Kakeya Conjecture and Conditional Kolmogorov Complexity

Kak ey a Conjecture and Conditional Kolmogoro v Complexit y Nic holas G. P olson ∗ Daniel Zan tedesc hi † W orking paper, Marc h 2026. Commen ts w elcome. Abstract This pap er dev elops an information-theoretic framew ork for algorithmic complexit y under regular identifiable fib ering. The central question is: when a decoder is given information ab out the fib er lab el in a fib ered geometric set, how muc h can the residual description length b e reduced, and when do es this reduction fail to bring dimension b elo w the am bien t rate? W e formulate a directional compression principle, prop osing that sets admitting regular, identifiable fiber decomp ositions should remain informa- tionally incompressible at ambien t dimension, unless the fib er structure is degenerate or adaptively c hosen. The principle is phrased in the language of algorithmic dimension and the p oint-to-set principle of Lutz and Lutz, which translates p oin twise Kolmogorov complexit y in to Hausdorff dimension. W e prov e an exact analytical result: under effec- tiv ely bi-Lipsc hitz and iden tifiable fibering, the complexit y of a p oin t splits additiv ely as the sum of fib er-lab el complexity and along-fiber residual complexity , up to logarithmic o v erhead, via the c hain rule for Kolmogoro v complexit y . The Kak eya conjecture (assert- ing that sets containing a unit segment in every direction ha v e Hausdorff dimension n ) motiv ates the framework. The conjecture w as recently resolved in R 3 b y W ang and Zahl [ 18 ]; it remains op en in dimension n ≥ 4, precisely b ecause adaptive fiber selection undermines the naive conditional split in the general case. W e isolate this adaptive- fib ering obstruction as the key difficult y and prop ose a formal researc h program con- necting geometric measure theory , algorithmic complexity , and information-theoretic compression. Keywor ds: Algorithmic dimension, conditional complexit y , fiber decomposition, geometric side information, Kakey a conjecture, Kolmogorov complexity , p oin t-to-set principle, source co ding. ∗ Bo oth Sc ho ol of Business, Universit y of Chicago. E-mail: ngp@chicagobooth.edu . † Muma College of Business, Univ ersity of South Florida. E-mail: danielz@usf.edu . 1 1 In tro duction Information theory has long studied how side information at the decoder reduces enco ding rate. In the classical Slepian–W olf and Wyner–Ziv frameworks, a compressor exploits cor- relation b et w een a source and the deco der’s side information to reduce transmission rate. This pap er develops a geometric v ariant: supp ose a p oint b elongs to a structured geometric ob ject organized by fibers (line segmen ts, curv es, or submanifolds) indexed b y some lab el. If the deco der is giv en side information ab out the fib er lab el, how m uch can the residual description length b e reduced? More formally , consider x ∈ R n decomp osed as x = ψ ( z , u ), where z is a fiber label and u is a co ordinate along the fib er. The enco der transmits a finite-precision description of x ; the deco der holds side information ab out z . The residual complexit y is the Kolmogoro v complexit y of u given z . The compression principle asks: under what conditions do es the total complexit y split additively , and when do es this additive split force the point to remain incompressible at the am bient informational dimension? 1.1 Kak eya as the Canonical Configuration The Kak eya conjecture pro vides the arc hetypal instance. A Kakey a set E ⊆ R n con tains a unit line segment in ev ery direction on S n − 1 . The conjecture asserts dim H ( E ) = n , despite the p ossibility of measure-zero constructions. The difficult y is fundamental: a com binatorial ric hness condition (a segment in every direction) m ust force a metrical conclusion (full Hausdorff dimension). The algorithmic viewp oin t makes the connection transparent. F or a p oint x = a ( e ) + te in a Kakey a set, the chain rule decomp oses description length at precision r in to three comp onen ts: K ( e ↾ r ) ≈ ( n − 1) r directions fill S n − 1 , K ( t ↾ r | e ↾ r ) ≈ r along-segmen t p osition, not determined b y e, K ( a ( e ) ↾ r | e ↾ r ) = O (log r ) base p oin t, under regularit y . Com bining: K ( x ↾ r ) ≈ ( n − 1) r + r = nr , so dim( x ) = n , and the p oint-to-set principle [ 1 ] then yields dim H ( E ) = n . The key subtlety is the regularit y of the base p oin t map e 7→ a ( e ). When the map is Lipsc hitz (the sticky Kakeya condition of Lutz and Stull [ 2 ]: | a ( e ) − a ( e ′ ) | ≲ | e − e ′ | ), the third term satisfies K ( a ( e ) ↾ r | e ↾ r ) = O (log r ) and the argument holds. In the non-sticky case, K ( a ( e ) ↾ r | e ↾ r ) can b e as large as O ( r ): the base p oin t is then informationally independent of the direction, providing an adversarial oracle with an extra dimension to exploit. F rom the 2 presen t p ersp ectiv e, this identifies a cen tral information-theoretic obstruction in dimensions n ≥ 4 (see Remark 2 ). 1.2 Algorithmic Dimension as the Unifying Language The precise language is provided by algorithmic dimension, defined as the limiting ratio of Kolmogoro v complexit y to precision. The p oint-to-set principle of Lutz and Lutz bridges p oin t wise algorithmic dimension to global Hausdorff dimension: for analytic sets, the Haus- dorff dimension equals the minimax v alue of oracle-relativized p oin twise dimension. This framew ork encodes the side-information question: an oracle represents a deco der with access to geometric side information, and proving full dimension requires sho wing that no oracle can simultaneously compress all p oints of the set b elow the am bien t rate. At the point wise lev el, the ambien t-dimension heuristic is an incompressibility claim, closely aligned with Martin-L¨ of-st yle algorithmic t ypicality under admissible side information. 1.3 Kolmogoro v Complexity and Conditional Complexit y The cen tral quantit y throughout the paper is Kolmo gor ov c omplexity . F or a binary string x , the Kolmogorov complexity K ( x ) is the length of the shortest program on a fixed universal prefix-free T uring mac hine that outputs x . Intuitiv ely , K ( x ) is the minimum n umber of bits needed to describ e x . The c onditional Kolmo gor ov c omplexity K ( x | y ) is the length of the shortest program that outputs x when giv en y as an auxiliary input. It measures how muc h information ab out x remains once y is known, the algorithmic analogue of conditional entrop y H ( X | Y ). The chain rule states K ( x, y ) = K ( y ) + K ( x | y ) + O (log K ( x, y )) , mirroring the Shannon iden tity H ( X , Y ) = H ( Y ) + H ( X | Y ). In this pap er, x is a p oin t in R n describ ed at finite precision r , z is a fib er lab el, and K ( x ↾ r | z ↾ r ) measures the residual description length once the fib er lab el is kno wn. The compression problem is to understand when and by how muc h this conditional complexity is reduced relative to K ( x ↾ r ). 