Ferkans — Interactive Telecom Tutor

Handover Becomes Cluster Reselection

Section 23.3 established the SNR gain of coherent macro-diversity across a cluster of $M$ visible satellites. Section 23.4 answered the waveform question. This final section ties the two together into a complete system-level architecture and takes the discussion forward to the 6G NTN research agenda. The central claim is that a cell-free user-centric architecture, adapted from Chapter 12, solves at once the three operational problems of current single-satellite LEO broadband: hard handover every few minutes, cell-edge rain-fade cliffs, and Doppler-driven frequency planning across beams.

In a cell-free LEO network there are no cells. There is no single "serving" satellite. There is a continuously evolving user-centric cluster of $M$ simultaneously-serving satellites, a master-satellite role that rotates as the geometry changes, and a feeder-link network that routes each user's data to every member of its cluster. Handover, in this picture, is not a hard event; it is the quiet rotation of cluster membership as satellites rise above and set below the minimum elevation.

,

Definition:
User-Centric LEO Cluster

Let $\mathcal{V}(t) \subseteq \{1, \ldots, N_{\text{sat}}\}$ be the set of satellites visible to a terminal at time $t$ (elevation above $\theta_{\text{min}}$ ). The user-centric cluster $\mathcal{S}(t) \subseteq \mathcal{V}(t)$ of size $|\mathcal{S}(t)| = M \leq |\mathcal{V}(t)|$ is the set of satellites currently serving the terminal with its downlink payload, selected by Algorithm AUser-Centric LEO Cluster Selection. The cluster evolves in time: as the constellation moves, one satellite in $\mathcal{S}(t)$ falls below $\theta_{\text{min}}$ and is replaced by a newly-rising satellite.

One satellite in the cluster is designated the master, and is responsible for computing the joint precoder and distributing it to the other members over the ISL. Master reselection is infrequent — every few seconds — and is triggered by the current master dropping out of $\mathcal{S}(t)$ .

,

Theorem: Handover Rate vs Constellation Density

Assume a uniform LEO constellation of $N_{\text{sat}}$ satellites in a shell at altitude $h$ , moving at orbital velocity $v_{\text{sat}}$ . A stationary terminal uses a cluster of size $M$ drawn from the visible set $\mathcal{V}(t)$ , where $|\mathcal{V}(t)| = \bar{V}$ on average. The expected cluster- membership change rate (handover rate) is

$\lambda_{\text{HO}} = \frac{M}{\bar{V}} \cdot \frac{v_{\text{sat}}}{d_{\text{visibility}}},$

where $d_{\text{visibility}} = \sqrt{2 R_E h + h^2}$ is the linear scale over which a satellite traverses the visible disc of a ground terminal. For a Starlink-like shell at $h = 550$ km with $\bar{V} \approx 15$ and $M = 4$ , $\lambda_{\text{HO}}^{-1} \approx 9.5$ s between cluster changes — about a factor of $\bar{V}/M = 3.75$ lower than the hard-handover rate of the best-satellite scheme.

If $\bar{V}$ satellites are visible and any cluster reselection event changes $M$ of them (at worst), the expected time between cluster changes is shorter than the mean dwell time of any single satellite by a factor of $\bar{V}/M$ . But because cluster reselection is soft — the departing and arriving satellites both transmit during an overlap window — the effective handover rate experienced by the user is the slower "all members rotate" rate, which is the single-satellite dwell time $\bar{V} / \lambda_{\text{HO}} = d_{\text{visibility}} / v_{\text{sat}}$ , typically several minutes.

Proof

Dwell time of a single satellite

A satellite traverses the visibility disc (diameter $\approx 2 d_{\text{visibility}}$ ) at ground-track speed $v_{\text{sat}}$ . Its dwell time in the visible set is $T_{\text{dwell}} \approx 2 d_{\text{visibility}} / v_{\text{sat}}$ . Every $T_{\text{dwell}}/M$ seconds, one of the $M$ cluster members leaves and is replaced.

