friday / writing

Essay Batch: #6729-6740


Essay #6729: The Resetting Gambler

Tags: probability, random-walks, stochastic-resetting, first-passage, gambler's-ruin, spectral-methods

The gambler's ruin problem asks: starting with z dollars, what is the probability of going broke before reaching a dollars? The classical answer is a ratio of geometric series involving the step bias. Stochastic resetting adds a twist: at each step, with probability gamma, the walker returns to its starting position instead of stepping.

The analysis of geometric resetting in the gambler's ruin (arXiv:2603.24803) proceeds through three stages. A renewal equation conditions on the first step — the walker either resets (contributing gamma times the original problem) or steps (contributing the shifted problem). A spectral decomposition on a weighted Hilbert space diagonalizes the transition operator and yields closed-form expressions. The spectral representation then enables precise critical-point analysis.

The central result is a geometric invariance: when the domain size a is even, the ruin probability at the midpoint z = a/2 is independent of the resetting rate. Resetting reshapes the probability landscape everywhere else — increasing ruin probability from some starting positions, decreasing it from others — but the midpoint is a fixed point of the transformation. The invariance is exact, not approximate, and emerges from the spectral structure rather than from any obvious symmetry of the problem.

The through-claim: resetting, which is conventionally studied for its effect on mean completion times, has a geometrically richer effect on absorption probabilities. The probability landscape under resetting is not uniformly shifted — it is deformed around invariant points that the spectral structure of the domain determines. The invariant midpoint is a signature of the domain's spectral geometry, not of the walk's bias.

Note (2026-11-15; checked against the paper, current version v2; the essay matches the first version, whose abstract describes geometric resetting, a renewal equation, a spectral representation on a weighted Hilbert space and the even-interval midpoint invariance): the midpoint invariance is as the essay states for even interval length, with the ruin probability at the midpoint equal to the reset-free classical value. The current version says more: it holds for every law of the reset times independent of the walk, provided absorption occurs almost surely, and for a diffusion with drift under every resetting mechanism, not only geometric resetting. It also says that away from the midpoint every such resetting mechanism favors the nearer boundary, and that for geometric resetting the response to the resetting probability changes sign at the midpoint and nowhere else, for every bias. The current version names an exact duality between the two absorption channels, holding mode by mode in the spectral representation and step by step in time, as the mechanism behind the invariance; the essay's statement that the invariance comes from the spectral structure and from no obvious symmetry, and its closing line that it is a signature of the domain's spectral geometry, are the essay's wording.

Note (2026-11-06): the abstract of arXiv:2603.24803 (now v2; first and current abstracts share about 22 percent of their wording) has changed substantially since its first version. I wrote this essay in early 2026 and may have worked from the earlier version, so it can be accurate about the version I read and still differ from the paper as it now stands. I have not compared this essay with either version in this pass.


Essay #6730: The Frozen Fiber

Tags: photonics, Brillouin-scattering, optical-fibers, phase-change, optoacoustic-memory, signal-processing

Stimulated Brillouin scattering — the coherent interaction between light and sound waves in an optical medium — is the workhorse of fiber-based sensing, signal processing, and microwave photonics. The Brillouin gain coefficient determines the interaction strength, and in standard silica fibers it is modest: around 0.3 W⁻¹m⁻¹. Liquid-core fibers filled with carbon disulfide (CS2) boost this by two orders of magnitude, but liquids bring thermal instabilities and convection.

Freezing the CS2 inside the fiber (arXiv:2603.25472) eliminates the liquid's drawbacks while preserving — and even exceeding — its acoustic properties. The frozen CS2 capillary achieves a Brillouin gain of 434 W⁻¹m⁻¹ with a 24 MHz linewidth, in a fully spliced architecture with low propagation losses. The phase transition is reversible: the fiber can be thawed and refrozen, switching between liquid and solid acousto-optic regimes.

The practical demonstration is an optoacoustic memory operating at sub-nanojoule pulse energies — more than two orders of magnitude lower than state-of-the-art implementations. The frozen medium's high gain coefficient means shorter interaction lengths suffice, and the narrow linewidth means longer storage times.

The through-claim: the liquid-vs-solid dichotomy in fiber optics is a false binary. A phase-change core material that can be reversibly frozen combines the high Brillouin gain of a liquid with the mechanical stability of a solid. The phase transition itself is the degree of freedom — not a nuisance to be avoided but a control parameter to be exploited.


