{"id":"db8d0938-e86b-40c4-a74b-028a32b088f2","arxiv_id":"2412.19836","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Parametric reduced-order models built by least-squares projection, including POD, reduced basis methods, and Gaussian process emulation, can be viewed as conditional expectations in a Bayesian updating framework.","lead":"This paper argues that many common ways of building reduced-order models, where a complex simulation is replaced by a low-dimensional approximation, are mathematically the same as computing a conditional expectation, the average of an unknown quantity given the data. The value is a unified view that lets engineers and statisticians treat modeling and numerical errors as extra uncertainties in one Bayesian framework.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (46) is true by stipulation: Eg in Eq. (44) is defined as the argmin of the same squared loss that defines the ROM, so the Bayesian-update identification is a reformulation, not a theorem, unless S_g is shown to be L^2(sigma(z)).","rationale":"The paper's Hilbert-space machinery in Sections 2 and Appendix B is standard and coherent; the linear-map encoding, RKHS, KLE, and SVD facts are correctly assembled and are genuine useful connections. The concern is not with those technical steps but with the central interpretive leap. The reader's strongest claim states that every least-squares ROM is an approximate Bayesian update. The paper supports this only through the generalized definition in Eq. (44), which is a definitional choice: if Eg is defined as the minimizer of the same loss that defines the ROM, then Eq. (46) is a tautology. If instead Eg is meant as the classical Kolmogorov conditional expectation of Section 3.2, then Eq. (46) requires exhibiting a sigma-algebra sigma(z) such that the ROM equals E(r(mu)|sigma(z)). The concrete two-point counterexample demonstrates that this cannot hold for a generic subspace projection: the only sigma-algebras on a two-point space yield either the prior mean or the full function, while the projection loses one coordinate without being constant. Thus the central claim is either a stipulated interpretation or an unproven theorem, and the 'Bayesian update' interpretation carries no predictive content unless the missing sigma-algebra or an explicit restriction on S_g is supplied. The Frechet extension in Eq. (54) is also asserted without the required norm or well-posedness analysis, but that is secondary. This assessment reinforces, rather than overturns, the reader's CONDITIONAL verdict: the paper should clarify the definitional status of Eq. (46), state conditions under which the ROM is a genuine conditional expectation, and ideally supply a worked example demonstrating the uncertainty-augmented workflow.","tokens_in":28282,"tokens_out":4838,"duration_ms":46632,"concrete_test":"Take P={0,1} with uniform probability, r(0)=(1,0), r(1)=(0,1) in U=R^2, and let the ROM be the orthogonal projection onto U_a=span{(1,0)}, so ra(0)=(1,0) and ra(1)=(0,0). Enumerate all sub-sigma-algebras of P: the trivial one and the full one. The corresponding Kolmogorov conditional expectations are E[r]=(1/2,1/2) and r(p), neither of which equals ra(p). Hence no random variable z satisfies ra=E(r|sigma(z)); Eq. (46) fails under the standard CEX definition. If confirmed, the claim that every least-squares ROM is a CEX is an artifact of the generalized definition in Eq. (44), not a consequence of Kolmogorov conditional expectation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 defines the 'generalized conditional expectation' Eg(x|z) in Eq. (44) as the argmin over S_g(z) of Psi_x(chi), with Psi_x the squared distance. Section 4.2 then observes that a parametric ROM solves the identical minimization (Eq. (45)) and concludes ra(mu)=Eg(r(mu)|z) in Eq. (46). This makes the central identification true by construction: whatever the ROM is, one can call it a CEX by choosing S_g(z) to be the ROM's approximation manifold. What would give the claim substance is a proof that S_g(z)=L^2(Omega, sigma(z)) for some observable z, i.e., that the ROM is the Kolmogorov conditional expectation of r(mu) given z. No such proof or construction of z is given. For a generic finite-dimensional subspace projection, no such z exists: conditional expectations are determined by sigma-algebras, and a projection onto an arbitrary subspace is generally not a function of any coarser observation that loses exactly the orthogonal component. The Frechet extension in Eq. (54) is asserted with no norm specified for the operator loss Psi_RN and no existence or uniqueness argument, but the more load-bearing issue is that