{"id":"72655d5e-0e2f-4ec2-ab65-0ea89357f0eb","arxiv_id":"2504.17594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Repeated JPEG compression converges to fixed-point images in finite steps, and such images expose later edits after a single recompression.","lead":"This paper proves that repeatedly compressing and decompressing a JPEG image eventually reaches a stable, unchanging image, and turns that stable state into a tamper-evident signature. It gives a mathematical proof of a phenomenon observed since 2010 and demonstrates an image forensics use.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof relies on D being orthonormal, but the DCT matrix P defined in §2.3 is not orthonormal: the DC row has norm √2, so D^⊤D≠I and the monotone-chain argument collapses.","rationale":"The reader's weakest assumption about finite-valued error terms does not hold up: because every x_t lies in Ω0, a finite set of 256^64 integer vectors, the set {D x_t} has at most 256^64 elements, so ϵ_t and η_t do take only finitely many values. The proof's monotonicity argument is therefore not refuted by that objection. The real flaw is more elementary: P as written is not orthonormal. The k=0 row of P has norm √2, so D^⊤D≠I; all isometry-based equalities in Theorem 2 are invalid. Theorem 3 relies on Theorem 2, so the central convergence proof is unsupported. The RGB subsampling concern in §2.5 is also legitimate and acknowledged by the authors as future work, but it concerns the application extension rather than the block-level theorem. Since the theorem may be salvageable by replacing P with the standard orthonormal DCT-II matrix, the appropriate verdict remains conditional: accept only after the normalization error and the RGB/subsampling justification are fixed. The paper provides empirical validation (Figs. 3–6), but no code or data artifacts, so independent reproducibility is limited.","tokens_in":7132,"tokens_out":13892,"duration_ms":134914,"concrete_test":"Numerically evaluate P_{kn}=0.5*cos(pi/8*(n+0.5)*k) for k,n=0..7 as defined in §2.3 and compute P^⊤P. If the first diagonal entry is 2 rather than 1, the orthonormality claim fails. Then re-derive Theorem 2 using the standard orthonormal DCT-II matrix (DC row 1/√8): check whether the equalities η_{t+1}=||D x_{t+1}-[D x_t]_q||=||x_{t+1}-D^⊤[D x_t]_q|| and the chain ϵ0≥η1≥ϵ1≥... are recovered. If the corrected matrix restores the chain, the error is a fixable normalization mistake; if not, the finite-step convergence claim needs a new proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing defect is in §2.3: the matrix P_{kn}=1/2 cos(π/8(n+1/2)k) for k,n=0..7 is not orthonormal. For k=0, P_{0,n}=1/2, so its squared Euclidean norm is 8×1/4=2, not 1; the standard orthonormal DCT-II uses DC row 1/√8 for k=0. Hence P^⊤P is a diagonal matrix with first entry 2, not the identity, and the claim \"P^⊤P=PP^⊤=I\" is false. Every subsequent equality that relies on D^⊤D=I is invalid: in Theorem 2, η_{t+1}=||D x_{t+1}-[D x_t]_q|| is asserted to equal ||x_{t+1}-D^⊤[D x_t]_q||, which requires D to be an isometry. Without that equality the monotone chain ϵ0≥η1≥ϵ1≥... does not follow. Theorem 3 depends on Theorem 2 for finite-step convergence of all blocks to a fixed point, so the central proof is unsupported as written. The reader's finite-valued objection is not the right issue: Ω0 is finite, so {D x_t} and the error terms take finitely many values; the actual gap is the false orthonormality of D.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims to provide the first formal proof that repeated JPEG compression/decompression, modeled as a transform T_q on 8×8 pixel blocks, has fixed points and that every block converges to a fixed point in finitely many steps. It then uses this property to propose tamper-evident JPEG images, in which any modification becomes detectable after a single additional JPEG round trip. The main proofs rely on viewing DCT, quantization, rounding, and truncation as Euclidean projections and on a monotone chain of error terms. The paper also states a theorem on distinct fixed points for sufficiently separated inputs and presents empirical demonstrations on 1M image blocks and on RGB images.","tokens_in":7444,"tokens_out":9359,"duration_ms":89328,"significance":"If the central theorem were correctly proved, it would give a rigorous explanation of a long-observed JPEG phenomenon and could be a useful building block for image forensics and integrity verification. The paper is self-contained and does not fit parameters to data; the finite-state monotone-chain strategy is a natural and potentially sound approach to the block-level problem. However, the current manuscript has a load-bearing error in the definition and properties of the DCT matrix, and the extension to RGB images rests on a false subsampling premise. With a corrected DCT normalization and a reworked proof, the block-level result is likely salvageable; the RGB application needs