Method for analytically calculating

BER (bit error rate) in presence of

non-linearity
Gaurav Malhotra
Xilinx
Outline

• Review existing methodology for calculating BER based

on linear system analysis.
– Link model with ISI, Crosstalk, Jitter, Noise.

• Model of nonlinearity based on power series.

• Modification of PDF in presence of nonlinearity.

• BER results for a typical high speed link.

• Link model with multiple linear & NL blocks.

 Linear system : Link Model
Crosstalk
AFE (CTLE + DFE)
d d’
[+1 -1] Tx Channel Rx [+1 -1]
- Jitter ‘enhancement’
- RJ - ISI / reflections AWGN - RJ
- DCD - DCD
- PSIJ - PSIJ
- SJ - SJ

Crosstalk’
Crosstalk(t)
Equivalent linear model: d’
• Signal and impairments
d
can be referred to the [+1 -1] h(n) Signal(t) x(t) [+1 -1]
slicer input. Tx + Channel + RxAFE
AWGN’  - RJ’
n(t) - DCD’
- PSIJ’
Joint pdf (per sampling phase): - SJ’
pdf(Signal)  pdf(AWGN’) 
pdf(xtalk’)  pdf(ISI)

• Objective is to determine joint PDF 𝐹𝑋 𝑥 of signal + impairments [ x(t) ] at the decision point.
𝑆𝑆
• 𝐵𝐸𝑅𝑘 = 𝑃 𝑒𝑟𝑟𝑜𝑟 = 𝑘𝑃 𝑒𝑟𝑟𝑜𝑟 𝑑𝑘 𝑃(𝑑𝑘 ), where 𝑃 𝑒𝑟𝑟𝑜𝑟 1 = 𝐹
−∞ 𝑋
𝑥|1 S

• Taking timing jitter into account : 𝐵𝐸𝑅 = 𝑘 𝐵𝐸𝑅𝑘 𝐹 𝑘

LTI Systems: Link BER methodology
Joint pdf for
Each phase

Bath Tub
curve

Voltage
noise

Timing
noise

Joint pdf (per sampling phase): Joint pdf :

pdf(AWGN)  pdf(xtalk)  pdf(ISI) pdf(RJ)  pdf(SJ)  ..

Conditional
pdf 𝐵𝐸𝑅 = 𝐵𝐸𝑅𝑘 𝐹 𝑘
𝑘
Recap goal
• If we can accurately determine the probability distribution at the
decision point, we can calculate BER.
– 𝐵𝐸𝑅𝑘 = 𝑃 𝑒𝑟𝑟𝑜𝑟 = 𝑘𝑃 𝑒𝑟𝑟𝑜𝑟 𝑑𝑘 𝑃(𝑑𝑘 ), where
𝑆𝑆
𝑃 𝑒𝑟𝑟𝑜𝑟 1 = 𝐹
−∞ 𝑋
𝑥|1 S

– Taking timing jitter into account: 𝐵𝐸𝑅 = 𝑘 𝐵𝐸𝑅𝑘 𝐹 𝑘

• GOAL: to determine PDF (overall/joint including all impairments

AND nonlinearity) at the decision point.
Modeling of nonlinearity

Actual Circuit
• Common model / System

of NL: I
LTI system
O
U
𝑛 N T
𝑌= 𝑛 𝑋 P Model P
𝑛 U U
T
• Observed to be
T

LTI system 𝑋 Power series 𝑌= 𝑛 𝑋 𝑛

very close to real

𝑛
Model polynomial

circuits.
Volterra series
Model
Modeling of nonlinearity
Circuit model
(Known )

Input Actual Circuit 𝑌

  𝑀 = [𝑋1 𝑋 2 … 𝑋 𝑛 ]
/ System
1
⋮ = 𝑀−1 𝑌
Linear Model 𝑋
NL 𝑌′ =
𝑛
𝑛 𝑋 𝑛
𝑛
H(f)  (?) Error = 𝒀 – 𝒀′
Design specification y2 / error2 (dB)
(Known )
No NL Up to 3rd Up to 5th Up to 7th
modeling order order order

11 23 46 51

• Design specification (say pole-zero model) is known.

• Input, X, Y , Y’ are time domain signals. Only NL terms {𝑛 } are unknown.

• Matrix inversion (zero forcing) though not optimum, but gives a good estimate of NL terms.
 Modeling of nonlinearity

No NL modeling Up to 3rd order Up to 5th order

No NL Up to 3rd Up to 5th Up to 7th

modeling order order order

11 23 46 51

• Adding higher order terms in estimation reduces error in modeling

due to NL.
Up to 7th order
Modification of PDF in presence of NL
• Let y = g(x) represent the output of a non-linear function whose input is x.
• The PDF of Y, FY(y) can be determined in terms of PDF of X as:
[Probability, Random variables and Stochastic Processes: Athanasios Papoulis, Section 5-2]

x g(x) y

y
g(x1) g(x2) g(x3)

x x1 x2 x3

pdf of y =
pdf of x =
𝐹𝑋 𝑥 g(x) 𝐹𝑌 𝑦 =
𝐹𝑋 𝑥1
+
𝐹𝑋 𝑥2
+⋯
𝐹𝑋 𝑥𝑛
|𝑔′ 𝑥1 | |𝑔′ 𝑥2 | |𝑔′ 𝑥𝑛 |

