Pith. sign in

REVIEW 3 major objections 5 minor 32 references

A Quantum of Learning: Using Quaternion Algebra to Model Learning on Quantum Devices

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read An adaptive quaternion-valued filter can learn a mapping between data and hypothesis distributions encoded on qubits.

desk verdict The quaternion adaptive-learning setup is a fresh combination, but Eq. (7) gives P=4 for measuring |0> against |0> and P=0 for |+x>, so the learning loop is not connected to quantum measurements as written. read the letter →

arxiv 2504.13232 v2 pith:27LLKHDM submitted 2025-04-17 quant-ph cs.LGmath.QAstat.ML

Sayed Pouria Talebi , Clive Cheong Took , Danilo P. Mandic This is my paper · ORCID
classification quant-phcs.LGmath.QAstat.ML MSC 81P6815A33 PACS 03.67.-a
keywords quaternionalgebraquantummachinelearningqubitrepresentationHR-calculusmeasurementadaptivefilteringgradientdescentBlochsphere
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 tries to establish that quaternion algebra is the right language for training quantum learning machines, replacing the complex-valued formulation of quantum computation. It models every single-qubit gate as a quaternion involution and every multi-qubit computation as a linear unitary map in an augmented quaternion space. On that footing it derives an adaptive algorithm, using quaternion gradients, that learns a mapping from a data distribution encoded in multi-qubit states to a hypothesis distribution encoded in measurement outcomes. If the model holds, a quantum circuit can be trained with the same adaptive-filtering toolkit used for classical quaternion-valued signals, complete with convergence and observability conditions.

What carries the argument

The load-bearing object is the augmented quaternion vector $q_a=[q,\,q^{\imath},\,q^{\jmath},\,q^{\kappa}]^{\mathrm{T}}$ and the identity that rotations of a quaternion become linear maps in this augmented space. Proposition 1 turns quantum gates into quaternion involutions, so multi-qubit computations become linear unitary maps in that space. The measurement model $P(q=h)=\tfrac{1}{4}\|\Re\{h_a^{\mathrm{T}}q_a\}\|^2$ converts those linear maps into probability distributions, and the HR-calculus gradient provides the update rule. Together these pieces make a quantum circuit look like a quaternion-valued linear system that classical adaptive-filtering theory can train.

What would settle it

Evaluate Eq. (7) for a single qubit: measuring the state $|0\rangle$ in the $|+x\rangle$ basis gives probability $1/2$ by quantum mechanics, while Eq. (7) gives $0$; measuring $|0\rangle$ against itself gives $4$ instead of $1$. If this mismatch persists in the paper's own model, the update rule's probabilities do not describe physical measurements.

Watch

Extended reading notes

Core claim

The central claim is that an adaptive quaternion-valued matrix $W$, updated by HR-calculus gradient descent on a measurement-based cost, learns the mapping $f: P(D_n\mid\zeta)\to P(\zeta\mid D_n)$ when data and hypothesis distributions are encoded in multi-qubit systems. Proposition 1 asserts that every quantum gate is a quaternion involution and every multi-qubit computation is a linear unitary transformation in augmented quaternion space, so the optimization problem becomes a quaternion-valued linear estimation problem. The paper further claims that the mapping is recoverable when expected measurement outcomes are available together with $3\lceil\log_2(z)\rceil$ linearly independent measurement vectors, and that, with an encoding satisfying Assumption 2, step sizes $\mu\in[0,(3\lceil\log_2(z)\rceil)^{-1})$ give convergent training. The numerical example on an 8-qubit system shows the cost decreasing and the estimated circuit reproducing the likelihood of held-out hypotheses.

Load-bearing premise

The training procedure stands on the measurement probability formula in Eq. (7), which is stated without derivation and whose probabilities must match quantum measurement outcomes for the learned circuit to mean anything on real hardware.

Editorial extensions

If this is right

  • A quantum learning machine can be trained by quaternion adaptive filtering, inheriting convergence and observability criteria from that theory.
  • The mapping between data and hypothesis distributions is learnable from expected measurement outcomes plus $3\lceil\log_2(z)\rceil$ linearly independent measurement vectors.
  • Training converges for step sizes $\mu\in[0,(3\lceil\log_2(z)\rceil)^{-1})$ when the encoding keeps $\Re\{h_a^{\mathrm{T}*} y_a^n\}>0$ for all measurement vectors and samples.
  • The trained circuit estimates $P(\zeta'\mid D_n)$ for held-out hypotheses, as demonstrated on an 8-qubit numerical example.

