Pith. sign in

REVIEW 2 major objections 4 minor 43 references

Path Portfolio Optimization: Defect, Lift, and the Price of Path Complexity

T0 review · 2 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read The paper establishes that excess growth is the signature lift gap, and that the sample-size floor for path portfolios belongs to unstructured estimation, not to path complexity.

desk verdict The excess-growth/lift-gap identity is clean, new, and worth citing; the sample-floor inversion is honest but conditional on correct specification, and the paper's own caveats mostly cover it. read the letter →

arxiv 2608.02355 v1 pith:YGBAMVO5 submitted 2026-08-03 q-fin.PM

classification q-fin.PM MSC 91G1060L1060H10
keywords pathsignaturesroughpathsshufflealgebraexpectedsignatureportfoliooptimizationstochastictheoryexcessgrowthrateestimationrisk
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 develops portfolio theory on a path-first basis: the price path's signature is the universal coordinate, and a portfolio is a linear functional of the signature. It claims that the covariance of signature coordinates—the defect form—is exactly risk, and that the gap between two lift conventions, contracted with portfolio weights, is exactly the excess growth rate of stochastic portfolio theory. It further claims that the antisymmetric, directional part of the level-two signature is unchanged by the lift choice, while variance signals and ruin depend on it. The empirical core is a dimensional trade-off: with the expected signature known, quadratic path functionals multiply certainty equivalent by up to about 60×, but estimating it freely requires roughly six observations per parameter, whereas estimating only the generator and rebuilding the expected signature recovers about 95% of oracle value at barely one observation per parameter. The paper concludes that the floor is a property of unstructured estimation, not of path complexity.

What carries the argument

The central object is the signature S(X) of the log-price path, with words indexing coordinates. The defect form D(u,v)=⟨u⊔⊔v,E[S]⟩−⟨u,E[S]⟩⟨v,E[S]⟩ is exactly the covariance of signature coordinates; because the shuffle product is adjoint to the coproduct, the whole risk form is the non-group-like part of the expected signature. The lift theorem states that the forward and geometric lifts of level-two signatures differ by half the quadratic covariation; contracted with diag(π)−ππᵀ this becomes excess growth. The estimation analysis uses the exact loss identity CE*−CE(ℓ)=γ/2‖D^{1/2}(ℓ−ℓ*)‖², which turns plug-in error into the spectrum of the inverse sample covariance, and a model-consistent

What would settle it

Run the estimation comparison on data generated by a misspecified driver—heavy-tailed increments, stochastic volatility, or a self-exciting process—at M/p≈1.19. If the model-consistent estimator's certainty equivalent falls well below about 0.95 of oracle or fails to improve as M grows, the conclusion that the floor is a property of unstructured estimation collapses. A second check: verify on a path with expected area whether the antisymmetric level-two block adds nonzero value; Proposition 22's zero-area attribution predicts it should, and a zero result would falsify the word-class decomposit

Watch

Extended reading notes

Core claim

At the paper's center is the identification of stochastic portfolio theory's excess growth rate with the geometricity defect of the signature. For fixed portfolio weights, the difference between logarithmic wealth and the level-one signature flow equals the contraction of the level-two lift gap with the tensor diag(π)−ππᵀ, which is exactly the excess growth rate integrated over time. The classical rebalancing premium is therefore not an artifact but the failure of the mean path to be group-like—the same defect as the covariance of signature coordinates. The paper also proves that the antisymmetric level-two block is lift-invariant pathwise, so directional signals are convention-free, while v

Load-bearing premise

The claim that the sample-size floor belongs to unstructured estimation rather than path complexity depends on the Gaussian driver being correctly specified for the model-consistent estimator; under misspecification its bias does not vanish with sample size, so the 0.95× oracle recovery at M/p=1.19 is not guaranteed.

Editorial extensions

If this is right

  • If excess growth is the lift gap, then choosing an execution convention is choosing which side of the rebalancing premium one can access; a geometric-lift theory has already spent the premium.
  • Variance and covariance swaps enter the investable universe only under the forward lift, so the lift choice is a decision about whether the volatility surface is tradable.
  • The antisymmetric block's lift-invariance means lead–lag estimates should not change when the integration convention changes; observed sensitivity would point to estimation or data, not to the integration convention.
  • The sample-size floor for signature-linear portfolios is an estimator property: naming a driver class and rebuilding the expected signature can recover nearly all attainable value at M/p≈1.19, so path complexity is not the binding constraint.
  • The word-class decomposition implies that, on drivers with zero expected area, the level-two premium is convexity in the terminal increment, not path-dependence; path-dependent area earns nothing unless the driver carries expected area.

Reading between the lines

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

  • Editorial inference: If excess growth is a lift gap, then any strategy systematically harvesting rebalancing premium is implicitly long the non-geometric component of the path; this may connect the variance risk premium to the defect form.
  • Editorial inference: The model-consistent estimator's success suggests a practical rule—fit the generator, not the full covariance—that should transfer to other path-dependent payoff classes; this is testable on non-Gaussian drivers.
  • Editorial inference: The entropy-production ceiling on one-step directional edges implies a general bound linking tradable lead–lag alpha to time-reversal asymmetry; a market with high measured self-excitation but no sign-carrying direction should show no cross-area P&L.
  • Editorial inference: Since the antisymmetric block is lift-invariant, empirical discrepancies in lead–lag estimates across execution conventions would indicate non-signature effects, not integration choices.
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

2 major / 4 minor

Summary. The paper develops a signature-based ('path-first') framework for mean-variance portfolio optimization. The expected signature supplies both the mean vector and, through the shuffle-product defect form, the covariance matrix of signature-linear payoffs; the optimizer is then the solution of a single linear system. The main structural results are: (i) the Marcus/forward lift gap at level two equals half the quadratic covariation, so contracting it with diag(pi)-pi pi' gives Fernholz's excess growth rate exactly; (ii) the antisymmetric level-two block is lift-invariant while the symmetric block and solvency are lift-dependent; (iii) a numerical study of estimation error shows a raw plug-in floor around six observations per parameter, a sign reversal of shrinkage between d=2 and d=20, and a model-consistent estimator that recovers most of the oracle value; (iv) Hawkes closures and cross-area lead-lag effects are analyzed with explicit caveats. The paper carefully tiers proved, conditional/measured, and conjectural claims.

Significance. If the central identification is correct, the paper gives a genuinely new reading of stochastic portfolio theory: excess growth is the geometricity defect of the portfolio map, i.e. the Marcus-forward lift gap contracted with a fixed portfolio tensor. The Hilbert-space existence theorem, the exact quadratic loss decomposition, and the careful claim-tiering are also strengths. The empirical work is unusually transparent: both calibrations are fully specified, dispersions are reported, and Section 11 explicitly lists non-claims, including the correct-specification caveat on the model-consistent estimator. These features make the core theoretical contribution credible and the paper worth publishing after revision.

