Metrical Determinism in Theory and Practice

• Metrical determinism is a multifaceted concept where metric and measurement structures impose constraints that determine admissible outputs and system behaviors.
• In online algorithms, barely-random strategies achieve near-optimal competitive ratios, balancing limited randomness with strong performance guarantees.
• Across fields like differential geometry, relativity, MBQC, and prosody, metrical determinism either secures unique outcomes or reveals evaluation illusions from aggregated metrics.

Metrical determinism is a polysemous research term used in several technically distinct senses. In online algorithms, it denotes minimizing randomness in metrical task systems while preserving near-randomized competitive guarantees (Cosson et al., 2024). In differential geometry and relativity, it denotes ways in which a metric or causal structure fixes geodesics, domains of dependence, or metric-preserving morphisms (Voellinger, 6 Mar 2026, Isenberg, 2016, Halvorson et al., 7 Mar 2025). In evaluation methodology for diffusion LLMs, it denotes the illusion of stability created when dataset-level metrics average away sample-level variability (Fang et al., 15 Apr 2026). In prosody, music, and controllable generation, it denotes the extent to which metrical constraints determine admissible outputs or hierarchical meter (Hu et al., 2023, Jiang et al., 2022, Rama et al., 2010). The term therefore does not name a single doctrine; it names a family of determinacy claims anchored in metric, metrical, or measurement structure.

1. Terminological scope

Current usage distributes the term across several literatures rather than a single unified field. The common pattern is that a formal structure—metric, metrical hierarchy, or evaluation metric—determines, or appears to determine, a class of admissible evolutions, outputs, or comparisons.

Domain	Sense of metrical determinism	Representative paper
Online algorithms	Very little randomness in MTS while keeping competitive guarantees	(Cosson et al., 2024)
Riemannian geometry	The metric determines an isometry $\Theta_g$	(Voellinger, 6 Mar 2026)
Relativity and formal theories	The metric fixes causal cones, geodesics, and unique evolution up to isomorphism	(Isenberg, 2016, Halvorson et al., 7 Mar 2025)
MBQC	Robust determinism of measurement patterns via graphical flow conditions	(Mhalla et al., 2022)
DLM evaluation	Aggregate metrics create an illusion of stability	(Fang et al., 15 Apr 2026)
Prosody and music	Compliance with formal metrical constraints or hierarchical meter	(Hu et al., 2023, Jiang et al., 2022, Rama et al., 2010)

This plurality is not merely terminological. In some domains, metrical determinism is a positive theorem about uniqueness or controllability. In others, it is a warning that the appearance of determinism is an artefact of representation or aggregation.

2. Online computation: metrical determinism in metrical task systems

In metrical task systems (MTS), the basic model is a finite metric space $(X,d)$ with $|X|=n$, an initial state $x_0$, and a sequence of tasks $\tau_t$ with service-cost vectors $c_t:X\to\mathbb{R}_{\ge 0}$. If the online algorithm moves from $x_{t-1}$ to $x_t$, then its cumulative cost is

$$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$

An algorithm is $\alpha$-competitive if

$(X,d)$0

with the randomized version interpreted in expectation over the random seed (Cosson et al., 2024).

Within this literature, metrical determinism means designing online algorithms that use very little randomness—ideally a constant number of bits independent of the request length $(X,d)$1—while preserving the competitive guarantees of fully randomized algorithms. The deterministic baseline is $(X,d)$2; in fact, on any $(X,d)$3-point metric, the optimal deterministic competitive ratio is $(X,d)$4. By contrast, on general metrics there is a randomized algorithm with competitive ratio $(X,d)$5, and that rate is matched by an $(X,d)$6 lower bound (Cosson et al., 2024).

The main barely-random reduction states that any $(X,d)$7-competitive randomized MTS algorithm on an $(X,d)$8-point metric can be converted into a randomized algorithm using only $(X,d)$9 random bits, with competitive ratio at most $|X|=n$0 and additive constant $|X|=n$1. Equivalently, there is an $|X|=n$2-competitive algorithm using at most $|X|=n$3 random bits. The proof passes through a fractional formulation using distributions $|X|=n$4, movement cost $|X|=n$5, a discretization to $|X|=n$6 with $|X|=n$7, and a potential-based rounding rule

$|X|=n$8

where

$|X|=n$9

This yields a factor-$x_0$0 loss in both service and movement up to an additive constant (Cosson et al., 2024).

