Unit 7.

Channel Estimation and

Equalization
ECE-GY 6023. WIRELESS COMMUNICATIONS
PROF. SUNDEEP RANGAN

1
Learning Objectives
Describe role of channel estimation in the receiver steps
Describe role and configuration of reference signals in OFDM
◦ Compute overhead
◦ Qualitatively relate density to the coherence time and bandwidth

Implement and simulate simple kernel-type channel estimators

Compute bias and variance of the kernel estimators
Optimize the kernel to minimize the bias + variance error
Compute the signal to distortion and noise ratio (SNDR) for a given estimator
Measure the SNDR via simulation
Implement a simple channel estimator and equalizer for 5G NR downlink

2
Outline
Role of Channel Estimation and Equalization
OFDM Channel Estimation via Kernel Regression
Bias-Variance Tradeoff

3
 Channel Estimation and Equalization
𝑟𝑟[𝑛𝑛, 𝑘𝑘]
𝑥𝑥[𝑛𝑛, 𝑘𝑘] h[𝑘𝑘, 𝑛𝑛] 𝑥𝑥� 𝑛𝑛, 𝑘𝑘 ≈ 𝑥𝑥[𝑛𝑛, 𝑘𝑘]

Fading channel ℎ[𝑛𝑛, 𝑘𝑘] Equalization

Channels fade in time and frequency

◦ In OFDM, can be described by time-varying frequency response: 𝑟𝑟 𝑛𝑛, 𝑘𝑘 = ℎ 𝑛𝑛, 𝑘𝑘 𝑥𝑥 𝑛𝑛, 𝑘𝑘 + 𝑤𝑤[𝑛𝑛, 𝑘𝑘]

Equalizers: Estimate TX symbols from faded symbols

𝑟𝑟[𝑛𝑛,𝑘𝑘] ℎ 𝑛𝑛,𝑘𝑘 ∗ 𝑟𝑟[𝑛𝑛,𝑘𝑘]
◦ Approximately invert the channel 𝑥𝑥� 𝑘𝑘, 𝑛𝑛 ≈ or MMSE 𝑥𝑥� 𝑘𝑘, 𝑛𝑛 =
ℎ[𝑛𝑛,𝑘𝑘] ℎ 𝑛𝑛,𝑘𝑘 2 +𝑁𝑁0

Key challenge: Equalizers need a channel and/or noise estimate

4
 Reference Symbols
TX reference RX reference
𝑥𝑥[𝑛𝑛, 𝑘𝑘] ℎ 𝑛𝑛, 𝑘𝑘 𝑥𝑥[𝑛𝑛, 𝑘𝑘]

h[𝑛𝑛, 𝑘𝑘] 𝑔𝑔[𝑛𝑛, 𝑘𝑘]

Fading channel Equalizer estimator

Equalizers require training or reference signals

TX sends a known reference sequence 𝑥𝑥[𝑛𝑛, 𝑘𝑘] on some locations (𝑘𝑘, 𝑛𝑛)
RX sees 𝑟𝑟 𝑛𝑛, 𝑘𝑘 = ℎ 𝑛𝑛, 𝑘𝑘 𝑥𝑥 𝑛𝑛, 𝑘𝑘 + 𝑤𝑤 𝑛𝑛, 𝑘𝑘
Estimates channel ℎ[𝑛𝑛, 𝑘𝑘] and/or equalizer 𝑔𝑔[𝑛𝑛, 𝑘𝑘] for 𝑟𝑟[𝑛𝑛, 𝑘𝑘]