major comments (2)
  1. [Section 9, Proposition 33] The claim that the forward lift grants N_2-L_2=d(d+1)/2 'additional directions ... spanned by the quadratic-covariation payoffs — the variance and covariance swaps' is not supported by the displayed algebra. A forward-lift symmetric level-two linear combination gives S_iS_j-[X_i,X_j]_T, not [X_i,X_j]_T alone; isolating the bracket would require access to S_iS_j, which is not a linear functional of the signature. Moreover, under a geometric lift the signature-linear payoff space is already N_m-dimensional (shuffle identities are nonlinear), so the geometric lift's free directions are not cut to L_m in the control space. The practical conclusion that the forward lift 'puts the volatility surface in the investable universe' at a cost of d(d+1)/2 additional parameters is therefore not established and should be corrected or carefully qualified.
  2. [Abstract, Section 1 contribution 3, Section 6.1 / Proposition 23] The headline inversion 'the sample-size floor belongs to unstructured estimation rather than to path complexity' is established only under a correctly specified Gaussian driver. Proposition 23 estimates only (b, Sigma) and rebuilds E[S]=exp(T(b+1/2 Sigma)); under misspecification the estimator's bias does not vanish with M and the reported 0.95x recovery at M/p=1.19 is not guaranteed. The caveats in Section 6.1 and Section 11 are clear, but the abstract and contribution list state the inversion without the correct-specification qualifier. Since this is a main empirical contribution, the abstract should say 'within the model class' or 'under correct specification'.
minor comments (4)
  1. [Abstract] Typo: 'This paper buildspath portfolio optimization' should be 'builds path portfolio optimization'.
  2. [Section 9] After correcting the major point, please align notation: use N_m for the number of control coordinates and L_m for the intrinsic dimension of the geometric signature, and avoid saying that shuffle relations 'cut the free directions' of the control space.
  3. [Corollary 13] The discrete identity has a minus sign (pi'u - log(...)= -gamma* + ...), while the text says the one-period return 'exceeds' the coordinate by gamma*. Both are consistent, but a one-sentence sign clarification would help readers.
  4. [Section 3.1] The statement that the antisymmetric component of the optimal control is 'exactly zero' is based on a numerical calibration; in view of the careful claim-tiering elsewhere, consider writing 'numerically zero to machine precision for this calibration'.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: Corollary 13 is proved from Itô's formula and Proposition 3 from the shuffle identity; the author's prior [39] supplies vocabulary and a closure that is independently checked, not load-bearing.

full rationale

The paper's core derivation chain is self-contained. Proposition 3 is exactly the pathwise shuffle identity plus expectation: 'The shuffle identity ⟨u,S⟩⟨v,S⟩=⟨u⊔⊔v,S⟩ holds pathwise; take expectations and subtract the product of means.' That makes the defect form a covariance by direct computation, not by importing the conclusion. Corollary 13 is proved in-text from Itô's formula: 'd log V_t = π⊤dX_t + γπ* dt', with the level-two lift gap SM−SF = 1/2[X,X]_T contracted against diag(π)−ππ⊤; the paper explicitly says 'The continuous identity is classical [13,14]; the placement is the point', so the excess-growth identification is a reinterpretation of a proved identity rather than a self-referential derivation. Theorem 7 is standard Hilbert-space spectral linear algebra; Proposition 9 follows algebraically from S_ij+S_ji=S_iS_j. The estimation claims are carefully tiered: Propositions 20–23 are labeled 'Conditional, measured', and Proposition 23's success is explicitly scoped in Section 6.1 — 'The caveat is exactly as large as the claim... under misspecification its bias does not vanish with M' — and again in Section 11 Non-claims. Thus the sample-size-floor inversion is not presented as a theorem derived from the framework, and its acknowledged conditioning on a correctly specified Gaussian driver is a stated limitation, not a hidden circular input. The author's prior [39] is used for the 'geometricity defect' vocabulary, the Hawkes closure in Section 7, and a conjectural extension (Remark 34 says 'nothing in this paper depends on it'). The Hawkes closure is independently checked against simulation in Table 4, and the central theoretical results do not rely on [39]. I find no equation in which an output equals a fitted input, no load-bearing self-citation chain, and no uniqueness theorem imported from the authors' prior work to force the conclusions. Score 2 reflects the self-citational framing, not any circular reduction.

Assumptions & free parameters 9 free parameters · 8 assumptions · 0 invented entities

The paper's central structural claims rest on standard signature-algebra facts plus explicit modeling assumptions (Gaussian drivers, geometric lift, zero expected area, correct specification). All numerical parameters are invented and internally consistent; no new entities are introduced.