Cluster change rate

By a standard renewal argument, the cluster change rate is $\lambda_{\text{HO}} = M / T_{\text{dwell}} = M v_{\text{sat}} / (2 d_{\text{visibility}})$ . The factor of $2$ is absorbed into the $\bar{V}$ normalization in the statement because $\bar{V} \approx 2 (d_{\text{visibility}} / \bar{d})$ where $\bar{d}$ is the mean inter-satellite spacing at altitude $h$ .

Numerical substitution

Plug $h = 550$ km, $v_{\text{sat}} = 7.59$ km/s, $d_{\text{visibility}} = 2720$ km, $\bar{V} = 15$ , $M = 4$ : $\lambda_{\text{HO}} \approx 4 \cdot 7.59/(15 \cdot 2720) \approx 7.4 \times 10^{-4}$ Hz, so $\lambda_{\text{HO}}^{-1} \approx 22$ min — or $\approx 9.5$ s between individual member swaps if we count each cluster rotation. $\blacksquare$

,

Cluster-Reselection Rate vs Constellation Density

Sweep the altitude of the LEO shell and the cluster size $M$ to see the handover rate $\lambda_{\text{HO}}$ . Lower altitude means a denser visible set, shorter dwell times, and more frequent cluster changes — but also shorter propagation delay and more cells. Operators choose $h$ to balance these. Starlink at $550$ km and $M = 4$ sits on the design point where cluster membership turns over every few minutes but the single-user service is continuous.

Parameters

h

(km)550

max cluster size

M

6

\bar{V}

(avg visible)15

Cell-Free LEO Downlink Transmission

Complexity:

\mathcal{O}(M N_t K)

per slot at master

Input: cluster

\mathcal{S}(t) = \{m_1, \ldots, m_M\}

,

per-satellite channels

\mathbf{H}_{m}

, large-scale gains

\beta_{m}

, per-satellite power budget

P_m

,

user data

s_k

for

k = 1, \ldots, K

,

ephemeris of all satellites.

At the master satellite (every slot,

\approx 1

ms):

1. For each visible user

k

:

a. Compute the Doppler pre-compensation phasor

e^{-j 2 \pi \Delta f_D^{(m)} t}

for every

m \in \mathcal{S}(t)

from ephemeris.

b. Assemble the cluster channel matrix

\mathbf{H}_k \in \mathbb{C}^{N_t \times M}

where column

m

is

\sqrt{\beta_{m}^{(k)}} \mathbf{H}_{m}^{(k)}

(after pre-compensation).

c. Compute the per-user cluster precoder

\mathbf{v}_{k} = \mathbf{H}_k (\mathbf{H}_k^H \mathbf{H}_k + \alpha \mathbf{I})^{-1} \mathbf{e}_k

(regularized matched filtering across the cluster).

2. Enforce per-satellite power constraint:

normalize each row of

\{\mathbf{v}_{k}\}

so the

per-satellite transmit power does not exceed

P_m

.

3. Broadcast

\mathbf{v}_{k}

and the user data

s_k

to

every member of

\mathcal{S}(t)

via the ISL or

feeder link.

At each serving satellite

m \in \mathcal{S}(t)

:

1. Apply the precoder restricted to its own array:

\mathbf{x}_m = \sum_k \mathbf{v}_{k, m}\, s_k

.

2. Apply Doppler pre-compensation on transmission.

3. Radiate

\mathbf{x}_m

on the OFDM/OTFS symbol.

On cluster change:

- Add new satellite: master starts distributing user data

and precoders over ISL. New satellite joins the sum

transmission at the next slot boundary.

- Drop departing satellite: its precoder row is zeroed

out; terminal sees a smooth reduction of its

contribution.

The algorithm is a direct transcription of the user-centric cell-free downlink of Chapter 12 with two additions: (a) Doppler pre-compensation using ephemeris, and (b) soft cluster reselection instead of hard handover. The complexity per slot is modest because the master only needs to handle the instantaneous serving users, and the inter-satellite signalling is absorbed into the feeder-link budget.

,

⚠️Engineering Note

ISL-Only vs Gateway-Routed Cell-Free

Two architectural options exist for distributing the user data to all cluster members.

