HAL Id: hal-02439210
https://hal.archives-ouvertes.fr/hal-02439210
Submitted on 14 Jan 2020
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
On the Capacity of MIMO Optical Wireless Channels
Longguang Li, Stefan Moser, Ligong Wang, Michèle Wigger
To cite this version:
Longguang Li, Stefan Moser, Ligong Wang, Michèle Wigger. On the Capacity of MIMO Optical Wireless Channels. 2018 IEEE Information Theory Workshop (ITW), Nov 2018, Guangzhou, China.
pp.1-5, �10.1109/itw.2018.8613496�. �hal-02439210�
On the Capacity of MIMO Optical Wireless Channels
Longguang Li
∗, Stefan M. Moser
†‡, Ligong Wang
§, and Michèle Wigger
∗∗
LTCI, Telecom ParisTech, Université Paris-Saclay, 75013 Paris, France
†
Signal and Information Processing Lab, ETH Zürich, Switzerland
‡
Institute of Communications Engineering, National Chiao Tung University, Hsinchu, Taiwan
§
ETIS—Université Paris Seine, Université de Cergy-Pontoise, ENSEA, CNRS, Cergy-Pontoise, France
Abstract—This paper investigates the capacity of the multiple- input multiple-output free-space optical intensity channel under a per-input-antenna peak-power constraint and a total average- power constraint over all input antennas. Our work considers the setup with more transmit than receive antennas, and charac- terizes capacity as an alternative optimization problem over the distribution of the input vector times the channel matrix. This alternative capacity expression is then used to obtain upper and lower bounds on the capacity, which match asymptotically in the high signal-to-noise ratio regime.
I. I
NTRODUCTION ANDC
HANNELM
ODELConsider a wireless optical intensity-modulation direct- detection (IM-DD) system where the transmitter is equipped with n
Tlight-emitting diodes (LEDs) and the receiver with n
Rphotodetectors. The photodetectors receive a superposition of the signals sent by the LEDs, and we assume that the crosstalk between the LEDs is constant. Hence, the channel output is
Y = Hx + Z, (1) where x = (x
1, . . . , x
nT)
Tdenotes the n
T-dimensional channel input vector, Z the n
R-dimensional noise vector with inde- pendent standard Gaussian entries, and H the deterministic n
R-by-n
Tchannel matrix with nonnegative entries, which we also write in the form
H = [h
1, h
2, . . . , h
nT]. (2) The channel inputs correspond to optical intensities sent by the LEDs, hence they are nonnegative:
x
k∈ R
+0, k = 1, . . . , n
T. (3) We assume the inputs are subject to a peak-power (peak- intensity) and an average-power (average-intensity) constraint:
Pr
X
k> A
= 0, ∀ k ∈ { 1, . . . , n
T} , (4a) E
k X k
1≤ E (4b)
for some fixed parameters A, E > 0, where k · k
1denotes the L
1-norm. Note that the average-power constraint is on the expectation of the channel input and not on its square. Also note that A describes the maximum power of each single LED, while E describes the allowed total average power of all LEDs together. We denote the ratio between the allowed average power and the allowed peak power by α:
α , E
A , (5)
where 0 < α ≤ n
T. For α ≥
n2T, by symmetry of the setup, the average-power constraint is inactive. Thus, when α ≥
n2T, the channel essentially reduces to the case where there is only a peak-power constraint.
In previous works, bounds on the capacity of the above- described wireless optical channel were derived for the single- antenna case (n
T= n
R= 1) [1]–[4], the no-crosstalk case (H is diagonal) [5], the multiple-input single-output (MISO) case (n
T> 1 and n
R= 1) [6], [7], the case with a full-rank square H (n
T= n
R> 1) [8], [9], and the case with a full-column-rank H [8].
In this paper we focus on the multiple-input multiple-output (MIMO) setup with more transmit than receive antennas:
n
T> n
R> 1. (6)
Inspired by our results on the MISO channel [7], we find the most energy-efficient signaling method for the setup (6) (Lemma 1 ahead). In the MISO case, antennas can be ordered according to decreasing channel gains, and the optimal signal- ing strategy is to rely as much as possible on stronger antennas.