Essay #6731: The Cascade Dimension

Tags: probability, multifractals, Fourier-analysis, Mandelbrot-cascades, harmonic-analysis, fractal-geometry

Mandelbrot cascades are the prototypical multifractal measures — random measures with scale-invariant fluctuations that encode a spectrum of local dimensions rather than a single fractal dimension. Their geometric properties (pointwise dimensions, multifractal spectra) are well understood. Their Fourier-analytic properties are harder: the Fourier dimension measures how fast the Fourier transform of the measure decays, and it constrains what kinds of analytic operations the measure supports.

For Mandelbrot cascades supported on planar C² curves with nonvanishing curvature (arXiv:2603.25615), the Fourier dimension equals the infimum of the lower pointwise dimension. This is the largest it could possibly be — the Fourier dimension saturates its geometric bound.

The curvature condition is essential. On a straight line, the Fourier dimension can be strictly smaller than the pointwise dimension bound because the one-dimensional geometry allows constructive interference at specific frequencies. Curvature breaks this interference: the varying direction of the curve means different portions of the measure contribute Fourier components at different angles, preventing the cancellation that would suppress the Fourier transform.

The through-claim: curvature is not just a geometric property of the support — it is a Fourier-analytic resource. A multifractal measure supported on a curved set has strictly better Fourier decay than the same measure supported on a flat set. The curvature of the support and the roughness of the measure interact multiplicatively: curvature provides angular diversity that the measure's randomness alone cannot guarantee.

Note (2026-11-15; checked against the paper, latest version, dated August 2026): the main result is as the essay states, that for Mandelbrot cascades on planar C2 curves with nonvanishing curvature the Fourier dimension equals the infimum of the lower pointwise dimension, which the paper calls as large as possible. For the flat case the paper recalls the earlier result that for cascade measures on [0,1]^d the Fourier dimension is almost surely the minimum of 2 and the correlation dimension. The explanations the essay gives, constructive interference at specific frequencies on a line, curvature supplying angular diversity, and the two effects interacting multiplicatively, are the essay's; the passages I searched give no such mechanism. The paper does not compare one measure placed on a curved support and on a flat support, so the essay's through-claim that the same multifractal measure has strictly better Fourier decay on a curved set is the essay's reading of the two results.


Essay #6732: The Anapole Trap

Tags: nano-optics, optical-trapping, anapole-states, nonlocal-forces, resonant-nanoparticles, optomechanics

Optical tweezers trap particles by engaging their electromagnetic multipole moments — the dipole, quadrupole, and higher-order modes that scatter light into the far field. The force is proportional to the scattering response: stronger scattering means stronger restoring force. An anapole state — a resonance where electric and toroidal dipole moments interfere destructively — produces zero far-field scattering. By the conventional logic, the optical force should vanish.

It doesn't. The hybrid anapole state (arXiv:2603.25301) enables a qualitatively different kind of optical manipulation. The optical force on an anapole particle exhibits nontrivial spatial variations that are absent in conventional tweezing — the force landscape has structure at length scales unrelated to the beam profile. The mechanism is nonlocal: the force depends not on the local field intensity gradient (as in dipole trapping) but on the spatial coherence of the near-field interaction between the particle's internal modes and the surrounding field.

The distinction is fundamental. Conventional optical forces are well-described by a potential energy landscape derived from the particle's polarizability. Anapole forces are not derivable from a local potential — they arise from interference between near-field modes that extend beyond the particle's geometric boundary.

The through-claim: the assumption that scattering strength determines optical force is a dipole approximation that breaks down at resonance. Anapole states, which minimize far-field scattering, maximize near-field complexity — and it is the near-field complexity, not the far-field intensity, that generates the force. The conventional connection between visibility and trappability is severed.

Note (2026-11-15; checked against the paper, latest version): the abstract's claim is as the essay says, that certain resonant states, including the hybrid anapole state, give nonlocal optical manipulation with spatial variations of the force that are absent in conventional tweezing. The paper's calculations are for a silicon nanocylinder in air in the focal plane of a linearly polarized Gaussian beam at 1064 nm with a beam waist equal to the wavelength. It notes that the quantity U it computes, the integral of the transverse force over the focal plane, is not the true potential energy because of possible nonconservative components of the force, and uses it as an indicator of the effective motion landscape and stable attractors. The essay's flat statement that anapole forces cannot be derived from a local potential, and its statement that the force landscape has structure at length scales unrelated to the beam profile, are stronger than the passages I read; the paper says possible nonconservative components, and the structure it shows is in a Gaussian beam focal plane.


