Research
What Does It Cost to Force-Close a Lightning Channel? Evidence from 974,000 Channel Closes, 2018-2026
When two Lightning nodes cannot agree, either can broadcast a pre-signed "commitment" transaction that closes the channel on-chain, then claim its balance through further pre-signed, time-locked…

- arXiv
- 2611.16311
- Published
- 2026-11-12
- Authors
- Saurabh Kumar
AI summary
Overview
Research area: Bitcoin and the Lightning Network — on-chain measurement of payment-channel closures, transaction-fee economics, and the security implications of delayed confirmation. Falls at the intersection of empirical blockchain forensics, protocol design, and Layer-2 security.
Technical level: Intermediate. The paper is readable without a protocol-engineering background, but it assumes familiarity with what a commitment transaction, an HTLC, an anchor output and a to_self_delay are.
Scope (one sentence): A chain-wide census of 973,581 Lightning channel closes — 523,815 of them force closes — from January 2018 to October 2026, tracing every output to its spend in order to price what a failed channel costs in fees, block space, locked time, and exposed payments across the fee regimes the old design lived through.
What This Paper Is About
When two Lightning nodes cannot agree, either can broadcast a pre-signed "commitment" transaction that closes the channel on-chain, then claim its balance through further pre-signed, time-locked transactions. Those transactions carry fees and delays, and a transaction that confirms too late can cost an operator a pending payment. That fee design is being replaced right now — Bitcoin Core 31.0 (April 2026) removed the CPFP carve-out that anchor channels were built around, and the Lightning specification adopted zero-fee commitments in May 2026. The paper's goal is to measure, from the chain itself, what the outgoing design actually cost, so the migration can be judged against a real baseline rather than forum argument.
Key Contributions
-
A census with rates, not just counts. The authors trace 973,581 closes (523,815 force closes), follow 1,346,588 force-close outputs to their spends, and convert raw counts into per-channel rates and cohort incidence — for example, 35% of public channels funded in 2019 were force-closed within a year, against 17% of those funded in 2025.
-
The price of a force close in three currencies. Fees, block space and locked time are measured together: a force close costs 2–6 times a cooperative close in fees, about three times its block space since 2022, and the closer waits a median of 1.7 days for its own balance. Counting the open, a public channel's whole life cost a median of 3,435 sat on-chain.
-
The anchor money trail. Every P2WSH anchor output is classified by who spent it: owners spent 81% of anchors in 2022 and 5% in 2026, third parties sweep the rest, and fees took 87% of the swept value in 2026. No P2WSH anchor was left unspent and no third party has swept a simple-taproot anchor.
-
A placebo-tested answer to "do fee spikes cause force closes?" plus a zero-fee counterfactual. Force closes in a spike week and the week after were 0.97 times the four weeks before, against 1.02 after ordinary weeks of the same years. Across the anchor force closes observed, zero-fee commitments would have saved 12% of the anchor mechanism's block space.
Main Findings
-
Fee spikes did not bring a wave of force closes. Organic public force closes in a spike week and the following week were 0.97 times their level in the four weeks before, versus 1.02 for placebo weeks of the same years. The result holds for pre-anchor channels, whose commitment fee is fixed at signing, as well as for anchor channels.
-
The cost is a ratio, not an absolute. The median public force close cost 925 sat in 2026, and the median was highest in 2020 at 14,318 sat, against 7,346 sat in 2023. Against a cooperative close of the same year, the ratio is 2–6 times in every year: 2.1 in 2019, 6.1 in 2022, 3.7 in 2023, 4.7 in 2026.
-
Block space roughly doubled with anchors. An anchor force close uses a median 639 vB against 292 vB for a pre-anchor one. Since 2022 a force close has taken 3.0–3.1 times the block space of a cooperative close.
-
The wait dominates the fee only in cheap-fee years. Closers' balances were locked for 68,036 BTC-days over the period; 49% of
to_self_delayvalues are 144 blocks and the slowest tenth of closes wait up to 22 days. The break-even return on locked funds was 2.7% in 2025 and 1.2% in 2026, but 6.1–54.1% a year before 2025, and 16.7% over the whole period. -
Force closing as a share of public closes fell, but less than counts suggest. The force share fell from 44% in 2023 to 20% in 2026, while the organic force-close rate fell from 41 to 26 per 1,000 open channels per month. Open public channels peaked at 90,340 and stood at 40,353 at the cutoff.
-
Most force closes now happen outside the public graph. In 2026, force closes of channels not in the public list outnumbered public ones 3.4 to one, and 23% of 2026's unlisted force closes were simple taproot channels. The largest burst in the data is unlisted: on 1 September 2024, 8,990 unlisted force closes holding 36.7 BTC confirmed within about 2 hours (18 blocks), 6,910 of them pre-anchor.
-
Mass-close events persist. Each year has 64–239 mass-close events (one node party to at least 20 public force closes on one UTC day), holding 19–32% of public force closes; the largest put 856 channels on-chain in one day.
-
Late HTLC timeouts grow with fees, and anchors help. Of 70,808 HTLC-timeout transactions, the median confirmed 1–3 blocks after expiry. More than 40 blocks late were 6.5–8.0% of timeouts each year 2019–2022, then 18.2% in 2023 and 16.4% in 2024. At a block median fee of 50 sat/vB or more, 30.5% of pre-anchor timeouts were more than 40 blocks late against 20.6% of anchor ones; at 5–20 sat/vB the gap was 13.2% against 4.7%. The exception is 2024, when anchor timeouts were late more often (12.8% against 6.3%).
-
Actual forwarding losses are tiny. Since 2021, 40 payments had a forwarding node's two HTLCs resolved in opposite ways, losing 0.48 BTC in all; in 38 of them (0.44 BTC) the upstream HTLC timed out while the downstream one was paid, and in 100% of those the preimage was already on-chain a median of 58 blocks before the upstream timeout confirmed. No organic loss has the ordering a replacement-cycling attack leaves; the 7 pairs from 2026 (0.0016 BTC) that do are a test one operator ran against its own nodes.
-
One payment can take many channels down. 25% of force closes with an HTLC resolved on-chain share a payment hash with another force close: 6,951 payments, up to 18 closes for one payment, a median of 140 blocks between the first and the last, with 96% timed out on every channel.
-
Penalties are rare except for one engineered week. 950 penalty transactions took 11.4 BTC over the whole period. Between 28 September and 5 October 2026, revoked commitments of 306 private anchor channels were broadcast and every one was penalized within 120 blocks; 305 (3.18 BTC) were splices opened a median of 5 blocks earlier by one funder. The transactions match a Core Lightning bug fixed in version 26.06.8, released 19 September 2026.
-
Anchor outputs became third-party dust. Owners spent 93.7% of anchors in 2021, 81.3% in 2022 and 4.8% in 2026. Third-party sweeps arrive at the first block they can from 2025, chain-wide sweep transactions rose to 10,390 in 2026, and miners took 87.0% of the swept value that year against the 6.38 sat/vB break-even rate.
-
Lightning is a rounding error in Bitcoin's fee market. All of Lightning's closes and required spends took at most 0.18% of Bitcoin's transaction fees in any year; the whole dataset accounts for 70.92 BTC of lifecycle fees across 692,525 public channels.
-
Zero-fee commitments are a small block-space win. The first zero-fee commitment closed on 29 October 2025, before the specification merged. Measured against observed anchor force closes, the format would save 12% of the anchor mechanism's block space — so the authors argue it should be judged by how promptly payments resolve when fees spike, not by saved bytes.
Methodology in Plain English
The authors ran their own Bitcoin Core node and scanned every block from block 503,816 (before Lightning's first channels) to block 974,041 (the last block before 1 November 2026), keeping per-block fee statistics. They assembled a list of public channels by joining their own gossip since May 2023, mempool.space's record of closed channels and close reasons, and the lnresearch gossip archive for earlier years, resolving it to 736,110 funding outpoints.
Force closes are identified by a fingerprint baked into every commitment transaction: the upper byte of its locktime is 0x20 and the upper byte of its single input's sequence is 0x80, because those fields encode the obscured commitment number. That fingerprint also finds force closes of channels outside the public list, which the authors call "unlisted." Each of the 1,346,588 outputs of the 523,815 force closes was followed to the transaction that spent it (1,212,596 had been spent by the cutoff) and classified from its witness script: the broadcaster's delayed sweep, the peer's sweep, an HTLC transaction or claim, an anchor spend by owner or third party, or a penalty.
Validation is two-pronged: a separate chain index built by a different program reproduces the closing transaction for all 698,476 comparable public closes and matches the force/cooperative label in every case; and mempool.space's close reasons, where given, agree with the authors
Authors’ abstract
Bitcoin's Lightning payment network returns the funds of a channel that cannot close cooperatively through pre-signed transactions whose fees and delays fall on its operators. That fee design is being replaced: Bitcoin Core 31.0 removed the mempool rule anchor channels relied on, and the Lightning specification adopted zero-fee commitments in May 2026. What the old design costs has never been measured across fee regimes. We trace every public channel close and every force…
The full abstract on arXiv