The same framework yields a collective interpretation. A team of $x_0$1 agents incurs average cost

$x_0$2

and for $x_0$3 there exists a deterministic collective algorithm with $x_0$4-competitiveness, while a single deterministic agent is exactly $x_0$5-competitive. The results also imply a deterministic algorithm using $x_0$6 advice bits with $x_0$7-competitiveness, and show that $x_0$8 random bits are near-optimal because any $x_0$9-competitive algorithm requires at least $\tau_t$0 random bits (Cosson et al., 2024).

Related work shows that the deterministic–randomized gap depends sharply on model details. Under parametrization by the number $\tau_t$1 of distinct requests, deterministic MTS on a uniform metric has competitive ratio $\tau_t$2, while the randomized ratio remains $\tau_t$3 for all $\tau_t$4; on a two-level HST, even $\tau_t$5 gives $\tau_t$6 and $\tau_t$7. For metrical service systems, both deterministic and randomized algorithms can exhibit gain for small $\tau_t$8, and on some families determinism is essentially as powerful as randomness (Bubeck et al., 2019). In metrical service systems with multiple servers, deterministic online algorithms face large combinatorial lower bounds, while on uniform metrics randomization achieves $\tau_t$9 against a deterministic lower bound of $c_t:X\to\mathbb{R}_{\ge 0}$0 (Chiplunkar et al., 2012).

3. Differential geometry: metric determination, metrical distortion, and revised Gauss lemmas

In differential geometry, metrical determinism appears in a substantially different sense. The paper on metrical distortion defines a canonical point-set association

$c_t:X\to\mathbb{R}_{\ge 0}$1

such that $c_t:X\to\mathbb{R}_{\ge 0}$2 is an isometry and “actually induces the given Riemannian geometry.” The governing geometric PDE is

$c_t:X\to\mathbb{R}_{\ge 0}$3

where $c_t:X\to\mathbb{R}_{\ge 0}$4 is the inner differential $c_t:X\to\mathbb{R}_{\ge 0}$5, and in coordinates the metric satisfies

$c_t:X\to\mathbb{R}_{\ge 0}$6

(Voellinger, 6 Mar 2026).

The paper’s central contrast is between the exponential map $c_t:X\to\mathbb{R}_{\ge 0}$7 and the metrical distortion $c_t:X\to\mathbb{R}_{\ge 0}$8. The classical exponential map is characterized by geodesically radial length preservation; the revised construction $c_t:X\to\mathbb{R}_{\ge 0}$9 is characterized by geodesically radial volume preservation. The classical Gauss lemma states that along $x_{t-1}$0,

$x_{t-1}$1

preserving radial lengths and orthogonality. The revised theorem asserts that there is a unique geodesically radial, infinitesimally volume-preserving mapping $x_{t-1}$2, and in dimension $x_{t-1}$3 this volume preservation is global (Voellinger, 6 Mar 2026).

A distinctive technical feature is the “differential slip,” a scalar gauge reparametrization connecting the affine parameter $x_{t-1}$4 in $x_{t-1}$5 with the Riemannian arc-length parameter $x_{t-1}$6 on $x_{t-1}$7. In the radial setting,

$x_{t-1}$8

and the exterior differential is defined by covariant gradient transport,

$x_{t-1}$9

rather than by an ordinary inner differential. This makes metrical determinism a statement not only about the metric tensor but about the induced isometry, its PDE, and the gauge factor connecting synthetical and Riemannian coordinates (Voellinger, 6 Mar 2026).

The two-sphere furnishes the worked example. For the exponential map at the north pole, $x_t$0 and the geodesic polar metric is

$x_t$1

For the metrical distortion,

$x_t$2

and the slip factor becomes

$x_t$3

The article’s notion of metrical determinism is therefore explicitly constructive: the metric determines $x_t$4, $x_t$5 determines the PDE, and the PDE determines the unique geodesically radial, volume-preserving isometry $x_t$6 (Voellinger, 6 Mar 2026).

4. Relativity and formal theories: causal structure, uniqueness, and model invariance

In general relativity, metrical determinism is tied to the causal structure induced by the Lorentzian metric $x_t$7. The metric determines timelike, null, and spacelike directions; the geodesic equation

$x_t$8

and the null condition

$x_t$9

fix free fall and characteristic propagation. Given suitable initial data on a spacelike hypersurface $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$0, determinism means that the evolution of geometry and matter is uniquely fixed within the domain of dependence $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$1. Global hyperbolicity secures this determinism; Cauchy horizons and closed timelike curves undermine it. The Taub region of Taub–NUT is globally hyperbolic but extendible in multiple inequivalent ways across a Cauchy horizon; the Gödel universe contains closed timelike curves through every point and has no globally hyperbolic region at all (Isenberg, 2016).