Reading between the lines

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

  • If the measurement formula in Eq. (7) were replaced by a map that matches the Born rule exactly, the same augmented-quaternion machinery would connect the algorithm to measurements on real quantum hardware; the paper's own simulation uses Eq. (7) rather than the Born rule.
  • Requiring the learned matrix $W$ to be unitary on the augmented space would turn the proposed algorithm into a training rule for parameterized quantum circuits, a constraint the paper asserts but never enforces.
  • A natural testable extension is to rerun the same update rule with measurement probabilities computed by the standard Born rule and compare the convergence behavior, since the cost function would then reflect physical measurements rather than Eq. (7).
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a quaternion-valued model for qubits, quantum gates, and measurement, then uses the augmented quaternion representation and HR-calculus to formulate an adaptive learning problem on multi-qubit systems. The central claim is that a quaternion matrix W, trained by gradient descent, can learn a mapping between probability distributions encoded in multi-qubit states, with a guaranteed convergence step-size range. The learning objective is built on a quaternion measurement probability formula in Eq. (7), and the convergence and observability conditions are imported from the authors' prior work on quaternion distributed filtering. A numerical example with an 8-qubit system is presented to demonstrate learning.

Significance. If the central measurement model were correct, the paper would contribute a new algebraic framework for quantum machine learning, connecting quaternion adaptive filtering to quantum systems and potentially enabling new training algorithms. The use of HR-calculus for quaternion gradients is a legitimate and potentially useful technique. However, the core measurement probability formula is not consistent with the Born rule on elementary states, and the learned transformation is not constrained to be unitary. Because the cost function, gradient, and convergence analysis all rest on this formula, the main claims are not supported. The numerical validation is generated within the same quaternion model, so it does not provide independent evidence.

major comments (3)
  1. [Section III, Eq. (7)] The measurement probability formula P(q=h) = (1/4)||Re(h_a^T q_a)||^2 contradicts the Born rule for simple states. For q = h = κ (the |0> state), the augmented vectors are q_a = h_a = [κ, -κ, -κ, κ]^T, so h_a^T q_a = -4 and Eq. (7) gives P = 4, whereas the Born rule gives P = 1. For h = ı (|+x>) and q = κ, h_a^T q_a is purely imaginary, yielding P = 0, whereas the Born rule gives 1/2. The factor 1/4 and the real-part extraction do not fix the normalization, since the quaternions involved are already unit-norm and purely imaginary. This formula is asserted without derivation, and because the cost function in (8), the gradient, and the convergence analysis in Section IV all use it, the learning algorithm is disconnected from actual quantum measurement statistics.
  2. [Section IV, Remark 4 and Proposition 1] Proposition 1 requires multi-qubit computations to be represented by linear unitary maps in the augmented quaternion space, and Remark 4 claims that multiplication by the learned matrix W models unitary rotations and synthesizable quantum operations. However, W is learned as an unconstrained quaternion matrix via gradient descent; no step in the algorithm imposes unitarity or any structural constraint derived from quantum circuits. Consequently, the learned W need not correspond to any valid quantum operation, and the claimed connection to quantum computation is not established. The numerical example does not test whether the learned W is unitary or physically realizable.
  3. [Section IV, Theorem 1 and convergence claim] The observability condition in Theorem 1 and the step-size bound µ ∈ [0, (3⌈log2 z⌉)^(-1)) are imported from the authors' prior work on quaternion distributed filtering ([32], and also [27]), rather than proved within this quantum setting. Since the learning problem is built on the measurement model of Eq. (7), which is not a valid quantum measurement rule, transferring these conditions is not justified. Additionally, the numerical example generates both the data and the target from the same quaternion model, so it does not validate the approach against an independent quantum measurement model or a standard simulation of the Born rule.
minor comments (5)
  1. [Throughout] There are several typographical errors that should be corrected: 'Theses' should be 'These' in the introduction, 'repressuring' should be 'representing' in Remark 3, 'effect' should be 'affect' in Remark 2, and 'Internatational' should be 'International' in reference [6].
  2. [Section III, Remark 3] The phrase 'An elegant solution for repressuring the state' appears to be a typo for 'representing'. The remark would be clearer if the proposed mapping from α and β to quaternion components were written explicitly with equations.
  3. [Section III, Eq. (7) and surrounding text] The notation P(q = h) for measurement outcomes is ambiguous, since quantum measurement outcomes are labeled by basis states, not by the full pre-measurement state. The authors should clarify the intended meaning and relate it to standard projection-valued measures.
  4. [Section IV, gradient expression] The displayed gradient expression contains a fraction: ∂d̂_s,n/∂W* = (1/8) Re{h_a^T W x_a} h_a x_a^H. The factor 1/8 is not derived; it likely depends on the specific measurement formula in (7). Since (7) is in question, the gradient would need to be re-derived if the measurement model is repaired.
  5. [Section V, numerical example] The numerical example reports convergence and a likelihood estimate, but it does not state the learning rate µ used, the number of iterations, or the initialization of W. These details are needed to reproduce the experiment. More importantly, the example never checks whether the learned W is unitary or whether the measurement statistics satisfy the Born rule.