additional justification.","major_comments":[{"comment":"The matrix P defined by P_{kn}=1/2 cos(π/8 (n+1/2)k) is not orthonormal: for k=0, P_{0n}=1/2 for all n, so row 0 has squared Euclidean norm 2. The standard orthonormal DCT-II uses 1/√8 for the DC entry. Consequently P^T P = diag(2,1,...,1) ≠ I, and D=P⊗P^T is not an isometry. In addition, for the 2D transform X↦PXP^T, the correct vectorization is vec(PXP^T)=(P⊗P)vec(X), not (P⊗P^T)vec(X). Since Eq. (1) and the proof of Theorem 2 use D^T D=I and D^T as an exact inverse, the equalities ε_t=||x_t−D^T[Dx_t]_q|| and η_{t+1}=||x_{t+1}−D^T[Dx_t]_q|| do not follow, and the monotone chain ε_0≥η_1≥ε_1≥... is unsupported. Theorem 3's finite-step convergence is the central mathematical claim of the paper, so this defect is load-bearing.","section":"Section 2.3, Eq. (1)"},{"comment":"The claim that chroma downsampling and upsampling 'do not interfere with the fixed-point computation' is incorrect for standard 4:2:0 JPEG. Downsampling combines 2×2 pixel neighborhoods, and upsampling interpolates from neighboring chroma samples, so the pixels in a given block do interact with pixels in neighboring blocks. The RGB tamper-evident demonstration in Fig. 6 therefore rests on an unproven premise. This is also inconsistent with Section 5, where extending the theory to color-space transforms, downsampling/upsampling, and boundary padding is listed as future work. The Section 2.5 statement and the RGB application should either be removed or supported by a separate argument.","section":"Section 2.5"}],"minor_comments":[{"comment":"The projection property for quantization is stated as [Dx_{t+1}]_q = arg min_{z∈[Ω0]_q} ||Dx_{t+1}−z||, but the quantization projection defined in Section 2.2 is with respect to [Z^d]_q, the set of all integer multiples of q. Since [Ω0]_q is generally a strict subset, the displayed equality is not justified. The inequality ε_{t+1}≤η_{t+1} can be recovered using the global projection set, so this is a repairable flaw rather than a fatal one.","section":"Theorem 2 proof"},{"comment":"The paper should define the vectorization convention explicitly and use the correct Kronecker-product expression for the 2D DCT, e.g., vec(PXP^T)=(P⊗P)vec(X).","section":"Section 2.3"},{"comment":"The statement that the quality degradation is 'less than 25dB below the threshold of human visual perception' is ambiguous; the PSNR values should be compared with a clearly stated threshold.","section":"Figure 5"},{"comment":"The caption uses 'tempering' where 'tampering' is meant.","section":"Figure 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The block-level fixed-point theorem is likely true for a correctly normalized orthonormal DCT, and the monotone projection argument can probably be repaired. The elementary DCT normalization error and the incorrect Kronecker factor cast doubt on the current proof text, but they are fixable within a revision. The more serious gap is the RGB extension, which is contradicted by the paper's own future-work paragraph; the RGB claims should be substantially scaled back unless a proper treatment of 4:2:0 subsampling is added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has a good intuition and a clever application, but the main theorem as written is not proven. The DCT matrix P defined in §2.3 is not orthonormal: for k=0, P_{0,n}=1/2, so the row norm is sqrt(2), not 1. Hence P^TP = diag(2,1,...,1), not the identity. The proof of Theorem 2 relies on D being an isometry to equate ||Dx_t - [Dx_t]_q|| with ||x_t - D^T[Dx_t]_q||. Without that, the monotone chain ϵ0 ≥ η1 ≥ ϵ1 ≥ ... does not follow, and Theorem 3, which depends on Theorem 2, is unsupported. The reader's finite-valued objection is not the real problem: Ω0 is finite, so the error terms do take finitely many values. The stress-test note is correct, and it is a load-bearing flaw.\n\nWhat the paper does well: it gives the first formal framing of JPEG fixed points as an explicit transform, and the tamper-evident image construction is a genuinely new use of the phenomenon, removing the need for external fingerprints or watermarks. The numerical experiments are simple but consistent, and the related work is honest about Huang et al. 2010 and Niu et al. 2021. The paper reads as a sincere attempt to formalize a known empirical fact.\n\nThe soft spots beyond the orthonormality error: Section 2.5 claims fixed points for full RGB images, but the conclusion says extending to RGB is future work, and the claim about downsampling/upsampling not interfering is dubious for 4:2:0 subsampling. Also, no code or data are provided, though the experiments are not hard to reproduce.\n\nWho this is for: researchers in image forensics and anyone using double-JPEG artifacts for tamper detection. The tamper-evident application could be useful if the theory is fixed.