Sleep and Dreams: The Nightly Voyage Out of Waking

The rules are written down, and they have been rewritten. From 1968 the standard was the Rechtschaffen and Kales manual, which cut sleep into five stages: NREM stages 1, 2, 3 and 4, plus REM. In 2007 the American Academy of Sleep Medicine published a revised scoring manual that reduced those five to four. Stage 1 became N1. Stage 2 became N2. Stages 3 and 4 were merged into a single stage, N3. REM became R. The AASM system is the standard used in clinical and research polysomnography today, which is why modern charts show four stages and older ones show five. Nothing about anybody's brain changed in 2007. What changed was the rule for cutting a continuous set of signals into named pieces, and two stages that had been counted separately under the older manual stopped being counted separately.

What the scorer is reading is a small set of recorded signals: chiefly the EEG trace of cortical electrical activity and the EOG record of eye movement, with muscle tone entering the definition of REM. Each stage has a signature. N1 is light sleep, carrying theta waves at 4 to 7 Hz. N2 carries sleep spindles, thalamocortical bursts at 12 to 14 Hz, together with K-complexes, and it takes the largest share of total sleep time. N3 is slow-wave sleep, delta oscillations from 0.5 to 4 Hz, and the highest-amplitude EEG activity of any state, asleep or awake. REM shows rapid eye movements, near-total skeletal muscle atonia, and a desynchronized low-amplitude EEG that looks remarkably like wakefulness. A full night typically runs through 4 to 6 of these ultradian cycles at roughly 90 minutes each, with slow-wave sleep concentrated in the first third of the night and REM periods lengthening toward morning.

Alexander Borbely formalized the modern account in 1982: sleep is driven by the interaction of two processes. Process S is homeostatic, the pressure that builds the longer you stay awake and discharges while you sleep. Process C is circadian, an internal oscillation that pushes for sleep and for waking at particular points in the day whatever your recent history. The framework has held up through decades of sleep deprivation studies, through forced desynchrony protocols, and through the molecular identification of adenosine accumulation as the biochemical substrate of Process S. Two clocks, one of them counting how long you have been up and the other counting the day.

One boundary is worth marking here, because two states look alike from outside and are not alike underneath. Sleep and general anesthesia are both reversible conditions in which a person stops responding, and there the resemblance ends. Sleep is endogenous and cyclical, governed by that two-process interaction: homeostatic pressure mediated by adenosine, working against a circadian oscillation of approximately 24.2 hours driven by the suprachiasmatic nucleus. General anesthesia is drug-induced, acts on specific molecular targets including GABA-A and NMDA receptors, and works by disrupting long-range cortical connectivity and thalamocortical information transfer. It has no ultradian cycling, no alternation of REM and NREM, and no relationship to time of day or to how long you have been awake. The mechanism of anesthesia has its own article elsewhere in this library. The point here is only that being put under is not being put to sleep.

Xie and colleagues demonstrated in Science in 2013 that the interstitial space in the mouse brain expands by approximately 60 percent during sleep, or under anesthesia, and that this expansion facilitates a cerebrospinal-fluid-driven clearance of metabolic waste, including amyloid-beta. This is the glymphatic system model, developed by Nedergaard's group, and it has since been confirmed in human subjects through MRI-based techniques. The brain appears to open its own drainage while the person using it is offline.

The glymphatic picture is not closed. The open question is how much of the transport is convective bulk flow and how much is ordinary diffusion, and whether what holds in a mouse brain holds in a much larger human one. Hladky and Barrand argued in 2014, in Fluids and Barriers of the CNS, that the evidence for bulk flow in human brains, as against mouse brains, remains inconclusive. That the brain clears waste during sleep is well supported. The specific hydraulics of how it does so in a human skull are still argued over.

Staresina and colleagues used intracranial recordings in epilepsy patients, published in Nature Neuroscience in 2015, to demonstrate a temporal coupling between three rhythms: neocortical slow oscillations, thalamic sleep spindles, and hippocampal sharp-wave ripples. The three nest hierarchically inside one another. This triple coupling is the strongest mechanistic support for the active systems consolidation model, in which memory is transferred during sleep. Note carefully what is solid here and what is not. The coupling is measured directly, in human brains, with electrodes in them. Exactly how that coupling carries any particular memory from one place to another is a further claim, and a less settled one.

Tononi and Cirelli proposed, in 2003 and again in 2006, that net synaptic strength increases across waking and decreases during slow-wave sleep, so that one function of deep sleep is renormalization: turning the gain back down before it saturates. The supporting evidence is varied and real. Molecular markers such as Homer1a and Arc track the predicted pattern, electrophysiological recordings show reduced synaptic responses after sleep, and de Vivo and colleagues reported in Science in 2017 that electron microscopy finds smaller synaptic interfaces after sleep than before it. The objection is equally real: Marcos Frank argued in 2012 that sleep also involves synaptic potentiation, not downscaling alone. Synaptic homeostasis is a leading account of what slow-wave sleep is for. It is not the settled answer to that question.

Siclari and colleagues reported in 2017 that dreaming during NREM sleep is associated with local decreases in low-frequency EEG activity in a posterior region of cortex, a 'hot zone.' That result cuts directly against the older REM-centric view. What it supports instead is that dreaming is a graded phenomenon occurring across sleep stages, with REM dreams being more vivid and more narrative, but not qualitatively unique. REM is a stage of sleep. Dreaming is something the brain does in more than one of them, and the two are not the same category of thing.

Here is the supporting statistic, which should be read as support and not as a contradiction of what was just said. Approximately 80 percent of awakenings from REM sleep produce a dream report, and approximately 20 to 50 percent of awakenings from NREM sleep also produce one. The NREM reports tend to be shorter, less vivid and more thought-like. REM reports are characterized by visual imagery, emotional intensity, especially fear, anxiety and anger, narrative structure, incorporation of recent waking experience, and rapid forgetting on waking. The right reading of those percentages is a difference of degree. A fifth to a half of NREM awakenings yielding a dream report is not the profile of a stage in which dreaming does not happen.

Every figure in the paragraph above, and in every study of dream content anywhere, comes from the same procedure: wake a sleeping person and ask what was going through their mind. The datum is the report. The report is a memory of an experience, produced by a brain that has just changed state, and the same research notes rapid forgetting on waking as a defining feature of the material being recalled. This is not a reason to dismiss the work, which is careful and has produced real, replicated numbers. It is a reason to keep the word 'reported' in the sentence. When a study says that dreams contain some proportion of a category of content, what has been measured is the proportion in what people said.

What can be measured without a report is the brain state underneath. PET and fMRI studies of REM sleep show high activation in the pontine brainstem, the thalamus, the visual association cortex, the amygdala, the hippocampus and the anterior cingulate cortex, together with low activation in the dorsolateral prefrontal cortex, the region most associated with executive control, working memory and reality-monitoring. Maquet and colleagues confirmed that pattern with PET imaging in work published in 1996 and 2000. It is a good explanation of two familiar features of dreaming: dreamers accept bizarre content uncritically, with the DLPFC offline, and dreams run emotionally hot, with limbic structures including the amygdala hyperactivated. The pattern also overlaps substantially with the default mode network. Note the shape of that explanation. It accounts for the character of dreaming. It does not read the dream.

One small case shows how much that report layer matters. The belief that people once dreamed in black and white and now dream in color is not a finding about dreams. Most people dream in color, confirmed by laboratory dream reports, though some do report grayscale dreams. The black-and-white claim appears to track the era of black-and-white television and film, which makes it a cultural influence on how dreams were reported rather than a change in what was dreamt. The reporting channel has a history of its own, and that history can be mistaken for the phenomenon.

Hobson and McCarley proposed the activation-synthesis hypothesis in 1977: dreams are generated when the cortex attempts to interpret what Hobson and McCarley described as random neural signals arriving from the brainstem, particularly the pontine reticular formation, during REM sleep. The dream, on this reading, is the narrative gloss the cortex puts on activity it did not itself originate. Hobson later revised the idea into the AIM model, organized around activation, input source and modulation. The criticism the source literature itself carries is aimed at one word. Describing the incoming activation as random is overstated, because dream content is not random: it shows consistent patterns reflecting waking concerns, emotional preoccupations, and memory processes.

Revonsuo proposed in 2000 that dreams evolved as an offline threat simulation system, rehearsing responses to threatening situations in a safe virtual environment. The evidence offered is that threatening events are overrepresented in dream content relative to waking life, across published cross-cultural findings on being chased, attacked, falling, and social threat. The criticism the source literature carries is direct and hard to wave away: not all dreams are threatening, and the theory has difficulty accounting for mundane, positive, or creative dreams. Note also what the supporting evidence is made of. It is content data, which means it is report data, with the limit named in the previous section attached to it.

The third account is that dreaming plays a role in memory consolidation and integration. Specific dream content correlates with recently learned material, and hippocampal replay of waking experience during sleep is well documented. Walker and Stickgold argued in 2004 that sleep benefits memory consolidation, and that REM sleep may particularly benefit emotional memory consolidation and creative problem-solving. The gap in this account is precise and should not be papered over: the relationship between neural replay and conscious dream content is unclear. That sleep consolidates memory is well supported, and that is a claim about sleep. That dreaming does the consolidating is a further claim about experience, and it is the part that is not established.

A near relative of that third account gets far more airtime than its evidence supports. Anecdotal reports of scientific discoveries arriving in dreams, Kekule's benzene ring in 1865 and Mendeleev's arrangement of the periodic table in 1869, are offered as evidence of a problem-solving function. Stickgold has provided experimental evidence, published in 2005 and 2009, that sleep facilitates insight and creative restructuring on memory tasks. That is a real finding about sleep. Whether dreaming specifically, as against sleep-stage physiology, is what drives it remains unclear. And the two famous stories are anecdotes, which is a description of their evidential class, not an insult to them.

The Freudian account, that dreams are disguised wish-fulfillments expressing unconscious desires, remains influential in psychoanalytic circles and lacks empirical validation. The part of it that survives is modest and real: some dream content clearly reflects emotional concerns. The specific machinery Freud proposed, the symbolism and the disguise, is not supported.

Pre-modern interpretive frameworks read dreams as prophetic messages from gods or spirits. Modern neuroscience identifies them instead as products of endogenous brain activity, with content shaped by recent experience. Freud named that in 1900 as day residue, and our research file states that approximately 65 percent of dream content incorporates fragments from the prior one to seven days. One caution on that figure, offered because this article would rather be exact than tidy: the file does not name the specific modern studies behind it, and it was not traced to a single named source when this article was researched. Take it as the file's own stated number, consistent with a well-established day-residue literature, rather than as a figure independently re-verified here.

Anecdotal reports of dreams predicting future events are common across cultures and have no controlled empirical support. The standard explanation is coincidence, selective memory, and the sheer number of dreams a person has across a lifetime. That explanation has not itself been proven in a controlled setting, which is why this sits where it does rather than being called closed. What can be said is that nothing in the controlled literature supports the alternative.

The overstatement runs in the other direction too, and it is the one an enthusiastic reader of activation-synthesis is most likely to make: that dreams are meaningless noise. They are not. Dream content is not random. It shows systematic relationships to waking concerns, emotional state, and recent experience. Dreams have structured content even where their generation involves bottom-up noise. A brain assembling a story partly out of noisy input still produces a story that is about something.

The population guideline is real, named and recent. Watson, Badr, Belenky and colleagues published a joint consensus statement of the American Academy of Sleep Medicine and the Sleep Research Society in 2015, in the Journal of Clinical Sleep Medicine. A panel of 15 experts reviewed over 5,300 scientific articles and concluded that adults should sleep 7 or more hours per night on a regular basis to promote optimal health, and that regularly sleeping less than that is associated with adverse health outcomes. Note the figure the consensus actually supports: seven or more, not a fixed eight. The popular round number is not what the panel wrote.

The laboratory dose-response finding is a different kind of claim, and a bounded one. Van Dongen and colleagues published in Sleep in 2003 that restricting sleep to 6 hours a night for 14 nights produced cognitive deficits equivalent to approximately 48 hours of total sleep deprivation, and that subjects were often unaware of their own impairment, which is the sleep-debt phenomenon in one sentence. Matthew Walker's 2017 book Why We Sleep carried that class of finding to a general audience. Read the study at its own size: a specific restriction schedule, in one cohort, against a defined outcome measure. It is real, replicated work on a dose-response relationship. It is not a statement that everybody needs exactly eight hours.

The famous extreme case is the third kind of claim, and the one that needs the most care. Randy Gardner, a 17-year-old high school student in San Diego, stayed awake for 264 hours, 11 days, from January 28 to February 8, 1964, as a science-fair project. It was monitored by Stanford sleep researchers, including William Dement, and by a US Navy physician. Cognitive decline, mood swings and hallucinations appeared well before the end. He recovered after roughly 15 hours of sleep with no documented lasting damage. The record was broken repeatedly afterward, until Guinness World Records stopped accepting sleep-deprivation attempts in 1997 for safety reasons. What this case is: one closely monitored subject. What it is not: a controlled study, or evidence about what anybody else can safely do.

The short-sleeper claim, that some people simply need very little sleep, is the rare piece of folk wisdom with a real genetic basis that still does not mean what it is usually used to mean. A mutation in the DEC2 gene, also called BHLHE41, identified by He, Jones, Fujiki and colleagues and published in Science in 2009, does produce genuine short sleepers. It affects fewer than approximately 1 percent of the population. For the vast majority of people, chronic sleep restriction below 7 hours produces measurable cognitive and health deficits. The variant is real, and the Tier 4 badge on this block grades the folk inference rather than the 2009 finding. The odds that any particular person who says they get by on very little is carrying that variant are small.

One more thing belongs in this section, because leaving it out would be exactly the sort of tidiness this article is against. Some of Matthew Walker's claims in Why We Sleep have been challenged, by Alexey Guzey in a 2019 piece and by others, for overstating effect sizes and citing the literature selectively. Walker has acknowledged some corrections while maintaining his core arguments about the risks of sleep deprivation. The underlying science of sleep deprivation is not overturned by that exchange. The popular presentation of it was, in places, sharper than the evidence, and that is worth knowing whenever you meet a confident sleep statistic out in the world.