Pith. sign in

REVIEW 6 major objections 6 minor 49 references

PRIME: Efficient Algorithm for Token Graph Routing Problem

T0 review · 6 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PRIME claims near-optimal, sub-second routing for token swaps by combining pruned graph search with a sign-gradient allocation method.

desk verdict Plausible new DEX routing framework with promising empirical gains over SOR, but the linear-convergence proof is invalid as written and the headline experiments lack error bars. read the letter →

arxiv 2603.08337 v2 pith:DEDKUZO7 submitted 2026-03-09 cs.DB

Haotian Xu , Yuqing Zhu , Yuming Huang , Jing Tang This is my paper · ORCID
classification cs.DB
keywords tokengraphroutingdecentralizedexchangeconstantfunctionmarketmakerconvexoptimizationgeneralizednetworkflowsigngradientmethodqueryliquidity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that token-swap routing on decentralized exchanges—splitting a trade across liquidity pools to maximize output—can be solved near-optimally and fast despite concave price functions, graphs with hundreds of thousands of tokens, and price spans of eleven orders of magnitude. Its algorithm, PRIME, first prunes the graph to a small core of high-liquidity hubs plus precomputed shortcuts, then finds candidate paths by accepting only routes whose marginal price beats a threshold, and finally allocates the trade with a sign-gradient optimizer. The paper reports that PRIME beats the incumbent Uniswap Smart Order Router by up to 8.42 basis points on large trades and by larger margins on some small trades, while cutting computation by up to 96.7%. It also reports near-optimality against a global convex solver once that solver's residual imbalances are corrected. If these claims are right, PRIME is a practical graph-query solution for high-frequency decentralized markets.

What carries the argument

The load-bearing mechanism is the pair of (1) marginal-price-threshold path discovery—a BFS pruning search on a Core Graph of hub tokens, augmented by a Shortcut Index of precomputed non-core paths, that accepts a new path only if its marginal price exceeds the current threshold—and (2) the Adaptive Sign Gradient Method (ASGM), an allocation optimizer on the probability simplex that repeatedly moves funds from the path with the lowest marginal price to the path with the highest marginal price, using only the sign of the gradient difference and a backtracking line search. The paper proves (Theorem 1) that the strongly concave allocation objective makes ASGM converge at a linear rate, and the

What would settle it

Run a large real swap (e.g., 1,000 WETH to USDC) on a snapshot where an exhaustive search up to depth 6 through non-core tokens finds an execution at least 3 basis points better than PRIME's output after the same dust-correction is applied to both; or construct a synthetic graph with a deep non-core chain carrying the best marginal price and show PRIME misses it.

Watch

Extended reading notes

Core claim

PRIME's central claim is that the Token Graph Routing Problem has a near-optimal solution computable in milliseconds by decomposing the graph into a Core Graph of top-liquidity tokens and a Shortcut Index of better-priced paths through non-core tokens; discovering paths with a pruned BFS that accepts only marginal-price improvements; and optimizing allocations with ASGM over paths that do not reuse the same pool. The paper demonstrates on real Ethereum snapshots that this beats the standard smart-order router by up to 8.42 basis points on large trades while reducing computation up to 96.7%, and that its output is within a few basis points of a global convex optimum after correcting that solv

Load-bearing premise

The pruning premise—that every materially better route goes through top-liquidity hubs or is captured by the precomputed shortcuts, within a few hops—is load-bearing; if the best execution ever lives on a long path through non-core tokens, PRIME will not see it.

Editorial extensions

If this is right

  • Large swaps on fragmented decentralized-exchange liquidity gain several basis points over incumbent heuristics while staying fast enough for sub-second execution.
  • Small trades also gain, because the Shortcut Index recovers better-priced long-tail pools that aggressive liquidity-filtering heuristics remove.
  • The equal-marginal-price condition is a principled optimality stop, because the allocation objective is strongly concave and has a unique optimum.
  • Computation stays near-constant as the graph grows, unlike convex-solver time which scales with the number of pools.
  • The 85–97% time reduction makes the approach viable for repeated routing queries as market states change.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the core-graph-plus-shortcuts decomposition should generalize to other markets with concave swap-like gain functions, such as order-book depth approximated as a concave curve; a synthetic order-book graph benchmark against full enumeration would test this.
  • Editorial inference: ASGM's sign-only, backtracking-line-search updates may transfer to other ill-conditioned convex allocation problems on the simplex, but the paper only establishes them for token allocation; a badly scaled synthetic quadratic would separate optimizer generality from the routing application.
  • Editorial inference: PRIME's path-discovery can act as a feasibility repairer for any global convex solver, not just as a standalone router—the paper's post-processing comparison already illustrates this, and a concrete test would feed arbitrary solver outputs through the repair step and measure dust elimination.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 6 minor

Summary. The paper formalizes the Token Graph Routing Problem (TGRP) on DEX swap graphs with concave edge functions, and proposes PRIME, a two-stage algorithm. Stage 0 preprocesses the large token graph into a hub-based Core Graph and a Shortcut Index; Stage 1 iteratively discovers candidate single-edge paths using a BFS pruning method; Stage 2 merges/expands the candidates into multi-edge paths and solves the resulting allocation problem with the proposed Adaptive Sign Gradient Method (ASGM). The paper claims ASGM converges linearly to the optimal allocation (Theorem 1) and reports experiments on real Ethereum data showing that PRIME outperforms Uniswap SOR by up to 8.42 bps in execution price while reducing computation by up to 96.7%. A comparison against a convex SOCP optimum on small graphs is used to support near-optimality, and the paper reports deployment in a hedge fund production environment.

Significance. If the claims hold, PRIME would be a practically valuable and scalable DEX router, with a clean formalization of the problem, a carefully engineered two-stage algorithm, and a strong empirical comparison against the industry standard. The paper's strengths include the use of real Ethereum data across three market regimes, ablation studies for the shortcut index and flow relaxation, and an honest attempt to compare against a convex-optimization upper bound. However, the central theoretical result — ASGM's linear convergence — is not established by the current proof, and the main empirical comparison lacks error bars and independent validation of the pruning assumptions. The significance is therefore conditional: the engineering contribution is plausible, but the advertised theoretical guarantee and the strength of the empirical claims need substantial revision.

