bechtel.explicating top-down.4.0.forpreprint


UC San Diego
UC San Diego Previously Published Works

Title
Explicating Top-Down Causation Using Networks and Dynamics

Permalink
https://escholarship.org/uc/item/34m389rq

Journal
PHILOSOPHY OF SCIENCE, 84(2)

ISSN
0031-8248

Author
Bechtel, William

Publication Date
2017-04-01

DOI
10.1086/690718
 
Peer reviewed

eScholarship.org Powered by the California Digital Library
University of California

https://escholarship.org/uc/item/34m389rq
https://escholarship.org
http://www.cdlib.org/


Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
	
  

Explicating	
  Top-­‐Down	
  Causation	
  Using	
  Networks	
  and	
  Dynamics	
  
	
  

William	
  Bechtel	
  
Department	
  of	
  Philosophy	
  and	
  Center	
  for	
  Circadian	
  Biology	
  

University	
  of	
  California,	
  San	
  Diego	
  
	
  
Abstract	
  
	
  
In	
  many	
  fields	
  in	
  the	
  life	
  sciences	
  investigators	
  refer	
  to	
  downward	
  or	
  top-­‐down	
  causal	
  
effects.	
  Craver	
  and	
  I	
  defended	
  the	
  view	
  that	
  such	
  cases	
  should	
  be	
  understood	
  in	
  terms	
  of	
  a	
  
constitution	
  relation	
  between	
  levels	
  in	
  a	
  mechanism	
  and	
  intra-­‐level	
  causal	
  relations	
  
(occurring	
  at	
  any	
  level).	
  We	
  did	
  not,	
  however,	
  specify	
  when	
  entities	
  constitute	
  a	
  higher-­‐
level	
  mechanism.	
  In	
  this	
  paper	
  I	
  appeal	
  to	
  graph-­‐theoretic	
  representations	
  of	
  networks,	
  
now	
  widely	
  employed	
  in	
  systems	
  biology	
  and	
  neuroscience,	
  and	
  associate	
  mechanisms	
  
with	
  modules	
  that	
  exhibit	
  high	
  clustering.	
  As	
  a	
  result	
  of	
  such	
  interconnections	
  mechanisms	
  
often	
  exhibit	
  complex	
  dynamic	
  behaviors	
  that	
  constrain	
  how	
  individual	
  components	
  
respond	
  to	
  external	
  inputs,	
  a	
  central	
  feature	
  of	
  top-­‐down	
  causation.	
  
	
  
Keywords:	
  constraints;	
  downward	
  causation;	
  endogenous	
  dynamics;	
  graph	
  
representations;	
  mechanistic	
  explanations;	
  negative	
  feedback	
  	
  
	
  
	
  
	
  
Contact	
  Information:	
  William	
  Bechtel,	
  Department	
  of	
  Philosophy,	
  University	
  of	
  California,	
  
San	
  Diego,	
  9500	
  Gilman	
  Drive,	
  La	
  Jolla,	
  CA	
  92093-­‐0119;	
  bechtel@ucsd.edu	
  
	
  
	
  
Acknowledgements	
  
	
  
I	
  thank	
  three	
  anonymous	
  referees	
  for	
  this	
  journal	
  for	
  their	
  very	
  helpful	
  comments	
  and	
  
suggestions.	
  I	
  also	
  thank	
  John	
  Norton	
  and	
  Visiting	
  Fellows	
  at	
  the	
  Center	
  for	
  Philosophy	
  of	
  
Science	
  at	
  the	
  University	
  of	
  Pittsburgh	
  in	
  2014-­‐2015,	
  especially	
  Sara	
  Green,	
  Raphael	
  Scholl,	
  
and	
  Maria	
  Serban,	
  for	
  their	
  spirited	
  discussion	
  of	
  an	
  earlier	
  draft	
  of	
  this	
  paper.	
  Likewise,	
  I	
  
thank	
  members	
  of	
  the	
  audience	
  at	
  the	
  Workshop	
  on	
  Levels	
  of	
  Organization,	
  Causality,	
  and	
  
Top-­‐Down	
  Relations	
  sponsored	
  by	
  the	
  IAS	
  Research	
  Center	
  for	
  Life,	
  Mind,	
  and	
  Society,	
  
University	
  at	
  the	
  Basque	
  Country,	
  San	
  Sebastian,	
  in	
  June	
  2015,	
  especially	
  Leonardo	
  Bich,	
  
Alvaro	
  Moreno,	
  and	
  Kepa	
  Ruiz-­‐Mirazo,	
  for	
  valuable	
  discussion	
  at	
  and	
  after	
  the	
  Workshop.	
  
Finally,	
  I	
  thank	
  Jason	
  Winning	
  for	
  many	
  productive	
  discussions	
  concerning	
  constraints	
  and	
  
mechanism.	
  
	
   	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
1.	
  Introduction	
  
	
  
States	
  of	
  whole	
  systems	
  often	
  constrain	
  the	
  behavior	
  of	
  their	
  parts.	
  Conditions	
  in	
  a	
  cell,	
  
such	
  as	
  its	
  phase	
  in	
  the	
  cell	
  cycle,	
  constrain	
  which	
  genes	
  are	
  expressed.	
  The	
  same	
  molecule	
  
can	
  have	
  different	
  effects	
  on	
  components	
  of	
  a	
  cell	
  (e.g.,	
  promoting	
  or	
  inhibiting	
  apoptosis)	
  
depending	
  on	
  the	
  conditions	
  in	
  the	
  cell.	
  Phenomena	
  such	
  as	
  these	
  are	
  often	
  characterized	
  
as	
  involving	
  downward	
  or	
  top-­‐down	
  causation	
  (Noble	
  2006).	
  They	
  are	
  contrasted	
  with	
  
cases	
  of	
  bottom-­‐up	
  causation	
  in	
  which	
  a	
  state	
  of	
  a	
  component	
  of	
  a	
  system	
  partially	
  
determines	
  the	
  state	
  of	
  the	
  whole	
  (e.g.,	
  a	
  genetic	
  mutation	
  impairs	
  fatty-­‐acid	
  metabolism	
  of	
  
a	
  cell	
  or	
  bonds	
  between	
  actin	
  and	
  myosin	
  produce	
  muscle	
  contraction).	
  While	
  most	
  
researchers	
  do	
  not	
  find	
  bottom-­‐up	
  causation	
  mysterious	
  (it	
  is	
  invoked	
  in	
  reductionistic	
  
explanation	
  of	
  properties	
  of	
  a	
  system	
  in	
  terms	
  of	
  its	
  parts),	
  many	
  have	
  found	
  top-­‐down	
  
causation	
  to	
  present	
  a	
  puzzle:	
  how	
  does	
  a	
  whole	
  system	
  have	
  effects	
  over	
  and	
  above	
  those	
  
of	
  each	
  of	
  its	
  components?	
  	
  
	
  
In	
  an	
  attempt	
  to	
  defuse	
  the	
  mystery,	
  Craver	
  and	
  I	
  (Craver	
  and	
  Bechtel	
  2007)	
  distinguished	
  
constitution	
  and	
  causation	
  and	
  proposed	
  treating	
  constitution	
  as	
  a	
  relation	
  between	
  lower-­‐
level	
  parts	
  and	
  higher-­‐level	
  mechanisms	
  and	
  causation	
  as	
  a	
  relation	
  between	
  entities	
  at	
  the	
  
same	
  level.	
  On	
  our	
  proposal,	
  parts	
  constitute	
  wholes;	
  they	
  don’t	
  cause	
  their	
  properties.	
  By	
  
viewing	
  causation	
  as	
  occurring	
  at	
  all	
  levels,	
  not	
  just	
  the	
  lowest	
  one,	
  we	
  presented	
  
themselves	
  as	
  offering	
  an	
  account	
  that	
  would	
  capture	
  what	
  is	
  characterized	
  as	
  top-­‐down	
  
causation	
  without	
  engendering	
  conceptual	
  problems	
  such	
  as	
  those	
  posed	
  by	
  Kim’s	
  (1998)	
  
exclusion	
  argument.	
  When	
  one	
  thing	
  acts	
  on	
  a	
  whole	
  mechanism,	
  and	
  its	
  components	
  are	
  
also	
  modified,	
  we	
  proposed	
  that	
  higher-­‐level	
  causation	
  was	
  involved	
  in	
  producing	
  the	
  effect	
  
on	
  the	
  whole	
  mechanism.	
  Since	
  the	
  parts	
  are	
  what	
  constitute	
  the	
  whole	
  mechanism,	
  one	
  or	
  
more	
  of	
  them	
  would	
  themselves	
  be	
  changed	
  in	
  that	
  process.	
  No	
  additional	
  causal	
  processes	
  
were	
  involved	
  between	
  the	
  whole	
  and	
  the	
  part,	
  although	
  additional	
  causal	
  processes	
  might	
  
then	
  ensue	
  in	
  the	
  mechanism	
  as	
  a	
  result	
  (thereby	
  altering	
  the	
  state	
  of	
  the	
  system	
  as	
  a	
  
whole).	
  	
  
	
  
However,	
  the	
  examples	
  Craver	
  and	
  I	
  introduced	
  fail	
  to	
  bring	
  out	
  in	
  what	
  sense	
  higher-­‐levels	
  
are	
  involved	
  in	
  producing	
  these	
  effects	
  at	
  the	
  lower	
  level.	
  In	
  offering	
  an	
  example	
  in	
  which	
  a	
  
person’s	
  activity	
  (playing	
  tennis)	
  results	
  in	
  a	
  change	
  within	
  the	
  person’s	
  body	
  (altered	
  
metabolism),	
  we	
  emphasized	
  the	
  role	
  of	
  lower-­‐level	
  causal	
  relations:	
  “In	
  this	
  and	
  many	
  
similar	
  cases,	
  a	
  change	
  in	
  the	
  activity	
  of	
  the	
  mechanism	
  as	
  a	
  whole	
  just	
  is	
  a	
  change	
  in	
  one	
  or	
  
more	
  components	
  of	
  the	
  mechanism	
  which	
  then,	
  through	
  ordinary	
  intra-­‐level	
  causation,	
  
causes	
  changes	
  in	
  other	
  components	
  of	
  the	
  mechanism.”	
  This	
  presents	
  the	
  challenge:	
  why	
  
should	
  one	
  treat	
  the	
  whole	
  mechanism	
  as	
  a	
  higher-­‐level	
  entity	
  rather	
  than	
  just	
  a	
  collection	
  
of	
  lower-­‐level	
  entities,	
  with	
  all	
  the	
  causality	
  operative	
  between	
  these	
  lower-­‐level	
  entities?	
  
This	
  is	
  a	
  version	
  of	
  Kim’s	
  exclusion	
  argument	
  against	
  top-­‐down	
  causation	
  and	
  a	
  number	
  of	
  
critics	
  (see,	
  for	
  example,	
  Soom	
  2012;	
  Rosenberg	
  2015)	
  has	
  objected	
  that	
  in	
  the	
  end	
  Craver	
  
and	
  me,	
  like	
  Kim,	
  only	
  allow	
  causation	
  at	
  the	
  lowest	
  level.	
  In	
  pushing	
  a	
  similar	
  objection,	
  
Fazekas	
  and	
  Kertész	
  (2011)	
  argue	
  that	
  constitution	
  should	
  be	
  expressed	
  in	
  an	
  identity	
  
claim,	
  and	
  this	
  undermines	
  any	
  sense	
  of	
  autonomy	
  for	
  higher-­‐level	
  causation.	
  	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
My	
  goal	
  in	
  this	
  paper	
  is	
  to	
  unpack	
  Craver	
  and	
  my	
  distinction	
  between	
  causation	
  and	
  
constitution	
  by	
  explicating	
  more	
  clearly	
  (1)	
  when	
  an	
  entity	
  or	
  activity	
  that	
  is	
  regarded	
  as	
  at	
  
a	
  higher-­‐level	
  enters	
  into	
  causal	
  relations	
  such	
  that	
  the	
  causality	
  should	
  be	
  treated	
  as	
  at	
  a	
  
higher-­‐level	
  and	
  (2)	
  the	
  relation	
  that	
  holds	
  between	
  the	
  state	
  of	
  the	
  mechanism	
  as	
  a	
  whole	
  
and	
  the	
  state	
  of	
  its	
  components.	
  While	
  bottom-­‐up	
  causation	
  has	
  seemed	
  less	
  mysterious,	
  
the	
  same	
  problem	
  arises	
  in	
  cases	
  of	
  supposed	
  bottom-­‐up	
  causation:	
  how	
  do	
  operations	
  of	
  
parts	
  of	
  a	
  system	
  have	
  effects	
  on	
  the	
  whole	
  when	
  what	
  they	
  seem	
  to	
  have	
  is	
  effects	
  on	
  other	
  
parts,	
  which	
  together	
  constitute	
  whole.	
  Whereas	
  top-­‐down	
  causation	
  raises	
  the	
  question	
  
why	
  the	
  whole	
  is	
  considered	
  the	
  cause,	
  bottom-­‐up	
  causation	
  raises	
  the	
  question	
  why	
  
effects	
  are	
  assigned	
  to	
  the	
  whole.	
  	
  
	
  
Crucial	
  to	
  any	
  account	
  of	
  inter-­‐level	
  causation	
  is	
  the	
  notion	
  of	
  level	
  that	
  is	
  invoked.	
  Craver	
  
and	
  I	
  dissociated	
  our	
  treatment	
  of	
  levels	
  from	
  many	
  in	
  the	
  literature,	
  such	
  as	
  the	
  notion	
  of	
  
levels	
  of	
  science	
  that	
  were	
  invoked	
  in	
  the	
  theory-­‐reduction	
  literature	
  or	
  notions	
  of	
  levels	
  
defined	
  in	
  terms	
  of	
  the	
  size	
  of	
  entities—e.g.,	
  molecules,	
  cells,	
  organs,	
  organisms	
  (for	
  an	
  
frequently	
  cited	
  example,	
  see	
  Churchland	
  and	
  Sejnowski	
  1992).	
  In	
  the	
  context	
  of	
  
discussions	
  of	
  top-­‐down	
  causation,	
  neither	
  of	
  these	
  senses	
  brings	
  out	
  what	
  many	
  find	
  to	
  be	
  
problematic.	
  There	
  is	
  nothing	
  problematic	
  with	
  objects	
  studied	
  in	
  one	
  discipline	
  having	
  
causal	
  effects	
  on	
  those	
  studied	
  in	
  other	
  disciplines	
  or	
  for	
  large	
  entities	
  to	
  have	
  causal	
  
effects	
  on	
  small	
  entities.	
  Craver	
  and	
  I	
  restrict	
  our	
  account	
  to	
  mechanistic	
  levels	
  in	
  which	
  a	
  
mechanism	
  consists	
  of	
  appropriately	
  organized	
  parts	
  performing	
  operations.	
  The	
  notion	
  of	
  
a	
  mechanism	
  causing	
  the	
  state	
  of	
  its	
  parts	
  does	
  bring	
  out	
  what	
  many	
  have	
  found	
  
problematic	
  in	
  talk	
  of	
  top-­‐down	
  causation	
  since	
  the	
  state	
  of	
  the	
  parts	
  and	
  the	
  state	
  of	
  the	
  
mechanism	
  are	
  not	
  independent	
  in	
  the	
  sense	
  required	
  for	
  one	
  to	
  cause	
  the	
  other.	
  
