Matthias Patzold
Agder University College
Mobile Communications Group
Grooseveien 36, NO4876 Grimstad, Norway
Email: matthias.paetzold@hia.no
Homepage: http://ikt.hia.no/mobilecommunications/
BEATS/CUBAN Workshop 2004
1/37
Contents
I. Introduction
II. Principles of Fading Channel Modelling
III. Modelling of MIMO Channels
IV. Simulation of MIMO Channels
V. Conclusion
BEATS/CUBAN Workshop 2004
2/37
I. Introduction
Mobile Communications Using Smart Antennas
Transmitter
Channel
) x , t , ( h J
Receiver
Mobile station Base station
1
2
M
R
01 1 01 0 01 1 01 0
Spacetime signal processing
Demodulation
Decoding
Spacetime coding
Modulation
1
2
M
T
Realistic spatiotemporal channel models are required for the design, test, and optimization
of mobile communication systems employing smart antenna technologies.
BEATS/CUBAN Workshop 2004
3/37
Meaning of Channel Modelling and Simulation for Mobile Communications
Reference system:
+
Channel
model
Noise
Data source Data sink
Transmitter
Receiver
d(i) x(t) y(t)
d(i)
Performance measure: Bit error probability P
b
= f
_
E
b
N
0
_
Simulation system:
+
Channel
Noise
Data source Data sink
Transmitter
Receiver
simulator
d(i) x(t) y(t)
d(i)
Performance measure: Bit error probability
P
b
= f
_
E
b
N
0
,
_
P
b
Model error
The model error describes the essential error which is introduced by the channel simulator.
Channel simulators are important for the test, the parameter optimization, and the perfor
mance analysis of mobile communication systems.
BEATS/CUBAN Workshop 2004
4/37
The Two Main Categories of Fading Channel Simulators
Simulators based on reference models: Simulators based on measurements:
Simulation model
Reference model
Measurement
Realworld
Snapshot
measurements
campaigns
channel
channel
Realworld
Simulation model
Problem: Reference models are not
close to realworld channels.
Problem: Finding of typical scenarios.
The dilemma of mobile fading channel modelling.
BEATS/CUBAN Workshop 2004
5/37
Some Requirements for Fading Channel Simulators
High precision w.r.t. a given reference model or w.r.t. measured channels
High eciency
The underlying stochastic channel simulator must be ergodic to reduce the number of trials
Easy to determine the model parameters
Easy to understand and easy to implement
Reproducibility for enabling fair performance comparisons
Easy to extend (frequency selectivity, spatial selectivity, multicluster propagation scenarios, number
of antenna elements, etc.)
Low complexity
BEATS/CUBAN Workshop 2004
6/37
II. Principles of Fading Channel Modelling
Two Fundamental Methods for Modelling of Coloured Gaussian Noise Processes
1. Filter method:
Stochastic process (t)
Reference model
2. Sumofsinusoids (SOS) method:
cos(2f
1
t +
1
)
cos(2f
2
t +
2
)
cos(2f
N
t +
N
)
c
1
c
2
c
N
.
.
.
.
.
.
.
.
.
.
.
.
e


e


e


+

(t)
Stochastic process (t)
Reference model
BEATS/CUBAN Workshop 2004
7/37
For a nite number of harmonic functions N, we obtain:
Parameters: ensemble
n
= Random variable (RV)
f
n
= const.
c
n
= const.
cos(2f
1
t +
1
)
cos(2f
2
t +
2
)
cos(2f
N
t +
N
)
c
1
c
2
c
N
.
.
.
.
.
.
e


e


e


+

(t)
Stochastic process (t)
Stochastic simulation model
Parameters: single realization
n
= const.
f
n
= const.
c
n
= const.
cos(2f
1
t +
1
)
cos(2f
2
t +
2
)
cos(2f
N
t +
N
)
c
1
c
2
c
N
.
.
.
.
.
.
e