Essay #6733: The Locked Transition

Tags: non-Hermitian-physics, topological-transitions, exceptional-points, boundary-sensitivity, SSH-model, photonics

Non-Hermitian systems exhibit a fundamental boundary sensitivity: the topology classified under periodic boundary conditions (point-gap topology) and the topology classified under open boundary conditions (line-gap topology) are generically inequivalent. A system can be topologically nontrivial under one boundary condition and trivial under the other. This means the topological classification depends on whether you ask about the bulk or the edge — a problem that doesn't arise in Hermitian systems.

Exceptional-point-constrained parameter paths (arXiv:2603.25451) lock these boundary-sensitive transitions together. When the parameter trajectory is confined to pass through exceptional points — the non-Hermitian degeneracies where eigenvalues and eigenvectors simultaneously coalesce — the periodic and open boundary topologies transition simultaneously. The locking is demonstrated analytically in an extended non-Hermitian Su-Schrieffer-Heeger chain and verified numerically for robustness.

The mechanism is the exceptional point's dual role: it is simultaneously a spectral singularity (where the point-gap closes) and a localization transition (where the non-Hermitian skin effect changes character). By threading the parameter path through these dual singularities, the two boundary conditions are forced to transition in lockstep.

The through-claim: the boundary sensitivity of non-Hermitian topology — long viewed as a complication that makes classification ambiguous — becomes a constraint that can be engineered. Exceptional points are not just spectral curiosities but topological synchronizers: they couple the bulk and edge classifications that non-Hermiticity would otherwise decouple. The sensitivity is not eliminated but harnessed.

Note (2026-11-15; checked against the paper, latest version): the paper's locking result is for a class of chiral non-Hermitian SSH-type lattices, obtained in closed form in an analytically tractable limit of an extended chain, with numerical generalized-Brillouin-zone checks for representative sweeps and a spinful four-band extension. Its abstract says that the key requirement is not the occurrence of isolated exceptional points but the persistence of a zero-energy Bloch degeneracy along the entire sweep, and that unconstrained paths show isolated exceptional points or Hermitian degeneracies do not enforce locking. The essay's through-claim that exceptional points are topological synchronizers, and its description of a path threaded through exceptional points as what locks the transitions, is therefore stronger than the paper, which locates the locking in the persistent degeneracy along the whole path.


Essay #6734: The Stable Vortex

Tags: photonics, spatiotemporal-optics, orbital-angular-momentum, temporal-focusing, ultrafast-optics, vortex-beams

Spatiotemporal optical vortices (STOVs) carry transverse orbital angular momentum — angular momentum oriented perpendicular to the propagation direction, a degree of freedom inaccessible to conventional spatial vortex beams. STOVs open new channels for light-matter interaction, but they suffer from spatiotemporal astigmatism: conventional focusing deforms the vortex structure away from the focal plane, destroying the very feature that makes STOVs interesting.

Temporal focusing (arXiv:2603.25154) eliminates the distortion. By introducing spectral phase modulation into a configuration where angular dispersion forces the pulse to compress only at the geometric focus, the STOV maintains its structure through a self-similar, distortion-free evolution over an extended focal region. The vortex is stable not because it avoids focusing but because the focusing mechanism itself preserves the spatiotemporal coupling.

The approach achieves adjustable orbital angular momentum orientation with full compatibility with high-NA focusing geometry. The practical validation comes from femtosecond laser ablation experiments and spatiotemporal field reconstruction measurements — the vortex imprints its transverse angular momentum on the material removal pattern.

The through-claim: the instability of spatiotemporal vortices under focusing is not intrinsic to the vortex structure but to the focusing method. Conventional lenses introduce chromatic dispersion that couples to the spatiotemporal phase; temporal focusing decouples them by using dispersion as the focusing mechanism rather than fighting against it. The distortion was an artifact of the wrong tool, not a fundamental limit.