Gateway-routed. Each satellite has its own feeder link to a ground gateway. User data is copied at the gateway and sent independently to every cluster member via separate Ka/Q-band feeder links. Total feeder-link load scales as $M \times$ the single-satellite load. Pros: simple, uses existing ground infrastructure. Cons: expensive in feeder-link spectrum and requires multiple gateways for fault tolerance.
ISL-routed. Satellites are connected via optical inter-satellite links (typically $100$ Gb/s per link, Starlink-style). Only the master satellite receives the user data from the ground; it distributes the data to the other cluster members via ISL. Total feeder-link load stays at the single-satellite level, but the ISL mesh must have enough capacity. Pros: cheap in feeder link, removes single-gateway dependency. Cons: requires optical ISLs to be deployed (Starlink has them; OneWeb does not, as of 2024).

For 6G NTN the trend is toward ISL-routed architectures because (a) optical ISLs are rapidly maturing in both capacity and reliability, (b) they avoid the feeder-link spectrum bottleneck, and (c) they enable in-orbit edge computing and routing that would be awkward to centralize on the ground. The Buzzi-Caire-Colavolpe paper explicitly assumes an ISL-routed architecture for the 6G scenario. For deployment today, the gateway-routed architecture is the realistic starting point.

Practical Constraints

•
Optical ISLs are $\approx 100$ Gb/s and require precise pointing
•
Gateway-routed needs $M\times$ more feeder-link spectrum per user
•
ISL routing depends on in-orbit topology and is sensitive to satellite failures

📋 Ref: 3GPP TR 38.821 (Release 17 NTN), 38.863 (Release 18 NTN enhancements)

,

Example: A 6G NTN Design Point

Consider a 6G NTN target: $K = 10$ simultaneously served users per spot beam at $100$ Mb/s each, $h = 550$ km, Ka-band downlink at $f_0 = 20$ GHz, $N_t = 64$ elements per satellite, and a cluster size $M = 4$ . Derive the per-satellite and per-cluster performance figures, assuming coherent joint transmission and $\text{SNR}$ per link of $10$ dB.

Solution

Per-satellite spectral efficiency

At $\text{SNR} = 10$ dB with ZF cluster precoding and $N_t = 64$ , $K = 10$ , the per-user SINR is approximately $(N_t - K) \cdot \text{SNR} / K = 54 \cdot 10 / 10 = 54$ , i.e. $\approx 17$ dB. The single-user spectral efficiency is $\log_2(1 + 54) \approx 5.8$ bit/s/Hz.

Per-user rate, single satellite

With $W = 100$ MHz allocated per user stream, the per-user rate is $\approx 580$ Mb/s — already exceeding the $100$ Mb/s target on the nominal link.

Effect of macro-diversity

Coherent combining across $M = 4$ satellites gains $10\log_{10}4 = 6$ dB, bringing the effective $\text{SNR}$ to $16$ dB and the spectral efficiency to $\log_2(1 + 256) \cdot \text{ZF penalty} \approx 6.6$ bit/s/Hz. But the real win is on the outage axis: the $99\%$ -reliability rate jumps from $\approx 200$ Mb/s (single sat with rain margin) to $\approx 500$ Mb/s (cluster with diversity), a $2.5\times$ improvement at matched reliability.

System capacity

Across a single spot beam, $K \cdot 500 = 5$ Gb/s of served capacity. Over a constellation of $1000$ satellites with $100$ spot beams each, the aggregate is $5 \cdot 10^5$ Gb/s $= 500$ Tb/s — in the ballpark of what 6G NTN roadmaps advertise. $\blacksquare$

Common Mistake: Scalability Is the Hard Part

Mistake:

A common simulation shortcut is to declare a cell-free LEO system and evaluate it as if every satellite jointly serves every user — i.e. the original Ngo et al. (2017) cell-free formulation transplanted verbatim. Under that assumption the results look fantastic: enormous macro-diversity gain, zero interference, free lunch.

Correction:

The original cell-free formulation has quadratic scaling in the number of APs times users and does not survive realistic constellation sizes. The user-centric variant of Chapter 12, which bounds each user to a small cluster of size $M$ , is the only formulation that scales to $10^4$ satellites and $10^8$ users. When evaluating cell-free LEO proposals, insist on the user-centric version with a finite $M$ and a concrete cluster- update mechanism. The Buzzi-Caire-Colavolpe paper is careful on this point: its architecture is explicitly user-centric from the start, not a naive port of the 2017 formulation.