In other words, if an antenna is used for active signaling in a channel use, then all stronger antennas should transmit at full power A and all weaker antennas should be silenced.
In the MIMO case, there is no clear order for the transmit antennas. Nevertheless, we show that it is optimal to restrict to signaling methods that, for each channel use, choose n
Rtransmit antennas as “active” signaling antennas, and set the remaining antennas to either full power A or to 0 according to a given rule. The “active” antennas can then be used in the same way as suggested in earlier works for MIMO channels that have a square full-rank channel matrix [8].
We introduce some further notation. We define U ,
I ⊆ { 1, . . . , n
T} : |I| = n
R, (7) and for every I = { i
1, . . . , i
nR} ∈ U with i
1< · · · < i
nR, we define
H
I,
h
i1, . . . , h
inR. (8) Throughout this paper, we assume that every H
I, I ∈ U , is of full rank n
R, i.e., that any n
Rcolumn vectors in H are linearly independent. To simplify derivations, we further assume that for all I ∈ U and j ∈ { 1 , . . . , n
T} \ I ,
1
TnRH
−I1h
j6 = 1, (9)
where 1
nRdenotes the n
R-dimensional vector (1, . . . , 1)
T.
1II. C
APACITY ANDM
INIMUM-E
NERGYS
IGNALINGThe capacity of the channel (1) is [10]
C
H(A, αA) = sup
QX
I(X; Y) (10) where the supremum is over all distributions Q
Xon X satisfying (3) and (4). Define
X ¯ , HX (11) and notice that X ¯ takes value in the zonotope
R (H) , (
nTX
k=1
λ
kh
k: λ
1, . . . , λ
nT∈ [0, A]
)
, (12) which has volume [11]:
V
H, A
nRX
I∈U
| det H
I| . (13) Define for each I ∈ U the parallelepiped
D
I, (
X
i∈I
λ
ih
i: λ
i∈ [0, A], ∀ i ∈ I )
; (14) the coefficient
s
I, X
j∈{1,...,nT}\I
1
1
TnRH
−I1h
j> 1 ; (15) and the vector
v
I, A X
j∈{1,...,nT}\I
1
1
TnRH
−I1h
j> 1 h
j; (16) where 1 {·} denotes the indicator function.
Lemma 1:
1) For any I , J ∈ U , I 6 = J , the intersection of v
I+ D
Iand v
J+ D
Jhas Lebesgue measure zero.
2) The image set R (H) is covered by the parallelepipeds v
I+ D
I I∈U,
[
I∈U
v
I+ D
I= R (H). (17)
3) For any I ∈ U , any point in v
I+ D
Iis achieved with minimum total input power by an input vector x satisfying, for every j ∈ { 1 , . . . , n
T} \ I ,
x
j= A · 1
1
TnRH
−1Ih
j> 1 . (18) Proof: Omitted.
Figure 1 shows the partitions of R (H) into the union (17) for two examples of channel matrices H.
Using the above lemma, we can characterize the capacity C
H(A, E) as an optimization problem over X. ¯
Proposition 2: The capacity of the MIMO optical intensity channel is given by
C
H(A, αA) = sup
QX¯
I( ¯ X; Y) , (19)
1In the full version of this paper, which is under preparation, we show that the condition (9) can be dropped.
h1
h2
h3
h4 D{1,2}
v{1,3}+D{1,3}
v{1,4}+D{1,4}
h1
h2
h3
h4 D{1,2}
v{1,3}+D{1,3}
v{1,4}+D{1,4}
R(H) R(H)
D{2,3}
Fig. 1. Partitioning of R(H)into the union (17) for two 4×2 MIMO examples. The example on the left is forH= [7,5,2,1; 1,2,2.9,3]and the example on the right forH= [7,5,2,1; 1,3,2.9,3].
where the supremum is over all distributions Q
X¯over R (H) subject to the power constraint
E
UAs
U+
H
−1UE X ¯ U
− v
U1
≤ αA, (20) where U is a random variable taking value in U such that U = I if, and only if, X ¯ lies in the interior of v
I+ D
I.