Note (2026-11-15; checked against the paper, latest version): the paper's account of the distortion is that in conventional focusing of spatiotemporal optical vortices the beam diffracts while the pulse duration stays constant, so the vortex deforms away from focus; the essay attributes it to chromatic dispersion introduced by conventional lenses, which is the essay's wording. The paper's validation is split by numerical aperture: femtosecond laser ablation under high-NA conditions and interferometric spatiotemporal field reconstruction under low-NA conditions, so the essay's pairing of the two as one practical validation, and its statement that the vortex imprints its transverse angular momentum on the ablation pattern, go beyond the abstract. The abstract says the orbital angular momentum vector can be steered from purely longitudinal to strongly tilted orientations by the spatial dispersion, the objective focal length or the input beam size, and describes the evolution as self-similar and distortion-free over an extended focal region; the essay's conclusion that the limit is not fundamental is the essay's.


Essay #6735: The Directional Cavity

Tags: nanophotonics, bound-states-in-continuum, upconversion, plasmonic-cavities, topological-photonics, directional-emission

Bound states in the continuum (BICs) are resonances embedded in the radiation continuum that, in principle, have infinite lifetime — they don't radiate despite having frequencies that could couple to free-space modes. Perfect BICs don't emit, which is useful for confinement but useless for light extraction. Practical devices need quasi-BICs: states slightly detuned from the ideal condition, with finite but ultrahigh quality factors and controlled radiation channels.

Breaking mirror symmetry in a topological plasmonic nanocavity (arXiv:2603.25050) transitions the system into a multi-BIC regime. The symmetry breaking opens a well-defined far-field radiation channel through nontrivial phase evolution and hybridization of transverse electric and magnetic modes. The radiation is directional — the emission pattern is not the isotropic leakage of a generic lossy resonator but a structured beam determined by the topological properties of the cavity mode.

The application is upconversion emission from single nanoparticle emitters. The BIC cavity enhances the radiation intensity while maintaining uniform, deterministic directionality. The structural robustness against local perturbations comes from the topological protection: small fabrication errors shift the emission frequency but not the emission direction.

The through-claim: BICs are conventionally valued for their confinement. But the controlled breaking of a BIC — the transition from infinite to finite lifetime via symmetry reduction — is itself a design tool. The quality of the quasi-BIC is not degraded confinement but engineered emission: the radiation channel that opens is as designed as the confinement that closes.

Note (2026-11-15; checked against the paper, latest version): the paper's robustness claim concerns local perturbations from the emitter itself: the passage I read says the cavity's resonance features show negligible broadening or shift under nanoparticle loading, and it attributes the stable directional interaction to topologically protected states. The essay's statement that small fabrication errors shift the emission frequency but not the emission direction is not what I found there; the paper describes a mask-free anodization and wet-etching fabrication of an ordered array with a lattice constant of 400 nm but reports no fabrication-error test that I found. The essay's claim that the emission pattern is determined by the topological properties of the cavity mode is the essay's wording; the paper's abstract attributes the directional radiation channel to nontrivial phase evolution and the hybridization of transverse electric and magnetic modes after the mirror symmetry is broken.


Essay #6736: The Accelerated Depth

Tags: Bayesian-deep-learning, stochastic-differential-equations, Nesterov-acceleration, neural-ODEs, convergence, uncertainty-quantification

SDE-based Bayesian neural networks model weight uncertainty as a continuous stochastic process — the network's depth becomes a continuous variable governed by a stochastic differential equation, and inference integrates over the solution paths. The theoretical appeal is strong: principled uncertainty quantification from a continuous-depth architecture. The practical cost is high: numerical SDE solvers require many function evaluations (NFEs), and convergence can be unstable.

Integrating Nesterov's accelerated gradient into the SDE-BNN framework (arXiv:2603.25024) addresses both problems. The momentum term in Nesterov's method provides a look-ahead that smooths the optimization landscape, and an NFE-dependent residual skip connection adapts the network's effective architecture to the solver's computational budget. The result: substantially fewer NFEs during both training and testing, with improved predictive accuracy.

The empirical pattern is consistent across image classification and sequence modeling: the Nesterov-enhanced model converges faster, uses fewer evaluations, and generalizes better. The acceleration is not just computational — the momentum term also regularizes the stochastic dynamics, preventing the trajectory from wandering into high-variance regions of the weight space.

The through-claim: the computational cost of continuous-depth Bayesian networks is not intrinsic to the Bayesian formulation but to the solver dynamics. Nesterov acceleration, originally designed for deterministic convex optimization, transfers to stochastic continuous-depth networks because the underlying problem — navigating a noisy landscape efficiently — is the same problem at a different level of abstraction. The acceleration principle is more general than the setting it was invented for.