Circularity Check

2 steps flagged · score 6.0 of 10

Numerical 'prediction' and the convergence guarantee reduce to the paper's own quaternion model: the 'correct likelihood' and the learned estimate are both computed from Eq. (7), and convergence/observability are imported from the authors' prior quaternion filtering papers.

  1. self definitional [Eq. (7), Section IV cost definitions (8)-(9), Section V numerical example]
    "P(q=h)=1/4||Re{h_a^T q_a}||^2, P(q≠h)=1/4||Im{h_a^T q_a}||^2 (7). This, allows quantum computations to be interpreted as functions on probability distributions. ... data points were generated from P(Dn|ζ′) where ζ′ were a selection of 8 hypothesis from the region representing yn in the previous experiment. The estimated circuit was used to find P(ζ′|Dn). The likelihood function of ζ′ is shown against its estimate in Fig. 4. Note that the estimated circuit closely estimated the correct likelihood for ζ′."

    The 'correct likelihood' used for evaluation is generated by the same Eq. (7) that defines the quaternion measurement model; the desired outputs ds,n = E{Mhs(ya_n)} in the cost are also defined through (7). Since Remark 4 leaves W unconstrained, the learned map is just a linear map in the augmented quaternion space, the same class used to synthesize xn from yn. The experiment therefore verifies that the model can fit its own encoding/measurement definition; it is not validated against Born-rule measurement statistics or any external quantum benchmark. The agreement in Fig. 4 is built into the problem formulation: the target distribution and the estimated distribution are both computed with the same Eq. (7).

  2. self citation load bearing [Section IV, Theorem 1 proof and step-size bound after the gradient derivation]
    "From the work in [32] it follows that the class of problem is (8) have a solution if the state is measurable along 3⌈log2 (z)⌉-independent vectors. ... where from classical results on quaternion-valued adaptive filtering it follows that µ ∈ [0, (3⌈log2 (z)⌉)−1) will yield convergent behaviour."

    The convergence and observability guarantees are not proved in this paper; they are imported verbatim from [32] and [27], both by the same authors (Talebi, Werner, Mandic; Talebi), after asserting that (8) is 'equivalent to the distributed state estimation problem in [32]'. That equivalence rests on the paper's own quaternion measurement model (7), whose validity is not independently established. The step-size bound is likewise taken from 'classical results on quaternion-valued adaptive filtering' (the first author's PhD thesis), making the central convergence claim an inheritance from the authors' prior framework rather than a theorem about quantum learning proved here.