2Proof: Notice that X ¯ is a function of X and that the Markov chain X ( − − X ¯ ( − − Y holds. Therefore, I ( ¯ X; Y) = I(X; Y). By Lemma 1, R (H) can be decomposed into the shifted parallelepipeds { v
I+ D
I} and the minimum energy required to achieve an image point ¯ x ∈ v
I+ D
Iis:
As
I+ k H
−1I(¯ x − v
I) k
1. (21) The proposition follows by taking expectation over (21).
III. C
APACITYR
ESULTSThis section presents bounds on capacity, as well as asymp- totic expressions at high signal-to-noise ratios (SNR). Let q be a probability vector on U with entries
q
I, | det H
I| V
H, I ∈ U . (22) Further define
α
th, n
R2 + X
I∈U
s
Iq
I. (23)
We shall see that the threshold α
thdetermines whether X ¯ can be made uniform over R (H) or not.
A. Lower Bounds
Theorem 3: For α ≥ α
th, C
H(A, αA) ≥ 1
2 log
1 + A
2nRV
2H(2πe)
nR. (24)
For α < α
th,
C
H( A, αA ) ≥ 1 2 log
1 + A
2nRV
2H(2πe)
nR· e
2ν, (25)
2To make the statement simpler, here we require thatX¯ be in the interior of somevI+DI. By the supremum in (19), this restriction causes no loss of optimality.
where
ν , sup
λ∈
(
max{0,n2R+α−αth},min{n2R,α})
n
R1 − log µ
1 − e
−µ− µ e
−µ1 − e
−µ− inf
p
D(p k q)
, (26) where µ is the unique solution to the following equation:
1
µ − e
−µ1 − e
−µ= λ
n
R, (27)
and where the infimum is over all probability vectors p on U such that
X
I∈U
p
Is
I= α − λ. (28)
B. Upper Bounds
For each I ∈ U and ` ∈ { 1 , . . . , n
R} , let σ
I,`be the square root of the `-th diagonal entry of the matrix H
−1IH
−IT. Define
σ
max, max
I∈U,
`∈{1,...,nR}
σ
I,`and σ
min, min
I∈U,
`∈{1,...,nR}
σ
I,`. (29)
Theorem 4: For an arbitrary α, C
H(A, αA) ≤ n
Rlog
σ
max+ A
√ 2πe
+ log V
H. (30) Theorem 5: For α < α
th,
C
H(A, αA) ≤ sup
p µ>0
inf
( n
Rlog
σ
max+ A
√ 2πe
1 − e
−µµ
+ log V
H+ µn
Rσ
maxA √ 2π
1 − e
− A2 2σmin
+ µ α − X
I∈U
p
Is
I!
− D(p k q) )
, (31) where the supremum is over all probability vectors p on U .
Theorem 6: For α < α
th, C
H(A, αA)
≤ sup
p δ,µ>0
inf
(
n
Rlog A
σ
min· e
µδA− e
−µ(1+Aδ)√ 2πeµ(1 − 2 Q (δ))
!
+ log V
H+ n
RQ δ
σ
max+ δ
√ 2πσ
mine
− δ2 2σmax
+ µn
Rσ
maxA √
2 π
e
− δ2
2σmax
− e
−(A+δ)22σmin+ µ α − X
I∈U
p
Is
I!
− D(p k q) )
, (32)
where Q ( · ) denotes the Q-function associated with the stan- dard Gaussian distribution, and the supremum is over all probability vectors p on U .
C. Asymptotic High-SNR Capacity Expressions Theorem 7: For α ≥ α
th,
A
lim
→∞C
H(A, αA) − n
Rlog A = 1 2 log
V
2H(2πe)
nR. (33) For α < α
th,
A
lim
→∞C
H(A, αA) − n
Rlog A
= 1 2 log
V
2H(2πe)
nR+ sup
λ∈
(
max{0,n2R+α−αth},min{n2R,α})
n
R1
− log µ
1 − e
−µ− µ e
−µ1 − e
−µ− inf
p
D(p k q)
, (34) where µ and p are the same as in Theorem 3.
Figure 2 depicts the derived lower and upper bounds for an example with α < α
th. The bounds match asymptotically at high SNR. For this example, upper bound (31) is always tighter than (32). That this does not hold in general can be seen in [12, Examples 2 and 3], which compare the two bounds in some single-input single-output and MISO special cases.
