Types of Communication Analog: continuous variables with noise ⇒ P{error = 0} = 0 (imperfect) Digital: decisions, discrete choices, quantized, noise ⇒ P{error} → 0 (usually perfect) Message S1, S2,…or SM Modulator ⇒ v(t) channel; adds noise and distortion M-ary messages, where M can be infinite Received message S1, S2,…or SM Demodulator; hypothesize H1…HN (chooses one) The channel can be radio, optical, acoustic, a memory device (recorder), or other objects of interest as in radar, sonar, lidar, or other scientific observations. Lec14.10-1 2/6/01 A1 Optimum Demodulator for Binary Messages Hypothesis: H1 Message: S1 OK Probability H2 a priori ERROR P1 S2 ERROR vb va E.G. V1 P2 v Demodulator design 2-D case OK vb v measured t V2 va v ∈ V ⇒ 1 " H1" v ∈ V2 ⇒ " H2 " Lec14.10-2 2/6/01 How to define V1,V2? vc A2 Optimum Demodulator for Binary Messages v vb E.G. V1 2-D case v measured va V2 vb t How to define V1,V2? va v ∈ V ⇒ " H " 1 1 v ∈ V2 ⇒ " H2 " vc ∆ Minimize Perror = Pe = P1∫ p {v S1} dv + P2 ∫ p {v S2 } dv V2 V1 ∫ replace with V 1 = P1 + ∫ ⎣⎡P2p {v S2 } − P1p {v S1}⎤⎦ dv V1 Note: Lec14.10-3 2/6/01 ∫V1 p{v S1} dv + ∫V2 p{v S1} dv = 1 A3 Optimum Demodulator for Binary Messages Pe = P1 + ∫ [P2p{v S2 } − P1p{v S1}] dv V1 To minimize Perror , choose V1 ∋ P1p {v S1} > P2p {v S2 } Very general solution ↑ [i.e., choose maximum a posteriori P (“MAP” estimate)] Lec14.10-4 2/6/01 A4 Example: Binary Scalar Signal Case ∆ ∆ S1 = A volts , S2 = O volts , ∴ p{v S1} = p{v S 2 } = 2 2N 1 − − ( v A ) e 2πN p{ v S 2 } If P1 = P2 : P2 (bias choise toward H2 and a priori information) noise 1 e − v 2N 2πN p{ v S1} P1 0 Decision threshold if P1 = P2 Lec14.10-5 2/6/01 2∆ σn = N , Gaussian A/2 v A p { v S} P2 A/2 0 H2 P1 A Threshold if P1 > P2 v A5 Rule For Defining V1 : (Binary Scalar Case) Choose V1 ∋ P1p {v S1} > P2p {v S2} 1 − v 2 2N p {v S2} = e 2πN (binary case) “Likelihood ratio” { } P A= > 2 ⇒" V1" p{v S2 } P1 or (equivalently) An A > An(P2 P1) ⇒" V1" ∆ p v S1 For additive Gaussian noise, [ ] ( An A = - (v - A )2 2N + v 2 2N = 2vA − A 2 ) ? 2N > An(P2 P1) A 2 + 2NAn (P2 P1) A N , or v > + A n (P2 P1) ∴ choose V1 if v > 2A 2 A Lec14.10-6 2/6/01 bias A6 Binary Vector Signal Case For better performance, use multiple independent samples: v(t) S1 ? ∆ p{ v S1} P2 > A= p{ v S2 } P1 t 0 1 2 m S 2 m Here P {v1,v 2 ,...,vm S1} = p {vi S1} (independent noise samples) π i =1 − (v i − S1i ) 2 2N 1 e Where p{v i Si } = 2πN ( i =1 m p { v Si} = Lec14.10-7 2/6/01 ( 1 2πN ) m e ) − ∑ vi − S1i 2 2N B1 Binary Vector Signal Case m p{ v Si } = ( 1 e ( ) − ∑ v i − S1i 2 2N i −1 2πN ) m Thus the test becomes: p {v S1} ? P2 A = > p {v S2} P1 ∆ m ? ⎡m P2 2 2⎤ 1 An A = vi − S2i ) − ∑ ( vi − S1i ) ⎥ > An ( ∑ ⎢ 2N ⎣i=1 P1 ⎦ i=1 v − S2 But v − S2 2 2 v − S1 2 2 − v − S1 = −2v • S2 + 2v • S1 + S2 • S2 − S1 • S1 ∗ ∗ S S S S • − • 2 2 + N An ⎛ P2 ⎞ Therefore v ∈ V1 iff v • (S1 − S2 ) > 1 1 ⎜P ⎟ 2 ⎝ 1⎠ Lec14.10-8 2/6/01 Bias = 0 if energy E1 = E2 Bias = 0 if P2 = P1 B2 Binary Vector Signal Case P V1 iff v1 • (S1 − S2 ) > S1 • S1 − S2 • S2 + N An ⎜⎛ 2 ⎞⎟ 2 ⎝ P1 ⎠ m ∑τ × S1 v i=1 + - operator H1 H2 m ∑τ × S2 i=1 Multiple hypothesis generalization: ? Choose Hi if fi = v • Si − Si • Si + N An Pi > all f j≠i 2 This “matched filter” receiver minimizes Perror ∆ Lec14.10-9 2/6/01 B3 Graphical Representation of Received Signals 2 3-D Case: Average energy = ∑ Si Pi 2 i =1 S1 V1 n V2 0 S2 S1( t ) S13 S11 n t S12 Lec14.10-10 2/6/01 B4 Design of Signals Si +1 E.G. consider S1 = −S 2 S1 = +2 vs. Average energy S 2 E.G. 2-D space for S1, S 2 , S3 , S 4 : = E =1 -1 S2 S2 b a S4 n S4 S1 S2 S1 0 decision boundaries Lec14.10-11 2/6/01 E = 2(P1 = P2 ) 3-D space (S1 = S11, S12 , S13 ) (S1 = S11, S12 ) S3 S2 = 0 Better b S3 b ratio a B5 Design of Signals Si 2-D space: S1,..., S16 Si2 19 Si1 16-ary signals or magnitude/phase vs equilateral triangle slightly lower average signal energy for same p{error} n-Dimensional sphere packing optimization unsolved Lec14.10-12 2/6/01 B6 ∆ Calculation of p{error} = Pe : Binary case: For additive Gaussian noise, optimum is 2 S1 − S2 " H1" if v • (S1 − S2 ) > 2 Where v = S + n ( ∆ 2 [ 2 P + N An 2 P1 ] N = n ( t ) = NoB = kTsB No 2 W Hz - 1 × 2B, double sideband ) 2 2 ⎫ ⎧ S S − 1 2 P ⎪ ⎪ Pe S1 = p⎨v • (S1 − S2 ) < + N An 2 ⎬ 2 P1 ⎪ ⎪⎩ ⎭ 2 ⎧⎪ − S1 − S2 P2 ⎫⎪ = p ⎨n • (S1 − S2 ) < + N An ⎬ = p {y < −b} 2 P1 ⎪⎩ ⎪⎭ Lec14.10-13 2/6/01 y • 2B[GRVZM] -b • 2B D1 Duality of Continuous and Sampled Signals ⎫ ⎧ 2 ⎪⎪ − S1 − S2 P2 ⎪⎪ (S1 − Pe S1 = p⎨ n• S2 ) < + N An ⎬ = p{y < −b} P1 ⎪ 2 ⎪y •2B[GRVZM] ⎪⎭ ⎪⎩ −b• 2B Conversion to continuous signals assuming nyquist sampling is helpful here, S1( t )[0 < t < T] ↔ S1 (2BT samples, sampling theorem) ∆ [ T y = ∫ n( t ) • [S1( t ) − S2 ( t )] dt No WHz −1 2 noise o ] T No ∆1 2 b = ∫ [S1( t ) − S2 ( t )] dt − An(P2 P1) 2o 2 -B 0 B f 2⎫ ⎧⎡ 2BT ⎤ ⎪ ⎪ 1 2∆ 2 σ y = E y = E⎨⎢ n j (S1j − S2 j )⎥ ⎬ ∑ ⎥⎦ ⎪ ⎪⎢⎣ 2B j=1 ⎩ ⎭ [ ] Lec14.10-14 2/6/01 D2 Calculation of Pe, continued 2⎫ ⎧⎡ 2BT ⎤ ⎪ ⎪ 1 2 ∆ 2 σ y = E y = E⎨⎢ n j (S1j − S2 j )⎥ ⎬ ∑ ⎥⎦ ⎪ ⎪⎢⎣ 2B j=1 ⎩ ⎭ [ ] )( 2 ⎧2BT 2BT 1 ⎛ ⎞ ⎪ = ⎜ ⎟ E⎨ ∑ ∑ nin j S1i − S2i S1j − S2 j ⎝ 2B ⎠ ⎪⎩ i=1 j=1 ( ) ⎫⎪ ⎬ ⎪⎭ where E ⎣⎡nin j ⎤⎦ = Nδij 2 T 2 No 2 2 ⎛ 1 ⎞ [ ] σ y = ⎜ ⎟ N S1 − S2 = S ( t ) − S ( t ) dt 1 2 ∫ 2 o ⎝ 2B ⎠ N oB T 2B ∫ [S1( t ) − S 2 ( t )] dt 2 o Lec14.10-15 2/6/01 D3 Calculation of Pe, continued 2 T 2 N ⎛ 1⎞ σ2y = ⎜ ⎟ N S1 − S2 = o ∫ [S1( t ) − S2 ( t )]2 dt 2 o ⎝ 2B ⎠ NoB T 2B ∫ [S1( t ) − S 2 ( t )] dt 2 o p( y ) = 1 2πσ2y e − y 2 2σ 2y (GRVZM) Therefore: P(y) Pe S1 = Lec14.10-16 2/6/01 −b ∫ −∞ 1 2πσ2y e − y 2 2σ2y dy -b 0 y D4 Definition of ERFC(A) A ∆ 1 “Error function” ERF( A ) = ∫e π −A −x2 dx σ= 1 2 “Complementary error function” ∆ ERFC ( A ) = 1 - ERF(A) x -A A 1 Then Pe S1 = ERFC ( A ), where A must be found 2 If we let x 2 = y 2 2σ2y then A ERF(A) = 1 ∫ e π −A − x2 dx = 1 Aσ y 2 ∫ 2πσ2y − Aσ y 2 e where the new limits Aσ 2 and factor 1 Lec14.10-17 2/6/01 − y 2 2σ2y dy 2πσ2y arise as follows: D5 Definition of ERFC(A) A 1 ERF(A) = e ∫ π −A −x2 dx = 1 Aσ y 2 ∫ e − y 2σ 2y 2πσ2y − Aσ y 2 where the new limits Aσ 2 and factor 1 dy 2πσ2y arise as follows: Since x = y σ y 2 , the limit x = A = y σ y 2 becomes a limit where y = Aσ y 2 Also, dx = dy σ y 2 so 1 π becomes 1 Lec14.10-18 2/6/01 2πσ2y D6 Solution for Pe for Binary Signals ( Pe S1 = 1 ERFC( A ) = 1 ERFC b σ y 2 2 2 (where the limit ) and Pe = P1Pe S1 + P2Pe S2 b = Aσ y 2 , so A = b σ y 2 ) If P1 = P2 = 1 , and since Pe S1 = Pe S2 , then 2 ( Pe = 1 ERFC b σ y 2 2 Lec14.10-19 2/6/01 ) T ⎡ ⎤ 1 [S (t ) − S (t)]2 dt ⎢ ⎥ 1 2 ∫ 2 ⎢ ⎥ 0 1 = ERFC ⎢ ⎥ 2 T ⎢ 2 ⎥ ⎢ 2 (No 2) ∫ [S1(t ) − S2 (t )] dt ⎥ ⎢⎣ ⎥⎦ 0 D7 Solution for Pe for Binary Signals T ⎡ ⎤ 2 1 [S (t) − S (t)] dt ⎢ ⎥ 1 2 ∫ 2 ⎢ ⎥ 0 1 Pe = ERFC ⎢ ⎥ 2 T ⎢ 2 ⎥ ⎢ 2 (No 2) ∫ [S1(t ) − S2 (t )] dt ⎥ ⎣ ⎦ 0 ⎡1 T ⎤ 1 2 [ ] − Pe = ERFC ⎢ S ( t ) S ( t ) dt No ⎥ 1 2 ∫ 2 ⎢⎣ 2 0 ⎥⎦ T T 2 If ∫ S1 ( t )dt + ∫ S 2 2 ( t )dt is fixed for P1 = P2 then 0 0 T ⎛ ⎞ 2 To minimize Pe , let S2 ( t ) = −S1( t ) ⎜ maximizes ∫ [S1( t ) − S2 ( t )] dt ⎟ ⎜ ⎟ ⎝ ⎠ 0 Lec14.10-20 2/6/01 D8 Examples of Binary Communications Systems ⎤ ⎡1 T 1 2 Pe = ERFC ⎢ ∫ [S1(t) − S 2 (t)] dt No ⎥ 2 ⎥⎦ ⎢⎣ 2 0 T 1 Assume P1 = P2 = and define ∫ s12 (t)dt ∆ =E 2 0 Modulation type s1(t) s2(t) Pe “OOK” (on-off keying) A cos ωo t 0 “FSK” (frequency-shift keying) A cos ω1t A cos ω2 t 1 ERFC E avg 2No 2 “BPSK” binary phaseshift keying) A cos ωt − A cos ωt 1 ERFC E avg No 2 Lec14.10-21 2/6/01 1 ERFC E 4N o 2 = 1 ERFC Eavg 2No 2 D9 Examples of Binary Communications Systems Note: Pe = f (E AVG No ) [J] [W Hz-1 = J] Cost of communications ∝ cost of energy, Joules per bit (e.g. very low bit rates imply very low power transmitters, small antennas) Lec14.10-22 2/6/01 D10 Probability of Baud Error Pe 1 10-1 10-2 10-3 10-4 10-5 10-6 10-7 6 dB OOK (coherent) FSK BPSK FSK non-coherent 3 dB 0 4 8 12 16 20 E/No (dB) Non-coherent FSK: carrier is unsynchronized so that both sine and cosine terms admitted, increasing noise. Such “envelope detectors have a different form of Pe(E/No). Lec14.10-23 2/6/01 Note how rapidly Pe declines for E No ~ > 12 − 16 dB D11