e


e


+

(t)
Deterministic process (t)
Deterministic simulation model
BEATS/CUBAN Workshop 2004
8/37
Relationships Between Stochastic Processes, Random Variables,
Sample Functions, and RealValued Numbers
Gaussian process: (t) = lim
N
N
n=1
c
n
cos(2f
n
t +
n
) (c
n
= const., f
n
= const.,
n
= RV)
Stochastic process: (t) =
N
n=1
c
n
cos(2f
n
t +
n
) (c
n
= const., f
n
= const.,
n
= RV)
simulation
model
Deterministic
Stochastic
simulation
model
model
Reference
Gaussian process
Stochastic process
Random
variable
Sample
function
Real number
t = t
0
n
= const. t = t
0
(t
0
) = (t
0
,
n
)
(t
0
) = (t
0
)
(t) = (t,
n
)
N < N
(t) = (t,
n
)
(t) = (t,
n
)
n
= const.
n
= RV
BEATS/CUBAN Workshop 2004
9/37
Principle of Deterministic Channel Modelling
Step 1: Starting point is a reference model derived from one, two or several Gaussian
processes.
Step 2: Derive a stochastic simulation model from the reference model by replacing each
Gaussian process by a sumofsinusoids with xed gains, xed frequencies, and
random phases.
Step 3: Determine the deterministic simulation model by xing all model parameters of the
stochastic simulation model, including the phases.
Step 4: Compute the model parameters of the simulation model by using a proper param
eter computation method.
Step 5: Simulate one (or some few) sample functions (deterministic processes).
BEATS/CUBAN Workshop 2004
10/37
Example: Derivation of a channel simulator for Rice processes
Step 1: Reference model Steps 2 & 3: Simulation model
(t)
(t)
2
1
(t)
f t+ m (t)=
)
2
sin(2
cos(2f t+ m (t)=
1
)
H (f)
H (f)
(t)
2
v
(t)
1
v
(t)
(t)
(t)
WGN
WGN
2
1
n=1
c
i,n
cos(2f
i,n
t +
i,n
)
Model parameters: gains frequencies phases
Historical overview
Methods Disadvantages
Rice method (Rice, 1944) Small period
Jakes method (Jakes, 1974) Special case only
Monte Carlo method (Schulze, 1988) Nonergodic
Method of equal distances (Patzold, 1994) Small period
Method of equal areas (Patzold, 1994) Slow convergency
Harmonic decomposition technique (Crespo, 1995) Small period
Meansquareerror method (Patzold, 1996) Small period
Method of exact Doppler spread (Patzold, 1996) Special case only
L
p
norm method (Patzold, 1996)
Method porposed by Zheng and Xiao (2002) Nonergodic
BEATS/CUBAN Workshop 2004
12/37
Classes of SOS Channel Simulators and Their Statistical Properties
Deterministic process:
i
(t) =
N
i
n=1
c
i,n
cos(2f
i,n
t +
i,n
)
Class Gains Frequencies Phases Firstorder Widesense Mean Autocor.
c
i,n
f
i,n
i,n
stationary stationary ergodic ergodic
I const. const. const.
II const. const. RV yes yes yes yes
III const. RV const. no/yes
a
no/yes
a
no/ yes
a
no
IV const. RV RV yes yes yes no
V RV const. const. no no no/yes
a
no
VI RV const. RV yes yes yes no
VII RV RV const. no/yes
a
no/yes
a
no/yes
a
no
VIII RV RV RV yes yes yes no
a
If certain boundary conditions are fullled.
BEATS/CUBAN Workshop 2004
13/37
Application of the Scheme to Parameter Computation Methods
Stochastic process:
i
(t) =
N
i
n=1
c
i,n
cos(2f
i,n
t +
i,n
)
Model parameters: gains frequencies phases (RVs)
Parameter computation methods Class FOS WSS
Mean
ergodic
Autocor.
ergodic
Rice method II yes yes yes yes
Monte Carlo method IV yes yes yes no
Jakes method II yes yes yes yes
Harmonic decomposition technique II yes yes yes yes
Method of equal distances II yes yes yes yes
Method of equal areas II yes yes yes yes
Meansquareerror method II yes yes yes yes
Method of exact Doppler spread II yes yes yes yes
L
p
norm method II yes yes yes yes
Method proposed by Zheng and Xiao IV yes yes yes no
BEATS/CUBAN Workshop 2004
14/37
Applications of the Principle of Deterministic Channel Modelling
Frequencynonselective channels (Rayleigh, Rice, ext. Suzuki, Nakagami, etc.),
Frequencyselective channels (COST 207, CODIT),
Spacetime wideband channels (COST 259),
Multiple crosscorrelated Rayleigh fading channels,
Multiple uncorrelated Rayleigh fading channels,
Perfect modelling and simulation of measured 2D and 3D channels,
Frequencyhopping fading channels,
Design of ultra fast channel simulators (tables system),
Design of burst error models,
Development of MIMO channels.
BEATS/CUBAN Workshop 2004
15/37
III. Modelling of MIMO Channels
Block Diagram of a MIMO Mobile Communication System
Transmitter Receiver
x
1
x
2
x
M
T
r
1
r
2
r
M
R
h
11
h
1
2
h
2
1
h
M
R
1
h
22
h
M
R
2
h
1
M
T
h
M
R
M
T
h
2
M
T
_
_
_
_
_
_