full rationale

The paper's internal algebra—quaternion state encoding, the HR-calculus gradient, and the gradient-descent recursion—is self-consistent given its premises. However, the load-bearing measurement formula (7) is introduced by assertion and then used both to define the training targets (ds,n) and to generate the 'correct' likelihoods in the numerical example. Thus the numerical validation is self-referential: the estimated circuit is fitted and evaluated with the same quaternion measurement rule, so agreement is a check of the model against itself rather than against quantum measurement data. In addition, the observability condition in Theorem 1 and the claimed convergence bound are lifted from the authors' earlier quaternion filtering work ([32], [27]) by asserting equivalence of the formulated problem to that prior setting; no independent derivation or external quantum benchmark is provided. These two features make the core empirical and theoretical claims partially circular, though the gradient computation and problem formulation retain independent mathematical content. Score 6 reflects one prediction that reduces by construction plus load-bearing self-citation for the convergence guarantee.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The framework rests on four mathematical or modeling postulates: the quaternion gate model, the nonstandard measurement formula, the positivity encoding, and the imported observability and convergence results.

free parameters (1)
  • Adaptation gain mu (step size) = not specified; range mu in [0, (3*ceil(log2 z))^(-1))
    Controls the gradient update in Section IV; the convergence range is asserted from [32], and no specific value is given in the numerical example.
assumptions (4)
  • domain assumption Quantum gates act as quaternion involutions, and multi-qubit computations are linear unitary maps in the augmented quaternion space (Proposition 1).
    True for single-qubit rotations via the quaternion double cover, but the extension to all multi-qubit circuits as an unconstrained augmented-quaternion matrix W is asserted, not derived; the learning algorithm optimizes W without enforcing unitarity.
  • ad hoc to paper Measurement probability is P(q=h) = (1/4)||Re(h_a^T q_a)||^2 and P(q neq h) = (1/4)||Im(h_a^T q_a)||^2 (Eq. 7).
    This formula is the foundation of the cost function and gradient. It is not derived from the Born rule and gives wrong values in simple cases: for |0> vs |+x>, Born probability is 1/2 but (7) gives 0; for |0> vs |0> it gives 4 rather than 1.
  • domain assumption Assumption 2: the target encoding satisfies Re(h_a^T* y_a^n) > 0 for all n and s.
    Introduced to make qubit states uniquely recoverable from measurements, described as an activation-like encoding choice; it restricts the class of learnable distributions.
  • ad hoc to paper The observability and convergence conditions from quaternion distributed filtering [32] transfer to the quantum learning problem.
    Theorem 1 equates problem (8) with the distributed state estimation problem of [32] and imports the 3*ceil(log2 z) condition; the equivalence is asserted, and [32] is authored by the same group.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Quantum of Learning: Using Quaternion Algebra to Model Learning on Quantum Devices." pith.science (2026). https://pith.science/paper/27LLKHDM

@misc{pith2026250413232,
  author       = {Pith},
  title        = {Pith review of: A Quantum of Learning: Using Quaternion Algebra to Model Learning on Quantum Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27LLKHDM}},
  note         = {Machine review of arXiv:2504.13232}
}
read the original abstract

This article considers the problem of designing adaption and optimisation techniques for training quantum learning machines. To this end, the division algebra of quaternions is used to derive an effective model for representing computation and measurement operations on qubits. In turn, the derived model, serves as the foundation for formulating an adaptive learning problem on principal quantum learning units, thereby establishing quantum information processing units akin to that of neurons in classical approaches. Then, leveraging the modern HR-calculus, a comprehensive training framework for learning on quantum machines is developed. The quaternion-valued model accommodates mathematical tractability and establishment of performance criteria, such as convergence conditions.

Figures

Figures reproduced from arXiv: 2504.13232 by the authors.

Figure 1
Figure 1. Schematic showing the right-hand rotation of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Bloch sphere representation of qubit |ψ⟩ modelled as the quaternion q = ı sin(θ) cos(ϕ) + ȷ sin(θ) sin(ϕ) + κ cos(θ) [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The cost, 10 log  1 8 P8 s=1 d  dˆs,n, ds,n, is shown for 103 independent trials of the derived algorithm. The cost across all trials lie within the light red region, while the solid red line showing the average over all trials. In order to demonstrate operations of the estimated circuit data points were generated from P (Dn|ζ ′ ) where ζ ′ were a selection of 8 hypothesis from the region representing yn in the … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Likelihoods of ζ ′ and its estimate. VI. CONCLUSION A quaternion-valued formulation of the adaptive learning problem on quantum computational devices was considered, where a quaternion-valued model for implementing computa￾tion and measurement operation on qubits was p…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [32]

