1: \documentclass[12pt,a4paper]{article}

2: \usepackage{graphicx}

3: %\usepackage{showkeys}

4: %\usepackage{tm-natbib}

5: \usepackage{natbib}

6: \usepackage{latexsym}

7: \usepackage{bm}

8: 

9: %\usepackage{twocolumn}

10: \bibpunct{(}{)}{;}{a}{,}{,}

11: 

12: \title{Egocentric Path Integration Models and their Application to

13:   Desert Arthropods}

14: 

15: \textheight25cm

16: \topmargin-2cm

17: 

18: %\renewcommand{\baselinestretch}{2}

19: 

20: \begin{document}

21: 

22: \maketitle

23: 

24: %\begin{mpage}

25: \begin{center}

26: Tobias Merkle, Martin Rost and Wolfgang Alt

27: \end{center}

28: 

29: \begin{center}

30: Theoretical Biology,\\

31: Mathematical and Natural Science Faculty,\\ 

32: University of Bonn\\

33: Kirschallee~1, D-53115~Bonn, Germany

34: \end{center}

35: %\end{mpage}

36: 

37: \begin{abstract}

38: Path integration enables desert arthropods to find back to their nest

39: on the shortest track from any position. To perform path integration

40: successfully, speeds and turning angles along the preceding outbound

41: path have to be measured continuously and combined to determine an

42: internal {\em global vector} leading back home at any time. A number

43: of experiments have given an idea how arthropods might use allothetic

44: or idiothetic signals to perceive their orientation and moving

45: speed. We systematically review the four possible model descriptions

46: of mathematically precise path integration, whereby we favour and

47: elaborate the hitherto not used variant of egocentric cartesian

48: coordinates. Its simple and intuitive structure is demonstrated in

49: comparison to the other models. Measuring two speeds, the forward

50: moving speed and the angular turning rate, and implementing them into

51: a linear system of differential equations provides the necessary

52: information during outbound route, reorientation process and return

53: path. In addition, we propose several possible types of systematic

54: errors that can cause deviations from the correct homeward

55: course. Deviations have been observed for several species of desert

56: arthropods in different experiments, but their origin is still under

57: debate. Using our egocentric path integration model we propose simple

58: error indices depending on path geometry that will allow future

59: experiments to rule out or corroborate certain error types. 

60: \\

61: \\

62: Key words: Path integration, desert arthropod, egocentric, cartesian coordinates

63: \\

64: \end{abstract}

65: 

66: \newpage

67: \section{Introduction}

68: \subsection{Path integration}

69: 

70: Desert arthropods display the ability to return from a foraging

71: excursion back to the nest on a straight way, often called the {\em

72:   home vector}. This ability to make a bee-line to the nest (or

73: another location, such as a feeding site) without orientation on

74: visible markers is based on an internal mechanism of {\em path

75:   integration} or {\em dead reckoning}, i.e.\ an integration of

76: walking speed and angular variation along the arthropod's walking

77: route. The result is a {\em global vector} that enables the arthropod

78: to determine distance and direction of its nest at any position and

79: time. After detecting and loading up the food, the arthropod just

80: unreels that vector and, therefore, stays on the right and shortest

81: track to its nest. Charles Darwin was the first to assume that animals

82: may navigate this way \citep{dar_73}. About a century later first

83: detailed investigations concerning arthropods

84: \citep{jan_57,goe_58,goe_66,weh_68,jan_70} were conducted. During the

85: last 30 years, the general interest has focused more and more on the

86: return path to the nest after foraging, and many investigations on both

87: arthropods \citep[e.g.][]{weh_86,mul_88,BIS99,col_99} and mammals

88: \citep[e.g.][]{mit_80,seg_93,ben_96,ben_97,seg_98} have been

89: performed.

90: 

91: Apart from path integration, it has been shown that many species are

92: capable of using landmarks to get their bearings

93: \citep{hof_85a,col_98,bis_03a,weh_03a}. These landmarks, often

94: referred to as {\em local vectors}, even seem to be preferred to the

95: global vectors \citep{weh_96a,col_98} as navigational tool. However,

96: before these vectors can be applied successfully, some information

97: about their position has to be stored by the arthropod. Moreover,

98: orientation with the aid of local vectors is error-prone, since

99: landmarks can disappear or change their appearance, and, of course, is

100: out of question for nests with no visible landmarks nearby.

101: 

102: The global vector gets updated on the complete trip, even if the

103: orientation is conducted by using landmarks

104: \citep[e.g.][]{weh_96a,col_98,col_03a}. Moreover, it has been

105: demonstrated that, after a sudden failure of the stored landmarks,

106: desert ants \emph{Cataglyphis fortis} revert to their global vector

107: for orientation. Even if not used for several days, desert ants keep

108: the global vector stored in their memory \citep{Zie_97}. Thus, the

109: relevance of global vectors as the main toolkit for homing seems to be

110: as clear as the evolutionary necessity to develop abilities to measure

111: the angular and linear components of the movements and to integrate

112: them for having a home vector available whensoever.

113: 

114: A number of models \citep{jan_57,mit_73,mul_88,ben_90,GAL90} try to

115: point out mechanisms of path integration \citep[for reviews,

116:   see][]{ben_95,mau_95,bie_00}. Whereas earlier models use {\em geocentric}

117: coordinates to represent the current global vector and the anticipated

118: home direction, the more recent models are based upon the assumption

119: that internal calculation of the global vector, pointing from the

120: animal's head to an anticipated nest position, should be performed in

121: an {\em egocentric} coordinate system. In this case, however,

122: mathematical computations and error estimations have so far been

123: carried out only in \emph{polar coordinates}, describing the

124: distance and direction towards the nest with respect to the animals

125: current position and walking direction.

126: 

127: Here we introduce a consistent model for {\em egocentric} path

128: integration using \emph{cartesian coordinates}. The mathematical

129: computation of current nest position relative to the moving arthropod

130: reduces to an inhomogeneous linear differential equation system of two

131: variables, namely \emph{forward moving speed} and \emph{angular

132:   turning rate}. Different ways of their estimation and further

133: processing implemented into this model provide several possibilities

134: to reproduce some observed reorientation phenomena of arthropods.

135: 

136: The article is organised as follows. We finish the introductory

137: Section reviewing more details on signal and information processing in

138: homing arthropods. The next Section 2 is devoted to previous models of

139: path integration. Section 3 presents our model approach and compares

140: it to those in the previous Section. In Section 4 we turn to

141: systematic errors in homing paths which have been studied extensively

142: and may give hints about internal information processing. Taking

143: advantage of the simple structure of our model we propose error tests

144: for future experiments. Finally we discuss our results and briefly

145: comment on possible neural realizations of egocentric cartesian path

146: integration and possible generalisations including global vectors for

147: feeding sites.

148: 

149: \subsection{Required information for path integration}  

150: 

151: Allothetic and idiothetic signals provide the arthropod with the

152: information required for path integration. During the following

153: description of these signals, we shall focus on desert arthropods and

154: here, in particular, on desert ants \emph{Cataglyphis fortis} and

155: \emph{Cataglyphis bicolor}, because more is known on them as compared

156: to other arthropods. Nevertheless, we shall also mention similar

157: investigations on other arthropods in order to firmly base the

158: modelling principles, mainly because neurobiological analyses have been

159: performed on larger and more easily accessible arthropods 

160: % AEND 0807 such as crickets or locusts 

161: \citep{weh_03a}.

162: 

163: \subsubsection{Allothetic signals}

164: 

165: Without doubt, the main allothetic signals to be considered are visual

166: inputs. Among them are landmarks which, however, have limited overall

167: value as described above. More reliable turn out to be optical sources

168: indefinitely far away: with regard to desert ants, spectral skylight

169: gradient, sun position, and the pattern of polarised skylight are the

170: most important cues

171: \citep{WEH97a,WEH97b,weh_01a,weh_03a,WEH03b}. These three can work

172: without help of each other, as was shown in experiments where one or

173: even two of them had been made inoperative \citep{mul_88,WEH97a}. In

174: all three the ants continuously use the sky as a reference to determine

175: their body axis orientation.

176: 

177: Whenever the arthropod applies {\em spectral cues}, it makes use of

178: the fact that light waves with their different wavelengths are not

179: equally distributed over the illuminated sky. This ability was shown

180: by \cite{WEH97a} for ants and \cite{ros_84} for bees.

181: 

182: Direct orientation with respect to the {\em azimuthal position} of the

183: sun (or any other light source) has been found in ants and many other

184: arthropods (e.g.\ bees, \citealp[von][]{fri_50}; or spiders,

185: \citealp{goe_58}).

186: 

187: The {\em skylight polarisation} pattern, also referred to as skylight 

188: compass, represents the most effective and stable means for

189: orientation in desert ants. Desert ants are able to see the e--vector

190: polarisation pattern that is produced by scattering of the sunlight at

191: air molecules in the atmosphere. This ability has been found in other

192: arthropods and vertebrates as well, first of all in bees

193: \citep[von][]{fri_49}, but also, for instance, in desert locusts

194: \emph{Schistocerca gregaria} \citep{EGG93}, desert isopods

195: \emph{Hemilepistus reaumuri} \citep{HOF84} or desert beetles 

196: \emph{Parastizopus armaticeps} \citep{BIS99}.

197: %When desert ants use the polarization compass, their

198: %sensory field of ommatidia, each carrying a particular orientation

199: %with respect to their body axis, perceives a relative template of the

200: %polarised skylight pattern, from which, for example, the angle between

201: %body orientation and direction of the e--vectors in the polarised

202: %light can be extracted. Experiments have shown that it is sufficient

203: A small visible section of the sky has been shown to be sufficient to

204: detect rotations \citep{fen_86,weh_94,WEH97b}.

205: 

206: Each rotation of the ant's body axis results in a corresponding change

207: of this orientational angle allowing the ant to measure not only its

208: current body direction relative to the allothetic skylight pattern,

209: but also the rate of its angular rotation, independent of whether it

210: is moving or turning on spot.