1.4 The Adaptiv e-Fib ering Obstruction The compression principle presupp oses a well-defined fib er assignment: eac h p oin t has a unique lab el and along-fiber co ordinate. In irregular Kak eya sets, ho wev er, a single p oint ma y lie on m ultiple line segmen ts, so the decomp osition is not unique. An adv ersarial oracle can exploit this m ultiplicity , selecting the fib er lab el achieving maxim um compression. This 3 adaptiv e-fib ering phenomenon iden tifies a cen tral obstruction from the presen t information- theoretic p ersp ectiv e. 1.5 Con tributions and Organization The con tributions are threefold. First, w e formulate a conditional-compression principle for fib ered geometric ob jects in information-theoretic language. Second, w e pro ve an exact analytical result: under effectively bi-Lipsc hitz and iden tifiable fib ering, the chain rule yields K A ( x ↾ r ) = K A ( z ↾ r ) + K A,z ( u ↾ r ) + O (log r ), with tw o corollaries connecting p oint wise dimension to Hausdorff dimension. Third, we identify the adaptiv e-fib ering obstruction as the fundamental difficulty and propose a formal researc h program at the in terface of geometric me asure theory , algorithmic complexit y , and source co ding. Bey ond the Kakey a conjecture itself, this framework addresses a recurring practical question in geometry-a ware coding systems. P oint-cloud compression uses directional priors; video co ding exploits motion v ectors; neural compression of high-dimensional data and retriev al-augmen ted generation rely on laten t fib er labels. In all these settings the deco der holds partial geometric side information, and the key question is: how m uch do es that lab el reduce residual description length? When lab els arise from a rich but non-unique fibering, the naiv e Shannon or chain-rule b ound ov erstates compression gains precisely b ecause of the adaptive-fibering obstruction identified here. The Kakey a problem is therefore not merely a case study but the canonical stress test for this class of questions. The minimax form ulation K r ( E ) and the Blackw ell preorder on side-information schemes (Section 6 ) pro vide a language for principled rate allo cation whenever fib er lab els are ambiguous or degraded. The pap er is organized as follows. Section 2 develops the abstract framework. Section 3 presen ts the central prop osition, pro of, and corollaries. Section 4 applies the framew ork to Kakey a sets. Section 5 discusses the adaptive-fibering obstruction. Section 6 bridges the framew ork to source co ding, metric en trop y , and Blackw ell comparisons. Section 7 illustrates practical implications via work ed examples. Section 8 outlines a research program of op en problems. Section 9 p ositions the work relative to existing literature. Section 10 concludes. App endices provide notation, a tec hnical sidebar, and finite-precision sc hematics. 4 2 Problem Setup: Compression with Geometric Side Infor- mation 2.1 Abstract Fib ered Sets Let X ⊆ R n b e compact and ( Z , d Z ) a metric space for fib er lab els. A fib ering is a family { F z } z ∈ Z of nonempt y compact subsets of X suc h that X = S z ∈ Z F z . A p oin t x ∈ X is decomp osed as x = ψ ( z , u ), where z ∈ Z is the fib er lab el and u is a co ordinate on F z . W e assume eac h fib er carries a metric structure so that the along-fib er co ordinate can b e describ ed at finite precision. 2.2 Finite-Precision Descriptions and Conditional Complexity Fix precision r > 0. W e describ e each quantit y to within 2 − r in the Euclidean metric. W rite z ↾ r for a finite-precision enco ding of z at resolution r , and similarly for x ↾ r and u ↾ r . The c onditional c omplexity of x giv en the fib er lab el z at precision r is K ( x ↾ r | z ↾ r ) ≈ K ( u ↾ r | z ↾ r ) , reflecting that once the fib er is identified, the remaining information is the along-fib er co ordinate. Heuristically , one expects a decomposition of the form K ( x ↾ r ) ≈ K ( z ↾ r ) + K ( u ↾ r | z ↾ r ) + K ( a ( z ) ↾ r | z ↾ r ) + O (log r ) , where K ( a ( z ) ↾ r | z ↾ r ) is the ov erhead from sp ecifying the fib er’s lo cation. The exact conditions under whic h this decomp osition holds as an equality up to O (log r ) are the sub ject of Prop osition 1 . 2.3 Iden tifiability and Ambiguit y Gain A fib ering is identifiable at x if { z : x ∈ F z } is a singleton; it is identifiable on X if iden tifiable at ev ery p oin t. The ambiguity gain at x is Γ r ( x ) := K ( x ↾ r ) − inf z : x ∈ F z K ( x ↾ r | z ↾ r ) , measuring the adv antage an adaptiv e compressor gains b y p oint wise fiber selection. In iden- tifiable fib erings with stable parametrization, Γ r ( x ) is sublinear in r . In irregular regimes, Γ r ( x ) can b e linear in r , indicating substantial adaptive compression. 5 3 Exact Results in Regular Fib ering Regimes 3.1 The Cen tral Prop osition Prop osition 1 (Effective additiv e decomp osition) . L et Z ⊆ R n − 1 and U ⊆ R b e c omp act sets, enc o de d at pr e cision r by standar d pr efix-fr e e binary descriptions of dyadic appr oxima- tions. L et X ⊆ R n b e c omp act, and let ψ : Z × U → X b e given by ψ ( z , u ) = a ( z ) + u · ϕ ( z ) , wher e: (i) Effectiv ely bi-Lipschitz: ther e exist c omputable c onstants L 1 , L 2 > 0 , e ach sp e cifiable in O (log r ) bits at pr e cision r , with L 1 ∥ ( z 1 , u 1 ) − ( z 2 , u 2 ) ∥ ≤ ∥ ψ ( z 1 , u 1 ) − ψ ( z 2 , u 2 ) ∥ ≤ L 2 ∥ ( z 1 , u 1 ) − ( z 2 , u 2 ) ∥ for al l ( z 1 , u 1 ) , ( z 2 , u 2 ) ∈ Z × U ; (ii) Iden tifiable: ψ is inje ctive on Z × U , so e ach x ∈ X has a unique pr eimage ( z ( x ) , u ( x )) ; and (iii) Computable with effectiv e modulus: ψ is a c omputable function, me aning ther e is a T uring machine that on input ( z ↾ r, u ↾ r, r ) outputs ψ ( z , u ) ↾ r in finitely many steps. Equivalently, the maps a : Z → R n and ϕ : Z → S n − 1 ar e c omputable with a c omputable mo dulus of uniform c ontinuity at e ach pr e cision r . Then for every or acle A ⊆ N and every x = ψ ( z , u ) ∈ X , K A ( x ↾ r ) = K A ( z ↾ r ) + K A,z ( u ↾ r ) + O (log r ) , wher e K A denotes pr efix-fr e e Kolmo gor ov c omplexity r elative to or acle A , K A,z denotes c omplexity with or acle A augmente d by the string z ↾ r , and the O (log r ) c onstant dep ends on L 1 , L 2 but not on x . Pr o of. Enco ding mo del. Throughout, z ↾ r denotes the r -bit dyadic appro ximation to z (the nearest p oint of (2 − r Z ) n − 1 ), and similarly for u ↾ r ∈ 2 − r Z and x ↾ r ∈ (2 − r Z ) n . Kolmogoro v complexit y is prefix-free and defined relativ e to a fixed universal oracle mac hine U . Upp er b ound. Let p z b e an optimal A -program for z ↾ r (so | p z | = K A ( z ↾ r )) and p u an optimal ( A, z ↾ r )-program for u ↾ r (so | p u | = K A,z ( u ↾ r )). Define program P + : (a) Deco de the self-delimiting pair ⟨ p z , p u ⟩ . The prefix-length encoding of | p z | uses 2 ⌈ log 2 ( | p z | + 1) ⌉ ≤ O (log r ) bits, since | p z | ≤ ( n − 1) r + O (1). (b) Run p z with oracle A to recov er z ↾ r ; run p u with oracle ( A, z ↾ r ) to recov er u ↾ r . 