Fetching contributors…
Cannot retrieve contributors at this time
1057 lines (842 sloc) 42.1 KB

BIP: ???
Layer: Peer Services
Title: Compact Client Side Filtering for Light Clients
Author: Olaoluwa Osuntokun <>
        Alex Akselrod <>
Comments: ???
Comments-URI: ???
Type: Standards Track
Created: 05-24-2017
License: CC0-1.0

Table of Contents


This BIP describes a new light client node type for Bitcoin as well as the modifications to current full-nodes required to support this new type of light client. The light client mode described in this BIP is meant to supersede BIP 37 as it provides a greater degree of privacy, utility, and also reduces the resources required for full-nodes to service this new light client mode compared to BIP 37[1]. The light client mode described in this BIP can be seen as a "reversal" of BIP 37[2]: rather than the light clients sending filters to full-nodes, full-nodes send filters to light clients. Unlike BIP 37, we don't utilize bloom filters. Instead, we utilize a compact filter (more efficient than bloom filters) which leverages Golomb-Rice coding for compression. Additionally, blocks are downloaded as a whole (from any source), rather than directly from peers as fragments with merkle-branches proving their authenticity.


Light clients in Bitcoin provide applications with a less resource intensive mechanism of validating the work of the most difficult chain and identifying entries in the blockchain's log which are relevant to said application. In order to accomplish the first, light clients download and verify the connectivity and work of only the block headers of the chain. Block headers are a constant 80-bytes, resulting in minimal bandwidth even for very long chains. In order to efficiently accomplish the second task (ascertaining relevant chain data) light clients require a mechanism to learn of relevant data in blocks.

BIP 37 is currently the most widely used light client execution mode within Bitcoin. In BIP 37, rather than fetching and fully validating all blocks in the chain, the light client instead verifies all headers and sends bloom filters containing relevant data to full-nodes. These full-nodes then service the light client by querying data within a block against the loaded bloom filter, if a transaction matches the filter, a merkle-branch for the matching transaction is sent and distinctly the transaction itself is sent.

However, BIP 37 has several downsides. Bloom filtering as widely implemented provides virtually zero privacy to wallets or other applications using this mechanism [3][4]. Additionally, applications are forced to carefully manage their false positive rates in order to not completely give away their set of interested items. Additionally, full-nodes can nearly undetectably lie by omission, causing a denial of service which can lead to undesirable failure modes in applications whose safety critically relies on responding to certain on-chain events. When faithfully servicing BIP 37 light clients, full-nodes may incur significant I/O and CPU resource usage due to maliciously crafted bloom filters, creating a denial-of-service vector.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

Design Rationale

In order to address the drawbacks of BIP 37 raised above, in this document we propose an alternative chain filtering mechanism for light clients. Our proposal sports a greater degree of privacy than BIP 37 as filtering is now done on the client side. Clients download a deterministically generated filter for a block and query it locally. If relevant items are found in the filter then the entire block will be fetched. The decoupling of filter querying from active communication with full-nodes enables light clients to fetch blocks from any source. Extremely privacy conscious light clients may opt to anonymously fetch blocks using cryptographic techniques such a Private Information Retrieval [5].

In order to reduce the size of the filter, we use a data structure capable of probabilistic set membership. We elide the selection of the bloom filter data structure in favor of utilising Golomb-Rice coding which allows us to generate filters more compact than bloom filters which approach the theoretical minimum size for probabilistic data structure.

Light clients operating using the method described in this document are able to verify the authenticity of filters received, thereby eliminating the ability for full-nodes to lie by omission. Such client side filtering also improves the utility of light clients for generic applications beyond simple wallets and usage of basic public key templates.

Finally, full-nodes only need to construct filters once as they're deterministically generated for each block. Once the index is built, no further active processing is required to serve light clients. Servicing light clients simply entails reading pre-computed filters and headers from disk and sending them over the network.


Before we specify the details of our proposal, we'll first go over a few preliminaries which will aid in the understanding our proposal.