H
H
H
H
H
H
H
H
H
H
H
Hj
S
S
S
S
S
S
S
S
S
S
S
S
S
S
S
Sw
*
H
H
H
H
H
H
H
H
H
H
H
Hj
3
Channel coecients: The channel coecient h
ij
(t) describes the transmission behavior of the
channel from the jth transmit antenna to the ith receive antenna.
Channel matrix: H(t) = [h
ij
(t)] C
M
R
M
T
Channel capacity: C(t) = log
2
det
_
I
M
T
+
S
BS, total
M
T
N
noise
H(t)H
H
(t)
_
[bits/s/Hz]
(M
R
M
T
)
BEATS/CUBAN Workshop 2004
16/37
Generalized Principle of Deterministic Channel Modelling
Step 1: Starting point is a geometrical model with an innite number of scatterers, e.g.,
(a) onering model (b) tworing model (c) elliptical model
BS MS
BS MS
BS MS
Step 2: Derive a stochastic reference model from the geometrical model.
Step 3: Derive an ergodic stochastic simulation model from the reference model by using
only a nite number of N scatterers.
Step 4: Determine the deterministic simulation model by xing all model parameters of the
stochastic simulation model.
Step 5: Compute the parameters of the simulation model by using a proper parameter
computation method, e.g., the L
p
norm method (LPNM).
Step 6: Generate one (or some few) sample functions by using the deterministic simulation
model with xed parameters.
BEATS/CUBAN Workshop 2004
17/37
Illustration of the Generalized Principle of Deterministic Channel Modelling
3 4 6
5
2 1
Innite complexity
(N )
sample functions sample functions
Nonrealizable Realizable
sample functions
Nonrealizable
Reference model
E E < >
Statistical properties
Deterministic SoS
simulation model
computation
Parameter
Simulation of
sample functions
Fixed
parameters
Stochastic SoS
simulation model
Finite complexity
One (or some few) Innite number of Innite number of
Finite complexity
(N 25) (N 25)
Lpnorm method (LPNM)
Geometrical model
BEATS/CUBAN Workshop 2004
18/37
Application of the Generalized Principle of Deterministic Channel Modelling
A Geometrical Model for a MIMO Channel
D
n2
D
n1
A
BS
1
BS
A
2
MS
MS
MS
0
MS
BS
n
BS
A
MS
1
D
2n
BS
max
S
n
v
y
0
BS
D
2
A
BS
MS
D
n
1n
x
R
v
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
h
12
(t) = lim
N
1
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
h
21
(t) = lim
N
1
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
h
22
(t) = lim
N
1
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
where the phases
n
are i.i.d. RVs and
a
n
= e
j(
BS
/)
[
cos(
BS
)+
BS
max
sin(
BS
) sin(
MS
n
)
]
b
n
= e
j(
MS
/) cos(
MS
n
MS
)
f
n
= f
max
cos(
MS
n
v
)
Channel matrix: H(t) = [h
ij
(t)] =
_
h
11
(t) h
12
(t)
h
21
(t) h
22
(t)
_
Channel capacity: C(t) = log
2
det
_
I
2
+
S
BS, total
2N
noise
H(t)H
H
(t)
_
[bits/s/Hz]
BEATS/CUBAN Workshop 2004
20/37
Statistical Properties of the Reference Model
Spacetime CCF:
