{"id":"e75fd6dc-6032-43c7-8474-c61db67e617b","arxiv_id":"2507.20686","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A subspace decomposition gives explicit conjugate-form formulas for the solution set of regularized least squares, yielding unified existence, compactness, and uniqueness conditions.","lead":"This paper derives explicit formulas for the solution set of regularized least-squares problems using a subspace decomposition. The formulas give unified conditions for when solutions exist, when they are bounded, and when they are unique.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 2.20(i) wrongly identifies the descent cone with its closed tangent cone; an explicit f, A, b makes Proposition 2.21(i) fail although the solution is unique.","rationale":"The reader's flagged Lemma 2.15 checks out under re-derivation: the strict-inequality direction does not require x0 to be arbitrary, because the negation of u ∈ ∂f(y) supplies a witness z, and the subgradient inequality at x* forces the claimed lower bound. The actual load-bearing flaw is in Section 2.4's cone unification. Lemma 2.20(i) equates a cone with its closure, and the proof's reverse inclusion silently uses the non-closed radial cone instead of the tangent cone defined in Definition 2.19(iii). The explicit example above is within the paper's assumptions, has a unique solution, and shows that Proposition 2.21(i) is not equivalent to the other criteria. This is a real correctness issue in an advertised contribution, though it does not invalidate the main solution-set characterization of Theorem 2.2 or the existence/compactness analysis. The verdict should be CONDITIONAL: the paper should be accepted only after Lemma 2.20 and Proposition 2.21 are corrected or qualified, e.g., by replacing T_{slev} with the unclosed radial cone (or descent cone) and adjusting the citation to [28, Theorem 5.1] accordingly.","tokens_in":30239,"tokens_out":28459,"duration_ms":303725,"concrete_test":"Check the claimed equivalence in Proposition 2.21 on f(x1,x2) = e^{x1} - x1 - 1 - x2, A = [0 1], b = -1. Compute the three cones at the unique solution x* = (0,0): Df(x*), T_{slev_{x*} f}(x*), and R_{∂f*(A^T r)}(x*), and intersect each with ker A. If T_{slev_{x*} f}(x*) ∩ ker A ≠ {0} while Df(x*) ∩ ker A = {0}, then Lemma 2.20(i) and Proposition 2.21(i) are refuted and the tangent-cone criterion must be replaced by the unclosed radial/descent cone.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The weakest link is not Lemma 2.15, which appears sound: for d outside C∞, the strict inequality follows by combining the failure of u ∈ ∂f(y) with the subgradient inequality at x*. The genuine problem is Lemma 2.20(i), which claims Df(x*) = T_{slev_{x*} f}(x*). By Definition 2.19(iii), T is the closure of the radial cone, while Df is the unclosed cone R_+(slev_{x*} f - x*). The proof of the reverse inclusion uses the false step that every point of the closure lies in the cone itself. Counterexample inside the paper's assumptions: f(x1,x2) = e^{x1} - x1 - 1 - x2 (proper lsc convex), A = [0 1], b = -1. The objective is e^{x1} - x1 - 1 - x2 + 1/2 (x2+1)^2 = e^{x1} - x1 - 1/2 + x2^2/2, with unique minimizer x* = (0,0). Its sublevel set is S = {x2 ≥ e^{x1} - x1 - 1}. One computes Df(x*) = {d2 > 0} ∪ {0}, while T_S(x*) = {d2 ≥ 0}. Since ker A = span{(1,0)}, we get Df(x*) ∩ ker A = {0}, but T_S(x*) ∩ ker A = R × {0}. Thus the tangent-cone condition of Proposition 2.21(i) falsely predicts non-uniqueness, so the claimed equivalence (i) ⇔ (ii) ⇔ (iii) is false as stated. The central Theorem 2.2 solution-set formula is not affected.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the regularized least-squares problem min f(x)+1/2||Ax-b||^2 and its equality-constrained analogue, using the orthogonal decomposition of R^n into ran A^T and ker A. The central result is an explicit solution-set formula X = x*_r + [(∂f*(A^T r)-x*_r) ∩ ker A] (Theorem 2.2), from which the authors derive existence, compactness, and uniqueness criteria in terms of ∂f*, and then connect these criteria to existing conditions based on recession cones, sublevel sets, restricted coercivity, and the radial/tangent/descent cones used in prior work. The constrained problem (1.2) is treated analogously and linked to (1.1) through infimal postcomposition. The paper also contains extensive worked examples, including lasso illustrations.","tokens_in":30586,"tokens_out":20809,"duration_ms":194567,"significance":"The core solution-set formula in Theorem 2.2 is clean, self-contained, and appears correct, and the separate treatment of existence versus compactness is a genuine clarification of earlier bundled criteria. The derivations are parameter-free and analytical, with no fitted constants or circular reliance on prior work. The infimal-postcomposition connection in Section 3.4 and the many worked examples are valuable. However, the claimed unification of uniqueness criteria contains false statements: Lemma 2.20(i) is wrong, Lemma 2.22 fails as a set equality, and Lemma 3.8(ii) is false. These errors are in the uniqueness/unification sections and therefore affect load-bearing claims, although the central solution-set derivation and most existence/compactness results remain sound.","major_comments":[{"comment":"The asserted equality Df(x⋆)=T_{slev_{x⋆}f}(x⋆) is false under the paper's own definitions: by Definition 2.19(iii), T is the closure of the radial cone, while Df is the unclosed cone R_+(slev_{x⋆}f−x⋆). The reverse inclusion in the proof assumes that every point of the closure is already in the radial cone, which is not true. A concrete counterexample is f(x1,x2)=e^{x1}−x1−1−x2, A=[0 1], b=−1. The unique minimizer is x⋆=(0,0); one computes Df(x⋆)={(d1,d2): d2>0}∪{(0,0)}, while T_{slev_{x⋆}f}(x⋆)={(d1,d2): d2≥0}. With kerA=span{(1,0)}, Df(x⋆)∩kerA={0} but T_{slev_{x⋆}f}(x⋆)∩kerA=R×{0}. Hence Proposition 2.21(i) predicts non-uniqueness even though the solution is unique, so the equivalence (i)⇔(ii)⇔(iii) in Proposition 2.21 fails as stated. The correct object for the equivalence is the radial cone, which equals Df(x⋆).","section":"§2.4.2, Lemma 2.20(i) and Proposition 2.21(i)"},{"comment":"The claimed set equality (∂f*(A^T r)−x⋆)∩kerA = (slev_{x⋆}f−x⋆)∩kerA = Df(x⋆)∩kerA is false. The paper's own Table 2.2, Example-4, already exhibits the failure: for f(x)=max{|x1|−1,0}, A=[0 1], b=1, x⋆=(0,1), the table gives Df(x⋆)∩kerA=R×{0} while (∂f*(0)−x⋆)∩kerA=[−1,1]×{0}. The inclusion Df∩kerA⊆(∂f*−x⋆)∩kerA fails because d∈Df only guarantees f(x⋆+ξd)≤f(x⋆) for some ξ>0, not for d itself. The correct identity is Df(x⋆)∩kerA=R_+((∂f*(A^T r)−x⋆)∩kerA). Since the positive hull of a set containing 0 is {0} exactly when the set is {0}, the uniqueness criteria in Proposition 2.21 survive as zero/nonzero tests, but the stated equalities and their proofs need to be revised.","section":"§2.4.2–2.4.3, Lemma 2.20(ii) and Lemma 2.22"},{"comment":"The equality R_{∂f*(y)}(x⋆)=∪_{v∈R^m: A^T v∈∂f(x⋆)}(∂f*(A^T v)−x⋆) is false. Take f=ι_{[-1,1]} on R, A=1, b=0, x⋆=0, and y=0∈∂f(x⋆)∩ranA^T. Then ∂f*=∂|·|, so ∂f*(0)=[−1,1] and R_{∂f*(0)}(0)=R_+([−1,1])=R, while the union over all v with A^T v∈∂f(x⋆)=R is ∂f*(R)=[−1,1]. Thus Lemma 3.8(ii) cannot serve as the bridge between the radial-cone condition and the union condition for (1.2). The proof of Theorem 3.9 should rely on Proposition 3.6, which remains a valid solution-set criterion, rather than on Lemma 3.8(ii).","section":"§3.3, Lemma 3.8(ii)"},{"comment":"The step from σ_{PkerA(dom f*)}(d)>0 for all d∈kerA\\{0} to 0∈ri PkerA(dom f*) is stated without proof or citation. This is standard support-function/relative-interior machinery, but since it supports the restricted-coercivity discussion, a one-sentence justification or a precise reference should be added.","section":"§2.3.4, Theorem 2.14(v)"}],"minor_comments":[{"comment":"In the proof of Lemma 2.1(iv), the sentence 'which yields (iii)' should read 'which yields (iv)'.","section":"Lemma 2.1(iv)"},{"comment":"The tables contain the placeholder '/reve' in most evaluation cells; these should be replaced with the intended check/cross symbols and the numerical values should be completed.","section":"Tables 2.1–3.2"},{"comment":"In the displayed formula for ∂f*, the last case condition uses '(x1,x2)' where '(u1,u2)' is meant.","section":"Table 2.2, Example-4 row for ∂f*"},{"comment":"There is a typo 'suth that' which should be 'such that'.","section":"Theorem 3.9 proof"},{"comment":"The condition 'inf_{t>0,d∈kerA} f(x+td)>−∞' would be clearer with a comma: 'inf_{t>0, d∈kerA} f(x+td)>−∞'.","section":"Proposition 2.12(ix)"}],"recommendation":"major_revision","confidential_remarks":"The core Theorem 2.2 and the existence/compactness analysis appear sound, and the paper is within scope for math.OC. The uniqueness-unification claims need substantive correction: the false Lemma 2.20(i), the false set equalities in Lemmas 2.20(ii)/2.22, and the false Lemma 3.8(ii) are all in sections where the authors claim to unify prior criteria. The authors' own Table 2.2 contains the data needed to see one of these failures, so the revision should also include a systematic check of all claimed equivalences against the examples."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my honest read. The core of the paper is the explicit solution-set characterization in Theorem 2.2: X = x*_r + (∂f*(A^T r) − x*_r) ∩ ker A, with x*_r = A†(I_m + A▷∂f)^{-1}(b). That is genuinely new and, as far as I can tell, correct. The separate treatment of existence (b ∈ ran(I_m + A▷∂f)) and compactness ((∂f*(A^T r))^∞ ∩ ker A = {0}) is useful and clean. The bridge in Theorem 2.16 to the older recession-cone criteria also looks sound, and the lasso corollary is a nice concrete application. The paper earns credit for that.\n\nBut the stress-test note is right, and the reader's flagged worry about Lemma 2.15 is not the real problem. Lemma 2.15 is fine. The actual flaw is Lemma 2.20(i), which claims Df(x*) = T_{slev_{x*} f}(x*). By the paper's own Definition 2.19(iii), T is the closure of the radial cone, while Df is the unclosed cone. The reverse inclusion in the proof incorrectly treats every point of the closure as a point of the cone. This matters because Proposition 2.21(i) then presents the tangent-cone condition as equivalent to uniqueness.\n\nThe counterexample in the stress test is valid: f(x1,x2) = e^{x1} − x1 − 1 − x2, A = [0 1], b = −1. The unique minimizer is x* = (0,0). The descent cone is {d2 > 0} ∪ {0}, while the tangent cone is its closure {d2 ≥ 0}. After intersecting with ker A = span{(1,0)}, the descent cone gives {0} but the tangent cone gives R × {0}. So the tangent-cone condition falsely rules out uniqueness. The central Theorem 2.2 formula is not affected, but the claimed equivalence (i)⇔(ii)⇔(iii) in Proposition 2.21 is false as stated. The descent-cone and radial-cone conditions remain correct; it is specifically the use of the closed tangent cone that breaks necessity.\n\nThere are also minor annoyances: several cross-reference typos, compressed steps in a few lemmas, and the table rendering in the arXiv version is hard to read. None of that changes the main math.\n\nBottom line: this deserves a serious referee. The solution-set formula and the existence/compactness analysis are worth publishing after a major revision that fixes the tangent-cone claim. I would bring it to a reading group, and I would cite the theorem for the solution-set formula.","headline":"The main solution-set formula is right, but the claimed unification of uniqueness criteria via the tangent cone is false as stated.","tokens_in":854,"tokens_out":1747,"would_cite":true,"duration_ms":60676,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["47H05","49M29","49M27","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A subspace split turns regularized least-squares solution properties into simple kernel-intersection checks.","keywords":["regularized least-squares","solution existence","solution compactness","solution uniqueness","restricted coercivity","recession function","conjugate subdifferential","lasso"],"falsifier":"A direct computation on the paper's paired examples would settle the central identity: take $f(x_1,x_2)=\\max\\{e^{x_2}-x_1,0\\}$ with $A=[1\\;0]$ and $b=0$ (predicted empty solution set) versus $A=[0\\;1]$ and $b=0$ (predicted non-empty, unbounded solution set). Evaluating the right-hand side of $X=x^\\star_r+[(\\partial f^*(A^T r)-x^\\star_r)\\cap\\ker A]$ and comparing it with the set of points satisfying Fermat's condition would refute the paper's claim if any minimizer is missed or any non-minimizer is included.","tokens_in":30039,"feed_emoji":"📐","tokens_out":9768,"duration_ms":84762,"temperature":0.7,"pith_summary":"The paper tries to establish that the solution set of $\\min_x f(x)+\\tfrac12\\|Ax-b\\|^2$ splits into a fixed range-space component and a kernel component selected by the conjugate subdifferential. If the central formula is right, the three basic solution properties—non-emptiness, boundedness, and uniqueness—become direct set-intersection checks instead of separate coercivity analyses. The paper also claims to unify earlier recession-cone, sublevel-set, descent-cone, and radial-cone criteria through a bridge lemma that matches the recession function of $f$ with the recession cone of $\\partial f^*(A^T r)$. A parallel treatment of the equality-constrained problem links the two formulations through exact infimal postcomposition. This matters because the resulting geometric conditions are concrete enough to apply directly to the lasso and other regularized estimators.","feed_headline":"Explicit formula pins down all solutions of regularized least-squares","feed_subtitle":"A subspace split turns existence, boundedness, and uniqueness of solutions into simple kernel-intersection checks.","key_machinery":"The machinery is the orthogonal decomposition $\\mathbb R^n=\\operatorname{ran}A^T\\oplus\\ker A$, applied through Fermat's optimality condition. Every candidate $x$ is written as $x_r+x_k$; the range component $x_r$ is forced by the least-squares term to be the resolvent point $A^\\dagger(I_m+A\\triangleright\\partial f)^{-1}(b)$, and the kernel component $x_k$ must lie in $(\\partial f^*(A^T r)-x_r)\\cap\\ker A$. The load-bearing bridge is Lemma 2.15, which states that the recession function obeys $f^\\infty(d)=\\langle A^T r,d\\rangle$ exactly when $d$ lies in $(\\partial f^*(A^T r))^\\infty$, with strict inequality otherwise; this is what converts all recession-cone and sublevel-set criteria into the conjugate-subdifferential language in which the solution set is written.","core_discovery":"For a proper lower semicontinuous convex $f$ and a linear map $A$, the solution set $X$ of (1.1) admits the explicit expression $X=x^\\star_r+[(\\partial f^*(A^T r)-x^\\star_r)\\cap\\ker A]$, where $x^\\star_r=A^\\dagger(I_m+A\\triangleright\\partial f)^{-1}(b)$ and $r=b-Ax^\\star_r$. From this identity the paper reads off: $X\\neq\\emptyset$ exactly when $b\\in\\operatorname{ran}(I_m+A\\triangleright\\partial f)$; $X$ is compact exactly when the recession cone $(\\partial f^*(A^T r))^\\infty$ meets $\\ker A$ only at the origin; and $X$ is a singleton exactly when $(\\partial f^*(A^T r)-x^\\star)\\cap\\ker A=\\{0\\}$. The paper further proves these conditions are equivalent to the classical criteria involving the recession cone $R_f$, the recession function kernel $\\ker f^\\infty$, sublevel sets, descent cones, and radial cones, thereby unifying previously separate results. The equality-constrained version is given the same subspace treatment, and its relation to (1.1) is expressed through exactness of the infimal postcomposition $A\\triangleright f$.","pith_inferences":["An implicit consequence is that algorithm designers can certify a unique residual and a unique range component even when the full solution set is unbounded, a distinction that bundled existence-compactness conditions do not expose.","The normal-cone picture for the lasso suggests a direct probabilistic route to lasso uniqueness: estimate the chance that the normal cone $N_{\\ell_\\infty}(A^T r)$ intersects $\\ker A$ nontrivially, without passing through descent cones.","The same subspace split should transfer to analysis formulations $f(Dx)+\\tfrac12\\|Ax-b\\|^2$ and to generalized lasso, with the kernel of $A$ replaced by the relevant subspace determined by both $A$ and $D$, as the paper's concluding remarks indicate.","The paired examples with the same $f$ but different orientations of $A$ show that emptiness versus unboundedness is a property of the pair $(f,A)$, not of $f$ alone; this invites a design-oriented question about which orientations make $A\\triangleright\\partial f$ maximal monotone."],"forward_implications":["All solutions of (1.1) share the same vector $Ax$, the same residual $r$, and the same range component $x^\\star_r$; only the kernel component can vary.","For any fixed $b$, existence holds exactly when $b\\in\\operatorname{ran}(I_m+A\\triangleright\\partial f)$, and existence for every $b$ is equivalent to maximal monotonicity of $A\\circ\\partial f^*\\circ A^T$.","The solution set is compact exactly when $(\\partial f^*(A^T r))^\\infty\\cap\\ker A=\\{0\\}$, and it is a singleton exactly when $(\\partial f^*(A^T r)-x^\\star)\\cap\\ker A=\\{0\\}$.","These conditions are equivalent to the classical criteria $R_f\\cap\\ker A=\\{0\\}$, $\\ker f^\\infty\\cap\\ker A=\\{0\\}$, and the descent-cone or radial-cone conditions used in earlier uniqueness theorems.","For the lasso ($f=\\|\\cdot\\|_1$), the solution set takes the explicit normal-cone form with respect to the $\\ell_\\infty$ unit ball, and a lasso solution exists for every $b$."],"supporting_citations":[{"why":"Supplies the monotone-operator calculus (resolvents, parallel composition, Minty's theorem, subdifferential rules) used throughout the proofs.","marker":"[4]"},{"why":"Defines the restricted-injectivity and sublevel-set criteria for solution properties that this paper generalizes and unifies.","marker":"[28]"},{"why":"Gives the radial-cone uniqueness theorem that the paper specializes and proves equivalent to its conjugate-subdifferential condition.","marker":"[14]"},{"why":"Establishes descent-cone uniqueness and the constrained-to-regularized connection that the paper extends via infimal postcomposition.","marker":"[15]"},{"why":"Provides the recession-function and coercivity toolkit anchoring the compactness equivalences.","marker":"[2]"},{"why":"Supplies the variational-analysis treatment of the lasso whose restricted-coercivity claims are refined by the new conditions.","marker":"[6]"},{"why":"Offers the gauge-function dual viewpoint on uniqueness that becomes a special case of the new characterization.","marker":"[20]"},{"why":"Provides the necessary-and-sufficient $\\ell_1$ uniqueness condition and the lasso example used to show the projection condition is sufficient but not necessary.","marker":"[34]"}],"fun_headline_variants":["All solutions of regularized least-squares in one identity","Subspace decomposition gives exact solution set for LS","Explicit formula characterizes every solution of regularized LS","New subspace split yields full solution set for regularized LS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Lemma 2.15, which asserts that the recession function satisfies $f^\\infty(d)=\\langle A^T r,d\\rangle$ exactly when $d$ lies in the recession cone of $\\partial f^*(A^T r)$, with strict inequality otherwise; because the proof fixes the minimizer $x^\\star$ as the reference point, the claimed unification with all recession-cone criteria collapses if this lemma fails.","fun_headline_variants_meta":{"raw":{"variants":["All solutions of regularized least-squares in one identity","Subspace decomposition gives exact solution set for LS","Explicit formula characterizes every solution of regularized LS","New subspace split yields full solution set for regularized LS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2819,"prompt_tokens":1009,"completion_tokens":1810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":1746}},"tokens_in":625,"tokens_out":1810,"duration_ms":13297,"temperature":1.0,"reasoning_tokens":1746,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:40:50.785504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct computation on the paper's paired examples would settle the central identity: take $f(x_1,x_2)=\\max\\{e^{x_2}-x_1,0\\}$ with $A=[1\\;0]$ and $b=0$ (predicted empty solution set) versus $A=[0\\;1]$ and $b=0$ (predicted non-empty, unbounded solution set). Evaluating the right-hand side of $X=x^\\star_r+[(\\partial f^*(A^T r)-x^\\star_r)\\cap\\ker A]$ and comparing it with the set of points satisfying Fermat's condition would refute the paper's claim if any minimizer is missed or any non-minimizer is included.","supporting_citations":[{"cited_title":"Bauschke and Patrick L","cited_arxiv_id":null,"evidence_quote":"Supplies the monotone-operator calculus (resolvents, parallel composition, Minty's theorem, subdifferential rules) used throughout the proofs."},{"cited_title":"Thesis, Universit´ e Paris–Dauphine, (2014)","cited_arxiv_id":null,"evidence_quote":"Defines the restricted-injectivity and sublevel-set criteria for solution properties that this paper generalizes and unifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes descent-cone uniqueness and the constrained-to-regularized connection that the paper extends via infimal postcomposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the recession-function and coercivity toolkit anchoring the compactness equivalences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational-analysis treatment of the lasso whose restricted-coercivity claims are refined by the new conditions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Offers the gauge-function dual viewpoint on uniqueness that becomes a special case of the new characterization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the necessary-and-sufficient $\\ell_1$ uniqueness condition and the lasso example used to show the projection condition is sufficient but not necessary."}],"review_version":2}