211: 

212: Although the polarised light pattern is changing with elevation of the

213: sun, desert ants are able to use a stereotypical projection that

214: resembles the skylight pattern at dawn or dusk, respectively, in their

215: memory \citep{WEH97a,WEH97b,weh_98,weh_01a}. Unlike for the

216: task to find a feeder at different daytimes (\citealp{WEH87};

217: \citealp{weh_93a}; \citealp{dye_94}) sun movement need not be

218: compensated to find back during an excursion, which normally lasts

219: only a few minutes, so the resulting error is negligible.

220: 

221: Although \cite{ron_95} have shown that frontal optic flow influences

222: the ants' odometer, the mechanisms to use allothetic cues for

223: detecting {\em directions} and {\em rotations} seem to be

224: inappropriate for measuring {\em distances} or {\em speeds}.

225: %, allothetic cues

226: %resp.\ visual signals do not seem to be sufficient for measuring

227: %distances or speeds.

228: Therefore, additional tools making use of idiothetic signals are needed.

229: 

230: 

231: \subsubsection{Idiothetic signals}

232: 

233: Far less is known about the use of idiothetic signals for the

234: orientation of desert arthropods. Compared to allothetic signals they

235: seem to be of minor or no importance for detecting

236: directions. Experiments with desert ants have shown that it is

237: possible to predict navigational errors by manipulating the visible

238: section of the sky \citep{WEH97a,WEH97b,weh_98,weh_01a}. Hence, the

239: ants obviously do not even revert to proprioceptive signals if the

240: polarisation compass as standard tool provides strange results. Also

241: desert beetles \emph{Parastizopus armaticeps} rely on the position of

242: the light source and the polarisation compass and seem not to revert

243: to proprioceptive cues in the case of ambiguities

244: \citep{BIS99}. Desert ants \emph{Cataglyphis fortis}, when captured at

245: a feeder and transferred to a test area in a dark flask without any

246: allothetic signals available, immediately after their release do

247: reorientate and set out into their stored home direction (T.~Merkle,

248: personal observation). Thus, any possibly existing idiothetic signals

249: do not have an effect on the ant's reorientation under such

250: conditions.

251: 

252: On the other hand, \cite{ron_95} found that desert

253: ants are able to estimate their walked distances without allothetic

254: signals. Therefore they proposed that ants use odometers that mainly

255: rely on proprioceptive signals. This is backed by investigations that

256: could eliminate energy consumption as possible cue for measuring

257: speeds or walking distances, when ants walk along slopes \citep{woh_02} or

258: with heavy load \citep{sch_93}. It seems quite obvious that such

259: proprioceptive signals derive from movements of the legs (for bristles

260: as mechanoreceptors cf.\ \citealp{kei_97}). 

261: 

262: \section{Previous models of path integration}

263: \label{prevmod}

264: 

265: \subsection{Geocentric models}

266: \label{prev_geocent}

267: 

268: The term path integration was established by \cite{mit_73}, referring

269: to a simple and evident mathematical algorithm, namely to determine

270: the ant's current position  by integrating its moving direction

271: $\bm{\theta}(s)$ along the migration path or, equivalently, by

272: integrating its velocity vector $\bm{V}(t) = v(t) \; \bm{\theta}(t)$ over

273: time $t$, where $v(t)$ denotes the ant's forward speed.

274: This yields the estimated final positional vector $\bm{P}$ from start

275: to end point of the path. The estimated global {\em home vector} is

276: then the inverse vector $\bm{G} = -\bm{P}$.

277: 

278: \subsubsection{Cartesian coordinates}

279: \label{prev_geo_cart}

280: 

281: \cite{mit_73,MIT82} use \emph{cartesian coordinates} to represent the

282: integrated positional vector $\bm{P}=(x,y)$ and the current

283: directional vector $\bm{\theta} = (\cos \phi, \sin \phi)$, where the

284: animal's angular orientation $\phi$ is given relative to some

285: reference direction, e.g.\ skylight polarisation. This is chosen as

286: the initial moving direction of the arthropod in their case study of

287: the spider \emph{Agelena labyrinthica}. In their model the arthropod

288: is supposed to estimate the current angle $\phi$ in two different

289: ways: by using an \emph{idiothetic azimuth storage} as integrated

290: value of its proprioceptively measured turning rates $\omega(t)$ along

291: the previous path (this idea of azimuth integration is related to the

292: earlier theory by \citealp{jan_57}), or by directly measuring the

293: \emph{allothetic azimuth value} $\phi$ of the current body axis with respect

294: to an exogenous direction. 

295: %of the sun or an experimental light source. 

296: Then, the resulting two inputs of the directional vector 

297: $\bm{\theta}$ are weighted and summed up for path

298: integration. With the aid of this model the authors could reproduce

299: typical two--segment experiments, during which a light source had been

300: turned by $90^\circ$.

301: 

302: It should be noticed that the described model of \cite{mit_73} is a

303: closed-loop control system (\emph{``Wirkungsgef\"uge''}) applicable to

304: any excursion of an arthropod, including the outbound route towards a

305: food source position $\bm{P}_0$, whose cartesian coordinates have to

306: be internally stored, or the homing route towards the origin

307: $\bm{P}_0=(0,0)$. In both cases the control system produces an

308: efferent motor signal for the turning angle $\omega$ being negatively

309: proportional to $|\bm{P}_0 - \bm{P}| \sin \delta$, where $\delta$ is

310: the deviation angle between current moving direction $\phi$ and the

311: global vector $\bm{G} = \bm{P}_0 - \bm{P}$.  By this steering

312: algorithm, the arthropod will turn into the direction of the global

313: vector and walk towards $\bm{P}_0$ until the global vector is

314: zero. Thus, the path integration mechanism is supposed to work during

315: the whole excursion.

316: 

317: This comprehensive navigation model has been adapted to experiments

318: with rodents \citep[e.g.][]{mit_80,ben_97} or humans \citep{mit_01}

319: and recently up--dated to be consistent with results on neural

320: activity patterns in the hippocampus of mammals \citep{mit_00}. It

321: provides important principles of information processing, path

322: integration, and motor control, and is mathematically easily

323: realizable in computer simulation programs.

324: Nevertheless, it remains

325: open, whether and how the necessary computational steps of calculating

326: trigonometric functions are physiologically performed within the

327: neural nets of arthropods or mammals. Moreover, this model supposes

328: two successive integrating mechanisms including the necessity to store

329: the computed variables, first $\phi = \int \omega \, dt$ and then

330: $\bm{P} = \int (\cos \phi, \sin \phi) \, ds$, if integration is 

331: over walked distance, $s$ denoting arc length, or, 

332: $\bm{P} = \int (v \bm{\theta})(t) \, dt$, if integration is over time. 

333: In the last case, besides determination

334: of the turning speed $\omega$, also that of forward speed $v$ is

335: required (see Table 1).

336: 

337: The extensive work on neural ``head cell'' and ``place cell'' dynamics

338: in mammals could be mentioned here, with the remarkable property that

339: a change in angular information, represented by head cells, can induce

340: a corresponding rotation of the two--dimensional activity pattern in

341: the imaginary chart represented by the array of place cells

342: \citep{sam_97,mit_00}. With regard to {\em desert ants}, \cite{weh_03a}

343: postulates the existence of a certain number of ``compass neurons'',

344: each with an accurately defined compass direction resulting in a

345: maximum firing rate of the respective neuron, whenever the arthropod

346: is heading into that direction. 

347: 

348: \subsubsection{Polar coordinates}

349: \label{prev_geo_pol}

350: 

351: Geocentric models in {\em polar coordinates} have been developed and

352: applied during the last 20 years by \cite{weh_86} and \cite{mul_88},

353: taken up by \cite{har_95} in connection with dynamical representations

354: by cyclical neural chains. In order to represent the actual position

355: vector $\bm{P} = r \, (\cos\nu, \sin\nu)$ of the arthropod, the

356: proposed model algorithms require the animal to compute, at least

357: approximately, distance $r$ from the nest and angle $\nu$ of the

358: position vector relative to an allothetic reference direction, 

359: determined by the sun or, most frequently, by polarised skylight.

360: Then, the global home vector is $\bm{G} = -\bm{P} = r \,

361: (\cos(\nu+\pi), \sin(\nu+\pi))$.

362: 

363: For a segmented path with step length $s_n$ (taken to be $1$ for

364: simplicity), \cite{mul_88} derive approximate recursive formulas for

365: updating the polar coordinates $(r_n,\nu_n)$ after the $n$th moving

366: step. The only additional input needed during each step, besides

367: knowing step length (or measuring forward speed), is the angle

368: $\tilde \delta_n = \phi_n - \nu_n$ between moving direction $\phi_n$

369: and the direction $\nu_n$  of the positional vector. In a continuous

370: description the corresponding general path integration formulas read

371: $r = \int \cos(\phi-\nu) \; ds$ and $\nu = \int \sin(\phi-\nu)/r \;

372: ds$  and are equivalent to a

373: system of nonlinear ordinary differential equations (see also Table 1)

374: \begin{eqnarray}

375: \frac{dr}{dt} & = & v \; \cos(\phi-\nu) \label{e01} \\

376: \frac{d\nu}{dt} & = & \frac{v}{r} \; \sin(\phi-\nu). \label{e02}

377: \end{eqnarray}

378: Again, as in Section 2.1.1, this mathematical integration algorithm 

379: requires the ability to

380: calculate the nonlinear trigonometric functions and, in addition, to

381: perform the division by distance $r$. \cite{mul_88} and \cite{har_95}

382: suggested the trigonometric functions could be approximated by

383: piecewise linear or polynomial functions. This led to a systematic

384: misestimation of increments in both variables, $r$ and $\nu$, for

385: moving directions not parallel or antiparallel to the position vector,

386: i.e.\ for $\tilde \delta = \phi - \nu \neq \pm \pi$. \cite{mul_88}

387: thus could remarkably well reproduce systematic errors in the angular

388: component $\nu$  of the global vector, observed in the classical

389: two--segment experiments, not only for desert ants but also for most

390: arthropods and mammals \citep[e.g.][]{bis_57,goe_58,mul_88,seg_93}. In

391: Section 4.3 we give a detailed analysis of this topic.