major comments (6)
  1. [Appendix A, Lemma 5 and Eqs. (14)–(17)] Theorem 1 is not established. Eq. (15) gives δ_{t+1} ≤ δ_t − ρδ_t², which is a quadratic decrease and yields at best sublinear tail convergence, not a uniform linear contraction. The proposed ρ′ = min{ρδ_0, αcΔ_0/J*} does not fix this: from δ_t ≤ δ_0 one gets 1−ρδ_t ≥ 1−ρδ_0, so the inequality cannot be turned into δ_{t+1} ≤ (1−ρ′)δ_t. Lemma 5 is also asserted without a valid proof: the two-coordinate SIGN direction is not a Frank-Wolfe vertex direction, and the duality-gap bound from [31] does not automatically transfer to this update. The abstract and contributions advertise linear convergence, so this must be either proved under explicit assumptions or downgraded to a sublinear rate.
  2. [Section III and Appendix A, Lemma 2] The paper claims that strict concavity on a bounded domain implies strong concavity. This is false in general; strong concavity requires a uniform lower bound on the second derivative (e.g., f'' ≤ −μ < 0). It happens to hold for the specific CPMM function when reserves are bounded away from zero, but it is not a consequence of Definition 1. Without an explicit μ-strong-concavity certificate for the composed path functions, the premise of Theorem 1 is not guaranteed. Please state the required curvature assumption and verify it for the swap functions used in the experiments.
  3. [Algorithm 3, lines 7–14] The ASGM update W_t ← W_{t−1} + SIGN·Δ is not feasible-preserving. If the worst path p− has current allocation W_{p−} < Δ, the update produces a negative weight. The Armijo backtracking checks only the objective increase, not nonnegativity or simplex feasibility, and no projection or clipping step is specified. As written, the algorithm can leave the feasible set. The paper should specify how feasibility is enforced (e.g., Δ ≤ W_{p−}, or projection onto the simplex) and analyze the effect on the convergence argument.
  4. [Table III and Figure 3] The main empirical comparison is based on a single one-hour window (00:00–01:00 UTC+8 on Jul 14, 2025) with no error bars or repeated trials, and Figure 3 reports daily averages without variance. The text claims PRIME 'consistently' outperforms SOR across bearish, stable, and bullish periods, but no standard errors, confidence intervals, or significance tests are provided. Given the small bp differences in many rows (e.g., 0.00–2.36 bp at 10 WETH), it is important to show the distribution over the 3,024 snapshots or at least per-period variance, and to state how many independent market states are used for each average.
  5. [Section VI-E, Table IV] The 'adjusted convex' optimum is produced by post-processing the raw SOCP solution with PRIME's own path discovery. This makes the near-optimality comparison partially circular: the repair procedure is a PRIME component, and it can bias the adjusted solution in either direction. The paper even reports that PRIME beats the adjusted convex solution in the 117-pool case. Please provide an independent feasibility repair (e.g., a different reserve-clearing algorithm), or bound the loss introduced by the repair, or report both raw and adjusted convex values with an explicit accounting of surplus/deficit.
  6. [Section V-A, VI-A, Algorithm 2] The near-optimality claim depends on three pruning choices: the hub set H selected as top-K by liquidity, the maximum hop count m≈3–4, and the removal of 96.65% of leaf tokens that connect only to WETH. No sensitivity analysis with respect to K and m is reported, and the convex-optimum comparison is limited to tiny V2 graphs (3–25 tokens). The reader cannot tell whether the pruned Core Graph plus Shortcut Index contains all materially better routes on the full graph. Please report performance as K and m vary, and evaluate against an exhaustive optimum on medium-size graphs to validate the pruning assumptions.
minor comments (6)
  1. [Table III caption] The caption refers to 'PRIME (No Shortcut)' but the table header uses 'PRIME-Core'. Use one consistent name.
  2. [Section III and Appendix A] The optimality gap is denoted both D(W_t) and δ_t. Pick one notation and use it consistently.
  3. [Section VI-A / Figure 3 caption] Figure 3 is described as 'daily average performance' while the text says market states were sampled at 10-minute intervals (3,024 snapshots). Clarify what is averaged and over what period.
  4. [Section V-C, Algorithm 3] The text says 'p+ reduces its allocation while p− increases its allocation', but Algorithm 3 assigns SIGN[p+] ← 1 and SIGN[p−] ← −1, which moves flow from p− to p+. Please fix the sign convention in the prose.
  5. [References] Reference [30] is cited for SPFA, but the cited Ahuja et al. paper may not be the source of the Shortest Path Faster Algorithm; please verify the reference.
  6. [Throughout] There are several typos, e.g., 'continusely', 'againts', and 'execution prices are presented as value ×10^3' for WBTC. A careful proofread is needed.

Circularity Check

1 steps flagged · score 2.0 of 10

Core derivation is not circular; only the 'theoretical optimum' benchmark is partly self-referential because PRIME's own path discovery repairs the convex solver's infeasible output. The ASGM proof gap is a soundness risk, not a circularity.

  1. other [Section VI-E (Comparison with Theoretical Optimum), Table IV and Figure 4]
    "To enable a valid comparison, we use our path discovery algorithm to post-process the raw convex output into a feasible solution. ... Ultimately, the adjusted convex output offers a negligible advantage of at most 3 bps, and PRIME even secures superior results in more complex 117-pool scenarios."