free parameters (9)
  • d=2 drift vector b = (0.06, 0.02)
    Invented to set the calibration; affects all oracle CE multiples and sample-size thresholds (Section 3.1, Section 6).
  • d=2 covariance Sigma = [[0.040, 0.012], [0.012, 0.090]]
    Invented covariance matrix for the pair; sets the defect form's eigenvalues and the 11.15x oracle gain.
  • d=20 drift vector b = equally spaced on [0.04, 0.10]
    Invented for the cross-section calibration; contributes to the 59.6x oracle gain.
  • d=20 covariance Sigma = volatilities on [0.22,0.35], equicorrelation 0.40
    Invented for the cross-section; determines condition number 1906 and effective dimension 7325.
  • risk aversion gamma = 3
    Chosen for all optimization experiments; CE values and shrinkage comparisons scale with gamma.
  • shrinkage intensity delta (fixed) = 0.25
    One of two ridge intensities; its flat behavior is used to claim fixed shrinkage is inconsistent.
  • Marchenko-Pastur ridge intensity delta = sqrt(p/M)
    Prescribed by random-matrix analogy, not tuned; used for the converging ridge policy.
  • Hawkes parameters = mu=0.5, beta=1, T=20, branching ratios 0.3-0.9
    Invented for the stationary-substitution experiment (Table 4).
  • Cross-area simulation parameters = mu=(0.4,0.4), beta=(1,1), T=20, alpha_off=0.6, tick=0.01, q in {0,0.85,1}
    Invented for the Hawkes lead-lag experiments (Tables 5).
assumptions (8)
  • standard math Signature universality: signature-linear functionals are dense in continuous functionals on compacta of unparameterized paths (Section 2).
    Inherited from Chen/Lyons signature theory; used to justify the strategy class.
  • standard math Shuffle identity pathwise: <u,S><v,S> = <u shuffle v,S>.
    Used in Proposition 3 to make the defect form a covariance.
  • domain assumption Geometric (Marcus) lift is group-like; forward lift breaks the shuffle identity by 1/2 [X,X] (Theorem 12).
    Execution convention is a modeling choice; the paper's 'lift is the execution convention' thesis depends on this characterization.
  • domain assumption Gaussian driver X_t = bt + sigma W_t, so E[S] = exp(T(b + 1/2 Sigma)).
    Used for the oracle calibrations and for the model-consistent estimator (Sections 3.1, 6).
  • domain assumption Correct model specification for Proposition 23.
    The 0.95x oracle recovery assumes the fitted generator family contains the true data-generating process; the paper states this caveat.
  • domain assumption Driver with zero expected area for Proposition 22.
    The claim that the antisymmetric block earns nothing is conditional on time-reversible/no-expected-area drivers; Section 11 states this.
  • domain assumption Stationary Markov chain and one-step measurability for Proposition 31.
    The entropy-production ceiling is conditional on these premises and is false without them.
  • domain assumption Marchenko-Pastur analogy: sample covariance spectrum concentrates on the MP bulk.
    Used to motivate delta = sqrt(p/M); paper flags it as 'a located analogy, not a theorem about D'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Path Portfolio Optimization: Defect, Lift, and the Price of Path Complexity." pith.science (2026). https://pith.science/paper/YGBAMVO5

@misc{pith2026260802355,
  author       = {Pith},
  title        = {Pith review of: Path Portfolio Optimization: Defect, Lift, and the Price of Path Complexity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YGBAMVO5}},
  note         = {Machine review of arXiv:2608.02355}
}
read the original abstract

This paper builds Path Portfolio Optimization: portfolio theory on a path-first framework in which the signature is the universal coordinate of the price path, and asks whether it survives estimation. A portfolio is a linear functional of the signature, so the control lives in a truncated tensor algebra, the covariance of signature coordinates is the non-group-like part of the expected signature --- a defect form --- and the whole mean--variance problem becomes a linear system in one tensor. Two structural results follow. The lift is the execution convention: the gap between the Marcus and forward lifts, contracted with portfolio weights, is Fernholz's excess growth rate exactly, so excess growth is the geometricity defect of the portfolio map. And the antisymmetric block at level two is lift-invariant pathwise, so directional signals are convention-free while variance signals and ruin are not. The empirical finding is a dimensional trade-off. With the expected signature known, quadratic path functionals raise the certainty equivalent elevenfold for a pair of assets and sixtyfold for a cross section of twenty; with it estimated, the unregularized policy is severely negative until the sample exceeds roughly six observations per parameter, and shrinkage flips from harmful in the pair to indispensable in the cross section. The entire gain sits in the symmetric block, which is convexity in the terminal increment rather than path-dependence; the path-dependent antisymmetric block earns nothing when the driver has no expected area. And the sample-size floor belongs to unstructured estimation rather than to path complexity: an estimator that fits only the generator of the driver and rebuilds the expected signature recovers almost all of the attainable value at barely one observation per parameter

Figures

Figures reproduced from arXiv: 2608.02355 by the authors.