By []byte we refer to a slice (or vector) of bytes. This value is typically expressed in C-like languages as an array of uint_8's.

By Var-Int we refer to the variable length integer encoding used widely within the Bitcoin p2p protocol as an efficient way to encode the number of items in a repeated series of items. The p2p message extensions in this proposal will utilize this variable-length integer encoding in an identical manner to the existing Bitcoin p2p messages.

By siphash(k, n) we refer to an invocation of the SipHash pseudo-random function with k as the 128-bit key, and n as the input to the PRF. We instantiate with the recommended parameters of c = 2 and d = 4.

We define the concept of an abstract bit stream instantiated by the function: new_bit_stream The bit_stream has two functions that operate on it, unary_encode(stream, n) and write_bits_big_endian(stream, n, k) where unary_encode(steam, n) emits n (an integer) to the stream in unary, and write_bits_big_endian(stream, n, k) emits the lower k bits of n to the stream using a big-endian binary encoding. For our unary encoding, we encode a series of 1's followed by a terminating 0.

Whenever we reference sorting, we refer to an ascending sorted order. The items in a sorted set should increase from smallest index to largest index.

We use a form of pseudo-code throughout the specification. In some areas we use pattern-matching to specify the details of an algorithm:

  • match(ITEM) denotes a clause which pattern matches on ITEM (similar to a switch statement in imperative languages).
  • Some denotes a non-empty ITEM, equivalent to a non-nil pointer or value
  • None denotes an empty ITEM, equivalent to a nil pointer or value


Compact Chain Filters

In this BIP, we propose that light clients be provided with compact filters which succinctly encode the contents of blocks. Instead of bloom filters, we instead employ a data structure which is a compressed version of the hashed values of the contents of blocks. Throughout the document, we refer to this data structure as a Golomb Coded Set (GCS). At a high level the set contains a list of sorted fixed size values. These values are then compressed using a type of run length encoding. In order to query the set, it must be decompressed.

We will now define simple functions for encoding and decoding integers using Golomb-Rice [6] coding. These functions will be used in the next section as a primitive in the construction of our compact filters.

golomb_encode(stream, n, k):
    let q = n >> k
    unary_encode(stream, q)
    write_bits_big_endian(stream, n, k)

golomb_decode(stream, k) -> int:
    let c = stream.read_bit()

    let n = 0
    while c == 0:
        c = stream.read_bit()

    let r = b.read_bits_big_endian(k)

    where read_bits_big_endian(k) decodes a fixed-length big-endian integer of

    c*m + r

With the two functions above, we're able to efficiently compress a single integer using Golomb-Rice coding. In the next section, we'll put everything together and use the primitives described above to construct our compact sets.

Golomb-Rice Coded Sets

Rather than insert items directly into our set, we instead first run the items through a PRF. This creates a set of uniformly distributed values. If we then sort each of these values, the delta between each of the values closely resembles a Geometric Distribution. We'll again leverage this fact to use Golomb-Rice coding to compresses our set by only encoding the delta between two successive elements in the set.

Golomb-Rice coded sets take two parameters:

  • N the number of items to be inserted into the set
  • P a value which is computed as 1/fp where fp is the desired false positive rate.
Given these two parameters, we can now construct our set.

Set Construction

Set construction takes three parameters: N, P and L

  • where L is a list of the raw items we wish to insert into the set
  • the type of L is assumed to be of []byte
NOTE: P must be a power of two as we target the specialized case of Golomb coding: Golomb-Rice coding.

Using N and P we compute F = N * P. F constricts the range of the hashed values accordingly in order to achieve our desired false positive rate.

In addition, to help optimize the algorithm, we use a fast range algorithm[7], multiplying the hashed value by F and taking only the top 64 bits. This fairly distributes the values over F without expensive division operations. In our domain, the operation will use 64-bit integers. As a result, one may need to manually compute the upper 64-bits of a 64-bits integers multiplication. This can be done with fewer cycles on CPUs that have 128-bit registers. We use 64-bits, as this is the outputs size of siphash(2, 4).