Eq. (46) relabels the ROM construction rather than establishing a Bayesian connection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unifying probabilistic interpretation of parametric reduced-order models (ROMs): any ROM obtained by a least-squares projection can be seen as a conditional expectation (CEX) of the full-order parametric map. Section 2 recalls the encoding of parametric objects r(p) into a linear map R, leading to RKHS, correlation operators, SVD/KLE expansions, and reduced representations. Section 3 reviews the variational definition of conditional expectation as an orthogonal projection and its role in Bayesian updating and filtering. Section 4 introduces a 'generalized conditional expectation' Eg defined as the squared-loss minimizer over a user-chosen manifold Sg(z), then claims that standard ROM constructions—GPE/Kriging, reduced basis methods, POD, truncated SVD, and low-rank tensor approximations—are instances of Eq. (46), ra(μ)=Eg(r(μ)|z). The paper also discusses how modeling, discretization, numerical, and reduction errors could be incorporated as additional random variables in the loss. The concluding sections acknowledge that the parametric map is assumed single-valued and that multiple-solution/bifurcation cases are deferred to future work.","tokens_in":28526,"tokens_out":3431,"duration_ms":34149,"significance":"If the central identification were a theorem, it would provide a genuinely useful bridge between model reduction and Bayesian inference, with the potential to justify probabilistic error models in ROMs and to motivate non-quadratic loss functions. The paper's background material is sound and competently assembled: the linear-map encoding, RKHS construction, correlation operator, SVD/KLE expansion, and the variational characterization of conditional expectation are standard and correctly stated. The concrete illustrations for GPE, RBM, POD, and low-rank tensors are instructive and show the broad intended scope. However, the load-bearing claim, Eq. (46), is not substantiated as stated; it is a definitional re-labeling rather than a proven equivalence. The paper would be acceptable as an interpretive/survey contribution if that status were made explicit, but in its current form it overstates the strength of the connection between standard ROM projection and Kolmogorov conditional expectation.","major_comments":[{"comment":"The central identification ra(μ)=Eg(r(μ)|z) is true by stipulation rather than by proof. In Eq. (44), Eg(x|z) is defined as the argmin of the squared-loss functional Ψx over an arbitrary manifold Sg(z); in Eq. (45) the ROM loss is the same functional; hence Eq. (46) restates the definition of the ROM as a minimization over its own approximation manifold. This does not establish that the ROM is a conditional expectation in the Kolmogorov sense, because no construction is given of a random variable z and a σ-algebra σ(z) such that Sg(z)=L²(Ω,σ(z)). To make the claim substantive, the paper must either prove that such a z exists for the ROM methods discussed, or explicitly redefine Eg as a new, non-Kolmogorov object and then avoid calling the result a 'Bayesian update'.","section":"Section 4.2, Eqs. (44)–(46)"},{"comment":"For projection-based ROMs such as POD, RBM, and truncated SVD, no observation z is specified. A genuine conditional expectation is always a function of the conditioning σ-algebra; a projection onto an arbitrary finite-dimensional subspace of L² is generally not realizable as conditioning on any coarser observation. Without exhibiting z (or showing that the chosen subspace equals L²(Ω,σ(z)) for some z), the assertion that every least-squares ROM 'may be seen as the CEX of r(μ)' is an analogy, not a mathematical equivalence. The paper should either provide such a construction for representative cases or temper the claim to a formal analogy, which is still of expository value.","section":"Section 4.3, 4.4; Section 4.2, paragraph after Eq. (46)"},{"comment":"The Fréchet-type extension in Eq. (54) is asserted without specifying the norm ∥·∥_L on L(Q,R^N). The accompanying text states that this is 'not a Hilbert norm', but no precise definition of the operator norm or of the underlying space is given, and no existence or uniqueness argument is provided for the argmin over the non-subspace manifold Sn,SVD(Zn). Since Eq. (54) is presented as a CEX analogue, this gap is load-bearing for the claim that the truncated SVD is an instance of the framework; either the norm should be specified (with existence/uniqueness addressed) or the statement should be clearly marked as heuristic.","section":"Section 4.4, Eq. (54)"}],"minor_comments":[{"comment":"The notation is inconsistent between R in Eq. (20) and RT in subsequent equations (e.g., Eq. (24) defines Ra via ra(p)^T w, while earlier the map is called R). The paper should standardize the notation for the linear map and its transpose.","section":"Section 2.6, Eq. (24) and surrounding text"},{"comment":"The displayed formula κN(μ1,μ2)=rN(μ1)^T rN(μ1) should presumably be rN(μ1)^T rN(μ2); otherwise the kernel is not symmetric in its two arguments.","section":"Section 4.3, Eq. (49)"},{"comment":"The dimension of KGPE is stated as R^{m×n}, but it should be R^{m×m}; also g(μ) is an m-vector so Eq. (51) is consistent with an m×m matrix.","section":"Section 4.3, before Eq. (51)"},{"comment":"There are several typos and missing words: 'tey' for 'they' in Example 1 of Section 2, 'in in' in Section 2.6, 'te' for 'the' near the end of Section 3, and 'by by' in Section 3. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The notation Et,F in Eq. (54) and Et in Eqs. (55), (56), (58), (59) is not consistently defined or explained; in particular, what the subscript F refers to is unclear. Please define all expectation operators explicitly before using them in the minimization statements.","section":"Section 4.4, Eqs. (54)–(56)"}],"recommendation":"major_revision","confidential_remarks":"This is an expository/perspective paper whose main theorem-as-stated is a tautology; the genuinely useful content is the collection of examples and the variational framing. If the journal accepts interpretive papers, a major revision that explicitly reframes Eq. (46) as a formal analogy (or provides a real construction of z) could make it publishable. The author's prior work is cited extensively; the contribution relative to those papers should be clarified in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good conceptual synthesis, not a new-results paper. The idea to view least-squares ROMs as conditional expectations and use that to fold modeling, discretization, and reduction errors into a Bayesian loss is clearly presented, and the survey of GPE, RBM, POD, and low-rank tensor methods is genuinely useful. The Hilbert-space projection framework and the linear-map encoding from the author's earlier work are laid out carefully, and the self-citations are appropriate. The soft spot is the status of Eq. (46). The generalized conditional expectation Eg is defined in Eq. (44) as the argmin over S_g of the squared loss. A ROM is itself the argmin of a squared loss over its approximation manifold, so Eq. (46) holds by construction. Calling the ROM a conditional expectation is a legitimate reading, but it does not establish a Bayesian connection unless S_g is shown to be L^2(Omega, sigma(z)) for some observation z. For a generic subspace projection no such z generally exists, and the paper gives no construction. So the central claim is a reformulation rather than a theorem. The Frechet extension in Eq. (54) is asserted without a norm on the operator loss or existence and uniqueness arguments, which is a secondary weakness. The author is upfront about the single-valued map assumption and defers bifurcations; that is honest. The paper has no numerical experiments, code, or data, but that is acceptable for an expository theory paper. The math that is there is standard and correctly applied. I would send it to peer review because the synthesis is worth having, but I would require the authors to qualify Eq. (46) as a definitional identification and, ideally, to give conditions under which a subspace projection is a genuine Kolmogorov conditional expectation. As it stands it reads like a stronger claim than it is. For whom: people working in reduced-order modeling who want a Bayesian umbrella for various methods, and people in UQ who want to see where ROMs sit relative to filtering. It would make a decent reading-group discussion piece, but I would not cite it as a source of new results. Verdict: maybe accept after major revision as an expository contribution.","headline":"A clear, honest synthesis paper that re-labels least-squares ROMs as conditional expectations; the label is defensible but largely definitional.","tokens_in":638,"tokens_out":878,"would_cite":false,"duration_ms":32010,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35B30","35R30","35R60","41A45","41A63","60G20","60G60","60H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A parametric reduced-order model built by least-squares projection is a conditional expectation, so every such ROM is a Bayesian update.","keywords":["reduced order models","conditional expectation","Bayesian updating","Karhunen-Loève expansion","uncertainty quantification","machine learning","least squares"],"falsifier":"Take a parametric problem where the map $p \\mapsto r(p)$ is multi-valued, for instance a pitchfork bifurcation, and attempt to construct the ROM through Eq. (46); the conditional expectation returns a single averaged state that matches none of the branches, so the identity $r_a(\\mu)=E_g(r(\\mu)|z)$ fails exactly where the single-valued assumption breaks.","tokens_in":28007,"feed_emoji":"🎲","tokens_out":8902,"duration_ms":69217,"temperature":0.7,"pith_summary":"The paper argues that the standard way of building parametric reduced-order models—projecting the full model's solution onto a low-dimensional subspace by least squares—is mathematically the same operation as computing a conditional expectation of a random variable. Concretely, it shows that a ROM $r_a(\\mu)$ computed this way satisfies $r_a(\\mu) = E_g(r(\\mu)|z)$, so the reduced state is the best least-squares prediction of the full state given the available parameter-dependent observations. If that identification is right, then proper orthogonal decomposition, Karhunen-Loève expansions, reduced-basis methods, Gaussian process emulation, low-rank tensor approximations, and much of machine learning are all instances of Bayesian updating with a mean-square-error loss. The payoff is practical: one can add the modeling, discretization, numerical, and reduction errors that are normally ignored when a ROM is built, and treat them as additional random variables in the same least-squares loss, producing reduced models that are optimal with respect to all these uncertainties at once.","feed_headline":"Least-squares ROMs are conditional expectations","feed_subtitle":"That identity lets modelers fold modeling, discretization, and numerical errors into a ROM as random variables.","key_machinery":"The mechanism that carries the argument is the encoding of the parametric family $\\{r(p)\\}$ into a linear map $R: U \\to Q$ between Hilbert spaces, combined with the variational definition of conditional expectation as a least-squares projection. The linear map turns an unstructured parameter set into a vector space, so the correlation $C_U = R^\\dagger R$ has a singular value decomposition whose truncation—the Karhunen-Loève expansion—gives the ROM; the conditional expectation, defined as the minimizer of $\\Psi_x(\\chi) = \\|x - \\chi\\|^2$ over a subspace of functions of the observations, is the same least-squares operation with the same Galerkin/Pythagoras structure. Identifying the training/ROM sample space with the conditioning $\\sigma$-algebra produces the identity $r_a(\\mu) = E_g(r(\\mu)|z)$, which is the bridge between model reduction and Bayesian updating.","core_discovery":"The central claim is that a parametric reduced-order model obtained by least-squares projection can be identified with a conditional expectation (CEX). Encoding the parametric map $p \\mapsto r(p)$ as a linear map $R$ and truncating its singular value expansion gives a ROM $r_a(\\mu)$; the paper shows, through the variational characterization of conditional expectation as a least-squares projection onto a subspace of functions of the observed variable $z$, that Eq. (46) holds: $r_a(\\mu) = E_g(r(\\mu)|z)$. In the author's formulation, every such ROM 'may be seen as the CEX of $r(\\mu)$', meaning the reduced model is the best mean-square prediction of the full state given the information $z$ under the chosen probability measure. This makes POD, KLE, RBM, Gaussian process emulation, low-rank tensor ROMs, and least-squares trained neural networks instances of one common Bayesian updating operation, and opens the way to include all the errors in the modeling chain as random variables in the loss.","pith_inferences":["A direct testable extension: if the CEX view is right, a ROM trained on samples drawn from one parameter distribution and then evaluated on another will be suboptimal exactly by the mismatch between the conditioning measure and the target measure; reweighting the loss to the target distribution should recover the optimal ROM on a standard parametric PDE benchmark.","For bifurcation problems, the conditional-expectation ROM would produce an averaged state between branches; a decision-oriented extension the paper does not develop would replace the mean-square loss with a risk-sensitive or multi-modal loss while keeping the projection-geometry viewpoint.","The Fréchet-mean generalisation suggests that non-Hilbertian losses such as robust or Wasserstein distances could replace mean-square error in ROM construction, but the paper only notes this possibility and does not analyse it."],"forward_implications":["Any ROM built from a least-squares projection—POD, KLE, reduced basis, Gaussian process emulation, low-rank tensor approximations—can be treated as an approximate Bayesian update, so filtering, sequential updating, and posterior error estimates apply to it.","The modeling, discretization, numerical, and reduction errors that are usually ignored when building a ROM can be included as random variables in the same squared-distance loss, making the reduced model optimal against total uncertainty rather than just the FOM-to-ROM error.","The identification gives a common vocabulary for reduced-order modeling and machine learning: training a surrogate by minimizing mean-square error on samples is an approximate conditional expectation, so error analyses and uncertainty tools transfer between the two fields.","Since any factorization of $C_U = R^\\dagger R$ yields a different linear re-parametrisation with the same SVD structure, the conditional-expectation interpretation is invariant across the many ROM algorithms, not tied to one particular expansion."],"supporting_citations":[{"why":"Introduces the encoding of parametric objects into linear maps and the factorisation / re-parametrisation framework that Section 2 builds on.","marker":"[35]"},{"why":"Extends the linear-map analysis of parametric models and underlies the SVD-based representation of ROMs.","marker":"[36]"},{"why":"Supplies the variational definition of conditional expectation as least-squares projection onto a subspace, the basis for Section 3.2.","marker":"[6]"},{"why":"Frames Bayesian inverse problems through conditional expectation, connecting CEX to Bayesian updating.","marker":"[13]"},{"why":"Presents Bayesian inversion via conditional expectation and filtering, informing the CEX machinery in Section 3.","marker":"[37]"},{"why":"Companion treatment of inverse problems in a Bayesian setting, supporting the CEX-BU connection.","marker":"[38]"},{"why":"Reduced basis methods, one of the ROM classes the paper re-interprets as conditional expectation.","marker":"[45]"},{"why":"Gaussian process emulation / Kriging, the worked example in Section 4.3.","marker":"[26]"}],"fun_headline_variants":["ROMs as conditional expectations: a unified theory","Least-squares ROMs are Bayesian updates","Conditional expectation explains every LS-ROM","POD, KLE, RBM: all conditional expectations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes the parametric map $p \\mapsto r(p)$ is single-valued for every $p$, so each parameter value has exactly one state; if a parameter value admits multiple solutions, the linear map $R$, the correlation $C_U = R^\\dagger R$, and the conditional-expectation identity are not defined.","fun_headline_variants_meta":{"raw":{"variants":["ROMs as conditional expectations: a unified theory","Least-squares ROMs are Bayesian updates","Conditional expectation explains every LS-ROM","POD, KLE, RBM: all conditional expectations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000736,"raw_usage":{"total_tokens":3296,"prompt_tokens":957,"completion_tokens":2339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2280}},"tokens_in":573,"tokens_out":2339,"duration_ms":16497,"temperature":1.0,"reasoning_tokens":2280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:55:53.602229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a parametric problem where the map $p \\mapsto r(p)$ is multi-valued, for instance a pitchfork bifurcation, and attempt to construct the ROM through Eq. (46); the conditional expectation returns a single averaged state that matches none of the branches, so the identity $r_a(\\mu)=E_g(r(\\mu)|z)$ fails exactly where the single-valued assumption breaks.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the encoding of parametric objects into linear maps and the factorisation / re-parametrisation framework that Section 2 builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends the linear-map analysis of parametric models and underlies the SVD-based representation of ROMs."},{"cited_title":"Bobrowski, Functional analysis for probability and stochastic processes, Cambridge University Press, Cambridge, 2005","cited_arxiv_id":null,"evidence_quote":"Supplies the variational definition of conditional expectation as least-squares projection onto a subspace, the basis for Section 3.2."},{"cited_title":"Ibrahimbegović, ed.), Computational Methods in Applied Sciences, vol","cited_arxiv_id":null,"evidence_quote":"Companion treatment of inverse problems in a Bayesian setting, supporting the CEX-BU connection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gaussian process emulation / Kriging, the worked example in Section 4.3."}],"review_version":1}