EE359 Discussion Session 3 Capacity of Flat and Frequency ...p g[i] ˘fading distribution E[jx[i]j2]...

Post on 13-Mar-2020

4 views 0 download

Transcript of EE359 Discussion Session 3 Capacity of Flat and Frequency ...p g[i] ˘fading distribution E[jx[i]j2]...

EE359 Discussion Session 3Capacity of Flat and Frequency Selective Channels

February 5, 2020

EE359 Discussion 3 February 5, 2020 1 / 25

Note about scattering functions

EE359 Discussion 3 February 5, 2020 2 / 25

For deterministic response

h(t, τ) =∑i

αi(t)e−j2π(t−τi(t))δ(t− τi(t))

With no movement (Doppler), h(t, τ)→ h(τ) becomes time-invariantresponse (LTI system)

t (Time) ρ (Doppler)

τ (Delay) f (Frequency)

EE359 Discussion 3 February 5, 2020 3 / 25

For deterministic response

h(t, τ) =∑i

αi(t)e−j2π(t−τi(t))δ(t− τi(t))

With no movement (Doppler), h(t, τ)→ h(τ) becomes time-invariantresponse (LTI system)

t (Time) ρ (Doppler)

τ (Delay) f (Frequency)

EE359 Discussion 3 February 5, 2020 3 / 25

What is Capacity?

Maximum achievable rate with no errors

Maximum mutual information over all input distributions

Usually found by proving matching lower (achievability) and upper(converse) bounds

Often easy to bound, but hard to prove

EE359 Discussion 3 February 5, 2020 4 / 25

Notions of Capacity in Wireless Systems

AWGN Capacity

Only CSI distribution known at TX and RX

CSI at RX onlyI With or without outage

CSI at TX and RXI With or without adaptationI With or without outage

EE359 Discussion 3 February 5, 2020 5 / 25

AWGN Capacity

Capacity of fixed channel with no fading.

C = B log2(1 + γ)

Can be used to bound other settings

EE359 Discussion 3 February 5, 2020 6 / 25

CSI Distribution Known

Really complicated

EE359 Discussion 3 February 5, 2020 7 / 25

Channel capacity with CSIR

Two notions of capacity

Ergodic capacity

C = B

∫log2 (1 + γ) p(γ)dγ

Achieved by coding over fading states γ

Outage capacity

Find out minimum SNR γmin needed to achieve outage prob Pout. Ifit violates power constraint then C = 0

Transmit at that SNR thereby achieving

C = (1− Pout)B log2(1 + γmin)

EE359 Discussion 3 February 5, 2020 8 / 25

Channel capacity with CSIR

Two notions of capacity

Ergodic capacity

C = B

∫log2 (1 + γ) p(γ)dγ

Achieved by coding over fading states γ

Outage capacity

Find out minimum SNR γmin needed to achieve outage prob Pout. Ifit violates power constraint then C = 0

Transmit at that SNR thereby achieving

C = (1− Pout)B log2(1 + γmin)

EE359 Discussion 3 February 5, 2020 8 / 25

Outage capacity (without CSIT)

0 2 4 6 8 10

−1

0

1

time

γ

Outage event

γ

Outage capacity

Find out minimum SNR γ0

needed to achieve outageprob Pout. If it violatespower constraint then C = 0

Transmit at that SNRthereby achieving

C = (1−Pout)B log2(1+γ0)

EE359 Discussion 3 February 5, 2020 9 / 25

Capacity with CSIR and CSIT

System model

y[i] =√g[i]x[i] + n[i]√

g[i] ∼ fading distribution

E[|x[i]|2] ≤ P̄

g[i] known at transmitter and receiver

What is capacity with fixed TX power?

Can capacity increase with rate and power adaptation?

EE359 Discussion 3 February 5, 2020 10 / 25

Capacity with CSIR and CSIT

System model

y[i] =√g[i]x[i] + n[i]√

g[i] ∼ fading distribution

E[|x[i]|2] ≤ P̄

g[i] known at transmitter and receiver

What is capacity with fixed TX power?

Can capacity increase with rate and power adaptation?

EE359 Discussion 3 February 5, 2020 10 / 25

Capacity with CSIR and CSIT

System model

y[i] =√g[i]x[i] + n[i]√

g[i] ∼ fading distribution

E[|x[i]|2] ≤ P̄

g[i] known at transmitter and receiver

What is capacity with fixed TX power?

Can capacity increase with rate and power adaptation?

EE359 Discussion 3 February 5, 2020 10 / 25

Towards optimally exploiting the CSI g[i]

Idea

Vary transmit power P and rate R as a function of g[i] or equivalently, of

γ = g[i]P̄N0B

Power P (γ)

Rate R = B log(

1 + γ P (γ)P̄

)

Notation

Fading distribution given by p(γ)

Assuming that Tx uses fixed average power P̄ .

EE359 Discussion 3 February 5, 2020 11 / 25

Towards optimally exploiting the CSI g[i]

Idea

Vary transmit power P and rate R as a function of g[i] or equivalently, of

γ = g[i]P̄N0B

Power P (γ)

Rate R = B log(

1 + γ P (γ)P̄

)Notation

Fading distribution given by p(γ)

Assuming that Tx uses fixed average power P̄ .

EE359 Discussion 3 February 5, 2020 11 / 25

Optimization problems for optimal CSIT and CSIR use

Optimization problem

maxP (γ)

∫B log

(1 +

P (γ)

P̄γ

)p(γ)dγ

s.t. E[P (γ)] ≤ P̄P (γ) ≥ 0 ∀ γ

Optimization problem (discrete γ)

maxP(γ)

∑i

B log

(1 +

P (γi)

P̄γi

)p(γi)

s.t. E[P (γ)] ≤ P̄P (γi) ≥ 0 ∀ i

EE359 Discussion 3 February 5, 2020 12 / 25

Optimization problems for optimal CSIT and CSIR use

Optimization problem

maxP (γ)

∫B log

(1 +

P (γ)

P̄γ

)p(γ)dγ

s.t. E[P (γ)] ≤ P̄P (γ) ≥ 0 ∀ γ

Optimization problem (discrete γ)

minP(γ)

∑i

−B log

(1 +

P (γi)

P̄γi

)p(γi)

s.t. E[P (γ)] ≤ P̄P (γi) ≥ 0 ∀ i

EE359 Discussion 3 February 5, 2020 13 / 25

Lagrangian Methods

Optimization problem

minx∈S

f0(x),

s.t. fi(x) ≤ 0, ∀i = 0, . . . ,m,

s.t. hi(x) = 0, ∀i = 0, . . . , p,

Lagrangian

L(x, λ, ν) = f0(x) +

m∑i=1

λifi(x) +

p∑i=1

νihi(x)

ν, λ are called Lagrange multipliers or dual variables

Process sometimes refered to as “regularlizing constraints”.

EE359 Discussion 3 February 5, 2020 14 / 25

Lagrangian Duality

Dual problem

g(λ, ν) = infx∈S

L(x, ν, λ)

= infx∈S

(f0(x) +

m∑i=1

λifi(x) +

p∑i=1

νihi(x)

)

Solving the dual problem

Find closed form expression for g(λ, ν): ∇xL(x, λ, ν) = 0

g(λ, ν) is now a convex optimization problem (in our case of onevariable)!

EE359 Discussion 3 February 5, 2020 15 / 25

Why consider dual problem?

Let p? be the optimal value of the original optimization problem:

f0(x̃) = p?, x̃ is feasible.

Let d? be the optimal value of the dual problem:

g(λ̃, ν̃) = d?

Lower Bound Property

p? ≥ d? as long as λ ≥ 0.

Strong Duality

p? = d? for many problems! True in our case

Take EE364a/b to understand when and why this is true

EE359 Discussion 3 February 5, 2020 16 / 25

Example: Least squares

min xᵀx

s.t Ax = b

Dual Function

L(x, v) = xᵀx+ νᵀ(Ax− b)∇xL(x, ν) = 2x+Aᵀν = 0⇒ x = −(1/2)Aᵀν.

Plug into L to get g:

g(ν) = L((−1/2))Aᵀν, ν) = −(1/4)νᵀAAᵀν − bᵀν

Lower Bound Property

p? ≥ −(1/4)νᵀAAᵀν − bᵀν ∀ν

EE359 Discussion 3 February 5, 2020 17 / 25

Example: Least squares

min xᵀx

s.t Ax = b

Dual Function

L(x, v) = xᵀx+ νᵀ(Ax− b)∇xL(x, ν) = 2x+Aᵀν = 0⇒ x = −(1/2)Aᵀν.

Plug into L to get g:

g(ν) = L((−1/2))Aᵀν, ν) = −(1/4)νᵀAAᵀν − bᵀν

Lower Bound Property

p? ≥ −(1/4)νᵀAAᵀν − bᵀν ∀ν

EE359 Discussion 3 February 5, 2020 17 / 25

Example: Least squares

min xᵀx

s.t Ax = b

Dual Function

L(x, v) = xᵀx+ νᵀ(Ax− b)∇xL(x, ν) = 2x+Aᵀν = 0⇒ x = −(1/2)Aᵀν.

Plug into L to get g:

g(ν) = L((−1/2))Aᵀν, ν) = −(1/4)νᵀAAᵀν − bᵀν

Lower Bound Property

p? ≥ −(1/4)νᵀAAᵀν − bᵀν ∀ν

EE359 Discussion 3 February 5, 2020 17 / 25

Deriving Waterfilling Expression

1 Form Lagrangian by relaxing Pj > 0:

minPj

∑j

−B log

(1 +

PjP̄γj

)p(γj)→ f0(Pj)∑

j Pj

P̄≤ 1⇒

∑j

Pj − P̄ = 0→ fj(Pj)

2 Solve ∂L/∂Pj = 0 for Pj/P . Let γ0 = λP .

Hint: Write Lagrangian in terms of Pj , P̄ , γ before taking derivative.

EE359 Discussion 3 February 5, 2020 18 / 25

Deriving Waterfilling Expression

1 Form Lagrangian by relaxing Pj > 0:

minPj

∑j

−B log

(1 +

PjP̄γj

)p(γj)→ f0(Pj)∑

j Pj

P̄≤ 1⇒

∑j

Pj − P̄ = 0→ fj(Pj)

2 Solve ∂L/∂Pj = 0 for Pj/P . Let γ0 = λP .

Hint: Write Lagrangian in terms of Pj , P̄ , γ before taking derivative.

EE359 Discussion 3 February 5, 2020 18 / 25

Deriving Waterfilling Expression

1 Form Lagrangian by relaxing Pj > 0:

minPj

∑j

−B log

(1 +

PjP̄γj

)p(γj)→ f0(Pj)∑

j Pj

P̄≤ 1⇒

∑j

Pj − P̄ = 0→ fj(Pj)

2 Solve ∂L/∂Pj = 0 for Pj/P . Let γ0 = λP .

Hint: Write Lagrangian in terms of Pj , P̄ , γ before taking derivative.

EE359 Discussion 3 February 5, 2020 18 / 25

Extensions of waterfilling idea

Can be applied to any system where the sum of logarithms need to beoptimized with a sum power and positivity constraints

Examples

Continuous fading states

Time-invariant frequency selective fading channel - waterfilling overfrequency

Time-varying frequency selective fading channel - waterfilling overtime and frequency (may not be optimal)

MIMO channels - waterfilling over spatial diversity

EE359 Discussion 3 February 5, 2020 19 / 25

Block fading vs. frequency-selective fading

Block Fading: ∑γi≥γ0

(1

γ0− 1

γi

)p(γi) = 1

Frequency-selective Fading:∑γi≥γ0

(1

γ0− 1

γi

)= 1

EE359 Discussion 3 February 5, 2020 20 / 25

Finding the optimal power allocation

1 Assume P (γi) > 0 ∀i

2 Solve for γ0:

1

γ0= 1 +

∑i

p(γi)

γior

N

γ0= 1 +

∑i

1

γi

3 If any P (γi) < 0 assume the P (γi) for lowest γi is zero and repeatprevious step

4 Given γ0, can compute P (γi) and C.

EE359 Discussion 3 February 5, 2020 21 / 25

“Waterfilling” interpretation of the solution P (γi)

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0γi

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0P(γ

i)

Cutoff γ0

Figure: “Waterfilling” with cutoff γ0EE359 Discussion 3 February 5, 2020 22 / 25

Suboptimal power adaptation schemes

Power adaptation buys you little in practice.

Other Ideas

Fix TX power but vary rate

Channel inversion: Received SNR is constant

Truncated channel inversion: Received SNR is constant, and do notuse channel if gain is too low

EE359 Discussion 3 February 5, 2020 23 / 25

Channel inversion in more details

Channel inversion

If target SNR is σ, transmit at σγ

Expected power constraint gives σ = P̄E[1/γ]

What happens for Rayleigh fading?

E[1/γ] =∞

Truncated channel inversion

Do not use the channel if γ < γ1 (outage)

If target SNR is σ, transmit at σγ

Expected power constraint gives σ = P̄Eγ>γ1 [1/γ]

EE359 Discussion 3 February 5, 2020 24 / 25

Channel inversion in more details

Channel inversion

If target SNR is σ, transmit at σγ

Expected power constraint gives σ = P̄E[1/γ]

What happens for Rayleigh fading? E[1/γ] =∞

Truncated channel inversion

Do not use the channel if γ < γ1 (outage)

If target SNR is σ, transmit at σγ

Expected power constraint gives σ = P̄Eγ>γ1 [1/γ]

EE359 Discussion 3 February 5, 2020 24 / 25

Question

How does capacity under truncated inversion behave with increasingoutage probability ?

EE359 Discussion 3 February 5, 2020 25 / 25