6 (c) Ev aluate ψ ( z ↾ r, u ↾ r ). By the Lipschitz upp er b ound, ∥ ψ ( z ↾ r , u ↾ r ) − x ∥ ≤ L 2 √ n · 2 − r , so rounding to precision r + c + , where c + = ⌈ log 2 ( L 2 √ n ) ⌉ = O (1), yields a string within 2 − r of x . Output this string as x ↾ r . The total program length is | p z | + | p u | + O (log r ), giving K A ( x ↾ r ) ≤ K A ( z ↾ r ) + K A,z ( u ↾ r ) + O (log r ) . (1) Lo wer bound. W e construct an explicit inv ersion pro cedure P − that computes ( z ↾ r , u ↾ r ) from x ↾ r with O (log r ) description ov erhead. Set c − = ⌈ log 2 (4 L 2 √ n/L 1 ) ⌉ + 1 (a fixed in teger, O (1), determined by L 1 , L 2 ). Run P − at grid resolution s = r + c − : (a) Enco de the parameter r as an auxiliary input to P − at cost ⌈ log 2 r ⌉ = O (log r ) bits. The constants L 1 , L 2 , c − are hardco ded in P − . (b) En umerate the finite dyadic grid G s =  Z ∩ 2 − s Z n − 1  ×  U ∩ 2 − s Z  in lexicographic order. Compactness of Z × U b ounds |G s | ≤ C 0 · 2 ns for a computable constan t C 0 . (c) F or each ( z ′′ , u ′′ ) ∈ G s , compute ψ ( z ′′ , u ′′ ) and test whether ∥ ψ ( z ′′ , u ′′ ) − x ↾ r ∥ ≤ 2 · 2 − r . Halt at the first match and output ( z ′′ ↾ r , u ′′ ↾ r ). Corr e ctness. Let ( z † , u † ) ∈ G s b e the nearest grid point to ( z , u ). Then ∥ ( z † , u † ) − ( z , u ) ∥ ≤ √ n · 2 − s , so by the Lipsc hitz upp er b ound and the triangle inequality , ∥ ψ ( z † , u † ) − x ↾ r ∥ ≤ L 2 √ n · 2 − s + 2 − r ≤  L 2 √ n · 2 − c − + 1  · 2 − r ≤ 2 · 2 − r , where the last step uses c − ≥ log 2 ( L 2 √ n ), so the search halts. F or any matching p oin t ( z ′′ , u ′′ ), the bi-Lipschitz lo w er b ound giv es ∥ ( z ′′ , u ′′ ) − ( z , u ) ∥ ≤ ∥ ψ ( z ′′ , u ′′ ) − x ↾ r ∥ + ∥ x ↾ r − x ∥ L 1 ≤ 3 · 2 − r L 1 . (2) Exact r e c overy. The bound ( 2 ) controls how far the grid match ( z ′′ , u ′′ ) can b e from the true ( z , u ), but the ratio 3 /L 1 ma y exceed 1, so ( z ′′ ↾ r , u ′′ ↾ r ) nee d not equal ( z ↾ r , u ↾ r ) in general. W e handle this as follows. Using flo or truncation, w ↾ r := ⌊ w i · 2 r ⌋ · 2 − r comp onen t wise, the v alue z ↾ r lies in the ℓ ∞ -ball of radius ⌈ 3 /L 1 ⌉ · 2 − r cen tered at z ′′ ↾ r . The num b er of r -scale dy adic cells within this ball is at most (2 ⌈ 3 /L 1 ⌉ + 1) n , a constan t M depending only on L 1 and n . The program P − therefore outputs ( z ′′ ↾ r, u ′′ ↾ r ) together with a ⌈ log 2 M ⌉ -bit auxiliary p ointer app ended as hardwired advice, not derived from the first matc h itself. The p ointer indexes z ↾ r within the list of at most M candidate cells; a companion subroutine chec ks each candidate against the original x ↾ r via ψ (using h yp othesis (iii)) and iden tifies the unique match. F rom this, ( z ↾ r, u ↾ r ) is recov ered exactly . The p oin ter length is O (1) (dep ending only on L 1 , n , b oth hardco ded), so it is absorb ed 7 in to the O (log r ) budget. Overhe ad ac c ounting. P − is a fixed program with L 1 , L 2 , c − , M hardco ded, and the only run time input is r (at cost O (log r ) bits), so the in version contributes K A ( z ↾ r, u ↾ r ) ≤ K A ( x ↾ r ) + O (log r ) . (3) By the chain rule for prefix-free complexit y (with K A ( z ↾ r ) ≤ ( n − 1) r + O (1)), K A ( z ↾ r ) + K A,z ( u ↾ r ) ≤ K A ( z ↾ r, u ↾ r ) + O  log K A ( z ↾ r )  ≤ K A ( x ↾ r ) + O (log r ) . (4) Com bining ( 1 ) and ( 4 ) yields K A ( x ↾ r ) = K A ( z ↾ r ) + K A,z ( u ↾ r ) + O (log r ), where the constan t dep ends on L 1 , L 2 , n but not on x or r . 3.2 Corollary: Additivity of Algorithmic Dimension Corollary 1. Under the hyp otheses of Pr op osition 1 , for any or acle A and x = ψ ( z , u ) ∈ X , dim A ( x ) = lim inf r →∞ K A ( z ↾ r ) + K A,z ( u ↾ r ) r . (The pr op osition gives a p ointwise O (log r ) c omp arison; existenc e of the limit is not asserte d. If b oth lim r →∞ K A ( z ↾ r ) /r and lim r →∞ K A,z ( u ↾ r ) /r exist individual ly, the liminf is a limit and e quals d Z + d F .) In p articular, if z has algorithmic dimension d Z r elative to A and u has dimension d F r elative to A and z , then dim A ( x ) = d Z + d F . R emark 1 (Incompressibilit y and Martin-L¨ of randomness) . The condition dim( x ) = n is a liminf statemen t: lim inf r →∞ K ( x ↾ r ) /r = n . This is distinct from—and w eaker than—the condition K ( x ↾ r ) ≥ nr − O (1) for al l r , which is the Levin–Schnorr characterization of Martin-L¨ of randomness. Martin-L¨ of random points ha ve full ambien t algorithmic dimension (dim( x ) = n ), but the con verse need not hold: a p oint with dim( x ) = n need not be Martin- L¨ of random. In our regular-fibering regime, Proposition 1 gives a lo wer b ound K A ( x ↾ r ) ≥ ( n − 1) r + r − O (log r ) = nr − O (log r ) for suitable oracle A , whic h is the o ( r )-error version consisten t with the Martin-L¨ of b enchmark. F rom this p ersp ectiv e, the chain-rule argument is an argument that p oin ts in a regular-fib ered Kak eya set are algorithmic al ly typic al at the ambien t rate: directional richness preven ts the set from resembling a low er-dimensional subspace. The oracle in the p oint-to-set principle plays the role of an adversary attempting to mak e x app ear compressible; the regular-fib ering case shows this adv ersary fails. Whether the stronger Levin–Sc hnorr b ound holds (i.e., whether K ( x ↾ r ) ≥ nr − O (1) along a dense subsequence) is a separate question not addressed here. 8 3.3 Corollary: Application to Kak eya Sets Corollary 2 (Kakey a regular-fib ering case) . L et E ⊆ R n b e a Kakeya set with a r e gular identifiable fib ering ψ : S n − 1 × [0 , 1] → E given by ψ ( e, t ) = a ( e ) + te , wher e the ful l map ( e, t ) 7→ a ( e ) + te satisfies the thr e e hyp otheses of Pr op osition 1 (effe ctively bi-Lipschitz, identifiable, and c omputable with effe ctive mo dulus). Then for any or acle A and any x ∈ E such that e has dimension n − 1 r elative to A and t has dimension 1 r elative to A and e , dim A ( x ) = n. This aligns the r e gular-fib ering r e gime with the p oint-to-set me chanism underlying ful l- dimensional lower b ounds. 