Lost-K Phenomenon: Structured Loss in Complex Systems

• Lost-K Phenomenon is a collection of domain-specific constructions where a missing element triggers a structured effect rather than generic degradation.
• It appears in contexts ranging from boarding models to BosonSampling, signal processing, and long-context AI, revealing invariant symmetries despite omitted components.
• The phenomenon underscores loss-tolerant hardness, nonmonotonic responses, and positional biases that inform both theoretical insights and practical applications.

The expression Lost-K Phenomenon does not denote a single standardized object across the literature surveyed here. Instead, it recurs as a family of domain-specific constructions in which a missing element, a suppressed structural component, or a disadvantaged position produces a sharply structured effect rather than a generic degradation. In combinatorial probability, the first $k$ passengers with lost boarding passes induce a boarding process with explicit blocking probabilities and independent occupancy events; in BosonSampling, exactly $k$ lost photons replace a single permanent-squared amplitude by an average over many permanent-squared terms while preserving classical hardness for constant $k$; in signal processing and scattering, “lost-K” refers either to disappearance of the K valley in non-standard BCG or to effective disappearance of a K-matrix pole; and in long-context AI, related lost-* phenomena identify systematic failures of retrieval or grounding after reasoning, in the middle of context, later in context, or at the end of retrieved evidence (Grimmett et al., 2019, Aaronson et al., 2015, Jiao et al., 11 Aug 2025, Svarc, 2012, Whitecross et al., 10 Apr 2026, Tao et al., 7 Jul 2025, Liu et al., 15 Jun 2026).

1. Terminological scope and recurrent structure

Across these works, the label is best understood as a family resemblance rather than a single theorem. The lost object may be a boarding pass, a photon, a waveform feature, a pole, a segment of context, a hidden test subset, or a tracked object. What makes the family coherent is that the “loss” typically does not erase structure; instead, it exposes a latent symmetry, a nonmonotonic mode reorganization, a positional bias, or a recovery mechanism.

Area	“Lost” object or position	Defining effect
Boarding model	first $k$ boarding passes	$\mathbb P(D_m)=\frac{k}{n-m+k+1}$
BosonSampling	exactly $k$ input photons	output probabilities average many permanent-squared terms
Non-standard BCG	K valley after narrow-band filtering	I-J-K morphology becomes incomplete
Hadron scattering	K-matrix pole	T-matrix pole can remain while K-pole is pushed far away
Long-context models	later, middle, or end evidence	retrieval or contextual grounding degrades

This comparison suggests that the central analytic question is not merely what is lost, but what remains invariant. In the boarding-pass model the invariant is symmetry over still-possible occupied seats; in lossy BosonSampling it is the recoverability of the permanent through interpolation; in non-Hermitian optics it is the mode structure of a coupled system driven through a critical point; and in long-context models it is the persistent dependence of output quality on evidence position (Grimmett et al., 2019, Aaronson et al., 2015, Peng et al., 2014, Duerinckx et al., 9 May 2026).

2. The probabilistic archetype: lost boarding passes and the Lost-$k$ formula

In the most classical use of the term, a plane has $n$ seats and $n$ passengers boarding in order, passenger $i$ being assigned seat $k$0. The first $k$1 passengers have lost their boarding passes and each chooses uniformly at random among the seats currently available. Every later passenger $k$2 sits in seat $k$3 if it is free and otherwise chooses uniformly among the remaining empty seats. If $k$4 denotes the event that passenger $k$5 finds seat $k$6 occupied, the main formula is

$k$7

For $k$8, this reduces to

$k$9

and in particular $k$0, the classical lost boarding pass result (Grimmett et al., 2019).

The mechanism is a symmetry argument. When passenger $k$1 boards, the seats $k$2 are already occupied, because if any of them had been free the corresponding passenger would have taken the assigned seat earlier. In addition, exactly $k$3 more seats are occupied, but their labels are undetermined by the prior history. These “mysterious” occupied seats form a $k$4-subset of

$k$5

which has size $k$6. By symmetry, every such $k$7-subset is equally likely, so seat $k$8 is occupied with probability $k$9. The last passenger therefore sits correctly with probability

$k$0

The familiar $k$1 is the special case $k$2 (Grimmett et al., 2019).

