The $\Theta(M/N)$ Scaling Law (Ji-Caire-Molisch)

The Headline Result

The central theorem of D2D caching theory — due to Ji, Caire, and Molisch (IEEE Trans. IT, 2016) — is that the per-user throughput scales as $\Theta(M/N)$ independent of the network size $n$ . This is strikingly different from every ad-hoc capacity result before it: in classical Gupta-Kumar analysis, per-user throughput vanishes as $n$ grows. Caching reverses this: the effective throughput is determined entirely by the memory ratio $M/N$ , not by the number of users.

This result is a CommIT contribution and one of the most fundamental in coded-caching theory. It provides the theoretical justification for D2D-based content delivery as a scalable alternative to infrastructure.

Theorem: Ji-Caire-Molisch Scaling Law

Consider a D2D caching network with $n$ users uniformly distributed in a unit-area region, per-user cache $M = \mu \cdot N$ files from a library of $N$ files. Under i.i.d. uniform demands and a protocol-model interference constraint, the per-user throughput satisfies $T_n(M) \;=\; \Theta(\mu), \quad n \to \infty,$ provided the per-user cache has constant $\mu$ (i.e., $M = \Theta(N)$ ).

Intuition in three moves:

With $n M$ total cached copies and $n$ users demanding random files, each file is cached $\sim nM/N$ times on average.
The probability that a random user's demand is within D2D range of a cached copy is $\Theta(1)$ (since neighbors are $\Theta(1)$ many and each has $M/N$ fraction of the library).
Spatial reuse allows $\Theta(n)$ simultaneous D2D transmissions. Aggregate throughput $\Theta(n)$ ; per-user $\Theta(1)$ ... but scaled by the hit probability $M/N$ , giving $\Theta(M/N)$ .

The miracle: throughput is independent of $n$ because both the supply (cached copies) and demand scale with $n$ .

Proof

Setup and interference radius

Choose interference radius $r(n) = \sqrt{c \log n / n}$ for a suitable constant $c$ , so that the resulting random geometric graph is connected with high probability and each user has $\Theta(\log n)$ neighbors.

Cache coverage

With random uniform placement of $M$ files per user, each user has $\Theta(\log n \cdot M/N)$ neighbors caching a given file on average. For $M/N$ constant, this is $\Theta(\log n)$ copies of any desired file in a user's D2D range.

Local hit probability

Under uniform demand, the probability that at least one of a user's $\Theta(\log n)$ neighbors caches its demanded file approaches 1 as $n \to \infty$ , even accounting for the $M/N$ fraction of the library each caches.

Spatial reuse

Transmissions at distance $\geq r(n)$ don't interfere. The number of simultaneous non-interfering transmissions is $\Theta(n / (r(n)^2)) = \Theta(n / (\log n / n)) = \Theta(n^2/\log n)$ . This is way more than the $n$ users; we are not pairing-limited.

Aggregate vs per-user

Each simultaneous link carries rate $C_{\text{link}}$ . Aggregate throughput: $\Theta(n)$ per-round throughput. Divided by $n$ users, each user gets $\Theta(1)$ per-round "opportunity" — but only $\Theta(M/N)$ of those opportunities are hits (local content serves demand). Hence per-user throughput = $\Theta(M/N)$ . $\blacksquare$

🎓CommIT Contribution(2016)

Fundamental Limits of Caching in Wireless D2D Networks

M. Ji, G. Caire, A. F. Molisch — IEEE Transactions on Information Theory, vol. 62, no. 2

The Ji-Caire-Molisch 2016 paper is the foundational CommIT contribution for D2D caching theory. Its key results:

Per-user throughput scales as $\Theta(M/N)$ , independent of $n$ . This is a constant scaling — the first throughput result in ad-hoc wireless that doesn't decrease with network size.
Caching converts the local D2D network into a globally distributed library: the aggregate cache has effective size $\sim nM/N$ copies of each file, ensuring near-certain local hit.
Order-optimality: the achievable rate matches a cut-set lower bound up to constants, so the scaling is the correct asymptotic answer.

