acceptodds
Under review as a conference paper at ICLR 2027

ConservGP: Certified Directional Compression of Gaussian Process Uncertainty

Abstract

Compressing a Gaussian process (GP) changes its predictive uncertainty, and the direction of the change matters: a planner needs the surrogate at least as uncertain as the teacher, a monitor no more uncertain. We present ConservGP, which fits a fixed-size sparse surrogate under a directional variance contract and certifies the contract over the continuous input domain, with every arithmetic step rounded outward on both models. Plain interval arithmetic loses the lower bound of the GP variance on most subintervals. The certifier instead encloses the mean coefficients through a verified linear solve and evaluates the variance in a form that keeps the cancellation between its two correlated terms. The certificate proves a signed inequality with a chosen margin, in either direction, between the teacher and the surrogate. Previous guarantees instead bound the range of one model or the size of an approximation error, or give a one-directional conservative posterior at a margin the construction fixes. In experiments, certified surrogates meet their contracts on every one-dimensional channel from synthetic data and two drone platforms, on the full domain volume of five problems with two to four inputs under certification-aware compression, and for six-output drone teachers under a matrix ordering on the joint covariance. On drone state channels the surrogate stores 54 times fewer parameters than the teacher and predicts about 34 times faster. In the selection step of safe Bayesian optimization the certified surrogate proposes only candidates inside the teacher's safe set, which unconstrained compression leaves in a fifth of the seeds. As a certified anomaly monitor it matches the teacher's discrimination at a matched false-alarm rate. For any input dimension, contracting the variance (the monitor contract) costs strictly more KL divergence than inflating it by the same margin (the envelope contract), with a gap that grows cubically at small margins and faster beyond them.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Gaussian processes, uncertainty quantification, model compression, sparse Gaussian processes, formal verification, monitoring, interval arithmetic, certified guarantees, safe Bayesian optimization, anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FseqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes to compress a GP teacher p into a fixed-size surrogate q made of K pseudo-observations with per-point noise. q is fitted by minimizing the grid-averaged pointwise KL(p‖q) under sufficient conditions. The paper’s value proposition is a small, fast GP surrogate that comes with a machine-checked, whole-domain proof that its predictive band sits on a chosen side of the teacher’s.

Critical strengths
  • The question of which side of the teacher’s uncertainty a compressed GP lands on is relevant for safety-critical deployment.

  • The proofs are correct, in my opinion.

Critical weaknesses
  • The manuscript reads very artificially. I recommend rewriting.

  • Theorems 1–2 reduce to the monotonicity of f(r) = ln r + 1/(2r²) − ½ in the ratio r = σ_q/σ_p. Equality holds at the trivial point μ_q = μ_p, σ_q = (1 ± δ)σ_p. “The optimizer matches the bound” therefore only shows σ_q ≈ 1.15σ_p everywhere.

  • The practical scope is weak: It is unclear to me whether this is a practical solution; the speed-up is not practically relevant. In the paper's example (Table 5), the teacher can be trained and queried on-board for predictions.

  • I am unsure how this methodology behaves in high-dimensional spaces. As written, it seems only dimensions 1 and 2 are practical.

  • All tables should report RMSE, NLL, and held-out coverage (in addition to speed-up) across baseline methods for the main experiments.

Questions for authors

What is the highest dimensionality you would consider being practical for the methodology? Why?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Fseq (1/2)

Weakness 1 (the manuscript reads artificially, rewriting recommended). We agree that the presentation was dense in places, and we have revised it. The abstract is rewritten. The introduction now says that the teacher grows with every trajectory that the robot observes. Contribution 2 and the paragraph after Theorem 2 now say in words when the bounds are attained and how close the fit comes. Contribution 3 states in one sentence what a certificate proves, and the captions of Figure 1 and Tables 1 and 4 now say in words what is shown.

Weakness 2 (Theorems 1 and 2 reduce to the monotonicity of f(r), and matching the bound only shows σq ≈ 1.15σp). We agree that both theorems follow from the Gaussian KL formula and the monotonicity of f(r). Their role in the paper is to give the experiments a reference value: no surrogate with σq ≥ (1 + δ)σp on the grid has KL below Φ(δ), so the excess of a fit's KL over Φ(δ) is the most that any other such surrogate could gain. A KL close to Φ(δ) therefore shows more than σq ≈ 1.15σp: it shows that K points reproduce the teacher's mean and the shape of its variance over the whole domain, and that the optimizer added no width beyond the margin that the contract requires.

How close our optimizer comes depends on K. At K ≥ 25 the constrained fit's KL, averaged over three seeds, is 2.1 to 6.8 % above Φ(0.15) at every teacher lengthscale that we tested. When K is too small for a short lengthscale (K ≤ 10 at lengthscale 0.05), the gap is large. The table of these gaps is in our reply to Reviewer F7tJ (weakness 2), and the revised contribution 2 states where the fit is close to the bound.

Weakness 3 (the speed-up is not practically relevant, and the teacher of Table 5 can run on board). Table 5 shows the teacher at one size, but the teacher does not stay at that size. It keeps every training point, so each trajectory that the robot observes adds to its memory and to its prediction time (caption of Table 5). The surrogate (POGP, K = 25) stays at 79 floats and a fixed prediction time. We measured the growth during the response period: with trajectories of 200 points each, one batch of 200 teacher predictions takes 2.8 ms with 7 trajectories and 3.4 s with 10,000, on one CPU core and with every factorization precomputed, while the fixed-size surrogate takes about 0.08 ms at both of these sizes. For scale, the public GeoLife collection holds 17,621 GPS trajectories with more than 48,000 hours of data. A robot that learns online leaves the size of Table 5 behind, and the fixed-size surrogate is what enables on-board prediction as the data grows. The introduction now opens with this motivation (item 2 at the end of our second comment).

Weakness 4 and the question (high dimensions, highest practical dimension). The paper's target is a robot model indexed by time: one input, with one or six outputs. During the response period, three selected scalar refits certified in 8.3 to 135.8 s with 16 processes on a 20-core desktop processor, permitting in-mission recertification of these channels (our reply to moUG gives the joint six-output timings separately). For more inputs, the highest practical number with the current certifier is three for offline updates and four for an overnight run. Two inputs took under 3 min on 8 workers, three took 16 min to 10.6 h, and four took 13.4 h (Table 2). For five inputs we have no measurement. A straight line fitted to log10 worker-seconds against the number of inputs (the five runs with two to four inputs) gives about 18-fold growth per added input. Extrapolated with ideal scaling over workers, five inputs would need about 156 h on 32 workers. A certificate within one hour would need about 15 workers at three inputs, 270 at four and 5,000 at five; within ten minutes, about 90, 1,650 and 30,000 (the fit's spread is a factor of 6.3 either way). The limit lies in the certificate: it tiles the input domain with boxes that are checked independently, so more workers shorten it (table in our reply to Reviewer moUG, weakness 1). The KL bounds and the constrained fit hold for any number of inputs.

AuthorsOpenReview ↗

Response to Reviewer Fseq (2/2)

Weakness 5 (RMSE, NLL and held-out coverage across baselines). We computed these metrics during the response period. In short, the constrained fit and post-hoc scaling keep the teacher's accuracy (RMSE within 2 % on every multi-input problem) and have wider bands. These show as coverage at or above the teacher's and, where the teacher's band was already wide enough, as a higher NLL. Coverage is the share of held-out observations inside a model's own µ ± 2σ band (95.4 % if calibrated). For the synthetic one-input problem of Table 4, with every baseline, we drew 1,000 fresh trajectories for each of 20 seeds, none of them seen by the teacher:

MethodRMSENLLHeld-out coverage of µ ± 2σ
Teacher1.12821.6320.891
Unconstrained surrogate1.12811.6320.891
Post-hoc scaling1.12811.5770.933
Penalty fit1.12571.5910.939
SQP, envelope1.12791.5780.932
SQP, monitor1.12801.7770.830
SQP, faithful1.12811.6320.891

RMSE differs by at most 0.0025 between rows. The envelope surrogate covers more held-out points than the teacher and has lower NLL, because the teacher under-covers on this problem. The monitor surrogate narrows the band by design, and its coverage drops.

For the five multi-input problems of Table 2 we scored the teacher, the constrained fit and post-hoc scaling on held-out data (teacher / constrained fit / post-hoc scaling). The surrogates are refits at the settings of Table 2 and carry no certificate. The variance includes the teacher's fitted observation noise for all three models:

Problem (inputs, held-out points)RMSENLLCoverage of µ ± 2σ
Meuse 2D (2, 31)0.339 / 0.333 / 0.3380.40 / 0.52 / 0.501.000 / 1.000 / 1.000
Synthetic 3D (3, 4,096)0.090 / 0.092 / 0.090−1.12 / −0.29 / −0.780.942 / 1.000 / 1.000
Meuse-3D (3, 31)0.320 / 0.326 / 0.3180.26 / 1.23 / 0.940.935 / 1.000 / 1.000
Ishigami (3, 4,096)0.572 / 0.568 / 0.5670.80 / 2.22 / 1.530.991 / 1.000 / 1.000
Synthetic 4D (4, 4,096)0.126 / 0.128 / 0.124−0.79 / 1.29 / 0.730.948 / 1.000 / 1.000

RMSE stays within 2 % of the teacher's on every problem. The surrogates' NLL is higher than the teacher's on all five, because their bands are wider than these data need.

On the 42 KIOS and Terra channels we held out one fifth of the trajectories of each regime and rebuilt the teacher from the others. The constrained fit's RMSE is 1.001 times the teacher's in the median (0.889 to 1.462), and post-hoc scaling's 1.000 times (0.963 to 1.040). With the variance as in the paper (no observation noise added), the 2σ band covers on average 98.9 % of the held-out observations for the teacher and 99.9 % for both surrogates. Both surrogates have a higher NLL than the teacher on 40 of the 42 channels.

During the response period, we also scored the penalty fit and both Appendix F SVGP baselines on the same 42 held-out channels, with equal weight per channel:

MethodMedian RMSE / teacherMedian NLL minus teacher (nats/observation)Mean 2σ coverage
Penalty1.0050.16399.9%
Post-hoc SVGP1.0000.14099.9%
Constrained SVGP1.0000.14199.9%

For the five multi-input problems, the penalty scores equal the constrained rows above because those refits used the released penalty recipe. We added a full-covariance SVGP with a separate lengthscale per input, trained for 2,000 Adam steps; this extends Appendix F's one-input baseline and uses the same held-out split and observation noise. The entries below are post-hoc SVGP / constrained SVGP:

ProblemRMSENLL2σ coverage
Meuse 2D0.335 / 0.3730.48 / 2.421.000 / 1.000
Synthetic 3D0.182 / 0.148-0.31 / -0.521.000 / 0.978
Meuse-3D0.339 / 0.4480.96 / 4.471.000 / 1.000
Ishigami0.551 / 0.5591.63 / 2.091.000 / 1.000
Synthetic 4D0.171 / 0.1670.93 / 0.461.000 / 1.000

These added fits have no certificates; iteration-limited or grid-infeasible fits remain included.

We have revised the paper accordingly:

  1. Table 4 has these columns, and Tables 29 and 30 report the held-out results, including the additional baseline scores.
  2. The opening sentence of the introduction now reads: "Deploying a Gaussian process (GP) trained on a robot's sensor data requires compressing the full posterior into a fixed-size surrogate, because the time and memory of one GP prediction grow with the number of training observations, and the teacher grows with every trajectory that the robot observes."
  3. Appendix J states that three inputs are practical for offline updates on the problems of Table 2, that four inputs need an overnight run, and that five or more do not fit an overnight run at these settings.
Reviewer_moUGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to control the direction of changes in predictive uncertainty after GP compression to meet the different requirements of planning and monitoring: a planner requires the surrogate to be at least as uncertain as the teacher, whereas a monitor requires it to be no more uncertain. The authors propose CONSERVGP, which fits a fixed-size sparse model through constrained optimization and then uses outward-rounded interval arithmetic to certify the constraints over the entire continuous input domain. Experiments include synthetic data, data from two drone platforms, and low-dimensional multi-input and multi-output settings.

Critical strengths
  1. The problem is clearly formulated and relevant to practical decision-making. A small KL divergence or average approximation error does not directly guarantee that the compressed model preserves the required uncertainty direction at every input. The paper expresses this requirement as an explicit contract between models and seeks to verify it over the continuous domain. The KL lower bounds for variance inflation and contraction also provide an intuitive reference for understanding the costs of different contracts.’

  2. The authors address loose interval bounds caused by cancellation between correlated terms in GP variance calculations through verified linear solves, affine expansions, and residual corrections. The paper reports certification results for 62 one-dimensional channels and further examines multi-input problems and matrix constraints on joint covariance.

Critical weaknesses
  1. Offline certification is expensive in some settings. Table 2 (page 6) reports 10.6 hours for the three-dimensional Ishigami problem and 13.4 hours for the four-dimensional synthetic problem. Although the paper acknowledges this limitation, the practical update frequency supported by these costs remains unclear.

  2. The handling of incomplete certification is not sufficiently clear. The sequential experiment on page 24 reports that 8 out of 500 steps did not obtain a certificate, but the compressed model was still used. The authors should explain why these steps were not certified and how such cases should be handled in deployment.

  3. The monitor may introduce substantial additional false alarms. The discussion following Table 8 (page 16) states that, when the teacher’s alarm threshold is retained, the certified monitor adds false alarms on 1,788 of 2,000 normal trajectories. Preserving the teacher’s alarms is a benefit, but the additional false alarms may affect practical usability.

Questions for authors
  1. In Table 2 (page 6), certification takes 10.6 and 13.4 hours for the three-dimensional Ishigami and four-dimensional synthetic problems, respectively. This overhead may limit applications requiring frequent model updates. Could the authors provide further analysis of this limitation?

  2. The sequential experiment on page 24 reports that 8 out of 500 steps did not obtain a continuous-domain certificate, but the compressed model was still used. Could the authors explain the specific reasons for these certification failures and how such cases should be handled in practice?

  3. The discussion following Table 8 (page 16) states that retaining the teacher’s threshold causes the certified monitor to add false alarms on 1,788 of 2,000 normal trajectories. Could these false alarms be reduced by decreasing the degree of interval contraction or recalibrating the threshold? How would these adjustments affect the detection rate and the preservation of the teacher’s alarms?

  4. The paper uses many symbols. Please consider adding a notation table that collects the definitions and meanings of the main symbols for ease of reference.

2 comments
AuthorsOpenReview ↗

Response to Reviewer moUG (1/2)

Weakness 1 and Question 1 (certification cost and update rate). The paper's main setting is a drone model indexed by time (one input). During the response period, three fixed scalar refits ran 10 to 11 times faster with 16 processes than with one, on a 20-core desktop processor. The table gives individual channels from this new measurement. These scalar channels permit recertification during a mission. The rates count the certificate only; fitting one surrogate took about 100 s on average in the synthetic experiment (Section 5). The 10.6 h and 13.4 h belong to the problems with three and four inputs, which suit updates between missions or overnight.

InputsProblemProcesses usedCertified updates/hour, measuredProjected to 32 processes (ideal scaling)
1Synthetic / KIOS / Terra, one scalar refit each16431.9 / 86.5 / 26.5no projection
1, six outputsTen drone models, one joint certificate each (5.5 min to 7.8 h)10.13 to 10.8no projection
1, six outputsThe fastest of these ten, refitted and certified again4 / 1683.5 / 138.6 (43.1 / 26.0 s)no projection
2Meuse82155
3Synthetic163.76.2
3Meuse-3D200.470.75
3Ishigami140.090.21
4Synthetic80.070.29

Rows 1 and 3 are medians of three runs and include startup and model loading. The rows with two to four inputs are the runs of Table 2, each with the number of processes of its original run. Their last column is computed, not measured: it spreads the measured certification work over 32 processes.

The certifier tiles the input domain and checks each tile independently, so workers share one certificate, and the cost grows about 18-fold per added input (fit over the five runs with two to four inputs, spread 6.3 times either way). From the fit, 32 workers would need about 0.5 h at three inputs, 8.6 h at four and about 156 h at five, an extrapolation beyond the data. More workers shorten these times in proportion while that scaling holds: by the same fit, 128 workers (edge or cloud compute) would need about 2.1 h at four inputs (Appendix J).

AuthorsOpenReview ↗

Response to Reviewer moUG (2/2)

Weakness 2 and Question 2 (the 8 uncertified safe-loop steps). None of the 8 steps was a violation. The loop refits the surrogate after every observation, 500 times in total. To keep each step fast, the submission did not run the outward-rounded certifier of Section 4 inside the loop. It ran a floating-point interval check at a fixed resolution, which can end without a decision. With no headroom (γ = 0) the surrogate's band lies very close to the contract boundary, and at 8 of the 500 steps the bounds of this fast check were too loose to decide. We have now run the outward-rounded certifier of Section 4 on all 500 updates. It certifies every one, including those 8, so the surrogate met the contract at every step of the loop.

In deployment, the surrogate replaces the teacher only once its certificate is obtained; until then the robot keeps the teacher. At the 8 steps the submission's loop used the surrogate before it had a certificate, and the revised Appendix N states this. A small headroom also lets the fast check decide every step: with γ = 0.10 it passes all 500 (Table 28).

Weakness 3 and Question 3 (the 1,788 added false alarms). Yes to both: a smaller contraction and a recalibrated threshold each reduce the false alarms. During the response period we swept the margin over 10 seeds, each scored on 200 normal and 200 anomalous trajectories. The teacher raises 100 false alarms and 1,671 detections. The table gives the surrogate's counts, summed over every fitted seed (certified or not), first with the alarm threshold that was set for the teacher and then with a threshold recalibrated for the surrogate:

Margin δCertified seedsFalse alarms / detections, teacher's thresholdFalse alarms / detections, recalibrated threshold
0.15 (paper)10 of 101,888 / 1,745100 / 1,671
0.0510 of 10584 / 1,692100 / 1,671
0.028 of 10216 / 1,681100 / 1,671
0.014 of 10157 / 1,675100 / 1,671

Smaller contraction: at the teacher's threshold, the false alarms added to the teacher's 100 fall from 1,788 at δ = 0.15 to 57 at δ = 0.01. Detections fall from 1,745 to 1,675, still at or above the teacher's 1,671, and fewer seeds certify. Recalibration: we set the surrogate's alarm threshold on its own scores for normal trajectories, so that it has the teacher's 5 % false-alarm rate. No false alarm is then added at any δ shown: the surrogate raises 100 false alarms and 1,671 detections, the same as the teacher. In both cases the surrogate raised every alarm that the teacher raised on the anomalous trajectories.

Question 4 (notation table). We have added one as the last appendix of the revised PDF (Appendix R). It lists each main symbol with its meaning and the place where it is defined.

We have revised the paper accordingly: Table 26 in Appendix J gives the update rates, Appendix J also gives the time range of the earlier joint certificates, and the conclusion states the 18-fold growth per added input. Appendix N reports the certificates of all 500 saved updates and the rule to keep the teacher until a certificate is obtained. The margin sweep is Table 9, next to Table 8.

Reviewer_F7tJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies GP compression under directional uncertainty constraints. This paper propose ConservGP, which fits a fixed-size surrogate and certifies that its predictive intervals contain, or lie within, those of the teacher over a continuous input domain. The paper also derives KL lower bounds under variance inflation and contraction constraints. Experiments on synthetic and drone data evaluate certification and compression efficiency, with additional studies on safe exploration and anomaly monitoring.

Critical strengths
  • The paper addresses a practically relevant problem in GP compression. The proposed contracts account for changes in both the mean and variance when specifying containment between the teacher’s and surrogate’s predictive intervals.

  • The method certifies interval containment between the teacher and surrogate over the continuous input domain, while accounting for dependencies and cancellation in GP variance calculations.

  • The one-dimensional experiments cover synthetic data and multiple channels from two drone platforms. The optimization ablations and measurements of model size and prediction time help assess both the fitting procedure and the compression benefits.

Critical weaknesses
  • Appendix F reports that an SVGP trained with the same pointwise KL objective, followed by global variance scaling, achieves comparable KL to ConservGP. POGP has reported advantages in model size and numerical stability, but the additional benefit of constrained fitting over this post-hoc alternative has not been fully evaluated under the same continuous-domain certification procedure.

  • Theorems 1–2 provide useful reference values, but their connection to the approximation limits of fixed-size GP compression remains limited. The bounds follow from the Gaussian KL formula and monotonicity in the standard-deviation ratio, without depending on the kernel, teacher complexity, or support size KK. The equality conditions in Corollary 1 also do not establish attainability by a fixed-KK surrogate.

  • Section 7 and Appendix G examine how additional headroom γ\gamma affects KL and verification cost. However, these analyses do not show how the same choices affect downstream decisions in the multidimensional setting.

Questions for authors
  • [Post-hoc comparison] Could the authors compare constrained fitting and pointwise-KL training followed by global variance scaling using the same continuous-domain certifier, and report certification success and cost?

  • [Theorems 1–2 and Corollary 1] Could the authors discuss when the KL bounds are attainable or approximately attainable with a fixed number of support points, and how this depends on the teacher or kernel?

  • [Headroom and downstream performance] Could the authors evaluate different values of γ\gamma under a fixed target contract and report certification effort together with task performance? This would help assess whether the reduction in certification effort justifies the additional conservatism.

3 comments
AuthorsOpenReview ↗

Response to Reviewer F7tJ (1/3)

Weakness 1 and Question 1 (post-hoc comparison). Appendix F made this comparison only for the SVGP architecture, on 1D synthetic data, and checked the result on a grid. We have now run both pipelines through the same continuous-domain certifier: for the SVGP architecture and for our surrogate architecture (POGP) on all 62 1D channels, and for POGP on the 2D to 4D problems of Table 2. In 1D and 2D, both pipelines certified, and our constrained fit is as close to the teacher as the post-hoc fit or closer. In 3D and 4D, the post-hoc fit is closer on three of four problems, but none of its certificates finished within the budget given below, while every constrained configuration (POGP) certified.

The post-hoc pipeline fits without constraints and then scales its variance by the smallest global factor that gives σq ≥ (1+δ+γ)σp on the grid, so both pipelines face the same variance target. We write r = KL(post-hoc) / KL(constrained). A value r > 1.05 means that our constrained fit is closer to the teacher, r < 0.95 means that the post-hoc fit is closer, and values in between count as a tie.

For 1D (δ = 0.15, and δ = 0.20 for Terra as in the paper, γ = 0), both pipelines certified every channel:

SurrogateData (channels)Constrained lower / tie / post-hoc lowerMedian r
POGPSynthetic (20)19 / 1 / 01.075
POGPKIOS (18)13 / 5 / 01.237
POGPTerra (24)13 / 11 / 01.055
SVGP, K = 25Synthetic (20)0 / 20 / 00.999
SVGP, K = 25KIOS (18)2 / 16 / 01.001
SVGP, K = 25Terra (24)2 / 22 / 01.001

The ratio r compares the two pipelines within one architecture; it does not rank POGP against SVGP. With our POGP, the constrained fit is closer or ties on every channel, and in the median the post-hoc KL is 5.5 to 23.7 % higher. With the SVGP architecture of Appendix F, the two pipelines tie on 58 of 62 channels (median r of 0.999 to 1.001), as Appendix F found on its grid.

For 2D to 4D (the γ of Table 2), each post-hoc certificate ran until it finished or reached its time budget: 3 hours on 3 workers for Meuse, 17 hours for synthetic 3D and 18 hours for the other three, on 5 to 10 workers. These budgets are 1.2 to 23 times the worker-seconds that our certificate of the same problem needed in total. The table gives the certification work as boxes checked, a count that does not depend on the machine, and the wall-clock time with the number of workers ('w'):

ProblemKL post-hoc / constrainedrOur certificate (Table 2)Post-hoc certificate
Meuse (2D)0.239 / 0.2251.063complete, 0.19 M boxes, 174 s (8w)complete, 0.23 M boxes, 352 s (3w)
Synthetic (3D)0.470 / 0.9440.498complete, 2.06 M boxes, 983 s (16w)stopped at 17.0 h (6w) after 48.5 M boxes, 69.5 % of the volume certified
Meuse-3D (3D)1.289 / 1.2861.002complete, 10.7 M boxes, 2.1 h (20w)stopped at 18.0 h (5w) after 12.9 M boxes, 1.6 % certified
Ishigami (3D)0.804 / 1.4730.546complete, 18.5 M boxes, 10.6 h (14w)stopped at 18.0 h (10w) after 14.1 M boxes, 0.5 % certified
Synthetic (4D)1.594 / 2.0580.775complete, 11.1 M boxes, 13.4 h (8w)stopped at 18.0 h (8w) after 14.7 M boxes (finer initial grid), 0.0097 % certified

Our times include building the surrogate. Our certificates come from earlier runs of the same configurations (Table 2); refitting changed the constrained KL by at most 0.022 from the values in this table.

So in 1D and 2D no r falls below the tie band: the constraint costs no closeness to the teacher, and both pipelines certify. In 3D and 4D, global scaling gives 22 to 50 % lower KL on three problems and ties on Meuse-3D, but none of its certificates finished, while every constrained configuration certified. The stopped certificates found no violation, and neither did a floating-point check at 200,000 random points, so they are unfinished, not failed. The revised paper states both results for 3D and 4D: the constrained fit is the pipeline whose certificate completed, and post-hoc scaling reaches lower KL on three of the four problems (Remark 1, Table 19).

AuthorsOpenReview ↗

Response to Reviewer F7tJ (2/3)

Weakness 2 and Question 2 (Theorems 1–2 and Corollary 1). Theorems 1 and 2 do follow from the Gaussian KL and monotonicity, and they are free of the kernel, the teacher and K by design: they give the smallest KL that any surrogate with the prescribed variance inflation or contraction can have. Question 2 is when a surrogate with a fixed K attains these bounds, and how much KL otherwise remains above the bounds.

Exact attainment: Equality needs µq = µp and σq = (1±δ)σp at every evaluation point (Corollary 1), so a K-point surrogate attains the bound exactly when it can reproduce the teacher's mean together with a scaled copy of its standard deviation. A POGP can do this whenever the teacher is a GP posterior with the surrogate's kernel on data at no more than K distinct inputs: place the pseudo-observations at those inputs and multiply the outputscale and every noise variance by (1±δ)². By Eq. (1) the mean stays µp and the variance becomes (1±δ)²σp². K = n is the simplest case. For K < n, the case of our experiments, equality is not reached in general, and the question becomes how large the gap is.

Approximate attainment (K < n): For teachers outside this family (more distinct inputs than K, or our Combined GP teachers, which average several trajectory posteriors), the gap is how well K points reproduce the teacher's mean and variance. This depends on the kernel and its hyperparameters; for the stationary kernels that we use, it depends mainly on the lengthscale relative to the domain. For variational sparse GPs, Burt et al. (2019) show that a number of inducing points that grows polylogarithmically with the data suffices for the squared-exponential kernel with Gaussian inputs, while for Matérn kernels (inputs uniform on an interval) the number that suffices in their analysis grows polynomially. The numerical K sweep in Table 22 of the revised PDF used 1D synthetic teachers on [0, 1] (three seeds, envelope contract, δ = 0.15, c = 2). Values are the mean constrained KL over Φ(0.15) over the three seeds. At K = 5, two fits at lengthscale 0.05 and one fit at lengthscale 0.15 did not complete, so those two entries are over one and two fits:

LengthscaleK = 5K = 10K = 15K = 25K = 40
0.053.882.901.661.0641.031
0.101.911.0971.0601.0291.024
0.151.3861.0401.0331.0311.024
0.301.0631.0381.0341.0441.044

During the response period we refitted this sweep and certified the fits with the outward-rounded certifier. All 24 constrained fits at K ≥ 25 certified, and their three-seed mean KL at each lengthscale and K was 2.1 to 6.8 % above Φ(0.15). At lengthscale 0.05, only one of three seeds gave a certified fit at K = 5 and one of three at K = 8, and one fit at each of these settings was far from the teacher (near-zero variance at some grid points): there K points are too few to follow the teacher.

During the response period, matched Matérn-3/2 and Matérn-5/2 fits at ℓ = 0.15 gave mean KL/Φ(0.15) of 1.107 and 1.057 at K = 25, and 1.074 and 1.031 at K = 40 (ten seeds each; the three-seed means in the revised appendix are 1.132 and 1.040 at K = 25, and 1.091 and 1.027 at K = 40). These fits enclose the teacher's band on a 2,001-point grid but do not meet the separate condition σq ≥ 1.15σp everywhere on it, so the ratios to Φ are numerical comparisons and the fits are not certified.

AuthorsOpenReview ↗

Response to Reviewer F7tJ (3/3)

Weakness 3 and Question 3 (headroom and downstream performance). The paper showed KL and verifier cost against γ (Appendix G) without a task. During the response period we added safe-set selection on Meuse 2D under one fixed contract (envelope, c = 2, δ = 0.15), where γ changes only the variance target (1+δ+γ)σp. On a 200 × 200 grid, a location counts as safe when µ + cσ of the predicted log lead lies at or below a threshold, set so that the teacher calls 50 % of the domain safe. "Kept" is the share of the teacher's safe locations that the surrogate also calls safe. At 1, the surrogate gives up no safe location. Below 1, the planner loses some locations that the teacher calls safe. The reverse cannot happen under the envelope contract: a surrogate that meets it calls no location safe that the teacher does not.

We evaluated ten γ values with teacher training splits 0, 1 and 2. The table selects five values and gives the median across the three splits as constrained fit / post-hoc scaling, with the better value in bold (Table 21 has the ranges across splits). "Boxes" counts every sub-box that the certifier examined, our measure of certification effort. All models in the table certified.

γKLBoxes (millions)Kept
00.0944 / 0.05200.586 / 0.9530.8026 / 0.8609
0.0250.1091 / 0.06060.537 / 0.8710.8029 / 0.8515
0.050.1074 / 0.06960.506 / 0.8060.7861 / 0.8431
0.10.1393 / 0.08830.363 / 0.7240.7639 / 0.8273
0.50.307 / 0.25530.209 / 0.4030.6506 / 0.7156

Headroom therefore buys a cheaper certificate at the price of safe locations. From γ = 0 to 0.1, the constrained fit's box counts fell by 24% to 42% across paired splits, while Kept fell by 3.03 to 8.25 percentage points. At γ = 0.5 they used 2.3 to 4.9 times fewer boxes than at γ = 0. The finer grid has local reversals. We certified 59 of the 60 models with the outward-rounded certifier; the constrained fit of split 1 at γ = 0.0125 was inconclusive at the depth limit. All 60 had zero false-safe points on the task grid. Post-hoc used more boxes in all 29 pairs with two certificates, while giving lower KL and higher Kept at every γ for splits 1 and 2.

The verifier ran in this sweep with two inputs only. For three and four inputs we ran the same task on a grid at each γ during the response period, on refits without a certificate. At the 50 % threshold the constrained refit keeps 54.1 % of the teacher's safe set at γ = 0 and 30.4 % at γ = 0.5 on synthetic 3D. It keeps 74.8 %, 52.3 % and 8.2 % at γ = 0, 0.5 and 3 on Meuse-3D, 56.2 %, 32.3 % and 8.7 % at γ = 0, 0.5 and 1.5 on Ishigami, and 30.5 %, 14.4 % and 0.0075 % at γ = 0, 0.5 and 4 on synthetic 4D. The largest γ of each problem is that of Table 2.

We have revised the paper accordingly:

  1. Appendix F has both post-hoc tables, with the certificate status of every row (Tables 18 and 19). Appendix G has the K sweep (Table 22) and the headroom task on Meuse 2D (Table 21), and Appendix N the same task with three and four inputs (Table 31).
  2. The second contribution now says where the optimizer comes close to the bounds: "The bounds hold for every kernel and every K. At the K of our experiments the constrained fit comes within a few percent of these bounds (Section 5.3), and Appendix G shows how the gap grows when K is too small for the teacher's lengthscale."
  3. After Theorem 2, "both are attained with equality for every δ" now reads: "Equality holds when µq = µp and σq = (1 ± δ)σp at every evaluation point (Corollary 1), which a fixed-K surrogate reaches only if it can reproduce the teacher's mean and a scaled copy of its variance. Appendix G measures how close K points come."
  4. Remark 1 now states the mixed result of the post-hoc comparison.

Comments on the paper

1 comment

We thank the three reviewers for constructive and useful feedback. Three questions run through the reviews: what the constrained fit gains over scaling the variance after an unconstrained fit, what the KL bounds say about a surrogate of fixed size, and what a certificate costs in time, in number of inputs and in task performance. We ran new experiments on each during the response period. The revised PDF has been updated with them as well as to improve on the readability of the paper. The added results are also present in the replies below to the reviewers.

New experiments

  • Post-hoc variance scaling through the same continuous-domain certifier as the constrained fit, on all 62 one-input channels and on the five problems with two to four inputs (Reviewer F7tJ; Tables 18 and 19).
  • A sweep over the surrogate size K against the KL bounds (Reviewers F7tJ and Fseq; Table 22, with certified refits in our reply to Reviewer F7tJ).
  • Headroom γ against a task: the certification effort and the share of the teacher's safe set that the surrogate keeps (Reviewer F7tJ; Tables 21 and 31).
  • Certified updates per hour with one to four inputs (Reviewers moUG and Fseq; Table 26).
  • The full certifier on all 500 updates of the safe-exploration loop. It certifies every one, including the 8 steps that the submission reported as not certified (Reviewer moUG; Appendix N).
  • A sweep of the monitor's margin δ, and a recalibrated alarm threshold (Reviewer moUG; Table 9).
  • RMSE, NLL and coverage on held-out data for the teacher, the surrogates and the baselines (Reviewer Fseq; Tables 4, 29 and 30).
  • The teacher's prediction time as its data grow, against the fixed-size surrogate (Reviewer Fseq; in our reply).

What changed in the paper's claims

  • The comparison with post-hoc scaling is mixed, and the paper now clarifies this (Remark 1, Appendix F). With one and two inputs both pipelines certify, and the constrained fit is as close to the teacher or closer. With three and four inputs the post-hoc fit is closer on three of four problems, but none of its certificates finished within its budget, while every constrained configuration certified.
  • The paper no longer says ambiguously that the KL bounds "are attained with equality for every δ". It states when equality holds, and Appendix G measures how close a surrogate with K points comes.
  • The paper states how far the certifier reaches: three inputs are practical for offline updates, and four need an overnight run (Appendix J and the conclusion).
  • Appendix N states the rule for deployment: the surrogate replaces the teacher only once its certificate is obtained.

Presentation.

A notation table is added (Appendix R). We also revised the abstract, the introduction and several captions for clarity and better reading flow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.