Asymptotic Capacity Bounds for Magnetic Recording. Raman Venkataramani Seagate Technology (Joint work with Dieter Arnold)

Transcription

1 Asymptotic Capacity Bounds for Magnetic Recording Raman Venkataramani Seagate Technology (Joint work with Dieter Arnold)

2 Outline Problem Statement Signal and Noise Models for Magnetic Recording Capacity Bounds Additive noise dominated channels Jitter noise dominated channels

3 Problem Statement How does the capacity of the system scale with the bit-width? T What is the optimum bit width from a purely information theoretic viewpoint?

4 Signal and Noise Models Signal: the magnetization in the medium is either ±1 Noise: electronic noise and transition noise.

5 Signal and Noise Models x(t) = k a k p(t kt ), p(t) = u(t) u(t T ) r(t) = k a k a k 1 2 h T (t kt t k ) + w(t) a k {+1, 1} t k : transition noise w(t) : electronic noise AWGN with PSD N 0 /2 h T (t) : transition ( 1 +1) response

6 Recording Modes Magnetization profile for longitudinal recording Magnetization profile for perpendicular recording

7 Transition Response Longitudinal recording h T (t) = Perpendicular recording A 1 + (2t/w) 2 h T (t) = A erf ( 2 log 2t ) w The pulse width is sometimes also written as PW50.

8 Capacity The fundamental quantity of interest is the areal capacity of the magnetic medium. In the Shannon sense, this is the number of bits per unit area that can be stored and recovered reliably. Assuming a fixed track width, we can focus on the linear density of bits or linear capacity: C = lim τ 1 τ max({x(t), t [0, τ]}; {r(t), t [0, τ]})

9 Problem Statement How is the optimal bit width capacity per unit length? T for maximum Misconceptions: Very small bit-widths are bad because... There is higher transition noise. There is longer ISI and lower SNR per bit. Truth: We can overcome all these problems with clever coding.

10 Thought Experiment No code T = 1 R = 1 Repetition code T = 0.5 R = 0.5 If we use better code, case 2 will outperform case 1.

11 1 Binary Entropy Function Repetition Code Code Rate Transition Probability Achievable region of code rate and transition probability.

12 A Note about the SNR We define: SNR = E iw 2 where E i = N 0 R Reasons for this definition: dh T (t) 2 dt dt It is finite for both perpendicular and longitudinal modes. It is a function of the head and medium alone. It is independent of or the code we use. T

13 Electronic Noise Dominated Channel x(t) = k a k p(t kt ), p(t) = u(t) u(t T ) r(t) = k (a k a k 1 )h T (t kt ) + w(t) = k a k h D (t kt ) + w(t) h D (t) = h T (t) h T (t T ) : dibit response w(t) : electronic noise AWGN with PSD N 0 /2

14 Sufficient statistics To be exact, we need to pass through a matched filter and sample at baud rate. T w r(t) If, it is almost optimal to use a simple low-pass filter and sample a baud rate:

15 Discrete-time Model We obtain an ISI channel with binary inputs: r n = k a k hn k + w n w n N(0, N 0 /2T ), i.i.d. where, hn h D (nt ) Let the above ISI channel have capacity. Then, the linear capacity is K B (T ) = 1 T C B(T ) C B (T )

16 Simple Upper Bound Suppose we relax the input constraint: {a k = ±1} {a k R, E a 2 k 1} We obtain a Gaussian channel, whose capacity is computed using the water-filling method. C B (T ) C G (T ) C G (T ) = max S x [ ] 1 2 such that 0.5 log (1 + S x[ν] H[ν] 2 S x [ν]dν = 1 N 0 /2T )

17 Lower Bounds Several researchers have computed information rates for binary ISI channels: Monte-Carlo method: compute the quantity I(x N ; r N )/N for a very long sequence of length N with the input generated using an appropriate Markov model. This method becomes unreliable or computationally intense for high recording densities.

18 Shamai-Laroia Conjecture Along any horizontal line, the SNR loss due to ISI is the same for both Gaussian and binary signalling.

19 Shamai-Laroia Conjecture Suppose that the input to the binary ISI channel is independent and uniformly (IUD) distributed C B (T ) C BP SK (2 I G(T ) 1) where I G (T ) is the i.i.d Gaussian information rate. Remark: This bound is also found to be surprisingly tight for non IUD Markov sources up to memory 6 and we have C B (T ) C BP SK (2 C G(T ) 1)

20 Optimal Input Distribution For Gaussian inputs, the optimal input spectrum is not flat. It becomes highly non i.i.d as T 0 For binary inputs, we could mimic the Gaussian water-filling spectrum, but it is neither optimal nor achievable in general. For the binary case, we optimize over firstorder Markov sources to get a lower bound.

21 Markov Source Model The first order model is specified by one probability parameter p T For each we pick the optimal probability to maximize the Shamai-Laroia bound. As T 0 we find that p 0

22 8 7 6 Gauss waterfilling MS1 optimized bits/use/t MS0 i.u.d T SNR 10 0 Linear capacity bounds for longitudinal recording

23 2.5 Gauss waterfilling 2 MS1 optimized bits/use/t MS0 i.u.d T SNR 10 0 Linear capacity bounds for perpendicular recording

24 Jitter Noise Dominated Channel Most general method: pulse-width modulation r(t) = n ( 1) n h T (t σ n w n ) σ n = n t k, w n N(0, σj 2 ) k=1

25 Capacity Since there is no additive noise, we can use a zero forcing equalizer to estimate ˆσ n = σ n + w n Thus, ˆt n = ˆσ n ˆσ n 1 = t n + w n w n 1 These are sufficient statistics and the channel is ISI! Linear capacity: C = max p(t N ) I( ˆT N ; T N ) E n T n

26 Lower Bound on Capacity We can show that the following is a lower bound on the linear capacity: K K 0 where K 0 is the capacity of the ISI-free channel: i.e., ˆt n = t n + w n I( K 0 = max ˆT ; T ) p(t ) E T Remark: This constant is not important. The bound behaves like. 1/σ j

27 Summary From a purely information theoretic viewpoint there are considerable gains at we higher recording densities. This ignores timing recovery and other implementation issues. In today s recording medium transition noise is a dominant noise source and σ j 0.1T We should lower potential gains. T by a factor of 10 to see