The result has been extended in many directions: to coded multicasting (Ch 11), to D2D with privacy (Ch 12), and to hybrid D2D/infrastructure networks. It is one of the most-cited coded-caching theory papers.

coded-cachingd2dcommitscaling-lawView Paper →

Per-User Throughput Scaling

Per-user throughput vs network size $n$ on log-log axes. D2D + caching: flat at $\mu$ (constant, $\Theta(M/N)$ ). D2D without caching (Gupta-Kumar): $\Theta(1/\sqrt{n \log n})$ , decreasing. Infrastructure (cellular, no caching): $\Theta(1/n)$ , decreasing faster. The D2D+caching advantage grows with $n$ .

Parameters

Memory ratio M/N0.3

Max network size n500

D2D Local Exchange and $\Theta(M/N)$ Scaling

Six users clustered in a D2D range. User 1 requests file

W_4

; neighbor user 4 serves it directly via short-range D2D link. No base station involved. Per-user throughput scales as

\Theta(M/N)

, independent of the total network size

n

— the Ji-Caire-Molisch scaling result.

Example: Scaling at Urban Scale

Compare per-user throughput for two urban scenarios: (a) 100 users in a 1 km² area, $\mu = 0.1$ . (b) 10,000 users in a 1 km² area (100x denser), same $\mu$ . For each, give the scaling-law prediction.

Solution

(a) Low-density

Per-user throughput $\propto M/N = 0.1$ (constant). 100 users in 1 km² means ~10m spacing; D2D range easily reaches ~10 neighbors. Hit probability high; throughput near $\Theta(0.1)$ (normalized to link capacity).

(b) High-density

Per-user throughput still $\propto M/N = 0.1$ . The 100x density increase doesn't improve per-user rate but maintains it. The total aggregate throughput scales 100x (since $n$ scaled 100x), consistent with $T_{\text{agg}} \propto n M/N$ .

Comparison to cellular

In a cellular network, 10000 users share one cell's ~10 Gbps capacity: 1 Mbps per user. In D2D + caching, per-user stays independent of density: potentially many Mbps per user.

Design insight

D2D is the scalable delivery architecture: density helps, not hurts. This justifies cache-aided D2D as a 6G delivery mechanism for dense urban / stadium / airport scenarios.

Key Takeaway

D2D caching achieves $\Theta(M/N)$ per-user throughput independent of $n$ . More users = proportionally more aggregate cache + proportionally more demand + spatial reuse of links. All three scale together, giving flat per-user performance. This is the Ji-Caire-Molisch 2016 result and the theoretical foundation for 6G D2D caching architectures.

Common Mistake: The Scaling Is Model-Dependent

Mistake:

Quoting $\Theta(M/N)$ as a universal D2D + caching rate without stating the model assumptions.

Correction:

The scaling law holds under specific assumptions: (i) random geometric graph / protocol interference model, (ii) random uniform demands, (iii) fixed memory ratio $M/N$ (not fixed $M$ ), (iv) $n \to \infty$ asymptotic regime.

For finite $n$ , the constant factor matters. For Zipf demands, the bound is replaced by a popularity-aware version (Ji-Caire-Molisch 2017 extensions). For physical interference models (SINR-based), constants change. Don't overclaim.

Caching Changes the Scaling Game

The classical wisdom (Gupta-Kumar 2000): "wireless ad-hoc per-user rate $\Theta(1/\sqrt{n})$ ; does not scale." This led to a decade of pessimism about scaling.

The coded caching perspective overturns this: if you store content at users (memory is cheap), you convert transmit traffic into retrieval traffic. Each user already has some of what others need. The network's delivery demand is correspondingly reduced. Per-user rate becomes $\Theta(M/N)$ , constant in $n$ .

This is a recurring theme: caching changes the scaling. In D2D networks, it turns a vanishing throughput into a constant one. In infrastructure, it turns a $1/K$ serial bound into a multiplicative $1 + KM/N$ gain. Whenever memory is abundant at users, caching rescales the fundamental limits.

The D2D Caching Network Model D2D vs Infrastructure Delivery

The Θ(M/N)\Theta(M/N)Θ(M/N) Scaling Law (Ji-Caire-Molisch)