Yoothana Suansook
Yoothana Suansook has a total of 1 co-authors. Here are the top co-authors by number of shared publications:
Kitti Paithoonwattanakij
Faculty of Engineering, King Mongkut’s Institute of Technology Ladkrabang, Chalongkrung Road, Ladkrabang, Bangkok 10520, Thailand
1 shared publication
Latest Publications
- Open Access
- 21 October 2026
- Published by MDPI AG in Preprints.org
The Judge Under Trial: Stress-Testing RART and the Limits of Automated Evaluation at Scale
The deployment of large language models (LLMs) as automated evaluators commonly termed LLM-as-a-Judge introduces a critical question: who evaluates the evaluator? This paper puts the lifecycle framework of Kong et al. [1] on trial, stress-testing its core claims for building, aligning, and monitoring LLM judges in production-scale recommendation explanation systems. We synthesize the core contributions: (1) the identification of a subtle failure mode termed ”right answer, wrong reason,” wherein a judge reaches the correct verdict through unsound reasoning that is invisible to label-agreement dashboards; (2) Reasoning-Aligned Rubric Tuning (RART), a procedure that iteratively refines rubric text using a Meta-Judge to compare judge rationales against human rationales; (3) a fourphase lifecycle-Birth, Training, Deployment, and Monitoring-that anchors LLM judges to human judgment at every stage; and (4) a formal mathematical model stress-tested via Monte Carlo simulation. We provide an open-source implementation of this simulation and analyze its visual output. The simulation confirms RART’s ability to stabilize reasoning alignment but visually reveals methodological vulnerabilities, including a stopping-rule artifact, a severe drift-band calibration problem, and the structural reality of blind-spot leaks. Furthermore, the framework relies on critical unvalidated assumptions-most notably, that the judge and generator do not share blind spots. We conclude that while RART effectively mitigates rubric degradation, it is not a closed-form mathematical solution to Al alignment and ultimately requires continuous human-in-the-loop audits.
- 21 October 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Judge Under Trial: Stress-Testing RART and the Limits of Automated Evaluation at Scale
The deployment of large language models (LLMs) as automated evaluators commonly termed LLM-as-a-Judge introduces a critical question: who evaluates the evaluator? This paper puts the lifecycle framework of Kong et al. [1] on trial, stress-testing its core claims for building, aligning, and monitoring LLM judges in production-scale recommendation explanation systems. We synthesize the core contributions: (1) the identification of a subtle failure mode termed ”right answer, wrong reason,” wherein a judge reaches the correct verdict through unsound reasoning that is invisible to label-agreement dashboards; (2) Reasoning-Aligned Rubric Tuning (RART), a procedure that iteratively refines rubric text using a Meta-Judge to compare judge rationales against human rationales; (3) a fourphase lifecycle-Birth, Training, Deployment, and Monitoring-that anchors LLM judges to human judgment at every stage; and (4) a formal mathematical model stress-tested via Monte Carlo simulation. We provide an open-source implementation of this simulation and analyze its visual output. The simulation confirms RART’s ability to stabilize reasoning alignment but visually reveals methodological vulnerabilities, including a stopping-rule artifact, a severe drift-band calibration problem, and the structural reality of blind-spot leaks. Furthermore, the framework relies on critical unvalidated assumptions-most notably, that the judge and generator do not share blind spots. We conclude that while RART effectively mitigates rubric degradation, it is not a closed-form mathematical solution to Al alignment and ultimately requires continuous human-in-the-loop audits.