	
  
What	
  is	
  crucial	
  to	
  the	
  mechanistic	
  conception	
  of	
  level	
  is	
  the	
  idea	
  of	
  an	
  entity	
  being	
  
constituted	
  by	
  its	
  parts	
  and	
  operations.	
  The	
  notion	
  of	
  constitution	
  is	
  what	
  must	
  be	
  
explicated	
  and	
  section	
  2	
  will	
  highlight	
  this	
  challenge	
  by	
  developing	
  an	
  example	
  biological	
  
mechanism	
  that	
  I	
  will	
  use	
  through	
  the	
  rest	
  of	
  the	
  paper.	
  Then,	
  in	
  section	
  3,	
  I	
  introduce	
  a	
  
framework	
  for	
  thinking	
  about	
  systems	
  that	
  has	
  been	
  extensively	
  applied	
  in	
  systems	
  biology	
  
(Barabasi	
  and	
  Oltvai	
  2004)	
  and	
  neuroscience	
  (Sporns	
  2010):	
  graph-­‐theoretic	
  
representations	
  of	
  networks.	
  Appealing	
  to	
  graph	
  theory	
  may	
  appear	
  paradoxical	
  as	
  graph-­‐
theoretic	
  representations	
  do	
  not	
  explicitly	
  advert	
  to	
  levels—all	
  nodes	
  are	
  represented	
  on	
  a	
  
plane.	
  Many	
  graphs	
  that	
  characterize	
  biological	
  systems,	
  though,	
  involve	
  modules	
  in	
  which	
  
there	
  is	
  high	
  clustering	
  of	
  nodes	
  often	
  around	
  one	
  or	
  more	
  hubs.	
  Many	
  of	
  these	
  modules,	
  on	
  
the	
  account	
  I	
  am	
  offering,	
  constitute	
  higher-­‐level	
  mechanisms.	
  One	
  must	
  not	
  only	
  show	
  that	
  
high-­‐level	
  mechanisms	
  can	
  be	
  identified,	
  but	
  that	
  the	
  effects	
  on	
  the	
  parts	
  when	
  the	
  whole	
  
mechanism	
  is	
  affected	
  correspond	
  to	
  what	
  has	
  led	
  to	
  talk	
  of	
  top-­‐down	
  causation.	
  That	
  is,	
  
the	
  condition	
  of	
  the	
  whole	
  mechanism	
  must	
  result	
  in	
  different	
  behavior	
  of	
  the	
  part	
  than	
  
would	
  occur	
  when	
  the	
  conditions	
  in	
  the	
  whole	
  mechanism	
  are	
  different.	
  This	
  requires	
  that	
  
we	
  turn	
  from	
  the	
  structure	
  of	
  modules	
  to	
  their	
  functioning.	
  The	
  clustered	
  nodes	
  in	
  
biological	
  networks	
  are	
  typically	
  not	
  ordered	
  sequentially;	
  rather,	
  they	
  are	
  connected	
  in	
  
such	
  an	
  interconnected	
  manner	
  that	
  one	
  can	
  identify	
  multiple	
  feedback	
  loops.	
  	
  When	
  the	
  
operations	
  corresponding	
  to	
  the	
  edges	
  in	
  graphs	
  are	
  non-­‐linear,	
  interconnected	
  modules	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

can	
  exhibit	
  complex	
  dynamic	
  behavior	
  such	
  as	
  oscillations.	
  As	
  a	
  result	
  of	
  dynamic	
  behavior	
  
within	
  the	
  module,	
  the	
  module	
  itself	
  does	
  not	
  respond	
  to	
  external	
  inputs	
  in	
  the	
  same	
  
manner	
  on	
  all	
  occasions.	
  How	
  it	
  responds	
  depends	
  on	
  the	
  current	
  conditions	
  of	
  the	
  module.	
  
This	
  is	
  a	
  diagnostic	
  symptom	
  of	
  top-­‐down	
  causation.	
  	
  
	
  
The	
  graph	
  representations	
  that	
  I	
  introduce	
  in	
  sections	
  3	
  and	
  4	
  help	
  make	
  it	
  clear	
  when	
  it	
  is	
  
appropriate	
  to	
  identify	
  mechanisms	
  as	
  higher-­‐level	
  components—mechanisms	
  are	
  highly	
  
interactive	
  modules	
  within	
  a	
  larger	
  network	
  that	
  are	
  capable	
  of	
  exhibiting	
  complex	
  
dynamics.	
  But	
  as	
  a	
  result	
  of	
  treating	
  all	
  interactions,	
  those	
  within	
  the	
  module	
  and	
  those	
  
between	
  modules,	
  as	
  edges,	
  the	
  graph	
  representation	
  does	
  not	
  make	
  manifest	
  why	
  the	
  
effects	
  of	
  mechanisms	
  on	
  their	
  constituent	
  components	
  are	
  different	
  from	
  the	
  propagation	
  
of	
  causal	
  effects	
  throughout	
  the	
  network.	
  Instead,	
  the	
  graph	
  representation	
  may	
  reinforce	
  
the	
  perception	
  that	
  all	
  causation	
  is	
  at	
  the	
  lowest	
  level	
  represented	
  by	
  individual	
  nodes	
  in	
  
the	
  graph.	
  To	
  address	
  this	
  issue,	
  in	
  section	
  5	
  I	
  will	
  deploy	
  a	
  distinction	
  between	
  constraints	
  
and	
  dynamical	
  laws	
  that	
  has	
  been	
  introduced	
  into	
  theoretical	
  biology	
  from	
  physics	
  (Hooker	
  
2013;	
  Pattee	
  1971).	
  Constraints	
  reduce	
  the	
  degrees	
  of	
  freedom	
  that	
  are	
  left	
  open	
  by	
  
dynamical	
  laws	
  alone	
  by,	
  for	
  example,	
  establishing	
  correlations	
  between	
  variables	
  or	
  
restricting	
  the	
  range	
  of	
  variables.	
  Although	
  the	
  imposition	
  of	
  constraints	
  and	
  the	
  
propagation	
  of	
  effects	
  of	
  constrained	
  systems	
  involve	
  diachronic	
  activity,	
  the	
  constraints	
  
exercised	
  by	
  a	
  whole	
  mechanism	
  on	
  its	
  parts	
  are	
  synchronic—they	
  are	
  realized	
  through	
  the	
  
organizational	
  relationships	
  between	
  the	
  components	
  at	
  a	
  time.	
  Given	
  the	
  organization	
  of	
  
the	
  mechanism,	
  the	
  responses	
  of	
  components	
  that	
  are	
  altered	
  when	
  the	
  mechanism	
  itself	
  is	
  
affected	
  by	
  external	
  causes	
  are	
  constrained	
  by	
  the	
  whole.	
  	
  
	
  
2.	
  When	
  Top-­‐Down	
  Causation	
  Seems	
  Problematic:	
  Levels	
  of	
  Mechanisms	
  	
  
	
  
The	
  notion	
  of	
  mechanism	
  invoked	
  in	
  the	
  conception	
  of	
  levels	
  I	
  am	
  addressing	
  arises	
  from	
  
the	
  practice	
  of	
  biologists	
  over	
  the	
  past	
  several	
  centuries.	
  In	
  the	
  life	
  sciences,	
  investigators	
  
developing	
  explanations	
  often	
  (1)	
  begin	
  by	
  identifying	
  the	
  mechanism	
  responsible	
  for	
  a	
  
specific	
  phenomenon	
  to	
  be	
  explained,	
  (2)	
  proceed	
  to	
  decompose	
  the	
  mechanism	
  into	
  its	
  
parts	
  and	
  the	
  operations	
  they	
  perform,	
  and	
  (3)	
  finally	
  recompose	
  the	
  mechanism	
  to	
  show	
  
how,	
  as	
  a	
  result	
  of	
  the	
  organized	
  parts	
  orchestrating	
  their	
  operations,	
  the	
  mechanism	
  
generates	
  the	
  phenomenon.	
  While	
  this	
  practice	
  has	
  been	
  pursued	
  for	
  several	
  centuries,	
  it	
  
has	
  assumed	
  a	
  central	
  place	
  in	
  philosophical	
  accounts	
  of	
  explanation	
  in	
  the	
  last	
  couple	
  
decades	
  (Bechtel	
  and	
  Richardson	
  1993/2010;	
  Bechtel	
  2006,	
  2008;	
  Machamer,	
  Darden,	
  and	
  
Craver	
  2000).	
  The	
  steps	
  of	
  decomposition	
  and	
  recomposition	
  are	
  what	
  invite	
  employing	
  
the	
  word	
  levels.	
  Parts	
  are,	
  in	
  a	
  rather	
  natural	
  sense,	
  at	
  a	
  lower	
  level	
  than	
  the	
  mechanism	
  
constructed	
  out	
  of	
  them.	
  	
  
	
  
This	
  conception	
  of	
  level	
  can	
  be	
  illustrated	
  using	
  an	
  example	
  to	
  which	
  I	
  will	
  return	
  
throughout	
  this	
  paper,	
  the	
  phenomenon	
  of	
  circadian	
  rhythmicity	
  in	
  mammals	
  (i.e.,	
  daily	
  
oscillations	
  in	
  behaviors	
  and	
  physiological	
  functions	
  that	
  are	
  endogenously	
  generated	
  but	
  
entrainable	
  to	
  day-­‐night	
  cycles	
  in	
  the	
  local	
  environment).	
  The	
  responsible	
  mechanism	
  
resides	
  in	
  individual	
  cells.	
  The	
  genes	
  Per	
  and	
  Cry	
  and	
  the	
  proteins	
  synthesized	
  from	
  them,	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

PER	
  and	
  CRY,	
  are	
  major	
  parts.1	
  PER	
  and	
  CRY	
  form	
  a	
  dimer	
  and,	
  after	
  being	
  transported	
  
back	
  into	
  the	
  nucleus,	
  inhibit	
  their	
  own	
  transcription	
  by	
  interfering	
  with	
  the	
  activators,	
  
BMAL1	
  and	
  CLOCK.2	
  These	
  genes	
  and	
  proteins	
  occupy	
  a	
  lower	
  level	
  than	
  the	
  mechanism	
  
itself.	
  	
  
	
  
Using	
  this	
  example,	
  we	
  can	
  now	
  see	
  what	
  is	
  sometimes	
  regarded	
  as	
  problematic	
  about	
  top-­‐
down	
  causation.	
  Whether	
  Per	
  and	
  Cry	
  are	
  transcribed	
  and	
  translated	
  into	
  the	
  proteins	
  PER	
  
and	
  CRY	
  depends	
  upon	
  the	
  phase	
  of	
  oscillation	
  the	
  host	
  cell	
  is	
  in.	
  The	
  oscillatory	
  phase,	
  a	
  
state	
  of	
  the	
  whole	
  cell,	
  is	
  determining	
  the	
  behavior	
  of	
  its	
  parts,	
  seemingly	
  in	
  accord	
  with	
  
accounts	
  of	
  top-­‐down	
  causation.	
  However,	
  the	
  phase	
  of	
  the	
  oscillator	
  at	
  a	
  time	
  just	
  is	
  the	
  
concentrations	
  of	
  PER,	
  CRY,	
  and	
  a	
  cadre	
  of	
  related	
  proteins	
  that	
  make	
  up	
  the	
  mechanism.	
  
Treating	
  the	
  concentrations	
  as	
  causing	
  the	
  phase,	
  or	
  vice	
  versa,	
  seems	
  to	
  violate	
  many	
  
intuitive	
  aspects	
  of	
  causation.	
  Causation	
  is	
  often	
  understood	
  as	
  involving	
  contact	
  action	
  or	
  a	
  
propagated	
  signal	
  (Hitchcock	
  2003).	
  Moreover,	
  causes	
  are	
  assumed	
  to	
  precede	
  their	
  
effects.	
  Both	
  conditions	
  require	
  that	
  causes	
  and	
  effects	
  be	
  wholly	
  distinct	
  (see	
  Lewis	
  2000,	
  ,	
  
for	
  detailed	
  arguments	
  as	
  to	
  why	
  causes	
  and	
  effects	
  must	
  be	
  distinct).	
  The	
  phase	
  of	
  the	
  
oscillation	
  in	
  the	
  cell	
  is	
  not	
  distinct	
  from	
  the	
  concentrations	
  of	
  PER	
  and	
  CRY.	
  More	
  
generally,	
  the	
  parts	
  and	
  wholes	
  of	
  a	
  mechanism	
  are	
  not	
  distinct—each	
  requires	
  the	
  
existence	
  of	
  the	
  other.	
  There	
  is	
  no	
  possibility	
  for	
  transmission	
  between	
  parts	
  and	
  wholes	
  
since	
  they	
  are	
  not	
  distinct.	
  Neither	
  the	
  states	
  of	
  the	
  parts	
  nor	
  the	
  state	
  of	
  the	
  whole	
  come	
  
before	
  the	
  other—they	
  occur	
  simultaneously.	
  Talking	
  of	
  top-­‐down	
  causation	
  in	
  these	
  cases	
  
also	
  seems	
  to	
  engender	
  the	
  problem	
  of	
  redundant	
  causation	
  highlighted	
  by	
  Kim	
  (1998).	
  
While	
  we	
  might	
  attribute	
  the	
  change	
  in	
  concentration	
  in	
  PER	
  or	
  CRY	
  to	
  the	
  phase	
  of	
  the	
  
oscillator,	
  it	
  can	
  also	
  be	
  attributed	
  to	
  specific	
  molecular	
  events,	
  viz.,	
  that	
  the	
  proteins	
  have	
  
attached	
  themselves	
  to	
  two	
  other	
  proteins,	
  BMAL	
  and	
  CLOCK,	
  and	
  removed	
  those	
  proteins	
  
from	
  the	
  E-­‐boxes	
  on	
  the	
  Per	
  and	
  Cry	
  genes,	
  terminating	
  the	
  synthesis	
  of	
  more	
  PER	
  and	
  CRY.	
  	