5
Receiver Steps in OFDM
Data-Reference
demux Equalization
Equalized
Data symbols symbols LLRs

Sync & OFDM Demodulator Decoder

AGC FFT

Reference
symbols
Channel & noise
estimate

Today’s unit

6
Frequency-Domain and Time-Domain EQ
Frequency-Domain Equalization:
◦ Typically, an OFDM channel 𝑟𝑟 𝑛𝑛, 𝑘𝑘 = ℎ[𝑛𝑛, 𝑘𝑘]𝑥𝑥 𝑛𝑛, 𝑘𝑘 + 𝑤𝑤[𝑛𝑛, 𝑘𝑘]
◦ Estimate frequency-domain channel ℎ[𝑛𝑛, 𝑘𝑘]
◦ Perform equalization with a filter 𝑥𝑥� 𝑛𝑛 = 𝑔𝑔[𝑛𝑛, 𝑘𝑘]𝑟𝑟 𝑛𝑛, 𝑘𝑘
◦ Used in 4G, 5G OFDM systems, most 802.11 versions

Time-Domain Equalization:
◦ Model channel in time-domain 𝑟𝑟 𝑛𝑛 = ∑𝑘𝑘 ℎ[𝑘𝑘] 𝑥𝑥 𝑛𝑛 − 𝑘𝑘 + 𝑤𝑤[𝑛𝑛]
◦ Estimate channel impulse response
◦ Perform equalization with a filter 𝑥𝑥� 𝑛𝑛 = ∑𝑘𝑘 𝑔𝑔[𝑘𝑘] 𝑟𝑟 𝑛𝑛 − 𝑘𝑘
◦ Used in 2G, 3G CDMA systems, 802.11ad single carrier version
◦ Do not discuss in this unit (except next few slides)

7
 802.11 ad Time-Domain Equalizer

Apply equalizer
802.11ad uses time-domain equalization
STF: Short training field
◦ AGC, detection of packet
◦ coarse estimate of timing CEF: Channel estimation field
◦ Process as previous lecture ◦ Train equalizer

8
802.11ad Preamble Details
Control packets
Longer STF.
Need to be decoded by all
stations. No directional gain

Data packets
Shorter STF

Based on complementary Golay codes [𝐺𝐺𝐺𝐺, 𝐺𝐺𝐺𝐺]

Have low auto-correlation
1
 = 1.76 Gsamp/S
𝑇𝑇𝑐𝑐
For data packets: STF = 1.23 𝜇𝜇s, CE= 0.654 𝜇𝜇s

9
Outline
Role of Channel Estimation and Equalization
OFDM Channel Estimation via Kernel Regression
Bias-Variance Tradeoff

10
OFDM Channel Estimation
OFDM Channel

𝑅𝑅 𝑛𝑛, 𝑘𝑘 = 𝐻𝐻 𝑛𝑛, 𝑘𝑘 𝑋𝑋 𝑛𝑛, 𝑘𝑘 + 𝑊𝑊[𝑛𝑛, 𝑘𝑘]

◦ Symbol 𝑘𝑘 and sub-carrier 𝑛𝑛

Channel 𝐻𝐻[𝑛𝑛, 𝑘𝑘] varies

◦ Over time 𝑘𝑘 due to Doppler spread
◦ Over frequency 𝑛𝑛 due to delay spread

Problem: Estimate 𝐻𝐻[𝑛𝑛, 𝑘𝑘]

Estimate needed to demodulate data 𝑋𝑋[𝑛𝑛, 𝑘𝑘]
51 RB, 15 kHz SCS, ~10 MHz bandwidth NR slot
500 ns delay spread, 200 Hz Doppler, ±60∘ AoA

11
Reference Signals
OFDM Channel
𝑅𝑅 𝑛𝑛, 𝑘𝑘 = 𝐻𝐻 𝑛𝑛, 𝑘𝑘 𝑋𝑋 𝑛𝑛, 𝑘𝑘 + 𝑊𝑊[𝑛𝑛, 𝑘𝑘]

5G NR Demodulation Reference Signals

◦ Called DM-RS
◦ Transmitted in downlink with data (PDCSH)
◦ Very configurable in terms of over-head