392: 

393: \subsection{Egocentric models}

394: \label{prev_egocent}

395: 

396: Another approach, which appears to be more adequate but came into

397: consideration much later, is to model the path integration process of

398: a moving animal in terms of a moving coordinate frame centred around

399: the animal's body, thus reflecting the fact that it perceives all

400: sensory inputs relative to its own position and

401: orientation. \cite{ben_90} chose \emph{polar coordinates} to represent

402: the global vector $\bm{G} = -\bm{P} = r \, (\cos \delta, \sin

403: \delta)$, where now the reference direction for $\delta = 0$ is the

404: body axis, serving as $X$--axis of the corresponding  cartesian

405: coordinate frame with the orthogonal lateral $Y$--axis, see

406: Fig. 1. Although the distance variable $r$ is the same as in the

407: geocentric polar model, the derived recurrent formulas for updating

408: $r_n$ and $\delta_n$ turn out to be much more complicated than any

409: other formula used before. First, egocentric cartesian coordinates

410: $(X_n,Y_n)$ are updated in terms of the former polar ones,

411: \begin{eqnarray}

412: X_{n+1} & = & r_n \, \cos(\delta_n - \omega_n) - s_n \label{e03} \\

413: Y_{n+1} & = & r_n \, \sin(\delta_n - \omega_n) \label{e04}

414: \end{eqnarray}

415: where $\omega_n$ denotes the change of the direction and $s_n$ the

416: length of the subsequent step. (An interchange of the order of stepping

417: and turning would give $s_{n+1}$ in (\ref{e03}) but no fundamental

418: change). Then, these equations are transformed into the new egocentric

419: polar coordinates:

420: \begin{eqnarray}

421: 	r_{n+1} & = & \sqrt{{X_{n+1}}^2 + {Y_{n+1}}^2} \label{e05}\\

422: 	\delta_{n+1} & = & \arctan \left( {\frac{Y_{n+1}}{X_{n+1}}}

423: 	\right) \label{e06}

424: \end{eqnarray}

425: (in order to calculate the correct values of $\delta_{n+1}$ the signs

426: of $X_{n+1}$ and $Y_{n+1}$ have to be considered, see

427: \citealp{ben_95}).

428: 

429: The advantage of this egocentric model is that now the only

430: \emph{input variables} are step length $s_n$ and turning angle

431: $\omega_n$, or, in the corresponding continuous path integration

432: model, forward speed $v$ and angular turning rate

433: $\omega$. \cite{GAL90} considered the corresponding differential

434: equation using egocentric polar coordinates in the continuous limit of

435: infinitely small time steps which, in the corrected formulation by

436: \cite{ben_95}, are

437: \begin{eqnarray}

438: \frac{dr}{dt} & = & - v \; \cos \delta  \label{e07} \\

439: \frac{d\delta}{dt} & = & v \; \frac{\sin \delta}{r} -

440: \omega. \label{e08}

441: \end{eqnarray}

442: These equations can directly be obtained form the corresponding polar

443: coordinate equations,  Eqs.~(\ref{e01}) and (\ref{e02}), by performing 

444: a simple angular transformation, $\delta = \nu + \pi - \phi$, so that 

445: again they require to compute division by

446: $r$ and trigonometric functions (see also Table 1).

447: 

448: In their simulation analysis, \cite{ben_90} studied the influence of

449: random errors on the estimation of changes of direction and walking

450: distance. With regard to directional changes, they distinguished

451: between allothetical and idiothetical orientation: they considered

452: idiothetical estimation as ``measuring the change of direction

453: itself'', whereas the allothetical estimation is defined as ``a

454: comparison between the heading of current and previous step relative

455: to some exteroceptive compass'' \citep{mau_95}. The different

456: estimation procedures were realized by providing the actual values

457: with normal distributed errors.

458: 

459: In their simulations random errors of allothetic signals had only

460: little influence, whereas those of idiothetic signals lead to

461: noticeable misestimation. \cite{ben_95} conclude that the egocentric

462: coding process is quite sensitive to idiothetic errors and organisms

463: relying on allothetic cues for measuring directional changes by far

464: outmatch those relying on idiothetic cues.

465: 

466: \section{Cartesian model for egocentric path integration}

467: \label{sectcartmod}

468: 

469: \subsection{System of linear differential equations for the global vector}

470: 

471: The physiological sensing and locomotion apparatus of any arthropod is

472: completely bound to its body. It is therefore naturally related to its

473: two symmetry axes, the posterior-anterior axis and the perpendicular

474: right-left axis. Thus, when identifying these symmetry axes with the

475: $X$ and $Y$ axis of a cartesian coordinate frame $(X,Y)$ and taking

476: the arthropod's body centre as the origin $(0,0)$, this

477: constitutes a proper planar moving coordinate frame for representing

478: the {\em relative position} of any object in the planar neighbourhood

479: of the arthropod, e.g.\ its nest. In this {\em egocentric cartesian}

480: model the global vector pointing from the arthropod's body to the

481: nest, relative to the animal's actual body axis orientation, is just

482: $\bm{G} = (X,Y)$, corresponding to the same vector as in Section

483: \ref{prev_egocent}, there only written in polar coordinates, see

484: Fig.~\ref{fig01}.  Notice that the original ansatz by \cite{ben_90}

485: already mentioned this cartesian coordinate system, but then switched

486: to polar coordinates for path integration (see Section

487: \ref{prev_egocent}). Indeed, the continuous version of 

488: Eqs.~(\ref{e03}) and (\ref{e04}), given the  arthropod's

489: {\em forward speed v} and {\em angular turning rate $\omega$}, 

490: yields the following model equations for a precise update of 

491: the global vector $(X,Y)$ during motion

492: \begin{eqnarray}

493: 	\frac{d X}{d t} & = & - v + \omega \cdot Y \label{e09}	\\

494: 	\frac{d Y}{d t} & = & - \, \omega \cdot X \label{e10}

495: \end{eqnarray}   

496: Compared to all other continuum equations or analogous discrete

497: recursion algorithms developed previously (see Section 2), this

498: two--dimensional differential equation system is remarkably simple: It

499: is linear in the two variable quantities $X$ and $Y$, and it just uses

500: the two speed input parameters as additive or multiplicative terms,

501: $v$ representing the rate of shifting the $X$ coordinate backwards,

502: $\omega$ the rate of rotating the $(X,Y)$ frame clockwise. These

503: operations can be easily performed by any suitable elementary analogue

504: circuit network like the one shown below in Fig.~\ref{fig03}. Possible

505: physiologically realizable neural representations are briefly

506: addressed in the Discussion Section~\ref{discussion}.

507: 

508: In conclusion, this {\em cartesian egocentric path integration model},

509: considered as a precise `dead reckoning' system, offers the

510: most simple computational scheme to determine the global vector and,

511: simultaneously, being related to a coordinate frame of the moving

512: arthropod. For comparison with the other models see

513: Table~\ref{table1}. 

514: 

515: Clearly, biological solutions of difficult problems can be complex and

516: the simplicity of our model does not make its realisation more likely

517: than that of other models. From a conceptional point of view the

518: existence of a simple solution is nevertheless striking, and in the

519: following we shall present its implementation and results.

520: 

521: 

522: \begin{figure}[!ht]

523: \begin{center}

524: \includegraphics[width = 10cm]{fig01.eps}

525: \end{center}

526: \caption{Theoretical scheme of egocentric path integration by means of

527:   cartesian coordinates $X$, $Y$ specifying the position of the nest

528:   relative to the arthropod's body axes and determining the global

529:   vector $\bm{G} = (X,Y)$, here with $X < 0$, $Y > 0$. In contrast,

530:   the corresponding model in polar coordinates (Section

531:   \ref{prev_egocent}) uses the two variables $r$, distance to the nest,

532:   and  $\delta$, angle between head orientation and nest direction.}

533: \label{fig01}

534: \end{figure} 

535: 

536: 

537: 

538: \begin{table}[!h]

539: \hspace{-2cm}

540: \begin{tabular}[t]{ll|cl|l}

541: & & Input variables & Internal variables & Global vector \\

542: \hline

543: \vspace{-3mm}

544: & & & & \\

545: \parbox{2cm}{Geocentric}

546: & 

547: \parbox{2cm}{Cartesian \\ {\it (Sect.~\ref{prev_geo_cart})} \\ Polar

548:   \\ {\it (Sect.~\ref{prev_geo_pol})} }

549: &

550: \parbox{2.5cm}{$\omega$,$v$ or $\omega_n$,$s_n$ \\ \\ $\phi$ \\ $s_n$ or $v$}

551: &

552: \parbox{3.5cm}{$\phi$, $\bm{P} = (x,y)$ \\

553:   $\langle$lin./nonlin.\ ODE$\rangle$ \\ $r$ and  $\nu$ \\

554:   $\langle$nonlinear ODE$\rangle$}

555: &

556: \parbox{4cm}{$\bm{G} = \bm{P}_0 - \bm{P}$ \\ \\ $\bm{G} = - r \; (\cos

557:   \nu,\sin \nu)$}

558: \\

559: \vspace{-3mm}

560: & & & & \\

561: \hline

562: \vspace{-3mm}

563: & & & & \\

564: \parbox{2cm}{Egocentric}

565: & 

566: \parbox{2cm}{Polar \\ {\it (Sect.~\ref{prev_egocent})} \\ Cartesian

567:   \\ {\it (Sect.~\ref{sectcartmod})} }

568: &

569: \parbox{2.5cm}{$\omega$,$v$ or $\omega_n$,$s_n$}

570: &

571: \parbox{3.5cm}{$r$ and $\delta$ \\ $\langle$nonlinear ODE$\rangle$ \\

572:   $X$ and  $Y$ \\ $\langle$linear ODE$\rangle$}

573: &

574: \parbox{4cm}{$\bm{G} = r \; (\cos \delta , \sin \delta )$ \\ \\

575:   $\bm{G} = (X,Y)$}

576: \\

577: \hline

578: \end{tabular}

579: \caption{Path integration models with the parameters and variables

580: used for input, internal calculation and output as global vector. For

581: notations and more details see the various model descriptions in the

582: text.

