acceptodds
Under review as a conference paper at ICLR 2027

The State Between Stages: Identifying Sufficient Statistics from Second Derivatives

Abstract

A deterministic differentiable program often lets one block of its inputs interact with the rest only through a few numbers, its interface, whose dimension is the state a staged solver must carry. Representation learning estimates such a summary from data; we compute it from second derivatives, even where no variable holds it. The blocks' cross-Hessian factors through the interface, so its rank at any set of second-block values bounds from below the dimension of every linear interface and every smooth one with a full-rank Jacobian. Under analyticity, constant rank, a full-rank summary map, and values drawn from a density, the rank almost surely equals the minimal dimension of an analytic interface, which can lie below that of any linear one. On a control rollout it reads 2 where the loop carries 4 numbers, recovering the zero-effort miss (where the mass would land if it coasted), which no line computes. Solving in stages through it reaches the joint optimum. On a published robot-arm benchmark it reads 2 against 3 carried numbers. Given only the objective, it recovers the interface of 200 of 200 random objectives and of 6 Feynman formulas, and two ratio interfaces as payoff differences.

Primary area:
optimization
Keywords:
sufficient statistic, state abstraction, interface dimension, cross-Hessian rank, identifiability, staged optimization, dynamic programming, differentiable programs, dimension reduction, functional separability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_XikeOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper takes optimization problems which can be written in two stages, and aims to find a compressed relationship between the first and second stage. This is done by evaluating the objective, using autodifferntiation to compute hessians, checking that the ranks of the hessians are consistent, and then finding a summary of the relationship between the stages. They give theoretical results which show that enough random probes reveal the full derivative span, under analyticity and other assumptions. If that span has constant dimension r nearby, then an r-dimensional nonlinear sufficient summary actually exists locally, and no smaller regular summary works. As far as I can tell, the contribution of this paper is theory which specifies conditions under which a low-rank hessian around observed points leads to a low-rank summary being guaranteed to exist. There are some computations done to verify this, but they are very difficult to follow.

Critical strengths

The introduction is well-written. I didn't really understand the purpose of the paper from the abstract but the gravitational example makes it simple and concrete.

Section 2 is mostly clear, and the theory all seems to be correct.

The problem considered in the paper is interesting and not one I had thought about before.

Critical weaknesses

The definition of 'interface' which seems to be completely crucial to the paper is buried in the introduction.

The examples in section 3 all seem to be control problems, but the dynamics are never explicitly written. The reader instead has to figure them out from the prose.

It is extremely hard to understand the experiments of section 4 from the text. There are no subsections, diagrams, results plots, and the details of each experiment are buried in a paragraph of text.

It took me quite a while to understand the actual purpose of the paper. A clearer statement of the aim of the paper near the start would be useful.

All examples considered in the computational experiments are tiny by the standards of optimization papers.

I think this paper needs a significant rewrite. It currently lacks a clear logical flow, and I had to repeatedly go back and forth, rereading parts over and over, to understand what is happening. Almost everything is in long paragraphs of text with no subsection breaks, diagrams or tables. The experimental sections are especially hard to read.

Questions for authors

What is it which prevents larger experiments (such as control problems with larger time horizons) from being possible?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Xike

Thank you for the detail of this review and for checking the theory. We worked from your list of where the paper lost you.

You noted that the definition of the interface is crucial to the paper and was hard to find in the introduction. Section 2 now states it as Definition 2.1, followed by the problem. The introduction gives it in words first.

You noted that the dynamics of the examples in Section 3 were never written out. Section 3.1 now displays the simulation's update and objective with every constant and derives the coast time τ\tau in the zero-effort miss q+τq˙−q∗\mathbf{q} + \tau \dot{\mathbf{q}} - \mathbf{q}^\ast (the displacement from the target q∗\mathbf{q}^\ast of the point where the mass would end if it coasted). Each of the three examples has its own subsection.

You found the experiments of Section 4 very hard to follow, with no subsections, diagrams, or results plots and each experiment's details inside one paragraph. Section 4 now opens with the procedure as a numbered list and has a subsection for each experiment (the Feynman formulas, the objectives with hidden interfaces, the two baselines). Table 1 gives each formula's context, probe, rank, and returned summary. To keep the page limit, the baselines' details moved to Appendix B and Section 4.3 summarizes them. We did not add a results plot.

You asked for a clearer statement of the aim near the start. The first paragraph of the introduction now states the aim. The second works it through the point mass of Figure 1. We state the formal problem after Definition 2.1.

We are glad that the gravitational example made the idea concrete. It is still in the first paragraph. Table 1 now lists it with its product m1m2m_1 m_2. The point mass leads because its code never computes its interface.

You noted that all our examples are tiny by the standards of optimization papers. They are as small as before, apart from the two runs of a longer simulation described below. We believe that the paper's claims do not rest on larger experiments. The theorems hold at any dimension in exact arithmetic. Each example tests one hypothesis or what the procedure returns. We agree that a demonstration on a large problem would have strengthened the paper. Section 6 says that the evidence lacks a decision program at scale that someone else wrote.

You asked what prevents larger experiments, such as control problems with longer horizons. The horizon enters the rank computation through the number of derivative evaluations. Our implementation takes one Hessian-vector product per probe coordinate (each at a small multiple of a gradient's cost), so a stack of mm probes of dimension pp costs about mpmp gradient evaluations however large the context is (Section 2). In the simulation pp counts the second phase's thrust coordinates, so it grows with the horizon. With a context of dimension 1000, a synthetic objective takes 3 probes (Section 3.3). In the simulation the zero-effort miss has two coordinates at every horizon. With phases of 100 and 400 steps, 204 context coordinates and 800 probe coordinates, the simulation's rank is still 2 (Appendix B, new).

The horizon also enters through conditioning. With the terminal velocity penalized as well at those phases, the interface is the four-number handoff state. Over 400 damped steps, however, two of the stack's four nonzero singular values are below our cutoff of 10−1010^{-10}, by less than a factor of two, and the computed rank is 2 (Section 6 and Appendix B).

The discovery procedure of Section 4 is more limited. Its first step differentiates in every context coordinate. Its closed-form fit searches over monomials, whose number grows combinatorially with the context dimension and the degree. The largest context we fit has 5 variables (Section 6). We have not yet run the method on a large program written by someone else.

We agree with you that a significant rewrite is due. This revision goes part of the way. It rewrites the abstract and the opening, adds the definition, the problem statement, the simulation's equations, the numbered procedure, the subsections, and Table 1, and simplifies the wording throughout. The order of the argument is as submitted, and several long paragraphs remain. Thank you again for the care in this review.

Reviewer_uHrmOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers differentiable programs. It is supposed to interact (the interface) with the rest of the world with few numbers which are learned from second derivatives. The paper seems to compute bounds on the interface.

Critical strengths

I have no idea what this paper is trying to do,

Critical weaknesses

I have no idea what this paper is trying to do,

Questions for authors
  1. Can you define the problem that are trying to solve?
  2. Can you define the contributions of the paper?
1 comment
AuthorsOpenReview ↗

Response to Reviewer uHrm

Thank you for reading the paper and for your two questions. The paper should have answered both on its first page.

You asked us to define the problem. Take a differentiable objective f(x,y)f(x, y) whose inputs we split into a context xx and a probe yy. An interface of dimension ss is a map ρ\rho from the context to Rs\mathbb{R}^s with f(x,y)=A(x)+H(ρ(x),y)f(x, y) = A(x) + H(\rho(x), y) near a base point (Definition 2.1, which is new). The problem is to find, from evaluations of ff and its derivatives alone, the least dimension of an interface and an interface that attains it. We take the least dimension within a class of maps, linear maps or smooth maps whose Jacobian has full rank at the base point, because without regularity a map can pack many numbers into one. We state the problem after the definition.

In the running example the context is a point mass's initial state and first-phase thrusts, while the probe is its second-phase thrusts. The interface is the point where the mass would end if it coasted, two coordinates that the simulation never computes.

You asked us to define the contributions. The introduction now lists three.

  1. The paper states and proves what the rank of the matrix of mixed second derivatives between xx and yy says about the interface (Section 2). The rank bounds from below the dimension of every linear interface and every smooth one with a full-rank Jacobian (Proposition 2.2 and Appendix A). Under analyticity, constant rank, and a full-rank summary map it equals the least dimension of an analytic interface (Theorem 2.5). The matrix stacked over enough random probe values attains that rank almost surely (Theorem 2.3).
  2. The matrix's singular vectors give the interface's directions even when the program does not compute the interface (Section 3). On the point mass, whose code passes position and velocity from step to step, the rank is 2 and the vectors recover the two coordinates of the coasting point. Solving in stages through that point matches the joint optimum.
  3. A procedure that is given only the objective returns the interface (Section 4). It recovers the interface of all 200 random objectives whose interfaces were hidden from us until we checked the answers. Of 9 physics formulas from the Feynman database, it returns a closed form, a product of two variables, for 6. For the 2 whose interface is a ratio it returns a difference of the objective between two probe values, which the program can evaluate but which the procedure does not reduce to a closed form. For the ninth it reports that the smallest interface is the whole context.

Thank you again for both questions.

Reviewer_Qrr6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper considers differentiable programs in which many inputs are given, but computations may depend on these inputs only through low-dimensional "interfaces". In particular, the paper gives a tool for identifying, at a node of the computational graph in the program, if the output only depends on a lower dimensional summary (sufficient statistic) of the inputs. They not only give algorithms for identifying such low-dimensional interfaces, but also for constructing the sufficient statistics, even if they are not already explicitly computed by the program (that is the program depends on these sufficient statistics only implicitly). This identification strategy is primarily based on the cross-Hessian. Performance is tested on many concrete objectives, and appears to give promising results in that low-dimensional interfaces are identified.

Critical strengths

Identification of low dimensional interfaces in a computational graph for a given program could have wide applicability. 

If this importance could be made more concrete, it could significantly strengthen the paper.

Evaluation on real objectives appears to show that the cross-Hessian does indeed identify low dimensional interfaces.

Critical weaknesses

This is exceptionally hard to read and understand as a non-expert in this area. For example, even in just the first paragraph the following terms are unclear: block, control rollout, thrust phase, handoff, loop. Any one of these could be inferred from context, but the density of such terms makes reading quite challenging. Unclear writing persists throughout the paper. It is possible that this low quality writing style comes from poor use of LLM tools. 



As another example, the paper often refers to nodes of the trace, but the trace (presumably a computation graph) and its nodes are never defined.

The paper lacks any figures, which is particularly troubling given the lack of clarity in the text. A figure could be particularly helpful for explaining the trace and what an interface is. 



Despite significant effort, I cannot understand this paper sufficiently to give any recommendation other than reject. I am happy to defer to reviewers with higher levels of confidence. I am regretful I cannot give a more informative review, but feel the authors have failed in their responsibility to make the work understandable.

Questions for authors

Can you describe in simple terms a practical pipeline for deploying this technique on real problems and the downstream benefits it would yield? 


What is the broader importance of solving this low-dimensional interface identification problem?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Qrr6

Thank you for this review and for saying plainly that you could not follow the paper. Your list of terms from the first paragraph showed us what to repair first.

You listed "block," "control rollout," "thrust phase," "handoff," and "loop" as unclear in the first paragraph. The opening now defines "block," describes the simulation in plain words, and defines "loop" and "handoff state" on it before using them. "Rollout" and "thrust phase" no longer appear.

You suggested that the writing may come from poor use of LLM tools. The paper's AI use statement discloses that we used LLM assistants to draft and edit the text. Our editing of the result was not enough. The undefined terms were our responsibility. We revised the opening from your list and then simplified the wording of the rest.

You noted that the trace and its nodes were never defined. The introduction now defines the trace as the graph whose nodes are the intermediate values the program computes and whose edges say which values each node is computed from. It is the computation graph, as you presumed.

You noted that the paper lacked figures. Figure 1 now shows, on the point-mass example, three handoff states that leave the second phase the same problem (panel a) and the singular values that give the rank (panel b). An interface is a set of numbers, computed from one group of the program's inputs, through which that group affects how the output depends on the rest. We did not draw the trace.

You asked how one would deploy the method in practice and what it would yield. The introduction's last paragraph now gives the steps in brief. In more detail:

  1. Write the objective as a differentiable program and split its inputs into the first stage's decisions (the context xx) and the second stage's (the probe yy).
  2. At a base point and a few random values of yy, compute the matrix of mixed second derivatives between xx and yy by automatic differentiation and stack the matrices.
  3. Under the hypotheses of Section 2, the stack's computed rank is a number of first-stage quantities that suffices for the second stage. Its top right-singular vectors give their directions.
  4. Find the quantities from those directions, in closed form or as differences of the objective between probe values (Section 4). The procedure stops without an answer when the rank changes across sample points.
  5. Build the second stage's optimal value (a table, a value function, or a planner) over those quantities in place of the full state the code passes, then solve the first stage against it.

On the point mass, the zero-effort miss is the point where the mass would end if it coasted, measured from the target. A table of the second phase's optimal value over the miss's two coordinates has 625 entries (25 per coordinate). Over the four coordinates of the state the code passes, it would need 390,625. Solving the first phase against the small table matches the joint optimum over all 16 thrust coordinates to within 10−910^{-9}.

You asked about the broader importance of the problem. Anyone who solves a deterministic problem in stages (by dynamic programming, model-predictive control, or a learned value function) chooses by hand what the first stage passes to the second, usually the variables the code already passes. Every object built for the second stage grows with that choice. A grid's size is exponential in the number of values passed. The method computes from the objective how many numbers suffice and which ones, even when the code never computes them. Section 6 states what it leaves out, chiefly a value function defined by an expectation over a transition kernel. Thank you again for your time and your candor.

Thank you for addressing the clarity concerns I pointed to. I still feel that clarity is a broad issue.

Although I aim for this to not impact my score, it seems worth noting that the original manuscript, the changes, and the authors response all seem entirely LLM generated. I sympathize with the other review saying "I have no idea what this paper is trying to do".

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers: summary of the revision

We thank the three reviewers for the time they gave this paper. The reviews agree that the paper is hard to follow. We are sorry for the effort our writing cost you.

The uploaded revision changes only what you raised, within its 10-page limit. Its theorems, experiments, and results are those of the submission, apart from the computations behind the new Figure 1 and two runs of a longer simulation that answer Reviewer Xike.

  • The abstract and the opening of the introduction are rewritten. They define each term they use (block, context, probe, interface, loop, handoff state, zero-effort miss) and work them through one example, a point mass pushed by thrusters in two phases. The introduction also defines the trace, its nodes, and a dominator where each first appears. Its last two paragraphs list three contributions and say how an analyst would use the method.
  • Figure 1 is new and shows three handoff states that leave the second phase the same problem (panel a) and the singular values that give the rank (panel b).
  • Section 2 states the interface as Definition 2.1 and then the problem.
  • Section 3 has a subsection for each example and writes out the simulation's dynamics and objective. Appendix B adds a simulation with phases of 100 and 400 steps.
  • Section 4 gives the procedure as a numbered list and has a subsection for each experiment. Table 1 lists the nine Feynman formulas with their splits, ranks, and returned summaries.
  • Section 6 has a new paragraph on what limits larger experiments.

To make room we moved to the appendices the construction for a chain of more than two blocks, the baselines' details, the cases in which the procedure stops without an answer, and part of the related work (Appendix C). Each leaves a pointer in the main text. We also simplified the wording throughout and split the longest paragraphs. Where the submission said that the cross-Hessian "reads" a rank and that the loop "carries" numbers, the revision says that the cross-Hessian has that rank and that the code passes those numbers.

The experiments are still small, Section 4 still lacks a results plot, and we did not draw the trace, which the introduction defines in a sentence.

We corrected twelve statements of the submission. Apart from these, its claims are unchanged.

  • The interface suffices for a staged solver (Definition 2.1). The submission said that a solver must carry it and that a value function or planner needs that many inputs.
  • Section 3.1: the first phase affects how the objective depends on the second phase's thrusts only through the zero-effort miss. The submission said that the first phase affects the objective only through the miss, but its effort also enters the objective.
  • Section 3.3: we removed the claim that no node's gradient at a single probe is an interface direction. One node's gradient is.
  • Section 2 and Appendix A: when constant rank fails, the rank is still a lower bound. The submission called it wrong.
  • Abstract and conclusion: the rank equals the least dimension for almost every draw of enough probes (Theorem 2.3).
  • Section 2: the sentence on the cost of derivatives describes the computations of Section 3. The procedure of Section 4 differentiates in every context coordinate.
  • Section 4.1: the second gap is the largest difference between the staged and joint values. The submission called it the distance between the two optima.
  • We removed a remark that the center-of-mass ratio is not a power product, because it has the level sets of m1/m2m_1/m_2.
  • Table 1's caption gives the ranges from which we sampled two probes, which the submission misstated. Appendix B lists three further computations where the submission counted two.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.