Compute-Equivalent Gain in Compute-Constrained Systems

• Compute-equivalent gain is defined as a metric that evaluates performance improvements after accounting for the compute cost required to achieve them.
• It formalizes trade-offs using compute-aware scoring, where training gain must exceed selection FLOPs, with parameters like λ capturing compute efficiency.
• The concept underpins diverse applications from data selection to metamaterial design and calibration, emphasizing effective performance relative to resource use.

Searching arXiv for the cited papers and closely related terminology. Compute-equivalent gain denotes a family of equivalence constructions in which an observed gain, amplification, or performance improvement is evaluated only after the cost, scaling, or transformation needed to obtain it has been accounted for. In the most explicit compute-constrained formulation, the term is not introduced as a named metric, but the underlying idea is formalized as a trade-off between selection FLOPs and downstream training performance: a method is useful only if its training gain exceeds the extra compute required to realize it (Yin et al., 2024). Closely related equivalence notions appear in temporal metamaterials, where a temporally multistepped gain/loss modulation is replaced by a single effective temporal step (Pacheco-Peña et al., 2021), in optical-fiber amplifier simulation, where a long amplifier is replaced by an equivalent short fiber with scaled gain terms (Drake et al., 2019), and in stochastic photomultiplication, where only a local operating-point-dependent effective gain remains meaningful (Armin, 8 Jun 2026). This suggests that compute-equivalent gain is best understood not as a universal scalar, but as an effective gain defined relative to a conserved observable, a resource budget, or a transformed model.

1. Conceptual scope

Within compute-constrained data selection, the relevant question is not whether a selector improves sample efficiency in isolation, but whether the improvement survives after selection cost is charged against the same total budget. The central formulation augments standard subset selection with a compute constraint,

$$\mathcal{S}^{*} = \argmax_{\mathcal{S} \subseteq \mathcal{D}} P(\mathcal{V}; T(\mathcal{S})) \quad \text{s.t.} \quad C_{T(\mathcal{S})} + \sum_{x \in \mathcal{D}} C_{v(x)} \le K,$$

where $C_{T(\mathcal{S})}$ is training FLOPs on the selected subset, $C_{v(x)}$ is the compute to score example $x$, and $K$ is the total compute budget (Yin et al., 2024). The paper states that its formalization is exactly a utility score discounted by the cost of obtaining it: a method is only useful if its training gain exceeds the extra selection compute required to realize that gain.

The same paper makes the interpretive step most closely aligned with the phrase compute-equivalent gain. It states that compute spent on selection can be “worth it” only if it yields enough additional model performance, and that the fitted parameter $\lambda$ in its compute-performance model captures how much performance is extracted per unit compute. A larger $\lambda$ therefore means the method converts compute into performance more efficiently (Yin et al., 2024).

A broader reading across the cited literature indicates that “equivalent gain” is usually a relational notion. In temporal metamaterials, equivalence means identical wave evolution under a single effective temporal step rather than a multistepped gain/loss pattern (Pacheco-Peña et al., 2021). In optical fibers, equivalence means approximately preserved signal and pump power distributions after shortening the fiber and rescaling gain/coupling (Drake et al., 2019). In trap-assisted photomultiplication, by contrast, the literature argues against any global scalar gain and restricts meaningful equivalence to a local small-signal operating point (Armin, 8 Jun 2026).

2. Formalization in compute-constrained data selection

The compute-constrained account begins from the standard subset-selection problem

$$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$

and then introduces compute-aware scoring by charging both training and selection against a common budget (Yin et al., 2024). In this setting, selection is not free preprocessing; it is part of the optimization problem itself.

To model the trade-off, the paper writes the total compute for choosing $k = |\mathcal{S}|$ points as

$C(k) = c \cdot k + \sum_x C_{v(x)},$

where $C_{T(\mathcal{S})}$0 is the fixed training cost per data point and $C_{T(\mathcal{S})}$1 is the total selection cost. The associated performance model is

$C_{T(\mathcal{S})}$2

with $C_{T(\mathcal{S})}$3 the zero-shot performance, $C_{T(\mathcal{S})}$4 the final upper bound after training on all data, $C_{T(\mathcal{S})}$5 the efficiency with which a method converts compute into performance, and $C_{T(\mathcal{S})}$6 the compute of full training (Yin et al., 2024).

Under this model, the operative equivalence is straightforward. A more expensive selector must behave like adding extra effective training compute; otherwise its apparent advantage under a fixed data budget disappears under a true end-to-end FLOP budget. The paper states this directly: if a method costs more selection compute, it must produce enough training gain to offset that cost; otherwise, a cheaper method or even random selection can be better under the same total FLOP budget (Yin et al., 2024).

This framing also clarifies why the phrase “cost-aware utility” should not be misread as a single predefined scalar. The paper does not define one abstract metric with that name. Rather, its utility is operational: a data point is valuable only if the performance boost it enables is worth the FLOPs spent to find it (Yin et al., 2024).

3. Cost regimes, selectors, and compute-optimality

The empirical analysis divides selection methods by cost and shows that compute-optimality changes with budget. The approximate selection FLOPs and rough forward-cost scalings reported in the paper are as follows (Yin et al., 2024).

Method	Approximate selection FLOPs	Rough forward-cost scaling
BM25	$C_{T(\mathcal{S})}$7	$C_{T(\mathcal{S})}$8
Embed	$C_{T(\mathcal{S})}$9	$C_{v(x)}$0
PPL	$C_{v(x)}$1	$C_{v(x)}$2
LESS	$C_{v(x)}$3	$C_{v(x)}$4

For gradient utility, the paper defines

$C_{v(x)}$5

with cost approximately

$C_{v(x)}$6

This makes explicit why gradient-based methods can perform well under a fixed selected-data budget yet fail to be compute-optimal under an end-to-end budget (Yin et al., 2024).

The paper’s principal empirical conclusion is that “many powerful data selection methods are almost never compute-optimal” and that “cheaper data selection alternatives dominate both from a theoretical and empirical perspective” (Yin et al., 2024). At small and medium compute budgets, BM25 and embedding-based methods dominate. At larger compute budgets, more expensive methods can become worthwhile: perplexity-based selection starts to become competitive, and gradient-based selection becomes optimal only when training compute is very large relative to selection compute.

The paper gives two concrete extrapolated thresholds. Perplexity-based selection becomes compute-optimal when the training model is about $C_{v(x)}$7 larger than the selection model, while gradient-based selection becomes compute-optimal when the training model is about $C_{v(x)}$8 larger than the selection model (Yin et al., 2024). It also states that, for 70B model size, PPL and LESS can outperform cheaper methods and become compute-optimal. By contrast, if selection cost is ignored and only training-budget performance is compared, LESS consistently outperforms other methods and PPL is often second-best (Yin et al., 2024). The contrast is central: a method can be strong under a fixed-data budget but poor under a true end-to-end compute budget.

4. Equivalent-gain constructions in wave and amplifier systems

A physically explicit equivalent-gain construction appears in temporal metamaterials with gain and loss. The system is a spatially unbounded, nonmagnetic medium with time-dependent relative permittivity

$C_{v(x)}$9

where $x$0 denotes loss, $x$1 denotes gain, and the real part is held fixed under a dispersionless approximation (Pacheco-Peña et al., 2021). Over a short total modulation time $x$2, the medium alternates between $x$3 and $x$4, with temporal filling factors $x$5. The multistepped temporal medium is then replaced by a single effective temporal step with

$x$6

This effective permittivity determines whether the equivalent medium is lossy, transparent, or gain-like (Pacheco-Peña et al., 2021).

The temporal-boundary physics is distinctive. Across a sudden temporal discontinuity $x$7, the wavenumber $x$8 is preserved while the frequency changes according to

$x$9

Because the effective permittivity can be complex, the effective frequency

$K$0

can also be complex, so wave amplitude decays for $K$1 and grows for $K$2 while $K$3 remains unchanged (Pacheco-Peña et al., 2021). The paper’s examples include equal-filling-factor cases with effective imaginary parts $K$4, $K$5, $K$6, $K$7, and $K$8, as well as duty-cycle tuning from a lossy medium at DC $K$9 with $\lambda$0 to a gain medium at DC $\lambda$1 with $\lambda$2 (Pacheco-Peña et al., 2021).

A computationally oriented analogue appears in optical-fiber amplifier simulation. There, a long fiber of length $\lambda$3 is replaced by a much shorter fiber of length $\lambda$4, while gain and coupling are scaled by $\lambda$5 so that power evolution is approximately preserved (Drake et al., 2019). The exact change of variables $\lambda$6 is mathematically equivalent but does not reduce cost because the phase factors oscillate faster by the same scaling factor. The computationally useful model therefore keeps the original beat-length oscillations while scaling the gain terms, yielding an equivalent short-fiber model whose speedup is approximately

$\lambda$7

For the example $\lambda$8 m and $\lambda$9 m, the observed speedup is about $\lambda$0 (Drake et al., 2019). This is a direct instance in which equivalence is engineered to preserve experimentally relevant observables—signal and pump power distributions—while substantially reducing computation.

5. Local effective gain and the limits of scalar descriptors

A recurring limitation across gain literatures is that an apparent gain measured at one operating point need not be a transferable descriptor. The strongest statement of this point is made for trap-assisted photomultiplication. That paper distinguishes apparent quantum efficiency, internal gain, chord gain, tangential or local small-signal gain, and effective gain, and argues that a single chord gain is not a meaningful device descriptor because the current–illumination relation is nonlinear (Armin, 8 Jun 2026). Its effective gain is

$\lambda$1

which is the mean number of carriers injected and collected per trap filling event before relaxation. The quantity is operating-point-dependent and self-limited, with

$\lambda$2

Only the local differential gain enters responsivity and detectivity (Armin, 8 Jun 2026).

The same paper shows why this matters physically. The injection process that amplifies current also accelerates relaxation of the gain-enabling trapped state, producing an inherently nonlinear response. After linearization,

$\lambda$3

so the mechanism that produces amplification simultaneously increases the relaxation rate (Armin, 8 Jun 2026). The paper further derives a strictly non-negative fluctuation penalty and concludes that gain cannot exceed the intrinsic detectivity limit of the underlying unity-gain photodiode: $\lambda$4 This makes explicit that effective gain is inseparable from the dissipative dynamics that sustain it.

A related but distinct usage appears in up-the-ramp detector calibration. There, the electronic gain $\lambda$5 is the conversion between accumulated photoelectrons and recorded digital numbers, with units $\lambda$6, and the covariance of adjacent read differences is modeled as

$\lambda$7

The paper does not introduce a separate named “compute-equivalent gain” metric, but it states that the likelihood-based $\lambda$8 is the gain parameter that makes the measured correlated read-difference statistics consistent with Poisson photon statistics plus Gaussian read noise (Brandt, 9 Dec 2025). This suggests a narrower equivalence notion: a computed gain can be called equivalent when it reconciles the statistical measurement model with the physical electron-counting model.

6. Algorithmic throughput, adjacent meanings, and terminological boundaries

In computational radio astronomy, gain equivalence appears not as physical amplification but as a reduction in the computational cost of achieving the same calibration quality. StEFCal recasts the standard least-squares gain-calibration problem

$\lambda$9

through an alternating direction implicit update that reduces per-iteration complexity from $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$0 for traditional LM/WALS methods to $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$1 (Salvini et al., 2014). The paper gives a scalar flop count of about $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$2 per iteration, reducible to about $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$3 by reusing dot products, and reports that the algorithm performs at or close to the Cramér–Rao bound. It also reports concrete pipeline gains: a reduction from about $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$4 s to $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$5 s per main iteration in the AARTFAAC WALS pipeline, an additional factor of about $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$6 when previous timeslice gains are reused as initial guesses, an overall improvement of more than $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$7 relative to the original setup, and a factor of $$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$8–$$\mathcal{S}^{*} = \arg\max_{\mathcal{S}\subseteq \mathcal{D}} P(\mathcal{T}; T(\mathcal{S})) \quad \text{s.t.} \quad |\mathcal{S}| \le K,$$9 speedup in LOFAR preprocessing while yielding practically identical calibration results (Salvini et al., 2014). A plausible implication is that, in algorithmic settings, compute-equivalent gain refers to calibration throughput gained at fixed statistical quality.

The literature also contains equivalence statements involving gain that are conceptually important but not compute-equivalent in the budgeted sense. In a gain-driven spin system, a ferromagnet with gain is mapped to an antiferromagnet with equal loss under the global spin inversion

$k = |\mathcal{S}|$0

which turns the ferromagnetic exchange term into an antiferromagnetic one and swaps gain/loss character in the LLG dynamics (Yang et al., 2018). This is an exact dynamical mapping, not a cost-adjusted performance equivalence.

Likewise, “potential gain” in graph theory is a navigability centrality rather than a resource-accounted gain measure. For the geometric variant,

$k = |\mathcal{S}|$1

and the paper proves that this is equivalent to degree centrality applied to Katz centrality, or, in its own interpretive language, “popularity $k = |\mathcal{S}|$2 similarity” (Meo et al., 2020). The term is therefore terminologically adjacent but substantively distinct.

Taken together, these boundaries are useful because they prevent a common misconception: not every “gain” accompanied by an “equivalence” relation is a compute-equivalent gain. In the compute-constrained sense, the defining feature is that the gain has been renormalized by the resources required to obtain it. Where that renormalization is absent, the literature generally speaks instead of effective gain, equivalent gain, dynamical mapping, or composite centrality, each with its own conserved quantity and interpretation.