4 Kak ey a as the Canonical Direction-Ric h Example 4.1 Definition and History Definition 1. A Kakeya set in R n is a compact set E ⊆ R n con taining a unit line segm en t in every direction: for each e ∈ S n − 1 , there exists a ( e ) ∈ R n suc h that { a ( e ) + te : t ∈ [0 , 1] } ⊆ E . Definition 2 (Kakey a conjecture) . Every Kakey a set E ⊆ R n satisfies dim H ( E ) = n . Besico vitch’s 1928 construction [ 16 ] yields Kakey a sets of Leb esgue measure zero in R n for all n ≥ 2, a remark able geometric ac hievemen t showing that direction-rich sets need not b e metrically thic k. Bey ond this construction, Besico vitch developed the the- ory of linearly measurable sets and pro jection results that directly anticipate Marstrand’s pro jection theorem—a foundational result in geometric measure theory establishing that the pro jections of sets of Hausdorff dimension s > 1 on to almost every line hav e p ositive length. Y et the Kakey a conjecture asserts such sets cannot b e dimensional ly small. The conjecture is established in the plane ( n = 2) by Da vies (1971). F or n ≥ 3 it has long resisted resolution, with low er b ounds of dim H ( E ) ≥ (2 n + 2) / 3 + ε n due to W olff, Katz– T ao, and Hickman–Rogers–Zhang [ 17 ]. Dvir (2008) resolved the finite-field analogue via the p olynomial metho d. Bourgain and Demeter’s decoupling theorem [ 15 ] pro vided a key harmonic-analytic to ol exploited in subsequen t dimension bounds. R emark 2 (Res olution in R 3 ) . W ang and Zahl [ 18 ] pro ved the Kakey a conjecture in R 3 ; a streamlined account app ears in [ 19 ]. The information-theoretic obstruction persp ective de- v elop ed here remains central for n ≥ 4: in higher dimensions the adaptive-fibering obstruc- tion is unresolv ed, and the compression-theoretic diagnosis (no oracle can simultaneously 9 reduce complexit y b elo w the ambien t rate for all p oints) provides a language for what an y future pro of m ust rule out. 4.2 Geometric Represen tation and Informational Decomp osition F or x = a ( e ) + te in a Kak eya set, the three geometric comp onen ts contribute: 1. Direction e ∈ S n − 1 : K ( e ↾ r ) ≈ ( n − 1) r bits. 2. Along-segmen t coordinate t ∈ [0 , 1]: K ( t ↾ r | e ↾ r ) ≈ r bits. 3. Basep oin t a ( e ): K ( a ( e ) ↾ r | e ↾ r ) = O (log r ) in the Lipsc hitz-regular case, or O ( r ) when irregular. Under regularity conditions, the total complexity b ecomes K ( x ↾ r ) ≈ ( n − 1) r + r + O (log r ) = nr + O (log r ) , suggesting that p oints in a Kakey a set retain full am bient complexity . 4.3 Application of the Central Prop osition The Kakey a fib ering ψ ( e, t ) = a ( e ) + te fits Prop osition 1 whenever the ful l parametrization ( e, t ) 7→ a ( e ) + te satisfies all three h yp otheses: (i) effectively bi-Lipsc hitz on S n − 1 × [0 , 1] join tly in ( e, t ), (ii) injective (iden tifiable), and (iii) computable with effective mo dulus. Regularit y of e 7→ a ( e ) alone is not sufficient; the needed condition is on the joint map ψ , and in particular the Lipsc hitz b ound m ust hold globally across both the direction and the fib er coordinate. Under these conditions, Corollary 2 yields dim A ( x ) = n for ev ery oracle A . The passage to Hausdorff dimension then uses the point-to-set principle: for ev ery oracle A there exists a p oint x ∈ E (sp ecifically one whose direction e and along- fib er coordinate t are algorithmically random relative to A ) achieving dim A ( x ) = n , so dim H ( E ) = min A sup x ∈ E dim A ( x ) = n . Key observ ation. In Kakey a sets with irregular basep oint map or non-unique fib er assignmen t, these conditions fail. An adv ersarial oracle can exploit fib er-choice freedom to reduce effectiv e description length below the am bient rate. This is the gap the Kakey a conjecture highlights. 10 a ( e ) x = a ( e ) + te t e K ( e ↾ r ) ≈ ( n − 1) r K ( t ↾ r | e ↾ r ) ≈ r K ( a ( e ) ↾ r | e ↾ r ) = O (log r ) (r e gular r egime) Figure 1: Decomp osition of a p oint on a directional fib er. Direction e con tributes ( n − 1) r bits, the along-fiber co ordinate con tributes r bits, and the basep oint o v erhead is logarithmic in the regular regime. T otal complexit y matc hes the ambien t dimension. 5 The Obstruction: Am biguous Fib ering and Deco der Ad- v an tage 5.1 Non-Uniqueness of Decomp ositions In an irregular Kak eya set, a p oint may lie on m ultiple segmen ts. If x = a ( e 1 ) + t 1 e 1 = a ( e 2 ) + t 2 e 2 for distinct e 1 , e 2 , the complexit y split is non-unique. The decomp osition via e 1 yields budget K ( e 1 ↾ r ) + K ( t 1 ↾ r | e 1 ↾ r ), the decomp osition via e 2 yields K ( e 2 ↾ r ) + K ( t 2 ↾ r | e 2 ↾ r ); these can differ substantially . A compressor minimizing description length will choose the smallest total. In the p oint- to-set framew ork, the oracle enco des the en tire fib er structure and adaptiv ely selects the b est decomp osition for eac h p oin t at eac h scale. The compression principle requires that no suc h adaptive c hoice reduces complexity b elow the ambien t rate. V erifying this uniformly is the central difficult y . 5.2 Adaptiv e Side Information and Oracle Adv an tage The p oin t-to-set principle frames dimension as a game: an adversarial oracle c ho oses side information to minimize complexity . In a fib ered geometry , the oracle’s most p ow erful strategy is to select, for each x and precision r , the fib er lab el e maximizing compression gain. Sp ecifically , the oracle can enco de: 1. the en tire fib er structure (basep oints, directions, segmen t endp oin ts); 2. an adaptiv e selection function ϕ r ( x ) returning the direction minimizing K A ( x ↾ r | ϕ r ( x ) ↾ r ) at each scale; and 3. full information ab out fib er-assignmen t uniqueness at each p oint. This adaptive freedom is the core obstruction. In the regular-fib ering regime of Prop osi- tion 1 , the fiber assignment is essentially unique and the oracle’s adv an tage is neutralized. In the general case, proving the Kak eya conjecture w ould require sho wing that directional ric hness imp oses a complexity low er bound no adaptiv e sc hem e can circumv en t. 11 Example 1 (Planar crossing: explicit Γ r calculation) . Consider the simplest case in R 2 : tw o line segmen ts crossing at a single p oint. Let e 1 = (1 , 0) (horizon tal) with fib er F e 1 = { ( t, c ) : t ∈ [0 , 1] } at height c , and e 2 = (0 , 1) (v ertical) with fib er F e 2 = { ( c ′ , s ) : s ∈ [0 , 1] } at abscissa c ′ . Their in tersection is x ∗ = ( c ′ , c ). F or any x = ( x 1 , x 2 ) with K ( x ↾ r ) ≈ 2 r (b oth coordinates algorithmically indep enden t and random at precision r ): • Horizontal fib er: lab el z 1 = x 2 , residual u 1 = x 1 ; K ( x ↾ r | z 1 ↾ r ) ≈ K ( x 1 ↾ r | x 2 ↾ r ) ≈ r . • V ertic al fib er: lab el z 2 = x 1 , residual u 2 = x 2 ; K ( x ↾ r | z 2 ↾ r ) ≈ r . The adaptive oracle selects the more compressiv e lab el: Γ r ( x ) = K ( x ↾ r ) − min  K ( x ↾ r | z 1 ↾ r ) , K ( x ↾ r | z 2 ↾ r )  ≈ 2 r − r = r . This linear gain Γ r ( x ) = r + O (log r ) is sharp but not pathological: the chain rule forces K ( x ↾ r | z ↾ r ) ≥ K ( x ↾ r ) − K ( z ↾ r ) − O (log r ) ≈ r for any direction lab el z with K ( z ↾ r ) ≤ r , so no oracle can drive the residual b elow r − O (log r ) for a single algorithmically random p oin t. The obstruction in the full Kakey a setting is therefore glob al and or acle-uniform : the c hallenge (Problem 3 in Section 8 ) is to show that a single oracle A cannot simultaneously reduce K A ( x ↾ r | z A ( x ) ↾ r ) b elow r − Ω( r ) for al l x ∈ E at once, not merely for one isolated in tersection. 5.3 Disin tegration of Complexit y and the Borel–Kolmogoro v Phenomenon The core structure of Prop osition 1 is an algorithmic disinte gr ation : the total description of x splits into the description of its fib er lab el z = π ( x ) plus the residual description of x given the lab el. Algorithmically , this is the c hain rule. The contin uous analogue is the form ula for algorithmic dimension under a pro jection π : dim( x ) = dim( π ( x )) + dim( x | π ( x )) , (5) where dim( x | π ( x )) denotes the dimension of x relative to the oracle π ( x ). Equation ( 5 ) mirrors the classical chain rule for Kolmogorov complexit y K ( x, y ) = K ( y ) + K ( x | y ) + O (log K ( x, y )) and the disintegration theorem in measure theory , which expresses a join t measure as a base measure on lab els in tegrated against conditional measures on fib ers. The algorithmic and measure-theoretic statements are not formally equiv alen t but share the same logical structure: a complex ob ject is analyzed by first sp ecifying a coarse lab el and then describing the residual. 12 The non-uniqueness of this decomposition has a sharp parallel in classical probabil- it y . The Borel–Kolmogoro v paradox demonstrates that conditioning on a measure-zero ev ent can yield different conditional distributions dep ending on the sigma-algebra used to define the limiting pro cedure. The conditional distribution of a p oint on a great circle, for example, dep ends on whether one takes the limit through latitude strips, longitude strips, or some other family of s ets. There is no canonical answ er without specifying the disin tegration. The algorithmic setting mak es this iden tification precise. A n adaptive fib ering is a choic e of sigma-algebr a, which is a choic e of or acle. More concretely: different pro jection maps π (i.e., different fib er-lab el functions x 7→ z ( x )) yield differen t disin tegrations of dim( x ) via equation ( 5 ), and none is canonical without additional regularity . The oracle in the p oin t-to-set principle enco des exactly this c hoice. When one asks “what is K A ( x ↾ r )?”, the oracle A implicitly selects the conditioning structure: which fib er family , which basep oin t map, which disintegration. The stic ky Kak eya condition— | a ( e ) − a ( e ′ ) | ≲ | e − e ′ | —eliminates the Borel–Kolmogorov freedom by making the fib ering essentially unique. Under stic kiness, all reasonable disinte- grations agree to within O (log r ), and the algorithmic chain rule is stable. This is precisely wh y the stic ky case is tractable: stic kiness con verts a non-canonical conditioning problem in to a canonical one. The full Kakey a conjecture is, in these terms, the statement that no c hoice of sigma- algebra—no c hoice of oracle—can reduce total description length b elow the am bient rate. Unlik e the sticky case, the irregular case requires a uniform b ound acr oss al l p ossible fib erings and al l p ossible or acles simultane ously . This uniformity requirement is why the conjecture resists the techniques that resolv e the stic ky case. A pro of is lik ely to require new tools at the in tersection of algorithmic information theory , harmonic analysis, and geometric measure theory—connecting oracle-relativized cov ering argumen ts with the decoupling inequalities (cf. [ 15 ]) that ha ve driven recent harmonic-analytic progress. 6 Information-Theoretic In terfaces 6.1 Source Co ding with Side Information The compression principle translates naturally into source co ding. In Slepian–W olf and Wyner–Ziv frameworks [ 7 , 8 ], a source X is enco ded at rate R = H ( X | Z ) when side information Z is a v ailable at the deco der. Co v er [ 11 ] observed that side information is essen tially Ba y esian conditioning: providing Z reduces en tropy from H ( X ) to H ( X | Z ). The presen t setting is the geometric and algorithmic analogue: the fib er lab el z is the conditioning v ariable, and K ( x ↾ r | z ↾ r ) plays the role of H ( X | Z ) for individual 13 Regular fib ering x e ( x ) unique fiber lab el stable c onditioning Adaptiv e fibering x e 1 e 2 e 3 oracle pic ks b est split m ultiple fiber lab els adaptive side information Figure 2: Regular vs. adaptiv e fib ering. Left: each p oin t lies on a unique fib er, yielding a stable conditional-complexity decomp osition. Righ t: a p oint lies at the intersection of mul- tiple fib ers; an oracle selects the most compressive representation p oin twise. This adaptive freedom is the cen tral obstruction. descriptions. The compression principle prop oses that this conditional rate cannot drop b elo w the intrinsic fiber dimension when the fib er structure is regular. This connection is further explored in the authors’ related w ork [ 12 , 13 ]. The parallel is instructive but not exact: source co ding concerns i.i.d. sequences and Shannon en tropy , while the present setting inv olv es individual strings and Kolmogoro v complexit y . 6.2 Metric En tropy and Co vering Complexity A second in terface connects to metric en trop y in the sense of Kolmogoro v–Tikhomiro v. The ε -en tropy H ε ( E ) measures the logarithm of the minimum n umber of ε -balls needed to co ver E . F or sets of Hausdorff dimension d , one exp ects H ε ( E ) ≍ d log(1 /ε ). The compression principle is interpretable as a statemen t ab out conditional cov ering complexit y . With fib er-lab el side information, the cov ering problem reduces to co vering a single fiber F e , with residual entrop y H ε ( F e ) ≍ d F log(1 /ε ). The total co vering complexity is approximately H ε ( E ) ≈ H ε ( Z ) + sup z ∈ Z H ε ( F z ) , whic h with ( n − 1)-dimensional directions and one-dimensional fib ers yields H ε ( E ) ≍ n log(1 /ε ), consistent with dim H ( E ) = n . 6.3 Blac kwell Comparisons, Garbling, and Stitc hed Fib erings A third p ersp ective places admissible side-information sc hemes in a partial order analogous to the Blac kwell ordering of statistical experiments [ 10 ]. Say that Z 1 is mor e informative 14 than Z 2 for residual compression of x if K ( x ↾ r | Z 1 ↾ r ) ≤ K ( x ↾ r | Z 2 ↾ r ) + o ( r ) uniformly o v er x ∈ E and all oracles A . This preorder captures which side-information sc hemes yield asymptotically smaller residual descriptions, and connects the present frame- w ork to classical comparison-of-exp erimen ts theory . W e use “preorder” informally here; a rigorous definition would fix an admissible class of co ding models and make the uniformit y conditions explicit. Garbling. A garbling of Z is a computable map G : Σ ∗ → Σ ∗ applied to the lab el stream: the deco der receiv es G ( Z ↾ r ) in place of Z ↾ r . Computable maps cannot increase the infor- mation conten t of their input, a fact that translates directly in to a complexit y inequalit y . R emark 3 (Garbling weakly increases residual complexity) . Let Z be a regular identifiable fib er-lab el sc heme and G a computable garbling. In the regime of Prop osition 1 , K A ( x ↾ r | G ( Z ) ↾ r ) ≥ K A ( x ↾ r | Z ↾ r ) − O (log r ) , uniformly o v er x ∈ E and all oracles A . That is, garbling weakly increases residual complex- it y up to logarithmic terms, a data-pro cessing inequality for K . Consequen tly , Z dominates G ( Z ) in the informativeness preorder whenever G is not O (log r )-close to the identit y on the relev ant lab el strings. Stitc hed fib erings. Global fib er-lab el schemes are often assem bled from lo cal charts. A stitche d fib ering consists of a collection { ψ α } of lo cal regular fib erings on ov erlapping op en sets { U α } co vering E , with computable transition functions on ov erlaps U α ∩ U β . Within eac h c hart, Prop osition 1 applies and the complexit y split is clean. On ov erlaps, how ever, a p oin t x ∈ U α ∩ U β ma y receive fiber lab els z α and z β that are not canonically iden tified, generating the ambiguit y gain Γ r ( x ) = K ( x ↾ r ) − inf z ∈ Z ( x ) K ( x ↾ r | z ↾ r ) ≥ 0. When the transition functions require more than O (log r ) bits to sp ecify , the stitc hed sc heme fails to dominate any single-chart sc heme, and Blackw ell comparisons b et w een ov erlapping c harts b ecome inc omp ar able rather than ordered. This incomparability is a signature of the adaptiv e-fib ering obstruction: no canonical global labeling exists, and the compression gains predicted by the single-c hart analysis cannot b e realized uniformly . 15 Geometry Fib ered set E family { F z } Enco der P oin t x ∈ E precision r Deco der / Oracle Side info z ↾ r adaptiv e fib ering Outcome Residual complexit y dim A ( x ) ? ≥ n admissible fibers enco ded description conditional compression p oin t-to-set dim H ( E ) = n ? Figure 3: Minimax framework for geometric compression. Geometry determines admissible fib er decomp ositions; an enco der selects a p oin t at finite precision; an oracle-lik e deco der c ho oses fav orable side information; the p oint-to-set principle con v erts surviving p oint wise complexit y into a Hausdorff dimension lo wer b ound. 6.4 Robust Conditional Compression and Minimax F orm ulation The compression question is fundamen tally adv ersarial. Define the r obust c onditional c om- plexity K r ( E ) := min fiberings F sup x ∈ E K ( x ↾ r | z F ( x ) ↾ r ) . A complete definition requires fixing an admissible class of fib erings, for example computable regular fib erings or stitc hed computable atlases; the minimax v alue dep ends on this choice. The compression principle proposes K r ( E ) ≥ d F · r − o ( r ) when the fiber family is sufficien tly ric h. This minimax formulation unifies differen t p ersp ectives: the oracle-based argumen t lo wer-bounds K r ( E ) via the p oint-to-set principle, the source-co ding argumen t interprets it as w orst-case conditional rate, and the metric-en tropy argumen t relates it to conditional co vering num b ers. 7 Practical Implications Structured side information is ubiquitous in compression, signal pro cessing, and mac hine learning: enco ders exploit geometric embeddings, manifold co ordinates, or learned latent factors, identifying a fib er structure and enco ding a p oint b y lab el plus residual. If the fib er family is genuinely ric h and fib ers retain indep enden t v ariability , side information ab out the fib er lab el cannot reduce residual description length b elow the in trinsic fib er dimension. The distinction betw een genuine conditional-rate reduction and ambiguit y-enabled artifacts 16 mirrors the distinction b et ween regular and adaptive fib ering. Garbled or coarsened fib er lab els (arising, for example, from quantized latent co des, noisy cluster assignments, or lossy index compression) act as computable degradations of the true side-information sc heme. By the data-pro cessing inequality for K (Remark 3 ), suc h garbling can only increase residual complexity , never reduce it: a compressed or noisy side-information signal provides w eakly less b enefit than the original. Similarly , laten t represen tations assem bled b y stitc hing lo cal charts (as in atlas-based or mixture-of-exp erts generativ e mo dels) ma y exhibit ambiguit y at c hart seams. The stitc hing obstruction of Section 6 predicts that ov erlap-induced m ultiplicity (and any transition functions requiring sup er-logarithmic specification) degrades the effectiv e side information and forfeits the full compression gains av ailable from a single globally identifiable fib ering. Rate allo cation in geometry-a w are comm unication systems must answ er: how many bits go to fib er index, how many to along-fib er co ordinate, and when is side information gen uinely useful? The framework, and in particular the Blackw ell preorder of Section 6 , pro vides a language for these questions. Example 2 (Blo c kc hain finalit y as structured side information) . A transaction or local ledger state x in a distributed blo ck chain system is in terpreted relative to structured con textual lab els—c hain identit y , blo ck depth, chec kp oin t status, or finality ep o c h, which serve as side information z . In a stable finalized regime, these lab els function like an identifiable fib er map: once the relev ant chain history is fixed, the residual uncertain t y ab out the even t’s status is sharply reduced, and K ( x ↾ r | z ↾ r ) ≪ K ( x ↾ r ). Under forks, reorganization risk, or delay ed finalit y , how ev er, the same transaction may remain compatible with several admissible histories, so the label z is no longer globally unique. The admissible lab el set Z ( x ) expands to co ver comp eting chain branc hes, and an y lab eling sc heme built by stitching lo cally coheren t no de views creates seam ambiguit y at the o verlap. A depth-only lab el is a garbling of ric her finality information: b y Remark 3 , it can only increase residual complexit y . In a fork ed regime the comp eting histories are Blac kwell-incomparable (Section 6 ): no single lab el dominates the others in the informa- tiv eness preorder, and apparent compression gains ma y reflect the c hoice of history lab el rather than genuine reduction in uncertaint y . In the pap er’s notation, the ledger ev ent pla ys the role of x , the history or finality lab el pla ys the role of z , and fork ambiguit y enlarges Z ( x ) from a singleton to a multi- elemen t admissible set. The example illustrates a general lesson: side information reduces residual description length only when the underlying represen tation is stable and globally iden tifiable: the pap er’s distinction b et w een regular and adaptiv e fib ering applies whenever con text lab els ma y fail to b e globally unique. Example 3 (Retriev al context as structured side information) . In retriev al-augmented sys- 17 tems, an output x (a generated answer, decision, or ranked result) is interpreted relative to auxiliary context such as retriev ed passages, memory entries, or prompt scaffolds. These ob jects act as structured side information z . A richer retriev al con text is naturally more informativ e than a coarsened summary or truncated