Diversity-Multiplexing Tradeoff

Can We Have Diversity and Multiplexing Simultaneously?

Sections 15.3--15.5 focused on maximising data rate (capacity, multiplexing gain). But MIMO also offers diversity gain --- protection against fading by transmitting redundant copies across spatial channels. The fundamental question is:

Can we simultaneously achieve high multiplexing gain AND high diversity gain, or must we trade one for the other?

The landmark result of Zheng and Tse (2003) answers this definitively with the diversity-multiplexing tradeoff (DMT), one of the most beautiful results in MIMO theory.

Historical Note: Zheng and Tse: Unifying Diversity and Multiplexing

2000s

Lizhong Zheng and David Tse published "Diversity and Multiplexing: A Fundamental Tradeoff in Multiple-Antenna Channels" in the IEEE Transactions on Information Theory in May 2003. Before this work, the MIMO literature was split into two camps: those pursuing diversity (space-time codes like Alamouti and Tarokh's designs) and those pursuing multiplexing (V-BLAST, Foschini's BLAST).

The DMT unified both perspectives into a single framework, showing that diversity and multiplexing are two endpoints of a continuous tradeoff curve. This paper is one of the most cited in information theory and fundamentally changed how MIMO systems are designed and evaluated.

Definition:

Diversity Gain

The diversity gain (or diversity order) dd of a MIMO scheme characterises how fast the error probability decays with SNR at high SNR:

d=βˆ’lim⁑SNRβ†’βˆžlog⁑Pe(SNR)log⁑SNRd = -\lim_{\text{SNR} \to \infty} \frac{\log P_e(\text{SNR})}{\log \text{SNR}}

A scheme with diversity gain dd has error probability that behaves as Pe∼SNRβˆ’dP_e \sim \text{SNR}^{-d} at high SNR. Higher diversity means steeper decay and more reliable communication.

The maximum achievable diversity for an ntΓ—nrn_t \times n_r channel is dmax⁑=ntβ‹…nrd_{\max} = n_t \cdot n_r (full diversity), achieved for example by the Alamouti code for nt=2n_t = 2.

The notation Pe≐SNRβˆ’dP_e \doteq \text{SNR}^{-d} (exponential equality) means lim⁑log⁑Pelog⁑SNR=βˆ’d\lim \frac{\log P_e}{\log \text{SNR}} = -d, ignoring polynomial and constant factors.

Definition:

Multiplexing Rate in DMT Framework

In the DMT framework, the multiplexing rate rr parameterises a family of rate-SNR operating points:

R(SNR)=rlog⁑2(SNR)bits/s/HzR(\text{SNR}) = r \log_2(\text{SNR}) \quad \text{bits/s/Hz}

where 0≀r≀min⁑(nt,nr)0 \leq r \leq \min(n_t, n_r). As SNR increases, the data rate grows as rr times the log of SNR.

  • r=0r = 0: fixed rate (does not grow with SNR) --- all resources devoted to diversity.
  • r=min⁑(nt,nr)r = \min(n_t, n_r): maximum multiplexing --- the rate grows as fast as channel capacity, leaving no room for extra diversity.

Note that rr need not be an integer. Fractional multiplexing rates correspond to intermediate operating points on the DMT curve.

Theorem: Zheng-Tse Diversity-Multiplexing Tradeoff

For an ntΓ—nrn_t \times n_r i.i.d. Rayleigh fading MIMO channel (block fading with coherence length β‰₯nt+nrβˆ’1\geq n_t + n_r - 1), the optimal diversity-multiplexing tradeoff is the piecewise-linear function connecting the points

(r,dβˆ—(r))forΒ integerΒ r=0,1,…,min⁑(nt,nr)(r, d^*(r)) \quad \text{for integer } r = 0, 1, \ldots, \min(n_t, n_r)

where

dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)d^*(r) = (n_t - r)(n_r - r)

At the extreme points:

  • r=0r = 0: dβˆ—(0)=ntnrd^*(0) = n_t n_r (full diversity)
  • r=min⁑(nt,nr)r = \min(n_t, n_r): dβˆ—(min⁑(nt,nr))=0d^*(\min(n_t, n_r)) = 0 (full multiplexing, zero diversity)

For non-integer rr, dβˆ—(r)d^*(r) is the linear interpolation between adjacent integer points.

Think of the ntΓ—nrn_t \times n_r MIMO channel as having ntβ‹…nrn_t \cdot n_r "resources" (fading coefficients). Allocating rr streams for multiplexing uses up resources, leaving (ntβˆ’r)(nrβˆ’r)(n_t - r)(n_r - r) for diversity protection. The product form arises because diversity requires redundancy across both transmit and receive dimensions simultaneously.