583: }

584: \label{table1}

585: \end{table}

586: 

587: \subsection{Modelling foraging excursions, reorientation, and homing}

588: \label{modelexcurs}

589: 

590: %A biologically important advantage of o

591: Our path integration model

592: %,also shared by the models in Sections 2.1.1 and 2.2, is that it only

593: depends on the values of $v$ and $\omega$, cf.~Table 1. 

594: These can be regarded as the elementary physiological control

595: variables which the arthropod uses to steer its locomotion, e.g.\ by

596: changing speed or frequency of striking leg motion on both sides or,

597: respectively, on one side relative to the other.

598: 

599: In order to model typical paths of directionally persistent random

600: walks, as observed for desert arthropods, one has to account for mean

601: values and standard deviations of speed $v$ and turning rate $\omega$

602: as well as for their temporal auto-correlations which can be extracted

603: from corresponding experimental time series

604: \citep[see][]{ALT90}. Discrete correlated random walk models sometimes

605: used \citep[e.g.][]{bye_01} are not adequate as they assume piecewise

606: constant walking directions $\phi_n$ and turning angles

607: $\omega_n$. The two speeds $v(t)$ and $\omega(t) = d\phi(t)/dt$,

608: however, being related to the physiologically controlled, relatively

609: fast leg movement on both sides of the arthropod, should better be

610: modelled as fluctuating continuous processes on an adequate smoothness

611: level. The simplest stochastic process of this kind is described by

612: the following two independent Ornstein--Uhlenbeck equations for first

613: order coloured noise, which have previously been used also for

614: modelling the systematic search of arthropods \citep{alt_95},

615: \begin{eqnarray}

616: 	d v & = & \frac{1}{T_v} \; (v_0 - v)  \; dt + \beta_v \; dW_t

617: 	\label{e11} \\

618: 	d \omega & = & \frac{1}{T_\omega} \; (\omega_0 - \omega) \; dt

619: 	+ \beta_\omega \; dW_t \label{e12}	 

620: \end{eqnarray}

621: Following the standard approach, random perturbations are expressed as

622: additive Wiener increments $dW_t$ \citep{ITO65}. In simulations one

623: uses a sequence of values $v$ and $\omega$ with {\em finite} time

624: differences $\tau$,

625: \begin{eqnarray}

626:   v_{t+\tau} & = & v_t + \frac{\tau}{T_v} \; (v_0 - v_t) +

627:   \beta_v \sqrt{\tau} \; \zeta \\

628:   \omega_{t+\tau} & = & \omega_t + \frac{\tau}{T_\omega} \; (\omega_0

629:   - \omega_t) + \beta_\omega \sqrt{\tau} \; \zeta

630: \end{eqnarray}

631: where $\zeta$ denotes a standard normally (${\cal N}(0,1)$) distributed

632: random variable, drawn independently at each step for each

633: variable. Eqs.~(\ref{e11}) and (\ref{e12}) are obtained in the limit

634: $\tau \to 0$. $v_0$ and $\omega_0$ are the preferred values of forward

635: and turning speed, respectively, and the $T_{v|\omega}$ denote the

636: corresponding mean persistence times of fluctuations with amplitudes

637: $\beta_{v|\omega}$. In case of a stationary time series they yield

638: variances of size $\sigma_{v|\omega}^2 = \beta_{v|\omega}^2

639: T_{v|\omega}/2$ by an equilibrium of perturbations with strength

640: $\beta_{v|\omega}$ and decay at rate $1/T_{v|\omega}$. In our

641: presentation we assume, for simplicity, that $T_v$ is negligibly

642: small, such that during locomotion the forward moving speed has a

643: constant value $v \equiv v_0$. The presented results also hold for the

644: general case of fluctuating forward speed.

645: 

646: 

647: \begin{figure}[!ht]

648: \begin{center}

649: \includegraphics[width = 10cm]{fig02.eps}

650: \end{center}

651: \caption{Model simulation of a natural outbound path and the

652:   successful return path back to the nest due to precise path

653:   integration, according to the 3-Phase-model, see text. Parameters

654:   used for calculations are, (1) for the outbound path $T_\omega =

655:   0.3$~s, $\beta_\omega = 1$~s$^{-3/2}$, and constant forward speed $v

656:   \equiv v_0 = 0.2$~m/s, (2) constant $\omega_{\rm rot} = 1$~s$^{-1}$

657:   during rotation and (3) for homing the same as in (1) but a feedback

658:   constant $c = 1/0.05$~s$^{-1}$ for beacon steering. (Top left): Plot

659:   of the actual path in cartesian $(x,y)$-coordinates of an

660:   observer. Position of the nest at $(0,0)$ is marked by a filled

661:   circle. (Top right): Corresponding plot of the nest position in the

662:   same scale in relative cartesian

663:   $(X,Y)$-coordinates, where the origin denoting the home position is

664:   marked by a filled circle. Note that the animal's head direction is

665:   the $X$--axis pointing upwards, while the lateral $Y$--axis points

666:   to the left. (Bottom): Corresponding plots of $X$ and $Y$ over time.

667: }

668: \label{fig02}

669: \end{figure}

670: 

671: 

672: % 00000000000000000000000000000000000000000000000000

673: %Noch einzufuegende Teile:

674: %

675: %see Fig.~\ref{fig01} 

676: %

677: %Thus, our model does not need to assume calculation of any angles,

678: %particularly, the angle $\delta$ between  body orientation and nest

679: %position, as it is supposed in the polar model (see Section

680: %\ref{prev_egocent}). Small angular deviations $\delta$ from the homing

681: %course are here represented by the equivalent deviation value $Y = r

682: %\sin \delta$.

683: % -------------------------------------------------------

684: 

685: In the following, we describe the three successive phases of an arthropod's

686: typical excursion, using the egocentric path integration system

687: Eqs.~(\ref{e09}), (\ref{e10}) and the physiological motion control system

688: Eqs.~(\ref{e11}) and (\ref{e12}). We explain the corresponding

689: dynamics of the global vector by means of a simulated example

690: presented in Fig.~\ref{fig02}.

691: 

692: \begin{itemize}

693:    \item{Phase 1:} {\bf (Foraging)} \\

694:      The arthropod starts foraging at the nest site, e.g.\ $(x_0,y_0)

695:      = (0,0)$, where the global vector is reset to zero $\bm{G} = (X,Y)

696:      = 0$. Holding the mean turning rate $\omega_0 = 0$, the animal

697:      approximately keeps its chosen initial direction, $\phi =

698:      \phi_0$, for some time, leading it almost straight away from the

699:      nest, corresponding to increasingly negative $X$ values of the

700:      internal global vector, while the $Y$ component stay close to

701:      zero. This initial behaviour is well expressed in the example of

702:      Fig.~\ref{fig02}, then followed by a random right-hand turn of

703:      the $(x,y)$-path, which corresponds to increasingly $Y$ values

704:      meaning that now the nest lies to the right side the animal.

705:    \item{Phase 2:} {\bf (Reorientation)} \\

706:      After finding food at some position $\bm{P}_1 = (x_1,y_1)$, the

707:      arthropod stops there in its current angular orientation, $\phi =

708:      \phi_1$, keeping the actual global vector $\bm{G}_1 = (X_1,Y_1)$

709:      internally stored (even during handling the food). Then the

710:      arthropod starts its reorientation phase by turning on spot, say

711:      with constant rotation speed $\omega \equiv \pm \omega_{\rm

712:      rot}$, depending on whether the stored global vector $\bm{G}_1$

713:      has positive or negative $Y_1$ value. During rotation the global

714:      vector $\bm{G}=(X,Y)$ also rotates according to path integration

715:      in Eqs.~(\ref{e09}) and (\ref{e10}), since now we set $v \equiv

716:      v_0=0$. Finally, the arthropod is assumed to stop its rotation

717:      ($\omega =0$) as soon as the condition $Y=0$ is fulfilled,

718:      meaning that now its head is oriented towards the nest and the

719:      actually positive $X$-value represents the arthropod's distance

720:      from the nest (cf. the scheme in Fig.~\ref{fig03}). For the

721:      example in Fig.~\ref{fig02}, see the counter-clockwise rotation

722:      circle ending on the positive $X$ axis. Notice that the condition

723:      ${Y=0}$ corresponds to ${\delta = 0}$ in egocentric polar

724:      coordinates (Section 2.2 and Fig.~\ref{fig01}) because of the

725:      equivalences  $Y = r \cdot \sin\delta$ ($\approx r \delta$ for

726:      small $\delta$) and $X = r \cdot \cos\delta$ ($\approx r +

727:      r\delta^2/2$ for small $\delta$).

728:    \item{Phase 3:} {\bf (Homing)} \\        

729:      The arthropod now returns back to the nest (as straight as

730:      possible) according to the actually stored global vector

731:      $\bm{G}=(X,0)$. Since during walking the global vector will be

732:      constantly updated and, due to inevitable random perturbations,

733:      the $Y$ component will eventually deviate from the zero value,

734:      the arthropod must tend to hold the internal {\em steering

735:      condition} $Y=0$ as closely as possible. This can be modelled by

736:      implementing a {\em counter-steering} turning rate $\omega_0 = -

737:      c Y$ into the stochastic differential equation (\ref{e12}). See

738:      Fig.~\ref{fig03} for an analogue circuit scheme describing this

739:      feedback control, which nonlinearly and cyclically couples the

740:      linear path integration system, Eq.~(\ref{e09}) and (\ref{e10}),

741:      to the linear motor control equation (\ref{e12}).