prompt, suggesting a Blackw ell-style comparison among con text-lab el schemes (Section 6 ); summaries or k eyword extracts ma y b e view ed as garblings of fuller contextual evidence. The complication is that practical systems assemble con text from m ultiple partially o verlapping sources (retriev ed passages, loc al memories, system prompts, in termediate sum- maries) that m ust be stitc hed in to a single conditioning structure. When those pieces do not define a unique global con textual represen tation, the same output x may remain compatible with sev eral admissible context lab els, enlarging Z ( x ). In such a regime, apparent reduc- tions in res idual description length reflect adaptive context selection rather than gen uine informational resolution. In the notation of the present pap er, the output pla ys the role of x , the retriev ed con text plays the role of z , and the admissible context family corresponds to Z ( x ). More generally , the v alue of contextual side information dep ends not only on how muc h con tent it carries but on whether the conditioning structure it defines is globally identifiable: the pap er’s distinction b etw een regular and adaptiv e fibering applies naturally to an y system whose outputs dep end on dynamically assembled con text. 8 A F ormal Researc h Program W e formulate six problems constituting a researc h program for the compression-theoretic study of directional geometries. Problem 1 (Geometric conditional rate function). Define a rigorous geometric conditional rate function for fib ered subsets of R n , interpolating b et ween p oint wise algorithmic com- plexit y and global metric en trop y . Problem 2 (Stabilit y of the conditional split). Iden tify sufficient conditions on { F z } and ψ under which K ( x ↾ r ) = K ( z ↾ r ) + K ( u ↾ r | z ↾ r ) + o ( r ) holds uniformly o ver x ∈ E and all oracles A . Problem 3 (Oracle-uniform lo w er b ounds). F or a fib ered set E , show that for every oracle A , sup x ∈ E dim A ( x ) ≥ d Z + d F , without regularit y assumptions on the fib er assignment. This is the core technical c hallenge underlying the full compression principle. Existing techniques (effectiv e Hausdorff measure, energy metho ds, F rostman-t yp e arguments) succeed when the fib ering is identifiable because 18 iden tifiability lets one separate the fib er-lab el bits from the along-fib er bits via the c hain rule uniformly . Without identifiabilit y the chain rule decouples only point wise, and no curren t metho d forces the oracle to “pa y twice” for the fib er-lab el o v erhead across all p oints sim ultaneously . A promising approac h is to track the effe ctive Hausdorff conten t of the fib er- lab el family as a function of precision, translating cov ering-num b er low er b ounds directly in to oracle-relativized complexit y b ounds. Problem 4 (Algorithmic dimension and metric en tropy). Establish quantitativ e relation- ships b etw een p oint wise algorithmic dimension and global metric en tropy for geometrically structured families. Problem 5 (Minimax side-information game). F ormalize the minimax game b etw een en- co der selecting x ∈ E and decoder selecting a fib er assignment. Characterize the game v alue for natural classes of fib ered sets. Problem 6 (Instantiation in specific geometries). In v estigate whether sp ecific geometric constructions instantiate the compression principle, including Kakey a configurations with con trolled basep oint maps and extractor-type constructions in additive com binatorics. 9 P ositioning Relativ e to Existing Literature This pap er sits at the intersection of three in tellectual traditions. Kakeya the ory and ge ometric me asur e the ory. The Kak eya conjecture has b een studied via harmonic analysis, com binatorial geometry , and additive com binatorics. Ma jor con tri- butions include dimension b ounds of W olff, Katz–T ao, and Hickman–Rogers–Zhang [ 17 ], and decoupling inequalities of Bourgain–Demeter [ 15 ]. Dvir’s finite-field resolution via the p olynomial metho d is striking. W ang and Zahl [ 18 ] recently prov ed the conjecture in R 3 , with a streamlined accoun t in [ 19 ]. The present pap er do es not contribute new geome tric b ounds; it prop oses an information-theoretic framework for the dimensional question in arbitrary dimension, where the conjecture remains open. L e onid L evin and the c omplexity-me asur e c onne ction. The bridge b et ween algorithmic randomness and complexity w as substan tially dev elop ed b y Leonid Levin, a student of Kol- mogoro v at Moscow State Universit y and later at Boston Univ ersit y . Levin’s w ork on uni- v ersal search, resource-b ounded complexit y , and the connections b etw een randomness tests and incompressibilit y help ed establish the mo dern framew ork in which measure-theoretic and algorithmic notions of “t ypicality” coincide. Lutz’s constructiv e Hausdorff dimension and the p oin t-to-set principle are direct descendants of this program: they translate the lo cal complexity rate of a single p oint in to global dimensional information ab out the set con taining it. 19 A lgorithmic dimension and the p oint-to-set principle. The framework of algorithmic dimension (Lutz, Ma yordomo, and collab orators) and the p oin t- to-set principle (J. Lutz, N. Lutz) provide our mathematical language. Geometric applications by Lutz and Stull demonstrate the p ow er of this approach. What is classical is the framework; what is pro- p osed is the isolation of adaptive-fibering as a general structural phenomenon. Information the ory: entr opy, c o ding, and extr action. Connections to source co ding (Slepian–W olf, Wyner–Ziv), metric entrop y (Kolmogorov–Tikhomiro v), and extractors are suggested but not formalized. F ormalizing these connections is part of the research program. More broadly , the pap er touches t wo distinct traditions b eyond the geometric core: algorithmic randomness, where Martin-L¨ of typicalit y supplies a natural incompressibility b enc hmark, and the foundations of conditioning, where the Borel–Kolmogorov phenomenon serv es as an in terpretive reminder that low er-dimensional conditional structure depends on the chosen mo de of disin tegration. In summary , this is a brief analytical and conceptual paper. It tak es an existing math- ematical framew ork, applies it to a classical geometric problem, and develops a general compression principle with in terfaces to information theory . 