{"id":"5e9c5493-d722-4031-93d9-6d415142d78c","arxiv_id":"2501.12896","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"π-Quant's core representation theorem fails: the curve e^{iθ}+e^{iπθ} is dense in the disk but does not cover it, and Lemma 3.2's angle formulas are internally inconsistent.","lead":"A new optimizer-state compression method, π-Quant, claims to store parameter pairs as a single angle using irrational rotations, cutting state bit-width to 3.32 bits per parameter. The central theorem behind this scheme is false, because the rotation curve is dense but not surjective, and the angle solver uses the wrong branch.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's surjectivity claim is false: f(θ)=e^{iθ}+e^{iπθ} is a dense 1D curve, not all of the disk, so exact pair-to-angle compression is impossible; this collapses the lossless 3.32-bit premise.","rationale":"The reader's REJECT verdict is well supported and should not be changed. I partially agree with the reader's weakest_assumption: the branch error in Lemma 3.2 is a genuine additional defect, but the more fundamental failure is Theorem 3.1's surjectivity. Without a surjective map, the entire concept of exactly representing an arbitrary parameter pair by a single angle is invalid, independent of the angle-solving equations. Appendix A's transition from non-periodicity and boundedness to 'traverses every point' is exactly the invalid step that creates the false theorem; the dense-curve counterexample is standard and directly applies. Lemma 3.2's branch error, although real, is secondary because even a correct branch computation could not make the curve cover the disk. The experimental section reports favorable training metrics, but it does not measure reconstruction error on random pairs, and no error bars or artifact checks are provided; thus the empirical results cannot rescue the theoretical guarantee. Because the central mechanism is mathematically unsound, keeping the reader's REJECT verdict is appropriate. No ad hominem is intended; the critique targets the argument's validity only.","tokens_in":15866,"tokens_out":7932,"duration_ms":82619,"concrete_test":"Take z=(1,0). Using f(θ)=2cos(((1−π)θ)/2)e^{i((1+π)θ)/2}, if f(θ)=(1,0) then ((1−π)/2)θ must be π/3+nπ or 2π/3+nπ. The correct direction condition then forces ((1+π)/(1−π))(n+1/3) or ((1+π)/(1−π))(n+2/3) to be an integer for some integer n, which is impossible since (1+π)/(1−π) is irrational. This analytic check settles that Theorem 3.1's surjectivity is false. A complementary numerical check: densely sample θ and compute the minimum distance from f(θ) to (1,0); the positive lower bound predicted by the algebra confirms that the lossless pair-encoding premise fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central load-bearing claim is Theorem 3.1, which asserts that every z with ||z||≤2 equals e^{iθ}+e^{iπθ} for a unique θ. This is false. The map f(θ) is a real-analytic curve: f(θ)=2cos(((1−π)θ)/2) exp(i((1+π)θ)/2). Its image is one-dimensional, dense in the disk when π is irrational, but it has empty interior; it cannot contain every point of the disk. The proof in Appendix A makes an invalid inference: non-periodic + continuous + bounded does not imply that the trajectory traverses every point; a dense non-closed curve is the canonical counterexample. For a concrete counterexample, f(θ)=(1,0) would require |2cos(((1−π)θ)/2)|=1 and direction ((1+π)θ)/2 ≡ 0 or π (mod 2π). Substituting θ=(π/3+nπ)·2/(1−π) or θ=(2π/3+nπ)·2/(1−π) forces ((1+π)/(1−π))(n+1/3) or ((1+π)/(1−π))(n+2/3) to be an integer, impossible because (1+π)/(1−π) is irrational. Thus generic pairs (x,y) have no exact θ. Since the algorithm computes θ from (x,y) and later reconstructs from θ, the advertised lossless halving of parameter scale in Eq. (3)/(11) cannot hold for arbitrary optimizer states. Lemma 3.2 compounds this: α−β=θ and α+β=πθ are only valid on a branch where cos(((1−π)θ)/2)>0; the method never tracks this branch, so even for curve points Eq. (11) can recover the wrong pair. The empirical results may indicate a useful approximate quantizer, but the paper's central theoretical guarantee—exact representation at 3.32 bits—is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes π-Quant, a training-oriented optimizer-state compression method. It claims a new theorem that every complex number z with ||z||≤2 can be written uniquely as z=e^{iθ}+e^{iπ̄θ}; based on this it represents each pair of optimizer-state entries