    The 'theoretical optimum' benchmark is supposed to be an independent upper bound, but the feasibility repair is performed by PRIME's own path-discovery component. The adjusted convex solution is therefore not an independent benchmark: it is a convex allocation completed by the algorithm under evaluation. The near-optimality conclusion drawn from Table IV and Figure 4 is thus partly self-referential, although the comparison is not equal by construction and the repair generally lowers the convex solution's value.

full rationale

The central derivation of PRIME is not circular. Path discovery (Algorithm 2) and the allocation optimizer (ASGM) are defined independently of the experimental outputs, and the main empirical claims are benchmarked against external baselines: OSP, the Uniswap Smart Order Router, and a Gurobi-solved SOCP convex relaxation. No fitted parameter is renamed as a prediction, and there are no load-bearing self-citations; the reference list contains no self-citations by the authors. The one genuinely circular-ish element is in Section VI-E: the comparison against the 'theoretical optimum' uses PRIME's own path discovery to post-process the raw convex solution into a feasible one, so the adjusted convex benchmark is no longer fully external. This weakens the independence of the near-optimality claim but does not reduce PRIME's output to the benchmark by construction. Separately, the proof of Theorem 1 in Appendix A contains a load-bearing gap: Lemma 5 imports a Frank-Wolfe style bound from [31] without verifying that ASGM's two-coordinate SIGN update is a Frank-Wolfe vertex direction, and Eq. 15's quadratic decrease bound, δ_{t+1} ≤ δ_t − ρδ_t^2, does not imply Eq. 17's uniform linear contraction δ_{t+1} ≤ (1 − ρ′)δ_t. This is a correctness/soundness problem, not a circularity, and therefore does not increase the circularity score. Overall, the paper's central claims retain independent content; the circularity score is 2.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the model and optimization assumptions above; the free parameters are engineering choices tuned on the same data, and no independent benchmark of the complete algorithm against prior specialized route optimizers is provided.

free parameters (6)
  • Hub set H (top-K by liquidity) = K unspecified in text
    Defines the Core Graph; if the true best route goes through non-hub/non-shortcut tokens, PRIME loses it. Empirically chosen in Sec. V-A.
  • Max hop count m = 3 or 4 (typical)
    Algorithm 2 prunes paths with len(p) >= m; justified only by empirical observation about slippage (Sec. V-B).
  • Armijo parameter alpha = 1e-4
    Chosen by ablation in Appendix B; reported insensitive.
  • Decay factor beta = 0.5
    Chosen by ablation in Appendix B as latency/precision tradeoff.
  • Input-time averaging window = 1 hour
    Main results use one-hour average starting 00:00 UTC+8 on 2025-07-14 (Sec. VI-C); a different window alters results.
  • Leaf-token pruning = removes 96.65% of tokens
    Tokens connecting exclusively to WETH are pruned (Sec. VI-A); any such token with a materially better price path would be missed.
assumptions (6)
  • domain assumption Swap functions are strictly concave, increasing, f(0)=0, and strongly concave on bounded domain.
    Definition 1 and Sec. III; true for constant-product pools on bounded intervals, but V3 concentrated liquidity is only piecewise concave and strong concavity is asserted, not proven.
  • domain assumption Composition and weighted sums of such functions preserve strong concavity.
    Lemma 2 (Appendix A) proves this by assertion ('strong concavity is maintained over the finite trade range'); no rigorous argument is given.
  • domain assumption J has L-Lipschitz gradients.
    Appendix A Lemma 3: bounded second derivatives on bounded domain; plausible for CFMMs.
  • ad hoc to paper There exists c>0 with ∇J(W_t)^T SIGN >= c·δ_t (Lemma 5).
    This is the crucial inequality for linear convergence; cited [31] covers Frank-Wolfe variants, but the specific bound for this SIGN direction and simplex geometry is not derived, and the subsequent proof step is invalid.
  • domain assumption Liquidity concentrates in a small hub set; top-K hubs capture the routing backbone.
    Sec. V-A 'liquidity in real-world markets often exhibits a clustering effect' [8]; used to justify pruning; not verified on the test periods.
  • domain assumption Pool-disjoint paths have independent swap functions.
    Sec. III Definition 4, 'consistent with the model proposed by [24]'; a modeling assumption that excludes shared-pool optimal flows (acknowledged via PRIME-Flow).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PRIME: Efficient Algorithm for Token Graph Routing Problem." pith.science (2026). https://pith.science/paper/DEDKUZO7