10 Conclusion The central idea is straightforw ard: geometric ob jects supp orting a rich family of direc- tional fib ers should b e informationally incompressible. The Kak ey a conjecture is the most prominen t instance. W e hav e form ulated this as a conditional-compression principle in the language of algo- rithmic dimension. The decomp osition K ( x ↾ r ) ≈ K ( e ↾ r ) + K ( t ↾ r | e ↾ r ) captures that total p oint complexit y is approximately the sum of directional and along-fib er complexities. Under regularity conditions (Lipschitz basep oin t maps, identifiable parametrizations), the decomp osition is stable and clarifies the mec hanism of kno wn lo wer-bound argumen ts. The real difficult y is the adaptive-fibering obstruction: when the fiber assignmen t is irregular, a point admits multiple decomp ositions, and an adv ersarial deco der exploits am- biguit y to reduce description length. W e iden tify this as the information-theoretic core of the Kakey a problem. The con tribution is a compression principle and researc h program, not a resolution of the Kakey a conjecture. W e outline concrete problems at the interface of geometric measure theory , algorithmic information theory , and source co ding. Viewed this w ay , the challenge is not merely to bound geometric dimension, but to understand whic h forms of structured side information truly reduce description length, and which do not. 20 A Notation and T erminology T able 1: Principal Notation Sym b ol Meaning R n Euclidean n -space S n − 1 Unit sphere in R n e ∈ S n − 1 Direction (fib er lab el, Kakey a set- ting) a ( e ) Basep oin t of segmen t in direction e t ∈ [0 , 1] Along-fib er coordinate x = a ( e ) + te P oint on segment in direction e z ∈ Z Fib er label, abstract setting u Along-fib er coordinate, abstract F z Fib er indexed b y z K ( σ ) Prefix-free Kolmogoro v complexity K A ( σ ) Complexit y relativ e to oracle A x ↾ r Precision- r approximation of x dim( x ) Algorithmic dimension: lim inf r →∞ K ( x ↾ r ) /r dim A ( x ) Oracle-relativized algorithmic di- mension dim H ( E ) Hausdorff dimension of E H ε ( E ) Kolmogoro v–Tikhomirov ε -entrop y K r ( E ) Robust conditional complexity at precision r Γ r ( x ) Am biguity gain: K ( x ↾ r ) − inf z K ( x ↾ r | z ↾ r ) B T ec hnical Sidebar: Heuristic Complexit y Decomp osition W e presen t a heuristic calculation illustrating the expe cted complexit y decomposition under fa vorable conditions. This is motiv ational, not a theorem. Setup. Let E ⊆ R n b e a Kak ey a set with basep oin t map a : S n − 1 → R n . Fix x = a ( e ) + te at precision r . Chain rule. By the c hain rule for Kolmogorov complexity , K ( x ↾ r ) = K ( e ↾ r ) + K ( a ( e ) ↾ r | e ↾ r ) + K ( t ↾ r | e ↾ r , a ( e ) ↾ r ) + O (log r ) . Directional con tribution. The direction e on S n − 1 satisfies K A ( e ↾ r ) ≥ ( n − 1) r − o ( r ) for algorithmically random e . 21 Basep oin t ov erhead. If a ( · ) is Lipsc hitz with constant L , then K ( a ( e ) ↾ r | e ↾ r ) = O (log r ). Without Lipschitz regularit y , this term can b e O ( r ), the source of the adaptiv e- fib ering difficult y . Along-fib er contribution. F or t algorithmically random in [0 , 1] relativ e to ( e, A ): K A,e ( t ↾ r ) ≥ r − o ( r ). Heuristic conclusion. Under the Lipsc hitz assumption, K A ( x ↾ r ) ≥ ( n − 1) r + r − o ( r ) = nr − o ( r ) , yielding dim A ( x ) ≥ n . The gap to a full theorem lies in con trolling basep oin t o verhead uniformly without Lipschitz regularit y . C Illustrativ e Finite-Precision Sc hematic T o visualize the distinction b etw een regular and adaptive fib ering, we compare schematic co de-length functions. Define toy co de lengths for a p oin t in R n at precision r : L reg ( r ) = nr + c log (1 + r ) , (6) L adapt ( r ) = nr − γ ( r ) , (7) where c > 0 captures basep oin t o v erhead and γ ( r ) ≥ 0 represents ambiguit y-enabled com- pression gain. In the regular regime, co de length tracks nr up to logarithmic correction. Under adaptive side information, γ ( r ) may reduce code length. The directional compression principle prop oses that for geometries with gen uine di- rectional ric hness and identifiable fib er structure, the adaptiv e gain satisfies γ ( r ) = o ( r ). The full Kakey a conjecture is that this holds ev en without identifiabilit y , explaining why it remains op en. 22 5 10 15 20 25 30 35 40 45 50 0 50 100 150 Precision r Sc hematic co de length Am bien t b enc hmark nr Regular: nr + c log (1+ r ) Adaptiv e (mild): nr − √ r Adaptiv e (substan tial): nr − 0 . 2 r Figure 4: Sc hematic co de length under regular and adaptiv e fib erings ( n = 3). In the regular regime, co de length trac ks nr up to logarithmic ov erhead. Under adaptive side in- formation, am biguity-enabled compression gains reduce description length. The directional compression principle prop oses that for direction-ric h geometries with identifiable structure, gains must b e sublinear. 23 References [1] J. H. Lutz and N. Lutz, “Algorithmic information, plane Kak eya sets, and conditional dimension,” ACM T r ans. Comput. The ory , v ol. 10, no. 2, pp. 7:1–7:22, 2018. [2] J. H. Lutz and D. M. Stull, “Bounding the dimension of p oints on a line,” Inform. Comput. , vol. 275, p. 104601, 2020. [3] R. O. Da vies, “Some remarks on the Kak eya problem,” Pr o c. Cambridge Philos. So c. , v ol. 69, pp. 417–421, 1971. [4] Z. Dvir, “On the size of Kak ey a sets in finite fields,” J. A mer. Math. So c. , v ol. 22, no. 4, pp. 1093–1097, 2009. [5] T. W olff, “An impro v ed b ound for Kakey a t yp e maximal functions,” R ev. Mat. Ib er o am. , vol. 11, no. 3, pp. 651–674, 1995. [6] N. H. Katz and T. T ao, “New b ounds for Kak ey a problems,” J. A nal. Math. , v ol. 87, pp. 231–263, 2002. [7] D. Slepian and J. K. W olf, “Noiseless coding of correlated information sources,” IEEE T r ans. Inf. The ory , v ol. 19, no. 4, pp. 471–480, Jul. 1973. [8] A. D. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the deco der,” IEEE T r ans. Inf. The ory , v ol. 22, no. 1, pp. 1–10, Jan. 1976. [9] J. H. Lutz, “The dimensions of individual strings and sequences,” Inform. Comput. , v ol. 187, no. 1, pp. 49–79, 2003. [10] D. Blackw ell, “Equiv alent comparisons of exp eriments,” Ann. Math. Stat. , v ol. 24, no. 2, pp. 265–272, 1953. [11] T. M. Cov er and J. A. Thomas, Elements of Information The ory , 2nd ed. Hob oken, NJ, USA: Wiley , 2006. [12] N. G. Polson and D. Zantedesc hi, “De Finetti + Sanov = Bay es,” Sep. 2025. [13] N. G. P olson and D. Zantedesc hi, “Ba yesian prediction under momen t conditioning,” arXiv:2510.20742, Oct. 2025. [14] A. N. Kolmogoro v and V. M. Tikhomiro v, “ ε -entrop y and ε -capacit y of sets in function spaces,” Amer. Math. So c. T r ansl. , v ol. 17, pp. 277–364, 1961. 24 [15] J. Bourgain and C. Demeter, “The proof of the ℓ 2 decoupling conjecture,” A nn. of Math. , vol. 182, no. 1, pp. 351–389, 2015. [16] A. S. Besico vitch, “On Kakey a’s problem and a similar one,” Math. Z. , vol. 27, no. 1, pp. 312–320, 1928. [17] J. Hickman, K. M. Rogers, and R. Zhang, “Improv ed b ounds for the Kakey a maximal conjecture in higher dimensions,” A mer. J. Math. , v ol. 144, no. 6, pp. 1511–1560, 2022. [18] H. W ang and J. Zahl, “V olume estimates for unions of conv ex se ts, and the Kakey a set conjecture in three dimensions,” arXiv:2502.17655, F eb. 2025. [19] L. Guth, “Outline of the W ang–Zahl pro of of the Kak eya conjecture in R 3 ,” arXiv:2508.05475, Aug. 2025. 25