by one rotation angle θ, computes θ through a \"geometric solution\" (Lemma 3.2), quantizes θ to λ decimal digits, and reports 3.32-bit optimizer states with full accuracy on TinyLlama and several benchmark tasks. The appendix attempts to prove Theorem 3.1, Lemma 3.2, and the quantization-error bounds. The paper also provides code and a linear-complexity pipeline.","tokens_in":16343,"tokens_out":6268,"duration_ms":62869,"significance":"If the central theorem were true, representing two real parameters by one angle would halve the optimizer-state tensor size and could enable low-bit training, which would be an important result. The paper provides code, covers relevant baselines, and tests on a range of tasks. However, the central mathematical premise is false: the trajectory of f(θ)=e^{iθ}+e^{iπθ} is a dense one-dimensional curve, not the full disk, so exact pair-to-angle compression is impossible; the geometric branch used in Lemma 3.2 is wrong for a large fraction of representable points; and the proof in Appendix A contains an invalid inference. The experiments lack error bars and use a small number of training steps, so they cannot rescue the unsupported lossless and full-accuracy claims. The paper may describe a workable approximate quantizer, but the advertised theoretical guarantee and the 3.32-bit lossless claim require a fundamentally different argument.","major_comments":[{"comment":"The surjectivity claim of Theorem 3.1 is false. Writing f(θ)=e^{iθ}+e^{iπθ} in polar form gives f(θ)=2cos(((1−π)θ)/2)exp(i((1+π)θ)/2). The image is a real-analytic curve with empty interior; when π is irrational it is dense in the disk, but density does not imply surjectivity, and most points of the disk have no preimage. The proof in Appendix A explicitly uses the invalid inference that the trajectory \"will traverse every point within the domain's circle, since it is non-closed and continuous.\" A standard counterexample is a dense curve on the torus, which is non-closed and continuous but not surjective. Consequently Eq. (3) and the reconstruction in Eq. (11) cannot guarantee exact recovery of an arbitrary pair (x,y), and the lossless 3.32-bit claim collapses.","section":"Theorem 3.1 and Appendix A"},{"comment":"The system in Eq. (4) uses the wrong branch of the geometry. From the polar form, the argument of f(θ) is ((1+π)θ)/2 only when cos(((1−π)θ)/2)>0 and the arctangent branch is chosen appropriately; otherwise the argument is shifted by π. Correspondingly, the identities α−β=θ and α+β=πθ in Eq. (4) hold only on the branch where cos(((1−π)θ)/2)>0; on the branch where this cosine is negative, the roles are interchanged, giving α−β=πθ and α+β=θ (mod 2π). The method never tests or stores this branch, so Algorithm 1 and the reconstruction in Eq. (11) can return the wrong pair even for points that do lie on the curve. This is load-bearing because m and g in Eqs. (8) and (10) are computed from α−β under the false branch assumption.","section":"Lemma 3.2, Eq. (4), Appendix B"},{"comment":"The proof of Lemma A.1 is invalid at the step \"Subtract Eq. (16) from Eq. (15)\": Eq. (15) contains the original functions cosθ and sinθ, while Eq. (16) contains their derivatives, so subtracting the two systems does not produce Eq. (17). Moreover, even if f were non-periodic, non-periodicity of a bounded continuous curve does not imply surjectivity; the visualization in Figure 6 only shows dense-looking coverage over long θ-ranges and provides no evidence of exact coverage of every point. Since the proof of Theorem 3.1 rests on this inference, the central existence and uniqueness statement is unproved and in fact false.","section":"Appendix A, proof of Lemma A.1"},{"comment":"The claimed quantization-error advantage is arithmetically false. The text states \"2 · (1 + ¯π) < π\", but for ¯π≈π we have 2(1+π)≈8.28 > π, so the bound in Eq. (13) is larger than 10^{-λ}, not smaller. Additionally, the step \"Since ||y′||<1 (See Eq. (9))\" in the proof of Eq. (12) is wrong: y′=sinθ′+sinπ̄θ′ is a sum of two unit-magnitude terms and can be as large as 2, while Eq. (9) normalizes the stored parameters, not the trigonometric terms in Eq. (24). Therefore Eqs. (12) and (13) do not establish that the method has lower quantization error than traditional uniform quantization.","section":"Section 3.3, Eqs. (12) and (13)"},{"comment":"The claimed 3.32-bit setting is not a quantization bit-width in the usual sense. Eq. (10) stores m and g as two separate λ-digit decimal