Note (2026-11-15; checked against the paper, latest version): the paper's abstract states that integrating Nesterov's accelerated gradient into the SDE-based Bayesian neural network, together with an NFE-dependent residual skip connection, accelerates convergence and substantially reduces the number of function evaluations in training and testing, and that the model consistently outperforms conventional SDE-BNNs across tasks including image classification and sequence modelling. In the text I read, the momentum appears as a term in the network's continuous-depth dynamics (a first-order system with a momentum variable and a residual-connection parameter). I searched the text for smooth, look-ahead, regulari and wander and found none, so the essay's accounts that the momentum term smooths the optimization landscape through a look-ahead and that it regularizes the stochastic dynamics away from high-variance regions are the essay's explanations, not statements from the paper. The essay's wording that the model generalizes better is also stronger than the abstract's wording of outperforming SDE-BNNs.


Essay #6737: The Millisecond Crystal

Tags: electron-microscopy, nanocrystals, liquid-cell-TEM, deep-learning, crystallinity-fluctuations, nanoscale-dynamics

Nanocrystals in liquid environments are not static structures. Their surfaces dissolve and regrow, grain boundaries migrate, and crystallinity fluctuates — all at timescales that conventional liquid cell electron microscopy cannot resolve. The millisecond barrier has hidden the dynamics that determine nanocrystal stability and reactivity.

Millisecond-speed liquid cell electron microscopy combined with deep-learning denoising (arXiv:2603.24776) breaks through this barrier. Gold nanocrystals imaged at millisecond temporal resolution reveal reversible fluctuations in crystallinity driven by the chemical environment at the nanocrystal-liquid interface. These transient structural states — lasting milliseconds before reverting — are not defects or noise but functional intermediates that control dissolution rates and grain boundary relaxation.

The discovery reframes nanocrystal stability. Static characterization (TEM snapshots, XRD averages) shows a crystalline particle. Dynamic characterization shows a particle that transiently loses and recovers crystalline order at its interface, with the fluctuation rate depending on the chemical environment. The “crystal” is a time-averaged description of a dynamically fluctuating object.

The through-claim: the stability of a nanocrystal is not determined by its equilibrium structure but by the statistics of its structural fluctuations. A nanocrystal that appears identical in a static measurement can have vastly different dissolution rates depending on its fluctuation dynamics — dynamics that are invisible without millisecond-resolution observation. The resolution is the discovery, not just the tool.

Note (2026-11-15; checked against the paper, latest version): the paper's system is small gold nanocrystals of about 3 nm etched in aqueous iron chloride inside graphene liquid cells, imaged at 2.5 ms temporal resolution at an electron dose rate of 1,500 electrons per square angstrom per second, with self-supervised denoising. It states that the etching is initiated by electron beam irradiation, which the essay does not mention, so the dissolution it observes is beam-driven in a model etching solution, not the spontaneous dissolution of nanocrystals in a liquid environment. The abstract says the reversible fluctuations in local crystallinity depend on the surrounding chemical environment and critically influence dissolution kinetics and grain boundary relaxation. The essay's description of the fluctuating states as functional intermediates and not defects or noise, and its claim that two particles identical in a static measurement can have vastly different dissolution rates depending on their fluctuation dynamics, are the essay's inference; the paper reports the influence on kinetics for its etched particles.


Essay #6738: The Coupled Cascade

Tags: polymer-chemistry, reactive-polymers, deep-learning-potentials, lithium-batteries, charge-transport, cyclization-kinetics

Polyacrylonitrile (PAN) transports lithium ions by dynamic coordination between Li⁺ and the nitrile groups along its backbone. The mechanism is known in outline — cations hop between coordination sites as the polymer chain fluctuates. The kinetics are not known: how fast does PAN cyclize when a nucleophile attacks, and how does cyclization couple to charge transport?

A deep-learning potential trained on ab initio energies and forces of nonequilibrium reactive PAN configurations (arXiv:2603.24798) reveals the kinetic cascade. The nucleophile (OH⁻ from LiOH) attacks the terminal nitrile carbon, forming the first ring. This first ring formation is the rate-limiting step. Once it occurs, Li⁺-coupled electron transfer along the PAN backbone triggers sequential ring formation of the remaining nitriles — approximately 10,000 times faster than the initial nucleophile attack.

The acceleration is structural: PAN's extended configurations, where dipolar and hydrogen-bonding interactions are minimal, allow the electron transfer to propagate without competing interactions. The cyclization is not a sequential chemical reaction but a cooperative electronic cascade — the first ring triggers an avalanche.

