Ferkans — Interactive Telecom Tutor

Zero CSIT — But Still Gain?

In the extreme no-CSIT regime, the transmitter knows nothing about the instantaneous channel beyond its statistics. Classical results (Davoodi-Jafar 2016) show that the Gaussian MIMO BC with no CSIT has DoF = 1 — spatial multiplexing collapses entirely. Single-beam multicasting is the best we can do.

Remarkably, coded caching retains a gain even here. The caching mechanism does not use channel information; XOR cancellation at receivers uses only the received signal and local cache. Hence the cache-aided no-CSIT DoF is $t + 1$ , strictly above the no-CSIT pure-MIMO DoF of 1. This "CSIT-free" DoF gain is sometimes called blind interference management: the cache provides a form of side information that removes receiver-side interference without requiring channel knowledge at the transmitter.

Theorem: Cache-Aided No-CSIT DoF

For the cache-aided $L$ -antenna BC with $K$ users and memory ratio $\mu$ , the no-CSIT DoF is $\mathrm{DoF}_{\text{no-CSIT}} \;=\; \min(t + 1, K),$ where $t = K \mu$ . The DoF is achieved by the MAN scheme with single-beam multicast delivery; the $L - 1$ extra antennas provide no DoF improvement without CSIT.

Without CSIT, the transmitter's only option is to broadcast a common signal — one stream, all receivers. The MAN scheme sends XOR messages at this broadcast layer; cached side information lets each XOR simultaneously satisfy $t+1$ users. Spatial multiplexing gain is lost; caching gain is preserved.

Proof

Upper bound

Without CSIT, the transmitter's beamformer is channel-independent; all users receive the same signal up to a scalar $\mathbf{h}_k^H \mathbf{v}$ . Effective channel is a single-stream link of rate $\log_2(1 + \text{SNR})$ . Cut-set bound: $R_{\text{sum}} \leq (t + 1) \log_2(1 + \text{SNR})$ at high SNR (MAN-type converse).

Achievability

Apply the single-antenna MAN scheme: split files into $\binom{K}{t}$ subfiles per MAN placement; broadcast $\binom{K}{t+1}$ XOR messages. Each XOR serves $t+1$ users. Rate in bits per channel use: $(t+1) \log_2(1 + \text{SNR})$ asymptotically. DoF = $t+1$ .

Tight converse

The Lampiris-Elia-Caire '18 paper shows this matches the no-CSIT upper bound. $\blacksquare$

DoF across CSIT Regimes

Bar chart comparing DoF in three CSIT regimes — full, delayed, and none. Cache-aided scheme (blue) retains a significant DoF in all three regimes. Pure MIMO (red) collapses to 1 at no CSIT. The gap between the two schemes grows as CSIT quality degrades — coded caching's value is largest where CSIT is most expensive.

Parameters

Users K10

Antennas L4

Memory ratio M/N0.3

Delayed CSIT: A Middle Regime

Between full CSIT and no CSIT sits the delayed CSIT regime, where the transmitter learns $\mathbf{h}_k$ after its coherence block (channel has already passed). Maddah-Ali-Tse (2012) and Yang et al. (2013) showed that pure MIMO with delayed CSIT achieves DoF $\frac{L(L-1)+L}{L^2} = \frac{2L-1}{L}$ per user... wait, more accurately: retrospective schemes achieve DoF approaching $L$ for $K \to \infty$ , scaling as $\frac{KL}{K+L-1}$ in the symmetric case.

For cache-aided delayed CSIT, the Lampiris-Caire scheme achieves approximately $t + \frac{2L}{L+1}$ per-round DoF, interpolating between the full-CSIT $t+L$ and the no-CSIT $t+1$ . The exact delayed-CSIT characterization for cache-aided BC is the subject of ongoing CommIT research.

Example: High-Mobility Example

A vehicular scenario: 5G NR at $f_c = 3.5$ GHz, $v = 100$ km/h. Coherence time $T_{\text{coh}} \approx 1/f_D \approx 3$ ms. At 1 MBaud symbol rate, $T_c = 3000$ symbols. But CSIT is 2–5 ms old when used — larger than $T_{\text{coh}}$ . Effectively no-CSIT. For $L = 16$ , $K = 20$ , $\mu = 0.2$ , compare DoF with/without caching.

Solution

Without caching (pure MIMO)

Effective CSIT = none. DoF = 1. Per-user DoF = 1/20 = 0.05.

With caching

$t = KM/N = 4$ . No-CSIT DoF = $t+1 = 5$ . Per-user DoF = 5/20 = 0.25.

Gain

5x improvement in DoF from adding caches — entirely due to the coded multicasting mechanism, not the spatial multiplexing the antennas could provide. Pure MIMO in this high-mobility regime is worthless; caching makes the system work.

Design implication

For high-mobility deployments (vehicular, drones, UAVs), coded caching is the primary DoF resource. The 5G NR vehicular extensions (V2X) could in principle benefit from cache-aided multicast, though deployment is in early stages.

⚠️Engineering Note

Deployed Systems That Exploit Blind Gain

The "CSIT-free caching gain" narrative maps to several deployed or near-deployed systems:

5G MBMS (Multimedia Broadcast Multicast Service). Traditional broadcast is CSIT-free; coded caching extensions would substantially improve its efficiency for popular content.
6G / B5G vision. Near-user caching (edge nodes) with multicast delivery is a planned feature in the 3GPP Rel-18+ pipeline.
Content delivery at mmWave. The short coherence time at mmWave frequencies makes CSIT expensive; cache-aided multicast could recover throughput.
LEO satellite broadcasting. Starlink and similar systems face rapid channel aging due to satellite motion; cached content at user terminals compensates.

These systems do not (yet) implement the Lampiris-Caire scheme verbatim. But the principle — rely on cache + multicast rather than CSIT-heavy beamforming — is increasingly accepted in system design.

Practical Constraints

•
3GPP MBMS supports multicast delivery to cached devices
•
mmWave coherence time 20-100 symbols at 30-100 km/h
•
LEO satellite channel ages ~10% within typical coherence block
•
Cache-aided multicast is Rel-18+ study item (as of 2025)

Why This Matters: Connection to Blind Interference Alignment

Classical blind interference alignment (Jafar 2012) designs transmit signals so that interference aligns in receive subspaces without CSIT — purely from receive-side structure (e.g., varying receive filters). The cache-aided no-CSIT scheme plays an analogous role: cached content is the "receive-side structure" that lets XOR cancellation work without transmit-side CSI.

This connection, first formalized by Shariatpanahi-Caire (2017+), puts cache-aided coded caching in the pantheon of CSIT-robust wireless techniques alongside Alamouti coding, space-time codes, and BIA. All share a design philosophy: exploit known receive-side structure rather than demand transmit-side channel knowledge.

Common Mistake: Distinguish Instantaneous vs Statistical CSIT

Mistake:

Calling something "no CSIT" when the transmitter knows the fading distribution (e.g., Rayleigh with known variance).

Correction:

No-CSIT in this chapter means no instantaneous channel knowledge. Statistical knowledge (distribution, variance, spatial correlation) is always assumed — without it, the problem is ill-posed. Practical systems typically have:

Statistical CSIT: always known.
Slow-varying CSIT (path loss, long-term shadowing): easy.
Instantaneous small-scale CSIT: hard.

The cache-aided no-CSIT regime refers to no instantaneous CSIT; statistics are still used for transmit power allocation and outage analysis. The DoF result $\mathrm{DoF} = t + 1$ holds for this definition.

Coded Caching as Blind Interference Management