- Open Access
- 20 October 2026
- Published by MDPI AG in Preprints.org
The Judge Under Trial: Monte Carlo Verification of the Reasoning-Aligned Rubric Tuning Lifecycle
The deployment of Large Language Models (LLMs) as automated evaluators—commonly termed LLM-as-a-Judge—introduces a critical vulnerability: the “right answer, wrong reason” failure mode, where a judge reaches the correct verdict through unsound reasoning that is invisible to standard label-agreement dashboards. Kong et al. [1] propose Reasoning-Aligned Rubric Tuning (RART) to address this, anchoring the judge’s rationale to human judgment via a Meta-Judge and iterative rubric refinement. This paper presents a reproducible Monte Carlo simulation (300 replicates, 40 iterations) that empirically verifies the RART lifecycle against a label-only baseline (Vanilla). We formally model the latent rubric state θ_R = (q_R, b_R), the noisy meta-judge detection layer, the biased reflector update, and the statistical drift monitor. Our results confirm the framework’s central thesis. While Vanilla achieves marginally higher final label accuracy (q_R = 0.956) than RART (q_R = 0.875), its reasoning alignment collapses entirely (b_R = −0.842), whereas RART stabilizes reasoning (b_R = +0.910). Furthermore, the simulation visually exposes the structural vulnerabilities of the framework: Vanilla consumes 2.6× the computational budget (39.4 vs 14.8 iterations) due to the stopping-rule artifact; the standard 2σ drift band produces an unacceptable 11.3% false-alarm rate at K = 3 raters; and the blind-spot leak between shared judge-generator architectures decays only when human oversight (p_H > 0) is present. The simulation serves as both a pedagogical tool and an empirical validation that the LLM judge is an amplifier of human judgment, not a replacement for it.
- 20 October 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Judge Under Trial: Monte Carlo Verification of the Reasoning-Aligned Rubric Tuning Lifecycle
The deployment of Large Language Models (LLMs) as automated evaluators—commonly termed LLM-as-a-Judge—introduces a critical vulnerability: the “right answer, wrong reason” failure mode, where a judge reaches the correct verdict through unsound reasoning that is invisible to standard label-agreement dashboards. Kong et al. [1] propose Reasoning-Aligned Rubric Tuning (RART) to address this, anchoring the judge’s rationale to human judgment via a Meta-Judge and iterative rubric refinement. This paper presents a reproducible Monte Carlo simulation (300 replicates, 40 iterations) that empirically verifies the RART lifecycle against a label-only baseline (Vanilla). We formally model the latent rubric state θ_R = (q_R, b_R), the noisy meta-judge detection layer, the biased reflector update, and the statistical drift monitor. Our results confirm the framework’s central thesis. While Vanilla achieves marginally higher final label accuracy (q_R = 0.956) than RART (q_R = 0.875), its reasoning alignment collapses entirely (b_R = −0.842), whereas RART stabilizes reasoning (b_R = +0.910). Furthermore, the simulation visually exposes the structural vulnerabilities of the framework: Vanilla consumes 2.6× the computational budget (39.4 vs 14.8 iterations) due to the stopping-rule artifact; the standard 2σ drift band produces an unacceptable 11.3% false-alarm rate at K = 3 raters; and the blind-spot leak between shared judge-generator architectures decays only when human oversight (p_H > 0) is present. The simulation serves as both a pedagogical tool and an empirical validation that the LLM judge is an amplifier of human judgment, not a replacement for it.
- Open Access
- 9 October 2026
- Published by MDPI AG in Preprints.org
The Mathematics of Deception: A Unified Framework for AI-Driven Honeypot Systems
Honeypots occupy a peculiar place in the security landscape: they are among the few mechanisms whose value is defined by their failure. A firewall that is never tested, an intrusion detection system that never alarms—these are successes. A honeypot that is never attacked is useless. This inversion of the usual security logic gives honeypots an unusually clean signal, but it has also kept their theoretical foundations fragmented. Game theorists model deception in isolation; information theorists treat intrusion detection generically; queuing theorists model traffic separately; and the recent wave of AI-driven honeypots—powered by Large Language Models and Reinforcement Learning—has proceeded largely without formal mathematical treatment.This paper attempts a synthesis. We develop a unified mathematical framework for AI-driven honeypot systems, building on six pillars: stochastic processes (the honeypot as a continuous time Markov chain), game theory (the honeypot as a Bayesian signaling device), detection theory (the honeypot as a near-ideal hypothesis test), information theory (the honeypot as a channel for intelligence), optimization (the honeypot as a constrained POMDP), and AI integration (LLM as observation kernel, RL as policy optimization).Along the way, we derive equilibrium conditions, detection thresholds, optimal deployment fractions, information-theoretic bounds, and LLM deception limits. We prove that the mixed strategy equilibrium is unique and Lyapunov stable, solve the POMDP for a reduced state space to obtain a threshold policy, and show that in the myopic case this threshold coincides exactly with the game-theoretic equilibrium under consistent payoff specifications. We generalize the Poisson detection assumption to a Weibull hazard model and derive a bound on the mutual information achievable under LLM-generated observations via the data processing inequality.We accompany the analytical development with numerical Monte Carlo experiments based on exact Gillespie simulation of the CTMC, replicator dynamics for the game-theoretic component, and parametric sampling for the detection and information-theoretic components. These experiments validate the closed-form predictions and expose the qualitative behavior of the system in transient and steady state. All quantitative claims are analytical; the simulations serve as numerical illustrations, not as empirical validation of real-world deployments.