The following routine computes the uncompressed set given the parameters above:

hashed_set_construct(N, P, raw_items, k): -> []uint64:
    let F = N * P

    let set_items = []
    for item in raw_items:
        let set_value = (siphash(k, item) * F) >> 64

    // Sorts in ascending order.


Using the routine above, we can transform our set of (possibly heterogeneous items) in to a list of uniformly distributed values. As a final step, these values are then sorted. When sorting then items MUST be ordered in ascending order.

Set Compression

Once the set of hashed items has been constructed (and sorted in ascending order) we then use Golomb-Rice coding to compress the set by encoding the delta value between each successive element within the set. As these values are uniformly distributed, the deltas between these values will be Geometrically Distributed, meaning that Golomb-Rice coding will be optimal for this use-case [8].

The following routine describes the compression process:

gcs_compress(sorted_set, fp) -> []byte:
    let stream = new_bit_stream()

    // P is equivalent to m, the size of a golomb code-word.
    let P = 1 << fp

    let last_value = 0
    for value in sorted_set:
        // Compute the difference between this value and the last value modulo
        // P.
        let remainder = (value - last_value) & (P - 1)

        // Compute the difference between this value and the last one, divided
        // by P. This is our quotient.
        let quotient = (value - last_value - remainder) >> fp

        // Write out the quotient value in unary into the bit stream.
        unary_encode(stream, quotient)

        // Finally, write the remainder into the bit stream using fp bits.
        write_bits_big_endian(stream, remainder, fp)

        // Track this value so we can use it compute the diff between this
        // value and the last.
        last_value = value


The routine above computes a compressed set using Golomb-Rice coding to encode the delta between elements within the set. Unlike a bloom-filter, this data-structure cannot be queried in its current form. Instead, one MUST first perform the reverse computation to decompress the items in the set, revealing the true values which can be queried against.

Set Querying/Decompression

Given a compressed Golomb-Rice coded set, one MUST first decompress the set itself in order to query items which have been included within the set. Decompression of a set follows the reverse procedure of encoding. To decode an element, we'll decode the encoded quotient and remainder of encoded delta. With the full delta re-constructed, we then add this value to the prior value in order to reconstruct the full value. Following this procedure we can incrementally decompress the set lazily without decompressing the entire filter.

Querying for a Single Item

The following routing describes how one queries a compressed set for a single item:

gcs_match(key: [16]byte, compressed_set: []byte, target: []byte, fp, N: int) -> bool:
    // First we'll map the item into the domain of our encoding.
    let item = (siphash(key, target) * (N * (1 << fp))) >> 64

    stream = new_bit_stream(compressed_set)

    // We initialize the initial accumulator to a value of zero.
    let last_value = 0

    // As the values in the set are sorted once the decoded values exceeds the
    // value we wish to query for, we can terminate our search early.
    for last_value < item:
        // Read the delta between this value and the next value which has been
        // encoded using Golomb-Rice codes.
        let decoded_value = golomb_decode(stream, fp)

        // With the delta computed, we can now reconstruct the original value.
        let set_item = last_value + decoded_value

        // If the values match up, then the target item _may_ be in the set, so
        // we return true.
        if set_item == item:

        last_value = set_item

    // If we reach this point, then the item isn't in the set.
Querying Against a Set of Items

For most applications, the common case will be attempting to match a list of items to the filter. In this case, we can perform a "zip" search against two sorted lists: the step-by-step decompressed values of the set, and the list of items we'd like to query.

The following routine will evaluate to true if any of the items in a target set are maybe within the original set of items (pre encoding):

gcs_match_any(key: [16]byte, compressed_set: []byte, targets [][]byte, 
              fp, N: int) -> bool:

    stream = new_bit_stream(compressed_set)

    // Once again, we'll map our set of target values into the domain our
    // encoding, sorting as a last step so we can zip through the values.
    let items = []
    for t in target:
        let item = (siphash(key, t) * (N * (1 << fp))) >> 64

    // Set up a set of accumulator values that we'll use to zip down the two
    // filters.
    let last_set_val, last_target_val = 0, 0 
    last_target_val = items[0]
    let = 1

    // We'll keep running until one of the values matches each other. If this
    // happens, then we have a match!
    while last_set_val != last_target_val:
        // Perform a pattern match to decide which filter we'll need to
        // advance.
            case last_set_val > last_target_val:
                // If we still have items let, advance the pointer by one.
                if i < len(items):
                    last_target_val = items[i]

                // Otherwise, we've ran our items in our target set, which
                // means nothing matched.

            case last_target_val > last_set_val:
                // In this case, we'll advance the filter we're querying
                // against. This entails decompressing the next element in the
                // set.
                let decoded_value = golomb_decode(stream, fp)

                // Accumulate the decoded delta value to the current value in
                // order to retrieve the current set item.
                last_set_val += decoded_value

    // If we reach this point, the two items in the set matched!

Peer to Peer Network Extensions

With the procedures to construct, compress, and query the sets explained, we'll now turn to the modifications to Bitcoin's p2p protocol required to support this new operating mode.

Peer to Peer Service Bit

To start, we reserve a currently unutilized service bit. This is required as light clients SHOULD preferentially peer to full-nodes that support the features outlined in this BIP.

The 6th service bit will now be dedicated to signaling support for the features described within this BIP:

  • SFNodeCF = 1 << 6

Filter Types

As this framework for client-side chain filtering is meant to be generic, in this document we define two filter types. A filter type denotes both the construction/querying for a filter as well as the contents of the filter.

At the time of writing of this BIP, two filter types are defined:

  • Normal (0x00)
  • Extended (0x01)
A Normal filter is intended to contain all the items that a light client needs to sync a basic Bitcoin wallet. In order to facilitate this use-case, for each transaction, normal filters contain:
  • The outpoints of each input within a transaction.
  • The data-pushes contained within the public key script of each output within the transaction.
  • The txid of the transaction itself.
An Extended filter contains extra data that is meant to facilitate the adoption of more advanced smart contracting applications by this BIP. For each transaction found in a block, an Extended filter contains:
  • Each item within the witness stack of an input (if the input has a witness).
  • Each data push of the signature script of an input.
Notably, this construction does not currently interpret P2SH scripts or witness scripts to extract data pushes from them; however, future filter types may be designed to do so.

Filter Construction

In order to ensure that filters are deterministically generated, we will use the first 16-bytes of the block hash of a Bitcoin block as the key to our siphash function. Full-nodes that support this BIP SHOULD treat the set of filters as an additional index of the blockchain. Once a new block arrives, both filter types SHOULD be constructed, and stored on disk. Full nodes MAY opt to dynamically construct the filters at runtime, trading off space for additional computation. Full-nodes that update to support this BIP once already synced, SHOULD upon start-up, re-index the chain, constructing filters for each block from genesis to current chain tip.

When indexing input and output scripts, we only index the push datas in the script. The function extract_push_datas returns a vector of byte slices that contain any pushed data found within the script. Pushed datas are the byte slices following: OP_PUSHDATA1, OP_PUSHDATA2, OP_PUSHDATA4, and the opcodes numbered 1 to 75. The set of returned values includes OP_O, but excludes OP_1 - OP_16. OP_O MUST be emitted as an empty byte slice. For the complete set of opcodes defined in Script, we refer the reader to [9].

Given a Bitcoin block, a full-node MUST construct a Normal compact filter as follows:

construct_normal_gcs_filter(block, fp) -> []byte:
    let siphash_key = block.hash()[:16]

    let P = 1 << fp

    let raw_items = []
    for tx in block.transactions:
        let txid = tx.hash()

        for output in tx.outputs:
            let output_bytes = extract_push_datas(output.script)
            for output_byte in output_bytes:

        if tx.is_coinbase():

        for input in tx.inputs:
            // Inputs serialized as they are on the wire in transactions.
            // Input index serialized in little-endian.
            let input_bytes = input.hash || input.index

    let N = len(raw_items)
    let F = N * P

    let hashed_items = []
    for raw_item in raw_items:
        let hashed_item = (siphash_key(siphash_key, raw_item) * F) >> 64


    gcs_compress(hashed_items, fp)