𝑆𝑖𝑚𝑝𝑙𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑓𝑜𝑟 𝑚𝑜𝑛𝑜𝑡𝑜𝑛𝑖𝑐 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛𝑠:

𝑑𝑥
𝐹𝑌 𝑦 = | 𝑑𝑦 | 𝐹𝑋 𝑥
Modification of PDF : AWGN Example
Y= X + X3
𝑑𝑥
𝐹𝑌 𝑦 = | | 𝐹𝑋 𝑥
𝑑𝑦

d d’
[+k -k] NL [+1 -1]

AWGN

X= Signal +AWGN
𝐹𝑋 𝑥 = pdf(Signal)  pdf(AWGN)

Note the
‘warping’ of
PDF in
accordance
𝑑𝑥
with | |
𝑑𝑦
AWGN Example : Simulation VS Analysis

Y = X + X3
𝑑𝑥
𝐹𝑌 𝑦 = | | 𝐹𝑋 𝑥
𝑑𝑦
1
=| 2 | 𝐹𝑋 𝑥
1+3X

d d’
[+0.5 -0.5] NL [+0.5 -0.5]

AWGN

X= Signal +AWGN
𝐹𝑋 𝑥 = pdf(Signal)  pdf(AWGN)

• PDF can be obtained analytically or by running a bit-by-bit simulation.

• Both methods give the same result.

• Analytically computing BER is much faster. This is the method we will adopt for this presentation.
AWGN Example : PAM2 VS PAM4
• In general we expect higher order
modulations to suffer more from NL.

• Outer points in constellation

dominate BER.

• *** Detection rule may be modified

to take advantage of (known) non-
linearity. This paper assumes that
same detection rule (minimum
distance) as is used for linear system
analysis is used for calculating BER in
presence of non-linearity.
 Link Model: typical high speed link
Crosstalk’
TX +
d
[+1 -1] package + CTLE NL
connector
AWGN’

Y= X + X3
𝑑𝑥
𝐹𝑌 𝑦 = | | 𝐹𝑋 𝑥
𝑑𝑦

Tx + package + card + Crosstalk

Equivalent model with NL:
connector + CTLE
• Linear components d d’
convolve.
[+1 -1] Channel NL [+1 -1]

AWGN - RJ [UI/64] (rms)

- DCD [UI/32]

X= Signal +AWGN
𝐹𝑋 𝑥 = pdf(Signal)  pdf(AWGN)
 pdf(xtalk)  pdf(ISI)

• CTLE (Analog front end) is a significant source of NL.

•  = -0.3; Output = Input – 0.3 * Input3

• CTLE output referred Xtalk : Xtalk_Out(f) = Xtalk_In(f) * CTLE(f)

Link Model: typical high speed link
Crosstalk’
TX +
d
[+1 -1] package + CTLE NL
connector Y = X + X3
AWGN’

BASELINE: PAM2 VS PAM4

Start with the same BER,

compare the effect of NL

PAM-4 PAM-2

• Bandwidth of insertion loss, crosstalk, AWGN and CTLE for PAM2 are half that of PAM4.
• Jitter is specified as a fraction of UI, so that automatically adjusts for signaling rate.
• Since the crosstalk channel is not flat, we had to make small adjustment on gain of crosstalk channel to make the
baseline BER (without NL) the same for both PAM2 & PAM4.
BER results: typical high speed link

Bandwidth UI BER without NL BER with NL

(Nyquist)

PAM2 FN 1/(2* FN) 1e-25 1e-23

PAM4 FN / 2 2/(2* FN) 1e-25 1e-20

Link Model: Multiple NL blocks

𝐹𝑌 𝑦 = 𝐹𝑍 𝑧 =
𝑑𝑥 𝐹𝐵 𝑏
| | 𝐹𝑋 𝑥 𝐹𝐴 𝑎 = 𝑑𝑏
𝑑𝑦 (LTI method) | | 𝐹𝐵 𝑏
𝐹𝐷𝑓𝑒 𝑑𝑓𝑒| − 1  𝐹𝑌 𝑦|1 𝑑𝑧

Tx + package + card + Crosstalk

connector + CTLE
Summer
d d’
[+1 -1] LTI1 NL1 LTI3 NL2 [+1 -1]
𝐹𝐷𝑓𝑒 𝑑𝑓𝑒
AWGN

LTI2 DFE
X= Signal +AWGN 𝐹𝐷𝑓𝑒 𝑑𝑓𝑒| − 1 =
𝐹𝑋 𝑥 = pdf(Signal)  pdf(AWGN) {Tap1 x pdf(d) }
 pdf(xtalk)  pdf(ISI) {Tap2 x pdf(d) } …
d’
[+1 -1]

PDF transformation
• Linear block : Convolution
• Nonlinear block : 𝐹𝑌 𝑦 =
𝐹𝑋 𝑥1
|𝑔′ 𝑥1 |
𝐹 𝑥
+ 𝑋 2 +
|𝑔′ 𝑥2 |
⋯
𝐹𝑋 𝑥𝑛
|𝑔′ 𝑥𝑛 |
Summary

• Presented methodology for calculating BER of a link in presence of

nonlinearity.
– Modification of PDF.

– Static nonlinearity model using power series polynomial considered.

• Work ongoing to model nonlinearity using Volterra series.

• Higher order modulations are more susceptible to NL.

• Quantified the loss for a typical NL, typical high speed link.