5 7 9 11 13 15 17 19 21 23 25
A (dB) 1
2 3 4 5 6 7 8 9 10 11
Capacity (nats per channel use)
Upper Bound (32), α = 0.9 Upper Bound (33), α = 0.9 Lower Bound (27), α = 0.9
Fig. 2. Bounds on capacity of 3×2MIMO channel with channel matrix H= [1,1.5,3; 2,2,1], and average-to-peak power ratioα= 0.9. Note that the threshold of the channel isαth= 1.476.
IV. P
ROOFSA. Derivation of Lower Bounds
For any choice of the random vector X ¯ over R (H) that satisfies (20), the following holds:
C
H(A, αA) ≥ I( ¯ X; ¯ X + Z) (35)
= h( ¯ X + Z) − h(Z) (36)
≥ 1 2 log
e
2h( ¯X)+ e
2h(Z)− h(Z) (37)
= 1
2 log 1 + e
2h( ¯X)(2πe)
nR!
, (38)
where (37) holds by the Entropy Power Inequality [13].
The lower bound (24) follows from (38) and by choosing X ¯ to be uniform over R (H). It can be verified that this choice satisfies the power constraint whenever α ≥ α
th.
If α < α
th, it can be shown that it is always possible to choose λ ∈ max { 0,
n2R+ α − α
th} , min {
n2R, α }
, p to satisfy (28), and µ as the unique solution to (27). For each I we then pick the conditional probability density function f
X¯|U=Ito be the n
R-dimensional product truncated exponen- tial distribution rotated by the matrix H
I:
f
X¯|U=I(¯ x) = 1 A
nR| det H
I| ·
µ 1 − e
−µ nRe
−µkH−1 I (¯x−vI)k1
A
.
(39) Note that this corresponds to the entropy-maximizing distri- bution under a total average-power constraint on H
−I1X. The ¯ resulting expression is finally optimized over the distribution Q
U.
B. Derivation of Upper Bounds
For any X ¯
?, we bound the mutual information in (19) as I( ¯ X
?; Y
?) = I( ¯ X
?; ¯ X
?+ Z) (40)
≤ I( ¯ X
?; ¯ X
?+ Z, U
?) (41)
= H(U
?) + I( ¯ X
?; ¯ X
?+ Z | U
?). (42) Moreover, for each size-n
Rsubset I ∈ U ,
I( ¯ X
?; ¯ X
?+ Z | U
?= I )
= I X ¯
?− v
I; ( ¯ X
?− v
I) + Z
U
?= I
(43)
= I H
−1I( ¯ X
?− v
I); H
−1I( ¯ X
?− v
I) + H
−1IZ
U
?= I (44)
= I( ¯ X
?I; ¯ X
?I+ Z
I| U
?= I ), (45) where we defined
Z
I, H
−I1Z, (46) X ¯
?I, H
−I1( ¯ X
?− v
I). (47) To further bound the term in (45), we then use the duality upper-bounding technique with a product output distribution
R
I(y
I) =
nR
Y
`=1
R
I,`( y
I,`) . (48) Denoting by W
I( ·| X ¯
?I) the transition law of the MIMO channel X ¯
?I7→ Y
I, ( ¯ X
?I+ Z
I) and by W
I,`( ·| X ¯
I,`?) the marginal transition law for its `-th component, we have:
I( ¯ X
?I; ¯ X
?I+ Z
I| U
?= I )
≤ E
X¯?I|U?=ID W
I( ·| X ¯
?I) R
I( · )
(49)
≤ − h X ¯
?I+ Z
IX ¯
?I, U
?= I
− E
X¯?I|U?=I"
nRX
`=1
E
WI(YI|X¯?I)[log R
I,`(Y
I,`)]
# (50)
= − n
R2 log 2πe + log | det H
I|
−
nR
X
`=1
E
QX?¯I,`|U ?=Ih
E
WI,`(YI,`|X¯I,`? )[log R
I,`(Y
I,`)] i , (51)
where the last equality holds because h( ¯ X
?I+ Z
I| X ¯
?I, U
?= I )
= h(Z
I) = 1
2 log (2πe)
nRdet H
−I1H
−IT. (52)
The upper bounds in this paper are then obtained by combining (42) with (45) and (51), by optimizing over the probability vector p on U , and by picking appropriate choices for the output distribution R
I,`( · ).
