{"id":"9ff08f8e-1bd3-4022-a123-a1dccd15e671","arxiv_id":"2501.12361","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Deflated and adaptive greedy reduced-basis algorithms that detect bifurcation points and certify multiple coexisting solution branches, validated on the Coanda effect in a sudden-expansion channel.","lead":"Two new algorithms build cheap, certified computer models of physical systems whose solutions split into several coexisting branches as a parameter varies. One algorithm finds the splitting point automatically; the other keeps every branch within a chosen error tolerance, demonstrated on a channel-flow test case.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Certification claim is unsupported: Eq. (12) is used as a guaranteed upper bound per branch, but the paper neither proves nor numerically verifies that the linear estimator dominates the true error.","rationale":"The reader's weakest assumption identifies exactly the load-bearing issue: the deflated-greedy certification inherits the single-solution BRR estimator theory, but the linear bound (12) is used without proof or numerical verification that it dominates the true error for each branch. This is the most critical point because the paper's novelty claim is precisely 'certifiable' multiple branches; if the estimator is only an indicator, the contribution reduces to an accurate but uncertified ROM, which is a considerably weaker statement. I agree with the reader that this warrants a CONDITIONAL verdict rather than rejection: the numerical campaign is credible, the actual errors shown in Figure 7 are small, and the authors are honest about limitations such as the failure of tau_N(mu) <= 1 in Appendix A and the absence of hyper-reduction in Remark 3. The secondary concern about the adaptive-greedy inf-sup heuristic is also valid but less central, since even a failure of automatic mu* detection would not invalidate the deflated-greedy machinery if mu* is known or supplied. The proposed concrete test is straightforward and decisive: if the estimator exceeds the true error everywhere, the certification claim is supported for the benchmark; if not, the paper should be revised to claim accuracy rather than certification. Since the reader already chose CONDITIONAL, my assessment does not change that verdict.","tokens_in":23878,"tokens_out":4826,"duration_ms":50168,"concrete_test":"With the authors' settings (Nmax = 35, eps = 1e-3, |Ph| = 51, mu0 = 2, final basis N = 25), rerun the deflated-greedy offline and, for every branch i and every mu in the test set Pte, compute in the same norm the linear estimator (12) evaluated at the reduced solution B u_N^i(mu) and the true error ||u_h^i(mu) - B u_N^i(mu)||. Then verify Delta_lin_N(mu) >= true error for all i and mu, and produce the estimator-versus-error overlay that is missing from Figure 7. If any violation occurs, the certification claim as stated is false and should be weakened to accuracy-without-guarantee; if no violation occurs across all 151 test parameters and all 3 branches, the concern is resolved for this benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: Sections 6.3 and 7 state that deflated-greedy certifies all admissible branches with errors below eps = 1e-3. This rests on the linear estimator Delta_lin_N(mu) = ||G(Bu_N; mu)|| / beta^h_N(mu) (Eq. 12), computed separately for each reduced branch in Algorithm 7. For nonlinear Navier-Stokes, Eq. (12) is not a proven upper bound: writing the residual identity G(u_N) - G(u_h) = J(u_N)e + Q(e) gives beta||e|| <= ||G(Bu_N)|| + C||e||^2, which does NOT imply ||e|| <= ||G||/beta. The rigorous BRR estimator (11) requires tau_N(mu) <= 1, which Appendix A reports is never met in the bifurcating regime; the paper switches to (12), whose reliability is exactly what the certification claim needs. The paper neither shows that each reduced branch solution lies in the local uniqueness ball of the corresponding HF branch, nor plots estimator against true error. Figure 7 reports only actual errors E(mu), and Figure 9 shows estimator values at one iteration without overlaying them with E(mu). Furthermore, Section 6.3 concedes that at the bifurcation point on the symmetric branch the certification is violated because reduced Newton converged to a different branch, which already conflicts with a literal reading of the 'all branches, all parameters' claim. Thus the headline claim is currently a statement about an unvalidated surrogate, not a demonstrated guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two RB greedy algorithms for bifurcating parametric nonlinear PDEs: an adaptive-greedy that refines the training set near the detected bifurcation point, and a deflated-greedy that discovers multiple coexisting branches via reduced deflation and computes branch-wise error estimators. The methods are tested on the Coanda effect governed by steady Navier-Stokes in a sudden-expansion channel, where three branches coexist below the pitchfork point (mu* ≈ 