The formal reconstruction in contemporary philosophy of physics recasts the issue in model-theoretic terms. Determinism is treated as a property of a theory’s models and morphisms, not as a possible-worlds slogan. For metric theories, “agreement” is isomorphism of initial segments, and the strongest criterion is Belot’s $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$2: any isomorphism $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$3 of admissible initial segments extends uniquely to an isomorphism $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$4 of total models. In metrically specialized form,

$$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$5

On this account, general relativity, formulated over globally hyperbolic Lorentzian spacetimes with metric-preserving morphisms, is $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$6-deterministic; the hole argument arises only after enriching the theory with haecceitistic structure it does not itself posit (Halvorson et al., 7 Mar 2025).

A different but closely related issue concerns counterfactuals. Because the metric in GR is dynamical, removing matter or altering sources does not leave a fixed background against which alternative scenarios can be compared. There is no canonical identification of “the same spacetime point” across distinct solutions, and vacuum solutions are not uniquely fixed by local source changes. In Curiel’s analysis, this is why counterfactuals involving metrical structure are peculiarly difficult in GR: the very feature that makes the metric “determined by physics” also deprives one of a canonical cross-model comparison scheme (Curiel, 2015).

Several broader physical proposals extend the notion beyond standard spacetime determinism. A model-invariance program argues that causal and metric structure, rather than deterministic or stochastic representation as such, is what survives empirically equivalent reformulations and therefore merits ontological commitment (Nolland, 27 Dec 2025). In invariant set theory, metrical determinism is determinism enforced by the metric and geometric constraints of a measure-zero fractal invariant set in state space (Palmer, 2013). In metric dynamics, the claim is stronger still: all observed dynamics can be represented as geodesic motion in a suitably chosen anisotropic metric space, so that “force” becomes an auxiliary notion derived from geodesic acceleration (Siparov, 2015). These views differ sharply in ontology, but all assign determinative power to metric or geometric structure rather than to an independent force law or stochastic rule.

5. Measurement-based quantum computing: robust determinism and Shadow Pauli Flow

In measurement-based quantum computing (MBQC), metrical determinism is robust determinism: the implemented quantum map is independent of random measurement outcomes, uniform in the measurement angles, and stepwise, so that every partial computation is also deterministic. The formal setting is an open graph $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$7 with graph-state entanglement, measurements $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$8, and classical Pauli corrections $$\mathrm{cost}_A(\sigma)=\sum_{t=1}^T \big(c_t(x_t)+d(x_{t-1},x_t)\big).$$9, $\alpha$0 conditioned on outcomes $\alpha$1 (Mhalla et al., 2022).

For plane-only measurements, Generalized Flow (GFlow) characterizes robust determinism. With Pauli measurements present, GFlow ceases to be necessary because some apparent back-action on already measured qubits is harmless when those qubits were measured in the corresponding Pauli basis. Pauli Flow (PF) extends GFlow and guarantees robust determinism, but PF is only complete in a weaker sense: an open graph can drive a deterministic computation if and only if it has a Pauli Flow, yet a specific correction strategy and measurement order need not be reflected by PF (Mhalla et al., 2022).

The paper’s main advance is Shadow Pauli Flow (SPF). Using the correction and impact sets

$\alpha$2

$\alpha$3

SPF allows regulated anachronistic impacts provided that they are absorbed by shadow correctors supported entirely in the past. The main theorem is exact: an MBQC is robustly deterministic if and only if its correction strategy is consistent with a Shadow Pauli Flow. Moreover, SPF can be computed in polynomial time (Mhalla et al., 2022).

This result separates two notions of determinism. PF is a resource-state criterion: some deterministic computation exists. SPF is a strategy-level criterion: a particular order, depth, and correction schedule is robustly deterministic. In that sense, SPF plays for MBQC a role analogous to unique-extension criteria in metric theories: it characterizes not only existence of a deterministic semantics but compatibility with the concrete operational structure.

6. Evaluation metrics: aggregation-induced determinism in diffusion LLMs

In diffusion LLMs (DLMs), metrical determinism denotes the illusion of stability created when evaluation collapses sample-level variability into a single dataset-level score. Let $\alpha$4 and let $\alpha$5 indicate whether configuration $\alpha$6 is correct on sample $\alpha$7. Dataset-level accuracy is

$\alpha$8

and dataset-level variability is

$\alpha$9

whereas sample-level non-determinism is

$(X,d)$00

Because

$(X,d)$01

weak dependence makes the aggregate variance shrink with $(X,d)$02, thereby masking structured instability (Fang et al., 15 Apr 2026).

The paper also defines a per-sample flip rate

$(X,d)$03

and introduces Factor Variance Attribution (FVA) to decompose variability across evaluation factors. With factor means $(X,d)$04, grand mean $(X,d)$05, between-factor and within-factor mean squares $(X,d)$06 and $(X,d)$07, FVA is

$(X,d)$08

FVA close to $(X,d)$09 indicates between-factor dominance; FVA near $(X,d)$10 indicates strong within-factor sensitivity (Fang et al., 15 Apr 2026).

Empirically, dataset-level metrics can be nearly constant while sample-level variability is large. For LLaDA on question answering, the precision factor yields PIQA accuracy $(X,d)$11 with dataset-level standard deviation $(X,d)$12 but sample-level standard deviation $(X,d)$13; ARC-Challenge gives $(X,d)$14, $(X,d)$15, and $(X,d)$16; WinoGrande gives $(X,d)$17, $(X,d)$18, and $(X,d)$19. Code generation is more sensitive: for LLaDA on HumanEval, varying diffusion steps gives dataset-level pass@1 standard deviation $(X,d)$20 and sample-level standard deviation $(X,d)$21. Pooled FVA is $(X,d)$22 for LLaDA and $(X,d)$23 for LLaDA-1.5, indicating dominant between-factor effects but still nontrivial within-factor sensitivity (Fang et al., 15 Apr 2026).

Here metrical determinism is not a property of the model but a pathology of evaluation design. Stable aggregate scores do not imply stable behavior. The recommended remedy is factor-aware reporting: sample-level variance, flip rates, factor settings, and FVA should accompany dataset-level metrics (Fang et al., 15 Apr 2026).

7. Prosody, music, and controllable generation

In controllable poetry generation, metrical determinism refers to the degree to which generated outputs satisfy prosodic and formal constraints across samples. PoetryDiffusion implements a soft, penalty-based version. A diffusion generator handles semantics, while an independently trained metrical controller imposes

$(X,d)$24

with the tone term omitted for sonnets. The controller operates throughout denoising rather than by hard projection or repair. This yields effectively deterministic format control for sonnets—Format $(X,d)$25—with Rhyme $(X,d)$26, and on SongCi it yields Format $(X,d)$27, Tone $(X,d)$28, and Rhyme $(X,d)$29. The limitation is explicit: the system does not enforce English stress patterns or syllable counts, so its metrical determinism for sonnets is limited to line count and rhyme scheme (Hu et al., 2023).

In self-supervised music analysis, metrical determinism means the extent to which hierarchical meter is determined by signal and constraints. The model predicts an eight-layer binary metrical tree from beat to section level using a CRF over joint hierarchical states. Hard regularity up to measures is implemented by setting $(X,d)$30 for $(X,d)$31, while higher levels use finite penalties. The resulting system is nearly deterministic up to the measure level under reliable beat alignment and strong binary regularity, but determinacy becomes contingent at hypermetrical levels. Reported downbeat detection reaches $(X,d)$32 F1 on MIDI and $(X,d)$33 on audio; performance at level $(X,d)$34 is poor because stable binary regularity at $(X,d)$35 measures is rare in popular music (Jiang et al., 2022).

In computational Sanskrit prosody, a verse is metrically deterministic when a fixed rule set for syllabification, guru/laghu assignment, sandhi, and the canonical catalogue of metres yields a unique metrical class. The classifier scans the transliterated text in $(X,d)$36, computes a binary guru/laghu pattern, represents varṇa metres by gaṇa sequences, and uses hash-based lookup for sama, ardhasama, and viṣama classes. Ambiguity is confined to recognized optionalities such as pādānta-guru or special consonant clusters; mandatory sandhi correction reduces spurious alternatives. In this setting, metrical determinism is literal classification uniqueness under an explicit phonological rule system (Rama et al., 2010).

Across these prosodic and musical settings, a recurrent distinction emerges between hard determinacy and guided compliance. PoetryDiffusion achieves near-deterministic adherence for some constraints but retains probabilistic guidance for others (Hu et al., 2023). Hierarchical music analysis is near-deterministic at lower levels and probabilistic at higher ones (Jiang et al., 2022). Sanskrit verse classification is deterministic when the rule set and input normalization close off optionality (Rama et al., 2010).

Metrical determinism is therefore best understood as a domain-dependent concept rather than a single theory. In online algorithms it concerns the reduction of randomness without forfeiting competitive guarantees; in geometry and relativity it concerns what the metric fixes, locally or globally, about evolution and structure; in MBQC it concerns correction strategies that neutralize measurement randomness; in evaluation methodology it names a false appearance of stability produced by aggregation; and in prosody and music it concerns the extent to which formal meter determines admissible outputs or analyses. What unifies these uses is not a shared formalism but a shared question: when does a metric, metrical, or measurement structure make behavior genuinely determinate, and when does it merely make it appear so?