Given a Bitcoin block, a full-node MUST construct an Extended compact filter as follows:

construct_extended_gcs_filter(block, fp) -> []byte:

    let siphash_key = block.hash()[:16]

    let P = 1 << fp

    let raw_items = []
    for tx in block.transactions:
        if tx.is_coinbase():

        for input in tx.inputs:
            for wit_elem in input.witness:

            let sig_script_pushes = extract_push_datas(input.sig_script)
            for push in sig_script_pushes:

    let N = len(raw_items)
    let F = N * P

    let hashed_items = []
    for raw_item in raw_items:
        let hashed_item = (siphash_key(siphash_key, raw_item) * F) >> 64

    // Sorted in ascending order.

    gcs_compress(hashed_items, fp)

Filter Capability Querying

As it's feasible that in the future, this document is extended to encompass additional filter encoding algorithms or filter contents, we define a new p2p message that allows light clients to ascertain which filters a node supports.

The getcftypes message is an empty message whose command string is: getcftypes

A full-node that receives a getcftypes message MUST respond with a cftypes message which is defined as follows:

Field Size Description Data Type Comments
Var-Int NumFilters uint64 The number of supported filters.
NumFilters SupportedFilters [NumFilterBytes]byte A byte slice with each byte denoting a supported filter type

Compact Filter Header Chain

As the filters described in this BIP are not consensus critical, meaning each filter is ``not`` validated by full-nodes and committed into blocks by miners, we require an alternative (albeit less-binding) method to allow light clients to identify and reject invalid filters. The purely p2p solution to this problem is to obtain a deterministic hash-chain of each filter. This hash chain or "filter header chain" is similar to the regular Bitcoin headers in that it allows a light client to verify the authenticity of a received filter.

The filter header chain for a particular filter type is described by the following recurrence:

filter_header(n: uint) -> [32]byte = 
   // The zero hash is 32 bytes of 0's.
   let zero_hash [32]byte = {0*32}

   if n == 0:
       double-sha-256(genesis_block.prevblock || filter(0))

   match filter(n): 
      // If the filter isn't empty, then we hash the filter itself into the
      // header chain.
      case Some:
          double-sha-256(filter_header(n-1) || double-sha-256(filter(n)))

      // Otherwise, if the filter is empty (created from a block with a single
      // coinbase transaction whose output script contains no push datas), then
      // we'll hash the zero_hash.
      case None:
          double-sha-256(filter_header(n-1) || double-sha-256(zero_hash))

   where filter(n) is the filter for block height n

The filter header for the genesis block uses the hash stored in the prevblock field of the genesis block header itself, as there's no prior filter header (by definition).

Due to the nature of filter construction, it's possible to construct a block such that an "empty" filter will be produced. This is the case of a coinbase transaction that has no data pushes in its output script. In this case, the "hash" of said filter is simply "32 zeroes".

We now introduce two new messages to support the fetching and verification of the filter header chain by light clients.

The getcfheaders message is defined as follows:

Field Size Description Data Type Comments
Var-Int NumBlockLocators uint64 Number of block locators.
NumBlockLocators * 32 BlockLocatorHashes [NumBlockLocators][32]byte Block locator hashes, with the same semantics as in getheaders.
32 HashStop [32]byte Hash to stop at.
1 FilterType byte Type of filter header being requested.

The BlockLocators within the message are to be interpreted identically to the BlockLocators within Bitcoin's getheaders and getblocks messages [10].

The cfheaders> message MUST be sent in response to a getcfheaders message for a particular block hash. The cfheaders message is defined as follows:

Field Size Descriptions Data Type Comments
32 StopHash []byte Block hash for the last filter header returned, for locating the filter headers in the blockchain.
1 FilterType byte Byte identifying the type of filter headers being returned.
Var-Int NumHeaders uint64 Hash to stop at.
NumHeaders * 32 HeaderHashes [NumHeaders][32]byte Slice of filter headers.

