On a binary symmetric channel with crossover probability p, what does p predict about a thousand-bit block over a radio link?
answer
- per-bit rate, not per-block
- both directions flip equally
- no third symbol, no warning
- expected flips equal n times p
- clean block chance is (1-p)^n
basics
~20 sA binary symmetric channel flips each transmitted bit independently with probability p, and the receiver cannot see which bits changed. With p = 0.001 a thousand-bit block averages one flip, yet only about 37 percent of blocks arrive clean.
solid answer
~50 sThe binary symmetric channel is the one-parameter model of a noisy digital link: input and output are both bits, every transmitted bit arrives inverted with probability `p` and intact with probability `1 - p`, the two directions are equally likely (that is the *symmetric* part), and each bit's fate is independent of every other bit's. Because `p` is a **per-bit** rate, a block of `n` bits carries `n*p` flips on average and arrives completely clean with probability `(1 - p)^n`. For `n = 1000` and `p = 0.001` that is one expected flip and only a 36.8 percent chance of a clean block — 63 percent of blocks are damaged. The other half of the model matters as much: the output alphabet is the same as the input alphabet, so a flipped bit looks exactly like a good one and the channel itself raises no alarm.
go deeper
Recall the one-line shape: a per-bit flip probability, both directions equal, and no indication at the receiver that a bit changed. Knowing that a bit error rate is per bit, not per message, is the useful takeaway.
Explain the conversion from the per-bit rate to block behaviour: expected flips are n times p, and a clean block has probability (1-p)^n. Be ready to say why one expected flip per block still leaves a third of blocks untouched.
Show that you treat each defining property as an assumption to validate against measured link data — symmetry, independence, and the absence of any confidence signal — and say what you would measure to test them before designing against the model.
The judgement is which model the organisation commits to. Choosing a single-parameter bit-flip model fixes the whole downstream redundancy budget and the failure mode operators will see, so the cost of the model being wrong belongs in the decision, not in a later incident.
## What the model actually asserts The **binary symmetric channel** (BSC) is the simplest model of a link that corrupts bits. It has one input symbol set `{0, 1}`, the same output symbol set, and one parameter, the **crossover probability** `p`. Four properties define it, and each one is a claim about the physical link that you are choosing to believe: - **Bit-level corruption.** Each transmitted bit is received inverted with probability `p` and unchanged with probability `1 - p`. - **Symmetry.** `P(receive 1 | sent 0)` equals `P(receive 0 | sent 1)`, both `p`. Neither value is more fragile than the other. - **Memorylessness.** What happens to one bit is independent of what happened to its neighbours. There is no notion of a fade, a scratch or a bad second. - **Silence.** The output alphabet equals the input alphabet, so there is no third symbol meaning "I could not read this". A corrupted bit is a perfectly ordinary bit with the wrong value. That last property is what makes the BSC the *expensive* channel to work with, and it is the one candidates skip. ## What p predicts about a block `p` is a rate per bit, so nothing about it is a per-block figure until you bring in the block length `n`. Two quantities follow directly: 1. The **expected number of flips** in a block is `n * p`. This is a count, not a probability — at `n = 1000`, `p = 0.001` it equals exactly 1, which is not "the block always fails". 2. The **probability the block arrives untouched** is `(1 - p)^n`, because the flips are independent. The number of flips per block follows a binomial distribution with parameters `n` and `p`, closely approximated by a Poisson distribution with mean `n * p` when `p` is small. For thousand-bit blocks the arithmetic is worth memorising as a shape: | crossover p | expected flips per block | block arrives clean | block has two or more flips | |---|---|---|---| | 0.0001 | 0.1 | 90.5% | 0.47% | | 0.001 | 1 | 36.8% | 26.4% | | 0.01 | 10 | 0.0043% | 99.95% | | 0.1 | 100 | about 1e-46 | essentially 1 | The interesting row is the second. A link described as "one bit error in a thousand" sounds nearly clean, and yet almost two blocks in three are damaged, and a quarter of all blocks carry more than one error. **Block length multiplies a per-bit rate into a per-block reality**, and the growth is exponential in `n`: doubling the block size squares the clean-block probability. ## Why the silence is the costly part Because the channel emits only ones and zeros, the receiver has no way, from the channel alone, to distinguish a flipped bit from a correct one. Everything a system does about that — any redundancy it adds, any check it runs — exists because the channel refuses to say. The practical consequence is that a repair has two halves: **finding** which position is wrong and **fixing** its value. On this channel you pay for both. A related subtlety: `p = 0.5` is the worst possible symmetric channel, not `p = 1`. At `p = 1` every bit is inverted, so the receiver inverts everything back and the link is perfect. At `p = 0.5` the output is statistically independent of the input and carries nothing at all. Candidates who think "higher p is monotonically worse" have not thought about the model. ## Where the model stops being true The BSC is a model, and each defining property is an assumption a real link can violate: - **Asymmetry.** Some physical media corrupt one signalling level far more readily than the other, so a single `p` misdescribes them; that needs two parameters. - **Correlation.** Real noise arrives in bursts — a fade, a mechanical defect, an interferer keying up. A memoryless model with the same average rate predicts a very different distribution of damage per block. - **Hard decisions.** A real demodulator usually produces a *confidence* alongside each bit. Feeding only the hard bit into a BSC model throws that confidence away; a model with a third "unreadable" output symbol keeps some of it. ## What interviewers listen for - That you state `p` as a per-bit probability and convert it to a per-block figure with `n` rather than eyeballing it. - That you name the silence: no marker, no third symbol, corrupted bits are indistinguishable. - That you know `(1 - p)^n` collapses fast, so "our error rate is tiny" is not an argument about blocks until the block size is on the table. - That you treat memorylessness as an assumption to be checked against measurements, not a law.
- What does the word symmetric in this channel model actually assert?That the two error directions are equally likely: a transmitted 0 arriving as 1 has the same probability as a transmitted 1 arriving as 0, both `p`. Real media can be lopsided — one signalling level may degrade toward the other far more readily — and a link like that needs two parameters. In the extreme, a channel where one direction never errs behaves quite differently from a symmetric one at the same average error rate.
- Does raising p from 0.001 to 0.01 make a thousand-bit block ten times more likely to be damaged?No. The expected flip count scales linearly, 1 to 10, but the clean-block probability is `(1 - p)^n`, which collapses from 36.8 percent to about 0.0043 percent. The probability a block is damaged goes from roughly 63 percent to roughly 99.996 percent — a factor of about 1.6, because it was already close to its ceiling. Rates scale linearly; block-level probabilities saturate.
- Is a crossover probability of 0.5 worse or better than one of 0.9?0.5 is worse, and it is the worst case. At `p = 0.9` the receiver can invert every bit and recover a channel with effective crossover 0.1, so the link still carries information. At `p = 0.5` the output is statistically independent of the input: no processing at the receiver can extract anything, because the received bits would look the same whatever was sent.
It is a proofreader who silently rewrites one character in a thousand and leaves no mark: each page looks perfectly typeset, and a thousand-character page is wrong about two times in three.
saying these in an interview costs you the question
- Reads p as the fraction of blocks that fail, not bits
- Says a flipped bit arrives marked as suspicious
- Multiplies p by block length and calls it a probability
- Claims a low bit error rate means most blocks arrive clean
- Thinks crossover 0.5 is a mildly noisy link, not the worst case
- Assumes one direction of flip is naturally likelier here