@misc{pith2026260308337,
  author       = {Pith},
  title        = {Pith review of: PRIME: Efficient Algorithm for Token Graph Routing Problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEDKUZO7}},
  note         = {Machine review of arXiv:2603.08337}
}
read the original abstract

Optimizing asset exchanges on blockchain-driven platforms poses a novel and challenging graph query optimization problem. In this model, assets represent vertices and exchanges form edges, recasting the graph query task as a routing problem over a large-scale, dynamic graph. However, the existing solutions fail to solve the problem efficiently due to the non-linear nature of the edge weights defined by a concave swap function. To address the challenge, we propose PRIME, a two-stage iterative graph algorithm designed for the Token Graph Routing Problem (TGRP). The first stage employs a pruned graph search to efficiently identify a set of high-potential routing paths. The second stage formulates the allocation task as a strongly convex optimization problem, which we solve using our novel Adaptive Sign Gradient Method (ASGM) with a linear convergence rate. Extensive experiments on real-world Ethereum data confirm PRIME's advantages over industry baselines. PRIME consistently outperforms the widely-used Uniswap routing algorithm, achieving up to 8.42 basis points (bps) better execution prices on large trades while reducing computation up to 96.7%. The practicality of PRIME is further validated by its deployment in hedge fund production environments, demonstrating its viability as a scalable graph query processing solution for high-frequency decentralized markets.

Figures

Figures reproduced from arXiv: 2603.08337 by the authors.

Figure 1
Figure 1. An Example Token Route from WETH to USDC [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The Workflow of PRIME Algorithm aims to identify a candidate paths set Psingle. We perform a Breadth-First Search (BFS) pruning algorithm (Algorithm 2) on the core graph Gcore. The algorithm iteratively adds a new path to the set only if it provides a marginal price improvement over the current threshold τ . This threshold is updated using the Adaptive Sign Gradient Method (ASGM) in each iteration. Finally, the algo… view at source ↗
Figure 3
Figure 3. Performance of PRIME relative to SOR across three distinct market periods: Bearish, Stable, and Bullish. The figure [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Scalability Comparison adjusted outputs. While raw theoretical output appears slightly higher, applying a corrective heuristic to ensure feasibility significantly reduces its yield.Ultimately, the adjusted convex output offers a negligible advantage of at most 3 bps, a…
Figure 5
Figure 5. Figure 5: Token swap execution patterns in production protocols. Uniswap V3 offers multiple pools for the same token pair, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Hyperparameter Ablation Study. (a) Execution time remains flat across [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 3 linked inside Pith

  1. [26]

    Optimal routing for constant function market makers,

    G. Angeris, A. Evans, T. Chitra, and S. Boyd, “Optimal routing for constant function market makers,” inProc. ACM EC, 2022, pp. 115– 128

    2022

  2. [29]

    Convex network flows,

    T. Diamandis, G. Angeris, and A. Edelman, “Convex network flows,” arXiv preprint arXiv:2404.00765, 2024

    arXiv 2024

  3. [28]

    An efficient algorithm for optimal routing through constant function market makers,

    T. Diamandis, M. Resnick, T. Chitra, and G. Angeris, “An efficient algorithm for optimal routing through constant function market makers,” inProc. FC. Springer, 2023, pp. 128–145

    2023

  4. [31]