- 9 October 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Mathematics of Deception: A Unified Framework for AI-Driven Honeypot Systems
Honeypots occupy a peculiar place in the security landscape: they are among the few mechanisms whose value is defined by their failure. A firewall that is never tested, an intrusion detection system that never alarms—these are successes. A honeypot that is never attacked is useless. This inversion of the usual security logic gives honeypots an unusually clean signal, but it has also kept their theoretical foundations fragmented. Game theorists model deception in isolation; information theorists treat intrusion detection generically; queuing theorists model traffic separately; and the recent wave of AI-driven honeypots—powered by Large Language Models and Reinforcement Learning—has proceeded largely without formal mathematical treatment.This paper attempts a synthesis. We develop a unified mathematical framework for AI-driven honeypot systems, building on six pillars: stochastic processes (the honeypot as a continuous time Markov chain), game theory (the honeypot as a Bayesian signaling device), detection theory (the honeypot as a near-ideal hypothesis test), information theory (the honeypot as a channel for intelligence), optimization (the honeypot as a constrained POMDP), and AI integration (LLM as observation kernel, RL as policy optimization).Along the way, we derive equilibrium conditions, detection thresholds, optimal deployment fractions, information-theoretic bounds, and LLM deception limits. We prove that the mixed strategy equilibrium is unique and Lyapunov stable, solve the POMDP for a reduced state space to obtain a threshold policy, and show that in the myopic case this threshold coincides exactly with the game-theoretic equilibrium under consistent payoff specifications. We generalize the Poisson detection assumption to a Weibull hazard model and derive a bound on the mutual information achievable under LLM-generated observations via the data processing inequality.We accompany the analytical development with numerical Monte Carlo experiments based on exact Gillespie simulation of the CTMC, replicator dynamics for the game-theoretic component, and parametric sampling for the detection and information-theoretic components. These experiments validate the closed-form predictions and expose the qualitative behavior of the system in transient and steady state. All quantitative claims are analytical; the simulations serve as numerical illustrations, not as empirical validation of real-world deployments.
- Open Access
- 9 October 2026
- Published by MDPI AG in Preprints.org
Fractional Calculus in Backpropagation Learning: Theory, Algorithms, and Conditional Convergence Analysis
We develop a mathematical framework for integrating fractional calculus into the backpropagation learning algorithm. We present two complementary approaches: (i) fractional-order gradient descent (FOGD), which replaces the integer-order derivative in the optimizer with a Caputo or Grünwald–Letnikov fractional derivative, inducing power-law memory in weight updates; and (ii) fractional-order neural networks (FONN), which employ fractional derivatives within the network’s activation and chain rule. We prove the equivalence of the Grünwald–Letnikov and Riemann–Liouville definitions for α ∈ (0, 1) under explicit regularity conditions, and cite the finite-part interpretation for α ≥ 1 as a classical result rather than a new proof. We establish the Mittag-Leffler solution to fractional gradient flow and derive its power-law asymptotics via complete monotonicity and Hankel contour integration. We state a corrected fractional Taylor proposition valid only for functions that are not classically C1 at the expansion point, and we explicitly note the well-known counterexample for standard smooth functions. We prove a conditional convergence result for stochastic fractional gradient descent on non-convex objectives at rate O(T^(−1/2)) + O(T^(−α)), where the memory-control condition is stated as Assumption 7.4 and the O(T^(−α)) bias term is derived explicitly from that assumption under an additional regularity hypothesis (Assumption 7.5). We give an explicit failure example showing that Assumption 7.4 can fail for small α and large truncation K, and we exhibit an admissible fractional activation (the Mittag-Leffler activation) for which the fractional chain rule of Section 6.3 is valid. Numerical verification confirms the theoretical predictions where the assumptions hold, and is reported with the regime of validity made explicit. The framework specializes to SGD, momentum, and Nesterov acceleration as the case α = 1, connects heuristically to anomalous diffusion on loss landscapes, and yields an emergent power-law learning-rate schedule.