To prove Theorem 4, we choose
R
I,`(y) =
1 A+√
2πeσI,`
e
−y2
2σI,`
if y < 0 ,
1 A+√
2πeσI,`
if y ∈ [0, A],
1 A+√
2πeσI,`
e
−(y−A)2
2σI,`
if y > A,
(53)
which yields, irrespectively of the value of x ¯
?I,`:
− E
WI,`(YI,`|x¯?I,`)[log R
I,`(Y
I,`)] ≤ log A + √
2πeσ
I,`. (54) To prove Theorem 5, we choose
R
I,`(y) =
µ√ e A(1−e−µ)+µ√
2πeσI,`
e
−y2 2σ2
I,`
if y < 0,
µ A(1−e−µ)+µ√
2πeσI,`
e
−µyAif y ∈ [0, A],
µ√ e A(1−e−µ)+µ√
2πeσI,`
e
−(y−A)2 2σ2
I,`
if y > A.
(55) Following the steps in [12, Appendix B], we obtain:
− E
WI,`(YI,`|x¯?I,`)[log R
I,`(Y
I,`)]
≤ log √
2πeσ
I,`+ A 1 − e
−µµ
+ µσ
I,`A √
2π 1 − e
−A2 2σ2
I,`
! + µ
A ¯ x
?I,`. (56) The proof is concluded by combining (56) with (42), (45), and (51), and by noting that
µ A E
QU ?"
nRX
`=1
E
QX?¯U,`|U ?X ¯
U??,`#
= µ
A E
QU ?H
−U1?E X ¯
?U
?− v
U?1
(57)
≤ µ α − E
QU ?[ s
U?]
, (58)
where (57) follows by (47), and (58) by (20).
To prove Theorem 6, we choose
R
I,`(y) =
√ 1
2πσI,`
e
−y2 2σ2
I,`
y < − δ,
µ
A
·
1−2Qδ
σI,`
eµδA−e−µ(1+Aδ)
e
−µyAy ∈ [ − δ, A + δ],
√ 1 2πσI,`
e
−(y−A)2 2σ2
I,`
y > A.
(59)
Following the steps in the proof of [3, Theorem 3], we obtain:
− E
WI,`(YI,`|¯x?I,`)
[log R
I,`( Y
I,`)]
≤ log
A
σ
I,`· e
µδA− e
−µ(1+Aδ)√ 2πeµ(1 − 2 Q
δ σI,`
)
+ Q δ
σ
I,`+ δ
√ 2πσ
I,`e
−δ2 2σ2
I,`
+ µσ
I,`A √ 2π e
−δ2 2σ2
I,`
− e
−(A+δ)2 2σ2
I,`
!
+ µ
A x ¯
?I,`. (60) The proof is concluded by combining (60) with (42), (45), (51), and (58).
C. Derivation of Asymptotic High-SNR Capacity
The proof for the case α ≥ α
this straightforward and omitted.
For α < α
th, the fact that the left-hand side of (34) cannot be smaller than its right-hand side follows directly from (25).
To show the reverse direction, we rely on (31). The proof sketch is as follows. Let
λ = α − X
I∈U
p
Is
I, (61)
and choose µ
∗to be the unique solution of equation (27) in µ.
Then, proceeding from (31), we obtain:
C
H( A, αA ) − n
Rlog A
≤ sup
p
1 2 log
V
2H(2 πe )
nR− D(p k q) + n
Rlog 1 − e
−µ∗µ
∗+ µ
∗λ + o(A)
(62)
= sup
p
1 2 log
V
2H(2πe)
nR− D(p k q) + n
R1 − log µ
∗1 − e
−µ∗− µ
∗e
−µ∗1 − e
−µ∗+ o(A)
. (63) The proof is concluded by taking A → ∞ .
