acceptodds
Under review as a conference paper at ICLR 2027

Synergistic Learning in Domain Adaptation with a Known Link: A Minimax Understanding

Abstract

When can combining source and target data yield a much better performance than using either one alone? Statistically, we address this issue from a minimax viewpoint: writing for the minimax risk attainable from source and target observations, we ask when decays *strictly faster* than both and . When this happens, we call it the *Synergistic Learning Phenomenon* (SLP) and quantify it by the *Synergistic Advantage Ratio* . In this work, we focus on domain adaptation with a *known link*: the source regression function is , the target regression function is for a known operator , and the two domains may also differ in their designs. Three links—the identity, integration, and differentiation—together with the behaviour of the density ratio determine whether synergy is possible, at what rate, and in which sample-size window. For the identity link (covariate shift) with a Beta source and a uniform target, SLP is possible only if , where is the smoothness of (Result I; Zhou et al., 2022). An integral link instead can yield the joint rate even for a bounded ratio (Result II), and under a Beta source it has a phase diagram with cuts at and : integral SLP for , an extra level-split rate for , and covariate-shift-type spatial SLP for (Result III). A derivative link synergizes only under a stronger singularity (Result IV), giving a sharp directional asymmetry under a bounded ratio. Across all three links, synergy arises from *role separation*: each dataset resolves a distinct bottleneck associated with the link and the design. We characterize the sample-size windows in which combining source and target data improves on both individual minimax rates, and how the SAR varies with the source and target sample sizes. Simulations support the predicted phase structure, and a mirrored integration–differentiation analysis of real vehicle and smartphone signals illustrates the directional asymmetry.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Synergistic Learning, Domain Adaptation, Transfer Learning, Functional Links, Minimax Rates

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qmuzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies the synergistic learning phenomenon (SLP) in nonparametric domain adaptation when the source and target regression functions are related by a known operator, g=Afg=Af. Synergy is characterized through the synergistic advantage ratio (SAR), which compares the joint minimax risk with the better of the two individual minimax risks. The paper considers three operators—identity, integration, and differentiation—under both bounded density ratios and unbounded density ratios induced by Beta source distributions.

The main theoretical contributions concern the integration and differentiation links. For integration, the authors show that SLP can occur even under bounded density ratios and establish minimax rates and sample-size regimes under singular source distributions. For differentiation, they establish a stronger singularity requirement for SLP and show that no asymptotic rate improvement is possible under bounded density ratios. These results are supported by matching minimax upper and lower bounds.

The theoretical results are further illustrated by numerical simulations and experiments on vehicle yaw-rate and heading data, which examine the predicted synergy regimes and the asymmetry between integration and differentiation.

Critical strengths
  1. Rigorous minimax characterization across different sampling regimes. The paper establishes matching minimax upper and lower bounds for integration and differentiation links under bounded density ratios and singular source distributions. It characterizes the sample-size regimes in which synergistic learning occurs and identifies how the optimal convergence rates and phase transitions depend on the sampling distributions.

  2. Interesting distinction between integration and differentiation. The paper shows that integration can produce synergistic learning even under bounded density ratios, whereas differentiation cannot. The theoretical analysis explains this asymmetry through the different statistical information provided by the two observation channels. This provides a clearer understanding of how functional relationships, beyond distribution shift alone, influence the possibility of synergistic learning.

Critical weaknesses
  1. The interpretation of SAR as a measure of synergy deserves further justification. SAR compares joint estimation using both datasets against the better single-domain estimator with its original sample size. While this is a mathematically valid comparison, it does not necessarily establish that combining the two observation channels is more statistically efficient than an optimal allocation of the same total sampling budget. For example, in the bounded-density integration setting, the paper establishes a diverging SAR when nT≫nn_T\gg n, although allocating all N=n+nTN=n+n_T observations to the source channel would achieve the faster rate N−1N^{-1}. Thus, SAR can diverge even when joint data collection is not optimal under an equal-cost budget. Whether SAR captures the benefit of jointly using two already available datasets or a more general statistical advantage of combining observation channels remains an important question regarding the interpretation of synergy.

  2. The generality and robustness of the results with respect to the functional link. The analysis is restricted to three exactly specified operators: identity, integration, and differentiation. It remains unclear whether the identified synergy mechanisms extend to more general operators and whether small perturbations or misspecification of the known operator AA could substantially affect the minimax rates or even the existence of SLP. The sensitivity of the theoretical conclusions to the functional relationship is an important concern regarding the general applicability.

Questions for authors
  1. Interpretation of SAR. In the bounded-density integration setting, SAR can diverge even when allocating the same total sampling budget entirely to the source domain achieves a faster rate. Does this limit the interpretation of SAR as a measure of synergy between observation channels?

  2. Relevance of the sample-size regimes. The introduction states that “Labeled target observations are often scarce while a related source domain supplies a far larger sample at lower cost”. However, the bounded-density integration result requires nT≫nn_T\gg n for SLP. How does this result relate to the motivating scenario?

  3. Robustness of the functional link. How sensitive are the minimax rates and SLP conclusions to small perturbations or misspecification of the assumed operator AA? Could such deviations substantially affect the observed synergy in the real-data experiments?

1 comment
AuthorsOpenReview ↗

Response to Reviewer qmuz

Thank you. We clarify SAR, sample-size regimes, and conditional robustness. The main substantive revisions are highlighted in cyan in the revised manuscript.

W1/Q1: SAR and budget allocation. Section 3 clarifies SAR's scope and cites related work; we elaborate here. The joint, source-only and target-only minimax risks optimize over estimators for the same target, linked class and L2(hT)L_2(h_T) loss: SAR(n,nT)=RS(n)∧RT(nT)RM(n,nT).\mathrm{SAR}(n,n_T)=\frac{R_S(n)\wedge R_T(n_T)}{R_M(n,n_T)}. Strict SLP means this ratio diverges as both counts grow: removing either channel, at unchanged remaining counts, loses the joint minimax rate. This is target-dependent complementarity. The same function and derivative observations yield SLP for the primitive but not its derivative (Appendix C.13; Section 4.2).

Budget allocation also optimizes counts: inf⁡m,mT≥0:,cSm+cTmT≤BRM(m,mT).\inf_{m,m_T\ge0:\\,c_Sm+c_Tm_T\le B}R_M(m,m_T). Here cS,cTc_S,c_T are costs and BB the budget; RMR_M already optimizes estimators. Harding et al. (2026) study budgeted population-mean estimation with known group compositions and shared group-conditional responses. Their effective-sample-size criterion balances quantity, coverage and cost; our operator analysis asks when channels complement each other at fixed counts.

Here an equal-cost bounded-design integral example distinguishes them. For N=n+nTN=n+n_T, all-source risk N−1N^{-1} beats mixed risk (nnT)−1/2(nn_T)^{-1/2} in the SLP window. Conversely, for n≫nTn\gg n_T, mixed risk matches the all-source order N−1N^{-1} while SAR remains of constant order. Thus strict SLP neither guarantees budget efficiency nor is required to match that order; constant-factor gains can still matter.

Q2: Relevance of the target-rich window. Sections 1, 4.1 and 5 now broaden the original source-rich motivation. Source/target roles specify the observed functions, not sample counts: direct rate measurements can be sparse while cumulative state logs are abundant. In Result II, target residual correction reduces the random error of an unbiased source pilot. At bandwidth bb, balancing b/n+1/(nTb)b/n+1/(n_Tb) gives (nnT)−1/2(nn_T)^{-1/2}. The full SLP window is n≪nT≪n(2β+3)/(2β+1).n\ll n_T\ll n^{(2\beta+3)/(2\beta+1)}. The lower endpoint requires enough targets to beat the source risk n−1n^{-1}; the upper endpoint is where the target-only rate nT−2(β+1)/(2β+3)n_T^{-2(\beta+1)/(2\beta+3)} matches the joint order. These boundaries follow from matching minimax bounds. Source-rich SLP is also possible under sufficiently singular source designs (Results I, III and IV).

W2/Q3: Generality and robustness. Identity, integration and differentiation preserve, raise and lower smoothness. They illustrate distinct information balances, but Section 5 gives no universal theorem for other operators: localization, nullspaces, boundaries and designs need separate analysis.

Appendix D.2 addresses robustness for bounded-design integration. For gr=Af+rg_r=\mathcal A f+r, residual smoothing gives Rr≤2R0+2Br,Br=E∣Sbr−r∣L2(hT)2.R_r\le2R_0+2B_r,\qquad B_r=\mathbb E\\|S_b r-r\\|_{L_2(h_T)}^2. Here R0,RrR_0,R_r are exact/perturbed estimator risks and SbS_b averages residuals (zero on empty windows). Uniformly over allowed functions and discrepancies, Br=O(R∗)B_r=O(R^{\ast}), R∗=(nnT)−1/2R^{\ast}=(nn_T)^{-1/2}, preserves the upper rate in the SLP window. This does not cover every fixed small error.

For fixed-radius residuals in H(γ,κr)\mathcal H(\gamma,\kappa_r) with γ>β+1\gamma > \beta+1, Proposition 14 yields this upper rate on the restricted subwindow in Corollary 15 using a local-polynomial correction with Gram-matrix and count stability checks. Under Gaussian noise, Corollary 15 gives the matching rate and strict SLP for the same pair class in all three risks, satisfying P0=(f,0):f∈H(β,κ)⊆P⊆H(β,κ)×H(γ,κr).\mathcal P_0=\\{(f,0):f\in\mathcal H(\beta,\kappa)\\}\subseteq\mathcal P\subseteq\mathcal H(\beta,\kappa)\times\mathcal H(\gamma,\kappa_r). Because P\mathcal P contains P0\mathcal P_0, the exact-link minimax lower bounds remain valid. A fixed misspecified operator graph (f,(A⋆−A)f)\\{(f,(\mathcal A_\star-\mathcal A)f)\\} need not contain it; even with uniformly smooth residuals, the upper bound alone does not imply SLP.

Table 4 (Appendix F.1) tests the existing empirical estimator on selected audited windows; intervals exclude selection uncertainty. At tested scale changes within ±10\pm10\\%, geometric SAR remains above one (1.251–1.391), while ratio-of-mean-risk intervals include one. At +20+20\\%, both point estimates fall below one (0.988 and 0.884), with intervals straddling one. Larger positive mismatch can therefore erode observed gains.

These sensitivity checks do not directly validate Corollary 15: scaling changes the residual by −δ∫0⋅f-\delta\int_0^\cdot f, guaranteed only β+1\beta+1 smoothness under our class assumptions, whereas the corollary requires γ>β+1\gamma > \beta+1. The empirical estimator also differs from its stable local-polynomial correction.

Reviewer_81qNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies when combining source and target observations yields a minimax convergence rate strictly faster than either dataset alone. It provides theoretical analysis on this.

Critical strengths

The paper is mathematically strong and presents sufficient theoretical analysis.

Critical weaknesses
  1. The connection between the theoretical setting and the motivating domain adaptation problem is insufficiently developed. The analysis assumes an exactly known functional relationship, and several constructions additionally require knowledge of the source density and smoothness parameters. These assumptions define a legitimate theoretical problem, but the paper should explain more clearly which practical transfer problems they capture and what insight the results provide for those problems.

  2. The novelty is not adequately positioned relative to joint estimation from function and derivative observations. The related-work discussion emphasizes covariate shift and single-experiment inverse problems, while overlooking directly relevant work such as Hall and Yatchew’s Nonparametric Estimation When Data on Derivatives Are Available and Dai’s Nonparametric Estimation via Partial Derivatives.

  3. The presentation places substantial emphasis on technical case distinctions without sufficiently developing the underlying insight. Precise minimax results naturally require mathematical detail. However, the main text repeatedly subdivides regimes even when the joint rate remains unchanged, as in the SL and TL branches of Theorem 1. This makes it difficult to distinguish a change in the estimation mechanism from a change in which individual baseline is better.

  4. The experiments provide limited independent validation of the claimed phase structure. Also, I would appreciate it if the authors validate the method on the widely used domain adaptation benchmarks.

Questions for authors

See my weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 81qN

Thank you. We clarified the scope and novelty, reorganized the findings, and separated empirical checks from asymptotic conclusions. The main substantive revisions are highlighted in cyan in the revised manuscript.

W1. Practical scope, insights, and assumptions

Section 1 now identifies structured transfer problems with a prescribed link g=Afg=\mathcal{A}f: yaw rate/heading, cost/marginal-demand systems, and simulated responses/gradients. Both functions remain unknown and are observed noisily under different designs and sample sizes. SAR asks whether combining the samples strictly improves on both single-domain minimax rates.

Sections 4-5 establish:

  1. Strict synergy with identical designs. Integration achieves the joint minimax rate (nnT)−1/2(nn_T)^{-1/2} on an intermediate sample-size window, strictly faster than both single-domain rates with identical designs (Theorem 1). Thus the response relationship can supply complementarity without covariate shift.

  2. Directional limits of synergy. When both design densities are bounded above and away from zero, derivative estimation admits no strict minimax rate improvement over the better single-domain rate for any sample-size sequence (Theorem 3). The same observations can yield synergy for the primitive but not its derivative (Appendix C.13): complementarity depends on the target.

  3. Singular designs reshape synergy. For Beta sources with a>2a > 2, target data estimate cumulative contributions poorly resolved by source data in the relevant level-split regimes; source data suffice for the remaining coefficients. For a>3a > 3, the samples can also divide estimation spatially (Section 4.3). Sufficiently strong singularity enables derivative synergy absent under bounded designs (Section 4.4). Design coverage changes both how synergy arises and whether it is possible.

Known (h,a,β)(h,a,\beta) enter weighting, localization, and tuning as an oracle benchmark; fully data-driven adaptation remains open (Section 7). Section 7 and Appendix D.2 give conditional robustness results for controlled link errors or sufficiently smooth errors with local-polynomial residual correction, preserving the Result II risk upper rate on stated sample-size windows.

W2. Prior work and novelty

Section 2 now acknowledges that Hall & Yatchew (2007) and Dai (2025) allow unequal sample sizes. Here we give the rate comparison: Hall & Yatchew express rates through the smallest channel size; Dai (Section 3.4) reports min⁡nT−2m/(2m+1),n−1\min\\{n_T^{-2m/(2m+1)},n^{-1}\\} in one dimension in an SS-ANOVA RKHS, with function data anchoring the integration constant. Our integral model uses a Hölder class and g(0)=0g(0)=0; m=β+1m=\beta+1 aligns smoothness indices, not parameter spaces or anchoring assumptions. We characterize joint and both single-domain minimax risks within this model to determine strict synergy, including reverse derivative targets and singular designs.

W3. Insight and regime presentation

The abstract, introduction, and conclusion follow these findings. Sections 4.1-4.2 place bounded-design integration and differentiation together; Sections 4.3-4.4 then examine singularity. Section 5 explains the mechanisms, and Section 3 defines the regime labels after SAR.

The guide to Section 4 separates a joint-rate change from a switch in the stronger single-domain baseline. Identical-rate SL/TL rows are merged, with the equalizing scales stated separately. Genuine rate changes retain separate branches: at nT≍nn_T\asymp n in Theorem 5 (2<a<32 < a < 3), the dominant contribution changes from target estimation of cumulative coefficients, nT−1n_T^{-1}, to joint estimation of the remaining coefficients, (nnT)−1/2(nn_T)^{-1/2}; both use the Appendix D.3 estimator.

W4. Independent evidence and benchmarks

Figure 2(b) and increasing-sample checks evaluate actual, unswitched joint-estimator risk, without replacing that risk by a minimum with single-domain risks or switching at theoretical boundaries (Appendix E.1). Integration shows an interior gain and declining outer-regime advantage; the derivative control remains near no gain. Beta-design regression and coefficient-model risk checks probe selected mechanisms (Appendices E.2-E.3). The theory establishes complete asymptotic boundaries and class-wide minimax claims.

IO-VNBD supplies synchronized CAN yaw rate and GPS heading with an approximate kinematic link; reversing their roles tests both directions on observed signals. DomainNet and WILDS do not directly supply the prescribed function/derivative observation structure; applying our constructions would require extra modeling. Section 6 and Appendix F.1 check qualitative directional gains using held-out target MSE, with intervals conditional on windows selected by link diagnostics. Since these temporal designs do not isolate an unbounded density ratio, Appendix F.2 uses controlled Beta coverage and latent functions with resampled IO-VNBD noise marginals to examine singularity under empirically calibrated noise.

Reviewer_jK94OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates when combining source and target datasets can achieve a faster estimation rate than using either dataset alone.

The authors assume that the two underlying regression functions satisfy a known relationship g=Afg = A f, and study three operators: identity, integration, and differentiation. They derive minimax convergence rates and characterize the conditions under which a synergistic learning phenomenon occurs. Numerical simulations and a vehicle-sensor experiment are presented to illustrate the theoretical findings.

Critical strengths

The paper addresses a mathematically interesting question about how different types of measurements can provide complementary information.

The comparison between integration and differentiation offers useful intuition into how measurement operators influence learning efficiency. The analysis of different sampling distributions and the resulting convergence-rate regimes provides a potentially valuable theoretical contribution.

Critical weaknesses

My main concern is the highly restrictive assumption that the relationship g=Afg = Af is exactly known in advance.

If my understanding is correct, the integration setting is essentially like combining noisy velocity and displacement measurements while already knowing their exact mathematical relationship. Although meaningful in carefully calibrated physical systems, this assumption seems too idealized for general domain adaptation, where the relationship between datasets is usually unknown or imperfect. The paper does not establish robustness to systematic discrepancies such as g=Af+rg = Af + r, unknown calibration, or timing errors.

Furthermore, the basic integration result requires more target samples than source samples, which contrasts with the usual motivation for transfer learning.

The novelty claims also appear overstated. Previous studies, including Hall and Yatchew (2007) and Dai (2025), have already investigated joint estimation using function and derivative observations. While the specific minimax rates and phase transitions may be new, the paper does not sufficiently distinguish its contributions from this literature.

Finally, the experimental evidence does not fully validate the theoretical claims. In Fig. 2, the candidate-risk construction guarantees an SAR of at least one, while prescribed estimator switching makes the outer no-synergy regions equal to one by design. The real-world experiment also relies on calibrated signals and selected windows. These results illustrate the proposed mechanisms but do not independently establish the predicted phase boundaries or demonstrate robustness in general real-world settings.

Overall, these limitations substantially weaken the paper's practical significance and the strength of its broader claims.

Questions for authors

Could the authors clarify how the exact, known relationship g=Afg=\mathcal A f can be justified in realistic domain-adaptation problems?

  • Would the claimed synergistic gains persist under an approximate relationship g=Af+rg = Af + r, unknown calibration, or operator misspecification?

  • How do the theoretical contributions differ from prior work by Hall and Yatchew (2007) and Dai (2025), particularly regarding convergence rates and sample-size assumptions?

  • Furthermore, since Fig. 2 uses a minimum over candidate risks and prescribed estimator switching, can the authors provide results for the actual joint estimator across both predicted synergy and no-synergy regions?

  • Finally, how should the requirement n_T >> n in the basic integration setting be reconciled with the usual source-rich, target-scarce motivation of transfer learning?

1 comment
AuthorsOpenReview ↗

Response to Reviewer jK94

Thank you for these comments. We have clarified the known-link scope and contributions, added conditional robustness results, and evaluated the actual joint estimator across the predicted regimes. The main substantive revisions are highlighted in cyan in the revised manuscript.

W1/Q1: Known links and robustness.

Our setting covers structured transfer with a pre-specified relation between unknown functions, as in rate/heading sensors, cost/marginal-demand systems, and simulation gradients (revised Sec. 1). A known link does not guarantee complementary information.

For the bounded-design integral model, write gr=Af+rg_r=\mathcal Af+r. App. D.2 gives Rr≤2R0+2BrR_r\le2R_0+2B_r for perturbed/exact estimator risks, where Br=E∣Sbr−r∣L2(hT)2B_r=\mathbb E\\|S_br-r\\|_{L_2(h_T)}^2 and SbS_b averages target residuals in bandwidth-bb windows (zero if empty). Uniformly controlling Br=O(R∗)B_r=O(R^{\ast}), R∗=(nnT)−1/2R^{\ast}=(nn_T)^{-1/2}, preserves the risk upper rate in the SLP window. A fixed small error need not meet this shrinking tolerance. Independently estimated calibration must have propagated prediction error on this scale; timing errors are covered when their residuals satisfy these conditions. Estimating an unknown time delay itself is not studied.

For fixed-radius r∈H(γ,κr)r\in\mathcal H(\gamma,\kappa_r) with γ>β+1\gamma > \beta+1, Prop. 14 uses a local-polynomial correction with stability checks to attain R∗R^{\ast} on the restricted subwindow of Cor. 15, n(2γ+1)/(2γ−1)⪯nT≪n(2β+3)/(2β+1)n^{(2\gamma+1)/(2\gamma-1)}\preceq n_T\ll n^{(2\beta+3)/(2\beta+1)}. Under Gaussian noise, Cor. 15 also gives strict minimax SLP if P\mathcal P lies within these source and discrepancy classes and contains the exact-link subclass P0=(f,0):f∈H(β,κ)\mathcal P_0=\\{(f,0):f\in\mathcal H(\beta,\kappa)\\}; all three risks use this same class. A fixed operator error rf=(A⋆−A)fr_f=(\mathcal A_\star-\mathcal A)f satisfying the uniform smoothness bound inherits the upper bound, but its graph need not contain P0\mathcal P_0, so SLP does not follow automatically (App. D.2).

W3/Q2: Prior work and contribution.

Sec. 2 now acknowledges that Hall & Yatchew (2007) and Dai (2025) allow unequal sample sizes. Both focus on primitive recovery under bounded designs. Here we compare rates: Dai's one-dimensional RKHS rate is min⁡nT−2m/(2m+1),n−1\min\\{n_T^{-2m/(2m+1)},n^{-1}\\}, with function data anchoring the integration constant. Our Hölder model fixes g(0)=0g(0)=0; m=β+1m=\beta+1 aligns smoothness, not parameter spaces.

Within our model, Result II gives (nnT)−1/2(nn_T)^{-1/2} strictly beyond both single-domain optima, even with identical designs. Theorem 3 rules out strict minimax derivative synergy under bounded designs for every sample-size sequence, while allowing constant-factor gains. These results now appear consecutively (Secs. 4.1-4.2). Results III-IV show how singularity changes the integral mechanism and enables derivative synergy. Thus our contribution concerns strict two-baseline gains, directional limits, and design dependence.

W2/Q4: Why target-rich samples?

Source-rich transfer is often motivated by cheaper source data; it does not define the sample ordering. Here, ordinary pooling in one-dimensional Hölder regression with g=fg=f and identical bounded designs and Gaussian noise has source-only, target-only and joint minimax risks RS≍n−pR_S\asymp n^{-p}, RT≍nT−pR_T\asymp n_T^{-p}, RM≍(n+nT)−pR_M\asymp(n+n_T)^{-p} for p=2β/(2β+1)p=2\beta/(2\beta+1). If n≫nTn\gg n_T, pooling improves the target-only rate but matches the source-only order: SAR remains of constant order.

In Result II, beating the source rate n−1n^{-1} requires n≪nTn\ll n_T; beating the target rate nT−2(β+1)/(2β+3)n_T^{-2(\beta+1)/(2\beta+3)} requires nT≪n(2β+3)/(2β+1)n_T\ll n^{(2\beta+3)/(2\beta+1)}. These define the SLP window. Sparse rate measurements and abundant state logs motivate this ordering; singular designs also permit source-rich synergy (revised Sec. 1; Secs. 4-5).

W4/Q3: Independent empirical checks.

Fig. 2(b) and App. E.1 evaluate the actual joint estimator throughout the scan, without replacing its risk by a minimum with single-domain risks or switching estimators at regime boundaries. Its raw empirical ratio, min⁡(R^S,R^T)/R^M\min(\widehat R_S,\widehat R_T)/\widehat R_M, peaks on the grid at the predicted equalizing scale: 2.865 with pointwise 95% interval [2.527,3.229]. It is below one at evaluated outer-region points; sample-size scaling runs also show declining outer-region advantage. These support selected predictions; complete asymptotic boundaries follow from the minimax bounds.

On IO-VNBD at (n,nT)=(140,480)(n,n_T)=(140,480), the geometric mean SAR across 11 tracks is 1.385 [1.126,1.699]; removing the fitted scalar calibration while retaining the coordinate sign gives 1.275 [1.038,1.560]. Both use the same windows selected through link diagnostics; intervals exclude selection uncertainty. The corresponding ratio-of-mean-risks intervals include one (Table 4, App. F.1).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.