  
	
  
Craver	
  and	
  I	
  sought	
  to	
  make	
  a	
  virtue	
  out	
  of	
  these	
  problems	
  for	
  top-­‐down	
  (or	
  comparable	
  
problems	
  for	
  bottom-­‐up)	
  causation	
  by	
  treating	
  the	
  inter-­‐level	
  relation	
  as	
  constitution	
  and	
  
noting	
  that	
  a	
  state	
  of	
  the	
  whole	
  mechanism	
  involves	
  at	
  least	
  some	
  of	
  its	
  parts	
  being	
  in	
  
appropriate	
  states.	
  Any	
  phenomenon	
  that	
  seems	
  to	
  exhibit	
  bottom-­‐up	
  or	
  top-­‐down	
  
causation	
  can	
  be	
  accommodated	
  by	
  viewing	
  causes	
  as	
  operating	
  either	
  (1)	
  between	
  the	
  
whole	
  mechanism	
  and	
  other	
  entities	
  outside	
  it	
  or	
  (2)	
  among	
  the	
  parts	
  within	
  the	
  
mechanism.	
  The	
  fact	
  that	
  the	
  parts	
  constitute	
  the	
  whole	
  has	
  the	
  consequence	
  that	
  the	
  state	
  
of	
  the	
  whole	
  mechanism	
  is	
  changed	
  whenever	
  the	
  state	
  of	
  one	
  of	
  its	
  parts	
  is	
  changed	
  and	
  
the	
  state	
  of	
  at	
  least	
  one	
  of	
  the	
  parts	
  is	
  changed	
  whenever	
  the	
  state	
  of	
  the	
  whole	
  is	
  changed.	
  
Craver	
  and	
  Bechtel	
  referred	
  to	
  how	
  the	
  constitution	
  relation	
  mediates	
  between	
  changes	
  to	
  
the	
  parts	
  and	
  changes	
  to	
  the	
  whole	
  as	
  yielding	
  mechanistically	
  mediated	
  effects.	
  	
  

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
1	
  There	
  are	
  two	
  paralogs	
  of	
  PER	
  and	
  CRY	
  in	
  the	
  mammalian	
  clock	
  mechanism,	
  but	
  since	
  
they	
  function	
  in	
  similar	
  ways,	
  I	
  am	
  omitting	
  this	
  detail.	
  	
  
2	
  There	
  is	
  now	
  some	
  doubt	
  as	
  to	
  whether	
  the	
  mechanism	
  requires	
  proteins	
  inhibiting	
  their	
  
own	
  transcription;	
  on	
  some	
  proposals,	
  the	
  oscillator	
  may	
  be	
  post-­‐translational.	
  For	
  
purposes	
  of	
  this	
  paper,	
  I	
  will	
  assume	
  that	
  the	
  core	
  mechanism	
  involves	
  transcription-­‐
translation	
  feedback	
  loops.	
  	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
A	
  major	
  shortcoming	
  in	
  Craver	
  and	
  my	
  account	
  is	
  that	
  by	
  merely	
  pointing	
  to	
  a	
  constitution	
  
relation	
  we	
  left	
  unspecified	
  what	
  it	
  is	
  for	
  parts	
  to	
  constitute	
  a	
  mechanism	
  such	
  that	
  the	
  
mechanism	
  is	
  at	
  a	
  higher	
  level.	
  The	
  need	
  to	
  address	
  this	
  question	
  can	
  be	
  seen	
  by	
  examining	
  
the	
  three-­‐tier	
  diagram	
  Craver	
  (2007)	
  uses	
  to	
  present	
  his	
  view	
  of	
  levels.	
  At	
  the	
  top	
  an	
  arrow	
  
on	
  the	
  left	
  terminates	
  at	
  a	
  darkened	
  oval	
  that	
  represents	
  a	
  mechanism.	
  Another	
  arrow	
  
leaves	
  from	
  the	
  mechanism.	
  This	
  is	
  intended	
  to	
  show	
  that	
  when	
  the	
  mechanism	
  receives	
  an	
  
input,	
  it	
  generates	
  an	
  output.	
  Dotted	
  lines	
  connect	
  that	
  oval	
  to	
  the	
  one	
  below,	
  indicating	
  
that	
  what	
  is	
  below	
  is	
  an	
  expansion	
  of	
  what	
  is	
  above.	
  In	
  this	
  middle	
  tier,	
  the	
  arrow	
  on	
  the	
  
left	
  is	
  shown	
  as	
  entering	
  into	
  the	
  mechanism	
  to	
  terminate	
  at	
  one	
  of	
  the	
  ovals	
  within	
  it.	
  The	
  
four	
  ovals	
  within	
  are	
  connected	
  by	
  arrows,	
  culminating	
  in	
  one	
  from	
  which	
  the	
  arrow	
  on	
  the	
  
right	
  now	
  exits.	
  One	
  of	
  the	
  ovals	
  in	
  the	
  mechanism	
  is	
  again	
  shown	
  in	
  black,	
  and	
  it	
  is	
  
expanded	
  in	
  the	
  third	
  tier	
  in	
  the	
  same	
  manner.	
  The	
  figure	
  makes	
  clear	
  the	
  sense	
  in	
  which	
  
this	
  account	
  of	
  mechanisms	
  is	
  supposed	
  to	
  give	
  rise	
  to	
  a	
  view	
  of	
  levels—the	
  inner	
  ovals	
  in	
  
tier	
  two	
  are	
  inside	
  the	
  larger	
  oval	
  which	
  is	
  then	
  shown	
  in	
  black	
  above	
  it	
  on	
  the	
  upper	
  tier.	
  
The	
  question	
  the	
  diagram	
  poses	
  is:	
  what	
  do	
  the	
  ovals	
  represent?	
  Put	
  another	
  way,	
  why	
  is	
  
the	
  oval	
  in	
  the	
  middle	
  tier	
  around	
  all	
  four	
  inner	
  ovals,	
  and	
  not	
  some	
  other	
  possible	
  
combination	
  of	
  ovals	
  (e.g.,	
  just	
  two	
  of	
  the	
  inner	
  ovals	
  in	
  tier	
  2,	
  or	
  one	
  of	
  these	
  and	
  another	
  
outside	
  of	
  the	
  oval	
  shown)?	
  	
  

	
  
Figure	
  1.	
  Craver-­‐style	
  representation	
  of	
  three	
  levels	
  of	
  a	
  mechanism.	
  Adapted	
  from	
  Craver	
  (2007).	
  

Motivating	
  the	
  ovals	
  in	
  Figure	
  1	
  requires	
  an	
  answer	
  to	
  the	
  original	
  question:	
  when	
  do	
  
components	
  constitute	
  a	
  mechanism?	
  A	
  mechanism	
  is	
  not	
  just	
  an	
  aggregation	
  of	
  parts,	
  each	
  
of	
  which	
  performs	
  an	
  operation.	
  In	
  their	
  characterization	
  of	
  mechanisms,	
  Machamer,	
  
Darden	
  and	
  Craver	
  (2000)	
  emphasize	
  productive	
  continuity	
  between	
  the	
  parts	
  involved	
  in	
  
generating	
  the	
  phenomenon.	
  This	
  is	
  critical,	
  but	
  not	
  sufficient.	
  A	
  pond	
  is	
  a	
  collection	
  of	
  
entities,	
  many	
  performing	
  operations,	
  with	
  productive	
  continuity	
  between	
  operations.	
  But	
  
it	
  is	
  not	
  what	
  scientists	
  would	
  call	
  a	
  mechanism.	
  One	
  thing	
  that	
  is	
  central	
  in	
  all	
  accounts	
  of	
  a	
  
mechanism	
  is	
  that	
  the	
  parts	
  are	
  those	
  entities	
  whose	
  activities	
  or	
  operations	
  are	
  
responsible	
  for	
  the	
  phenomenon	
  attributed	
  to	
  the	
  whole	
  mechanism.	
  In	
  the	
  mechanism	
  
literature,	
  the	
  specification	
  of	
  the	
  phenomenon	
  to	
  be	
  explained	
  is	
  often	
  invoked	
  to	
  
determine	
  which	
  entities	
  belong	
  to	
  a	
  mechanism	
  and	
  which,	
  even	
  though	
  they	
  are	
  located	
  
among	
  the	
  other	
  entities,	
  do	
  not.	
  With	
  respect	
  to	
  the	
  pond,	
  if	
  one	
  identifies	
  a	
  phenomenon,	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

such	
  as	
  maintenance	
  of	
  pH,	
  then	
  one	
  can	
  search	
  for	
  the	
  mechanism—the	
  productively	
  
linked	
  operations	
  of	
  entities	
  in	
  the	
  pond	
  that	
  contribute	
  to	
  that	
  phenomenon.	
  	
  
	
  
As	
  useful	
  as	
  appealing	
  to	
  a	
  phenomenon	
  to	
  be	
  explained	
  is	
  in	
  identifying	
  candidate	
  parts	
  of	
  
a	
  mechanism,	
  it	
  is	
  insufficient	
  to	
  fix	
  the	
  boundaries	
  and	
  hence	
  pick	
  out	
  a	
  mechanism	
  as	
  a	
  
higher-­‐level	
  entity	
  existing	
  as	
  such	
  in	
  the	
  world.	
  One	
  reason	
  is	
  that	
  the	
  same	
  entity	
  may	
  be	
  
involved	
  in	
  the	
  production	
  of	
  several	
  different	
  phenomena.	
  To	
  accommodate	
  this,	
  we	
  
require	
  an	
  account	
  in	
  which	
  a	
  mechanism	
  can	
  share	
  parts	
  with	
  other	
  mechanisms.	
  Another	
  
reason	
  is	
  that	
  the	
  range	
  of	
  entities	
  or	
  activities	
  that	
  can	
  affect	
  a	
  given	
  phenomenon	
  is	
  not	
  
sharply	
  bounded.	
  Often	
  entities	
  very	
  distant	
  in	
  time	
  and	
  space	
  can	
  play	
  critical	
  roles	
  in	
  the	
  
generation	
  of	
  a	
  phenomenon	
  (Bechtel	
  2015).	
  The	
  first	
  entity	
  in	
  tier	
  two	
  of	
  Craver’s	
  diagram	
  
has	
  an	
  arrow	
  coming	
  into	
  it	
  from	
  somewhere	
  else,	
  but	
  that	
  entity	
  is,	
  for	
  some	
  unexplained	
  
reason,	
  not	
  counted	
  as	
  part	
  of	
  the	
  mechanism.	
  	
  
	
  
The	
  issue	
  of	
  where	
  to	
  draw	
  boundaries	
  around	
  a	
  mechanism	
  is	
  in	
  fact	
  a	
  crucial	
  issue	
  in	
  
biology.	
  As	
  research	
  proceeded	
  on	
  circadian	
  rhythms	
  researchers	
  identified	
  multiple	
  
feedback	
  loops	
  in	
  addition	
  to	
  that	
  involving	
  PER	
  and	
  CRY.	
  For	
  example,	
  BMAL1,	
  which	
  
serves	
  as	
  an	
  activator	
  of	
  Per	
  and	
  Cry	
  transcription,	
  is	
  produced	
  by	
  a	
  feedback	
  loop	
  in	
  which	
  
its	
  synthesis	
  is	
  regulated	
  by	
  the	
  nuclear	
  receptors	
  RORα	
  and	
  REV-­‐ERBα,	
  while	
  it	
  together	
  
with	
  CLOCK	
  are	
  activators	
  of	
  their	
  synthesis.	
  Figure	
  2	
  presents	
  the	
  conception	
  of	
  the	
  core	
  
clock	
  mechanism	
  as	
  it	
  was	
  understood	
  around	
  2005.	
  It	
  presents	
  the	
  circadian	
  clock	
  
mechanism	
  as	
  a	
  well-­‐bounded	
  set	
  of	
  components.	
  (The	
  figure	
  itself	
  does	
  not	
  show	
  any	
  
inputs	
  or	
  outputs,	
  but	
  in	
  fact	
  there	
  are	
  input	
  signals	
  that	
  serve	
  to	
  entrain	
  the	
  clock	
  to	
  the	
  
day/night	
  cycle	
  and	
  output	
  signals	
  that	
  serve	
  to	
  regulate	
  the	
  expression	
  of	
  a	
  wide	
  variety	
  of	
  
other	
  genes.)	
  Since	
  then	
  the	
  emergence	
  of	
  new	
  techniques,	
  such	
  as	
  knocking	
  down	
  
expression	
  of	
  genes	
  through	
  use	
  of	
  small	
  interfering	
  RNAs,	
  has	
  revealed	
  more	
  than	
  300	
  
additional	
  genes	
  that	
  are	
  both	
  expressed	
  in	
  a	
  circadian	
  manner	
  and	
  exert	
  effects	
  on	
  the	
  
phase	
  or	
  amplitude	
  of	
  circadian	
  rhythms	
  (Zhang	
  et	
  al.	
  2009).	
  Many	
  of	
  these	
  genes	
  were	
  
already	
  identified	
  as	
  components	
  of	
  other	
  cellular	
  mechanisms.	
  The	
  last	
  dozen	
  years	
  has	
  
revealed	
  extensive	
  interactions,	
  for	
  example,	
  between	
  components	
  of	
  basic	
  metabolic	
  
mechanisms	
  and	
  core	
  components	
  of	
  the	
  circadian	
  clock.	
  The	
  questions	
  of	
  where	
  to	
  draw	
  
the	
  boundary	
  of	
  the	
  clock	
  mechanism	
  and	
  why	
  to	
  draw	
  it	
  there	
  have	
  assumed	
  prominence	
  
(for	
  examples,	
  see	
  Bechtel	
  2015).	
  If	
  Craver’s	
  account	
  is	
  to	
  provide	
  a	
  principled	
  
characterization	
  of	
  entities	
  in	
  terms	
  of	
  mechanistic	
  levels,	
  one	
  needs	
  a	
  procedure	
  for	
  
limning	
  the	
  boundaries	
  of	
  the	
  mechanism	
  and	
  distinguishing	
  it	
  from	
  other	
  entities	
  that	
  
affect	
  the	
  phenomenon	
  but	
  are	
  not	
  identified	
  as	
  parts	
  of	
  the	
  mechanism.	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
Figure	
  2.	
  The	
  major	
  parts	
  and	
  operations	
  of	
  the	
  mammalian	
  circadian	
  clock	
  as	
  
understood	
  circa	
  2005.	
  Not	
  shown	
  are	
  the	
  various	
  kinases	
  that	
  phosphorylate	
  the	
  
proteins	
  and	
  determine	
  whether	
  they	
  are	
  broken	
  down	
  or	
  transported	
  into	
  the	
  
nucleus.	
  	