V. C
ONCLUSIONIn this paper, we derive upper and lower bounds on the capacity of the MIMO free-space optical intensity channel when the transmitter has more antennas than the receiver (n
T> n
R). In our model, channel inputs are subject to a peak-power constraint for each antenna and an average- sum-power constraint over all antennas. The bounds match asymptotically in the high-SNR regime and show that the high- SNR asymptotic capacity saturates in the total average power
αA for all α ≥ α
th. The reason is that this threshold suffices to attain a uniform distribution over the image set R (H) produced by the channel matrix H.
To derive our capacity bounds, we provide an alternative expression for the capacity. It is based on the insight that the optimal (most energy-efficient) signaling strategy is to choose a set I ∈ U , set all inputs { x
j: j ∈ { 1, . . . , n
T} \ I} either to 0 or to full power A according to the rule in (18), and signal with the n
Rantennas in D
Ias for a full-rank n
R× n
Rchannel matrix H
I[8]. In the MISO case, this means that, if a given antenna is used for signaling, all stronger antennas need to send at full power A [7].
A
CKNOWLEDGMENTThe work of L. Li and M. Wigger was supported by the ERC under grant agreement 715111.
R
EFERENCES[1] Michèle A. Wigger, “Bounds on the capacity of free-space optical intensity channels,” Master’s thesis, Signal and Inf. Proc. Lab., ETH Zürich, Switzerland, Mar. 2003.
[2] Andrew L. McKellips, “Simple tight bounds on capacity for the peak- limited discrete-time channel,” inProc. IEEE Int. Symp. Inf. Theory, Chicago, IL, USA, Jun. 27 – Jul. 2, 2004, p. 348.
[3] Amos Lapidoth, Stefan M. Moser, and Michèle A. Wigger, “On the capacity of free-space optical intensity channels,” IEEE Trans. Inf.
Theory, vol. 55, no. 10, pp. 4449–4461, Oct. 2009.
[4] Andrew Thangaraj, Gerhard Kramer, and Georg Böcherer, “Capacity bounds for discrete-time, amplitude-constrained, additive white Gaussian noise channels,”IEEE Trans. Inf. Theory, vol. 63, no. 7, pp. 4172–4182, Jul. 2017.
[5] Anas Chaaban, Zouheir Rezki, and Mohamed-Slim Alouini, “Funda- mental limits of parallel optical wireless channels: capacity results and outage formulation,”IEEE Trans. Commun., vol. 65, no. 1, pp. 296–311, Jan. 2017.
[6] Stefan M. Moser, Ligong Wang, and Michèle Wigger, “Asymptotic high- SNR capacity of MISO optical intensity channels,” inProc. IEEE Inf.
Theory Workshop, Kaohsiung, Taiwan, Nov. 6–10, 2017, pp. 86–90.
[7] Stefan M. Moser, Ligong Wang, and Michèle Wigger, “Capacity results on multiple-input single-output wireless optical channels,”
2018, to app. in IEEE Trans. Inf. Theory. Available: http://moser- isi.ethz.ch/publications.html
[8] Stefan M. Moser, Michail Mylonakis, Ligong Wang, and Michèle Wigger, “Asymptotic capacity results for MIMO wireless optical com- munication,” inProc. IEEE Int. Symp. Inf. Theory, Aachen, Germany, Jun. 25–30, 2017, pp. 536–540.
[9] Anas Chaaban, Zouheir Rezki, and Mohamed-Slim Alouini, “MIMO intensity-modulation channels: Capacity bounds and high SNR char- acterization,” inProc. IEEE Int. Conf. Commun., Paris, France, May 21–25, 2017, pp. 1–6.
[10] Claude E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, pp. 379–423 and 623–656, Jul. and Oct. 1948.
[11] Geoffrey C. Shephard, “Combinatorial properties of associated zono- topes,”Canad. J. Math., vol. 26, no. 2, pp. 302–321, Feb. 1974.
[12] Longguang Li, Stefan M. Moser, Ligong Wang, and Michèle Wigger,
“The MISO free-space optical channel at low and moderate SNR,” in Proc. Annu. Conf. Inf. Sci. Syst., Princeton, NJ, USA, Mar. 21–23, 2018, pp. 1–6.
[13] Thomas M. Cover and Joy A. Thomas,Elements of Information Theory, 2nd ed. New York, NY, USA: Wiley, 2006.