Proof
Achievability (outline)

Zheng and Tse construct a coding scheme based on a min⁑(nt,nr)Γ—nt\min(n_t, n_r) \times n_t space-time code (an extension of the approximate universality result). For integer rr, they show that a scheme transmitting rr data streams with a specific space-time code achieves diversity order exactly (ntβˆ’r)(nrβˆ’r)(n_t - r)(n_r - r).

The key idea is that with rr streams at rate log⁑2(SNR)\log_2(\text{SNR}) each, the outage event {I(H)<R}\{I(\mathbf{H}) < R\} has probability Pr⁑[outage]≐SNRβˆ’(ntβˆ’r)(nrβˆ’r)\Pr[\text{outage}] \doteq \text{SNR}^{-(n_t - r)(n_r - r)}.

Converse (outline)

The converse shows no scheme can exceed dβˆ—(r)d^*(r). The proof analyses the outage probability Pr⁑[log⁑det⁑(I+SNRntHHH)<rlog⁑2(SNR)]\Pr[\log\det(\mathbf{I} + \frac{\text{SNR}}{n_t}\mathbf{H}\mathbf{H}^{H}) < r\log_2(\text{SNR})] by computing the exponential order of this probability using the Wishart eigenvalue distribution.

By a change of variables Ξ»i=SNRβˆ’Ξ±i\lambda_i = \text{SNR}^{-\alpha_i} (exponential parameterisation), the outage event and its probability are characterised in terms of {Ξ±i}\{\alpha_i\}, yielding the optimisation

dβˆ—(r)=infβ‘βˆ‘i(1βˆ’Ξ±i)+<rβˆ‘i(nβˆ’2i+1)Ξ±id^*(r) = \inf_{\sum_i (1-\alpha_i)^+ < r} \sum_i (n - 2i + 1)\alpha_i

whose solution evaluates to (ntβˆ’r)(nrβˆ’r)(n_t - r)(n_r - r) at integer rr. β– \blacksquare

,

The Diversity-Multiplexing Tradeoff

Watch how the DMT curve dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)d^*(r) = (n_t - r)(n_r - r) changes as the antenna configuration varies from 1Γ—11 \times 1 to 8Γ—88 \times 8. Observe how adding antennas expands both the maximum diversity (height at r=0r = 0) and maximum multiplexing (width at d=0d = 0).
The tradeoff frontier expands as the number of antennas grows. Known schemes (Alamouti, V-BLAST, Golden Code) are marked on the curve for the 2Γ—22 \times 2 case.

Diversity-Multiplexing Tradeoff Curve

Plot the optimal DMT curve dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)d^*(r) = (n_t - r)(n_r - r) for different antenna configurations. Compare the tradeoff achievable by specific schemes (Alamouti, V-BLAST, etc.).

Parameters
4
4

Overlay operating points of Alamouti, V-BLAST, etc.

Example: 2Γ—22 \times 2 Diversity-Multiplexing Tradeoff

For a 2Γ—22 \times 2 MIMO channel:

(a) Compute the DMT curve dβˆ—(r)d^*(r) for r=0,1,2r = 0, 1, 2.

(b) What diversity order does V-BLAST achieve at multiplexing gain r=2r = 2?

(c) What diversity order does the Alamouti code achieve at r=1r = 1?

(d) Is there a scheme that achieves the optimal DMT at all points?

Solution
Compute DMT at integer points

Using dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)=(2βˆ’r)2d^*(r) = (n_t - r)(n_r - r) = (2-r)^2:

rr dβˆ—(r)=(2βˆ’r)2d^*(r) = (2-r)^2 Interpretation
0 4 Full diversity (no multiplexing)
1 1 Balanced tradeoff
2 0 Full multiplexing (no diversity)

The DMT curve is the piecewise-linear interpolation: dβˆ—(r)=(2βˆ’r)2d^*(r) = (2-r)^2 for r∈[0,2]r \in [0, 2].

V-BLAST at $r = 2$

V-BLAST transmits 2 independent streams, achieving r=2r = 2. The DMT says dβˆ—(2)=0d^*(2) = 0, and V-BLAST with MMSE-SIC or ML detection achieves this --- it has zero diversity gain. In fact, V-BLAST with ZF detection has d=0d = 0 at r=2r = 2, confirming it is DMT-optimal at the full-multiplexing point.

