You are on page 1of 7

LMMSE ( Linear Minimum Mean

Square Estimation) Detection


Ref: Jerry R. Hampton, “Introduction to MIMO Communications”,
Cambridge University Press 2014
( Section 1.9.2 and 8.3.3)
LMMSE Detection
The LMMSE detection method is based on the assumption that each
transmitted signal can be estimated at the receiver by linearly combining
weighted versions of the received signals.
We consider a MIMO system of 𝑁𝑡 transmit and 𝑁𝑟 receive antennas.
For the 𝑘𝑡ℎ transmission,
𝐬ො 𝑘 = Wො𝐫 𝑘
For 𝑝 number of transmissions,
S෠ 𝑘 = WR
In LMMSE, we try to find the matrix W that minimizes the mean square error between S and S෠
Let W0 be the optimal matrix.
2
W0 ≜ arg min E S − S෠ = arg min E S − WR 2𝐹
𝐹
W W

𝐻
W0 = arg min E Tr[ S − WR S − WR ] = arg min E Tr[SS𝐻 − SR𝐻 W𝐻 − WRS 𝐻 + WRR𝐻 W𝐻 ]
W W
To find the value of 𝑊 that minimizes the RHS of the equation,
𝜕 𝜕 𝜕 𝜕
𝐸 Tr SS 𝐻 − Tr SR𝐻 W𝐻 − Tr WRS 𝐻 + Tr WRR𝐻 S𝐻 =0
𝜕W 𝜕W 𝜕W 𝜕W

SS 𝐻 is not a function of W, so the first term is zero. Using the matrix relations (given as appendix) , the
equation above can be reduced to:
E RR𝐻 W𝐻 = E[RS 𝐻 ]
Substituting R = 𝜌HS + Z and expanding and evaluating the expectation, we get,
E {ρHSS𝐻 H 𝐻 + ZZ 𝐻 }W𝐻 = E[ 𝜌HSS𝐻 ]
We make the following assumptions:

H is fixed over a frame length of 𝑝 modulation symbols


Transmitted symbols and noise are uncorrelated
E SS𝐻 = 𝜎𝑠2 𝐼𝑁𝑡
E ZZ𝐻 = 𝜎𝑧2 𝐼𝑁𝑟

For normalization, 𝜎𝑠2 = 1Τ𝑁𝑡 and 𝜎𝑧2 = 1

The equation E {ρHSS𝐻 H 𝐻 + ZZ 𝐻 }W𝐻 = E[ 𝜌HSS𝐻 ] can be written as


2 𝐻 𝐻
𝜌
𝜌𝜎𝑠 𝐼𝑁𝑡 HH + 𝐼𝑁𝑟 W = H
𝑁𝑡

𝜌 𝐻 𝐻
𝜌
HH + 𝐼𝑁𝑟 W = H
𝑁𝑡 𝑁𝑡
−1
𝜌 𝐻 𝜌
W0 = H HH 𝐻 + 𝐼𝑁𝑟
𝑁𝑡 𝑁𝑡
Alternatively,
−1
𝜌 𝜌 𝐻
W0 = H H + 𝐼𝑁𝑡 H𝐻
𝑁𝑡 𝑁𝑡
In LMMSE detection, the received matrix R is pre multiplied by 𝑊0

෩ = 𝑊0 R = 𝜌𝑊0 HS + 𝑊0 𝑍 = 𝜌𝑊0 HS + 𝑍෩
R
෩ can be written as
The 𝑘𝑡ℎ element of the 𝑗𝑡ℎ row of R

𝑁𝑟 𝑁𝑡

𝑟෥𝑗 𝑘 = 𝜌 ෍ 𝑤𝑗𝑚 ෍ ℎ𝑚𝑛 𝑠𝑛 𝑘 + 𝑧෥𝑗 (𝑘)


𝑚=1 𝑛=1
This can be written as

𝑁𝑟 𝑁𝑟 𝑁𝑡

𝑟෥𝑗 𝑘 = 𝜌 ෍ 𝑤𝑗𝑚 ℎ𝑚𝑗 𝑠𝑗 𝑘 + 𝜌 ෍ 𝑤𝑗𝑚 ෍ ℎ𝑚𝑛 𝑠𝑛 𝑘 + 𝑧෥𝑗 (𝑘)


𝑚=1 𝑚=1 𝑛≠𝑗
Therefore, we can write
𝑁 𝑁
𝜌 σ𝑚=1
𝑟
𝑤𝑗𝑚 σ𝑛≠𝑗
𝑡
ℎ𝑚𝑛 𝑠𝑛 𝑘 𝑧෥𝑗 (𝑘)
𝑟ෝ𝑗 𝑘 = 𝑠𝑗 𝑘 + +
𝜌 σ𝑁
𝑚=1 𝑤𝑗𝑚 ℎ𝑚𝑗
𝑟 𝑁𝑟
𝜌 σ𝑚=1 𝑤𝑗𝑚 ℎ𝑚𝑗

Signal Noise
interference

2
𝑠ෝ𝑗 (k)= arg min 𝑟ෝ𝑗 𝑘 − 𝑠𝑗 𝑘
𝑠𝑗 (𝑘)
𝑗 = 1, 2, ⋯ 𝑁𝑡 𝑘 = 1, 2, ⋯ 𝑝

This algorithm, unlike ZF, does not suffer from noise amplification.
Appendix

You might also like