\n\nRecommendation: I would send this to peer review because the idea has real merit and the forensic community would benefit from a correct proof. But the authors need to fix the DCT normalization error, rework the proof with the correct orthonormal transform (or handle the scaling explicitly), and either prove or retract the RGB claims. As it stands, the central theorem is unproven, so the paper needs major revision before acceptance.","headline":"The tamper-evident construction is a nice idea, but the proof of the central fixed-point theorem collapses because the DCT matrix as defined is not orthonormal.","tokens_in":7930,"tokens_out":2725,"would_cite":false,"duration_ms":24908,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Repeated JPEG compression converges to a stable fixed image, and that fixed image can expose any later edit on its own.","keywords":["JPEG fixed points","tamper-evident image","double JPEG compression","image forensics","discrete cosine transform","fixed point convergence","quantization table","image integrity authentication"],"falsifier":"For a fixed quality table, take a large set of random 8x8 blocks and compute $T_q$ repeatedly, recording the first $t$ with $x_{t+1}=x_t$ and the number of distinct values of the quantization error $\\epsilon_t$ before that. If any block runs for ten thousand iterations without stabilizing, or the error keeps taking new distinct values after ten thousand steps, the paper's finite-step theorem is false; if all blocks stabilize within a few hundred iterations across many quality tables, the practical claim is confirmed.","tokens_in":6936,"feed_emoji":"🔍","tokens_out":12996,"duration_ms":121196,"temperature":0.7,"pith_summary":"This paper tries to prove a long-observed phenomenon: repeating the same JPEG compression and decompression cycle on an image eventually leaves it unchanged, at a stable JPEG fixed point. It models one cycle as a transform $T_q$ on 8x8 pixel blocks and proves that every block reaches such a fixed point in finitely many iterations while the reconstruction error decreases monotonically. The authors then turn this into a tamper-evident image: compress until stable, and any later edit will make one more JPEG cycle change the edited blocks, exposing both the fact and the location of tampering without watermarks or stored fingerprints. If the theorems are right, any image distributed as JPEG can carry its own proof of authenticity and flag malicious edits after a single verification compression.","feed_headline":"Every JPEG converges to a fixed point that reveals tampering","feed_subtitle":"Recompress with the same settings: any tampered block changes, so a photo can prove its own authenticity.","key_machinery":"The load-bearing object is the vectorized JPEG transform $T_q(x) = ([D^{\\top}[Dx]_q])^{255}_{0}$, where $D = P \\otimes P^{\\top}$ is a 64x64 orthonormal discrete-cosine-transform matrix and $[\\cdot]_q$ is coefficient-wise quantization to multiples of the quantization step. The proof exploits orthonormality: multiplying by $D$ preserves Euclidean distance, so the quantization error $\\epsilon_t = \\|Dx_t - [Dx_t]_q\\|$ and the rounding and truncation error $\\eta_{t+1} = \\|x_{t+1} - D^{\\top}[Dx_t]_q\\|$ satisfy $\\epsilon_t \\ge \\eta_{t+1} \\ge \\epsilon_{t+1} \\ge 0$. Because each operation is a projection onto a finite set of integer blocks, the paper argues the chain can descend only finitely often and therefore ends at a fixed point. For the tamper-evident application, the same mechanism means the transform is idempotent at the fixed point: recompressing with the same table changes nothing unless the pixels have been edited.","core_discovery":"The central claim is that the JPEG transform, written in vectorized form as $T_q(x) = ([D^{\\top}[Dx]_q])^{255}_{0}$, is a projection process whose error decreases monotonically: quantization followed by rounding and truncation gives a non-increasing chain that must terminate at a point $x$ with $x = T_q(x)$. Theorem 3 states that the fixed-point set is exactly the terminal layer of the nested sequence of reachable sets, and every starting block lands in it after finitely many steps; Theorem 4 adds that blocks already separated by more than the threshold $2\\delta$ cannot converge to the same fixed point. The paper reports experiments on one million random 8x8 blocks showing that consecutive-image differences vanish in a few cycles and that the stable image remains visually close to an ordinary single-compressed JPEG. On that basis it constructs a tamper-evident RGB image whose altered blocks are localized by one recompression with the same quantization table.","pith_inferences":["Going beyond the paper: the same fixed-point distance could be used as a continuous tamper score rather than a binary detector, since the deviation from the fixed point should grow with the strength of the edit; the paper does not quantify this scaling.","Going beyond the paper: the full-RGB extension assumes chroma subsampling does not couple neighboring pixels, but for 4:2:0 averaging subsampling those pixels do