numbers inside one θ, so the total storage is 2λ decimal digits per θ, i.e., 3.32λ bits per parameter because one θ covers two parameters. The choice λ=1 gives 3.32 bits by construction, not by an information-theoretic analysis of the state distribution. The memory comparison in Table 3 also mixes this decimal-digit representation with FP32/FP16 baselines; the paper should report the actual byte-level layout before claiming a 41.9% memory reduction from the 3.32-bit setting.","section":"Section 3.2, Eq. (10) and Table 3"}],"minor_comments":[{"comment":"The notation ||∆x|| for a scalar quantity is misleading; use |∆x| throughout the error-bound discussion.","section":"Eq. (12)"},{"comment":"There are typos in the appendix and table headers: \"Appedix D\" should be \"Appendix D\" and \"belu scores\" should be \"BLEU scores.\"","section":"Appendix D and Table 4"},{"comment":"The sentence \"It is represented by multiplying a complex number of the form i.e., cos θ + i sin θ\" contains a duplicated \"i.e.,\" and should be rephrased.","section":"Section 2"},{"comment":"The initialization m0=Quant(0) and v0=Quant(0) is underspecified: Algorithm 1 expects a tensor with nontrivial X and Y splits, and the scaling step w=max(|X,Y|) is undefined for a zero tensor; specify how zero states are quantized.","section":"Algorithm 2"},{"comment":"The fractional-part notation {·} is used before being defined; define it explicitly when it first appears in Eq. (5).","section":"Eqs. (5)-(8)"}],"recommendation":"reject","confidential_remarks":"The central theorem of the paper is false and the proof in Appendix A is invalid at a load-bearing step, so this is not a case where additional experiments or polishing could fix the manuscript. A valid version would need to replace the surjectivity claim with a careful density or approximation statement and re-derive the algorithm and error analysis from scratch, likely with a substantially different method. I would advise against encouraging resubmission in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper's headline claim is wrong. Theorem 3.1 asserts that every point in the disk |z| ≤ 2 can be written uniquely as e^{iθ}+e^{iπθ}. That map is a continuous one-dimensional curve: it is dense when π is irrational, but its image has empty interior, so most points have no exact representation. The appendix 'proof' jumps from 'non-periodic' to 'traverses every point,' which is precisely the gap between dense and surjective. Lemma 3.2 also has a branch error: the relations θ = α−β and πθ = α+β only hold when cos(((π−1)/2)θ) > 0; when that cosine is negative the roles swap. The method never tracks the branch, so the reconstruction in Eq. (11) can output the wrong pair. What is genuinely new is the idea of using an irrational rotation coefficient to pack two real values into one angle. That is not present in the cited optimizer-quantization literature, and the masked-digit construction for approximating m is clever. The empirical evaluation on TinyLlama and other tasks follows a standard setup, and the reported numbers look competitive if the method is treated purely as a heuristic quantizer with a fixed codebook. But the paper frames everything around exact lossless compression, and without that the contribution is much weaker. The bit-width formula (3.32λ) only makes sense once you assume exact recovery, and the error bound in Eq. (13) misuses the triangle inequality. The experimental section lacks error bars, and the code link is unverified, but those are minor relative to the mathematical failure. My recommendation is clear: reject. The load-bearing theorem is false, and a serious referee would likely catch it quickly. I would desk reject and, if the authors resubmit, urge them to drop the exactness claim and present the method as an approximate quantization scheme with measured error bars.","headline":"The paper's central theorem is false: e^{iθ}+e^{iπθ} traces a dense 1D curve, not the whole disk, so the lossless 3.32-bit compression guarantee collapses, even though the empirical method may work as a heuristic.","tokens_in":16782,"tokens_out":4951,"would_cite":false,"duration_ms":49261,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that any pair of parameters inside a disk of radius 2 can be stored as a single rotation angle, and that quantizing these angles to 3.32 bits preserves full optimizer accuracy.","keywords":["optimizer state compression","low-bit optimizers","irrational complex