Basic channel estimation idea

◦ Transmit know 𝑋𝑋[𝑛𝑛, 𝑘𝑘] on DM-RS.
�0 𝑛𝑛, 𝑘𝑘 ≈ 𝑅𝑅 𝑛𝑛, 𝑘𝑘 ⁄𝑋𝑋 𝑛𝑛, 𝑘𝑘 on DM-RS symbols
◦ Estimate 𝐻𝐻
◦ Interpolate to data symbols

12
Configuring in DM-RS
Example configuration
◦ DM-RS on one symbol in RB
◦ Occupies 6 out of 12 REs

Also added a phase-tracking reference signal

◦ Tracks phase variations over time
◦ Useful for carrier frequency offset and
phase noise (more on this later)

Many configurations possible

See excellent MATLAB 5G tutorial with videos

13
 Configuring in MATLAB
MATLAB 5G Toolbox
◦ Provides all config possibilities
◦ Easy to use

See excellent MATLAB 5G tutorial

with videos:

14
Estimating Channel in a Single Symbol
First consider estimation in one symbol
True channel is ℎ[𝑛𝑛]
◦ 𝑛𝑛 = sub-carrier index

RX: 𝑟𝑟 𝑛𝑛 = ℎ 𝑛𝑛 𝑥𝑥 𝑛𝑛 + 𝑤𝑤 𝑛𝑛

Subcarrier
DM-RS 𝑥𝑥[𝑛𝑛] known
Example to right:
◦ 10 MHz bandwidth
◦ 200 ns delay spread
◦ 10 dB SNR on DM-RS

15
Raw Channel Estimate
Channel: 𝑟𝑟 𝑛𝑛 = ℎ 𝑛𝑛 𝑥𝑥 𝑛𝑛 + 𝑤𝑤 𝑛𝑛
On DM-RS positions, 𝑥𝑥[𝑛𝑛] is known
Raw channel estimate:

𝑟𝑟 𝑛𝑛
ℎ� 0 𝑛𝑛 =
𝑥𝑥[𝑛𝑛]

◦ Can be computed in every DM-RS position

We see estimate is noisy

16
Raw Channel Estimate Error
Drop sub-carrier index (since we are looking at one RE)
2
Channel: 𝑟𝑟 = ℎ𝑥𝑥 + 𝑤𝑤, 𝑤𝑤~𝐶𝐶𝐶𝐶 0, 𝑁𝑁0 , 𝑥𝑥 = 𝐸𝐸𝑥𝑥
𝑟𝑟 𝑤𝑤
Raw channel estimate: ℎ� 0 = =ℎ+
𝑥𝑥 𝑥𝑥
Error is
2 1 𝑁𝑁0
𝐸𝐸 ℎ − ℎ� 0 = 2 𝐸𝐸 𝑤𝑤 2
=
𝑥𝑥 𝐸𝐸𝑥𝑥
Relative error is: 2
𝐸𝐸 ℎ − ℎ� 0 𝑁𝑁0 1
= =
ℎ2 𝐸𝐸𝑥𝑥 ℎ 2 𝛾𝛾
◦ 𝛾𝛾 =SNR on the symbol
Relative error is the inverse of the SNR

17
Smoothing the Estimate
Raw channel estimates are:
◦ Noisy
◦ Only available on DM-RS locations
Often smooth the estimate
◦ Fit over all locations
Many options
◦ Linear interpolation, splines
◦ Parametric
We look at Kernel regression
◦ Works well
◦ Easy to implement and analyze

18
Kernel Regression
Problem: Want to estimate ℎ 𝑛𝑛 , 𝑛𝑛 = 1, … , 𝑁𝑁
◦ Given noisy values ℎ� 0 𝑛𝑛 = ℎ 𝑛𝑛 + 𝑤𝑤 𝑛𝑛
◦ At locations 𝑛𝑛 ∈ 𝐼𝐼 ⊂ 0, … . , 𝑁𝑁 − 1