  

	
  
3.	
  Flattening	
  Levels:	
  Using	
  Graphs	
  to	
  Identify	
  Mechanisms	
  
	
  
I	
  will	
  return	
  to	
  the	
  question	
  of	
  what	
  the	
  ovals	
  represent	
  in	
  Craver’s	
  diagram	
  below.	
  First,	
  I	
  
turn	
  to	
  a	
  second	
  question	
  posed	
  by	
  his	
  diagram.	
  At	
  the	
  top	
  level,	
  arrows	
  just	
  contact	
  the	
  
black	
  oval	
  representing	
  the	
  mechanism.	
  At	
  the	
  middle	
  tier,	
  the	
  corresponding	
  arrows	
  
contact	
  the	
  smaller	
  ovals	
  inside	
  the	
  larger	
  oval.	
  At	
  the	
  lowest	
  tier,	
  the	
  arrows	
  penetrate	
  the	
  
larger	
  oval	
  that	
  corresponds	
  to	
  one	
  of	
  those	
  small	
  ovals	
  and	
  contacts	
  yet	
  inner	
  ovals.	
  What	
  
the	
  penetration	
  of	
  arrows	
  into	
  the	
  mechanism	
  at	
  lower	
  tiers	
  suggests	
  is	
  that	
  the	
  causal	
  
effect	
  is	
  not	
  on	
  the	
  higher-­‐level	
  unit	
  as	
  a	
  whole,	
  but	
  on	
  one	
  (or	
  possibly	
  several)	
  of	
  its	
  
components.	
  If	
  in	
  the	
  bottom	
  tier,	
  instead	
  of	
  expanding	
  an	
  inner	
  oval,	
  Craver	
  had	
  chosen	
  to	
  
expand	
  the	
  initial	
  oval	
  from	
  the	
  middle	
  tier,	
  the	
  point	
  would	
  have	
  been	
  even	
  clearer.	
  Then	
  
on	
  the	
  three	
  tiers	
  the	
  arrow	
  representing	
  input	
  to	
  the	
  mechanism	
  would	
  connect,	
  
respectively,	
  with	
  the	
  highest-­‐level	
  oval	
  (the	
  mechanism),	
  one	
  of	
  the	
  inner	
  ovals	
  (a	
  
component	
  of	
  the	
  mechanism),	
  and	
  one	
  of	
  the	
  ovals	
  within	
  that	
  oval	
  (a	
  component	
  of	
  the	
  
component).	
  At	
  the	
  lowest	
  tier,	
  the	
  arrows	
  in	
  and	
  out	
  are	
  no	
  different	
  than	
  those	
  between	
  
components	
  and	
  the	
  idea	
  of	
  causation	
  at	
  multiple	
  levels	
  seems	
  to	
  be	
  lost.	
  In	
  many	
  cases,	
  
such	
  reduction	
  to	
  lower	
  levels	
  might	
  seem	
  to	
  be	
  appropriate—when	
  light	
  exposure	
  
entrains	
  circadian	
  rhythms,	
  photons	
  affect	
  the	
  melanopsin	
  molecules	
  in	
  the	
  intrinsically	
  
photoreceptive	
  retinal	
  ganglion	
  cells.	
  As	
  a	
  result,	
  when	
  neurotransmitters	
  are	
  released	
  they	
  
initiate	
  a	
  signaling	
  cascade	
  within	
  cells	
  in	
  the	
  suprachiasmatic	
  nucleus,	
  resulting	
  in	
  
increased	
  Per	
  transcription.	
  All	
  of	
  these	
  events	
  seem	
  to	
  be	
  at	
  a	
  single	
  level.	
  But	
  then	
  in	
  what	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

sense	
  is	
  the	
  clock,	
  a	
  higher-­‐level	
  entity,	
  entrained?	
  And	
  how	
  does	
  its	
  phase	
  as	
  a	
  result	
  of	
  
entrainment,	
  affect	
  the	
  behavior	
  of	
  its	
  components?	
  
	
  
This	
  exegesis	
  of	
  Craver’s	
  diagram	
  suggests	
  that	
  the	
  critics	
  who	
  viewed	
  Craver	
  and	
  my	
  
account	
  as	
  rendering	
  higher	
  levels	
  epiphenomenal	
  were	
  right.	
  It	
  suggests	
  a	
  highly	
  
reductionistic	
  picture	
  of	
  levels	
  according	
  to	
  which	
  causal	
  relations	
  that	
  were	
  supposed	
  to	
  
be	
  between	
  entities	
  at	
  higher	
  levels	
  of	
  organization	
  dissolve	
  into	
  causal	
  interactions	
  at	
  the	
  
lowest	
  level	
  considered.	
  To	
  bring	
  out	
  this	
  point,	
  I	
  have	
  presented	
  a	
  flattened	
  representation	
  
of	
  Craver’s	
  diagram	
  in	
  Figure	
  3.	
  It	
  preserves	
  the	
  relationships	
  between	
  components	
  at	
  the	
  
middle	
  and	
  lowest	
  tier	
  in	
  Craver’s	
  diagram,	
  and	
  fills	
  in	
  possible	
  decompositions	
  of	
  the	
  three	
  
ovals	
  in	
  the	
  middle	
  tier	
  that	
  were	
  not	
  expanded	
  in	
  Craver’s	
  diagram.	
  To	
  identify	
  what	
  were	
  
supposed	
  to	
  be	
  higher-­‐levels	
  entities,	
  I	
  have	
  included	
  dotted	
  ovals	
  around	
  the	
  lower-­‐level	
  
ovals	
  and	
  a	
  dashed	
  oval	
  around	
  the	
  whole	
  set.	
  	
  	
  

	
  
Figure	
  3.	
  A	
  flattened	
  redrawing	
  and	
  elaboration	
  of	
  Craver’s	
  diagram,	
  with	
  dotted	
  ovals	
  
grouping	
  nodes	
  that	
  correspond	
  to	
  units	
  at	
  the	
  middle	
  and	
  top	
  level	
  in	
  Craver’s	
  diagram.	
  

	
  
If	
  one	
  ignores	
  the	
  dotted	
  ovals,	
  Figure	
  3	
  corresponds	
  to	
  a	
  graph-­‐theoretic	
  representation	
  of	
  
a	
  network.	
  A	
  graph	
  representation	
  uses	
  nodes	
  (ovals)	
  to	
  represent	
  entities	
  and	
  edges	
  to	
  
represent	
  relations	
  (perhaps	
  causal)	
  between	
  nodes.	
  In	
  constructing	
  a	
  graph	
  
representation,	
  one	
  has	
  to	
  settle	
  on	
  which	
  entities	
  are	
  to	
  be	
  represented	
  as	
  nodes.	
  Often	
  in	
  
systems	
  biology	
  the	
  nodes	
  represent	
  relatively	
  low-­‐level	
  entities	
  such	
  as	
  genes	
  or	
  proteins.	
  
Here	
  I	
  have	
  simply	
  taken	
  the	
  entities	
  in	
  the	
  bottom	
  level	
  in	
  Craver’s	
  diagram	
  as	
  nodes.	
  This	
  
does	
  not	
  presuppose	
  that	
  there	
  is	
  a	
  lowest	
  level	
  (or,	
  as	
  I	
  will	
  discuss	
  below,	
  that	
  all	
  entities	
  
are	
  at	
  the	
  same	
  level).	
  As	
  inquiry	
  proceeds,	
  researchers	
  may	
  elect	
  to	
  treat	
  the	
  entity	
  
represented	
  as	
  a	
  node	
  as	
  a	
  mechanism	
  and	
  replace	
  the	
  node	
  with	
  a	
  set	
  of	
  nodes	
  
constituting	
  its	
  constituents.	
  	
  Edges	
  may	
  represent	
  either	
  structural	
  connections	
  or	
  
functional	
  connections	
  (for	
  purposes	
  of	
  this	
  paper	
  I	
  will	
  assume	
  they	
  are	
  functional).	
  In	
  
graph	
  representations	
  more	
  generally,	
  the	
  edges	
  in	
  a	
  graph	
  may	
  be	
  directed	
  or	
  undirected,	
  
but	
  in	
  a	
  graph	
  representation	
  of	
  a	
  mechanism,	
  the	
  edges	
  are	
  directed	
  and	
  represent	
  
operations	
  in	
  which	
  one	
  node	
  exercises	
  a	
  causal	
  effect	
  on	
  another.	
  
	
  
By	
  showing	
  all	
  edges	
  as	
  between	
  nodes,	
  the	
  graph	
  representation	
  brings	
  out	
  what	
  is	
  
challenging	
  in	
  explicating	
  top-­‐down	
  or	
  bottom-­‐up	
  causation—there	
  does	
  not	
  seem	
  to	
  be	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

any	
  principled	
  criterion	
  for	
  identifying	
  levels.	
  The	
  dotted	
  and	
  dashed	
  ovals	
  that	
  correspond	
  
to	
  entities	
  at	
  higher	
  levels	
  in	
  Craver’s	
  diagram	
  appear	
  to	
  be	
  purely	
  arbitrary	
  impositions	
  on	
  
the	
  graph	
  representation.	
  One	
  could	
  draw	
  ovals	
  that	
  group	
  nodes	
  in	
  different	
  ways.	
  
Moreover,	
  since	
  they	
  are	
  not	
  the	
  endpoint	
  of	
  edges,	
  the	
  dotted	
  and	
  dashed	
  ovals	
  or	
  
whatever	
  they	
  represent	
  seem	
  to	
  be	
  causally	
  inert.	
  Rather	
  than	
  providing	
  an	
  account	
  of	
  
causation	
  at	
  multiple	
  levels,	
  it	
  appears	
  that	
  all	
  higher	
  levels	
  appear	
  to	
  have	
  been	
  rendered	
  
epiphenomenal.	
  	
  
	
  
I	
  will	
  argue,	
  however,	
  that	
  there	
  is	
  a	
  way	
  to	
  use	
  graph-­‐theoretic	
  analysis	
  to	
  identify	
  
structures	
  within	
  networks	
  that	
  correspond	
  to	
  the	
  sort	
  of	
  entities	
  traditionally	
  viewed	
  as	
  
residing	
  at	
  higher	
  levels	
  and	
  to	
  show	
  in	
  what	
  respect	
  these	
  entities	
  constrain	
  causal	
  
processes.	
  To	
  do	
  so,	
  I	
  need	
  to	
  introduce	
  three	
  measures	
  graph	
  theorists	
  have	
  introduced	
  to	
  
characterize	
  graphs:	
  

1. Mean	
  shortest	
  path-­‐length	
  is	
  a	
  measure	
  of	
  the	
  average	
  number	
  of	
  edges	
  that	
  must	
  
be	
  traversed	
  on	
  the	
  shortest	
  path	
  between	
  two	
  nodes.	
  	
  

2. The	
  clustering	
  coefficient	
  is	
  a	
  measure	
  of	
  how	
  connected	
  to	
  each	
  other	
  the	
  nodes	
  
linked	
  to	
  a	
  given	
  node	
  are.	
  	
  

3. Degree	
  distribution	
  is	
  a	
  measure	
  of	
  how	
  the	
  number	
  of	
  connections	
  from	
  
different	
  nodes	
  is	
  varied.	
  	
  

	
  
In	
  terms	
  of	
  these	
  measures,	
  one	
  can	
  describe	
  different	
  network	
  topologies.	
  A	
  randomly	
  
connected	
  network	
  will	
  have	
  a	
  small	
  mean	
  shortest	
  path-­‐length	
  (there	
  are,	
  on	
  average,	
  
short	
  paths	
  between	
  any	
  two	
  units),	
  allowing	
  for	
  rapid	
  transmission	
  between	
  nodes,	
  but	
  
exhibit	
  little	
  clustering	
  of	
  units.	
  Lattice	
  or	
  near-­‐neighbor	
  structures	
  exhibit	
  high	
  clustering,	
  
allowing	
  nodes	
  to	
  combine	
  operations	
  to	
  produce	
  collective	
  effects,	
  but	
  have	
  long	
  mean	
  
shortest	
  path-­‐lengths.	
  Highly	
  clustered	
  units	
  are	
  often	
  referred	
  to	
  as	
  modules.	
  Networks	
  
that	
  Watts	
  and	
  Strogratz	
  (1998)	
  designated	
  small	
  worlds	
  retain	
  the	
  short	
  mean	
  shortest	
  
path-­‐length	
  of	
  random	
  networks	
  but	
  contain	
  modules	
  of	
  clustered	
  units	
  more	
  typical	
  of	
  
near-­‐neighbor	
  networks.	
  Figure	
  4	
  presents	
  a	
  toy	
  example	
  network	
  in	
  which	
  nodes	
  
represent	
  entities	
  and	
  arrows	
  causal	
  interactions	
  between	
  them.	
  It	
  exhibits	
  small-­‐world	
  
organization—there	
  are	
  distinct	
  modules	
  of	
  interconnected	
  units	
  but	
  a	
  number	
  of	
  
connections	
  linking	
  components	
  in	
  different	
  modules.	
  As	
  discussed	
  further	
  below,	
  the	
  
interconnections	
  within	
  these	
  modules	
  enables	
  them	
  to	
  function	
  as	
  higher-­‐level	
  units	
  in	
  
which	
  the	
  response	
  to	
  an	
  input	
  from	
  outside	
  depends	
  on	
  the	
  state	
  of	
  the	
  module.	
  	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

	
  
Figure	
  4.	
  A	
  graph	
  representation	
  of	
  modules	
  organized	
  around	
  provincial	
  hubs	
  and	
  integrated	
  
through	
  a	
  connector	
  hub.	
  

	
  
When	
  these	
  network	
  topologies	
  were	
  being	
  explored,	
  most	
  researchers	
  assumed	
  that	
  node	
  
degree	
  would	
  be	
  distributed	
  normally.	
  However,	
  Barabási	
  and	
  Albert	
  (1999)	
  found	
  that	
  in	
  
many	
  real	
  world	
  networks,	
  degree	
  distribution	
  corresponds	
  to	
  or	
  approximates	
  a	
  power	
  
law:	
  most	
  units	
  have	
  few	
  connections	
  to	
  other	
  units,	
  but	
  some	
  units	
  are	
  very	
  highly	
  
connected	
  to	
  other	
  units.	
  Networks	
  with	
  power-­‐law	
  degree	
  distribution	
  are	
  referred	
  to	
  as	
  
scale-­‐free	
  as	
  there	
  is	
  no	
  characteristic	
  scale	
  on	
  which	
  to	
  describe	
  the	
  network.	
  Highly	
  
connected	
  nodes	
  are	
  commonly	
  referred	
  to	
  as	
  hubs.	
  Provincial	
  hubs	
  are	
  ones	
  that	
  exhibit	
  
high	
  clustering	
  and	
  so	
  serve	
  as	
  integrators	
  of	
  activity	
  in	
  modules.	
  Connector	
  hubs,	
  on	
  the	
  
other	
  hand,	
  often	
  have	
  low	
  clustering	
  and	
  serve	
  to	
  integrate	
  activity	
  between	
  modules.	
  
