A rollup's security rests on anyone being able to reconstruct its state from data posted to a base chain. If that data is withheld, the guarantees fall apart even when the proofs are valid.
Withheld data breaks verification
A rollup operator commits to a new state root and publishes the transactions that produced it. Anyone can re-run those transactions and check the result.
If the operator publishes the root but not the data, nobody can check anything, and nobody can build the proof needed to challenge a fraudulent claim.
Users cannot even exit on their own terms, because constructing a withdrawal requires knowing the balances the withheld data encodes. Publication is therefore a security requirement, not a convenience.
Checking publication is expensive for individuals
The simple way to confirm data was published is to download it. That works, but it makes every verifier carry the full bandwidth of every rollup they care about.
Requiring that defeats the point of light clients, which exist so participants with modest hardware can still verify rather than trust.
The problem is asymmetric in an awkward way: proving data is available is easy, while proving that a small fraction is missing is not.
Erasure coding changes what withholding requires
The data is first expanded with erasure coding, so the original can be reconstructed from any sufficiently large subset of the expanded pieces.
This means a publisher cannot hide a small slice. To make reconstruction impossible they must withhold a substantial fraction of the whole, which is far more conspicuous.
Hiding a single transaction is no longer an option. The choice becomes publishing properly or visibly withholding a large share.
Random sampling detects large omissions
A light participant then requests a handful of randomly chosen pieces. If they all arrive, a large withheld fraction becomes very unlikely.
Each additional sample sharpens the conclusion, and confidence rises quickly with only a few requests. The participant never downloads the full data.
With many independent samplers drawing different pieces, the network collectively touches nearly everything while each individual touches almost nothing.
The guarantee is probabilistic and social
Sampling never proves availability outright; it makes withholding detectable with high probability. That is enough, because the response to detection is refusal to accept the block.
It also depends on there being enough samplers making independent requests. A network with few light clients gives a withholder better odds of serving every request it receives.
This is why data availability is treated as its own layer with its own participation requirements, rather than as a property that follows automatically from posting bytes somewhere.