Compact Filters

The last set of messages we introduce are for fetching the compact filters themselves. Light clients can SHOULD use these two messages to request a compact filter for a particular block hash.

The getcfilter message is defined as follows:

Field Size Description Data Type Comments
32 BlockHash [32]byte Block hash of the Bitcoin block for which the client wishes to fetch a filter.
1 FilterType byte Byte identifying the type of filter requested.

The cfilter message MUST be sent in response to a getcfilter message for a particular block hash.The cfilter message is defined as follows:

Field Size Description Data Type Comments
32 BlockHash [32]byte Block hash of the Bitcoin block for which the filter is being returned.
1 FilterType byte Byte identifying the type of filter being returned.
Var-Int NumFilterBytes uint64 A variable length integer encoding the number of bytes of the filter in the following field.
NumFilterBytes FilterBytes [NumFilterBytes]byte The raw compressed compact filter for this block.

The BlockHash field is included in both messages as this allows easily matching requests against responses, as the responses aren't self-identifying like block headers are (via own hash).

The parameters N (the number of elements in the filter) and P (1 << false_positive_rate) are required by the light client in order to properly incrementally decode, query, and validate (reconstruct from Bitcoin block) a compact filter. The parameter N cannot be known ahead of time, therefore we define the serialization of a compact filter of type 0x00 and 0x01 as:

N || raw_filter_bytes
where N is serialized as a 32-bit big-endian integer.

However, there exists a special case of a null filter. This this case an empty byte slice is transmitted rather than consuming 4-bytes to encode the size of zero.

However, as the parameter P MUST be globally agreed upon (for a particular filter type), we define this value statically for filter types: 0x00 and 0x01. For the two aforementioned filter types, the false positive rate MUST be: 20, meaning the parameter P is: 2^20, meaning fp=20. This value was chosen as during simulations it was the value that minimized the bandwidth utilized by the expected number of blocks downloaded due to false positives, and the bandwidth used to download the filters themselves. The code along with a demo used for the parameter tuning can be found [here]

Protocol Version Bump

As this BIP defines new peer-to-peer behavior, we bump the protocol version by one in order to distinguish the newly defined behavior. Full-nodes implementing this BIP should advertise a protocol version of: 70016.

New Wallet Capabilities Enabled

The new light client mode enables wallet to maintain a very compact client-side index of (possibly) the entire chain. Such an index provides a great degree of utility for wallets, as they're now able to perform tasks such as private key imports and full HD-seed imports without the need of a trusted third-party server. Additionally, the compact client-side chain index also opens up the door to smart contract applications which require agent action in response to on-chain events. Examples of such applications include Lightning.

Backwards Compatability

This light client protocl is NOT backwards compatible with BIP 37. Full nodes MAY implement both protocols to serve both types of light clients.

Implementation Notes

This filter header chain SHOULD be utilized by light clients to gain a greater degree of security against bamboozling full-nodes during their initial chain sync. In addition to fetching all the bitcoin headers, light clients implementing this BIP SHOULD also fetch all the filter headers from each of their connected peers. With these headers, light clients SHOULD efficiently detect nodes that advertise a conflicting filter chain history. To do this, light clients MUST ensure that all nodes return the ``same` filter header hash for a particular block header hash.

Light clients MAY use the filter header chain to verify purported filter authenticity when fetching the next set of headers from chain tip. Light clients MAY use the following algorithm to more efficiently verify the authenticity of filters (the naive version would fetch the entire filter from each peer, this version saves bandwidth):

verify_from_tip(tip_block_hash: [32]byte):
    let filter_types = {supported_filter_types...}
    let connected_peers = {list_of_connected_full_nodes...}

    for filter_type in filter_types:

        let filter_headers = set()
        for peer in connected_peers:
            let filter_header = peer.fetch_filter_header(tip_block_hash)

        if len(filter_headers) != 1:
            // Peers have conflicting filters. The light client should fetch
            // each unique filter from the set of peers AND fetch the block. The
            // light client can then verify which filter header is correct, and
            // BAN the offending peers.

