traffic (video).

Realizable trade-offs are more nuanced than the GDoF pentagon — finite SNR, imperfect CSIT, and scheduling delays all affect the actual operating point. But the GDoF framework provides a clean baseline to benchmark against.

Practical Constraints

•
5G NR supports separate resource pools for unicast and multicast
•
JLEC separation time-shares cacheable and uncacheable traffic
•
CSIT quality affects the spatial multiplexing gain of the uncacheable phase
•
Cache churn must not exceed peak-vs-off-peak time ratio

Historical Note: From Mohajer-Tuninetti to JLEC

2015–2020

The mixed-traffic coded-caching question was posed informally around 2015–2016 as a natural practical extension of MAN. Mohajer and Tuninetti gave partial answers in 2017 for specific $(K, M)$ regimes. The clean GDoF-optimal separation result came from the CommIT collaboration: Joudeh, Lampiris, Elia, Caire at EURECOM and TU Berlin. The 2019 ISIT paper was later expanded into a 2020 journal paper with additional converse details.

A related stream is demand-aware coded caching (Wan-Caire 2021+), which exploits the distinction between "popular" (cacheable) and "rare" (uncacheable) content statistically rather than operationally. This is the topic of Chapter 13.

The JLEC separation result is significant because it tells system designers that simple time-sharing is sufficient — no exotic joint coding is needed in the asymptotic regime. This is a useful "separation theorem" in the classical information-theoretic sense.

Common Mistake: GDoF Is Not a Rate at Any Finite SNR

Mistake:

Using the GDoF region directly as a finite-SNR rate region.

Correction:

GDoF is an asymptotic quantity: $\mathrm{GDoF}_k = \lim_{\rho \to \infty} R_k/\log_2 \rho$ . At any finite SNR, actual rates are smaller than $\mathrm{GDoF} \cdot \log_2 \rho$ by an $O(1)$ term that captures pre-log constants and leakage. For system design at realistic SNRs (10–30 dB), the GDoF region is an over-estimate; the exact rate depends on finite-SNR details.

The GDoF region is still valuable as a ceiling against which finite-SNR schemes can be benchmarked. The separation theorem (JLEC '19) establishes the shape of this ceiling and the nature of the optimal scheme structure.

GDoF-Optimal Separation Scheme