11,22
(
BS
,
MS
, ) := Eh
11
(t)h
22
(t + )
= lim
N
1
N
N
n=1
a
2
n
b
2
n
e
j2f
n
a
2
n
(
BS
)b
2
n
(
MS
)e
j2f
n
p(
MS
)d
MS
Time ACF: r
h
11
() := Eh
11
(t)h
11
(t + )
=
e
j2f
max
cos(
MS
v)
p(
MS
) d
MS
Relationships: r
h
11
() = r
h
12
() = r
h
21
() = r
h
22
() =
11,22
(0, 0, )
2D space CCF: (
BS
,
MS
) :=
11,22
(
BS
,
MS
, 0)
=
a
2
n
(
BS
)b
2
n
(
MS
)p(
MS
)d
MS
BEATS/CUBAN Workshop 2004
21/37
The Stochastic Simulation Model
Channel coecients:
h
11
(t) =
1
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
The phases
n
are i.i.d. RVs
Channel matrix:
H(t) =
_
h
11
(t)
h
12
(t)
h
21
(t)
h
22
(t)
_
Stochastic channel matrix
Channel capacity:
C(t) = log
2
det
_
I
2
+
S
BS, total
2N
noise
H(t)
H
H
(t)
_
[bits/s/Hz]
Stochastic process
The analysis of
C(t) has to be performed by using statistical aver
ages, e.g., C
s
= E
C(t).
BEATS/CUBAN Workshop 2004
22/37
Statistical Properties of the Stochastic Simulation Model
Spacetime CCF:
11,22
(
BS
,
MS
, ) := E
h
11
(t)
22
(t + )
=
1
N
N
n=1
a
2
n
(
BS
)b
2
n
(
MS
)e
j2f
n
Time ACF: r
h
11
() := E
h
11
(t)
11
(t + )
=
1
N
N
n=1
e
j2f
max
cos(
MS
n
v
)
Relationships: r
h
11
() = r
h
12
() = r
h
21
() = r
h
22
() =
11,22
(0, 0, )
2D space CCF: (
BS
,
MS
) :=
11,22
(
BS
,
MS
, 0)
=
1
N
N
n=1
a
2
n
(
BS
)b
2
n
(
MS
)
=
1
N
N
n=1
e
j2(
BS
/)[cos(
BS
)+
BS
max
sin(
BS
) sin(
MS
n
)]
e
j2(
MS
/) cos(
MS
n
MS
)
BEATS/CUBAN Workshop 2004
23/37
The Deterministic Simulation Model
Channel coecients:
h
11
(t) =
1
N
N
n=1
a
n
b
n
e
j(2f
n
t+
n
)
The phases
n
are constant quantities
Channel matrix:
H(t) =
_
h
11
(t)
h
12
(t)
h
21
(t)
h
22
(t)
_
Deterministic channel matrix
Channel capacity:
C(t) = log
2
det
_
I
2
+
S
BS, total
2N
noise
H(t)
H
H
(t)
_
[bits/s/Hz]
Deterministic process
The analysis of
C(t) has to be performed by using time averages,
e.g., C
t
=<
C(t) >.
BEATS/CUBAN Workshop 2004
24/37
Statistical Properties of the Deterministic Simulation Model
Spacetime CCF:
11,22
(
BS
,
MS
, ) := <
h
11
(t)
22
(t + ) >
=
1
N
N
n=1
a
2
n
(
BS
)b
2
n
(
MS
)e
j2f
n
Time ACF: r
h
11
() := <
h
11
(t)
11
(t + ) >
=
1
N
N
n=1
e
j2f
max
cos(
MS
n
v
)
Relationships: r
h
11
() = r
h
12
() = r
h
21
() = r
h
22
() =
11,22
(0, 0, )
2D space CCF: (
BS
,
MS
) :=
11,22
(
BS
,
MS
, 0)
=
1
N
N
n=1
a
2
n
(
BS
)b
2
n
(
MS
)
=
1
N
N
n=1
e
j2(
BS
/)[cos(
BS
)+
BS
max
sin(
BS
) sin(
MS
n
)]
e
j2(
MS
/) cos(
MS
n
MS
)
BEATS/CUBAN Workshop 2004
25/37
Parameter Computation Method
Parameters: The model parameters to be determined are the discrete AOAs
MS
n
(n = 1, 2, . . . , N).
Aim: Determine the model parameters
MS
n
such that
11,22
(
BS
,
MS
, )
11,22
(
BS
,
MS
, ) .
Solution: L
p
norm method
E
(p)
1
:=
_
1
max
max
_
0
[r
h
11
() r
h
11
()[
p
d
_
1/p
