User scheduling in MIMO systems decides which users transmit in each time-frequency slot and, in MU-MIMO, which users share it. A scheduler that uses channel state information can exploit fading: with 10 users at 10 dB mean SNR, serving the strongest user at each instant raises throughput by about 65% over taking turns. Proportional fair scheduling keeps most of that gain while still serving weak users, and semi-orthogonal user selection extends the idea to several users at once.
This page compares the main scheduling rules with simulated numbers. Channel state information covers how the CSI reaches the scheduler, and MIMO precoding covers how the chosen users are separated.
What user scheduling in MIMO systems does
At each scheduling interval, typically about 1 ms or less in cellular systems, the base station has a list of users with data waiting and a recent quality report from each. It selects users, assigns resource blocks, picks a rank and a modulation and coding scheme, then sends. The aim is to balance three things that pull apart: total throughput, fairness between users and latency.
The key idea is that fading is not only a problem. If each user’s channel fades independently, then at any moment one of them is probably near a peak. This is multiuser diversity.
Multiuser diversity
Assume K users with the same mean SNR, in Rayleigh fading. Each instantaneous SNR is exponentially distributed, and the average of the largest of K such values is:
E[max] = 1 + 1/2 + 1/3 + … + 1/K
For K = 10 that sum is 2.93, a gain of 4.7 dB in SNR with no extra power or antennas.
Worked example: 10 users at 10 dB
At 10 dB mean SNR, a single user averages 2.91 b/s/Hz in Rayleigh fading. Note that this is below the 3.46 b/s/Hz of a steady 10 dB channel, because of the concavity of the log. A max-rate scheduler serving the best of 10 users averages 4.81 b/s/Hz. That is a 65% gain, and it comes entirely from choosing when to transmit.
Multiuser diversity needs real fading. In a static channel, or with users in line of sight, there are no peaks to ride. It also interacts with massive MIMO: channel hardening removes the fading that the scheduler would otherwise exploit, so the gain from selecting users moves from timing towards spatial separation.
Round robin, max rate and proportional fair
Three rules define the range of choices.
- Round robin serves users in turn and ignores the channel. It is simple and needs no CSI, and gives equal time but not equal throughput.
- Max rate serves the user with the highest instantaneous rate Rk(t). It gives the highest throughput but starves weak users.
- Proportional fair (PF) serves the user with the highest ratio of rate to average throughput, Rk(t)/Ak(t).
The PF average is updated each slot, with a time constant tc:
Ak(t+1) = (1 – 1/tc) Ak(t) + (1/tc) Rk(t) if user k is served
A user is therefore scheduled when it is near its own peak, not when it is the strongest in the cell. A weak user at a peak beats a strong user at its average. The rule also has a formal meaning: it maximises the sum of the logs of user throughputs.
Worked example: five users with unequal SNR
Real cells have large differences in mean SNR. This simulation has five users at 20, 15, 10, 5 and 0 dB, in Rayleigh fading, with tc = 100 slots. The Jain fairness index is (Σx)2 / (nΣx2), which is 1 for perfect equality.
| Scheduler | User 1 (20 dB) | User 2 | User 3 | User 4 | User 5 (0 dB) | Total | Jain index |
|---|---|---|---|---|---|---|---|
| Round robin | 1.18 | 0.87 | 0.58 | 0.34 | 0.17 | 3.14 | 0.75 |
| Max rate | 4.86 | 1.24 | 0.16 | 0.01 | 0.00 | 6.27 | 0.31 |
| Proportional fair | 1.60 | 1.23 | 0.88 | 0.57 | 0.31 | 4.59 | 0.80 |
All rates are in b/s/Hz. Compared with round robin, PF gives every user more, and raises the total by 46%. The 0 dB user almost doubles, from 0.17 to 0.31 b/s/Hz. Max rate has a higher total, 6.27 b/s/Hz, but user 4 and user 5 are effectively cut off. This is why PF, or a variant with added quality-of-service weights, is the usual starting point.
User selection in MU-MIMO
With M base station antennas, up to M users can share a slot, which is a different kind of scheduling. Rates now depend on the set of users, not on each user alone: two users with nearly parallel channels interfere, even if both are strong. ZF has to remove that overlap, and doing so costs gain.
The aim is to pick users who are strong and close to orthogonal. Testing all subsets takes far too long, so practical methods are greedy:
- Start with the strongest user.
- For each remaining user, estimate the sum rate if it is added.
- Add the best one, and stop when no user raises the sum rate.
Semi-orthogonal user selection (SUS) is a popular simplification. It keeps only users whose channels lie within a set angle of orthogonal to those already chosen, then picks the strongest of them. Cellular systems often use a table of user pairings based on reported beams, which is cheaper still.
Worked example: four antennas with up to 64 users
This simulation uses M = 4 and ZF with equal power, with a greedy search across K candidates:
| Candidates K | Random 4 users, 20 dB | Greedy, 20 dB | Random 4 users, 5 dB | Greedy, 5 dB |
|---|---|---|---|---|
| 4 | 16.3 | 19.4 | 2.9 | 6.0 |
| 8 | 16.1 | 23.1 | 3.0 | 7.0 |
| 16 | 16.2 | 25.5 | 2.8 | 7.9 |
| 32 | 16.0 | 27.1 | 2.9 | 8.8 |
| 64 | 16.1 | 28.2 | 2.9 | 9.6 |
At 20 dB, selecting from 8 candidates adds 43% over a random group, and 64 candidates add 75%. At 5 dB, the gain is over 100% even with few candidates, partly because the greedy search can choose fewer than four streams. Random full loading at low SNR wastes power on weak, poorly separated users. For comparison, serving only the single best user with all four antennas gives 9.2 to 10.0 b/s/Hz at 20 dB, so MU-MIMO with good selection delivers nearly three times that.
Scheduling and CSI quality
Every scheduler depends on CSI, and its quality limits the gain in three ways.
- Feedback overhead. Reporting every user’s channel costs uplink capacity. Threshold reporting, where only users above a level report, cuts this with little loss, because the scheduler only wants the peaks anyway.
- CSI delay. Reports are several slots old when used. Fast fading can move in that time, so peak-riding is most effective for slow users.
- Estimation error. MU-MIMO needs accurate nulls. Channel estimation errors shows how quickly sum rate falls, and deep learning CSI feedback looks at compressing the reports.
Power allocation also interacts with user choice, as shown in CSI-based power allocation.
Frequently asked questions
What is user scheduling in MIMO?
It is the decision of which users get each time-frequency resource, and which of them share it spatially. It sits above precoding and uses CSI and quality reports to balance throughput against fairness.
What is proportional fair scheduling?
It serves the user with the highest ratio of current rate to average throughput, so users transmit near their own channel peaks. In the five-user example, it improved every user compared with round robin, with a total of 4.59 against 3.14 b/s/Hz.
What is multiuser diversity?
It is the throughput gain from having many independently fading users to choose among. With 10 users, the average best-user SNR is 4.7 dB above the mean, and the rate gain in the example was 65%.
How many users can MU-MIMO serve at once?
At most about as many streams as there are base station antennas, and usually fewer. Pairing compatible users matters more than filling every stream, particularly at low SNR.
Related topics
Massive MIMO spectral efficiency shows how many users a large array can carry, and beamforming techniques in cellular systems shows how reports are produced. MIMO capacity gives the single-user limit.