    On the global linear convergence of frank-wolfe optimization variants,

    S. Lacoste-Julien and M. Jaggi, “On the global linear convergence of frank-wolfe optimization variants,”Advances in neural information processing systems, vol. 28, 2015

    2015

  5. [1]

    Dex to cex volume ratio hits new high of 20%: Trading implications,

    B. News, “Dex to cex volume ratio hits new high of 20%: Trading implications,” March 2025. [Online]. Available: https: //blockchain.news/flashnews/dex-to-cex-volume-ratio-hits-new-high-o f-20-trading-implications

    2025

  6. [2]

    Liquidity fragmentation on decentralized exchanges,

    A. Lehar, C. Parlour, and M. Zoican, “Liquidity fragmentation on decentralized exchanges,”arXiv preprint arXiv:2307.13772, 2023

    arXiv 2023

  7. [3]

    Dexs volume in ethereum,

    DefiLlama, “Dexs volume in ethereum,” 2025. [Online]. Available: https://defillama.com/dexs/chains/ethereum

    2025

  8. [4]

    R. K. Ahuja, T. L. Magnantl, and J. B. Orlin,Network flows: theory, algorithms, and applications. Prentice-hall, 1993

    1993

Show all 49 references
  1. [5]

    Uniswap v2 core,

    H. Adams, N. Zinsmeister, and D. Robinson, “Uniswap v2 core,” 2020

    2020

  2. [6]

    Cryptocurrency 24h trading volume,

    Coingecko, “Cryptocurrency 24h trading volume,” 2025. [Online]. Available: https://www.coingecko.com

    2025

  3. [7]

    Numerical Stability Challenges in Uniswap V3 and Curve Pools,

    liewmanchoi, “Numerical Stability Challenges in Uniswap V3 and Curve Pools,” 2024, accessed: 2023-10-25. [Online]. Available: https://github.com/bcc-research/CFMMRouter.jl/issues/29

    2024

  4. [8]

    Power laws in economics and finance,

    X. Gabaix, “Power laws in economics and finance,”Annu. Rev. Econ., vol. 1, no. 1, pp. 255–294, 2009

    2009

  5. [9]

    Truemper,Optimal flows in networks with positive gains

    K. Truemper,Optimal flows in networks with positive gains. Case Western Reserve University, 1973

    1973

  6. [10]

    Minimization of functions having lipschitz continuous first partial derivatives,

    L. Armijo, “Minimization of functions having lipschitz continuous first partial derivatives,”Pacific Journal of mathematics, vol. 16, no. 1, pp. 1–3, 1966

    1966

  7. [11]

    Routing a swap,

    Uniswap, “Routing a swap,” 2024. [Online]. Available: https: //docs.uniswap.org/sdk/v3/guides/swaps/routing

    2024

  8. [12]

    Bitcoin: A peer-to-peer electronic cash system,

    S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” 2008

    2008

  9. [13]

    Ethereum: A secure decentralised generalised trans- action ledger,

    G. Woodet al., “Ethereum: A secure decentralised generalised trans- action ledger,”Ethereum project yellow paper, vol. 151, no. 2014, pp. 1–32, 2014

    2014

  10. [14]

    Formalizing and securing relationships on public networks,

    N. Szabo, “Formalizing and securing relationships on public networks,” First monday, 1997

    1997

  11. [15]

    Decentralized finance: On blockchain-and smart contract- based financial markets,

    F. Sch ¨ar, “Decentralized finance: On blockchain-and smart contract- based financial markets,”Federal Reserve Bank of St. Louis Review, vol. 103, no. 2, pp. 153–174, 2021

    2021

  12. [16]

    Defi protocols for loanable funds: Interest rates, liquidity and market efficiency,

    L. Gudgeon, S. Werner, D. Perez, and W. J. Knottenbelt, “Defi protocols for loanable funds: Interest rates, liquidity and market efficiency,” in Proceedings of the 2nd ACM Conference on Advances in Financial Technologies, 2020, pp. 92–112