Figure 1
Figure 1. Median realized certainty equivalent as a fraction of the oracle, against sample size in [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Oracle certainty equivalent of each level- [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Table 4 drawn: the first two moments of [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 7 linked inside Pith

  1. [1]

    Hawkes processes in finance.Market Microstructure and Liquidity, 1(1), 2015

    Emmanuel Bacry, Iacopo Mastromatteo, and Jean-Fran¸ cois Muzy. Hawkes processes in finance.Market Microstructure and Liquidity, 1(1), 2015

  2. [2]

    Cleaning large correlation matrices: Tools from random matrix theory.Physics Reports, 666:1–109, 2017

    Jo¨ el Bun, Jean-Philippe Bouchaud, and Marc Potters. Cleaning large correlation matrices: Tools from random matrix theory.Physics Reports, 666:1–109, 2017

  3. [3]

    Integration of paths, geometric invariants and a generalized Baker–Hausdorff formula.Annals of Mathematics, 65(1):163–178, 1957

    Kuo-Tsai Chen. Integration of paths, geometric invariants and a generalized Baker–Hausdorff formula.Annals of Mathematics, 65(1):163–178, 1957. 20

  4. [4]

    Characteristic functions of measures on geometric rough paths.Annals of Probability, 44(6):4049–4082, 2016

    Ilya Chevyrev and Terry Lyons. Characteristic functions of measures on geometric rough paths.Annals of Probability, 44(6):4049–4082, 2016

  5. [5]

    Thomas M. Cover. Universal portfolios.Mathematical Finance, 1(1):1–29, 1991

  6. [6]

    Cover and Joy A

    Thomas M. Cover and Joy A. Thomas.Elements of Information Theory. Wiley, 2nd edition, 2006

  7. [7]

    Signature methods in stochastic portfolio theory

    Christa Cuchiero and Janka M¨ oller. Signature methods in stochastic portfolio theory. arXiv:2310.02322, 2023

  8. [8]

    Signature-based models: Theory and calibration.SIAM Journal on Financial Mathematics, 14(3):910–957, 2023

    Christa Cuchiero, Guido Gazzani, and Sara Svaluto-Ferro. Signature-based models: Theory and calibration.SIAM Journal on Financial Mathematics, 14(3):910–957, 2023

Show all 43 references
  1. [9]

    A dynamic contagion process.Advances in Applied Probability, 43(3):814–846, 2011

    Angelos Dassios and Hongbiao Zhao. A dynamic contagion process.Advances in Applied Probability, 43(3):814–846, 2011

  2. [10]

    Optimal versus naive diversification: How inefficient is the 1 /n portfolio strategy?Review of Financial Studies, 22(5):1915–1953, 2009

    Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal. Optimal versus naive diversification: How inefficient is the 1 /n portfolio strategy?Review of Financial Studies, 22(5):1915–1953, 2009

  3. [11]

    Engl, Martin Hanke, and Andreas Neubauer.Regularization of Inverse Problems

    Heinz W. Engl, Martin Hanke, and Andreas Neubauer.Regularization of Inverse Problems. Kluwer, 1996

  4. [12]

    Goldberg

    Eymen Errais, Kay Giesecke, and Lisa R. Goldberg. Affine point processes and portfolio credit risk.SIAM Journal on Financial Mathematics, 1(1):642–665, 2010

  5. [13]

    Robert Fernholz.Stochastic Portfolio Theory

    E. Robert Fernholz.Stochastic Portfolio Theory. Springer, 2002

  6. [14]

