Abstract

Relaxed locally decodable codes (RLDCs) are error-correcting codes in which individual bits of the message can be recovered by querying only a few bits from a noisy codeword.

For uncorrupted codewords, and for every bit, the decoder must decode the bit correctly with high probability. However, for a noisy codeword, a relaxed local decoder is allowed to output a ``rejection’’ symbol, indicating that the decoding failed.

We study the power of adaptivity and two-sided error for RLD Cs. Our main result is that if the underlying code is linear, adaptivity and two-sided error do not give any power to relaxed local decoding. We construct a reduction from adaptive, two-sided error relaxed local decoders to non-adaptive, one-sided error ones. That is, the reduction produces a relaxed local decoder that never errs or rejects if its input is a valid codeword and makes queries based on its internal randomness (and the requested index to decode), independently of the input. The reduction essentially maintains the query complexity, requiring at most one additional query. For any input, the decoder’s error probability increases at most two-fold. Furthermore, assuming the underlying code is in systematic form, where the original message is embedded as the first bits of its encoding, the reduction also conserves both the code itself and its rate and distance properties

We base the reduction on our new notion of additive promise problems. A promise problem is additive if the sum of any two YES -instances is a YES -instance and the sum of any NO-instance and a YES -instance is a NO-instance. This novel framework captures both linear RLD Cs and property testing (of linear properties), despite their significant differences.

We prove that in general, algorithms for any additive promise problem do not gain power from adaptivity or two-sided error, and obtain the result for RLD Cs as a special case. The result also holds for relaxed locally correctable codes (RLCCs), where a codeword bit should be recovered.

As an application, we improve the best known lower bound for linear adaptive RLD Cs. Specifically, we prove that such codes require block length of $n \geq k^{{1+\Omega(1/q}2)}, where k denotes the message length and q denotes the number of queries.