Holographic Inversion That Can Say "No Dual Exists"

Holography is the conjecture that some quantum field theories are exactly equivalent to a theory of gravity in one extra dimension. Recent work treats the "holographic inverse problem" as a learning problem. It takes measurable boundary data, such as two-point correlation functions (Green functions), and trains a Neural ODE (a network whose depth is a continuous variable integrated by an ODE solver) so that its depth plays the role of the extra radial dimension. The learned network weights then encode the bulk spacetime geometry. These methods share a flaw that visual-grounding researchers have already named the existence assumption bias. Asked to locate "the blue elephant" in a kitchen photo, a standard grounding model draws a box anyway, because it was only ever trained on queries that have an answer. A holographic inversion network does the same thing with physics: given any boundary data, it bends its metric until the loss is small. It never reports that no healthy gravitational dual exists.

We propose to make bulk reconstruction able to abstain, and to make the abstention mean something physical. We drop the question "what geometry fits these data?" and ask instead "does any admissible geometry fit them?" A geometry is admissible if it is smooth, satisfies an energy condition, and carries no tachyonic instability. A tachyonic instability means a field with effectively negative mass-squared, which grows exponentially instead of oscillating; in anti-de Sitter space this happens when the mass falls below the Breitenlohner–Freedman (BF) bound. Our abstention signal is the gap between the best constrained fit and the best unconstrained fit. When abstention fires, the radius at which the constraint becomes binding tells us where in the bulk the would-be dual breaks down.

Core Claim

Let $G$ be boundary Green-function data and $h$ the learned bulk metric function. Let $L(h;G)$ be the boundary reconstruction loss, and let $\mathcal{A}$ be the set of admissible metrics. We define the admissibility gap

$$ \Delta(G) = \min_{h \in \mathcal{A}} L(h;G) \;-\; \min_{h} L(h;G), $$

and declare a null target ("no healthy bulk dual") when $\Delta(G)$ exceeds a threshold $\tau(\sigma)$. The threshold is calibrated on known-positive data at boundary noise level $\sigma$.

We make two claims. First, $\Delta$ separates data that have duals from data that do not, including deceptive near-miss cases that the network fits to near-zero loss with a geometry that is quietly tachyonic. Neither a loss threshold nor a check on the single fitted geometry separates these reliably. Second, the location where the admissibility constraint is active localizes the obstruction in the radial direction.

The second claim matters because of non-uniqueness. A free fit can land on an inadmissible geometry even when an admissible one also fits. A check on that one solution then wrongly rejects data that do have a dual. Searching over $\mathcal{A}$ removes this failure mode.

How This Sits Relative to Prior Work

The method-supplying papers are the two Neural ODE holography works. Neural ODE and Holographic QCD maps network depth to the radial coordinate and learns a metric from lattice-QCD chiral condensates, discovering a horizon with an adaptive dopri5 solver. Its only notion of rejection is a binary classifier that separates physical data from random noise. Learning geometries beyond asymptotic AdS uses holographic Wilsonian renormalization-group flow equations to reconstruct AdS, Lifshitz, and hyperscaling-violating metrics from Green functions, reporting loss near $10^{-10}$. For the SYK model, a solvable quantum many-body system often used as a toy dual of black holes, it finds an analytic critical curve $\epsilon = 1/(1+2\pi\bar T/\tilde v)$ that bounds where a bulk dual exists. That curve is our best ground truth for where a "null target" begins. Neither paper abstains, and neither checks stability of fields on the learned geometry. We reuse their pipelines unchanged as the unconstrained fit and add the constrained search and the gap.

The stability constraints come from the tachyonic-instability literature, which plays a constraining role here. Unveiling the Inconsistency of the Proca Theory with Nonminimal Coupling to Gravity and On the Kalb-Ramond field with non-minimal coupling to gravity show that curvature couplings generate runaway modes that are invisible in naive background analyses. Symmetries for Galileons and DBI scalars on curved space shows that curved backgrounds can force effective masses negative. Two lessons follow for us. A pointwise check that the effective mass stays above the BF bound at each radius is not the right criterion. And the effective mass of a probe field depends on the learned curvature through couplings of the form $m_{\rm eff}^2 = m^2 + \xi R$. We therefore define stability globally: the lowest eigenvalue of the radial wave operator for the probe field on the learned geometry must be non-negative. This eigenvalue can be computed differentiably at each training step.

On classical de Sitter solutions and parametric control is the closest analogue of a "null result" in gravity. It searches compactifications for a stable vacuum and repeatedly finds tachyons, but it does so by manual scans and issues no calibrated rejection.

The grounding literature motivates the design. GRES: Generalized Referring Expression Segmentation and GREx show that a dedicated no-target pathway beats forcing every query onto some region. Both also identify deceptive near-miss queries as the hardest remaining cases. Rex-Thinker quantifies the bias: base Qwen2.5-VL scores 7.1 on HumanRef rejection, versus 68.2 after explicit verification. We borrow three things from this work: the framing, the deceptive-case benchmark design, and the mixed-accuracy style of evaluation. We do not borrow any vision model.

