{"id":"ea075580-0a20-4c25-b7a6-d5dbfb4acf3c","arxiv_id":"2411.11242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors design a functional mirror descent for reproducing kernel Banach spaces and prove conditional linear and O(1/√t) convergence, with a finite-center p-norm RKBS instantiation.","lead":"This paper proposes a mirror-descent algorithm for learning functions in reproducing kernel Banach spaces, a generalization of standard kernel Hilbert spaces. It proves a conditional linear convergence rate and slower O(1/√t) guarantees, then builds a concrete p-norm Banach space with an explicit kernel to run the method.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 20 has an invalid contraction step around Eq. (30), and the stated learning-rate min leaves the claimed exponent unchecked; the theorem is likely repairable but is not proven as written.","rationale":"The reader's Hilbertizability concern is real: Lemma 19 makes Assumption 2 essentially a Hilbertizability condition, so the linear-rate theorem does not reach the genuinely non-Hilbertian p-norm RKBS that motivate the paper. This is a scope limitation, but the paper states the condition explicitly, so I do not treat it as a correctness failure. My independent read found a more immediate correctness issue in the proof of the headline theorem: Eq. (30) has a subgradient-index mismatch that prevents the displayed contraction from following from Eq. (29), and the final rate silently selects only the η = ν/γ branch of the stated min. The theorem is very likely repairable by reindexing and by using strong convexity at f_{k-1}, and the constrained O(1/√t) result appears standard. Since the central proof needs revision, the conditional verdict remains appropriate; I would not move to accept or reject without the authors addressing this step.","tokens_in":31350,"tokens_out":33461,"duration_ms":315021,"concrete_test":"Re-derive the proof of Theorem 20 by replacing Eq. (30) with the strong-convexity inequality at f_{k-1}: minimize L(f_{k-1}) + ⟨y−f_{k-1}, ∂L(f_{k-1})⟩ + (µ/2)||y−f_{k-1}||²_B over y to obtain ||∂L(f_{k-1})||²_{B*} ≥ 2µD_{k-1}. Combine this with Eq. (29) to get D_k ≤ [1 − µ(2ν−γη)κ²η]D_{k-1}. Then plug in both branches of η = min{ν/γ, 1/(2µνκ²)} and verify whether D_k ≤ D_0e^{−k·µν²κ²/γ} holds for all admissible constants. If the second branch violates the claimed exponent, the theorem statement must be revised; if it holds, the proof still needs the index correction to Eq. (30).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is in the proof of Theorem 20, not in the assumptions themselves. After Eq. (29), the proof has a descent estimate L(f_k) ≤ L(f_{k-1}) − A||∂L(f_{k-1})||²_{B*}, and it must convert this into a contraction in D_k = L(f_k) − L(f*). The correct step is to use µ-strong convexity of L at f_{k-1}, which gives ||∂L(f_{k-1})||²_{B*} ≥ 2µD_{k-1}. Instead, Eq. (30) derives a lower bound on L(f*) from strong convexity at f_k and then writes the subgradient as ∂f_{k-1}L. With ∂f_{k-1}L, the inequality has the wrong direction of substitution and cannot be combined with Eq. (29) to yield the displayed contraction; with ∂f_kL, it belongs to iteration k+1, so the indices must be shifted. Additionally, the final rate e^{−k·µν²κ²/γ} is obtained by setting (2ν−γη)η = ν²/γ, which holds only when η = ν/γ, but the theorem states η = min{ν/γ, 1/(2µνκ²)} and the other branch is never analyzed. These issues are likely fixable by reindexing and a branch analysis, but as written the central linear-rate theorem is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies mirror descent on reproducing kernel Banach spaces (RKBS). It proposes a functional mirror descent algorithm whose dual updates are expressed through the reproducing kernel, and it analyzes convergence. The main theoretical claims are: (i) a linear convergence rate for unconstrained mirror descent when the RKBS is reflexive and Hilbertizable and the loss and mirror map are strongly convex and smooth (Theorem 20), with an extension under the Polyak–Łojasiewicz condition (Theorem 22); (ii) a negative result (Lemma 19) stating that a functional that is both strongly convex and smooth exists only on spaces isomorphic to Hilbert spaces; and (iii) an O(1/√t) rate for projected mirror descent in a constrained setting (Theorem 25). The paper also constructs a finite-center p-norm RKBS with an explicit reproducing kernel and ℓp mirror map, and reports numerical experiments.","tokens_in":31649,"tokens_out":9809,"duration_ms":86551,"significance":"The contribution is potentially useful: it brings the mirror-descent framework to RKBS optimization, gives a concrete finite-dimensional p-norm RKBS with an explicit kernel and dual map, and provides a standard constrained O(1/√t) analysis. The explicit updates in Algorithm 1 and the connection to representer theorems are strengths. However, the headline linear-rate theorem is narrower than the abstract suggests: by the paper's own Lemma 19, the smooth-and-strongly-convex assumptions force the space to be isomorphic to a Hilbert space, so the linear-rate result does not cover genuinely non-Hilbertian RKBS. Moreover, the proof of Theorem 20 has a repairable but load-bearing gap around Eq. (30). With those points fixed, the paper would make a solid contribution to the optimization side of RKBS.","major_comments":[{"comment":"The contraction step in the proof of Theorem 20 is misindexed. Strong convexity at f_{k-1} gives L(f*) ≥ L(f_{k-1}) − (1/(2µ))||∂f_{k-1}L||²_{B*}, while strong convexity at f_k gives L(f*) ≥ L(f_k) − (1/(2µ))||∂f_k L||²_{B*}. The displayed equation instead writes an infimum around f_k and then uses ∂f_{k-1}L; with that subgradient the inequality does not follow, and with ∂f_kL it belongs to iteration k+1 and cannot be combined with Eq. (29). The displayed contraction factor is therefore not established as written, and Theorem 22 inherits this gap.","section":"§3.2, Eq. (30)"},{"comment":"The constant κ is undefined. The proof states 'there exists a scalar κ > 0 such that κ := 1/L3' but L3 never appears in Proposition 21 or elsewhere. Proposition 21 has constants L2 and L5; the proof needs to define κ explicitly (for p = 2, presumably κ = 1/L5) and show that κ depends only on ρ, since the rate e^{−k·µν²κ²/γ} is otherwise not quantitative.","section":"§3.2, after Proposition 21"},{"comment":"The final rate is obtained by setting (2ν−γη)η = ν²/γ, which requires η = ν/γ, but the theorem states η = min{ν/γ, 1/(2µνκ²)}. The second branch is never analyzed, and one must also verify that η < 2ν/γ so that the coefficient in Eq. (28) is positive. A complete proof needs a branch analysis and, if necessary, a different contraction factor for the second branch.","section":"§3.2, Theorem 20"},{"comment":"The scope of the linear-rate result is narrower than the abstract suggests. Lemma 19 implies that any space satisfying Assumption 2 is isomorphic to a Hilbert space, so Assumption 2 already implies Assumption 1; moreover, since Theorem 20 independently assumes that L and Φ are strongly convex and smooth, Assumption 2 is redundant as stated. The authors should state explicitly that the linear-rate guarantee applies only to Hilbertizable RKBS and does not cover genuinely non-Hilbertian RKBS, and they should restructure the assumptions accordingly.","section":"§3.1–3.2, Lemma 19 and Assumption 2"}],"minor_comments":[{"comment":"The convexity definition has a typo: 'F(λf + (1 + λ)f′)' should read 'F(λf + (1−λ)f′)'.","section":"§2.1, Definition 6"},{"comment":"The Jensen inequality direction is reversed in the line 'using convexity 1/t ∑ L(f_t) − L(f*) ≤ L(1/t ∑ f_t) − L(f*)'; it should be '≥'. The desired bound follows from L(1/t∑f_t) − L(f*) ≤ (1/t)∑(L(f_t) − L(f*)) together with the averaged regret bound.","section":"§3.3, Theorem 25 proof"},{"comment":"In Eqs. (61)–(62) the argument of F in the second-order expansion and in the smoothness/strong-convexity inequalities is written inconsistently as f′; these should be 'F(f + λf′)' and '⟨λf′, g*⟩' with the subgradient taken at f.","section":"§A.2, Lemma 19 proof"},{"comment":"The text 'for q = 1 + q/p' is incorrect; the intended relation is 1/p + 1/q = 1, equivalently q = p/(p−1).","section":"§4"},{"comment":"The initial values α(0) and β(0) are listed in the Data line but never specified; the algorithm should state an initialization, for example β(0) = 0 and α(0) = 0.","section":"§4, Algorithm 1"},{"comment":"The text says the training error is plotted over 500,000 iterations while the caption mentions up to 100,000 steps; the numbers should be made consistent.","section":"§4, Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a machine learning journal. The main theorem's proof has a real but repairable gap, and the advertised linear-rate result is conditional and narrower than the abstract suggests. I see no novelty or attribution concerns. I would not reject: the constrained result and the p-norm RKBS construction are largely sound, and the linear-rate theorem is likely repairable by reindexing Eq. (30) and adding a learning-rate branch analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, the paper has a real and useful contribution: a functional mirror descent algorithm for RKBS that writes dual updates as kernel evaluations, and a concrete finite-center p-norm RKBS with an explicit dual map and kernel. Second, the headline linear-rate result (Theorem 20) is not proven as written; the proof has an index mismatch around Eq. (30), an undefined constant in the definition of κ, and the displayed rate only follows for the learning rate η = ν/γ, not for the min of the two branches stated in the theorem. These are fixable, but they are genuine gaps.\n\nWhat works: the constrained O(1/√t) result is a standard mirror descent argument and appears correct. The p-norm RKBS construction is clean and genuinely useful. Lemma 19—that a functional that is both smooth and strongly convex forces the space to be isomorphic to a Hilbert space—is a good sanity check, and the authors deserve credit for stating the 'Hilbertizable' condition rather than burying it. That said, the assumption means the linear rate only applies in a near-Hilbert setting, so the non-Hilbertian selling point is weaker than the abstract suggests. The experiments are minimal and the comparison against a NumPy solve with a Gaussian kernel is not controlled, since the kernel and the space differ.\n\nWhere the soft spots are: Theorem 20's proof should be reindexed so the strong-convexity lower bound applies to f_{k-1} to combine with the descent estimate; as written Eq. (30) mixes ∂f_k and ∂f_{k-1}, and the contraction does not follow. The learning-rate issue is easy to miss: the stated min makes the exponent unverified in the second branch. The definition of κ via an undefined L3 is a presentation slip but should be fixed.\n\nOverall, the paper deserves refereeing. The core algorithm and RKBS construction are worthwhile, and the proof problems are repairable. A referee should be asked to check Theorem 20 line by line, and the authors should either fix it or restrict the rate claim to the case where the analysis goes through.","headline":"Useful RKBS construction and a clean constrained-rate proof, but the headline linear-rate theorem is not proven as written and the Hilbertizability assumption undercuts the non-Hilbertian pitch.","tokens_in":32182,"tokens_out":4566,"would_cite":true,"duration_ms":39110,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46B10","46B20","46E22","68Q32","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Mirror descent on reproducing kernel Banach spaces converges linearly only in the near-Hilbert case; elsewhere a sublinear rate is the provable best.","keywords":["reproducing kernel Banach spaces","kernel methods","linear convergence","mirror descent","Hilbertizable Banach spaces","p-norm RKBS","strong convexity","Polyak–Łojasiewicz inequality"],"falsifier":"Take the paper's own $p$-norm RKBS with $p\\neq 2$ (which is not Hilbertizable) and try to construct, or numerically locate, a functional that is both $\\mu$-strongly convex and $\\gamma$-smooth with respect to its norm; Lemma 19 predicts none exists, so exhibiting one would refute the negative result and reopen the possibility of linear rates beyond Hilbertizable spaces.","tokens_in":31128,"feed_emoji":"📉","tokens_out":11160,"duration_ms":88391,"temperature":0.7,"pith_summary":"This paper tries to establish that mirror descent, a first-order method designed for non-Euclidean geometry, can be transplanted from finite-dimensional vectors to reproducing kernel Banach spaces (RKBS), where the gradient step is taken in the dual space using the reproducing kernel. Its central theoretical claim is a linear convergence rate $L(f_k)-L(f^*)\\le (L(f_0)-L(f^*))e^{-k\\mu\\nu^2\\kappa^2/\\gamma}$ under the assumptions that the loss is $\\mu$-strongly convex and $\\gamma$-smooth and the mirror map is $\\nu$-strongly convex and $\\rho$-smooth on a reflexive, Hilbertizable RKBS. The paper also proves that those smooth-and-strongly-convex assumptions are impossible on any Banach space not isomorphic to a Hilbert space, so the linear-rate regime is confined to near-Hilbert geometry. For constrained optimization, it proves the standard $O(1/\\sqrt{t})$ rate under only convexity and Lipschitz assumptions, and it constructs a concrete $p$-norm RKBS family with an explicit kernel and dual map, so the algorithm reduces to coefficient updates on a similarity matrix.","feed_headline":"Linear speed on kernel Banach spaces only if nearly Hilbert","feed_subtitle":"Fast rates for mirror descent hold exactly when the Banach space is Hilbert-like — plus a p-norm family to run it.","key_machinery":"The carrying object is the functional mirror descent update $g_t\\leftarrow g_{t-1}-\\eta\\,\\partial_{f_{t-1}}L$, $f_t\\leftarrow(\\partial\\Phi)^{-1}(g_t)$, where $\\Phi$ is a strictly convex mirror potential and its Gâteaux derivative $\\partial\\Phi$ maps the primal RKBS onto its dual; the reproducing kernel supplies the concrete dual-space gradient $2\\sum_i(f_{t-1}(x_i)-y_i)K(x_i,\\cdot)$, so the iteration runs on coefficient vectors. The linear-rate proof is driven by a convex-conjugate lemma (from Zălinescu) that converts $\\rho$-smoothness of $\\Phi$ into a lower bound $\\kappa$ on the dual norm of its differential, together with strong convexity of $L$ to sandwich the loss decrease. The negative result Lemma 19 uses second-order Taylor expansion points (Borwein–Noll) and Kwapień's generalized parallelogram-law characterization to conclude that simultaneous strong convexity and smoothness forces the Banach norm to be equivalent to a Hilbert norm.","core_discovery":"On its own terms, the paper's discovery is Theorem 20: on a reflexive reproducing kernel Banach space (RKBS, a Banach space of functions whose point evaluations are continuous, so a kernel exists) that is Hilbertizable—isomorphic, not necessarily isometrically, to a Hilbert space—and admits a functional that is both $\\mu$-strongly convex and $\\gamma$-smooth with respect to the Banach norm, mirror descent with learning rate $\\eta=\\min\\{\\nu/\\gamma,1/(2\\mu\\nu\\kappa^2)\\}$ converges linearly at rate $e^{-k\\mu\\nu^2\\kappa^2/\\gamma}$, where $\\nu$ and $\\rho$ are the strong-convexity and smoothness constants of the mirror map and $\\kappa$ depends on $\\rho$. The companion Lemma 19 states that no Banach space not isomorphic to a Hilbert space admits any such functional, which means the linear-rate assumptions cannot hold in genuinely non-Hilbertian RKBS; the same rate is shown under the Polyak–Łojasiewicz inequality, and a $O(1/\\sqrt{t})$ rate for constrained problems under convexity and Lipschitzness. To make the method concrete, the paper introduces a family of $p$-norm RKBSs ($p\\neq 2$) with an explicitly given reproducing kernel and a dual map whose inverse is the standard $\\ell^q$ gradient, reducing each iteration to vector updates of coefficients in $\\mathbb{R}^n$.","pith_inferences":["Extension: the negative Lemma 19 suggests that the practical gains of $p\\neq 2$ RKBS must come from approximation or sparsity, not from faster optimization; any claim of linear rates on genuinely non-Hilbertian spaces would need a different smoothness/strong-convexity notion than the paper's Banach-norm definitions.","Extension: the concrete $p$-norm construction invites a testable study in which small $p$ produces sparser coefficients and possibly better generalization on high-dimensional data, at the cost of leaving the linear-rate (near-Hilbert) regime for the constrained $O(1/\\sqrt{t})$ regime; the paper's step-function experiments show error decreasing with smaller $p$ but do not measure this trade-off.","Extension: the rate's constant $\\kappa$ is only said to 'depend on $\\rho$' via Proposition 21; an explicit expression for $\\kappa$ in terms of $\\rho$ and the isomorphism constants would let practitioners predict the actual contraction factor, and a numerical check on a Hilbertizable RKBS with known constants could test whether the bound is tight."],"forward_implications":["If the theorem is right, first-order optimization on RKBS can match the Euclidean linear-rate guarantee $e^{-k\\mu\\nu^2\\kappa^2/\\gamma}$ whenever the space is Hilbertizable and the loss and mirror map satisfy the smoothness/strong-convexity pair.","Because Lemma 19 forces Hilbertizability whenever both assumptions hold, the analysis draws a clean boundary: genuinely non-Hilbertian RKBS admit no fast-rate regime of this type, and existing representer theorems for such spaces come with no corresponding fast optimization guarantee.","The linear rate survives when the loss's strong convexity is relaxed to the Polyak–Łojasiewicz inequality, so the same algorithm is fast for losses that are not convex in the classical sense.","In the constrained setting, projected mirror descent converges at $O(1/\\sqrt{t})$ under only convexity and Lipschitzness of the loss plus strong convexity of the mirror map, with no smoothness or Hilbertizability required.","The explicit $p$-norm RKBS family turns each mirror step into $\\beta^{(t)}\\leftarrow \\beta^{(t-1)}-\\eta\\,\\hat{H}^\\top(\\hat{H}\\alpha^{(t-1)}-Y)$ and a dual-map normalization, meaning the method ports directly to similarity-matrix-based kernel code."],"supporting_citations":[{"why":"Supplies Corollary 3.5.7 (Proposition 21) and Theorem 3.5.13: the convex-conjugate pair that converts mirror-map smoothness into a dual-norm lower bound, and the reflexivity criterion used to justify the linear-rate proof.","marker":"Zălinescu (2002)"},{"why":"Theorem 4.1 guarantees second-order Taylor expansion points for continuous convex functions on separable Banach spaces, the starting point of Lemma 19.","marker":"Borwein and Noll (1994)"},{"why":"Proposition 3.1 (generalized parallelogram law) characterizes Banach spaces isomorphic to Hilbert spaces, which is exactly the conclusion Lemma 19 derives.","marker":"Kwapień (1972)"},{"why":"Introduces mirror descent, the method the paper adapts to the functional RKBS setting.","marker":"Nemirovski and Yudin (1983)"},{"why":"Provides the standard mirror descent convergence analysis that the paper's setup extends to Banach function spaces.","marker":"Beck and Teboulle (2003)"},{"why":"Theorem 4.2 supplies the proof template for the constrained $O(1/\\sqrt{t})$ rate in Theorem 25.","marker":"Bubeck (2015)"},{"why":"Demonstrates linear convergence under the Polyak–Łojasiewicz inequality in the Euclidean setting, motivating Theorem 22.","marker":"Karimi et al. (2016)"},{"why":"Origin of the RKBS framework and its unique kernel representation, the target geometry for the mirror descent updates.","marker":"Zhang et al. (2009)"},{"why":"Provides the generic RKBS definitions and construction framework adopted in Section 2.","marker":"Lin et al. (2022)"},{"why":"Lemma 9 gives the strong convexity parameter $(p-1)$ of $\\frac12\\|\\cdot\\|_p^2$ for $p\\in(1,2)$, justifying the mirror map used in Algorithm 1.","marker":"Kakade et al. (2012)"}],"fun_headline_variants":["Fast rates in kernel Banach spaces only if nearly Hilbert","Mirror descent on RKBS: linear speed only for Hilbert-like","Nearly Hilbert RKBS admits linear mirror descent","Fast mirror descent on RKBS requires Hilbert-like structure","Mirror descent on kernel Banach spaces speeds only when Hilbert-like"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 2, that the RKBS admits a functional that is both $\\mu$-strongly convex and $\\gamma$-smooth with respect to the Banach norm—which, by the paper's own Lemma 19, forces the space to be isomorphic to a Hilbert space, so the headline linear rate applies only in a near-Hilbert setting.","fun_headline_variants_meta":{"raw":{"variants":["Fast rates in kernel Banach spaces only if nearly Hilbert","Mirror descent on RKBS: linear speed only for Hilbert-like","Nearly Hilbert RKBS admits linear mirror descent","Fast mirror descent on RKBS requires Hilbert-like structure","Mirror descent on kernel Banach spaces speeds only when Hilbert-like"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000494,"raw_usage":{"total_tokens":2481,"prompt_tokens":1059,"completion_tokens":1422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":1341}},"tokens_in":675,"tokens_out":1422,"duration_ms":25802,"temperature":1.0,"reasoning_tokens":1341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:46:27.559878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the paper's own $p$-norm RKBS with $p\\neq 2$ (which is not Hilbertizable) and try to construct, or numerically locate, a functional that is both $\\mu$-strongly convex and $\\gamma$-smooth with respect to its norm; Lemma 19 predicts none exists, so exhibiting one would refute the negative result and reopen the possibility of linear rates beyond Hilbertizable spaces.","supporting_citations":[],"review_version":1}