Although	
  the	
  number	
  of	
  nodes	
  is	
  too	
  small	
  to	
  produce	
  a	
  truly	
  scale-­‐free	
  network,	
  the	
  graph	
  
in	
  Figure	
  4	
  illustrates	
  how	
  modules	
  can	
  be	
  organized	
  around	
  provincial	
  hubs	
  and	
  
integrated	
  together	
  through	
  a	
  connector	
  hub.	
  	
  
	
  
What	
  is	
  important	
  for	
  present	
  purposes	
  is	
  that	
  the	
  modules	
  in	
  these	
  networks	
  are	
  
differentiated	
  on	
  principled	
  grounds—the	
  nodes	
  within	
  a	
  module	
  are	
  more	
  interconnected	
  
with	
  each	
  other	
  than	
  they	
  are	
  with	
  nodes	
  elsewhere.	
  The	
  extensive	
  interactions	
  between	
  
the	
  nodes	
  constituting	
  a	
  module	
  enable	
  them	
  to	
  work	
  together	
  as	
  a	
  unit.	
  This	
  does	
  not	
  
mean	
  that	
  the	
  connections	
  to	
  components	
  outside	
  the	
  module	
  are	
  not	
  important;	
  inputs	
  
from	
  them	
  may	
  figure	
  critically	
  in	
  regulating	
  some	
  behaviors	
  of	
  the	
  module	
  and	
  enabling	
  
the	
  output	
  of	
  the	
  mechanism	
  to	
  affect	
  other	
  processes.	
  But	
  the	
  enhanced	
  connectivity	
  
enables	
  greater	
  coordination	
  within	
  the	
  module,	
  allowing	
  the	
  components	
  to	
  work	
  together	
  
and	
  justifies	
  treating	
  them	
  as	
  constituting	
  a	
  mechanism	
  when	
  they	
  produce	
  a	
  phenomenon	
  
we	
  seek	
  to	
  explain.	
  To	
  return	
  to	
  Zhang	
  et	
  al.’s	
  (2009)	
  discovery	
  of	
  an	
  additional	
  300	
  genes	
  
that	
  affect	
  circadian	
  rhythms	
  when	
  knocked	
  down,	
  it	
  is	
  noteworthy	
  that	
  they	
  did	
  not	
  view	
  
them	
  as	
  core	
  clock	
  components	
  since	
  they	
  are	
  not	
  nearly	
  as	
  interconnected	
  as	
  the	
  
components	
  shown	
  in	
  Figure	
  2.	
  Rather,	
  they	
  viewed	
  the	
  proteins	
  expressed	
  by	
  these	
  genes	
  
as	
  external	
  factors	
  that	
  interact	
  with	
  the	
  mechanism.	
  One	
  of	
  the	
  interesting	
  consequences	
  
of	
  the	
  application	
  of	
  network	
  analyses	
  in	
  systems	
  biology	
  is	
  that	
  often	
  researchers	
  identify	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

clusters	
  or	
  modules	
  within	
  them	
  that	
  correspond	
  more	
  or	
  less	
  closely	
  to	
  mechanisms	
  that	
  
have	
  been	
  identified	
  using	
  more	
  traditional	
  approaches	
  in	
  cell	
  and	
  molecular	
  biology.	
  
Ravasz	
  et	
  al.	
  (2002)	
  provide	
  one	
  example	
  of	
  this.	
  The	
  researchers	
  constructed	
  an	
  overlap	
  
matrix	
  for	
  the	
  metabolic	
  network	
  in	
  E.	
  coli	
  and	
  identified	
  substrates	
  that	
  form	
  clusters	
  or	
  
modules	
  based	
  on	
  edges	
  corresponding	
  to	
  metabolic	
  interactions.	
  The	
  modules	
  identified	
  
through	
  this	
  analysis	
  closely	
  approximated	
  those	
  traditionally	
  identified	
  as	
  constituting	
  
mechanisms,	
  but	
  often	
  included	
  components	
  not	
  previously	
  identified.	
  	
  

	
  
4.	
  From	
  Graphs	
  to	
  Dynamics:	
  Coordinated	
  Dynamics	
  in	
  Mechanisms	
  
	
  
So	
  far	
  I	
  have	
  tried	
  to	
  show	
  that	
  biological	
  mechanisms	
  more	
  closely	
  correspond	
  to	
  
interconnected	
  modules	
  in	
  scale-­‐free	
  small-­‐world	
  networks	
  (Figure	
  4)	
  than	
  to	
  ovals	
  in	
  
Craver’s	
  diagram	
  (Figures	
  1	
  and	
  3).	
  To	
  appreciate	
  the	
  importance	
  of	
  interconnected	
  
modules,	
  we	
  must	
  move	
  beyond	
  graph-­‐theoretic	
  representations	
  and	
  consider	
  the	
  types	
  of	
  
dynamic	
  behavior	
  such	
  organization	
  supports.	
  When	
  one	
  starts	
  with	
  one	
  part	
  and	
  traces	
  
connections	
  in	
  interconnected	
  networks,	
  such	
  as	
  those	
  shown	
  in	
  Figures	
  2	
  and	
  4,	
  one	
  often	
  
discovers	
  that	
  the	
  parts	
  affected	
  by	
  the	
  operation	
  themselves	
  perform	
  operations	
  that	
  
directly	
  or	
  indirectly	
  affect	
  the	
  operation	
  of	
  the	
  part	
  from	
  which	
  one	
  started.	
  That	
  is,	
  the	
  
operations	
  feed	
  back	
  onto	
  parts	
  that	
  were	
  envisioned	
  as	
  earlier	
  in	
  the	
  process.	
  Such	
  
feedback	
  does	
  not	
  involve	
  backwards	
  causation,	
  since	
  the	
  effects	
  are	
  not	
  on	
  current	
  but	
  
future	
  operations	
  of	
  the	
  part,	
  but	
  over	
  time	
  it	
  can	
  give	
  rise	
  to	
  coordinated	
  dynamical	
  
behavior	
  that	
  constrains	
  the	
  behavior	
  of	
  the	
  parts	
  of	
  the	
  module.	
  	
  
	
  
Understanding	
  feedback	
  has	
  proven	
  challenging	
  for	
  humans.	
  Although	
  we	
  know	
  it	
  was	
  
already	
  employed	
  by	
  Ktesibios	
  of	
  Alexandria	
  in	
  the	
  3rd	
  century	
  CBE	
  to	
  regulate	
  the	
  flow	
  of	
  
water	
  in	
  his	
  water	
  clock,	
  negative	
  feedback	
  did	
  not	
  become	
  recognized	
  as	
  a	
  general	
  design	
  
principle	
  until	
  the	
  20th	
  century,	
  when	
  cyberneticists	
  such	
  as	
  Wiener	
  (1948;	
  Rosenblueth,	
  
Wiener,	
  and	
  Bigelow	
  1943)	
  presented	
  it	
  as	
  a	
  principle	
  for	
  enabling	
  engineered	
  and	
  natural	
  
systems	
  to	
  achieve	
  target	
  outcomes.	
  Prior	
  to	
  that	
  it	
  had	
  been	
  rediscovered	
  many	
  times,	
  
including	
  by	
  Watt,	
  who	
  employed	
  it	
  in	
  his	
  governor	
  for	
  the	
  steam	
  engine,	
  which	
  inspired	
  to	
  
Maxwell’s	
  mathematical	
  analysis	
  of	
  governors	
  (Mayr	
  1970).	
  But	
  even	
  when	
  negative	
  
feedback	
  was	
  recognized	
  as	
  a	
  design	
  principle	
  for	
  regulating	
  systems	
  to	
  pursue	
  target	
  
outcomes,	
  many	
  theorists	
  did	
  not	
  attend	
  to	
  the	
  fact	
  already	
  recognized	
  by	
  engineers	
  that	
  
negative	
  feedback	
  could	
  generate	
  oscillations.	
  	
  
	
  
Oscillations	
  in	
  biological	
  organisms	
  were	
  often	
  concealed	
  by	
  such	
  techniques	
  as	
  examining	
  
mean	
  behavior	
  and	
  not	
  attending	
  to	
  time-­‐series.	
  Physiological	
  processes	
  were	
  envisioned,	
  
in	
  accord	
  with	
  Machamer,	
  Darden,	
  and	
  Craver	
  (2000),	
  as	
  proceeding	
  “from	
  start	
  or	
  set-­‐up	
  
to	
  finish	
  or	
  termination	
  conditions.”	
  Variation	
  in	
  activity	
  was	
  regarded	
  as	
  noise.	
  But	
  
increasingly	
  through	
  the	
  20th	
  century	
  researchers	
  came	
  to	
  recognize	
  that	
  a	
  broad	
  range	
  of	
  
physiological	
  processes,	
  from	
  glycolysis	
  to	
  neural	
  processing,	
  generate	
  oscillations.	
  
Following	
  on	
  models,	
  such	
  as	
  one	
  advanced	
  by	
  Goodwin	
  (1965),	
  negative	
  feedback	
  was	
  
recognized	
  as	
  a	
  design	
  principle	
  for	
  generating	
  endogenous	
  oscillations	
  and	
  when	
  
oscillations	
  were	
  discovered	
  in	
  biological	
  organisms,	
  investigators	
  proposed	
  negative	
  
feedback	
  mechanisms.	
  Circadian	
  rhythms	
  were	
  no	
  exception,	
  and	
  circadian	
  researchers	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

were	
  on	
  the	
  hunt	
  for	
  a	
  feedback	
  mechanism	
  for	
  several	
  decades	
  before	
  Hardin,	
  Hall,	
  and	
  
Rosbash	
  (1990)	
  provided	
  critical	
  empirical	
  evidence.3	
  Using	
  cloning,	
  they	
  demonstrated	
  
that	
  the	
  mRNA	
  and	
  proteins	
  produced	
  by	
  the	
  first	
  identified	
  circadian	
  gene,	
  Period	
  (Per)	
  
oscillated	
  with	
  a	
  circadian	
  period,	
  with	
  the	
  concentration	
  of	
  the	
  protein	
  peaking	
  several	
  
hours	
  after	
  the	
  mRNA.	
  On	
  this	
  basis,	
  they	
  proposed	
  the	
  feedback	
  loop	
  described	
  above	
  in	
  
which	
  PER	
  inhibits	
  the	
  transcription	
  of	
  Per.	
  Mental	
  simulation	
  of	
  the	
  operations	
  proposed	
  
reveals	
  how	
  the	
  mechanism	
  could	
  oscillate:	
  when	
  PER	
  levels	
  are	
  low,	
  synthesis	
  of	
  new	
  PER	
  
proceeds,	
  but	
  as	
  PER	
  accumulates,	
  it	
  inhibits	
  further	
  synthesis.	
  Only	
  when	
  sufficient	
  PER	
  
has	
  broken	
  down	
  can	
  synthesis	
  resume.	
  Mental	
  simulation	
  cannot	
  determine	
  whether	
  the	
  
oscillations	
  will	
  dampen	
  or	
  be	
  sustained	
  indefinitely	
  (assuming	
  a	
  sufficient	
  supply	
  of	
  free	
  
energy);	
  accordingly,	
  Goldbeter	
  (1995)	
  constructed	
  a	
  computational	
  model,	
  inspired	
  by	
  
Goodwin’s,	
  that	
  demonstrated	
  that	
  the	
  proposed	
  negative-­‐feedback	
  mechanism	
  could	
  
produce	
  sustained	
  oscillations	
  under	
  physiological	
  conditions.	
  	
  
	
  
What	
  is	
  important	
  for	
  thinking	
  about	
  top-­‐down	
  causation	
  is	
  that	
  the	
  overall	
  state	
  of	
  the	
  
circadian	
  mechanism	
  determines	
  how	
  components	
  within	
  it	
  behave	
  and	
  how	
  they	
  will	
  
respond	
  to	
  perturbations	
  arising	
  outside	
  the	
  mechanism.	
  This	
  state	
  of	
  the	
  whole	
  
mechanism	
  might	
  be	
  described	
  in	
  terms	
  of	
  the	
  states	
  of	
  some	
  components	
  of	
  the	
  
mechanism	
  (for	
  example,	
  PER	
  concentrations	
  in	
  the	
  nucleus	
  are	
  high).	
  The	
  effect	
  on	
  Per	
  
transcription	
  at	
  that	
  time	
  is	
  determined	
  by	
  conditions	
  generated	
  within	
  the	
  mechanism,	
  not	
  
from	
  outside	
  it.	
  This	
  point,	
  however,	
  extends	
  far	
  beyond	
  the	
  circadian	
  example.	
  Any	
  
network	
  in	
  which	
  the	
  edges	
  are	
  not	
  all	
  in	
  one	
  direction	
  is	
  subject	
  to	
  complex	
  dynamical	
  
behavior,	
  such	
  as	
  oscillation,	
  in	
  which	
  the	
  behavior	
  of	
  parts	
  is	
  constrained	
  by	
  the	
  behavior	
  
of	
  other	
  parts	
  in	
  the	
  network.	
  	