742: 

743:      Finally, in our model the arthropod is assumed to stop its return

744:      phase as soon as the $X$ value of its global vector becomes

745:      zero. The resulting home path in the simulated example of

746:      Fig.~\ref{fig02} clearly shows, how some random perturbations

747:      lead to small deviations in the homing direction of the $(x,y)$

748:      path and corresponding small $Y$ deviations of the global vector,

749:      while the $X$ component is almost linearly decreasing to zero.

750: \end{itemize}

751: 

752: Notice that, according to this modelling scheme, the path integration

753: system in Eqs.~(\ref{e09}) and (\ref{e10}) is supposed to work

754: constantly in the arthropod's neural system during foraging,

755: reorienting, and homing, except when the animal is seriously perturbed

756: and not any more able to `measure' and `control' its forward motion

757: and directional turning. If this happens, the animal is assumed to

758: instantaneously halt the path integration system and keep the actual

759: value of the egocentric global vector and the orientational angle

760: (with respect to an allothetic visual cue) stored until it can proceed

761: in an unperturbed way.

762: 

763: Let us emphasise, that the presented {\em internal} dynamics of the

764: global vector $\bm{G}=(X,Y)$ determined by the simple linear system

765: differential equations (\ref{e09}) and (\ref{e10}), could equivalently

766: be described in polar coordinates $\bm{G} = r \, (\cos \delta, \sin

767: \delta)$ using the more complicated nonlinear differential equations

768: (\ref{e07}) and (\ref{e08}), including the `rotation stop condition'

769: $\delta = 0$ and the counter-steering term $\omega_0 = - c \, r \cos

770: \delta$ or a stronger variant like  $\omega_0 = - \tilde{c} \,

771: \delta$. However, there is an important difference in modelling the

772: `nest stop condition': In polar coordinates, the obvious termination

773: criterion would be chosen as $r = 0$ meaning that the global vector

774: $\bm{G}$ becomes exactly zero. It remains to be proven, which

775: counter-steering rule could guarantee that this condition is

776: attainable for stochastically perturbed random paths.

777: 

778: In contrast, the proposed termination criterion $X = 0$ in cartesian

779: coordinates would, for randomly perturbed return paths, generically

780: result in a non-vanishing small $Y$ value, then representing the

781: lateral distance of the arthropod to the nest. Thus, depending on this

782: value and on the current orientational angle of the animal, the

783: realized `stop position' can fluctuate around the true nest position,

784: even in the so far considered case of {\em precise path

785:   integration}. The size of this random error increases with the

786: length of the home vector, i.e.\ the distance between food and

787: nest. This corresponds to experimental observations (e.g.\ for desert

788: ants {\em C.~fortis} T.~Merkle, unpublished data) which furthermore

789: show that the length of the outbound path also contributes to such a

790: positional error. Therefore, other errors in path integration, being

791: accumulated  along the path, have also to be considered, which is the

792: topic of the following Section.

793: 

794: \begin{figure}[!ht]

795: \begin{center}

796: \includegraphics*[width = 14cm]{fig03.eps}

797: \end{center}

798: \caption{Analogue circuit scheme of the egocentric cartesian path

799:   integration model: dynamics of the two variables $X$ and $Y$

800:   according to the differential equations~(\ref{e09}) and (\ref{e10})

801:   and its coupling to the physiological control parameters

802:   represented by the two speed parameters for turning, $\omega$, and

803:   forward locomotion, $v$. In {\bf Phase 1}, $\omega$ and $v$ are

804:   externally controlled (random search, trained path towards feeder,

805:   \dots). During {\bf Phase 2} and {\bf Phase 3} the $X$ and $Y$

806:   values feed back into the speed control conditions such that, by

807:   counter--steering with respect to the `internal beacon' $Y=0$ in the

808:   latter case, equations~(\ref{e09}), (\ref{e10}) and (\ref{e12})

809:   constitute a coupled nonlinear control system along the homing

810:   path.}

811: %  two differential equations at

812: %  the heart of our model. $X$ and $Y$ influence each other's

813: %  differential equation by coupling via $\omega$. Additionally, $X$ is

814: %  changed by $v$. Different inputs activate the  `Go home!' unit,

815: %  which steers the animal to $X = Y = 0$.}

816: \label{fig03}

817: \end{figure}

818: 

819: 

820: \section{Systematic errors}

821: \label{errorsection}

822: 

823: It has been shown that 

824: many arthropods \citep[e.g.][]{bis_57,goe_58,hof_85b,weh_86,mul_88}

825: but also mammals \citep[e.g.][]{seg_93,eti_96,seg_98} exhibit errors

826: in determining the exact homing direction. In general, we have to

827: distinguish between random errors and systematic errors during path

828: integration. There is evidence that random errors, in addition to the

829: home vector steering error mentioned above, can originate from

830: inaccurate measurements of angles or distances, whereas systematic

831: errors probably arise at the neural level of the organism

832: \citep{ben_90,seg_98}. Orientation is less error prone if allothetic

833: reference frames are available, as polarised skylight for

834: arthropods \citep[e.g.][]{weh_98,weh_03a}, but a more difficult task

835: if not, as for mammals \citep{eti_04}. Systematic errors play an

836: important role, as the classical two--leg experiments (L-shaped

837: angular turning tests) have shown in both mammals

838: \citep[e.g.][]{mau_95,eti_96} and arthropods

839: \citep[e.g.][]{mul_88,BIS99}. Apart from  mistakes that concern

840: directional aberrations, there occur also errors 

841: by underestimation of distances

842: \citep{som_04}.

843: 

844: From an evolutionary point of view the presence of systematic homing

845: errors is interesting and has not been explained to date. It may have

846: an advantage that a homing animal typically assumes a shorter path and

847: ends up {\em in front of} its nest. It would then avoid an overshoot

848: and find familiar features that it has just passed on the outgoing

849: path which might help to reach the nest's entrance.

850: 

851: In this Section we implement two types of systematic errors into our

852: model. One concerns the estimation of nest distance following

853: \cite{som_04}, the other exhibit different variations in processing

854: the turns during path integration. All of them predict systematic

855: deviations from correct homeward courses and are based on feasible

856: neural assumptions, or reproduce behaviours that have been observed

857: during experiments. They may serve as a basis for ongoing and future

858: studies of systematic deviations (T.~Merkle, unpublished data).

859: 

860: \subsection{Underestimation of turning angles}

861: 

862: \cite{mul_88} trained desert ants to run through two channels of 10m

863: and 5m length and varied the connecting angle between them in several

864: steps from $0$ to $180^\circ$, see Figs.~\ref{fig04} and

865: \ref{fig05}. The ants miscalculated their covered outbound route and,

866: after leaving the second channel's end, turned about an angle which

867: was {\em larger} than the correct one leading home. The authors

868: reproduced this error very well by a simple formula, which accumulates

869: systematic miscalculations in path integration whenever the animal

870: walks different from the direct inbound and outbound directions. In

871: Fig.~\ref{fig05} the angular aberration function $\varepsilon$ is

872: shown, computed according to the approximative path integration model

873: by \cite{mul_88}. In general, this function fits quite well

874: observations in other arthropods and mammals

875: \citep{seg_93,BIS99}. Here we show that other error models can also

876: reproduce these data. For evaluation and fitting of the corresponding

877: error functions (see plots in Fig.~\ref{fig05}) we use the

878: advantages of our egocentric path integration system.

879: 

880: As a first error mechanism we consider a {\em systematic

881: underestimation} of  body axis rotation. In principle, that error

882: could occur during the estimation of $\omega$, i.e.\ by simply

883: perceiving a value lower than the actual value, or on the neural

884: level. The high accuracy concerning the ability of desert arthropods

885: measuring rotations makes it very likely that this error may be

886: created on the neural level. Therefore, we assume the animal perceives

887: the correct value $\omega_{\rm real}$, but uses a different value

888: $\omega_{\rm proc}$ for processing the path integration according to

889: the differential equations (\ref{e09}) and (\ref{e10}).

890: 

891: In a first choice the underestimation is taken to be a linear function

892: of the real value,

893: \begin{equation}

894: \omega_{\rm proc} = \lambda \; \omega_{\rm real}

895: \label{1a_underest}

896: \end{equation}

897: with a factor $\lambda \! < \! 1$ (error {\sf LU} in Table~\ref{table2}). In

898: a second variant, the fully saturated underestimation, $\omega_{\rm

899:   real}$ is processed correctly for small values but saturates towards

900: a certain maximal turning rate $\omega_{\rm c}$ (error {\sf NLUs} in

901: Table~\ref{table2}),

902: \begin{equation}

903: \omega_{\rm proc} = \frac{\omega_{\rm c}}

904: {\omega_{\rm c} + |\omega_{\rm real}|} \; \omega_{\rm real}

905: \label{1b_underest}

906: \end{equation}

907: A linear combination of both is given by the error {\sf NLU} in

908: Table~\ref{table2},

909: \begin{equation}

910: \omega_{\rm proc} = \left( \lambda + \frac{1 - \lambda}{\omega_{\rm c} +

911:   |\omega_{\rm real}|} \; \omega_{\rm c}\right) \; \omega_{\rm real}

912: \label{1c_underest}

913: \end{equation}

914: which again processes small values correctly.

915: 

916: For a related choice of errors we assume a temporal delay $\tau_{\rm

917:   del}$ in processing the information of $\omega_{\rm real}(t)$; 

918: the same underestimation could, in principle, be assumed also 

919: for $v_{\rm real}(t)$, but here variation on

920: natural outbound paths is rather low (T.~Merkle, personal observations

921:   on desert ants \emph{Cataglyphis fortis}). Phenomenologically such a

922: delay is implemented by a linear ordinary differential equation

923: representing a first order filtering process, namely

924: \begin{equation}

925: \frac{d \omega_{\rm proc}}{dt} = \frac{\omega_{\rm real} - \omega_{\rm

926:     proc}}{\tau_{\rm del}}

927: \label{1d_delay}

928: \end{equation}

929: such that $\omega_{\rm proc}$ is {\em smeared out} on a scale of

930: $\tau_{\rm del}$ as compared to $\omega_{\rm real}$ (error {\sf PD} in

931: Table~\ref{table2}).