Research Program

We build a benchmark first, because none exists. It is modeled on gRefCOCO's split into single-target, multi-target, and no-target queries, and has three parts.

The positive set contains Green functions generated, across a range of noise levels, from the geometry families in Cheng et al. (AdS, Lifshitz, hyperscaling-violating), plus SYK large-$q$ data on the dual-exists side of the critical curve.

The clear-negative set contains boundary data with no plausible dual. One family has spectral functions with negative weight, which violates unitarity. The other is SYK data well beyond the critical curve.

The deceptive set is the heart of the benchmark. We take smooth geometries and add an interior bump that makes a probe scalar, or a curvature-coupled scalar with $\xi \neq 0$, develop a negative-energy bound state; alternatively, the bump violates the null energy condition. We then compute their exact boundary Green functions. Such data look ordinary at the boundary, an unconstrained network fits them well, and yet no stable static dual exists. Some deceptive cases are deliberately chosen so that an admissible geometry fits almost as well as the generating one. These test whether the gap correctly accepts them.

The method has three parts. First, we parameterize $\mathcal{A}$, using monotone network heads where an energy condition reduces to monotonicity of a warp-factor derivative. Second, we impose a differentiable penalty on the lowest radial eigenvalue for stability. Third, we compute $\Delta$ from paired constrained and unconstrained fits. Each fit gets multiple random restarts so that a large $\Delta$ is not just an optimization failure.

We calibrate $\tau(\sigma)$ on held-out positives to a target false-rejection rate. We read out the obstruction location as the radius where the penalty's gradient concentrates, and compare it with the bump location that generated the data.

Baselines

The naive reference point is the unconstrained Neural ODE of Learning geometries beyond asymptotic AdS, run with its released settings, with a single threshold on final loss. Data whose fit loss exceeds the threshold count as "no dual". The threshold is tuned on the same positive calibration split we use for $\tau$. The relevant reported number is reconstruction loss of about $10^{-10}$ on AdS, Lifshitz, and hyperscaling-violating data. We searched Lacuna and found no reported abstention or rejection result for any holographic inversion method. This baseline is beaten if, at matched false-rejection rate on positives, our method rejects substantially more of the deceptive set.

The strongest existing competitor is a post-hoc physics check. We run the same unconstrained fit, then compute the null energy condition and the lowest radial eigenvalue on the single geometry it returns, and reject if either fails. This is the honest version of what a physicist would do by hand with linear perturbation theory. It gets an identical fitting budget: the same restarts and the same solver tolerances.

The grounding-style competitor is an amortized classifier: a network with a dedicated null output trained directly on Green functions to predict "dual exists", in the spirit of the ReLA null token. As a concrete number for that mechanism, GREx reports 63.98% gIoU on gRefCOCO val for ReLA. It is trained on the same benchmark training split.

The key ablation removes the gap and simply thresholds the constrained loss. This isolates whether the comparison with the free fit is what absorbs noise-level variation.

Throughout, we beat a baseline when we achieve higher mixed accuracy over positives and negatives (the analogue of $R_{\rm mix}$) and a lower false-acceptance rate on deceptive cases at equal positive acceptance.

Why not just…?

Why not just threshold the fit loss? The deceptive set is built so that this fails by construction: a flexible metric fits tachyonic-dual data to small loss. The grounding evidence that forced-output decoders produce confident false positives (Rex-Thinker's 7.1 rejection score for the base model) suggests the same failure here. The first experiment measures it directly, and if loss thresholding does well on deceptive data, the benchmark is too easy and we will make it harder.

Why not just check the fitted geometry for tachyons and energy-condition violations afterwards? This is the real threat. If the free fit's solution is usually close to the admissible one, the post-hoc check will match our method at lower cost, and we will report that. Our argument rests on non-uniqueness: when several geometries fit, a free fit may land on an inadmissible one and trigger a false rejection. The near-degenerate deceptive cases measure how often this happens. The proposal's value claim stands or falls on that number.

Why not just train a classifier with a null output on Green functions? The ReLA null token works because grounding has millions of labeled examples drawn from a fixed distribution. A classifier trained on our synthetic families may learn surface features of those families, give no physical reason for its decision, and fail on new geometry types. We test this with a held-out-family split: train on AdS and Lifshitz, test on hyperscaling-violating and SYK. We expect the classifier to degrade there while the gap does not, because the gap's decision rests on a physical criterion rather than on learned features.

Decisive Checks

Four results would settle the question. The first is whether the rejection boundary that $\Delta$ traces across SYK parameter space recovers the analytic critical curve of Cheng et al. without being told it. This is the cleanest test, because the ground truth is known exactly. The second is whether, on the deceptive set, the gap beats the post-hoc check specifically on the near-degenerate cases, with the obstruction location landing near the planted bump. The third is whether the lattice-QCD condensate data of Neural ODE and Holographic QCD are accepted. Rejecting them would signal an over-restrictive $\mathcal{A}$ rather than a discovery. The fourth is whether those same data, after we perturb their spectral content into negative weight, are rejected.

We will treat a failure of the first check, or a draw with the post-hoc check on the second, as evidence against the proposal.