A striking strengthening is that the occupancy events are independent. For $k$3, the events $k$4 are independent; the same argument yields independence of $k$5 for general $k$6. Thus the occupancy indicators form an independent family with explicit marginals $k$7. The paper also gives a backwards random-color representation: seats are independently colored red with probabilities $k$8, and displaced passengers follow the chain of red seats. For $k$9, this links directly to the independence of record events $\mathbb P(D_m)=\frac{k}{n-m+k+1}$0 for i.i.d. $\mathbb P(D_m)=\frac{k}{n-m+k+1}$1, where $\mathbb P(D_m)=\frac{k}{n-m+k+1}$2. In the large-$\mathbb P(D_m)=\frac{k}{n-m+k+1}$3 limit, the normalized displacements, after decreasing reordering, converge to the Poisson–Dirichlet distribution, and for $\mathbb P(D_m)=\frac{k}{n-m+k+1}$4 the rescaled diagram converges to $\mathbb P(D_m)=\frac{k}{n-m+k+1}$5 independent copies of the $\mathbb P(D_m)=\frac{k}{n-m+k+1}$6 limit (Grimmett et al., 2019).

3. Loss-tolerant hardness: the BosonSampling version of Lost-$\mathbb P(D_m)=\frac{k}{n-m+k+1}$7

In BosonSampling, the ideal no-collision output probability is

$\mathbb P(D_m)=\frac{k}{n-m+k+1}$8

up to the dropped factorial normalization in the no-collision setting. The lossy model studied in the relevant work assumes that exactly $\mathbb P(D_m)=\frac{k}{n-m+k+1}$9 of the $k$0 input photons are lost at the sources, with the lost subset unknown. For a detected no-collision output $k$1 containing only $k$2 photons, the probability becomes an average over all surviving $k$3-photon subsets. The paper abstracts this by defining

$k$4

where the sum is over all $k$5 proper submatrices $k$6 of an $k$7 Gaussian matrix $k$8. The lossy distribution is thus governed not by a single $k$9, but by an averaged quantity the paper calls the “temperament” $k$0 (Aaronson et al., 2015).

The main complexity-theoretic claim is that constant loss does not destroy hardness. The paper states that if $k$1 out of $k$2 photons are lost, then one cannot sample classically from a distribution that is $k$3-close, in total variation distance, to the ideal distribution, unless a $k$4 machine can estimate the permanents of Gaussian matrices in $k$5 time. In particular, if $k$6 is constant, simulating lossy BosonSampling is hard for a classical computer under exactly the same complexity assumption used for the original lossless case (Aaronson et al., 2015).

The reduction proceeds by embedding an $k$7 Gaussian matrix $k$8 into a larger $k$9 matrix $n$0 whose last $n$1 columns are multiplied by a scalar $n$2. Then $n$3 becomes a degree-$n$4 polynomial in $n$5, whose constant term is exactly $n$6. By evaluating this polynomial at $n$7 values of $n$8 very close to $n$9 and using polynomial interpolation or least squares, one can recover the permanent-squared term from the lossy average. The loss therefore changes the representation of the hard object, not the identity of the hard object itself. A plausible implication is that this version of Lost-$n$0 is a robustness statement: averaging over lost subsets obscures the permanent structure, but does not eliminate it when $n$1 (Aaronson et al., 2015).

4. Loss-induced nonmonotonicity in non-Hermitian and quantum optics

A different use of the same motif appears in coupled optical systems, where increasing loss can first suppress and then revive an optical or quantum effect. In a passive system of two directly coupled silica whispering-gallery-mode microtoroid resonators, one resonator is subjected to controllable extra loss by a chromium-coated nanofiber tip. The coupled-mode eigenfrequencies are

$n$2

and the exceptional point occurs when

$n$3

As the extra loss increases, the two resonances first approach each other, coalesce at the exceptional point, and then enter a weak-coupling regime with identical resonance frequency but different linewidths. The total intracavity intensity decreases at first, reaches a minimum near the exceptional point, and then increases again for larger $n$4. In the Raman-lasing realization, an existing Raman laser is extinguished by added loss and then revives when still more loss is introduced, because one supermode becomes increasingly localized in the less lossy resonator (Peng et al., 2014).

The fully quantum analogue is developed for a nonlinear optical-molecule system consisting of a Kerr nonlinear whispering-gallery-mode resonator $n$5 coupled to a linear resonator $n$6. In the one-photon subspace, the non-Hermitian eigenvalues are

$n$7

and the Hamiltonian exceptional point is reached when

$n$8

Below critical loss values, increasing $n$9 suppresses both intensity and nonclassicality; above the upper quantum critical point, the equal-time correlation $i$0 decreases again below $i$1, and at the exceptional point the single-photon blockade is recovered. The revived regime can also support two-photon blockade, characterized by $i$2 together with $i$3 (Zuo et al., 2022).

This interpretation is not uncontested. A separate analysis argues that loss-induced lasing is “basically the lasing mechanism that occurs in loss-coupled distributed feedback lasers,” dates back to the 1970s, and “does not rely on the physics of exceptional points.” In that account, the key principle is that threshold depends on the effective loss of the relevant mode rather than the average loss of the structure. In a waveguide with uniform gain $i$4 and periodic loss grating $i$5, the loss terms also provide distributed feedback, so increasing loss can initially lower threshold. Exceptional points are therefore not necessary for loss-induced lasing, even if EP-like pole dynamics can arise in some Fabry–Perot–assisted variants (Longhi et al., 2015).

5. Disappearing K-features: waveform morphology and scattering poles

In non-standard ballistocardiography, the Lost-K phenomenon is morphological and signal-processing specific. The standard I-J-K complex is degraded by narrow-band J-band filtering. The paper notes that the usual J-band filter keeps only about $i$6 Hz or, in its notation, $i$7 Hz for bcj; while this can suppress slurs and notches, it can also discard higher-frequency components needed to form the K valley. Lost-K is defined as the case in which the K valley becomes weak, blurred, or disappears after such filtering, making the standard I-J-K morphology incomplete. The stated cause is signal energy being discarded in the $i$8 Hz range (Jiao et al., 11 Aug 2025).

The same work proposes three transforms. The curvature-based transform

$i$9

and the inverted second-derivative transform

$k$00

are intended to recover or enhance lost K valleys, whereas the coarse-signal-based rising transform $k$01 is aimed mainly at reducing the shorter-J phenomenon. On a time-aligned ECG-BCG dataset with 40 subjects, the paper reports that bcc and bcd substantially improve J-K delay by recovering visible K valleys, with several records improving from more than $k$02 ms to less than $k$03 ms, and that bcd generally performs better than bcc for recovering lost K (Jiao et al., 11 Aug 2025).

In hadron scattering, by contrast, Lost-K refers to a K-matrix pole rather than a waveform valley. The central relation is

$k$04

so a K-matrix pole does not in general coincide with a T-matrix pole. The paper calls it the lost-K phenomenon when a visible T-matrix resonance pole exists but the corresponding K-matrix pole is shifted far away into an unphysical region, sometimes effectively to infinity. A simple background-phase transformation,

$k$05

implies a K-pole at

$k$06

so for $k$07 the pole is at $k$08. The physical resonance remains encoded in the T-matrix even when the K-pole is no longer phenomenologically nearby. This leads to the broader distinction among bare poles, K-matrix poles, and T-matrix poles, with T-matrix poles being the most physical and least model dependent (Svarc, 2012).

These two usages share the letter $k$09 but not the ontology. One concerns loss of a morphological valley under filtering; the other concerns displacement of a pole under unitarization and background dressing. The shared lesson is structural rather than semantic: a diagnostic feature associated with $k$10 may disappear from the most immediate representation while remaining indirectly recoverable.

6. Positional loss in long-context and retrieval-augmented AI

Long-context language modeling has generated a cluster of closely related lost-* phenomena. In “lost-in-thought,” the defining failure mode is that more reasoning hurts later in-context retrieval. A synthetic key-value benchmark separates direct retrieval from reasoning-retrieval, where the model must solve a math problem to determine the key before retrieving the value. Across five open-source $k$11B models and context lengths from $k$12K to $k$13K tokens, direct retrieval accuracy is much higher than retrieval after reasoning, while math accuracy remains high. A follow-up injection experiment shows that even after giving the model the correct key and exact lexical prefix of the answer, it still often hallucinates the value. RecaLLM addresses this by interleaving reasoning with explicit in-context retrieval, using recall spans and constrained decoding so that recalled text is a contiguous substring of the searchable context. On RULER, the reported averages are $k$14 for RecaLLM-Qwen2.5-7B and $k$15 for RecaLLM-Llama-3.1-8B, with gains persisting up to $k$16K tokens despite training samples of at most $k$17K tokens (Whitecross et al., 10 Apr 2026).

