{"id":"4516e5dc-9cb0-40d2-8eba-93d569cf913a","arxiv_id":"1908.07247","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A bounded-variable nonlinear least-squares solver with sparse recursive QR updates solves parameter-varying MPC problems without rebuilding the optimization problem at runtime.","lead":"This paper describes a fast optimization method for model predictive control on small embedded computers. It removes the step of building a new optimization problem when the model or controller settings change, and tests show large speedups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Section 3 admits the single-large-penalty guarantee is only proven for LTI models; for the general nonlinear model (1) the extension is explicitly stated to be non-immediate and non-deterministic, so the central claim that the solver solves (6) and thereby (5) is not yet supported.","rationale":"The reader identified exactly this gap as the weakest assumption: the penalty-parameter guarantee is proven only for linear models in [29, Theorem 1], and the manuscript explicitly flags the nonlinear extension as risky and non-deterministic. I agree with that reading. The strongest claim about eliminating construction and handling linear, nonlinear, and adaptive MPC depends on (6) being a trustworthy surrogate for (5). The paper's own Section 3 acknowledges that no deterministic bound exists for the general model (1), and Theorem 1 only establishes convergence for the penalized problem. The numerical example with fixed \\sqrt{\\rho}=10^4 and residual 3e-5 is encouraging but is a single, uncontrolled illustration rather than evidence of a uniform bound or scaling rule. This is a correctness risk, not a matter of differing from consensus; the reformulation may work in practice for many problems, but the central claimed generality is not established. Therefore the CONDITIONAL verdict is appropriate: the contribution is plausible and the algorithmic machinery is well-developed, but acceptance should be conditional on closing or narrowing this gap, e.g., by giving a deterministic penalty bound for a subclass of nonlinear models or adding a numerical study that systematically sweeps \\rho and reports Jacobian conditioning. The proposed concrete test directly probes whether the single large-penalty step can simultaneously achieve constraint satisfaction and numerical well-posedness.","tokens_in":21408,"tokens_out":1821,"duration_ms":16767,"concrete_test":"Construct or identify a small nonlinear MPC benchmark (e.g., a two-state nonlinear system with output constraints and Np=Nu=10) and run Algorithm 1 on formulation (6) for a geometric sequence of penalty values \\rho = 10^2,10^4,10^6,10^8. Record the worst-case equality residual max_k\\|h_k(z_k,\\phi_k)\\| and the condition number (or reciprocal condition estimate) of the BVLS Jacobian J=[W;Jh] at the solution. If the required residual accuracy (say <=1e-6) is only reached at a \\rho for which the subproblem condition number exceeds ~1/eps_machine, the single-penalty reformulation is numerically untenable for that problem; if a safely scaled \\rho achieves the target residual across all tested benchmarks, the concern is substantially refuted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that one construction-free BVNLLS optimizer solves linear, nonlinear, and adaptive MPC by replacing the equality constraints h_k(z_k,\\phi_k)=0 in problem (5) with a single quadratic penalty 1/(2\\rho)\\|h_k\\|^2 in problem (6). The paper itself (Section 3) limits the supporting theoretical guarantee to the linear case: it cites [29, Theorem 1] for quadratic cost with linear equalities, and states the extension to general nonlinear h_k 'is not immediate and thereby poses a risk given that the bound is not deterministic.' No theorem in this manuscript proves that for a fixed, finite \\rho the solution of (6) is near the feasible optimum of (5), or that a \\rho large enough to make the residual small can be chosen while keeping the least-squares Gauss-Newton system numerically full-rank. Algorithm 1 is a Gauss-Newton method for (6); its global convergence (Theorem 1) is convergence to a stationary point of the penalized problem, not a solution of the constrained problem (5). The numerical CSTR example uses \\sqrt{\\rho}=10^4 and reports equality residuals below 3e-5, but this is one example with one tuning value, and Figure 3 shows residual levels that are acceptable empirically but provide no scaling law or worst-case bound. Thus the load-bearing assumption for nonlinear/adaptive MPC is that a sufficiently large, fixed, well-scaled \\rho exists for the intended problem class; the manuscript provides no deterministic argument for this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an embedded-MPC solver that avoids the usual online construction of the optimization problem. The constrained MPC problem (5) is replaced by a box-constrained nonlinear least-squares problem (6) in which the model equality constraints are eliminated by a single large quadratic penalty with parameter rho. The authors then present a bounded-variable nonlinear least-squares algorithm (BVNLLS, Algorithm 1), an operator abstraction of the Jacobian so that the