rotations","rotation angle quantization","memory-efficient training","Adam optimizer","non-uniform quantization","pi-Quant"],"falsifier":"Sample a large set of random points from the disk $x^2+y^2\\le 4$, run the paper's Algorithm 1 with a large $\\lambda$ (e.g., $\\lambda=8$), reconstruct each point via Eq. (11), and measure the maximum Euclidean error between original and reconstructed values. If the error does not shrink toward zero as $\\lambda$ grows for points that are not on the curve, the exact-representation claim fails; if it reaches machine precision for the sampled points, that claim passes this test.","tokens_in":15647,"feed_emoji":"🥧","tokens_out":18595,"duration_ms":156522,"temperature":0.7,"pith_summary":"$\\pi$-Quant proposes to compress optimizer states — the momentum tensors kept by Adam and similar optimizers — by storing two real parameters as a single rotation angle. The paper claims a mathematical guarantee: for any pair $(x,y)$ inside the disk of radius 2, there is a real angle $\\theta$ such that $x+iy=e^{i\\theta}+e^{i\\bar\\pi\\theta}$, so the pair can be reconstructed exactly from $\\theta$. A system of geometric equations computes $\\theta$ in linear time, and each angle is then quantized to a few decimal digits; at $\\lambda=1$ the effective storage cost is about 3.32 bits per parameter. In experiments on a 1.1B-parameter language model and several other tasks, the method reports accuracy equal to or better than full-precision Adam while cutting GPU memory by about 40%. If the identity is exact as claimed, this would give a theoretically grounded way to halve optimizer-state size and push low-bit training below 4 bits without search-based quantization kernels.","feed_headline":"Two floats become one angle, cutting optimizer states to 3.32 bits","feed_subtitle":"π-Quant reports 3.32-bit Adam states with full accuracy and about 40% less GPU memory.","key_machinery":"The load-bearing object is the complex rotation identity $x+iy=e^{i\\theta}+e^{i\\bar\\pi\\theta}$, where $\\bar\\pi$ is constructed from $\\pi$ by shifting its digits so that $\\bar\\pi=10^{-\\lambda}+10^{-2\\lambda}\\cdot 0.3589793238\\ldots$; this digit layout lets one stored integer encode both the branch index $m$ and the fractional part $g$ of the angle. To solve for $\\theta$, the paper gives a geometric system (Lemma 3.2) that forms the triangle of the two unit vectors and derives $\\alpha=\\arctan(y/x)$, $\\beta=\\arccos(\\sqrt{x^2+y^2}/2)$, and then reduces the branch search to matching the fractional part of $m\\bar\\pi$ against $\\Omega=(\\alpha(1-\\bar\\pi)+\\beta(1+\\bar\\pi))/(2\\pi)$. The quantization step stores $\\lfloor(\\theta/2\\pi)\\cdot 10^\\lambda\\rfloor$, so the first $\\lambda$ digits carry $m$ and the last $\\lambda$ digits carry the remainder; the error analysis bounds the coordinate error by $\\Delta\\theta$ in each axis and the average error by $2(1+\\bar\\pi)\\,10^{-\\lambda}/\\pi$.","core_discovery":"On its own terms, the central discovery is the identity $z=e^{i\\theta}+e^{i\\bar\\pi\\theta}$: with an irrational coefficient $\\bar\\pi$ in the second rotation, the curve traced by $\\theta\\mapsto e^{i\\theta}+e^{i\\bar\\pi\\theta}$ is non-periodic, and the paper claims it passes through every point of the disk $\\|z\\|\\le 2$, so every two-dimensional pair $(x,y)$ can be assigned a unique angle $\\theta$ with $x+iy=e^{i\\theta}+e^{i\\bar\\pi\\theta}$. From this identity the authors derive that a pair of real numbers can be stored as one angle, halving the element count of optimizer state tensors, and that the angle can be quantized to $\\lambda$ decimal digits with a provable error bound, giving an effective bit-width of $3.32\\lambda$ bits. The paper further claims that with $\\lambda=1$ (a 3.32-bit state) training accuracy is fully preserved on language modeling and several downstream tasks, with memory savings of roughly 40% relative to full-precision Adam.","pith_inferences":["Read as a compression codec rather than an exact bijection, the scheme is a member of a larger family: any two real values can be mapped to an angle on a dense curve, with the bit budget set by $\\lambda$; this suggests applying the same idea to weight tensors, activations, or KV cache entries, not just momentum states.","The digit-shifting construction of $\\bar\\pi$ is one of infinitely many 'almost-rational' coefficients; choosing a different digit layout changes both the matching error for $m$ and the effective non-uniformity