Kernel function 𝑤𝑤 ℓ
2 /(2𝜎𝜎 2 )
◦ Example radial basis function: 𝑤𝑤 ℓ = 𝑒𝑒 −ℓ
◦ 𝜎𝜎 = “Bandwidth”
𝑧𝑧 [𝑛𝑛]
Kernel estimate: ℎ� 𝑛𝑛 = 1
𝑧𝑧0 [𝑛𝑛]
◦ 𝑧𝑧1 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 ℎ� 0 𝑘𝑘 ,
◦ 𝑧𝑧0 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘

Weights raw samples by the distance 𝑤𝑤 𝑛𝑛 − 𝑘𝑘

Takeda, Farisu, Milanfar, Kernel Regression for Image Processing and Reconstruction, 2007

19
Example
RBF Kernel
◦ Bandwidth 𝜎𝜎 = 7
◦ Length 𝐿𝐿 = 3𝜎𝜎 on each side

Estimate is visibly smoother

20
MATLAB Implementation
Easy to implement in software or hardware
See demo

21
Extending the Estimate to Other Symbols
Up to now:
◦ Estimated at channel at a single OFDM symbol
� 𝑘𝑘0 ] on OFDM symbol 𝑘𝑘 = 𝑘𝑘0
◦ 𝐻𝐻[𝑛𝑛,

Can be extended to other symbols in slot:

� 𝑛𝑛, 𝑘𝑘 = 𝐻𝐻
� 𝑛𝑛, 𝑘𝑘0

Subcarrier
◦ 𝐻𝐻
◦ Assumes channel is constant over time
◦ Coherence time ≫ Slot time

Can also have multiple RS symbols

◦ Useful when there is time variations in slot

See lab

Channel estimate at
𝑘𝑘 = 𝑘𝑘0

22
In-Class Exercise

23
Outline
Role of Channel Estimation and Equalization
OFDM Channel Estimation via Kernel Regression
Bias-Variance Tradeoff

24
Mean Squared Error with Kernel Regression
Problem: True channel ℎ 𝑛𝑛 , 𝑛𝑛 = 1, … , 𝑁𝑁
◦ Given noisy values ℎ� 0 𝑛𝑛 = ℎ 𝑛𝑛 + 𝑣𝑣 𝑛𝑛 at sub-carriers 𝑛𝑛 ∈ 𝐼𝐼.
𝑧𝑧 [𝑛𝑛]
Kernel estimate: ℎ� 𝑛𝑛 = 1
𝑧𝑧0 [𝑛𝑛]
◦ 𝑧𝑧1 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 ℎ� 0 𝑘𝑘 , 𝑧𝑧0 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘
Mean-squared error:
2
𝑀𝑀𝑀𝑀𝑀𝑀 𝑛𝑛 = 𝐸𝐸 ℎ� 𝑛𝑛 − ℎ 𝑛𝑛

◦ Measures the accuracy of the estimate

◦ For now, model ℎ[𝑛𝑛] as deterministic
◦ Average over noise 𝑣𝑣[𝑛𝑛]
� 𝑛𝑛 −ℎ 𝑛𝑛 2
𝐸𝐸 ℎ
May also be interested in the normalized MSE: 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 𝑛𝑛 =
ℎ 𝑛𝑛 2

25
 Bias and Variance Error
Error: 𝑒𝑒 𝑛𝑛 = ℎ 𝑛𝑛 − ℎ� 𝑛𝑛
Since ℎ� 𝑛𝑛 is linear in ℎ� 0 can write 𝑒𝑒 𝑛𝑛 = 𝑒𝑒𝑠𝑠 𝑛𝑛 − 𝑒𝑒𝑣𝑣 [𝑛𝑛]
1
◦ Signal component: 𝑒𝑒𝑠𝑠 𝑛𝑛 = ℎ 𝑛𝑛 − ∑ 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 ℎ 𝑘𝑘
𝑧𝑧0 [𝑛𝑛] 𝑘𝑘∈𝐼𝐼
1
◦ Noise component: 𝑒𝑒𝑣𝑣 𝑛𝑛 = ∑ 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 𝑣𝑣 𝑘𝑘
𝑧𝑧0 [𝑛𝑛] 𝑘𝑘∈𝐼𝐼

MSE can be written as two components:

𝑀𝑀𝑀𝑀𝑀𝑀 𝑛𝑛 = 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 2 n + 𝑉𝑉𝑉𝑉𝑉𝑉 𝑛𝑛

◦ Bias error: 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 2 n = 𝑒𝑒𝑠𝑠 𝑛𝑛 2

2
◦ Variance error: 𝑉𝑉𝑉𝑉𝑉𝑉 𝑛𝑛 = 𝐸𝐸 𝑒𝑒𝑣𝑣 𝑛𝑛

26
Variance Error
1
Noise term 𝑒𝑒𝑣𝑣 𝑛𝑛 = ∑ 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 𝑣𝑣 𝑘𝑘
𝑧𝑧0 [𝑛𝑛] 𝑘𝑘∈𝐼𝐼

2 𝑁𝑁0
Recall 𝐸𝐸 𝑣𝑣 𝑛𝑛 = ,
𝐸𝐸𝑥𝑥
◦ 𝑁𝑁0 = noise on the reference signals, 𝐸𝐸𝑥𝑥 = Reference signal energy

Variance on the kernel estimate is:

2
∑𝑘𝑘∈𝐼𝐼 |𝑤𝑤 𝑛𝑛 − 𝑘𝑘 |2 𝑁𝑁0
𝑉𝑉𝑉𝑉𝑉𝑉 𝑛𝑛 = 𝐸𝐸 𝑒𝑒𝑣𝑣 𝑛𝑛 = 2
∑ 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 𝐸𝐸𝑥𝑥
𝑘𝑘∈𝐼𝐼

Proof:
2 ∑𝑘𝑘∈𝐼𝐼 |𝑤𝑤 𝑛𝑛−𝑘𝑘 |2 𝐸𝐸 𝑣𝑣 𝑘𝑘 2
◦ Since 𝑣𝑣[𝑛𝑛] are uncorrelated, 𝐸𝐸 𝑒𝑒𝑣𝑣 𝑛𝑛 =
|𝑧𝑧0 𝑛𝑛 |2
2 𝑁𝑁0
◦ Substitute expressions for 𝑧𝑧0 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 and 𝐸𝐸 𝑣𝑣 𝑘𝑘 =
𝐸𝐸𝑥𝑥

27
Variance Error for a Uniform Kernel
Example: Uniform kernel
1 if ℓ ≤ 𝐿𝐿
𝑤𝑤 ℓ = �
0 else
◦ Length 2𝐿𝐿 + 1

Suppose that there are 𝐾𝐾 reference signals in window

𝑁𝑁0
Variance error: 𝑉𝑉𝑉𝑉𝑉𝑉 𝑛𝑛 =
𝐾𝐾𝐾𝐾𝑥𝑥
2 ∑𝑘𝑘∈𝐼𝐼 |𝑤𝑤 𝑛𝑛−𝑘𝑘 |2 𝑁𝑁0 𝐾𝐾 𝑁𝑁0
◦ Why? 𝑉𝑉𝑉𝑉𝑉𝑉 𝑛𝑛 = 𝐸𝐸 𝑒𝑒𝑣𝑣 𝑛𝑛 = 2 =
∑ 𝑤𝑤 𝑛𝑛−𝑘𝑘 𝐸𝐸𝑥𝑥 𝐾𝐾 2 𝐸𝐸𝑥𝑥
𝑘𝑘∈𝐼𝐼
−𝐿𝐿 𝐿𝐿
Conclusion: Variance error decreases with kernel width
In general, increasing kernel width, decreases variance error 𝐾𝐾 ref signal
locations in window