- 9 October 2026
- Published by MDPI AG in Preprints.org
Open AccessFractional Calculus in Backpropagation Learning: Theory, Algorithms, and Conditional Convergence Analysis
We develop a mathematical framework for integrating fractional calculus into the backpropagation learning algorithm. We present two complementary approaches: (i) fractional-order gradient descent (FOGD), which replaces the integer-order derivative in the optimizer with a Caputo or Grünwald–Letnikov fractional derivative, inducing power-law memory in weight updates; and (ii) fractional-order neural networks (FONN), which employ fractional derivatives within the network’s activation and chain rule. We prove the equivalence of the Grünwald–Letnikov and Riemann–Liouville definitions for α ∈ (0, 1) under explicit regularity conditions, and cite the finite-part interpretation for α ≥ 1 as a classical result rather than a new proof. We establish the Mittag-Leffler solution to fractional gradient flow and derive its power-law asymptotics via complete monotonicity and Hankel contour integration. We state a corrected fractional Taylor proposition valid only for functions that are not classically C1 at the expansion point, and we explicitly note the well-known counterexample for standard smooth functions. We prove a conditional convergence result for stochastic fractional gradient descent on non-convex objectives at rate O(T^(−1/2)) + O(T^(−α)), where the memory-control condition is stated as Assumption 7.4 and the O(T^(−α)) bias term is derived explicitly from that assumption under an additional regularity hypothesis (Assumption 7.5). We give an explicit failure example showing that Assumption 7.4 can fail for small α and large truncation K, and we exhibit an admissible fractional activation (the Mittag-Leffler activation) for which the fractional chain rule of Section 6.3 is valid. Numerical verification confirms the theoretical predictions where the assumptions hold, and is reported with the regime of validity made explicit. The framework specializes to SGD, momentum, and Nesterov acceleration as the case α = 1, connects heuristically to anomalous diffusion on loss landscapes, and yields an emergent power-law learning-rate schedule.
- Open Access
- 28 September 2026
- Published by MDPI AG in Preprints.org
From Spikes to Attention: The Mathematical Odyssey of Neural Networks
This paper presents a comprehensive history of neural networks, tracing the journey from the earliest philosophical speculations about intelligence to modern large language models and diffusion models. It covers biological foundations, the McCulloch-Pitts neuron, Hebb’s rule, the Perceptron, the AI Winter, backpropagation, convolutional neural networks, recurrent networks, the Transformer revolution, and contemporary developments in generative AI, multimodal systems, and AI safety. This revised edition corrects factual errors, fills significant omissions, and expands the coverage of modern developments including scaling laws,RLHF, AlphaFold, vision transformers, and state-space models. Throughout, the mathematical theories underlying each development are explicitly presented, demonstrating that the history of neural networks is inseparable from the evolution of linear algebra, calculus, probability, and optimization.
- 28 September 2026
- Published by MDPI AG in Preprints.org
Open AccessFrom Spikes to Attention: The Mathematical Odyssey of Neural Networks
This paper presents a comprehensive history of neural networks, tracing the journey from the earliest philosophical speculations about intelligence to modern large language models and diffusion models. It covers biological foundations, the McCulloch-Pitts neuron, Hebb’s rule, the Perceptron, the AI Winter, backpropagation, convolutional neural networks, recurrent networks, the Transformer revolution, and contemporary developments in generative AI, multimodal systems, and AI safety. This revised edition corrects factual errors, fills significant omissions, and expands the coverage of modern developments including scaling laws,RLHF, AlphaFold, vision transformers, and state-space models. Throughout, the mathematical theories underlying each development are explicitly presented, demonstrating that the history of neural networks is inseparable from the evolution of linear algebra, calculus, probability, and optimization.