        // Otherwise, syncing continues as normal: fetch filter to see if it
        // matches any relevant items.

Light clients MAY persistently commit all filter headers to disk, as when lazily fetching filters (due to a historical re-scan or chain analysis), they're able to verify the authenticity of any fetched filters.

Full-nodes MAY persistently compute and persist the filter header chain on-disk, just as the regular filters.

Zero-length filters are sent without an N value, allowing us to save 4-bytes. Clients are able to verify that a filter will be null before requesting it (as it will just be the prior filter header hashed with zero bytes). Clients can take this fact into account in order to save a round trip for null headers.

Light clients implementing this proposal SHOULD utilize the sendheaders message. This allow quicker syncing from tip, as rather than sending an inv to announce a new blocks, nodes will instead directly send the headers message. With this, nodes can save a round-trip and immediately request cfheaders from each of their connected peers before ultimately fetching the filter itself.

Light client syncing MAY be run in reverse, meaning fetching the regular Bitcoin and filter headers from the end of the chain and working backwards. This allows a client to nearly instantly be useable if an application doesn't require immediate access to historical filter data.

If fetching blocks directly over the p2p network (rather than via a distinct channel), light clients SHOULD fetch blocks from multiple peers in order to mitigate transaction intersection analysis.

Full nodes serving the light clients described in this BIP can be implemented in a completely stateless manner. Implementations can also make a space/time tradeoff by only holding filters on disk to some particular block horizon. If a filter beyond this horizon is requested, then it can be reconstructed in real time. This allows nodes to save on disk-space, as it's likely the older filters would only be fetched for historical rescans. It is also likely that the past few hundred filters will be fetched mostly frequently by smaller devices (phones, laptops, etc) periodically coming back online after a period of inactivity.

It is possible to implementations of this BIP to serve other implementations to a degree. All headers (regular bitcoin, regular, extended) can be served, and also any on-disk filters can also be served to other light clients.

Key import and rescan: with lazy filters fetching, having a start block is VERY important to avoid fetching all cfilters starting with block 1 (assume you generate your own filters/headers for the genesis block).


We would like to thank bfd (from the bitcoin-dev mailing list) for bringing the basis of this BIP to our attention, Greg Maxwell for pointing us in the direction of Golomb-Rice coding and fast range optimization, Joseph Poon for suggesting the filter header chain scheme, and Pedro Martelletto for writing the initial indexing code for btcd.

We would also like to thank Dave Collins, JJ Jeffrey, Eric Lombrozo for useful discussions.

Reference Implementation

Light client:

Full-node indexing:

Golomb-Rice Coded sets:

Appendix A

Mathematical Background

In the following sections, borrowing from techniques typically used in image and video processing, we describe our chosen encoding for the hash fingerprints of the items in our set of relevant items. In order to compress the items of the set in a lossy manner (creating data-structure capable of probabilistic set membership), we utilize Golomb-Rice codes to encode the delta between successive hash items within our set. This results in a very compact probabilistic set-membership structure.

With a goal of building relevant initiation in the minds of the readers of this document, we first start from the bottom of the abstraction ladder, describing the fundamental components our set encoding relies on.

Run-Length Encoding

Run-Length Encoding (or RLE) is typically used in the video/image processing space to losslessly compresses images, or video frames. RLE works by omitting the encoding of repeated values in a data stream. This achieves lossless compression as repeated items simply aren't transmitted. Instead, a value which represents the number of times a value repeats is transmitted.