interact, so RGB fixed points may be only approximate; a direct test would be to build the fixed point with real subsampling and measure residual drift over cycles.","Going beyond the paper: if finite-step convergence ever failed for some quantization table, the tamper-evidence construction would still work approximately, with verification based on a small tolerance instead of exact equality; the paper does not discuss this fallback."],"forward_implications":["Any 8x8 block compressed repeatedly with a fixed quantization table becomes stable after finitely many cycles, so a JPEG that has already converged is invariant under one more identical compression.","Typical images reach their fixed point within a few cycles, so producing a tamper-evident image costs only a small number of JPEG recompressions.","If the fixed-point image is edited, one additional JPEG cycle with the same table produces nonzero differences exactly in the edited blocks, localizing the tampering.","The scheme needs no watermark embedding and no externally stored fingerprint; the image itself is the verifier, as long as the verification quantization table matches the one used to create it.","Blocks that are far enough apart before convergence are guaranteed to converge to different fixed points, so distinct image regions do not accidentally merge into a single stable state that hides edits."],"supporting_citations":[{"why":"Defines the JPEG compression and decompression pipeline and the orthonormal 8-point discrete cosine transform that the paper formalizes as the transform T_q.","marker":"[1]"},{"why":"First observed the fixed-point phenomenon for same-quantization JPEG compression, the empirical fact the paper proves and generalizes.","marker":"[3]"},{"why":"Introduced the tamper-evident image idea based on fixed points of a Gaussian transform, the application the paper re-implements using JPEG fixed points.","marker":"[4]"},{"why":"Supplies the RGB-to-YUV color conversion used in the full-image fixed-point extension.","marker":"[6]"}],"fun_headline_variants":["JPEG loops end in a fixed point that exposes tampering","Recompress to reveal: JPEG fixed points flag any photo edit","One recompression can show if a JPEG was tampered with","JPEG's stable loop: a natural tool for tamper detection","Tamper-evident images from JPEG's convergence property"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small errors introduced at each JPEG round can take only finitely many different values, which forces the decreasing error sequence to stabilize after finitely many rounds; the frequency transform uses irrational matrix entries, so this finite-value premise is not automatic, and the RGB version also assumes chroma subsampling does not couple neighboring pixels.","fun_headline_variants_meta":{"raw":{"variants":["JPEG loops end in a fixed point that exposes tampering","Recompress to reveal: JPEG fixed points flag any photo edit","One recompression can show if a JPEG was tampered with","JPEG's stable loop: a natural tool for tamper detection","Tamper-evident images from JPEG's convergence property"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1651,"prompt_tokens":844,"completion_tokens":807,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":460,"tokens_out":807,"duration_ms":7582,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:35:37.730726+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed quality table, take a large set of random 8x8 blocks and compute $T_q$ repeatedly, recording the first $t$ with $x_{t+1}=x_t$ and the number of distinct values of the quantization error $\\epsilon_t$ before that. If any block runs for ten thousand iterations without stabilizing, or the error keeps taking new distinct values after ten thousand steps, the paper's finite-step theorem is false; if all blocks stabilize within a few hundred iterations across many quality tables, the practical claim is confirmed.","supporting_citations":[{"cited_title":"Tamper-evident Image using JPEG Fixed Points","cited_arxiv_id":"2504.17594","evidence_quote":"Defines the JPEG compression and decompression pipeline and the orthonormal 8-point discrete cosine transform that the paper formalizes as the transform T_q."},{"cited_title":"Digital images have been subject to edits since their inception, ranging from harmless enhancements to deliberate manipulations with malicious intent","cited_arxiv_id":null,"evidence_quote":"First observed the fixed-point phenomenon for same-quantization JPEG compression, the empirical fact the paper proves and generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced the tamper-evident image idea based on fixed points of a Gaussian transform, the application the paper re-implements using JPEG fixed points."},{"cited_title":"The jpeg still picture compression stan- dard,","cited_arxiv_id":null,"evidence_quote":"Supplies the RGB-to-YUV color conversion used in the full-image fixed-point extension."}],"review_version":1}