  
	
  
The	
  constraining	
  effect	
  of	
  the	
  whole	
  on	
  the	
  parts	
  is	
  clearly	
  seen	
  in	
  how	
  the	
  parts	
  respond	
  to	
  
external	
  inputs	
  differentially	
  depending	
  upon	
  the	
  state	
  of	
  the	
  mechanism.	
  This	
  is	
  manifest	
  
in	
  the	
  process	
  by	
  which	
  the	
  circadian	
  clock	
  is	
  entrained	
  to	
  local	
  conditions,	
  especially	
  light	
  
conditions.	
  Light	
  has	
  different	
  effects	
  at	
  different	
  times	
  of	
  day,	
  as	
  exhibited	
  in	
  the	
  phase-­‐
response	
  curve	
  shown	
  in	
  Figure	
  5	
  (time	
  is	
  shown	
  in	
  Circadian	
  Time,	
  according	
  to	
  which	
  0	
  
corresponds	
  to	
  dawn).	
  During	
  the	
  early	
  part	
  of	
  the	
  night	
  a	
  light	
  pulse	
  delays	
  the	
  phase	
  of	
  
the	
  oscillation,	
  but	
  a	
  light	
  pulse	
  late	
  at	
  night	
  advances	
  the	
  phase.	
  During	
  daytime	
  light	
  
pulses	
  have	
  no	
  effect.	
  Once	
  details	
  of	
  the	
  clock	
  mechanism	
  in	
  mammals	
  were	
  worked	
  out,	
  
researchers	
  determined	
  that	
  entrainment	
  to	
  light	
  resulted	
  when	
  a	
  light	
  signal	
  from	
  the	
  
retina	
  to	
  the	
  suprachiasmatic	
  nucleus	
  functions	
  to	
  increase	
  Per	
  transcription.	
  Now	
  
researchers	
  could	
  understand	
  why	
  light	
  exposure	
  had	
  different	
  effects	
  on	
  the	
  phase	
  
depending	
  on	
  time	
  of	
  day.	
  If	
  Per	
  transcription	
  is	
  already	
  at	
  its	
  maximum,	
  as	
  it	
  would	
  be	
  
during	
  anticipated	
  daytime,	
  light	
  input	
  could	
  have	
  no	
  further	
  effect.	
  At	
  dusk	
  and	
  early	
  night,	
  
as	
  Per	
  concentration	
  is	
  diminishing	
  due	
  to	
  the	
  feedback	
  mechanism,	
  a	
  light	
  signal	
  can	
  
counter	
  the	
  increasing	
  inhibition	
  and	
  keep	
  transcription	
  going	
  longer.	
  This	
  has	
  the	
  effect	
  of	
  
delaying	
  the	
  phase	
  of	
  the	
  oscillator.	
  On	
  the	
  other	
  hand,	
  later	
  in	
  the	
  night	
  as	
  dawn	
  is	
  
approaching,	
  Per	
  concentration	
  is	
  again	
  beginning	
  to	
  increase	
  due	
  to	
  the	
  endogenous	
  
	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
3	
  They	
  used	
  fruit	
  flies	
  as	
  their	
  model	
  organism,	
  but	
  within	
  the	
  decade	
  the	
  homologs	
  of	
  Per	
  
were	
  found	
  in	
  mammals.	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

oscillation.	
  A	
  light	
  signal	
  at	
  this	
  phase	
  will	
  speed	
  up	
  transcription,	
  resulting	
  in	
  reaching	
  
daytime	
  levels	
  in	
  PER	
  concentration	
  earlier	
  and	
  advancing	
  the	
  phase	
  of	
  the	
  oscillator.	
  	
  

	
  
Figure	
  5.	
  Typical	
  phase	
  response	
  curve	
  for	
  rodents	
  kept	
  for	
  total	
  darkness	
  and	
  then	
  
administered	
  light	
  pulses	
  at	
  different	
  times.	
  Circadian	
  time	
  0	
  is	
  the	
  time	
  of	
  expected	
  
dawn.	
  	
  

	
  
The	
  circadian	
  example	
  makes	
  clear	
  the	
  functional	
  importance	
  of	
  the	
  type	
  of	
  integration	
  of	
  
nodes	
  found	
  in	
  modules	
  in	
  network	
  representations	
  of	
  biological	
  systems.	
  The	
  operations	
  
of	
  individual	
  components	
  (e.g.,	
  transcribing	
  genes)	
  depend	
  on	
  the	
  state	
  of	
  other	
  parts	
  of	
  the	
  
mechanism	
  (e.g.,	
  the	
  concentrations	
  of	
  the	
  PER	
  protein	
  in	
  the	
  nucleus).	
  As	
  a	
  result	
  of	
  the	
  
interconnectivity	
  of	
  the	
  parts,	
  especially	
  the	
  feedback	
  loops,	
  the	
  module	
  identified	
  as	
  the	
  
mechanism	
  functions	
  as	
  a	
  unit,	
  with	
  the	
  operations	
  of	
  the	
  individual	
  parts	
  of	
  the	
  
mechanism	
  determined	
  by	
  other	
  parts	
  of	
  the	
  mechanism.	
  If	
  there	
  were	
  no	
  feedback	
  loops,	
  
the	
  parts	
  that	
  receive	
  the	
  input	
  would	
  not	
  be	
  sensitive	
  to	
  conditions	
  elsewhere	
  and	
  would	
  
always	
  respond	
  in	
  the	
  same	
  manner	
  to	
  the	
  input.	
  There	
  would	
  also	
  be	
  little	
  reason	
  to	
  
identify	
  the	
  parts	
  as	
  constituting	
  a	
  module.	
  The	
  feedback	
  loops	
  are	
  responsible	
  for	
  the	
  state	
  
of	
  the	
  module	
  modulating	
  the	
  responsiveness	
  of	
  its	
  parts	
  to	
  external	
  input.	
  Even	
  though	
  the	
  
input	
  to	
  the	
  mechanism	
  may	
  only	
  affect	
  one	
  or	
  a	
  few	
  components,	
  the	
  mechanism	
  as	
  a	
  
whole	
  is	
  the	
  relevant	
  unit	
  due	
  to	
  the	
  interconnections	
  that	
  run	
  through	
  the	
  mechanism.	
  	
  	
  
	
  
As	
  network	
  analyses	
  have	
  been	
  pursued	
  in	
  fields	
  such	
  as	
  systems	
  biology	
  and	
  neuroscience,	
  
it	
  has	
  become	
  apparent	
  that	
  modules	
  often	
  form	
  a	
  hierarchy:	
  modules	
  often	
  have	
  far	
  more	
  
connections	
  to	
  some	
  modules	
  than	
  others.	
  The	
  toy	
  network	
  in	
  Figure	
  4	
  shows	
  
interconnections	
  among	
  all	
  the	
  modules	
  that	
  render	
  the	
  whole	
  into	
  a	
  module.	
  The	
  
interconnections	
  that	
  give	
  rise	
  to	
  these	
  larger	
  modules	
  also	
  result	
  in	
  dynamics	
  that	
  affect	
  
the	
  behavior	
  of	
  the	
  individual	
  modules.	
  A	
  last	
  circadian	
  example	
  provides	
  an	
  illustration	
  of	
  
such	
  a	
  hierarchy	
  and	
  how	
  it	
  contributes	
  to	
  control	
  over	
  individual	
  components	
  multiple	
  
levels	
  below.	
  An	
  important	
  feature	
  of	
  oscillatory	
  systems,	
  already	
  recognized	
  by	
  Huygens	
  in	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

the	
  case	
  of	
  pendulum	
  clocks,	
  is	
  that	
  a	
  very	
  weak	
  signal	
  is	
  capable	
  of	
  causing	
  the	
  oscillations	
  
to	
  synchronize.	
  I	
  noted	
  above	
  that	
  the	
  suprachiasmatic	
  nucleus	
  (SCN),	
  a	
  bilateral	
  structure	
  
with	
  about	
  10,000	
  neurons	
  on	
  each	
  side,	
  serves	
  as	
  the	
  master	
  circadian	
  clock	
  in	
  mammals.	
  
By	
  dispersing	
  SCN	
  neurons,	
  Welsh	
  et	
  al.	
  (1995)	
  determined	
  that	
  individual	
  neurons	
  sustain	
  
oscillations	
  but	
  with	
  widely	
  varying	
  periods.	
  When	
  they	
  are	
  connected	
  in	
  the	
  SCN	
  or	
  in	
  
slices,	
  they	
  synchronize	
  with	
  each	
  other,	
  thereby	
  enabling	
  sloppy	
  oscillators	
  to	
  generate	
  a	
  
highly	
  regular	
  signal	
  to	
  transmit	
  to	
  the	
  rest	
  of	
  the	
  organism	
  (Aton	
  and	
  Herzog	
  2005).	
  
Synchronization	
  of	
  components	
  depends	
  on	
  the	
  long-­‐distance	
  connections	
  exhibited	
  in	
  
Figure	
  4.	
  It	
  begins	
  with	
  operations	
  in	
  which	
  molecules	
  such	
  as	
  the	
  hormone	
  VIP	
  are	
  
synthesized	
  in	
  individual	
  cells.	
  These	
  are	
  dispersed	
  out	
  of	
  the	
  cell	
  to	
  bind	
  to	
  receptors	
  on	
  
other	
  cells,	
  initiating	
  a	
  signaling	
  cascade	
  in	
  those	
  cells	
  that	
  ultimately	
  affects	
  the	
  processes	
  
of	
  Per	
  transcription	
  and	
  translation	
  within	
  these	
  cells.	
  	
  
	
  
To	
  understand	
  synchronization,	
  one	
  must	
  identify	
  modules	
  at	
  two	
  different	
  levels,	
  and	
  
understand	
  how	
  conditions	
  in	
  modules	
  at	
  each	
  level	
  affect	
  both	
  their	
  own	
  components	
  and	
  
those	
  at	
  still	
  lower	
  levels.	
  The	
  interactions	
  between	
  SCN	
  neurons	
  determines	
  the	
  rate	
  and	
  
efficiency	
  of	
  synchronization	
  among	
  them,	
  while	
  the	
  interactions	
  within	
  neurons	
  
determines	
  how	
  they	
  respond	
  when	
  VIP	
  binds	
  to	
  one	
  of	
  their	
  receptors.	
  Although	
  the	
  
individual	
  operations	
  are	
  described	
  at	
  the	
  molecular	
  level,	
  they	
  are	
  constrained	
  by	
  the	
  
current	
  conditions	
  in	
  the	
  large-­‐scale	
  modules.	
  
	
  
5.	
  Top-­‐Down	
  Effects	
  Due	
  to	
  Constraints	
  Imposed	
  by	
  Networks	
  Dynamics	
  
	
  
My	
  strategy	
  in	
  the	
  previous	
  sections	
  has	
  been	
  to	
  use	
  network	
  analyses	
  to	
  identify	
  modules	
  
and	
  the	
  dynamic	
  behavior	
  that	
  arises	
  in	
  some	
  modules	
  with	
  rich	
  interactive	
  connections	
  to	
  
characterize	
  the	
  constitutive	
  relation	
  that	
  holds	
  between	
  parts	
  and	
  a	
  mechanism.	
  When	
  
network	
  analyses	
  are	
  applied	
  to	
  biological	
  systems,	
  one	
  frequently	
  finds	
  small-­‐world	
  
organization	
  in	
  which	
  node	
  degree	
  is	
  distributed	
  according	
  to	
  a	
  power	
  law.	
  This	
  has	
  the	
  
result	
  of	
  creating	
  modules	
  of	
  highly	
  interconnected	
  nodes	
  in	
  which,	
  nonetheless,	
  several	
  
nodes	
  still	
  receive	
  inputs	
  from	
  outside	
  the	
  module.	
  These	
  modules	
  often	
  correspond	
  to	
  
biological	
  mechanisms	
  that	
  have	
  been	
  identified	
  through	
  more	
  classical	
  techniques.	
  	
  
	
  
In	
  Craver’s	
  diagram	
  (Figures	
  1	
  and	
  3),	
  the	
  ovals	
  drawn	
  around	
  mechanisms	
  seemed	
  
arbitrary,	
  but	
  on	
  the	
  network	
  account	
  there	
  are	
  principled	
  reasons	
  for	
  picking	
  these	
  out	
  as	
  
modules.	
  What	
  renders	
  a	
  group	
  of	
  entities	
  into	
  a	
  module	
  is	
  the	
  interconnections	
  and	
  
interactions	
  between	
  them.	
  When	
  the	
  entities	
  work	
  together	
  to	
  produce	
  a	
  phenomenon,	
  
they	
  count	
  as	
  a	
  mechanism.	
  The	
  interconnections	
  and	
  interactions	
  often	
  yield	
  dynamic	
  
behavior	
  in	
  which	
  components	
  in	
  a	
  mechanism	
  behave	
  differently	
  at	
  different	
  times	
  due	
  to	
  
activity	
  elsewhere	
  in	
  the	
  mechanism.	
  As	
  a	
  result,	
  determining	
  the	
  organization	
  and	
  
dynamics	
  of	
  the	
  module	
  is	
  crucial	
  for	
  understanding	
  the	
  behavior	
  of	
  its	
  parts	
  when	
  they	
  
receive	
  external	
  inputs.	
  
	
  
A	
  concern	
  I	
  raised	
  earlier	
  about	
  Craver	
  and	
  my	
  treatment	
  of	
  top-­‐down	
  causation	
  is	
  that	
  it	
  
rendered	
  all	
  causal	
  relations	
  at	
  the	
  lowest	
  level.	
  To	
  first	
  appearances,	
  graph	
  theoretic	
  
representations	
  of	
  networks	
  seem	
  to	
  reinforce	
  that	
  concern.	
  In	
  cases	
  in	
  which	
  things	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

outside	
  the	
  mechanism	
  exercise	
  effects	
  on	
  the	
  mechanism,	
  it	
  will	
  typically	
  be	
  by	
  affecting	
  
one	
  or	
  more	
  parts	
  of	
  the	
  mechanism.	
  These	
  altered	
  parts	
  then	
  causally	
  modify	
  other	
  
components	
  of	
  the	
  mechanism.	
  Even	
  endogenous	
  activities	
  such	
  as	
  oscillations	
  are	
  
explained	
  in	
  terms	
  of	
  the	
  feedback	
  between	
  components	
  of	
  the	
  mechanism.	
  The	
  network	
  
representation	
  seems	
  to	
  favor	
  a	
  highly	
  reductionistic	
  account	
  that	
  represents	
  all	
  activity	
  at	
  
one	
  lowest	
  level.	
  	
  
	
  
This	
  interpretation,	
  however,	
  is	
  mistaken.	
  First,	
  the	
  nodes	
  in	
  a	
  network	
  need	
  not	
  belong	
  to	
  
a	
  common	
  level	
  in	
  any	
  of	
  the	
  standard	
  senses.	
  In	
  some	
  cases	
  a	
  graph	
  representation	
  of	
  a	
  
network	
  is	
  developed	
  by	
  identifying	
  interactions	
  between	
  entities	
  that	
  might	
  be	
  taken	
  to	
  be	
  
at	
  a	
  common	
  level	
  on	
  grounds	
  such	
  as	
  common	
  size	
  or	
  common	
  type	
  of	
  entity.	
  Gene	
  
regulation	
  and	
  protein	
  interaction	
  networks	
  are	
  examples.	
  In	
  other	
  cases	
  the	
  nodes	
  
correspond	
  to	
  entities	
  that	
  are	
  structurally	
  or	
  functionally	
  connected	
  independently	
  of	
  
whether	
  they	
  are	
  situated	
  at	
  what	
  is	
  regarded	
  as	
  a	
  common	
  level.	
  The	
  only	
  sense	
  of	
  level	
  
that	
  is	
  explicitly	
  embodied	
  in	
  a	
  network	
  diagram	
  is	
  between	
  nodes	
  and	
  modules	
  that	
  
comprise	
  them.	
  This	
  is	
  the	
  familiar	
  mechanistic	
  sense	
  of	
  level.	
  As	
  Craver	
  and	
  I	
  argued,	
  
however,	
  this	
  provides	
  only	
  a	
  very	
  local	
  conception	
  of	
  level—the	
  parts	
  of	
  a	
  mechanism	
  that	
  
interact	
  are	
  at	
  the	
  same	
  level.	
  This	
  account	
  provides	
  no	
  guidance	
  for	
  judging	
  whether	
  the	
  
nodes	
  of	
  different	
  modules	
  correspond	
  to	
  entities	
  at	
  the	
  same	
  level	
  or	
  whether	
  the	
  sub-­‐
parts	
  of	
  the	
  parts	
  belong	
  to	
  the	
  same	
  level.	
  On	
  this	
  conception	
  of	
  level,	
  there	
  are	
  no	
  grounds	
  
for	
  treating	
  the	
  nodes	
  in	
  a	
  given	
  graph	
  as	
  at	
  a	
  common	
  level.	
  
	
  
Second,	
  although	
  in	
  any	
  graph	
  representation	
  there	
  will	
  be	
  a	
  set	
  of	
  nodes	
  that	
  correspond	
  
to	
  what	
  are	
  taken	
  as	
  the	
  basic	
  entities,	
  they	
  should	
  not	
  be	
  treated	
  as	
  representing	
  entities	
  at	
  
some	
  base	
  level.	
  At	
  best	
  they	
  represent	
  the	
  entities	
  at	
  which	
  the	
  graph	
  representation	
  
bottoms	
  out.	
  Just	
  as	
  researchers	
  investigating	
  a	
  mechanism	
  have	
  a	
  choice	
  as	
  to	
  whether	
  to	
  
characterize	
  the	
  behavior	
  of	
  the	
  mechanism	
  as	
  it	
  interacts	
  with	
  other	
  entities,	
  many	
  of	
  
them	
  mechanisms	
  (thereby	
  elaborating	
  the	
  characterization	
  of	
  the	
  phenomenon)	
  or	
  to	
  
decompose	
  it	
  and	
  appeal	
  to	
  its	
  parts	
  and	
  operations	
  to	
  explain	
  its	
  behavior,	
  those	
  
developing	
  a	
  graph	
  representation	
  have	
  a	
  choice	
  as	
  to	
  whether	
  to	
  represent	
  a	
  whole	
  
mechanism	
  as	
  a	
  node,	
  or	
  decompose	
  it	
  into	
  other	
  nodes	
  representing	
  the	
  parts	
  of	
  the	
  
mechanism.	
  On	
  many	
  occasions	
  researchers	
  seek	
  to	
  decompose	
  one	
  part	
  of	
  the	
  mechanism,	
  
leaving	
  others	
  untouched.	
  The	
  graph	
  representation	
  will	
  show	
  the	
  components	
  into	
  which	
  
the	
  one	
  mechanism	
  has	
  been	
  decomposed	
  interacting	
  with	
  the	
  other	
  mechanisms	
  that	
  have	
  
not	
  been	
  decomposed.	
  	
  
	
  
Third,	
  in	
  developing	
  a	
  graph	
  representation,	
  one	
  might	
  deliberately	
  represent	
  a	
  set	
  of	
  
entities	
  as	
  a	
  single	
  node.	
  Researcher	
  may	
  have	
  already	
  identified	
  the	
  components	
  of	
  a	
  
mechanism,	
  but	
  deem	
  them	
  not	
  to	
  be	
  relevant	
  to	
  their	
  analysis.	
  For	
  example,	
  in	
  analyzing	
  
the	
  interactions	
  between	
  SCN	
  neurons,	
  researchers	
  may	
  choose	
  to	
  treat	
  the	
  neurons	
  as	
  
units,	
  ignoring	
  the	
  interactions	
  of	
  molecules	
  within	
  them.	
  The	
  graph	
  representation	
  will	
  
employ	
  units	
  for	
  neurons	
  and	
  edges	
  for	
  the	
  connections	
  between	
  them.	
  There	
  are	
  contexts	
  
in	
  which	
  the	
  whole	
  SCN	
  becomes	
  a	
  single	
  node	
  and	
  the	
  edges	
  are	
  the	
  connections	
  between	
  
the	
  SCN	
  and	
  the	
  various	
  organs	
  that	
  the	
  SCN	
  regulates	
  and	
  which,	
  in	
  many	
  cases,	
  send	
  
inputs	
  back	
  to	
  the	
  SCN.	
  The	
  graph	
  representation	
  format	
  does	
  not	
  privilege	
  a	
  lowest	
  level,	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

but	
  represents	
  as	
  nodes	
  those	
  entities	
  whose	
  interactions	
  are	
  deemed	
  relevant	
  for	
  
understanding	
  the	
  phenomenon	
  of	
  interest.	
  
	
  
Although	
  a	
  graph	
  representation	
  provides	
  a	
  convenient	
  way	
  to	
  identify	
  modules	
  in	
  
networks,	
  by	
  using	
  the	
  same	
  type	
  of	
  arrow	
  to	
  relate	
  nodes	
  within	
  a	
  module	
  and	
  those	
  
between	
  nodes	
  in	
  different	
  modules,	
  it	
  glosses	
  over	
  an	
  important	
  distinction	
  relevant	
  to	
  
understanding	
  top-­‐down	
  causation	
  in	
  mechanisms.	
  Looking	
  diachronically	
  both	
  the	
  edges	
  
between	
  nodes	
  inside	
  a	
  module	
  and	
  those	
  between	
  modules	
  may	
  represent	
  causal	
  
relations.	
  But	
  in	
  the	
  moment	
  when	
  the	
  mechanism	
  receives	
  causal	
  input	
  from	
  outside	
  by	
  
having	
  the	
  state	
  of	
  one	
  or	
  more	
  of	
  its	
  parts	
  altered,	
  the	
  relation	
  between	
  the	
  parts	
  and	
  the	
  
mechanism	
  as	
  a	
  whole	
  is	
  not	
  diachronic	
  but	
  synchronic.	
  At	
  a	
  given	
  time,	
  the	
  mechanism	
  is	
  
constituted	
  of	
  its	
  various	
  parts.	
  The	
  natural	
  language	
  to	
  use	
  to	
  talk	
  about	
  how	
  
synchronically	
  the	
  parts	
  are	
  affected	
  by	
  the	
  whole	
  is	
  that	
  of	
  constraint—being	
  situated	
  in	
  
the	
  mechanism	
  constrains	
  the	
  behavior	
  of	
  the	
  part.	
  	
  
	
  
The	
  notion	
  of	
  constraint	
  has	
  roots	
  in	
  mechanics.	
  Fundamental	
  laws	
  such	
  as	
  those	
  proposed	
  
by	
  Newton	
  characterize	
  possible	
  ways	
  a	
  system	
  might	
  evolve	
  given	
  initial	
  conditions.	
  But	
  in	
  
a	
  given	
  situation	
  constraints	
  restrict	
  those	
  possibilities,	
  foreclosing	
  some	
  while	
  leaving	
  
others	
  open	
  (Hooker	
  2013).	
  Constraints	
  such	
  as	
  an	
  inclined	
  plane	
  limit	
  the	
  motion	
  a	
  marble	
  
can	
  take	
  as	
  a	
  result	
  of	
  the	
  gravitational	
  attraction	
  between	
  it	
  and	
  the	
  earth.	
  In	
  terms	
  of	
  
mathematical	
  representations	
  of	
  the	
  evolution	
  of	
  a	
  system,	
  constraints	
  are	
  captured	
  in	
  
relations	
  added	
  to	
  dynamical	
  equations	
  that	
  limit	
  the	
  degrees	
  of	
  freedom	
  available	
  through	
  
which	
  the	
  system	
  might	
  evolve.	
  	
  
	
  
Some	
  constraints,	
  such	
  as	
  an	
  inclined	
  plane	
  or	
  an	
  electric	
  wire,	
  are	
  fixed	
  with	
  respect	
  to	
  the	
  
system	
  in	
  question.	
  They	
  provide	
  a	
  minimalist	
  notion	
  of	
  top-­‐down	
  causation.	
  The	
  path	
  a	
  
marble	
  takes	
  after	
  being	
  deposited	
  on	
  an	
  inclined	
  plane	
  is	
  governed	
  by	
  the	
  angle	
  of	
  the	
  
plane.	
  Other	
  constraints	
  change,	
  often	
  in	
  response	
  to	
  other	
  activities	
  in	
  the	
  system.	
  A	
  
switch	
  in	
  an	
  electrical	
  system	
  can	
  direct	
  electricity	
  along	
  different	
  paths	
  in	
  a	
  circuit.	
  More	
  
interesting	
  sets	
  of	
  constraints,	
  and	
  more	
  interesting	
  examples	
  of	
  top-­‐down	
  causation,	
  arise	
  
as	
  parts	
  of	
  a	
  system	
  constrain	
  each	
  other	
  in	
  ways	
  that	
  change	
  dynamically.	
  In	
  the	
  much-­‐
discussed	
  example	
  of	
  Bénard	
  cells,	
  after	
  the	
  application	
  of	
  heat,	
  molecules	
  begin	
  to	
  move	
  
and	
  exert	
  constraints	
  on	
  each	
  other.	
  Eventually	
  coordination	
  between	
  the	
  molecules	
  results	
  
in	
  a	
  macro-­‐scale	
  convection	
  pattern	
  in	
  which	
  individual	
  molecules	
  are	
  constrained.	
  
	
  
The	
  notion	
  of	
  constraint	
  has	
  been	
  applied	
  to	
  biology	
  by	
  a	
  number	
  of	
  authors	
  (Pattee	
  1971;	
  
Rosen	
  1985;	
  Hooker	
  2013;	
  Moreno	
  and	
  Mossio	
  2015).	
  A	
  key	
  foundation	
  of	
  their	
  thinking	
  is	
  
that	
  biological	
  organisms	
  exist	
  far	
  from	
  thermodynamic	
  equilibrium	
  and	
  in	
  order	
  to	
  build	
  
and	
  maintain	
  themselves,	
  they	
  must	
  constrain	
  the	
  flow	
  of	
  free	
  energy	
  to	
  perform	
  work	
  that	
  
builds	
  the	
  structures	
  that	
  perform	
  the	
  constraint	
  role.	
  The	
  different	
  mechanisms	
  
constituting	
  a	
  biological	
  organism	
  each	
  play	
  a	
  role	
  in	
  such	
  a	
  process	
  by	
  restricting	
  the	
  range	
  
of	
  activities	
  that	
  can	
  occur.	
  The	
  circadian	
  mechanism	
  that	
  provided	
  my	
  example	
  throughout	
  
this	
  paper	
  figures	
  in	
  such	
  a	
  network	
  of	
  constraints	
  as	
  it	
  constrains	
  various	
  physiological	
  
processes	
  (typically	
  by	
  regulating	
  expression	
  of	
  other	
  genes)	
  to	
  occur	
  at	
  times	
  when	
  they	
  
can	
  work	
  together	
  to	
  maintain	
  the	
  organism.	
  The	
  parts	
  of	
  the	
  mechanism	
  itself	
  are	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

generated	
  by	
  the	
  operation	
  of	
  the	
  mechanism	
  (the	
  protein	
  PER	
  is	
  synthesized	
  when	
  BMAL1	
  
binds	
  to	
  the	
  promoter	
  on	
  the	
  Per	
  gene)	
  and	
  perform	
  operations	
  in	
  it	
  (PER	
  inhibits	
  the	
  
ability	
  of	
  BMAL1	
  to	
  activate	
  the	
  transcription	
  of	
  the	
  Per	
  gene).	
  	
  
	
  
The	
  notion	
  of	
  constraint	
  provides	
  a	
  way	
  to	
  understand	
  how	
  the	
  constitution	
  of	
  a	
  
mechanism	
  results	
  in	
  the	
  phenomena	
  referred	
  to	
  as	
  top-­‐down	
  causation.	
  Appealing	
  to	
  
graph	
  theoretic	
  representations	
  of	
  systems,	
  I	
  have	
  identified	
  mechanisms	
  as	
  dynamical	
  
systems	
  that	
  can	
  arise	
  in	
  modules.	
  The	
  degrees	
  of	
  freedom	
  available	
  to	
  entities	
  of	
  such	
  
interactive	
  dynamical	
  systems	
  are	
  reduced	
  and	
  so	
  they	
  behave	
  differently	
  than	
  they	
  would	
  
if	
  not	
  part	
  of	
  the	
  system.	
  Especially	
  when	
  the	
  constraints	
  are	
  changing	
  as	
  a	
  result	
  of	
  the	
  
dynamical	
  activity,	
  these	
  entities	
  may	
  exhibit	
  different	
  behavior	
  on	
  different	
  occasions.	
  	
  
	
  
Thinking	
  in	
  terms	
  of	
  constraints	
  also	
  facilitates	
  a	
  response	
  to	
  the	
  assumption	
  of	
  the	
  
closedness	
  of	
  the	
  physical	
  that	
  lies	
  at	
  the	
  foundation	
  of	
  Kim’s	
  exclusion	
  argument.	
  
Fundamental	
  dynamical	
  laws	
  do	
  apply	
  universally	
  and	
  as	
  long	
  as	
  one	
  can	
  specify	
  initial	
  
conditions	
  for	
  each	
  entity	
  in	
  the	
  system,	
  one	
  can	
  determine	
  its	
  behavior	
  by	
  invoking	
  these	
  
laws.	
  But	
  even	
  to	
  address	
  problems	
  of	
  analytic	
  mechanics,	
  physicists	
  recognize	
  the	
  need	
  to	
  
invoke	
  constraints.	
  These	
  constraints	
  are	
  not	
  derived	
  from	
  the	
  laws	
  themselves	
  but	
  must	
  
be	
  ascertained	
  empirically	
  and	
  added	
  to	
  the	
  laws	
  to	
  determine	
  behavior.	
  The	
  set	
  of	
  possible	
  
constraints	
  is	
  not	
  closed.	
  Yet,	
  only	
  in	
  terms	
  of	
  constraints	
  can	
  one	
  predict	
  or	
  explain	
  
outcomes.	
  Moreover,	
  they	
  are	
  determined	
  locally.	
  In	
  biology,	
  the	
  constraints	
  imposed	
  in	
  a	
  
mechanism	
  are	
  specific	
  to	
  the	
  conditions	
  in	
  the	
  living	
  system.	
  From	
  this	
  perspective,	
  the	
  
physical	
  is	
  far	
  from	
  closed	
  but	
  rather	
  is	
  extremely	
  open-­‐ended.	
  Wherever	
  one	
  finds	
  a	
  set	
  of	
  
components	
  organized	
  into	
  a	
  module	
  with	
  sufficient	
  interactions,	
  one	
  will	
  encounter	
  
constraints	
  that	
  limit	
  the	
  behavior	
  of	
  the	
  components	
  and	
  how	
  they	
  respond	
  to	
  external	
  
inputs.	
  The	
  phenomenon	
  described	
  as	
  top-­‐down	
  causation	
  is	
  not	
  unusual,	
  but	
  common.	
  	