28
Bias Error
Signal error:
1
◦ 𝑒𝑒𝑠𝑠 𝑛𝑛 = ℎ 𝑛𝑛 − ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛 − 𝑘𝑘 ℎ 𝑘𝑘 = ℎ 𝑛𝑛 − ℎ� 𝑛𝑛
𝑧𝑧0 𝑛𝑛

ℎ� 𝑛𝑛 = local average of ℎ 𝑘𝑘

𝑤𝑤 𝑛𝑛−𝑘𝑘
◦ Write as: ℎ� 𝑛𝑛 = ∑𝑘𝑘∈𝐼𝐼 𝛼𝛼𝑛𝑛−𝑘𝑘 ℎ 𝑘𝑘 , 𝛼𝛼𝑛𝑛−𝑘𝑘 = ∑
𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛−𝑘𝑘
◦ Weighted average of values of ℎ[𝑘𝑘]
◦ Weights 𝛼𝛼ℓ sum to one
2
Bias error, 𝑒𝑒𝑠𝑠 𝑛𝑛 , is large when:
◦ Average over region where ℎ[𝑛𝑛] changes significantly
◦ Window ≥ Coherence bandwidth / time
Low bias error High bias error
In general, increasing kernel increases the bias error Kernel width Kernel width
≪ Coherence BW ≥ Coherence BW

29
 Bias-Variance Tradeoff

Decreasing Increasing
kernel width kernel width

High variance High bias

Too low averaging Too much averaging

30
 Bias-Variance Tradeoff
High variance High bias
Small kernel width High kernel width
Too low averaging Too much averaging

This example:
◦ Mean delay spread, 𝛿𝛿 = 200 ns
1
◦ Coherence BW, 𝑊𝑊𝑐𝑐𝑐𝑐𝑐 = = 2.5 MHz
2𝛿𝛿
◦ Optimal width:
𝜎𝜎 = 9 subcarriers
𝜎𝜎 = 135 kHz
◦ We see: 𝜎𝜎 ≪ 𝑊𝑊𝑐𝑐𝑐𝑐𝑐
Optimal

31
Modeling Channel Estimation Error
Channel estimation error: ℎ� = ℎ + 𝑣𝑣
2
Channel estimation error has two components: 𝐸𝐸 𝑣𝑣 = 𝐵𝐵𝐵𝐵𝐵𝐵𝑠𝑠 2 + 𝑉𝑉𝑉𝑉𝑉𝑉
◦ Each term depends on level of averaging in the kernel estimator

Bias error:
◦ Typically scales with the channel power
◦ Since medium is linear, increasing TX power will not change the relative kernel error
◦ Assume 𝐵𝐵𝐵𝐵𝐵𝐵𝑠𝑠 2 = 𝑐𝑐𝑏𝑏 ℎ 2
◦ Constant 𝑐𝑐𝑏𝑏 increases with the kernel width

Variance error:
𝑁𝑁 ∑𝑘𝑘∈𝐼𝐼 |𝑤𝑤 𝑛𝑛−𝑘𝑘 |2
◦ From earlier, we can write 𝑉𝑉𝑉𝑉𝑉𝑉 = 𝑐𝑐𝑣𝑣 0 , 𝑐𝑐𝑣𝑣 = 2
𝐸𝐸𝑥𝑥 ∑𝑘𝑘∈𝐼𝐼 𝑤𝑤 𝑛𝑛−𝑘𝑘
◦ Constant 𝑐𝑐𝑣𝑣 decreases with the kernel width

32
Effective Noise in Equalization
Consider channel on a single resource element:
2
◦ Channel: 𝑟𝑟 = ℎ𝑥𝑥 + 𝑤𝑤, 𝑤𝑤~𝐶𝐶𝐶𝐶 0, 𝑁𝑁0 , 𝐸𝐸 𝑥𝑥 = 𝐸𝐸𝑥𝑥
𝐸𝐸𝑥𝑥 ℎ 2
◦ SNR is 𝛾𝛾0 =
𝑁𝑁0

Simple model for channel estimation error: ℎ� = ℎ + 𝑣𝑣, 𝑣𝑣~𝐶𝐶𝐶𝐶(0, 𝑁𝑁𝑣𝑣 )

2 𝑐𝑐𝑣𝑣 𝑁𝑁0
◦ Bias+ variance channel estimation error: 𝑁𝑁𝑣𝑣 = 𝑐𝑐𝑏𝑏 ℎ +
𝐸𝐸𝑥𝑥

Consider equalized estimate:

𝑧𝑧 = ℎ� ∗ 𝑟𝑟 = ℎ + 𝑣𝑣 ∗
ℎ𝑥𝑥 + 𝑤𝑤 = ℎ 2 𝑥𝑥 + 𝑛𝑛,
◦ Total 𝑛𝑛 = (ℎ + 𝑣𝑣)∗ 𝑤𝑤 + 𝑣𝑣 ∗ ℎ𝑥𝑥
◦ Neglecting cross terms: 𝑛𝑛 ≈ ℎ∗ 𝑤𝑤 + 𝑣𝑣 ∗ ℎ𝑥𝑥
Total effective noise, 𝑛𝑛, has two terms:
◦ Original noise: ℎ∗ 𝑤𝑤
◦ Distortion from channel estimation error: 𝑣𝑣 ∗ ℎ𝑥𝑥

33
Signal-to-Noise and Distortion (SNDR)
Equalized symbol:
𝑧𝑧 = ℎ� ∗ 𝑟𝑟 = ℎ + 𝑣𝑣 ∗
ℎ𝑥𝑥 + 𝑤𝑤 = ℎ 2 𝑥𝑥 + 𝑛𝑛, 𝑛𝑛 ≈ ℎ∗ 𝑤𝑤 + 𝑣𝑣 ∗ ℎ𝑥𝑥
Signal energy: 𝐸𝐸𝑠𝑠 = ℎ 4 𝐸𝐸𝑥𝑥
Total effective noise energy:
2
◦ Error energy is 𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡 = 𝐸𝐸 𝑛𝑛 = ℎ 2 𝑁𝑁0 + 𝑁𝑁𝑣𝑣 ℎ 2 𝐸𝐸𝑥𝑥
𝛾𝛾0
Signal-to noise and distortion (SNDR): 𝛾𝛾 =
1+𝑐𝑐𝑣𝑣 +𝑐𝑐𝑏𝑏 𝛾𝛾0

𝐸𝐸𝑠𝑠 ℎ 4 𝐸𝐸𝑥𝑥 ℎ 2 𝐸𝐸𝑥𝑥

◦ Proof: 𝛾𝛾 = = =
𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡 ℎ 2 𝑁𝑁0 + ℎ 2 𝐸𝐸𝑥𝑥 𝑁𝑁𝑣𝑣 𝑁𝑁0 +𝑁𝑁𝑣𝑣 𝐸𝐸𝑥𝑥
𝑐𝑐𝑣𝑣 𝑁𝑁0
◦ Substitute in 𝑁𝑁𝑣𝑣 = + 𝑐𝑐𝑏𝑏 ℎ 2
𝐸𝐸𝑥𝑥

34
SNDR Regimes
𝛾𝛾0
SNDR: 𝛾𝛾 =
1+𝑐𝑐𝑣𝑣 +𝑐𝑐𝑏𝑏 𝛾𝛾0

Low SNR regime:

𝛾𝛾0
◦ 𝛾𝛾 ≈ High SNR
1+𝑐𝑐𝑣𝑣
1 Low SNR Bias error saturation
◦ Loss in SNR of Variance error loss 𝛾𝛾 ≈ 1⁄𝑐𝑐𝑏𝑏
1+𝑐𝑐𝑣𝑣
1⁄(1 + 𝑐𝑐𝑣𝑣 )
◦ Limited by variance error, 𝑐𝑐𝑣𝑣

High SNR regime:

1
◦ 𝛾𝛾0 ≈
𝑐𝑐𝑏𝑏
◦ Limited by bias error, 𝑐𝑐𝑏𝑏

35
Estimating Effective Noise
For equalization and LLR calculations we need to estimate the effective noise
Noise estimation can be performed via residual error:
◦ Reference symbols: 𝑟𝑟𝑛𝑛 = ℎ𝑛𝑛 𝑥𝑥𝑛𝑛 + 𝑤𝑤𝑛𝑛 , 𝑛𝑛 = 1, … , 𝑁𝑁𝑅𝑅𝑅𝑅
◦ Obtain channel estimate: ℎ� 𝑛𝑛 (e.g., via kernel regression)
1 2
�𝑅𝑅𝑅𝑅 =
◦ Effective noise estimate: 𝑁𝑁 ∑ 𝑟𝑟 − ℎ� 𝑛𝑛 𝑥𝑥𝑛𝑛
𝑁𝑁𝑅𝑅𝑅𝑅 𝑛𝑛 𝑛𝑛

Residual error estimate captures noise and distortion

◦ If ℎ� 𝑛𝑛 = ℎ𝑛𝑛 + 𝑣𝑣𝑛𝑛 , then 𝑟𝑟𝑛𝑛 − ℎ� 𝑛𝑛 𝑥𝑥𝑛𝑛 = 𝑟𝑟𝑛𝑛 − ℎ𝑛𝑛 𝑥𝑥𝑛𝑛 + 𝑣𝑣𝑛𝑛 𝑥𝑥𝑛𝑛 = 𝑤𝑤𝑛𝑛 + 𝑣𝑣𝑛𝑛 𝑥𝑥𝑛𝑛

But the residual error estimate is biased:

◦ Under-estimates the variance error in the distortion
◦ Measuring error on the symbols used in training
◦ Can compensate by scaling the noise estimate before using in the equalizer or demodulator

36
Measuring True Effective Noise
From previous slide, residual error noise estimate is biased
◦ Under-estimates the variance error
◦ Uses training symbols for error estimation

To more accurately measure noise, use separate set of test symbols
◦ Get test symbols: 𝑟𝑟𝑛𝑛 = ℎ𝑛𝑛 𝑥𝑥𝑛𝑛 + 𝑤𝑤𝑛𝑛 , 𝑛𝑛 = 1, … , 𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡
◦ These symbols are distinct from the RS used in training
1 2
�𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 =
◦ Error estimate on the test: 𝑁𝑁 ∑𝑛𝑛 𝑟𝑟𝑛𝑛 − ℎ� 𝑛𝑛 𝑥𝑥𝑛𝑛
𝑁𝑁𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡

Rarely used during actual operation of the receiver

◦ Adding test symbols adds excessive overhead

But, the test symbols can be done in test and simulation

◦ For example, populate data symbols with known symbols

37
Noise Estimation Simulation
Simulation parameters as before
◦ Mean delay spread, 𝛿𝛿 = 200 ns
1
◦ Coherence bandwidth, 𝑊𝑊𝑐𝑐𝑐𝑐𝑐 ≈ = 2.5 MHz
2𝛿𝛿
◦ One RS per 4 sub-carriers: 60 kHz

“True” noise measured on data symbols

◦ Not used in training

“Estimated” noise measured on RS symbols

We see:
◦ Estimated noise is slightly optimistic at low SNRs
◦ Both true and estimated noise levels off at high SNR
◦ Limited by bias SNR

38
SNDR in Practice
Estimated the SNDR from effective noise estimate
In this simulation:
◦ About 0.4 dB loss at low SNRs
◦ About 27 dB saturation

39
In-Class Exercise

40