E
(p)
2
:=
_
1
BS
max
MS
max
BS
max
_
0
MS
max
_
0
[(
BS
,
MS
) (
BS
,
MS
)[
p
d
MS
d
BS
_
1/p
Two alternatives: (i) Joint optimization: E
(p)
= w
1
E
(p)
1
+ w
2
E
(p)
2
w
1
, w
2
: weighting factors
(ii) Using two independent sets
/
MS
n
and
MS
n
in r
h
11
() and (
BS
,
MS
), respectively.
Orthogonalization of the optimization problem.
BEATS/CUBAN Workshop 2004
26/37
Structure of the Simulation Model
+
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . . . . .
. . .
.
.
.
.
.
.
+
.
.
.
.
.
.
1/
N
1/
N
1/
N
s
2
(t)
r
1
(t)
r
2
(t)
b
N
2
b
1
b
N
b
2
b
1
a
1
a
2
a
N
s
1
(t)
a
N
a
2
a
1
e
j(2f
N
t +
N
)
e
j(2f
2
t +
2
)
e
j(2f
1
t +
1
)
BEATS/CUBAN Workshop 2004
27/37
Performance Evaluation
Comparison of the time ACFs r
h
11
() (reference model) and r
h
11
() (simulation model)
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
0.5
0
0.5
1
Time separation, (s)
T
i
m
e
a
u
t
o
c
o
r
r
e
l
a
t
i
o
n
f
u
n
c
t
i
o
n
Reference model
Simulation model
Simulation
max
Reference model: Based on the onering model with uniformly distributed AOAs
MS
.
Simulation model: Designed by using the L
p
norm method (N = 25 , p = 2,
max
= 0.08 s, f
max
= 91 Hz).
BEATS/CUBAN Workshop 2004
28/37
Reference model Simulation model
2D space CCF (
BS
,
MS
) 2D space CCF (
BS
,
MS
)
0
5
10
15
20
25
30
0
0.5
1
1.5
2
2.5
3
0.5
0
0.5
1
2
D
s
p
a
c
e
c
r
o
s
s
c
o
r
r
e
l
a
t
i
o
n
f
u
n
c
t
i
o
n
0
5
10
15
20
25
30
0
0.5
1
1.5
2
2.5
3
0.5
0
0.5
1
2
D
s
p
a
c
e
c
r
o
s
s
c
o
r
r
e
l
a
t
i
o
n
f
u
n
c
t
i
o
n
MS
/
BS
/
MS
/
BS
/
Reference model: Based on the onering model with uniformly distributed AOAs
MS
.
Simulation model: Designed by using the Lpnorm method with N = 25
(p = 2,
BS
max
= 30,
MS
max
= 3,
BS
=
MS
= 90
,
BS
max
= 2
).
BEATS/CUBAN Workshop 2004
29/37
IV. Simulation of MIMO Channels
Channel capacity:
C(t) := log
2
det
_
I
2
+
S
BS, total
2N
noise
H(t)
H
H
(t)
_
[bits/s/Hz]
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
1
2
3
4
5
6
7
8
9
10
11
Time, t (s)
S
i
m
u
l
a
t
e
d
c
h
a
n
n
e
l
c
a
p
a
c
i
t
y
,
(
b
i
t
s
/
s
/
H
z
)
Capacity
Mean Capacity
Parameters: N = 25, M
BS
= M
MS
= 2,
BS
=
MS
= /2, SNR = 17 dB
BEATS/CUBAN Workshop 2004
30/37
Scattering Scenario
Geometrical onering model:
scatterers
Cluster of
BS
y
x
MS
BS
max
v
Parameters:
M
BS
= M
MS
= 2
BS
=
MS
= /2
BS
max
= 2
v
=
MS
0
= 180
f
max
= 91 Hz
Reference model: assuming the von Mises distribution for the AOA
MS
:
p(
MS
) =
1
2I
0
()
e
cos(
MS
MS
0
)
= 0 : p(
MS
) = 1/(2) (isotropic scattering)
: p(
MS
) (
MS
MS
0
) (extremely nonisotropic scattering)
Simulation model: designed by using the LPNM with N = 25.
BEATS/CUBAN Workshop 2004
31/37
PDF of the Channel Capacity
C(t)
0 2 4 6 8 10