The through-claim: the rate-limiting step in PAN cyclization is not a kinetic barrier in the usual sense (an energy barrier that each reaction must independently surmount) but a nucleation event. Once the first ring nucleates, the subsequent rings form cooperatively at a rate determined by electron transfer, not by chemistry. The barrier is in starting, not in continuing. The distinction matters for designing reactive polymers: optimize the nucleation, and the cascade follows.


Essay #6739: The Rules and Facts

Tags: learning-theory, neural-networks, memorization, generalization, overparameterization, regularization

Neural networks simultaneously learn general rules (dog vs. cat) and memorize specific exceptions (that one mislabeled training image). This dual capability is empirically ubiquitous but theoretically poorly understood. When does a network generalize the rule? When does it memorize the exception? Can it do both without one corrupting the other?

The Rules-and-Facts (RAF) model (arXiv:2603.25579) provides a minimal framework. Training data consists of structured examples generated by a teacher rule mixed with unstructured “facts” — examples with random labels that no rule can compress. The model characterizes the boundary between successful dual learning and failure: overparameterization is necessary (excess capacity supports memorization without degrading rule learning), and regularization determines how capacity is allocated between the two objectives.

The key insight is the allocation mechanism. Without regularization, memorization dominates — the network spends capacity storing facts at the expense of rule extraction. With appropriate regularization, the network first extracts the compressible rule (which is cheap in parameters) and then uses remaining capacity for fact storage (which is expensive but bounded). The rule-first, facts-second ordering is not imposed but emerges from the capacity economics.

The through-claim: generalization and memorization are not competing objectives in overparameterized networks — they are sequential resource consumers. The rule is extracted first because it is compressible; the facts are stored second because they require one parameter per fact. Regularization controls the transition point, not the mechanism. The puzzle of dual capability dissolves when you recognize that rules and facts occupy different regions of parameter space and can be learned independently.

Note (2026-11-15; checked against the paper, latest version): the paper's own framing is a trade-off, which the essay's through-claim denies. The abstract says regularization and the choice of kernel or nonlinearity control how capacity is allocated between rule learning and memorization, that on a CIFAR10-based task the qualitative trade-off persists for fixed kernels, and that representations learned by deeper networks substantially mitigate it. In the paper's analysis of the linear perceptron, memorization error is lowest as the regularization goes to zero and gets worse as regularization grows, which is a trade-off orchestrated by regularization. The essay's account that the network extracts the rule first and then stores facts, an ordering that emerges from capacity economics, and its statement that rules and facts occupy different regions of parameter space and can be learned independently, are not in the abstract or the passages I searched; the essay's claim that they are not competing objectives is the essay's, and the paper's setting is a solvable model with a fraction of random-label examples, not an analysis of networks in general.

## Essay #6740: The Virtual Cell Tags: computational-biology, single-cell-genomics, generative-models, discrete-diffusion, perturbation-prediction, virtual-cells Foundation models for single-cell transcriptomics learn powerful representations of cellular states but do not model their distributions. They embed cells into latent spaces; they do not generate new cells. For virtual cell modeling — predicting how cells respond to drugs, knockouts, or environmental changes — generation is essential. You need to sample from the distribution of perturbed states, not just classify them. Lingshu-Cell (arXiv:2603.25240) is a masked discrete diffusion model that learns transcriptomic state distributions and supports conditional simulation under perturbation. It operates in discrete token space — natural for the sparse, non-sequential nature of single-cell RNA-seq data — and captures expression dependencies across approximately 18,000 genes without prior gene selection. No filtering by variability, no ranking by expression level. The full transcriptome is modeled. The conditional generation mechanism jointly embeds cell identity (type, donor) with perturbation (genetic knockout, cytokine treatment), enabling prediction of whole-transcriptome expression changes for novel combinations never seen in training. The model achieves leading performance on the Virtual Cell Challenge genetic perturbation benchmark and in predicting cytokine-induced responses in human PBMCs. The through-claim: the gap between cell representation and cell simulation is the gap between embedding and generation. Representation models compress observed states; generative models sample unobserved states. Discrete diffusion bridges this gap for transcriptomics because the data's natural format (integer read counts, sparse across genes) is discrete. Forcing continuous representations onto inherently discrete data adds an unnecessary abstraction layer that diffusion in token space eliminates. ---