Typically RLE takes the form of encoding repeated values in a binary stream. A simple RLE scheme works as follows:

  • Encode the run length (number of occurrences) of 0's using k bits.
    • k acts as fixed length encoding for the length of a run.
    • This value acts as the maximum encodable run-length.
  • Transmission of runs of 1's is omitted.
  • Two 1's in a row are denoted by a zero-length run of zero.
As an example, consider the following sequence of bits:
{0}^14 1 {0}^9 11 {0}^20 1 {0}^30 11 {0}^11

The RLE of the bit stream above would be:

1110 1001 0000 1111 0101 1111 1111 0000 0000 1011

RLE allows one to efficiently encode a data stream in a lossless manner. Due to the encoding of runs, RLE works best when encoding a set with a high degree of redundancy. A careful reader will notice that by using a fix-length encoding for the size of runs, efficiency is lost. Therefore, rather than using a fix-length encoding for the size of a run, we can instead use a variable length encoding for the size of a run. This allows us to compress runs of a large size. To do so, we'll now turn to Golomb-Rice Coding.

Golomb-Rice Coding

RLE works well when encoding a data stream that has a high degree of redundancy. However, in our case due to the hashing of items within the compact filter, we'll be dealing with items that are uniformly distributed. We can use this fact to leverage a more efficient encoding scheme based on the distribution of the length of a run. The Distribution represents the probabilities of a number of failures before the first success in a series of Bernoulli trials (yes/no experiments). If our values are i.i.d (independent, identically distributed) distributed of the run-length r can be represented as [6]:

P(r = n) = p^n * (1-p)
Intuitively, this calculates the probability of N zeroes (a run) followed by a single 1 (end of a run). Golomb coding takes advantage of this relationship to efficiently encode integers using a two-tuple. Given a group size of m one can encode an integer as:
n = (q*m) + r
  where q is (n / m)
   and  r is n % m

Golomb Coding encodes the two values (q and r for a given integer n as a two-tuple. The first value q is encoded using unary, and the second value r is encoded using a fixed-length series of bits. If m = 2^k for some k then this encoding is a specialized sub-set of Golomb encoding known as Golomb-Rice encoding. In this case, r (the remainder) is the k least-significant-bits of n

In this case "runs", can be seen as the number of multiples of m that divide into n If an encoded integer is close to the value of m then few bits (in unary) will be used to encode each value.

To aid in understanding we provide the following examples of using Golomb-Rice encoding to code integers given m=5

n  = (q, r) = c
0  = (0, 0) = 0 00
1  = (0, 1) = 0 01
2  = (0, 2) = 0 10
3  = (0, 3) = 0 110
4  = (0, 4) = 0 111
5  = (1, 0) = 10 00
6  = (1, 1) = 10 01
7  = (1, 2) = 10 10
8  = (1, 3) = 10 110
9  = (1, 4) = 10 111
10 = (2, 0) = 110 00

The P value can also be interpreted as the parameter to our Geometric Distribution. Intuitively, to achieve a false positive rate of 1/32 (1/2^5), in a series of queries of items which aren't in the set, we expect to receive a "NO" (false) 32 times, before getting a "YES" (true, our false positive). Once again, P MUST be a power of two.

Appendix B


A number of alternative set encodings were considered before Golomb-Rice coded sets were settled upon. In this appendix section, we'll list a few of the alternatives along with our rationale for not pursuing them.

Cryptographic Accumulators

Cryptographic accumulators [11]are a cryptographic data structure that enables (amongst other operations) a one way membership test. One advantage of accumulators are that they are constant size, independent on the number of elements inserted into the accumulator. However, current constructions of cryptographic accumulators require an initial trusted set up. Additionally, accumulators based on the Strong-RSA Assumption require mapping set items to prime representatives in the associated group which can be preemptively expensive.

Matrix Based Probabilistic Set Datastructures

There exist data structures based on matrix solving which are even more space efficient compared to bloom filters [12]. We instead opted for our GCS-based filters as they have a much lower implementation complexity and are easier to understand.

Appendix C

Test Vectors

A set of test vectors using real blocks from Bitcoin's testnet3 can be found here. The vectors come in the form of CSV files, which for a given testnet height and block hash contain the: basic filter header, extended filter header, basic filter, and extended filter.


  1. ^
  2. ^
  3. ^
  4. ^
  5. ^
  6. ^
  7. ^
  8. ^
  9. ^
  10. ^
  11. ^
  12. ^


This document is licensed under the Creative Commons CC0 1.0 Universal lisence.