0.96). The authors report that the deflated-greedy achieves actual reduced errors below epsilon = 1e-3 for all branches and parameters, except for one documented violation at the bifurcation point on the symmetric branch, and that the adaptive-greedy detects mu* starting from a coarse four-point training grid. The paper argues these are the first greedy-based strategies for certified approximation of bifurcating PDEs.","tokens_in":24123,"tokens_out":3830,"duration_ms":37371,"significance":"If the certification claim is substantiated, the work is a meaningful advance: prior POD-based ROMs for bifurcations require a-priori knowledge of the branching structure and do not provide error bounds, while vanilla-greedy misses coexisting branches. The numerical study is honest and fairly thorough: results are compared against independently computed high-fidelity solutions and against the literature value mu* ≈ 0.96, POD and vanilla-greedy baselines are included, and Appendix A explicitly documents that the nonlinear BRR estimator is inapplicable (tau_N > 1) in the bifurcating regime. The main contribution, deflated-greedy with branch-wise estimators, is conceptually novel and clearly presented. However, the central word 'certified' currently rests on an unvalidated linear estimator, and the paper itself concedes an exception to the 'all branches, all parameters' claim, so the significance can only be assessed after these gaps are closed.","major_comments":[{"comment":"The branch-wise certification is based entirely on the linear estimator Delta_lin_N(mu) = ||G(Bu_N; mu)|| / beta^h_N(mu). For a nonlinear PDE this is not a proven upper bound: the residual identity yields beta||e|| <= ||G(Bu_N)|| + C||e||^2, which does not imply ||e|| <= ||G||/beta. The rigorous nonlinear estimator (11) requires tau_N(mu) <= 1, and Appendix A reports that this condition is never met in the bifurcating regime. The paper neither proves that each reduced branch solution lies in the local uniqueness ball of the corresponding HF branch nor provides numerical verification that Delta_lin dominates the true error. Since the headline claim in Sections 6.3 and 7 is that the deflated-greedy 'certifies all branches', this missing reliability check is load-bearing and must be addressed.","section":"Section 5.2, Eq. (12)"},{"comment":"The text states that 'for all branches and all parameters in Pte, the reduced error is reliably below the tolerance bound', and then immediately concedes: 'The unique point in the parametric space that seems to violate the certification is indeed the one corresponding to the bifurcation point while reconstructing the symmetric branch'. This is an explicit contradiction of the literal 'all branches, all parameters' claim. At minimum, the claim must be restricted to parameters outside a neighborhood of mu*, or the algorithm must be modified so that the symmetric branch at mu* is also certified. As written, the certification statement in Section 7 ('can certify all the branches') is inaccurate.","section":"Section 6.3, Figure 7g"},{"comment":"The adaptive-greedy detection of mu* relies on the assumption that 'for large enough N, the reduced inf-sup (9) is a good approximation of the high-fidelity inf-sup (5)', but no convergence statement or quantitative bound is provided. Since 'detecting the bifurcation point starting from scarce information' is a claimed contribution, the paper should either supply a convergence analysis for mu_bif as N increases or explicitly label the detection as a heuristic and validate it with a systematic study (e.g., showing that mu_bif approaches mu* as N grows). The current evidence is a single numerical example, which is suggestive but not sufficient for a general claim.","section":"Section 4, Eq. (9)"},{"comment":"The certification claim concerns the estimator, yet the paper never plots the estimator against the true error. Figure 7 reports only the actual error E(mu), and Figure 9 shows estimator values at one iteration without overlaying them with E(mu). A direct empirical check would be to plot Delta_lin_N(mu) (or the branch-wise max Delta) together with E(mu) on the test set Pte for the final deflated-greedy basis, for each of the three branches. Such an overlay would either confirm that the estimator dominates the true error or expose the gap that currently undermines the word 'certified'.","section":"Section 6.3, Figures 7 and 9"}],"minor_comments":[{"comment":"The relative error E(mu) is defined as a sum of velocity and pressure relative errors rather than a norm of the combined error; since the two fields have different scales, this additive combination should be justified or replaced by a single norm to avoid masking large errors in one component.","section":"Section 6.3, Eq. (24)"},{"comment":"The procedure for handling the case where the same parameter maximizes the estimator multiple times is described only in prose; a small pseudocode listing (e.g., a while loop over the sorted estimator values) would make the algorithm unambiguous and easier to reproduce.","section":"Remark 2, Section 5.2"},{"comment":"The inset panel in Figure 5 is very small and the axes labels are difficult to read; enlarging the inset and using clearly marked ticks would improve readability.","section":"Figure 5"},{"comment":"The sentence 'for some problems, e.g., the bifurcating ones, such a condition could be difficult to obtain, and the linear estimator is exploited for all greedy iterations' is vague; the paper should specify in advance which problems and under which conditions the linear estimator is used, and state that it is not a proven bound in general.","section":"Section 2.3, Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is largely honest: Appendix A documents the failure of the nonlinear estimator in the bifurcating regime, and Section 6.3 admits a violation of the certification claim at the bifurcation point. These admissions are to the authors' credit, but they also reveal that the abstract's 'certifying multiple coexisting branches simultaneously' overstates what is demonstrated. The reference list is heavily weighted toward the authors' prior work; while those works are relevant, the novelty claims would be strengthened by engaging with a broader set of recent external works on branch-wise and local ROMs for bifurcations. The core idea is promising and the numerical evidence is supportive, so I recommend major revision rather than rejection, conditional on closing the estimator-reliability gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is the first greedy-based ROM framework I know of that detects a pitchfork bifurcation without prior knowledge and tries to certify each coexisting branch. The adaptive-greedy (refine training set around the minimizer of reduced inf-sup) and deflated-greedy (reduced deflation plus per-branch error estimators) are genuinely new. And the paper is honest: Appendix A shows the BRR nonlinear estimator fails (tau_N > 1) in the bifurcating regime, and Remark 3 flags the absence of hyper-reduction. The numerical campaign on the Coanda effect is credible: errors for all three branches sit below 1e-3 away from mu*, adaptive detection lands near 0.96, and the POD/vanilla-greedy comparisons show the expected failures.\n\nThe load-bearing problem is the certification claim in Sections 6.3 and 7. The linear estimator (12) is not proven to be an upper bound in the multi-branch setting. The residual identity gives beta||e|| <= ||G|| + C||e||^2, which does not imply ||e|| <= ||G||/beta. The paper neither proves each reduced branch lies in the local uniqueness ball of the corresponding HF branch nor plots estimator against true error. The lone concession in Section 6.3—that at mu* the symmetric-branch certification fails because reduced Newton converged to another branch—already conflicts with the literal 'all branches, all parameters' claim. So 'certifies all branches' should be read as 'errors stay below tolerance in the numerics', not as a proven guarantee. The adaptive-greedy detection also rests on the heuristic that the reduced inf-sup minimizer tracks mu*, with no convergence statement. Minor: 'mu_guess' in Algorithm 4 is undefined, and the mesh is not specified.\n\nNone of this kills the contribution. The algorithms are sensible, the evidence supports their behavior, and the limitations are disclosed. But a serious referee should ask for either a proof sketch of the estimator bound under a local uniqueness assumption per branch, or a direct estimator-vs-error verification plot (and an honest rewording of the certification claim).\n\nFor anyone working on ROM for bifurcation problems, this is worth reading and citing. I would send it to review—the idea is important and the honest numerical work deserves referee time. My own verdict would be conditional acceptance, with the certification claim tightened.","headline":"A genuinely new certified-ROM idea that deserves refereeing, but the headline branch-certification claim is not actually established by the numerics—the linear estimator is used as a guaranteed bound without proof or verification.","tokens_in":24761,"tokens_out":1962,"would_cite":true,"duration_ms":19679,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N15","37M20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a deflated-greedy reduced-basis algorithm can certify the approximation error on every coexisting branch of a bifurcating nonlinear PDE, without prior knowledge of the bifurcation.","keywords":["reduced basis methods","greedy algorithm","deflation","bifurcation","a posteriori error estimation","Navier-Stokes equations","Coanda effect","model order reduction"],"falsifier":"At the $N=25$ deflated-greedy configuration reported in Section 6.3, compute the relative error $E(\\mu)$ and the linear estimator $\\Delta_{\\rm lin}^N(\\mu)$ on the full test set for each of the three branches; if the estimator falls below $E$ for any branch in the bifurcating regime, the certification claim is false. A complementary check is to run the same algorithm on a pitchfork bifurcation with an odd number of unequal branches and see whether reduced deflation actually converges to every branch.","tokens_in":23577,"feed_emoji":"🌊","tokens_out":10485,"duration_ms":97460,"temperature":0.7,"pith_summary":"This paper tries to establish that the reduced-basis greedy framework can be extended to nonlinear parametric PDEs whose solution manifold branches, removing the two obstacles that have kept certification out of reach: the need for a priori knowledge of the coexisting states, and the breakdown of the error estimator near bifurcation points. It introduces the adaptive-greedy algorithm, which starts from a coarse parameter grid and uses the parameter minimizing the reduced inf-sup constant to refine the sampling until the bifurcation point is located. It then introduces the deflated-greedy algorithm, which uses deflation to discover several coexisting snapshots at the same parameter value, builds reduced solutions on every branch, and computes a separate error estimator for each of them. In the Coanda-effect channel-flow test case, where the Navier-Stokes equations admit three coexisting solutions, the paper reports that only the deflated-greedy strategy keeps the reduction error below the tolerance on all branches.","feed_headline":"Deflated greedy method certifies every branch of a bifurcating PDE","feed_subtitle":"In a Coanda-effect channel flow, only the deflated strategy keeps all three coexisting branches under 1e-3 error","key_machinery":"The machinery has two load-bearing parts. The first is the reduced inf-sup constant $\\beta_h^N(\\mu)$, the stability factor of the reduced Jacobian evaluated at a reduced solution, which the adaptive-greedy uses as a bifurcation detector and the estimator uses as a denominator; its vanishing at the critical parameter is exactly why standard certification fails near $\\mu^*$. The second is the deflation operator $\\mathcal{M}(y,u) = (\\|y-u\\|^r + \\sigma)^{-1} I$, which multiplies the residual so that Newton iterations are repelled from already discovered solutions, allowing the algorithm to collect multiple coexisting states at one parameter value. The deflated-greedy wires these together: it computes one estimator value per reduced branch solution, selects the parameter and branch with the largest estimator, adds the corresponding high-fidelity snapshot to the basis, and then deflates the high-fidelity system to find any remaining coexisting snapshots at that same parameter. Continuation gives the reduced Newton solver initial guesses from the previously computed parameter, so the branch identity is carried from one parameter step to the next.","core_discovery":"The central claim is that both known workarounds—POD with a complete, hand-chosen snapshot set, and vanilla-greedy restricted to a single branch—can be replaced by algorithms that need no prior knowledge of the bifurcation. The adaptive-greedy treats the reduced inf-sup constant $\\beta_h^N(\\mu)$ as a stability indicator: because the high-fidelity inf-sup constant vanishes at the critical parameter $\\mu^*$, the parameter that minimizes the reduced constant is taken as an approximation of the bifurcation point, and new training points are inserted around it. The deflated-greedy applies the deflation operator $\\mathcal{M}(y,u) = (\\|y-u\\|^r + \\sigma)^{-1} I$ at both levels: at the high-fidelity level it collects multiple coexisting snapshots for the same parameter, and at the reduced level it solves for reduced solutions on different branches. For each reduced branch solution it evaluates the linear a posteriori estimator $\\Delta_{\\rm lin}^N(\\mu) = \\|G(B u_N; \\mu)\\|/\\beta_h^N(\\mu)$, then enriches the basis with the snapshot belonging to the worst-approximated branch. For the sudden-expansion channel, the paper reports that with $N=25$ basis functions the deflated-greedy keeps the relative reduction error below $\\varepsilon=10^{-3}$ for all three branches, whereas POD and vanilla-greedy approximate only the symmetric branch and lose the asymmetric ones.","pith_inferences":["The paper does not plot the estimator against the true error for each branch; verifying the effectivity ratio $\\Delta_{\\rm lin}^N(\\mu)/E(\\mu)$ branch by branch would be the natural next check before applying the certification claim to a new problem.","The authors note the adaptive strategy could serve as a preprocessing step for deflation, restricting deflated solves to the multi-solution regime; implementing that coupling would directly reduce the offline cost of the deflated-greedy.","Because the framework is advertised as agnostic to the type of bifurcation, a natural test is a transcritical or Hopf bifurcation, where branches are not created in symmetric pairs and reduced deflation may or may not converge to every admissible state.","If the branch-wise certification transfers to other PDEs, the same one-estimator-per-discovered-branch structure could be applied to time-periodic or stochastic bifurcations, where coexisting attractors play the role of coexisting steady solutions."],"forward_implications":["If the claim is right, a reduced model built by deflated-greedy can be used to query any parameter in the trained range and return an accurate solution on any branch the user asks for, not just the stable branch.","The adaptive-greedy result means offline sampling can begin from a very coarse parameter grid, with the computational budget concentrated around the detected bifurcation point instead of spread over the uniqueness region.","Because the same parameter value can contribute several basis functions, the reduced basis encodes qualitatively different states, which is exactly the information that vanilla-greedy's estimator-driven sampling misses.","The deflated estimator's maximum over branches gives a stopping criterion that is meaningful for the whole solution ensemble, so convergence of the reduced model cannot be declared while an entire branch is still unrepresented."],"supporting_citations":[{"why":"Supplies the deflation method that the high-fidelity and reduced solvers use to discover multiple coexisting solutions for the same parameter.","marker":"[21]"},{"why":"Provides the nonlinear a posteriori error-estimation theory from which the linear estimator used for certification is derived.","marker":"[8]"},{"why":"Supplies the standard reduced-basis greedy framework and error-bound theory that the two new algorithms extend.","marker":"[51]"},{"why":"Justifies the use of the linear error bound for parametrized Navier-Stokes flows and supplies the non-bifurcating comparison test in the appendix.","marker":"[37]"},{"why":"Documents the branch-wise and global POD approaches for bifurcating PDEs that the new algorithms are compared against.","marker":"[40]"},{"why":"Provides POD-based detection strategies for steady bifurcations in the Coanda-effect setting, the baseline the adaptive detector is designed to beat.","marker":"[48]"},{"why":"Establishes the symmetry-breaking pitchfork bifurcation in the sudden-expansion channel that serves as the numerical test case.","marker":"[50]"},{"why":"Supplies one of the references for the known bifurcation point value $\\mu^*\\approx 0.96$ against which the adaptive detector is assessed.","marker":"[41]"}],"fun_headline_variants":["Deflated-greedy certifies every coexisting branch of a bifurcating PDE","Adaptive greedy detects bifurcation point without prior knowledge","Deflation turns greedy method into a certified all-branch solver","Coanda flow: deflated greedy keeps all three branches under 1e-3"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The certification depends on the standard single-solution error bound still being a true upper bound when applied branch by branch, which requires each reduced branch solution to stay close enough to the corresponding true branch that the single-solution uniqueness argument applies to that branch.","fun_headline_variants_meta":{"raw":{"variants":["Deflated-greedy certifies every coexisting branch of a bifurcating PDE","Adaptive greedy detects bifurcation point without prior knowledge","Deflation turns greedy method into a certified all-branch solver","Coanda flow: deflated greedy keeps all three branches under 1e-3"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3546,"prompt_tokens":1076,"completion_tokens":2470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":2391}},"tokens_in":692,"tokens_out":2470,"duration_ms":17174,"temperature":1.0,"reasoning_tokens":2391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:15:20.803363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At the $N=25$ deflated-greedy configuration reported in Section 6.3, compute the relative error $E(\\mu)$ and the linear estimator $\\Delta_{\\rm lin}^N(\\mu)$ on the full test set for each of the three branches; if the estimator falls below $E$ for any branch in the bifurcating regime, the certification claim is false. A complementary check is to run the same algorithm on a pitchfork bifurcation with an odd number of unequal branches and see whether reduced deflation actually converges to every branch.","supporting_citations":[{"cited_title":"Brezzi, J","cited_arxiv_id":null,"evidence_quote":"Provides the nonlinear a posteriori error-estimation theory from which the linear estimator used for certification is derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the branch-wise and global POD approaches for bifurcating PDEs that the new algorithms are compared against."},{"cited_title":"Pitton, A","cited_arxiv_id":null,"evidence_quote":"Provides POD-based detection strategies for steady bifurcations in the Coanda-effect setting, the baseline the adaptive detector is designed to beat."},{"cited_title":"Quaini, R","cited_arxiv_id":null,"evidence_quote":"Establishes the symmetry-breaking pitchfork bifurcation in the sudden-expansion channel that serves as the numerical test case."}],"review_version":1}