- Open Access
- 28 September 2026
- Published by MDPI AG in Preprints.org
The Gauge Dagger: A Gauge-Theoretic Reconstruction of Quantum Mechanics and Attention
We present a gauge-theoretic reconstruction of the structural identity between quantum mechanics and the attention mechanism of the Transformer. The original observation—that the conjugate transpose, or dagger, is the defining operation of both theories—is here derived from a single principle: local U(dk) gauge invariance. We formulate attention as a lattice gauge theory on the sequence bundle E = M × Cdk , with the attention weights as the gauge connection, the score matrix as the parallel transport phase, and the attention output as the covariant derivative. The softmax is identified as a gauge-fixing term that breaks the U(1) phase symmetry, and the phase-preserving complex softmax is identified as the gauge-invariant normalization. The density matrix and the Born rule are recovered as the gauge-invariant measurement postulate. We derive the Yang–Mills analogue of attention, introduce the attention curvature Fijk, write down the full gauge-theoretic Lagrangian, and derive the equations of motion. Standard attention is shown to be the real, flat limit of the gauge theory. Complex attention is shown to be the gauge-covariant extension. We make falsifiable predictions, identify open problems, and argue that the dagger is not merely an analogy between quantum mechanics and deep learning but the generator of local gauge transformations in both. The theory is rigorous, derivable from first principles, and reduces to the original formulation in the appropriate limit.
- 28 September 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Gauge Dagger: A Gauge-Theoretic Reconstruction of Quantum Mechanics and Attention
We present a gauge-theoretic reconstruction of the structural identity between quantum mechanics and the attention mechanism of the Transformer. The original observation—that the conjugate transpose, or dagger, is the defining operation of both theories—is here derived from a single principle: local U(dk) gauge invariance. We formulate attention as a lattice gauge theory on the sequence bundle E = M × Cdk , with the attention weights as the gauge connection, the score matrix as the parallel transport phase, and the attention output as the covariant derivative. The softmax is identified as a gauge-fixing term that breaks the U(1) phase symmetry, and the phase-preserving complex softmax is identified as the gauge-invariant normalization. The density matrix and the Born rule are recovered as the gauge-invariant measurement postulate. We derive the Yang–Mills analogue of attention, introduce the attention curvature Fijk, write down the full gauge-theoretic Lagrangian, and derive the equations of motion. Standard attention is shown to be the real, flat limit of the gauge theory. Complex attention is shown to be the gauge-covariant extension. We make falsifiable predictions, identify open problems, and argue that the dagger is not merely an analogy between quantum mechanics and deep learning but the generator of local gauge transformations in both. The theory is rigorous, derivable from first principles, and reduces to the original formulation in the appropriate limit.
- Open Access
- 24 September 2026
- Published by MDPI AG in Preprints.org
The Dagger in the Garden (Revised Edition) A Narrative Scientific Report on Complex-Valued Attention, Phase Interference, and the Learning of Ambiguity in a Single Poem
This paper reports a complete end-to-end experiment in complex-valued language modeling. We take the mathematical theory developed in The Dagger and the Query: A Unified Mathematical Theory of Complex, Unitary, and Quantum Attention (Suansook, 2026) and instantiate it as a working system: a character-level ComplexGPT trained on a single 595-character poem, The Rose Is Not a Rose. The poem was chosen because it exhibits simultaneous lexical, syntactic, semantic, and pragmatic ambiguity --- precisely the structure that the imaginary part of a complex attention score is designed to represent. We make three contributions. First, we prove the mathematical and implementation correctness of the complex attention machinery: all six verification tests pass, including the interference identity \( O_i = 2w\cos\theta \cdot v\ \) for symmetric phases, and the closed-form parameter count \( P_{\mathrm{non\text{-}emb}} = 4Vd + L(4d + 24d^2) + 2d \) matches the observed count of 53,824 for \( V=34,\ d=32,\ L=2 \). Second, we demonstrate trainability: training loss falls from 3.55 to 1.44, and validation loss falls from 3.56 to a minimum of 2.54 before rising mildly --- a signature of overfitting on a small corpus, not of architectural failure. Third, we produce recognizable English with the poem's vocabulary (rose, wind, garden, word, the, and), though not cohesive poetry. We are explicit about what this paper does not show. It does not prove that complex attention outperforms real-valued attention. No real-valued control baseline was trained. It does not measure whether the imaginary part of the attention score is actually used — no diagnostic was instrumented, so we cannot say whether the model has collapsed to real-valued attention or is genuinely exploiting phase. It does not establish that phase carries linguistic meaning. These are empirical questions that remain open. We position the paper as a software and mathematical proof-of-concept, not a machine learning hypothesis paper. Its contribution is correctness and trainability, not benchmark superiority. We close by specifying the experiments that would elevate it to the latter: (1) a real-valued control baseline, (2) an instrumented phase diagnostic, (3) a larger and more diverse corpus, and (4) a comparison of validation loss and sample quality across the two architectures. Until those are run, the paper's central hypothesis — that phase-preserving attention provides an advantage on phase-sensitive tasks — remains unproven.- 24 September 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Dagger in the Garden (Revised Edition) A Narrative Scientific Report on Complex-Valued Attention, Phase Interference, and the Learning of Ambiguity in a Single Poem
This paper reports a complete end-to-end experiment in complex-valued language modeling. We take the mathematical theory developed in The Dagger and the Query: A Unified Mathematical Theory of Complex, Unitary, and Quantum Attention (Suansook, 2026) and instantiate it as a working system: a character-level ComplexGPT trained on a single 595-character poem, The Rose Is Not a Rose. The poem was chosen because it exhibits simultaneous lexical, syntactic, semantic, and pragmatic ambiguity --- precisely the structure that the imaginary part of a complex attention score is designed to represent. We make three contributions. First, we prove the mathematical and implementation correctness of the complex attention machinery: all six verification tests pass, including the interference identity \( O_i = 2w\cos\theta \cdot v\ \) for symmetric phases, and the closed-form parameter count \( P_{\mathrm{non\text{-}emb}} = 4Vd + L(4d + 24d^2) + 2d \) matches the observed count of 53,824 for \( V=34,\ d=32,\ L=2 \). Second, we demonstrate trainability: training loss falls from 3.55 to 1.44, and validation loss falls from 3.56 to a minimum of 2.54 before rising mildly --- a signature of overfitting on a small corpus, not of architectural failure. Third, we produce recognizable English with the poem's vocabulary (rose, wind, garden, word, the, and), though not cohesive poetry. We are explicit about what this paper does not show. It does not prove that complex attention outperforms real-valued attention. No real-valued control baseline was trained. It does not measure whether the imaginary part of the attention score is actually used — no diagnostic was instrumented, so we cannot say whether the model has collapsed to real-valued attention or is genuinely exploiting phase. It does not establish that phase carries linguistic meaning. These are empirical questions that remain open. We position the paper as a software and mathematical proof-of-concept, not a machine learning hypothesis paper. Its contribution is correctness and trainability, not benchmark superiority. We close by specifying the experiments that would elevate it to the latter: (1) a real-valued control baseline, (2) an instrumented phase diagnostic, (3) a larger and more diverse corpus, and (4) a comparison of validation loss and sample quality across the two architectures. Until those are run, the paper's central hypothesis — that phase-preserving attention provides an advantage on phase-sensitive tasks — remains unproven. - Open Access
- 24 September 2026
- Published by MDPI AG in Preprints.org
The Quantum Dagger: How Quantum Mechanics and Attention Share a Mathematical Core
This paper tells the story of a mathematical object — the dagger — and its quiet, persistent presence in two of the most consequential theories of the twentieth and twenty-first centuries. The dagger, or conjugate transpose, is the defining operation of quantum mechanics: it turns a state into a dual, an operator into an observable, and a symmetry into a conserved quantity. In 2017, when Vaswani and colleagues introduced the Transformer [1], they wrote down an equation whose core operation —the scaled dot-product — contained the dagger in disguise, hidden behind the innocent transpose of a real-valued matrix. This paper is a narrative exploration of that coincidence, and an argument that it is not a coincidence at all. We trace the dagger from Born and Jordan’s matrix mechanics in 1926 [11] to the attention mechanism of modern artificial intelligence; we watch it split into real and imaginary parts, we follow its phase as it survives or fails to survive the softmax bottleneck, and we witness its reunion with unitarity and the Born rule [5]. The story is not merely one of analogy. It is a story of structural identity — a single mathematical operation applied to two different substrates, revealing a unity that was there all along, waiting for the field to become complex enough to see it.
- 24 September 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Quantum Dagger: How Quantum Mechanics and Attention Share a Mathematical Core
This paper tells the story of a mathematical object — the dagger — and its quiet, persistent presence in two of the most consequential theories of the twentieth and twenty-first centuries. The dagger, or conjugate transpose, is the defining operation of quantum mechanics: it turns a state into a dual, an operator into an observable, and a symmetry into a conserved quantity. In 2017, when Vaswani and colleagues introduced the Transformer [1], they wrote down an equation whose core operation —the scaled dot-product — contained the dagger in disguise, hidden behind the innocent transpose of a real-valued matrix. This paper is a narrative exploration of that coincidence, and an argument that it is not a coincidence at all. We trace the dagger from Born and Jordan’s matrix mechanics in 1926 [11] to the attention mechanism of modern artificial intelligence; we watch it split into real and imaginary parts, we follow its phase as it survives or fails to survive the softmax bottleneck, and we witness its reunion with unitarity and the Born rule [5]. The story is not merely one of analogy. It is a story of structural identity — a single mathematical operation applied to two different substrates, revealing a unity that was there all along, waiting for the field to become complex enough to see it.
- Open Access
- 23 June 2015
- Published by Hindawi Limited in Mathematical Problems in Engineering
- Vol. 2015, 1-13
Fractional Dynamics in Calcium Oscillation Model
The calcium oscillations have many important roles to perform many specific functions ranging from fertilization to cell death. The oscillation mechanisms have been observed in many cell types including cardiac cells, oocytes, and hepatocytes. There are many mathematical models proposed to describe the oscillatory changes of cytosolic calcium concentration in cytosol. Many experiments were observed in various kinds of living cells. Most of the experimental data show simple periodic oscillations. In certain type of cell, there exists the complex periodic bursting behavior. In this paper, we have studied further the fractional chaotic behavior in calcium oscillations model based on experimental study of hepatocytes proposed by Kummer et al. Our aim is to explore fractional-order chaotic pattern in this oscillation model. Numerical calculation of bifurcation parameters is carried out using modified trapezoidal rule for fractional integral. Fractional-order phase space and time series at fractional order are present. Numerical results are characterizing the dynamical behavior at different fractional order. Chaotic behavior of the model can be analyzed from the bifurcation pattern.
- 23 June 2015
- Published by Hindawi Limited in Mathematical Problems in Engineering
- Vol. 2015, 1-13
Open AccessFractional Dynamics in Calcium Oscillation Model
The calcium oscillations have many important roles to perform many specific functions ranging from fertilization to cell death. The oscillation mechanisms have been observed in many cell types including cardiac cells, oocytes, and hepatocytes. There are many mathematical models proposed to describe the oscillatory changes of cytosolic calcium concentration in cytosol. Many experiments were observed in various kinds of living cells. Most of the experimental data show simple periodic oscillations. In certain type of cell, there exists the complex periodic bursting behavior. In this paper, we have studied further the fractional chaotic behavior in calcium oscillations model based on experimental study of hepatocytes proposed by Kummer et al. Our aim is to explore fractional-order chaotic pattern in this oscillation model. Numerical calculation of bifurcation parameters is carried out using modified trapezoidal rule for fractional integral. Fractional-order phase space and time series at fractional order are present. Numerical results are characterizing the dynamical behavior at different fractional order. Chaotic behavior of the model can be analyzed from the bifurcation pattern.