{"id":"b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c","title":"How Far Can a Trainable Frequency Move?","argument":"Michael Poppel, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien and Jonas Stein study how trainable-frequency quantum models can move encoding-gate frequencies toward a target. They argue that starting frequency matters: from the unit start used in prior work, training rarely reaches targets more than one resolution step away. Their analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach a target from within half a step; with a fixed coefficient, a phase-dependent barrier can stop it. A ternary prefactor grid puts a model frequency within half a step of every band frequency, using logarithmically many encoding gates. In experiments, this start raised successful fits on synthetic targets from 69% to 84% compared with the best random start. The results also qualify that improvement: with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not distinguish these series from the others. The authors say the starting-point dependence also appears in a classical model with independently trainable frequencies, and frame the work as guidance on training, not as evidence of a circuit advantage.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports mixed outcomes on real series: held-out R² improved on four of seven, and the classical screen did not distinguish those series from the others. The authors report that with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. The supplied source is the arXiv abstract and page text; it does not include the paper’s full methods or experimental details.","speaker":{"name":"Poppel, Michael, Baumann, Markus, Wölckert, Sebastian, Linnhoff-Popien, Claudia, Stein, Jonas","roleAtSource":"Article author"},"source":{"title":"How Far Can a Trainable Frequency Move? Resolution, Grids and Coupling in Trainable-FrequencyQuantum Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2602.23409","publishedAt":"2026-02-26","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors’ analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach its target from within half a step; a fixed coefficient can encounter a phase-dependent barrier.","excerpt":"under gradient flow, an isolated real frequency whose coefficient adapts optimally reaches its target from within half a step","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"The authors propose a ternary prefactor grid that places a model frequency within half a step of every band frequency, using logarithmically many encoding gates.","excerpt":"A ternary prefactor grid places a model frequency within half a step of every band frequency, with logarithmically many encoding gates.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"On synthetic targets, the grid start raised successful fits from 69% to 84% compared with the best random start.","excerpt":"the grid start raised successful fits on synthetic targets from 69% to 84% against the best random start.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"For 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","excerpt":"with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not separate those series from the others.","excerpt":"training the frequencies improved held-out $R^2$ at equal budget by 0.010--0.045 on four of seven series, but a classical screen did not separate these from the others.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"The authors say a classical model with independently trainable frequencies shows the same dependence on the start, and describe their results as guidance on training rather than evidence of a circuit advantage.","excerpt":"our results describe how to train such models, not an advantage of the circuit.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.23409","checkedAt":"2026-11-09T10:14:07.495Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What explains why the classical screen did not distinguish the four real series with improved held-out R² from the others?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How Far Can a Trainable Frequency Move?","summary":"Michael Poppel, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien and Jonas Stein study how trainable-frequency quantum models can move encoding-gate frequencies toward a target. They argue that starting frequency matters: from the unit start used in prior work, training rarely reaches targets more than one resolution step away. Their analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach a target from within half a step; with a fixed coefficient, a phase-dependent barrier can stop it. A ternary prefactor grid puts a model frequency within half a step of every band frequency, using logarithmically many encoding gates. In experiments, this start raised successful fits on synthetic targets from 69% to 84% compared with the best random start. The results also qualify that improvement: with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not distinguish these series from the others. The authors say the starting-point dependence also appears in a classical model with independently trainable frequencies, and frame the work as guidance on training, not as evidence of a circuit advantage.","question":"What explains why the classical screen did not distinguish the four real series with improved held-out R² from the others?","claims":[{"text":"The authors’ analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach its target from within half a step; a fixed coefficient can encounter a phase-dependent barrier.","excerpt":"under gradient flow, an isolated real frequency whose coefficient adapts optimally reaches its target from within half a step","support":"source-stated"},{"text":"The authors propose a ternary prefactor grid that places a model frequency within half a step of every band frequency, using logarithmically many encoding gates.","excerpt":"A ternary prefactor grid places a model frequency within half a step of every band frequency, with logarithmically many encoding gates.","support":"source-stated"},{"text":"On synthetic targets, the grid start raised successful fits from 69% to 84% compared with the best random start.","excerpt":"the grid start raised successful fits on synthetic targets from 69% to 84% against the best random start.","support":"source-stated"},{"text":"For 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","excerpt":"with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better.","support":"source-stated"},{"text":"On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not separate those series from the others.","excerpt":"training the frequencies improved held-out $R^2$ at equal budget by 0.010--0.045 on four of seven series, but a classical screen did not separate these from the others.","support":"source-stated"},{"text":"The authors say a classical model with independently trainable frequencies shows the same dependence on the start, and describe their results as guidance on training rather than evidence of a circuit advantage.","excerpt":"our results describe how to train such models, not an advantage of the circuit.","support":"source-stated"}],"limitations":["The abstract reports mixed outcomes on real series: held-out R² improved on four of seven, and the classical screen did not distinguish those series from the others.","The authors report that with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","The supplied source is the arXiv abstract and page text; it does not include the paper’s full methods or experimental details."],"mediaId":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","correctionReason":null,"id":"b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c","publicationId":"f1a0b531-58be-4b7f-883b-ed76acd70721","version":1,"sourceUrl":"https://arxiv.org/abs/2602.23409","sourceTitle":"How Far Can a Trainable Frequency Move? Resolution, Grids and Coupling in Trainable-FrequencyQuantum Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-02-26","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:12:30.093Z","createdAt":"2026-11-08T17:10:42.200Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","kind":"article","url":"https://arxiv.org/abs/2602.23409","sourceUrl":"https://arxiv.org/abs/2602.23409","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T17:09:42.362Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.23409","checkedAt":"2026-11-09T10:14:07.495Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","confidence":0.6699999999999999,"provider":"typesafe","model":"jev-1.13.0","inputHash":"7a78d2e1a04a483d409177eae5db30956fa654b886c5ad49f594974beabc07cb","decidedAt":"2026-11-08T17:09:43.224Z"},"provenance":{"textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observationId":"d942256999bb5095ce1e0c48eb86c26d937566bce58c6d2b254ccc0b628a312e","retrievedAt":"2026-11-08T17:09:42.230Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"4da419fd-36c8-4c3a-8da3-3ef910737540"},"reviewer":{"model":"gpt-6-luna","runId":"b65730ea-52b7-4562-85cf-33819a72efe6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T17:10:42.200Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Poppel, Michael","url":null,"kind":"unknown"},{"name":"Baumann, Markus","url":null,"kind":"unknown"},{"name":"Wölckert, Sebastian","url":null,"kind":"unknown"},{"name":"Linnhoff-Popien, Claudia","url":null,"kind":"unknown"},{"name":"Stein, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Poppel, Michael","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Poppel, Michael","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Baumann, Markus","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Baumann, Markus","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Wölckert, Sebastian","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wölckert, Sebastian","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Linnhoff-Popien, Claudia","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Linnhoff-Popien, Claudia","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Stein, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Stein, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2602.23409","snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z","basis":"publisher-page-metadata","sourceLinks":52,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c/source"}},"canonicalUrl":"https://escapemeridian.com/readings/how-far-can-a-trainable-frequency-move-b92ff6fb","formats":{"html":"/readings/how-far-can-a-trainable-frequency-move-b92ff6fb","markdown":"/readings/how-far-can-a-trainable-frequency-move-b92ff6fb.md","json":"/readings/how-far-can-a-trainable-frequency-move-b92ff6fb.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c","revision":"f1a0b531-58be-4b7f-883b-ed76acd70721","revisionKind":"publication-id","title":"How Far Can a Trainable Frequency Move?","url":"https://escapemeridian.com/readings/how-far-can-a-trainable-frequency-move-b92ff6fb"},"source":{"type":"source","id":"https://arxiv.org/abs/2602.23409","revision":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","revisionKind":"sha256-source-text","title":"How Far Can a Trainable Frequency Move? Resolution, Grids and Coupling in Trainable-FrequencyQuantum Models","url":"https://arxiv.org/abs/2602.23409"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c","title":"How Far Can a Trainable Frequency Move?","argument":"Michael Poppel, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien and Jonas Stein study how trainable-frequency quantum models can move encoding-gate frequencies toward a target. They argue that starting frequency matters: from the unit start used in prior work, training rarely reaches targets more than one resolution step away. Their analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach a target from within half a step; with a fixed coefficient, a phase-dependent barrier can stop it. A ternary prefactor grid puts a model frequency within half a step of every band frequency, using logarithmically many encoding gates. In experiments, this start raised successful fits on synthetic targets from 69% to 84% compared with the best random start. The results also qualify that improvement: with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not distinguish these series from the others. The authors say the starting-point dependence also appears in a classical model with independently trainable frequencies, and frame the work as guidance on training, not as evidence of a circuit advantage.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports mixed outcomes on real series: held-out R² improved on four of seven, and the classical screen did not distinguish those series from the others. The authors report that with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. The supplied source is the arXiv abstract and page text; it does not include the paper’s full methods or experimental details.","speaker":{"name":"Poppel, Michael, Baumann, Markus, Wölckert, Sebastian, Linnhoff-Popien, Claudia, Stein, Jonas","roleAtSource":"Article author"},"source":{"title":"How Far Can a Trainable Frequency Move? Resolution, Grids and Coupling in Trainable-FrequencyQuantum Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2602.23409","publishedAt":"2026-02-26","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors’ analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach its target from within half a step; a fixed coefficient can encounter a phase-dependent barrier.","excerpt":"under gradient flow, an isolated real frequency whose coefficient adapts optimally reaches its target from within half a step","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"The authors propose a ternary prefactor grid that places a model frequency within half a step of every band frequency, using logarithmically many encoding gates.","excerpt":"A ternary prefactor grid places a model frequency within half a step of every band frequency, with logarithmically many encoding gates.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"On synthetic targets, the grid start raised successful fits from 69% to 84% compared with the best random start.","excerpt":"the grid start raised successful fits on synthetic targets from 69% to 84% against the best random start.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"For 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","excerpt":"with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not separate those series from the others.","excerpt":"training the frequencies improved held-out $R^2$ at equal budget by 0.010--0.045 on four of seven series, but a classical screen did not separate these from the others.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"},{"text":"The authors say a classical model with independently trainable frequencies shows the same dependence on the start, and describe their results as guidance on training rather than evidence of a circuit advantage.","excerpt":"our results describe how to train such models, not an advantage of the circuit.","url":"https://arxiv.org/abs/2602.23409","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.23409","checkedAt":"2026-11-09T10:14:07.495Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What explains why the classical screen did not distinguish the four real series with improved held-out R² from the others?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How Far Can a Trainable Frequency Move?","summary":"Michael Poppel, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien and Jonas Stein study how trainable-frequency quantum models can move encoding-gate frequencies toward a target. They argue that starting frequency matters: from the unit start used in prior work, training rarely reaches targets more than one resolution step away. Their analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach a target from within half a step; with a fixed coefficient, a phase-dependent barrier can stop it. A ternary prefactor grid puts a model frequency within half a step of every band frequency, using logarithmically many encoding gates. In experiments, this start raised successful fits on synthetic targets from 69% to 84% compared with the best random start. The results also qualify that improvement: with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors. On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not distinguish these series from the others. The authors say the starting-point dependence also appears in a classical model with independently trainable frequencies, and frame the work as guidance on training, not as evidence of a circuit advantage.","question":"What explains why the classical screen did not distinguish the four real series with improved held-out R² from the others?","claims":[{"text":"The authors’ analysis says that, under gradient flow, an isolated real frequency with an optimally adapting coefficient can reach its target from within half a step; a fixed coefficient can encounter a phase-dependent barrier.","excerpt":"under gradient flow, an isolated real frequency whose coefficient adapts optimally reaches its target from within half a step","support":"source-stated"},{"text":"The authors propose a ternary prefactor grid that places a model frequency within half a step of every band frequency, using logarithmically many encoding gates.","excerpt":"A ternary prefactor grid places a model frequency within half a step of every band frequency, with logarithmically many encoding gates.","support":"source-stated"},{"text":"On synthetic targets, the grid start raised successful fits from 69% to 84% compared with the best random start.","excerpt":"the grid start raised successful fits on synthetic targets from 69% to 84% against the best random start.","support":"source-stated"},{"text":"For 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","excerpt":"with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better.","support":"source-stated"},{"text":"On real series, training frequencies improved held-out R² at equal budget on four of seven series, but a classical screen did not separate those series from the others.","excerpt":"training the frequencies improved held-out $R^2$ at equal budget by 0.010--0.045 on four of seven series, but a classical screen did not separate these from the others.","support":"source-stated"},{"text":"The authors say a classical model with independently trainable frequencies shows the same dependence on the start, and describe their results as guidance on training rather than evidence of a circuit advantage.","excerpt":"our results describe how to train such models, not an advantage of the circuit.","support":"source-stated"}],"limitations":["The abstract reports mixed outcomes on real series: held-out R² improved on four of seven, and the classical screen did not distinguish those series from the others.","The authors report that with 8 and 24 target frequencies, training the fixed grid ten times longer fitted better than training prefactors.","The supplied source is the arXiv abstract and page text; it does not include the paper’s full methods or experimental details."],"mediaId":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","correctionReason":null,"id":"b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c","publicationId":"f1a0b531-58be-4b7f-883b-ed76acd70721","version":1,"sourceUrl":"https://arxiv.org/abs/2602.23409","sourceTitle":"How Far Can a Trainable Frequency Move? Resolution, Grids and Coupling in Trainable-FrequencyQuantum Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-02-26","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:12:30.093Z","createdAt":"2026-11-08T17:10:42.200Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","kind":"article","url":"https://arxiv.org/abs/2602.23409","sourceUrl":"https://arxiv.org/abs/2602.23409","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T17:09:42.362Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2602.23409","checkedAt":"2026-11-09T10:14:07.495Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"a30a1287c93aafcccd7b23d2d9f05580cc6c0a174c08c414bde0c55cd69697b2","confidence":0.6699999999999999,"provider":"typesafe","model":"jev-1.13.0","inputHash":"7a78d2e1a04a483d409177eae5db30956fa654b886c5ad49f594974beabc07cb","decidedAt":"2026-11-08T17:09:43.224Z"},"provenance":{"textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observationId":"d942256999bb5095ce1e0c48eb86c26d937566bce58c6d2b254ccc0b628a312e","retrievedAt":"2026-11-08T17:09:42.230Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"4da419fd-36c8-4c3a-8da3-3ef910737540"},"reviewer":{"model":"gpt-6-luna","runId":"b65730ea-52b7-4562-85cf-33819a72efe6"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T17:10:42.200Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Poppel, Michael","url":null,"kind":"unknown"},{"name":"Baumann, Markus","url":null,"kind":"unknown"},{"name":"Wölckert, Sebastian","url":null,"kind":"unknown"},{"name":"Linnhoff-Popien, Claudia","url":null,"kind":"unknown"},{"name":"Stein, Jonas","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Poppel, Michael","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Poppel, Michael","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Baumann, Markus","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Baumann, Markus","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Wölckert, Sebastian","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wölckert, Sebastian","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Linnhoff-Popien, Claudia","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Linnhoff-Popien, Claudia","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}},{"actor":{"name":"Stein, Jonas","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Stein, Jonas","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2602.23409","locator":"head/meta[name=author|citation_author]"},"assertionId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T10:14:07.495Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","sourceUrl":"https://arxiv.org/abs/2602.23409","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2602.23409","snapshotId":"55070ad2a7b41adf63caac290b74e7fe2e87c3898470b5e56a3eb693c3b5a506","contentSha256":"753b2adcd6dccee80fc4c38913c846c9b69dcdd0a809b733424e412b25c4a15c","textHash":"11948ddd0aa92caff37c037d0fa6d0fba825511cd0dc7d740a8c7a3fd8d45c91","observedAt":"2026-11-09T10:14:07.495Z","basis":"publisher-page-metadata","sourceLinks":52,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b92ff6fbbec9b3bf8de1409caf41ca3705beba59674265c0ea45308bbba0b72c/source"}}}],"audits":[{"publicationId":"f1a0b531-58be-4b7f-883b-ed76acd70721","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The five authors match the arXiv page. The findings and the framing as training guidance rather than a circuit advantage are attributed to them, not presented as established fact or as Escape Meridian's view.","excerpt":"Authors:Michael Poppel, Markus Baumann, Sebastian Wölckert, Claudia Linnhoff-Popien, Jonas Stein"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful but hard to follow for a reader without specialist background. Central terms are never explained in plain language: trainable-frequency quantum models, encoding gates, prefactors, resolution step, gradient flow, ternary prefactor grid, band frequency, held-out R² and classical screen. It also never says why moving a frequency matters for fitting data. Smallest repair: add one plain sentence saying the model fits signals by tuning frequencies and that the paper asks how far training can shift them. Keep the technical terms and attach tooltips. The question also relies on the unexplained term classical screen.","excerpt":"Trainable-frequency quantum models train the prefactors of their encoding gates so that the circuit's frequencies can move onto the target's."},{"dimension":"date","status":"pass","explanation":"originalPublished is 2026-02-26, which matches the v1 submission date. The 5 Nov 2026 date belongs to the later v4 revision and the feed date is later still, so neither is used.","excerpt":"[Submitted on 26 Feb 2026 (v1), last revised 5 Nov 2026 (this version, v4)]"},{"dimension":"support","status":"uncertain","explanation":"Most claims match the abstract verbatim. However, the reach result drops a stated condition: it holds for every target at least one step from zero frequency. The summary and claim 1 omit this; restore it. The phrase 'qualify that improvement' links the 69% to 84% grid-start result to a separate many-frequency experiment, which the source does not frame that way. Repair: describe the separate regime where least squares on the fixed grid fits near-perfectly but the circuit does not reach the coefficients. The real-series gain size (0.010 to 0.045) is also omitted.","excerpt":"reaches its target from within half a step, for every target at least one step from zero frequency"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-08T17:10:42.200Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}