solver does not need to form or store problem matrices (Algorithms 2 and 3), and sparse recursive thin QR update rules for the inner active-set least-squares solver. Numerical results on a CSTR example report large speedups over fmincon and IPOPT for increasing prediction horizons. The paper's own Section 3 limits the theoretical support for the single-penalty reformulation to linear models, and Section 4.2 proves only an Armijo decrease property rather than a full global convergence theorem. The benchmark comparison warm-starts the proposed solver but does not report warm-starting the baseline solvers.","tokens_in":21783,"tokens_out":4224,"duration_ms":46969,"significance":"If the central claims held, the paper would be a useful contribution to embedded MPC: the operator abstraction in Section 5 genuinely eliminates a repeated problem-construction step for parameter-varying models, the sparse QR structure analysis in Section 6 is nontrivial, and the library-free C implementation with the reported speedups would be attractive for real-time applications. The paper also gives a clean proof of the equivalence between the bound-constrained Lagrangian method and an NLLS-box problem (Proposition 1), and it is honest about the provenance of several results in the authors' earlier conference papers. However, the significance is currently limited by the gap between what is proved and what is claimed: for nonlinear models the key equivalence between (5) and (6) has no deterministic penalty bound, and the 'global convergence' theorem does not establish convergence to a constrained optimum.","major_comments":[{"comment":"The reformulation of the constrained MPC problem (5) as the box-constrained penalized problem (6) is load-bearing for nonlinear and adaptive MPC, but the paper does not supply a supporting theorem for this case. The text states that the guarantee of [29, Theorem 1] applies to quadratic cost with linear equality constraints, and that the extension to general h_k 'is not immediate and thereby poses a risk given that the bound is not deterministic.' No theorem in this manuscript provides a deterministic bound on the penalty parameter rho for the general model (1), nor does any result show that a rho large enough to make the equality residual small can be chosen while keeping the Jacobian of the residual full rank. The CSTR example in Section 7 uses sqrt(rho)=1e4 and reports residuals below 3e-5, but that is one problem instance and one tuning value. As written, the paper establishes that Algorithm 1 solves the penalized problem (6), not that it solves the intended constrained problem (5). I ask the authors either to prove a deterministic penalty-bound result under explicit regularity assumptions or to restrict the claims accordingly.","section":"Section 3, Eqs. (5)-(6)"},{"comment":"The theorem is titled 'Global convergence of BVNLLS,' but the proof establishes only that for each iteration there exists a sufficiently small alpha satisfying the Armijo decrease condition (21), using a Taylor expansion. This is a local descent argument; it does not prove that the sequence of iterates converges, nor that its limit points are stationary points of (6) or (5). A standard global convergence proof would need assumptions such as boundedness of the level set, continuity of r and its Jacobian, and a limit-point argument, together with a treatment of the line-search termination. Please either rename and qualify the result (e.g., 'Armijo decrease property') or provide a genuine global convergence theorem with explicit assumptions.","section":"Section 4.2, Theorem 1"},{"comment":"The computational comparison is not symmetric. The text notes that BVNLLS 'exploits warmstarts based on the previously computed solution which is provided from the second instance onwards,' while fmincon and IPOPT are not reported to be warm-started with comparable initial guesses. Since warm-starting from the previous sampling instant is natural in MPC, the reported two-orders-of-magnitude speedup conflates algorithmic speed with initialization advantage. Please either warm-start the baseline solvers in the same way and re-report the timings, or present cold-start timings for all solvers so that the comparison isolates the algorithmic contribution.","section":"Section 7.2, Figures 4-5"}],"minor_comments":[{"comment":"The output line says 'Local minimum z* of (5)', but Algorithm 1 solves the penalized problem (6); the comment should refer to (6).","section":"Algorithm 1, Step 12"},{"comment":"Lemma 2 invokes Lemma 4 before Lemma 4 is stated; the ordering or the cross-references should be adjusted for readability.","section":"Section 4.2, Lemma 2"},{"comment":"The computation of beta and eta is described only in prose ('at most one integer division'); giving closed-form expressions would improve reproducibility and reduce the risk of indexing errors.","section":"Section 5.2, Algorithms 2-3"},{"comment":"The discussion of the predicted nonzero structure says that the predicted structure 'rarely contains indices of zero elements' and that R may become dense after reorthogonalization; this caveat is important and should be repeated prominently wherever the sparse QR speedups are claimed.","section":"Section 6.4"},{"comment":"The paper relies heavily on the authors' own prior results [29]-[31] for the penalty equivalence and for the dense BVLS solver; since those results are load-bearing, the manuscript should state precisely which parts of Sections 2-4 are newly proved here and which are quoted from [29]-[31].","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central algorithmic machinery is an extension of the authors' own conference and journal papers, and the two load-bearing pillars (the single-QPM accuracy result and the dense BVLS solver) are prior results rather than new contributions in this manuscript. The sparse recursive QR update and the operator abstraction are the genuinely new parts. The main risk for the journal is that the nonlinear/adaptive MPC claim is supported neither by a deterministic penalty-bound theorem nor by a symmetric benchmark comparison; both issues are fixable with additional analysis or honest restriction of the claims. I would support asking for a reproducibility artifact, since the numerical section refers to unpublished settings in [31] for the benchmark configuration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what's new: the operator abstraction Jix/JtiX in Section 5 and the sparse recursive thin QR update with predicted fill-in in Section 6. Those are real contributions—cleanly argued, and directly relevant to embedded and adaptive MPC where rebuilding the QP/NLP at each sample is a bottleneck. The sparsity analysis is the strongest part: Theorem 2 and Corollaries 1–3 give rigorous predicted-nonzero-structure bounds, letting the active-set solver skip zero flops without sparse-matrix indexing overhead. The CSTR benchmark is one example, but it shows order-of-magnitude gains over the dense variant and much larger gaps versus fmincon/IPOPT.\n\nThree soft spots, in increasing order.\n\nFirst, the benchmark warm-starts BVNLLS from the previous solution but starts fmincon and IPOPT cold. The headline speedup ratios are inflated by that asymmetry. The controlled comparison—custom-sparse versus dense BVNLLS—still shows a solid order-of-magnitude gain, so the qualitative claim survives, but the two-orders-of-magnitude figures should not be taken at face value.\n\nSecond, Theorem 1, titled \"Global convergence of BVNLLS,\" proves that an Armijo step exists at each iteration. That tells you the line search is well-defined; it does not prove the iterates converge to a stationary point. Standard line-search theory would supply the rest under mild assumptions, so this is overselling rather than an error—but the title should change.\n\nThird, and most important: the single-penalty reformulation (6) has a supporting theorem only for linear models, from their earlier [29]. Section 3 of this paper says the extension to general nonlinear h_k \"is not immediate and thereby poses a risk given that the bound is not deterministic.\" For the nonlinear and adaptive cases—the paper's main selling point—the claim that the solver solves (5) rests on a hand-tuned ρ and one CSTR example with sqrt(ρ)=10^4. There is no deterministic argument that a single finite ρ can simultaneously keep equality residuals small and keep the Gauss-Newton system full-rank. The authors are transparent about this, and the linear case is fully supported; but the gap is load-bearing for the nonlinear claim and should weigh heavily in a revision.\n\nOn the citation pattern: yes, the paper leans on the authors' own [29,30,31], but it says so explicitly and draws a clear line around what is new. That is honest scholarship, and the novelty claims check out.\n\nWorth reading for anyone implementing embedded or adaptive MPC without code generation. It deserves a serious referee—send it out.","headline":"The sparse QR-update machinery and operator abstraction are the real contributions; the nonlinear/adaptive claim rests on a penalty-parameter guarantee the authors themselves admit they do not have.","tokens_in":22244,"tokens_out":6674,"would_cite":true,"duration_ms":59101,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49M37","65K05","90C30","93B40"],"pacs":[],"model":"deepseek-v4-flash","headline":"By expressing MPC problem matrices as abstract functions of the model and tuning parameters, this paper eliminates the problem-construction step and lets a single bounded-variable nonlinear least-squares solver handle linear, nonlinear…","keywords":["model predictive control","bounded-variable nonlinear least squares","quadratic penalty method","active-set methods","sparse recursive QR factorization","parameter-varying systems","embedded optimization","problem-construction elimination"],"falsifier":"Run the algorithm on a nonlinear model with a stiff or poorly scaled Jacobian and increase $\\rho$ until constraint residuals meet a target; if at some $\\rho$ the least-squares subproblem becomes numerically rank-deficient (or the solver converges to a point with equality residual above the specified tolerance), the single-penalty step has failed to solve the intended constrained MPC problem. A concrete test: for the CSTR benchmark with $N_p=N_u=160$, measure the worst-case $\\|h(z,\\phi)\\|_2$ while increasing $\\sqrt{\\rho}$ from $10^2$ to $10^6$ and record the smallest $\\rho$ at which the residual stops decreasing or the solver's condition number grows beyond, say, $10^{12}$.","tokens_in":21227,"feed_emoji":"🎛️","tokens_out":6495,"duration_ms":58677,"temperature":0.7,"pith_summary":"This paper proposes an algorithm for model predictive control (MPC) that removes the usual first step of constructing a fresh optimization problem whenever the model or tuning parameters change. Instead of forming and storing the problem matrices, the algorithm represents every matrix-vector product through two abstract operators whose expressions are parameterized by the model order, horizons, and weights, so the same compiled code handles linear, nonlinear, and adaptive models. The equality constraints from the prediction model are folded into the cost as a single quadratic penalty, turning the constrained problem into a box-constrained nonlinear least-squares problem, which is solved by an active-set method with sparse recursive QR updates. A sympathetic reader would care because this promises real-time embedded MPC with a small memory footprint and no code regeneration when the controlled plant or the MPC settings change.","feed_headline":"MPC without problem construction: one least-squares solver does it all","feed_subtitle":"Expressing matrices as abstract functions lets one embedded solver handle changing models and tuning in real time.","key_machinery":"The load-bearing identity is the quadratic-penalty reformulation (6) that replaces the model equality constraints $h_k(z_k,\\phi_k)=0$ by the single penalty term $\\tfrac{1}{2\\rho}\\|h_k\\|^2$, converting constrained MPC into a box-constrained nonlinear least-squares problem $\\min_{p_k\\le z_k\\le q_k} \\tfrac{1}{2}\\|r_k(z_k)\\|_2^2$. The second mechanism is the pair of abstract operators $J_{ix}$ and $J_{tiX}$, which compute the $i$-th column of the Jacobian scaled by a scalar and the inner product of the $i$-th row of $J^\\top$ with a vector, using only the model coefficients and the integers defining the sparsity pattern. These operators make the solver parameter-free: the compiled code does not change when the model or tuning parameters change. The third mechanism is the sparse recursive thin QR factorization of Section 6, whose predicted nonzero structure is characterized in Corollaries 2 and 3, reducing the cost of active-set bound flips to operations on nonzeros only.","core_discovery":"The central claim is that the problem-construction phase of MPC can be eliminated systematically. For the class of models expressible as a differentiable input-output equation $M(Y_k,U_k,S_k)=0$, the Jacobian of the residual in the least-squares formulation (6) has a sparsity pattern that depends only on a few integers: model orders $n_a,n_b$, output and input dimensions $n_y,n_u$, and horizons $N_p,N_u$. By ordering decision variables appropriately, every column of the Jacobian can be generated on the fly by two operators $J_{ix}$ and $J_{tiX}$ that read off the appropriate model coefficients, so neither the matrix $J$ nor the MPC problem itself is ever stored. The authors prove global convergence of the resulting bounded-variable nonlinear least-squares algorithm (BVNLLS) via an Armijo-type backtracking line search, and they derive rigorous upper bounds on the nonzero structure of the thin QR factors, which lets the inner active-set solver update its factorization using Givens rotations only on nonzero entries. With this machinery, one implementation solves linear MPC in a single least-squares pass and nonlinear or adaptive MPC by iterating the same solver, with numerical experiments on a CSTR benchmark showing roughly two orders of magnitude speedup over generic nonlinear programming solvers on small problems.","pith_inferences":["The two-operator abstraction is a template that likely transfers to any least-squares problem whose Jacobian columns are shifts of a short parameter-dependent kernel, such as structured low-rank or convolutional inverse problems, not just MPC.","Because the paper only proves the single-penalty near-optimality for linear models, a natural testable extension is an adaptive scheme that raises $\\rho$ only when the observed equality residual exceeds a threshold, or combines the penalty with a few bound-constrained Lagrangian iterations to certify constraint satisfaction at convergence.","The reordering of decision variables suggested in Section 6.4 to move the columns of $u_{k+N_u-1}$ last is a fill-reducing heuristic; one could benchmark it against a symbolic minimum-degree ordering of the same sparsity pattern to see whether further gains are available.","If the prediction model is replaced by a neural network fitted to data, the same operators could be used with autodiff-generated Jacobians, making the approach a candidate for data-driven adaptive MPC with no per-iteration problem reformulation."],"forward_implications":["A single embedded solver, with no code regeneration, can handle linear MPC, nonlinear MPC, and adaptive MPC as long as the prediction model fits the differentiable input-output form (1) and the cost is a sum of squares.","Because the problem matrices are never formed, memory use stays nearly independent of the prediction horizon, growing only with the stored model coefficients and the QR factors of the current active set.","Linear MPC is solved in one least-squares solve with no backtracking, since the Gauss-Newton subproblem is exact; nonlinear MPC iterates the same solver with warm starts from the previous sampling instant.","The sparsity-analysis results (Theorem 2 and Corollaries 1-3) guarantee that the predicted nonzero structure of the QR factors is determined only by the model and horizon parameters, so a library-free C implementation can skip zero flops without sparse-matrix storage.","For feasible problems with general nonlinear inequality constraints, the same NLLS-box solver can be used inside a bound-constrained Lagrangian method, at the price of solving multiple instances rather than one."],"supporting_citations":[{"why":"Introduced the quadratic-penalty formulation of MPC as a bounded-variable least-squares problem and proved the single-penalty near-optimality result for linear models that the present paper extends.","marker":"[29]"},{"why":"Supplies the inner bounded-variable least-squares solver based on recursive thin QR updates, which the present paper specializes to the sparse, parameter-dependent structure.","marker":"[30]"},{"why":"Provided the dense BVNLLS variant and the CSTR nonlinear MPC benchmark whose simulation scenarios and timing comparisons are reused in Section 7.","marker":"[31]"},{"why":"Provides the thin QR factorization and Gram-Schmidt orthonormalization theory used to derive the sparsity bounds on the Q and R factors in Section 6.","marker":"[16]"},{"why":"Supplies the active-set, quadratic-penalty, bound-constrained Lagrangian, and line-search convergence framework that underpins Algorithm 1 and Theorem 1.","marker":"[23]"},{"why":"Provides the large-scale interior-point solver used as a benchmark in the numerical comparisons of Figures 4 and 5.","marker":"[35]"}],"fun_headline_variants":["Skip MPC problem assembly, hit the solver directly","One least-squares solver handles linear, nonlinear, and adaptive MPC","Abstract functions eliminate MPC construction phase","Embedded MPC without storing the problem matrix","From parameter sets to solutions in one least-squares pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that a single, hand-tuned penalty parameter $\\rho$ can be chosen large enough to keep the model-equality violations negligible while still leaving the least-squares Jacobian numerically full-rank; the paper proves this is safe only for linear models, and states that the extension to the general nonlinear case 'is not immediate and thereby poses a risk given that the bound is not deterministic'.","fun_headline_variants_meta":{"raw":{"variants":["Skip MPC problem assembly, hit the solver directly","One least-squares solver handles linear, nonlinear, and adaptive MPC","Abstract functions eliminate MPC construction phase","Embedded MPC without storing the problem matrix","From parameter sets to solutions in one least-squares pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2932,"prompt_tokens":981,"completion_tokens":1951,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1878}},"tokens_in":597,"tokens_out":1951,"duration_ms":14508,"temperature":1.0,"reasoning_tokens":1878,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:21:59.953358+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a nonlinear model with a stiff or poorly scaled Jacobian and increase $\\rho$ until constraint residuals meet a target; if at some $\\rho$ the least-squares subproblem becomes numerically rank-deficient (or the solver converges to a point with equality residual above the specified tolerance), the single-penalty step has failed to solve the intended constrained MPC problem. A concrete test: for the CSTR benchmark with $N_p=N_u=160$, measure the worst-case $\\|h(z,\\phi)\\|_2$ while increasing $\\sqrt{\\rho}$ from $10^2$ to $10^6$ and record the smallest $\\rho$ at which the residual stops decreasing or the solver's condition number grows beyond, say, $10^{12}$.","supporting_citations":[{"cited_title":"Saraf and A","cited_arxiv_id":null,"evidence_quote":"Introduced the quadratic-penalty formulation of MPC as a bounded-variable least-squares problem and proved the single-penalty near-optimality result for linear models that the present paper extends."},{"cited_title":"Saraf and A","cited_arxiv_id":null,"evidence_quote":"Supplies the inner bounded-variable least-squares solver based on recursive thin QR updates, which the present paper specializes to the sparse, parameter-dependent structure."},{"cited_title":"Saraf, M","cited_arxiv_id":null,"evidence_quote":"Provided the dense BVNLLS variant and the CSTR nonlinear MPC benchmark whose simulation scenarios and timing comparisons are reused in Section 7."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the thin QR factorization and Gram-Schmidt orthonormalization theory used to derive the sparsity bounds on the Q and R factors in Section 6."},{"cited_title":"W¨ achter and L.T","cited_arxiv_id":null,"evidence_quote":"Provides the large-scale interior-point solver used as a benchmark in the numerical comparisons of Figures 4 and 5."}],"review_version":1}