A complementary theoretical account is given by a kinetic theory for Transformers. Causal self-attention is modeled as a non-exchangeable interacting particle system with triangular dependency graph and ALiBi-like causal weights. The mean-field limit alone does not explain retrieval-position effects; the decisive object is the next-order cross-correlation between the final token and the source token. For iid uniformly distributed tokens, the resulting correlation equation can be solved in closed form, and under an explicit smallness condition the token retrieval profile is $k$18-shaped, with primacy, recency, and a unique global minimum in the interior. The paper thus gives a rigorous explanation of the empirical lost-in-the-middle phenomenon in terms of prefix accumulation, recency bias, and triangular causal propagation of correlations (Duerinckx et al., 9 May 2026).

Another line of work identifies lost-in-the-later, a positional bias in which LLMs increasingly underuse or ignore information that appears later in the provided context. Using the CoPE framework and the multilingual MultiWikiAtomic dataset, the paper finds that earlier context segments contribute most to the response, later segments contribute least, and the imbalance worsens as context length grows up to $k$19 atomic sentences. CK usage peaks at about $k$20, and reasoning models as well as non-reasoning models prompted with chain-of-thought use context even less than non-reasoning models without CoT. The proposed CK Prompt improves contextual grounding and makes recall more even across context quartiles (Tao et al., 7 Jul 2025).

The same positional vulnerability appears in long-input, long-output generation. LongInOutBench constructs $k$21 synthetic multi-document scientific summarization samples from $k$22 arXiv papers, and the analysis shows that answer quality drops most for the middle-position paper. RAL-Writer mitigates this by explicitly retrieving and restating important chunks that are both semantically relevant and likely to be overlooked because of their position. The importance score is

$k$23

where $k$24 is an embedding-based relevance score and $k$25 is a position score that is minimal in the middle and maximal at both ends. The method improves consistency and quality relative to plan-and-write baselines (Zhang et al., 10 Mar 2025).

In multimodal KB-VQA, the shape changes again. “Lost at the End” reports that the usual U-shaped lost-in-the-middle effect flips to primacy: gold evidence placed first in the prompt outperforms gold evidence placed last by $k$26 to $k$27 percentage points across all reader-by-benchmark cells at $k$28. Text-only controls show that multimodality amplifies an already-present text-mode primacy bias by $k$29 to $k$30 times, while image-position and distractor-shuffle ablations localize the main failure to prompt slot $k$31 of the instruction-tuned reader. Retrieval-side interventions such as MMR, oracle reranking, and rank-based reordering leave the gap intact, so recall@$k$32 is not an adequate proxy for end-to-end performance in deployed multimodal KB-VQA (Liu et al., 15 Jun 2026).

7. Adjacent analogues and interpretive limits

Several neighboring literatures extend the same structural intuition even when the label is not literally “Lost-K.” In transfinite computability, a lost melody is a real $k$33 such that the singleton $k$34 is decidable while $k$35 itself is not computable. The phenomenon occurs for resetting ITRMs, does not occur for unresetting infinite time register machines where recognizability equals computability, and exhibits a finer model-dependent classification for parameter-free OTMs, parameterized OTMs, and $k$36-register machines (Carl, 2014).

In benchmark evaluation, the “lost MNIST digits” work reconstructs the full original $k$37-sample test set, including the $k$38 test images that were never distributed. Error rates on the recovered $k$39K are usually slightly worse than on the standard $k$40K test set, but classifier ordering and model selection remain broadly reliable. The authors attribute this robustness to the pairing benefits of comparing classifiers on the same digits, implying that shared test data can remain comparatively informative even when absolute performance estimates drift (Yadav et al., 2019).

In multi-object tracking, detector misses create lost tracks that may never be recovered under standard tracking-by-detection. The proposed Compensation Tracker treats the lost set as a recoverable candidate pool via motion compensation and object selection, without additional networks or retraining. On MOT2020, the reported result is $k$41 of MOTA and $k$42 of IDF1, with substantial reductions in identity switches relative to the baseline (Zou et al., 2020).

A final caution comes from fractional modeling of circuits and neurons. One paper argues that if charged-particle transport is described using Riemann–Liouville or Caputo derivatives, then objectivity is lost, because such derivatives are nonlocal and anchored to a lower integration limit. Under the paper’s observer-shift criterion, classical integer-order models remain objective but the fractional versions generally do not (Balint et al., 2018).

Taken together, these adjacent cases suggest that the most informative use of the Lost-K label is comparative rather than universal. It names a recurrent pattern in which loss, omission, suppression, or positional disadvantage does not merely remove information; it reorganizes the effective state space, often leaving behind a simpler invariant, a displaced hard object, a nonmonotonic threshold, or a measurable retrieval bias. The precise content of the phenomenon is therefore inseparable from the representation in which the loss is defined.