A Compositional Atlas for Algebraic Circuits

Circuits based on sum-product structure have become a ubiquitous representation to compactly encode knowledge, from Boolean functions to probability distributions. By imposing constraints on the structure of such circuits, certain inference queries become tractable, such as model counting and most probable configuration. Recent works have explored analyzing probabilistic and causal inference queries as compositions of basic operators to derive tractability conditions. In this paper, we take an algebraic perspective for compositional inference, and show that a large class of queries - including marginal MAP, probabilistic answer set programming inference, and causal backdoor adjustment - correspond to a combination of basic operators over semirings: aggregation, product, and elementwise mapping. Using this framework, we uncover simple and general sufficient conditions for tractable composition of these operators, in terms of circuit properties (e.g., marginal determinism, compatibility) and conditions on the elementwise mappings. Applying our analysis, we derive novel tractability conditions for many such compositional queries. Our results unify tractability conditions for existing problems on circuits, while providing a blueprint for analysing novel compositional inference queries.

Paper

References (59)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer oPB26/10 · confidence 3/52024-07-08

Summary

This paper unifies two lines of work: 1) a compositional approach to tractability of queries over logical [12], probabilistic [35] and causal circuits [36] and 2) a (commutative) semiring-based perspective over different computational tasks ([12] among many others). The benefit is the characterization of certain probabilistic and causal tasks in terms of combinations of simple operations on algebraic circuits. Concretely, this approach is used to derive new complexity results in second-level algebraic model counting (including highlighting a wrong claim in [20]) and causal inference, among others. **update after rebuttal** The response clarified a bit the scope of the contribution. I think that this is solid work overall and that adding parts of their response to the main text would make the paper more accessible. I raised my score accordingly.

Strengths

- The paper is overall well-written, albeit dense (which I think it is a necessity due to the limited space) - The novel properties of AC introduced here could lead to more effective compilation algorithms

Weaknesses

- Novel and existing notions/definitions are mixed together, making it hard to give proper credit where it is due. - The high-level contributions are not very clear from the text (in particular in the first sections). [35] clearly listed their contributions at the end of their introduction, in contrast, I found lines 39-42 more vague in this sense, giving the (wrong) idea that the compositional approach is a novelty. - I think that the practical impact of this contribution is not be very clear to a general AI/ML audience.

Questions

1. Could you clearly outline what are the novel notions that enable this generalized compositional framework? 2. Could you similarly outline what are the key ideas / existing work this work builds upon? 3. Could you provide examples (for the general NeurIPS audience rather than experts in AMC and AC) of the practical impact of this contribution? ## Minors Multiple typos on Def. 1

Rating

6

Confidence

3

Soundness

4

Presentation

3

Contribution

2

Limitations

The limitations of this work are adequately addressed in the conclusion.

Reviewer oGJN7/10 · confidence 3/52024-07-10

Summary

The paper mainly focuses on the problem of deriving tractability conditions for compositional operations over circuits as to solve a number of queries. By fixing the language of circuits over semirings, the paper generalizes known results and introduce novel ones regarding the tractability of queries solved via algebraic circuits. In addition, it analyzes scenarios where milder structural conditions might be required, when compared to previous results.

Strengths

The results of this paper are part of a series of many theoretical results regarding the tractability of complex operations over circuits. I think the paper positively contributes to the current research area, with a more general understanding of already known results but also new ones. For instance, the exponential separation (Theorem 9) provides an example where lifting the X-first property not needed for computing 2AMC queries on particular semirings might provide an exponential expressiveness advantage. However, note that I lack background regarding the causal probabilistic queries being considered in Section 4.2, and I am not able to fully check the claims regarding their complexity. I believe the paper is very well written: the notation is very clear, most of the claims are well motivated from the beginning, and there are concise examples across the paper. The proofs in the appendix are also easily readable once having understood the structural properties being used. The reason of the lower scores related to soundness and contribution are listed in the weaknesses section. However, note that I am somewhat confident that the authors can solve the following issues during the rebuttal phase and therefore potentially improving my score.

Weaknesses

I think that some results requiring particular semirings and homomorphisms require so many properties that make me wonder their actual contribution. Tractable mapping (Theorem 4) requires the following combination of properties to be both true: 1. $\tau$ is additive OR $C$ is deterministic 2. $\tau$ is multiplicative OR ($\tau$ maps the multiplicative identity AND for each product $\alpha$, one of its input must output zero or one as a constant). I think that the authors should have stressed the limitations of the given properties on $\tau$. E.g., if $C$ is not deterministic and there are no product units satisfying the above, then which functions $\tau$ are we restricted to in practice? See the question 1 below for more on this. Theorem 3 claims that one can compute the product of support compatible circuits linearly in the maximum circuit size for both time and memory. However, I think the authors should state that this complexity assumes that (i) we already know the isomorphism between nodes, and (ii) we already know how to "match" the supports of sum nodes in the algorithm. I think that without (ii) one would definitely need linear space and quadratic time instead. In line L298-L300, the authors claim that for particular semirings and mappings, one could relax the determinism property from the circuit for tractable 2AMC. However, this does not seem to be reflected in table 2, where X-Det is present in all 2AMC queries. I think the authors should either make a milder claim or explicitly show in table 2 a case of such a query. Looking at Algorithm 2 and 3, I think there are corner cases that are not covered neither by them nor by the proofs. For example, the case of multiplying a sum node and a product node is missing. The case of multiplying input nodes and sum/product nodes is also missing. In fact, these algorithms are way too small when compared to the algorithm of the product showed in [35] (Algorithm 3 in the appendix). I believe the cases must be covered for completeness. I believe a minor weakness of the presentation of this paper is that it mixes two kinds of theoretical results. It contains some results that leverage algebraic circuits and specific properties regarding semirings and morphisms (e.g., Theorem 4). However, it seems the majority of the other results do not actually need any specific property of the semiring being chosen, and could have been part of a work following the compositional atlas paper on circuits [35]. When reading the paper, I often found myself asking whether the theorems and proof I was looking at were the result of a property of semirings I was missing. This is because of the title and the definition of algebraic circuits is the first definition appearing, and I was expected the rest of the paper being very dependent on it. Instead, it seems many theorems can stand on their own with the usual circuit definition given in [35]. Although minor, I think it would be beneficial if the authors stressed at the beginning that the chosen semiring is not really important for many of the results being shown.

Questions

1. If we restrict $\tau$ to be both additive and multiplicative homomorphisms, then does this imply it can only be identity function? It looks like it can only be identity function as for Lemma 3.8 in the appendix of [35]. Do the authors agree on this? If yes, then I think it is necessary to refine the conditions required by Theorem 4 as to include these results about the satisfaction of Cauchy functional equations. 2. Following Question 1, are there examples of both additive and multiplicative homomorphisms for some particular choices of source and destination semirings? 3. It seems the support-compatibility definition does not consider how scopes of products are factorized. What is the relationship between support compatibility between circuits and compatibility + determinism? Is structured decomposability PLUS determinism in a circuit more restrictive than support-compatibility with itself?

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

I have discussed the presence of limitations in the weaknesses section.

Reviewer 41Ra5/10 · confidence 4/52024-07-11

Summary

The paper presents sufficient conditions under which certain problems (e.g. 2AMC are tractable when performed on circuits.

Strengths

The paper is generally well-written and technically sound. It also makes a solid effort in trying to unify tractability conditions in the context of algebraic circuits.

Weaknesses

1) My main problem with the paper is that it only discusses sufficient conditions (also acknowledged by the authors). This is constrast to teh probabilistic circuit atlas [1]. This in itself is not a problem, however, the paper does not discuss the problem of eighted model integratio, which has also been formulated as an AMC problem with integration as an extra operation [2]. In general WMI is intractable, however, it has been shown that there exist tractable fragments [3]. If I understand the paper correctly, the tractable fragment of WMI does not fall within the proposed framework. This is quite limiting as the framework does not even cover all tractable problems that can be formulated as 2AMC. 2) Tractable algebraic circuits seem to be highly related to performing tractable inference on first-order circuits. [4] This relationship is not discussed at all. [1] Vergari, Antonio, et al. "A compositional atlas of tractable circuit operations for probabilistic inference." Advances in Neural Information Processing Systems 34 (2021): 13189-13201. [2] Dos Martires, Pedro Zuidberg, Anton Dries, and Luc De Raedt. "Exact and approximate weighted model integration with probability density functions using knowledge compilation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. No. 01. 2019. [3] Zeng, Zhe, et al. "Probabilistic inference with algebraic constraints: Theoretical limits and practical approximations." Advances in Neural Information Processing Systems 33 (2020): 11564-11575. [4] Van den Broeck, Guy, et al. "Lifted probabilistic inference by first-order knowledge compilation." IJCAI. 2011.

Questions

Could the authors comment on the two points I raised under "weaknesses".

Rating

5

Confidence

4

Soundness

4

Presentation

3

Contribution

2

Limitations

covered.

Reviewer x2Dq6/10 · confidence 3/52024-07-12

Summary

They investigate algebraic circuits on semi-rings (with sums and products). They give criterions that allow efficient combinations of circuits and aggregation of the variables of one circuit (e.g. the sum over all inputs of the circuit). They give algorithms and hardness results for Algebraic Model Counting.

Strengths

they give an unified view on different important problems which I did not expect to be related foundamental research efficient algorithms well-written

Weaknesses

a lot of technical definitions the conditions under which the algorithms are efficiently applicable seem rather restrictive

Questions

what is an "atlas"? the word is never used in the text the classic circuit classes have an uniformity constraint, on the complexity of an algorithm that computes the circuit for a certain input length. do these algebraic circuits not need uniformity ? >Theorem 2 why not just add one new product node as root node, and then make both C and C' children of this new node. it is a much simpler construction and would also compute the product.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

yes

Reviewer oGJN2024-08-09

Thank you for the very good responses to all my questions. I have decided to raise my overall score to full acceptance.

Reviewer oPB22024-08-11

Response to the authors

Thank you for clarifying some aspects of your contribution, I don't have further questions at the present time.

Reviewer 41Ra2024-08-11

**On WMI and AMC** Indeed the method by Zeng et al. solves WMI via a message passing scheme. As it is well-known that message passing is essentially performs a series of sum product evaluations this can be cast as an AMC problem, which was done by Zuidberg et al. Given these two works it should hence be possible to cast WMI as an inference problem in the AMC atlas. **On first order circuits** You state that "you work focuses on propositoinal formulas". However, if you had pure propositional formulas you would not have any variables to perform the extra operation over. I.e. the max in a marginal map problem. In this regard you are not dealing with propositional formulas but with (a fragment of) weighted first-order logic. **In conclusion** I still believe the connections to WMI circuits and FO logic circuits should be studied more carefully if one were to claim an atlas of algebraic circuits.

Authorsrebuttal2024-08-12

Thank you for the discussion. **On WMI and AMC** It is true that standard message-passing on discrete graphical models can be easily translated to circuits. However, the algorithm of Zeng et al. for WMI is different; for instance, integration is not even efficient if the diameter of the tree is large. As far as we are aware, a circuit interpretation of their algorithm does not currently exist, and would involve novel research. In particular, in Zuidberg et al., the *logical* (Boolean) part of the WMI problem is compiled to a circuit (which, as an AMC problem, is captured in our atlas), and has the standard exponential-in-treewidth complexity; while the *integration* uses an external symbolic integration engine (PSI-Solver). **On first-order circuits** We think the reviewer may be confusing propositional variables, of which we have many, with logical variables, of which we have none. Logical variables, when instantiated as arguments to predicates, give rise to propositional variables in the form of atoms in the work of Van den Broeck et al. Our work has no predicates and no logical variables, and is therefore only tangentially related. Our semiring operations, such as max, are over propositional variables, not logical variables. In weighted first-order model counting as defined in Van den Broeck et al., one sums a *product of weights* over *grounded predicates satisfying the first-order formula*. In contrast, in algebraic circuits, we sum/aggregate a *function* over *assignments of the propositional variables*.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC