{"id":"75cda78f-a0ae-4db8-a90b-bb067e6011ef","arxiv_id":"2502.00851","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The elbow point is defined as the k minimizing the tangent of the angle between adjacent segments of the SSE curve under a slope-increase condition, using only addition, subtraction, multiplication, and division.","lead":"This paper proposes an arithmetic formula for choosing the number of clusters in k-means by measuring the sharpest bend in the error curve. The authors aim to replace the subjective visual elbow heuristic with an objective, easily programmed rule.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proposed 'objective' elbow criterion is not scale-invariant: multiplying SSE by a constant changes the slopes in Eq. (4) nonlinearly and can change which k minimizes tan ψ_k, so the central claim of objectivity fails.","rationale":"Reading the paper in good faith, the central claim is that the minimizer in Theorem 2.5 gives the optimum k objectively. The trigonometric identity is derived correctly, and the exclusion of non-flattening points is at least well-defined. However, even if one grants the 'closest-to-90°' heuristic, the criterion is not invariant under the natural scaling operation of multiplying all data (and hence all SSE values) by a constant. The explicit two-sequence example above shows the selected k changes; therefore the proposed method cannot be the claimed improvement over visual, scale-dependent elbow plots. This is a direct falsification of the central claim, not merely a disagreement with the heuristic, and it warrants rejection. The reader's verdict already identifies scale-dependence as one of two fragile premises; my stress-test makes it the primary, decisive objection, so the rejection remains unchanged.","tokens_in":8480,"tokens_out":7971,"duration_ms":82652,"concrete_test":"Run the paper's Algorithm 1 (or Eq. (4)) on the two SSE arrays A = [100, 60, 40, 30, 28] and B = [1, 0.6, 0.4, 0.3, 0.28] (B = 0.01·A). A yields k = 4 as the minimum tangent; B yields k = 2. If the method's output changes under this scaling, the criterion is not objective or scale-invariant. Optionally verify that B is exactly the SSE array obtained after multiplying all coordinates of a dataset with SSE array A by 0.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is the scale dependence of the criterion in Theorem 2.5. The formula Eq. (4) uses raw slopes m_lk = SSE(k+1) − SSE(k). If all SSE values are multiplied by c > 0, every slope is multiplied by c and tan ψ_k becomes c(m_{k−1}−m_k)/(1 + c^2 m_{k−1}m_k). Since m_{k−1} < m_k < 0, the denominator changes nonlinearly, so the ordering of the tan ψ_k values is not invariant. This is not merely a plotting artefact: rescaling every data coordinate by a constant multiplies the k-means SSE by c^2 while leaving cluster assignments unchanged, so any scale-invariant clustering criterion must give the same k. A concrete counterexample inside the theorem's domain is SSE = (100, 60, 40, 30, 28). For c = 1, the admissible tangents are tan ψ_2 ≈ −0.0250, tan ψ_3 ≈ −0.0498, tan ψ_4 ≈ −0.381, so the method selects k = 4. For c = 0.01 (the same SSE curve after data are rescaled by a factor 0.1), the tangents are tan ψ_2 ≈ −0.185, tan ψ_3 ≈ −0.098, tan ψ_4 ≈ −0.0798, so the method selects k = 2. The flattening condition m_lk > m_lk−1 holds in both cases. Thus Theorem 2.5 returns different 'optimum' k for the same underlying clustering problem. The paper provides no normalization or canonical scale before applying Eq. (4), despite motivating the method by distorted scaling (Figures 3–4).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an automated elbow method for k-means clustering. It models the sequence (k, SSE(k)) as a polyline, defines at each interior point the upward-facing angle between the two adjacent segments, and selects as the elbow the point minimizing tan(ψ_k) among points where the consecutive SSE drops are decreasing (Theorem 2.5). The authors claim this makes the elbow method objective while using only arithmetic operations. They provide pseudocode, a Python implementation, and a single eight-point 2D example, from which they conclude that the optimal number of clusters is 6.","tokens_in":8851,"tokens_out":8740,"duration_ms":90532,"significance":"If the proposed criterion were valid, it would offer a simple and objective rule for choosing the number of k-means clusters, and the use of only arithmetic operations is a practical advantage over inverse-trigonometric or logarithmic alternatives. The paper is transparent about its derivation, ships no fitted parameters, and includes reproducible code, which are strengths. However, the central claim is undermined by a scale-invariance failure: the selected k changes when every SSE value is multiplied by a constant, even though such a rescaling leaves the k-means cluster assignments unchanged. Since the paper explicitly motivates the method by the distortions caused by arbitrary axis scaling, this is a load-bearing defect. The empirical validation is also limited to one illustrative dataset with visual confirmation rather than external criteria.","major_comments":[{"comment":"The proposed elbow criterion is not invariant under uniform rescaling of the SSE values, and this directly contradicts the paper's motivating problem of distorted axis scaling. Under the scaling SSE(k) -> c*SSE(k), every slope m_k becomes c*m_k, so tan(ψ_k) in Eq. (4) becomes c*(m_{k-1} - m_k)/(1 + c^2*m_{k-1}*m_k). Because the denominator changes nonlinearly, the ordering of the tangents can change. For example, take SSE = (100, 60, 40, 30, 28). For c = 1 the admissible tangents are approximately -0.025, -0.050, and -0.381 at k = 2, 3, 4, so the method selects k = 4. For c = 0.01, which corresponds to multiplying every SSE by 0.01 and hence to rescaling all data coordinates by 0.1, the tangents become approximately -0.185, -0.098, and -0.080, so the method selects k = 2. The condition m_k > m_{k-1} is satisfied in both cases. Uniform scaling of all coordinates leaves k-means assignments unchanged and multiplies SSE by c^2, so the same clustering problem yields different 'optimum' k. The manuscript provides no canonical normalization before applying Eq. (4), so the claimed objectivity fails.","section":"Theorem 2.5, Eq. (4)"},{"comment":"The optimality premise that the elbow is the corner whose upward-facing angle is closest to 90 degrees is assumed rather than derived. Theorem 2.5 proves only a formula for that geometric quantity under the additional flattening condition; it does not connect the criterion to cluster quality, stability, separation, or any external notion of optimal k. The simulation in Section 2.3 uses the algorithm's own smallest tangent at k = 6 as evidence that k = 6 is correct, and the text then asserts that the k = 6 clustering in Figure 14 is better than the k = 3 clustering in Figure 15. There is no comparison with silhouette analysis, gap statistics, ground-truth labels, or any quantitative validation. Consequently, the manuscript's claim that the method gives 'a more reliable cluster determination' is not empirically supported.","section":"Sections 2.1 and 2.3"}],"minor_comments":[{"comment":"The final equivalence in the proof contains a sign error: the displayed expression should read SSE(k) - SSE(k + 1) >= SSE(k - 1) - SSE(k), not SSE(k) - SSE(k + 1) >= -SSE(k - 1) - SSE(k). The surrounding text uses the correct inequality.","section":"Proposition 2.4"},{"comment":"The pseudocode for the SSE(k) function is incorrect as written: it loops only i = 1 to k over centroids and sums d(Xi, Ci), but it should sum over all n data points, assigning each point to its nearest centroid. The Python implementation uses kmeans.inertia_, so the code is correct, but the pseudocode does not match it.","section":"Algorithm 1"},{"comment":"The f-string print statement 'Tanpsi(i+1) = tanpsi[i]' will not produce the intended values when copied literally, and the variable name 'optimal k' should be 'optimal_k' for valid Python. These are presentation issues in an otherwise useful code listing.","section":"Algorithm 2"},{"comment":"Figure 15 has the placeholder caption 'Figure example', and there are several typographical errors throughout, including 'clusering', 'one oh such programming languages', 'Is is answered', and 'choses'. These should be corrected before publication.","section":"Figures and text"},{"comment":"The theorem assumes SSE(k+1) <= SSE(k), but k-means can produce equal SSE values across neighboring k; the algorithm sets tanpsi[k] = 0 in such cases, and the first such index is then returned as the optimum, which is an arbitrary tie-breaking rule. The claim that the algorithm 'considers every possibility of the elbow method graph behaviour' is therefore not fully justified.","section":"Theorem 2.5 assumptions"}],"recommendation":"reject","confidential_remarks":"The scale-invariance counterexample is decisive and lands directly on the main claim: the method returns different cluster counts for the same clustering problem under uniform data rescaling. This is not a local fix within the current scope, since the manuscript proposes the raw-SSE tangent formula as the objective criterion and offers no normalization or external validation that could repair the claim. The paper may be of interest as a heuristic after substantial reworking, but in its present form it does not meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely useful part is a closed-form, arithmetic-only rule for reading an elbow off an SSE plot: choose k that minimizes tan(psi_k) = (-SSE(k+1) + 2SSE(k) - SSE(k-1)) / (1 + (SSE(k)-SSE(k-1))(SSE(k+1)-SSE(k))) subject to mlk > mlk-1. That is a legitimate simplification of Shi et al.'s cosine-rule approach, and the derivation from the standard angle-between-lines identity is correct. No fitted parameters, no inverse trig, and the pseudocode and Python implementation are concrete and reproducible on the toy example.\n\nThe soft spots are load-bearing. The closest-to-90-degree criterion is assumed, not derived from cluster quality, stability, or any accepted cluster validity index. Worse, the angle measure is not invariant to vertical rescaling of the SSE axis. Multiply every SSE by a constant c and the tangents change nonlinearly, so the minimizing k can change. A concrete counterexample within the theorem's domain: SSE values (100, 60, 40, 30, 28) give k = 4; the same curve scaled by 0.01 gives k = 2. Since rescaling the data by a constant multiplies SSE by c^2 without changing cluster assignments, an objective criterion cannot behave that way. The paper's motivating concern is distorted plots, but this method is itself vulnerable to scaling distortion.\n\nThe validation is also weak. There is one toy dataset, and the conclusion that k = 6 is correct is based on the same visual flattening the method is meant to replace. No comparison to silhouette, gap statistic, or known ground truth. The facing-downward corner exclusion in Theorem 2.5 is post hoc, introduced after seeing the failing case, and is not independently validated.\n\nThe citation pattern is ordinary; no red flags there. But the central claim of objectivity is unsupported, and the scale dependence alone would sink the method as stated.\n\nThis paper could work as a classroom exercise in analytic geometry or real analysis, maybe a short teaching note. As a clustering methodology it is not ready. I would desk-reject in current form, and tell the authors to fix the scale issue and validate on multiple datasets with known cluster structure if they want a fair referee later.","headline":"The rational-form elbow formula is a neat pedagogical simplification, but the method misses its stated goal because the angle criterion is scale-dependent and the validation is circular.","tokens_in":9339,"tokens_out":2565,"would_cite":false,"duration_ms":27374,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that the elbow point in k-means can be identified exactly by minimizing the tangent of the upward-facing corner angle, computed from three consecutive SSE values.","keywords":["k-means clustering","elbow method","number of clusters","angle between lines","sum of squared errors","cluster validation","analytical geometry","objective criterion"],"falsifier":"Run the formula on a synthetic dataset whose true cluster count is known, then repeat after multiplying every $SSE(k)$ by a constant; if the selected $k$ changes, the method depends on arbitrary vertical scaling, contradicting its motivation. A simpler check: if the argmin of $\\tan(\\psi_k)$ is not the number of true components in a well-separated mixture, the angle criterion itself fails.","tokens_in":8263,"feed_emoji":"📐","tokens_out":6515,"duration_ms":60263,"temperature":0.7,"pith_summary":"K-means clustering requires the user to fix the number of clusters in advance, and the standard elbow method—looking for the point where the SSE curve bends most sharply—is a visual, subjective heuristic. This paper claims to turn that heuristic into an objective rule: the elbow point is the k whose upward-facing corner angle is closest to 90 degrees, and that condition can be computed exactly with a simple formula for the tangent of the angle using only three consecutive SSE values. A monotonicity property of the tangent function lets the authors select the closest-to-90-degree corner by minimizing the tangent, and an extra slope condition removes corners where the SSE drop is still growing, which would otherwise be false elbows. Because the resulting algorithm uses only addition, subtraction, multiplication, and division, it can run in any programming language without special math libraries. If the claim is right, choosing k becomes a deterministic arithmetic step rather than a matter of eyeballing a plot.","feed_headline":"A tangent formula finds k-means' true elbow","feed_subtitle":"The angle closest to 90 degrees is now computed from SSE values, not guessed from a plot.","key_machinery":"The carrying object is the tangent of the upward-facing corner angle, $\\tan(\\psi_k)$, defined for each interior point $k$ by the formula above. It comes from the angle-between-two-lines identity, and the derivative $\\sec^2 \\psi$ shows that $\\tan \\psi$ is strictly increasing on $(90^\\circ, 180^\\circ)$, so the minimizing tangent equals the angle closest to 90 degrees. The additional constraint $m_{l_k} > m_{l_{k-1}}$, where $m_{l_k} = SSE(k+1) - SSE(k)$, filters out corners that face downward in the sense of growing SSE drops, preventing the algorithm from selecting a point after which the SSE still falls sharply.","core_discovery":"The central result is Theorem 2.5: with $SSE(k)$ monotonically decreasing, the elbow point is $(k, SSE(k))$ that minimizes $\\tan(\\psi_k)$ over $k = 2, \\ldots, n-1$ satisfying $m_{l_k} > m_{l_{k-1}}$, where $$\\tan(\\psi_k) = \\frac{-SSE(k+1) + 2SSE(k) - SSE(k-1)}{1 + (SSE(k)-SSE(k-1))(SSE(k+1)-SSE(k))}.$$ The angle $\\psi_k$ is the upward-facing angle at the corner formed by connecting consecutive SSE points; it always lies between 90 and 180 degrees under the flattening assumption, and $\\tan$ is increasing on that interval, so the smallest (most negative) tangent is the angle closest to 90 degrees. The slope condition $m_{l_k} > m_{l_{k-1}}$ excludes corners at which the drop from $k$ to $k+1$ is at least as large as the drop from $k-1$ to $k$, because those points are not the start of a flattening tail. The authors derive the formula from the analytic-geometry identity for the angle between two lines, using the difference of slopes divided by one plus the product of slopes.","pith_inferences":["One consequence the paper leaves implicit is that the tangent formula is not invariant under vertical rescaling of SSE, so a preprocessing normalization (for example, dividing every $SSE(k)$ by $SSE(1)$) would make the chosen $k$ independent of plotting scale; the paper motivates its method with scaling distortion but does not include such a step.","A testable extension is to benchmark the formula on synthetic data with known cluster structure against the silhouette and gap statistics, which the paper does not do; its own example only demonstrates a single data set.","The algorithm as written breaks ties by returning the first $k$ that attains the minimum tangent; in cases where two $k$ values give very close tangents, a stability check across random k-means initializations would be a practical addition."],"forward_implications":["Choosing $k$ reduces to evaluating a rational expression in three consecutive SSE values, so any system that can compute SSE can implement the method.","The method automatically rejects candidates where the SSE drop is still increasing, avoiding the false elbow that can appear when a plot is drawn with a distorted vertical scale.","Because no trigonometric inverse or logarithm is involved, the algorithm ports to programming languages and embedded environments that lack math libraries.","The same tangent array can be plotted as a diagnostic: the minimizing $k$ is the elbow, and the tangent values show how much each candidate misses the 90-degree ideal.","If the formula is adopted, cluster-count selection for k-means becomes reproducible: two analysts running the same SSE sequence get the same $k$."],"supporting_citations":[{"why":"proves the angle-between-lines tangent identity that Theorem 2.1 and the final formula are built on.","marker":"[3]"},{"why":"supplies the definitions of k-means clustering and SSE that define the points being connected.","marker":"[2]"},{"why":"provides the elbow method and k-value selection background that the paper formalizes.","marker":"[16]"},{"why":"presents the prior quantitative elbow-point method based on the cosine rule, the main alternative the new formula improves on.","marker":"[12]"},{"why":"offers an unsupervised k-means variant whose logarithmic operations motivate the paper's design goal of using only standard arithmetic.","marker":"[13]"}],"fun_headline_variants":["Tangent formula pinpoints k-means elbow","Exact elbow point via tangent angle","K-means elbow: no more guessing","Tangents make k-means elbow exact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the upward-facing angle closest to 90 degrees—computed from raw SSE slopes with no normalization—marks the true elbow, a heuristic that is stated rather than derived from cluster quality or stability.","fun_headline_variants_meta":{"raw":{"variants":["Tangent formula pinpoints k-means elbow","Exact elbow point via tangent angle","K-means elbow: no more guessing","Tangents make k-means elbow exact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000443,"raw_usage":{"total_tokens":2253,"prompt_tokens":967,"completion_tokens":1286,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":1244}},"tokens_in":583,"tokens_out":1286,"duration_ms":9462,"temperature":1.0,"reasoning_tokens":1244,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:30:08.154204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the formula on a synthetic dataset whose true cluster count is known, then repeat after multiplying every $SSE(k)$ by a constant; if the selected $k$ changes, the method depends on arbitrary vertical scaling, contradicting its motivation. A simpler check: if the argmin of $\\tan(\\psi_k)$ is not the number of true components in a well-separated mixture, the angle criterion itself fails.","supporting_citations":[{"cited_title":"Pearson, 1992","cited_arxiv_id":null,"evidence_quote":"proves the angle-between-lines tangent identity that Theorem 2.1 and the final formula are built on."},{"cited_title":"Everitt.Cluster Analysis","cited_arxiv_id":null,"evidence_quote":"supplies the definitions of k-means clustering and SSE that define the points being connected."},{"cited_title":"Research on k-value selection method of k-means clustering algorithm","cited_arxiv_id":null,"evidence_quote":"provides the elbow method and k-value selection background that the paper formalizes."},{"cited_title":"A quantitative discriminant method of elbow point for the optimal number of clusters in clustering algorithm","cited_arxiv_id":null,"evidence_quote":"presents the prior quantitative elbow-point method based on the cosine rule, the main alternative the new formula improves on."},{"cited_title":"Sinaga and Miin-Shen Yang","cited_arxiv_id":null,"evidence_quote":"offers an unsupervised k-means variant whose logarithmic operations motivate the paper's design goal of using only standard arithmetic."}],"review_version":1}