    Quaternion-valued dis- tributed filtering and control,

    S. P. Talebi, S. Werner, and D. P. Mandic, “Quaternion-valued dis- tributed filtering and control,” IEEE Transactions on Automatic Control, vol. 65, no. 10, pp. 4246–4257, 2020

    work page 2020

  2. [27]

    Adaptive filtering algorithms for quaternion-valued sig- nals,

    S. P. Talebi, “Adaptive filtering algorithms for quaternion-valued sig- nals,” Ph.D. dissertation, Imperial College London, 2016

    work page 2016

  3. [1]

    A logical calculus of the ideas immanent in nervous activity,

    W. S. McCulloch and W. Pitts, “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics , no. 2, pp. 115–133, December 1943

    work page 1943

  4. [2]

    Adaptive switching circuits,

    B. Widrow and M. Hoff, “Adaptive switching circuits,” IRE WESCON Convention Record, pp. 96–104, August 1960

    work page 1960

  5. [3]

    Reliable, trainable networks for comput- ing and control,

    B. Widrow and J. B. Angell, “Reliable, trainable networks for comput- ing and control,” Aerospace Engineering , vol. 21, no. 9, pp. 78–123, September 1962

    work page 1962

  6. [4]

    D. P. Mandic and J. A. Chambers, Recurrent Neural Networks for Prediction: Learning Algorithms, Architectures and Stability . John Wiley & Sons, 2001

    work page 2001

  7. [5]

    From the quantum Moore’s law toward silicon based universal quantum computing,

    E. Prati, D. Rotta, F. Sebastiano, and E. Charbon, “From the quantum Moore’s law toward silicon based universal quantum computing,” IEEE International Conference on Rebooting Computing , pp. 1–4, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:24877662

    work page 2017

  8. [6]

    Simulating physics with computers,

    R. Feynman, “Simulating physics with computers,” Internatational Journal of Theoretical Physics , vol. 21, no. 6/7, 1982

    work page 1982

Show all 32 references
  1. [7]

    A novel approach to nonlinear/non-Gaussian Bayesian state estimation,

    N. J. Gordon, D. J. Salmond, and A. F. M. Smith, “A novel approach to nonlinear/non-Gaussian Bayesian state estimation,” In IEE Proceedings on Radar and Signal Processing , vol. 140, pp. 107–113, 1993

    1993

  2. [8]

    On sequential Monte Carlo sampling methods for Bayesian filtering,

    A. Doucet, S. Godsill, and C. Andrieu, “On sequential Monte Carlo sampling methods for Bayesian filtering,” Statistics and Computing , vol. 10, no. 3, pp. 197–208, 2000

    2000

  3. [9]

    Quantum signal processing,

    Y . C. Eldar and A. V . Oppenheim, “Quantum signal processing,” IEEE Signal Processing Magazine , vol. 19, no. 6, pp. 12–32, 2002

    2002

  4. [10]

    An introduction to quantum filtering,

    L. Bouten and M. R. J. R. Van Handel, “An introduction to quantum filtering,” SIAM Journal on Control and Optimization , vol. 46, no. 6, pp. 2199–2241, 2007

    2007

  5. [11]

    Quantum computing through the lens of control: A tutorial introduction,

    J. Berberich and D. Fink, “Quantum computing through the lens of control: A tutorial introduction,” IEEE Control Systems , vol. 44, no. 6, pp. 24–49, 2024

    2024

  6. [12]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, September 2017

    2017

  7. [13]

    On quantum methods for machine learning problems part I: Quantum tools,

    F. Ablayev, M. Ablayev, J. Z. Huang, K. Khadiev, N. Salikhova, and D. Wu, “On quantum methods for machine learning problems part I: Quantum tools,” Big Data Mining and Analytics , vol. 3, no. 1, pp. 41– 55, 2020

    2020

  8. [14]

    On quantum methods for machine learning problems part II: Quantum classification algorithms,

    ——, “On quantum methods for machine learning problems part II: Quantum classification algorithms,” Big Data Mining and Analytics , vol. 3, no. 1, pp. 56–67, 2020

    2020

  9. [15]

    Training deep quantum neural net- works,

    K. Beer, D. Bondarenko, T. Farrelly, T. J. Osborne, R. Salzmann, D. Scheiermann, and R. Wolf, “Training deep quantum neural net- works,” Nature Communications, vol. 11, no. 1, p. 808, 2020