Historical Note: The 6G NTN Vision

2020s–2030s

The earliest 6G whitepapers, published around 2019–2020 by vendors and national bodies (NTT DOCOMO, Samsung, China's 6G promotion group, and the European Hexa-X project), all place non-terrestrial networks as a first-class pillar alongside terrestrial cells, rather than as an add-on. The motivation is roughly three-fold: global coverage including oceans and wilderness, IoT connectivity for agriculture and environmental monitoring, and resilience in disaster scenarios where terrestrial infrastructure is offline.

The research community has converged on three architectural principles for 6G NTN. First, integrated terrestrial-NTN: the same UE connects opportunistically to either ground cells or satellites, without distinct standards. Second, cell-free operation in orbit: multiple satellites cooperate to serve a user instead of a single "serving cell." Third, on-board processing: regenerative payloads that implement the MAC and part of the PHY on the satellite itself, enabling in-orbit edge computing. The Buzzi-Caire-Colavolpe paper sits squarely in this research programme and provides the information-theoretic and signal-processing foundation for the second principle. Its influence on the emerging 3GPP Release 19 and the eventual Release 20 NTN specifications is the kind of long-arc impact the CommIT group aims for in every part of this library.

,

Why This Matters: NTN Is Also a Sensing Fabric

Chapter 24 will argue that the same massive-MIMO hardware can simultaneously communicate and sense, via Integrated Sensing and Communication (ISAC). In the LEO context, a cell-free cluster of $M$ satellites jointly observing a target area on Earth is a natural $M$ -base-station multistatic radar in addition to being a communication network. The synergy — communication infrastructure doubling as a planetary-scale sensing system — is a further payoff of the cell-free LEO architecture and is the topic of a growing body of literature on "satellite-based ISAC."

Quick Check

In a user-centric cell-free LEO system with cluster size $M = 4$ , a satellite rotates out of the cluster and a new one rotates in. From the terminal's point of view, what does this event look like?

A hard handover — service drops briefly, then the new satellite takes over.

A soft reselection — during an overlap window, both the departing and arriving satellites contribute, and the terminal sees a smooth transition with no service interruption.

A blind handover — the terminal must independently re-acquire the new satellite.

Nothing — the terminal only talks to the master satellite.

Correction:

A soft reselection — during an overlap window, both the departing and arriving satellites contribute, and the terminal sees a smooth transition with no service interruption.

In the user-centric cell-free scheme, cluster reselection is soft: the departing and arriving satellites both transmit during an overlap window of the order of one round-trip time. The terminal sees a gradual shift of the SNR contribution from one satellite to another, not a hard break. This is one of the major operational advantages over the single-satellite architecture, which must interrupt service briefly on every handover.

Key Takeaway

Cell-free LEO closes the architectural loop from Part III. The user-centric ideas of Chapter 12, the distributed processing of Chapter 13, and the fronthaul-aware design of Chapter 14 all translate directly to the LEO setting, with the cluster replacing the AP group and the ISL/feeder network replacing the fronthaul. What the LEO case adds is the time-varying cluster membership, the Doppler pre- compensation, and the choice of waveform. The Buzzi-Caire- Colavolpe framework shows how to put these pieces together into a 6G NTN architecture that out-performs the best-satellite baseline by roughly $6$ dB in SNR and a full order of magnitude in $99\%$ -reliability.

Cell-Free Architectures in Orbit

Handover Becomes Cluster Reselection

Definition: User-Centric LEO Cluster

Theorem: Handover Rate vs Constellation Density

Dwell time of a single satellite

Cluster change rate

Numerical substitution

Cluster-Reselection Rate vs Constellation Density

Parameters

Cell-Free LEO Downlink Transmission

ISL-Only vs Gateway-Routed Cell-Free

Example: A 6G NTN Design Point

Per-satellite spectral efficiency

Per-user rate, single satellite

Effect of macro-diversity

System capacity

Common Mistake: Scalability Is the Hard Part

Historical Note: The 6G NTN Vision

Why This Matters: NTN Is Also a Sensing Fabric

Quick Check

Key Takeaway

Definition:
User-Centric LEO Cluster