Searcharxiv⌕ Search

arXiv subjects

Jack Humphries

Publications and source records attributed to Jack Humphries.

5 recordsLinked to original sources

A Log-Timescale Perturbation Bound for Selective State Space Models at Every Sequence Length

Mamba models deployed on streaming data, such as limit order book updates, run far past the sequence lengths they were trained on. A guarantee on how much a small change of input can move the hidden state therefore has to hold at every length. The closest existing bound we know of measures the change of the timescale additively, and it grows linearly with the sequence length unless the timescale is bounded away from zero. In this paper, we measure the change on a log scale instead. The key observation is that the decay factor of each step then moves by at most a fixed fraction of that step's input weight, so the per-step errors cannot pile up. Using this, we prove that two runs of a selective-scan cell from a common start stay within a bound that holds at every length, shrinks to zero with the size of the change and needs no lower bound on the timescale. We give a construction showing that, for small changes, this bound is within a factor of about 2 of the best possible, and we extend it to the update used in Mamba's released code when the timescale is bounded above. We then evaluate block-level bounds built from the theorem on two small models trained on one day of order book data. With a bounded timescale, a bound built on the existing analysis is tighter than ours for at most the first 8 steps. With Mamba's softplus timescale, whose certified lower bound is as small as $10^{-21}$, it is looser at every length. Our bounds hold for every input to a block, but they remain more than five orders of magnitude above the changes that gradient attacks find.

cs.LG↗

TW Hya: an old protoplanetary disc revived by its planet

Dark rings with bright rims are the indirect signposts of planets embedded in protoplanetary discs. In a recent first, an azimuthally elongated AU-scale blob, possibly a planet, was resolved with ALMA in TW Hya. The blob is at the edge of a cliff-like rollover in the dust disc rather than inside a dark ring. Here we build time-dependent models of TW Hya disc. We find that the classical paradigm cannot account for the morphology of the disc and the blob. We propose that ALMA-discovered blob hides a Neptune mass planet losing gas and dust. We show that radial drift of mm-sized dust particles naturally explains why the blob is located on the edge of the dust disc. Dust particles leaving the planet perform a characteristic U-turn relative to it, producing an azimuthally elongated blob-like emission feature. This scenario also explains why a 10 Myr old disc is so bright in dust continuum. Two scenarios for the dust-losing planet are presented. In the first, a dusty pre-runaway gas envelope of about 40 Earth mass Core Accretion planet is disrupted, e.g., as a result of a catastrophic encounter. In the second, a massive dusty pre-collapse gas giant planet formed by Gravitational Instability is disrupted by the energy released in its massive core. Future modelling may discriminate between these scenarios and allow us to study planet formation in an entirely new way -- by analysing the flows of dust and gas recently belonging to planets, informing us about the structure of pre-disruption planetary envelopes.

astro-ph.EP↗

On the origin of wide-orbit ALMA planets: giant protoplanets disrupted by their cores

Recent ALMA observations may indicate a surprising abundance of sub-Jovian planets on very wide orbits in protoplanetary discs that are only a few million years old. These planets are too young and distant to have been formed via the Core Accretion (CA) scenario, and are much less massive than the gas clumps born in the classical Gravitational Instability (GI) theory. It was recently suggested that such planets may form by the partial destruction of GI protoplanets: energy output due to the growth of a massive core may unbind all or most of the surrounding pre-collapse protoplanet. Here we present the first 3D global disc simulations that simultaneously resolve grain dynamics in the disc and within the protoplanet. We confirm that massive GI protoplanets may self-destruct at arbitrarily large separations from the host star provided that solid cores of mass around 10-20 Earth masses are able to grow inside them during their pre-collapse phase. In addition, we find that the heating force recently analysed by Masset and Velasco Romero (2017) perturbs these cores away from the centre of their gaseous protoplanets. This leads to very complicated dust dynamics in the protoplanet centre, potentially resulting in the formation of multiple cores, planetary satellites, and other debris such as planetesimals within the same protoplanet. A unique prediction of this planet formation scenario is the presence of sub-Jovian planets at wide orbits in Class 0/I protoplanetary discs.

astro-ph.EP↗

Constraining the initial planetary population in the gravitational instability model

Direct imaging (DI) surveys suggest that gas giants beyond 20 AU are rare around FGK stars. However, it is not clear what this means for the formation frequency of Gravitational Instability (GI) protoplanets due to uncertainties in gap opening and migration efficiency. Here we combine state-of-the-art calculations of homogeneous planet contraction with a population synthesis code. We find DI constraints to be satisfied if protoplanet formation by GI occurs in tens of percent of systems if protoplanets `super migrate' to small separations. In contrast, GI may occur in only a few percent of systems if protoplanets remain stranded at wide orbits because their migration is `quenched' by efficient gap opening. We then use the frequency of massive giants in radial velocity surveys inside 5 AU to break this degeneracy - observations recently showed that this population does not correlate with the host star metallicity and is therefore suspected to have formed via GI followed by inward migration. We find that only the super-migration scenario can sufficiently explain this population whilst simultaneously satisfying the DI constraints and producing the right mass spectrum of planets inside 5 AU. If massive gas-giants inside 5 AU formed via GI, then our models imply that migration must be efficient and that the formation of GI protoplanets occurs in at least a tens of percent of systems.

astro-ph.EP↗

Changes in the metallicity of gas giant planets due to pebble accretion

We run numerical simulations to study the accretion of gas and dust grains onto gas giant planets embedded into massive protoplanetary discs. The outcome is found to depend on the disc cooling rate, planet mass, grain size and irradiative feedback from the planet. If radiative cooling is efficient, planets accrete both gas and pebbles rapidly, open a gap and usually become massive brown dwarfs. In the inefficient cooling case, gas is too hot to accrete onto the planet but pebble accretion continues and the planets migrate inward rapidly. Radiative feedback from the planet tends to suppress gas accretion. Our simulations predict that metal enrichment of planets by dust grain accretion inversely correlates with the final planet mass, in accordance with the observed trend in the inferred bulk composition of Solar System and exosolar giant planets. To account for observations, however, as much as ~30-50% of the dust mass should be in the form of large grains.

astro-ph.EP↗