  
	
  
6.	
  Conclusion	
  
	
  
Craver	
  and	
  I	
  proposed	
  that	
  bottom-­‐up	
  as	
  well	
  as	
  top-­‐down	
  causation	
  could	
  be	
  understood	
  
by	
  limiting	
  causal	
  processes	
  to	
  within	
  levels	
  of	
  mechanisms	
  and	
  treating	
  the	
  constitution	
  of	
  
mechanisms	
  as	
  mediating	
  between	
  levels.	
  We	
  did	
  not,	
  however,	
  provide	
  an	
  account	
  of	
  what	
  
it	
  was	
  about	
  constitution	
  of	
  mechanisms	
  that	
  motivates	
  appeals	
  to	
  effects	
  or	
  causes	
  at	
  other	
  
levels.	
  Invoking	
  graph	
  representations,	
  in	
  this	
  paper	
  I	
  have	
  characterized	
  mechanisms	
  as	
  
modules	
  that	
  appear	
  in	
  many	
  graphs	
  of	
  biological	
  networks.	
  Modules	
  consist	
  of	
  nodes	
  that	
  
are	
  highly	
  clustered	
  and	
  in	
  that	
  respect	
  are	
  distinguished	
  from	
  other	
  nodes	
  (with	
  which	
  
they	
  still	
  have	
  a	
  number	
  of	
  edges).	
  Particularly	
  important	
  in	
  integrating	
  nodes	
  into	
  modules	
  
that	
  can	
  exhibit	
  what	
  are	
  taken	
  to	
  be	
  top-­‐down	
  effects	
  are	
  feedback	
  connections	
  that	
  are	
  
common	
  in	
  biological	
  systems.	
  With	
  feedback,	
  modules	
  are	
  capable	
  of	
  endogenous	
  activity	
  
that	
  results	
  in	
  them	
  being	
  in	
  different	
  states	
  at	
  different	
  times.	
  When	
  this	
  happens,	
  some	
  of	
  
their	
  components	
  are	
  in	
  different	
  states.	
  Depending	
  on	
  the	
  state,	
  the	
  component	
  will	
  
produce	
  different	
  responses	
  to	
  inputs.	
  This	
  is	
  the	
  sort	
  of	
  behavior	
  that	
  has	
  
characteristically	
  motivated	
  talk	
  of	
  top-­‐down	
  causation.	
  In	
  developing	
  this	
  account,	
  I	
  have	
  
employed	
  flat	
  graph	
  representations	
  of	
  systems	
  in	
  which	
  modules	
  can	
  be	
  identified.	
  Their	
  
dynamics	
  can	
  then	
  be	
  analyzed.	
  This	
  has	
  allowed	
  me	
  to	
  pick	
  out	
  mechanisms	
  in	
  which	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

phenomena	
  traditionally	
  characterized	
  as	
  top-­‐down	
  causation	
  can	
  be	
  understood	
  without	
  
directly	
  appealing	
  to	
  levels.	
  But	
  this	
  explication	
  of	
  levels	
  also	
  allows	
  us,	
  if	
  desired,	
  to	
  
reintroduce	
  reference	
  to	
  levels	
  and	
  characterize	
  what	
  is	
  meant	
  by	
  top-­‐down	
  or	
  bottom-­‐up	
  
causation	
  in	
  non-­‐problematic	
  ways.	
  Finally,	
  by	
  characterizing	
  how	
  a	
  mechanism	
  affects	
  
parts	
  in	
  terms	
  of	
  constraint	
  I	
  have	
  differentiated	
  how	
  the	
  mechanism	
  restricts	
  the	
  behavior	
  
of	
  its	
  parts	
  from	
  the	
  way	
  the	
  parts	
  are	
  affected	
  by	
  external	
  inputs	
  and	
  provided	
  a	
  way	
  to	
  
resist	
  the	
  assumption	
  that	
  the	
  closedness	
  of	
  the	
  physical	
  underlying	
  Kim’s	
  exclusion	
  
argument.	
  	
  
	
  
	
  
References	
  	
  
	
  
Aton,	
  Sara	
  J.,	
  and	
  Erik	
  D.	
  Herzog.	
  (2005).	
  "Come	
  Together,	
  Right...Now:	
  Synchronization	
  of	
  

Rhythms	
  in	
  a	
  Mammalian	
  Circadian	
  Clock."	
  Neuron	
  48:531-­‐534.	
  
Barabási,	
  Albert-­‐László,	
  and	
  Réka	
  Albert.	
  (1999).	
  "Emergence	
  of	
  Scaling	
  in	
  Random	
  

Networks."	
  Science	
  286:509-­‐512.	
  
Barabasi,	
  Albert-­‐Laszlo,	
  and	
  Zoltan	
  N.	
  Oltvai.	
  (2004).	
  "Network	
  Biology:	
  Understanding	
  the	
  

Cell's	
  Functional	
  Organization."	
  Nature	
  Reviews	
  Genetics	
  5:101-­‐113.	
  
Bechtel,	
  William.	
  (2006).	
  Discovering	
  Cell	
  Mechanisms:	
  The	
  Creation	
  of	
  Modern	
  Cell	
  Biology,	
  

Cambridge:	
  Cambridge	
  University	
  Press.	
  
———.	
  (2008).	
  Mental	
  Mechanisms.	
  Philosophical	
  Perspectives	
  on	
  Cognitive	
  Neuroscience,	
  

London:	
  Routledge.	
  
———.	
  (2015).	
  "Can	
  Mechanistic	
  Explanation	
  Be	
  Reconciled	
  with	
  Scale-­‐Free	
  Constitution	
  

and	
  Dynamics?"	
  Studies	
  in	
  History	
  and	
  Philosophy	
  of	
  Science	
  Part	
  C:	
  Studies	
  in	
  History	
  
and	
  Philosophy	
  of	
  Biological	
  and	
  Biomedical	
  Sciences.	
  

Bechtel,	
  William,	
  and	
  Robert	
  C.	
  Richardson.	
  (1993/2010).	
  Discovering	
  Complexity:	
  
Decomposition	
  and	
  Localization	
  as	
  Strategies	
  in	
  Scientific	
  Research,	
  Cambridge,	
  MA:	
  
MIT	
  Press.	
  1993	
  edition	
  published	
  by	
  Princeton	
  University	
  Press.	
  

Churchland,	
  Patricia	
  S.,	
  and	
  Terrence	
  J.	
  Sejnowski.	
  (1992).	
  The	
  Computational	
  Brain,	
  
Cambridge,	
  MA:	
  MIT	
  Press.	
  

Craver,	
  Carl	
  F.	
  (2007).	
  Explaining	
  the	
  Brain:	
  	
  Mechanisms	
  and	
  the	
  Mosaic	
  Unity	
  of	
  
Neuroscience,	
  New	
  York:	
  Oxford	
  University	
  Press.	
  

Craver,	
  Carl	
  F.,	
  and	
  William	
  Bechtel.	
  (2007).	
  "Top-­‐Down	
  Causation	
  without	
  Top-­‐Down	
  
Causes."	
  Biology	
  and	
  Philosophy	
  22:547-­‐563.	
  

Fazekas,	
  Peter,	
  and	
  Gergely	
  Kertész.	
  (2011).	
  "Causation	
  at	
  Different	
  Levels:	
  Tracking	
  the	
  
Commitments	
  of	
  Mechanistic	
  Explanations."	
  Biology	
  &	
  Philosophy	
  26:365-­‐383.	
  

Goldbeter,	
  Albert.	
  (1995).	
  "A	
  Model	
  for	
  Circadian	
  Oscillations	
  in	
  the	
  Drosophila	
  Period	
  
Protein	
  (Per)."	
  Proceedings	
  of	
  the	
  Royal	
  Society	
  of	
  London.	
  B:	
  Biological	
  Sciences	
  
261:319-­‐324.	
  

Goodwin,	
  Brian	
  C.	
  (1965).	
  "Oscillatory	
  Behavior	
  in	
  Enzymatic	
  Control	
  Processes."	
  Advances	
  
in	
  Enzyme	
  Regulation	
  3:425-­‐428.	
  

Hardin,	
  Paul	
  E.,	
  Jeffrey	
  C.	
  Hall,	
  and	
  Michael	
  Rosbash.	
  (1990).	
  "Feedback	
  of	
  the	
  Drosophila	
  
Period	
  Gene	
  Product	
  on	
  Circadian	
  Cycling	
  of	
  Its	
  Messenger	
  Rna	
  Levels."	
  Nature	
  
343:536-­‐540.	
  



Copyright	
  Philosophy	
  of	
  Science	
  2016	
  
Preprint	
  (not	
  copyedited	
  or	
  formatted)	
  
Please	
  use	
  DOI	
  when	
  citing	
  or	
  quoting	
  

Hitchcock,	
  Christopher.	
  (2003).	
  "Of	
  Humean	
  Bondage."	
  British	
  Journal	
  for	
  the	
  Philosophy	
  of	
  
Science	
  54:1-­‐25.	
  

Hooker,	
  Clifford	
  A.	
  (2013).	
  "On	
  the	
  Import	
  of	
  Constraints	
  in	
  Complex	
  Dynamical	
  Systems."	
  
Foundations	
  of	
  Science	
  18:757-­‐780.	
  

Kim,	
  Jaegwon.	
  (1998).	
  Mind	
  in	
  a	
  Physical	
  World,	
  Cambridge,	
  MA:	
  MIT	
  Press.	
  
Lewis,	
  David.	
  (2000).	
  "Causation	
  as	
  Influence."	
  Journal	
  of	
  Philosophy	
  97:182-­‐197.	
  
Machamer,	
  Peter,	
  Lindley	
  Darden,	
  and	
  Carl	
  F.	
  Craver.	
  (2000).	
  "Thinking	
  About	
  

Mechanisms."	
  Philosophy	
  of	
  Science	
  67:1-­‐25.	
  
Mayr,	
  Otto.	
  (1970).	
  The	
  Origins	
  of	
  Feedback	
  Control,	
  Cambridge,	
  MA:	
  MIT	
  Press.	
  
Moreno,	
  Alvaro,	
  and	
  Matteo	
  Mossio.	
  (2015).	
  Biological	
  Autonomy:	
  A	
  Philosophical	
  and	
  

Theoretical	
  Inquiry,	
  Dordrecht:	
  Springer.	
  
Noble,	
  Denis.	
  (2006).	
  The	
  Music	
  of	
  Life:	
  Biology	
  Beyond	
  the	
  Genome,	
  Oxford:	
  Oxford	
  

University	
  Press.	
  
Pattee,	
  Howard	
  Hunt.	
  (1971).	
  "Physical	
  Theories	
  of	
  Biological	
  Co-­‐Ordination."	
  Quartery	
  

Review	
  of	
  Biophysics	
  4:255-­‐276.	
  
Ravasz,	
  Erzsébet,	
  A.	
  L.	
  Somera,	
  D.	
  A.	
  Mongru,	
  Zoltan	
  N.	
  Oltvai,	
  and	
  Albert-­‐Laszlo	
  Barabasi.	
  

(2002).	
  "Hierarchical	
  Organization	
  of	
  Modularity	
  in	
  Metabolic	
  Networks."	
  Science	
  
297:1551-­‐1555.	
  

Rosen,	
  Robert.	
  (1985).	
  "Organisms	
  as	
  Causal	
  Systems	
  Which	
  Are	
  Not	
  Mechanisms:	
  An	
  
Essay	
  into	
  the	
  Nature	
  of	
  Compleixty,"	
  In	
  Robert	
  Rosen,	
  ed.,	
  Theoretical	
  Biology	
  and	
  
Complexity:	
  Three	
  Essays	
  on	
  the	
  Natural	
  Philosophy	
  of	
  Complex	
  Systems,	
  165-­‐203.	
  
New	
  York:	
  Academic	
  Press.	
  

Rosenberg,	
  Alex.	
  (2015).	
  "Making	
  Mechanism	
  Interesting."	
  Synthese:1-­‐23.	
  
Rosenblueth,	
  Arturo,	
  Norbert	
  Wiener,	
  and	
  Julian	
  Bigelow.	
  (1943).	
  "Behavior,	
  Purpose,	
  and	
  

Teleology."	
  Philosophy	
  of	
  Science	
  10:18-­‐24.	
  
Soom,	
  Patrice.	
  (2012).	
  "Mechanisms,	
  Determination	
  and	
  the	
  Metaphysics	
  of	
  Neuroscience."	
  

Studies	
  in	
  History	
  and	
  Philosophy	
  of	
  Science	
  Part	
  C:	
  Studies	
  in	
  History	
  and	
  Philosophy	
  
of	
  Biological	
  and	
  Biomedical	
  Sciences	
  43:655-­‐664.	
  

Sporns,	
  Olaf.	
  (2010).	
  Networks	
  of	
  the	
  Brain,	
  Cambridge,	
  MA:	
  MIT	
  Press.	
  
Watts,	
  Duncan,	
  and	
  Steven	
  Strogratz.	
  (1998).	
  "Collective	
  Dynamics	
  of	
  Small	
  Worlds."	
  

Nature	
  393:440-­‐442.	
  
Welsh,	
  David	
  K.,	
  Diomedes	
  E.	
  Logothetis,	
  Markus	
  Meister,	
  and	
  Steven	
  M.	
  Reppert.	
  (1995).	
  

"Individual	
  Neurons	
  Dissociated	
  from	
  Rat	
  Suprachiasmatic	
  Nucleus	
  Express	
  
Independently	
  Phased	
  Circadian	
  Firing	
  Rhythms."	
  Neuron	
  14:697-­‐706.	
  

Wiener,	
  Norbert.	
  (1948).	
  Cybernetics:	
  Or,	
  Control	
  and	
  Communication	
  in	
  the	
  Animal	
  and	
  the	
  
Machine,	
  New	
  York:	
  Wiley.	
  

Zhang,	
  Eric	
  E.,	
  Andrew	
  C.	
  Liu,	
  Tsuyoshi	
  Hirota,	
  Loren	
  J.	
  Miraglia,	
  Genevieve	
  Welch,	
  Pagkapol	
  
Y.	
  Pongsawakul,	
  Xianzhong	
  Liu,	
  Ann	
  Atwood,	
  Jon	
  W.	
  Huss,	
  Jeff	
  Janes,	
  Andrew	
  I.	
  Su,	
  
John	
  B.	
  Hogenesch,	
  and	
  Steve	
  A.	
  Kay.	
  (2009).	
  "A	
  Genome-­‐Wide	
  Rnai	
  Screen	
  for	
  
Modifiers	
  of	
  the	
  Circadian	
  Clock	
  in	
  Human	
  Cells."	
  Cell	
  139:199-­‐210.