932: 

933: \subsection{Underestimation of distance to the nest}

934: 

935: The error due to distance underestimation, which we consider here, has

936: previously been referred to as {\em leaky integrator} by

937: \cite{som_04}. This idea can be implemented into our egocentric cartesian

938: path integration model in a straightforward way: with

939: a constant rate the integrated global vector ``leaks'' or decays from

940: the memory. Thus, the two--dimensional model Eqs.~(\ref{e09}) and

941: (\ref{e10}) are varied by simply adding a proportional decay term

942: in each equation (error {\sf LI} in Table 2)

943: \begin{eqnarray}

944: \dot X & = & - v + \omega_{\rm real} Y - \frac{X}{\tau_{\rm L}}

945: \label{leak_int1} \\

946: \dot Y & = & - \omega_{\rm real} X - \frac{Y}{\tau_{\rm L}}

947: \label{leak_int2}

948: \end{eqnarray}

949: with mean decay time $\tau_{\rm L}$. Also in egocentric polar

950: coordinates, c.f.\ Eqs.~(\ref{e07}) and (\ref{e08}), the leaky

951: integrator is easily expressed by a proportional decay of radial distance

952: $r$, see Table~\ref{table2}.

953: 

954: In the case of a one--dimensional path, e.g.\ always walking along $x$

955: without any turns ($\omega_{\rm real} \equiv  0$),

956: Eqs.~(\ref{leak_int1}) and (\ref{leak_int2}) lead to an exponential underestimation ($x_{\rm

957:   ue}$) of the actual walking distances ($x$) as

958: \begin{equation}

959: x_{\rm ue} = \xi_{\rm L} \left( 1 - \exp(-x/\xi_{\rm L}) \right).

960: \label{sat_dist}

961: \end{equation}

962: The estimated distance $x_{\rm ue}$ saturates at a length $\xi_{\rm

963:   L} = v \tau_{\rm L}$ in the limit of long walking distances $x$, but

964: for short paths $x \ll \xi_{\rm L}$ the error is small and $x \approx

965: x_{\rm ue}$.

966: 

967: This is precisely the best fit to the experiments performed by

968: \cite{som_04}. The authors trained desert ants to walk through linear

969: channels to a feeder. Afterwards the ants were captured at the feeder

970: and released in a linear test channel. They headed off in homeward

971: direction and performed a back and forth search around their assumed

972: nest position. By extracting $x_{\rm ue}$ from the search behaviour,

973: Sommer and Wehner found the relation of Eq.~(\ref{sat_dist}).

974: 

975: In a truly two--dimensional path the leaky integrator of

976: (\ref{leak_int1}) and (\ref{leak_int2}) may also lead to an {\em

977:   angular deviation} of the search path from the true homeward 

978: direction. Earlier sections of the outgoing path have decayed in the

979: memory more than later ones. If the animal has turned in--between,

980: this will result in a {\em different} misestimation of related

981: directions and, consequently, in a homing angle misestimation. In

982: Fig.~\ref{fig04} this is explained for the classical two--leg

983: experiment of \cite{mul_88}.

984: 

985: There is, however, a quantitative mismatch between the fit of the LI

986: equations (\ref{leak_int1}) and (\ref{leak_int2}) to the experiments

987: of \cite{mul_88} and to those of \cite{som_04}. In the latter case one

988: obtains $\xi_{\rm L} \approx 90$m which is substantially different

989: from the value of $18$m of the fit to M\"uller's and Wehner's two--leg

990: experiments. At present we conclude that most likely some part of the

991: error occurs during the turn. To fully answer this contradiction, one

992: would have to take into account more details of the ants' walks, such

993: as, e.g., walking speed or waiting times.

994: 

995: 

996: \begin{table}[!h]

997: \begin{tabular}[c]{l|l}

998: \hline

999: Error & Differential Equations\\

1000: \hline

1001: \hline

1002: {\bf Turning rate underestimation} & \\

1003: \vspace{.1cm}

1004: \begin{tabular}[c]{ll}

1005: {\sf LU} & linear underestimation of $\omega$ \\

1006: &  $ \omega_{\rm proc} = \lambda \; \omega_{\rm real}, \; \; \; \; 0 <

1007: \lambda < 1 $ \\

1008: {\sf NLUs} & fully saturated underestimation \\

1009: & $ \omega_{\rm proc} = \displaystyle\frac{\omega_{\rm c} \omega_{\rm

1010:     real}}{\omega_{\rm c} + | \omega_{\rm real} |} \mbox{ with }

1011: \omega_{\rm c} > 0$\\

1012: {\sf NLU} & partially saturated underest. \\

1013: & $\omega_{\rm proc} = \omega_{\rm  real} \left( \lambda + 

1014: \displaystyle\frac{1 -

1015:   \lambda}{\omega_{\rm c} + | \omega_{\rm real} |} \; \omega_{\rm c}

1016: \right)$ \\

1017: {\sf PD} & processing delay \\

1018: & $\dot \omega_{\rm proc} = \displaystyle\frac{\omega_{\rm real} - \omega_{\rm

1019:     proc}}{\tau_{\rm del}}$

1020: \end{tabular}

1021: &

1022: \begin{minipage}[c]{5cm}

1023: \begin{eqnarray}

1024: \dot X & = & - v + \omega_{\rm proc} \; Y \nonumber \\

1025: \dot Y & = & - \omega_{\rm proc} \; X \nonumber

1026: \end{eqnarray}

1027: \end{minipage}

1028: \\

1029: \hline

1030: {\bf Nest distance underestimation} & \\

1031: \vspace{0.1cm}

1032: \begin{tabular}[b]{ll}

1033: {\sf LI} & Leaky integrator \\

1034: & \\

1035: & egocentric cartesian \\

1036: & \\

1037: & \\

1038: & egocentric polar \\

1039: &

1040: \end{tabular}

1041: &

1042: \begin{minipage}[b]{5.5cm}

1043: \begin{eqnarray}

1044: \dot X & = & - v + \omega_{\rm real} \; Y - X/\tau_{\rm L}

1045: \nonumber \\

1046: \dot Y & = & - \omega_{\rm real} \; X - Y/\tau_{\rm L}

1047: \nonumber

1048: \end{eqnarray}

1049: \begin{eqnarray}

1050: \dot r & = & - v  \cos \delta - r/\tau_{\rm L}

1051: \nonumber \\

1052: \dot \delta & = & v \sin \delta / r - \omega_{\rm real} \nonumber

1053: \end{eqnarray}

1054: \end{minipage}

1055: \\

1056: \hline

1057: \end{tabular}

1058: \caption{Error types in path integration and their respective

1059:   formulae.}

1060: \label{table2}

1061: \end{table}

1062: 

1063: 

1064: Up to now we have proposed different elementary error

1065: mechanisms. Based on current knowledge of sensoric and neural

1066: processes it is not possible to prove or refuse their validity. They

1067: nevertheless lead to non--trivial consequences which are not

1068: immediately visible but can be seen in simulated realizations of outbound routes

1069: together with the path integration procedure.

1070: 

1071: \subsection{Resulting deviations}

1072: 

1073: First, we document the outcome of a simulated experiment as in

1074: \cite{mul_88}. The left side of Fig.~\ref{fig04} shows a sketch of

1075: the two--leg experimental setup which in \cite{mul_88} had lengths $a

1076: = 10$m and $b = 5$m. The ant starts at the nest (open circle), turns

1077: after distance $a$ by an angle $0 \le \alpha \le \pi$ to the right,

1078: leaves the channel after another walked distance $b$ (at the black

1079: circle), thereby overcompensating its turn to the correct home

1080: direction by an angular deviation $\varepsilon$.

1081: 

1082: \begin{figure}[!ht]

1083: \begin{center}

1084: \includegraphics*[width = 10cm]{fig04.eps}

1085: \end{center}

1086: \caption{Angular deviation $\varepsilon$ in the two--leg experiment

1087:   with channels of length $a$ and $b$, respectively, and clockwise

1088:   connecting angle $\alpha$. (Left side): Experimental situation as

1089:   observed in geocentric $(x,y)$--coordinates. (Right side):

1090:   Representation of the global vector in internal

1091:   $(X,Y)$--coordinates; correct representation indicated by thick

1092:   black line, dashed during counterclockwise turn about angle

1093:   $\alpha$. {\em Angle underestimation} leads to turn by $\alpha' <

1094:   \alpha$ and angular aberration $\varepsilon$ after leaving the

1095:   second channel (black dot--dashed line). {\em Leaky integrator} is

1096:   shown in grey, also leading to angular deviation. For more details

1097:   see text.}

1098: \label{fig04}

1099: \end{figure}

1100: 

1101: The right side of Fig.~\ref{fig04} represents this path in the

1102: internal $(X,Y)$--coordinates. First consider a correct processing

1103: without any systematic error: $X$ decreases to $-a$ (thick black

1104: line) and $Y = 0$, then the animal turns by an angle $\alpha$, such

1105: that now $X = - a \cos \alpha$ and $Y = - a \sin \alpha$ (black dashed

1106: line), finally $X$ decreases further to $X = - a \cos \alpha - b$,

1107: whereas $Y$ remains constant (black solid line with black circle).

