{
  "status": "shipping",
  "advantages": [],
  "limitations": [
    {
      "text": "Erasure recovery does not authenticate symbols or detect every corruption; protect the object and its decoding metadata separately.",
      "citationIds": [
        "rfc6330"
      ]
    },
    {
      "text": "The linked calculation is a constructed binary-matrix demonstration, not an LT degree-distribution simulation or a RaptorQ implementation.",
      "citationIds": [
        "guide"
      ]
    }
  ],
  "related": [
    "erasure-coding",
    "lt-code",
    "maximum-distance-separable-code",
    "rateless-iblt",
    "raptor-code",
    "raptorq"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-21",
  "slug": "fountain-code",
  "term": "Fountain code",
  "aliases": [
    "Rateless erasure code"
  ],
  "definition": "A fountain code generates additional encoding symbols on demand from a source block, allowing receivers to recover the block from a sufficiently informative received set.",
  "note": "Rateless describes how redundancy is generated, not a guarantee that any k received symbols recover k inputs. LT codes mix selected inputs by XOR; Raptor constructions add a precode before the LT stage. The precode permits recovery of inputs that the LT stage leaves unresolved. Degree distribution, decoding algorithm and reception overhead affect recovery. [raptor2006]\n\nRFC 6330 specifies the distinct, systematic RaptorQ scheme. For supported extended block sizes K', its decoder requirements bound average failure by 1/100, 1/10,000 and 1/1,000,000 at K', K'+1 and K'+2 received symbols respectively, with encoding-symbol identifiers sampled independently and uniformly. K' is the supported padded block size, not automatically the original K. These are conditional requirements, not universal disk-loss probabilities. [rfc6330]\n\nIn Hesela's three-input example, three distinct XOR equations can have rank two. Another three-equation set has full rank but no degree-one equation, so pure peeling stalls. An added singleton starts recovery without increasing rank. Receiving more equations, gaining independent information and enabling a particular decoder are separate events. [example]",
  "citations": [
    {
      "id": "raptor2006",
      "title": "Amin Shokrollahi. Raptor Codes. IEEE TIT 52(6), 2551-2567 (2006), sections III-IV. DOI: 10.1109/TIT.2006.874390.",
      "url": "https://www.qualcomm.com/content/dam/qcomm-martech/dm-assets/documents/Raptor_Codes_IEEE_technical_analysis.pdf",
      "accessed": "2026-11-21"
    },
    {
      "id": "rfc6330",
      "title": "RFC 6330 (2011): RaptorQ, sections 5.3.1, 5.8 and 6.",
      "url": "https://www.rfc-editor.org/rfc/rfc6330.html",
      "accessed": "2026-11-21"
    },
    {
      "id": "example",
      "title": "Hesela: exact synthetic XOR equation, rank and peeling observations",
      "url": "https://hesela.dev/research/fountain-rank/observations.json",
      "accessed": "2026-11-21"
    },
    {
      "id": "guide",
      "title": "Hesela: worked decoding example and reproduction instructions",
      "url": "https://hesela.dev/research/fountain-rank/README.txt",
      "accessed": "2026-11-21"
    }
  ],
  "provenance": [
    "The 2006 paper gives code constructions and mathematical analysis; RFC 6330 gives a later interoperable scheme. Neither is treated as a storage-fleet measurement.",
    "The example uses fixed, explicitly listed coefficient rows and known-correct equations from one source block."
  ]
}