Alamouti at $r = 1$

The Alamouti code for nt=2n_t = 2 achieves diversity order d=2nr=4d = 2n_r = 4 at rate R=log⁑2(1+SNR)R = \log_2(1 + \text{SNR}), which corresponds to r=1r = 1. But dβˆ—(1)=1d^*(1) = 1, not 4!

Wait --- the Alamouti code actually achieves d=2nr=4d = 2n_r = 4 which is the full diversity ntnr=4n_t n_r = 4, but at multiplexing gain r=1r = 1 the optimal diversity is only 1. The resolution: Alamouti achieves d=4d = 4 at a fixed rate (r=0r = 0 in the DMT sense, since the rate does not grow as log⁑2(SNR)\log_2(\text{SNR})). At r=1r = 1, Alamouti's diversity drops to d=1d = 1, matching the DMT. β– \blacksquare

DMT-optimal scheme

For 2Γ—22 \times 2, the approximately universal (AU) space-time codes (e.g., the Golden code) achieve the optimal DMT dβˆ—(r)=(2βˆ’r)2d^*(r) = (2-r)^2 at all points r∈[0,2]r \in [0, 2]. These codes achieve both full diversity and full rate. β– \blacksquare

Quick Check

For a 3Γ—23 \times 2 MIMO channel, what is the maximum achievable diversity gain dβˆ—(0)d^*(0)?

dβˆ—(0)=ntβ‹…nr=6d^*(0) = n_t \cdot n_r = 6

dβˆ—(0)=min⁑(nt,nr)=2d^*(0) = \min(n_t, n_r) = 2

dβˆ—(0)=nt+nrβˆ’1=4d^*(0) = n_t + n_r - 1 = 4

dβˆ—(0)=(ntβˆ’1)(nrβˆ’1)=2d^*(0) = (n_t - 1)(n_r - 1) = 2

Correction:
dβˆ—(0)=ntβ‹…nr=6d^*(0) = n_t \cdot n_r = 6

At r=0r = 0 (fixed rate, full diversity): dβˆ—(0)=(ntβˆ’0)(nrβˆ’0)=3Γ—2=6d^*(0) = (n_t - 0)(n_r - 0) = 3 \times 2 = 6.

Common Mistake: DMT Does Not Apply at Fixed Rate

Mistake:

Applying the DMT formula to evaluate a scheme that operates at a fixed data rate (constant in SNR), claiming it has "diversity gain dβˆ—(r)d^*(r)" for some r>0r > 0.

Correction:

The DMT characterises the tradeoff for schemes whose rate scales as rlog⁑2(SNR)r\log_2(\text{SNR}). A fixed-rate scheme has r=0r = 0 in the DMT framework and can potentially achieve the full diversity dβˆ—(0)=ntnrd^*(0) = n_t n_r. The DMT is a high-SNR asymptotic framework; for finite SNR analysis, use exact PEP (pairwise error probability) bounds instead.

Common Mistake: DMT Requires Sufficient Coherence Time

Mistake:

Applying the Zheng-Tse DMT result to channels with very short coherence time (fast fading) without checking the coherence requirement.

Correction:

The Zheng-Tse DMT requires the coherence length Tβ‰₯nt+nrβˆ’1T \geq n_t + n_r - 1 symbols. For shorter coherence times, the available DoF are reduced (some dimensions must be spent on training/channel estimation), and the DMT curve changes. Zheng and Tse's "non-coherent DMT" shows that with unknown CSI, the DoF drops to mβˆ—(1βˆ’mβˆ—/T)m^*(1 - m^*/T) where mβˆ—=min⁑(nt,nr,⌊T/2βŒ‹)m^* = \min(n_t, n_r, \lfloor T/2 \rfloor).

DMT-Achieving MIMO Schemes

SchemeRate rrDiversity ddDMT-optimal?Complexity
Repetition coding0ntnrn_t n_rYes (at r=0r=0)Low
Alamouti (nt=2n_t=2)12nr2n_rOnly at r=0r=0Low
V-BLAST (ZF)min⁑(nt,nr)\min(n_t,n_r)0Only at r=min⁑(nt,nr)r=\min(n_t,n_r)Moderate
V-BLAST (ML)min⁑(nt,nr)\min(n_t,n_r)0Only at r=min⁑(nt,nr)r=\min(n_t,n_r)High
Golden Code (2Γ—22\times 2)0≀r≀20 \leq r \leq 2(2βˆ’r)2(2-r)^2Yes (all rr)High
CDA codes0≀r≀min⁑(nt,nr)0 \leq r \leq \min(n_t,n_r)(ntβˆ’r)(nrβˆ’r)(n_t-r)(n_r-r)Yes (all rr)Very high
πŸŽ“CommIT Contribution(2006)

DMT-Optimal Code Constructions

P. Elia, K. R. Kumar, S. A. Pawar, P. V. Kumar, G. Caire β€” IEEE Transactions on Information Theory, vol. 52, no. 9, pp. 3869--3884

While Zheng and Tse established the optimal DMT curve dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)d^*(r) = (n_t - r)(n_r - r), they did not provide explicit code constructions achieving it for general ntΓ—nrn_t \times n_r. Elia, Kumar, Pawar, Kumar, and Caire resolved this by constructing explicit algebraic space-time codes based on cyclic division algebras (CDA) that achieve the optimal DMT for any number of antennas.