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
P
D
F
o
f
t
h
e
c
a
p
a
c
i
t
y
C
(
t
)
Level, r
= 0
= 10
= 100
/
Observations: inuences the PDF p
C
(r) of the capacity
C(t)
If ; angular spread of the AOA
MS
; Var
C(t)
BEATS/CUBAN Workshop 2004
32/37
Normalized LCR of the Channel Capacity
C(t)
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Level, r
N
C
(
r
)
/
f
m
a
x
= 0
= 3
= 5
= 10
= 100
/
Observations: has a strong inuence on the LCR
N
C
(r) of the capacity
C(t)
If ; angular spread of the AOA
MS
;
N
C
(r)
BEATS/CUBAN Workshop 2004
33/37
Normalized ADF of the Channel Capacity
C(t)
0 2 4 6 8 10
0
50
100
150
200
250
300
350
400
450
500
Level, r
T
C
(
r
)
f
m
a
x
= 0
= 10
= 100
/
Observations: has a strong inuence on the ADF
T
C
(r) of the capacity
C(t)
If ; angular spread of the AOA
MS
;
T
C
(r) .
BEATS/CUBAN Workshop 2004
34/37
The Eect of on the BER Performance
4 5 6 7 8 9 10
10
4
10
3
10
2
10
1
10
0
E
b
/N
0
(dB)
B
i
t
e
r
r
o
r
r
a
t
e
= 0
= 20
= 30
= 40
Coding system: 4D 16QAM linear 32state spacetime trellis code with 2 transmit and 2 receive
antennas (
BS
= 30 and
MS
= 3).
Observations:
; angular spread ; BER
If decreases from 40 to 0, then a gain of ca. 1.5 dB can be achieved at a BER
of 2 10
3
.
BEATS/CUBAN Workshop 2004
35/37
The Eect of the Antenna Spacing on the BER Performance
4 5 6 7 8 9 10
10
4
10
3
10
2
10
1
10
0
E
b
/N
0
(dB)
B
i
t
e
r
r
o
r
r
a
t
e
BS
/ = 1/2,
MS
/ = 1/2
BS
/ = 1 ,
MS
/ = 1
BS
/ = 30 ,
MS
/ = 3
BS
/ = 30 ,
MS
/ = 1/2
BS
/ = 1/2,
MS
/ = 3
BS
/ = 3 ,
MS
/ = 1
Coding system: 4D 16QAM linear 32state spacetime trellis code with 2 transmit and 2 receive
antennas.
Observations:
BS
,
MS
; spatial correlation ; BER
A large spacing among the antennas at the BS provides a better performance
than a large antenna spacing at the MS.
BEATS/CUBAN Workshop 2004
36/37
VI. Conclusion
a
The principle of deterministic channel modelling has been generalized.
a
The generalized concept has been applied on the geometrical onering scattering model.
a
It has been shown how the parameters of the simulation model can be determined by using the
L
p
norm method (LPNM).
a
The designed deterministic MIMO channel simulator with N = 25 scatterers has nearly the
same statistics as the underlying stochastic reference model with N = scatterers.
a
The new MIMO spacetime channel simulator is very useful in the investigation of the PDF,
LCR, and ADF of the channel capacity C(t) from timedomain simulations.
a
Finally, it has been demonstrated that the system performance decreases if the spatial correlation
increases.
BEATS/CUBAN Workshop 2004
37/37
Much more than documents.
Discover everything Scribd has to offer, including books and audiobooks from major publishers.
Cancel anytime.