    Robert Fernholz and Ioannis Karatzas

    E. Robert Fernholz and Ioannis Karatzas. Stochastic portfolio theory: An overview. In Handbook of Numerical Analysis: Mathematical Modeling and Numerical Methods in Finance, pages 89–167. Elsevier, 2009

  7. [15]

    Signature trading: A path-dependent extension of the mean–variance framework with exogenous signals.Quantitative Finance, 25 (2):197–226, 2025

    Owen Futter, Blanka Horvath, and Magnus Wiese. Signature trading: A path-dependent extension of the mean–variance framework with exogenous signals.Quantitative Finance, 25 (2):197–226, 2025. arXiv:2308.15135

  8. [16]

    Kernel learning for mean–variance trading strategies.arXiv:2507.10701, 2025

    Owen Futter, Nicola Muca Cirone, and Blanka Horvath. Kernel learning for mean–variance trading strategies.arXiv:2507.10701, 2025

  9. [17]

    Portfolio selection with parameter and model uncertainty: A multi-prior approach.Review of Financial Studies, 20(1):41–81, 2007

    Lorenzo Garlappi, Raman Uppal, and Tan Wang. Portfolio selection with parameter and model uncertainty: A multi-prior approach.Review of Financial Studies, 20(1):41–81, 2007

  10. [18]

    Extracting information from the signature of a financial data stream.arXiv:1307.7244, 2013

    Lajos Gergely Gyurk´ o, Terry Lyons, Mark Kontkowski, and Jonathan Field. Extracting information from the signature of a financial data stream.arXiv:1307.7244, 2013

  11. [19]

    Hardiman, Nicolas Bercot, and Jean-Philippe Bouchaud

    Stephen J. Hardiman, Nicolas Bercot, and Jean-Philippe Bouchaud. Critical reflexivity in financial markets: A Hawkes process analysis.European Physical Journal B, 86:442, 2013

  12. [20]

    Tibshirani

    Trevor Hastie, Andrea Montanari, Saharon Rosset, and Ryan J. Tibshirani. Surprises in high-dimensional ridgeless least squares interpolation.Annals of Statistics, 50(2):949–986, 2022. 21

  13. [21]

    Alan G. Hawkes. Spectra of some self-exciting and mutually exciting point processes. Biometrika, 58(1):83–90, 1971

  14. [22]

    Michael E. Hoffman. Quasi-shuffle products.Journal of Algebraic Combinatorics, 11(1): 49–68, 2000

  15. [23]

    Optimal execution with rough path signatures.SIAM Journal on Financial Mathematics, 11(2):470–493, 2020

    Jasdeep Kalsi, Terry Lyons, and Imanol Perez Arribas. Optimal execution with rough path signatures.SIAM Journal on Financial Mathematics, 11(2):470–493, 2020

  16. [24]

    Optimal portfolio choice with parameter uncertainty

    Raymond Kan and Guofu Zhou. Optimal portfolio choice with parameter uncertainty. Journal of Financial and Quantitative Analysis, 42(3):621–656, 2007

  17. [25]

    Kelly, Semyon Malamud, and Kangying Zhou

    Bryan T. Kelly, Semyon Malamud, and Kangying Zhou. The virtue of complexity in return prediction.Journal of Finance, 79(1):459–503, 2024

  18. [26]

    Kir´ aly and Harald Oberhauser

    Franz J. Kir´ aly and Harald Oberhauser. Kernels for sequentially ordered data.Journal of Machine Learning Research, 20(31):1–45, 2019

  19. [27]

    A well-conditioned estimator for large-dimensional covari- ance matrices.Journal of Multivariate Analysis, 88(2):365–411, 2004

    Olivier Ledoit and Michael Wolf. A well-conditioned estimator for large-dimensional covari- ance matrices.Journal of Multivariate Analysis, 88(2):365–411, 2004

  20. [28]

    Analytical nonlinear shrinkage of large-dimensional covariance matrices.Annals of Statistics, 48(5):3043–3065, 2020

    Olivier Ledoit and Michael Wolf. Analytical nonlinear shrinkage of large-dimensional covariance matrices.Annals of Statistics, 48(5):3043–3065, 2020

  21. [29]

    Learning from the past, predicting the statistics for the future, learning an evolving system.arXiv:1309.0260, 2013

    Daniel Levin, Terry Lyons, and Hao Ni. Learning from the past, predicting the statistics for the future, learning an evolving system.arXiv:1309.0260, 2013

  22. [30]

    Terry Lyons and Andrew D. McLeod. Signature methods in machine learning. arXiv:2206.14674, 2022

  23. [31]

    Expected signature of Brownian motion up to the first exit time from a bounded domain.Annals of Probability, 43(5):2729–2762, 2015

    Terry Lyons and Hao Ni. Expected signature of Brownian motion up to the first exit time from a bounded domain.Annals of Probability, 43(5):2729–2762, 2015

  24. [32]

    Nonparametric pricing and hedging of exotic derivatives.Applied Mathematical Finance, 27(6):457–494, 2020

    Terry Lyons, Sina Nejad, and Imanol Perez Arribas. Nonparametric pricing and hedging of exotic derivatives.Applied Mathematical Finance, 27(6):457–494, 2020

  25. [33]

    Lyons and Zhongmin Qian.System Control and Rough Paths

    Terry J. Lyons and Zhongmin Qian.System Control and Rough Paths. Oxford University Press, 2002

  26. [34]

    Steven I. Marcus. Modeling and approximation of stochastic differential equations driven by semimartingales.Stochastics, 4(3):223–245, 1981

  27. [35]

    Portfolio selection.Journal of Finance, 7(1):77–91, 1952

    Harry Markowitz. Portfolio selection.Journal of Finance, 7(1):77–91, 1952

  28. [36]

    Marˇ cenko and Leonid A

    Vladimir A. Marˇ cenko and Leonid A. Pastur. Distribution of eigenvalues for some sets of random matrices.Mathematics of the USSR-Sbornik, 1(4):457–483, 1967

  29. [37]

    Richard O. Michaud. The Markowitz optimization enigma: Is optimized optimal?Financial Analysts Journal, 45(1):31–42, 1989

  30. [38]

    Kernel mean embedding of distributions: A review and beyond.Foundations and Trends in Machine Learning, 10(1–2):1–141, 2017

    Krikamol Muandet, Kenji Fukumizu, Bharath Sriperumbudur, and Bernhard Sch¨ olkopf. Kernel mean embedding of distributions: A review and beyond.Foundations and Trends in Machine Learning, 10(1–2):1–141, 2017. 22

  31. [39]

    A general theory of paths: Signatures, jump lifts, and expected signatures of self-exciting processes.arXiv:2606.28869 [math.PR], 2026

    Miquel Noguer i Alonso. A general theory of paths: Signatures, jump lifts, and expected signatures of self-exciting processes.arXiv:2606.28869 [math.PR], 2026

  32. [40]

    Free Lie algebras.London Mathematical Society Monographs, 7, 1993

    Christophe Reutenauer. Free Lie algebras.London Mathematical Society Monographs, 7, 1993

  33. [41]

    The signature kernel is the solution of a Goursat PDE.SIAM Journal on Mathematics of Data Science, 3 (3):873–899, 2021

    Cristopher Salvi, Thomas Cass, James Foster, Terry Lyons, and Weixin Yang. The signature kernel is the solution of a Goursat PDE.SIAM Journal on Mathematics of Data Science, 3 (3):873–899, 2021

  34. [42]

    Joel A. Tropp. User-friendly tail bounds for sums of random matrices.Foundations of Computational Mathematics, 12(4):389–434, 2012

  35. [43]

    Tropp.An Introduction to Matrix Concentration Inequalities, volume 8 ofFoundations and Trends in Machine Learning

    Joel A. Tropp.An Introduction to Matrix Concentration Inequalities, volume 8 ofFoundations and Trends in Machine Learning. Now Publishers, 2015. 23

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.