    2020

  13. [17]

    SoK: Decentralized Finance (DeFi),

    S. M. Werner, D. Perez, A. Klages-Mundt, D. A. Zetzsche, and R. P. Buckley, “SoK: Decentralized Finance (DeFi),”IEEE S&P, pp. 1–1, 2022

    2022

  14. [18]

    Uniswap v3 core,

    H. Adams, N. Zinsmeister, M. Salem, R. Keefer, and D. Robinson, “Uniswap v3 core,”Tech. rep., Uniswap, Tech. Rep., 2021

    2021

  15. [19]

    Stableswap-efficient mechanism for stablecoin liquidity,

    M. Egorov, “Stableswap-efficient mechanism for stablecoin liquidity,” Retrieved Feb, vol. 24, p. 2021, 2019

    2021

  16. [20]

    A non-custodial portfolio manager, liquidity provider, and price sensor,

    F. Martinelli and N. Mushegian, “A non-custodial portfolio manager, liquidity provider, and price sensor,”URl: https://balancer.finance/whitepaper, 2019

    2019

  17. [21]

    An analysis of uniswap markets,

    G. Angeris, H.-T. Kao, R. Chiang, C. Noyes, and T. Chitra, “An analysis of uniswap markets,” 2021

    2021

  18. [22]

    S. P. Boyd and L. Vandenberghe,Convex optimization. Cambridge university press, 2004

    2004

  19. [23]

    Curve - pool - 3pool,

    C. DAO, “Curve - pool - 3pool,” 2024. [Online]. Available: https://curve.fi/dex/ethereum/pools/3pool/swap/

    2024

  20. [24]

    Global order routing on exchange networks,

    V . Danos, H. E. Khalloufi, and J. Prat, “Global order routing on exchange networks,” inProc. FC Workshop. Springer, 2021, pp. 207–226

    2021

  21. [25]

    An algorithm for finding shortest routes from all source nodes to a given destination in general networks,

    J. Y . Yen, “An algorithm for finding shortest routes from all source nodes to a given destination in general networks,”Quarterly of applied mathematics, vol. 27, no. 4, pp. 526–530, 1970

    1970

  22. [27]

    Constant function market makers: Multi-asset trades via convex optimization,

    G. Angeris, A. Agrawal, A. Evans, T. Chitra, and S. Boyd, “Constant function market makers: Multi-asset trades via convex optimization,” in Handbook on Blockchain. Springer, 2022, pp. 415–444

    2022

  23. [30]

    Faster algorithms for the shortest path problem,

    R. K. Ahuja, K. Mehlhorn, J. Orlin, and R. E. Tarjan, “Faster algorithms for the shortest path problem,”Journal of the ACM (JACM), vol. 37, no. 2, pp. 213–223, 1990

    1990

  24. [32]

    Sushiswap,

    SushiSwap, “Sushiswap,” https://www.sushi.com, 2025, accessed on May 24, 2025

    2025

  25. [33]

    Pancakeswap,

    PancakeSwap, “Pancakeswap,” https://pancakeswap.finance, 2025, accessed on May 24, 2025

    2025

  26. [34]

    (2024) Fee-on-transfer & rebase tokens: An erc-20 security bug you need to know

    0xnolo. (2024) Fee-on-transfer & rebase tokens: An erc-20 security bug you need to know. Medium. [Online]. Available: https: //medium.com/@0xnolo/fee-on-transfer-rebase-tokens-an-erc-20-secur ity-bug-you-need-to-know-f4e5badea1ee

    2024

  27. [35]

    A line graph-based framework for identifying optimal routing paths in decentralized exchanges,

    Y . Zhang, Y . Li, and C. Tessone, “A line graph-based framework for identifying optimal routing paths in decentralized exchanges,”arXiv preprint arXiv:2504.15809, 2025

    2025 arXiv

  28. [36]

    W. H. Press,Numerical recipes 3rd edition: The art of scientific computing. Cambridge university press, 2007

    2007

  29. [37]

    Gurobi Optimization,Gurobi Optimizer Reference Manual, 2025

    L. Gurobi Optimization,Gurobi Optimizer Reference Manual, 2025. [Online]. Available: https://www.gurobi.com

    2025

  30. [38]

    A next-generation on-chain liquidity provider powered by pro-active market maker algorithm,

    D. Team, “A next-generation on-chain liquidity provider powered by pro-active market maker algorithm,” 2020

    2020

  31. [39]

    (2024) Fusion swap documentation

    1inch. (2024) Fusion swap documentation. Accessed: April 23, 2024. [Online]. Available: https://docs.1inch.io/docs/fusion-swap/introduction/

    2024

  32. [40]

    (2024) Classic swap documentation

    ——. (2024) Classic swap documentation. Accessed: April 23, 2024. [Online]. Available: https://portal.1inch.dev/documentation/apis/swap/c lassic-swap

    2024

  33. [41]

    Centralized exchanges vs. decentralized exchanges in cryp- tocurrency markets: A systematic literature review,

    S. H ¨agele, “Centralized exchanges vs. decentralized exchanges in cryp- tocurrency markets: A systematic literature review,”Electronic Markets, vol. 34, no. 1, p. 33, 2024

    2024

  34. [42]

    Pcsp: Efficiently answering label- constrained shortest path queries in road networks,

    L. Wang and R. C.-W. Wong, “Pcsp: Efficiently answering label- constrained shortest path queries in road networks,”Proceedings of the VLDB Endowment, vol. 17, no. 11, pp. 3082–3094, 2024

    2024

  35. [43]

    Efficient label- constrained shortest path queries on road networks: A tree decompo- sition approach,

    U. Zhang, L. Yuan, W. Li, L. Qin, and Y . Zhang, “Efficient label- constrained shortest path queries on road networks: A tree decompo- sition approach,”Proceedings of the VLDB Endowment, 2021

    2021

  36. [44]

    Approximate skyline index for constrained shortest pathfinding with theoretical guarantee,

    Z. Liu, L. Li, M. Zhang, W. Hua, and X. Zhou, “Approximate skyline index for constrained shortest pathfinding with theoretical guarantee,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 2024, pp. 4222–4235

    2024

  37. [45]

    An experimental evaluation and guideline for path finding in weighted dynamic network,

    M. Zhang, L. Li, and X. Zhou, “An experimental evaluation and guideline for path finding in weighted dynamic network,”Proceedings of the VLDB Endowment, vol. 14, no. 11, pp. 2127–2140, 2021

    2021

  38. [46]

    Linear programming and extensions,

    G. B. Dantzig, “Linear programming and extensions,” 2016

    2016

  39. [47]

    A strongly polynomial algorithm for generalized flow maximization,

    L. A. V ´egh, “A strongly polynomial algorithm for generalized flow maximization,” inProceedings of the forty-sixth annual ACM symposium on Theory of computing, 2014, pp. 644–653

    2014

  40. [48]

    Maximum network flows with concave gains,

    M. Shigeno, “Maximum network flows with concave gains,”Mathemat- ical programming, vol. 107, no. 3, pp. 439–459, 2006. APPENDIXA CONVERGENCERATEANALYSIS This appendix provides a detailed analysis of the conver- gence rate of the Adaptive Sign Gradient Method (ASGM) for the Tok...

    2006

  41. [49]

    sweet spot

    Main Proof: Linear Convergence Rate:Using the Armijo condition (Lemma 4), the improvement in the objective value is: J(W t+1)−J(W t) =J(W t + ∆t ·SIGN)−J(W t) ≥α∆ t · ∇J(Wt)⊤SIGN. (12) Substituting the lower bound for∆ t from Lemma 6: J(W t+1)−J(W t)≥α·min ∆0, 2(1−α) L · ∇J(W ...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.