Their construction uses the theory of central simple algebras over number fields to design space-time code matrices with the non-vanishing determinant (NVD) property: det⁑(X1βˆ’X2)β‰₯Ξ΄>0\det(\mathbf{X}_1 - \mathbf{X}_2) \geq \delta > 0 for all distinct codewords X1,X2\mathbf{X}_1, \mathbf{X}_2. This NVD property is both necessary and sufficient for DMT optimality.

A related construction by El Gamal, Caire, and Damen (the LAST codes) achieves the DMT using lattice-based space-time codes, providing an alternative algebraic framework.

DMTspace-time-codesCDAalgebraic-codesView Paper β†’
πŸŽ“CommIT Contribution(2004)

LAST Codes: Lattice Space-Time Codes for the DMT

H. El Gamal, G. Caire, M. O. Damen β€” IEEE Transactions on Information Theory, vol. 50, no. 6, pp. 968--985

El Gamal, Caire, and Damen showed that lattice space-time (LAST) codes achieve the optimal DMT of the MIMO channel. The key insight is that lattice codes, when combined with minimum-distance (lattice) decoding, naturally provide both the coding gain needed for diversity and the rate scaling needed for multiplexing.

The LAST code framework provides:

  • An explicit construction based on algebraic number theory
  • A connection between the Hermite constant of the lattice and the diversity gain
  • A practical decoding approach via lattice reduction and successive cancellation

This work, together with the CDA codes of Elia et al., closed the gap between the information-theoretic DMT and practical code design.

DMTlattice-codesLASTspace-time-codesView Paper β†’
⚠️Engineering Note

MIMO Layer Limitations in 5G NR

While the DMT and DoF results assume arbitrary code design, practical MIMO systems are constrained by standards and hardware:

  • 5G NR downlink: maximum 8 MIMO layers (streams) per UE, even with 32 or 64 antenna ports at the gNB. The bottleneck is UE antenna count and RF chain cost.
  • 5G NR uplink: maximum 4 MIMO layers per UE. Most UEs support only 1--2 uplink layers due to power amplifier constraints.
  • Rank adaptation: the gNB dynamically selects the number of layers (1 to 8) based on channel rank, SNR, and UE capability. The rank indicator (RI) is fed back by the UE every 5--80 ms.
  • Codebook constraints: practical precoding uses finite codebooks (Type I: 1 beam; Type II: linear combination of beams), not the ideal SVD precoder. The codebook quantisation loss is typically 1--3 dB relative to perfect CSIT.
  • Receiver complexity: ML detection is infeasible for more than ∼4\sim 4 layers. 5G NR UEs use MMSE-IRC (interference rejection combining), which is suboptimal but practical.
Practical Constraints
  • β€’

    5G NR: max 8 DL layers, max 4 UL layers per UE

  • β€’

    Codebook quantisation loss: 1-3 dB vs perfect SVD precoding

  • β€’

    ML detection infeasible beyond ~4 layers; MMSE-IRC used in practice

  • β€’

    Rank adaptation latency: 5-80 ms RI feedback period

πŸ“‹ Ref: 3GPP TS 38.214, Β§5.2

Diversity gain

The negative exponent of the error probability decay with SNR: d=βˆ’lim⁑log⁑Pelog⁑SNRd = -\lim \frac{\log P_e}{\log \text{SNR}}. Higher diversity means more reliable communication.

Related: Multiplexing gain, Diversity-multiplexing tradeoff (DMT)

Diversity-multiplexing tradeoff (DMT)

The optimal tradeoff curve dβˆ—(r)=(ntβˆ’r)(nrβˆ’r)d^*(r) = (n_t - r)(n_r - r) between diversity gain and multiplexing rate for an ntΓ—nrn_t \times n_r MIMO channel, established by Zheng and Tse (2003).

Related: Diversity gain, Multiplexing gain

Degrees of Freedom and Multiplexing GainSummary