1108: {\em Angle underestimation} would result in a turn by $\alpha' <

1109: \alpha$ such that in the end $X = - a \cos \alpha' - b$ (dot--dashed

1110: line with black circle) which lies off the true direction to the nest

1111: by an error angle $\varepsilon$. A similar result is obtained by the 

1112: {\em leaky integrator} (fat grey lines): First $X$ decreases from 0 to

1113: $- \xi_{\rm L} \left( 1 - \exp (-a/\xi_{\rm L}) \right)$ and $Y$ remains

1114: 0 (thick grey line), then $X$ and $Y$ are turned by an angle $\alpha$

1115: (dashed grey line). The turn occurs so fast that ``leakage'' can be

1116: neglected ($\tau_{\rm turn} \approx r_{\rm turn}/v \ll \tau_{\rm

1117:   L}$, see also below). During the final decrease of $X$, both the

1118: values of $X$ and $Y$ ``leak'' such that finally

1119: \begin{eqnarray}

1120: X_{\rm L} & = & - \xi_{\rm L} \; \left[e^{-b/\xi_{\rm L}} \left( 1 -

1121:   e^{-a/\xi_{\rm L}} \right) \cos \alpha  + \left( 1 - e^{-b/\xi_{\rm

1122:   L}} \right) \right] \\

1123: Y_{\rm L} & = & -  \xi_{\rm L} \; e^{-b/\xi_{\rm L}} \left( 1 -

1124:   e^{-a/\xi_{\rm L}} \right) \sin \alpha

1125: \end{eqnarray}

1126: which is indicated by the filled grey circle, again resulting in an

1127: angular deviation $\varepsilon$ (in the sketch, for simplicity, the

1128: same as for angle underestimation).

1129: 

1130: The three upper curves of Fig.~\ref{fig05} show theoretical

1131: predictions for the angle error $\varepsilon$ in the two--leg

1132: experiment of M\"uller and Wehner: the error according to the formula

1133: of \cite{mul_88} calculated numerically as a function of the angle $0

1134: \le \alpha \le \pi$ between the outgoing channels in radian units

1135: (solid line); the best fit of turning underestimation (dashed line),

1136: i.e.\ $\lambda = 0.87$ in Eq.~(\ref{1a_underest}), and for the leaky

1137: integrator (dash--dotted line), $\tau_{\rm L} = 90$ s in

1138: Eqs.~(\ref{leak_int1}) and (\ref{leak_int2}). The same model errors

1139: are applied to a Z--shaped channel and shown in the three lower curves

1140: with the same coding (solid, dashed, dotted). Note that errors are

1141: smaller, but deviations cancel only partially.

1142: 

1143: Nonlinear underestimation of the angular turning rate,

1144: Eqs.~(\ref{1b_underest}) and (\ref{1c_underest}), does not show any

1145: different behaviour from (\ref{1a_underest}), because we can assume a

1146: {\em constant} turning rate $\omega_{\rm turn} = v/r_{\rm turn} = 4$

1147: s$^{-1}$ given by the ratio of the walking speed $v = 0.2$ m/s and

1148: radius of the turn in the channel $r_{\rm turn} = 0.05$ m, which is

1149: half the wall to wall distance, because in experiments deserts ants

1150: tend to keep equal distance to both channel walls (T.~Merkle, personal

1151: observation). However, nonlinear underestimation will lead to

1152: different results for arbitrarily curved paths as we will see next.

1153: 

1154: \begin{figure}[!ht]

1155: \begin{center}

1156: \includegraphics*[width = 9cm]{fig05.eps}

1157: \end{center}

1158: \caption{Error angles $\varepsilon$ as a function of the

1159:   intermediate turning angle in the two--leg experiment (three upper

1160:   curves) and a Z--shaped channel with three parts of 5~m length each

1161:   (three lower curves) in radian units. Solid lines ({\sf MW}):

1162:   Deviation following \cite{mul_88}; dashed ({\sf LU}): linear

1163:   underestimation of turning rate with $\lambda = 0.87$; dotted ({\sf

1164:   LI}): leaky integrator with $\tau_{\rm L} = 90$ s (resp.\ $\xi_{\rm

1165:   L} = 18$ m). Linear underestimation cannot account for correct path

1166:   integration under full turns ($\alpha = \pi$) but does well for $0

1167:   \le \alpha \le (5/6) \, \pi$. In the Z--shaped channel errors are

1168:   smaller than for single turn and experimentally visible (if at all)

1169:   only for angles around 150$^\circ$ ($= 5 \pi/6$) with errors {\sf

1170:   MW} and {\sf LU}.

1171: }

1172: \label{fig05}

1173: \end{figure}

1174: 

1175: \begin{figure}[!ht]

1176: \begin{center}

1177: \includegraphics*[width = 9cm]{fig06.eps}

1178: \end{center}

1179: \caption{Homing error for different path integration error types as

1180:   function of integrated curvature $\int \omega(t) \, dt = \phi_{\rm

1181:   end} - \phi_0$ (in radians, $2 \pi$ for full turn) of 1000 simulated

1182:   random outbound runs of length 20 m each. (Left columns): angular

1183:   deviation $\varepsilon$ in radians; (right columns): euclidean

1184:   distance $\Delta$ from nest in metres. (Top row, {\sf MW}): error of

1185:   \cite{mul_88}; (2nd row, {\sf LI}): leaky integrator with $\tau_{\rm

1186:   L} = 300$ s; (3rd row, {\sf LU}): linear underestimation of

1187:   $\omega(t)$ with $\lambda = 0.87$; (4th row, {\sf NLU}): nonlinear

1188:   underestimation; (bottom row, {\sf PD}): perception delay with

1189:   $\tau_{\rm del} = 0.3$ s exhibits no systematic dependence on

1190:   curvature.}

1191: \label{fig08}

1192: \end{figure}

1193: 

1194: To investigate how well the different types of systematic errors fit

1195: random outbound paths we simulated runs, as they might be performed by

1196: an untrained ant searching for food without any knowledge on food

1197: sources (see Fig.~\ref{fig06}, upper panel). In particular, we

1198: considered a fluctuating turning rate $\omega(t)$ with a persistence

1199: time $T_\omega$ as in Eq.~(\ref{e12}) of the model in

1200: Section~\ref{modelexcurs}. For 1000 such runs we extracted

1201: correlations between characteristic indicators of the path, such as

1202: its integrated curvature $\phi_{\rm end} - \phi_0 = \int \omega(t) \,

1203: dt$, and the two most direct measures for homing deviation: the {\em

1204:   angular misestimation} $\varepsilon$ between calculated and correct

1205: homeward course, and the {\em euclidean distance} $\Delta$ between

1206: supposed and real nest positions.

1207: 

1208: The results are shown in Fig.~\ref{fig08}. There is a clear

1209: correlation between curvature $\int \omega$ and the directional

1210: mismatch of the homing vector, for all error mechanisms except for

1211: perception delay (see left panels in column). In particular, all

1212: mechanisms tend to {\em overcompensate} turns effectuated during the

1213: outbound path, as there is a {\em positive correlation} between $\int

1214: \omega$ and the deviation angle. Remind that all predict

1215: overcompensation for the two--leg experiment of M\"uller and Wehner as

1216: well.

1217: 

1218: There is a striking difference between the leaky path integrator 

1219: and the approximative integration formula of M\"uller and Wehner on

1220: one side, and turning rate underestimation on the other side: The

1221: first two predict a larger euclidean distance from the nest for paths

1222: where left and right turns compensate ($\int \omega \approx 0$) and

1223: come closer to the nest when there is a substantial net turn,

1224: resulting in $\wedge$--shapes in the right hand panels of rows 1 and 2

1225: in Fig.~\ref{fig08}. On the other hand, turning rate underestimation

1226: predicts smaller distance to the nest for compensated turns, and

1227: larger euclidean mismatch for paths with higher turns, leading to

1228: $\vee$--shapes of lines 3 and 4 in the right hand panels. Experiments

1229: which cover both the {\em full return path} and the {\em systematic

1230:   search}, additionally to the initial direction analysed by

1231: \cite{mul_88} may be able to decide the type of homing error mechanism

1232: in desert ants (T.~Merkle, in preparation).

1233: 

1234: \begin{figure}[!ht]

1235: \begin{center}

1236: \includegraphics*[width = 5cm]{fig07.eps}

1237: \end{center}

1238: \caption{Homing error in euclidean distance $\Delta$ between supposed

1239:   and  real nest position as function of distance between starting

1240:   and end point of foraging path, i.e.\ distance of feeding site from

1241:   nest, $0 \le d \equiv |\bm{P}(t_1) - \bm{P}x(0)| \le 20$. Simulated

1242:   paths had arch length 20~m, so $d = 20$~m means a perfectly straight

1243:   path. (Top): The {\em leaky integrator} predicts increasing $\Delta$

1244:   with $d$, whereas  according to \cite{mul_88} $\Delta$ has a maximum

1245:   for intermediate $d$ (middle). (Bottom): {\em Angle underestimation}

1246:   leads to an opposite relation, $\Delta$ decreasing with $d$. Same

1247:   parameters as in Fig.~\ref{fig08}.}

1248: \label{fig08a}

1249: \end{figure}

1250: 

1251: Fig. \ref{fig08a} shows the homing errors produced by different error

1252: types as functions of the distances $d$ between start and end points

1253: of foraging trips that had the same overall path lengths. Thus, the

1254: values of $d$ indicate the sinuousity of the different paths: straight

1255: paths have large, winded paths small $d$. Roughly speaking the error

1256: of the {\em leaky integrator} increases with $d$ and becomes maximal

1257: for perfectly straight paths. The error postulated by \cite{mul_88}

1258: also increases with $d$ over a wide range, but {\em decreases} for

1259: very large values for almost straight paths. {\em Angle

1260:   underestimation} yields an opposite picture, the deviation decreases

1261: over the entire range of $d$, although large fluctuations may

1262: obfuscate measurements. Clearly these findings have to be further

1263: developed in comparison to real experiments, but they indicate how

1264: field work can enable an observer to differentiate between various

1265: error types.