of the quantizer, so the coefficient itself becomes a tunable hyperparameter.","A testable extension is to measure the reconstruction error floor on random float pairs with large $\\lambda$; the paper's error analysis assumes the pair lies on the curve, so random pairs would reveal whether the scheme behaves as exact compression or as a bounded-error approximate codec in practice.","The non-uniform error profile the authors plot (small errors near zero, larger at the periphery) matches a curve whose density is higher near the origin; this could be exploited by pairing the codec with per-tensor scales that push the value distribution toward the high-density region."],"forward_implications":["Optimizer state tensors drop to half their element count (two reals stored as one angle), and with the 3.32-bit angle format this yields roughly a 75% reduction in parameter scale.","At $\\lambda=1$ the paper reports accuracy at or above full-precision Adam on language modeling and on ARC-Challenge, Hellaswag, Lambada, and PIQA, while GPU memory falls by about 40% (from 19.47 GB to 11.32 GB for a 1.1B-parameter model).","Because the method is arithmetic rather than search-based, it needs no custom GPU kernels and can set any bit-width simply by changing $\\lambda$.","The same angle representation applies to any tensor that can be split into two equal halves, so the authors suggest it can compress model weights and, as future work, KV caches in transformers."],"supporting_citations":[{"why":"Defines the Adam optimizer whose first- and second-moment states are the tensors the paper compresses.","marker":"(Kingma & Ba, 2015)"},{"why":"Supplies the 8-bit block-wise optimizer baseline that π-Quant must match or beat in the main experiments.","marker":"(Dettmers et al., 2021)"},{"why":"Supplies the 4-bit search-based optimizer baseline, the nearest competitor at low bit-width.","marker":"(Li et al., 2023)"},{"why":"Provides the PG-19 dataset used for the continual pre-training evaluation that supports the accuracy claims.","marker":"(Rae et al., 2020)"},{"why":"Provides the TinyLlama checkpoint used as the backbone in the main language modeling experiments.","marker":"(Zhang et al., 2024)"}],"fun_headline_variants":["One angle, two floats: π-Quant slashes optimizer states to 3.32 bits","π-Quant stores two floats as one angle, hitting 3.32-bit optimizer","Irrational complex rotations pack two floats into one 3.32-bit angle","Two floats, one angle: π-Quant cuts optimizer memory 40% at full accuracy","π-Quant uses irrational rotations to hit 3.32-bit optimizer states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that every parameter pair stored in an optimizer state lies exactly on the curve traced by the two rotating unit vectors, and that the branch relations used to compute the angle recover the same pair when the angle is inverted; pairs not on that curve, or lying on a different branch, are not reconstructed without error.","fun_headline_variants_meta":{"raw":{"variants":["One angle, two floats: π-Quant slashes optimizer states to 3.32 bits","π-Quant stores two floats as one angle, hitting 3.32-bit optimizer","Irrational complex rotations pack two floats into one 3.32-bit angle","Two floats, one angle: π-Quant cuts optimizer memory 40% at full accuracy","π-Quant uses irrational rotations to hit 3.32-bit optimizer states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000752,"raw_usage":{"total_tokens":3333,"prompt_tokens":921,"completion_tokens":2412,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2301}},"tokens_in":537,"tokens_out":2412,"duration_ms":16238,"temperature":1.0,"reasoning_tokens":2301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:42:59.066516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample a large set of random points from the disk $x^2+y^2\\le 4$, run the paper's Algorithm 1 with a large $\\lambda$ (e.g., $\\lambda=8$), reconstruct each point via Eq. (11), and measure the maximum Euclidean error between original and reconstructed values. If the error does not shrink toward zero as $\\lambda$ grows for points that are not on the curve, the exact-representation claim fails; if it reaches machine precision for the sampled points, that claim passes this test.","supporting_citations":[{"cited_title":"W., Potapenko, A., Jayakumar, S","cited_arxiv_id":null,"evidence_quote":"Provides the PG-19 dataset used for the continual pre-training evaluation that supports the accuracy claims."}],"review_version":1}