    2020

  10. [16]

    M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, 10th ed. Cambridge University Press, 2011

    2011

  11. [17]

    Quantum computing: Lecture notes,

    R. de Wolf, “Quantum computing: Lecture notes,” arXiv:1907.09415, 2023

    1907 arXiv

  12. [18]

    A universal training algorithm for quantum deep learning,

    G. Verdon, J. Pye, and M. A. Broughton, “A universal training algorithm for quantum deep learning,” arXiv:1806.09729v1, 2018

    2018 arXiv

  13. [19]

    A framework for approx- imating qubit unitaries,

    V . Kliuchnikov, A. Bocharov, M. Roetteler, J. Yard, A. Bocharov, J. Yard, M. Roetteler, and V . Kliuchnikov, “A framework for approx- imating qubit unitaries,” Tech. Rep., October 2015, arXiv preprint arXiv:1510.03888

    2015 arXiv

  14. [20]

    A framework for exact synthesis,

    V . Kliuchnikov and J. T. Yard, “A framework for exact synthesis,” Tech. Rep., April 2015, arXiv preprint arXiv:1504.04350

    2015 arXiv

  15. [21]

    The HR-calculus: Enabling information processing with quaternion algebra,

    D. P. Mandic, S. P. Talebi, C. C. Took, Y . Xia, D. Xu, M. Xiang, and P. Bourigault, “The HR-calculus: Enabling information processing with quaternion algebra,” arXiv:2311.16771, 2024

    2024 arXiv

  16. [22]

    Geometric alge- bra quantum convolutional neural network: A model using geometric (Clifford) algebras and quantum computing,

    G. Altamirano-Escobedo and E. Bayro-Corrochano, “Geometric alge- bra quantum convolutional neural network: A model using geometric (Clifford) algebras and quantum computing,” IEEE Signal Processing Magazine, vol. 41, no. 2, pp. 75–85, 2024

    2024

  17. [23]

    Quaternion-based arithmetic in quantum information processing: A promising approach for efficient color quantum imaging,

    A. M. Grigoryan and S. S. Agaian, “Quaternion-based arithmetic in quantum information processing: A promising approach for efficient color quantum imaging,” IEEE Signal Processing Magazine , vol. 41, no. 2, pp. 64–74, 2024

    2024

  18. [24]

    Quaternion involutions and anti- involutions,

    T. A. Ell and S. J. Sangwine, “Quaternion involutions and anti- involutions,” Computers & Mathematics with Applications , vol. 53, no. 1, pp. 137–143, 2007

    2007

  19. [25]

    A quaternion gradient operator and its applications,

    D. P. Mandic, C. Jahanchahi, and C. C. Took, “A quaternion gradient operator and its applications,” IEEE Signal Processing Letters , vol. 18, no. 1, pp. 47–50, 2011

    2011

  20. [26]

    Augmented statistics of quaternion random variables: A lynchpin of quaternion learning machines,

    C. C. Took, S. P. T. ans R. M. Fernandez Alcala, and D. P. Mandic, “Augmented statistics of quaternion random variables: A lynchpin of quaternion learning machines,” IEEE Signal Processing Magazine , vol. 41, no. 3, pp. 72–87, 2024

    2024

  21. [28]

    R. A. Silverman, Modern Calculus and Analytical Geometry . New York: Macmillan, 1969

    1969

  22. [29]

    A quaternion frequency estimator for three-phase power systems,

    S. P. Talebi and D. P. Mandic., “A quaternion frequency estimator for three-phase power systems,” In Procceding of IEEE International Conference on Acoustics, Speech and Signal Processing , pp. 3956– 3960, 2015

    2015

  23. [30]

    Generalized variational inference,

    J. Knoblauch, J. Jewson, and T. Damoulas, “Generalized variational inference,” arXiv:1904.02063, 2019

    1904 arXiv

  24. [31]

    Probability distribution function map- ping method,

    D. H. Hoitsma and S. J. Engel, “Probability distribution function map- ping method,” US Patent, US20110010140A1 , January 2011, current Assignee; Northrop Grumman Systems Corp

    2011

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.