1266: 

1267: \begin{figure}[!ht]

1268: \begin{center}

1269: \includegraphics*[width = 10cm]{fig08a.eps}

1270: \includegraphics*[width = 10cm]{fig08b.eps}

1271: \end{center}

1272: \caption{Supposed  nest locations for different error types. (Upper

1273:   panel): Two simulated random outbound paths both starting in

1274:   direction $\phi_0 = 0$ and having arch length 20 m (black and grey)

1275:   with supposed nest locations according to M\"uller and Wehner ($+$),

1276:   leaky integrator with $\tau_{\rm L} = 300$ s ($\times$),

1277:   underestimation of $\omega$ with $\lambda = 0.87$ and $\omega_{\rm

1278:   c} = 0.2$ s$^{-1}$ (linear: $\Box$, nonlin.: $\Diamond$) and

1279:   processing delay of $\omega$ with $\tau_{\rm del} = 0.3$ s

1280:   ($\triangle$). Note that ($+$) and ($\times$) are relatively

1281:   close. (Lower panel): End points of 100 simulated outbound paths

1282:   ($\circ$) with supposed nest positions after M\"uller and Wehner

1283:   ($+$), leaky integrator (grey $\times$), and

1284:   $\omega$--underestimation ($\Box$). Real nest marked by filled white

1285:   circle. Note that ($+$) and ($\times$) coincide well, in front of

1286:   real nest. ($\Box$) are grouped closely around the nest. Units are

1287:   in metres.}

1288: \label{fig06}

1289: \end{figure}

1290: A difference is also visible in the predictions for the supposed nest

1291: positions, as presented in Fig.~\ref{fig06}. In its upper panel it

1292: shows the endpoints of two random outbound runs of length 20 m (one in

1293: black, the other one in light grey), together with the respective

1294: supposed nest locations under different error mechanisms: the formula

1295: of M\"uller and Wehner (MW, marked by $+$), linear (LU, $\Box$) and

1296: nonlinear (NLU, $\Diamond$) underestimation with $\lambda = 0.87$ and

1297: $\omega_{\rm c} = 0.2$ s$^{-1}$ and turning perception delay with

1298: $\tau_{\rm del} = 0.3$ s (PD, $\triangle$). The time constant for the

1299: leaky integrator (LI, $\times$), $\tau_{\rm L} = 300$ s, was chosen

1300: such that it best fitted the results of the phenomenological error

1301: formula of \citep{mul_88}. In the lower panel of Fig.~\ref{fig06} the

1302: same is shown for 100 paths (without the paths themselves), where all

1303: runs start in the same initial direction $\phi_0 = 0$, such that the

1304: end points $(\circ)$ lie in a sickle shaped domain to the right. Again

1305: there is a striking coincidence between the leaky integrator and

1306: M\"uller's formula as opposed to the results of turning rate

1307: underestimation. Notice that only the two first error mechanisms lead

1308: to a home vector pointing to a location {\em in front of} the actual

1309: nest (see small sickle--shape domain to the right of nest position).

1310: 

1311: \section{Discussion}

1312: \label{discussion}

1313: 

1314: In this work we have presented a very simple model for path

1315: integration using egocentric cartesian coordinates. In contrast to all

1316: previous models, including the egocentric one using polar coordinates

1317: (Section~\ref{prev_egocent}), in this model the arthropod does not

1318: need to perform complicated calculations such as applying

1319: trigonometric or other non--linear functions, but rather updates two

1320: cartesian coordinate values of the relative global vector

1321: $\bm{G}=(X,Y)$ by computing a simple system of linear differential

1322: equations. Moreover, although we assume that neither the actual

1323: relative angle $\delta$ nor the distance $r$ to the nest have to be

1324: calculated or stored at any time, solely by using the internal

1325: $\bm{G}$--vector information the arthropod has the ability to orient

1326: towards the nest position at any time along its path and to hold this

1327: orientation during the home run. Keeping $Y = 0$ serves as an

1328: `internal beacon' for home orientation, where the simple

1329: counter--steering mechanism can be realized as an elementary negative

1330: feedback control of the turning rate by the internal variable $Y$,

1331: until the second internal variable $X$ reaches the desired zero

1332: value. Thus, the path integration values $(X,Y)$ do not only provide a

1333: record of the arthropod's positional movement, but can also be used as

1334: information input for orientation. Moreover, accumulated information

1335: on the whole internal $(X,Y)$-path, as depicted in Fig.~\ref{fig02},

1336: for example, may be used by the arthropod to guide its observed

1337: systematic search for the `true' nest position after failure. This

1338: could be done by estimating the probability of how far away the nest

1339: is located, depending on the probability of the accumulated path

1340: integration error. For previous experiments and theories see

1341: \cite{HOF90}.

1342: 

1343: In order to prepare the theoretical background for further

1344: investigations, we have restricted this first presentation to (i)

1345: construct our model, (ii) implement various error mechanisms and (iii)

1346: evaluate their predictions for future comparison with experiments by

1347: stochastic simulations. Figs.~\ref{fig08} and \ref{fig08a} suggest

1348: simple checks to accept or exclude one or another error mechanism and

1349: therefore can serve as a guide for future experiments and modelling.

1350: 

1351: The mechanisms suggested here could also be tested in experiments with

1352: specially designed channels for outbound runs. Different types of

1353: angle misestimation would e.g.\ lead to different systematic

1354: deviations of the home run in a channel where left turns are sharper

1355: than right turns. Differences between the leaky integrator and

1356: angle--misestimation should become apparent in a comparison between

1357: two different two--leg experiments, e.g.\ both with $\alpha = \pi/2$,

1358: but two different values $a_1 \neq a_2$ which both would be

1359: substantially larger than $b$.

1360: 

1361: Physiological realizations of the integration procedure itself and the

1362: underlying fundamental neural mechanisms are far from being

1363: clarified. Besides, a clarification of the neural processes working in

1364: the brain and the locomotory  control apparatus of desert arthropods

1365: does not seem to be within reach in the near future. We share this

1366: problem with all other existing models of path integration.

1367: 

1368: On the other hand, anatomic features of neurons, their interactions,

1369: their integration, and their cooperation within networks have been

1370: known for a long time. \cite{har_95} have developed a simple and

1371: efficient neural network for path integration in desert ants. In a

1372: particular form it even incorporates the systematic errors observed by

1373: \cite{mul_88} on a neural level. Based on their model of an

1374: incremental encoding one could easily construct a neural architecture

1375: for path integration in cartesian variables $X$ and $Y$ as in

1376: Eqs.~(\ref{e09}) and (\ref{e10}), where the required estimation of

1377: speeds could most likely be encoded by spike rates of afferent

1378: neurons.

1379: 

1380: A further, much simpler way would be to represent the internal

1381: cartesian coordinates $(X,Y)$ directly by the deviations of two

1382: non--spiking interneuron activities $N^X$ and $N^Y$ from their basal

1383: activity values $N^X_0$ and $N^Y_0$  supposed to be attained when the

1384: animal is `at home', i.e. for $(X,Y) = (0,0)$. Regulation of these

1385: interneurons as well as their mutual interactions could then be

1386: realized by suitably defined dendritic synapses of a neural net akin

1387: to the scheme presented in Fig.~\ref{fig03}. Again, the obvious

1388: simplicity of this `linear'  control network may favour the egocentric

1389: cartesian path integration model as candidate for a most elementary

1390: neural realization in the arthropod, compared to other, more

1391: complicated models. Although mathematical simplicity is not an ad--hoc

1392: argument to explain natural evolution of biological control systems,

1393: it is tempting to ``grow'', i.e.\ let develop by evolutionary

1394: algorithms, neural networks for the task of orientation and analyse

1395: their mathematical structure post--hoc, as it has been done for robot

1396: motion control \citep{pas_01}.

1397: 

1398: The ability to find successful feeding sites, also observed for

1399: beetles, for instance \citep[von][]{FRI65}, was particularly

1400: investigated in detail for desert ants

1401: \citep[][]{WEH83b,WEH87,col_99}. It is therefore a natural question to

1402: ask, whether a similar simple rule as that of keeping $Y = 0$ may help

1403: to find a previously known feeding site. Consider all trajectories in

1404: geocentric coordinates which keep the egocentric $Y \equiv 0$ constant:

1405: they are the radii around the nest position. On the way home they all

1406: {\em converge}, and if by random fluctuations the animal switches over

1407: to a trajectory in its neighbourhood it nevertheless is guided towards

1408: the nest by the beacon condition $Y = 0$. But for outbound routes they

1409: {\em diverge}, and random errors are not corrected on their own. It

1410: would even be better to follow a {\em fixed compass direction},

1411: because trajectories of the same direction are parallel to each other,

1412: and randomly accumulated errors will not be enhanced during the

1413: course. In the light of our model it seems natural to suggest that the

1414: relative position of a feeding site is internally stored as another

1415: global vector $\bm{G}_{\rm f} = (X_{\rm f},Y_{\rm f})$ which is

1416: updated simultaneously with $(X,Y)$. Depending on whether the animal

1417: steers towards the feeder or home, either $Y_{\rm f} = 0$ with $X_{\rm

1418:   f} > 0$ is the beacon condition, or $Y = 0$ with $X>0$. To our

1419: knowledge no experiments with obstacles on the way to a trained feeder

1420: have been performed corresponding to those described in \cite{weh_03a}

1421: for homing paths with obstacles (see Fig.~2B therein). If the animals

1422: are able to compensate forced deviations on the path to the feeder in

1423: the same manner as on paths leading home, this would indicate a

1424: similar internal processing for both positions.

1425: 

1426: Moreover, a better understanding of mechanisms by which global vector

1427: information is combined or substituted with local information 

1428: from landmarks requires efficient mathematical models and

1429: simulations that are able to reproduce experimental data. With our

1430: egocentric path information system for the two relative cartesian

1431: coordinates we have presented a most simple modelling tool that can

1432: help to evaluate and discriminate various hypotheses on orientation,

1433: random or systematic errors, and possible neural representations.

1434: 

1435: 

1436: \subsection*{Acknowledgements}

1437: 

1438: This work was partly supported by Research Group {\em Wissensformate}

1439: of Bonn University and Special Research Program SFB 611 of Deutsche

1440: For\-schungs\-ge\-mein\-schaft. We thank G.~Hoffmann for helpful

1441: comments on the manu\-script, the anonymous referees for many detailed

1442: suggestions to improve the manuscript, and R.~Wehner for introducing

1443: T.M.\ to {\em Cataglyphis} and including him in his research

1444: project on path integration in desert ants.

1445: 

1446: \bibliographystyle{JTB}

1447: 

1448: \bibliography{JTB}

1449: 

1450: 

1451: \end{document}

1452: 

1453: