Distributed vs. Centralized Precoding in Cell-Free Systems: Impact of Realistic Per-AP Power Limits

This paper demonstrates that while centralized precoding in cell-free massive MIMO is theoretically superior, its performance advantage vanishes under realistic per-AP power constraints, rendering distributed precoding a more robust and practical alternative.

The Gist

Imagine a massive orchestra trying to play a perfect symphony for a crowd of 100 people scattered across a park. This is the world of Cell-Free Massive MIMO, a cutting-edge technology designed to give everyone in a city or stadium a perfect, uninterrupted internet connection.

In this orchestra:

• The Musicians are the Access Points (APs)—small antennas scattered everywhere.
• The Audience are the Users (your phone, laptop, etc.).
• The Conductor is the Central Processing Unit (CPU) that tells everyone what to play.

The Big Debate: One Conductor vs. Many Local Leaders

For a long time, engineers believed the best way to run this orchestra was Centralized Precoding. In this scenario, one super-smart Conductor (the CPU) listens to the entire room, calculates exactly how every single musician should play to cancel out noise and hit the perfect note for every audience member, and sends those instructions to everyone.

The theory says this should be perfect. Because the Conductor sees the whole picture, they can create a "beam" of sound that hits every person perfectly while ignoring others.

However, this paper argues that in the real world, this "Super Conductor" approach has a fatal flaw.

The Flaw: The "Power Hungry" Soloist

Here is the problem: In a real city, some musicians are right next to a listener, while others are far away.

When the Super Conductor tries to calculate the perfect sound, they realize that to make the sound reach the person far away without drowning out the person nearby, the distant musician needs to play extremely loudly, while the nearby musician plays very softly.

In the math world, the Conductor says, "Okay, we have a total budget of 100 watts of power for the whole orchestra. Let's give 90 watts to that one distant musician and 10 watts to the rest."

The Reality Check:
In the real world, every musician (AP) has a physical limit. Their amplifier (like a speaker) can only handle so much volume before it breaks or distorts.

• If the Conductor tells a musician to play at 90 watts, but their speaker can only handle 20 watts, the speaker blows up (or the signal gets distorted).
• To fix this, the Conductor has to turn the volume down for everyone so the loudest musician doesn't break. But now, the whole orchestra is playing so quietly that no one can hear the music.

This is what the paper calls the "Power Concentration Effect." The centralized plan demands that a few specific APs do all the heavy lifting, exceeding their hardware limits, which forces a massive reduction in power for the whole system.

The Solution: Distributed Precoding (The Local Leaders)

The paper suggests that instead of one Super Conductor, we should use Distributed Precoding.

In this approach, each musician (AP) is a bit more independent. They only listen to the people sitting right next to them. They don't try to coordinate a perfect global symphony with the whole orchestra. Instead, they just play their own part as best they can for their immediate neighbors.

The Analogy:

• Centralized: One conductor trying to manage 50 musicians from a single podium. They get the math right, but they ask one musician to scream so loud they break their voice, forcing everyone else to whisper.
• Distributed: Each musician manages their own small section. They don't have the perfect global view, but they never ask anyone to break their voice. Everyone plays at a safe, healthy volume.

What the Paper Found

The researchers ran simulations (computer experiments) to see which method actually works better when you respect the physical limits of the speakers.

1. The "Perfect Math" Trap: When they forced the Centralized method to respect the physical limits (by turning down the volume or distorting the instructions), its performance crashed. It lost its "theoretical superiority."
2. The Robust Winner: The Distributed method, which was already known to be "good enough," turned out to be better in the real world. It was robust, didn't break any hardware, and provided a more consistent experience for everyone.

The Takeaway

The paper teaches us a valuable lesson about engineering: Theoretical perfection often fails when it ignores physical reality.

Just because a plan looks perfect on a piece of paper (or a computer screen) doesn't mean it will work in the real world if it asks your equipment to do the impossible. Sometimes, a slightly less "perfect" plan that respects the limits of your tools (Distributed Precoding) is actually the superior choice.

In short: Stop trying to make one musician scream the whole song. Let the whole orchestra play together at a volume they can actually handle.

Technical Summary

Here is a detailed technical summary of the paper "Distributed vs. Centralized Precoding in Cell-Free Systems: Impact of Realistic Per-AP Power Limits."

1. Problem Statement

The paper addresses a critical gap between theoretical performance and practical implementation in Cell-Free (CF) massive MIMO systems.

• Theoretical Context: Information theory suggests that centralized precoding (leveraging system-wide Channel State Information) should significantly outperform distributed precoding (where Access Points act independently) because the set of feasible distributed precoders is a subset of centralized ones.
• The Practical Limitation: Most existing theoretical analyses derive centralized precoders under a sum-power constraint (total system power is fixed and can be arbitrarily distributed among Access Points).
• The Core Issue: In realistic CF deployments, path loss varies drastically between APs and users. To achieve optimal global spatial directivity, centralized algorithms often concentrate transmission power on a small subset of APs closest to a user. This "power concentration effect" forces specific APs to exceed their instantaneous per-AP hardware limits (saturation points of Power Amplifiers).
• The Consequence: When realistic per-AP power constraints are enforced, the theoretical superiority of centralized precoding collapses, potentially making distributed precoding the superior choice.

2. Methodology

The authors analyze the downlink of a CF-mMIMO system with $L$ APs and $K$ single-antenna users, employing a user-centric dynamic clustering strategy.

System Model

• Channel: Spatially correlated Rician fading (including Line-of-Sight and Non-LoS components).
• Estimation: MMSE channel estimation is used, accounting for pilot contamination.
• Strategies:
  • Distributed: Each AP computes local precoding vectors independently.
  • Centralized: A Central Processing Unit (CPU) computes global precoding vectors based on full CSI, which are then split into segments for each AP.

Precoding Design & Normalization

The study focuses on standard closed-form linear precoders (Conjugate Beamforming/MR, Regularized Zero-Forcing/RZF, and MMSE). The core analysis involves how these precoders are normalized to satisfy power constraints:

1. Distributed Normalization: Uses short-term normalization (normalizing the vector norm to 1 for every channel realization) to strictly satisfy the instantaneous per-AP power limit ($P_{max} \leq P_a$).
2. Centralized Normalization (The Conflict):
   • Standard centralized designs use a sum-power constraint (total power $P_s$).
   • When applied to CF, this leads to power concentration, where some APs transmit at levels far exceeding their individual limits ($P_l \gg P_a$).
   • To enforce realistic limits, the authors propose and analyze two heuristic correction methods:
     • Global Power Scaling (PS): If any AP exceeds its limit, all power coefficients are scaled down globally. This preserves spatial directivity but drastically reduces total transmit power.
     • Local Normalization (LN): Each AP's segment of the precoding vector is normalized independently. This satisfies power limits but distorts the intended spatial directivity, destroying the interference cancellation benefits of centralized processing.

3. Key Contributions

• Identification of the "Power Concentration Effect": The paper demonstrates that the theoretical gains of centralized precoding are an artifact of unrealistic sum-power assumptions. In reality, the path-loss disparity in CF systems causes centralized algorithms to demand impossible power levels from specific APs.
• Performance Degradation Analysis: The authors prove that applying simple, low-complexity heuristics (Global Scaling or Local Normalization) to enforce per-AP constraints causes the performance of centralized precoding to drop significantly, often below that of distributed schemes.
• Robustness of Distributed Precoding: The study highlights that distributed precoding, while theoretically suboptimal, is inherently robust to hardware constraints because it naturally respects per-AP power limits without requiring complex re-normalization.
• Evaluation of Closed-Form Solutions: The work focuses on practical, closed-form solutions (ZF, MMSE) rather than intractable iterative optimizations, making the findings relevant for real-time implementation.

4. Simulation Results

Simulations were conducted with 50 APs (4 antennas each) and 10 users under Rician fading and pilot contamination. Key findings include:

• MR (Conjugate Beamforming):
  • Distributed MR with Max-Min power control showed robust performance and high fairness.
  • Centralized MR with Local Normalization (cent-LN) suffered severe Spectral Efficiency (SE) degradation.
  • Centralized MR with Global Scaling (cent-PS) improved fairness slightly but still lagged behind distributed schemes in overall SE.
• RZF (Regularized Zero-Forcing):
  • Distributed RZF consistently achieved higher SE probabilities.
  • Centralized RZF with Local Normalization performed the worst, with CDF curves shifted significantly to the left (lower SE).
• MMSE (Minimum Mean Square Error):
  • Crucial Finding: Contrary to prior literature (e.g., [3]) which claimed centralized MMSE vastly outperforms distributed, this study found distributed MMSE (dist-MM) dominated.
  • The 95%-likely SE for distributed MMSE was 2.38 bps/Hz, whereas centralized MMSE with Local Normalization dropped to 1.23 bps/Hz (Equal Power) and 1.72 bps/Hz (Max-Min).
  • Even with Global Scaling, centralized MMSE failed to outperform distributed schemes in terms of fairness and robustness.

5. Significance and Conclusion

• Paradigm Shift: The paper challenges the prevailing assumption that centralized precoding is always superior in CF-mMIMO. It argues that under realistic instantaneous per-AP power constraints, the complexity and fronthaul overhead of centralized processing are not justified by performance gains.
• Practical Implication: For real-world deployments where APs have strict hardware limits, distributed precoding is the more robust and reliable strategy. It avoids the "power concentration" pitfall and maintains performance without complex, delay-inducing optimization.
• Future Directions: The authors suggest that future research should focus on designing low-complexity precoders specifically aware of per-AP constraints, rather than trying to adapt sum-power-constrained centralized designs.

In summary, the paper concludes that while centralized precoding is theoretically optimal under idealized power assumptions, distributed precoding is the practical winner in cell-free systems due to its inherent compliance with hardware limitations and resilience to power concentration effects.